关键词: 语言模型, 机器翻译, 古汉语翻译, Siku-RoBERTa, Siku-GPT, Siku-Trans

Abstract: This study aims to construct and validate a Chinese ancient-modern translation model based on pre-trained language models, providing strong technical support for the research of ancient Chinese and the inheritance and dissemination of cultural heritage. The study selected a total of 300,000 pairs of meticulously processed parallel corpora from the "Twenty-Four Histories" as the experimental dataset and developed a new translation model—Siku-Trans. This model innovatively combines Siku-RoBERTa(as the encoder) and Siku-GPT(as the decoder), designed specifically for translating ancient Chinese, to build an efficient encoder-decoder architecture. To comprehensively evaluate the performance of the Siku-Trans model, the study introduced three models as control groups: OpenNMT, SikuGPT, and SikuBERT_UNILM. Through comparative analysis of the performance of each model in ancient Chinese translation tasks, we found that Siku-Trans exhibits significant advantages in terms of translation accuracy and fluency. These results not only highlight the effectiveness of combining Siku-RoBERTa with Siku-GPT as a training strategy but also provide important references and insights for in-depth research and practical applications in the field of ancient Chinese translation.

Key words: Language model, Machine translation, Ancient Chinese translation, Siku-RoBERTa, Siku-GPT, Siku-Trans