Learning Domain-Specific Information Extraction Patterns from the Web

Proceedings of the Workshop on Information Extraction Beyond The Document, pages 66–73, Sydney, July 2006. c⃝2006 Association for Computational Linguistics Learning Domain-Specific Information Extraction Patterns from the Web Siddharth Patwardhan and Ellen Riloff School of Computing University of Utah Salt Lake City, UT 84112 {sidd,riloff}@cs.utah.edu Abstract Many information extraction (IE) systems rely on manually annotated training data to learn patterns or rules for extracting in- formation about events. Manually anno- tating data is expensive, however, and a new data set must be annotated for each domain. So most IE training sets are rel- atively small. Consequently, IE patterns learned from annotated training sets of- ten have limited coverage. In this paper, we explore the idea of using the Web to automatically identify domain-specific IE patterns that were not seen in the training data. We use IE patterns learned from the MUC-4 training set as anchors to identify domain-specific web pages and then learn new IE patterns from them. We compute the semantic affinity of each new pattern to automatically infer the type of informa- tion that it will extract. Experiments on the MUC-4 test set show that these new IE patterns improved recall with only a small precision loss. 1 Introduction Information Extraction (IE) is the task of identi- fying event descriptions in natural language text and extracting information related to those events. Many IE systems use extraction patterns or rules to identify the relevant information (Soderland et al., 1995; Riloff, 1996; Califf and Mooney, 1999; Soderland, 1999; Yangarber et al., 2000). Most of these systems use annotated training data to learn pattern matching rules based on lexical, syntactic, and/or semantic information. The learned patterns are then used to locate relevant information in new texts. IE systems typically focus on information about events that are relevant to a specific domain, such as terrorism (Sundheim, 1992; Soderland et al., 1995; Riloff, 1996; Chieu et al., 2003), man- agement succession (Sundheim, 1995; Yangarber et al., 2000), or job announcements (Califf and Mooney, 1999; Freitag and McCallum, 2000). Supervised learning systems for IE depend on domain-specific training data, which consists of texts associated with the domain that have been manually annotated with event information. The need for domain-specific training data has several disadvantages. Because of the manual la- bor involved in annotating a corpus, and because a new corpus must be annotated for each domain, most annotated IE corpora are relatively small. Language is so expressive that it is practically impossible for the patterns learned from a rela- tively small training set to cover all the different ways of describing events. Consequently, the IE patterns learned from manually annotated train- ing sets typically represent only a subset of the IE patterns that could be useful for the task. Many recent approaches in natural language processing (Yarowsky, 1995; Collins and Singer, 1999; Riloff and Jones, 1999; Nigam et al., 2000; Wiebe and Riloff, 2005) have recognized the need to use unannotated data to improve performance. While the Web provides a vast repository of unannotated …

Content truncated to 500 words