Leveraging MuRIL for Enhanced Hindi Fake News Detection on a Comprehensive Dataset

Today, with the growing use of the internet, social media platforms have become a vital source of news with many users relying on them for real-time updates. This highlights the need for efficient fake news detection techniques to encourage widespread information dissemination. While a large amount of research has focused on resourceful languages like English, languages like Hindi that have less natural language processing resources remain understudied. This paper addresses this gap by evaluating transformer models: IndicBERT and MuRIL for fake news detection in Hindi on a large-scale dataset containing 286,157 training samples and 71,530 test samples, comparing them against baseline machine learning and deep learning approaches. The experiments reveal that MuRIL outperforms all other approaches, achieving an accuracy and F1 score of 0.98. The findings indicate that MuRIL performs well at detecting Hindi misinformation, establishing it as a promising resource for social-media platforms and further studies in lessresourced languages.