From Scarcity to Sufficiency: Machine Translation Techniques for Low-Resource LLMs Enhancement

This study addresses the significant performance disparities in large language models between high-resource languages like English and low-resource languages such as Thai, particularly in chat performance and attack defense. We introduce MT-LLE (Machine Translation for Low-resource LLMs Enhancement), a novel three-step approach leveraging machine translation techniques to transfer knowledge from resource-rich to resource-poor languages. The method comprises machine translation continual pre-training (MT-CPT), machine translation supervised fine-tuning (MT-SFT), and machine translation chain-of-thought (MT-CoT). Experimental results demonstrate that MT-LLE significantly enhances single and multi-round chat performance, attack performance, localization understanding, and even machine translation capabilities in low-resource languages.