A new multilingual, multimodal Natural Language Processing (NLP) framework has been developed to provide early detection of misinformation and violence-prone dynamics linked to social unrest. The system integrates XLM-RoBERTa for text representation and CLIP for visual embeddings, enhanced by a multi-head attention mechanism and auxiliary features like sarcasm and geospatial metadata.
- A fused dataset containing 138,256 Bangla and English samples was created by combining multiple benchmark datasets.
- Experiments conducted on a stratified 30% subset of the data achieved 98% test accuracy with strong precision and recall.
- The framework utilizes geospatial signals to help anticipate real-world escalation of violence driven by false information.
The outcomes demonstrate the efficacy of multimodal approaches in early misinformation detection, highlighting the specific value of incorporating geospatial data for anticipating real-world harm.