Abstract

Arabic Text Classification (ATC) remains challenging due to the Arabic language’s morphological richness and semantic complexity. This paper proposes ABPC-Net, a hybrid framework integrating a frozen Arabic Transformer encoder, a Bidirectional LSTM, parallel multi-scale CNN branches, and a lightweight capsule-inspired vector projection head for hierarchical feature integration. Evaluated on the SANAD dataset and its subsets (AlArabiya, AlKhaleej, and Akhbarona) over five independent runs, ABPC-Net achieves mean accuracies of 97.00±0.04%, 99.14±0.10%, 98.40±0.10%, and 95.59±0.12%, respectively. Under identical experimental conditions, the proposed framework consistently outperforms re-implemented frozen and fully fine-tuned AraBERT and MARBERT baselines. Cross-dataset evaluation on BBC Arabic and CNN Arabic further provides evidence of intra-domain transferability and rapid few-shot adaptability across Arabic news sources. The reported results are scoped to Modern Standard Arabic news classification.

Key words: Arabic Text Classification, Deep Learning, Transformer Models, Capsule Networks, Natural Language Processing (NLP).