Gender Agreement Errors in Machine-Translated Arabic

“I am a teacher and I am tired” has two correct Arabic translations, and the English sentence contains nothing that chooses between them. Most reported Arabic gender errors are this: not a mistake, but a forced guess about information the source did not carry.

The output you are looking at

English   I am a teacher and I am tired.

Arabic    أنا مدرّس وأنا متعب.        (speaker male)
          أنا مدرّسة وأنا متعبة.       (speaker female)

English   Are you sure you want to delete your account?

Arabic    هل أنت متأكد أنك تريد حذف حسابك؟      (addressee male)
          هل أنتِ متأكدة أنكِ تريدين حذف حسابكِ؟  (addressee female)

In the second pair, five things change: the pronoun أنت / أنتِ, the predicate adjective متأكد / متأكدة, the pronoun suffix on أنك, the verb form تريد / تريدين, and the possessive suffix on حسابك. The English carries one word, you, that determines all five and specifies none of them.

How many slots one sentence opens

Arabic agreement is denser than in the European languages usually cited for this problem. Four categories of slot:

• Finite verbs agree with their subject in gender in the second and third person, in both the perfect and the imperfect: كتب (he wrote) / كتبت (she wrote), تكتب (you m. write / she writes) / تكتبين (you f. write). Every verb in the sentence is a slot, not just the main one.
• Adjectives and participles agree with their noun, usually by the feminine marker ة (tāʾ marbūṭa): كبير / كبيرة. Predicate adjectives, attributive adjectives and passive participles are all separate slots.
• Attached pronoun suffixes. Possession and object marking are suffixes, and the second person singular suffix is ـكَ masculine and ـكِ feminine — a difference carried entirely by a short vowel mark that is usually not written. This is worth pausing on: in undiacritised text the two are the same letters, so the distinction is present in speech and in careful writing and absent from the string. A model that produces the unvocalised form has not resolved the question; it has hidden it.
• Number interacts. Arabic has a dual as well as a plural, and the plural of human nouns is gendered while the dual is marked separately again — so the addressee axis is not two-way but أنتَ, أنتِ, أنتما, أنتم, أنتنّ. Five second-person forms, one English you. The number side of this is covered in Arabic dual and plural forms in generated text.

Where the source is silent, the output skews masculine, and that direction is documented rather than anecdotal: Stanovsky, Smith and Zettlemoyer (ACL 2019) built the WinoMT challenge set specifically to measure whether translation systems resolve gender from context or fall back to a stereotype, and found systematic masculine defaults across the commercial and academic systems they tested, including into Arabic. The finding to take from it is the direction and the mechanism; a per-system error rate from 2019 is not a number to quote about a model shipped since.

The second person, and why UI strings are unsolvable

For prose you can supply the missing fact. For interface strings you often cannot, because the addressee is whoever is logged in and your product may not know their gender — and asking for it to render a confirmation dialog is not a trade most teams want to make.

The strategies that actually ship, in rough order of preference:

• Rewrite to avoid the second person. Arabic UI localisation leans heavily on verbal nouns (maṣdar) for exactly this reason: حذف الحساب — “deletion of the account” — addresses nobody and needs no gender. Buttons, menu items and titles take this form naturally, and it is what professional Arabic localisation of major software already does.
• Use the impersonal or the plural. Formal address in the masculine plural is a recognised convention in institutional Arabic and reads as neutral rather than as wrong, though it is a choice with critics.
• Store gender as a variable when you legitimately have it. If it is already a profile field, treat it as a message-selection parameter like a plural category, not as something to interpolate. ICU MessageFormat has a select construct for precisely this.
• Do not solve it with a slash. متأكد/ة occurs informally but is not acceptable in most professional copy and interacts badly with right-to-left rendering.

The rule models get wrong even with full information

One Arabic agreement rule is not about missing information at all, and it is the one to check first in generated Arabic because it is counter-intuitive to speakers of European languages.

A plural of a non-human noun takes feminine singular agreement. “The books are new” is الكتب جديدة — the adjective is feminine singular, not masculine plural, because كتب denotes non-rational entities. Likewise the verb: “the files arrived” is وصلت الملفات, feminine singular, not وصلوا. Human plurals behave differently and take real plural agreement.

A model translating from English, where the plural is a plural, has to apply a rule that depends on a semantic property of the noun — rational versus non-rational — that the English marks nowhere. Generated Arabic that reads fluently and uses masculine plural agreement with inanimate subjects is one of the clearest signatures of machine translation, and it is worth putting at the top of a review checklist because it is frequent, systematic and easy to spot once you know it.

What to do

1. Separate the two failure classes before you debug. If the correct answer is not derivable from the input, no prompt fixes it and you are choosing a policy, not a model. If the answer is in the input — a name, a previous sentence, a metadata field — then it is a context problem and is fixable.
2. Pass gender as labelled data, not as a hint. A line such as speaker: female; addressee: unknown above the text is far more reliable than hoping the model infers it from a name, since many Arabic and non-Arabic names are ambiguous and the inference is a second guess stacked on the first.
3. Ask for both forms when the answer is unknown and let the calling code choose, rather than accepting an arbitrary default that will be wrong for roughly half of your readers.
4. Translate paragraphs, not sentences, so that a gender established in sentence one is still visible at sentence four.
5. Review for non-human plural agreement first. It is the highest-yield single check on Arabic output and it requires no knowledge of your content.