10.46243/jst.2022.v7.i09.pp65-82 registered

The widespread circulation of false information via online platforms is a growing cause for alarm because of the havoc it may wreak. Several machine learning strategies have been proposed for spotting hoaxes. However, the vast majority of them concentrated on a certain category of news (like politics), raising the issue of dataset bias in the used models. Here, we provide the results of a benchmark study that compares three datasets to determine which machine learning technique performs best. To the best of our knowledge, we are the first to investigate and evaluate the performance of many state-of-the- art pre-trained language models for false news detection, alongside the performance of classical and deep learning models. When it comes to detecting false news, we discover that BERT and other comparable pre-trained models perform the best, even when working with a tiny dataset. Because of this, these models are a much superior choice for languages with few electronic contents (i.e., training data). Additionally, we analyzed the models' efficacy, article topics, and article lengths, and shared our findings and insights. We hope that our benchmark study will encourage additional investigation in the field of false news identification and enable news sites and blogs choose the most effective approach.