Enhanced Bert Framework for Information Veracity Analysis

Abstract

The exponential increase of fake information distribution through online platforms also highlights the necessity of acuteness of need the creation of accurate but interpretable fake news detectors tools. The following paper presents a new architecture to the identification of fake news with an optimized BERT model with enhanced attention processes and incorporation of handmade features. The proposed architecture is unique in its usage of a higher order process mechanism that unites the, of the power of deep learning methodology and interpretable algorithmic. Also in use are techniques, as well as the use of a specialized few-shot learning mechanism. The suggested architecture was tested with the help of typical standards, with outstanding level of performance Classification Accuracy 99.92 with all the precision, recall, and the F1 values were found to be greater than 98.5. The proposed architecture was found to be performing well in different fields of content, with a greater than 96 percent accuracy with regard to political, health, and content in entertainment, with 92 dataspare coverage scenarios. Addressing the problem of interpretability of SHAP-based visualization, we have added neural networks. The techniques that give an understanding of the decision-making processes and structural optimizations which minimize inference latency by 65 percent, thus making it easier to deploy to the real world. The system has included justifiable prejudice redressing reductions in prediction dissimilarities of 42, making sure Accuracy of 91 percent against adversarial attacks, thus capturing the field to the following level of credible misinformation management.