Exploiting Argument Information to Improve Event Detection via Supervised Attention Mechanisms

Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics, pages 1789–1798 Vancouver, Canada, July 30 - August 4, 2017. c⃝2017 Association for Computational Linguistics https://doi.org/10.18653/v1/P17-1164 Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics, pages 1789–1798 Vancouver, Canada, July 30 - August 4, 2017. c⃝2017 Association for Computational Linguistics https://doi.org/10.18653/v1/P17-1164 Exploiting Argument Information to Improve Event Detection via Supervised Attention Mechanisms Shulin Liu1,2, Yubo Chen1,2, Kang Liu1 and Jun Zhao1,2 1 National Laboratory of Pattern Recognition, Institute of Automation, Chinese Academy of Sciences, Beijing, 100190, China 2 University of Chinese Academy of Sciences, Beijing, 100049, China {shulin.liu, yubo.chen, kliu, jzhao}@nlpr.ia.ac.cn Abstract This paper tackles the task of event detection (ED), which involves identi- fying and categorizing events. We ar- gue that arguments provide significant clues to this task, but they are either completely ignored or exploited in an indirect manner in existing detection approaches. In this work, we propose to exploit argument information explicitly for ED via supervised attention mech- anisms. In specific, we systematically investigate the proposed model under the supervision of different attention strategies. Experimental results show that our approach advances state-of- the-arts and achieves the best F1 score on ACE 2005 dataset. 1 Introduction In the ACE (Automatic Context Extraction) event extraction program, an event is repre- sented as a structure comprising an event trigger and a set of arguments. This work tackles event detection (ED) task, which is a crucial part of event extraction (EE) and fo- cuses on identifying event triggers and cate- gorizing them. For instance, in the sentence “He died in the hospital”, an ED system is expected to detect a Die event along with the trigger word “died”. Besides, the task of EE also includes event argument extraction (AE), which involves event argument identi- fication and role classification. In the above sentence, the arguments of the event include “He”(Role = Person) and “hospital”(Role = Place). However, this paper does not focus on AE and only tackles the former task. According to the above definitions, event ar- guments seem to be not essentially necessary to ED. However, we argue that they are capa- ble of providing significant clues for identifying and categorizing events. They are especially useful for ambiguous trigger words. For exam- ple, consider a sentence in ACE 2005 dataset: Mohamad fired Anwar, his for- mer protege, in 1998. In this sentence, “fired” is the trigger word and the other bold words are event arguments. The correct type of the event triggered by “fired” in this case is End-Position. How- ever, it might be easily misidentified as At- tack because “fired” is a multivocal word. In this case, if we consider the phrase “for- mer protege”, which serves as an argumen- t (Role = Position) of the target event, we would have more confidence in predicting it as an End-Position event. Unfortunately, most existing methods per- formed event detection individually, where the annotated arguments in training set are totally ignored (Ji and Grishman, 2008; Gupta and …

Content truncated to 500 words