CTrL-FND: content-based transfer learning approach for fake news detection on social media

Online social network platforms are utilized efficiently by massive users to read and disseminate the news in the form of text, image, audio and video. So, it is necessary to validate the genuineness of the news at an initial stage to avoid spreading fake news. Many existing works focused on textual content, they employed a pretrained word embedding and language models to capture the semantic and contextual information, respectively, for fake news identification. Though the existing text-based models achieve better predictions, still it has some limitations as follows: lacuna in extracting the efficient context-based features, pretrained on smaller corpus and static-masking utilization. To address this, we propose a C ontent-based Tr ansfer L earning framework for F ake N ews D etection (CTrL-FND) which contains a word embedding block (WEB) and a classification block (CLB). In WEB, a transfer learning pretrained model, named RoBERTa, is employed for efficient context-based word representation since it is pretrained on larger corpus, eliminates the next sentence prediction loss and incorporates a dynamic masking pattern. The enriched contextual feature vector of WEB is passed as an input to the CLB block, which has a feed forward neural network to classify the news article into fake or legitimate. The proposed model has been evaluated using two standard datasets namely Politifact and Gossipcop, achieved an accuracy of 92.77% and 91.78%, respectively. Experimental results exhibit that the CTrL-FND model outperforms the other state-of-the-art (SoTA) techniques, especially achieved an average accuracy of 10.49% and 14.53% improvements compared to the SoTA methods on Politifact and Gossipcop, respectively.