Comparative analysis of deep learning and transfer learning techniques for dialect identification in the Gulf and Saudi regions

Purpose This paper investigates the automatic identification of Arabic dialects within the Gulf Cooperation Council (GCC) countries and regional dialects across Saudi Arabia. Dialect identification (DI) is critical for enhancing natural language processing (NLP) applications in the Arab world, where significant dialectal variations exist. Design/methodology/approach Using two newly constructed Twitter corpora, GAC-6 (˜ 1.7M tweets across six Gulf dialects) and SDC-5 (˜ 790K tweets across five Saudi dialects), we develop a multi-DI framework that leverages deep learning and transfer learning. We train long short-term memory (LSTM), BiLSTM and a fine-tuned Arabic bidirectional encoder representations from transformers (AraBERT) model and benchmark them under a single, fair protocol against strong classical baselines – logistic regression, linear support vector machine, multinomial Naïve Bayes, Random Forest and a soft-voting ensemble – trained on character n-gram (2–5) and word n-gram (1–3) features using bag-of-words/ term frequency and inverse document frequency (TF-IDF) representations. Findings Transformer-based transfer learning yields the best performance: AraBERT achieved superior performance, with macro-F1 score of 90.1% on GAC-6 and 90.6% on SDC-5. This exceeded the strongest non-transformer baselines by +2.5 macro-F1 (vs. the ensemble) and +1.2 (vs. BiLSTM) on the respective datasets and by wider margins against other classical models. To assess robustness under domain shift, we also report results on public benchmarks (NADI, QADI), where the transformer remains state-of-the-art among the compared systems. Originality/value The paper introduces a novel approach by focusing on Arabic dialects in a specific region and using advanced NLP techniques for dialect differentiation. It is the first to use the large, specialized datasets GAC-6 and SDC-5 for DI.