基于跨度和图卷积网络的篇章级事件抽取模型  

Document-level Event Extraction Model Based on Span and Graph Convolutional Network

作　　者：廖涛[1] 牛冰宇 LIAO Tao;NIU Bingyu(College of Computer Science and Engineering,Anhui University of Science and Technology,Huainan 232001,China)

机构地区：[1]安徽理工大学计算机科学与工程学院,安徽淮南232001

出　　处：《湖北民族大学学报(自然科学版)》2025年第1期108-113,共6页Journal of Hubei Minzu University:Natural Science Edition

基　　金：国家自然科学基金项目(62076006);安徽省高校协同创新项目(GXXT-2021-008)。

摘　　要：为解决现有的事件抽取方法在实体抽取子任务中难以充分利用上下文信息,导致事件抽取精度较低的问题,提出了基于跨度和图卷积网络的篇章级事件抽取(document-level event extraction based on span and graph convolutional network, DEESG)模型。首先,设计中间线性层对编码的向量进行线性处理,并结合标注信息计算最佳跨度,通过提升对跨度开始位置和结束位置判断的准确度来提高实体抽取的精度;接着,提出异构图的构建方法,使用池化策略将实体与句子表示为图的节点,根据提出的建边规则构建异构图,以此建立全局信息的交互,并利用多层图卷积网络(graph convolutional network, GCN)对异构图进行卷积,获得具有上下文信息的实体表示和句子表示,以此解决上下文信息利用不充分的问题;然后,利用多头注意力机制进行事件类型的检测;最后,为组合中的实体分配论元角色,完成事件抽取任务。在中文金融公告(Chinese financial announcements, ChFinAnn)数据集上进行实验。结果表明,与拥有追踪器的异构图交互模型(graph-based interaction model with a tracker, GIT)相比,DEESG模型的F1分数提升了1.3个百分点。该研究证实DEESG模型能有效应用于篇章级事件抽取领域。In order to solve the problem of existing event extraction methods in entity extraction subtasks,which made it difficult to fully utilize contextual information and led to low event extraction accuracy,the document-level event extraction based on span and graph convolutional network(DEESG)model was proposed.Firstly,an intermediate linear layer was designed to linearly process the encoded vectors,and annotation information was combined to calculate the optimal span.By improving the accuracy of determining the start and end positions of the span,the precision of entity extraction was enhanced.Secondly,a method for constructing heterogeneous graphs was designed,using a pooling strategy to represent entities and sentences as nodes of the graph.Heterogeneous graphs were constructed based on the proposed edge-building rules to establish global information interaction.Multi-layer graph convolutional networks(GCN)were used to convolve heterogeneous graphs,obtaining entity and sentence representations with contextual information,thus solving the problem of insufficient utilization of contextual information.Thirdly,a multi-head attention mechanism was used to detect event types.Finally,argument roles were assigned to the entities in the combination to complete the event extraction task.Experiments were conducted on the Chinese financial announcements(ChFinAnn)dataset,and the results showed that DEESG model improved the F1 score by 1.3 percentage points compared to graph-based interaction model with a tracker(GIT).The research confirmed that the DEESG model could be effectively applied in the field of document-level event extraction.

关 键 词：事件抽取 跨度 实体抽取 异构图 图卷积网络 上下文信息 

分 类 号：TP391.1[自动化与计算机技术—计算机应用技术]