Emotion recognition in conversations (ERC), which involves identifying the emotional state of each utterance within a dialogue, plays a vital role in developing empathetic artificial intelligence systems. In practical applications, such as video-based recruitment interviews, customer service, health monitoring, intelligent personal assistants, and online education, ERC can facilitate the analysis of emotional cues, improve decision-making processes, and enhance user interaction and satisfaction. Current multimodal emotion recognition research faces several challenges, such as ineffective emotional information extraction from single modalities, underused complementary features, and inter-modal redundancy. To tackle these issues, this paper introduces a cross-modal gated attention mechanism for emotion recognition. The method extracts and fuses visual, textual, and auditory features to enhance accuracy and stability. A cross-modal guided gating mechanism is designed to strengthen single-modality features and utilize a third modality to improve bimodal feature fusion, boosting multimodal feature representation. Furthermore, a cross-modal distillation loss function is proposed to reduce redundancy and improve feature discrimination. This function employs a dual-supervision mechanism with teacher and student models, ensuring consistency in single-modal, bimodal, and trimodal feature representations. Experimental results on the IEMOCAP and MELD datasets indicate that the proposed method achieves higher accuracy and comparable F1 scores than existing approaches, highlighting its effectiveness in capturing multimodal dependencies and balancing modality contributions.
Zhao et al. (Sat,) studied this question.