Evaluation of the effect of back-translation direction on machine translation quality for low-resource sundanese-english language

Machine translation capabilities for widely spoken languages are advanced, but many low-resource languages like Sundanese still lack effective tools, limiting global communication and information access. This study investigates how the direction of back-translation, a common data augmentation technique, influences the performance of two state-of-the-art multilingual models—mT5-Base and NLLB-200-600M—on low-resource Sundanese-English translation tasks. Our methodology involved fine-tuning both models on three datasets: a 3.5k pure baseline corpus, a 5.6k corpus augmented by back-translating Sundanese into English, and a 7k corpus augmented by back-translating English into Sundanese. Experiments, conducted with FP32 precision on Kaggle’s P100 GPU environment, revealed that while mT5-Base showed modest gains (BLEU 0.10 to 13.52 E-S; 0.76 to 23.77 S-E on 7k corpus), NLLB-200 consistently achieved substantially higher quality. Its baseline BLEU scores of 15.36 (E-S) and 21.07 (S-E) significantly improved, peaking at 14.63 (E-S) and 37.04 (S-E) on the 7k dataset, with strong ROUGE-1 scores. Our results demonstrate that back-translation direction and volume critically affect low-resource machine translation: augmenting with synthetic English benefits Sundanese-to-English tasks, while synthetic Sundanese aids English-to-Sundanese translations. Moreover, NLLB-200’s architecture better leverages augmented data, outperforming mT5-Base by up to 16 BLEU points on the largest corpus. These findings highlight the critical role of both data augmentation strategies and robust model design for advancing translation quality for underrepresented language pairs. Future work should focus on continually expanding and diversifying datasets, experimenting with emerging state-of-the-art models, and incorporating deeper linguistic features of Sundanese to achieve further breakthroughs.