Enhancing Emotion Recognition in Conversation with Dialogue Discourse Structure and Commonsense Knowledge

Emotion Recognition in Conversations (ERC) is the task of identifying the emotions of utterances from speakers in a conversation, which is beneficial to many applications. In this paper, we introduce two kinds of external knowledge, i.e., dialogue discourse structure and social commonsense knowledge implied in dialogue to enhance representation and emotional reasoning. The dialogue discourse structure directly reveals the adjacent or long-distance dependencies between utterances and provides prior knowledge for the semantic interaction between utterances. Implicit commonsense knowledge in utterances can serve as emotional inference cues to model deeper inter-utterance emotional interactions. Specifically, we construct a discourse structure and commonsense knowledge enhanced graph structure over the conversation and use graph convolutional networks to incorporate historical context and commonsense knowledge for utterances. Experimental results show that incorporating discourse structure and commonsense knowledge can effectively improve the performance of the model.