Few-Shot Document-Level Event Argument Extraction

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 8029–8046 July 9-14, 2023 ©2023 Association for Computational Linguistics Few-Shot Document-Level Event Argument Extraction Xianjun Yang Yujie Lu Linda Petzold {xianjunyang,petzold}@ucsb.edu yujielu10@gmail.com Department of Computer Science University of California, Santa Barbara Abstract Event argument extraction (EAE) has been well studied at the sentence level but under-explored at the document level. In this paper, we study to capture event arguments that actually spread across sentences in documents. Prior works usually assume full access to rich document supervision, ignoring the fact that the available argument annotation is usually limited. To fill this gap, we present FewDocAE, a Few-Shot Document-Level Event Argument Extraction benchmark, based on the existing document- level event extraction dataset. We first define the new problem and reconstruct the corpus by a novel N-Way-D-Doc sampling instead of the traditional N-Way-K-Shot strategy. Then we adjust the current document-level neural mod- els into the few-shot setting to provide base- line results under in- and cross-domain settings. Since the argument extraction depends on the context from multiple sentences and the learn- ing process is limited to very few examples, we find this novel task to be very challenging with substantively low performance. Consider- ing FewDocAE is closely related to practical use under low-resource regimes, we hope this benchmark encourages more research in this direction. Our data and codes will be available online1. 1 Introduction Event argument extraction (EAE), a sub-task of event extraction, is a fundamental task for many downstream NLP applications in the IE community. For example, events and arguments play an impor- tant role in the knowledge base population from unstructured data (Ge et al., 2018; Li et al., 2021). And the real world public affairs management re- lies on recognizing the event arguments from daily news and social media (Yuan et al., 2018; Ritter et al., 2012). Although tremendous progress has been made under the supervised setting, current 1https://github.com/Xianjun-Yang/FewDocAE neural models typically rely on large-scale human- annotated data, which is not reliable considering the huge amounts of novel events and arguments emerging in many fields every day. Few-shot learn- ing (FSL) (Fei-Fei et al., 2006) is proposed to tackle such limitations to make machine learning models more applicable given limited annotated examples and has been used a lot in the IE area (Han et al., 2018; Ding et al., 2021; Lai et al., 2021). Previous research (Yang et al., 2019; Tong et al., 2020) mainly focuses on sentence-level event ex- traction, such as the popular ACE2005 (Dodding- ton et al., 2004) dataset. In recent years, researchers have begun to realize that the complete event and arguments actually spread in a full document or paragraph (Li et al., 2021; Ebner et al., 2020; Li, 2022). And the focus starts to turn into document- level event extraction motivated by the newly pro- posed large-scale document-level corpora, namely the WikiEvents (Li et al., 2021), RAMS (Ebner et al., 2020) and the recent largest corpus DocEE (Li, …

Content truncated to 500 words