Gains:
- Ability to distinguish between light and full post-editing levels and choose the right level based on the purpose and visibility of the text
- Ability to implement an efficient post-editing flow that corrects only what is needed, avoiding both overcorrection and undercorrection
- Ability to run a separate check round for number, negation, and term and measure post-editing efficiency with time and edit distance
The most common way professional translation works today is not translation from scratch, but MTPE (Machine Translation Post-Editing): the machine produces a raw draft, the human translator corrects it and makes it ready for publication. In this unit, you'll learn the two formal levels of post-editing (light and full), how to apply which to which text, an efficient post-editing flow, and the "over-editing" trap. The goal is to become a balanced post-editor who ensures quality while preserving the time the machine gives you.
Two levels: light and full post-editing
Post-regulation in the industry is defined at two standard levels (ISO 18587 is the international standard for this business):
Light post-editing: The goal is to make the text clear and accurate — not perfect. You correct errors that distort meaning, incorrectly translated information, serious grammatical errors; But you don't make any stylistic touches because "it would be more beautiful". Suitable for: internal use, quick information, low-visibility text, gist translations.
Full post-editing: The aim is for the output to be indistinguishable from human translation. Meaning, terminology, tone, flow, tone — all must be of broadcast-ready quality. Suitable for: all texts with brand visibility that will be published, delivered to the client.
Mixing the two levels creates problems: if you apply light post-editing to a text to be published, the quality will be poor; If you do full post-editing on an internal note, you'll lose time and money. Clarify with the client what level is expected before work begins.
Tip: When the client says "machine translation, cheap", most of the time they want "light" but are talking about a text that will be "published" — discuss this contradiction up front. If the expectation is not clear, when the job is done you will ask "why isn't this perfect?" or “why did this take so long?” An argument ensues.
Over-editing trap
MTPE's most insidious enemy of efficiency is overcorrection: rewriting a sentence that the machine gave you that is correct but not your choice, just because "that's how I would write it." This destroys the entire time advantage of MTPE. The rule is clear: If the machine output is accurate, understandable and appropriate, don't touch it — only correct it if it's wrong. Changes of taste are limited in full post-editing and almost prohibited in light post-editing.
The opposite is also dangerous: under-editing — passing the machine's fluent but incorrect sentence as "looks good." Balance is the discipline of "correct what is necessary, leave what is unnecessary."
Caution: Overcorrection often occurs in novice post-editors, while undercorrection occurs due to fatigue and time pressure. To stay away from either extreme, "does this change improve meaning/accuracy/tone, or is it just my taste?" ask.
Efficient post-editing flow
A practical step-by-step flow:
- Scan first, fix later. Read through the text once; see the whole, the tone, and the repeating errors.
- Flag high risk items. Numbers, dates, negations, proper names, units — give these a separate tour.
- Correct semantic errors. Anything that breaks fidelity to the source comes first.
- Align term and consistency. Match Termbase and previous translations.
- Fix flow and tone (in full post-editing). Naturalize sentences that smell like translation.
- Final reading. Without looking at the source, just read the target text like a reader; If you get stuck, fix it.
You can use the LLM as a “second eye” in this flow: by feeding the machine output to the LLM and telling it to “mark only the risky areas,” you reduce your blind spots — but the final decision is yours.
three mini cases
Case 1 — The right level saved time. A company had its 40,000-word internal knowledge base articles translated so that employees could understand them. The translator applied light post-editing: corrected 60 errors that distorted the meaning, left no stylistic touches. The job was finished 55% faster than full post-editing and the objective was met.
Case 2 — Overcorrection billing. A novice post-editor rewrote the correct sentences to make them "flow more smoothly." The measurement showed that 70% of the edited sentences were already acceptable; The additional time spent extended the job 2 days ahead of schedule and erased the profit. The team wrote down the "only fix it if it's wrong" rule.
Case 3 — Insufficient correction caught. In a medical device manual, the machine translated the phrase "do not immerse" fluently but incompletely, saying only "clean" instead of "do not immerse." The high-risk item round caught this security bug; This showed how inadequate correction can turn into danger.
Four copyable templates
1) Post-edit level determination:
Your role: MTPE project consultant. I will describe the purpose and visibility of the text. Tell me whether light or full post-editing is appropriate for this job, the rationale, and what to fix or leave at each level.Text purpose/visibility: [explain]
2) Second eye inspection (to LLM):
Below is the source and machine translation. Flag HIGH RISK errors that a post-editor might miss: ambiguity, omitted/added information, wrong number/date/name, negation error, inconsistent term. Making stylistic suggestions.Source: [...] | Translation: [...]
3) Light post-editing instruction:
SLIGHTLY post-edit the machine translation below: correct only any confusing, inaccurate or unclear parts. Don't change your style or taste. Leave every sentence that is correct and understandable as it is. Briefly list what you changed and why. Translation: [...] | Source: [...]
4) Overcorrection control (to your own correction):
Below is the machine's sentence and my correction. For each correction, say: does this change (a) correct meaning/correctness, (b) require terminology/consistency, or (c) is just stylistic preference? (c) mark those as "may be unnecessary".Machine: [...] | My fix: [...]
Weak prompt / Strong prompt
Weak: "Fix this translation." (Level unclear; LLM rewrites everything, overcorrects.)
Strong: "Full post-edit this machine translation: make the meaning, terminology and fluency ready for publication, but stay true to the source, do not add. Do not change correct sentences unnecessarily. Label the reason for each change (meaning/term/flow) with a word."
Difference: strong prompt gives level, limit and justification requirement; prevents both overcorrection and undercorrection.
Level comparison chart
Size
Light post-editing
Full post-editing
target
Accurate and understandable
Human translation quality
semantic errors
is corrected
is corrected
Term consistency
basic level
full
tone/style
untouchable
Ready to publish
Stylistic improvement
not done
Limited, if necessary
Appropriate text
internal, temporary, gist
published, brand
Time/cost
low
high
Common mistakes
- Not clarifying the level with the customer. The expectation of "it should be cheap" and "it should be perfect" conflict.
- Overcorrection. Rewriting the correct sentence for pleasure and destroying the time advantage.
- Insufficient correction. Passing a fluent but wrong sentence; The most dangerous are safety/count errors.
- Not making separate rounds for high risk items. Number and negation errors are overlooked in the flow.
- Skipping the last reading. Reading the target text without looking at the source reveals any remaining roughness.
Measuring efficiency: edit distance and duration
Use two simple indicators to measure how “worth” the post-edit is. The first is time: the difference between translating the same volume from scratch and post-editing; Does MTPE actually speed up, or are you bogged down by overcorrection? The second is edit distance: it is the ratio of how much of the machine output you change. Too low edit distance (almost no touching) is a sign of poor editing; If the edit distance is too high (almost rewriting), either the text is not suitable for the machine or you are over-correcting. If you follow these two indicators in several jobs, you will concretely see in which type of text MTPE really brings profit, in which it is wiser to translate from scratch. You can't improve a process you don't measure; Look at these two numbers instead of intuition.
In summary
MTPE is the main way translation works today and has two levels: light post-editing, which makes the text accurate and understandable, and full post-editing, which brings it to human quality. The master post-editor clarifies the level before starting work; corrects only what is necessary and avoids both overcorrection and undercorrection; does a separate round of checking for number, negation, and term; He uses the LLM as a second set of eyes but makes the final decision himself.
Application task
Get a machine translation of a text of 150-200 words. Apply light post-editing first and measure the time; then apply full post-editing to the same text (on clean copy) and compare time and number of changes. Finally, use the “overcorrection control” template to determine how many of the changes you make are actually necessary and how many are stylistic.
checklist
- [ ] I clarified the light/full level with the client.
- [ ] I only corrected it if it was wrong; I did not change the correct sentences unnecessarily.
- [ ] I did a separate round of auditing of the number, date, name and negatives.
- [ ] I confirmed the security/critical sentences with the source against the risk of insufficient correction.
- [ ] I proofread without looking at the source.