🗣️ 37. Register and variation

The same French grammar is used differently in a law, a news bulletin, a hotel review, a tweet and a story told to friends. Some constructions cluster in some situations (register), some belong to one country or region (dialect), some to one period, and some to the rules of one publisher or one norm. This chapter treats the variation of use: the label a text can carry (language, dialect, style of speech), the French tradition of niveaux de langue from soutenu to argotique, the spoken and written forms counted across the UD French treebanks, the formal, official and literary registers, web and computer-mediated French, song lyrics as a style with a grammar of their own, and the norms and house styles that fix the written standard. The national varieties themselves (Belgium, Switzerland, Québec, Acadia, Africa) are described in chapter 8; here they appear only as the dialect part of the label. Every treebank count was made on UD 2.18 on 2026-11-06.

Contents

1. The label of a text: language, dialect, style
2. The styles of speech
3. Register as co-occurring features
4. Spoken and written French
5. Questions as a register marker
6. Official, administrative and legal French
7. Literary French and the passé simple
8. Web and computer-mediated French
9. Song lyrics: a style with its own grammar
10. The dialect modifier
11. Norms and house styles
12. How Mova handles it

1. The label of a text: language, dialect, style

1.1 The terms

Mova labels a text by three of these, in the order in which they modify one another: the language is the base; a dialect modifier narrows it to a variety; a style of speech says how the text uses that variety. fr + CA + colloquial is everyday spoken Québec French; fr + none + official is the French of a French ministry; fr + BE + press is a Brussels newspaper. A publication's house style is a style label too (§11): the same news story set by two newspapers differs in capitals, numerals and quotation marks, not in grammar. Genre is not a separate slot: a genre usually fixes the style (a law is official, a review is neutral to colloquial) and can be written into the style label where it matters (encyclopedic). §12.1 shows how the Eel uses this label on its French autoencoder today.

1.2 The French tradition: a common language with two directions

French grammars and dictionaries describe register as a scale of niveaux de langue around a neutral centre. The clearest early statement is Bally's (1909, bally-1909-traite-de-stylistique-fran, vol. I §§ 225–227). An expression either belongs to the common language (langue commune), which evokes no particular social setting, or leaves it in one of two directions. Upward it reaches the langue « écrite », which Bally says need not be written at all: a lecturer who improvises still speaks it, and it shows best when it is used in conversation, where it sounds bookish. Écrite is also not the same as littéraire: every literary expression is "written", not every written expression is literary. Downward it reaches the langue parlée, the language of real conversation, from the merely familiar through the populaire and the vulgar to argot. The value of a word in this scale is its effect by evocation: fourrer for mettre (Où as-tu fourré ton chapeau ?) affects the hearer not by its meaning but because it evokes the setting where it is natural (vol. I § 181). Bally also observes that speakers classify themselves by this choice, and that adapting to a "higher" setting often fails: one slang word gives the speaker away (§ 223).

The modern dictionary labels keep Bally's scale. The usual set, with courant as the unmarked centre:

Two refinements from recent work. Mekki et al. (2021, mekki-2021-tremolo-tweets-multi-label-corpus) and Mekki et al. (2025, mekki-2025-automatic-characterization-french-language-registers), following Gadet, define register against a norm that has two parts: a prescriptive norm (the rules of the grammar) and an objective norm (the rules of actual usage in a medium). A text is formal when it follows both, neutral (standard) when it follows them partly, casual when it breaks the grammatical norm. On tweets this means that a hashtag can be standard usage and a missing ne a deviation in the same text. Second, the levels are not fixed: mec "guy" was slang and is now familier; bouquin "book" is familier but no longer social marking. A register lexicon has to be dated.

2. The styles of speech

2.1 Markers by level

The % sign marks a form that is the norm of informal speech and not of the written standard; it is not an error in its register.

2.2 Slang, verlan and youth speech

Argot began as the secret language of criminals and trades (Bally treats argot, § 227, apart from the jargons of the trades, § 240) and has become a general label for group speech. Its regular processes:

• verlan, the inversion of syllables: meuf (femme), relou (lourd "annoying"), chelou (louche "shady"), ouf (fou), vénère (énervé), teubé (bête); a verlan word can be inverted again (beur → rebeu);
• clipping: ado (adolescent), appart (appartement), dico, exam, resto, with the suffix -o (apéro, proprio) and -oche, -os (cinoche, calmos);
• suffixes that mark the register: -ard (politicard), -ouze (partouze), -ax (relax); Mekki et al. (2025) find word endings in -ouze and -o among the patterns that set casual tweets apart;
• borrowings: from Arabic (kiffer, wesh, seum), Romani (chourav, bédo), English (crush, chiller, je suis dead), and the African languages; the français des cités of the suburbs is where many enter.

Slang dates fast. The discourse marker genre and the connective du coup are recent enough to show between two spoken treebanks recorded about ten to twenty years apart (§4.2).

3. Register as co-occurring features

Biber's multidimensional method — count many features per text, find the factors along which they vary together, read each factor as a function — has not, as far as the research library shows, been applied to French in an open publication. The open French work takes other routes to the same question. Mekki et al. (2021) annotated 3,269 tweets by hand for casual, neutral and formal register, with each label justified by one of 47 linguistic descriptors; the annotators agreed on casual and neutral (73% and 76%) far more than on formal (36%), whose confusions are with neutral. Mekki et al. (2025) then mined the whole 228,505-tweet corpus for emergent sequential patterns of lemma, part of speech, morphology and syntactic function without choosing the features in advance, and recovered 12 of the 28 register features of the earlier literature that their representation could express — among them ça for cela, negation without ne, nous against on, and y for impersonal il (y a). Christodoulides (2020, christodoulides-2020-variation-prosodique-des-styles-de) ran a principal component analysis of prosodic measures over 300 hours of the CEFC spoken corpus (2,500 speakers): no single prosodic parameter separates speaking styles, the first two components explain 25% of the variance and eight explain 71%; media presentations are homogeneous, conversations and sociolinguistic interviews vary most.

The UD treebanks show the co-occurrence in miniature. Rates per 1 000 words (syntactic words without punctuation), all splits, 2026-11-06:

"ne kept" is the share of negated finite clauses (a finite verb, or a predicate with a finite auxiliary or copula, governing pas, jamais, rien, personne, point or guère) that also have ne: ParisStories 395 such clauses, Rhapsodie 355, Sequoia 251, GSD 1 018, PUD 86. The PUD discourse rate comes from 21 sentence-initial mais annotated as discourse; the FQB dislocated rate is an annotation choice in Qu'est-ce qu'un espace personnel ?, where the noun phrase is dislocated. Neither is a register fact.

The features move together, as in Biber's "involved" dimension: where there are many interjections and fillers there are many self-repairs, first- and second-person subjects, on and ça, and few ne. The written treebanks sit at the other pole, and within them the speech-like texts stand out: Sequoia's first- and second-person rate (7.6) comes from the Europarl debates (26.7 per 1 000 words: nous 146 times as subject, on 25) and the drug leaflets (vous devez), not from its Wikipedia (0.7).

3.1 Inside one spoken treebank

Rhapsodie gives each sample metadata for its situation: planning, interactivity, social context, channel, task. Counted by those fields (2026-11-06):

The same speakers' French moves with the situation. Planned public speech (radio news, political speeches, sermons, the info kiosk) keeps ne in three negations out of five, private conversation in one out of four; fillers halve from private to public. Planning and publicity work in the same direction but are not the same variable: a public interview is spontaneous. Planned speech still dislocates:

(1) Rhapsodie Rhap_M2006-7 (radio news, planned): de la crise aux Antilles, il ne devrait pas en être question au sommet social de l'Élysée cet après-midi. — a dislocated topic (de la crise aux Antilles, resumed by en) in broadcast news, with ne kept.

4. Spoken and written French

4.1 The marks of the spoken register

(All splits, 2026-11-06. Passé simple: finite indicative forms with Tense=Past; the imperfect is Tense=Imp. Bare faut: faut/fallait not preceded by il and not in inversion faut-il.)

Read together:

• On stands for "we" in speech. In ParisStories on (in all its uses) outnumbers subject nous 782 to 4; in Rhapsodie's public samples nous returns (60 of its 78). In writing nous holds, and on is often the generic "one".
• Ça has replaced cela in speech (309 : 2 in ParisStories); in writing cela still leads (142 : 48 in GSD), and the GSD ça are mostly in reviews and quoted speech.
• Ne is dropped in 86% of negations in the stories among friends. The written figure is not 100%: the count finds 26 negated clauses without ne in GSD. About half are not clausal negation at all (pas mal de "quite a lot of", si jamais "if ever", rien à voir), so the measure slightly understates retention everywhere; the other dozen are real drops, nearly all in reviews and quoted speech (13 euros c'est pas énorme, fr-ud-train_01264; on connait pas davantage, fr-ud-train_01498; J'ai jamais vu ça.). Edited written French keeps ne practically always.
• The impersonal il of il faut is dropped in about half the cases in ParisStories, but il y a is almost always transcribed with il (216 : 2), although speakers usually say [ja]. The transcription conventions of the spoken treebanks restore some forms the speakers reduced; counts of reduced forms in transcripts are lower bounds.
• The passé simple is absent from speech; it is a written tense (§7).

Two ParisStories sentences show several marks at once:

(2) ParisStories_2022_7_chienGourmand__29: mon père, il veut pas trop. — dislocated subject resumed by il, ne dropped.

(3) ParisStories_2022_19_contenteDeOuf__41: et euh, et du coup, bah, à chaque fois, genre, euh, on était là, tous les jeudis. — the filler euh, the connective du coup, the markers bah and genre, on for "we", and a self-repair (et euh, et).

UD French gives spoken constructions their own relations and subtypes: reparandum for the abandoned part of a repair, discourse for fillers and markers, dislocated with the subtypes dislocated:subj (505 in the two spoken treebanks), dislocated:obj (55), dislocated:obl (9) and dislocated:mod (32), and parataxis:parenth, parataxis:insert, conj:dicto for parentheticals, insertions and the repetitions of reformulation. The principles of segmenting speech into units for UD are set out by Kahane et al. (2021, kahane-2021-annotation-guidelines-ud-sud-treebanks); the Rhapsodie treebank, its prosodic layer and its macrosyntactic segmentation are described by Botero-Garcia et al. (2025, garcia-2025-intonosyntactic-treebank-spoken-french-what). Both spoken treebanks are punctuated by their annotators, not by the speakers.

4.2 Discourse markers and generation

The most frequent discourse words, per 1 000 words (2026-11-06):

ParisStories was recorded between 2017 and 2021 by linguistics students, of friends and relatives telling anecdotes (its README). Rhapsodie's samples were gathered by a project begun in 2008 (Botero-Garcia et al. 2025), many from older corpora; where an age is given, its speakers are 15 to 81. The markers that dominate ParisStories — du coup (230 occurrences against 1), genre (78 against 0), bah rather than ben — are those that observers of French in the 2010s describe as spreading among younger speakers; Rhapsodie has the older ben and hein. The two corpora also differ in genre (anecdotes among friends against a mixture with much public speech), so the table shows a co-occurrence of age, date and situation, not a measured change in time. The markers themselves are described in chapter 38.

5. Questions as a register marker

The three ways to ask a yes/no question and the three ways to place a question word (§2.1) are the classic register variable of French syntax. Counting sentences that end in a question mark, by the form they contain (2026-11-06):

"Clitic inversion" is a subject clitic after the verb (viens-tu, a-t-il); "neither" holds intonation questions (tu viens ?), questions on the subject (qui a écrit… ?), nominal stylistic inversion and fragments. Inversion is the written form: 43% of the GSD questions, one question in 145 among friends. In speech the work is done by est-ce que and by intonation. FQB, whose questions are search queries and public-service FAQs written in the written norm, has inversion in a quarter and a large share of subject questions (Qui…, Quel est…).

Eshkol-Taravella, Barbedette, Liu & Soumah (2022, taravella-2022-classification-automatique-de-questions-spontanees) separate spontaneous from prepared questions in the ESLO2 and ACSYNT interview transcripts. Their indicators of a prepared question are subject inversion, the question word at the start, rheme before theme, and an announcement of the question (alors maintenant je vais vous poser des questions); of a spontaneous one, the question word at the end (elle est constituée comment ?), theme before rheme (mais alors là vous êtes en France pour combien de temps maintenant ?), disfluencies, requests for clarification and repeated names. A logistic regression on these criteria alone separated the two kinds with an F-score of about 0.75; the interviewer's prepared questions use the written syntax even when spoken.

6. Official, administrative and legal French

Bally (1909, vol. I § 238) defines the langue administrative as the language that states "in a precise and impersonal way" the obligations of social life, from police regulations to the articles of the code and the constitution. It is scientific in essence but always in contact with ordinary life, which is why it evokes its setting more than scientific terminology does. His specimen is an ordinance on rabid dogs (Les chiens qui présenteront des symptômes de rage devront être abattus ou séquestrés. Les maires donneront à cet effet les ordres nécessaires. Les contrevenants … seront passibles des peines de police, sans préjudice des peines plus graves …), set beside a plain rewording (Si un chien a l'air d'avoir la rage, il faudra le tuer ou le tenir enfermé …); the rewording is less precise, and it no longer sounds like authority speaking.

The marks of the register, in his example and in any French code:

The Civil Code's article on tort shows several of them: Tout fait quelconque de l'homme, qui cause à autrui un dommage, oblige celui par la faute duquel il est arrivé à le réparer. (Code civil, art. 1240, formerly 1382; French legislation is an official text, freely reusable.) A generic subject (tout fait quelconque), the present of rule, autrui, a pied-piped relative par la faute duquel.

The treebanks Mova holds have no modern legal text. Sequoia's Europarl debates are formal parliamentary speech (nous 146 times as subject against on 25, ne kept in 85 of 86 negations); its drug leaflets (EMEA) address the reader as vous in instructions (vous devez). ALTS is legal French, but of the sixteenth century (trial records of Guernsey, 1563–1569, and Terrien's commentary on the custom of Normandy, 1578), and serves here only as a contrast: it keeps ne in 93% of its 192 negated finite clauses, and its on (210) is the generic "one" of the procedure.

7. Literary French and the passé simple

The soutenu register has two morphological marks that speech lacks: the passé simple and the imperfect subjunctive. Counts of finite indicative forms with Tense=Past (the passé simple), 2026-11-06:

The most frequent forms in GSD are fut (415), furent (84), fit (40), prit (27), eut (24), devint (23). The passé simple is not only literary: it is the tense of historical narration in encyclopedic and journalistic prose, where a biography is written Il naquit … il devint …. The imperfect subjunctive (Mood=Sub|Tense=Imp) appears 10 times in all of GSD and once in PUD; outside literature and the most formal prose it is replaced by the present subjunctive (chapter 13).

The written norm of the classical language was defined in its time against court usage. Vaugelas (1647, Remarques sur la langue françoise, ed. Chassang 1880, vaugelas-1880-remarques-sur-la-langue) set le bon usage 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", and against it the mauvais usage of the majority. The literary register of today descends from that norm, and the distance between the literary and the spoken language that Bally (vol. I § 245) calls variable is large in French.

8. Web and computer-mediated French

Web French writes in a written medium with a conversational involvement; French linguists call the result parlécrit. Mekki et al. (2021) chose tweets for register study because their short format holds one register at a time, and collected 228,505 tweets from the Paris area (August 2020, 6.2 M words), labelled casual, neutral or formal by a CamemBERT classifier trained on the hand-labelled seed (F1 0.85, 0.84 and 0.95 on the seed's test part). Of the seed, 51% was neutral, 39% casual and 10% formal. The descriptors the annotators gave most often as reasons:

The CMC-specific forms (hashtags, @-mentions, emoji, which Paveau calls technomorphèmes) are not casual in themselves: a hashtag used as an ordinary word inside a well-formed sentence (Les violences vécues en #France ne sont pas des #incivilites) marks the formal register, and emoji placed to frame information mark it too. Casual tweets use pictograms for emotion, as a substitute for prosody. Mekki et al. (2025) confirm by pattern mining that first- and second-person singular forms (j', t', toi), the marker en gros, doubled subjects (Westbrook il est …) and electronic spellings are the strongest marks of casual against standard tweets.

GSD's web part (reviews and blogs) is much closer to the written norm. Emoticons are rare (four :), three ;), one :-) in all of GSD, annotated SYM and attached as parataxis), and there is no mdr, lol or ptdr. What GSD does record is the spelling errors of quick writing: 941 words carry Typo=Yes with the intended form in CorrectForm in MISC, most of them homophone and agreement errors: a for à (17), ou for où (8), coeur for cœur (8), un/une swapped (15), serais for serai (3), trés for très (6). The UD French annotation keeps the written form and records the correction, so a tree of a review is a tree of what was written; normalisation, which replaces non-standard words by standard ones before parsing, is the alternative surveyed by Nishimwe (2023, nishimwe-2023-normalisation-lexicale-de-contenus-generes); it helps a parser trained on standard text, and it removes exactly the forms a register study counts.

(4) GSD fr-ud-train_02431 (a restaurant review): On y va au moins une fois tous les 15 jrs et jamais déçu :) — jrs for jours, a verbless second clause, jamais without ne, and an emoticon.

9. Song lyrics: a style with its own grammar

A song is a style of speech in its own right. Its text is shaped by a tune before it is shaped by the grammar: the line has a fixed number of sung syllables, it ends on a rhyme, its stresses fall on the beats, and it is heard once, not reread. Lyrics therefore depart from the written norm routinely and without being errors, in every period and genre — the traditional song, the chanson of the cabaret and the music hall, rap and slam. A parser trained on newspapers and reviews meets in lyrics a grammar it has not seen; a register score computed on lyrics mixes the norm of the genre with the norm of speech. This is why song is a separate style label in this manual and not a kind of colloquial.

The examples below are from public-domain songs (traditional songs, and La Marseillaise, 1792) or constructed; copyrighted lyrics are not quoted.

9.1 Metre: the syllable count decides the form

French verse counts syllables, and the sung line counts them as the classical verse does: a mute e between consonants is sung as a full syllable, a final mute e before a consonant too; only an e before a vowel is elided, and an e at the end of the line is not counted in the metre, though the tune usually gives it a short note. The traditional song exploits both sides of the rule:

• Prê-te-moi ta plu-me / Pour é-cri-r(e) un mot (Au clair de la lune): the -e of prête is sung as a full syllable, so prête-moi has three sung syllables where speech has two (/pʁɛt.mwa/); the -e of écrire is elided before un, and the -e of plume at the line end is the unstressed feminine ending that the tune gives a short note. The same word has a different syllable count in speech and in song.
• Il é-tait un pe-tit na-vi-re / Qui n'a-vait ja-ja-ja-mais na-vi-gué: a syllable is repeated to fill the measure. A tokenizer reads ja-ja-jamais as one hyphenated form or three; it is one word jamais with a sung stutter, the musical counterpart of a reparandum (chapter 38).
• Sur le pont d'A-vi-gnon / L'on y dan-se: l'on for on, a written-formal variant kept here because it adds a syllable; elsewhere l'on marks the soutenu (§2.1).

The opposite operation, dropping syllables the written norm keeps, is the mark of modern song, and of rap above all. A rapper writes the elisions that speech makes and the spelling hides: j'sais pas for je ne sais pas, t'as for tu as, y a for il y a, j'te l'dis for je te le dis, chuis or j'suis for je suis, p'tit for petit. The written apostrophe here records a schwa that is not pronounced, not the elision of the norm (which applies only before a vowel; chapter 1). The UD treebanks of speech normalise these forms back to the standard spelling (ParisStories and Rhapsodie write je sais pas, il y a), so no treebank Mova holds contains them as written; a parser meets them first in lyrics and in web French (§8).

9.2 Rhyme and metre move the words

The rhyme falls on the last stressed syllable of the line, so a word is moved to the end to rhyme, and a complement is moved to the front to fit the count. The traditional song and the anthem use orders that prose has lost:

None of these is wrong; most are literary or archaic outside song (§7), and the song keeps them alive because its listeners learn the text by heart. For a parser, they produce non-projective or rare attachments (a nmod before its head across a verb; a null subject that the UD French guidelines do not expect), and a single anthem line may defeat a model that has never seen them.

9.3 The grammar of speech, fixed in writing

Lyrics written after the 1950s take up the spoken forms of §4 as their norm rather than as a deviation: ne is absent from almost every negation (j'sais pas, c'est pas grave); on is the first person plural; ça replaces cela; dislocation is constant (moi, j'ai rien dit; la ville, elle dort jamais); questions are intonation questions (tu viens ?) or have the question word in place (tu vas où ?). A song that keeps ne and nous is marked as formal or old-fashioned by that choice alone. The marks counted in §4.1 for speech are therefore expected in lyrics at least as often as in the spoken treebanks, without the hesitations and repairs of real speech: a song is planned speech written to sound unplanned.

9.4 Rap and slam: verlan, borrowings, the flow

Rap adds three layers on top of the spoken grammar.

• Verlan and argot (§2.2): meuf, relou, chelou, vénère, ouf, tess (cité), keuf (flic), reuch (cher). Verlan words take the grammar of their class: une meuf, ils sont relous, c'est trop ouf.
• Borrowings, with French morphology: from English (bitch, money, flow, mic, chiller, kiffer — the last from Arabic), from Arabic (wesh, wallah, seum), from Romani (poucave "snitch"), from Spanish (igo "mate"), from the languages of West and Central Africa. Zurbuchen & Voigt (2024, zurbuchen-2024-computational-analysis-exploration-linguistic-borrowings) annotated a lexicon of 741 borrowings and counted them in more than 8,000 French rap songs (1991–2024, 4.6 M tokens, 98,740 borrowing tokens). They find that borrowings have roughly doubled since the 1990s, to about 2% of all words in the lyrics; that shorter borrowings spread more; that terms of music and the arts are the one semantic class positively associated with usage in all three of their regressions (word count, songs, artists), although their share has declined over time; and that borrowings from the Niger-Congo languages grow fastest, those from Arabic stay high and stable, and interjections and set expressions gain over nouns. Their lexicon was curated by hand because off-the-shelf borrowing detection did not work well enough for French. The borrowed verbs are conjugated as first-group verbs (on chill, ça me saoule, il a ghosté), the nouns take French determiners and plurals, and English words may keep their English plural -s in spelling only.
• The flow. The rhythmic line of rap is longer and freer than the sung line of the chanson, but internal rhymes and assonance (multisyllabiques) drive the choice of words as strongly as the metre drives the traditional song. Word order inside the clause is less affected than in the old song: rap usually keeps the order of speech and plays instead on the lexicon, on enjambment across lines, and on interjections and ad-libs (yeah, eh, gang) placed between lines.

A constructed line shows the combined grammar: Wesh, la tess elle dort pas, moi j'chill avec ma meuf, c'est trop ouf — an interjection, a dislocated subject resumed by elle, ne dropped, j' before a consonant, an English verb with French agreement, two verlan words, a degree adverb trop in its colloquial sense "very". Each element is a feature of §§2–4; their density in one line is the song's.

Slam, spoken poetry performed without music, keeps the rhyme and the play on sounds but not the metre; its grammar is closer to written literary French, often with full negation, and its marks are repetition, anaphora (the figure) and wordplay (chapter 34).

9.5 Lyrics as data

Three properties of lyrics matter before any measurement.

• The unit is the line, not the sentence. Lyrics are usually written without final punctuation, one line per musical phrase, and a sentence runs across several lines or a line holds two sentences. A UD sentence segmenter trained on prose has no cue to follow; segmenting by line and by sentence give different trees.
• The spelling is the transcriber's. Lyrics sites transcribe by ear; the same song exists with j'suis, chuis and je suis, with or without the verlan spelled consistently. Zurbuchen & Voigt added every variant spelling seen five times or more to their lexicon for this reason.
• Rights. Almost all modern lyrics are copyrighted. The public domain holds the traditional songs, the songs of the nineteenth century (Béranger), and the chanson réaliste of authors dead more than seventy years (Aristide Bruant's argot songs, for instance). A song corpus that can be built into Mova is a corpus of these; modern rap and chanson can only be counted and tested on, under the rules of Data and model licences.

10. The dialect modifier

The national and regional varieties are the subject of chapter 8. As the dialect part of a text's label they are short codes; the values proposed here follow the region subtags of BCP 47 language tags where a country is meant:

Such codes are a convention of the configuration, not a closed list in the code (§12.1). The dialect and the style are independent: Québec French has its own soutenu, courant and familier levels, and Lancien (2020, lancien-2020-variation-stylistique-en-francais-quebecois) measures, in a preliminary analysis of two couples playing a picture-difference game, that Québec vowels are shorter and more centralised when they speak to their spouse than to an experimenter from the same community or from Europe: style shifts within the variety. A variety can also differ in the way it writes the same register: Dankova (2024, dankova-2024-lexical-richness-french-quebec-journalistic) finds that political and cultural articles in French newspapers of France have more adjectives, more varied adjectives, and more participles and gerunds than their Québec counterparts. For Québec French, the library also holds QFrBLiMP, a benchmark of grammatical minimal pairs (beauchemin-2026-qfrblimp-quebec-french-benchmark-linguistic).

The French UD treebanks Mova holds are all from France (PUD translated from English), and a regional marker in them is rare: the GSD review (5) has the Belgian softener une fois.

(5) GSD fr-ud-train_01264: Parmi mes connaissances beaucoup sont du même avis que moi, alors allez y une fois, 13 euros c'est pas énorme, et faites vous votre idée. — allez y une fois "do go there", a Belgian une fois after an imperative, in the same sentence as a dropped ne.

PoitevinDIVITAL, filed among the French treebanks, is Poitevin-Saintongeais, a langue d'oïl of its own and not a dialect of French.

11. Norms and house styles

The written standard is fixed less by grammar than by norms that writers choose to follow. Each is a candidate for the style part of a label when a corpus is known to follow it.

The 1990 rectifications are a norm option, not a replacement: the text asks that they be taught "it being clearly understood that the forms now in use will of course remain admitted", and the Académie noted that both spellings of every modified word stay accepted until usage decides (cslf-1990). A corpus can therefore be in the traditional spelling, in the rectified spelling, or mixed, and a spelling checker or a tokenizer that knows only one rejects correct French in the other. The typographic norms differ by country (the space before ; ! ? is a French-of-France and Imprimerie nationale rule that Swiss practice often omits; chapter 7). Like the English style guides, these norms describe the formal written pole; the parser annotates what is written, whatever norm it follows.

12. How Mova handles it

12.1 The Eel's condition: a label, not a detector

The Eel (tools/eel/) is Mova's learner for its native AST and pragmatics. On its French branch (commit of 2026-11-06, not yet merged into main), the Eel can run a French autoencoder round (ae-fr), and each autoencoder section may carry a condition, checked in tools/eel/src/autoencoder.rs (fn condition):

"condition": {"language": "fr", "dialect": null, "style": "encyclopedic"}

What the code does:

• It accepts only the three keys language, dialect and style; any other key (a test tries register) is refused before the round starts.
• language is required and must be two or three lowercase letters (an ISO 639 code by convention; the code checks the shape, not the list); dialect is null or 2–8 ASCII letters and digits (CA, BE, acadian; a hyphenated tag such as fr-CA is refused, because the language is already in its own field); style is null or 1–64 characters of lowercase letters, digits and : _ . - (encyclopedic, colloquial, house:lemonde, norm:1990).
• The condition is written into the state directory's dev.json with the pinned dev and test samples on the first round. A later round whose condition differs is refused (the tables belong to another condition) and the Eel pauses; the English autoencoder's tables are never touched by a French round, and two round kinds cannot share a state directory.
• The round's receipt records the condition.

What it does not do:

• It does not detect a dialect or a register. The condition is selected by the data: the configuration names a corpus of that language, dialect and style, and the label says what that corpus is. No code checks that the corpus matches its label.
• There is no list of allowed dialect or style values; only their shape is checked. The values of §1.2, §10 and §11 are the conventions this manual proposes.
• The condition does not change how the tables are trained; it keeps tables of different conditions apart, so that a model learned on encyclopedic French is not silently continued on colloquial French.

A song corpus under the condition. Song is a style value like any other: {"language": "fr", "dialect": null, "style": "song"}, or a narrower song:trad, song:chanson, song:rap, all of which pass the shape check. What the label would do today is what any condition does: a French song corpus would get its own autoencoder state directory, its own pinned dev and test samples, and tables that are never continued on encyclopedic or colloquial prose, nor the reverse. The song grammar of §9 (sung syllables, fronted complements, null subjects of the old songs, rap elisions and borrowings) would then be learned from songs and measured on songs, instead of being averaged into, and counted as errors against, the prose model. What the label would not do: it does not change the tree parser, which has one model for all French and no style input; it does not segment lyrics into lines or sentences (§9.5); and it does not supply a corpus. None is configured on 2026-11-06, and a buildable one would have to be public-domain songs (§9.5). Conditioning the parser itself on the style, or measuring it on lyrics, is not implemented.

12.2 Register measurement

world register-stats (world/src/register.rs) measures which words a register corpus uses far more than a reference corpus (log-odds with an informative Dirichlet prior, after Monroe, Colaresi & Quinn 2008, z-scored). Its tokenizer takes lowercase alphabetic runs of two letters or more, so it would run on French books unchanged (l'homme gives homme; accented letters are alphabetic). But the three lexicons compiled into global — register-legal.tsv, register-tale.tsv, register-children.tsv — are measured on English public-domain books, and global::register_z knows only those three names. No French register lexicon exists. The obvious first one, an administrative and legal lexicon from public-domain French official texts against a reference of French books, has not been built.

world domains and the expression lens (world/src/domains.rs, prag/src/lens.rs) read register labels (slang, formal, humorous …) from the English Wiktionary's expression base through global::register_domain; nothing reads the French Wiktionary's labels (familier, populaire, argot, soutenu, Belgique, Québec). Its raw dump is on disk, CC BY-SA, for learning and tests only (Data and model licences).

12.3 Pragmatics

Mova's sentence labels (prag/src/schema.rs: form, act, indirect, hedge, polarity, voice) have no register or style field; register lives at the document level, in the planned document passport (domain, style, narrator, purpose, mode; Context, layers and memory, Document level), which is not implemented. The labels are closed codes with language-neutral values, but every labeller and witness behind them reads English.

French pragmatics is being built for the Eel. On the same French branch the Eel's schedule has a round kind prag-fr (tools/eel/src/schedule.rs), run as a gated native round (tools/eel/src/native.rs): a pinned executable measures the installed state and a candidate on a pinned dev sample, the Eel checks hashes and the separation of train and dev keys, and installs only on a dev-metric gain or records a measurement whose negative control must fail. The French pragmatics executable itself is being written and has produced no labels or measurements; nothing in it reads register yet. Above the French and English layers, a language-independent meta-pragmatics is planned: one inventory of acts, stances and situational features (including the style of speech of §1) into which each language's pragmatics maps, so that a French familier and an English informal are the same value. It is a design direction; no code implements it.

12.4 What the counts of this chapter could become

Every feature of §§3–5 is read from a UD tree with no lexicon beyond a dozen lemmas: the rates of discourse, reparandum and dislocated, first- and second-person subjects, on/nous, ça/cela, ne retention in negated finite clauses, the passé simple, the question forms. They would give a French register score per document from the trees alone, as the treebank tables show they separate speech from writing and private from public speech. They are counted here by hand on the gold trees; Mova computes none of them. On its own parses they would be weaker: Mova's tree parser reads French through en xpos-fill, whose English tags carry no mood or tense (chapter 13), so the passé simple and the subjunctive must be read from FEATS of a French tagger Mova does not have.

12.5 Not there yet

• No French register lexicon, no French register labels from the French Wiktionary.
• No register, dialect or norm detection: the Eel's condition is a declared, pinned label.
• No register-conditioned parsing: one model parses written and spoken French; its score on the spoken treebanks is not measured.
• No song corpus and no song-conditioned parsing: the style song is a valid label with nothing behind it yet (§12.1, §9).
• No normalisation layer for web French, and no use of Typo=Yes/CorrectForm in training or evaluation.
• No French pragmatic labels (the prag-fr round has no executable yet) and no meta-pragmatics.

In Mova

• The Eel's condition and rounds (French branch, 2026-11-06): tools/eel/src/autoencoder.rs (condition, pin_heldout), tools/eel/src/schedule.rs (KINDS), tools/eel/src/native.rs, the test the_french_autoencoder_keeps_its_own_state_and_condition in tools/eel/src/lib.rs.
• Register lexicons: world/src/register.rs, global/data/register-legal.tsv, register-tale, register-children; global::register_z, global::register_domain in global/src/lib.rs.
• Domains and the lens: world/src/domains.rs, prag/src/lens.rs.
• Sentence labels: prag/src/schema.rs; the document passport: Context, layers and memory.
• The English counterpart: 37. Register and variation.

Sources

• Bally, C. 1909. Traité de stylistique française, vol. I. Heidelberg: Winter. §§ 223, 225–227 (the common language and its two directions, argot), § 238 (the administrative language), § 245. bally-1909-traite-de-stylistique-fran (public domain)
• Vaugelas, C. F. de. 1647. Remarques sur la langue françoise, ed. A. Chassang, 1880. Preface (le bon usage). vaugelas-1880-remarques-sur-la-langue (public domain)
• Conseil supérieur de la langue française. 1990. Les rectifications de l'orthographe. Journal officiel, 6 December 1990. cslf-1990 (official text)
• Mekki, J., Lecorvé, G., Battistelli, D., Béchet, N. 2021. TREMoLo-Tweets: A Multi-Label Corpus of French Tweets for Language Register Characterization. RANLP. mekki-2021-tremolo-tweets-multi-label-corpus (CC BY 4.0)
• Mekki, J., Béchet, N., Battistelli, D., Lecorvé, G. 2025. Automatic characterization of French language registers: illustration on tweets. TAL 67(1). mekki-2025-automatic-characterization-french-language-registers (CC BY 4.0)
• Christodoulides, G. 2020. Variation prosodique des styles de parole et interface syntaxe-prosodie. JEP. christodoulides-2020-variation-prosodique-des-styles-de (CC BY 4.0)
• Eshkol-Taravella, I., Barbedette, A., Liu, X., Soumah, V.-G. 2022. Classification automatique de questions spontanées vs. préparées dans des transcriptions de l'oral. TALN. taravella-2022-classification-automatique-de-questions-spontanees (CC BY 4.0)
• Nishimwe, L. 2023. Normalisation lexicale de contenus générés par les utilisateurs sur les réseaux sociaux. nishimwe-2023-normalisation-lexicale-de-contenus-generes (CC BY 4.0)
• Lancien, M. 2020. Variation stylistique en français québécois : l'effet de l'identité de l'interlocuteur. JEP. lancien-2020-variation-stylistique-en-francais-quebecois (CC BY 4.0)
• Dankova, N. 2024. Lexical Richness of French and Quebec Journalistic Texts. dankova-2024-lexical-richness-french-quebec-journalistic (CC BY 4.0)
• Zurbuchen, L., Voigt, R. 2024. A Computational Analysis and Exploration of Linguistic Borrowings in French Rap Lyrics. ACL Student Research Workshop. zurbuchen-2024-computational-analysis-exploration-linguistic-borrowings (CC BY 4.0)
• Traditional songs (Au clair de la lune, Il était un petit navire, Sur le pont d'Avignon, Malbrough s'en va-t-en guerre, À la claire fontaine) and Rouget de Lisle, La Marseillaise (1792): public domain.
• Beauchemin, D. et al. 2026. QFrBLiMP: a Quebec-French Benchmark of Linguistic Minimal Pairs. beauchemin-2026-qfrblimp-quebec-french-benchmark-linguistic (CC BY 4.0)
• Kahane, S. et al. 2021. Annotation guidelines of UD and SUD treebanks for spoken corpora: A proposal. kahane-2021-annotation-guidelines-ud-sud-treebanks (CC BY 4.0)
• Botero-Garcia, M. P. et al. 2025. An intonosyntactic treebank for spoken French: What is new with Rhapsodie? garcia-2025-intonosyntactic-treebank-spoken-french-what (CC BY 4.0)
• Code civil, art. 1240 (official text).
• Cards only (not open, not quoted): Dictionnaire de l'Académie française, 9th ed. (paris-2024-dictionnaire-de-lacademie-fran); Lexique des règles typographiques en usage à l'Imprimerie nationale (paris-2002-lexique-des-regles-typographiques); OQLF, Banque de dépannage linguistique (oqlf-2026-banque-de-depannage-linguistique).
• Monroe, B. L., Colaresi, M. P., Quinn, K. M. 2008. Fightin' Words. Political Analysis 16(4). As cited in world/src/register.rs.
• Universal Dependencies French treebanks, UD 2.18: GSD, ParisStories, Rhapsodie (CC BY-SA 4.0; short quotes with sent_id), Sequoia, FQB (LGPL-LR; counts only), PUD (CC BY-SA 3.0), ALTS (CC BY-SA 4.0; sixteenth-century, contrast only). Counted on 2026-11-06 from UD_French-*, all splits, syntactic words without punctuation as the denominator; Rhapsodie's groups from its sentence metadata (planning_type, social_context).

This page in Markdown → https://md.mova.elamur.ai/free/en/docs/languages/fr/37-register-and-variation/