Fake News Prediction in Hindi and English Robust Transformer Model

The ever-growing consumption of information from digital platforms such as social media has paved the way for information dissemination like never before, but not necessarily all the information or news delivered to us is authentic or holds true to reality. A huge amount of information is created and shared by social media users, some of which are manipulative and have no influence on reality, and can cause harm in a variety of ways. Implementation of an automated or AI-based system to classify text articles as misinformation is a challenging task as identifying the truth is complicated. An AI-based fake news prediction system is proposed to classify fake news from real ones by understanding the context of news and making decisions accordingly. To identify fake news by examining the context of the news, a Bi-directional Encoder Representations from Transformers (BERT) model, a pretrained model with a more potent feature extractor, has been implemented. BERT is well suited for tasks like predicting fake news because of its extensive contextualizing capabilities. The novelty of this paper is that the proposed system is trained, tested, and evaluated in 2 languages, English and Hindi to develop an overall reliable system to predict fake news in both languages. Two models of BERT namely DeBERTa and XLM-RoBERTa are chosen for the prediction of fake news. Further improvement in the model is achieved by using the additional news data to pre-train this model. Also, the model is fine-tuned to fit the requirements and to get the optimum result possible. The performance comparison of the chosen models is made and it is observed that the XLM-RoBERTa has improved accuracy over DeBERTa while testing for both languages.