Incremental Global Event Extraction

Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 2279–2289, Osaka, Japan, December 11-17 2016. Incremental Global Event Extraction Alex Judea and Michael Strube Heidelberg Institute for Theoretical Studies gGmbH Schloss-Wolfsbrunnenweg 35 69118 Heidelberg, Germany (alex.judea|michael.strube)@h-its.org Abstract Event extraction is a difficult information extraction task. Li et al. (2014) explore the benefits of modeling event extraction and two related tasks, entity mention and relation extraction, jointly. This joint system achieves state-of-the-art performance in all tasks. However, as a system oper- ating only at the sentence level, it misses valuable information from other parts of the document. In this paper, we present an incremental approach to make the global context of the entire docu- ment available to the intra-sentential, state-of-the-art event extractor. We show that our method robustly increases performance on two datasets, namely ACE 2005 and TAC 2015. 1 Introduction But the strikes prove controversial. In many cases, it is not sufficient to look at one sentence when extracting events. In our example, strikes has no potential event arguments, and the context is not sufficient to disambiguate it correctly: Is it the trigger of an ATTACK event because it actually means bomb strikes? Or is it not a trigger at all because it refers to the industrial action? Soon after dawn on this fourth day, confirmation of the ship’s first strike arrived . . . This is the first of an unknown number of strikes we’ll conduct during our watch in “operation enduring freedom” . . . But the strikes prove controversial. Given the other sentences it is easier to infer that strikes is indeed the trigger of an ATTACK event. In this paper, we present a system that makes the global context of a document available to a state-of-the-art event extractor. Looking at a broader context also benefits argument detection. For example, within one document entities play coherent roles in different events. Consider the following text: Sam Waksal was sentenced to seven years and three months in federal prison . . . He’s being released on his own cog in a sans before he’s to report to jail. Looking only at individual sentences, it is hard to predict that report triggers an ARREST-JAIL event, which in turn makes it hard to predict that he is the Person of this event. If the system looks at the entire document and knows that he and Sam Waksal are coreferent, it can better infer that the Defendant of a SENTENCE event can be the Person of an ARREST-JAIL event, which in turn makes it easier to infer that he is this person. In this paper, we present a method to incorporate the global, document-wide context into the decision process of a system that predicts entity mentions, events, and relations jointly. We use features that are based on the ‘one sense per discourse’ assumption (Gale et al., 1992), a concept widely used in Word Sense Disambiguation (e.g., Navigli and Lapata (2007)), and on the coherence of roles an …

Content truncated to 500 words