| In recent years.deep learning has become the focus of many research areas.For the machine translation task of natural language processing field,the appearing of sequence to sequence neural machine translation system breaks the situation of traditional machine translation system which consists of multi modules.The integrated structure and satisfying translation results attract many researchers' attention.Later.attention-based neural network improves the model.even making better translation than traditional statistical machine translation system,and makes it one of the main translation systems over the world.In this paper.we take attention-based neural network as the research background.Combining the recent relevant scientific researches,we study attention-based Mongolian-Chinese neural network machine translation system from three aspects as follows:(1)Pre-training of Mongolian word vector:vectors are the representation forms of words which are directly involved in model training.and the quality of vectors is closely related to the quality of final translation model.so we explore three Mongolian word vector pre-training methods to enhance the quality of translation.(2)Dictionary-based Mongolian word segmentation:In the case of serious data sparse problem caused by the characteristics of the Mongolian word formation,we cut Mongolian words with different particles base on stem,suffix and case dictionaries.(3)Mongolian feature extraction:stems.suffixes and cases are the language characteristics of the word formation for Mongolian,we extract these language features still base on dictionaries,and put them into the model training in order to improve the translation results.Finally.we construct and improve a complete attention-based Mongolian-Chinese neural machine translation system.Experiments show that based on our methods.our system can achieve sansrying transtation results.The best model can make the BLEU value of translation result reach 30.19. |
PDF Full Text Request