Language technology
Machine Translation Post-Editing Services
Our machine translation post-editing services pair private neural MT engines with specialized human post-editors working to ISO 18587. You get 40–60% cost reduction on high-volume technical content, and a straight answer about which content should never go near an engine.
Every project is scoped as light or full post-editing per the ISO 18587:2017 definitions, with the deliverable quality level written into the quote before work starts.
Your files never touch free public MT. We run access-controlled engines under a no-retention agreement, so your text cannot train anyone else's model.
We classify your content by consequence of error first. Safety warnings and patent claims are routed to full human translation, and we tell you why in writing.
What Our Machine Translation Post-Editing Service Actually Is
Machine translation post-editing, or MTPE, is a defined production method: a neural MT engine produces the first draft, and a professional post-editor with subject-matter background corrects it segment by segment against the source. It is not a discount label stuck on unchecked machine output, and it is not a translator glancing at MT suggestions. The method has an international standard, ISO 18587:2017, which specifies what post-editors must do, what qualifications they need, and what the two service levels mean. We scope every MTPE project against that standard, which is unusual; most vendors use the term loosely and deliver whatever the deadline allows.
Techniwords has offered MTPE as a managed service since neural engines became viable for technical content, building on 15 years of technical translation work for US industrial and software clients. The post-editors are the same specialized linguists who translate for us, not a cheaper second bench. That matters because a post-editor who cannot translate the sentence from scratch cannot reliably judge whether the engine got it right. MTPE sits alongside our other production methods in the full services catalog, and on many accounts we run it in parallel with conventional translation, each method assigned to the content it suits.
Light vs. Full Post-Editing Under ISO 18587
ISO 18587 defines two service levels, and the difference is contractual, not cosmetic. Light post-editing produces text that is accurate and comprehensible: the post-editor fixes mistranslations, omissions, and additions, but leaves awkward phrasing alone if the meaning is correct. Full post-editing produces text comparable to human translation: terminology is enforced against your termbase, style is corrected, and the output must be indistinguishable in use from a conventionally translated document. Your quote names the level explicitly, so there is never a debate at delivery about what was promised.
| Criterion | Light post-editing | Full post-editing |
|---|---|---|
| Goal per ISO 18587 | Comprehensible, accurate text | Output comparable to human translation |
| Terminology | Corrected only where meaning fails | Enforced against the approved termbase |
| Style and fluency | Left as generated if understandable | Edited to publication quality |
| Typical content | Support tickets, internal wikis, RFI drafts | Published docs, UI strings, customer-facing text |
| Typical saving vs. translation | 50–60% | 30–45% |
Most first-time buyers ask for light post-editing and actually need full. If the text will be read by customers, auditors, or regulators, light PE is a false economy: the reputational cost of one clumsy paragraph outweighs the per-word saving. We will say so during scoping rather than let you find out from your distributor in Monterrey.
Content Triage: What We Post-Edit and What We Refuse To
The most valuable thing we do on an MTPE account happens before any engine runs: we sort your content by the consequence of an undetected error. Neural MT fails in a specific way. It produces fluent, confident sentences that are occasionally wrong, and fluency makes the errors hard to spot. For some content that risk is acceptable because a post-editor will catch what matters. For other content the risk is structurally unacceptable, because a single missed negation can injure someone or void a patent claim.
- Good MTPE candidates: support tickets and knowledge-base articles, internal engineering wikis, e-commerce and parts listings, supplier correspondence, and repetitive structured content such as the part descriptions we handle in bill of materials translation.
- Case-by-case: release notes, requirement specifications, tender annexes, and developer-facing content of the kind covered on our page about translating software documentation, where placeholders and code samples need protection but the prose tolerates MTPE well.
- Never through MT at our shop: safety warnings and hazard statements, patent claims, informed consent text, and certified documents. These go to full human translation with independent revision, whatever the volume.
This triage is written into your quote as a content map. Clients in logistics and supply chain operations often see the sharpest split: shipment exception messages and carrier updates are ideal MTPE material, while dangerous goods declarations are not, even though both live in the same TMS export.
Private Engines and Confidentiality
We never run client content through free public MT portals. Free tools pay for themselves with your data: text submitted to them can be retained and used to train future models, which for unpublished specifications or pre-launch documentation is a disclosure you cannot take back. Our production engines are commercial systems accessed under enterprise agreements with contractual no-retention terms, running in access-controlled environments. Where an account justifies it, we train custom engine profiles on your own approved translations, which raises raw output quality measurably in narrow domains.
Engine choice is also a language decision. Output quality varies widely by pair and direction: English into Spanish or German starts strong, while English into Simplified Chinese needs more careful engine selection and heavier post-editing effort, a point we cover in depth on our Chinese technical translation page. We benchmark two or three engines on a sample of your real content before committing, because the best engine for automotive German is rarely the best engine for semiconductor Chinese.
One number worth knowing: on a well-matched engine, a qualified post-editor sustains 7,000–9,000 words per day, versus 2,000–2,500 for conventional translation. That throughput difference, not a cheaper hourly rate, is where legitimate MTPE savings come from.
The Real Economics: Where 40–60% Savings Come From, and When They Evaporate
Honest MTPE pricing starts from throughput. Because a post-editor moves three to four times faster than a translator on suitable content, we can price MTPE at 40–60% below conventional translation and still pay qualified specialists properly. The savings are real and repeatable, but only inside the conditions that create them: clean source text, a well-matched engine, a language pair the engine handles well, and content where light or full PE is genuinely sufficient.
The savings evaporate in predictable situations, and we would rather list them than surprise you. When source text is poorly written or riddled with ambiguity, the post-editor slows to translation speed. When a language pair produces weak raw output, full PE effort approaches retranslation and the discount shrinks to 10–15%, at which point we will recommend conventional translation instead. Heavily tagged files, embedded screenshots, and creative marketing copy all erode throughput. And on small projects under roughly 5,000 words, engine setup and QA overhead eat most of the margin, so MTPE only makes sense as part of an ongoing program. Combined with the discount grid from our translation memory services, mature MTPE accounts often see blended per-word costs fall by two thirds against their year-one baseline, because TM matches and MT output compound.
How We Measure Quality: Sampling and MQM
Saying "a human checked it" is not quality assurance. On MTPE accounts we run structured sampling: a second linguist audits a randomized 10% sample of delivered segments against the MQM (Multidimensional Quality Metrics) error typology, logging errors by category (accuracy, terminology, fluency, style, locale conventions) and severity. Each project gets a score, scores are tracked over time, and a drop below the agreed threshold triggers a full re-review at our cost. This is the same framework our reviewers apply when auditing third-party translations, which is described in detail on this page.
MQM data also feeds back into production. If terminology errors cluster, the termbase gets tightened or the engine profile retrained. If accuracy errors cluster in one content type, that type gets moved up the triage ladder to full PE or human translation. One electronics client saw exactly this: fluency scores were excellent across 60,000 words of component documentation, but sampling caught recurring polarity errors in capacitor specifications, so datasheet tables were reclassified as human-translation content while the surrounding prose stayed on MTPE. Firms in the electronics and semiconductor sector tend to adopt this measured approach fastest, because they already think in terms of sampling plans and defect rates.
A Typical MTPE Project, Step by Step
- Content audit and triage. We review a representative sample, classify content by risk, and map what goes to MTPE, what goes to human translation, and at which PE level.
- Engine benchmark. Two or three candidate engines translate a 2,000-word sample of your real content; a specialist scores the raw output and we pick the winner per language pair.
- Asset preparation. Your termbase and translation memory are connected so the engine and post-editor work against approved terminology from segment one.
- Post-editing. Specialized post-editors work in professional CAT tools at the contracted ISO 18587 level, with TM matches taking precedence over MT output wherever they exist.
- Sampled QA. Independent MQM audit of a randomized sample, plus automated checks on numbers, units, tags, and untranslatables.
- Delivery and feedback loop. Files return in the source format; audit findings update the termbase, the engine profile, and the triage map for the next batch.
Our process follows ISO 17100 for the human stages, meaning the people doing the work are qualified translators with documented subject expertise, and as a member of ATA and GALA we staff post-editing to the same standard as translation. Steps one and two are typically complete within a week; from then on, batches flow on agreed turnarounds.
File Formats and Pipeline Integration
MTPE lives or dies on file handling, because the content that suits it usually arrives in bulk from a system rather than as a tidy Word document. We ingest XLIFF 1.2 and 2.0, bilingual exports from Phrase, memoQ, and Trados, JSON and CSV dumps from helpdesk platforms such as Zendesk and Salesforce Knowledge, XML and DITA from component content systems, and structured exports from PIM and ERP systems for product data. Before anything reaches an engine, a preparation pass locks placeholders, inline tags, product codes, and untranslatable strings, so the engine cannot mangle a variable and the post-editor never wastes minutes rebuilding markup. Deliveries return in the same structure they arrived in, ready for reimport without manual patching on your side.
For recurring programs we set up standing pipelines rather than one-off transfers. A helpdesk client on a monthly cadence, for example, drops an export into a monitored exchange folder; the batch is analyzed against the TM, machine translated, post-edited, sample-audited, and returned on a fixed day each month, with a one-page batch report showing volume, match distribution, MQM score, and spend. API-based handoffs are available where your platform supports them. The point is to remove per-batch project management from your desk entirely: after setup, the program runs on schedule and you read a report instead of writing briefs.
Pricing and Turnaround
MTPE is quoted per source word at the contracted PE level, after TM analysis. A typical structure on an established account: full PE at roughly half the conventional translation rate, light PE lower still, with TM matches discounted on top under the standard grid. Engine benchmarking and setup are one-time line items on program accounts and waived above volume thresholds. Turnaround runs three to four times faster than conventional translation for the same team size, which is often the deciding factor: a 100,000-word knowledge base that would take a translation team five weeks moves through MTPE in under two.
What never changes with the method is transparency. Your quote itemizes the content map, the PE level per content type, the engine used, the sampling rate, and the per-word rate for each bucket. If we think MTPE is the wrong tool for part of your content, that bucket is priced as human translation and the quote says why. From our Texas base we cover all US time zones, and batch programs can be scheduled around your release calendar.
Machine Translation Post-Editing FAQ
When is MTPE a good fit, and when is it not?
MTPE fits high-volume, repetitive, lower-risk content: support articles, internal documentation, parts listings, supplier correspondence. It is a poor fit for safety-critical text, patent claims, certified documents, marketing copy, and small one-off projects under about 5,000 words, where setup overhead cancels the savings. The deciding question is the consequence of an undetected error. If a missed negation could injure someone or create legal exposure, we route that content to full human translation and say so in the quote.
What is the difference between light and full post-editing?
Both are defined in ISO 18587:2017. Light post-editing delivers accurate, comprehensible text: errors of meaning are fixed, but awkward phrasing stays if it is understandable. Full post-editing delivers text comparable to human translation, with terminology enforced against your termbase and style edited to publication quality. Light PE suits internal or ephemeral content; anything customers, auditors, or regulators will read should be full PE. Your quote states the level explicitly so the deliverable is never ambiguous.
Is our content safe, or does it end up training someone's public model?
We never use free public MT portals, whose terms typically allow retention and reuse of submitted text. Our engines are commercial systems under enterprise agreements with contractual no-retention terms, running in access-controlled environments, and NDAs cover every linguist on your account. Where volume justifies it, we train private engine profiles on your own approved translations; those profiles are yours and are never shared across clients. Unpublished specifications stay unpublished.
How much do we actually save with MTPE?
On well-suited content, 40–60% against conventional translation: light PE toward the top of that range, full PE typically 30–45%. The savings come from throughput, since a post-editor sustains 7,000–9,000 words per day versus 2,000–2,500 for translation. They shrink when source text is messy, the language pair produces weak raw output, or files are heavily tagged. When projected savings fall below about 15%, we recommend conventional translation instead, because at that margin you are paying MTPE prices for translation effort.
How do you control quality on machine-translated content?
Through measurement, not assurances. A second linguist audits a randomized 10% sample of every delivery against the MQM error typology, scoring accuracy, terminology, fluency, style, and locale conventions by severity. Scores are tracked per project and per content type; falling below the agreed threshold triggers a full re-review at our cost. Automated checks cover numbers, units, tags, and untranslatables on 100% of segments. Audit findings feed back into the termbase, the engine profile, and the content triage map.
Can you post-edit machine translation output we generated ourselves?
Yes, with one condition: we first benchmark a sample of your raw output. If your engine produces reasonable quality for the language pair and domain, we quote post-editing at the appropriate ISO 18587 level. If the raw output is below the threshold where post-editing beats retranslation, we show you the sample comparison and quote translation instead. Roughly a third of client-generated MT we evaluate falls into that second category, usually on harder language pairs or specialized vocabulary.