Localising Creative Work, Not Just Translating It

Machine translation is now good enough that its failures are concentrated in one place: everything that was doing more than conveying information. Idiom, humour, register and cultural reference all break, and the automatic metrics used to judge translation quality actively reward the broken version.

Translation, localisation, transcreation

Confusion between the first and third is where projects fail. A tagline is a transcreation job, and briefing it as translation guarantees a literal rendering that is technically correct and commercially dead.

The four things that break

Idiom

An idiom is a fixed expression whose meaning is not the sum of its parts. Machine output handles common ones and produces literal nonsense on the rest, and the failure is silent: the output is grammatical, so nothing flags it.

The harder case is the half-idiom, where a writer has played on an expression. “The ball is in your court, and the court is on fire” needs a target-language expression that can be extended the same way. There may not be one, in which case the correct answer is a different image entirely, which no sentence-level system will propose.

Humour whose mechanism is the language

Jokes that turn on ambiguity, sound or spelling cannot survive translation, because the mechanism is in the source language rather than in the meaning.

SOURCE (a bakery feature)
  Headline: "Bread Winners"

LITERAL TRANSLATION
  A phrase meaning "people who win bread". True, and not a joke.

WHAT IS ACTUALLY REQUIRED
  Identify the mechanism: a compound noun ("breadwinner", a household
  earner) split back into its parts, one of which is the subject.
  Then find a compound in the target language whose parts can be split
  the same way and which touches baking, family income or both. It will
  not be about bread, and it should not be.

That analysis is the job. A model can help with it — asking for twenty target-language compounds containing a food word is a good use — but the decision about which one lands requires knowing what is funny in that language, which is not a property of the source text.

Register and address

English hides a decision that most languages force. “Are you sure you want to delete this?” must choose du or Sie, tu or vous, and in Japanese must select a politeness level. That choice is a brand decision made once and applied everywhere, and a translation system left to itself will make it inconsistently across a product — formal in the error messages, informal in the onboarding, because those strings resemble different training data.

The related failure is direction. A cheerful imperative register that reads as friendly in American English can read as presumptuous or childish elsewhere, and the translation will be accurate while the product feels wrong. In narrative work the same problem appears as character voice, where every speaker converges on one register — the mechanism is identical and only the deliverable differs.

Culture-bound reference

Sports metaphors are the clearest case because native speakers do not notice they are using them: out of left field, ballpark figure, sticky wicket, own goal. Each translates literally into a sentence about a sport the reader may not follow. The same applies to school systems, holidays, food as shorthand for class, and historical events used as adjectives.

The fix is not a target-language equivalent metaphor, which is the reflex answer. It is usually to ask what the metaphor was doing — signalling informality, compressing an explanation, creating a shared wink — and to do that job with whatever the target culture actually has.

Why the metrics penalise the correct answer

This is the mechanism that explains why automated pipelines systematically converge on the flat version, and it is not widely understood outside localisation teams.

BLEU, the long-standing automatic translation metric, scores a candidate by how many n-grams it shares with a reference translation. Neural metrics such as COMET are considerably better at judging adequacy and fluency, but they too are trained against human reference translations and reward correspondence with the source.

Now consider a correct transcreation of the bakery headline. It shares almost no n-grams with any literal reference; its meaning, compared sentence to sentence, is not the source’s meaning. It will score badly on n-gram overlap and mediocre on adequacy, while the flat literal version scores well on both.

• Any system tuned on these metrics is being tuned away from transcreation. That is not a bug in a particular tool; it is what the objective says.
• The correct evaluation is target-side only. Show the target text to target readers and ask whether it is funny, whether it sounds like a person, whether they would say it. The source does not appear in the test.
• Back-translation is a check for meaning drift and nothing else. It will catch a mistranslated number or a reversed negation. It will not catch a register error, and it will flag every successful transcreation as a failure.

A workflow that uses the machine properly

1. Segment by function, not by file. Sort every string into informational (error messages, field labels, legal text), register-bearing (onboarding, empty states, marketing body), and creative (taglines, headlines, jokes, character voice). The proportions are usually about 70, 25 and 5 per cent, and the last 5 per cent carries most of the risk.
2. Write intent notes for everything above informational. For each string: what it must accomplish, what the joke is doing, how formal it is, what it must not imply. This is the artefact that makes transcreation possible, and it is normally missing entirely.
3. Fix the register decision once, globally. Formality, person, and whether the product refers to itself. Put it in the glossary before any string is translated, as a constraint rather than a preference — the same rewrite that turns house style into checkable rules applies, and a glossary violation is exactly the sort of thing a script can fail on.
4. Machine-draft the informational tier. With the glossary and the register decision supplied as context.
5. Human-transcreate the creative tier, briefed from the intent notes rather than from the source string. A model is useful here as an option generator — twenty candidates for a native speaker to choose between — and not as the chooser.
6. Review the register-bearing tier in context, in the actual interface or layout, by a native speaker. Out of context this tier reads as fine and in context it reads as wrong.
7. Test target-side. No reference, no back-translation, no source. Just target readers and the question of whether it works.

Where machine output is genuinely fine

• Register-neutral informational text with a glossary and a fixed formality decision. Error messages, field labels, status text.
• High-volume, low-stakes, user-facing content where the alternative is no translation at all — product listings, support articles, archives. Something imperfect that exists beats perfection that does not.
• First-pass subtitles for review, where a human fixes names, idiom and line breaks. The timing is the expensive part and the machine does it well.
• Comprehension for internal use. Reading an incoming document to decide whether it needs a real translation.
• Terminology consistency checking across a large existing translation, which is a mechanical search job over supplied text and one people do badly.

The general boundary is the same one that runs through this whole cluster: it is safe where the source supplies everything the output needs and unsafe where the output must supply something the source did not have. Idiom, humour and register are all cases where the target text has to contain something the source text does not. The mechanics of machine translation quality covers the underlying models in more detail.

The sign-off rule

One rule, and it is about accountability rather than quality: a named native speaker of the target language signs off on anything that will be read by a customer. Not reviews — signs off, in the sense of being the person whose judgement it was.

The reason is that localisation errors are found by users and reported publicly, and the cost is not the error but the impression that nobody in the company speaks the language. A named reviewer produces a different quality of attention than a review queue, and it is the only control that reliably survives a deadline.