Multi-granular Information Fusion Approach to Graph Convolutional Network Based Conversational Emotion Recognition
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 conversa-tional emotion recognition.First,a feature extraction module uses the pre-trained language model RoBERTa to ob-tain 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.
万方数据
维普
