Hybrid deep learning approach for Indonesian hoax detection: a comparative evaluation with IndoBERT

Abstract

The spread of hoaxes in Indonesia has escalated significantly, with over 12,547 cases recorded between 2018 and 2023. Low public literacy and uncontrolled information flow contribute to the rapid dissemination of false content that fuels disinformation and social unrest. Previous studies have utilized artificial intelligence (AI) approaches such as Indonesia bidirectional encoder representations from Transformers (IndoBERT) and deep learning models like long short-term memory (LSTM), bidirectional LSTM (BiLSTM), convolutional neural network (CNN), and Transformer-based methods. However, most relied on a single modeling paradigm and did not address the trade-offs between classification performance and computational efficiency. This study proposes a hybrid architecture that integrates IndoBERT with bidirectional gated recurrent unit (BiGRU) and BiLSTM to enhance Indonesian hoax detection. Using 4,312 news articles and 10-fold cross-validation, we compare the performance of IndoBERT–BiGRU, IndoBERT–BiLSTM, and the proposed hybrid IndoBERT–BiGRU BiLSTM model. Evaluation metrics include accuracy, precision, recall, F1 score, and training time. The hybrid model achieved the best performance with 98.73% accuracy, 99.01% recall, 98.04% precision, and 98.98% F1 score, while also reducing training time compared to single models. These findings demonstrate that combining BiGRU and BiLSTM within the IndoBERT framework effectively balances performance and efficiency, making it a robust solution for Indonesian text classification.