Is post-editing only correcting grammar?

No. It includes checking whether the translation preserves the source's meaning and meets the agreed requirements.

Does automatic post-editing replace human review by definition?

No. It is an automated revision task; whether further review is needed depends on the workflow and quality requirement.

Do fewer edits always mean less effort?

No. Reading, understanding, research, and difficult decisions may take time even when the final text changes very little.

AI Glossary / Natural language and speech / Beginner

Post-editing

Definition

Post-editing is the revision of machine-generated output after it has been produced. In machine translation, its established use means reviewing and correcting a machine-translated draft so it meets a specified purpose and quality requirement. A human post-editor checks the target text against the source, rather than assuming that fluent wording accurately preserves the original meaning.

What it means

The starting point is a machine translation, but the accepted result depends on the editing brief. Some workflows require publication-ready language, while others define a more limited internal-use requirement. The required outcome should be explicit before work begins; fewer edits do not automatically mean that the result is better or sufficiently accurate.

Post-editing differs from translating from scratch and from pre-editing, which changes the source before translation. Automatic post-editing is a related research task in which another system revises machine output. That automated stage should not be silently described as completed human review.

How it works

A translator or reviewer compares the draft with the source and identifies mistranslations, omissions, additions, terminology problems, and unsuitable wording. They also inspect document-level consistency, formatting, names, numbers, and relationships that cannot be judged reliably from isolated target sentences.

Edits should serve the agreed brief. A fluent sentence can need a small but crucial correction, while an awkward sentence may require extensive rewriting. Sometimes the most efficient response is to replace a poor draft rather than preserve its structure. The workflow should permit that judgment rather than reward minimal changes at the expense of correctness.

Evaluation can measure final quality, time, and editing activity separately. Edit distance describes differences between versions but does not fully capture reading, research, or reasoning effort. A fair productivity comparison must also consider the material, language pair, quality target, and reviewer conditions, rather than treating every machine-generated draft as equally useful.

Three practical examples

1. Product documentation: A translator revises machine-translated instructions using the source and an approved terminology resource before the text reaches its intended audience.

2. Internal correspondence: A team defines the level of review required for translated messages and checks that meaning-critical details remain accurate.

3. Localization evaluation: A project compares post-editing with another translation workflow on representative material, measuring accepted quality as well as time spent.

Think of it this way

Think of post-editing as inspecting and repairing a first draft while keeping the original specification beside it. A smooth-looking draft can still fail that specification. The analogy is limited because translation is not mechanical component replacement: several different phrasings can convey the source appropriately, and deciding among them requires linguistic and contextual judgment.

Worked example

Consider an original miniature example with a Spanish source: “La entrega no incluye instalación.” A faulty machine draft says, “The delivery includes installation.” An appropriate corrected target is “The delivery does not include installation.” The missing negation changes the commitment conveyed by the sentence, even though the draft is fluent.

Now consider a separate hypothetical productivity comparison on matched 600-source-word assignments meeting the same acceptance requirement. If post-editing takes 30 minutes and translation from scratch takes 40 minutes, the rates are 1,200 and 900 source words per hour respectively.

The time reduction is (40−30)/40=25%, while the throughput increase is (1,200−900)/900≈33.3%. These are different percentages with different denominators. The figures are invented and independently checked, not evidence that post-editing generally produces those gains. The corrected sentence and arithmetic illustrate why accuracy, elapsed time, and reporting conventions all matter.

Illustrative case study

Fictional case: Alder Guides prepares multilingual setup notes. Reviewers notice that some machine drafts read naturally but omit a limitation present in the source. The team adds explicit source-comparison checks for conditions and negation instead of relying on target-language fluency alone.

It records final-quality findings separately from editing time and compares results across representative document types. The case illustrates a quality-control workflow; it does not claim that any real company achieved the hypothetical productivity figures or that a particular machine system caused the example error.

Common misconceptions and mistakes

Do not equate post-editing with proofreading the target text in isolation. A grammatically polished mistranslation still needs correction. Nor should a low edit count be treated as proof that the reviewer worked efficiently: it may reflect a strong draft, inadequate review, or a different brief.

Avoid claiming that post-editing is always faster than translation from scratch. Its usefulness depends on the draft and working conditions. Also distinguish human post-editing, automatic revision, and subsequent independent quality review, since they provide different forms of evidence about the accepted output.

Business use

For businesses, post-editing can form part of a translation workflow when machine drafts are useful starting points and qualified reviewers can assess them against the source. The decision should be based on relevant material and the required outcome.

Define terminology, acceptance criteria, responsibility for review, and what time measurements include. Compare workflows at equivalent quality rather than purchasing apparent speed through weaker checking. Preserve source, machine draft, and accepted version when appropriate so recurring problems can be analyzed and improvements evaluated systematically.

Check your understanding

Why are the example's 25% time saving and 33.3% throughput increase both correct?

Answer: The first compares minutes saved with the original 40 minutes. The second compares the higher hourly output with the original 900 words per hour. Their denominators differ.

Related terms

Machine translation
Edit distance
Localization
human review
pre-editing
automatic post-editing
translation quality
terminology management

Learn next

Compare a source, machine draft, and corrected translation. Classify meaning and style changes separately, then design a small evaluation that records accepted quality, time, and the exact productivity denominator.

Sources

Sources: Sanchez-Torron and Koehn, Machine Translation Quality and Post-Editor Productivity, study setup and measurement definitions: https://aclanthology.org/2016.amta-researchers.2.pdf . Gaspari and colleagues, Perception vs. reality: measuring machine translation post-editing productivity: https://aclanthology.org/2014.amta-wptp.5.pdf . Primary study passages were consulted. Automatic post-editing task record: https://aclanthology.org/C16-1241/ . Examples and numerical outcomes here are original and hypothetical; no study's findings are presented as universal business guarantees.

Previous: Possible world Next: Post hoc analysis