Multidirectional recurrent neural network machine translation model training method and device

Abstract

The invention discloses a multi-directional recurrent neural network machine translation model training method and device, and relates to the field of machine translations, the multi-directional recurrent neural network machine translation model comprises initial translation, recurrent source end translation and recurrent target end translation, during training, the characteristics of parallel data are utilized, in a training stage, a source end sentence sequence and a target end sentence sequence are regenerated through a translation model, and part of parameters of an initial translation model are optimized by calculating loss of the sentence sequence generated through reconstruction so as to improve performance in the initial translation model. calculating the similarity between the context vector of the sentence sequence obtained by different reconstructions and the context vector of the source end sentence or the context vector of the target end sentence output by the initial translation model.

Description

technical field [0001] The invention relates to the field of machine translation, in particular to a multi-directional recurrent neural network machine translation model training method and device. Background technique [0002] The current training scheme of the translation model is to provide each sentence with a real target sequence, and use the cross-entropy loss method to approximate the probability distribution of the generated characters to the probability distribution of the characters in the real target sentence, and then use various search strategies to obtain the probability distribution of the characters The probability distribution is processed to get the target sentence. However, there is a deviation between the semantics contained in the sentences obtained by this method and the features predicted by the decoder. In addition, the size of the data, the noise contained in the data, and the modeling ability of the model may make the model unable to fully fit the ...

AI Technical Summary

Problems solved by technology

However, there is a deviation between the semantics contained in the sentences obtained by this method and the features predicted by the decoder.

In addition, the size of the data, the noise contained in the data, and the modeling ability of the model may make the model unable to fully fit the training data, resulting in the semantics contained in the target sentence translated by the model and the semantics contained in the real target sentence. difference

Images