一种基于迁移学习的藏英神经机器翻译方法
A Tibetan-English Neural Machine Translation Method Based on Transfer Learning
随着全球化背景下跨文化交流需求的日益增长,藏英翻译在促进我国藏族文化的传播、加强与外界的联系以及提升藏语言数字化应用等方面具有重要的现实意义。然而,藏英平行语料资源的严重匮乏,极大地制约了藏英神经机器翻译研究进展,导致系统在处理低频词汇和复杂句式时表现欠佳。针对这一问题,文章系统探讨了迁移学习技术在藏英神经机器翻译中的应用潜力。通过分析父模型的数据规模、目标语言特性以及参数冻结等因素对藏英翻译模型性能的影响基础上,借助其他语言的丰富语料和预训练模型,探索迁移学习策略在藏英神经机器翻译中的潜在优势。研究结果显示,引入迁移学习策略后藏英神经机器翻译性能得到了很大的提升,相比于传统策略BLEU值提高了1.88个点。该方法不仅能够有效缓解藏英语句对的不足,还能有效增强藏英神经机器翻译模型的翻译质量和泛化能力。
With the increasing demand for cross-cultural communication in the context of globalization, Tibetan-English translation plays a crucial role in promoting the spread of Tibetan culture in China, strengthening ties with the outside world, and improving the digital application of Tibetan language. However, the serious shortage of Tibetan-English sentence parallel corpus resources has greatly limited the progress of research in Tibetan-English neural machine translation, resulting in poor performance of the system when dealing with low-frequency vocabulary and complex sentence patterns. In view of this problem, this paper systematically discusses the application potential of transfer learning technology in Tibetan-English neural machine translation. Based on the analysis of the influence of the data size of the parent model, the characteristics of the target language, and the parameter freezing on the performance of the Tibetan-English translation model, this study explores the potential advantages of employing a transfer learning strategy in Tibetan-English neural machine translation with the help of rich corpus and pre-training model of other languages. The results show that incorporating the transfer learning strategy significantly improves Tibetan-English neural machine translation compared to traditional methods, and the BLEU value is improved by 1.88 compared with the traditional strategy. This method can not only effectively make up for the deficiency of Tibetan-English sentence pairs, but also enhance the translation quality and generalization ability of Tibetan-English neural machine translation models.
| [1] |
严松思,珠杰,汪超, |
| [2] |
完么扎西.基于短语的藏英统计机器翻译关键技术研究[D].拉萨:西藏大学,2018. |
| [3] |
戴光荣,刘思圻.神经网络机器翻译:进展与挑战[J].外语教学,2023,44(1):82-89. |
| [4] |
李洪政,冯冲,黄河燕.稀缺资源语言神经网络机器翻译研究综述[J].自动化学报,2021,47(6):1217-1231. |
| [5] |
|
| [6] |
田静,贾智勇. 基于语言特征和迁移学习的机器翻译系统研究[J].自动化与仪器仪表,2022(8):213-216. |
| [7] |
|
| [8] |
|
| [9] |
卓玛扎西. 基于组块分割的无监督藏文句法分析研究[D].西宁:青海师范大学,2022. |
| [10] |
刘鑫鹏,栾悉道,谢毓湘, |
| [11] |
乌尼尔. 面向多任务知识迁移的低资源神经机器翻译研究[D].呼和浩特:内蒙古大学,2022. |
| [12] |
|
| [13] |
|
| [14] |
|
| [15] |
陈炜,汪洋,钟敏.结合生成式对抗网络的神经机器翻译研究[J]. 计算机与数字工程,2020,48(5):1164-1167. |
| [16] |
|
| [17] |
格桑加措,阿卜杜热西提·热合曼,尼玛扎西, |
| [18] |
基肖尔·帕皮内尼,萨利姆·鲁科斯,托德·沃德, |
| [19] |
|
新一代人工智能国家科技重大专项(2022ZD0116100)
国家自然基金重点项目(62436006)
西藏自治区科技计划技术创新引导项目(XZ202501JX0004)
/
| 〈 |
|
〉 |