Machine Translation Post-Editing Services (MTPE)

When you have a large volume of content and a tight timeline, fully human translation of every word isn’t always the right tool — but raw machine translation is a real risk for anything a reader relies on. Machine translation with post-editing (MTPE) is the middle path: the speed and cost of MT, made trustworthy by a qualified human editor.

Taika Translations delivers MTPE across 300+ languages, with quality levels matched to how each piece of content will actually be used. And when your content shouldn’t go through MTPE at all — certified documents, legal, medical, marketing — we’ll say so and route it to standard human translation or certified translation instead. Tell us your volume, languages, and timeline and we’ll recommend the mix.

What is MTPE? Machine translation post-editing, defined

MTPE is machine translation post-editing: a machine translation engine produces the first draft, and a qualified human linguist then edits that draft to a defined quality level. You will see the same workflow written several ways — post-editing machine translation, post-edited machine translation (PEMT), MT post-editing, or simply post-editing. They describe one thing.

What separates MTPE from “we ran it through a translation tool” is accountability and method. A post-editor works with the source text open, not just the output. They are qualified in the language pair, they work against your glossary and translation memory, and they are briefed on which level of editing the content needs before they start. The deliverable is a human-approved translation, not a machine draft that someone skimmed.

It is also a genuinely different skill from translating from scratch. A translator builds meaning; a post-editor diagnoses it — reading machine output critically enough to spot a sentence that is fluent, confident, and wrong. Fluency is exactly what modern engines are best at, which is why unreviewed output is dangerous: the errors that survive are the ones that read well.

The standard behind it: ISO 18587

MTPE is one of the few parts of the language industry with a dedicated international standard, and knowing what it does and does not cover is the single most useful thing a buyer can take into a vendor conversation.

ISO 18587:2017 — Translation services — Post-editing of machine translation output — Requirements (ISO/TC 37/SC 5) sets out requirements for the post-editing process and for post-editors’ competences. It applies only to content produced by a machine translation system, and its requirements are written for full post-editing — the level intended to produce a result comparable to professional human translation.

Two things about it are worth knowing before you compare quotes:

• Light post-editing sits in an informative annex, not in the requirements. The standard describes it, but does not specify it as a conformance level the way it does full post-editing. A vendor can therefore describe a project as following the standard while delivering either level of work. So the quote should name the level, not just say “post-edited.”
• The standard is being rewritten for AI. A second edition is in development as ISO/DIS 18587, at draft international standard ballot stage with ISO members, and it will replace the 2017 edition. It is retitled Translation services — Post-editing of non-human translation output — Requirements. The shift from “machine translation” to “non-human translation” is deliberate: the revision is being drafted to cover output from generative AI systems, not only from conventional MT engines. If your content is now being drafted by a large language model rather than a translation engine, that is the standard that will govern editing it.

Taika’s post-editing practice follows the levels and process discipline this standard describes — post-editors qualified in their language pair, a named editing level agreed before work starts, and no raw engine output delivered to a client. Taika meets and exceeds ISO 17100, 9001, 27001, and 18587 standards. (That describes how we work. Unlike Section 508, third-party ISO 18587 certification does exist and some providers hold it — so to be plain: this is process alignment, not a certificate we hold.)

Light post-editing vs full post-editing — and how to tell which one you are buying

This is where most MTPE disappointment comes from. The two levels are not “good” and “better.” They are aimed at different outcomes, and buying the cheaper one for a publication job does not produce a slightly rougher publication — it produces something that should not be published.

Light post-editing aims at text that is comprehensible and accurate, for gisting and internal use. The post-editor retains as much of the raw machine output as possible, ensures nothing has been added or omitted, removes inappropriate content, and restructures sentences whose meaning is wrong or unclear. Style is explicitly not the goal. The result reads like a machine draft that is now reliably correct.

Full post-editing aims at a result comparable to professional human translation: accurate, consistent with your terminology, stylistically appropriate for the audience, and fit to publish under your name.

Four questions settle which you need, and they are worth asking before you request a quote:

1. Who reads it, and does it carry your brand? Anything a customer, constituent, or regulator reads under your name is a full post-editing job.
2. Is it a decision document? If someone acts on it — a procedure, a safety instruction, an eligibility notice — the answer is full, regardless of volume.
3. Does it have a shelf life? Content that will sit in a help centre for three years is worth editing once, properly.
4. Is it being read, or searched? Internal archives and research material that people scan for a single fact are a legitimate light post-editing case.

If a quote does not tell you which level it prices, ask. “Post-edited” on its own is not a specification, and the gap between the two levels is the largest single variable in an MTPE price.

What actually decides whether MTPE saves you money

MTPE is sold as a discount and is often assumed to be one automatically. It isn’t. The saving comes from how much of the machine draft survives editing, and that varies far more than most buyers expect. Four things drive it:

• The language pair. Machine translation quality is a function of how much training data exists. High-resource pairs such as English–Spanish or English–German produce drafts a post-editor can largely keep. Lower-resourced languages produce drafts that take longer to repair than to redo — and those are the same languages where human capacity is scarcest. The awkward arithmetic of MTPE is that it saves least exactly where budgets are tightest.
• How the source is written. Clean, consistent, plainly written source produces good machine output. Long sentences, undefined acronyms, internal shorthand, marketing register, and text extracted from slides or tables all degrade it sharply. Editing the source before translation is often the cheapest intervention available, and it improves every target language at once.
• Subject matter and terminology. An engine has no way to know your term for a component, a benefit programme, or a product tier. Without a glossary it will translate each occurrence plausibly and inconsistently, and the post-editor pays for that in every segment. A maintained glossary and translation memory is the difference between MTPE getting cheaper each cycle and costing the same every time.
• Content type. Repetitive, structured, high-volume content is what MTPE was designed for. Short content is the worst case: the overhead of setting up and quality-checking the workflow doesn’t amortise across a few hundred words.

There is a real failure mode here worth naming: negative saving, where the post-editor spends longer diagnosing and rebuilding bad output than they would have spent translating the source properly. It happens most often in low-resource pairs, in specialist technical or clinical content, and in anything persuasive. When your content looks like that case, we say so at quoting and recommend standard human translation instead. That recommendation costs us the cheaper sale, and it is the reason we can be trusted on the projects where MTPE genuinely is the right call.

This is also why MTPE is quoted per project rather than from a rate card. A vendor who quotes an MTPE rate before seeing your content is pricing an assumption about the four variables above.

What should never go through MTPE

Some content is excluded not because the workflow could not process it, but because the consequences of a surviving error are not recoverable:

• Certified documents — birth and marriage certificates, diplomas, immigration paperwork. A certified translation carries a signed attestation of completeness and accuracy that a receiving authority relies on. It is produced by human translation with independent review.
• Legal filings and contracts. Meaning is load-bearing at the clause level, and an engine has no model of what a negation, a defined term, or a limitation of liability is doing. See legal translation.
• Medical and clinical content. Dosages, contraindications, and patient instructions. See medical translation.
• Marketing, brand, and persuasive copy. MTPE optimises for correctness, not for effect. Copy that has to persuade needs translation that can depart from the source deliberately — which is close to the opposite of what a post-editor is instructed to do.
• Anything with a regulatory filing attached, where a reviewer will compare source and target line by line.

Taika flags content in these categories at quoting and routes it to full human translation with the two-linguist translate–edit–proofread process. This is scoping, not upselling — we will also tell you when content you assumed needed full human translation is a perfectly good MTPE candidate.

Where MTPE fits in a bigger programme

MTPE is rarely the whole answer on its own. For organisations translating continuously rather than in one batch, it works best as one tier inside a programme that also carries full human translation and certified work — with shared terminology across all of it, so the glossary and translation memory improve every tier at once. See managed localization for the ongoing-programme version, enterprise localization services for large multi-language deployments, and website and software localization when the content lives in a product rather than in documents.

Tell us your volume, languages, timeline, and where the content will appear, and we will recommend the tier mix — including the parts that should not be post-edited at all.