Bootstrapped Training of Event Extraction Classifiers

Proceedings of the 13th Conference of the European Chapter of the Association for Computational Linguistics, pages 286–295, Avignon, France, April 23 - 27 2012. c⃝2012 Association for Computational Linguistics Bootstrapped Training of Event Extraction Classifiers Ruihong Huang and Ellen Riloff School of Computing University of Utah Salt Lake City, UT 84112 {huangrh,riloff}@cs.utah.edu Abstract Most event extraction systems are trained with supervised learning and rely on a col- lection of annotated documents. Due to the domain-specificity of this task, event extraction systems must be retrained with new annotated data for each domain. In this paper, we propose a bootstrapping so- lution for event role filler extraction that re- quires minimal human supervision. We aim to rapidly train a state-of-the-art event ex- traction system using a small set of “seed nouns” for each event role, a collection of relevant (in-domain) and irrelevant (out- of-domain) texts, and a semantic dictio- nary. The experimental results show that the bootstrapped system outperforms previ- ous weakly supervised event extraction sys- tems on the MUC-4 data set, and achieves performance levels comparable to super- vised training with 700 manually annotated documents. 1 Introduction Event extraction systems process stories about domain-relevant events and identify the role fillers of each event. A key challenge for event extrac- tion is that recognizing role fillers is inherently contextual. For example, a PERSON can be a perpetrator or a victim in different contexts (e.g., “John Smith assassinated the mayor” vs. “John Smith was assassinated”). Similarly, any COM- PANY can be an acquirer or an acquiree depending on the context. Many supervised learning techniques have been used to create event extraction systems us- ing gold standard “answer key” event templates for training (e.g., (Freitag, 1998a; Chieu and Ng, 2002; Maslennikov and Chua, 2007)). How- ever, manually generating answer keys for event extraction is time-consuming and tedious. And more importantly, event extraction annotations are highly domain-specific, so new annotations must be obtained for each domain. The goal of our research is to use bootstrap- ping techniques to automatically train a state-of- the-art event extraction system without human- generated answer key templates. The focus of our work is the TIER event extraction model, which is a multi-layered architecture for event extrac- tion (Huang and Riloff, 2011). TIER’s innova- tion over previous techniques is the use of four different classifiers that analyze a document at in- creasing levels of granularity. TIER progressively zooms in on event information using a pipeline of classifiers that perform document-level classi- fication, sentence classification, and noun phrase classification. TIER outperformed previous event extraction systems on the MUC-4 data set, but re- lied heavily on a large collection of 1,300 docu- ments coupled with answer key templates to train its four classifiers. In this paper, we present a bootstrapping solu- tion that exploits a large unannotated corpus for training by using role-identifying nouns (Phillips and Riloff, 2007) as seed terms. Phillips and Riloff observed that some nouns, by definition, refer to entities or objects that play a specific role in an event. …

Content truncated to 500 words