Localisation Is Not Translation: Cultural Rewriting for Multi-Market Games

A party card can read correctly in six languages and still fall flat in five of the markets it ships to. For humour, social risk and anything close to a taboo, we think the right operation is cultural rewriting. What gets localised is the effect the card is meant to have on the player, and the sentence on the card is only one way of getting that effect. When translation is the default, the usual result is text that is accurate and does nothing at the table. The sections below cover how to sort a corpus by what will not survive transfer, how to choose between domestication and foreignisation on purpose, how to run a rewriting pass with native writers in charge instead of native reviewers, and how to judge the result by how players in each market behave, which translation-review scores cannot tell you.

The functional unit of localisation: card effect versus card text

The framing comes from functional translation theory. Skopos theory (Reiss & Vermeer, 1984) holds that a translation is adequate if it serves the purpose of the target text in the target situation, and that this purpose, more than equivalence with the source, is what it should be judged by. Nida's earlier distinction between formal and dynamic equivalence (Nida, 1964) gets to the same place from another direction. He judged a rendering by the response it produces in its readers, and structural correspondence to the original came second.

Interactive content makes this easy to formalise, because you can watch a card do its job at the table. Treat a card as a function from a group state to an intended effect, and describe the effect with four components:

effect(card, players) = (recognition, risk, compliance, transgression)

localisation goal:  effect(c_B, players_B) ≈ effect(s_A, players_A)
translation goal:   semantics(c_B) ≈ semantics(s_A)

Recognition is whether players understand the situation without explanation. Risk is how exposed the answering player feels. Compliance is whether the action gets performed, and transgression is how far past ordinary conversation the card goes. For instructional text the two goals coincide, but once a card rests on shared assumptions they pull far apart. Take a card whose job is to make one player admit something mildly embarrassing at a controlled level of risk. It is working when players hesitate for a moment and then answer. A lexically perfect translation of it can land as bland, where nobody hesitates, or as hostile, where nobody answers, and you cannot see either failure by reading the translated string.

So the design question comes before the language question. Before any locale work starts, each card needs its function written down: rule type, difficulty tier, risk tier. Those stay fixed, while the wording is only one way of meeting them in one culture and is free to change.

A taxonomy of content that fails to travel between cultures

Sort the corpus by fragility before commissioning locale work. It tells you what can be translated cheaply and what has to be rewritten at real cost.

Wordplay and form-dependent humour

If a joke works through the surface form of the source language (rhyme, homophony, an idiom taken literally, a pun on a brand name), there is nothing in it to transfer. What has to be recreated is the mechanism, usually from unrelated material.

Register and politeness distance

Languages differ in how much social distance their grammar makes a speaker declare. German and Turkish force a choice between familiar and formal address in nearly every sentence. English does not, so a translator working from English has to invent one, and whatever they pick changes the game's relationship with the player. Formal address in a drinking game reads as parody or as a telling-off. Familiar address in a market that expects distance comes across as presumptuous.

Taboo proximity and social risk

Money, family, religion, sexual history, political affiliation and appearance rank differently from one culture to the next, and the penalty for crossing a line changes too. Exposure is a separate axis. Two cards can touch an equally acceptable topic and still carry very different risk, depending on who has to answer, whether the answer is public, and whether refusing costs anything. Mechanics that concentrate attention on a single player, such as a spotlight or a duel, amplify whatever miscalibration is already in the text.

Culturally bound references

Television, sport, snacks, school rituals, transport, local celebrities. These fail loudly and are easy to catch. The trouble is the usual fix, which is to swap one proper noun for another that is supposedly just as famous. A reference that is just as well known can still do a different job in the sentence, and the job is what has to carry over.

Grammar the source never had to declare

A template reading {name} owes {count} sips hides at least three decisions: the plural category, the gender of any adjective agreeing with the named player, and the case the name takes in a language with case marking. English marks none of them, so a template written in English quietly leaves out information that every other locale needs.

Domestication versus foreignisation in game localisation

Venuti (1995) describes two orientations. Domestication adapts a text to the target culture until you can no longer tell it was translated. Foreignisation keeps the strangeness of the source. Neither is right in general. The common failure in practice is never making the choice at all, so the corpus domesticates some cards, foreignises others, and reads as inconsistent. For a party game the default should be domestication, because a card has to be understood the moment it is read, and a card that needs a footnote has already lost the round. Where the source culture is part of the appeal, the choice is open. Make it per content class and write it into the brief.

Our own case is GUZZL, live on Google Play with more than 1,080 handcrafted cards across six languages. None of the six versions is a translation of another. They share rule types, difficulty tiers and pacing structure, but the sentences in each language were written for that culture. We took on the cost knowingly: six corpora to maintain, six sets of weekly rotation content, and no help from translation memory. In return, each card was written to work in its own locale, without having to match an English original.

A cultural rewriting protocol for multi-market card content

Start from an effect brief. Each card enters localisation with its rule type, difficulty tier, risk tier and intended effect. The English text comes along as one example implementation and is clearly marked as non-binding.

Put native rewriters in charge. A reviewer can only correct the text in front of them, and the failures above do not show up at sentence level, because the sentence itself is fine. Authorship has to sit with someone who plays in the target market.

Write a local replacement instead of swapping nouns. When a reference cannot travel, write a new one that does the same job locally, and accept that it will look nothing like the source when the two are side by side. A one-to-one swap usually keeps the shape of the sentence and loses the reason it was there.

Freeze the structural invariants. Rule type, difficulty tier and risk tier must come through rewriting unchanged. In GUZZL they carry mechanical weight: fourteen distinct rule types, boss rounds every ten cards, fifteen lives, and duel and spotlight mechanics that change how much exposure a card imposes. A rewriter who quietly shifts a card's risk tier changes the pacing of the boss round it feeds.

Test through play in the market. Back-translation only checks semantics, and the rewrite was allowed to change those. What you need to know is whether players hesitate, laugh, comply, or quietly skip.

The same split applies outside games. RelationCRM, which is in development, drafts messages in six tones across ten languages. There the tone has to stay fixed while the phrasing changes per language. A warm but brief message in another language is not a literal rendering of a warm but brief English sentence, and if you model it that way, native speakers find the output correct and unusable.

Localisation infrastructure: ICU messages, CLDR plurals and layout

Cultural rewriting can fail on infrastructure just as easily as on judgement.

ICU message formatting with real selection. Templates must express plural and gender selection declaratively, without string concatenation:

{count, plural, one {# sip} other {# sips}}
{gender, select, female {she} male {he} other {they}}

CLDR plural categories, read from the data. The Unicode Common Locale Data Repository (Unicode Consortium, Unicode Technical Standard #35) defines which categories each language actually uses. English needs two, Arabic uses six (including zero and two), and Chinese and Japanese resolve to one. Turkish has one and other, but Turkish nouns usually drop the plural suffix after a numeral, so a template can pick the right category and still read wrongly.

A string-expansion budget in the UI. As an illustrative planning figure, assuming short imperative card text and no compounding, we size German and Spanish card bodies at up to 1.3× the English character count and reserve layout for that. The exact factor is less important than choosing one before layout starts. Overflow is also an accessibility failure. WCAG 2.2 requires content to reflow and to survive increased text spacing, so a fixed-height card that clips its text fails a conformance criterion as well as looking wrong.

Right-to-left readiness before it is needed. That means logical layout properties instead of physical ones, mirrored directional icons, and bidirectional isolation around interpolated player names. Retrofitting all of this after a corpus exists costs substantially more. Nielsen's heuristic of matching the system to the real world (Nielsen, 1994) applies here, and for a right-to-left reader the reading direction is part of that world.

Measuring localisation quality by in-market outcomes

Translation-review scores measure fidelity to a source the rewrite deliberately moved away from, so they cannot catch the failures that matter. Behavioural proxies taken from play can.

Measure these per card and per locale, and compare a card's locale variants with each other instead of with a global average. If a card is skipped by a small minority of players in one locale and by a third of them in another, it has a localisation defect, whatever its review score says. Weekly content rotation makes this manageable. Each rotation is a natural cohort boundary, and a card that underperforms in exactly one locale should go back to that locale's rewriter before anyone considers retiring it.

Diacritic loss as a production defect class

We have shipped stripped diacritics to production, and the class of defect behind that incident matters more than the incident itself. It rarely starts with a translator. The damage usually happens somewhere in the pipeline: a spreadsheet export with a default codepage, an ASCII-folding step written for URL slugs whose output leaks into display strings, a font subset with missing glyphs, or a case transformation that ignores locale. Turkish shows the last one clearly. Correct casing needs the dotted and dotless i handled under Turkish rules, and a case fold run under a default locale converts them wrongly.

This is more than cosmetic. In Spanish, "años" and "anos" are different words, and in Turkish a dropped diacritic can change both the meaning and the register around it. A game that depends on calibrating social risk cannot afford that, because one mangled word is enough to make a card look cheap.

The class is cheap to gate, though. Each locale declares the non-ASCII characters it expects. Continuous integration asserts that rendered strings contain characters from that set and that no display string has gone through the folding path. Casing operations carry a locale tag at the call site. This catches the problem before release. Sentence-level review usually does not, because a reviewer reading a stripped string tends to read straight through the damage.

Limitations

We have not run a controlled comparison, with a translated corpus and a rewritten one side by side in the same market, so the case here rests on the functional argument and on what we saw in production, with no measured effect sizes. The first-party figures describe what we built. They do not show that the approach caused any result. The behavioural metrics are proposals. Nobody has tested how sensitive they are with small player populations, and a difference in skip rate between locales may come from who is playing and not from the card. The scope is also narrow. Party games are the fragile extreme, and running this protocol on transactional text would not be justified. Where we lean on cultural dimensions at all, note that Hofstede's framework (Hofstede, 1980) describes national averages from one population at one point in time. It can suggest where a card might miscalibrate, but it cannot predict how an individual player will react and does not replace asking people in the market.

References

• Hofstede, G. (1980). Culture's Consequences: International Differences in Work-Related Values.
• Nida, E. A. (1964). Toward a Science of Translating.
• Nielsen, J. (1994). Usability Engineering.
• Reiss, K., & Vermeer, H. J. (1984). Grundlegung einer allgemeinen Translationstheorie.
• Unicode Consortium. Unicode Technical Standard #35: Locale Data Markup Language (LDML).
• Venuti, L. (1995). The Translator's Invisibility: A History of Translation.
• W3C. (2023). Web Content Accessibility Guidelines (WCAG) 2.2.