基于角色分离的中文事件抽取研究  

Chinese Event Extraction Based on Role Separation

作　　者：张旭 朱艳辉[2] 曾志高 欧阳康[1,2] 孔令巍 ZHANG Xu;ZHU Yanhui;ZENG Zhigao;OUYANG Kang;KONG Lingwei(School of Computer,Hunan University of Technology,Zhuzhou412007,China;Hunan Key Laboratory of Intelligent Information Perception and Processing Technology,Hunan University of Technology,Zhuzhou412007,China)

机构地区：[1]湖南工业大学计算机学院,湖南株洲412007 [2]湖南工业大学智能信息感知及处理技术湖南省重点实验室,湖南株洲412007

出　　处：《山西大学学报（自然科学版）》2022年第4期936-946,共11页Journal of Shanxi University(Natural Science Edition)

基　　金：国家自然科学基金(61871432);湖南省自然科学基金(2020JJ6089);湖南省教育厅重点项目(19A133)。

摘　　要：端到端的事件抽取涉及两个子任务:事件检测和事件元素抽取,当前大多数研究工作并未解决元素抽取中角色重叠的问题,并且忽略了数据集提供的静态信息。文章提出一种基于两个独立编码器的事件抽取模型,通过词典增强的序列标注方法检测触发词和事件类型,再将事件类型、实体类型和边界信息作为元素抽取阶段的补充特征,使用多个角色二分类网络实现“分离式”的元素识别和角色分类,最后将基于静态数据统计得到的角色辨识度对所有分类器的损失函数加权求和,作为最终的损失函数,进而更好地拟合数据集特点。在两个公开的中文数据集ACE2005和DuEE上,实验F1值分别达到57.1%和76.1%,并且该方法有效提高了对角色重叠事件的识别准确率。End-to-end event extraction involves two subtasks:event detection and event argument extraction.Most current researches did not address the problem of overlapped roles in argument extraction and ignore the static information provided by the dataset.This paper proposes an event extraction model based on two independent encoders,which detects trigger words and event types by lexicon-enhanced sequence labeling,then uses event types,entity types and boundary information as additional features in the argument extraction stage.The argument recognition and role classification are achieved by using multiple binary classification networks of roles.The loss functions of all classifiers are added by weight based on the scores of roles’recognitions obtained from statistics of the dataset to better fit the dataset’s characteristics.The experimental F1 values on two public Chinese datasets ACE2005and DuEE are 57.1%and 76.1%respectively,and the method can significantly improve the accuracy of events with overlapped roles.

关 键 词：事件抽取 元素抽取 角色辨识度 角色重叠 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]