Machine translation method applied to stable deep machine translation model

Abstract

The invention discloses a machine translation method applied to a stable deep machine translation model, which comprises the following steps of: in the training process of a deep model, setting two matrixes mu and sigma for a gradient matrix G of partial parameters of an encoder in the deep model when each back propagation calculation is carried out according to an input source statement X; respectively representing a mean value and a standard deviation of the gradient matrix G by using matrixes mu and sigma, and calculating values of the matrixes mu and sigma according to the gradient matrix G; standardizing the gradient matrix G by using matrixes mu and sigma in each back propagation process during model training; the standardized use stage and position are controlled through a preset strategy and priori knowledge, rapid and accurate training of a machine translation model is carried out, and finally the translation quality of an output translation during decoding is improved. According to the method, the training process can be more stable and is not limited by specific models and tasks, the method has good universality, and the effect of the model can be more easily reproduced on different devices.

Description

Technical field [0001] The present invention relates to a machine translation model training method, specifically a machine translation method applied to a stable deep machine translation model. Background [0002] Machine Translation is a classic task in the field of natural language processing, the goal of which is to translate the source language (the language to be translated) into the target language (the translated language) through a computer. In today's interconnected world, the meaning of language is particularly important, including politics, business, education, medical and other fields, language is an irreplaceable medium for information transmission. In the face of the massive amount of data in today's Internet, it is difficult for humans to analyze this information by themselves, so machine translation technology as a powerful means of translation has been widely accepted. [0003]Although as early as the 17th century, scholars have proposed the idea of using machin...

AI Technical Summary

Problems solved by technology

However, the use of half-precision number calculations will inevitably affect the calculation accuracy to a certain extent, and because there are calculation errors in the calculation of half-precision numbers, when using half-precision numbers for training, it will also aggravate the neural network during reverse calculation. Vanishing gradient and exploding gradient problems

Images