Research And Application Of Key Technologies In Neural Machine Translation

In the era of economic globalization,the importance of translation services is gradually reflected.Compared with human translation services,machine translation is faster and more able to meet the needs of society.Neural machine translation(NMT),as a technology of training neural networks using large-scale bilingual parallel corpus,has become the mainstream method of machine translation,and is widely used in the translation work between various languages.In this thesis,the key technologies of neural machine translation are studied and applied.Neural machine translation usually only learns translation knowledge through parallel corpora and ignores the prior features of language itself.In addition,in the mainstream neural machine translation model,only the output of the top-level encoder is utilized,and other deep information is ignored,which also limits the performance of the translation model.In order to solve the above problems,the following work has been completed:(1)A Chinese English neural machine translation model with multi granularity morphological features has been proposed.Chinese words are composed of Chinese characters,which can be further divided into components.The morphological information of these parts is closely related to the semantics of words.The fine-grained feature is obtained by the component n-gram based Chinese word embedding model proposed in this thesis.The coarse-grained features are obtained through the word embedding layer of the BERT pretrained language model.The BLEU-4 score of the proposed model is 0.78 higher than that of the baseline model Transformer.In the task of word vector evaluation,the Chinese word embedding model proposed in this thesis has better performance than the mainstream Chinese word vector model,which verifies the ability of the model to extract Chinese morphological features.(2)A neural machine translation model based on deep encoder information is proposed.In order to solve the problem that the decoders only decode the output of the top layer encoder in the deep translation model,this thesis designs three methods to make the output information of the deep layer encoder fully utilized by the model,which are: the utilization of the parallel layer encoder information,the utilization of the multi-layer encoder information,and the utilization of the dynamic deep layer encoder information.Three translation models based on deep encoder information are trained according to the three methods respectively,and compared with Transformer in the Chinese English translation task.The experimental results show that the highest BLEU-4 score of the proposed models is 0.89 higher than that of the benchmark model,which verifies the effectiveness of the proposed methods.(3)Based on the improved neural machine translation model,a neural machine translation web system is designed and implemented.Based on the B/S architecture,the system is composed of user interaction layer,core service layer and model processing layer.The system has a simple and easy-to-use user interface and can provide accurate machine translation services.