PGIF: A Personality-Guided Iterative Feedback Graph Network for Multimodal Conversational Emotion Recognition

Multimodal emotion recognition in conversation (MERC) aims to identify emotions in target utterances from multimodal records, drawing significant attention for its value in conversational artificial intelligence. While early research focuses on exploring conversational context, recent efforts emphasize integrating multimodal cues. Existing methods focus on modeling the impact of conversational context on emotion recognition while neglecting the role of the speaker's personality factors. Furthermore, these approaches often suffer from inefficiencies in information transfer due to full-utterance connectivity and fail to leverage multiple fusion modes for complementary benefits. To address these issues, we propose a personality-guided iterative feedback graph network (PGIF) for MERC. PGIF incorporates personality information as a specialized modality to enhance the feature space for emotional inference. We utilize a graph network to model information flow, integrating interlocutor-aware contextual information by considering interlocutor dependencies between utterances. Additionally, we employ a dialogue discourse parser to directly model semantic relationships between utterances. Our iterative feedback fusion mechanism explicitly simulates emotional interactions between feature-level and decision-level modalities, improving inference without ground truth labels through iterative refinement. PGIF demonstrates improvements of 1.94% and 1.42% over state-of-the-art methods on the IEMOCAP and MELD datasets, respectively. Ablation studies further validate the effectiveness of PGIF's mechanisms, while its manipulation of input features and global fusion strategies ensures compatibility with existing approaches.