A chatbot that understands darija: what it adds to the work

A shopkeeper in Bab El Oued receives about a hundred messages a day and wants to answer faster. He asks for an assistant "that understands darija", because that is what his customers write. The first question put to him is not which of the three languages he wants, it is to show us a hundred real messages.

That is where the subject becomes concrete. A modern artificial intelligence system copes with a great deal of darija; what it does not do is guess how your customers write the words that matter to you. The extra work is there, and it is measured in days of tuning rather than in a feature to buy.

This article describes what darija adds to the work, what it breaks when added badly, and the three cases where French alone is enough. It does not repeat what an assistant must refuse to answer, which is the prerequisite, and it does not discuss price: the reason is in section 5.

Written darija has no standard, and that is not a defect: it is a spoken language its speakers write by ear. So the same word arrives in several Latin spellings, several Arabic ones, with or without digits, with vowels written or not, split or joined.

For a system the consequence is not that it fails to understand: it is that it fails to match. Two customers asking for the same thing produce two different forms, the system recognises one and not the other, and the log of unanswered questions fills with one request written six ways.

The work is therefore to build a list of variants for the words that decide things in your business — products, commercial gestures, the words for availability and price — and then to set a matching threshold. It is not a translation list and must not be mistaken for one: it ties spellings to an intent, not a word to a word.

The surprising part is that the list cannot be guessed. Three people who speak darija perfectly will produce three different lists, and none will look like what your customers actually write. That is why the previous section comes before this one.

Understanding darija and replying in it are two separate decisions, and the second is not technical. A reply written in Latin darija reads as close and familiar to some of your customers and as careless to others — the same sentence, read by an institutional client or a supplier, says the business writes the way it speaks.

The position we advise is asymmetric and easy to hold: understand all three languages, reply in French or in standard Arabic according to the message’s dominant language, in a short and concrete register. The customer does not need to read you in darija; they need to be understood when they write it.

There is a clear exception, and it is the transcribed voice note or the very informal exchange on a messaging app where your brand is already familiar. There, replying in the customer’s register is coherent — and the channel already imposes constraints of its own before that one. It is a decision taken once, written down, and applied everywhere — not a per-channel setting.

What must never happen is the system choosing its own register message by message. The variation is visible to a customer who writes twice, and it gives the impression of talking to two different businesses.

Commercial vocabulary moves: a product launches, a promotion creates a phrase, an expression spreads, a neighbourhood acquires a nickname. A frozen list of variants loses reach every season without anything saying so — the log fills a little more, and nobody makes the connection.

The useful rhythm is quarterly and the work is short: reread the unanswered questions since last time, take the new forms from them, add them, replay the test set. An hour or two, provided the log exists and somebody owns it.

That ownership has to be named inside the business, and it is the part most readily and most badly delegated. A supplier can hold the mechanism; they will not know that a new expression has come to mean your product among your customers, because they are your customers.

A business that cannot name anybody for that quarterly hour should stay with the three cases in section 7 and handle its messages by hand. That is an acceptable answer and it beats a system that degrades.

We read your hundred messages with you, take the variant list and the test set from them, set the matching threshold on your data, and come back after two weeks of real use. We leave the list and the test set with you, in a format somebody else can pick up.

We refuse to announce a comprehension rate. We have no annotated corpus and no comparison between models, and a percentage offered without them would measure something other than what it claims. The only honest figure is the one your test set produces on your messages, and it belongs to you.

We also refuse to deliver an automatic reply in darija on a channel where your brand speaks otherwise, even on request: the shift in register shows within two messages and costs more than the closeness gains.

What you can do without us is the first half of the work: the three columns in section 11, on a hundred messages, in one afternoon. Many businesses discover there that their difficulty is the catalogue rather than the language, and what the list of what an assistant must refuse says will help them more than darija.