HC-CoT: A hierarchical causal chain-of-thought framework for multimodal sarcasm detection

Multimodal sarcasm detection relies on identifying semantic incongruity between visual and textual modalities. However, existing methods typically model incongruity through monolithic feature fusion or shallow interactions, neglecting the hierarchical structure of sarcasm, which manifests distinctively at entity, attribute, and scene levels. Consequently, these models often rely on spurious correlations rather than genuine causal dependencies, resulting in limited robustness against modality imbalance and distribution shifts. To address these challenges, we propose the Hierarchical Causal Chain-of-Thought (HC-CoT) framework, where Chain-of-Thought refers to a structured inference trace over the hierarchical latent variables of our H-SCM (rather than LLM-style natural-language rationales), a hierarchical causal reasoning framework that models sarcasm with a three-level Hierarchical Structural Causal Model (H-SCM) whose bottom-up causal structure (Entity  →  Attribute  →  Scene) is learned under explicit sparsity and acyclicity constraints. Over this SCM, HC-CoT performs bidirectional inference: bottom-up evidence aggregation forms scene hypotheses, while top-down contextual refinement re-evaluates lower-level states without introducing reverse causal edges. Training combines supervised learning with (i) missing-modality consistency regularization and (ii) counterfactual augmentation with explicit label policies, improving robustness without relying on heuristic shortcut cues. Extensive experiments on the MMSD and MMSD2.0 benchmarks demonstrate that HC-CoT achieves new state-of-the-art performance, exhibiting significant gains in accuracy, robustness, and interpretability.