Predicting Emotion Before an Utterance: Replacing Contextual Information with Knowledge-Based Prediction

Takumi Fujimoto, Takayuki Ito

Abstract: As textual conversation becomes more mainstream, emotional understanding that contributes to collective intelligence and team performance plays an essential role. However, most studies of emotion recognition in conversation (ERC) have focused on analyzing emotion after a participant speaks, and few have addressed predicting emotion in ongoing conversations before an utterance is made. In this paper, we tackle predicting emotion before a speaker expresses their emotion and propose utilizing knowledge-based prediction as an appropriate replacement for contextual information. We develop a context model to capture conversations and replace future conversational data that cannot be exploited in actual predictions with appropriate knowledge-based predictions. Our model outperformed existing models in predicting conversational emotions on a dataset composed of conversations between multiple participants.