Multi-View Decoupling Framework for Emotion Recognition in Conversations

Emotion Recognition in Conversations (ERC) has always been a highly focused research topic in Human-Computer Interaction (HCI). The emotions of dialogues are not only dependent on the current discourse, but are also influenced by the context. Although recent studies have put forward various methods to address the issue, they are still insufficient to capture emotion shifts and context dependencies in conversations, leading to limited accuracy in emotion classification. To address this challenge, we propose an innovative emotion recognition framework, Multi-View Emotion Recognition Framework(MVERF), based on Multi-View Encoder and Decoupled Contrastive Learning. Specifically, an optimistic perspective encoder and a pessimistic perspective encoder are designed to independently capture optimistic and pessimistic emotions respectively, to ensure that both perspectives’ emotions are preserved and complementary. The unified emotion representation, as well as context embedding in dialogues, is then fed into a multi-head attention mechanism to capture multi-layered information, allowing for a better understanding of context and enhancing the accuracy in identifying emotion shifts. Decoupled contrastive Learning is exploited to avoid mutual interference between positive and negative samples, ensuring the purity and differentiation of emotion representation. In addition, a momentum contrast mechanism is raised to maintain a dynamic sample queue and to increase the number and diversity of negative samples. Extensive experimental results indicate that our proposed model exhibits excellent performance on the ERC task.