Multimodal NLP framework claims 98% accuracy for early detection of misinformation tied to violence

Researchers describe a multilingual, multimodal NLP framework aimed at early detection of misinformation and violence-prone dynamics on social platforms. The authors note that rapid information sharing on platforms like Facebook and WhatsApp has accelerated the spread of fake news, manipulated content, and provocative narratives, which are increasingly linked to social unrest, political instability, and mob violence.

To build the system, the team created a fused dataset of 138,256 Bangla and English samples by combining multiple benchmark datasets. The framework integrates XLM-RoBERTa for multilingual text representation, CLIP for visual embedding, and a multi-head attention mechanism for multimodal fusion. Auxiliary features such as sarcasm detection and geospatial metadata are also incorporated.

In experiments on a stratified 30% subset of the data, the framework achieved 98% test accuracy with strong precision and recall. The authors report that multimodal approaches like this can improve early misinformation detection and that geospatial signals add value for anticipating real-world escalation.

The work has been accepted as a book chapter for publication by Taylor & Francis in 2026.