FC-VIB: fact-checking with variational information bottleneck

The proliferation of false information, such as fake news, political deception, and online rumors, has emerged as a critical societal issue, necessitating automated fact-checking methods to mitigate their negative impact. Existing studies primarily follow a three-step pipeline: document retrieval, evidence sentence selection, and claim verification, with significant attention given to retrieval and multi-evidence reasoning. However, current approaches often suffer from noise propagation and oversmoothing issues in graph neural networks, which undermine their ability to capture diverse perspectives from evidence. To address these challenges, we propose a novel fact verification framework, FC-VIB, grounded in the variational information bottleneck (VIB) theory. FC-VIB introduces an adaptive evidence graph structure learning method that controls the flow of information into the reasoning model, balancing prediction and compression to filter task-irrelevant information. The framework includes four stages: (1) encoding claims and evidence using visual and text encoders to initialize node features; (2) learning the evidence graph structure adaptively to construct the IB-Graph (GIB); (3) utilizing a graph neural network to obtain the evidence graph representation distribution; and (4) sampling from the learned distribution via the reparameterization trick for claim verification. The model is trained with supervised classification and KL-divergence loss for the IB objective. Extensive experiments on a public text-based fact verification benchmark and two multimodal benchmarks demonstrate that FC-VIB effectively constructs high-quality evidence graphs, achieving superior performance compared to state-of-the-art baselines.