结合BERT数据增强的基于词切分的蒙汉神经机器翻译系统  

Mongolian-Chinese neural machine translation system based on word segmentation with BERT data enhancement

作　　者：何乌云 秀芝 包晶晶 陈美兰 王斯日古楞 HE Wuyun;XIU Zhi;BAO Jingjing;CHEN Meilan;WANG Siriguleng(Collage of Computer Science and Technology,Inner Mongolia Normal University,Hohhot 011500,China)

机构地区：[1]内蒙古师范大学计算机科学技术学院,内蒙古呼和浩特011500

出　　处：《厦门大学学报（自然科学版）》2022年第4期667-674,共8页Journal of Xiamen University：Natural Science

基　　金：国家自然科学基金(61762072);内蒙古自治区科技计划项目(2021GG0139);内蒙古自治区高等学校科学研究项目(NJZY21546);内蒙古师范大学研究生科研创新基金资助项目(GXJJS20129)。

摘　　要：神经机器翻译是目前机器翻译领域主流研究方法,但是蒙汉平行语料的稀缺使得蒙汉神经机器翻译性能难以提升.本文针对基于Transformer的蒙汉神经机器翻译系统,利用深度学习模型对蒙古文词切分方法进行研究,分析了蒙古文部分切分、BPE子词切分和BiLSTM-CNN-CRF神经网络切分方法对于蒙汉机器翻译模型的影响,并在此基础上利用基于BERT(bidirectional encoder representations from Transformers)中文语义相似度计算的数据增强技术去扩充蒙汉机器翻译训练数据.在CCMT2019提供的数据集上进行对比实验,实验结果表明,数据增强方法的BLEU值相较于基线实验提升显著,且BLEU4值达到了75.28%.Although neural machine translation(NMT)is currently regarded as the mainstream research method in the field of machine translation,the scarcity of Mongolian-Chinese(M-C)parallel corpus retards improvements of M-C NMT performances.In this study,aiming at the M-C NMT systems based on Transformer,we use the deep-learning model to study the Mongolian word segmentation method.Mongolian partial segmentation,BPE subword segmentation and BiLSTM-CNN-CRF neural network segmentation method for the influence of the Mongolian-Chinese machine translation model are compared among one another.On this basis,the data enhancement technology based on bidirectional encoder representations from Transformers(BERT),Chinese semantic similarity calculation is used to expand the M-C machine translation training data.In comparison of experimental results on the dataset provided by CCMT2019,ours show that the BLEU value of the data enhancement method is significantly improved when compared with that of baseline experiments,and the BLEU4 value reaches 75.28%.

关 键 词：蒙汉神经机器翻译 Transformer神经网络 BERT 语义相似度 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]