MAG-EMN-KD: Multimodal Attention Gating with Emotion Memory Network and Knowledge Distillation for conversational emotion recognition

Emotion Recognition in Conversations (ERC) aims to recognize utterance-level emotions by modeling multimodal contextual information. However, existing methods often suffer from modality and noise, as weak modalities like audio and video may introduce irrelevant information. To address these challenges, we propose MAG-EMN-KD, a text-dominant multimodal framework for conversational emotion recognition. The proposed model leverages context-aware prompt learning to extract highly discriminative emotional representations from the text modality and adopts a two-stage knowledge distillation strategy to transfer high-level emotional knowledge from text to weaker modalities. Building on these enhanced representations, an Emotion Memory Network (EMN) is introduced to capture conversation-level emotion dynamics by maintaining global emotional memory across dialogue history. Furthermore, a Cross-Modal Emotion Attention (CMEA) mechanism performs text-guided selective fusion, where EMN-enhanced text representations attend to semantically aligned audio cues while suppressing weakly discriminative visual information. Experiments on IEMOCAP and MELD show that MAG-EMN-KD consistently outperforms strong baselines, achieving weighted F1 scores of 71.88% and 65.34%, respectively.