| The Mongolian-Chinese machine translation quality has been improved along with the rapid development of neural network machine translation.The work of this thesis put forward methods for Mongolian-Chinese machine translation research.However,low resource of bilingual sentence pair limits the performance of MongolianChinese machine translation which has contributed to performance issues for the attention weight and word embedding,and the method of neural networks machine translation on other tasks cannot be applied to Mongolian-Chinese machine translation.To solve the above problems,we proposed a model of Mongolian-Chinese neural machine translation with prior information,which used prior information to enrich the available features of the model to improve the translation performance.Firstly,we proposed guided alignment model based Mongolian-Chinese machine translation.We reordered the target sentence to reduce the differences in word order between the Chinese and Mongolian to improve the quality of the word alignment.Afterwards,using the word alignment in statistical machine translation guides neural network training.Secondly,replace the out-of-vocabulary method based words semantic similarity.We use word embedding to calculate the similarity of words,then replace the out-of-vocabulary by the most similar words which are covered by the target vocabulary.Finally,Pre-training model.Training word embedding using large scale monolingual corpus initial word embedding of machine translation model and we added part-of-speech feature.Experiment show that guided alignment model has a significant improvement in the Mongolian-Chinese translation tasks with 31.98 BLEU points,and improvement of 2.29 BLEU points over the baseline.Experiment show that the method of replacing the out-of-vocabulary based words semantic similarity can relieve the complexity of training of machine translation.Pre-training word embedding has improvement of 2.68 BLEU points over the baseline. |
PDF Full Text Request