Hierarchical Multi-View Transformer Architecture with Adversarial Regularization for Robust Fake News Detection

Abstract

The ever growing distribution of fake news on internet sites has heightened the need to have robust automated fake news detecting functions. The current paper proposes a different type of architecture known as Hierarchical Multi-View Transformer (HMV-Transformer) that integrates multi-scale semantic representations with DeBERTa-v3-large and Longformer encoders, followed by the mixture of attention features and adversarial regularization.Though transformer architectures such as BERT, RoBERTa, and DeBERTa have demonstrated satisfactory performance in their text classification tasks, certain challenges still remain to be addressed in the framework of long form articles, interrelations between the headline and the body, and resistance The proposed model will comprise of 5-fold stratified cross-validation, cross-validation of ensembles, and SHAP based interpretability analysis. With the Kaggle Fake and Real News dataset, experimental evaluation demonstrates a higher score compared to state-of-the-art transformer baselines with a score of 99.7% and F1-score of 0.997. The significance of improvements of the statistics (p < 0.05) is verified. The proposed framework makes misinformation resilient, extensive and interpretable.