Implementation Of Scenario-level Mongolian-chinese Machine Translation System Under Meta-learning Framework

Currently,the neural network machine translation model has performed well in various language translation tasks.However,because there is no parallel corpus,it is difficult to improve the translation quality of a language such as Mongolian.While transfer learning improves the translation quality of low-resource languages using high-resource languages,the selection of parameters directly restricts the improvement of translation performance.Consequently,this thesis proposes a scenario-level Mongolian-Chinese machine translation model under the meta-learning framework.Through multi language learning,a meta parameter with better generalization performance is used to initialize the Mongolian-Chinese machine translation model.In order to avoid language differences affecting the generalization of meta parameters,this thesis embeds the multilingual words in the same metric space.Finally,the Mongolian-Chinese machine translation system is realized,which is convenient for users to use.The specific work of this thesis is as follows.(1)Constructing a multilingual word embedding model.First,a bilingual word embedding space is constructed using a translation language model based on parallel corpora of different languages,and then all the bilingual words embedding spaces are mapped into an extra metric space by using the generalized procrustes Analysis method to implement multilingual word alignment.(2)Constructing a scenario-level Mongolian-Chinese machine translation model under the meta-learning framework.The construction process is divided into two phases: meta-training and meta-testing.In the meta-training phase,in order to get the meta parameter with better generalization,each update of the meta parameter depends on all translation scenarios.In order to solve the problem of the large amount of calculation and memory consumption caused by the dependence of meta parameter update on the display optimization path of each translation scenario,this thesis uses implicit differential to calculate the meta gradient,so that the calculation of meta gradient is only related to the solution of each translation scenario.In the meta-testing phase,this thesis uses the Mongolian-Chinese parallel corpus to train on the obtained meta parameter.(3)The realization of Mongolian-Chinese machine translation system.This thesis mainly realizes the function of Mongolian-Chinese bilingual translation,and also adds the function of clearing and copying content,so as to improve the convenience of users using the system.The experimental results show that the multilingual word embedding model can effectively resolve the problem of language differences.At the same time,the Mongolian-Chinese machine translation model based on scenario-level meta-learning method can obtain a BLEU value of 36.81% on the Mongolian-Chinese parallel corpus of CCMT2019.The experimental results also show that the meta parameter obtained by the scenario-level meta-learning method can also show better generalization on smaller data sets.