How to Build a Neural Machine Translation System for a Low-Resource Language

A neural machine translation system was developed for Dongxiang, a low-resource language, by fine-tuning Meta's NLLB-200 model. The project utilized a bilingual dataset of over 42,000 Dongxiang-Chinese sentence pairs to train two separate, direction-specific models. The methodology covers data preprocessing, tokenizer evaluation, and training with an Adafactor optimizer on an A100 GPU. This guide provides a reproducible pipeline, including code and models available on GitHub and Hugging Face, to support translation for languages not covered by mainstream systems.