Transformer Models for Hausa Personal Name Spelling Correction: A Comparative Evaluation of Byt5, Flan-T5, and AfroLlama

Misspelling of personal names presents a significant challenge to data integrity, record linkage, identity verification, and information retrieval, particularly in low-resource linguistic environments. Hausa personal names exhibit substantial spelling variation arising from phonetic, orthographic, transliteration, and typographical differences, yet limited research has investigated Transformer-based approaches for their automatic correction. This study investigates the effectiveness of Transformer-based language models for automatically correcting misspelt Hausa personal names. A corpus comprising 148,710 spelling instances derived from 2,061 Hausa personal names was used, including 19,472 real human spelling attempts and synthetically generated variants. Three models, ByT5-small, Flan-T5-small, and AfroLlama, were evaluated under a three-phase cumulative curriculum-learning framework in which progressively larger edit-distance ranges were introduced during training. To reduce the possibility of canonical-name-level data leakage, the 2,061 canonical names were partitioned into mutually exclusive training, validation, and test sets containing 1,690, 185, and 186 canonical names, respectively. Model performance was evaluated using exact-match accuracy, character-level precision, recall and F1, SacreBLEU, and character error rate (CER), with McNemar's test used to assess differences in paired prediction outcomes. Fine-tuned ByT5-small achieved the strongest overall performance, correctly restoring 3,328 of 13,558 test instances (24.55%), with character precision of 92.45%, recall of 80.68%, F1 of 86.16%, SacreBLEU of 69.7286, and CER of 0.2048. Fine-tuned Flan-T5-small achieved 11.85% exact-match accuracy and 83.14% character F1. In contrast, AfroLlama declined from 6.84% to 4.46% exact-match accuracy following fine-tuning, accompanied by a reduction in character F1 from 74.93% to 69.70%. McNemar's tests indicated statistically significant differences between each pair of evaluated models (p