Arabic Dual and Plural Noun Forms in AI-Generated Text

English has two grammatical numbers and one rule. Arabic has three numbers, three plural formations, and a counting system where the noun goes back to the singular above ten. A model that has learned the English mapping of “more than one means plural” produces Arabic that a reader recognises as wrong immediately.

Singular, dual, plural

Arabic marks the dual as a grammatical category of its own, not as a special case of the plural. It is a suffix on the noun, and it is not optional: for exactly two of something you use the dual, and the numeral اثنان is normally omitted because the noun already says two.

noun: كتاب  kitāb  "book" (masculine)

singular   كتاب        kitāb        one book
dual       كتابان      kitābān      two books   (nominative)
           كتابين      kitābayn     two books   (accusative / genitive)
plural     كتب         kutub        books

noun: مدينة  madīna  "city" (feminine)

singular   مدينة       madīna
dual       مدينتان     madīnatān
plural     مدن         mudun

The dual is not confined to nouns. Adjectives take dual agreement, verbs have dual conjugations, and there is a dual pronoun هما distinct from both the singular and the plural. A sentence about two people is dual throughout, and a generated sentence that uses a dual noun with a plural verb is wrong in a way that is immediately audible.

Three kinds of plural

• Sound masculine plural. A suffix, ـون in the nominative and ـين otherwise. Applies mostly to male human nouns and participles: معلم (teacher) becomes معلمون.
• Sound feminine plural. The suffix ـات, replacing the feminine ending: معلمة becomes معلمات. Also used for many loanwords and abstract nouns regardless of gender.
• Broken plural. The word is restructured internally rather than suffixed, following one of dozens of patterns: كتاب → كتب, بيت → بيوت, رجل → رجال, قلم → أقلام. Which pattern a noun takes is lexical information stored per word, not derivable from the singular.

The broken plural is the majority case for ordinary nouns, and it is the one where generation fails in a specific way: a model that has not seen a particular noun’s plural frequently enough will invent a plausible pattern. The result is a word that follows a real Arabic template and is not the plural of that noun, which reads as an error rather than as a typo and is hard to catch with a spell checker because the invented form is often a real word.

The counting rules, which invert

This is the part that has no English analogue at all. Which form of the noun follows a numeral depends on the numeral’s value, and the rule reverses above ten.

count      noun form                       example
-------    ---------------------------     ----------------------
1          singular (numeral optional)     كتاب واحد     one book
2          DUAL (numeral omitted)          كتابان        two books
3–10       PLURAL, genitive                ثلاثة كتب     three books
11–99      SINGULAR, accusative            أحد عشر كتاباً  eleven books
                                           عشرون كتاباً    twenty books
                                           تسعة وتسعون كتاباً  ninety-nine
100, 1000  SINGULAR, genitive              مئة كتاب      a hundred books
           and multiples                   ألف كتاب      a thousand books
0          plural or singular, genitive    لا كتب        no books

Twenty books is literally “twenty book”, with the noun in the singular carrying an accusative ending. This construction is called tamyīz, the specifier, and it is not a stylistic option. A model that writes عشرون كتب — twenty with the plural — has produced the sentence an English-trained pattern predicts and the sentence Arabic does not allow.

There is a second inversion inside the 3–10 range. The numeral itself takes the opposite gender marking from the noun it counts: three books, with masculine كتاب, uses the feminine-marked ثلاثة; three cities, with feminine مدينة, uses ثلاث without the ending. This polarity applies to 3 through 10 and to the unit digit of 13 through 19, and it is a rule generated text drops constantly, because it is exactly the kind of agreement that looks like an error when it is correct.

CLDR encodes this counting structure directly. Arabic is one of the few languages using all six plural categories, and the boundaries map onto the rules above: zero for 0, one for 1, two for 2, few for n mod 100 in 3–10, many for n mod 100 in 11–99, and other for everything else including 100 and 1000. The category list is published in the CLDR language plural rules chart, which is the thing to check rather than any prose description including this one.

Non-human plurals take feminine singular

One more rule, and it is the one that most often makes otherwise good generated Arabic read as translated. When a plural noun refers to something non-human, everything agreeing with it — the adjective, the verb, the pronoun — takes feminine singular agreement rather than plural.

The new books        الكتب الجديدة
                     al-kutub al-jadīda
                     plural noun + FEMININE SINGULAR adjective

The books arrived    وصلت الكتب
                     waṣalat al-kutub
                     FEMININE SINGULAR verb + plural noun

but, for people:

The new teachers     المعلمون الجدد
                     al-muʿallimūn al-judud
                     plural noun + PLURAL adjective

A model producing الكتب الجديدون has applied plural agreement to a non-human plural. Grammatically it is a clear error; stylistically it is the single most recognisable marker of machine-translated Arabic.

Where generated text goes wrong

• Plural for two. ٢ كتب instead of كتابان. The most common failure, and the one that appears whenever a count is interpolated into a template rather than generated in context.
• Plural above ten. ١١ كتب instead of أحد عشر كتاباً. Predictable from the same English pattern.
• Missing numeral polarity. ثلاث كتب instead of ثلاثة كتب.
• Invented broken plurals for uncommon or borrowed nouns.
• Plural agreement on non-human plurals.
• Dropped case endings. The tamyīz construction depends on an accusative ending that is often unwritten in informal text, so a model trained heavily on informal Arabic omits it in formal registers where it is required.

The mitigation is structural rather than prompt-level. Do not ask a model to inflect a noun for a count you already know: put the count into an ICU message with the six categories and have the six forms authored once by somebody who speaks Arabic. Reserve generation for prose where the count is part of the content, and have that reviewed. Where a count and a noun meet in a string template, the plural library is right by construction and the model is right by luck.