Multimodal Biases Mitigation: Multimodal Multi-Granularity Causal Reasoning Framework for Multimodal Fake News Detection

Multimodal fake news detection aims to automatically judge the veracity of multimodal news according to its textual and visual contents. However, existing multimodal fake news detection models unavoidably suffer from unintended biases, such as, textual, visual and entity bias, which usually results in seriously degrading their generalization abilities. Previous studies overlooked biases while learning multimodal global and local fusion, which leads to learn spurious label correlation of multimodal news. Therefore, we propose a model-agnostic debiasing framework, called Multimodal Multi-granularity Causal Reasoning (MMCR), which makes the first attempt to mitigate global and local biases for fake news detection. The MMCR can be readily combined with existing multimodal models and further enhance their detection performance. Specifically, the MMCR mitigates global biases by exploiting a counterfactual multimodal scenario to reverse textual and visual contents for estimating the direct influences of multimodal global contents. Different from previous studies, the MMCR designs an entity causal reasoning strategy by causally removing the direct local biases of the multimodal entities for prediction. The extensive experimental results show that the MMCR could improve the detection performance of multimodal models. Studies also confirm that our MMCR can improve generalization ability of multimodal models.