基于深度学习的篇章级事件抽取综述  

Document-Level Event Extraction Based on Deep Learning:A Survey

作　　者：胡蓉[1,2] 万常选[1,2] 万齐智 刘德喜[1,2] 刘喜平[1,2] HU Rong;WAN Chang-Xuan;WAN Qi-Zhi;LIU De-Xi;LIU Xi-Ping(School of Computer and Artificial Intelligence,Jiangxi University of Finance and Economics,Nanchang 330032;Key Laboratory of Data and Knowledge Engineering,Jiangxi University of Finance and Economics,Nanchang 330013)

机构地区：[1]江西财经大学计算机与人工智能学院,南昌330032 [2]江西财经大学数据与知识工程江西省高校重点实验室,南昌330013

出　　处：《计算机学报》2025年第2期381-406,共26页Chinese Journal of Computers

基　　金：国家自然科学基金项目(62272205,62272206,62076112,62462034);江西省教育厅科学技术研究项目(GJJ210531,GJJ2400411);江西省研究生创新专项资金项目(YC2023-B188);江西省自然科学基金(20242BAB25119,20212ACB202002,20232ACB202008);江西省主要学科学术和技术带头人培养计划领军人才项目(20213BCJL22041)资助。

摘　　要：篇章级事件抽取是自然语言处理的重要任务且富有挑战,当前涌现了很多优秀的研究成果。尽管国内外存在少量篇章级事件抽取综述,但存在一些局限:(1)按文献采用的具体技术或任务实现步骤对现有研究成果进行分类,未深入分析现有研究成果间的关联与区别,未深刻理解现有研究成果分别致力于解决哪些问题;(2)简单介绍现有数据集,未能正确认识每个数据集的特点及带来的任务挑战。由于每个数据集侧重点不同,研究者们致力于解决不同的问题,因此现有梳理方式未能清晰地展示不同数据集下不同研究问题的研究进展。为此,本文重新梳理篇章级事件抽取的2个(子)任务的研究成果。首先,针对2个任务,分别明确任务目标,分析解决任务的基本思路,总结现有研究进展(基于哪些数据集解决了哪些问题)。然后,总结对应数据集的特点,归纳任务面临的挑战,再深入分析具体研究方法,并图示化展示推进情况。最后,结合有待继续攻破的问题,讨论篇章级事件抽取未来发展趋势。Document-level Event extraction(DEE)is an important and challenging task in natural language processing,and numerous outstanding research achievements have emerged.DEE mainly focuses on two tasks,namely Document-level Event Identification and Argument Extraction(DocEI&AE),Document-level Event Argument Extraction(DocEAE).DocEI&AE indicates the complete document-level event extraction,that is,judging what types of events exist in a given document,identifying all events under each event type,and extracting arguments of corresponding roles.DocEAE refers to the event argument extraction,that is,given the event types and event triggers contained in each document,extracting event arguments of the corresponding roles triggered by each trigger.The goals of two tasks are different,and the task steps are also not exactly the same.Furthermore,the corresponding datasets also have different characteristics and focus on causing different research problems.Although there are a few surveys on document-level event extraction,they share the following two limitations.(1)The classification of existing researches is usually conducted based on specific techniques or task steps adopted in the literatures,without an in-depth analysis of the correlations and differences between the methods or a profound understanding of the issues each of them aims to address.(2)The description of datasets for DEE is simple and fails to understand the characteristics and task challenges of each dataset.Due to the different concerns of each dataset,the issues that researchers strive to solving vary.Therefore,current reviewing methods fail in clearly demonstrating the research progress on different issues under diverse datasets.This paper reorganizes the results of DocEI&AE and DocEAE tasks in document-level event extraction.Firstly,the task objectives and the common approach to solving the tasks are clarified and analyzed,then the current research progress(solving which issues based on the datasets)is summarized.Specifically,there are two ways to implement

关 键 词：篇章级事件抽取 信息抽取 事件抽取数据集 事件论元抽取 深度学习 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]