Whether AI Models Can Generate Cantonese Instead of Mandarin

#ai #llm #machinelearning #programming

Cantonese and Mandarin are not mutually intelligible, and Cantonese has its own grammar, not just its own pronunciation. Asking a model for Cantonese usually returns Standard Written Chinese — Mandarin-based writing that a Cantonese speaker can read aloud in Cantonese but would never say.

Model capability on Cantonese is exactly the kind of claim that ages, so this page does not assert what any specific model can do today. It gives you the grammatical test instead, which stays true. The diagnostic below takes about two minutes and answers the question for the model you actually intend to use.

What the question actually asks

There are three distinct things people mean by “Cantonese output”, and separating them is most of the work:

Traditional characters. A script question, not a language question. Covered in Simplified versus Traditional Chinese, and models generally handle it.
Standard Written Chinese, read in Cantonese. This is what almost all formal Hong Kong writing is: newspapers, government notices, contracts. The grammar and vocabulary are Mandarin-based; a Cantonese speaker reads the characters with Cantonese pronunciation. This is what a model will usually give you, and for many purposes it is the correct output.
Written vernacular Cantonese. Cantonese grammar and vocabulary written down, using a set of characters that mostly do not appear in Mandarin writing. This is what appears in Hong Kong internet forums, informal messaging, comics, advertising copy and some subtitling. It is what people mean when they are dissatisfied with the previous answer.

Only the third is a language-capability question, and only the third is hard.

The markers of genuine written Cantonese

These substitutions are near-total: written Cantonese uses the left-hand form where Standard Written Chinese uses the right. If a response contains none of the left column, it is not Cantonese regardless of what script it is in.

function            Cantonese   SWC / Mandarin
------------------------------------------------
"to be" (copula)     係          是
negation             唔          不
"not have"           冇          沒有
genitive / 's        嘅          的
"at, located in"     喺          在
perfective aspect    咗          了
he / she / it        佢          他 / 她 / 它
plural suffix        哋          們
"what"               乜嘢 / 咩    什麼
"this / these"       呢           這
"that / those"       嗰           那
"to eat"             食           吃
"to drink"           飲           喝
"to give"            畀           給

Beyond substitution, three structural features are harder to fake and therefore better tests:

Double-object order is reversed. Cantonese puts the direct object before the indirect: 畀本書我 (“give book me”). Mandarin puts the recipient first: 給我一本書. A model that has substituted 畀 for 給 but kept Mandarin word order has produced Mandarin in disguise, and this is the most common way the output fails.
Classifiers can stand without a numeral. Cantonese allows 本書 for “the book” where Mandarin generally requires 這本書 or 一本書. The bare classifier carries definiteness.
Sentence-final particles carry meaning. 啦, 咩, 㗎, 喎, 囉, 嘅啫 and others encode mood, evidentiality and politeness. Their absence is the surest sign that the output is Mandarin structure with Cantonese vocabulary sprinkled on.

Why the corpus is the constraint

Cantonese has roughly 85 million speakers, which by speaker count puts it among the larger languages in the world — comparable to German. Its written corpus is nothing like that size, and the reason is sociolinguistic rather than technological.

Everything formal in Hong Kong and Guangdong is written in Standard Written Chinese. School instruction in writing is in SWC. Newspapers, legal texts, official documents, academic work and most books are SWC. Vernacular Cantonese writing is confined to registers that are informal by definition, and informal text is exactly the text least likely to be well represented in a curated pretraining corpus. So a language with 85 million speakers contributes a written corpus more typical of a language with a few million.

Two technical consequences follow. Cantonese-specific characters such as 嘅, 咗, 喺 and 冇 sit outside the frequency range that earns tokenizer merges, so they fragment — the mechanism in the tokenizer vocabulary bottleneck. And some characters used in Hong Kong writing come from the Hong Kong Supplementary Character Set, whose coverage in fonts and normalisation pipelines is uneven. Cantonese has its own ISO 639-3 code, yue, which is worth using in metadata and in prompts precisely because it removes the ambiguity that “Chinese” carries.

A diagnostic you can run

This takes two minutes and tells you what any given model does, without relying on anything asserted here.

Ask the model, in Cantonese, to say “I don’t have that book, give it to him” in written vernacular Cantonese, and to use no Standard Written Chinese.
Check the negation. Genuine Cantonese uses 冇 for “don’t have”. If the output has 沒有, it is SWC.
Check the pronoun and the copula. Look for 佢 rather than 他, and 係 rather than 是.
Check the double-object order. The Cantonese form puts the book before the recipient — 畀本書佢. If the recipient comes first, the model substituted vocabulary and kept Mandarin syntax.
Check for a sentence-final particle. A natural Cantonese sentence in this register usually carries one; a bare declarative with none is a translation artefact.
Repeat with a longer passage, 300 words or so. Cantonese adherence typically decays with length in the same way every other variety instruction in this cluster does, and a model that passes on one sentence may not hold across a paragraph.

What to expect, and what to do

The realistic expectation is a spectrum rather than a yes or no. Models will generally produce Traditional characters on request, usually produce recognisable Cantonese vocabulary when asked explicitly for vernacular Cantonese, and are least reliable on the structural features — word order, bare classifiers, particles — which is precisely the order you would predict from a corpus that contains a lot of Cantonese words embedded in Mandarin-structured text.

Three practical responses. Ask for it by name and by code: “write in vernacular written Cantonese (粵語書面語, yue), not Standard Written Chinese” — naming the thing you do not want matters here more than usual, because SWC is the default that has to be displaced. Supply several sentences of genuine Cantonese as few-shot examples, since the structural features are learned from examples far better than from description. And check the output with the markers above before it reaches a reader, because the failure mode is fluent and a non-speaker cannot see it.

For anything formal — terms of service, official notices, documentation aimed at Hong Kong — Standard Written Chinese in Traditional characters is not a failure to produce Cantonese. It is the correct register, and vernacular Cantonese would be wrong.