Every lie has a grain of truth: disentangling deception from authentic content for fake news detection

Fake news detection remains a pressing challenge due to the exponential growth of digital content. This paper introduces Semantic Divergence and Alignment Learning (SDAL) method, that explicitly decomposes news into three complementary components: topic-generalizable features capturing common representations within the same news topic, content-specific features isolating manipulative contents, and auxiliary features integrating external knowledge. Through four objective functions of detection, separation, consistency, and reconstruction, our method maximizes the separability between misleading and factual content while preserving topic semantic, a critical aspect often overlooked by existing methods. Empirical evaluations on multiple benchmark datasets demonstrate that SDAL consistently outperforms state-of-the-arts in both detection accuracy and interpretability.