Learning Deep Fake-News Detectors from Scarcely-Labelled News Corpora

Nowadays, news can be rapidly published and shared through several different channels (e.g., Twitter, Facebook, Instagram, etc.) and reach every person worldwide. However, this information is typically unverified and/or interpreted according to the point of view of the publisher. Consequently, malicious users can leverage these unofficial channels to share misleading or false news to manipulate the opinion of the readers and make fake news viral. In this scenario, early detection of this malicious information is challenging as it requires coping with several issues (e.g., scarcity of labelled data, unbalanced class distribution, and efficient handling of raw data). To address all these issues, in this work, we propose a Semi-Supervised Deep Learning based approach that allows for discovering accurate and effective Fake News Detection models. By embedding a BERT model in a pseudo-labelling procedure, the approach can yield reliable detection models also when a limited number of examples are available. Extensive experimentation on two benchmark datasets demonstrates the quality of the proposed solution.