Research On Network Optimization Of Neural Machine Translation System

The demand for cross-language communication is increasing,how to solve the communication barriers between different languages with low cost and high efficiency has become a difficult problem.In response to this problem,machine translation came into being.At present,neural network-based neural machine translation system has exceeded statistical machine translation in performance,but neural network-based neural machine translation system has network modeling problems such as insufficient utilization of multi-layer net-work information,large parameter scale,poor interpretability,etc.Therefore,from the perspective of network optimization,this paper proposes an information fusion method to solve the problem of insufficient utilization of multi-layer network information.Aiming at the problem of large parameter scale,this paper proposes a model compression method based on progressive semi-knowledge distillation.Aiming at the problem of poor interpretability in the process of neural machine translation,the quality of word alignment based on neural machine translation is improved.1.Multi-layer information fusion optimization method.Aiming at the problem of insufficient information utilization between layers in the multi-layer network of the existing machine translation system,an information fusion method is proposed.The output information of each layer is fused as supplementary information to make up for the deficiency of the original output information,and three different fusion methods are proposed to make full use of the information of the middle layer and improve the quality of translation2.Model compression optimization method based on progressive semi-knowledge distillation.A network optimization method is proposed to solve the problems of difficult optimization and large demand for model storage resources in neural machine translation.This topic proposes network optimization methods of semi-knowledge distillation and progressive semi-knowledge distillation.Through semi-knowledge distillation,the machine translation model is first compressed by half,and then the other half of neural machine translation is progressively compressed,thus compressing the whole model into a translation model with simplified parameters,thus solving the problem of large scale of neural machine translation parameters.3.Optimization method of word alignment for neural machine translation model.Aiming at the problem that Transformer is difficult to generate high-quality alignment information,this paper proposes to optimize the alignment network based on Transformer system by using monolingual corpus and regular items in alignment set to improve the alignment quality generated by the alignment network,so as to improve the interpretability of neural machine translation.Experiments based on several different orders of magnitude and different language pairs show that the network optimization method studied in this topic can effectively solve the existing problems in machine translation,such as insufficient information utilization,large scale of model parameters,and difficulty for Transformer to obtain high-quality word alignment,and comprehensively optimize the overall performance of neural machine translation.