基于篇章级语义图的对话一致性检测  

Examining Dialogue Consistency Based on Chapter-Level Semantic Graph

作　　者：李霏 邓凯方 范茂慧 滕冲[1] 姬东鸿[1] Li Fei;Deng Kaifang;Fan Maohui;Teng Chong;Ji Donghong(Key Laboratory of Aerospace Information Security and Trusted Computing,Ministry of Education,School of Cyber Science and Engineering,Wuhan University,Wuhan 430072,China)

机构地区：[1]武汉大学国家网络安全学院空天信息安全与可信计算教育部重点实验室,武汉430072

出　　处：《数据分析与知识发现》2024年第5期18-28,共11页Data Analysis and Knowledge Discovery

基　　金：教育部人文社会科学研究青年基金项目(项目编号:21YJCZH064)的研究成果之一。

摘　　要：【目的】通过融合包含共指链以及抽象语义表示等语义信息的对话篇章级语义图,提高对话一致性检测的准确性。【方法】首先,利用预训练语言模型BERT编码对话上下文和知识库;其次,构建包含共指链和抽象语义表示等语义信息的对话篇章级语义图,利用多关系图卷积神经网络捕获语义图中的语义信息;最后,构建多个分类器预测多种对话不一致现象。【结果】基于CI-ToD基准数据集,与现有对话不一致检测模型进行实验对比,本文模型在F1值或准确率指标上较之前的最优模型取得0.01以上的提升。【局限】所提模型不能很好地处理对话中存在的共指实体省略问题。【结论】融合共指链以及抽象语义表示等多种类别的语义信息能够有效提升对话一致性检测的效果。[Objective]This paper integrates chapter-level semantic graphs to improve the accuracy of dialogue consistency detection.[Methods]First,we used the pre-trained language model BERT to encode the dialogue context and knowledge base.Then,we constructed a dialogue chapter-level semantic graph containing coreference chains and abstract meaning representations.Third,we captured the semantic information of the constructed graph using a multi-relation graph convolutional network.Finally,we built multiple classifiers to predict dialogue inconsistency.[Results]We examined our new model on the CI-ToD benchmark dataset and compared its performance with the existing models.The proposed model’s F1 value improved by more than 1% over the optimal models.[Limitations]The proposed model cannot address the co-referential entity omission in dialogues.[Conclusions]Integrating various types of semantic information,such as coreference chains and abstract meaning representations,can effectively improve the performance of dialogue consistency detection.

关 键 词：对话系统 一致性检测 共指链 抽象语义表示 图卷积神经网络 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]