Towards boosting unlabeled text corpora for Arabic dialect identification

Arabic dialect identification (ADI) aims to automatically determine the specific regional dialect of a given Arabic text. State-of-the-art ADI solutions often rely on fine-tuning Arabic-specific pre-trained language models (PLMs). Although effective, these PLMs are predominantly trained on modern standard Arabic (MSA), which limits their performance on dialectal data. Furthermore, the high degree of similarity among Arabic dialects makes it difficult to learn accurate dialect-specific representations without large volumes of labeled data. However, labeling such data is both costly and labor-intensive, particularly for low-resource languages like Arabic. To address these challenges, we propose a self-training neural approach that independently learns Dialectal Indicators . Our method leverages unlabeled data to construct a matrix that captures dialectal tokens frequently co-occurring in similar contexts. This matrix provides dialect-specific representations, which are integrated with PLM outputs to enhance ADI performance. We evaluate our approach on multiple ADI and related datasets. Results show that our method significantly improves PLM performance over direct fine-tuning, achieving gains of up to 36.2% in accuracy and 11.52% in macro-F1-score.