Transformer-Based Potential Emotional Relation Mining Network for Emotion Recognition in Conversation

Abstract

Emotion recognition in conversation (ERC) has attracted much attention due to its widespread applications in the field of human communication analysis. Compared with the vanilla sentiment analysis of the single utterance, the ERC task which aims to judge the emotion labels of utterances in the conversation requires modeling both the contextual information and the speaker dependency. However, previous models are limited in exploring the potential emotional relation of the utterances. To address the problem, we propose a novel transformer-based potential emotion relation mining network (TPERMN) to better explore the potential emotional relation and integrate the emotional clues. First, we utilize the global gated recurrence unit to extract the situation-level emotion vector. Then, different speaker GRU are assigned to different speakers to capture the intra-speaker dependency of the utterances and obtain the speaker-level emotion vector. Second, a potential relation mining transformer called PERformer is devised to extract the potential emotional relation and integrate emotional clues for the situation-level and speaker-level emotion vector. In PERformer, we combine graph attention and multi-head attention mechanism to explore the deep semantic information and potential emotional relation. And an emotion augment block is designed to enhance and complement the inherent characteristics. After multi-layer accumulation, the updated representation is obtained for emotion classification. Detailed experiments on two public ERC datasets demonstrate our model outperforms the state-of-the-art models.