Machine translation post-editing (MTPE) is the process in which a qualified human linguist reviews and corrects text produced by a machine translation engine so that it meets a defined quality target. Among AI applications in artificial intelligence, it is one of the most mature examples of human-in-the-loop production: the neural engine drafts at high speed, and the post-editor supplies the judgment on meaning, terminology and tone that the model cannot guarantee.
MTPE is not proofreading a finished translation and not translating from scratch. It sits between the two, and its economics depend on how good the raw output is, how much quality the client actually needs and how carefully the effort is measured.
How the post-editing workflow operates
A post-editing workflow passes source text through an engine, then routes the raw machine output to a linguist who edits it inside a computer-assisted translation (CAT) tool. The tool shows source and target segments side by side, together with translation memory matches and approved terminology, so the editor can compare each machine suggestion against trusted references.
Modern engines are almost always neural machine translation (NMT) systems or general large language models prompted to translate. Their output is usually fluent, which is precisely the danger: fluent errors such as a dropped negation, a wrong number or an invented term read naturally and are easy to miss. The post-editor therefore checks meaning against the source first and style second.
The typical stages are shown below, from preparation to delivery.
| Stage | What happens | Who or what does it |
|---|---|---|
| Pre-processing | Cleaning source files, protecting tags and placeholders | Project manager, scripts |
| Engine selection | Choosing a generic or domain-adapted engine per language pair | Localization engineer |
| Raw translation | Engine translates segments not covered by memory | NMT engine or language model |
| Post-editing | Linguist corrects meaning, terminology and style | Professional post-editor |
| Quality assurance | Automated checks for numbers, tags, terminology and consistency | QA tool plus reviewer |
| Feedback loop | Edited segments feed memories and engine retraining | Localization team |
Light and full post-editing levels
Post-editing is defined by its target quality level, agreed before work starts. The international standard ISO 18587 sets requirements for full post-editing, and industry guidelines distinguish it from light post-editing, which aims only at understandable and accurate text.
The distinction matters because the two levels differ in effort, price and suitable content. Mixing them up is a frequent cause of disputes between buyers and vendors.
| Level | Quality goal | Typical content | What the editor ignores |
|---|---|---|---|
| Light post-editing | Accurate and understandable meaning | Internal documents, knowledge base triage, gisting | Awkward but correct style, minor repetition |
| Full post-editing | Quality comparable to human translation | Product interfaces, contracts, published documentation | Nothing that affects meaning, style or terminology |
| Transcreation review | Adapted message for a target market | Marketing slogans, campaigns | Literal fidelity, when creativity requires it |
For regulated or high-risk text, such as medical labeling, legal agreements or safety instructions, full post-editing is normally followed by an independent second review. A single pass over machine output is rarely sufficient where an error can cause harm or liability.
Measuring post-editing effort and productivity
Post-editing effort is measured with time-based and edit-based metrics, because the number of words alone says nothing about how much work the machine output required. Pricing, engine evaluation and vendor comparison all depend on these measurements.
| Metric | What it measures | How it is used |
|---|---|---|
| Edit distance | Share of characters or words changed by the editor | Comparing engines and adjusting rates |
| HTER | Human-targeted translation edit rate against the final edit | Research and engine benchmarking |
| Throughput | Words edited per hour or per day | Planning capacity and deadlines |
| Keystroke and pause data | Typing activity and hesitation in the CAT tool | Detecting cognitively difficult segments |
| Error annotation | Errors classified by type and severity | Final quality acceptance |
Productivity gains vary widely. For well-supported language pairs and repetitive technical content, editors often process roughly one and a half to three times the volume of translation from scratch. For low-resource languages, creative text or poor source quality, the gain can shrink to nothing, and editing may even be slower than translating.
How MTPE differs from related practices
MTPE is one of several ways to combine machines and people in translation, and each assigns responsibility differently. The table separates it from neighboring approaches that are often confused with it.
| Approach | Starting point | Human role | Main risk |
|---|---|---|---|
| Raw machine translation | Engine output only | None or spot checks | Undetected meaning errors |
| Machine translation post-editing | Engine output | Systematic correction to a target level | Anchoring to the machine wording |
| Interactive translation | Engine suggestions while typing | Translator writes, engine predicts | Over-acceptance of suggestions |
| Human translation with memory | Previously approved segments | Translator writes new text | Higher cost and longer turnaround |
| Pre-editing | Simplified source text | Writer adapts source for the engine | Unnatural source style |
Related AI applications such as AI dubbing and AI writing assistants reuse the same pattern: a model drafts, and a human edits to an agreed standard. MTPE is distinctive because translation quality can be judged against an explicit source text.
Choosing content and engines for post-editing
MTPE pays off when the content is repetitive, structured and terminologically stable, and when the engine has been tested on representative samples. A short pilot with measured edit distance is the most reliable way to decide whether a language pair and content type are suitable.
| Situation | Recommended approach | Reason |
|---|---|---|
| Technical manuals in a major language pair | Full post-editing | High engine quality and stable terminology |
| Customer support tickets for internal triage | Light post-editing or raw output | Speed matters more than polish |
| Marketing campaign copy | Human translation or transcreation | Tone and cultural fit dominate |
| Clinical trial documents | Full post-editing plus second review | Regulatory exposure and patient safety |
| Low-resource language pair | Pilot first, then decide | Engine quality is unpredictable |
Domain adaptation usually improves results more than switching vendors. Training or prompting an engine with the client’s translation memories and glossaries reduces terminology errors, which are among the most time-consuming corrections for editors.
Limitations and common mistakes
The main weaknesses of MTPE come from human factors and from the nature of neural output rather than from the editing method itself. Experienced teams watch for the following problems.
- Priming and anchoring: editors tend to keep machine wording that is acceptable but unidiomatic, which flattens style across large projects.
- Hidden omissions: neural engines sometimes skip a clause entirely, and a fluent target sentence hides the gap.
- Unfair pay models: flat discounts per word ignore segments that need complete retranslation, which damages editor motivation.
- Confidentiality: sending client text to a public engine can breach data protection obligations and contracts.
- Cognitive fatigue: correcting many near-correct sentences is tiring, and attention drops in long sessions.
A sound program addresses these risks with secure engine deployment, effort-based compensation and automated QA checks for numbers, negations and terminology.
Stripped to the essentials
A computer can translate a page in seconds, but it does not truly understand what it reads. It may get a number wrong, drop a word like “not” or use the wrong technical term while still sounding confident. Post-editing means a trained translator reads the computer’s version next to the original and fixes these mistakes.
Think of it like a first draft written by a fast assistant. The machine saves typing; the human guarantees meaning. How much the human polishes depends on what the text is for: a quick internal note needs only correct meaning, while a published manual needs to read as if a professional wrote it.
Where it is applied
A manufacturer of industrial pumps needs its 300-page maintenance manual in German, Polish and Spanish within three weeks. The localization team runs a pilot on 2,000 words per language with an engine adapted using its existing translation memories, and measures an edit distance of about 20 to 25 percent for German and Spanish, but over 40 percent for Polish.
The team assigns full post-editing for German and Spanish, where editors deliver roughly twice their usual daily volume, and chooses conventional translation with memory support for most of the Polish content. Every language then goes through automated checks for torque values, part numbers and warnings. The manuals ship on schedule, and the corrected segments are fed back into the memories to improve the next update.