Analysis of Language Model Role in Improving Machine Translation Accuracy for Extremely Low Resource Languages

Abstract

—Several previous studies have suggested using statistical machine translation instead of neural machine translation for extremely low-resource languages. We could translate texts from 12 different regional languages into Indonesian using machine translation experiments. We increased the accuracy of machine translation for 12 extremely low-resource languages by using several monolingual corpus sizes on the language model’s target side. Since many Indonesian sources are available, we added this corpus to improve the model’s performance. Our study aims to analyze and evaluate the impact of different language models trained on various monolingual corpus on the accuracy of machine translation. The increase in accuracy when enlarging the monolingual corpus is not observed every time, according to our experiments. Therefore, it is necessary to perform several experiments to determine the monolingual corpus to optimize the quality. Experiments showed that Melayu Pontianak achieved the highest bilingual evaluation understudy improvement point. Specifically, we found that by adding a monolingual corpus of 50–100K, they performed a bilingual evaluation understudy improvement point of 2.15, the highest improvement point they reached for any of the twelve languages tested.