Dual-Debias: A counterfactual inference framework for causally robust fact-checking

The rapid dissemination of misinformation on social media has raised significant concerns regarding information credibility and public trust. Fact-checking, which aims to verify the veracity of claims based on evidence, has therefore become an essential task in automatic misinformation detection and analysis. However, most existing fact-checking models rely heavily on spurious associations between claim-evidence and veracity labels rather than genuine causal reasoning. Our empirical analysis reveals that current fact-checking datasets contain substantial distributional biases, where superficial features such as topics or keywords are spuriously associated with specific veracity labels, leading to deceptively high model accuracy. To address this issue, we reformulate fact-checking as a counterfactual inference problem, enabling the estimation of causal effects between claims, evidence, and veracity labels. Based on this perspective, we propose Dual-Debias, a novel counterfactual reasoning framework that introduces a dual-branch architecture with a Jensen–Shannon divergence–based dynamic constraint to adaptively mitigate bias while preserving causal information. Extensive experiments are conducted on four datasets, PolitiHop, symmetric-PolitiHop, CHEF, and our newly constructed symmetric-CHEF, covering both English and Chinese. The symmetric datasets serve as counterfactual benchmarks, constructed to balance distributional bias and eliminate spurious correlations between claim content and veracity labels. Experimental results demonstrate that Dual-Debias consistently outperforms existing baselines across datasets and languages, achieving superior robustness, generalization, and interpretability under distributional shifts.