Detecting Misinformation on Telegram Anti-vaccine Communities

Due to the substantial volume of misinformation regarding COVID-19 in Brazil, this paper proposes the application of machine learning methods to identify false or harmful information in anti-vaccine communities on Telegram. To this end, we developed a dataset of 1,500 messages labeled by experts according to three aspects: veracity level, potential for harm, and category. The labeling process achieved an agreement score of 81%. Experiments were conducted using state-of-the-art algorithms such as XGBoost, a BERT-based classifier, and a GPT-based classifier. The models trained on the labeled dataset achieved an F1-Score of 0.83 for detecting falsehood and 0.92 for potential harm, indicating their effectiveness in identifying misinformation in this context.