Machine Translation Post-Editing (MTPE): Termbase, Sampling Rates, and Scorecard

Bottom line first: Machine translation plus human post-editing (MTPE) can cut the per-piece cost of multilingual content to 30–40% of fully manual translation. There is only one condition: build the termbase first. Spend 2–3 days building 100–200 industry terms, drop the sampling rate from 100% on the first piece down to 10%–20%, and set the pass mark at 4.0 on a five-point scale. Regulatory text, warranty terms, quotations, and contracts stay out of the MTPE pipeline — if it needs a human, use a human.

1. Without fixed terminology, post-editing is just repeated work

How to set localization standards for multilingual content is covered in Three Criteria for Localization; which gates to clear before launch is covered in Quality Gates for Multilingual Launches; and how to handle numbers and units in tables and charts is covered in i18n for Multilingual Charts and Tables. This article covers only the step in between:once machine translation produces a draft, how a human takes over, how much to sample, and how to score and archive quality.

The MTPE math has to be done up front. A 1,500-word technical article takes about 6–8 hours to translate manually; the machine draft takes 1 minute and post-editing takes 2–3 hours, saving more than half the work hours. But that math only holds if the post-editor does not agonize over wording when they get the machine draft — every “servo system,” “die-cutting accuracy,” and “EU Declaration of Conformity” has exactly one correct translation, and only sentence structure and word order change. If the same batch of articles renders servo system three different ways, the editor has to stop and verify every time, and post-editing degrades into inefficient manual translation that costs more than getting it right the first time.

So the order of operations is non-negotiable:termbase first, machine translation second, post-editing third. Do not start batch MTPE before the termbase has at least 100 entries. Run one test piece first to calibrate the process.

2. The Minimum Viable Termbase and a Build Checklist

You do not need to buy a system for your termbase. A three-column online spreadsheet is the minimum viable form:

Chinese term	English term (final)	Rejected translations
Servo system	servo system	servo motor system; servo drive
Die-cutting accuracy	die-cutting accuracy	cutting precision; die cutting tolerance
EU Declaration of Conformity	EU Declaration of Conformity	CE certificate; CE report
Lead time	lead time	delivery period; production cycle
Tolerance class	tolerance class	tolerance grade; tolerance level

The “Do Not Use” column is just as important as the “Approved” column next to it — it keeps out the wrong translations that machine translation engines and earlier drafts have already put into circulation. If a spot check turns up a banned translation, the piece fails outright. No discussion needed.

Start your termbase with these four categories, in priority order:

Term category	Suggested volume	Source	Who verifies
Products and model numbers	40–60 entries	Product catalogs, spec sheets	Technical engineers
Process and component names	40–60 entries	Production and process documents	Technical engineers
Certifications and standards	20–40 entries	Original certificates, standard texts	Quality or customs
Commercial and delivery terms	30–50 entries	Contract templates, quotes	Export manager

That’s 130–210 entries in total, and you can build it in two or three days. Take the official English wording from certificates and contracts wherever possible — don’t let translators invent their own. If the certificate says EU Declaration of Conformity, the termbase has to say the same, and editors have no authority to change it.

3. Which terms belong in the termbase, and which ones you don’t translate at all

There’s only one test for whether a term belongs in the termbase:If a mistranslation would be noticed by the customer or create commercial ambiguity, it goes in. A parameter off by one digit, a certification name written wrong, a vague delivery term — these three kinds of errors cost you trust the moment an overseas buyer reads your email. By contrast, purely rhetorical wording doesn’t need to go in the termbase; leave that to the post-editor’s ear.

The counterpart to the termbase is the do-not-translate list. These items must be locked and kept as-is before machine translation:

Do not translate	Why it stays as-is	Post-editing note
Product and part numbers	Change one character and it’s a different product	Check character by character against the source
Trademarks and brand names	Legally protected marks; translating them risks infringement	Match the official capitalization and spelling
Standard and certification numbers	EN 60335-1, UL 94 V-0 — get them wrong and the claim is false	Keep the spaces and hyphens in the number format
Units of measure	Unit conversion follows its own rules; it isn’t translation	Space between number and unit follows the target language convention
Company names and registered addresses	Follow the business license and official website	Match the website, including punctuation

There are two ways to implement the do-not-translate list: use your translation engine’s term-locking feature, or replace the listed items with placeholders before machine translation and restore them afterward. The first is less work but depends on engine support; the second is safer but you have to make sure no placeholders survive into the final draft — a leftover placeholder in a spot check is a critical defect, and the piece goes straight back for rework.

4. The three levels of post-editing and their time factors

“Post-editing” isn’t one action — it’s three levels of depth. When you get a machine-translated draft, decide which level applies first. Pick the wrong one and you either waste hours or miss the quality bar:

Level	What we do	What you don’t do	Time factor	Where it applies
Light	Fix only mistranslations, omissions, word order, and numbers and units	Leave sentence structure and word choice alone	0.3–0.4 (relative to full human translation)	Parameter pages, spec descriptions, catalog entries
Medium	On top of light, rewrite sentences and unify tone	No localization rewriting	0.5–0.6	Standard blog posts, news, product intros
Deep	Rewrite for the target market’s phrasing on top of medium editing, adapting examples and metaphors	Facts and specs stay unchanged	0.7–0.8	Landing pages, industry insights, bylined articles

What the three coefficients mean: medium post-editing takes about 2.5–3 hours for a 1,500-word article, while full human translation takes 6–8 hours — that gap is where MTPE’s cost savings come from. The cost comparison across the three paths is summarized below:

Path	Hours per article (1,500 words)	Cost level (vs. full human translation)	Quality consistency
Full human translation	6–8 hours	100%	Depends on the translator, low variance
Light post-editing	2–2.5 hours	30%–35%	Consistent terminology, flat phrasing
Medium post-editing	2.5–3 hours	45%–50%	Consistent terminology, phrasing close to original
Deep post-editing	4–5.5 hours	70%–80%	Close to original writing, close to human translation cost

The deep tier costs nearly as much as full human translation, so use it only when an article will be excerpted by AI or run on paid traffic — paying deep-tier hours for a category page is a waste. For most sites, the steady-state setup is: light for category and spec pages, medium for blog posts, and for landing pages pick either deep or original writing depending on budget.

You also need to tier within a batch: in a batch of 10 articles, running deep before light is the wrong order and will cause problems. The right order is run 1–2 articles at the deep tier first to calibrate terminology and tone, then put the rest through medium or light. The phrasing locked in during the calibration articles becomes new termbase entries, and the later articles get faster to edit each time.

5. Sampling rate: from 100% on the first article down to 10%–20%

Sampling is the core of MTPE quality control. The rate isn’t a constant — it steps down as the partnership progresses:

Stage	Sampling scope	Sampling rate	Purpose
First article	Every sentence, full text	100%	Calibrate terminology and tone baseline
First 10 articles	First half of each article plus 2 random passages	50%	Confirm consistency
Articles 11–30	Random passages	20%	Maintain deterrence
Steady stage	Random passages	10%	Routine monitoring
Change translator or engine	Back to first-article standard	100%	Recalibrate

Where you sample matters:Sampling must cover the opening 200 words, the first sentence under each heading level, and every sentence containing a number. Machine translation and post-editing errors aren’t evenly distributed — the first sentence under a heading is where editors most often cut corners, and number sentences are where errors cost the most. Hit rates in these two spots are usually more than 3 times higher than in random passages.

Both downgrades and upgrades go by score: 5 consecutive articles scoring above 4.5 drops the sampling rate one tier; a single article scoring below 3.5 means the whole article is redone and that translator goes back to 50% sampling for the next 5 articles. Put these rules in your contract with the translator or translation team — that works better than asking verbally.

6. The five-dimension scorecard and passing line

Sampling needs one consistent ruler: five dimensions, 1–5 points each, weighted total of 4.0 is the passing line:

Dimension	Weight	5-point standard	3-point line (triggers corrective action)
Terminology accuracy	30%	All final translations and whitelist terms hit, zero banned translations	1–2 terminology deviations
No factual errors	25%	Numbers, units, model numbers and standard numbers match the source	1 number or model number deviation
Sentences read smoothly	20%	No translationese; sentence breaks follow target-language conventions	A few sentences need a second read
Consistent tone	15%	Tone matches published content; person and honorifics are consistent	Noticeable tone shift
Links and formatting	10%	Internal and external links resolve, anchor text is correct, formatting follows site conventions	1 broken link or leftover placeholder

The pass mark is a weighted score of 4.0, and the no-factual-errors item must not score below 4— a piece that clears the total but fails the factual dimension still goes back for rework. The reason this is a hard veto is simple: terminology and fluency are quality issues, but a factual error is a business incident. Score records are filed per piece in each translator’s profile, so after three months you can see which dimension each translator is weak in and give them targeted samples and feedback. That is far more useful than saying “improve quality” in general terms.

7. Closing the loop on issues, where machine translation is off-limits, and three cases where you should skip it entirely

Issues found in spot checks have to go into a closed loop, or the same mistakes will repeat every batch. The loop has three steps:

Steps	Action	Timeframe	Output
Update the terminology base	Add new mistranslations to the banned list and missing terms to the approved list	Same day	Terminology base version +1
Rescan the batch	Search all delivered files from the same batch for the new banned terms	Within 3 days	Rescan record
Log it in the issue database	Record the error type, translator and resolution	Same week	Material for the monthly quality report

The rescan step is the one most often skipped, and it costs the most. When the same mistranslation appears in 20 published articles, fixing them one by one after the fact costs more than 5 times what it would have cost to catch it at delivery, and pages already live are cached by Google and AI engines.

Four content types must never enter the MTPE pipeline: legal and compliance statements, warranty and liability clauses, quotes and contract text, and safety warnings. What these four have in common is that every word carries legal consequences. Machine translation plus post-editing cannot cover the boundaries of liability, so translate them by hand and have the business owner review and sign off.

Finally, three cases where you should skip machine translation altogether:The target language is the source content(there is no source text to machine translate, such as a local feature written directly for the German market, which follows the original writing workflow);High-value landing pages(pages on the conversion path, where deep rewriting should already be budgeted as original writing);Key copy under 150 words(headlines, button copy, one-line selling points, where editing machine output takes longer than writing it from scratch). For how to work out the plugin and architecture costs of a multilingual site, see the accounting approach in The Hidden Costs of Multilingual Plugins.

FAQ

How many terms does a terminology base need before it is good enough?

There’s no magic number — look at coverage instead. Pull 5 articles in a row; if terminology hits 95% or better, that’s enough. For industrial exporters, 150–250 entries usually cover the backbone of technical docs and blog posts. If your product lines span several industries, build a separate glossary per industry and maintain each one independently. Once you blow past 500 entries, you’ve usually let in a pile of rhetorical phrasing — that kind of term is better left to the post-editor, and putting it in the glossary is over-management.

We’re switching to a new machine translation engine. Do we have to rerun the whole process?

Recalibrate, yes — but you don’t have to rerun everything. After the switch, take 3 past articles, re-run MT and post-editing on them, and compare terminology hit rate and edit volume against the old versions. If the new versions need 20% less editing, switch over and go back to normal sampling. If the difference is small or the new engine is worse, stay on the old one. For the month you switch, put sampling back up to 50%, then drop it back down after 10 stable articles in a row.

What do we do with articles that fail sampling?

For scores between 3.5 and 4.0, flag the problem paragraphs and send them back for revision, then recheck only the revised sections. Below 3.5, or if a factual error triggers an automatic fail, rework the whole article and re-sample the full text. Put the clause that rework hours aren’t paid extra in your contract. Log every case in the issue database and the translator’s file. If the same translator scores low on the same dimension in two consecutive batches, pause the collaboration or replace them.

Our multilingual site has a dozen languages. Does every language need this whole process?

Tier them by commercial weight. Main languages — usually English plus your two biggest markets — get the full MTPE process. Secondary languages run a lighter version: sampling cut in half, tone weighted lower, but terminology and factual accuracy stay at full strength, because a wrong spec or certification is the same disaster in any language. For languages with zero inquiries, stop publishing first — don’t keep a pipeline running just to hit a metric.

Do post-editors need technical knowledge?

For light and medium editing, no — the glossary is enough. For deep editing, yes, because you have to judge whether a machine-translated sentence like “the guide roller pressure is set to 0.3MPa” actually makes sense. Someone with only a language background will accept it as-is. In practice: language editors handle the wording, and each batch gets one technical reviewer who checks only numbers, model codes and spec sentences — 1–2 hours per batch, no need to be involved the whole way through.

Should we archive machine-translated drafts?

Yes. Keep three things, filed by article: the raw MT draft, the post-edited final version, and the scoring sheet. Keeping the raw MT draft serves two purposes: it gives you a baseline for comparison when you switch engines or translators, and if a factual dispute comes up, you can pinpoint whether the error came from the machine or the person. Archive for 12 months, stored alongside your sampling records — this is also what you settle and reconcile against with your translation team.

Final thoughts

MTPE isn’t a way to “save money with machines” — it’s a division of labor that puts human time where the machine gets things wrong. The termbase is the foundation of that division, the sampling rate is the dial on trust, and the scorecard is the arbiter of disputes. Get all three in place and machine translation becomes controllable productivity; leave any one out and the hours you save come back doubled in rework and customer complaints.

Where to start: this week, build a three-column glossary and get it to 100 entries. Next week, take a mid-difficulty article through the full workflow — machine translation, post-editing, 100% sampling, scorecard on file — and log the actual hours at each step. That’s your own cost model.

One line worth remembering: the tighter you pin down your terms, the freer your post-editors are — the quality ceiling of MTPE is set the day you build the termbase.