Syntactic structure fused Tibetan and Chinese language neural machine translation method

Abstract

The invention relates to a syntactic structure fused Tibetan and Chinese language neural machine translation method, and belongs to the technical field of machine translation and feature fusion application. The method aims to introduce syntactic structures of more languages in a neural machine translation framework so as to help to improve the machine translation quality, and a relative position coding method is provided by optimizing the position coding technical problem of transformers so as to fuse syntactic structure information. According to the method, a transformer is optimized by extracting a structure position coding method based on a dependency relationship, and finally the purpose of improving the Tibetan-Chinese neural machine translation quality is achieved. According to the method, the association efficiency of the self-attention neural network between two languages can be effectively improved, the problems caused by different syntactic structures of the two languages are relieved, the time complexity of the algorithm is reduced, the context information loss problem caused by absolute position coding adopted by a traditional model is solved, and mistaken translation and missed translation of low-resource neural machine translation are reduced.

Description

technical field [0001] The present invention relates to a Tibetan-Chinese language neural machine translation method that integrates syntactic structures, in particular to a self-attention Tibetan-Chinese language neural machine translation method based on relative position encoding of dependent syntactic structures, belonging to the technical field of machine translation and feature fusion applications . Background technique [0002] In recent years, neural machine translation has achieved the best performance in multiple translation tasks, and the translation model trained on large-scale corpus can be comparable to human translation. Neural machine translation differs from traditional statistical machine translation in that it no longer uses a rule-based approach, but instead uses a deep learning approach. Specifically, in the bilingual corpus based on the end-to-end "encoder-decoder" framework, the correspondence between each chunk after word segmentation is calculated. ...

AI Technical Summary

Problems solved by technology

[0006] The purpose of the present invention is to address the existing problems and deficiencies in the prior art. In the face of Tibetan-Chinese neural machine translation, due to the differences in the syntactic structure of the language itself, it is difficult to train, and a new Tibetan-Chinese fusion syntactic structure is proposed. Linguistic Neural Machine Translation Approaches

Images