Machine Translation Post-Editing for Technical Documentation: When Is It Good Enough?
Machine translation post-editing (MTPE) is good enough for technical documentation when the content is repetitive, the terminology is fixed and a qualified linguist checks every sentence against the source before release. It is not enough on its own for safety warnings, unclear source text or new terminology. DEMA Solutions decides the level of post-editing per document type, not per project, because the risk differs from one document to the next.
Who this guide is for
This guide is for the people who sign off on how documentation gets translated: technical documentation and localization managers, technical writers, and the procurement and quality directors who approve the budget, at manufacturers of HVAC, lifting, packaging, process, automation or rail equipment. They handle manuals, spare parts catalogues, HMI strings and CE technical files in several languages.
The question is rarely “should we use machine translation?” A supplier has usually already offered it. The real questions are which documents can safely start from a machine draft, which post-editing level is being quoted, how much is really saved, what happens to confidential content and how to test it without risking a launch.
What machine translation post-editing actually is
Post editing translation is a workflow in two steps. An engine produces a first draft from the source text, and a qualified linguist then reads every sentence of that draft against the source and corrects it until it meets an agreed standard. The linguist, not the engine, is responsible for the result.
That is where most of the confusion starts. Raw machine translation is the engine’s output with nobody checking it; MTPE always includes a human who compares source and target. Some offers use the “MTPE” label for a quick read-through, or a monolingual polish with no look at the source. A monolingual reader can make a sentence sound right; only a bilingual post-editor can see that it says the wrong thing. Always ask whether the post-editor works from the source.
The international standard for post-editing, ISO 18587, sets requirements for the process of full post-editing of machine output and for the competences of the post-editors who do it. It describes full post-editing as producing a result comparable to human translation, and treats light post-editing, which aims only at an understandable text, as a different service outside its main requirements. That distinction is the starting point for every sensible MTPE decision.
Four options, not two
The real choice is between four levels of human involvement, and the useful decision is which level each document deserves.
Raw machine translation. Fast and almost free, with nobody accountable for the result. Acceptable for understanding an incoming document internally, never for anything a customer, technician or authority will read. Video platforms offer the same bargain with their automatic subtitles, and our guide compares AI subtitles with professional subtitling for published company video.
Light post-editing. A linguist corrects what is wrong or incomprehensible, without working on style or consistency. It suits internal, short-lived documents.
Full post-editing. A linguist brings the text to the standard of a human translation. This is the minimum level for customer-facing technical documentation.
Human translation with revision. A linguist translates from the source using your memory and glossary, and a second qualified linguist revises the result. The right route where a machine draft costs more to fix than it saves, or where a plausible error is too risky.
The mistake is choosing one option for everything. A supplier who proposes one route for your whole documentation set has not looked at it.
How machine output fails on technical content
Modern engines produce fluent text, and that is exactly the problem. In technical documentation, the errors that matter are often the ones that read well. A post-editor who knows where to look will find them; a reader skimming for style will not.
Terminology drifts. An engine does not know that your company calls a component a “discharge valve” and never a “relief valve” or an “outlet valve”. It picks whatever is statistically likely, and it may pick differently in chapter three and chapter seven. In a spare parts catalogue, that means a technician cannot match the manual to the part list. When the engine meets a term that is not in your glossary at all, it invents a translation that looks perfectly reasonable.
Numbers and units change silently. Decimal separators are switched, thousands separators appear or vanish, units are converted or abbreviated differently, and ranges such as “10 to 15 Nm” lose one of their numbers. Tolerances, torque values, pressures and temperatures are the details that a technician will act on without questioning, which is why they need a systematic check rather than a read-through.
Negation and modal verbs go wrong. Safety warnings depend on small words. “Do not open the cover before the machine has stopped” and “Open the cover before the machine has stopped” differ by one word, and engines occasionally drop it, especially in long sentences or where the source uses an unusual construction. “Must”, “should” and “may” carry different obligations, and a draft that turns an obligation into a recommendation changes what the user is told to do.
Tags, variables and placeholders break. Manuals exported from a CCMS or DITA source, InDesign files and HMI resource files contain tags, cross-references and variables. Engines can move them, merge them or translate what should stay untouched, such as a parameter name or an error code. The file may still open, but the layout breaks, a link points nowhere or an interface string shows a variable name to the operator. Protecting placeholders and keys is one reason why the software localization process includes internationalization checks and pseudo-localization before any string reaches a translator.
Short strings lose their context. HMI strings are often one or two words with strict length limits. “Stop” can be a command or a status; the engine guesses, and the guess has to be checked against the screen.
Unclear source is amplified. Where the source is ambiguous, the engine does not ask. It picks one reading and produces a confident sentence, and the post-editor has to notice that there was a choice at all.
Light vs full post-editing, in practice
The difference between the two levels is not just “more checking”. It is a different definition of when the job is done.
In light post-editing, the post-editor fixes what is wrong in meaning and leaves what is merely awkward. Terms may be phrased differently from one paragraph to the next and style is not harmonised with your previous documents. Nothing should mislead the reader, but nothing is polished either.
In full post-editing, the post-editor treats the machine draft as raw material. Terminology must match the approved glossary everywhere, clumsy sentences are rewritten, the text must be consistent with the approved translations in your memory, and formatting, tags and numbers are checked systematically. The output should be indistinguishable in quality from a human translation.
The practical consequence for buyers is simple. If a quote says “MTPE” without saying which level, ask. If it says “light”, check that the content really is internal and short-lived. And if the text will reach a customer, a technician or an authority, nothing below full post-editing is appropriate.
Which documents suit MTPE: a decision rule by risk
A useful rule is to ask two questions of every document type: who will act on this text, and what happens if a sentence is plausible but wrong? Volume and repetition decide whether a machine draft saves time; risk decides how much human work sits on top of it. The starting point below is for manufacturers, to adapt to your own risk assessment.
- Spare parts lists and catalogues: short, repetitive and terminology-heavy, a good fit for full post-editing provided your glossary controls the engine so part names never vary.
- Service bulletins and maintenance manuals: recurring structure and stable terms, a good fit for full post-editing with numbers, units and tolerances checked systematically.
- Installation and operating manuals (instructions for the user): suitable for full post-editing in the descriptive chapters, with safety chapters handled separately.
- Safety warnings, labels and safety chapters: low volume and high risk; human-led translation, or full post-editing followed by revision by a second linguist.
- HMI and software strings: short, context-poor and length-limited; machine drafts often need so much editing that the gain is smaller than it looks.
- CE technical files and declarations of conformity: documents that an authority or notified body may assess; human-led translation with revision.
- Marketing and sales collateral: tone and persuasion matter more than terminology; human translation or transcreation.
- Internal documents read for understanding only: light post-editing, or raw machine translation if nobody outside the company will read them.
For machinery, there is also a legal reason to treat instructions carefully. The Machinery Regulation (EU) 2023/1230, which applies from 20 January 2027, requires instructions and safety information in a language easily understood by users in the country where the machine is placed on the market. Our article on what the Machinery Regulation means for translated instructions covers those obligations, and the broader rules for manuals across EU product legislation are in our guide to user manual translation. No regulation forbids machine translation; every one of them makes the manufacturer responsible for what the translated text says.
Why controlled machine translation works better
The quality of a machine draft depends heavily on what the engine is allowed to draw on. A generic public engine has never seen your documentation. A controlled setup starts from your own language assets.
The first layer is your translation memory. Approved sentences are reused as they are, and only new or changed sentences go to the engine, so for a manual updated with each product variant, the machine may draft only a fraction of the text.
The second layer is your glossary. When approved terms are enforced, the draft uses “discharge valve” from the first sentence, and the post-editor spends time on meaning rather than replacing the same wrong term again and again. That only works if the glossary exists, is approved and is maintained, which is why keeping terminology consistent across translators and suppliers is a precondition for MTPE rather than an optional extra.
The third layer is the feedback loop. Every correction the post-editor makes, once approved, should go back into your translation memory and glossary, so the same error does not return in the next release. Without that loop, you pay for the same correction again and again.
If your existing translations are scattered across old files and previous suppliers, they can still be put to work. Translation memory alignment rebuilds a usable memory from source and target documents you already have, so the machine starts from your approved wording instead of a blank page.
What MTPE really costs, and the pricing traps
MTPE is usually cheaper and faster than human translation, but not by as much as many buyers expect, and the way it is priced can hide the difference between a real post-edit and a quick read.
Post-editing is not proofreading. A post-editor reads every source sentence, compares it with the machine output and corrects meaning, terms, numbers and tags. For technical content that is real professional work. A rate that only pays for skimming will buy skimming.
Discounts on top of discounts hide work. When machine translation is combined with translation memory matches, some pricing models apply one discount on top of another: a reduced rate for partial memory matches, then a further reduction because a machine draft exists. The quote looks low, but the effort needed to make the text publishable has not disappeared; it is underpaid, and that usually shows in the result. A partial match still needs a linguist to compare the old approved sentence, the new source and the machine suggestion, which is more thinking, not less.
Your memory changes the equation more than the engine does. Exact matches from your own approved memory are the cheapest and safest text in any project, because they have already been checked. The more of each update comes from reuse, the less the engine has to produce. That reuse only stays yours if you own your translation memory in a portable format, rather than having it sit with whichever supplier did the last job.
Ask for leverage reporting. A transparent quote shows how many words come from exact memory matches, how many from partial matches, how many are new, and which route each part takes. Procurement can then compare suppliers on the same basis, rather than on a per-word rate that means different things in different quotes. Our breakdown of what drives the cost of technical manual translation goes into that mix in more detail.
Quality checks and revision in an MTPE workflow
A post-edited text needs the same quality controls as a human translation, and in some places more, because the errors a machine makes are different from the errors a person makes.
Automated checks catch what humans miss through fatigue: numbers that differ between source and target, missing or moved tags, terms that break the glossary, inconsistent translations of the same sentence and length overruns in HMI strings. They should run on every file, with results reviewed by a linguist rather than accepted or ignored in bulk.
Revision by a second qualified linguist is the other half. For safety-relevant content, a second person comparing source and target is the most reliable defence against the fluent but wrong sentence. For lower-risk content, sample-based review can be enough, if the acceptance criteria were agreed in advance. Reused text deserves a second look in sensitive projects too: a sentence approved years ago for a previous product may no longer match the current machine, and because it looks familiar, nobody questions it. The questions worth asking a supplier about all of this are collected in how to check a supplier’s quality before you hire them.
Data confidentiality and public engines
Technical documentation often contains material a manufacturer would never publish: unreleased designs, process parameters, customer-specific configurations. Pasted into a free public engine, that text leaves your control, may be stored outside your contracts and, depending on the terms of service, may be used to improve the engine.
That is an argument for knowing which engine is used, under what contract, where data is processed, whether it is stored and whether it is used for training, and for getting those answers in writing. Internally, agree a simple rule with engineering and service teams about what may be pasted into public tools, because the most common risk is not the supplier but a colleague in a hurry.
How to pilot MTPE without risking a launch
The safest way to introduce machine translation post-editing is a small, measured pilot on real content, not a policy decision based on a demonstration.
Choose one document type that is a good candidate: recurring, terminology-heavy, customer-facing but not safety-critical, such as a maintenance manual for one product family. Pick one or two language pairs where you have reviewers you trust, ideally colleagues at a subsidiary or distributor who already read your documentation in that language.
Prepare the assets first. Make sure the glossary is approved and the translation memory up to date, and write down the acceptance criteria: which errors count, which are critical and what the reviewer checks. Agree full post-editing for anything customer-facing.
Run the pilot alongside your normal workflow, so a disappointing result does not delay a release. If budget allows, have the same content done through human translation with revision as well, and let your reviewer assess both without knowing which is which. Record turnaround, the number and type of corrections, and the time your team spent on queries.
Then review honestly. Terminology problems mean the glossary needs work before scaling; wrong numbers or tags mean the automated checks need tightening. If quality was equivalent and turnaround shorter, extend MTPE to the next document type, and keep safety chapters and authority documents on their own route.
What to ask a supplier about MTPE
Most disappointments with post editing come from expectations that were never written down. Ask which level of post-editing is included for each document type, and whether the post-editor works from the source. Ask which engine is used, how your glossary and translation memory control it, and what happens to your data. Ask who owns the memory and glossary at the end of the project, and in which format you receive them. Ask how partial memory matches and machine drafts are priced, and whether the quote shows leverage per category. Ask which automated checks run on numbers, tags and terminology, who revises safety-relevant content, and how a complaint is handled. And ask how approved corrections flow back into your assets, so the next release starts from better information.
A supplier who answers all of these clearly is managing machine translation. One who cannot is selling it. Choosing a new supplier more broadly? These ten questions for choosing a technical translation company cover the rest of the relationship.
How DEMA Solutions decides the level for each document
At DEMA Solutions, machine translation is always managed by people and is never the party responsible for the work. For each content type we look at risk, volume and intended use, and then set the route. The workflow follows five steps:
- Prepare the source and assets. Translation memory, glossary and style requirements are in place before any machine output is generated.
- Choose the route per document type. Raw output for internal understanding only, full post-editing for customer-facing documentation, human-led translation where the machine does not help.
- Post-edit with qualified linguists who know the subject and are responsible for the final text, working from the source.
- Run systematic checks on terminology, numbers, tags and formatting, and in sensitive projects check reused translations again.
- Keep the corrections. Approved changes go back into your translation memory and glossary, so the next release starts from better information.
Our machine translation and post-editing service describes the full workflow, and our quality page explains how it fits our ISO 17100 certified process, in which a second qualified linguist revises every translation. Your memories and glossaries belong to you and are delivered in portable formats (TMX and TBX), and our in-house DTP team returns post-edited files in their original formats.
We have post-edited high-volume technical, financial and legal content for years, and our translators have held a recognised post-editing certification since 2019. That experience is why we decide the route per document type rather than per project: the same client can need full post-editing for its maintenance procedures and human-led translation for its safety chapter. It is part of how we handle technical document translation as a whole.
Where TUBE fits
For manufacturers with continuous documentation, TUBE keeps machine translation under your control. Your glossary and translation memory control machine translation, so your terminology and approved wording apply from the first draft; every approved translation goes into a memory you own and can export; and leverage reporting shows how much of each update comes from text you have already approved.
Frequently asked questions
It is a workflow in which an engine produces a first translation and a qualified linguist corrects it against the source for meaning, terminology, numbers, formatting and fluency before it is used. The linguist is responsible for the final text.
Light post-editing makes the text accurate and understandable without polishing style or consistency. Full post-editing brings it to the standard of human translation, which is what customer-facing documentation needs.
For recurring manual content with stable, approved terminology, full post-editing by a qualified linguist usually works well. Safety chapters, warnings, new terminology and unclear source text are safer with human-led translation or full post-editing followed by revision.
It depends on the content, the language pair and how much of the text is already in your translation memory. Be cautious with quotes that stack machine translation and memory discounts on top of each other, because the post-editing effort does not disappear.
You can, but the output is not controlled by your glossary or memory, and confidential content may leave your control. Check the engine's terms on storage and training before sending anything sensitive.
ISO 17100 covers translation services with human translation and revision. Post-editing of machine output is described in a separate standard, ISO 18587, which sets requirements for full post-editing and for post-editor competences.
Choose the right route for each document
If you are being offered machine translation for your documentation, or already use it and are unsure about the results, tell us which documents you translate and into which languages. We will tell you where post-editing fits, where it does not, and what it would take to pilot it on your own content.
How can we help?
Tell us what you need and we’ll get back to you with the solutions.