RAFAH: RoBERTa Fine-Tuned Model and a Hybrid Features Set for Fake News Detection

The problem of fake information is increasingly prevalent on multiple internet sources, where it is spread in different forms including text, images, audio, and video. Many people use social media platforms like Facebook, Twitter, and Instagram for news, which makes them susceptible to misinformation. While efforts are being made to detect online fake news automatically through the machine and deep learning techniques, further research is needed on detecting fake news using transformer-based architectures. To address this gap, we propose the RoBERTa-based fake news detection model with hybrid features - RAFAH. First, we extract a hybrid feature set (emotional, linguistic, psycholinguistic, and readability) from the datasets. Later, we optimize and fine-tune RoBERTa model parameters to identify online fake news. We compare the performance of our proposed methodology using a variety of machine and deep learning techniques. Our comparison reveals that our proposed methodology outperforms all models on two benchmark datasets (BuzzFeed and ISOT). Compared to state-of-the-art models, the proposed methodology attains an accuracy of 0.89 on the BuzzFeed dataset and 0.99 on the ISOT dataset achieving an improvement of 1% and 10% on the ISOT and BuzzFeed dataset, respectively.