Modeling Sentiment-Speaker-Dependency for Emotion Recognition in Conversation

Emotion Recognition in Conversations (ERC) plays a crucial role in the development of human-machine interaction. Conversations are a multi-party, multi-emotion, and multi-turn process of information propagation. However, existing works, which focus on designing models and algorithms for better learning representations of dialogue context and speakers, but rarely care about the key element of the strong correlation and inseparable interdependence of emotional states on the sentiment polarity in the process. To address this issue, we propose a novel model, named S2D-ERC (Sentiment-Speaker-Dependency for Emotion Recognition in Conversation), for ERC task. The proposed model constructs a conversation as a directed acyclic graph and represents both speaker- and sentiment-dependencies between utterances with heterogeneous edges. Additionally, to capture the information interaction dynamics in conversation context, we employ a cross-attention mechanism where latent representations of speaker and sentiment are learned with two different directions of information flow. The experimental results on two benchmarks, compared with state-of-the-art models, demonstrate the superiority and effectiveness of our model.