RTBERT: A transformer based approach for improved rumor classification from tweet

Abstract

In an era where both the general public and established news outlets increasingly rely on social media for real-time information, the abundance of rumors offers a huge difficulty. False information can have far-reaching implications, affecting individuals, communities, and even entire countries. To address this issue, a low-cost, self-regulating, and forward-thinking rumor detection technique is required. This research performs an intensive analysis of the performance of three robust machine learning algorithms, including XGBoost, SVM, and Random Forest, as well as two deep learning-based transformers, namely BERT and DistilBert. In this research, the models are trained and evaluated on a combined dataset comprising data from Twitter15 and Twitter16 datasets. Support Vector Machine (SVM) and Random Forest exhibit the best accuracy among classical machine learning models, reaching 89.05%. In comparison, among the transformer-based deep learning models, BERT achieves the best accuracy of 90.20%. In conclusion, the learning-based transformers beat its competitors in terms of accuracy, recall, precision, and F-measure, proving its efficacy in minimizing the detrimental impact of rumors on people, communities, and society as a whole.