针对音变问题改进的维汉神经网络机器翻译鲁棒性方法研究  

Research on the Robustness of Improved Uyghur-Chinese Neural Network Machine Translation Method for Phonetic Change

作　　者：阿里木·赛买提 沙丽瓦尔·阿里木 卡哈尔·卡迪尔 刘轩铭 凌震华[3] 刘俊华 ALIMU·Saimaiti;SHALIWAER·Alimu;KAHAER·Kadier;LIU Xuan-ming;LING Zhen-hua;LIU Jun-hua(Xinjiang Branch,IFLYTEK Co.,Ltd.,Urumqi,Xinjiang,830046,China;Academic Affairs Office of Xinjiang University,Urumqi,Xinjiang,830046,China;National Engineering Research Center for Speech and Language Information Processing,University of Science and Technology of China,Hefei,Anhui,230000,China;AI Research Institute of IFLYTEK Co.,Ltd.Hefei,Anhui,230000,China)

机构地区：[1]科大讯飞股份有限公司新疆分公司,新疆乌鲁木齐830046 [2]新疆开放大学教务处,新疆乌鲁木齐830046 [3]中国科学技术大学语音及语言信息处理国家工程研究中心,安徽合肥230000 [4]科大讯飞股份有限公司AI研究院,安徽合肥230000

出　　处：《新疆师范大学学报（自然科学版）》2023年第4期1-8,共8页Journal of Xinjiang Normal University(Natural Sciences Edition)

摘　　要：由于维吾尔语具有音变差异,在记录时容易混淆,这对机器翻译的容错能力提出了挑战。从词粒度研究的角度出发,针对维吾尔语中具有明显语音特性的音变词的翻译问题,采用字符级神经网络机器翻译(Character level Neural Network Machine Translation,CharNMT)结构,并结合覆盖率机制进行优化,这一方法不仅提高了音变词的翻译质量,还展现出较好的容错能力。在基础模型取得了1.25 Blue分的提升,并在训练和解码速度方面也具有优势。The Uyghur language's phonetic variation introduces challenges in machine translation due to the potential for recording errors.This study focuses on the translation of phonetically variant words with unique phonetic attributes from a word granularity perspective.By leveraging the Character level neural Network Machine Translation(CharNMT)framework and integrating it with a coverage attention mechanism,we enhance the translation quality of such words while also showcasing robust fault tolerance.Foundational model achieved a 1.25 point improvement in BLEU score and demonstrated superior training and decoding speeds.

关 键 词：维吾尔语音变词 神经网络机器翻译技术 容错能力 字符级神经网络机器翻译 

分 类 号：TP391.2[自动化与计算机技术—计算机应用技术] TP183[自动化与计算机技术—计算机科学与技术]