Fake news proliferation on social media platforms threatens public discourse, democratic processes, and public health. Traditional machine learning approaches fail to capture deep contextual semantics and sequential deception patterns. This paper proposes HybridBERT-LSTM-Attention, a novel architecture that synergistically combines BERT’s bidirectional contextual embeddings, LSTM’s temporal sequence modelling, and an Attention mechanism for interpretability. Evaluated on LIAR and FakeNewsNet datasets, the proposed model achieves 98.2% accuracy and 0.98 F1-score, outperforming standalone BERT by 3.1% and Bi-LSTM by 7.4%. The attention layer provides crucial explainability by visualising words that contribute most to the” fake” classification, addressing the black-box problem in content moderation. Furthermore, we employ focal loss to mitigate class imbalance and DistilBERT for computational efficiency. Extensive ablation studies confirm that all three components are essential. Results demonstrate that modelling both semantics and narrative structure is key to detecting modern fake news.
Fake News Detection, BERT, LSTM, Attention Mechanism, Natural Language Processing, Deep Learning.
Shivam Kumar, Dr. Harsh Lohiya. Using Deep Learning Models to Detect Fake News: An Innovative Hybrid BERT-LSTM-Attention Method. Indian Journal of Modern Research and Reviews. 2026; 4(6):189-192
Download PDF
