Beyond Transformers: Introducing a Transformer-Mamba Hybrid for Multiclass Misinformation Quantification

The spread of misinformation across digital platforms poses significant challenges to public trust, governance, and decision-making. Although deep learning models, particularly transformers, have advanced misinformation detection, most existing methods are still limited to binary or coarse-grained classification. In this work, we introduce a unified framework that reframes misinformation quantification as a multiclass relation extraction (RE) task across diverse domains. The framework combines semi-supervised learning to reduce annotation effort with topic modeling to uncover fine-grained, domain-specific categories of misinformation, thereby supporting improved interpretability and adaptability. Furthermore, we present the first evaluation of Mamba, a state-space model, for multiclass RE, demonstrating its potential as a scalable alternative to transformers. To further enhance performance, we propose a Hybrid (DeBERTa + Mamba) model that integrates the contextual strengths of transformers with the sequential efficiency of Mamba, resulting in higher accuracy and F-scores across all domains. Overall, our framework combines scalability with domain-level interpretability to quantify fine-grained misinformation.