Context-Aware Fake News Classification with BERT-LSTM and Bahdanau Attention

Abstract

The rapid growth of the Internet and the World Wide Web has transformed the way information is created, shared, and consumed, while the unchecked spread of fake news has emerged as a significant global challenge with social, political, and economic consequences. The difficulty of automatically distinguishing misleading content from legitimate news is compounded by the limitations of conventional approaches in capturing contextual and sequential information from textual data. Motivated by the need for more effective and reliable automated fake news detection, this study presents a comparative evaluation of deep learning approaches using the ISOT Fake News Dataset. Four models were developed and evaluated under the same experimental setting: a Long Short-Term Memory (LSTM) network, a Bidirectional Encoder Representations from Transformers (BERT) model, a hybrid BERT-LSTM model, and a hybrid BERT-LSTM model with Bahdanau Attention. Model performance was assessed using accuracy, precision, recall, and F1-score. The proposed Hybrid BERT-LSTM with Bahdanau Attention achieved the best overall performance, with an accuracy of 99.49%, precision of 99.65%, recall of 99.27%, and F1-score of 99.46%. The model also reduced the total number of classification errors from 84 for the Hybrid BERT-LSTM baseline to 46, including a reduction in fake-news instances incorrectly classified as real from 47 to 15. These results indicate that contextual representations from BERT, sequential dependency modelling through LSTM, and adaptive feature weighting through Bahdanau Attention provide complementary capabilities for fake news classification. The proposed architecture also outperformed the previously published standalone BiLSTM model and several recent benchmark approaches evaluated on the ISOT dataset. However, the present results establish predictive effectiveness rather than computational or scalability superiority, as dedicated computational profiling and repeated-run statistical testing were not conducted. Future work will focus on multilingual and multimodal fake news detection, explainable artificial intelligence, computational benchmarking, and evaluation on Nigerian-specific misinformation datasets.