Distance-Aware Hypergraph and Attention Network With Unimodal Assistance for Multimodal Sentiment Analysis

Multimodal sentiment analysis (MSA) has advanced with deep learning, yet some limitations persist. Early methods can capture intra-sample patterns but fail to model inter-sample structural correlations. While Transformers improved inter-modal modeling, their high computational cost and confinement to pairwise relationships prevent them from capturing the complex, group-wise cues inherent in human expression. To address these gaps, we propose a novel Distance-Aware Hypergraph and Attention Network with Unimodal Assistance (DHAN-UA). Our framework's core is a semantic-proximity-aware hypergraph hybrid convolution. We first introduce an inverse-distance weighting strategy to redefine the hypergraph incidence matrix, ensuring a node's influence within a hyperedge is proportional to its semantic closeness to the hyperedge's conceptual center. This approach embeds crucial categorical and semantic priors into the unimodal features. Building upon this, a synergistic hierarchical attention framework first applies intra-modal self-attention to capture long-range dependencies in the topologically-aware features, then uses an inter-modal attention module for efficient weighted fusion. This design achieves both feature refinement and effective fusion. Finally, to enhance model stability, a unimodal-assisted training strategy acts as a regularizer, ensuring the model retains essential information from all modalities. Extensive experiments demonstrate that our model achieves competitive performance on benchmark datasets while maintaining superior computational efficiency.