Dynamic temporal modeling and distribution-calibrated fusion for multimodal sentiment analysis

Existing multimodal sentiment analysis (MSA) methods still lack dynamic modeling of unstable emotional information. Emotional expressions in audio and visual modalities often exhibit different temporal rhythms, while most existing recurrent state update mechanisms propagate hidden states at fixed temporal progression scales. In addition, some cross-modal associations merely reflect statistical co-occurrence patterns, and such incidental correlations do not always correspond to genuine emotional dependencies. Moreover, the reliability and contribution of each modality may dynamically vary across real-world scenarios. To address these issues, this paper proposes a multimodal sentiment analysis framework based on dynamic temporal modeling and distribution-calibrated fusion. Specifically, adaptive temporal modeling is performed on audio and visual sequences through dynamic temporal granularity selection, enabling the model to capture both short-term emotional fluctuations and long-range emotional dependencies while estimating temporal confidence. Furthermore, Gaussian distribution modeling and perturbation consistency constraints are introduced to characterize uncertainty in modality representations, thereby preventing over-reliance on unreliable associations during fusion. Temporal confidence and distributional uncertainty are further combined into a unified modality confidence, which dynamically regulates cross-modal fusion and representation alignment to improve semantic consistency of fused representations. Experimental results on the CMU-MOSI, CMU-MOSEI, and CH-SIMS datasets demonstrate the effectiveness of the proposed framework.