Cross-Modal Probabilistic Representation Learning for Conversational Emotion Recognition

Multimodal Emotion Recognition in Conversations (ERC) targets the identification of emotions across textual, acoustic, and visual modalities. However, most existing approaches rely on deterministic representations. This limitation is critical, as psychological studies increasingly reveal human emotional expression to be an inherently probabilistic process, sensitive to the statistical distribution of cues. To address this issue, we propose a Cross-Modal Probabilistic Representation (CPR) framework that models emotional states as multivariate Gaussian distributions. This approach explicitly captures both the central tendency (mean) and the uncertainty (covariance) of emotional cues across modalities. This probabilistic paradigm, however, requires a specialized alignment strategy. We introduce a novel Bifocal Probabilistic Alignment objective that synergistically aligns distribution centers using contrastive learning while calibrating their uncertainty scales via a consistency regularization. Finally, an uncertainty-aware graph fusion mechanism leverages these aligned distributions to dynamically weight information based on modality-specific confidence, leading to more robust fusion. Extensive experiments on the benchmark datasets show that CPR significantly outperforms state-of-the-art methods, improving the weighted F1 score by 3.93-10.16%. Ablation studies confirm the efficacy of each proposed component, particularly in resolving ambiguous emotions. Code is released here.