Machine translation post-editing, explained

Post-editing is the human correction of machine translation (MT) output. A machine produces a draft, and a translator fixes it instead of translating from scratch. The international standard for the work is ISO 18587:2017, which sets requirements for full, human post-editing and for the competences of post-editors (ISO). Two published studies found that post-editing saved time: a 2010 company test measured an average throughput gain of 74 percent, and a 2013 Stanford experiment found shorter times and better quality in three language pairs. The effort itself is usually measured by counting edits.

What ISO 18587 covers

The ISO abstract page lists ISO 18587:2017 as edition 1, published in April 2017, 15 pages, from ISO technical committee TC 37/SC 5. The abstract says the standard is for translation service providers, their clients, and post-editors, and that it applies only to content processed by MT systems. For translation services in general, it points to ISO 17100.

The standard is no longer static. As of October 2026 the page marks it as an International Standard to be revised, stage 90.92. A draft successor, ISO/DIS 18587, carries a new title: post-editing of non-human translation output. A contract that cites ISO 18587 should name the edition.

A 2026 monograph by Márta Lesznyák, published by Akadémiai Kiadó, summarizes the standard's definition of post-editing as the correction of raw machine translation by a human translator.

Light and full post-editing

The ISO abstract names only full post-editing. Mara Nunziatini, writing at the 2020 conference of the European Association for Machine Translation, says the standard sets requirements for both full and light post-editing (EAMT 2020). Full post-editing should produce output equivalent to human translation. On light post-editing the standard is less precise, but it directs the post-editor to focus on accuracy and to disregard style. For both levels, sentences are to be restructured only when the meaning is unclear.

Nunziatini adds that definitions of light and full post-editing are inconsistent across the literature and across language service providers. One industry guideline the paper summarizes says light post-editing fixes major errors only, accepts minor ones, and should allow a fast turnaround.

Lesznyák describes the older division the same way: full post-editing aimed at human-translation quality for publication, and light post-editing at a comprehensible text. Citing Rico Pérez (2024), the monograph argues that neural MT output is often already comprehensible, so the distinction has lost its validity. That volume uses post-editing to mean full post-editing.

What the productivity studies found

Plitt and Masselot, of the software company Autodesk, ran a two-day test with twelve translators working from English into French, Italian, German, and Spanish (PBML 93, January 2010). The MT engine was Moses, trained only on the company's own data. The test covered 144,648 source words, with times recorded for each sentence. Throughput ranged from 400 to 1,800 words per hour. MT made every translator faster, by 20 to 131 percent, and 74 percent on average, which saved 43 percent of translation time. Slower translators gained more. The quality team, not told which texts were post-edited, rated every reviewed job average or good, and found more mistakes in translated jobs than in post-edited ones. The authors warn that their numbers may be of limited use to other MT users.

Green, Heer, and Manning hired 16 self-described professional translators for each of Arabic, French, and German, translating 27 English sentences from Wikipedia. Suggestions came from Google Translate in March 2012, and the translators were not told the source. Post-editing reduced time and improved quality in all three languages. The quality effect was strongest for Arabic and French (p < 0.001) and weaker for German (p < 0.05). Afterward, 69 percent of the translators said post-editing was the faster interface.

Measuring effort with edit distance

Snover and colleagues defined translation edit rate (TER) in 2006 as the number of edits needed to make the system output match a reference, divided by the average reference length. An edit is an insertion, deletion, or substitution of a word, or a shift of a word sequence, and each costs one. Human-targeted TER (HTER) asks a fluent speaker to edit the output until it is fluent and means the same as the reference, then measures TER against that new version. In effect, HTER scores a post-edit.

On 100 sentences from the 2004 Arabic MT evaluation, HTER averaged 33.5 against 49.6 for TER with four ordinary references, a 33 percent reduction. Building each targeted reference took 3 to 7 minutes per sentence. HTER correlated with human quality judgments at 0.630, higher than any other metric tested; the two human judges correlated with each other at only 0.478.

Edits and time are not the same thing. In the Autodesk test, the post-editor who made the most changes was also the fastest, and the authors found no clear correlation between edit distance and throughput.

What to check

• Does the contract specify light or full post-editing, and where is each level written down?
• Which edition of ISO 18587 is cited, given that the 2017 edition is marked to be revised?
• Was productivity measured against each translator's own unaided baseline, with recorded times?
• Did reviewers judge quality without knowing which text was post-edited?
• Is effort reported as an edit rate such as HTER, as time, or both, and do the two agree?

Azcuna's BABEL reads, translates, and searches foreign-language documents entirely on-device, with no cloud and no connection, and keeps every claim tied to the passage it came from.

Sources

1. 1ISO 18587:2017, Translation services - Post-editing of machine translation output - Requirements · International Organization for Standardization (abstract page)
2. 2Machine Translation Post-Editing Levels: Breaking Away from the Tradition and Delivering a Tailored Service · Mara Nunziatini, EAMT 2020, Lisbon (ACL Anthology)
3. 3A longitudinal investigation of translation trainees' translation and post-editing competence development I, section 1.1.3 Post-editing · Márta Lesznyák, Akadémiai Kiadó (2026)
4. 4A Productivity Test of Statistical Machine Translation Post-Editing in a Typical Localisation Context · The Prague Bulletin of Mathematical Linguistics, No. 93 (January 2010)
5. 5The Efficacy of Human Post-Editing for Language Translation · Green, Heer, and Manning, CHI 2013 (Stanford NLP Group copy)
6. 6A Study of Translation Edit Rate with Targeted Human Annotation · Snover, Dorr, Schwartz, Micciulla, and Makhoul, AMTA 2006 (ACL Anthology)

Where this shows up in our work

• BABEL, offline multilingual translation and search translates and searches foreign-language documents on-device, every fact tied to its source passage.