文本中事件因果关系识别与应用技术综述  

A Survey on Event Causality Identification and Applications in Texts

作　　者：李顺航 周刚 卢记仓 李志博 黄宁博 陈静[1,2] LI Shunhang;ZHOU Gang;LU Jicang;LI Zhibo;HUANG Ningbo;CHEN Jing(Strategic Support Force Information Engineering University,Zhengzhou,Henan 450001,China;State Key Laboratory of Mathematical Engineering and Advanced Computing,Zhengzhou,Henan 450001,China)

机构地区：[1]战略支援部队信息工程大学,河南郑州450001 [2]数字工程与先进计算国家重点实验室,河南郑州450001

出　　处：《中文信息学报》2024年第9期1-23,共23页Journal of Chinese Information Processing

基　　金：河南省科技攻关计划项目(222102210081);河南省自然科学基金(222300420590)。

摘　　要：事件因果关系是一类重要的逻辑关系,其揭示了事件发展的动因与规律。通过自然语言处理技术对事件之间蕴含的因果关系进行识别,能够帮助形成事件因果知识库,进而促进诸如事件预测、智能问答等下游任务性能提升与可解释性增强,具有重要理论与实践价值。基于此,该文围绕事件因果关系识别与应用展开综述。首先,介绍文本中事件因果关系、因果关系识别等基本概念与任务定义,明确研究范畴;随后,总结归纳因果关系识别任务常用数据集与评测指标,并对典型评测数据集进行探索分析,进而充分挖掘任务固有难点;然后,按照基于规则挖掘、基于特征工程和基于深度学习三个类别对因果关系识别相关模型与方法进行划分,并给出系统阐释、对比和总结,并对事件因果关系支撑的下游应用场景与方法进行了概述,进一步说明了事件因果关系的重要应用价值;最后,针对文本中事件因果关系识别任务的现有挑战和未来技术方向进行了讨论与展望。Causality is an important type of logical relation between events that expresses high-level logical information and reveals event development patterns.The identification of event causality contained in texts via natural language processing methods is important in providing interpretability and robustness for various downstream applications such as question answering and event prediction.Therefore,we comprehensively review both the identification and application of event causality.First,to clarify the research scope,the basic concept of causality and the task definition of event causality identification(ECI)are introduced.Then,common-used datasets for ECI are summarized and further explored to figure out the inherent difficulties.Subsequently,following the technology development timeline,related ECI methods fall into three categories:rule mining,feature engineering,and deep learning.Based on this,a systematic and structured introduction,comparison,and summary are provided.Moreover,a brief overview of the application scenario of event causality is given to further show the significant application value of causal knowledge.Finally,the existing challenges and future research directions on ECI are discussed.

关 键 词：因果关系识别 自然语言处理 深度学习 数据增强 知识提升 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]