Document-Level Event Argument Extraction With a Chain Reasoning Paradigm

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 9570–9583 July 9-14, 2023 ©2023 Association for Computational Linguistics Document-Level Event Argument Extraction With A Chain Reasoning Paradigm Jian Liu1∗, Chen Liang1∗, Jinan Xu1, Haoyan Liu2 and Zhe Zhao2 1 Beijing Jiaotong University 2 Tencent AI Lab {jianliu, 21120367, jaxu}@bjtu.edu.cn {haoyanliu, nlpzhezhao}@tencent.com Abstract Document-level event argument extraction aims to identify event arguments beyond sen- tence level, where a significant challenge is to model long-range dependencies. Focusing on this challenge, we present a new chain rea- soning paradigm for the task, which can gen- erate decomposable first-order logic rules for reasoning. This paradigm naturally captures long-range interdependence due to the chains’ compositional nature, which also improves in- terpretability by explicitly modeling the rea- soning process. We introduce T-norm fuzzy logic for optimization, which permits end-to- end learning and shows promise for integrating the expressiveness of logical reasoning with the generalization of neural networks. In ex- periments, we show that our approach outper- forms previous methods by a significant mar- gin on two standard benchmarks (over 6 points in F1). Moreover, it is data-efficient in low- resource scenarios and robust enough to de- fend against adversarial attacks. 1 Introduction Identifying event arguments (i.e., participants of an event) is a crucial task for document-level event understanding (Ebner et al., 2020; Li et al., 2021). In this task, the major challenge is to model long- range dependencies between event triggers and ar- guments, as an event expression can span multiple sentences (Ebner et al., 2020; Liu et al., 2021; Li et al., 2021). Consider the event expressed by a trigger detonated (type=Attack) in Figure 1. To locate its argument Tartus (semantic role=Place), a model should capture a large context window of three sentences and 178 words to support the reasoning process. Currently, it still remains an open problem for effectively capturing such dependencies (Liu et al., 2021, 2022c). Prior research has proposed to model ∗Equal contribution. The second explosion was a sucicde bomber who detonated his belt as people rushed … … 178 words are omitted … … The Arzunah Bridge on a double explosion at the entrance to the city of Tartus, at a … The second explosion was a sucicde bomber who detonated his belt as people rushed … … 178 words are omitted… … The Arzunah Bridge on a double explosion at the entrance to the city of Tartus, at a … Place r1= Target r2= LocatedIn Place(T=detonated, ?) Place(T, ?) r1(T, Ar. B.) ∧ r2(Ar. B., Tartus) Figure 1: Illustration of the document-level EAE task (�) and our chain-of-reasoning paradigm (■). beyond-sentence clues by incorporating hierarchi- cal encoding mechanisms (Du and Cardie, 2020a), generative paradigms (Li et al., 2021; Ma et al., 2022; Du et al., 2022), and document-level induc- tive bias (Wei et al., 2021; Pouran Ben Veyseh et al., 2022; Liu et al., 2022b). Nevertheless, such meth- ods do not explicitly characterize the reasoning patterns underlying the document context, which potentially …

Content truncated to 500 words