Roberta-Based Framework for Fake News Detection using Multi-Source Text and Visual Analytic

The Misinformation erodes public trust and harms decision-making across domains such as health and politics. This research presents a RoBERTa-based fake-news detection framework trained on a hybrid corpus that combines fact-checked articles and misinformation-labelled short texts. The pipeline standardizes text inputs, tokenizes with Byte-Pair Encoding, and fine-tunes RoBERTa using class balancing and label smoothing to mitigate overfitting on minority classes. It implements a lightweight inference service and a user interface that explains predictions with confidence, confusion-matrix drill-downs, and precision, recall diagnostics. In comparative evaluations against LSTM, BERT, XLNet, and DistilBERT, the proposed system achieves 94.2% accuracy, with precision, recall, and F1 of 94.0%, 93.8%, and 93.9%, respectively, on a held-out test set. To ensure reproducibility, it reports dataset composition, train/validation/test splits, random seeds, and hardware settings. The results indicate that (i) RoBERTa provides consistent gains over non-contextual baselines, (ii) balanced sampling improves minority-class recall, and (iii) simple explanation views aid practitioner interpretation without degrading throughput. These findings support the use of transformer encoders with careful training protocol and transparent reporting for reliable fake-news detection.