Comparative analysis of advanced transformer architectures for online rumor veracity detection

This paper aims to explore the effects of social networking sites and the rapid dissemination of rumors that flood the internet, raising a serious question about the credibility of data, as well as the information credibility and accuracy. This applied to veracity detection in the context of rumor detection requires precise identification and categorization of data that might fall under the “rumor” or “not rumor” category. This paper aims to review popular transformer models that are used for detecting such data. We researched on the improved form of pre-training models comprising Bidirectional Encoder Representations from Transformers (BERT), a Robustly Optimized BERT Pretraining Approach (RoBERTa), and XLNet models. These models are assessed according to the capability to learn identifying features and subtle patterns that separate rumors from the rest of the information. BERT, RoBERTa, and XLNet are found to be impressive solutions in enhancing the accuracy of the veracity detection because they employ bidirectional context comprehension, minor changes in layer configuration and a new language permutation, respectively. Hence, we present a comparative analysis among these popular transformer models to guide researchers in selecting the appropriate model for their work. We have used multiple evaluation metrics like sensitivity, specificity, precision, net present value (NPV), false positive rate (FPR), false negative rate (FNR), false discovery rate (FDR), accuracy, F1 score, and Matthews correlation coefficient to evaluate the performance of these models. Experimental results reveal that the XLNet model shows the highest performance by achieving an accuracy of 96.7% and an F1 score of 96.5%. This shows that the generalized autoregressive (AR) pre-training method of XLNet significantly enhances the performance of the model.