Detecting Disinformation: Enriching the COVAX Reality Dataset with BERT for Stance Detection

Abstract

Social media news pages and online news portals have significantly grown in popularity among the users across the globe for accessing the latest and important news. Unlike traditional newspapers, online news portals and social media provide the news round the clock and can be accessed from any location free of cost. However, the internet sources, especially the social media, have also become a dangerous tool for spreading misinformation and fake news causing severe damages to the life and property in the society. It has become imperative to find ways to distinguish authentic news from the misinformation on the internet to avoid social unrest and negative consequences. The current study makes the stance detection in news possible by extending the state-of-the-art COVAX reality dataset through the addition of two key features of a news body and a news summary. In addition, our study contributes a novel method comprising three different text summarization techniques (LEAD-3, sequence-to-sequence (Seq2Seq), and Bidirectional Encoder Representations from Transformers (BERT)) to authenticate the online news content. The BERT model consistently outperformed the other two models by achieving the best precision value (89.57%), Recall (91.84%), F-measure (90.64%) and BERT score (0.833). Our findings also prove the feasibility of using the transformer-based model, ELECTRA-Large, to combat the spread of fake news with 89.07% accuracy.