A Multilingual Machine Translation Model Based on Neural Networks

The strategy of the currently proposed machine translation method is still based on a certain a priori assumption, that is, the error distribution of the translation model exhibits a certain regularity. If the model is under different training levels, the above regular error distribution may not be completely applicable. In this paper, a scheduling sampling algorithm based on the decoding step is proposed, which can largely fit the error distribution of the model in the test scenario, thereby further reducing the inconsistency between training and testing of the translation model. When designing the scheduling sampling strategy, we can further consider the current real-time capability of the model, and jointly determine the model's scheduling sampling strategy according to the decoding steps and the confidence of the model.