Research On Chinese And English Machine Translation Model Based On Deep Nerual Network

Under the background of economic globalization,with the rapid development of Internet technology,the international communication of various industries has become more frequent,and the demand for cross language communication is more and more obvious.As an efficient tool,Machine Translation can retain the original semantics and achieve the equivalent transformation between different languages.It has great practical significance.In recent years,deep learning technology has developed rapidly.Not only in the field of speech recognition and image processing,the related research in the field of Natural Language Processing has also achieved some good results.In this thesis,the Chinese and English Machine Translation model based on the deep neural network is studied.In this thesis,the end to end encoder decoder framework is used to construct the neural Machine Translation model,which enables the machine to learn the feature automatically.Transforming corpus data into word vectors by means of distributed representation,and the neural network is used to map the source language and the target language directly.A neural Machine Translation model is constructed for different neural network structures.First,three neural network structures are used to construct the neural Machine Translation model.The RNN network structure can handle the indefinite long sequence,but there is a gradient explosion and the gradient disappearance.The LSTM network structure alleviates the gradient failure and improves the ability to deal with the long distance sequence through the gate valve mechanism.The structure of the GRU network is simplified on the basis of LSTM,which reduced the training complexity and improve the translation performance.In view of the problem that any length source language sequence of the encoder is encoded into a fixed dimension background vector,the attention mechanism is introduced to dynamically adjust the influence degree of source language end context to the target language sequence.In order to better reflect the context information,this thesis further proposes a Machine Translation model based on bidirectional GRU to compare and analyze a variety of translation models to verify the effectiveness of the model performance improvement.In view of the problem that neural machine translation can not make good use of linguistic knowledge,the neural machine translation model which joins word sequence information is proposed.On the basis of the bidirectional GRU translation model with attention mechanism,Stanford Parser is used for syntactic analysis to obtain word sequence information,and it is used in the form of bidirectional encoding.The encoder part of the translation model is used to form the background vector by vector stitching.Experiments show that the addition of part of speech sequence information can improve the performance of the translation model.