Conflict-aware commonsense knowledge distillation with hypergraph reasoning for robust multimodal sentiment analysis

Existing graph-structured multimodal sentiment analysis (MSA) methods generally rely on heuristic topology construction and shallow statistical correlation modeling, making them susceptible to spurious semantic cues and modality bias under implicit sarcasm, semantic reversal, and cross-modal conflict scenarios. To address these limitations, this paper proposes CKD-MHR, a robust framework based on conflict-aware commonsense knowledge distillation and population manifold hypergraph reasoning. Specifically, the framework leverages a multimodal large language model (MLLM) to construct a commonsense knowledge elicitation (CKE) mechanism that extracts structured semantic rationales and conflict-aware signals from image–text pairs in a zero-shot manner. The same MLLM-derived semantic supervision jointly guides conflict-aware pathway allocation, rationale-derived intent-prototype hypergraph construction, and rationale-aligned contrastive learning. This coordinated design connects intra-sample interaction with cross-sample high-order semantic propagation while reducing reliance on misleading cross-modal associations. Experimental results show that CKD-MHR improves the F1 score by 2.89% over the strongest baseline on MVSA and achieves the best AUPRC, ACC, and F1 on Hateful Memes, while remaining competitive in MMSD2.0, indicating improved robustness to cross-modal conflicts and misleading cross-modal associations.