An Enhanced Hybrid BERT-BiLSTM Learning Model for Arabic News Classification
Roua A. Abou Khachfeh, Islam Elkabani, Ziad Ahmad Osman · 2025
Text classification is one of the core tasks in natural language processing. Its goal is to classify textual input into predefined groups. Three deep learning architecture, convolutional neural networks (CNN), bidirectional gated recurrent units (BiGRU), and bidirectional long short-term memory (BiLSTM) are examined in this study for classifying Arabic news. We evaluate models for their performance on both small and big sizes of the Arabic news dataset previously collected from news portals in order to assess how well these models can cope with volume variation. Furthermore, for improving the performance of classification, a hybrid approach combining the best of the BiLSTM model with those of BERT, especially AraBERT, is used. Our hybrid model achieves an F1 score of 95%, outperforming existing state-of-the-art (SOTA) methods for Arabic news classification.