The Research Of Deep Learning Method For Machine Translation Modeling

Nowadays,there are great developments in natural language processing.Machine translation has been focused for a long haul by researchers.Neural machine translation technology has simpler translation model,and it is easy to operate and it doesn’t need lots of expertise,so it has been the main translation model now.However,there are still some problems for us to solve.On one thing,because translation model are trained on large scale corpus by neural networks,the training period is quite long.On another thing,the translation result is still not so good when using neural translation models.There are still some mistakes,such as missing translation,wrong translation.The two main problems in neural machine translation in that it has long training terms and the translation result is not ideal.In order to solve the two problems,this paperimproved translation model both in encoding and decoding.Proposed a new loss function based on decay weight and a group embedding model(1)Proposed a kind of loss function based on decay weight help model’s training during decoding.Traditional method in neural machine translation is taking former translated word as part of inputs in translation of next word.So the previous words in a sequence have more importance during translating.The new loss function distributes large weight to words that appeared earlier,which made model prefer translate previous words precisely.This paper had experimented using IWSLT dataset in German-English translation task.Experiment result shows,in German to English translation,compared withtraditional loss function,the models with decay weight loss function improve 1.63%in Bleu score.(2)We proposed a kind of encoder which based on group embedding.The input of traditional natural language model is the word vector of each word.This kind of input could only carry information in training corpus.Group embedding model can using extra linguistic information for encoder as some input information.This information is beyond the limit of training corpus,which supervised the model to learn well.In English-German translation test,compared with traditional method,group embedding model shortened the convergence time by 35.29%.The idea of group embedding can be used in any model only if the input could be word vector.Experiment result shows,either in sentiment analysis or named entity recognition,the model can improve recognition precision effectively,which indicates the universality of group embedding in natural language tasks.