A Unified Model of Phrasal and Sentential Evidence for Information Extraction

Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing, pages 151–160, Singapore, 6-7 August 2009. c⃝2009 ACL and AFNLP A Unified Model of Phrasal and Sentential Evidence for Information Extraction Siddharth Patwardhan and Ellen Riloff School of Computing University of Utah Salt Lake City, UT 84112 {sidd,riloff}@cs.utah.edu Abstract Information Extraction (IE) systems that extract role fillers for events typically look at the local context surrounding a phrase when deciding whether to extract it. Of- ten, however, role fillers occur in clauses that are not directly linked to an event word. We present a new model for event extraction that jointly considers both the local context around a phrase along with the wider sentential context in a proba- bilistic framework. Our approach uses a sentential event recognizer and a plausible role-filler recognizer that is conditioned on event sentences. We evaluate our system on two IE data sets and show that our model performs well in comparison to ex- isting IE systems that rely on local phrasal context. 1 Introduction Information Extraction (IE) systems typically use extraction patterns (e.g., Soderland et al. (1995), Riloff (1996), Yangarber et al. (2000), Califf and Mooney (2003)) or classifiers (e.g., Freitag (1998), Freitag and McCallum (2000), Chieu et al. (2003), Bunescu and Mooney (2004)) to extract role fillers for events. Most IE systems consider only the immediate context surrounding a phrase when deciding whether to extract it. For tasks such as named entity recognition, immediate context is usually sufficient. But for more complex tasks, such as event extraction, a larger field of view is often needed to understand how facts tie together. Most IE systems are designed to identify role fillers that appear as arguments to event verbs or nouns, either explicitly via syntactic relations or implicitly via proximity (e.g., John murdered Tom or the murder of Tom by John). But many facts are presented in clauses that do not contain event words, requiring discourse relations or deep structural analysis to associate the facts with event roles. For example, consider the sentences below: Seven people have died . . . and 30 were injured in India after terror- ists launched an attack on the Taj Hotel. . . . in Mexico City and its surrounding sub- urbs in a Swine Flu outbreak. . . . after a tractor-trailer collided with a bus in Arkansas. Two bridges were destroyed . . . in Baghdad last night in a resurgence of bomb attacks in the capital city. . . . and $50 million in damage was caused by a hurricane that hit Miami on Friday. . . . to make way for modern, safer bridges that will be constructed early next year. These examples illustrate a common phenomenon in text where information is not explicitly stated as filling an event role, but readers have no trou- ble making this inference. The role fillers above (seven people, two bridges) occur as arguments to verbs that reveal state information (death, destruc- tion) but are not event-specific (i.e., …

Content truncated to 500 words