Detection of Truthful, Semi-Truthful, False and Other News with Arbitrary Topics Using BERT-Based Models

Abstract

Easy and uncontrolled access to the Internet provokes the wide propagation of false information, which freely circulates in the Internet. Researchers usually solve the problem of fake news detection (FND) in the framework of a known topic and binary classification. In this paper we study possibilities of BERT-based models to detect fake news in news flow with unknown topics and four categories: true, semi-true, false and other. The object of consideration is the dataset CheckThat! Lab proposed for the conference CLEF-2022. The subjects of consideration are the models SBERT, RoBERTa, and mBERT. To improve the quality of classification we use two methods: the addition of a known dataset (LIAR), and the combination of several classes (true + semi-true, false + semi-true). The results outperform the existing achievements, although the state-of-the-art in the FND area is still far from practical applications.

Keywords

• fake news detection bert-based models multi-class classification