Revisiting Event Argument Extraction: Can EAE Models Learn Better When Being Aware of Event Co-occurrences?

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 12542–12556 July 9-14, 2023 ©2023 Association for Computational Linguistics Revisiting Event Argument Extraction: Can EAE Models Learn Better When Being Aware of Event Co-occurrences? Yuxin He1 and Jingyue Hu1 and Buzhou Tang1,2,† 1Department of Computer Science, Harbin Institute of Technology, Shenzhen, China 2Peng Cheng Laboratory, Shenzhen, China 21S051047@stu.hit.edu.cn tangbuzhou@gmail.com Abstract Event co-occurrences have been proved effec- tive for event extraction (EE) in previous stud- ies, but have not been considered for event argument extraction (EAE) recently. In this paper, we try to fill this gap between EE re- search and EAE research, by highlighting the question that “Can EAE models learn better when being aware of event co-occurrences?”. To answer this question, we reformulate EAE as a problem of table generation and extend a SOTA prompt-based EAE model into a non- autoregressive generation framework, called TabEAE, which is able to extract the argu- ments of multiple events in parallel. Under this framework, we experiment with 3 differ- ent training-inference schemes on 4 datasets (ACE05, RAMS, WikiEvents and MLEE) and discover that via training the model to ex- tract all events in parallel, it can better dis- tinguish the semantic boundary of each event and its ability to extract single event gets substantially improved. Experimental results show that our method achieves new state-of- the-art performance on the 4 datasets. Our code is avilable at https://github.com/ Stardust-hyx/TabEAE. 1 Introduction Event argument extraction (EAE) is an essential subtask of event extraction (EE). Given an input text and trigger(s) of target event(s), the EAE task aims to extract all argument(s) of each target event. Recently, substantial progress has been reported on EAE, thanks to the success of pre-trained language models (PLMs). Previous studies on EE commonly take event co- occurrences into account. However, recent works on EAE (Ebner et al., 2020; Zhang et al., 2020; Xu et al., 2022; Du and Cardie, 2020; Wei et al., 2021; Liu et al., 2021; Li et al., 2021; Du et al., †Corresponding Author. Figure 1: An illustration of EE and EAE. The triggers are in red and the arguments are underlined. EE models aim at extracting all events concurrently, whereas main- stream EAE models are trained to extract the arguments for one event trigger at a time. 2021; Lu et al., 2021; Ma et al., 2022) only con- sider one event at a time and ignore event co- occurrences (as illustrated in Figure 1). In fact, event co-occurrences always exisit in text and they are useful in revealing event correlation and con- trasting the semantic structures of different events. For the instance in Figure 1, there exist two events in the same context. The two events are triggered by “leaving”, “become” respectively, and share the same subject “Davies”. It is clear that there ex- ists a strong causal correlation between the two events. However, mainstream works on EAE split the instance into two samples, which conceals this correlation. In this …

Content truncated to 500 words