Conversational emotion prediction based on appraisal theory

Perceiving, understanding, and expressing human emotions have always been challenging problems in the field of machine learning. Traditional research on emotion prediction has mainly focused on individual emotional changes and interactions between only two individuals. They overlook the fact that individuals adjust and control their emotional responses based on the environment, and the impact of the same environment on different individuals may vary. Moreover, these approaches are not applicable in multi-party conversation environments. In this paper, we propose a deep learning model for conversation-level emotion prediction based on the appraisal theory. The model consists of three components: BERT, primary appraisal, and secondary appraisal. The BERT layer is used to extract rich semantic information from the utterances of the speakers. The primary appraisal component comprehensively captures the entire conversation process and individual conversational characteristics using LSTM. The secondary appraisal component simulates the process by which speakers regulate and control their emotional responses through appraising themselves and others. The experimental results demonstrate that the proposed model outperforms five other methods on three manually annotated datasets, with F1 scores reaching 0.6004, 0.5603, and 0.5108, respectively. Further analysis validates the effectiveness of the model.