Improving Multiclass Classification of Fake News Using BERT-Based Models and ChatGPT-Augmented Data

Given the widespread accessibility of content creation and sharing, false information proliferation is a growing concern. Researchers typically tackle fake news detection (FND) in specific topics using binary classification. Our study addresses a more practical FND scenario, analyzing a corpus with unknown topics through multiclass classification, encompassing true, false, partially false, and other categories. Our contribution involves: (1) exploring three BERT-based models—SBERT, RoBERTa, and mBERT; (2) enhancing results via ChatGPT-generated artificial data for class balance; and (3) improving outcomes using a two-step binary classification procedure. Our focus is on the CheckThat! Lab dataset from CLEF-2022. Our experimental results demonstrate a superior performance compared to existing achievements but FND’s practical use needs improvement within the current state-of-the-art.