Abstract: The existing Tibetan-Chinese machine translation corpora exhibit significant domain data imbalance,resulting in inconsistent translation performance of trained models across different domains.Back-translation,as a common data augmentation method,enhances model performance by generating diverse pseudo-parallel data.However,traditional back-translation approaches struggle to fully account for the domain imbalance in the data distribution,leading to limited improvements in translation perfor-mance for resource-scarce domains,even as overall performance increases.This paper proposes a strategy that involves an in-depth analysis of the distribution of long-tail words in existing corpora,and targeted selection of monolingual data using these long-tail words from the existing Tibetan-Chinese bilingual corpora.By generating pseudo-data through back-translation,it performs data augmentation.This strategy aims to improve the overall performance of Tibetan-Chinese machine translation models while enhancing translation performance in data-scarce domains.Experiment results demonstrate that by fully considering domain data imba-lance and incorporating long-tail word data augmentation,the translation performance of machine translation models in resource-scarce domains can be effectively improved,providing a targeted approach to address the issue of domain data imbalance.

Key words: Long-tail words, Data augmentation, Tibetan-Chinese machine translation, Domain data imbalance