Research On Low-Resource Machine Translation Based On Teacher-Student Model

Neural machine translation has made a breakthrough with sufficient data.However,the performance suffers from the data scarcity problem in the low-resource condition.To a certain extent,the pivot method based on the student-teacher model overcomes the lack of parallel data.It has become an effective solution for low-resource neural machine translation,but it still has bottlenecks in terms of efficiency and data scale.To solve the efficiency problem,this paper proposes the concept of multivariate generated data and uses it to guide the student model to imitate the teacher model.To solve the problem of data scale,this paper proposes different improvement methods from the perspective of data diversity and system robustness.From the perspective of model robustness,the noise of human writing errors is introduced into the pseudo data and the structure of the student model is improved with the adversarial generative network.From the perspective of data diversity,this paper introduces a unilateral data set and sentence similarity matching method to match and replace the source sentences of parallel data.Through experiments in six translation tasks and complexity analysis,this paper verifies that multi-generation data can improve the efficiency of student model learning teacher model,and reduce the excessive inference.The comparative experiments of three translation tasks proved that diversified data enhancement can improve the model.Finally,in a comparative experiment with noise,it is found that the model for adversarial training is more robust than the original model.The above experiments show that the proposed method has a positive effect on the student-teacher model.