Machine Translation Post-Editing: Light vs. Full PE

Machine translation post-editing (MTPE) is the process of a human editor correcting machine-translated text to a defined quality target, instead of translating the source from scratch. There are two recognized levels — light and full post-editing — and they call for very different amounts of intervention. Only one of them, full post-editing, has a binding international standard (ISO 18587:2017) behind it. This guide covers what separates the two levels, what that standard actually requires, how a standards-aligned MTPE workflow runs end to end, and when the whole exercise is worth it.

What is machine translation post-editing?

Post-editing is not the same job as translating, and it isn't the same job as editing human-written text either. ISO 18587:2017 defines it narrowly: to "post-edit" is to edit and correct machine translation (MT) output. The process works with three texts at once — the original source text, the raw MT output, and the final target text the post-editor produces — which is why post-editing has its own skill set. A post-editor has to judge the source, judge the MT output, and decide how much of the raw text survives, segment by segment.

The practice is decades old. Institutions such as the European Commission and the Pan American Health Organization were already assigning humans to correct machine-generated output by the late 1970s, and the first academic studies on post-editing followed in the 1980s, as the field's documented history shows. What changed the calculus more recently is the underlying MT technology: neural machine translation produces noticeably more fluent raw output than the statistical and rule-based systems that preceded it, which is a large part of why post-editing moved from a niche correction task to a mainstream part of the localization workflow. For more on how the field's ambitions have shifted over time, see our piece on machine translation between dream and reality.

Light post-editing vs. full post-editing

Not every MTPE job aims for the same finish line. ISO 18587 defines the two levels precisely: full post-editing is the process of post-editing "to obtain a product comparable to a product obtained by human translation," while light post-editing is the process of post-editing "to obtain a merely comprehensible text without any attempt to produce a product comparable to a product obtained by human translation." In plain terms: full PE has to read as if a human wrote it — correct grammar, natural style, verified terminology. Light PE only has to be accurate and understandable; awkward phrasing is acceptable as long as the meaning is intact.

The split isn't always clean in practice. The original TAUS/CNGL post-editing guidelines frame it as a function of two variables rather than a fixed checklist: how good the raw MT output already is, and how good the client actually needs the final text to be. Very poor raw output may never reach "good enough" quality even after light post-editing, while very good raw output might need only light touches to become publishable — the labels describe a target, not a guaranteed amount of work.

What ISO 18587 actually requires

ISO 18587:2017, "Translation services — Post-editing of machine translation output — Requirements," is the standard most guides cite when they mention MTPE — but its scope is narrower than it's usually presented. The standard states its own scope directly: it "provides requirements for the process of full, human post-editing of machine translation output and post-editors' competences," and it is "only applicable to content processed by MT systems." Light post-editing is defined in the terms and definitions section and described further in an informative annex, but the standard sets no binding requirements for it — a provider can be assessed against ISO 18587 for full post-editing work, not for light post-editing work.

For the process it does govern, ISO 18587 requires a translation service provider to, among other things:

• Assess whether the source content is genuinely suitable for MT and post-editing before accepting the job, since combined MT/post-editing efficiency depends on the engine, language pair, domain, and style of the source.
• Finalize and document the agreement with the client, including the target quality level and project specifications, in writing.
• Integrate translation memories and terminology management so the MT output and the post-editor's corrections both stay consistent with approved client terminology.
• Assign a post-editor with documented competence — in translation, in the relevant domain and language pair, in post-editing judgment specifically, and in using terminology and translation-memory tools.
• Run a verification step before delivery and record feedback on the MT engine's recurring errors.

That competence requirement is worth underlining: the standard treats post-editing as a distinct skill from translation, not an automatic byproduct of being bilingual. Part of a post-editor's documented task is deciding, segment by segment, when to patch the MT output and when to discard it and retranslate instead.

The MTPE workflow, step by step

Stripped of certification language, a standards-aligned MTPE workflow runs through the same stages whether the target is light or full post-editing:

1. Assess suitability

Not all content is a good MTPE candidate. The provider checks the MT engine's likely output quality for that language pair and domain, the content's purpose, and the turnaround window before committing to post-editing over translation from scratch.

2. Set the target and specifications

Client and provider agree on light or full post-editing, the intended audience, any "do not translate" terms, and formatting requirements — documented up front rather than inferred once the post-editor has already started.

3. Prepare the source and run MT

Where practical, the source text is pre-edited for clarity — fixing ambiguity or non-standard formatting — since cleaner input produces cleaner raw MT output and less post-editing work downstream. The text then goes through the MT engine; general-purpose engines like DeepL and Google Translate (compared in our DeepL vs. Google Translate guide) handle straightforward content well, while regulated or high-volume workflows more often use a domain-tuned or custom-trained engine instead.

4. Integrate translation memory and terminology

Existing translation memories and glossaries are loaded so the MT output and the post-editor's corrections both stay aligned with terms the client has already approved — including product names and other proper nouns a generic MT engine has no way to recognize as untranslatable on its own.

5. Post-edit against the agreed level

The post-editor works through the MT output against the light or full target: correcting meaning errors, added or dropped information, and offensive or culturally inappropriate content in every case, and additionally restructuring sentences and polishing style when the target is full post-editing.

6. Verify, deliver, and feed back

A reviewer checks the edited text against the original specification before delivery. Many providers also log which error types the MT engine produced most often on that job, since that feedback is what lets both the engine and the next post-editing batch improve.

When MTPE makes sense — and when it doesn't

MTPE tends to work well for content where accuracy and turnaround matter more than creative voice: technical documentation, product manuals, support articles and knowledge bases, internal communications, and the large volumes of e-commerce or user-generated content that would never get translated by hand at all. It also suits content with a short shelf life, where "good enough, fast" beats "polished, late."

It tends to work poorly for marketing and advertising copy, where tone and cultural resonance matter more than literal accuracy; for legal, regulatory, or contractual text, where a single mistranslated clause carries real liability; and for literary or highly creative writing, where the goal is closer to re-creation than translation. In those cases, industry guidance consistently points toward human translation and revision instead, regardless of how fluent the raw MT output looks.

The decision isn't purely a judgment call, either — see how MT output quality itself gets measured, and where automated scoring stops being reliable, in our guide to evaluating machine translation quality.

How MTPE rates are usually structured

Post-editing is rarely billed the way translation from scratch is, and anyone researching post-editing rates is usually trying to understand the model rather than a number — actual rates vary too much by language pair, domain, volume, and provider for any single figure to be meaningful here. Three structures are common across the industry:

• Discounted per-word rate. The post-editor is paid a fraction of the standard translation per-word rate, on the logic that starting from MT output takes less time than translating from a blank page. The discount is usually steeper for light post-editing than for full.
• Hourly rate. Used when raw MT quality is inconsistent enough that per-word pricing would be unfair to the post-editor in either direction, or for smaller, non-repetitive jobs.
• Effort- or tier-based pricing. Some providers price by measured edit distance or a raw-MT quality score, so a clean segment that needs almost no correction costs less than one the post-editor effectively has to retranslate.

Whichever structure is used, post-editing rates tend to run lower than rates for translation performed entirely from scratch — a point professional bodies such as the International Association of Professional Translators and Interpreters have pushed back on publicly, arguing that post-editing effort is harder to predict and price fairly than translation is. Treat any "cheaper than translation" figure a vendor quotes as a negotiating starting point, not a fixed fact about the work involved.

Tools that support MTPE

Post-editing happens almost entirely inside CAT (computer-assisted translation) tools rather than as a standalone step. The features that matter specifically for MTPE are an integrated MT connector so raw output appears next to the source segment instead of in a separate window, translation memory and terminology matching that flags approved terms automatically, and — increasingly — machine translation quality estimation, which scores each segment's likely reliability so post-editors and project managers can prioritize where human attention goes first. See our comparison of machine translation software for how the major engines and CAT-integrated platforms differ on exactly these points.

Frequently asked questions about machine translation post-editing

Is machine translation post-editing worth the effort?

It depends on the content and the quality bar you need. Post-editing is generally reported as faster than translating from scratch, but the size of the time saving is disputed: industry sources have reported savings around 40%, while controlled academic studies have found actual savings under real working conditions as low as 0-20%, and occasionally none at all when the raw MT output needs heavy correction. The safest approach is testing MTPE on a representative sample of your own content before committing a full project to it.

What are the guidelines for machine translation post-editing?

The two reference points the industry uses are ISO 18587:2017, which sets binding requirements for full post-editing and for post-editor competence, and the original TAUS/CNGL post-editing guidelines, which describe target quality levels for "good enough" and human-translation-level output. Neither sets one universal quality bar — both frame the target quality as something the client and provider negotiate per project.

What are the three main types of machine translation?

Rule-based machine translation (RBMT) relies on hand-built grammar and dictionary rules; statistical machine translation (SMT) generates output from probability patterns learned across large bilingual text collections; and neural machine translation (NMT), the current standard approach, uses neural networks trained end-to-end to produce a translation directly. Most engines post-editors work with today, including the major consumer and enterprise tools, are neural.

What is SMT and NMT?

SMT (statistical machine translation) builds translations by statistically modeling how words and phrases correspond across large volumes of previously translated text. NMT (neural machine translation) replaced it as the dominant approach, using deep neural networks to model translation as a single learned function from source to target language — which generally produces more fluent, though not automatically more accurate, output than SMT did.

What skills does a machine translation post-editor need?

ISO 18587 describes the post-editor as needing translation competence, domain and language competence matched to the content, judgment specific to post-editing — deciding how much of the MT output to keep, correct, or discard — and the ability to use terminology and translation-memory tools, as a breakdown of the standard's competence records lays out. In practice most post-editors are trained translators, though the profession is still relatively young, and there's no settled industry consensus on whether a translation background or bilingual domain expertise makes for a better post-editor.

For the tools that make the workflow above practical at scale, start with our guide to the best machine translation software, then see how MT output actually gets measured in machine translation quality evaluation.