[Purpose/significance] In the process of machine translation from ancient Chinese to modern Chinese, due to the significant differences in vocabulary composition, syntax and flexible use of parts of speech between ancient Chi⁃ nese and modern Chinese, and the lack of open word segmentation data of ancient Chinese, the understanding and pro⁃ cessing ability of machine translation system is biased, which affects the translation quality to some extent. [Method/ process] Firstly, this paper puts forward an unsupervised thesaurus construction method. Based on UniLM model, it is combined with BERT, RoBERTa, RoFormer and RoFormerV2 pre-training models respectively and fine-tuned the model. With the help of UniLM model, the language relationship between the source language and the target language is generated into an intermediate language representation, and the pre-training model is used to learn the context-relat⁃ ed language representation, so as to increase the relevance between semantics, thus improving the machine translation of ancient and modern times. [Result/conclusion] The experimental results show that the BLEU value of machine translation of ancient Chinese prose, which integrates the knowledge characteristics of ancient Chinese prose, is in⁃ creased by 0.27 to 1.12 on BERT, RoBERTa, RoFormer and RoFormerV2 pre- training models respectively, which proves the effectiveness of the proposed method.