Joint Event Extraction via Structured Prediction with Global Features

Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics, pages 73–82, Sofia, Bulgaria, August 4-9 2013. c⃝2013 Association for Computational Linguistics Joint Event Extraction via Structured Prediction with Global Features Qi Li Heng Ji Liang Huang Departments of Computer Science and Linguistics The Graduate Center and Queens College City University of New York New York, NY 10016, USA {liqiearth, hengjicuny, liang.huang.sh}@gmail.com Abstract Traditional approaches to the task of ACE event extraction usually rely on sequential pipelines with multiple stages, which suf- fer from error propagation since event trig- gers and arguments are predicted in isola- tion by independent local classifiers. By contrast, we propose a joint framework based on structured prediction which ex- tracts triggers and arguments together so that the local predictions can be mutu- ally improved. In addition, we propose to incorporate global features which ex- plicitly capture the dependencies of multi- ple triggers and arguments. Experimental results show that our joint approach with local features outperforms the pipelined baseline, and adding global features fur- ther improves the performance signifi- cantly. Our approach advances state-of- the-art sentence-level event extraction, and even outperforms previous argument la- beling methods which use external knowl- edge from other sentences and documents. 1 Introduction Event extraction is an important and challeng- ing task in Information Extraction (IE), which aims to discover event triggers with specific types and their arguments. Most state-of-the-art ap- proaches (Ji and Grishman, 2008; Liao and Gr- ishman, 2010; Hong et al., 2011) use sequential pipelines as building blocks, which break down the whole task into separate subtasks, such as trigger identification/classification and argument identification/classification. As a common draw- back of the staged architecture, errors in upstream component are often compounded and propagated to the downstream classifiers. The downstream components, however, cannot impact earlier deci- sions. For example, consider the following sen- tences with an ambiguous word “fired”: (1) In Baghdad, a cameraman died when an American tank fired on the Palestine Hotel. (2) He has fired his air defense chief. In sentence (1), “fired” is a trigger of type Attack. Because of the ambiguity, a local classifier may miss it or mislabel it as a trigger of End-Position. However, knowing that “tank” is very likely to be an Instrument argument of Attack events, the cor- rect event subtype assignment of “fired” is obvi- ously Attack. Likewise, in sentence (2), “air de- fense chief” is a job title, hence the argument clas- sifier is likely to label it as an Entity argument for End-Position trigger. In addition, the local classifiers are incapable of capturing inter-dependencies among multiple event triggers and arguments. Consider sentence (1) again. Figure 1 depicts the corresponding event triggers and arguments. The dependency be- tween “fired” and “died” cannot be captured by the local classifiers, which may fail to attach “camera- man” to “fired” as a Target argument. By using global features, we can propagate the Victim ar- gument of the Die event to the Target argument of the Attack event. As another example, …

Content truncated to 500 words