Arabic NLP
Why sentiment models break on Najdi — and what fixes them
Most tools translate first and classify second. In Gulf Arabic that loses the sarcasm, the code-switching and half the spellings.
Ask a general-purpose sentiment model what «الخدمة طايرة» means and it will tell you a service is flying. In Najdi it is a complaint. The model is not wrong about the words; it is wrong about the room.
Three ways translation-first fails
First, dialect. Gulf conversation is not Modern Standard Arabic with an accent; verbs, negation and intensifiers differ, and sarcasm rides on them. Second, script. The same brand is written مرقب, مِرقَب and Mirqab in the same thread, and a keyword match that knows one of them misses two thirds of the volume. Third, code-switching. A post that starts in Arabic and ends in English is normal in Riyadh and Dubai, and a model that classifies one half is guessing about the whole.
What fixes it
Models trained on the conversation itself, not on translations of it; variant matching that treats spellings, diacritics and dialect forms of a term as one; and analysts who can see why a post was classified the way it was, and correct it. The correction is the point — every fix feeds the next month’s model.
The test is simple: take last month’s hundred most-shared posts about you and read them next to their labels. If the sarcasm is filed under positive, the model is reading a different room.