Gated Transformer with Dual Attention for Rumour Category Detection on Social Platforms

The spread of misinformation and unverified content on social media has become a major concern in the digital age. Accurate categorization of rumours—into support, denial, query, or comment—is essential to assess the credibility and impact of online discourse. In this work, we propose a simplified yet effective deep learning framework that combines a Transformer-based encoder with a dual attention mechanism. The model captures fine-grained word-level semantics and post-level contextual relevance within conversation threads. By integrating gated recurrent units (GRUs) with word and post-level attention, the framework enhances the ability to distinguish rumour types based on linguistic and discourse cues. Experimental evaluations on two benchmark datasets, PHEME and RumourEval-19, demonstrate strong performance in terms of accuracy and F1-score. Furthermore, attention visualizations provide interpretability, making the model’s predictions more transparent and trustworthy.