Hausa Fake News Detection Using Lightweight Transformer Models with Adaptive Fine-Tuning in Low-Resource Settings

The rapid growth of digital communication technologies and online news platforms has enhanced global information sharing but also accelerated the spread of fake news, leading to political polarization, reduced public trust, and social instability. Despite the success of transformer-based models in high-resource languages, their application to low-resource languages such as Hausa remains limited due to insufficient annotated data and computational constraints. This study aims to address this gap by developing a labelled dataset of 12,601 Hausa news articles and evaluating lightweight transformer models for fake news detection. A quantitative methodology was adopted, incorporating cross-lingual transfer learning to fine-tune three models: DistilBERT, MobileBERT, and MiniLM. Performance was assessed using Accuracy, Precision, Recall, F1-score, and ROC-AUC. The results indicate strong performance across all models, with DistilBERT achieving 95.12% accuracy (AUC 0.9786), MobileBERT 95.87% accuracy (AUC 0.9864), and MiniLM outperforming others with 96.49% accuracy (AUC 0.9923). The findings demonstrate that lightweight transformer models are effective for Hausa fake news detection in resource-constrained environments. It is recommended that MiniLM be adopted for practical deployment, while future research should explore larger datasets and hybrid architectures to further improve performance.