Health Misinformation Detection Using BERT and RoBERTa: A Comparative Study of Machine Learning and Deep Learning Approaches

Abstract

This expansion of fake news over the internet platforms is currently a significant challenge in the digital age that has led to misinformation in the political arena, society, and in the health sector. Deep learning and machine learning methods of automatic detection of fake news have limited research interest over the past few years. In this research paper, a comparative analysis of six classification models is presented to identify a fake news and they are Logistic Regression, Random Forest Tree, Support Vector Machine (SVM), Bidirectional Long Short-Term Memory (BiLSTM), Robustly Optimized BERT Pretraining Approach (RoBERTa) and Bidirectional Encoder Representations from Transformers (BERT).  The Fake news and Real news Dataset (44, 898 news articles) is experimented on. Preprocessing functions are applied to the dataset thus lowering the alphabets, stop words removal, punctuation removal and the URL filtering; after which the classification process is done using the dataset. All the six models are trained and tested in the same experimental conditions and with performance measures of accuracy, recall, F1-score and precision. The experimental results indicate that the BERT model achieves the highest classification rate of 99.95 percent and precision, F1-score and recall of 1.00, which is higher than other models. RoBERTa and Random Forest score 99.80 then SVM at 99.52, then BiLSTM at 99.00 and lastly, Logistic Regression at 98.95. The comparative analysis reveals that the transformer models are superior to machine learning and deep learning models in detecting fake news, and can be applied in future investigation of detecting misinformation.