You submit a post-edited document. 24 hours later, the client’s editor sends it back with 17 comments. You fixed every error the MT engine made - mistranslations, grammar problems, missing segments. So where did 17 comments come from?
Usually from the gap between “technically corrected” and “publishable.” Those are two different places, and most post-editing complaints live in the space between them.
Here’s what publishable actually means, what the standards require, and how to know when you’ve genuinely reached it.
What “publishable quality” means¶
Publishable quality in full post-editing means the output is indistinguishable from what an expert human translator would produce for the same text. Not just grammatically correct - not just “no mistranslations” - indistinguishable.
ISO 18587:2017, the international standard for post-editing of machine translation output, defines it directly:
“Full post-editing shall produce a quality output comparable to that of a human translation performed in accordance with ISO 17100.”
ISO 17100 is the standard for professional human translation. So your post-edited output has to meet the same bar as translation done from scratch by a professional. That’s the official definition.
In practice, publishable means a fluent reader of the target language can’t detect machine origin. They don’t stumble over awkward phrasing. They don’t notice terminology shifting between chapters. They don’t get the sense of something mechanically generated - because after full post-editing, it shouldn’t be.
This is different from light post-editing, where the goal is intelligibility: the reader can understand without being misled. That’s a lower bar - and it’s designed to be.
Full post-editing vs light post-editing: the quality gap¶
The MTPE process distinguishes two levels, and the quality targets are completely different.
Light post-editing (LPE) targets functional accuracy. You fix: - Mistranslations that change meaning - Anything that blocks comprehension - Terminology violations against required glossaries - Missing or added content
What you deliberately skip in light PE: stylistic awkwardness, non-critical word choices, phrasing that sounds mechanical but stays understandable. Light PE is appropriate for internal reports, raw data exports, knowledge base drafts. Not for anything published externally.
Full post-editing (FPE) targets publishable quality. You fix everything in LPE, plus: - Unnatural phrasing (grammatically fine, but sounds machine-generated) - Register mismatches (formal text accidentally rendered too casual, or vice versa) - Terminological inconsistency (same concept, three different renderings) - Culturally awkward expressions - Number, date, and measurement formatting for the target locale - Structural issues the MT pass left intact
The rate difference reflects the work difference. Full PE rates run $0.08-0.15 per word. Light PE is as low as $0.02-0.05/word. That’s not arbitrary - full PE requires more judgment calls, more domain expertise, and multiple review passes.
One translator summarized the difference in an industry discussion:
“Light PE is triage - you’re finding the bleeds. Full PE is surgery - you’re fixing everything that’s wrong, including the things that aren’t visibly wrong yet.”
That’s accurate. Light PE is faster because the scope is narrow. Full PE is better-paid because you’re closing every quality gap, not just the critical ones.
The error taxonomy: critical, major, and minor¶
The MQM framework (Multidimensional Quality Metrics), now incorporated into ISO 5060:2024, classifies translation errors by severity. Understanding this taxonomy tells you exactly which errors block publishable quality and which are just polish.
Critical errors - these make the translation unusable: - Mistranslations that change meaning in a dangerous or misleading way - Missing content (a sentence, a clause, a key condition) - Additions not in the source - Terminology violations in regulated domains (medical, legal, financial)
Any critical error means the translation is not publishable. Full stop.
Major errors - these significantly degrade quality: - Noticeable grammar errors (agreement, tense, word order) - Terminological inconsistency across the document - Register mismatches that feel jarring to a native reader - Culturally inappropriate phrases - Incorrect number or date formatting
A publishable translation has zero critical errors and very few major errors. Industry practice typically allows a maximum of 1-3 major errors per 1,000 words depending on content type. Legal and medical content often requires zero major errors.
Minor errors - noticeable but don’t significantly impair quality: - Slightly awkward phrasing - Inconsistent punctuation style within accepted conventions - Non-critical word choice preferences - Very minor formatting issues
A publishable translation can contain a small number of minor errors - native writers of the same document would make similar choices. The issue is when minor errors accumulate into a pattern that signals machine origin.
The DQF-MQM framework used by TAUS and major LSPs adds a fourth category: neutral observations - issues to flag but not count as errors, typically stylistic preferences where both choices are defensible.
Five quality dimensions every post-editor needs to cover¶
ISO 18587 evaluates quality across five dimensions. If any single dimension fails, the translation isn’t publishable - even if the other four are perfect.
1. Accuracy Does the translation convey the full, correct meaning of the source? No mistranslations, no omissions, no unintended additions. Accuracy is the non-negotiable floor - below accurate is already a critical error.
Specific watch points: negations (MT engines frequently drop “not”), numbers (0 vs. 00 is easy to miss in a long document), conditionals (“if/when/unless”), and passive vs active voice shifts that change meaning.
2. Fluency Does the translation read naturally in the target language? Not just grammatically correct - fluent means a native speaker reads it without pausing or rereading. This is where MT output most commonly fails the publishable bar even after accuracy is fixed.
Practical check: close the source and read the target alone. Mark every place you pause or re-read. Those are fluency problems. Then check whether the issue is from the MT draft or from over-literal editing that preserved an awkward source structure.
3. Style and register Does the translation match the tone and formality level the audience expects? A technical user manual for engineers needs different language than the same manual adapted for general consumers. MT engines tend to produce a neutral-formal register that often doesn’t fit the actual target text type.
Ask: what’s the content type, and what register does it require? Read 3-5 sentences and ask whether a native editor for that content type would have written exactly this.
4. Terminology Are domain-specific terms used correctly and consistently throughout? Correct means the right term for the concept. Consistent means the same term every time. MT engines are inconsistent - “contract” might be rendered three different ways in the same document when it appears in different sentence positions.
Dedicated terminology pass: after the main editing pass, search for each key term and verify consistency. For client-supplied glossaries, check every item on the list was applied correctly.
5. Formatting and conventions Numbers, dates, currencies, measurements, punctuation - correct for the target locale. German uses periods as thousand separators and commas as decimal (1.234,56); English uses the opposite (1,234.56). French adds a space before colons and question marks; English doesn’t. MT engines frequently copy source formatting without adapting it.
What ISO 18587 actually requires¶
ISO 18587:2017 goes beyond output quality. It specifies requirements for the post-editor, the process, and the project brief.
Post-editor competencies required by the standard: - Linguistic competency in both source and target language at professional level - Translation competency - understands equivalence, register, cultural adaptation - Technical competency - knows how MT systems work and what they typically get wrong - Domain knowledge - understands the subject matter enough to catch content errors
Worth emphasizing: ISO 18587 requires post-editors to have the same language competencies as professional translators under ISO 17100. This matters because many MTPE offers in the market come with lower competency requirements and lower rates. The standard doesn’t support that framing.
Process requirements: - Pre-assessment of whether MT output is suitable for post-editing before the project starts (some output is so poor that full PE takes longer than translation from scratch) - A documented brief including quality specifications, style guide, glossary, and format requirements - A final quality check before delivery
On the gap between the standard and industry practice, researchers in JosTrans found:
“Current post-editing guidelines across the industry reflect significant variation in how ‘full post-editing’ is defined operationally - a challenge that the ISO standard addresses at the principle level but not at the granular operational level.”
ISO tells you what publishable quality means. The client brief tells you the project-specific details of how to get there.
A draft revision, ISO/DIS 18587.2, is currently under development to address AI translation tools beyond traditional MT systems. The core quality principles are expected to remain.
The publishable checklist: 6 passes before you submit¶
Here’s a practical sequence for closing the gap from “corrected MT” to “publishable.” The order matters.
Pass 1: Accuracy (source vs target in parallel) Read source and target segment by segment. Mark anywhere meaning diverged. Don’t fix fluency yet - just accuracy. If the MT engine was decent, this pass is fast.
Pass 2: Completeness Verify every source segment is present in the target. MT engines occasionally drop segments, particularly across paragraph breaks or with complex tables. In your CAT tool, check for empty or untranslated segments.
Pass 3: Terminology Work through your glossary or term list. Verify every required term was used and used consistently. If the client supplied a glossary - check every item. This is the pass most post-editors skip under time pressure, and where most terminology-related client complaints come from.
Pass 4: Fluency (target only - source closed) Close the source. Read the target as a native reader would. Mark anywhere it reads unnaturally. Then fix. This pass is what separates “corrected MT” from “publishable” - you can’t do it in parallel with the source because source proximity makes you read what you think it says, not what it actually says.
Pass 5: Formatting and conventions Go through every number, date, currency, and measurement. Check punctuation conventions for the target locale. Check header capitalization rules (title case in English, sentence case in most other languages). Units and abbreviations.
Pass 6: Final read Read the whole document quickly, as a reader - not an editor. Does it read like it was written by a human? If you’d attach a caveat to the client - it’s not done. If you’d put your name on it without hesitation - it’s publishable.
Common MT errors that survive post-editing unnoticed¶
Most post-editing mistakes aren’t missed corrections. They’re patterns that become invisible through familiarity with the source text. The ones that consistently pass through full PE reviews:
False friends in the target language MT engines pick technically correct words that don’t fit domain usage. In legal text going into French, “actuel” means “current” but looks like “actual” - a classic trap that survives because the word individually looks fine. These slip through because post-editors familiar with the source domain read through context, not form.
Register drift in long documents MT engines aren’t consistent across 30-50 pages. The first 10 pages might render formal; the middle section slightly informal; the end formal again. Segment-by-segment review makes this drift invisible. The target-only read-through in pass 4 is the only way to catch it.
Number formatting in mixed-language documents When a document contains source-language terms mid-text (common in legal and technical content), MT engines sometimes copy source number formatting into the translated sections. A German document with English-style number formatting (1,234.56 instead of 1.234,56) - obvious to a native reader, easy to miss in a bilingual review.
Pronoun and agreement errors spanning sentences MT handles grammatical gender at the segment level. It makes agreement errors when references span sentence boundaries. In Spanish, French, German, Ukrainian - adjectives referring back to nouns in the previous sentence often carry the wrong gender. Segment-by-segment review doesn’t catch cross-sentence agreement.
Literal idioms MT translates idioms literally when no clean equivalent exists. The result is grammatically correct but strange to a native reader. Post-editors reading for meaning (which you should) don’t flag it as wrong - you understand what was meant. A native reader without the source context would find it odd.
Where “good enough” ends and “publishable” starts¶
Honestly, the exact line depends on content type and client specs. But here’s a workable framework:
Publishable = a native-speaker editor for that content type would accept it with few or no changes.
Not publishable = that editor would mark up the margins. Not for missed errors - but because the text has subtle patterns of machine origin that a professional would notice.
The gap is almost always fluency and register. Most post-editors catch accuracy errors. Fewer catch fluency issues in their own working direction - you read what the sentence should mean, not what it literally says. Register drift across a long document is genuinely hard to notice when you’re reviewing segment by segment.
Two practical tests:
- The stranger test: would you send this to a client you’ve never worked with before, with zero caveats? If you’d add “please check the style” - it’s not publishable.
- The reversal test: could someone who’s never seen the source tell this was translated? If yes - it’s not publishable yet.
Full PE rates exist because reaching this bar consistently requires experience, multiple review passes, and real domain knowledge. In 2025, about 50% of professional translators declined MTPE at discounted rates, citing quality expectations that don’t match the time required to reach publishable standard. That gap between what “post-editing” pays and what publishable quality actually costs is an ongoing tension in the industry.
FAQ¶
What’s the difference between full post-editing and proofreading?¶
Proofreading assumes a human translation as the starting point - you’re checking for typos, grammar slips, and minor inconsistencies in an already-quality text. Full post-editing starts from MT output, which has a fundamentally different error profile: potentially missing content, structural translation errors, terminology inconsistency, and register mismatches throughout. Full PE is a more extensive intervention because the starting material has more systematic problems.
How many errors are allowed in a publishable post-edited translation?¶
Industry practice (based on MQM-weighted scoring) typically allows zero critical errors, a maximum of 1-3 major errors per 1,000 words for general content, and minor errors at a level a native editor wouldn’t consistently flag. For legal, medical, or regulatory content, the threshold is stricter - often zero critical or major errors.
How long does full post-editing take compared to translating from scratch?¶
For good MT output from modern neural engines, experienced post-editors average 4,000-6,000 words per day at publishable quality, versus 1,500-2,500 words per day for translation from scratch. For poor MT output, the speed advantage collapses - if editing takes more than 60% of your translation-from-scratch time, it’s often not worth post-editing at all.
Does ISO 18587 post-editing qualify as ISO 17100 translation?¶
No. ISO 17100 explicitly excludes machine-translated content, even when fully post-edited. ISO 18587 is the separate standard that governs post-edited MT output. For legal, regulatory, or medical content where clients specify ISO 17100 compliance, post-editing - even full PE - is not a substitute for human translation from scratch.
What should be in a post-editing brief from the client?¶
A complete brief for full PE should specify: target register and tone, mandatory glossary, terminology to avoid, number and date formatting conventions, punctuation conventions for the target locale, how to handle untranslatable source-language terms, and quality thresholds (maximum weighted error score per 1,000 words or pass/fail criteria).
At what point is MT output not worth post-editing?¶
The pre-assessment required by ISO 18587 is supposed to catch this. The practical threshold: if full PE would take 60%+ of the time you’d need to translate from scratch, it’s usually not worth it - you’re spending more time on repair than original translation would take, at a lower per-word rate. This happens most often with rare language pairs, highly creative or legal content, and very short documents where setup overhead dominates.