Fine-Tuning Multilingual Khmer Neural Machine Translation

Abstract

Google Translate remains a strong baseline machine translation (MT) tool for Khmer. However, as a proprietary tool, it does not allow flexible deployment, customization, or improvement. In contrast, “No Language Left Behind” (NLLB) is an open-source MT solution, but its translation performance for Khmer is significantly weaker than that of Google Translate. Given the low-resource nature of the Khmer language, this paper pragmatically presents a robust machine translation model for translating Khmer to and from English, Thai, Vietnamese, and Laotian. This model is developed by fine-tuning a base NLLB model on a high-quality multilingual parallel corpus. The fine-tuned model achieves performance competitive to Google Translate while significantly outperforming the base NLLB model and the previous studies.