Abstract:
The rapid growth of digital communication platforms has intensified the spread of misinformation, posing significant social, political, and economic risks. This challenge is particularly severe in low-resource languages such as Sinhala, where the lack of advanced NLP tools and annotated datasets limits the development of reliable automated fake news detection systems. Addressing this gap, this research proposes an enhanced fake news detection framework tailored for the Sinhala language by integrating a fine-tuned transformer-based model with Explainable Artificial Intelligence (XAI) techniques to promote transparency and trust.
The study employs sinBERT, a pre-trained transformer model optimized for Sinhala linguistic patterns, which is further fine-tuned using a curated dataset of Sinhala news articles labeled across four categories: Credible, Partial, Uncertain, and Wrong. The data undergo rigorous preprocessing, including cleaning, tokenization, class balancing, and augmentation to address noise and imbalance. The model’s performance is evaluated using standard metrics such as accuracy, precision, recall, and F1 score, demonstrating strong predictive capability in identifying misleading news content.
To improve interpretability—an essential factor for user trust—the system incorporates XAI methods, specifically SHAP-based explanations, enabling users to visualize the most influential words that contribute to each prediction. This transparency helps bridge the gap between automated decision-making and human understanding.
Overall, this research contributes to the advancement of Sinhala NLP by providing a robust detection model, a valuable labeled dataset, and an interpretable decision-support framework. It highlights the potential of transformer architectures and XAI in combating misinformation in low-resource linguistic environments and sets a foundation for future improvements and multilingual extensions.
