Machine Translation Post-Editing: Levels, Workflow, Pricing


Machine translation post-editing (MTPE) is the work a professional translator does on raw machine output: reading it against the source, fixing errors, and shaping the text until it meets an agreed quality bar. Done well, it gives you machine speed with a human guarantee of accuracy. The key decision is the level: light post-editing for text that only has to be understood, full post-editing for text that has to read as if a person wrote it.
This guide covers what post-editing is, how the two levels differ, what a clean workflow looks like, which machine errors an editor hunts for first, and how to match the level to your content.
TL;DR
Post-editing means a human linguist corrects machine-generated translation. Revision is the term for correcting a human translation.
ISO 18587 defines two levels: light (comprehensible text) and full (quality comparable to human translation).
Choose the level by asking who reads the text and what happens if a sentence is wrong.
A reliable workflow runs source review, machine pass, accuracy edit, terminology and style, then a dedicated check of numbers, names and formatting.
We offer full MTPE from €0.04 per word and light MTPE from €0.03 per word, handled by one experienced translator.
Table of Contents
What machine translation post-editing is
Post-editing is the correction of machine-generated translation by a human. The European Commission’s translation knowledge centre separates it from revision: revision is applied to human translations, post-editing to machine output. In both cases the aim is to correct errors, improve fluency and adapt the text for its audience.
The practical difference is the material in front of the editor. ISO 18587:2017 points out that post-editing involves three texts, not two: the source text, the machine output and the final target text. The editor judges the machine’s proposal against the source, sentence by sentence, instead of translating from a blank page.
The idea is older than most people think. The professional association AIETI notes that the term dates back to the 1950s and has always been tied to machine translation. What has changed is the quality of the raw material. Neural and AI-based engines now produce fluent text, and that fluency is exactly why a trained human eye matters: a fluent sentence can still carry a wrong number, a dropped clause or the wrong term.
Sharon O’Brien, a leading post-editing researcher, describes the task as fundamentally a revision job. The editor must understand the source sentence and the machine’s proposal, identify errors, decide how to fix them and carry the fixes out, ideally with as few edits as possible while fully conveying the source meaning.

Light vs. full post-editing
ISO 18587 defines both levels. Full post-editing aims for a product comparable to one obtained by human translation. Light post-editing aims for a text that is merely comprehensible, with no attempt to make it comparable to human translation.
Light post-editing | Full post-editing | |
Goal | Understandable, accurate in meaning | Reads like a human translation |
What the editor fixes | Meaning errors, wrong terms, omissions, critical numbers | All of that, plus grammar, style, tone, consistency, polish |
Typical content | Internal documents, support tickets, reference material, large archives | Websites, marketing, manuals, contracts, regulatory and medical text, software UI, games |
Reader’s tolerance | Awkward phrasing is acceptable | Awkward phrasing is a defect |
Our published rate | from €0.03 per word | from €0.04 per word |
A useful rule: light post-editing protects meaning, full post-editing protects meaning and reputation. If a customer, regulator, patient or player will read the text, it needs the full level. If a colleague just needs to grasp what a document says, light is usually enough.
Neither level is a shortcut around accuracy. Even in a light pass, numbers, dates, dosages, names and units are checked, because those are the errors that cost money.
How to choose the right level
The cleanest test is two questions: who reads the output, and what happens if a sentence is wrong. Apply them to your own content:
Internal, low risk. Meeting notes, internal knowledge bases, supplier emails for understanding. Light post-editing fits.
External, brand-facing. Product pages, newsletters, app stores, onboarding flows, e-learning. Full post-editing, because tone and consistency are part of the product. If you localize courses, see our guide to e-learning translation.
Regulated or safety-relevant. Medical, legal, clinical and chemical safety documents. Full post-editing, with terminology locked down beforehand. Our guides to clinical trial translation and safety data sheet translation show what these documents demand.
Creative. Video game dialogue, slogans, humor. Here the machine draft is a starting point for cultural adaptation, which is localization rather than editing alone.
Mixed projects are common. A software company might send the interface strings and help center for full post-editing, and a ten-year backlog of support tickets for light post-editing. Splitting the work this way puts the budget where readers are.
A step-by-step MTPE workflow
Providers describe their pipelines a little differently, but a dependable one always contains the same stages. ISO 18587 also covers the work before production: deciding whether the source content suits machine translation at all, and optionally pre-editing the source to improve the raw output.
1. Scope and source check
Agree the level (light or full), the audience, the deadline and any glossary or style guide. Look at the source: clean, consistent, plainly written text produces better machine output. Ambiguous sentences, unfinished drafts and heavy wordplay do not.
2. Machine pass
The text goes through the engine, ideally inside a CAT tool, so that segments sit with the source on one side and the machine draft on the other. Where you already have a translation memory, approved segments are reused rather than re-machined. ISO 18587 describes both fully machine-translated and integrated environments where machine output and translation memory work together with consistent terminology management.
3. Segment-by-segment decision
For every segment the editor makes one of three calls: accept it, edit it, or retranslate it. A good editor does not polish what is already right, and does not patch a sentence that needs rebuilding.
4. Accuracy first
The first editing priority is meaning: omissions, additions, mistranslations and invented content. Roy Green, writing from the Commission of the European Communities, concluded that accuracy should be the top consideration in every situation. Everything else is secondary to that.
5. Terminology, grammar, style
Next come the client’s approved terms, then grammar and syntax, then style and tone. This is where full post-editing separates from light: a light pass stops once the text is correct and understandable. See our practical guide to terminology management for how glossaries keep this consistent across projects.
6. Verification pass
A separate read looks only at the things machines get wrong without anyone noticing: numbers, dates, proper nouns, URLs, units, punctuation, tags and formatting. Automated QA in the CAT tool flags mismatched numbers, untranslated segments, double spaces and odd length ratios. A human then confirms each flag.
7. Delivery and feedback
The final file is delivered in the original format. Recurring machine errors go into notes for next time, so the glossary and instructions improve with every project.
Machine translation errors a post-editor looks for
“Bad machine translation” is rarely random. Researchers have catalogued the same families of error for decades. According to O’Brien, engines produce grammatical and syntactic errors, unnecessary additions or omissions, wrong lexical or terminological choices, and errors of collocation or style.
Green’s classic study grouped errors by how much trouble they cause the editor:
Minor: a wrong article, preposition or pronoun, or a slightly off word choice. Quick fixes.
Major: idioms translated word for word, or words misread as the wrong part of speech. These need a rewrite of the clause.
Grey areas: doubtful translations and near misses. Green found these cause the most trouble, because the editor has to decide subjectively whether to change them and how far to go.
That last category matters more with modern AI output. Today’s engines make fewer obvious mistakes and more near misses: a sentence that sounds right, reads smoothly and is subtly wrong. The remedy is a trained specialist who knows the field. A medical editor recognizes that a drug name or dosage unit looks plausible but is off. A legal editor notices when a term of art has been replaced with an everyday synonym.
Typical problem spots
Numbers and units. A wrong figure reads perfectly well, which makes it the most dangerous error.
Terminology drift. The same source term rendered three different ways across one document.
Dropped or added content. Whole clauses omitted, or invented filler added.
Register and tone. Formal where the brand is friendly, or the reverse.
Tags and placeholders. Broken variables, markup or UI strings in software and games.
Cultural mismatch. Jokes, idioms, units of measure, date formats and forms of address that do not travel.
Confidentiality, style and cultural adaptation
Three questions come up on almost every MTPE project.
Where does the text go? Public, free machine translation tools may not be suitable for contracts, patient data or unreleased products. Decide in advance which engine handles the material and who sees it. We cover the practical side in our guide to confidential translation services.
Whose style wins? Give the editor a glossary, a style guide and a few approved sample texts. AIETI points out that post-editing needs clear criteria for how the task is carried out and for the effort involved, without lowering the quality of the final text. Written instructions are what make a result repeatable.

Does it need adapting, not just correcting? Marketing copy, game dialogue and e-learning often need localization: changing examples, humor, references and formats for the target market. Post-editing fixes the machine’s errors. Localization decides what the audience should actually read.
Perspective: why one specialist makes MTPE work
We see the same pattern in nearly every MTPE project: the machine draft is the easy part, and the judgment is the work. A single experienced translator who knows the subject, whether it is a technical manual, a clinical document or a game script, catches the near misses that a generalist reads straight past. Because one person reads the whole project, terminology stays consistent from the first segment to the last, and nobody has to reconcile five editors’ choices at the end. We work in modern CAT tools with built-in quality assurance, which lets us take on high-volume projects without trading away accuracy. We also separate the passes: meaning first, then terminology and style, then a dedicated check of numbers, names and formatting. Our clients often have urgent deadlines, so we are reachable beyond regular business hours and we quote deadlines we can keep. Confidentiality is part of the job, not an add-on. Bernadett
Our MTPE and translation services
We handle post-editing and translation for technical, medical, legal, automotive, financial, marketing, software, website, multimedia and video game content, working between Hungarian and English, and also German and French into and out of Hungarian. Our published rates start at €0.04 per word for full post-editing and €0.03 per word for light post-editing. Human translation into Hungarian starts at €0.05 per word, and into English, German or French at €0.07 per word. Localization is quoted per project.

If you have a machine-translated file, a backlog to process or a launch with a tight deadline, send it over through our MTPE and translation page and tell us the audience and the level you have in mind. We will confirm the approach and a realistic deadline.
Getting the best results from MTPE
Send clean source text. Final, consistent source files give better machine output and fewer edits.
Share reference material. Glossaries, previous translations and style notes are worth more than any setting in the engine.
Name the audience. Customers, regulators, engineers and players all need different registers.
Pick the level per content type. Do not apply one level to everything.
Keep a feedback loop. Recurring errors become glossary entries and instructions.
Keep the same editor where you can. Familiarity with your product and terms speeds up every later project and improves consistency.
FAQ
What is the difference between post-editing and proofreading?
Proofreading checks a finished text for surface mistakes. Post-editing compares machine output with the source and corrects meaning, terminology, grammar and style. A proofreader may never look at the source text, while a post-editor works with it at every segment.
Is light or full post-editing better?
It depends on the reader and the risk. Light post-editing makes text accurate and understandable, which suits internal and reference material. Full post-editing targets quality comparable to human translation, which is the right choice for anything customers, patients, regulators or players will read.
Can AI-translated documents be used as they are?
For private understanding, often yes. For anything published, signed or relied on, a professional should review it, since fluent AI output can still contain wrong numbers, omissions or the wrong term. Post-editing checks exactly those points before the text goes out.
How is MTPE priced?
Most providers price per word, with a lower rate than full human translation. Our rates start at €0.03 per word for light and €0.04 per word for full post-editing. Final quotes depend on the content type, the quality of the machine draft and the deadline.
Sources
Revision and post-editing, Knowledge Centre on Translation and Interpretation, European Commission
ISO 18587:2017, Translation services: Post-editing of machine translation output
How to deal with errors in machine translation, Sharon O’Brien
The MT errors which cause most trouble to posteditors, Roy Green
Recommended

Comments