Mongolian-Chinese non-autoregressive machine translation method based on knowledge graph

Abstract

A Mongolian-Chinese non-autoregressive machine translation method based on a knowledge graph comprises the following steps: constructing a Mongolian-Chinese bilingual named entity semantic web through a knowledge graph triple, and constructing a named entity dictionary based on multi-feature alignment; then representing Mongolian rare word composition distribution by means of a knowledge graph triple, performing rare word segmentation, performing interactive enhanced generative adversarial training in the process, and adding interactive information to guide generative adversarial to obtain a comprehensive rare word knowledge graph triple set; performing knowledge distillation on the processed Mongolian-Chinese parallel corpus, and constructing a knowledge graph of the Mongolian-Chinese parallel corpus when the teacher model extracts corpora for the student model through knowledge distillation; when non-autoregressive machine translation is carried out, adopting a knowledge graph as a projection matrix training set, obtainng a projection matrix through bidirectional embedding mapping joint training, and carrying out decoding information retouching. According to the method, the translation quality of machine translation can be improved on the premise that the translation rate is improved.

Description

Technical field [0001] The present invention belongs to the technical field of machine translation, Mongolian and Chinese in particular, it relates to non-knowledge based machine translation map autoregressive method. Background technique [0002] Machine translation (MT) is the use of the computer will automatically translate one language into another language, and the conversion process to maintain the same meaning. Research existing machine translation model based on multi-regression translations from machine translation model (AT) study, for example, it is now the hotspot model Transformer, appears the model of its efficient coding efficiency, the effect of superimposed layers of the attention of quality improved significantly. But the Transformer model itself, there are: Exposure Bias (transitive error) and the decoder can not decode the parallel case, which greatly waste of GPU computing resources, so that the translation effect of less than ideal state researchers. [0003...

AI Technical Summary

Problems solved by technology

However, the Transformer model itself has: Exposure Bias (error transitivity) and the situation that the decoder cannot decode in parallel, which greatly wastes the computing resources of the GPU, and makes the translation effect not reach the ideal state of the researcher.

[0003] Considering that the existing autoregressive machine translation model cannot make full use of the computing power of the GPU, based on this, a non-autoregressive machine translation model (NAT) is proposed, which can make full use of the computing power of the GPU and increase the translation speed and efficiency by 7.2 about times

However, the non-autoregressive machine translation model itself also causes continuous repeated translation and missed translation due to the excessively fast translation. inter-sequence dependencies

Images