A study of BERT-based bi-directional Tibetan-Chinese neural machine translation
Yidong Sun, Dan Yang, Changyuan Yu, Andi Dong, Cuo Yong · 2022
Tibetan-Chinese bidirectional translation can promote the development of Tibetan-Chinese scientific and cultural exchanges as well as educational and cultural undertakings, but it has been facing the dilemma of low resource language parallel corpus scarcity. In this paper, we hope to provide a research idea for Tibetan-Chinese machine translation under low-resource conditions through the study of pre-trained language models, and to alleviate the problems existing in Tibetan-Chinese machine translation such as scarce corpus and unsatisfactory translation performance, so as to promote the development of Tibetan society. In this paper, by introducing pre-trained language models BERT and ALBERT into Tibetan-Chinese bidirectional translation, the translation effect is greatly improved by using large-scale monolingual data. The translations before and after use are also analyzed. Through the tests, the average improvement of the Chinese-Tibetan translation is 2.13 BLEU values, and the average improvement of the Tibetan-Chinese translation is 2.68 BLEU values.