Context-Aware Hierarchical Graph Fusion for Emotion Recognition in Conversations

Emotional Recognition in Conversation (ERC) aims to improve user experience by identifying emotions in dialogue signals. It has extensive uses across different domains. Yet, the existing research into ERC is mainly focused on directly modeling the relationships in multimodal features once they have been fused, often overlooking the complex dependencies present in conversations. To address these issues, this approach proposes Context-Aware Hierarchical Graph Fusion (CA-HGF). The proposed method uses hierarchical fusion graphs to effectively capture both speaker-specific and cross-modal temporal dependencies. Additionally, the approach introduces two key innovations: a context filtering module that combines similarity and information entropy matrices to filter out low-importance or weakly correlated utterances, allowing the Graph Neural Network to better capture intrinsic relationships, while processing multifrequency signals to capture both emotional commonalities and differences, thus facilitating the exploration of diverse emotional relationships. Results from experiments conducted on IEMOCAP and MELD indicate CA-HGF surpasses current methods, offering a more effective and robust method for recognizing emotions in conversations.