How the dictionary is written

sa9si’s entries are written by language models, without human review. This page explains what is checked before a word is published, what is not, and why some words are missing.

Where the words come from

No word is invented. Every entry starts from a word attested in a public source. For Moroccan Darija, the list comes from the lexicon of DODa, the Darija Open Dataset, which also gives conjugations, feminine and plural forms, and real translated sentences. For Algerian Darja, which has no open equivalent, the words come from Wiktionary categories and from a word list published on Hugging Face, of which only the words themselves are used; a corpus of Algerian texts ranks them by frequency and keeps the Moroccan words that are also used in Algeria. For Tunisian Derja, the words come from Wiktionary, the lemmas of the TArC corpus, frequent words in a corpus of Tunisian texts, and Moroccan and Algerian words common in Tunisia; a corpus word is kept only if it appears at least thirty times, and a model first sets aside conjugated forms and forms carrying a pronoun, to keep only base words. Each entry’s sources are named at the foot of its page and described on the sources and licences page.

Three models, two independent opinions

Each entry is first written in French. A first model, Claude Opus, writes the full entry: part of speech, meanings, usage note and example. A second model, Claude Sonnet, receives the same word without seeing that work and only says whether it is a common word of the dialect and what it means. A third, Claude Haiku, compares the two. A word is published only if both writers recognise it as an ordinary word of the dialect and the judge finds their meanings agree. At the slightest disagreement the word is dropped rather than corrected by hand.

Proper nouns are also set aside, as are words one of the models assigns to another language or dialect, vulgar words, and entries whose usage note is too short to teach anything. That is why common words can be missing: their absence means the check did not succeed, not that the word does not exist.

The English translation

The English text is a translation of the French entry, never a new one. Claude Sonnet translates the headword, the meanings, the usage note and the example; Darija words and their spellings are kept as they are. A different model, Claude Haiku, then compares the English meanings with the French ones, position by position. If any meaning has changed, gone missing or moved, the word gets no English page. So the English dictionary has 2,876 of the 2,898 Moroccan words, 1,900 of the 1,919 Algerian words and 3,514 of the 3,539 Tunisian words published in French.

Examples

When DODa or, for Tunisian, TArC, holds a real sentence that illustrates the first meaning, that sentence is quoted, with a translation written by the model. Otherwise the model writes a short sentence. Either way, the judge checks that the sentence contains the word, uses it in one of the given meanings, and that the translation is faithful. A rejected example is removed and the entry appears without one. Real sentences are labelled as such.

Spellings

Latin-letter spellings offered by the models and the sources are kept only if the site’s Arabic keyboard links them back to the word in Arabic script. This check removes French or English words slipped into a list by mistake, and spellings of another word. Conjugations and plural forms reproduce DODa’s, in its spelling, which often leaves out short vowels.

The keyboard

The keyboard was measured on real pairs of words written in Latin and Arabic letters, taken from DODa for Moroccan and from the Algerian word list for Algerian, without giving it the answers. For a Moroccan word, the right Arabic spelling comes first in 53% of cases and within the first five in 76%; for an Algerian word, 46% and 70%; for a Tunisian word, measured on TArC’s pairs, 52% and 80%. That is why the keyboard shows its other suggestions: Arabizi has no fixed spelling and one Latin word can match several Arabic ones.

What to take from this

Two independent writers reduce errors without removing them, and a checked translation adds one more step where meaning can shift. A speaker remains the best reference, especially for figurative meanings, registers and regional variation within a country.