Research On Machine Translation Method Based On Deep Neural Network

In recent years,more and more countries have actively participated in global cooperation.In this context of interconnection,machine translation models based on deep neural networks play a very important role in language exchanges and text translation between countries.The rapid development of the Internet has accumulated a lot of parallel corpus data,which provides more possibilities for the development of neural machine translation.The encoder-decoder is currently the most commonly used framework for neural machine translation.The encoder receives the source sentence sequence and maps the source sentence sequence information to a semantic space.The decoder receives and processes the information in the semantic space and calculates the probability distribution at the current moment,and output the target sequence step by step according to the encoding algorithm.Under this framework,a large number of translation models have emerged,all of which have their own advantages and disadvantages.Thesis makes innovations and changes on the basis of the two models to improve the performance of the translation system.The main work is as follows:(1)Generally,the encoder-decoder model based on RNN has problems such as being unable to mine sentence semantics well and translate long text sequences.In thesis,a neural machine translation model based on LSTM enhanced attention is proposed,and multi-head attention is introduced on the RNN baseline model.and multi-hop attention mechanism.By adding multi-head attention calculation,the grammatical information of different subspaces can be mined,and then the semantic information in the sentence sequence can be paid attention to,and then multiple attentions are calculated on each head respectively,which can effectively deal with long-distance dependency problems and in long sentences.(2)Although the Transformer model can process data in parallel and speed up the computational efficiency of the model,its local representation ability is poor.Based on the Transformer model,thesis proposes a neural machine translation model with ELMo representation fused with dynamic mask attention.In the word embedding module,the ELMo model is used to obtain the dynamic word vector representation of the word,so as to obtain rich word representation and be able to deal with the problem of polysemy.On the multi-head attention module,the multi-head attention calculation is mixed,and multiple heads are combined.Through a linear transformation,the attention of each head fully considers the attention information of each head,thereby improving the expressive ability of self-attention.On the encoder side,a dynamic mask attention layer is added to focus on the important information around the current word and discard the unimportant information,thereby improving the local modeling ability of the model.These practices enable the model to fully acquire local information and sentence semantic information,and translate sentences that are more in line with human natural language.