AI-Powered Fake News Detection Tool for Nepali Media

The rapid proliferation of misinformation poses significant challenges to public discourse in Nepal, particularly given the prevalence of bilingual content and varied literacy levels. This study introduces a robust, bilingual (Nepali/English) fake-news detection framework that integrates a web application and a browser extension to facilitate on-demand credibility assessment. The system’s backend is implemented in Django, while the frontend leverages React with Vite for efficient rendering. Incoming articles are ingested via URL scraping (Beautiful Soup) or direct text input, then subjected to Unicode-based filtering, custom Nepali stemming, and sequence tokenization. Feature extraction combines TF-IDF vectors (10,000-token vocabulary) and dense embeddings (100-dimensional), which feed an ensemble of five classifiers: a bidirectional LSTM (2 layers, hidden size 256, dropout 0.5), logistic regression, gradient boosting, random forest, and an isolation-forest one-class detector. Model outputs are aggregated by weighted averaging—weights optimized to maximize validation ROC-AUC. Evaluated on a balanced Nepali-English news corpus, the ensemble achieved 91.2 % accuracy, 0.93 precision, 0.92 recall, and 0.96 ROC-AUC, outperforming single-model baselines. User studies (n = 15) confirmed high usability and effectiveness in flagging fabricated content. These results demonstrate the tool’s potential to enhance digital literacy and mitigate the spread of fake news across Nepal’s media ecosystem.