Method and device for evaluating translation quality

Abstract

The embodiment of the invention discloses a method and device for evaluating translation quality. A specific embodiment of the method comprises the steps of obtaining a to-be-evaluated translation anda reference translation; inputting the to-be-evaluated translation and the reference translation into a pre-trained replication encoding model to obtain semantic similarity of the to-be-evaluated translation and the reference translation, the replication encoding model being a neural network used for calculating a probability that a pair of sentences is replicated sentences; respectively analyzing the translation to be evaluated and the reference translation into two syntactic trees through a syntactic analyzer; calculating the similarity of the two syntax trees as the text similarity of thetranslation to be evaluated and the reference translation; and taking the weighted sum of the semantic similarity and the text similarity as a translation quality score. According to the embodiment, by combining the scores of the indexes, comprehensive evaluation of the current translation in the aspects of semantics, syntax, lexical and overall information intelligibility can be reflected, and the situation that the automatic evaluation score is high, but the intelligibility is low is avoided.

Description

technical field [0001] The embodiments of the present disclosure relate to the field of computer technology, and in particular to a method and device for evaluating translation quality. Background technique [0002] Traditional automatic translation quality evaluation methods generally use BLEU (bilingual evaluation understudy, bilingual mutual translation quality auxiliary tool), METEOR (An Automatic Metric for MT Evaluation with Improved Correlation with Human Judgments, an automatic machine translation with improved correlation with human judgments). Evaluation method) and other indicators to evaluate the translation, focusing on the degree of matching between the original text and the target reference translation. For example, BLEU is more concerned with the matching degree of continuous strings, that is, n-grams. Some indicators such as METEOR are supplemented on the basis of BLEU, which can not only match strings with the same text, but also identify strings with diffe...

AI Technical Summary

Problems solved by technology

In this case, adopting the traditional automatic evaluation method may easily lead to low scores and inaccurate evaluation results due to too much missing translation information.

For example, using traditional methods to evaluate, the translation rate of an interpreter with many years of experience is only about 70%, and the translation acceptability is lower than 85%.

[0004] Traditional translation quality assessment methods have high labor costs, and at the same time, manual assessment is limited by the translation level of different human evaluators, resulting in large fluctuations in assessment results

It is necessary to construct a reference translation separately for the interpretation scene, and at the same time, the traditional method cannot distinguish the important information in the translation, and the quality of the translation is measured completely according to the fidelity

Images