Research On Low Resource Neural Machine Translation Based On Transfer Learning

With the development of society and the progress of science and technology,the demand for cross language communication is increasing day by day.The high cost and low efficiency of manual translation can no longer adapt to the development of the times.Neural machine translation uses advanced technologies such as deep learning to improve the efficiency and accuracy of translation,which can better meet the needs of a large number of translation business.However,neural machine translation relies on large-scale parallel corpus,and when the translation field is different,the generalization ability of the model is poor.In reality,most languages have some problems,such as the scarcity of data resources,the scarcity of fields and types of parallel corpora,transfer learning method can transfer the knowledge from the old domain to the similar new domain,which can effectively alleviate the above problems.Based on the research of neural machine translation based on transfer learning,this paper proposes a hybrid transfer learning model of similar multi language domain fusion,aiming at improving the translation quality of low resource languages.ased on this model,a low resource language translation system is constructed.The specific work is as follows:(1)Aiming at the problem of data resource scarcity,this paper proposes a related multilingual mixed transfer learning method.This method selects a variety of high resource languages similar to low resource languages as the parent objects to solve the problem of scarcity of low resource language corpus.Compared with the mixed transfer learning method,the proposed method improves the translation quality by 1.2 BLEU,it improves the translation quality of low resource languages.(2)Aiming at the problem of low generalization ability of model,this paper proposes a domain fusion mixed transfer learning method.By adjusting the low resource language model twice,we can solve the problem of translation quality degradation when the training data and test data are different.Compared with the mixed transfer learning method,the proposed method improves 1.3 BLEU and the generalization ability of the low resource language model.(3)This paper proposes a mixed transfer learning model based on related multilingual domain fusion.The model is composed of related multilingual mixed transfer learning method and domain fusion mixed transfer learning method.When the model is applied to the Uzbek to Chinese translation task,compared with the related multilingual mixed transfer learning method alone,it improves the BLEU by 0.9,compared with the mixed transfer learning method of domain fusion alone,it improves the BLEU value by 1.2,and the translation quality is further improved,which proves the feasibility of the proposed mixed transfer learning model of related multilingual domain fusion Effectiveness.(4)Building a low resource language translation system.Using the mixed transfer learning model of related multilingual domain fusion proposed in this paper,the Uzbek to Chinese machine system based on B / S architecture is constructed.The design and implementation of the system are introduced in detail,and the translation results are displayed.In view of the scarcity of low resource language corpus and the low generalization ability of the model,this paper proposes a mixed transfer learning model based on related multilingual domain fusion,which can effectively improve the translation quality of low resource languages.At the same time,based on this model,the Uzbek to Chinese translation system is constructed,which has a certain application value,and is helpful to the improvement of the level of information technology of ethnic minority and the mutual exchange of various nationalities.