Multimodal emotion recognition in conversations via a dynamic emotion-adaptive heterogeneous graph network with CLIP-based visual anchoring

Multimodal Emotion Recognition in Conversations (MERC) has gained increasing attention due to its importance in affective computing and conversational analysis. A key challenge in MERC is modeling dynamic emotional dependencies across dialogue turns, where inter-speaker interactions and intra-speaker temporal evolution exhibit distinct characteristics. To address this challenge, we propose a Dynamic Emotion-Adaptive Heterogeneous Graph Network (DEAHGN) that explicitly decomposes conversational dependencies into inter-speaker and intra-speaker relational subgraphs. Each utterance is represented as a graph node, while emotional and temporal relationships are encoded as edges and modeled using graph attention networks. This structured formulation enables targeted modeling of cross-speaker interactions and within-speaker emotional continuity. Furthermore, we introduce a Contrastive Language–Image Pre-training (CLIP)-based Emotion-Informed Visual Anchoring Mechanism (EIVAM), which constructs emotion-discriminative visual anchors from facial representations and employs probabilistic sampling to enhance cross-modal semantic alignment and robustness. By jointly modeling structured conversational dynamics and emotion-aware visual alignment, DEAHGN provides complementary relational and semantic cues for robust emotion recognition. Experiments on three benchmark datasets demonstrate the effectiveness of the proposed framework, achieving 82.57% W-Acc and 83.14% W-F1 on IEMOCAP, 72.71% W-Acc and 71.26% W-F1 on MELD, and 87.58% Weighted-F1 and 60.96% Micro-F1 on DailyDialog, demonstrating competitive performance across the evaluated benchmarks.