Context- and Sentiment-Aware Networks for Emotion Recognition in Conversation
Cheng Liu,E. Cambria,Geng Tu,Dazhi Jiang,Jintao Wen
DOI: https://doi.org/10.1109/TAI.2022.3149234
2022-10-01
IEEE Transactions on Artificial Intelligence
Abstract:Emotion recognition in conversation (ERC) has promising potential in many fields, such as recommendation systems, man–machine interaction, and medical care. In contrast to other emotion identification tasks, conversation is essentially a process of dynamic interaction in which people often convey emotional messages relying on context and common-sense knowledge. In this article, we propose a context- and sentiment-aware framework, termed Sentic GAT, to solve this challenge. In Sentic GAT, common-sense knowledge is dynamically represented by the context- and sentiment-aware graph attention mechanism based on sentimental consistency, and context information is captured by the dialogue transformer (DT) with hierarchical multihead attention (HMAT), where HMAT is used to obtain the dependency of historical utterances on themselves and other utterances for better context representation. Additionally, we explore a contrastive loss to discriminate context-free and context-sensitive utterances in emotion identification to enhance context representation in straightforward conversations that directly express ideas. The experimental results show that context and sentimental information can promote the representation of common-sense knowledge, and the intra- and inter-dependency of contextual utterances effectively improve the performance of Sentic GAT. Moreover, our Sentic GAT using emotional intensity outperforms the most advanced model on the tested datasets.
Computer Science