APP下载

基于神经网络的维汉翻译系统实现

2018-12-14张胜刚艾山·吾买尔吐尔根·依布拉音买合木提·买买提米尔夏提·力提甫

现代电子技术 2018年24期
关键词:小语种

张胜刚 艾山·吾买尔 吐尔根·依布拉音 买合木提·买买提 米尔夏提·力提甫

关键词: 小语种; 机器翻译; theano; 神经网络; 开源系统; 负载均衡

中图分类号: TN711?34                         文献标识码: A                       文章编号: 1004?373X(2018)24?0157?05

Implementation of Uyghur?Chinese translation system based on neural network

ZHANG Shenggang1,2, Hasan Wumaier1,2, Tuergen Yibulayin1,2, Mahmut Maimaiti1,2, Mirxat Litip1

(1. School of Information Science and Engineering, Xinjiang University, Urumqi 830046, China;

2. Xinjiang Laboratory of Multi?Language Information Technology, Urumqi 830046, China)

Abstract: The machine translation of minority languages is generally to achieve online service system on the basis of the open source system due to reasons such as development costs and user scale. At present, the source codes provided by neural machine translation are mostly written by the theano, but the machine translation written by the theano cannot meet users′ needs due to its slow translation speed. How to realize the stable Uyghur?Chinese neural network machine translation system based on the theano is taken as the research object. The multi?layer bi?directional network framework of the ALU neuron is used as the translation model. The django is used to realize the translation service interface. The nginx+uwsgi is selected to achieve load balancing, so as to improve the translation speed. The experimental results show that, the translation speed of the system constituted by 5 translation engines is 1.3~1.55 times higher than that of the system constituted by 10 translation engines. The research results in this paper have an important reference value for using the open source system to rapidly realize the translation system that can meet daily visits of less than 10 million times per day.

Keywords: minority language; machine translation; theano; neural network; open source system; load balancing

0  引  言

隨着“一带一路”倡仪的提出,为了使得沿线的所有民族之间都可以无障碍自由的沟通,机器翻译变得日益重要。机器翻译(Machine Translation,MT)指的是利用机器实现自动从一种语言转换为另一种相应的语言,在翻译过程中可以分为源语言和目标语言。机器翻译的历史可以大致分为4个阶段。在第一个阶段中机器翻译主要是基于情报学中的“加密/解密”方法,但这一方法并没有获得成功。在第二个阶段中机器翻译主要利用语言学家手工书写翻译规则,然而语言学家编写的规则有限并且成本巨大,若某个句子不符合翻译规则则该句就无法正常翻译。第三阶段的机器翻译主要是基于概率统计的方法[1?3]。通俗来讲,基于统计的机器翻译的基本思想是利用概率统计方法找到大规模的平行语料的翻译规律从而构建相应的机器翻译模型。更具体来说就是将源语言的句子分割为词单元或者短语单元等,然后将其输入到翻译模型中并选择最好的目标语言句子进行输出。第四阶段也就是目前阶段主要是基于神经网络的机器翻译(神经机器翻译,NMT)[4?7]。通过多个不同的非线性处理单元,基于神经网络的方法能够自动对任意数据进行特征抽取和学习并拟合任意的非线性关系。目前基于神经网络的机器翻译一般采用encoder?decoder模型,encoder端主要将源语言句子编码为一个稀疏稠密的实数向量,decoder端则根据编码器端的结果生成目标语言句子。……

登录APP查看全文

猜你喜欢

小语种
小语种 大世界
“一带一路”背景下小语种专业面临的挑战和对策
小语种国际化人才培养模式初探
中国周边区域研究文献的需求与保障
中国周边区域研究文献的需求与保障
小语种就业现状调查与分析
高职单招班小语种优质课堂教学探究
全球化背景下小语种语言的发展
小语种对大学生(非小语种专业)就业竞争力影响调查