Abstract: In recent years,the rapid development of language models has greatly promoted the model effect of supervised machine translation.However,the performance of supervised machine translation is highly dependent on the quality of parallel corpora.In view of the lack of high-quality Chinese-Burmese parallel corpora resources,this paper proposes a corpus construction method based on pivot optimization self-training.Firstly,the initial machine translation model is trained with a small-scale high-quality Chinese-Burmese parallel corpus.Then,a pseudo-parallel corpus from Burmese to Chinese is generated based on this model.At the same time,an English-Burmese parallel corpus with English as the pivot language is introduced,and the pivot English is translated into Chinese using existing high-quality English-Chinese translation tools to construct a second set of Chinese-Burmese pseudo-parallel corpora.To further improve the quality of the pseudo-parallel corpus,it designs a cross-lingual representation scoring mechanism to select higher quality sentence pairs from the two sets of pseudo-parallel corpora based on semantic similarity.Finally,the initial translation model is iteratively optimized and trained using the selected high-quality pseudo-parallel corpora.Experimental results show that the proposed method achieves an average 8.32 BLEU value improvement in the Chinese-Burmese machine translation task.Detailed analysis experiments prove that the pivot language optimization method can effectively enhance the model self-training effect and gradually improve the quality of pseudo-parallel corpus when the initial model performance is weak.In addition,this study constructs 700 000 high-quality Chinese-Burmese parallel corpus to further promote the development of Chinese-Burmese machine translation.

Key words: Parallel corpus construction, Machine translation, Self-training, Pivotal language, Chinese, Burmese