Explainable Fake News Detection via Multi-Aspect Semantic Discovery

The proliferation of misinformation through social networks necessitates sophisticated automated fact-checking systems capable of providing transparent explanations. Current approaches face critical methodological limitations in evidence processing: (1) tendency to equate evidence volume with veracity strength, leading to frequency-based bias; (2) potential oversight of crucial context due to filtering, context window constraints; and (3) oversimplification of complex semantic relationships between evidence. These limitations are particularly pronounced when analyzing complex news items where multiple interrelated aspects influence truthfulness. In this paper, we propose FACETS (explainable FAke news deteCtion via multi-AspEcT Semantic discovery), a novel framework for multi-aspect fact verification that employs semantic discovery techniques to address these limitations. FACETS processes evidence by first employing evidence clustering with unsupervised contrastive learning-enhanced embeddings to organize related evidence by distinct claim aspects. To gain clear and concise insights, a prompt-based module refines each cluster into logical justifications that highlight the most critical aspects of the evidence. Finally, we propose an inference module to determine the claim’s veracity through systematic reasoning over the generated justifications. Extensive experiments on two real-world benchmarks demonstrate that FACETS outperforms state-of-the-art baselines, highlighting the effectiveness of our proposed framework.