Bilingual fake-news detection in low-resource media: A Transformer-based framework for Nepali–English content

Abstract

The spread of misinformation in Nepal, especially across the bilingual Nepali–English media landscape, represents a significant threat to informed public discourse. Existing solutions for low-resource languages often rely on traditional machine learning or simple recurrent neural networks, which struggle with the morphological complexity of Nepali and the semantic nuances of code-switched content. This paper presents a robust, production-ready framework for fake news detection that transitions from traditional ensembles to a state-of-the-art transformer-based architecture. We fine-tune XLM-RoBERTa (XLM-R), a multilingual model optimized for cross-lingual transfer, on a newly curated and balanced corpus of 16,000 articles (8,000 real, 8,000 fake). Unlike “black-box” approaches, we integrate SHAP (SHapley Additive exPlanations) to provide word-level interpretability, allowing users to understand why an article is flagged. The model achieves superior performance with an accuracy of 99.53% and an F1-score of 0.995, significantly outperforming baseline Bi-LSTM ensembles. The system is deployed as a scalable web application and browser extension using a microservices architecture (Django/React) backed by PostgreSQL, ensuring high concurrency and data persistence. We further address the ethical implications of automated detection by implementing strict privacy protocols and bias mitigation strategies for sensitive political content.