Objective : To improve the performance of the Uyghur-Chinese translation model, we use the method of knowledge distillation to obtain a more robust Chinese- Uyghur translation model. The reverse translation method generates the generated data with better quality based on the Chinese-Uyghur model. Different data screening methods are used to Get high-quality generated data further. Then, a high-performance Uyghur-Chinese neural machine translation system is obtained using the existing parallel and generated data. In the CCMT2021 Uyghur-Chinese evaluation task, the impact of the above method on the quality of Uyghur-Chinese translation is verified, and it is the first place in this translation task.
Methods : In this paper, various data filtering methods are studied. The generated data are extracted and mixed into actual data according to the scores of generated data obtained by different methods. Then different models are trained using the data obtained by different data screening methods, and these models are integrated to obtain the best translation system. In addition, to obtain a better reverse translation model, the data obtained by the knowledge distillation method are added to the actual data to enhance the robustness of the Chinese-Uygur translation model. Different data screening methods are used to obtain the generated data, and a multi-branch model training method is used to obtain the translation model.
Results : The experiment uses 160,000 Uyghur - Chinese sentence pairs and 6 million Chinese monolingual data. These data are obtained by processing the data provided by the Uyghur - Chinese translation task. This paper conducts a comparative study of different data screening methods, uses different data screening methods to process the generated data, and conducts experiments on different methods with the same amount of data. The experimental results of different data screening methods are similar and are significantly better than those of the baseline model, data distillation, and reverse translation. This result may be due to the method of adding labels when training the translation model with generated data. This method can indicate which data is generated and which is actual, thus making the results of different data filtering methods similar. And it is found in this paper that after fine-tuning the model obtained by data enhancement only with parallel data, the performance of the model decreases, which may be because only training sets are used to fine-tune the model. Result Using the model ensemble method is better than that of a single model, which indicates that the model with better performance can be obtained by an ensemble of the models obtained by different data selection methods. In addition, this paper compares the depth and breadth of models and the impact of different model architecture integration on performance. It is found that deeper models may not lead to better results under low resource conditions. The integration of different model structures is better. In addition, this paper compares the depth and breadth of models and the impact of different model architecture integration on performance. It is found that deeper models may not lead to better results under low resource conditions. The integration of different model structures is better. In the CCMT2021 Uyghur-Chinese translation task, compared with the baseline system, reverse translation, and other systems for the same task, our system trained has better translation results, and it is the first place in this translation task.
Conclusions : To improve the performance of the Uyghur - Chinese translation model, we use reverse translation, knowledge distillation, fine-tuning, model averaging, and Ensemble integration. At the same time, this paper also compares different data filtering methods, hoping to continue to study data enhancement methods through these comparative experiments. Due to the instability of interpolation methods for fusion methods, this paper hopes to combine domain similarity and data quality filtering from the root rather than just using the interpolation method to fuse different methods. This paper hopes to combine the advantages of dynamic convolution and Transformer models in the following research to achieve better results in low-resource translation tasks. It also hopes to try deeper models on the Transformer model and integrate models with different structures.