Semantic Structure Enhanced Event Causality Identification

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 10901–10913 July 9-14, 2023 ©2023 Association for Computational Linguistics Semantic Structure Enhanced Event Causality Identification Zhilei Hu, Zixuan Li∗, Xiaolong Jin∗, Long Bai, Saiping Guan, Jiafeng Guo and Xueqi Cheng School of Computer Science and Technology, University of Chinese Academy of Sciences; CAS Key Laboratory of Network Data Science and Technology, Institute of Computing Technology, Chinese Academy of Sciences. {huzhilei19b, lizixuan, jinxiaolong, bailong18b, guansaiping}@ict.ac.cn {guojiafeng, cxq}@ict.ac.cn Abstract Event Causality Identification (ECI) aims to identify causal relations between events in un- structured texts. This is a very challenging task, because causal relations are usually expressed by implicit associations between events. Exist- ing methods usually capture such associations by directly modeling the texts with pre-trained language models, which underestimate two kinds of semantic structures vital to the ECI task, namely, event-centric structure and event- associated structure. The former includes im- portant semantic elements related to the events to describe them more precisely, while the lat- ter contains semantic paths between two events to provide possible supports for ECI. In this paper, we study the implicit associations be- tween events by modeling the above explicit semantic structures, and propose a Semantic Structure Integration model (SemSIn). It uti- lizes a GNN-based event aggregator to integrate the event-centric structure information, and em- ploys an LSTM-based path aggregator to cap- ture the event-associated structure information between two events. Experimental results on three widely used datasets show that SemSIn achieves significant improvements over base- line methods. 1 Introduction Event Causality Identification (ECI) is an impor- tant task in natural language processing that seeks to predict causal relations between events in texts. As shown in the top of Figure 1, given the unstruc- tured text and event pair (shot, protect), an ECI model needs to identify that there exists a causal re- lation between two events, i.e., protect cause −→ shot. ECI is an important way to construct wide causal connections among events, which supports a va- riety of practical applications, such as event pre- diction (Hashimoto, 2019), reading comprehen- sion (Berant et al., 2014), and question answer- ing (Oh et al., 2013, 2017). ∗Corresponding authors. protect-01 shoot-02 attempt-01 person :ARG0 shoot-02 protect-01 :time -1 Event-associated Structure :ARG1 -1 :ARG1 -1 protect-01 person gunfire person attempt-01 :ARG1 Event-centric Structure :ARG2 -1 :ARG1 -1 :ARG0 -1 shoot-02 protect-01 attempt-01 person gunfire person student name cause-01 die-01 instant “Horton” :time :ARG1 :ARG0 :ARG1 :ARG2 :ARG0 :ARG0 :name :op1 :ARG1 :ARG0 :ARG1 :manner :ARG1 Semantic Graph Unstructured Text: Horton was shot and died instantly when she attempted to protect students from gunfire. ECI input: (text, shot, protect) ECI output: causality Figure 1: An example of the ECI task, as well as the se- mantic graph and semantic structures corresponding to the unstructured text. The orange nodes denote events. ECI is challenging because most causal relations are expressed by texts implicitly, which requires the model to understand the associations between two events adequately. Existing methods …

Content truncated to 500 words