Research On Machine Translation Based On Deep Learning

Machine translation is one of the hot issues in the field of artificial intelligence.The information age promotes the development of machine translation,but it also puts forward higher requirements for the quality of machine translation.Although the machine translation model is constantly developing,it has encountered bottlenecks such as missing translation,sparse data,common sense errors,poor text translation and so on.The pre training models such as Bert,ernie1.0 and ernie2.0 have achieved breakthrough results in various NLP tasks.However,in the field of machine translation,which requires high grammatical integrity and semantic accuracy,the pre training model is not effective.However,many ideas in the pre training model are worth learning.This paper synthesizes the ideas of Bert,ernie1.0 and ernie2.0,and designs a bi-directional enhanced multi task machine translation model based on seq2 seq,which uses multi-layer bi-directional transformer as encoder and multi-layer LSTM stack as decoder.The main contents include:(1)In order to alleviate the problem of data sparsity,two pre training machine translation models from source language to target language and from target language to original language are trained by aligning bilingual corpus in both positive and negative directions on the basis of seq2 seq.Then the two pre training models are used to translate monolingual data of the original language and monolingual data of the target language respectively to obtain two groups of pseudo aligned bilingual data,Then,the pseudo alignment data and the original bilingual alignment data with a small amount of data are put together to form a certain number of new alignment data,and the new alignment data is used for expectation maximization(EM)iteration.(2)Ernie1.0 uses three levels of masking strategies: word mask,phrase mask and entity mask to implicitly learn entity attributes,entity relations and other knowledge information.Learning from the idea of ernie1.0,entity mask is added to transformer to enhance knowledge.(3)Based on the idea of ernie2.0,the task of generating summary is introduced.Before implementing the machine translation task with the seq2 seq model,the seq2 seq model is pre trained with the text summary data set,and then the machine translation task is initialized with the pre trained parameters.Finally,a machine translation model with incremental learning on the summary generation task is obtained.