Robust Arabic tweet NER via label-aware data augmentation and AraBERTv2

Abstract

Named entity recognition (NER) is vital for turning unstructured social media text into structured information. However, Arabic tweets pose distinct challenges; informality, brevity, dialectal variation, and inconsistent orthography. This study targets those challenges by coupling targeted data augmentation with a transformer model, bert-base-arabertv2. We design a lightweight augmentation pipeline—synonym replacement, name and location replacement, and deletion of third-person Arabic names—to expand linguistic variety and reduce overfitting under limited annotation. The approach is simple, but deliberate: preserve labels when substituting entities with type-consistent alternatives; remove corresponding tags when deleting names; and keep tweet semantics intact where possible. We then fine-tune bert-base-arabertv2 on the combined original and augmented data and evaluate on a held-out set of tweets. The result is a substantial gain in overall performance: F1=0.93 with augmentation versus 0.72 without. These findings indicate that controlled, label-aware augmentation can improve robustness and generalization for Arabic tweet NER, where data scarcity and linguistic variability otherwise degrade accuracy. Beyond empirical gains, our work offers a practical recipe—clear augmentation heuristics and a standard transformer backbone—that can be replicated and adapted to similar low-resource, noisy domains. This contributes to more reliable Arabic social media analysis and downstream information extraction.