Intelligent Arabic News Classification Systems Using AraBERT Transformer for Digital Media Engineering

Abstract

The rapid expansion of Arabic digital news has created a pressing need for accurate and scalable automatic news categorization. Arabic natural language processing remains challenging because of the morphological richness of the language, its complex syntax, the prevalence of dialectal variation, and the near-universal absence of diacritics in online text. This paper proposes a transformer-based framework for Arabic news classification centered on fine-tuned AraBERT, a bidirectional encoder pre-trained exclusively on large-scale Arabic corpora. The framework incorporates Arabic-specific text preprocessing, subword tokenization via the AraBERT tokenizer, and a single fully connected softmax classifier appended to the contextual [CLS] representation. Experiments are conducted on the SANAD benchmark dataset, which contains approximately 194,797 Modern Standard Arabic news articles distributed across seven topical categories. The proposed model achieves an accuracy of 98.4%, a macro-averaged precision of 99.1%, a macro-averaged recall of 99.8%, and a macro-averaged F1-score of 99.0%, outperforming fine-tuned multilingual baselines mBERT and XLM-R by substantial margins. Detailed error analysis via confusion matrix and per-class classification reports confirms strong generalization across all categories, with only minor confusion between thematically adjacent domains such as Politics and Finance. The results validate that Arabic-focused pre-training is decisive for high-quality Arabic news categorization and establish a reproducible, scalable pipeline for future research.