Generative AI for Low-Resource NLP: Evaluating Khmer Data Generation for Homophone Spelling Correction
Résumé
This study demonstrates that generative AI can transform data scarcity from a persistent obstacle into a powerful opportunity for advancing NLP in low-resource languages, with Khmer as the primary focus. The shortage of high-quality, domain-specific datasets, compounded by Khmer's complex syntax and abundance of homophones, has long limited the development of effective language technologies. To address this, we evaluated five state-of-the-art generative AI models: Qwen3-235B-A22B, ChatGPT's GPT-4o, Claude Sonnet 4, DeepSeek-V3 (R1), and Gemini 2.5 Pro, on their ability to produce high-quality Khmer text for training Sor-Ser, a homophone spelling correction system. Our evaluation measured accuracy, diversity, grammatical precision, semantic relevance, and redundancy. GPT-4o consistently emerged as the top performer, generating diverse, syntactically precise, and contextually accurate sentences with minimal redundancy, making it the most suitable for this task. Claude Sonnet 4 and Qwen3-235B-A22B also showed strengths in specific dimensions, suggesting their potential in targeted applications. These results provide strong evidence that generative AI can produce scalable, high-quality datasets tailored to linguistic and domain-specific needs. By turning data scarcity into opportunity, this work lays the groundwork for bridging the gap between high-and low-resource languages, driving more inclusive, equitable, and sustainable NLP development.
Domaines
| Origine | Fichiers produits par l'(les) auteur(s) |
|---|---|
| Licence |