Saudi Arabian Multi-Dialects Identification Using Ensemble Transformer-based Model

Dialect identification is a critical task in natural language processing (NLP), particularly for linguistically rich and diverse languages such as Arabic. The complexity of the identification of Arabic dialects is amplified by the high degree of regional variation, with Saudi dialects distinguished by their internal diversity and distinctive linguistic characteristics. In this study, we investigate the effectiveness of several transformer-based models: SaudiBERT, AraBERTv2, CamemBERT-da, MARBERTv2, and ARBERTv2–through individual fine-tuning and ensemble techniques for the identification of Saudi Arabian dialects. We further propose an ensemble approach that combines these models using both hard voting and soft voting methods to enhance performance. Experimental results demonstrate that the soft voting ensemble achieves superior results compared to individual models, reaching an accuracy of 76.6%, weighted F1 score of 73.5%, precision of 75.5%, and recall of 76.6%. These findings underscore the benefits of model ensembling for improving performance in the Saudi Arabian multi-dialect identification task.