融合情绪标签和原型网络的对话情绪识别  

Emotion recognition in conversation integrating emotion labels and prototypical networks

作　　者：张洪通 王健[1] 徐博[1] 杨亮[1] 林鸿飞[1] ZHANG Hongtong;WANG Jian;XU Bo;YANG Liang;LIN Hongfei(School of Computer Science and Technology,Dalian University of Technology,Dalian 116024,China)

机构地区：[1]大连理工大学计算机科学与技术学院,辽宁大连116024

出　　处：《大连理工大学学报》2025年第3期313-320,共8页Journal of Dalian University of Technology

基　　金：国家自然科学基金资助项目(62006034).

摘　　要：对话情绪识别的目标是识别对话中话语的情绪,尽管已经有一些先进的监督对比学习方法被提出用于区分不同的情绪类别,但情绪标签中的内在信息仍未得到充分利用.情绪标签包含特定的语义和复杂的相互关系,可以被用作对比学习的样本,以提高情绪识别的效果.提出了一种标签引导的原型对比学习方法用于对话情绪识别,该方法设计了一个对比目标,其中情绪标签被视为正负样本,构建的标签嵌入参与了对比训练过程,有效地丰富了对比样本.此外,该方法利用原型网络关注数据的整体分布和平均特性.在3个广泛使用的基准数据集上的实验表明,所提方法在对话情绪识别上的性能超越了现有方法.The goal of emotion recognition in conversation is to identify the emotions of the utterances within a dialogue.Although several advanced supervised contrastive learning methods have been proposed to distinguish different emotion categories,the intrinsic information in emotion labels is not fully utilized.Emotion labels contain specific semantics and complex relationships.They can be utilized as samples for contrastive learning to facilitate emotion recognition.A label-guided prototypical contrastive learning method is proposed for emotion recognition in conversation.This method designs a contrastive objective in which the emotion labels are treated as positive/negative samples and the constructed label embeddings are involved in the contrastive training process,which effectively enriches the contrastive samples.Additionally,this method utilizes a prototypical network to focus on the overall distribution and average characteristics of the data.Experiments on three widely used benchmark datasets show that the proposed method outperforms existing approaches for emotion recognition in conversation.

关 键 词：情绪识别 对比学习 标签信息 原型网络 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]