Skip to content

The page's dialect does not fit the country or the rest of the page

dialect-registerminor
On this page

Why it matters

  • Dialect is a brand's identity to a local reader. A Saudi store that writes «إزاي» and «عايز» sounds Egyptian, and an Egyptian store that writes «وايد» and «شلون» sounds Gulf; the visitor takes that as a sign the text was copied, or written for another market.
  • Colloquial Arabic in commercial writing is a choice of tone, not a mistake, but a tone that changes between the heading and the text, or two dialects on one page, looks unintended.
  • Search engines and AI assistants match the language of the question to the language of the page: someone searching in their dialect finds pages written in it first.

Example

Wrong
<p>شلونك يا غالي؟ وايد عندنا عطور حلوة، وإذا تبغى عود ممتاز الحين تواصل معنا ونوصله لك بسرعة.</p>
Right
<p>ازاي تختار العطر المناسب؟ عندنا عطور حلوة اوي، ولو عايز عود ممتاز دلوقتي كلمنا ونوصله لك بسرعة.</p>

How to fix

  • Decide one voice for the brand: plain Modern Standard suits every market, a local dialect suits one.
  • If you choose a dialect, choose the one of the country you serve, and use the same words in headings and text.
  • Review text copied from another country's site or from a translator, which may carry its dialect's words with it.

How we detect

  1. We read the visible text of an Arabic page (code tags left out) and separate the headings from the rest.
  2. We count words from a small list: words only speakers of one dialect write, such as إزاي, عايز and دلوقتي (Egyptian), وايد, شلون and ابغى (Gulf), هلق, شو and بدي (Levantine), بزاف, واش and ديال (Maghrebi). Words two dialects share are left out, and hamza forms and ta marbuta are folded before matching.
  3. With fewer than 80 Arabic words we give no verdict. Otherwise a dialect is called when at least 3 of its words are found and they are a fifth of all the telling words (dialect and Modern Standard); otherwise the text is Modern Standard.
  4. We compare the dialect with the country the page is written for, named by two different kinds of evidence at least (domain, language, currency, calling code); with no confirmed country we do not compare.
  5. Where the Arabic NLP service runs, we also read the text's sentences with CAMeL's dialect model: if the list found no dialect, the page has 80 words or more, and the model gives one variety 70% or more of the probability over the sentences and that variety wins at least half of them, the model's verdict is used. Without the service the list alone decides.
  6. The list is a sample, not a dictionary, and the model errs on short or mixed text. It is a minor finding.