融合多语义特征的精读式文档级事件抽取  

Intensive reading document-level event extraction based on multi-feature semantics

作　　者：赵梦瑶 刘大明 ZHAO Meng-yao;LIU Da-ming(College of Computer Science and Technology,Shanghai University of Electric Power,Shanghai 201306,China)

机构地区：[1]上海电力大学计算机科学与技术学院,上海201306

出　　处：《计算机工程与设计》2024年第6期1903-1909,共7页Computer Engineering and Design

基　　金：甘肃省自然科学基金项目(SKLLDJ032016021)。

摘　　要：为解决文档级事件抽取任务依赖实体识别、忽略先验语义和参数分散的问题,提出一种融合多语义特征的精读式抽取方法。结合“三阶段”阅读特点,根据事件与角色交互、角色类型及释义特征构建外部语义模板,提出窗口切分算法切割文档语义;基于预训练模型BERT融合外部与窗口语义;多轮精读文档避免实体依赖,设计记忆网络对精读结果建模,完成跨句定位参数和事件路径扩展。引入噪声扰动防止模型过拟合。实验结果表明,该模型性能优于当前主流方法,验证了其可行性和有效性。To solve the problem that document-level event extraction tasks rely on entity recognition,ignore prior semantics and parameters dispersion,an intensive reading extraction method integrating multiple semantic features was proposed.Combined with the characteristics of three stages reading,the external semantic template was constructed according to the interaction between events and roles,role types and interpretation features,and the window segmentation algorithm was proposed to cut document semantics.The external semantics and window semantics were integrated based on the pre-training model BERT.Multi-round intensive reading documents avoided entity dependence.A memory network was designed to model the intensive reading results,and cross-sentence positioning parameters and event paths expansion were completed.The noise disturbance was introduced to prevent overfitting of pretrained language models(PLMs).Experimental results show that the performance of the model is superior to the current mainstream methods,and its feasibility and effectiveness is verified.

关 键 词：实体依赖 参数分散 语义特征融合 窗口切分算法 预训练模型 多轮精读 记忆网络 噪声扰动 

分 类 号：TP391.1[自动化与计算机技术—计算机应用技术]