基于多粒度融合的图卷积网络会话情感分析  

Multi-granular Information Fusion Approach to Graph ConvolutionalNetwork Based Conversational Emotion Recognition

作　　者：王佳 朱小飞 唐顾 黄贤英[1] WANG Jia;ZHU Xiaofei;TANG Gu;HUANG Xianying(College of Computer Science and Engineering,Chongqing University of Technology,Chongqing 400054,China)

机构地区：[1]重庆理工大学计算机科学与工程学院,重庆400054

出　　处：《中文信息学报》2024年第5期136-145,共10页Journal of Chinese Information Processing

基　　金：国家自然科学基金(62141201);重庆市自然科学基金(CSTB2022NSCQ-MSX1672);重庆市教育委员会科学技术研究计划重大项目(KJZD-M202201102)。

摘　　要：会话情感分析指对一段会话中的每句话进行情感分类,目前大部分会话情感分析模型不仅忽略了对话中内部信息的相互影响,而且没有考虑到日常对话中存在的隐性背景情感。为了有效解决这些问题,该文提出了一个基于多粒度融合的图卷积神经网络,其主要包括两个模块,即特征提取模块和星图增强的图学习模块。首先,特征提取模块使用预训练语言模型RoBERTa获取会话中语句之间粗粒度的上下文信息,同时结合句法依赖树获取词之间细粒度的句法信息,从而将多粒度特征信息引入到会话情感建模。然后,在星图增强的图学习模块中建模会话的背景情感信息和会话中不同说话者之间的交互信息,从而增强会话情感分析的准确性。实验结果表明,该文提出的模型与其他基线模型相比,其准确性以及度量指标F 1值在所有数据集上均有显著提升。Conversation sentiment analysis refers to the classification of emotions for each sentence in a conversation.To capture the hidden background emotions and the interaction of internal information in the conversation,this paper proposes a multi-granular information fusion approach to graph convolutional neural network based conversational emotion recognition.First,a feature extraction module uses the pre-trained language model RoBERTa to obtain coarse-grained contextual information between statements in conversation,and applies the syntactic dependence tree to obtain fine-grained syntactic information between words.Then,a star graph learning module enhances the accuracy of conversation sentiment analysis by modeling the contextual sentiment information of the conversation and the interaction information between different speakers in the conversation.Experimental results show that the accuracy of the proposed model and the value of the metric F 1 are significantly improved in all data sets compared with other baselines.

关 键 词：会话情感分析 多粒度融合 句法依赖树 图卷积网络 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]