Evaluating neural machine translation using error analysis in English-Arabic texts

Abstract EN

The aim of this study was to evaluate the output of Neural Machine Translation of translating texts from English into Arabic using error analysis.

Google Translate was taken as an example as the leading neural machine translations.

Most of the studies done on machine translation were on rule-based and statistical machine translation rather than neural machine translation.

Texts were selected based on the American Translator Association criteria which is used in their examinations.

Three texts were selected to represent three types of texts: general, financial, and scientific.

Error analysis then was used to analyze the results of the translation and compare them with each other and with that in the literature.

105 errors were discovered in the three texts with an average of 1.9 error per sentence.

27 of the errors were syntactic errors, while 14 of the total errors are grammatical errors, and 64 of the errors are semantic errors.

Although there is a clear improvement in Google Translate, especially in the grammar part, since it was shifted to a neural system, more has to be done to improve it in general and in the semantic part in particular.