Discriminative Reasoning with Sparse Event Representation for Document-level Event-Event Relation Extraction

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 16222–16234 July 9-14, 2023 ©2023 Association for Computational Linguistics Discriminative Reasoning with Sparse Event Representation for Document-level Event-Event Relation Extraction Changsen Yuan1, Heyan Huang1,∗, Yixin Cao2, Yonggang Wen3 1Beijing Institute of Technology, Beijing, China 2Singapore Management University, Singapore 3Nanyang Technological University, Singapore {yuanchangsen, hhy63}@bit.edu.cn caoyixin2011@gmail.com, ygwen@ntu.edu.sg Abstract Document-level Event-Event Relation Extrac- tion (DERE) aims to extract relations between events in a document. It challenges conven- tional sentence-level task (SERE) with difficult long-text understanding. In this paper, we pro- pose a novel DERE model (SENDIR) for better document-level reasoning. Different from exist- ing works that build an event graph via linguis- tic tools, SENDIR does not require any prior knowledge. The basic idea is to discriminate event pairs in the same sentence or span multi- ple sentences by assuming their different infor- mation density: 1) low density in the document suggests sparse attention to skip irrelevant in- formation. Our module 1 designs various types of attention for event representation learning to capture long-distance dependence. 2) High den- sity in a sentence makes SERE relatively easy. Module 2 uses different weights to highlight the roles and contributions of intra- and inter- sentential reasoning, which introduces support- ive event pairs for joint modeling. Extensive experiments demonstrate great improvements in SENDIR and the effectiveness of various sparse attention for document-level representa- tions. Codes will be released later. 1 Introduction Event-Event Relation Extraction (ERE) is the task of identifying the relation between two events in texts. As shown in Figure 1, for any pair of events1, e.g., (Services, downtime), it shall make the classi- fications for which relation type it holds. Clearly, event pairs may be in the same sentence (SERE) (Kadowaki et al., 2019a; Liu et al., 2020; Kad- owaki et al., 2019b), or scattered across the entire document (DERE) (Phu and Nguyen, 2021). In practice, DERE can benefit a wider range of ap- plications, such as knowledge graph construction (Chen et al., 2019) and future event forecasting ∗Corresponding author 1Event is defined as the trigger word in this area. [3] Vodafone Egypt said some users are continuing to face slight problems due to damage in Telecom’, “Egypt‘s”, ’cables which began in central Cairo . [4] Vodafone , along with the country‘s three telecom providers has been affected by a damage in one of Telecom Egypt’s fiber cables … 14,21,35,41,38 (6, 11) 6-2-Serv (6, 9)9-2-down (6,7)7-2-disrup (7, 24) 24-2-co (7, 14)14-4-dam (7, 12)12-3-dam 13-4-affected 14-4-damage [2] Services on‘, “Egypt’s”, ‘three telecom operators were disrupted in parts of Alexandria on Tuesday afternoon , and some users in Cairo experienced slower internet connections due to “downtime”. … Figure 1: Example of document-level ERE from EventStoryLine. Solid lines denote causal relations; some events and relations are omitted for clarity. The red boxes are indicator words for causal relations. (Hashimoto, 2019) but it remains challenging due to the difficulty in long-text understanding. In this paper, we propose to improve the rea- …

Content truncated to 500 words