Mongolian-Chinese neural machine translation method based on multiple constraint terms

Abstract

A Mongolian-Chinese neural machine translation method based on multiple constraint terms comprises the steps that firstly, a model training process based on reinforcement learning is constructed for Mongolian-Chinese neural machine translation tasks, and then constraint conditions are further improved for a training optimization target on the basis of a reinforcement model; comprising the following steps: adding a semantic constraint module to relieve the problem of poor translation fluency caused by a single BLEU value evaluation system; performing parameter constraint on the training process to improve the efficiency of model training; and performing word list constraint on the corpus to reduce the number of unregistered words in the translation. According to the method, by adjusting the overall constraint mode, the problems that in a low-resource Mongolian-Chinese machine translation task, a model is poor in sequence structure analysis capacity and low in training efficiency are solved, and meanwhile the problem that variance is large due to intensive training is solved; according to the method, a mean value reward and pruning column search method is adopted to effectively relieve the negative influence caused by the enhanced training.

Description

technical field [0001] The invention belongs to the technical field of machine translation, in particular to a Mongolian-Chinese neural machine translation method based on multiple constraints. Background technique [0002] When using a deep learning model to process Mongolian texts, other indicators other than the standard BLEU value need to be trained and constrained to ensure the quality of Mongolian translations with scarce resources. [0003] In the translation model of the prior art, in the training process, due to the lack of Mongolian language resources, it is easy to generate training sample distribution deviation and semantic loss, the model training efficiency is not high, and it is easy to lead to inconsistent training sample distribution deviation and evaluation indicators in the training inference stage. question. SUMMARY OF THE INVENTION [0004] In order to overcome the shortcomings of the above-mentioned prior art, the purpose of the present invention is ...

AI Technical Summary

Problems solved by technology

[0003] In the translation model of the prior art, during the training process, due to the lack of Mongolian resources, it is easy to generate training sample distribution deviation and semantic loss, the model training efficiency is not high, and it is easy to cause training sample distribution deviation and inconsistent evaluation indicators in the training inference stage question

Images