Training a Bespoke Grammatical Error Correction Model for Azerbaijani EFL Learners: A Low-Resource NLP Innovation for Educational Enhancement

Résumé

This paper presents a custom-trained grammatical error correction (GEC) system tailored to the specific L1 interference patterns of Azerbaijani English-as-Foreign-Language (EFL) learners. By collecting and annotating 3,000 learner sentences (incorrect-correct pairs) and fine-tuning a modern large language model (LLM), we demonstrate that a localized GEC model outperforms off-the-shelf tools like Grammarly and GPT-3.5. The custom model achieved a precision of 0.78 and F₀.₅ score of 0.74, compared to 0.59 for Grammarly and 0.68 for GPT-3.5 (Table 1). Notably, it corrected errors in articles, prepositions, verb tenses, and subject–verb agreement more accurately (Table 2). These results underscore the impact of L1-specific data on model effectiveness. We discuss implications for EFL pedagogy, local NLP development, and equitable AI, noting that incorporating learners’ native language patterns (L1 transfer) can significantly improve automated feedback. Ethical considerations such as data privacy and algorithmic fairness are addressed. This work illustrates how focused NLP innovation in low-resource settings can create practical AI tools that support language learning more effectively than generic systems.