Research Of Short Sequence's Machine Translation Based On Recurrent Neural Network

In the era of globalization,the relations among various countries and peoples are getting closer and closer,and translation also naturally becomes a kind of urgent need.Although artificial translation may be accurate,the cost is unaffordable for many needs.Therefore,the importance of machine translation is self-evident.However,machine translation can not meet the automatic and high-quality requirements yet.So,it still has important academic and industrial value that explore more efficient machine translation algorithms.Deep learning is widely used and has made breakthrough achievements in the field of machine translation.However,neural machine translation often uses words as the basic unit of input.Not only the process is more complicated and the imperfect segmentation algorithm may cause errors.The dictionary is too large which will cause the network model's dimension become too high.In addition,semantic will be complex and it is difficult to deal with.In view of the above problems,this paper takes character level bilingual data as input and further improves neural translation network model.After all,the character is different from the word,input form's conversion will result in the decline of the original model.Therefore,under the overall framework of the encoder-decoder,this topic further improves the neural machine translation model,emphasizing the ability of express words.In the word vector generation module,the original word characterization is rough extracted by convolution,and the word sequence is formally cut by recurrent neural network,which enhances the ability of character representation.In addition,the hard-cut term will increase the difficulty of model's optimization.This subject add an information supplement process,strengthening the characte's expression ability.The classical attention mechanism only focuses on the current output and the overall input information.By integrating the historical information,the overall value of the weight value can be enhanced and the word alignment effect can be further strengthened.Through experimental verification,the CRNN-embed of this subject has been further improved compared to RNN-embed and RNN-search models based on the character level input.By comparing different models based on different corporas,we can find that the growth of corpus scale has a significant impact on the model.Besides,by comparing the model effects of different input forms of characters and words,it is proved that the character-level model is more challenging.