基于循环生成对抗网络的机器翻译方法研究

Research on Machine Translation Method Based on Cyclic Generation Countermeasure Network

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%.

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%.