Saliency as Evidence: Event Detection with Trigger Saliency Attribution

Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 4573 - 4585 May 22-27, 2022 c⃝2022 Association for Computational Linguistics Saliency as Evidence: Event Detection with Trigger Saliency Attribution Jian Liu, Yufeng Chen, Jinan Xu Beijing Jiaotong University, School of Computer and Information Technology, China {jianliu, chenyf, jaxu}@bjtu.edu.cn Abstract Event detection (ED) is a critical subtask of event extraction that seeks to identify event triggers of certain types in texts. Despite sig- nificant advances in ED, existing methods typ- ically follow a “one model fits all types” ap- proach, which sees no differences between event types and often results in a quite skewed performance. Finding the causes of skewed performance is crucial for the robustness of an ED model, but to date there has been lit- tle exploration of this problem. This research examines the issue in depth and presents a new concept termed trigger salience attribu- tion, which can explicitly quantify the un- derlying patterns of events. On this founda- tion, we develop a new training mechanism for ED, which can distinguish between trigger- dependent and context-dependent types and achieve promising performance on two bench- marks. Finally, by highlighting many dis- tinct characteristics of trigger-dependent and context-dependent types, our work may pro- mote more research into this problem. 1 Introduction Event detection (ED) is the first and a crucial step of event extraction, which aims to identify events of certain types in plain texts (Ahn, 2006; Nguyen and Grishman, 2015; Mitamura et al., 2017). Previous methods to ED typically adopt a “one model fits all types” approach, seeing no difference between event types and using a single model to address them all (Ji and Grishman, 2008; Li et al., 2013; Chen et al., 2015; Lin et al., 2020). However, such approaches produce quite skewed performance on different types. Tasking the ACE benchmark as an example, we note the state-of-the-art ED model (Wadden et al., 2019) can strike 90% in F1 for the type DIVORCE, yet only 50% for the type START- POSITION, and it is more surprising that the training set of DIVORCE is eight times smaller than that of START-POSITION. Finding the causes underlying S1: The couple divorced four years later. S2: He became the first US minister to England. [Divorce] [Start-Position] Figure 1: Two typical event instances of DIVORCE and START-POSITION (taken from the ACE 2005 bench- mark), where the trigger words are colored. the skewed performance is crucial to the robustness of an ED model; however, this problem is still understudied in current research. In this study we take a fresh look at above prob- lem and for the first time attribute the skewed per- formance to the contextual patterns of events. Let consider the two typical instances of DIVORCE and START-POSITION shown in Figure 1. Intu- itively, they demonstrate distinct patterns: the DI- VORCE event is more trigger-dependent, and the trigger word (i.e., “divorced”) is very indicative of the event’s occurrence; by contrast, the START- POSITION event …

Content truncated to 500 words