Post-Editing of Machine Translation

Abstract

This paper explores the practice of post-editing translations produced by Neural Machine Translation (NMT), with a particular emphasis on Arabic translations from Google Translate.. The core of the research involves a practical analysis of five different English texts—technical, literary, news, marketing, and legal—translated into Arabic using NMT. Each translation is evaluated for errors using a five-category framework: semantic/contextual, morphosyntactic, lexical/terminological, and stylistic/fluency. This paper tackles a deficiency in the research concerning post-editing Neural Machine Translation (NMT) for Arabic, as current studies primarily concentrate on European language pairs and do not provide genre-specific, measurable error analysis. Utilizing Google Translate as the NMT system, we post-edited five English–Arabic translations that span technical, literary, news, marketing, and legal genres. A mixed-method approach was used, where two professional translators independently identified errors using a framework consisting of five categories:  semantic/contextual,morphosyntactic, lexical/terminological, and stylistic/fluency. The analysis reveals that while NMT can produce fluent translations, it often results in subtle yet significant errors that require thorough post-editing. The findings indicate that post-editing demands greater cognitive effort and sharper analytical skills than traditional translation. This paper add to the growing area of Machine Translation Post-Editing (MTPE) by providing specific examples of NMT errors in Arabic and proposing practical solutions for post-editing.