Exploiting the Entity Type Sequence to Benefit Event Detection

Proceedings of the 23rd Conference on Computational Natural Language Learning, pages 613–623 Hong Kong, China, November 3-4, 2019. c⃝2019 Association for Computational Linguistics 613 Exploiting the Entity Type Sequence to Benefit Event Detection Yuze Ji,1,2 Youfang Lin,1,2 Jianwei Gao,1,2 Huaiyu Wan1,2∗ 1School of Computer and Information Technology, Beijing Jiaotong University, Beijing, China 2Beijing Key Laboratory of Traffic Data Analysis and Mining, Beijing, China {yuzeji, yflin, gaojianwei, hywan}@bjtu.edu.cn Abstract Event Detection (ED) is one of the most im- portant tasks in the field of information extrac- tion. The goal of ED is to find triggers in sen- tences and classify them into different event types. In previous works, the information of entity types are commonly utilized to benefit event detection. However, the sequential fea- tures of entity types have not been well utilized yet in the existing ED methods. In this paper, we propose a novel ED approach which learns sequential features from word sequences and entity type sequences separately, and com- bines these two types of sequential features with the help of a trigger-entity interaction learning module. The experimental results demonstrate that our proposed approach out- performs the state-of-the-art methods. 1 Introduction Event Extraction (EE) is one of the essential tasks of Information Extraction, which aims to extract structured events from unstructured texts. Accord- ing to ACE (Automatic Context Extraction) event annotation guideline1, an event is represented by an event trigger, which is often a single verb or noun, and a set of event arguments, the partici- pants of the event. Event Detection (ED), as a crucial step in EE task, focuses on finding event trigger words and classifying them into different event types. As pointed out in (Liu et al., 2019; Ritter et al., 2012), the ambiguity in natural lan- guages makes ED a challenging task. On the one hand, various expressions can be used to represent the same event type; on the other hand, the same event triggers, when placed in different context, can be categorized in totally different event types. To illustrate the second phenomenon, we present ∗Corresponding author: hywan@bjtu.edu.cn 1https://www.ldc.upenn.edu/sites/www.ldc.upenn.edu/ files/english-events-guidelines-v5.4.3.pdf two examples from the widely used ACE 20052 dataset: 1) A Russian Soyuz capsule (VEH) dropped (Transport) the astronauts (PER) off this morning. 2) U.S. planes (VEH) dropped (Attack) a bomb (WEA) near northern Iraq. Notice that the same annotated event trigger dropped has different meanings in the two sentences above and thus evokes entirely different event types. In the first sentence, dropped evokes a Transport event, but in the second sentence, dropped represents an Attack event. According to Liu et al. (2018a), in the ACE 2005 dataset, 57% of the event triggers are ambiguous. How to alleviate the ambiguity of event triggers has become a crucial problem in the ED task. In several previous works, researchers have proved that entity mentions could play a positive role on alleviating the ambiguity of event triggers (Hong et al., 2011; Li et al., 2013; Feng et al., 2016; Liu et al., 2017, 2019, 2018a). An entity mention, …

Content truncated to 500 words