提升中华传统文化文本机器翻译性能的研究

摘要/Abstract

摘要： 中华传统文化文本承载着中华文化的核心价值观与思想观念,其英译对推动中华传统文化传播、创新发展及现代性对话具有重要意义。这类文本多以文言书写,句法修辞独特,文化内涵丰富,导致英译面临人工翻译效率低、机器翻译质量欠佳的双重困境,其中的低资源文本更是如此。传统对联作为其中的典型文本,尤其面临此类翻译挑战。本文以对联翻译为研究对象,从前期积累的对联英译库中抽样,选用机器翻译中的MarianMT、SMaLL-100和mT5三种预训练模型,通过全局和层次微调进行精细调整,以提升对联英译性能。结果表明,通过参数精细调整,各模型均展现出译联性能的显著提升,SMaLL-100和MarianMT模型的提升幅度较mT5更显著,而mT5模型在每种策略下的评价损失都显著降低,全局微调对大多数指标的提升效果普遍较好。本研究对于进一步训练调整预训练模型,提升其英译中华传统文化文本的性能有重要的参考价值。

关键词: 预训练模型, 精细调整, 对联, 机器翻译, 人工智能

中图分类号:

H059

. 提升中华传统文化文本机器翻译性能的研究[J]. 上海翻译, 2025, 184(5): 19-24.

参考文献

[1] Chang, E. et al. Time-aware ancient Chinese text translation and inference [J/OL].arXiv preprint arXiv:2107.03179, 2021. https://arxiv.org/pdf/2107.03179 (2021-07-07) [2024-12-26].
[2] Church, K. W. et al. Emerging trends: A gentle introduction to fine-tuning [J].Natural Language Engineering, 2021, 27(6): 763-778.
[3] Devlin, J. et al. Bert: Pre-training of deep bidirectional transformers for language understanding [A]. Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies [C]. Minneaplis: Association for Computational Linguistics, 2019: 4171-4186.
[4] Gao, R. Y. et al. Machine translation of Chinese classical poetry: A comparison among ChatGPT, Google Translate, and DeepL Translator [J].Humanities and Social Sciences Communications, 2024, 11(1): 1-10.
[5] Li, J. H. et al. Eliciting the translation ability of large language models via multilingual finetuning with translation instructions [J]. Transactions of the Association for Computational Linguistics, 2024, 12: 576-592.
[6] Liu, Dayiheng et al. Ancient-modern Chinese translation with a new large training dataset [J]. ACM Transactions on Asian and Low-Resource Language Information Processing, 2019, 19(1): 1-13.
[7] Ray, P. P. ChatGPT: A comprehensive review on background, applications, key challenges, bias, ethics, limitations and future scope [J].Internet of Things and Cyber-Physical Systems, 2023, 3: 121-154.
[8] Reinauer, R. et al. Neural machine translation models can learn to be few-shot learners [J/OL]. arXiv preprint arXiv:2309.08590, 2023.https://arxiv.org/pdf/2309.08590 (2023-09-15) [2024-11-29].
[9] Richburg, A. and M. Carpuat. How multilingual are large language models fine-tuned for translation? [J/OL]. arXiv preprint arXiv:2405.20512, 2024.https://arxiv.org/pdf/2405.20512 (2024-05-30) [2024-12-29].
[10] Su, T. et al. Unlocking parameter-efficient fine-tuning for low-resource language translation [J/OL].arXiv preprint arXiv:2404.04212, 2024.https://arxiv.org/pdf/2404.04212? (2024-04-05) [2024-12-28].
[11] Vaswani, A. et al. Attention is all you need [J/OL]. arXiv: 1706.03762, 2017.https://proceedings.neurips.cc/paper/2017/file/3f5ee243547dee91fbd053c1c4a845aa-Paper.pdf (2023-08-02)[2024-11-27].
[12] Yang, Z. C. et al. Generating classical Chinese poems from vernacular Chinese [A]. Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing [C]. Hong Kong: Association for Computational Linguistics, 2019: 6155-6164.
[13] Yang, Z. N., K. J. Chen, and J. Q. Chen. Guwen-UNILM: Machine translation between ancient and modern Chinese based on pre-trained models [A]. In Lu Wang et al. (eds.). Natural Language Processing and Chinese Computing (Part I)[C]. Cham: Springer International Publishing, 2021: 116-128.
[14] 王世钰.中华民族共同体视域下的海外楹联英译研究[J].上海翻译,2023(1):43-48.
[15] 吴梦成等.基于预训练语言模型的汉语古现翻译模型构建[J].信息资源管理学报,2024(6):143-155.
[16] 许乾坤等.基于UniLM模型的古文到现代文机器翻译词汇共享研究[J].情报资料工作, 2024(1): 89-100.
[17] 赵霞,王钊.结合对比学习和双流网络融合知识图谱摘要模型[J].计算机应用研究,2025(3): 720-727.
[18] 朱献珑,周丽.社会翻译学视域下的全人译者能力研究再思考[J].上海翻译, 2025(4): 52-58.