Machine translation post-editing explained: what MTPE actually buys you

If you've read a language services tender lately, you've seen the acronym. MTPE, machine translation post-editing, has moved from industry jargon to a standard line item in government and enterprise procurement in the space of about three years. Panels now ask providers to quote it. Providers now lead with it. And a lot of the buyers writing it into requirements documents are specifying a service they haven't had defined for them.

That's a problem worth fixing, because MTPE done properly and MTPE done cheaply are different products that happen to share a name, and the difference doesn't show up until the content is published.

What is machine translation post-editing?

MTPE is exactly what the words say: a machine translates the content first, then a qualified human editor works through the output and corrects it. The machine provides speed and cost efficiency. The editor provides the accuracy and judgment the machine can't guarantee.

The commercial logic is straightforward. Machine translation output in major language pairs is now good enough that most sentences need light correction or none. Paying a professional to edit machine output costs less per word than paying them to translate from scratch, and it's faster. For high-volume content, the saving is large enough that procurement teams noticed, which is why the acronym is in your tender template.

What the tender template usually doesn't say is which kind of post-editing is being bought. There are two, and the gap between them is the whole game.

Light vs full post-editing: the distinction that decides everything

The international standard for post-editing, ISO 18587, was published in 2017 and draws the line that matters:

• Light post-editing produces output that's comprehensible and accurate in meaning. The editor fixes errors that change the meaning and leaves everything else. The result reads like what it is: corrected machine output. Grammatically odd phrasing survives if it doesn't mislead.
• Full post-editing produces output comparable to human translation. The editor corrects meaning, terminology, grammar, style and consistency. The result should be indistinguishable from content a professional translated from scratch.

Light post-editing has legitimate uses: internal documents, gist understanding, content with a short shelf life and a tolerant audience. Full post-editing is the minimum for anything published, anything carrying legal or health weight, and anything representing your organisation to the public.

The procurement trap: a tender that says "MTPE" without specifying which one has invited quotes for both. The cheapest quote is usually pricing light post-editing. The buyer is usually imagining full. Both parties are technically honest, and the community reading your published content pays the difference. A vaccination reminder that's "comprehensible but stylistically odd" reads, to the person receiving it, like your organisation didn't care enough to write to them properly. The Sydney COVID-era research we've covered elsewhere on this Journal showed translated government material scoring worse than English for the very audiences it targeted. Underspecified post-editing is one of the ways that happens.

Why is MTPE in tenders now?

Three forces converged:

1. The engines improved. Neural machine translation raised baseline quality to the point where editing beats retranslating for most general content in well-resourced language pairs.
2. Budgets didn't. Multicultural communication obligations keep expanding. Translation budgets don't expand with them. MTPE is how a language services line stretches across more languages and more content.
3. It became measurable. Edit-distance metrics like TER estimate how much correction machine output will need before the job starts, which lets providers price post-editing effort per word instead of guessing. Measurable effort became quotable effort, and quotable effort became a tender line.

The result: MTPE per-word rates typically sit well below from-scratch human translation, with the exact gap depending on language pair, content type and the machine output's starting quality. That last variable matters more than buyers expect. Post-editing terrible machine output costs more than translating from scratch, which is why serious providers assess output quality before quoting rather than publishing a flat MTPE rate for every language.

Where MTPE goes wrong

Four failure patterns account for most MTPE disappointment:

1. Fluent-but-wrong output. Modern engines produce confident, natural-sounding sentences that are occasionally wrong in ways that don't look wrong. An editor skimming fluent output misses errors a translator working from the source would never make. This is the specific skill ISO 18587 expects of post-editors: reading against the source, not just polishing the target.
2. The depth was never agreed. Light delivered where full was assumed. Covered above, still the most common failure.
3. Wrong content routed to MTPE. Post-editing suits informational content. Content where cultural fit carries the risk, campaign creative and community engagement material above all, needs the message to feel originally theirs. That content needs adaptation or creation, not correction. Routing it to MTPE saves money and loses the audience.
4. Editor credentials unspecified. ISO 18587 requires post-editors to hold translation qualifications and competence in both languages and the subject domain. A tender that doesn't ask inherits whatever the cheapest quote includes. For certified document streams, the editing must be done by the credentialled translator, not merely reviewed by one.

What to specify when you buy MTPE

Six lines that belong in any tender or contract that includes MTPE:

1. Post-editing depth: light or full, per content category, in ISO 18587 terms.
2. Post-editor qualifications: professional translation credentials in the language pair, and NAATI certification wherever the content feeds official use.
3. Quality measurement: which metrics are run on output, and what score triggers escalation to deeper review. A provider who can't answer this is grading their own homework by feel.
4. Content routing rules: which content types are eligible for MTPE and which route to human translation, adaptation or community review instead.
5. Terminology assets: whose glossary and translation memory the work runs against, and who owns the memory the work creates. The memory is an asset; don't leave it in the provider's hands by silence.
6. Sampling and audit rights: a percentage of delivered content independently reviewed each quarter.

A communications manager who puts those six lines in a tender has done more for output quality than any evaluation criterion about the provider's technology.

How this works on SourceTarget

On the platform, MTPE is the middle rung of the ladder rather than a separate product. Content is machine translated instantly with a confidence score. That score, together with the content type, decides what happens next: publish, route to expert post-editing, or escalate to community review. The post-editing tier is priced per word by review depth, the depth is explicit before payment, and every delivered job carries a quality score with a plain-English explanation of what was checked.

The routing is the point. The question isn't whether MTPE is good or bad. It's which content deserves it, and the honest answer is: a lot of yours does, and some of yours never will. A platform's job is to know the difference before the invoice, not after the complaint.