基于迁移学习的汉越神经机器翻译
黄继豪,余正涛*,于志强,文永华

(昆明理工大学信息工程与自动化学院,云南省人工智能重点实验室,云南 昆明 650500)

神经机器翻译; 迁移学习; 注意力机制; 汉语-越南语

Chinese-Vietnamese neural machine translation based on transfer learning
HUANG Jihao,YU Zhengtao*,YU Zhiqiang,WEN Yonghua

(Key Laboratory of Artificial Intelligence of Yunnan Province,School of Information Engineering and Automation,Kunming University of Science and Technology,Kunming 650500,China)

DOI: 10.6043/j.issn.0438-0479.201908037

备注

针对汉语-越南语(简称汉越)平行语料受限的问题,提出了一种基于迁移学习的汉越神经机器翻译(TLNMT-CV)模型.在训练汉语-英语、英语-越南语的翻译模型的基础上,通过迁移学习方法,利用训练得到的汉语端编码器和越南语端解码器,分别对汉越翻译模型的编码器与解码器参数进行初始化,并使用小规模汉越语料进行微调优化,得到TLNMT-CV模型.实验表明,TLNMT-CV模型能够快速地实现新模型的初始化,提高模型的参数质量,从而提高翻译性能.相比Transformer,TLNMT-CV模型的双语互译评估(BLEU)值提升了1.16个百分点.
Aiming at the problem of limited Chinese-Vietnamese parallel corpus,we propose a Chinese neural machine translation model based on transfer learning(TLNMT-CV).First,we train Chinese-English and English-Vietnamese translation models,and then use the learned Chinese-side encoder and Vietnamese-side decoder through the transfer learning method to respectively encode and decode the new Chinese-Vietnamese translation model.Parameters are initialized and finally fine-tuned and optimized using a small-scale Chinese-Vietnamese corpus to obtain the TLNMT-CV model.Experimental results show that the TLNMT-CV model training can quickly achieve the initialization of the new model,improve the parameter quality of the model,and thus improve the translation performance.Compared with the baseline system Transformer,the BLEU value of the TLNMT-CV model increased by 1.16 percentage points.