MM-NodeFormer: Node Transformer Multimodal Fusion for Emotion Recognition in Conversation

Emotion Recognition in Conversation (ERC) has great prospects in human-computer interaction and medical consultation. Existing ERC approaches mainly focus on information in the text and speech modalities and often concatenate multimodal features without considering the richness of emotional information in individual modalities. We propose a multimodal network called MM-NodeFormer for ERC to address this issue. The network leverages the characteristics of different Transformer encoding stages to fuse the emotional features from the text, audio, and visual modalities according to their emotional richness. The module considers text as the main modality and audio and visual as auxiliary modalities, leveraging the complementarity between the main and auxiliary modalities. We conducted extensive experiments on two public benchmark datasets, IEMOCAP and MELD, achieving an accuracy of 74.24% and 67.86%, respectively, significantly higher than many state-of-the-art approaches.