Enhancing Misinformation Detection on Twitter with a Content-Based Multi-Lingual Bert Model

The rapid spread of misinformation during global crises like COVID-19 has severely impacted public health, governance, and social trust. Social media platforms such as Twitter have amplified this issue, underscoring the urgent need for multilingual, real-time misinformation detection. The proposed Content-based Attention Multi-lingual BERT (CA-BERT) model addresses this challenge by enhancing the standard BERT framework with a content-based attention mechanism that assigns adaptive weights to semantically important tokens often linked to false or misleading content. This attention enables deeper contextual understanding of misinformation cues across diverse linguistic contexts. Using the LIME interpretability method, CA-BERT provides transparent explanations of its predictions, supporting accountable decision-making for policymakers and content moderators. Leveraging multilingual BERT (mBERT) allows the model to handle multiple languages simultaneously, ensuring robust cross-lingual applicability. Evaluations using a balanced multilingual tweet dataset on COVID-19 topics demonstrate that CA-BERT outperforms baseline models such as RoBERTa, DANN, and HANN, achieving 96% recall for true information and 95% for misinformation in English, with F1 Scores of 93% and 92%, respectively. The model maintains strong cross-lingual generalization, especially for Dutch (75% F1) and Spanish (72% F1), with slightly lower performance for Arabic due to tokenization and dialectal complexity. These results highlight CA-BERT’s adaptability while underscoring the need for improved handling of low-resource, morphologically rich languages. Future work involves region-specific preprocessing, cross-lingual transfer learning, and multimodal misinformation detection, aiming to transform CA-BERT into a core component of multilingual real-time disinformation monitoring systems.