Enhancing Arabic NER with Targeted Data Augmentation and a Specialized BERT Model

Named Entity Recognition (NER) is central to many natural language processing (NLP) applications, yet it remains challenging for low-resource languages such as Arabic. This work addresses Arabic NER on a limited corpus (the NEWS dataset) by applying three targeted data augmentation techniques—synonym replacement, person-name (PERS) tag replacement, and deletion of every third person-name mention—which collectively increased the corpus size by 213%. Subsequently, an Arabic BERT model (asafaya/bert-base-arabic) was fine-tuned on the augmented data with early stopping, achieving a validation F1 score of 0.867 by epoch 5. This approach outperformed baseline methods, demonstrating the effectiveness of systematic data augmentation and the use of a domain-specific pretrained model for Arabic NER.