Enhancing Arabic Named Entity Recognition with CAMeLBERT, CRF, and Data Augmentation

Abstract

Arabic Named Entity Recognition (NER) remains challenging due to the language’s complex morphology and limited annotated data. This paper proposes an enhanced NER approach that addresses these challenges by combining targeted data augmentation, a language-specific transformer model, and a sequence optimization layer. We expand a small Arabic NER corpus by over 200% using carefully designed augmentation techniques, then fine-tune a specialized Arabic BERT model (CAMeLBERT) with a Conditional Random Field (CRF) decoding layer. Our approach achieved an F1-score of 86.7% on the validation set, substantially outperforming baseline models. The results demonstrate that strategic augmentation coupled with an Arabic-optimized transformer and CRF yields state-of-the-art performance for Arabic NER. This work provides a detailed technical rationale for each design choice and includes a comparative analysis against other deep learning NER architectures to underscore the effectiveness of the proposed method.