Bangla Counterfeit News Identification: Using the Power of BERT

Many individuals in today's digital age rely on the internet for news consumption. However, on occasion, these platforms distribute false information, which can have serious negative effects on both society and consumers. Fake news can be combated with tools, but most of them are only available to English speakers, which leaves millions of Bangla speakers out in the cold. This research highlights the potential of using an advanced transformer model and how to leverage YouTube as an easily accessible data warehouse to gather a diverse range of Bangla news content to enhance fake news detection for Bangla speakers. Our created dataset includes 1913 fake and 1995 authentic news titles, collected from fact-checking websites and YouTube. We evaluate several machine learning models for their ability to detect fake news using TF-IDF and BoW feature extraction techniques, as well as transformer models, including Bangla-bert-base and mBERT. Our best-performing model, Bangla-bert-base, achieved an accuracy of 84.19% and an F1 score of 85.03%.