A Text-Guided Two-Stage Multimodal Fusion Framework for Dialogue Emotion Recognition

Multimodal Emotion Recognition in Conversation (MERC) focuses on identifying speakers' emotions by jointly exploiting textual, acoustic, and visual cues in dialogue. Although existing methods have made considerable progress, they still face two main challenges: modeling effective cross-modal interactions while reducing the influence of noisy non-verbal information, and learning robust multimodal representations from heterogeneous modalities. To overcome these limitations, we propose TTMFF, a Text-Guided Two-Stage Multimodal Fusion Framework. Specifically, the Text-Guided Audio-Visual Fusion (TAVF) module uses textual information as semantic guidance to enhance acoustic and visual representations and capture contextual dependencies. The Dynamic Dual-modal Polynomial Fusion (DDPF) module further integrates enhanced textual features and fused audio-visual features by modeling both linear semantic information and second-order cross-modal interactions. In addition, a lightweight Adaptive Trust-Aware Distillation (ATAD) strategy is employed as auxiliary regularization during training to provide confidence-gated guidance for unimodal branches. The results on IEMOCAP and MELD indicate that TTMFF performs competitively against representative methods, while the ablation experiments confirm the contribution of the proposed fusion design.