Effective Information Extraction with Semantic Affinity Patterns and Relevant Regions

Proceedings of the 2007 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning, pp. 717–727, Prague, June 2007. c⃝2007 Association for Computational Linguistics Effective Information Extraction with Semantic Affinity Patterns and Relevant Regions Siddharth Patwardhan and Ellen Riloff School of Computing University of Utah Salt Lake City, UT 84112 {sidd,riloff}@cs.utah.edu Abstract We present an information extraction system that decouples the tasks of finding relevant regions of text and applying extraction pat- terns. We create a self-trained relevant sen- tence classifier to identify relevant regions, and use a semantic affinity measure to au- tomatically learn domain-relevant extraction patterns. We then distinguish primary pat- terns from secondary patterns and apply the patterns selectively in the relevant regions. The resulting IE system achieves good per- formance on the MUC-4 terrorism corpus and ProMed disease outbreak stories. This approach requires only a few seed extraction patterns and a collection of relevant and ir- relevant documents for training. 1 Introduction Many information extraction (IE) systems rely on rules or patterns to extract words and phrases based on their surrounding context (e.g., (Soderland et al., 1995; Riloff, 1996; Califf and Mooney, 1999; Yan- garber et al., 2000)). For example, a pattern like “<subject> was assassinated” can reliably identify a victim of a murder event. Classification-based IE systems (e.g., (Freitag, 1998; Freitag and McCal- lum, 2000; Chieu et al., 2003)) also generally de- cide whether to extract words based on properties of the words themselves as well as properties associ- ated with their surrounding context. In this research, we propose an alternative ap- proach to IE that decouples the tasks of finding a rel- evant region of text and finding a desired extraction. In a typical pattern-based IE system, the extraction patterns perform two tasks: (a) they recognize that a relevant incident has occurred, and (b) they iden- tify and extract some information about that event. In contrast, our approach first identifies relevant re- gions of a document that describes relevant events, and then applies extraction patterns only in these rel- evant regions. This decoupled approach to IE has several po- tential advantages. First, even seemingly good pat- terns can produce false hits due to metaphor and id- iomatic expressions. However, by restricting their use to relevant regions of text, we could avoid such false positives. For example, “John Kerry attacked George Bush” is a metaphorical description of a ver- bal tirade, but could be easily mistaken for a physi- cal attack. Second, IE systems are prone to errors of omission when relevant information is not explicitly linked to an event. For instance, a phrase like “the gun was found...” does not directly state that the the gun was used in a terrorist attack. But if the gun is mentioned in a region that clearly describes a terror- ist attack, then it can be reasonably inferred to have been used in the attack. Third, if the IE patterns are restricted to areas of text that are known to be rel- evant, then …

Content truncated to 500 words