融合依存句法和图注意力网络的事件检测模型  

Event Detection Based on Graph Attention Network with Dependency Syntactic

作　　者：郑志蕴[1] 裴晓波 李钝[1] 张行进[1] 王军锋[1] ZHENG Zhi-yun;PEI Xiao-bo;LI Dun;ZHANG Xing-jin;WANG Jun-feng(School of Computing and Artificial Intelligence,Zhengzhou University,Zhengzhou 450001,China)

机构地区：[1]郑州大学计算机与人工智能学院,郑州450001

出　　处：《小型微型计算机系统》2023年第11期2477-2484,共8页Journal of Chinese Computer Systems

基　　金：国家重点研发计划公共安全专项项目(244)资助;科学基金项目(17BXW065)资助。

摘　　要：事件检测是信息抽取的关键子任务,目的是识别文本中特定类型的事件实例.图注意力网络在图结构上使用自注意力机制,能实现高质量的事件检测,但现有图注意力网络存在忽略依存标签信息、上下文信息获取模型复杂等问题,导致模型准确率下降、计算资源开销较大.本文提出依存边信息嵌入的图注意力网络模型(EIEGAT),设计依存边信息嵌入模块,将依存标签信息嵌入到图的邻接矩阵中,使模型在构造图时同时考虑节点和依赖边的表示,提升事件检测准确率.使用结构和计算更简单的门控循环单元捕获单词的上下文信息,在维持性能的同时简化网络结构、节省内存空间.实验表明,EIEGAT有效提高事件检测的总体性能,在ACE2005英文语料集上事件识别与事件分类的F1值分别提高5%与0.9%.Event detection is a critical sub-task of information extraction,aiming to identify specific types of event instances in natural language.Graph Attention Network model uses self-attention mechanism on graph structure to achieve high quality event detection.However,the existing Graph Attention Network suffers from ignoring dependent label information and complex obtaining model of context information,which leads to reduced model accuracy and computational resource overhead.This paper proposes a Graph Attention Network model with dependency edge information embedding(EIEGAT).A dependency edge information embedding module is designed to embed the label information of the dependency edge into the adjacency matrix of graph,so that the model considers both node and dependent edge representations when constructing graph to improve the accuracy of event detection.Using the Gated Recurrent Unit with simpler structure and computation to capture the context information of words can simplify the network structure and save the memory space while maintaining the performance.The experimental results show that the EIEGAT can effectively improve the overall performance of event detection,and the F1 values of event recognition and event classification on ACE2005 English corpus are increased by 5%and 0.9%respectively.

关 键 词：事件检测 图神经网络 依存句法 图注意力网络 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]