Event Detection with Dual Relational Graph Attention Networks

Proceedings of the 29th International Conference on Computational Linguistics, pages 1979–1989 October 12–17, 2022. 1979 Event Detection with Dual Relational Graph Attention Networks Jiaxin Mi 1,2,3, Po Hu 1,2,3∗, Peng Li 4 1Hubei Provincial Key Laboratory of Artificial Intelligence and Smart Learning, Central China Normal University, Wuhan, Hubei, China 2School of Computer Science, Central China Normal University, Wuhan, Hubei, China 3National Language Resources Monitoring & Research Center for Network Media, Central China Normal University, Wuhan, Hubei, China 4Institute for AI Industry Research (AIR), Tsinghua University, China {mjx, phu}@mail.ccnu.edu.cn, lipeng@air.tsinghua.edu.cn Abstract Event detection, which aims to identify in- stances of specific event types from pieces of text, is a fundamental task in information ex- traction. Most existing approaches leverage syntactic knowledge with a set of syntactic re- lations to enhance event detection. However, a side effect of these syntactic-based approaches is that they may confuse different syntactic rela- tions and tend to introduce redundant or noisy information, which may lead to performance degradation. To this end, we propose a sim- ple yet effective model named DualGAT (Dual Relational Graph Attention Networks), which exploits the complementary nature of syntactic and semantic relations to alleviate the problem. Specifically, we first construct a dual relational graph that both aggregates syntactic and se- mantic relations to the key nodes in the graph, so that event-relevant information can be com- prehensively captured from multiple perspec- tives (i.e., syntactic and semantic views). We then adopt augmented relational graph atten- tion networks to encode the graph and optimize its attention weights by introducing contextual information, which further improves the per- formance of event detection. Extensive exper- iments conducted on the standard ACE2005 benchmark dataset indicate that our method significantly outperforms the state-of-the-art methods and verifies the superiority of Dual- GAT over existing syntactic-based methods. 1 Introduction Event Detection (ED) aims to identify event trig- gers from a given text and classify them into pre- defined event types (Chen et al., 2015), playing an important role in information extraction. As shown in Figure 1, an event detection method is expected to identify the event trigger “thrust” from the example sentence and classify it into the event type Transport. Event detection can be used to facilitate various natural language processing * Corresponding author (NLP) applications, such as adverse drug event dis- covery (Wei et al., 2020; Liu et al., 2018a), court decision event identification (Filtz et al., 2020), fi- nancial event extraction (Zheng et al., 2019; Liang et al., 2020) and so on. Various event detection methods have been ex- plored, including traditional feature-based (Hong et al., 2011; Li et al., 2013) and deep learning meth- ods (Chen et al., 2015; Nguyen et al., 2016). Due to the powerful feature representation extraction capa- bility of deep neural networks, deep learning meth- ods have outperformed traditional feature-based methods in most cases and have become popular. Most deep learning methods have recently paid more attention to exploiting syntactic relations in event detection. These methods adopt Graph Neu- ral Networks (GNNs) such as …

Content truncated to 500 words