亚词及单词深度维汉机器翻译模型对比  

Contrastive study of Uyghur-Chinese deep machine translation models based on sub word and word

作　　者：张胜刚 艾山·吾买尔[1,2] 吐尔根·依布拉音[1,2] 买合木提·买买提[1,2] ZHANG Sheng-gang;Hasan·Wumaier;Tuergen·Yibulayin;Mahmut·Maimaiti(School of Information Science and Engineering,Xinjiang University,Urumqi 830046,China;Xinjiang Laboratory of Multi-Language Information Technology,Xinjiang University,Urumqi 830046,China)

机构地区：[1]新疆大学信息科学与工程学院,新疆乌鲁木齐830046 [2]新疆大学新疆多语种信息技术重点实验室,新疆乌鲁木齐830046

出　　处：《计算机工程与设计》2019年第8期2326-2330,共5页Computer Engineering and Design

基　　金：新疆多语种信息技术实验室开放课题基金项目(2016D03023);国家自然科学基金项目(61662077)

摘　　要：针对基于神经网络的维汉机器翻译中的集外词问题和随着网络层数的加深训练和优化模型会变得更加困难这两个问题,在对相关工作研究后,提出基于深层神经网络的亚词及单词的维汉机器翻译模型。在翻译单元上将基于词的翻译单元替换为基于词和亚词的混合翻译单元,将基于GRU的神经非线性单元替换为基于ALU的非线性单元,缓解训练和优化模型的难度并提高译文质量。通过实验发现该模型相比基准系统提高了近13个BLEU值,该研究对形态丰富黏着语言与汉语的机器翻译具有借鉴意义。As to the problem of the extra-words in Uyghur-Chinese machine translation based on neural network and the problem that as the number of network layers increases,training and optimizing models becomes more difficult,after researching on rela- ted work,a deep neural network of sub word and word Uyghur-Chinese machine translation model was proposed.Word-based translation units were replaced with mixed translation units based on words and subwords,and GRU-based neuro-nonlinear units were replaced with ALU-based nonlinear units to ease the difficulty of training and optimizing the model and to improve the quality of the translation.Through experiments,it is found that the proposed model improves nearly 13 BLEU values compared with the benchmark system.This research has certain reference significance to the form-rich adhesive language and Chinese machine translation.

关 键 词：维汉机器翻译 深层神经网络 亚词切分 集外词问题 BLEU值 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]