[1]夏 珺,周湘贞,隋 栋.基于循环生成对抗网络的机器翻译方法研究[J].南京师大学报(自然科学版),2022,45(01):104-109.[doi:10.3969/j.issn.1001-4616.2022.01.015]
　Xia Jun,Zhou Xiangzhen,Sui Dong.Research on Machine Translation Method Based on Cyclic Generation Countermeasure Network[J].Journal of Nanjing Normal University(Natural Science Edition),2022,45(01):104-109.[doi:10.3969/j.issn.1001-4616.2022.01.015]
点击复制

基于循环生成对抗网络的机器翻译方法研究()
分享到：

《南京师大学报（自然科学版）》[ISSN:1001-4616/CN:32-1239/N]

卷:
第45卷
期数:
2022年01期
页码:
104-109
栏目:
·计算机科学与技术·
出版日期:
2022-03-15

文章信息/Info

Title:
Research on Machine Translation Method Based on Cyclic Generation Countermeasure Network
文章编号:
1001-4616(2022)01-0104-06
作者:
夏 珺1; 周湘贞2; 隋 栋3
(1.黔南民族师范学院外国语学院,贵州 都匀 558000)(2.马来西亚国立大学信息科学与技术学院,马来西亚 雪兰莪 43600)(3.北京建筑大学电气与信息工程学院,北京 102406)
Author(s):
Xia Jun1; Zhou Xiangzhen2; Sui Dong3
(1.School of Foreign Languages,Qiannan Normal University for Nationalities,Duyun 558000,China)(2.Faculty Information Science and Technology,National University of Malaysia,Selangor 43600,Malaysia)(3.School of Electrical and Information Engineering,Beijing University of Civil Engineering and Architecture,Beijing 102406,China)
关键词:
语音识别; 语言翻译; 循环对抗网络; 长短时记忆模块
Keywords:
speech recognition; language translation; cyclic countermeasure network; long-short memory module
分类号:
TP391
DOI:
10.3969/j.issn.1001-4616.2022.01.015
文献标志码:
A
摘要:
近几年来,智能语言处理在语言学习方面已经得到了广泛的应用,但是由于在处理语言中往往会存在网络模型优化困难、强制对其的标记数据会出现精度偏差,与以往大多数使用判别模型结合HMM混合模型进行声学模型训练的系统相比,本文提出了一种基于循环生成对抗网络的机器翻译方法,该方法主要结合生成对抗网络来训练机器翻译模型. 首先,将一段语音输入神经机器翻译模块进行离散,预先变换得到MFCC特征; 然后,将经过预处理的语音输入到特征提取模块并结合长时短时记忆网络循环提取语音特征; 最后,将网络模型输出的语音与人工翻译的语音进行对比,并判别网络模型输出的语音特征与人工翻译的语音是否匹配,如果不匹配则继续优化生成网络. 实验结果表明,我们的网络与传统的高斯核混合模型相比有明显的提升. 本文方法在CSDN口令集、Rockyou口令集、Tianya口令集和Yahoo口令集中均取得了优越的结果,其中在Yahoo口令集中单词错误率降至19.5%.
Abstract:
In recent years,intelligent language processing has been widely used in language learning. However,due to the difficulty of network model optimization and the accuracy deviation of its labeled data,compared with most previous systems using discriminant model combined with HMM hybrid model for acoustic model training,This paper proposes a machine translation method based on cyclic generation countermeasure network. This method mainly combines generation countermeasure network to train machine translation model. Firstly,a speech is input into the neural machine translation module for discrete pre transformation to obtain MFCC features; Then,the preprocessed speech is input to the feature extraction module,and the speech features are extracted circularly combined with the long-term and short-term memory network; Finally,the speech output from the network model is compared with the artificially translated speech,and whether the speech features output from the network model match the artificially translated speech is judged. If not,the network is optimized. The experimental results show that our network is significantly improved compared with the traditional Gaussian kernel mixture model. This method has achieved excellent results in CSDN password set,rockyou password set,Tianya password set and Yahoo password set,and the word error rate in Yahoo password set is reduced to 19.5%.

参考文献/References:

[1] DAS D,DAS A K,PAL A R,et al. Meta-heuristic algorithms-tuned elman vs. jordan recurrent neural networks for modeling of electron beam welding process[J]. Neural processing letters,2021,53(2):1647-1663.
[2]BRUNO J H,JARVIS E D,LIBERMAN M,et al. Birdsong learning and culture:analogies with human spoken language[J]. Annual review of linguistics,2021,7(1):89-97.
[3]BOER B D,THOMPSON B,RAVIGNANI A,et al. Evolutionary dynamics do not motivate a single-mutant theory of human language[J]. Scientific reports,2020,10(1):22-31.
[4]MUKHERJEA A,ALI S,SMITH J A. A human rights perspective on palliative care:unraveling disparities and determinants among asian american populations[J]. Topics in language disorders,2020,40(3):278-296.
[5]LI K,PAN W,LI Y,et al. A method to detect sleep apnea based on deep neural network and hidden Markov model using single-lead ECG signal[J]. Neurocomputing,2018,294(6):94-101.
[6]张旭东,黄宇方,杜家浩,等. 基于离散型隐马尔可夫模型的股票价格预测[J]. 浙江工业大学学报,2020,48(2):148-153.
[7]OUISAADANE A,SAFI S. A comparative study for Arabic speech recognition system in noisy environments[J]. International journal of speech technology,2021,11(3):1-10.
[8]PIKHART M. Human-computer interaction in foreign language learning applications:applied linguistics viewpoint of mobile learning[J]. Procedia computer science,2021,184:92-98.
[9]YAO Q,UBALE R,LANGE P,et al. Spoken language understanding of human-machine conversations for language learning applications[J]. Journal of signal processing systems,2020,92(3):78-89.
[10]HINTON G E. A practical guide to training restricted boltzmann machines[J]. Momentum,2012,9(1):599-619.
[11]DEWANGAN S,ALVA S,JOSHI N,et al. Experience of neural machine translation between Indian languages[J]. Machine translation,2021,35(1):71-99.
[12]SHTERIONOV D,SUPERBO R,NAGLE P,et al. Human versus automatic quality uation of NMT and PBSMT[J]. Machine translation,2018,32(3):217-235.
[13]HOCHREITER S,SCHMIDHUBER J. Long short-term memory[J]. Neural computation,1997,9(8):1735-1743.
[14]GRAVES A,JAITLY N. Towards end-to-end speech recognition with recurrent neural networks[C]//International Conference on Machine Learning,Beijing,2014:1764-1772.
[15]DENG L,YU D. Deep learning for signal and information processing[J]. Now publishers,2013,12(8):218-227.
[16]WU S,LI G,DENG L,et al. Li-norm batch normalization for efficient training of deep neural networks[J]. IEET transations on neural and learning systems,2018,30(7):2043-2051.
[17]MAAS A L,QI P,XIE Z,et al. Building DNN acoustic models for large vocabulary speech recognition[J]. Computer speech & language,2014,41(C):195-213.
[18]CUI X,ZHANG W,FINKLER U,et al. Distributed training of deep neural network acoustic models for automatic speech recognition:a comparison of current training strategies[J]. IEEE signal processing magazine,2020,37(3):39-49.
[19]TAI K S,SOCHER R,MANNING C D. Improved semantic representations from tree-structured long short-term memory networks[J]. Computer science,2015,5(1):36-41.
[20]XU Y,DU J,DAI L R,et al. An experimental study on speech enhancement based on deep neural networks[J]. IEEE signal processing letters,2013,21(1):65-68.
[21]SALMELA L,TSIPINAKIS N,FOI A,et al. Predicting ultrafast nonlinear dynamics in fibre optics with a recurrent neural network[J]. Nature machine intelligence,2021,12(8):1-11.
[22]POLIAK A,RASTOGI P,MARTIN M P,et al. Efficient,compositional,order-sensitive n-gram embeddings[C]//Conference of the European Chapter of the Association for Computational Linguistics,London,2017:503-508.
[23]黄光许,田垚,康健,等. 低资源条件下基于 i-vector 特征的LSTM递归神经网络语音识别系统[J]. 计算机应用研究,2017,34(2):392-396.
[24]舒帆,屈丹,张文林,等. 采用长短时记忆网络的低 资源语音识别方法[J]. 西安交通大学学报,2017,51(10):120-127.

备注/Memo

备注/Memo:
收稿日期:2021-10-11.
基金项目:贵州省教育厅人文社科项目(2019zc116)、国家自然科学青年基金项目(61702026).
通讯作者:夏珺,副教授,研究方向:机器学习,智能翻译,自然语言处理. E-mail:273976230@qq.com
更新日期/Last Update: 1900-01-01