嵌入句法信息在汉蒙非自回归机器翻译应用  

Application of Embedding Syntactic Information in Chinese Mongolian Non-Autoregressive Machine Translation

作　　者：程永坤 苏依拉[1] 王涵 仁庆道尔吉[1] CHENG Yong-kun;SU Yi-la;WANG Han;RENQING Dao'erji(College of Information Engineering,Inner Mongolia University of Technology,Hohhot Inner Mongolia 010080,China)

机构地区：[1]内蒙古工业大学信息工程学院,内蒙古呼和浩特010080

出　　处：《计算机仿真》2023年第5期486-490,共5页Computer Simulation

基　　金：国家自然科学基金(61966027,61966028);内蒙古自然科学基金(2021MS06028);内蒙古自治区攻关项目(2021GG0329)。

摘　　要：针对当前的语言研究模型多为自回归神经机器翻译模型,存在Exposure Bias现象和不并行解码问题,提出非自回归神经机器翻译模型进行汉蒙翻译研究。借助生成对抗网络对汉蒙语料进行对抗训练,然后利用教师模型Transformer对得到的语料进行知识蒸馏处理,为学生模型提供高精度汉蒙语料,最后利用图卷积神经网络学习句子中的句法信息,并将句法信息融入到词嵌入层中。通过仿真结果证明,所提研究模型结合实验所用方法,在保证了模型翻译速度大幅度提升的前提下,同时译文的翻译质量也呈现出提高的效果。In view of the fact that most of the current language research models are autoregressive neural machine translation models,which have the phenomenon of exposure bias and non-parallel decoding,a non-autoregressive neural machine translation(NAT)model is proposed for Chinese Mongolian translation research.The generative adversarial network(GAN)was used to train the Chinese and Mongolian corpus,and then the teacher model transformer was used to distill the knowledge of the obtained corpus,so as to provide high-precision Chinese and Mongolian corpus for the student model.Finally,the graph convolution network(GCN)was used to learn the syntactic information in the corpus,and the syntactic information was integrated into the word embedding layer.The simulation results show that the proposed research model combined with the methods used in the experiment can not only ensure a significant improvement of the model translation speed,but also improve the translation quality.

关 键 词：非自回归神经机器翻译 生成对抗网络 知识蒸馏 图卷积神经网络 句法信息 

分 类 号：TP3[自动化与计算机技术—计算机科学与技术]