Breaking Language Barriers with Neural MT and Human QA Harmony: The Hybrid Strategy for Global Reach

Updated · 7 min read
On this pageWhat ISO 18587 and TAUS actually defineLight vs full post-editing, side by sideChoosing a level by risk, not by budgetEnglish-to-Korean errors a post-editor fixes every dayHow to write a post-editing briefMaking the hybrid work

A project manager sends 40,000 words of machine-translated Korean to a linguist with one line: "Please post-edit." The linguist can read that as "make it understandable" or as "make it publishable". The two levels of machine translation post-editing differ in time, price and result by a wide margin, and nobody finds out which one was done until the text is live.

Machine translation plus human editing is now the default for a lot of Korean content. The hybrid works when everyone agrees what the human part is supposed to achieve. This post sets out the two standard levels of post-editing, what each one fixes and leaves alone, how to choose between them, and the English-to-Korean errors a post-editor will meet on almost every job.

What ISO 18587 and TAUS actually define

ISO 18587:2017 (Translation services — Post-editing of machine translation output — Requirements) is the international standard for this work. Its scope is narrower than many vendor pages suggest: it provides requirements for the process of full, human post-editing of machine translation output, for translation service providers, their clients and post-editors. It covers the process, post-editor competences and qualifications, and the requirements for full post-editing. It does not evaluate MT engines. ISO currently flags the 2017 edition as one to be revised, so a contract that cites it should name the edition.

The TAUS post-editing guidelines, produced with the CNGL research centre, are the industry reference for the two levels most buyers talk about. They define two target quality levels: "good enough" quality, meaning comprehensible and accurate but not stylistically polished, and quality similar or equal to human translation.

In practice the industry maps these to light post-editing (good enough) and full post-editing (human-comparable). The rules overlap more than people expect. Both levels require fixing meaning errors, restoring omitted or deleting added information, and removing offensive or inappropriate content. The difference is everything beyond meaning.

Light vs full post-editing, side by side

Light post-editingFull post-editing
GoalReader understands the text correctlyText is publishable, comparable to human translation
Meaning errors, omissions, additionsFixedFixed
SpellingFixedFixed, plus punctuation and spacing
Grammar and sentence structureChanged only if it blocks understandingCorrected throughout
Client terminologyUsually not enforcedEnforced from the glossary
Style, tone, naturalnessLeft as isBrought to the brief's style
Formatting, tags, placeholdersOnly if brokenMatch the source
Typical useInternal reading, triage, knowledge base drafts with short lifeCustomer-facing, published, contractual or regulated text

One rule from the TAUS guidance applies to both levels and is the one most often broken: use as much of the raw output as possible, and do not rewrite correct output just because you would have phrased it differently.

Choosing a level by risk, not by budget

Pick the level from what the text will do, then price it. Three questions settle most cases.

Who reads it? If the only readers are your own staff trying to understand a Korean supplier email or a competitor's product page, light post-editing, or even raw MT, may be enough. If customers, regulators or a court will read it, start at full.

How long does it live? A support answer that will be replaced next week can tolerate awkward phrasing. A manual, a contract template or a product page will be reused and copied into other documents, so its errors multiply.

What does an error cost? Light post-editing is designed to leave imperfect text behind. That is fine when an imperfection costs a second of reading time and dangerous when it changes a dosage, a price or the scope of a claim. For patent and legal text, see why post-editing and human translation are not interchangeable for patents; for the wider business risk of skipping the human step, what raw machine translation does to market entry.

Light is not "full, but faster". If you order light post-editing, accept text that is correct but plainly machine-made. Asking for a light rate and then rejecting the result for style is the most common post-editing dispute, and the brief is where it is prevented.

Confidentiality is a separate decision from quality level. Before any text goes into an MT engine, check where it is processed and stored; the data exposure risks of machine translation covers what to ask.

English-to-Korean errors a post-editor fixes every day

Korean and English differ in ways that neural engines still handle unevenly. These are the patterns worth naming in a brief, with the level that should catch each one.

Politeness level

Inconsistent speech level

Source: "Enter your password." / "Your order has shipped."Risk: one string comes out formal, the next casual or overly deferential.Korean marks the relationship to the reader in verb endings; English does not. Research on automatic translation of Korean honorifics found that some models systematically overuse polite forms compared with human judgments, and that output shifts when the addressee is specified. Full post-editing enforces the one level the style guide sets. Light post-editing usually leaves it.
Pronouns

"You" and "it" carried over literally

Source: "If you lose it, you can reset it from your account."Risk: every pronoun translated, or "it" tied to the wrong noun.Korean usually leaves out subjects and objects that are clear from context, so a word-for-word rendering of every "you" sounds stilted, and an engine guessing what "it" refers to can guess wrong. A wrong referent is a meaning error and is fixed at both levels; the stilted repetition is fixed only in full.
Number and articles

"A device" vs "the devices"

Source: "The controller sends a signal to the sensors."Risk: singular and plural, and "a" vs "the", collapse into one form.Korean has no articles and often leaves plural unmarked. In specifications, claims and contracts it can change scope, so the brief should tell post-editors to mark number and definiteness wherever the source depends on them.
Terminology

Same term, three translations

Source: "workspace" in the UI, the help center and the release notesRisk: each segment translated on its own, with a different equivalent each time.Engines mostly work sentence by sentence. Without an enforced glossary, consistency is luck. Full post-editing applies the client glossary; light usually does not. See how terminology management keeps product language consistent.
Modifiers

What does "open" attach to?

Source: "Check the valve on the pump that is open."Risk: the engine picks one reading and makes it explicit.Korean places modifiers before the noun, so the translation must commit to whether the valve or the pump is open. If the source is ambiguous, the post-editor should query, not guess.
Negation

"Not all" becomes "none"

Source: "Not all models support this feature."Risk: rendered as "no model supports this feature".Scope of negation is a classic meaning error and must be fixed at any level.

Add the mechanical checks to every job: numbers, units, dates, product names left in English as the style guide requires, and placeholders such as {count} or HTML tags that the engine moved or translated.

How to write a post-editing brief

A good brief fits on one page and removes the guesswork. Include:

  1. Level: light or full, with the target stated in words ("understandable for internal engineers" or "publishable on the Korean website").
  2. Audience and purpose: who reads it, and what they do after reading.
  3. Speech level and tone: one politeness level for the whole project, with an example sentence in English describing the register.
  4. Glossary and do-not-translate list: product names, UI labels, legal terms, and whether they are mandatory at this level.
  5. What not to touch: no preferential rewrites; leave correct output alone.
  6. High-risk elements: numbers, negation, singular and plural, safety warnings, prices, claims. These are checked against the source at both levels.
  7. Query route: who answers questions about ambiguous source text, and how fast.
  8. Acceptance check: how the result will be reviewed, by whom, and what counts as a failure.

The acceptance check matters most. Agree it before work starts so that a reviewer measures the delivery against the level ordered, not against an ideal human translation. A sampling approach works well for high-volume light post-editing; see risk-based QA methods that cut review cost. For fast-moving content such as news and statements, where human post-editing still protects reputation shows where the line sits. For how engines and the rules around them changed over the past year, see our review of 2025 translation technology trends.

Making the hybrid work

The hybrid model is a set of choices about where people spend attention. Route each content type on purpose, write the level into the brief, and check the result against what was ordered. The machine then handles volume while people handle meaning, tone and risk.

Not sure which level your Korean content needs?

We map your content types to raw MT, light or full post-editing, or human translation, and write the brief and acceptance checks with you. Talk to us → or see our services.

Work with us

Machine speed, human judgment.

Post-editing and review of machine-translated Korean content, scoped to the risk of each document.

Get a free quote