Analysis of Utterance Combinations for Emotion Recognition in Conversation

Abstract

Emotion recognition in conversation is an important step for developing empathetic systems in diverse areas such as healthcare, education, and business. Recent work demonstrates that utterance-level conversational context modeling leads to high-performance emotion recognition. In this paper, we combine the utterances of conversations where the same person speaks continuously with the same emotions, and and train with DialogueRNN to further improve emotion recognition performance. According to comparison of DialogueRNN models that trained by combining different numbers of utterances, the combination of three utterances achieves the highest F1-score of 63.90%, and improved the F1-score of 1.63% compared to baseline.