Definitive neural network machine translation method and storage medium

Abstract

A method and a storage medium for machine translation of a definite neural network comprise that following step: 1, training a basic NMT model; 2, using that NMT model to translate a verification setand output the word alignment information in the ATTENTION LAYER; 3, using a part-of-speech tagger to carry out automatic part-of-speech tagging, and then automatically generate (UNK2) corpus by utilizing word alignment information; 4, adding (UNK2) corpus into that original corpus data and retraining; 5, replacing the words to be replaced with the (UNK2) corpus by using the substitution function,decoding, and replacing the words with the word alignment information; 6, completing that UNK replacement decoding process. The invention can solve the shortcoming that the technical effect of NMT for UNK replacement is not ideal, and the pertinent flow can greatly increase the successful probability of replacement while maintaining high quality translation quality. The invention applies the UNKreplacement function to ensure the accuracy of noun translation.

Description

technical field [0001] The invention discloses a translation method, in particular a limited neural network machine translation method and a storage medium. Background technique [0002] Most people who are engaged in translation related work will encounter the following problems: [0003] (1) Are you troubled by the inexplicable translation results of the current neural network machine translation? [0004] (2) Are you troubled by the digital problems in the translation process? [0005] (3) Have you been blamed by customers for mistranslation of keywords such as terms? [0006] (4) Are you discouraged by the sensitivity of political news translation? [0007] (5) Do you want to fully intervene in the pre-processing and post-compilation of machine translation to achieve the goal of efficient translation? [0008] (6) Do you want to automatically add the precious term base you have kept in your hand to get the translation result you want? [0009] Translation is a proce...

AI Technical Summary

Problems solved by technology

However, it is believed that the various feature values ​​set, and the error rate of each step that requires N kinds of natural language processing technologies (word segmentation, part-of-speech tagging, syntactic analysis, etc.), will lead to a large deviation in the final result

And the biggest problem is that the ordering model is not perfect and cannot cover the global features, resulting in the fluency of statistical machine translation is not very good.

Images