Translation matrix model selection system based on OpenKiWi

Abstract

The invention provides a translation matrix model selection system based on OpenKiWi, a method for evaluating a plurality of translation results output by a plurality of translation engines based on an evaluation engine of an OpenKiWi core, and a translation terminal. The selection system comprises a to-be-translated source text input component, a GMM clustering engine, an evaluation engine basedon an OpenKiWi core and a selection output interface. The evaluation engine based on the OpenKiWi core performs sentence-level evaluation and word-level evaluation on a translation result, obtains a sentence-level evaluation matrix and a word-level evaluation matrix, combines the sentence-level evaluation matrix and the word-level evaluation matrix to obtain an N-order evaluation matrix PN, performs iterative computation on the N-order evaluation matrix PN, and outputs an optimal translation result based on an iterative computation result.

Description

technical field [0001] The invention belongs to the technical field of translation, and in particular relates to a translation matrix model selection system based on OpenKiWi, a method for evaluating multiple translation results output by multiple translation engines by an evaluation engine based on OpenKiWi core, and a translation terminal. Background technique [0002] Machine Translation (MT), also known as automatic translation, is the use of computer programs to translate text from one natural language (source language) into another natural language (target language). It is a branch of Natural Language Processing and is inseparably related to Computational Linguistics and Natural Language Understanding. [0003] Generally speaking, a formed statistical machine translation system uses a large-scale bilingual corpus for training, and these data often come from different fields, including politics, finance, military, science and technology, sports, education, entertainment...

AI Technical Summary

Problems solved by technology

This difference in domain and style brings two problems to the research of statistical machine translation: first, the machine translation system trained by using bilingual corpus in a certain domain or style has the best effect in translating texts in the same domain, while When translating texts in other fields, the results are often unsatisfactory, and sometimes the translation results are even unacceptable; second, although the machine translation system trained by using multi-field mixed bilingual corpus has good versatility, due to the interaction of data in multiple fields Influence and mutual restriction, the system often cannot achieve the best effect in a certain field

[0008] However, in this process, the memory effect of the translation engine matrix selection will lead to extreme polarization in the accuracy of the translation results, that is, if the initial selection is appropriate, the subsequent selection will become more and more accurate; otherwise, the subsequent selection will become more and more accurate. more and more unrealistic

Usually, in order to avoid this situation, R&D personnel are required to introduce manual feedback, which reduces the automation of the translation engine matrix and greatly reduces work efficiency; and this manual feedback process relies on the a priori rules of the R&D personnel, leading to accurate results At the same time as the accuracy decreases, the objectivity will also decrease

Images