融合事件类型的中文事件抽取方法  

Gated Fusion Framework for Chinese Event Extraction

作　　者：王瑾睿 李劼 WANG Jinrui;LI Jie(School of Computer Science,Beijing University of Posts and Telecommunications,Beijing 100876,China)

机构地区：[1]北京邮电大学计算机学院,北京100876

出　　处：《计算机工程与应用》2025年第1期214-220,共7页Computer Engineering and Applications

基　　金：国家自然科学基金(U22B2019)。

摘　　要：事件抽取技术是自动化地从文本信息中获得结构性数据的重要手段,也是自然语言处理领域的重点研究方向之一。事件抽取包含两个子任务,事件类型检测与事件论元抽取。近年来的事件抽取研究引入了预训练语言模型作为文本的语义表征,然后采用序列标注BIO完成抽取任务,但此类方法容易存在标签歧义问题。因此又有学者提出在事件文本特征中融入特征知识以避免歧义,可现有的融合方法忽略了事件抽取各个子任务间的依赖关系。为解决以上问题,针对事件抽取任务采用联合学习的算法框架,通过门融合机制将事件类型信息作为新知识融入事件的文本表示中,再进行事件触发词抽取与事件论元抽取。实验结果证明该算法模型在论元抽取任务上较之基线方法表现更优异。Event extraction is an essential task which aims to extract structural event information from unstructured text.Event extraction consists of two sub-tasks,event type detection and event argument extraction.Recent prior work introduced pre-trained language model to get semantic representation of the text.Some works formulate EE into a sequence labeling task(BIO),which is prone to label ambiguity.While others integrate feature knowledge into event text representation to avoid ambiguity.However,the existing methods of integration ignore the dependencies between each sub-task of event extraction.This paper proposes a joint learning method for the event extraction task.The event type feature is integrated into the text representation as new knowledge through gate fusion mechanism.Experimental results show that this method performs better than the baselines on event argument extraction.

关 键 词：文本事件抽取 预训练语言模型 门融合 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]