基于GAN的中英翻译算法

Chinese-English Translation Algorithm Based on GAN

神经网络机器翻译算法中,大部分人为设定的目标函数是尽可能提高n元词精确率,但这无法保证译文更加自然和准确。因此,将对抗生成网络(GAN)应用于机器翻译算法,使得网络自己学习目标函数。GAN中,生成模型采用Transformer模型生成假样本;判别模型采用基于卷积神经网络的二分类器判别真假样本。训练过程中,判别模型给出的分类结果和双语评估基础值(BLEU)目标函数一起评估假样本,并将结果反馈给生成模型,引导其进行参数更新及优化。采用AI-Challenger大赛提供的中英语料库和从全国机器翻译研讨会(CWMT)数据中随机抽取的英中语料库,分别在循环神经网络(RNN)、Transformer和Transformer+GAN模型上进行训练。Transformer和RNN模型相比,训练时长缩短了34h,BLEU提升了4.9分,而加入GAN训练后,BLEU又有0.02分的提升。

Among the neural network machine translation algorithms,the most artificial objectives targets are to improve the n -gram precisions as much as possible,but this cannot guarantee a more natural and accurate translation.The generative adversarial network (GAN) is applied to the machine translation algorithm to make the network learn the objective function automatically.In GAN,the generation model applies Transformer model to generate the false sample,the discriminant model uses a bi-classifier based on convolutional neural network to discriminate true and false samples.During the training,the classification results of the discriminant model and the objective function of the bilingual evaluation understudy (BLEU) evaluate the false samples,and the results feed back to the generation model to optimize the parameters.Using the Chinese-English dataset provided by the AI-Challenger competition and English-Chinese dataset randomly selected from the data of the China workshop on machine translation (CWMT) to train respectively on the recurrent neural network (RNN) model,Transformer model and Transformer+GAN model.Compared to RNN model,the training duration of Transformer model was reduced by 34 h and BLEU was improved by 4.9 points,thus BLEU has also increased by 0.02 points by adding GAN training.