Fake News Detection Using BERT and Bi-LSTM with Grid Search Hyperparameter Optimization

Abstract

Listen

Translate article

Fake news and misinformation disseminated on social media can significantly distort public perception and behav-ior, leading to serious issues. These deceptive contents have the potential to increase societal polarization by caus-ing individuals to make decisions based on false information. During crises, the spread of fake news can endanger public health, destabilize the economy, and undermine trust in democratic institutions. To address this critical issue, numerous studies today employ machine learning and deep learning models. In this study, the transformer architec-ture, widely used in natural language processing, was utilized. To process longer texts more reliably, Bidirectional LSTMs were hybridized with the transformer architecture in the model. For easier detection of fake tweets, the target categories in the dataset were balanced, and the TomekLinks algorithm was employed to enhance classification performance. To improve model performance, a parameter pool was established, and Grid Search was used to identi-fy parameters yielding the most successful results. In our tests, all top 10 models achieved an accuracy of 99%. The highest-performing model achieved an impressive accuracy of 99.908%.