Why most chatbots fail at Lebanese Arabizi — and what fixes it
Arabizi is Levantine Arabic written in Latin letters with digits for Arabic sounds. Most AI handles it badly for a specific, fixable reason. Here is what actually goes wrong.
What Arabizi is
Arabizi — also called Franco-Arabic, Franco, or Arabish — is Arabic written in Latin letters, using digits for sounds Latin has no letter for. The 3 stands in for ع, 7 for ح, 2 for ء or ق, 5 for خ.
"shu 3ndkon" is "what do you have". "ade l delivery" is "how much is delivery". "bdk shi tene" is "do you want anything else". Across Lebanon and much of the Levant this is simply how people type, and a business that only understands formal Arabic or English is not understanding its customers.
Why models get it almost right, which is worse
A large language model has seen enough Arabizi to understand it. Understanding is not the problem.
The problem is producing it. Arabizi has no standard spelling — everyone writes it slightly differently, and the differences are regional and personal. A model trained on whatever Arabizi happened to be on the internet reaches for the commonest forms it saw, which are frequently Egyptian or Gulf rather than Lebanese.
The result reads as fluent and foreign at the same time. A Lebanese customer sees "baddak", "addesh", "aywa", "tamam" and knows immediately that whoever wrote it is not from here — the same way an English reader clocks a translated sentence that is grammatically perfect and obviously not written by a native.
The second failure: invented words
Worse than the wrong spelling is the word that does not exist. Asked to say something it has no vocabulary for — an order status, a shipping update — a model will transliterate from Arabic it half-knows and produce a word no speaker has ever used.
A spelling correction cannot catch this, because you cannot list every word that does not exist. The only fix is vocabulary: give the model the actual phrase for the actual situation, rather than leaving it to invent one.
What actually fixes it
- Measure how the business really writes, rather than guessing. A correction table built from a real person's messages beats one built from general knowledge of the dialect.
- Enforce it, rather than requesting it. A style instruction in a prompt is a suggestion the model can ignore, and does. A deterministic correction applied to the reply before it is sent cannot be ignored.
- Close the vocabulary gaps deliberately. Every state an order or a booking can be in needs a real phrase written by a speaker, not one composed on the spot.
How GhostAgent does it
GhostAgent's Arabizi spellings are measured from a real Lebanese business's own messages rather than assumed, and the corrections are applied to every outgoing reply, not merely requested in the prompt.
Order and appointment states have written phrases rather than improvised ones, and any phrase without an attested Lebanese form keeps the English word inside an Arabizi sentence — which is what people genuinely write — rather than inventing a Levantine one.
It also handles English, Modern Standard Arabic, French and Spanish, and follows whichever language the customer switches into mid-conversation.
Common questions
- What is Arabizi?
- Arabic written in Latin letters with digits substituting for Arabic sounds — 3 for ع, 7 for ح, 2 for ء or ق, 5 for خ. It is also called Franco-Arabic, Franco or Arabish, and it is how a great many people in Lebanon and the Levant type every day.
- Can AI understand Arabizi?
- Understanding it is largely a solved problem. Writing it convincingly is not: without correction, a model produces the commonest forms it saw in training, which are often Egyptian or Gulf rather than Levantine, and invents words outright when it lacks the vocabulary for a situation.
- Does the bot switch between Arabizi and English?
- It follows the customer. Lebanese conversations move between Arabizi, English and Arabic within a single exchange, often within a single sentence, and a reply that insists on one language reads as a form rather than a person.