Multi-View Contrastive Parsing Network for Emotion Recognition in Multi-Party Conversations

Recent Emotion Recognition in Conversation (ERC) works significantly outperform large language models, represented by ChatGPT, in the dyadic conversation environment by introducing knowledge and adjusting training strategies. However, Multi-Party Conversations (MPCs) are more complex due to their multi-thread nature, low information density, and general long-range dependencies. In addition, previous studies have overlooked the phenomenon of utterance polysemy. To address these challenges, this paper proposes a Multi-View Contrastive Parsing Network (MuVCPN). Specifically, we first parse the entire conversation and extract emotion-related cues from independent sub-conversation views. Then, we update the utterance distance based on the parsing results and use a discourse structure-aware self-attention mechanism to capture the conversational information flow from the global view. At the same time, we adopt supervised contrastive learning to group utterances from the same sub-conversation together. Extensive experiments on four benchmarks show that the proposed MuVCPN model outperforms baseline models on the ERC task. Additionally, experimental results indicate that utilizing different views and sub-conversation level contrastive learning can improve performance in the MPCs environment.