Deep Stylometric-Semantic Fusion Network for Robust Fake News Detection

The pervasive spread of fake news poses a significant global challenge, undermining public trust. While traditional detection methods and advanced large language models show promise, they often miss subtle non-semantic features like writing style, emotional tone, and lexical choices. This paper introduces the Deep Stylometric-Semantic Fusion Network (DSSFN), a novel end-to-end framework for enhanced fake news detection. DSSFN deeply integrates powerful semantic representations, from models like RoBERTa-large, with an extensive set of advanced stylometric features encompassing diverse linguistic dimensions. A core innovation is its hierarchical multi-modal fusion module, based on a Transformer architecture with cross-attention layers. This module facilitates iterative, context-aware interaction between modalities, yielding a comprehensive enhanced representation. Evaluated on the widely recognized WELFake news dataset, DSSFN achieves state-of-the-art performance, outperforming strong baselines. Experiments validate the critical contributions of both stylometric features and the fusion mechanism. Interpretability analyses and future research directions are also discussed.