HMCAT: Hierachical Multi-task and Commonsense-Aware Transformer for Emotion Recognition in Conversation.

Understanding emotion is a prerequisite to expressing emotion. Emotion recognition in Conversation (ERC) has attracted more and more attention because of its significance in developing empathetic dialogue systems. Recently, many approaches have focused on using neural networks to model semantics in the context. However, these methods are inadequate in understanding emotions due to a lack of ability to extract and integrate emotion cues. In this paper, we propose a Hierarchical Multi-task and Commonsense-Aware Transformer (HMCAT) to handle the challenges. Based on the introduction of commonsense knowledge, we explore the interaction patterns between cognitive factors and emotional factors in the process of dialogue advancement so as to fully understand the emotional cues in the context of dialogue. The model has been experimented on two datasets in ERC and another two auxiliary tasks, demonstrating its superiority empirically over existing strong baselines. The experimental results show that HMCAT also has excellent ability in dialogue action recognition and emotion prediction.