Document-Level Multi-Event Extraction with Event Proxy Nodes and Hausdorff Distance Minimization

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 10118–10133 July 9-14, 2023 ©2023 Association for Computational Linguistics Document-Level Multi-Event Extraction with Event Proxy Nodes and Hausdorff Distance Minimization Xinyu Wang1,2, Lin Gui2, Yulan He1,2,3 1Department of Computer Science, University of Warwick 2Department of Informatics, King’s College London 3The Alan Turing Institute Xinyu.Wang.11@warwick.ac.uk {lin.1.gui, yulan.he}@kcl.ac.uk Abstract Document-level multi-event extraction aims to extract the structural information from a given document automatically. Most recent ap- proaches usually involve two steps: (1) model- ing entity interactions; (2) decoding entity inter- actions into events. However, such approaches ignore a global view of inter-dependency of multiple events. Moreover, an event is decoded by iteratively merging its related entities as ar- guments, which might suffer from error prop- agation and is computationally inefficient. In this paper, we propose an alternative approach for document-level multi-event extraction with event proxy nodes and Hausdorff distance min- imization. The event proxy nodes, representing pseudo-events, are able to build connections with other event proxy nodes, essentially cap- turing global information. The Hausdorff dis- tance makes it possible to compare the similar- ity between the set of predicted events and the set of ground-truth events. By directly mini- mizing Hausdorff distance, the model is trained towards the global optimum directly, which im- proves performance and reduces training time. Experimental results show that our model out- performs previous state-of-the-art method in F1-score on two datasets with only a fraction of training time. 1 1 Introduction Event extraction aims to identify event triggers with certain types and extract their corresponding argu- ments from text. Much research has been done on sentence-level event extraction (Du and Cardie, 2020; Lin et al., 2020; Lu et al., 2021). In recent years, there have been growing interests in tackling the more challenging task of document-level multi- event extraction, where an event is represented by a cluster of arguments, which may scatter across multiple sentences in a document. Also, multiple events in the same document may share some com- mon entities. For example, as shown in Figure 1, 1Code is available at https://github.com/xnyuwg/procnet ... [5] Shenkai Petrochemical Co., Ltd. received the receipt from the company's shareholder, Yexiang Investment Management Co., Ltd. on the evening of November 15, 2016 Regarding the notice of the shares being frozen. ... [8] On November 14, 2016, Yexiang Investment received the Notice of Litigation Preservation from the People's Court of Binjiang District, and granted a total of 47,577,481 shares held by Yexiang Investment will be frozen, and the freezing period is from October 31, 2016 to October 30, 2019 ... [10] Yexiang Investment is … holding 47,577,481 shares of the company, accounting for 13.07% of the company's total share capital. ... [12] On February 2, 2016, the 42,000,000 shares held by it are pledged to Haitong Securities Co., Ltd., and the repurchase transaction date was February 1, 2017. ... Event #1: Equity Pledge Pledger: Yexiang Investment Management Co., Ltd. Pledgee: Haitong Securities Co., Ltd. TotalHoldingShares: 47,577,481 …

Content truncated to 500 words