Elamur: **Mova** · [Parse](https://mova.elamur.ai/free/en/parse/) · [Dictionary](https://md.dictionary.elamur.ai/free/en/) · [Translate](https://md.translate.elamur.ai/free/en/) · [Research](https://md.research.elamur.ai/free/en/) · [Groupants](https://md.groupants.elamur.ai/free/en/)

[Mova](https://md.mova.elamur.ai/free/en/) · [this page in HTML](https://mova.elamur.ai/free/en/docs/languages/fr/08-national-varieties/)

[home](https://md.mova.elamur.ai/free/en/) / [docs](https://md.mova.elamur.ai/free/en/docs/) / [languages](https://mova.elamur.ai/free/en/docs/languages/) / [fr](https://md.mova.elamur.ai/free/en/docs/languages/fr/) / 08. National and regional varieties of French

# 08. National and regional varieties of French

**On this page**

1. [1. What counts as a variety, and what does not](#1-what-counts-as-a-variety-and-what-does-not)
2. [2. Metropolitan France](#2-metropolitan-france)
3. [3. Belgium](#3-belgium)
4. [4. Switzerland](#4-switzerland)
5. [5. Luxembourg, Monaco, the Aosta Valley, the Channel Islands](#5-luxembourg-monaco-the-aosta-valley-the-channel-islands)
6. [6. North America](#6-north-america)
7. [7. The Caribbean, French Guiana and the creoles](#7-the-caribbean-french-guiana-and-the-creoles)
8. [8. Africa, the Indian Ocean and the Pacific](#8-africa-the-indian-ocean-and-the-pacific)
9. [9. Varieties in the data Mova holds](#9-varieties-in-the-data-mova-holds)
10. [10. Cues for a variety label](#10-cues-for-a-variety-label)
11. [11. How Mova handles national varieties](#11-how-mova-handles-national-varieties)
12. [Sources for this chapter](#sources-for-this-chapter)

French is the first or usual language of people in Europe, North America, the Caribbean and Africa, and the language of school, law and administration in many more places. The written standard is nearly the same everywhere; the differences between countries lie in the sounds, in the vocabulary and in a few points of grammar. This chapter describes those differences country by country (§§2–8), says what the treebanks Mova holds contain of them (§9), and gives, in §10, the **cue table** from which a variety label (`fr-FR`, `fr-BE`, `fr-CH`, `fr-CA`, `fr-x-africa`) can be learned, in the way `tools/langseg` learns the Hebrew and Aramaic varieties from the cues of their manuals. The register differences inside one variety (spoken against written, familiar against formal) belong to the chapter on register ([37. Register and variation](https://md.mova.elamur.ai/free/en/docs/languages/fr/37-register-and-variation/)) and are touched here only where they decide how a variety is seen.

Three kinds of statement are kept apart. What the public-domain phoneticians and grammarians of about 1900 observed (Passy 1906, Nyrop 1899–1902, Martinon 1913, Dauzat 1908) is attributed to them and dated. What recent open research has measured is given with its finding. Lexical labels are checked against the regional labels of the French Wiktionary (raw dump of 2026-09-01, CC BY-SA: used as evidence and for tests, never copied); a word whose label we could not check in any source we hold is marked "(general description)". Treebank counts are dated 2026-11-06.

## 1. What counts as a variety, and what does not

### 1.1 The standard and its norms

French has an old tradition of a single norm: the "good usage" of the seventeenth-century court, which Vaugelas defined as the way of speaking of the soundest part of the Court in conformity with the way of writing of the soundest part of the authors of the time (vaugelas-1880; [09](https://md.mova.elamur.ai/free/en/docs/languages/fr/09-old-middle-classical-french/) §6). The tradition left a habit of seeing everything else as a deviation, and a well-documented **linguistic insecurity** among speakers of non-Parisian varieties, which Dankova (dankova-2024) recalls for Quebec. The written norm of formal Quebec, Belgian, Swiss and African French is close to the "international" written French of France, with regional marks that are few and mostly lexical; the press of the countries is therefore hard to tell apart by machine (§10.1).

The countries have their own language bodies (described in [00. Introduction: what French is, and how this manual writes it](https://md.mova.elamur.ai/free/en/docs/languages/fr/00-introduction/) §4): the Office québécois de la langue française and its *Banque de dépannage linguistique* (oqlf-2026, card only), the Service de la langue française of the Fédération Wallonie-Bruxelles (moreau-2014, card only), the Swiss federal Chancellery (suisse-2000, card only). Their recommendations are known to us from the cards and are not quoted.

### 1.2 Other languages of France

A speaker of Picard, Poitevin or Alsatian speaks a language that is not French. Dauzat (dauzat-1908, dauzat-1922) places the difference between a "dialect" and a "patois" at the social, not the linguistic, level. The *langues d'oïl* besides French are Picard, Walloon, Norman, Gallo, Poitevin-Saintongeais, Burgundian, Lorrain and Champenois; beside them Occitan, Franco-Provençal, Breton, Basque, Catalan, Corsican and the Germanic dialects of Alsace and Moselle are spoken in France. The COLaF project (sagot-2025) plans corpora and tools for the languages of France, the overseas languages and the immigrant languages, and for French "in all its geographic diversity"; it starts from the observation that speech technology loses accuracy as soon as the speaker moves away from an ideal standard.

The one treebank of such a language that Mova holds is UD French PoitevinDIVITAL: 239 sentences, 5 484 words, test file only, CC BY-SA, with a French gloss in the `MISC` column and the metadata `language_variety = Poitevin (poit1240)`:

> **contesmarais46.6** · UD French PoitevinDIVITAL *Le sant trtouts içhi.* "they are all here" (the treebank's translation: *ils sont tous ici*)

It is not French and is not treated as a variety of French. It does show what a French parser does on a related *langue d'oïl*: Mova's parser trained on GSD reaches LAS **24.85** on it with its own tags and **59.01** with the treebank's gold tags (`en parse-eval`, GSD train → PoitevinDIVITAL test, after `en xpos-fill`, 2026-11-06; the same run gives 87.00 / 89.78 on GSD dev). The tagger, which has never seen *sant* or *içhi*, is where most of the loss lies.

### 1.3 Creoles

The French-based creoles (Haitian, Guadeloupean, Martinican, Guianese, Reunionese, Mauritian, Seychellois) are languages of their own with their own grammar, not dialects of French (§7).

## 2. Metropolitan France

### 2.1 Where the standard comes from

The standard pronunciation is the educated pronunciation of the Paris region; the pronunciations of Lyon, Toulouse, Marseille, Lille or Strasbourg differ. In the descriptions of 1906–1913 the "North" is the Paris norm and the "South" (*le Midi*) the main departure from it. The regional accents of France are accents of one language, in which the regional language beneath (Occitan, Alsatian, Picard, Breton) has left a mark.

### 2.2 The Midi

The features below are the ones Passy, Nyrop and Martinon describe, and they are still the features by which a southern accent is recognised:

- **The *e* of spelling is pronounced** at the end of words and inside them: *une table* [ynə tablə], *la porte* [la pɔʁtə]. Martinon notes that the mute *e*, heard almost nowhere else, is heard "naturally" in the Midi (martinon-1913).
- **Nasal vowels** are weakly nasal and followed by a nasal consonant, velar at the end of a word: *pain* [pɛ̃ŋ], *lampe* [lɑ̃mpə] (general description; see [02. Vowels](https://md.mova.elamur.ai/free/en/docs/languages/fr/02-vowels/) §6.5).
- **The law of position is applied strictly**: a close-mid vowel in an open syllable, an open-mid vowel in a closed one, so that *fée* and *fait* are both [fe] and *rose* is [ʁɔzə] (general description; [02. Vowels](https://md.mova.elamur.ai/free/en/docs/languages/fr/02-vowels/) §4).
- **The palatal lateral** [ʎ] of *soleil, œil, briller* was still heard in the Midi and in Switzerland in Passy's day, where the north already had /j/; northerners who tried to imitate it said [lj] and confused *souiller* with *soulier* (passy-1906, §§195, 233).
- **Words** of regional use, such as *chocolatine* for *pain au chocolat* in the Southwest (general description).

### 2.3 Other regions

- **The North (Nord, Pas-de-Calais)**: Picard words under the French (*chicon* for *endive*, *drache* "downpour", both labelled "Nord de la France" as well as "Belgique" in the Wiktionary); the treebanks have no text of this kind.
- **Alsace and Lorraine**: the accent and some constructions carry the Germanic substrate; Dauzat counts the Alsatian dialects among the languages of France outside the Romance ones (dauzat-1908). The Wiktionary labels *drache* and *cornet* "bag" Lorraine too.
- **Paris and the *banlieue***: the speech of the *cités*, with borrowings from Arabic, Romani and African languages and with *verlan*, belongs to register and generation, not to a region ([37. Register and variation](https://md.mova.elamur.ai/free/en/docs/languages/fr/37-register-and-variation/)).

### 2.4 Mergers and distinctions by region

| distinction | educated Paris | elsewhere |
| --- | --- | --- |
| /a/–/ɑ/ (*patte–pâte*) | merged for most speakers | kept in Belgium, Switzerland, Québec |
| /ɛ̃/–/œ̃/ (*brin–brun*) | merged | kept in Belgium, Switzerland, the Midi, Québec |
| /e/–/ɛ/ in a final open syllable (*fée–fait*) | kept by many, weakening | lost in the Midi (one mid vowel by position) |
| /ɛ/–/ɛː/ (*mettre–maître*) | lost | kept in Belgium, Switzerland, Québec |
| long final vowel of a feminine (*amie* vs *ami*) | lost | kept in Belgium and Switzerland (passy-1906, §114; kahane-2025 for Belgium) |
| /ʎ/ (*soleil*) | /j/ | kept in the Midi and Switzerland in 1906 (passy-1906, §195); now rare |
| final /ə/ | silent | pronounced in the Midi |

Passy writes that in Switzerland and Belgium the vowels formerly followed by an *e* stayed long, so that *venue, amie, boue, fermée* are distinct from *venu, ami, bout, fermé*, and that Parisian declamation often does the same (passy-1906, §114). A century later the spoken treebank ParisStories, which annotates what can be heard, marks 1 006 words `HasSpokenGender=NotInThisDialect` (*vrai* / *vraie*, *compliqué* / *compliquée*): their gender is not audible in the Paris speech of the corpus, though it is in a variety like Belgian French, where the feminine vowel is lengthened (kahane-2025; [04. Connected speech: enchaînement, liaison, elision, the mute e and assimilation](https://md.mova.elamur.ai/free/en/docs/languages/fr/04-connected-speech/) §9).

## 3. Belgium

French (in Wallonia and Brussels) is one of the three official languages. The Belgian variety has a well-identified vocabulary, a few sounds and a few constructions.

- **Numerals**: *septante* (70) and *nonante* (90) for *soixante-dix* and *quatre-vingt-dix*; *quatre-vingts* (80) as in France. The words are old French: Nyrop traces *septante* to Gallo-Roman *settanta* and notes that it was replaced in the literary language after the fifteenth century by *soixante-dix*; he quotes Ménage's ruling that one says *soixante-dix, quatre-vingts, quatre-vingt-dix* in ordinary speech, and Littré's regret that *nonante* "has aged" while staying in use in Switzerland, Savoy and the Midi (nyrop-1899, vol. II, numerals). The Wiktionary labels both *septante* and *nonante* Belgium, Switzerland, Acadia and Congo-Kinshasa.
- **Meals**: *déjeuner* (breakfast), *dîner* (lunch), *souper* (dinner), against the *petit-déjeuner, déjeuner, dîner* of France (general description; the same in Switzerland and Québec).
- **Vocabulary** (each labelled "Belgique" in the Wiktionary): *GSM* "mobile phone", *kot* "student room" (from Flemish), *drache* "downpour", *jobiste* "student with a job", *farde* "folder", *essuie* "towel", *chicon* "endive", *pistolet* "bread roll", *endéans* "within (a time)", *tantôt* "this afternoon, later today", and *savoir* in the sense "be able to" (% *Je ne sais pas venir* "I cannot come").
- **Phrases** (general description): *une fois* as a softening particle (% *Viens une fois ici*), *s'il vous plaît* when handing something over ("here you are"), *à tantôt*.
- **Sounds**: /ɛ̃/–/œ̃/, /a/–/ɑ/ and the long vowels are kept (§2.4); the feminine vowel is lengthened (*amie* [amiː]); final voiced obstruents may be devoiced in Brussels, a Dutch-contact feature, and the *w* of *wagon* is /w/ where France has /v/ (general description).

The treebank record is thin: *septante* never occurs in the treebanks we hold, *nonante* once (in ALTS, a sixteenth-century Norman text), *kot* once in GSD (2026-11-06).

## 4. Switzerland

French (*le romand*) is spoken in the west of the country: Geneva, Vaud, Neuchâtel, Jura and the French-speaking parts of Fribourg, Valais and Bern.

- **Numerals**: *septante* (70) and *nonante* (90) everywhere; *huitante* (80) in Vaud, Valais and Fribourg (the Wiktionary labels *huitante* "Suisse" only), *quatre-vingts* in Geneva, Neuchâtel and Jura (general description for the cantonal split). *Huitante* continues Old French *(h)uitante*; Nyrop says its learned doublet *octante* was long "relegated to the patois" (nyrop-1899).
- **Vocabulary** (labelled "Suisse" in the Wiktionary unless marked): *votation* "vote in a referendum", *panosse* "floor cloth", *s'encoubler* "to trip", *cornet* "(plastic) bag", *natel* "mobile phone", and *poutser* "to clean" (general description); *déjeuner, dîner, souper* as in Belgium.
- **Sounds**: Passy observed long final vowels (*amie* [amiː]) and the palatal [ʎ] in Switzerland (passy-1906, §§114, 195), that *fille* and *ail* end there in a lateral rather than in /j/, and so do not lengthen their vowel (§115), and "infinitely more liaisons" in the Suisse romande than around Paris, which made Swiss speech sound "more or less affected" to a Parisian (passy-1906, liaison chapter). His book prints a sample of Swiss pronunciation beside the Parisian one.
- **Typography**: the Swiss house rule usually sets `; ! ?` without a space before them ([07](https://md.mova.elamur.ai/free/en/docs/languages/fr/07-punctuation-typography/) §2); the treebanks normalise spacing, so the cue is lost there.

## 5. Luxembourg, Monaco, the Aosta Valley, the Channel Islands

- **Luxembourg**: French is one of the three official languages beside Luxembourgish and German, the language of law and administration. The Wiktionary labels *GSM*, *farde* and *endéans* Luxembourg as well as Belgium.
- **Monaco**: French is the official language.
- **The Aosta Valley** (Italy): French is co-official; the local speech is Franco-Provençal.
- **The Channel Islands**: the Norman dialects of Jersey and Guernsey. Sixteenth-century Norman legal French from Guernsey is in the UD treebank ALTS ([09](https://md.mova.elamur.ai/free/en/docs/languages/fr/09-old-middle-classical-french/)).

## 6. North America

### 6.1 Québec

Québec French is the largest variety outside Europe and the only one with a normative body of its own; it keeps close to the French of France in writing and differs in speech.

**Written journalism.** Dankova (dankova-2024) compares the press of France (*Le Monde, Le Figaro, France-Soir*) with that of Québec (*Le Devoir, La Presse*) and finds the French texts richer in adjectives, in different adjectives and in participles and gerunds, with verbs of perception and cognition 2.5 times more frequent (360 occurrences against 143); she concludes that the Quebec texts are lexically less rich while sharing the same grammar and syntax. It is a finding about written journalism, not about speech.

**Sounds.**

- /t/ and /d/ are **affricated** before /i y/ and the glides /j ɥ/: *petit* [pətsi], *dur* [dzyʁ], *tu* [tsy] (general description; [01. Consonants](https://md.mova.elamur.ai/free/en/docs/languages/fr/01-consonants/) §2.1).
- The close vowels /i y u/ have **lax** allophones [ɪ ʏ ʊ] in a syllable closed by a non-lengthening consonant (*vite* [vɪt], *jupe* [ʒʏp], *route* [ʁʊt]). Lancien's pilot study (lancien-2020a) asks whether laxing also occurs in clipped words that end in a lengthening consonant (/v z ʒ ʁ/); with two speakers reading short texts, the high vowels were shorter there than in matching minimal pairs, which the author treats as a preliminary indication.
- **Vowel length** is kept: /ɑ/ is distinct from /a/, the long /ɛː/ of *fête, père* is long or diphthongised (*père* [paɛ̯ʁ]), and vowels are long before the lengthening consonants /v z ʒ ʁ/. Riverin-Coutlée (coutlee-2020) measured absolutely final fricatives: voiced /v z ʒ/ are shorter than voiceless /f s ʃ/, voiced for a smaller share of their duration than older reports say, and they **lengthen the preceding stressed vowel even when it is intrinsically long**.
- **Style**: Lancien (lancien-2020b) recorded two couples and found shorter, more central vowels when the speakers talked to their partner than to a stranger.
- **Prominence**: Lamontagne and Goad (lamontagne-2022), on corpus speech from Saguenay, find that phrasal prominence is drawn away from the final syllable to a heavy one (a closed syllable or a long vowel), above all when the final syllable is open; they distinguish a lexical prominence marked by duration and amplitude from a phrasal one marked by pitch. The Paris-based account of [03](https://md.mova.elamur.ai/free/en/docs/languages/fr/03-syllable-stress-intonation/) §2, prominence on the last syllable of the group, is thus not the whole story for this variety.
- The nasal /ɑ̃/ is fronted towards [ã] and /ɛ̃/ raised towards [ẽ] (general description; [02. Vowels](https://md.mova.elamur.ai/free/en/docs/languages/fr/02-vowels/) §6.5).

**Grammar and function words** (casual speech; general description unless a source is named):

- The **particle *-tu*** marks a yes/no question with any subject: % *Tu viens-tu ?*, % *C'est-tu prêt ?* ([26. Questions](https://md.mova.elamur.ai/free/en/docs/languages/fr/26-questions/)).
- *Être après* + infinitive for the progressive: % *Je suis après manger* "I am eating".
- *Icitte* (*ici*), *moé, toé* (*moi, toi*), *pis* (*puis*, "and then, and"); the 1930 *Glossaire* already has *icite* "here" and *moé* (canada-1930).
- *Pantoute* "not at all" (canada-1930: *J'veux pas pantoute*).
- A *pas* inside some superlative and universal clauses, as in % *C'est le plus beau blé qu'il y a pas sur le marché*. Labelle (labelle-2023) shows that this *pas* is not expletive: it is an ordinary negation over a silent comparative ("than which there is not [a more beautiful one]") and widens the domain of the superlative or universal.
- Some nouns have a different gender in casual use (*une job*), and *ils* is often used for *elles* (general description).

**Vocabulary.** The official terms that Québec coins for English loans (*courriel* "e-mail", *clavarder* "to chat", *magasiner* "to shop", *fin de semaine* "weekend", *stationnement* "car park", *cellulaire* "mobile phone") and the old words of the country (*char* "car", *chum* "close friend, boyfriend", *blonde* "girlfriend", *tuque* "wool cap", *poudrerie* "blowing snow", *achaler* "to annoy", *dépanneur* "corner shop", *breuvage* "drink"). The *Glossaire du parler français au Canada* (canada-1930), public domain, records *magasiner*, *chum*, *blonde* ("young woman courted, sweetheart"), *achaler*, *poudrerie*, *pantoute*, and gives *char* the meaning "railway carriage" (*embarquer dans les chars*). Dauzat already notes that Canadians say *un char* where France takes the English *wagon* (dauzat-1908). The Wiktionary labels *chum, magasiner, clavarder, pantoute, poudrerie* "Québec" and *blonde, breuvage, char, tuque* "Canada". *Courriel* has since been adopted in France as the official term, so it is now a weaker cue than it was. Swearing uses the *sacres*, words from the Catholic liturgy (*tabarnak, câlice, ostie*) (general description).

**Meals**: *déjeuner* breakfast, *dîner* lunch, *souper* dinner. **Thanks**: *bienvenue* "you're welcome" (general description).

**The norm as data.** Two recent corpora take the OQLF norm as their reference: QFrCoLA (beauchemin-2025), sentences marked acceptable or not by the norm, and QFrBLiMP (beauchemin-2026), 1 761 minimal pairs for 20 grammatical phenomena, judged by twelve native Quebec speakers. The authors find that fine-tuned models do better than zero-shot large models, that all models fail on phenomena that need deep semantic understanding, and that on Quebec French most models score lower than on the European-French equivalent, the strongest ones within the statistical margin (beauchemin-2026). A third study uses idioms to measure the dialect gap of language models (beauchemin-2026b). All three are CC BY 4.0 and could serve Mova as tests (§11).

### 6.2 Acadia and the Maritimes

Acadian French (New Brunswick, Nova Scotia, Prince Edward Island) keeps older features: *je* with a plural verb (% *j'avons, je parlons*), the ending *-ont* in the third person plural (% *ils parlont*), older words (*asteur* "now", from *à cette heure*), and the *passé simple* still in speech in some communities (general description). The Wiktionary labels *achaler* and *septante, nonante* Acadia as well.

### 6.3 Ontario, the West, Louisiana and New England

Franco-Ontarian, Franco-Manitoban and other Western Canadian French share the Québec features in different proportions, with heavy contact with English. Louisiana French (Cajun) and the French of New England are in decline; the Wiktionary labels *blonde, char, icitte* Louisiana as well as Canada or Haiti. No treebank in our library covers them.

## 7. The Caribbean, French Guiana and the creoles

In Guadeloupe, Martinique, French Guiana, Saint-Martin and Saint-Barthélemy French is the official language beside a French-based creole. The French spoken is close to the European standard, with the accent of the islands and local words.

**Haiti** has two official languages, French and Haitian Creole. Haitian Creole is a separate language whose lexicon is mostly French: *mwen pa konnen* "I don't know", *li te manje* "he ate", with the tense markers *te* (past), *ap* (progressive), *pral* (future) before an uninflected verb and the determiner after the noun (*kay la* "the house") (general description). Research on the creoles matters to Mova because it shows how much a parser can transfer from French:

- Mompelat, Dakota and Kübler (mompelat-2022) built a parser for **Martinican Creole** from 80 sentences of the creole plus UD French GSD. A simple concatenation of the two treebanks was best: LAS 71.77 with POS embeddings on the Martinican development set, against 51.95 for the French treebank alone and 62.86 for the 80 creole sentences alone. The lexifier helps, but cannot do the job alone.
- Havard and colleagues (havard-2024) trained a speech model for Haitian from the 499 field recordings of the *Atlas linguistique d'Haïti*; the best model was a French-pretrained speech model with continued pre-training on Haitian.
- Mompelat and colleagues (mompelat-2025) describe translation systems between French and Martinican and Guianese Creole for WMT 2025.

A creole is **not a variety of French** for Mova: it gets its own package and its own language code, with this manual as the lexifier reference.

## 8. Africa, the Indian Ocean and the Pacific

### 8.1 The Maghreb

French is a second language and a language of administration, higher education and business in Morocco, Algeria and Tunisia. Speech mixes Maghrebi Arabic, Berber and French, often within one sentence. Toughrai and colleagues (toughrai-2025) show the difficulty for tools: North African Arabic is written both in Arabic and in Latin letters, and a phoneme-based normalisation that maps both scripts, and French, into one simple Latin representation lets models trained only on standard Arabic and French work on dialect tasks. A French segment inside such text is a language switch (`fr` inside `ar`), not a variety of French; the label `fr-x-africa` of §10 is not meant for it.

### 8.2 West and Central Africa

French is the language of school, administration and the city in Senegal, Mali, Burkina Faso, Côte d'Ivoire, Cameroon, the two Congos, Gabon and elsewhere, and in some cities (Abidjan, Libreville, Kinshasa) the usual language of daily life. The reference description, the *Inventaire des particularités lexicales du français en Afrique noire* (ifa-1983), is a card only; the facts below are checked against the Wiktionary labels where it has them.

- **Vocabulary**: *essencerie* "petrol station" (labelled "Afrique"), *maquis* "open-air eating place" (Côte d'Ivoire, Burkina Faso), *gâter* "to break (a device)" (Côte d'Ivoire), *go* "girl, girlfriend" (Côte d'Ivoire, Cameroon), and *cadeauter* "to give as a present" and *enceinter* "to make pregnant" (both labelled "Afrique"), and the widely cited *deuxième bureau* "mistress" (general description). The Wiktionary labels *septante, nonante, drache* and *farde* Congo-Kinshasa too: the French of the DR Congo shares Belgian words.
- **Grammar** (general description): *pour* as a possessive (% *C'est pour moi* "it's mine"), *-là* after the noun as a familiar definite marker (% *la femme-là*), calques from the national languages.
- **Sounds** (general description): a tapped or trilled /r/, fewer nasal vowels in some regions, and the vowels and tones of the substrate language.
- ***Nouchi*** in Côte d'Ivoire is a youth argot mixing French and local languages.

Ngom and colleagues (ngom-2025) extract place names from 15 000 French news articles on Burkina Faso with CamemBERT and a zero-shot model; it shows what open text there is for the French of West Africa: mostly press, which is the variety of Africa closest to the written standard. Ouedraogo and colleagues (ouedraogo-2026) work on speech-to-speech translation between French and Mooré (Burkina Faso), the kind of pair in which the French side is the local variety.

### 8.3 The Indian Ocean and the Pacific

French is official in Madagascar (with Malagasy), the Comoros and the Seychelles, and is the language of the press beside English in Mauritius; Réunion and Mayotte are French territories (the Wiktionary labels *GSM* Réunion too). In Réunion, Mauritius and the Seychelles French coexists with a French-based creole. French is the language of New Caledonia (beside the Kanak languages), French Polynesia and Wallis and Futuna; their languages are among the aims of COLaF (sagot-2025).

### 8.4 Elsewhere

French has an old presence in Lebanon, in Vietnam, Laos and Cambodia, and in Egypt, as a school and cultural language. These varieties have not been studied on our corpora.

## 9. Varieties in the data Mova holds

### 9.1 The treebanks

| treebank (UD 2.18) | variety | genre | licence |
| --- | --- | --- | --- |
| GSD | France (written) | blogs, news, reviews, Wikipedia | CC BY-SA 4.0 |
| Sequoia | France | Europarl, regional press, Wikipedia, medicine | LGPL-LR |
| FQB | France | questions | LGPL-LR |
| PUD | translated from English | news, Wikipedia | CC BY-SA 4.0 |
| ParisStories | spoken, Paris | stories told | CC BY-SA 4.0 |
| Rhapsodie | spoken, France | several genres, prosody | CC BY-SA 4.0 |
| ALTS | sixteenth-century Normandy and Guernsey | legal | CC BY-SA 4.0 |
| ParTUT | France (and Italy) | legal, web | CC BY-NC-SA 4.0: counts only |
| PoitevinDIVITAL | Poitevin-Saintongeais (not French) | tales | CC BY-SA 4.0 |

There is no Belgian, Swiss, Canadian or African French treebank, and no creole treebank.

### 9.2 What the data shows of the variety markers

Counted over the nine French treebanks, all files, lower-cased word forms (2026-11-06):

| marker | count |
| --- | --- |
| *septante, huitante, octante* | 0 |
| *nonante* | 1 (ALTS, sixteenth century) |
| *soixante-dix* / *quatre-vingt-dix* as one token | 5 / 2 (GSD, Rhapsodie, PUD) |
| *GSM, natel, courriel, magasiner, chum, tuque, pantoute, essencerie* | 0 |
| *clavardage* | 2 (PUD) |
| *votation* | 2 (GSD) |
| *kot* | 1 (GSD) |
| *dépanneur* | 1 (Sequoia) |
| *cellulaire* / *portable* | 3 / 10 in GSD (most *cellulaire* are the biological adjective) |
| *stationnement* / *parking* | 4 / 11 in GSD, 0 / 2 in Sequoia, 0 / 1 in ParisStories |
| *week-end* / *weekend* | 15 / 4 in GSD; *fin de semaine* is not separable by form from other *fin* |
| *petit-déjeuner* / *déjeuner* / *dîner* / *souper* | GSD 2 / 6 / 5 / 2 |
| *tantôt* | GSD 4, Sequoia 2 (the standard sense "now… now" as well as the Belgian one) |

The treebanks are the French of France, with a few Swiss, Belgian and Quebec words that enter through the press (*votation* of Swiss news, *clavardage* in a translation). A tool trained on them treats *septante* and *nonante* as unknown words and does not know that they are numerals.

## 10. Cues for a variety label

`tools/langseg` cuts a text into language segments and decides between close varieties with **cues named in the manuals**, whose weights it learns from a labelled pool (`tools/langseg/src/cues.rs`: each cue is a form, a short name and the variety the manual expects it to point to; `langseg cues` then shows which variety each cue was actually learned to favour, so a cue the manual over-rates is visible). Today langseg knows English, transliteration and the Hebrew-script varieties; it has no French. The table below is the French cue list in the same spirit, for the sub-labels `fr-FR`, `fr-BE`, `fr-CH`, `fr-CA` and `fr-x-africa` (Africa south of the Sahara; Maghreb French is better treated as a switch inside Arabic text, §8.1).

### 10.1 What to expect

Two open studies set the expectations low. In the French Cross-Domain Dialect Identification task of VarDial 2022, news paragraphs from France, Belgium, Switzerland and Canada, with every named entity replaced by `$NE$` to remove country bias, were to be told apart across domains; the best submitted system, a CamemBERT baseline, reached a macro F1 of 0.40, and the SUKI team's adaptive naive Bayes on character 8-grams 0.27, with Belgian, Swiss and French texts largely confused (per-variety F1 of its best run: Belgium 43.1, Switzerland 51.9, France 8.2, Canada 3.3) (jauhiainen-2022). Choudhry (choudhry-2022) collected web text by national top-level domain and found that the decision tree that classified it best relied on names of cities and countries for more than half of its nodes, a cue to the country of origin rather than to the dialect. Edited written French has few variety marks per paragraph; a label should come from the cues below when they fire, and be "unknown" otherwise.

### 10.2 The cue table

The **strength** column is the manual's expectation: *strong* = rare outside the named varieties; *medium* = used elsewhere but much more often there; *weak* = a tendency only, or lost in normalised text. The weights themselves must be learned, as in langseg. "W" = labelled in the French Wiktionary (2026-09-01 dump); "GD" = general description, no source in our library; the treebank column gives the count in all nine French treebanks (2026-11-06).

**Numerals and dates**

| cue | name | points to | strength | evidence | treebanks |
| --- | --- | --- | --- | --- | --- |
| *septante* | `num_70` | fr-BE, fr-CH (also Acadia, DR Congo) | strong | W; nyrop-1899 | 0 |
| *nonante* | `num_90` | fr-BE, fr-CH (also Acadia, DR Congo) | strong | W; nyrop-1899 (Littré) | 1 (ALTS) |
| *huitante* | `num_80h` | fr-CH (Vaud, Valais, Fribourg) | strong | W | 0 |
| *septante* with *quatre-vingts* in one text | `num_70_80` | fr-BE (or Geneva, Neuchâtel, Jura) | medium | GD | – |
| *soixante-dix, quatre-vingt-dix* | `num_70_fr` | fr-FR, fr-CA, fr-x-africa | medium (also BE/CH in formal texts) | GD | 7 |

**Lexicon: Belgium**

| cue | name | points to | strength | evidence | treebanks |
| --- | --- | --- | --- | --- | --- |
| *GSM* | `be_gsm` | fr-BE (also Luxembourg) | strong | W | 0 |
| *kot, koter, kotteur* | `be_kot` | fr-BE | strong | W | 1 |
| *drache, dracher* | `be_drache` | fr-BE (also Nord, Lorraine, DR Congo) | strong | W | 0 |
| *jobiste* | `be_jobiste` | fr-BE | strong | W | 0 |
| *farde* | `be_farde` | fr-BE (also Luxembourg, DR Congo) | strong | W | 0 |
| *essuie* (noun) | `be_essuie` | fr-BE | medium | W | 0 |
| *chicon* | `be_chicon` | fr-BE (also Nord) | medium | W | 0 |
| *endéans* | `be_endeans` | fr-BE (also Luxembourg) | strong | W | 0 |
| *savoir* + infinitive meaning "can" | `be_savoir` | fr-BE | medium (needs the parse: *je ne sais pas venir*) | W | – |
| *à tantôt* | `be_tantot` | fr-BE | medium (*tantôt… tantôt* is standard) | W (*tantôt*) | 6 (*tantôt*) |
| *une fois* after an imperative | `be_unefois` | fr-BE | weak | GD | – |

**Lexicon: Switzerland**

| cue | name | points to | strength | evidence | treebanks |
| --- | --- | --- | --- | --- | --- |
| *votation* | `ch_votation` | fr-CH | strong | W | 2 |
| *panosse* | `ch_panosse` | fr-CH | strong | W | 0 |
| *(s')encoubler* | `ch_encoubler` | fr-CH | strong | W | 0 |
| *natel* | `ch_natel` | fr-CH | strong | W | 0 |
| *poutser* | `ch_poutser` | fr-CH | strong | GD | 0 |
| *cornet* "bag" | `ch_cornet` | fr-CH (also Lorraine) | weak (*cornet* is also standard "cone") | W | – |

**Lexicon: Canada (Québec, Acadia)**

| cue | name | points to | strength | evidence | treebanks |
| --- | --- | --- | --- | --- | --- |
| *magasiner, magasinage* | `ca_magasiner` | fr-CA | strong | W; canada-1930 | 0 |
| *clavarder, clavardage* | `ca_clavarder` | fr-CA | strong | W | 2 (PUD) |
| *pantoute* | `ca_pantoute` | fr-CA | strong | W; canada-1930 | 0 |
| *chum* | `ca_chum` | fr-CA | strong | W; canada-1930 | 0 |
| *achaler* | `ca_achaler` | fr-CA (also Acadia) | strong | W; canada-1930 | 0 |
| *poudrerie* "blowing snow" | `ca_poudrerie` | fr-CA | medium (also "powder works") | W; canada-1930 | 0 |
| *tuque* | `ca_tuque` | fr-CA | strong | W | 0 |
| *dépanneur* "corner shop" | `ca_depanneur` | fr-CA | medium (also "repairman" everywhere) | GD (W labels another sense "Suisse") | 1 |
| *breuvage* "drink" | `ca_breuvage` | fr-CA | medium | W | 0 |
| *char* "car", *blonde* "girlfriend" | `ca_char` | fr-CA | weak (standard words in other senses) | W; canada-1930 | – |
| *courriel* | `ca_courriel` | fr-CA | weak (now official in France) | W | 0 |
| *cellulaire* "mobile phone" | `ca_cell` | fr-CA | weak (also the biological adjective) | GD | 3 |
| *fin de semaine* "weekend" | `ca_findesemaine` | fr-CA | medium (multiword) | GD | – |
| *stationnement* "car park" | `ca_stationnement` | fr-CA | weak (also standard) | GD | 4 |
| *icitte, moé, toé* | `ca_icitte` | fr-CA (casual; also Louisiana, Haiti) | strong | W; canada-1930 | 0 |
| question particle *-tu* after a non-2sg subject | `ca_tu` | fr-CA (casual) | strong (needs the parse) | GD | – |
| *être après* + infinitive | `ca_apres` | fr-CA (casual) | medium (needs the parse) | GD | – |
| *présentement* "now" | `ca_presentement` | fr-CA, fr-x-africa | weak | GD | 3 |

**Lexicon: Africa south of the Sahara**

| cue | name | points to | strength | evidence | treebanks |
| --- | --- | --- | --- | --- | --- |
| *essencerie* | `af_essencerie` | fr-x-africa (Senegal) | strong | W | 0 |
| *maquis* "eating place" | `af_maquis` | fr-x-africa (Côte d'Ivoire, Burkina Faso) | weak (standard "scrubland, Resistance") | W | – |
| *go* "girl" | `af_go` | fr-x-africa (Côte d'Ivoire, Cameroon) | weak (an English word in French text) | W | – |
| *gâter* "break down (a device)" | `af_gater` | fr-x-africa | weak (standard "spoil") | W | – |
| *deuxième bureau* "mistress" | `af_deuxbureau` | fr-x-africa | medium (multiword) | GD (ifa-1983 card) | – |
| *cadeauter, enceinter* | `af_cadeauter` | fr-x-africa | strong | W ("Afrique") | 0 |
| *-là* after a noun with no *ce* before it | `af_la` | fr-x-africa (also casual Québec) | weak | GD | – |

**Lexicon: France**

| cue | name | points to | strength | evidence | treebanks |
| --- | --- | --- | --- | --- | --- |
| *portable* "mobile phone" | `fr_portable` | fr-FR | medium | W ("France") | 12 |
| *petit-déjeuner* with *déjeuner* "lunch" | `fr_meals` | fr-FR | medium (needs context) | GD | 3 |
| *week-end, parking, mail* (English loans the other bodies replace) | `fr_loans` | fr-FR | weak (used everywhere) | GD | 37 (*week-end, weekend, parking*) |
| *chocolatine* | `fr_sw` | fr-FR (Southwest) | strong | GD | 0 |

**Typography and spelling**

| cue | name | points to | strength | evidence |
| --- | --- | --- | --- | --- |
| a space before `? ! ;` | `ty_space` | fr-FR, fr-BE | weak (lost in normalised text; the corpora strip U+202F) | [07](https://md.mova.elamur.ai/free/en/docs/languages/fr/07-punctuation-typography/) §2 |
| no space before `? ! ;`, a space before `:` | `ty_nospace` | fr-CH, fr-CA | weak (also web typing everywhere) | [07](https://md.mova.elamur.ai/free/en/docs/languages/fr/07-punctuation-typography/) §2 |
| feminised titles in *-eure* (*professeure, auteure*) | `sp_eure` | fr-CA, fr-CH, fr-BE (older texts) | weak (now spread in France) | GD (moreau-2014, suisse-2000 cards) |

**Rules of thumb**, to be checked against the learned weights: (1) one strong numeral cue outweighs any number of weak cues; *huitante* alone decides fr-CH; (2) *septante/nonante* without *huitante* is fr-BE or fr-CH, and the lexical cues of the two countries decide between them; (3) a text with no cue gets no label, not fr-FR: fr-FR is the default only in the sense that the treebanks are of France; (4) proper names (towns, newspapers, parties) are cues to the country a text is about, not to the variety it is written in (choudhry-2022), and should be masked as the VarDial task did; (5) the cues of casual speech (*-tu*, *icitte*, *une fois*) need a parse or a spoken source; in edited text they are rare.

## 11. How Mova handles national varieties

**What exists.** Mova reads French through the English parser with a tag map (`en xpos-fill`); the training data is UD French GSD, the written French of France, with LAS 87.00 (own tags) and 89.78 (gold tags) on GSD dev (2026-10-22, reproduced 2026-11-06). The only measurement on a text that is not the standard of France is the Poitevin one of §1.2: LAS 24.85 / 59.01 (2026-11-06). `tools/langseg` has the machinery a French variety label needs (cue features from a manual, naive Bayes weights learned from a pool, a Viterbi pass with a switch cost, the cue report), but no French label and no French cue: `src/cues.rs` and `src/lib.rs` list only the Hebrew-script varieties.

**What does not exist.**

- No variety label, no variety detection.
- No French lexicon, so no lexicon with variety marks. The French Wiktionary carries the labels used in §10 (*Belgique, Suisse, Québec, Canada, Afrique, Côte d'Ivoire, Congo-Kinshasa…*) and its raw dump is on disk (CC BY-SA, for learning and tests only).
- No labelled pool of French by country. A pool has to be built from open text whose country is known (national press under open licences, Wikipedia pages on local topics, public administration texts) before langseg can learn weights; QFrCoLA and QFrBLiMP (CC BY 4.0) give Québec sentences, and nothing comparable is held for Belgium, Switzerland or Africa.
- No test on Quebec, Belgian, Swiss or African text.
- The constructions of Québec speech (*-tu*, *être après*, *pas* in superlatives) are unknown to a parser trained on GSD.
- Creoles: a transfer experiment like Mompelat's would need a creole treebank; none is held.

**What would come first.** (1) Add the cue table of §10.2 to langseg as a French cue list, with `fr-*` labels. (2) Build a labelled pool from open text by country and learn the weights; report `langseg cues` for French so that the strengths above can be corrected. (3) Add *septante, huitante, nonante* to the numerals the parser and the lexicon know. (4) Run the GSD-trained parser on QFrCoLA's acceptable sentences to measure the drop on Québec French.

## Sources for this chapter

- Passy, Paul. 1906. *Les sons du français*, 6th ed. Public domain (`passy-1906`). §§114–115 (length, Switzerland and Belgium), §195 (*l mouillé*), liaison chapter (Swiss liaisons), the Swiss sample.
- Nyrop, Kristoffer. 1899–1930. *Grammaire historique de la langue française*, vol. II (numerals). Public domain (`nyrop-1899`). *Septante, octante, nonante*.
- Martinon, Philippe. 1913. *Comment on prononce le français*. Public domain (`martinon-1913`). The Midi *e*.
- Société du parler français au Canada. 1930. *Glossaire du parler français au Canada*. Public domain (`canada-1930`).
- Dauzat, Albert. 1908. *La langue française d'aujourd'hui*; 1922. *La géographie linguistique*. Public domain (`dauzat-1908`, `dauzat-1922`).
- Vaugelas, Claude Favre de. 1647, ed. Chassang 1880. *Remarques sur la langue françoise*. Public domain (`vaugelas-1880`).
- Kahane, Sylvain et al. 2025. Status of morphosyntactic features: illustration with written and spoken French UD treebanks. CC BY 4.0 (`kahane-2025`).
- Dankova, Natalia. 2024. Lexical richness of French and Quebec journalistic texts. CC BY 4.0 (`dankova-2024`).
- Riverin-Coutlée, Josiane. 2020. Sur le voisement des consonnes fricatives finales en français du Québec. CC BY 4.0 (`coutlee-2020`).
- Lancien, Mélanie. 2020. De la possibilité d'un relâchement des voyelles hautes… (`lancien-2020a`); Variation stylistique en français québécois (`lancien-2020b`). CC BY 4.0.
- Lamontagne, Jeffrey & Heather Goad. 2022. Weight sensitivity and prominence in Laurentian French. CC BY 4.0 (`lamontagne-2022`).
- Labelle, Marie. 2023. Apparent expletive negation, silent comparatives, and domain widening in Quebec French superlative and universal constructions. *Glossa*. CC BY 4.0 (`labelle-2023`).
- Beauchemin, David & Richard Khoury. 2025. QFrCoLA (`beauchemin-2025`); Beauchemin et al. 2026. QFrBLiMP (`beauchemin-2026`); Beauchemin et al. 2026. Idiom understanding as a tool to measure the dialect gap (`beauchemin-2026b`). CC BY 4.0.
- Mompelat, Ludovic, Daniel Dakota & Sandra Kübler. 2022. How to parse a creole: when Martinican Creole meets French. CC BY 4.0 (`mompelat-2022`); Mompelat et al. 2025, WMT 2025 CreoleMT (`mompelat-2025`).
- Havard, William et al. 2024. Technologies de la parole et données de terrain : le cas du créole haïtien. CC BY 4.0 (`havard-2024`).
- Jauhiainen, Tommi, Heidi Jauhiainen & Krister Lindén. 2022. Italian language and dialect identification and regional French variety detection using adaptive naive Bayes. VarDial 2022. CC BY 4.0 (`jauhiainen-2022`).
- Choudhry, Omar Shaur et al. 2022. Data collection and analysis of French dialects. CC BY 4.0 (arXiv) (`choudhry-2022`).
- Sagot, Benoît et al. 2025. COLaF : corpus et outils pour les langues de France et variétés de français. CC BY 4.0 (`sagot-2025`).
- Toughrai, Yassine et al. 2025. Modeling North African dialects from standard languages. CC BY 4.0 (`toughrai-2025`).
- Ngom, Charles Abdoulaye et al. 2025. Analyse textuelle et extraction géospatiale… CC BY 4.0 (`ngom-2025`); Ouedraogo, Fayçal S. A. et al. 2026. Contributing to speech-to-speech translation for African low-resource languages: the French–Mooré pair. CC BY 4.0 (`ouedraogo-2026`).
- French Wiktionary, raw dump 2026-09-01 (20260901), CC BY-SA: regional labels of the cue words checked, nothing copied.
- Cards only, no claim rests on them: OQLF *Banque de dépannage linguistique* (`oqlf-2026`); *Inventaire des particularités lexicales du français en Afrique noire* (`ifa-1983`); Moreau 2014 (`moreau-2014`); Swiss Confederation 2000 (`suisse-2000`).
- UD 2.18 French treebanks (GSD, Sequoia, FQB, PUD, ParisStories, Rhapsodie, ALTS, ParTUT for counts only, PoitevinDIVITAL), counted 2026-11-06.
- Mova: `tools/langseg/src/cues.rs`, `tools/langseg/src/lib.rs`, [langseg](https://md.mova.elamur.ai/free/en/docs/tools/langseg/); `en parse-eval` and `en xpos-fill` (`en/src/main.rs`), runs of 2026-11-06.
- [← previous 07. Punctuation, typography, capitals, abbreviations and numbers](https://md.mova.elamur.ai/free/en/docs/languages/fr/07-punctuation-typography/)
- [next → 09. Old, Middle and Classical French, and the literary tenses](https://md.mova.elamur.ai/free/en/docs/languages/fr/09-old-middle-classical-french/)
- [↑ back to Modern Standard French (fr) — a reference manual for Mova](https://md.mova.elamur.ai/free/en/docs/languages/fr/)