Method and device for language conversion based on syntactic drive

Abstract

The application discloses a sequential-driven language conversion method and device, which relate to the field of NLP. The specific implementation plan is: use a preset encoder to encode the source sentence to be converted, determine the first vector and the second vector corresponding to the source sentence; determine the current mask vector according to the preset rule, wherein, The mask vector is used to modify the vector output by the encoder; the third vector is determined according to the target language characters corresponding to the source characters before the first source character; the first vector, the second The second vector, the mask vector and the third vector are decoded to generate a target character corresponding to the first source character. This method introduces a mask vector in the attention layer of the decoder, and corrects the vector output by the encoder through the mask vector, so as to control the attention in the translation model, which can ensure that the translation result has the alignment information of the positive diagonal, Then, the sentence-by-sentence translation result is obtained, and the translation delay is reduced.

Description

technical field [0001] The present application relates to the technical field of computers, in particular to the technical field of Natural Language Processing (NLP), and in particular to a method and device for language conversion based on a sequential drive. Background technique [0002] In the current translation model, in order to improve the accuracy of the translation results, multiple translation results are usually generated during the translation process, and then according to the alignment information corresponding to each translation result, the alignment information closest to the positive direction is selected from the multiple translation results. The result of the diagonal is output as the final translation result. [0003] However, since most of the training corpus of the translation model is translation data with inverted sentences, even if the translation results obtained by expanding the translation candidates, when sentences such as inverted sentences and...

AI Technical Summary

Problems solved by technology

[0003] However, since most of the training corpus of the translation model is translation data with inverted sentences, even if the translation results obtained by expanding the translation candidates, when sentences such as inverted sentences and passive voice appear, they are usually in the form of Translate in an inverted way, which not only causes a delay in sentence translation, but also affects user experience due to non-sentence-driven translation

Images