Why Japanese Has No Grammatical Plural and What That Means for AI Text

本 is “book” and “books” and “the book” and “some books”. Japanese does not mark number on nouns at all, which is why a plural suffix appearing in translated Japanese is nearly always an error — and why the specific suffix models reach for changes the meaning rather than just sounding odd.

Number is simply not marked

In English, number is obligatory: you cannot say “I bought book” without committing to singular or plural. Japanese has no such requirement and no morphology to express it. A bare noun is unspecified for number, and the reader takes the count from context, or does not need it.

ファイルを削除しました。
"Deleted the file."  /  "Deleted the files."  — both, equally

昨日、友達に会った。
"I met a friend yesterday."  /  "I met friends yesterday."

This is not a gap in the language. It is a different division of labour: where English forces number on every noun and leaves definiteness to articles, Japanese leaves both unmarked and supplies them with particles, context, and counting expressions when they matter. The relevant consequence for anyone building software is that translating into Japanese loses information harmlessly, and translating out of Japanese loses it destructively — an English target sentence must choose a number the Japanese source never stated.

CLDR reflects this directly: Japanese has exactly one cardinal plural category, other, as do Chinese, Korean, Thai, Vietnamese and Indonesian. A Japanese message needs one form for every count, which is the reason a naive localisation of an English two-branch string into Japanese looks fine and a naive localisation of a Japanese string into Polish does not. The full picture is in how many plural forms a language actually needs.

How number is expressed instead

When the count matters, Japanese states it with a numeral and a classifier — a counter word chosen by the shape, category or animacy of the thing counted. The noun does not change.

本を三冊買った。      three books        冊 (さつ)  bound volumes
紙を二枚ください。    two sheets         枚 (まい)  flat thin things
ビールを一本         one bottle         本 (ほん)  long cylindrical things
犬が四匹いる。        four dogs          匹 (ひき)  small animals
人が五人来た。        five people        人 (にん)  people
車を二台             two cars           台 (だい)  machines, vehicles

Note 本 doing double duty as the noun “book” and as the counter for cylindrical objects, and note the sound changes: 一匹 is ippiki, 三匹 is sanbiki, 六匹 is roppiki. Counters are where generated Japanese most often goes wrong on number, and the failure is usually choosing the wrong counter for the noun rather than mishandling the numeral — 三個 where 三冊 belongs is grammatical, comprehensible and wrong in the way a native speaker notices instantly.

～たち is an associative plural

Japanese does have suffixes that look like plurals — ～たち, ～ら, ～ども, and reduplication with the iteration mark 々 — but they behave differently from an English plural in two important ways.

• They are optional. 学生が来た is fine for any number of students. 学生たち adds emphasis on plurality; it does not become required when there is more than one.
• They attach to animates, chiefly people. 私たち, 子供たち, 彼ら are natural. 犬たち is acceptable in an affectionate register. 本たち and ファイルたち are not Japanese — they occur only as deliberate personification.
• They are associative, not additive. This is the part that matters most. 田中さんたち does not mean “several people called Tanaka”. It means “Tanaka and the people with him” — a group identified by one member. Applied to a common noun the reading is looser, but the associative sense is the core one, and it is why the suffix cannot be used as a general pluraliser.

Reduplication is a separate, closed device: 人々 (people), 山々 (mountains), 国々 (countries), 時々 (sometimes), 我々 (we). It applies to a fixed list of nouns and often carries a connotation of variety or extent rather than plain count. It is not productive: you cannot invent 本々.

The over-correction error

The characteristic machine error in Japanese is not omitting number marking — it is adding it. A model translating from a language where plurality is obligatory has an English plural morpheme in front of it, and the strongest available correspondence for “plural” in its Japanese vocabulary is ～たち. So it emits one.

English   All the files were deleted.
wrong     すべてのファイルたちが削除されました。
right     すべてのファイルが削除されました。

English   Please review the documents.
wrong     ドキュメントたちを確認してください。
right     ドキュメントをご確認ください。

English   Our users love it.
odd       私たちのユーザーたちは...
right     ユーザーの皆様には...

Two things make this worse than a typical fluency error. It is systematic, so it appears on every plural noun in a document rather than occasionally. And it survives review by anyone checking against the English source, because the Japanese now matches the English more literally than correct Japanese would — the reviewer sees a faithful rendering of “files”.

The reverse error exists too and is quieter: translating out of Japanese, the model must invent a number, and it will pick one confidently. ファイルを削除しました becomes “the file was deleted” or “the files were deleted”, and if the sentence is a log message or a confirmation, the wrong choice is a factual claim about what happened. Where this matters, keep the Japanese and render the count from your own data rather than from the translation.

Consequences for strings and templates

For interface text, the single-category rule makes Japanese easy in one respect and awkward in another. Easy, because one string covers every count. Awkward, because an English source string that reads naturally with a plural branch usually needs restructuring rather than translating: 3 items selected becomes 3件選択中, using a counter, not a plural. Ask a model for the counter that fits the noun before you build the template, and put the counter in the template rather than in the interpolated value.

Two further points that catch people. The counter for a noun is a per-noun fact, so a generic template shared across entity types cannot be right for all of them — 件 is the safe general counter for records and cases and is what most Japanese software uses for exactly this reason. And numbers in Japanese text may appear as ASCII digits, full-width digits or kanji numerals depending on context and vertical layout, which is a normalisation question rather than a grammar one; it sits with the wider script-mixing problem covered in Japanese script mixing across kanji, kana and romaji.