Analysis of Back-Translation Methods for Low-Resource Neural Machine Translation

Back translation refers to the method of using machine translation to automatically translate target language monolingual data into source language data, which is a commonly used data augmentation method in machine translation tasks. Previous researchers’ works on back translation only focus on rich resource languages, while ignoring the low resource language with different quality. In this paper, we compare various monolingual selection methods, different model performance, pseudo-data and parallel corpus ratios, and different data generation methods for the validity of pseudo-data in machine translation tasks. Experiments on Lithuanian and Gujarati, two low-resource languages have shown that increasing the distribution of low-frequency words and increasing data diversity are more effective for models with sufficient training, while the results of insufficient models are opposite. In this paper, different back-translation strategies are used for different languages, and compared with common back-translation methods in WMT news tasks of two languages, and the effectiveness of the strategies is verified by experiments. At the same time, we find that combined back-translation strategies are more effective than simply increasing the amount of pseudo-data.