说话者特征融合的对话情感识别模型  

Speaker Feature Fusion Model for Emotion Recognition in Conversation

作　　者：刘欣雨 夏鸿斌 刘渊 LIU Xinyu;XIA Hongbin;LIU Yuan(School of Artificial Intelligence and Computer,Jiangnan University,Wuxi 214122,China;Jiangsu Key Laboratory of Media Design and Software Technology,Wuxi 214122,China)

机构地区：[1]江南大学人工智能与计算机学院,江苏无锡214122 [2]江苏省媒体设计与软件技术重点实验室,江苏无锡214122

出　　处：《小型微型计算机系统》2025年第3期571-577,共7页Journal of Chinese Computer Systems

基　　金：国家自然科学基金项目(61972182)资助.

摘　　要：对话情感识别旨在预测对话中话语的情感.目前的方法大多只针对上下文信息进行建模,忽略了对说话者的建模.同时,语境在对话情感识别中起着至关重要的作用.为此,本文提出了说话者特征融合的对话情感识别模型.首先,将说话者名字作为标签进行特征提取,并在构造图结构时为说话者设置单独的节点.其次,分别构建全局对话、说话者在对话中对自己的情感影响和对其他说话者情感影响的图结构.然后,通过多头注意力获得体现对话语境的全局特征,将其与图卷积及门控循环单元融合获得分类特征.最后,通过前馈网络对话语情感进行分类.在IEMOCAP、MELD、EmoryNLP这3个基准数据集上的实验结果表明,该模型在性能指标上较其他基线模型均有一定提升.Emotion recognition in conversation aims to predict the emotion of utterances in conversations.Most of the current methods only model contextual information,ignoring the modeling of speakers.At the same time,context plays a crucial role in emotion recognition in conversation.Therefore,this paper proposes a model which named Speaker Feature Fusion Model for Emotion Recognition in Conversation.Firstly,the speaker′s name was used as a label for feature extraction,and a separate node was set for the speaker when constructing the graph structure.Secondly,the graph structure of complete conversation,intra-and inter-speaker′s emotional impact in the conversation were constructed separately.Then,the global features reflecting the conversation context were obtained through multi-head attention,and the classification features were obtained by fusing them with Graph Convolutional Network and Gate Recurrent Unit.Finally,the utterance sentiment was classified through a feed-forward network.The experimental results on IEMOCAP,MELD and EmoryNLP show that the performance of the proposed model is improved compared with other baseline models.

关 键 词：对话情感识别 上下文建模 说话者建模 图卷积网络 注意力机制 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]