Research And Implementation Of Neural Machine Translation Based On Generative Adversarial Networks

In today’s rapidly evolving information age,machine translation applications are becoming more widespread and machine translation models are moving forward and making breakthroughs.Neural Machine Translation models have achieved advanced performance on a variety of benchmarks.However,these models often rely on massively parallel corpora for training and exhibit performance degradation and data sparsity on low-resource languages.In addition,existing neural machine translation models use a unidirectional decoding order from left to right or right to left,ignoring contextual semantic information and suffering from output imbalance.To address the above problems,this paper proposes a neural machine translation model based on generative adversarial networks,with the following details:(1)To address the data sparsity problem faced by machine translation,a simple data augmentation method is proposed to train the translation model by modifying the target-side sentences to add noise and creating new pseudo-parallel sentence pairs in combination with the source-side utterances.In order to make the generated final translated translations diverse,the data augmentation method is used in combination with the decoding stage to construct pseudo-sentence pairs using a sampling decoding strategy,and different baseline model methods and decoding strategies are selected for comparison experiments.The experimental results verify that the data augmentation method proposed in this study can effectively a lleviate the problem of insufficient generalisation ability of the neural machine translation model,thus improving the sentence representation ability and the performance of neural machine translation.(2)To address the unidirectional nature of the decoder order of the neural machine translation model and the output imbalance problem,a neural machine translation model based on generative adversarial network optimisation is proposed,and a new generative network is designed to improve the decoder order of the original transformer model,which improves the output imbalance problem of the original model and enhances the ability to capture contextual semantic information at the same time.In the generative adversarial network framework,the generative capability of the generative network is continuously strengthened by adversarial training of the generator and discriminator,prompting the generator to generate higher quality translations.The policy gradient approach in reinforcement learning is also used to improve the disadvantages of discrete text data in natural language processing.The experimental results show that the adversarial learning network model proposed in this study has a good effect on the ability to capture semantic information and the fluency of translations.(3)Finally,a neural machine translation system platform is built based on the improved neural machine translation model.The test results prove that the proposed method can translate higher quality translations in context and has c ertain application value.