融入说话者位置异构图的多方对话实体关系抽取  

Multi-party Dialogue Entity Relation Extraction Incorporating Speaker Position Heterogeneous Graph

作　　者：郑思源 ZHENG Siyuan(School of Computer Science and Mathematics,Fujian University of Technology,Fuzhou 350118,China)

机构地区：[1]福建理工大学计算机科学与数学学院,福建福州350118

出　　处：《现代信息科技》2024年第15期42-46,共5页Modern Information Technology

摘　　要：对话关系抽取旨在预测对话文本中的实体对之间的关系。在多方对话场景中,由于角色信息的不确定性,传统的实体关系抽取方法面临准确性挑战。针对角色信息不明的问题,模型构建了对话文本异构图,以捕捉多方对话的复杂结构。此外,为了有效处理不同对话轮次中的说话者信息,引入了长短时记忆网络进行特征处理,并将这些特征嵌入到相应的句子节点中。在DialogRE数据集上的实验显示,该模型展现了出色的性能,具体体现在F1和F1c分别达到了66.8%和62.2%,从而验证了其优越性。Dialogue Relation Extraction aims to predict the relationship between entity pairs in dialogue texts.In multiparty dialogue scenarios,traditional methods of entity relation extraction face accuracy challenges due to the uncertainty of role information.To address the issue of unclear role information,the model constructs a heterogeneous graph of dialogue texts to capture the complex structure of multi-party dialogue.Moreover,in order to handle the speaker information effectively in different rounds of dialogue,the Long Short-Term Memory network is introduced for feature processing and these features are embedded into the corresponding sentence nodes.Experiments on the DialogRE dataset demonstrate the outstanding performance of model,specifically reflected in achieving F1 and F1c scores of 66.8%and 62.2%respectively,thus confirming its superiority.

关 键 词：关系抽取 多方对话 角色信息 异构图 

分 类 号：TP391.1[自动化与计算机技术—计算机应用技术]