Enhancing Arabic Biomedical Named Entity Recognition Using Transformer-Based Representations and CRF Sequence Labeling

Electronic health records have witnessed tremendous growth in recent years. To make these documents useful for decision-making, high-performance natural language processing (NLP) systems are essential. Named entity recognition (NER) is a critical task for many biomedical NLP applications that contribute to improving patient care, drug discovery, and disease surveillance. However, despite its status as an official language in more than 22 countries, Arabic is largely neglected in this field. Only limited work has been done and there is little well-annotated public dataset. This work tackles these issues by proposing an NER model capable of recognizing entities such as diseases, symptoms, and organs from biomedical Arabic text. To achieve this, an annotated dataset was first developed, followed by fine-tuning the CAMeLBERT model, a BERT-based model, in conjunction with a conditional random field (CRF) layer. The evaluation results indicate that the CAMeLBERT+CRF model achieves the best overall F1-score of 90%, surpassing other base models such as CAMeLBERT and AraBERT. This study demonstrates the effectiveness of the hybrid approach and underscores the importance of transfer learning techniques for low-resourced and morphologically rich languages like Arabic.