Smarter edits? Post-editing with error highlights and translation suggestions

A study by Alina Karakanta finds that automatic post-editing (APE) error highlights and correction suggestions do not improve translator productivity or translation quality compared to regular post-editing, though they do enhance user experience [1]. The research, submitted on 20 May 2026 and revised on 16 June 2026, compared professional translators working from English to Dutch under three conditions: regular post-editing, post-editing with quality estimation (QE) error highlights, and post-editing with APE-derived highlights and correction suggestions [1][3]. No condition yielded measurable productivity or quality gains over standard post-editing [1]. The study noted that while QE highlights had a higher overlap with oracle edits, APE highlights received higher perceived accuracy and usefulness scores from the translators [3]. Correction suggestions, when provided alongside highlights, were seen as more informative than error highlights alone [3]. The work also found that QE highlights negatively affected translator confidence, particularly when the highlights were erroneous. However, when an erroneous highlight was accompanied by a correction suggestion, translators reframed it as a preferential edit rather than a distraction [3]. This finding underscores a nuanced interaction between automated support and translator agency. The study aligns with earlier research that questioned the practical benefits of QE features in post-editing workflows [3]. In 2021, a demonstration of the IntelliCAT system reported that QE and translation suggestions could significantly accelerate post-editing, achieving a 52.9% speedup in translation time compared to translating from scratch [4]. The contrast highlights the evolving and sometimes contradictory evidence base for these tools. More recent frameworks, such as TranslationCorrect, have integrated LLM-based error detection models like GPT-4o to highlight errors and allow translators to make fine-grained edits, aiming to streamline both translation workflows and data collection [5]. Karakanta's study concludes that evaluating LLM-based support tools requires measuring productivity, quality, and perception together, and recommends preserving translator agency through flexible, opt-in support [3].