Machine Translation 5 September 2026
What Is Post Editing? Post-Editing Machine Translation Output
Correcting machine output is a different job from translating from scratch. Unless the level is defined at the outset, both quality and cost become impossible to predict.
When machine translation entered the industry, a new kind of work was born with it: correcting the text the machine produces. The industry calls this work post-editing; in Turkish it goes by son düzeltme or makine çevirisi sonrası düzeltme.
Contents
The short answer to the question what is post editing is this: it is the correction of machine output by a human, checked against the source text. The long answer is more interesting, because this work is not merely faster than translating from scratch, it is different in kind.
This article looks at the levels of post-editing, how the process should be set up, when it pays off and when it costs more than it saves, and what measurement pricing ought to rest on.
What is post editing, and how does it differ from translation?
When translating from scratch, the translator reads the source text, builds the meaning and writes it again in the target language. The mental flow runs towards production. In post-editing the flow is reversed: a finished text is already in front of you, and your task is to verify it.
That difference explains why the work is tiring. Verification demands closer attention than production, because you do not know where the error is and you have to read every sentence with suspicion. Fluent output lowers that attention, and so raises the risk.
The second difference is that a ready-made text steers your thinking. With a sentence already in front of you, producing a solution wholly independent of it becomes harder. The “blind faith” problem discussed in the literature in connection with translation memory applies here too: a ready-made suggestion invites acceptance without question (Bowker 2005).
The third difference concerns wholeness. Someone correcting sentence by sentence may never see the text as a single whole. This resembles the problem the translation tools literature calls “collage translation” (Mossop 2006), and it threatens consistency.
Two levels: light and full post-editing
The industry distinguishes two levels, and the distinction needs to appear in the contract. Otherwise the two parties will understand very different things by the phrase “post-edited text”.
| Criterion | Light post-editing | Full post-editing |
|---|---|---|
| Purpose | A correct and comprehensible text | A text indistinguishable from human translation |
| Errors of meaning | Corrected | Corrected |
| Style and fluency | Left untouched | Rewritten |
| Term consistency | For critical terms | Throughout the text |
| Texts it suits | Internal correspondence, reading for information | Any text to be published |
In light post-editing the rule is clear: if a sentence is correct and comprehensible, it stays as it is, even where it could be written more elegantly. The rule is hard on the person doing the work, because the professional reflex is to improve the text.
In full post-editing the criterion is the result: the reader must not be able to tell that the text has been through a machine. The effort spent at this level often approaches that of translating from scratch, and the expected saving shrinks accordingly.
How is the process set up?
A well-built process starts before the post-editing stage. The first step is preparing the source text: broken formatting is cleaned up, scanned documents are converted into text, and very long sentences are split where necessary.
The second step is preparing terminology and memory. The organisation’s approved term list, and its translation memory where one exists, are brought into the process. These two resources remove the repetitive decisions that consume the most time during post-editing.
The third step is the post-editing itself, and it must always be done against the source. A check made by reading the output alone cannot catch omissions and negation errors; both look entirely natural in the output.
The fourth step is record-keeping. If the errors found are noted by type, after a few projects it becomes clear which type keeps recurring. That record guides both the choice of engine and the checklist, and it makes the process cheaper over time.
When does it pay off?
The cases in which post-editing pays off are well known, and they all have to do with the nature of the text. The more repetitive, formulaic and narrow in domain a text is, the faster the post-editing goes.
The cases in which it loses money are just as well known. Where the engine is weak in the domain, where the source is written in complex sentences, or where style determines the outcome, post-editing can take longer than translating from scratch. This is what experienced translators mean when they say “it would be quicker to translate this from the beginning”.
The decision should therefore rest on measurement rather than guesswork. Both methods are tried on a small sample and the actual times are compared. A few hours of this work sets the direction of a process that runs for months.
There is also a starting point for that measurement. Industry research has put average translator productivity at 2,684 words a day, with the industry benchmark taken as 2,500 words a day (DePalma 2012). Setting your own figures against that frame shows you whether your results are plausible.
Pricing and the measurement problem
Pricing post-editing is one of the most contested subjects in the industry. The common practice is to apply a discount to the translation unit price. The problem is that the discount rate usually rests on habit rather than on the actual state of the text.
A sounder method is pricing based on measurement. Post-editing time is measured on a small sample, the real effort per unit is calculated, and the price is set accordingly. This method protects buyer and supplier alike.
The unit of measurement also needs to be settled. The common practice in Türkiye is counting characters with spaces; a page of 1,800 characters with spaces corresponds to roughly 240 words (Pietrzak & Kornacki 2021, p. 83). Whether words or characters are being counted must be stated plainly in the quotation.
Finally, the level of post-editing is part of the price. Expecting full post-editing at the price of light post-editing is the most common cause of quality disputes. A quotation that does not state which level will be applied is incomplete.
Who should do this work?
Whether post-editing is a separate skill was debated for a long time. The view now settled in the industry is that it is a skill added on top of translation competence. Being a good translator is a necessary condition, but not a sufficient one.
The added skills are these: recognising the typical error patterns of machine output, resisting unnecessary improvement, and keeping up the pace of decision-making. The third is often the hardest, because the urge to beautify a text is a professional habit.
There is a question of attitude as well. A translator who keeps tools at arm’s length may struggle more when working with machine output. The literature examines this under the heading of the cognitive friction created by tools (Ehrensberger-Dow & O’Brien 2015). If the process is poorly designed, technology can slow the work down instead of speeding it up.
The practical upshot for the buyer is this: accept that this work cannot be handed to just anyone. Asking your supplier who will do the post-editing, and what experience that person has in the field, is as legitimate a question as asking the price.
Standards and contract language
There is an international standard for post-editing: ISO 18587 sets out the requirements for the human post-editing of machine translation output. The standard covers the levels of post-editing, the competences of the person who will carry it out, and the documentation of the process.
Citing the standard lightens the definitional load in a contract. The phrase “full post-editing under ISO 18587” makes sure both parties understand the same thing and offers a framework to fall back on in a dispute.
Another heading that belongs in the contract is confidentiality. If machine support is to be used, the contract should state which engine is used, where the data is sent and whether it is retained. This information is rarely shared unless it is asked for.
The third heading is transparency. Declaring whether the delivered text was produced with machine support is an expectation that is steadily taking hold. The buyer needs that information in order to decide how much internal checking to do.
The Türkiye context
In the Türkiye market, post-editing is often not defined as a separate service line. Whether a job was done with machine support or from scratch goes unstated unless somebody asks, which makes price comparisons misleading. The gap between two quotations may be a difference of method rather than a difference of quality.
On the language side, Turkish has a structure that increases the post-editing burden. Because the language is agglutinative, the meaning of a sentence can turn on a single suffix: when a negation suffix drops out, the output stays fluent but the meaning is reversed. Subject dropping, for its part, leads to errors of person. These two error types should be looked for deliberately during post-editing.
There is a layer of liability as well. In documents that require sworn translation and notarisation, the accuracy of the output is the responsibility of the translator who signs it; the fact that a machine was used in the process does not alter that responsibility. Furthermore, if a document contains personal data, sending it to a cloud-based engine has to be assessed separately under the KVKK. This article is for general information and is not legal advice.
What does this mean for you?
Set up properly, post-editing is a genuine saving; set up badly, it is a hidden quality risk. What makes the difference is that the decision has been defined from the start.
- When you request a quotation, ask plainly whether machine support will be used.
- If it will be, have the level of post-editing put in writing: light or full?
- Settle the unit of measurement: words, or characters with spaces?
- On the first project, ask for time and error measurements on a small sample, and tie the price to that measurement.
- For confidential documents, obtain written confirmation that the engine can be switched off.
In our quotations we state plainly whether machine support is in play, and the level of post-editing is written out project by project. We can work through together which of your document types this approach makes sense for: upload your document, or look at our pricing page to see how the pricing logic works.
Frequently asked questions
What is post-editing?
It is the correction of machine translation output by a human, checked against the source text. What sets it apart from translating from scratch is that the starting text already exists; that does not always mean less effort, however.
What is the difference between light and full post-editing?
In light post-editing the aim is for the text to be correct and comprehensible, and flaws of style are left alone. In full post-editing the result must be indistinguishable from a human translation. The two take different amounts of time and carry different prices.
Is post-editing always cheaper?
No. If the machine output is weak, post-editing can take longer than translating from scratch. The gain depends on how repetitive the text is, how narrow the domain is and how well the engine performs in that domain.
How is post-editing priced?
The common practice is a discounted per-word rate or an hourly rate. The sound approach is to measure the actual post-editing time on a small sample first and to base the price on that measurement.
References
- ISO 18587:2017. Translation services – Post-editing of machine translation output – Requirements. International Organization for Standardization.
- Pietrzak, P. & Kornacki, M. (2021). Using CAT Tools in Freelance Translation. Routledge Focus, p. 83; Chapter 3.
- DePalma, D. A. (2012). Translation Productivity Benchmarks. Common Sense Advisory.
- Bowker, L. (2005). Productivity vs Quality? A Pilot Study on the Impact of Translation Memory Systems. Localisation Focus, 4(1), pp. 13–20.
- Ehrensberger-Dow, M. & O’Brien, S. (2015). Ergonomics of the Translation Workplace. Translation Spaces, 4(1), pp. 98–118.
- Mossop, B. (2006). Has Computerization Changed Translation? Meta, 51(4), pp. 787–793.