Machine Translation Post-Editing for Researchers
Pubrica’s post-editors correct machine-translated and AI-generated text against the source document for researchers, institutions and regulatory teams. We work to the two levels defined in ISO 18587:2017: full post-editing, which produces output comparable to human translation, and light post-editing, which produces a comprehensible text. You receive the post-edited file with the level applied stated in writing and, at full level, a post-editing report and your AI and machine-use declaration.
Full post-editing is $0.06 a word and light is $0.035, from $180. Send us 500 words with the source document and we will tell you, free of charge, which level you need or whether fresh translation would be more suitable.
Meaning matches the sourceBoth levels
No content added or omittedBoth levels
Terminology, grammar, registerFull level only
Reads as natural EnglishFull level only; the most demanding part
ISO 18587 certification for light post-editingNot available, as Annex B is informative
Raw output → compare to source → correct → declare it
- Checked against the source document
- Level confirmed in writing before payment
- Engine and version recorded for your declaration
- Published prices for both levels
- Advice when fresh translation costs less
Service Scope
Six Components of Our Post-Editing Service
We offer two post-editing levels with separate prices, each suited to a different purpose, together with the checks and documents that support them.
Who We Serve
Post-Editing for Researchers, Institutions and Regulatory Teams
Compliance and Standards
Seven Standards Governing Post-Editing and Disclosure
Post-editing is governed by its own standard, separate from the translation standard. Pubrica works in compliance with the guidance below, and we state clearly where a level cannot be certified.
ISO 18587 Post-editing (2017, under revision)
The standard this service follows, which defines both post-editing levels. It has been under ballot as ISO/DIS 18587 since 4 September 2026 and is being retitled to cover “non-human translation output”.
ISO 18587 Annex B (informative, not normative)
Light post-editing is defined in this informative annex, so no provider can be certified for it. The accurate description is “to the light post-editing definition in ISO 18587”.
ISO 17100 Translation services (2015, excludes MTPE)
Listed here because of what it does not cover: clause 1 explicitly excludes machine translation plus post-editing, so this standard cannot be claimed for post-editing work.
ISO 5060 Evaluation of translation output (2024)
Provides an error typology with weights and severity levels. Our independent quality check follows this approach.
ICMJE Recommendations (January 2026)
The January 2026 revision adds a section on artificial intelligence in publishing. It does not mention translation, so we recommend disclosure.
Publisher AI disclosure policies (read July 2026)
Taylor & Francis lists translation among supported uses and asks for the tool’s name and version; Springer Nature places it in its permitted tier; Elsevier and Wiley require disclosure without naming translation.
ISO 9001 Quality management (2015, certified)
Pubrica is certified to this standard, cert. H3CR9TCU20251212INDQM10N. This is a quality management certification, not a certification against a post-editing standard.
Responsible Use of AI
Responsible Use of AI and Disclosure in Post-Editing
AI use matters particularly in editing, because improving a sentence is an editor’s task and writing it is an author’s. Journals do not prohibit AI assistance, but they do prohibit undisclosed assistance, and several now ask authors about it directly.
Why We Do Not Publish AI Accuracy Figures
Accuracy figures for AI tools are usually reported by the companies that sell them.
Pubrica therefore does not quote its own accuracy figures, as they would carry the same conflict of interest. Instead, we tell you which tool was used on your project, at which stage, and what a specialist changed afterwards.
How It Works
Our Post-Editing Process in Six Stages
Full post-editing takes four to seven business days and light post-editing three to five.
Sample-Based Provenance Triage Before Quotation
We read 300 to 500 words against the source and tell you what we think produced the text. We then recommend a level, and we tell you if translating afresh would cost less than post-editing the file.
Post-Editing Level Agreed in Writing
The level, full or light, is named on the quote rather than implied by the price. This prevents the most common dispute in post-editing, in which a client expects full post-editing but has ordered light.
Terminology Set Before Post-Editing Begins
We use your term base if you have one and build one if you do not. Agreeing the terms first makes the terminology work faster and more consistent.
Sentence-by-Sentence Post-Editing Against the Source
The post-editor works with the source beside the target throughout, which is what distinguishes post-editing from editing. At full level, grammar, register and terminology are also brought to publication standard; at light level they are not.
Independent Quality Check and Disclosure Statement
A second linguist checks a sample of the post-edited text against the source using an error typology, following the ISO 5060 approach. At the same time, we prepare your machine-use declaration.
Delivery with Term Base and Free Re-Edit Where Required
You receive the post-edited file, the term base, the declaration and a note of any points we could not resolve without you. If the work does not meet the level on your order, we re-edit it within the agreed scope at no added cost.
Request a Free Provenance Triage of Your Translation
We assess a sample of your document free of charge and tell you what we think produced it, which level it needs, what it would cost and whether fresh translation would be more suitable. The assessment takes two business days.
We ask for the engine name because engines and language models produce different errors. An engine tends to drop or invert content, while a language model tends to improve the text, for example by turning an awkward sentence into a clearer claim than the author made, or by adding a connective that asserts a causal relation the source only implied. In a research manuscript this second type of error is a data-integrity problem, and knowing which tool produced the draft tells the post-editor what to look for.
Our Experts
Your Post-Editing Team
A named team of subject specialists stays with your project from scoping to delivery. Your project coordinator introduces each team member by name before work begins.
Reviews the scientific content for accuracy and confirms that each conclusion is supported by the evidence. Each expert is matched to your field of research.
Edits for clarity, grammar, structure and consistency while preserving the author voice and scientific meaning.
Checks the finished work independently against the agreed scope and the applicable checklist. Findings are logged and resolved before delivery.
Manages the timeline, keeps you informed at each stage and coordinates the team so that questions are answered and deadlines are met.
Sample Work
Sample Post-Editing Report
This sample shows the report produced by a full post-edit of a clinical manuscript: each meaning error found against the source, the terminology work, a comparison with what light post-editing would have left, and the declaration supplied for the journal.
Post-editing report
Japanese to English · 4,800 words · clinical methods paper
Meaning · terminology · full vs light · declaration
31 meaning-level corrections4 changed what the paper reported
- Terms rendered inconsistently by the engine
- 23 instances across 9 terms, where the same source term appeared as two or three different English terms. This was the largest part of the work in the full post-edit.
- A plausible synonym for a fixed term
- An assay name had been replaced with a near-synonym from a different subfield. The sentence was grammatically correct but technically wrong, and a reviewer would notice the substitution.
- Term base built during the job
- 38 terms. The client had no term base, so we built one from their earlier English papers and the target journal’s conventions and handed it over.
- Institution and instrument names
- The engine had transliterated names that have official English forms. We replaced them with the registered names, verified on the institutions’ own websites.
- A statistical term used in the local sense
- We rendered it to the English convention after confirming the intended meaning with the author.
- Handover
- The full 38-term base, supplied as TMX and spreadsheet files. It belongs to the client and is not added to a shared resource.
Full post-editing deliveredLevel written on the order
- What the declaration says
- The engine and model, the version where knowable, which sections each touched, the level of post-editing applied, and that the authors retain full responsibility for accuracy and integrity.
- Why the version number is in there
- Taylor & Francis asks for the tool’s full name and version number, how it was used and why. A declaration without this information has to be rewritten at submission.
- Why we disclose where the policy is silent
- Elsevier and Wiley both require AI declarations but neither names translation, and the ICMJE’s January 2026 AI section does not mention it. We therefore recommend disclosure.
- What the declaration does not do
- It does not transfer responsibility. An AI tool cannot be an author or be accountable, so accountability remains with the authors.
- Recorded per section
- The authors had used two engines and one language model over several months, so we recorded the tools section by section. A single general statement would have been inaccurate.
- Handed to you as text
- A paragraph you paste into the submission form, rather than a certificate issued by us.
Redacted extract · field, findings and every figure changedSample report
The meaning tab shows why human post-editing is needed. Four of the thirty-one corrections changed what the paper reported: a hedged statement rendered as a finding, an omitted clause, a dosage converted at the wrong magnitude and a claim strengthened by a language model. Two rounds of internal review in English had not detected any of them.
The full-versus-light tab shows what the additional $120 paid for on this project and what it did not; accuracy was included at both levels.
Post-editing · sample report
Japanese to English, clinical methods paper, 4,800 words
An illustrative, anonymised report.
1Provenance Triage Before Quotation
| Section | What produced it | What that implied |
|---|---|---|
| Abstract, introduction, methods | A neural machine translation engine, one pass | Likely errors: omissions, negation and unit errors, and terminology drift. Sentences are usually fluent, so few errors look wrong on a first reading. |
| Results | The same engine, then edited in English by a co-author | The most difficult case. Monolingual editing removes the surface signs of machine origin but leaves the meaning errors underneath. The text reads better but is no more accurate. |
| Discussion, 900 words | A large language model | Likely errors are additions rather than omissions: smoothed claims, resolved ambiguities and inserted causal connectives. These were confirmed on reading. |
| References | Manual, by the authors | Out of scope for post-editing; checked for format only. |
The triage changed the quotation. The results section was priced 20% above the rest, because machine output that has already been edited monolingually takes longer to post-edit than raw output: the errors are the same, but the signs of them have gone. A per-word rate quoted without reviewing the file cannot account for this.
2Meaning-level corrections, against the source
| What the source said | What the machine produced | As post-edited, and why |
|---|---|---|
| “A tendency toward improvement was observed” | “Improvement was observed” | Hedge restored. The output stated a finding that the authors had reported as a tendency, in the results section. This cannot be detected by reading the English alone. |
| A methods sentence with an exclusion-window clause | The same sentence without it | Clause restored. The output was complete and grammatical but described a different protocol. Omission without a visible gap is a characteristic error of neural machine translation. |
| A dosage in source units | The same figure, converted at the wrong magnitude | Corrected and queried. The value was ten times the source and had passed two internal reviews in English. |
| An awkward hedged sentence in the discussion | A clean sentence asserting a causal relation | A characteristic language model error. The output was not a mistranslation but an unrequested improvement. We restored the author’s weaker claim and flagged it for their decision. |
| Two negated conditionals | Positive conditionals | Restored. Negation across clause boundaries is a known weakness of machine translation, and both sentences read naturally in English. |
| 25 further instances | — | Mistranslated conjunctions, tense shifts in the methods, and missing subjects restored where the source omitted them, which is common in Japanese but not possible in English. |
31 meaning-level corrections. Four changed what the paper reported. This ratio is typical. Most post-editing is repair work, and its main value lies in the few places where the machine output said something the authors did not.
3Terminology reconciliation
| Item | Handling |
|---|---|
| Inconsistent renderings | 23 instances across 9 terms, typically rendered one way in the methods and another in the discussion. Each form was correct in isolation, and the inconsistency was visible only in a term-by-term review. |
| Plausible-synonym substitution | One assay name had been replaced by a near-synonym from an adjacent subfield. Restored. |
| Term base | No term base existed. We built 38 terms from the authors’ earlier English papers and the journal’s conventions, applied them throughout and handed them over as TMX and spreadsheet files. |
| Institution and instrument names | Transliterated by the engine. We replaced them with registered English forms verified on the institutions’ websites. |
| One statistical term | Rendered to the English convention after a query to the authors rather than on our judgement, because the two readings imply different analyses. |
| Ownership | The term base is the client’s and is not added to a shared resource used for other clients. |
4What full post-editing bought over light
| Work | Light | Full |
|---|---|---|
| All 31 meaning corrections | Included | Included |
| No content added or omitted | Included | Included |
| Plainly wrong output removed | Included | Included |
| 23 terminology inconsistencies reconciled | Not included | Included |
| 47 awkward but comprehensible sentences rewritten | Not included | Included |
| Register brought to the journal’s conventions | Not included | Included |
| Formatting to the target journal | Not included | Included |
| Price, 4,800 words | $170 | $290 |
Accuracy is included at both levels. Light post-editing is not a lower standard of correctness; it stops once the text is correct and comprehensible. The $120 difference paid for consistency, register and format. That work is unnecessary for a document you only need to understand but essential for a manuscript going to peer review.
5The declaration, as supplied
| Element | What we wrote, and why |
|---|---|
| Tools named | Each engine and model by full name, with version numbers where the tool exposes one and a note where it does not. |
| Scope per section | Recorded section by section, because three tools had been used across the manuscript. A single general statement would have been inaccurate. |
| Level of human post-editing | Full post-editing against the source by a qualified post-editor, with independent quality checking. This is stated as a description of the work done, not as a certification. |
| Accountability | A sentence confirming that the authors retain full responsibility. Springer Nature states that human accountability cannot be transferred, and the ICMJE position is equivalent. |
| Reason for disclosure | The target journal’s publisher does not mention translation in its AI policy. An unnecessary disclosure has no cost, whereas a missing one can put acceptance at risk. |
6Deliverables and Exclusions
| Item | Note |
|---|---|
| The post-edited file, clean and tracked | Every change can be reviewed and rejected. |
| This report | Including the four corrections that changed what the paper reported, which the authors checked against their data. |
| The query log | Nine queries, with the authors’ answers recorded alongside. |
| The 38-term base | TMX and spreadsheet files, reusable on future papers. |
| The declaration text | A paragraph to paste into the submission form. |
| Not included: restructuring the discussion | The discussion had a structural weakness. We described it in a note and left it unchanged, because restructuring is substantive editing and is priced separately. |
| Not included: reconciling two disagreeing figures | A figure in the text disagreed with one in a supplementary table. We queried it rather than changing it. |
| Not included: claiming ISO 17100 conformity | ISO 17100 excludes machine translation plus post-editing from its scope. This project was completed to ISO 18587, and the documentation states this. |
This report does not claim that the post-edited text is perfect or that the paper will be accepted. ISO 18587 defines full post-editing as output “comparable to a product obtained by human translation”, which is a test of comparability rather than perfection. Our commitment is specific: if the work does not meet the level written on the order, we re-edit it within the agreed scope at no added cost.
Benefits
Benefits of Human Post-Editing Against the Source
Machine output is usually fluent, which makes its errors hard to detect. Human post-editing against the source finds the errors that a monolingual editor cannot see.
Scope and Pricing Tool
Choosing Between Full and Light Post-Editing
Enter what produced the draft, where the text is going, the quality of the raw output and its length. The tool recommends a level and estimates the price, and in some cases it advises that translating afresh would cost less than post-editing.
01
Recommends full or light post-editing, explains what each level covers, estimates the price and indicates when fresh translation is the better option.
Runs in your browser · nothing sent anywhereChoosing a level. Accuracy is included at both levels: the meaning is correct, nothing is added or omitted and plainly wrong output is corrected. Full post-editing also brings terminology, grammar, register and formatting to publication standard. Light post-editing stops once the text is correct and comprehensible. If a supplier quotes an acronym such as MTPE or PEMT rather than a level, ask which level it means, because neither term appears in the standard and full post-editing costs roughly twice as much as light.
Sources, standards and versions
Packages
Post-Editing Packages and Pricing
We price both ISO 18587 levels. Where a project is better handled as translation, we price it as translation.
Light Post-Editing for Comprehension
Documents you need to understand rather than publish
From $120$0.035/word · Annex B definition
Delivery 3–5 business days
- ✓Post-edited against the source document
- ✓Meaning checked: nothing added or omitted
- ✓Plainly wrong or offensive output removed
- ✓Numbers, units and names verified
- ✓A note of anything we could not resolve without you
Style is not edited at this level. Light post-editing cannot be certified, because Annex B of ISO 18587 is informative.
View a sample Select LightFull Post-Editing for Publication
Documents for journals, regulators, clients or the public
From $180$0.06/word · normative definition
Delivery 4–7 business days
- ✓Everything in Light
- ✓Terminology reconciled across the whole document
- ✓Grammar, syntax, punctuation and register to publication standard
- ✓A post-editing report listing each meaning-level change and its reason
- ✓Independent quality check by a second linguist
- ✓Your AI and machine-use declaration, written for you
- ✓iThenticate similarity report on request, $80
Free re-editing, within the agreed scope, if the work does not meet the level written on your order.
View a sample Select FullFull Post-Editing with Journal Submission Pack
Manuscripts going straight to a named journal
From $380$0.06/word + $200 journal pack
Delivery 7–10 business days
- ✓Everything in Full
- ✓Subject-matter review of the English by an editor in your field
- ✓Formatted to the target journal: word limits, headings, reference style
- ✓Cover letter and submission checklist
- ✓Free re-editing after reviewer comments, within 60 days and within the agreed scope
- ✓Pre-submission consultation with the post-editor before you file
The journal pack is the same one on our manuscript editing page, at the same price.
View a sample Select Full plus journalVolume Post-Editing for Books, Dossiers and Document Sets
Long or repeated work in which consistency across documents matters most
$0.055 a wordabove 25,000 words · term base included
Delivery From 15 business days
- ✓Everything in Full, across the whole set
- ✓One term base governing every document and every post-editor
- ✓A named project editor start to finish
- ✓Consistency pass over the finished whole
- ✓Free 1,000-word pilot before the project is agreed
Where CAT match-rate data can be exported, we price on it rather than treating everything as raw output.
View a sample Select VolumeOptional additions for any package: free provenance triage before quotation; monolingual editing where no source exists, described as editing rather than post-editing; subject-matter review by an editor in your field; journal formatting and the submission pack; and translation from the source where the triage shows that post-editing would cost more. Re-editing after reviewer comments is free within 60 days, within the agreed scope.
Quality Assurance
Our Service Commitments and Limitations
Our revision commitment
Free Re-Edit If the Agreed Post-Editing Level Is Not Met
The level, full or light, is written on the order in the standard’s own terms before we start, so the commitment can be checked objectively.
01
Post-editing always against the source
If we cannot read the source, we do not describe the work as post-editing. Monolingual work on machine output is priced and described as editing.
02
The level is stated on the order
Full or light is recorded on the order form, which prevents the most common dispute in this service.
03
Claims are kept at their original strength
Where the source hedges, the post-edited text hedges. Where a language model has strengthened a claim, we restore the original wording and tell you.
04
Advice when fresh translation is better
Where the raw output is poor and the text is for publication, post-editing costs more than translating afresh. We tell you this before quoting.
What we do not promise. Perfect accuracy: ISO 18587 defines full post-editing as output “comparable to” human translation, which is a test of comparability rather than perfection, and ISO 5060:2024 evaluates this work using an error typology with weights and severity levels. We also do not promise journal acceptance.
We claim no ISO 17100 conformity, because that standard excludes machine translation plus post-editing from its scope, and no certification for light post-editing, because light sits in informative Annex B with nothing there to certify against. We do not post-edit without a source, do not use a public machine translation engine, do not treat an earlier human translation as machine output in order to charge for post-editing when the job is revision, and do not pool your term base for other clients.
We recommend the lower-cost level when it is appropriate
If the document is for internal use, light post-editing at $0.035 a word is the appropriate purchase and full post-editing is unnecessary. Accuracy is included at both levels. The additional cost of full post-editing covers terminology consistency, register and formatting, which are essential for a submission but not needed for a document you are reading for information.
Your document is not processed in a public engine
A non-disclosure agreement is signed before any file moves, and post-editors, quality checkers and project staff are bound by internal confidentiality agreements. Your source, output and term base are held under the UK GDPR on access-controlled systems within a quality management system certified to ISO 9001:2015, cert. H3CR9TCU20251212INDQM10N — kept only as long as the project needs them and never used to train anything. Where machine translation is part of a job at all, it happens in a closed environment agreed with you in advance.
Client Stories
Client Feedback on Post-Editing Projects
Testimonials are rephrased from client feedback.
“A dosage was out by a factor of ten and two of us had read the English twice. It only shows when someone opens the Japanese beside it.”
“You told us the machine version was too poor to rescue and quoted translation instead. Nobody else had said that.”
“The journal asked for the tool and the version number at submission. It was already written for us.”
FAQs
Frequently Asked Questions About Post-Editing
If your question is not answered here, ask an expert. No quote is required.
Post-editing is the human correction of machine-translated or AI-generated text, carried out against the source document to a level agreed in advance. The international standard is ISO 18587:2017, which sets requirements for the process and for the competence of post-editors.
Working against the source is what distinguishes post-editing from editing. The most serious errors in machine output, such as a dropped negation, an omitted clause, a hedge rendered as a finding or a unit converted at the wrong magnitude, all produce fluent, grammatical English and cannot be detected by reading the target text alone.
The difference is a qualified human review of the output against the source.
Raw machine output is fluent, immediate and unverified. No reviewer checks it against the source before delivery, which is why it is inexpensive and why it is not suitable for publication.
Post-edited translation is the same output after a post-editor has compared every sentence with the source. The difference between $0.01 and $0.06 a word reflects this human review, which on the sample published here produced thirty-one meaning-level corrections, four of which changed what the paper reported.
Machine output is usually fluent, and that fluency makes its errors difficult to detect.
Neural engines produce confident, grammatical target text that can invert a negation, drop a hedge, omit a clause without a visible gap or convert a unit incorrectly. None of these errors looks like an error in English.
Language models add a different type of error: they can strengthen the source, turning an awkward sentence into a clearer claim than the author made or inserting a connective that asserts causation. In a research manuscript this is a data-integrity problem rather than a language error. It is also why ISO/DIS 18587, under ballot since 4 September 2026, is being retitled for “non-human translation output”.
ISO 18587 defines full post-editing as producing output “comparable to a product obtained by human translation” and light post-editing as producing “a merely comprehensible text”.
Accuracy is included at both levels. Light post-editing is not a lower standard of correctness: the meaning is correct and nothing is added or omitted, but the work stops once the text is correct. Full post-editing adds terminology consistency, grammar and register to publication standard, and formatting.
On the sample published here, the difference was $120: light post-editing would have delivered all thirty-one meaning corrections but left twenty-three terminology inconsistencies. Choose full post-editing for text you will publish and light post-editing for text you need to understand. Light post-editing is described in informative Annex B, so no provider can be certified for it.
Post-editing is the wrong choice when the raw output is poor and the text must reach publication standard. In that case post-editing costs more than translating from the source and produces a weaker result.
There are two reasons. A post-editor working on poor output must diagnose what the machine did to the meaning sentence by sentence, which involves all the work of translation plus the time spent correcting wrong renderings. Heavily post-edited poor output also keeps its original sentence structure, so it continues to read as a translation.
Output that a colleague has already edited monolingually is the most difficult case, because the editing removes the surface signs of machine origin without correcting the meaning errors underneath. We price this work above raw output.
There is no uniform rule across publishers, so we recommend disclosure.
Taylor & Francis lists “copyediting, language improvement and translations” among supported uses and requires the tool’s name and version. Springer Nature places translation in its permitted tier, disclosure encouraged. Elsevier (July 2026) permits AI only for language and readability and requires a declaration, without naming translation. Wiley (July 2026) requires disclosure, exempts spelling and grammar tools, and is also silent on translation. The ICMJE’s January 2026 AI section does not mention translation either.
Two of the four publishers address translation explicitly and two do not mention it, so we recommend disclosure in every case and write the declaration for you. A declaration does not transfer responsibility: an AI tool cannot be an author, so accountability for accuracy remains with you.
Related Services
Related Translation and Editing Services
When the triage shows that one of these services is more suitable for your document, we will recommend it.
1ALTERNATIVESwhen the triage recommends them
Translation with editingISO 17100, two linguists, $0.12 a word. Less costly than post-editing poor output. Scientific editingFor manuscripts in which the argument, rather than the language, needs work.2AFTER POST-EDITINGgetting it submission-ready
Manuscript editingJournal style, structure and the submission pack. ProofreadingThe final check before submission, at $0.025 a word.Insights
Further Reading on Machine Translation Post-Editing
Request a Free Post-Editing Triage on a 500-Word Sample
We assess a sample of your document free of charge within two business days and tell you what produced the draft, which level it needs, what it would cost and whether fresh translation would be more suitable.