Event Argument Identification on Dependency Graphs with Bidirectional LSTMs

Proceedings of the The 8th International Joint Conference on Natural Language Processing, pages 822–831, Taipei, Taiwan, November 27 – December 1, 2017 c⃝2017 AFNLP Event Argument Identification on Dependency Graphs with Bidirectional LSTMs 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 In this paper we investigate the perfor- mance of event argument identification. We show that the performance is tied to syntactic complexity. Based on this find- ing, we propose a novel and effective sys- tem for event argument identification. Re- current Neural Networks learn to produce meaningful representations of long and short dependency paths. Convolutional Neural Networks learn to decompose the lexical context of argument candidates. They are combined into a simple system which outperforms a feature-based, state- of-the-art event argument identifier with- out any manual feature engineering. 1 Introduction Event extraction is a difficult information extrac- tion task. The 2005 Automatic Content Extrac- tion evaluation (ACE 2005) defines three challeng- ing sub-tasks: Entity mention detection, the task of finding mentions of predefined entity types like persons and organizations; event trigger detection, the task of finding words, mostly verbs or nomi- nalizations, indicating an event from a set of pre- defined event types; and event argument identifi- cation, the identification of entity mentions1 play- ing a role in the events, as well as the identification of the roles they play. When we look at the evaluations in three of the most influential recent event extraction papers (Li et al., 2013, 2014; Chen et al., 2015; Nguyen et al., 2016) we note that argument identification performance is low, ranging from 52.7 to 55.4 F1. There are multiple reasons for the low perfor- mance. First, argument identification suffers from 1In this work we make no distinction between ‘entity’, ‘time’, and ‘place’ for the sake of simplicity. error propagation. Missed or spurious event trig- gers or entity mentions may lead to missed or spu- rious event arguments. Second, event structure is complex. Multiple entities can play the same role in the same event. Additionally, one entity can play different roles across events (and thus cause multiple event arguments). Consider the following example. A Palestinian boy as well as his brother and a sister were wounded late Wednes- day by Israeli gunfire. Here, the three entity mentions (in bold) are all Victims of the INJURE event triggered by ‘wounded’ as well as Targets of the ATTACK event triggered by ‘gunfire’. Such structures can become even more complex when more events and more entities are involved. The third reason for low argument identification performance is syntactic complexity. Many argu- ments are syntactically far away from their trig- gers, making it hard to construct meaningful syn- tactic features. Section 3 shows that performance is tied to the length of the shortest dependency path connecting trigger and argument. Error sources one and two were already targeted by systems which jointly infer triggers and their arguments. To the best of our knowledge, no pre- vious work identified …

Content truncated to 500 words