🗣️ 07. Punctuation, typography, capitals, abbreviations and numbers

French uses the marks of the other Latin-script languages and sets them differently. The differences matter to a reader of the printed page, and they matter more to a program that cuts a string into tokens: the space that sits before a question mark, the apostrophe that leaves a one-letter word behind, the hyphen that joins a verb to its inverted pronoun, the space that groups the digits of a number. This chapter describes the conventions of the French standard, the points where Switzerland, Belgium and Canada differ, what the open treebanks actually contain, and what Mova's tokenizer does with it. It replaces the older "12. Punctuation and typography" of this manual; its numbers are re-counted on the UD 2.18 release (2026-11-06).

The authoritative statement of the French rules is the Lexique des règles typographiques en usage à l'Imprimerie nationale (paris-2002), which is copyrighted and which we know only by its card (paris-2002-lexique-des-regles-typographiques). No sentence below is attributed to it as a quotation. The rules given here are the ones that every French typesetting manual and word processor share, checked against what the corpora do (§12); where we could not check a rule in the data, the text says so.

1. The marks and their tags

2. The high marks and the spaces

French has four high marks: the semicolon, the colon, the exclamation mark and the question mark. Each is separated from the word before it by a space.

The space is non-breaking so that a line can never begin with a mark that belongs to the word before. In practice typists put a plain space, many keyboards give no easy way to type U+202F, and text from the web mixes plain space, U+00A0 and no space. The standard is a space; a program must read all of these.

What the corpora show (UD 2.18, SpaceAfter=No analysis, 2026-11-06). A mark counts as "spaced" if the token before it has no SpaceAfter=No:

In the written treebanks the French rule holds in 99.7% of the cases in GSD; the six exceptions are typing slips. In ParisStories, which transcribes spoken stories with a mixed tradition, 21% of high marks are glued, and the comma is never separated by a space (5 378 glued against 10 spaced). The comma in GSD is glued 18 647 times and spaced 3 times. The treebank text keeps ordinary spaces: the corpus text lines contain no U+00A0 and no U+202F (neither character occurs anywhere in the French UD files, checked on all lines of the nine French treebanks, 2026-11-06), so the non-breaking character is lost before it reaches any tool.

Other languages' practice. Switzerland usually writes the question mark, the exclamation mark and the semicolon directly after the word, with a space only before the colon and inside the guillemets; Belgium follows France. Québec, following the practice of the Office québécois de la langue française, puts a non-breaking space before the colon and inside the guillemets but, in much printed usage, none before ; ! ?. These are practices, not different standards: the Imprimerie nationale rule is the one a French publisher uses, and a text can mix them.

3. The marks one by one

3.1 The full stop

It ends a declarative sentence and is followed by a capital. It ends an abbreviation that is cut inside a word (M., p., av.: §8). It is not put after a title, a heading or a caption, nor after the last abbreviation point of a sentence that already ends with one. The full stop is not used as a thousands separator in the standard (§9.1) nor in a time of day (15 h 30).

3.2 The comma

The French comma is used less than the English one in several places and more in others.

• It separates the members of a list (du pain, du vin, du fromage), but is not put before the last element when a coordinator introduces it: du pain, du vin et du fromage. There is no "Oxford comma"; a comma before et is a mistake unless it separates two clauses with different subjects (Il pleut, et nous sortons quand même).
• It sets off an apposition, an incise and a vocative (Paris, capitale de la France, est grand).
• It follows a fronted complement of some length (Dans la nuit du 3 au 4 août, la Chambre vota…), and often a subordinate clause before its main clause.
• It is not put between a subject and its verb, nor before que in a complement clause (Je pense que tu as raison), nor before a restrictive relative: les enfants qui dorment (only the children who sleep) against les enfants, qui dorment (all the children, and they sleep). The comma is the whole difference between a restrictive and an appositive relative in writing (25. Relative clauses).
• It is the decimal mark: 3,14 (§9.1).

3.3 Semicolon, colon, exclamation and question marks

The semicolon joins two independent clauses that are closely linked, or separates the elements of a list when they themselves contain commas. The colon introduces an enumeration, an explanation, a quotation (Il a dit : « non ») or a consequence; after a colon the word begins with a lower-case letter unless it is a proper name or opens a quoted sentence. The exclamation mark follows an interjection or an exclamative sentence; the question mark follows a direct question and is not put after an indirect question (Je me demande s'il viendra.). Two marks at the end may be combined (Quoi ?!) but a tidy text avoids it; the sequence ?! occurs in the open treebanks only in web text.

3.4 Ellipsis

The ellipsis (points de suspension) is three dots, typed as three full stops or as the single character U+2026. It follows the word directly and is followed by a space. It marks a break, a trailing-off or an omission in a quotation, then written between square brackets: […]. It is not followed by a full stop and it is not combined with etc.: etc… is wrong, etc. is enough. In GSD the three full stops ... occur 181 times (as one token) and the character … never; in ParisStories the transcriptions have … 67 times and ... 5 times; Rhapsodie has … 250 times and ... never (all 2026-11-06). The written web text of GSD types three points, the spoken transcriptions use the single character: a tokenizer must read both as one token.

3.5 Parentheses, brackets, dashes, slash

Parentheses and brackets have no space inside and a space outside. A pair of dashes (tirets) sets off an incise as commas or parentheses would; in the printed text they are the long dash with spaces, in a hurried text a hyphen with spaces. A single dash at the beginning of a line opens a turn of dialogue (§4). The slash is used for km/h, et/ou (a calque on and/or that stylists dislike), dates (22/10/2026) and line breaks in quoted verse (Le vent se lève / il faut tenter de vivre), in which it is spaced.

4. Quotations and dialogue

4.1 The guillemets

The quotation marks of French are the guillemets « », with a non-breaking space inside each: Il a dit : « Je viens. » A quotation that contains a second quotation uses the English double quotes inside: « Le mot “liberté” est usé ». Further nesting uses guillemets again, or single quotes. Titles of articles and of short poems are put between guillemets; titles of books, films and journals are put in italics (§10). Scare quotes are put between guillemets.

The final point of the quoted sentence goes inside the guillemets only when the quotation is a full sentence of its own: Il a dit : « Je viens. » but Elle trouve « absurde » cette idée. A point that belongs to the framing sentence goes after the closing guillemet.

Counts. In GSD the guillemets are used (955 opening, 894 closing, 2026-11-06) beside the straight double quote " (956 tokens). The typographic English quotes “ ” do not occur in any French treebank we hold (0 in GSD, Sequoia, PUD, FQB, ParisStories, Rhapsodie, ALTS). GSD puts a space inside the guillemets in 952 of the 955 openings (the corpus text normalises the non-breaking space to a plain one), and the closing guillemet is preceded by a space in all 894 cases. Sequoia and FQB use only the straight double quote (497 and 362 tokens) and no guillemet. The straight quote is therefore as common in French web text as the guillemet and a tokenizer must handle both.

4.2 Dialogue

In novels dialogue is set with a dash (tiret) at the head of each speaker's turn, with or without opening guillemets at the very beginning:

— Tu viens ? — Non, dit-il. — Pourquoi ?

The incise of the narrator (dit-il, répondit Marie) stays inside the turn, after the quoted words, with the subject inverted and no closing mark between: « Je viens, dit-elle, mais pas tout de suite. » Within a turn the stage direction is set off by commas or dashes. The English custom of opening and closing quotation marks around every turn is not the French one. The transcriptions of spoken French in the open treebanks (ParisStories, Rhapsodie) do not use the dash at all; they have no dialogue markup, only utterances.

5. The apostrophe

The apostrophe marks elision, the dropping of a final vowel before a word that begins with a vowel or a mute h (04. Connected speech: enchaînement, liaison, elision, the mute e and assimilation §5): l'arbre, j'ai, qu'il, s'il, n'est, d'abord, c'est. The word aujourd'hui contains an apostrophe that is part of the word (it is one token in UD: 143 times in GSD, 2026-11-06; so are quelqu'un, 205 times, and presqu'île). No space follows the apostrophe, and none precedes it.

The typographic apostrophe ’ (U+2019) is standard in print; the typewriter ' (U+0027) is what keyboards give. In the French treebanks the apostrophe is always the typewriter one: in GSD 16 937 tokens contain ' and none contains ’ (2026-11-06); the same holds for Sequoia (3 270 against 0), PUD (1 131), FQB (1 047) and ParisStories (2 271). The corpora were normalised, so text taken from the printed or word-processed web (which has U+2019) will differ from them: a program trained on the treebanks sees a character it has never seen unless it maps the apostrophe first.

The apostrophe is also used in the written forms of colloquial speech (t'as vu, j'sais pas, y'a, p'tit, c't'à dire), which are not standard spelling but occur in dialogue and informal text.

6. The hyphen

The hyphen (trait d'union) is a mark inside a word, not between words. It joins:

The rectifications of 1990 hyphenate all compound numerals and solder many compounds (weekend, portemonnaie, and contrecourant as a recommendation) (cslf-1990 part II); they do not touch the hyphen of inversion. Writers disagree on a number of compounds (week-end/weekend, socio-culturel/socioculturel): the hyphen is the least stable part of French spelling.

Counts. In GSD the letter-hyphen-letter pattern occurs in 3 175 tokens (2026-11-06; for example États-Unis, vingt-deux, arc-en-ciel, peut-être). Tokens that start with a hyphen are 572 plus 72 isolated hyphens: they are the second element of an inverted pronoun, split off by the treebank rule (-t-il 59, -t-elle 18, -là 13 and others). The treebank thus cuts a-t-il into the verb and a token written with a leading hyphen, -t-il, while it keeps peut-être and États-Unis whole.

7. Capital letters

French uses fewer capitals than English.

Accented capitals. Correct typography keeps the accents on capitals: État, À bientôt, École, ÉCOLE, Œuvre. Older printing and many web texts leave them off (Etat, A bientôt, Ecole), because the French keyboard has no direct key for them; the Académie française recommends keeping them (its position is known to us from its card only). In GSD, À occurs 209 times and the preposition written A (unaccented) 86 times (the latter when a capital is required at the start of a sentence): 29% of capital prepositions are written without the accent. Of the nouns État(s), École, Église, Économie, Écosse 235 are accented and 58 are not (2026-11-06). The consequence is the ambiguity A (verb) / À (preposition) in such text.

Months, days, nationalities. In GSD 2 091 tokens are the lower-case names of months and days and the adjectives français, anglais, allemand; the capitalised months and days that are not proper nouns are 30 (nearly all at the start of a sentence), and the capitalised names of nationalities (Français, Anglais, Allemands, Italiens, Espagnols and their feminines) are 83. The English habit of capitalising months and languages is therefore an error in French (it is the rule in English).

8. Abbreviations and acronyms

An abbreviation that stops before the last letter takes a point; one that keeps the last letter does not:

Acronyms (sigles) are written in capitals without points: SNCF, ONU, TVA, RER. When read as a word they may take only an initial capital: Unesco, Insee, Otan. Their plural is not marked in the standard (les ONG); the gender is that of the head noun (la SNCF, le CNRS).

Counts. GSD (2026-11-06): M. 51, etc. 29, av. 21, Me 14, n° 14, No 8, Mme 8, Dr 4, Mlle 1, St 1, Ste 1. Tokens of two or more capital letters (acronyms, titles in capitals) are 1 667 in GSD. The ellipsis after etc. does not occur (0 of 29).

9. Numbers, dates, units

9.1 Decimal mark and grouping

• Decimal comma: 3,14; 1,5 million; 0,5 %. In GSD 323 tokens have the shape digits-comma-digits and 58 digits-point-digits (the latter are mostly section numbers, versions and English-style decimals in loans).
• Thousands are grouped by a space, thin and non-breaking: 1 000 000, 25 785. The standard forbids the point and the comma as grouping (a decimal comma is a decimal comma). Four-digit numbers may be written without a space (1500, 2026). In GSD 175 tokens are space-grouped numbers (the treebank puts the space inside a single NUM token, as the English manual has in EWT), and 45 are point-grouped (1.000); the years (a NUM of four digits) are 4 123.
• Ranges are written with a dash or à: 1914-1918; GSD has 294 digit-hyphen-digit tokens.
• Fractions and percentages: un demi, 3/4, 10 % (with a space before the percent sign, which GSD writes as a separate token % 296 times).

9.2 Words or figures

Numbers up to sixteen (or to nine, by house style) and round numbers are written in words in running text: trois enfants, deux cents ans; larger ones in figures: 45 000 habitants, 1 200 mètres. A figure must not begin a sentence in a careful text. Ordinal numbers: 1er, 1re, 2e (not 2ème, 2nd, 2de, although 2ème is very common); centuries XIXe. In GSD 60 tokens are 1er, 226 are a Roman numeral plus e or ème (XXe, XIXe), 178 an Arabic numeral plus e or ème (2e, 3ème), and 20 of all these use the non-standard ème; the treebank has no superscript.

9.3 Dates and times

Dates: le 22 octobre 2026, le 1er mai; numerically 22/10/2026 (day first), also 22.10.2026 and 22-10-2026; ISO 2026-10-22 is used in computing. The day is not preceded by an article in a date line (Paris, le 22 octobre). Times: 15 h 30, 15h30 or 15:30; hours are l'heure, midi, minuit, 24-hour clock in writing (18 h). In GSD no token and no sentence text has the slashed date shape, and no token has the shape 15h30: dates in this corpus are written with the month in words (le 1er mai).

9.4 Currency, units and symbols

Units and currency follow the number with a space: 25 km, 10 %, 3 €, 12 °C, 5 m². The symbol precedes only in other languages ($5). The names of units are not capitalised and are not marked by an s in the plural when abbreviated (10 km, 10 kilomètres). In GSD the units km (76), m (80), kg (21), € (45), $ (13) and euros (66) occur as separate tokens after the number. The decimal comma stands before the decimals, and the unit follows: 1,5 kg.

10. Other conventions of the page

• Italics mark titles of books, films, periodicals, ships and works of art (Le Monde, Les Misérables), foreign words not yet assimilated (a priori, software), and emphasis. Underlining is not used in print.
• Foreign words keep their spelling but not their plural: the rectified spelling regularises borrowed plurals (des scénarios: 06 §6).
• Lists use a dash, a bullet or a letter; each item begins with a capital if it is a sentence, with a lower-case letter and a semicolon at the end if the list continues one sentence.
• Addresses and codes: the postal code before the town (75001 Paris); phone numbers in pairs of digits (01 23 45 67 89); the thousands space is not used there.
• URLs and e-mail: written without internal space. GSD has 8 tokens that look like URLs or addresses. A URL followed by a French sentence-final mark is the case that confuses tokenizers (the final point or ? is not part of the address).
• Inclusive writing (écriture inclusive): the middle dot or point médian joins the masculine and feminine endings (les étudiant·e·s), also written with a full stop, a hyphen or parentheses (étudiant.e.s, étudiant-e-s, étudiant(e)s). The Académie française has condemned it; official French administrative use forbids it; publishers vary. The research on the matter (flesch-2025) found in a corpus of non-binary authors' posts 96 non-binary grammatical-gender markers produced by 24 people (17.52% of the authors in the corpus), the large majority of them formed with a punctuation sign; another group of authors uses one gender consistently and another both. A rewriting system for gender-neutral French (doyen-2025) replaces masculine generic nouns with human collective nouns, which have a fixed gender (les lecteurs → le lectorat), rather than writing the dot. In the modern French treebanks the middle dot occurs in one GSD sentence, the form (e) in one more (2026-11-06; the many middle dots of the Old French treebanks are the medieval punctus): inclusive writing is practically absent from the annotated text.
• Line breaks and words: a word is broken at the end of the line at a syllable boundary with a hyphen; the apostrophe is never broken from its word (l'|ami is not allowed to break at the apostrophe in many manuals; some allow it).

11. The keyboard and the web

The standard French keyboard (AZERTY) gives no key for «, », É, À, Œ or for the non-breaking spaces; word processors insert some of these automatically. This explains the facts of the corpora: web text often has straight quotes, a plain space before the high marks (or none), capital letters without accents, and the typewriter apostrophe. On social media the high marks are often glued to the word. An annotated corpus (GSD, Sequoia) keeps the plain-space habit of the page.

12. How Mova handles punctuation and typography

Raw text enters through the English tokenizer. Mova has no French tokenizer. en/src/tok.rs is a finite automaton (a table of states by character classes) with dictionaries learned from the training treebank: abbreviations with a point, hyphenated words kept whole or split, words with an initial apostrophe, numeric suffixes (1990s). The file's description is the English one. Applied to French text it behaves as follows:

• Letters. Accented letters are alphabetic (char::is_alphabetic), so été is one word.
• Apostrophe. ' and ’ are the same class. Inside a word the automaton goes on: l'arbre, qu'il, jusqu'alors, c'est are one token each. The English clitic list (n't 's 're 've 'll 'd 'm) cuts nothing in these words. UD cuts them (l' + arbre), so every elision is a boundary error.
• Hyphen. A hyphen between letters stays inside the word, and the dictionary of the training treebank decides whether to split. For French GSD the dictionary contains 1 994 hyphenated words kept whole and 46 split, and the default is to keep. An inverted pronoun like a-t-il is therefore kept whole where the treebank splits it.
• Space and the non-breaking space. char::is_whitespace is true for U+00A0 and U+202F, so a number grouped with a non-breaking space (25 785) is cut into two tokens by this tokenizer, where UD keeps one; the plain space is cut the same way. The treebank's own groupings (175 tokens in GSD, 2026-11-06) are therefore always wrong for raw text.
• Guillemets. « » ¡ ¿ fall in the "other punctuation" class and are separate tokens; so are , ; : ! ? ( ). The space before ? is a separator and does no harm. This part agrees with UD.
• Final point. A word that ends in a point is kept if the dictionary knows it as an abbreviation (69 abbreviations learned from French GSD), otherwise the point is cut off.
• Contractions. au, aux, du, des are never split; UD splits them (à + le): GSD has 4 390 du, 3 298 des, 2 477 au and 808 aux (all cases together) as range lines, 11 022 range lines in all (2026-11-06).

Measured. en tok-eval with the dictionary learned from GSD train and tested on GSD dev, 2026-11-06 (binary built 2026-10-22; the run was repeated and gave the same numbers): of 1 476 dev sentences, 26.1% are tokenized exactly as the treebank has them (with the spaces: 26.1%); token precision 92.19%, recall 85.72%, F1 88.84%. On the test file (416 sentences) the exact rate is 21.9% and F1 88.16%. The errors printed by the tool are all of the types listed above: L' + oasis (we: L '), l'obscurité (kept whole), jusqu'alors (kept whole), aux (UD: à les), C'est (kept whole). This is the measured cost of reading French through the English tokenizer: about one in nine tokens is wrong, and nearly three sentences in four are not cut right. The measure is not a parse score; with gold tokens the parser's LAS is 87.00 (own tags) and 89.78 (gold tags) on GSD (PACKAGE.md, 2026-10-22).

Tags and the punctuation of the trees. en xpos-fill (en/src/main.rs, xpos_of) maps French punctuation to the English tag set. «, " and “ go to the opening-quote tag, » and ” to the closing-quote tag; , to the comma tag; . ? ! … ... to the sentence-final tag; ( [ and ) ] to bracket tags; - – — to the hyphen tag; every other punctuation sign (; : /) to the colon tag. The English parser therefore treats ; and : alike and does not distinguish the decimal comma (inside a number token it is not punctuation).

What a French tokenizer needs. Eleven rules cover almost all the differences above: (1) split a proclitic with an apostrophe from the following word (l', d', j', n', s', c', m', t', qu', jusqu', lorsqu', puisqu'), except for aujourd'hui, quelqu'un, presqu'île, prud'homme; (2) split the contracted articles and relatives au, aux, du, des, auquel, duquel… into two syntactic words, keeping the surface form as a range; (3) split the inverted pronoun after a hyphen (-t-il, -vous, -je) and -ce; (4) keep -ci and -même attached and split -là (as GSD does); (5) keep the hyphenated compounds and numerals whole, which needs a dictionary; (6) read U+00A0 and U+202F inside a number as part of it, with a space-grouped number as a single token; (7) read a decimal comma as part of the number; (8) keep abbreviations with their points and a list for French ones (the learned 69 is a start); (9) map U+2019 to ' before cutting; (10) treat « » as quotation marks whether or not spaces follow; (11) do not read a final point or question mark as part of a URL. None of this exists in code. The data to learn it from is the GSD range lines and the SpaceAfter columns of the French treebanks (CC BY-SA 4.0 for GSD, ParisStories, PUD, ALTS; LGPL-LR for Sequoia and FQB: see Sources).

Not handled at all. The 1990 rectified spelling of the hyphenated numerals is read as unknown words (06); inclusive writing (étudiant·e·s) is cut at the dot; the non-breaking space is not preserved on output, and nothing generates French typography (no space before ?, no guillemets) when French is written.

Sources for this chapter

• Paris: Imprimerie nationale. 2002. Lexique des règles typographiques en usage à l'Imprimerie nationale, 6th ed. Copyrighted; known by card only (paris-2002-lexique-des-regles-typographiques). No fact is attributed to it.
• Conseil supérieur de la langue française. 1990. Les rectifications de l'orthographe, Journal officiel, 6 December 1990 (cslf-1990). Official text. Library: books/c/cs/cslf-1990-les-rectifications-de-lorthographe.
• Flesch, Marie & Heather Burnett. 2025. Characterizing non-binary French: a first step towards debiasing gender inference (flesch-2025). CC BY 4.0. aclanthology.org/2025.gebnlp-1.16
• Doyen, Enzo & Amalia Todirascu. 2025. GeNRe: a French gender-neutral rewriting system using collective nouns (doyen-2025). CC BY 4.0. aclanthology.org/2025.findings-acl.411
• Bruno Guillaume, Marie-Catherine de Marneffe & Guy Perrier. 2019. Conversion et améliorations de corpus du français annotés en Universal Dependencies (guillaume-2019). CC BY 4.0. The tokenization conventions of GSD.
• UD French GSD, Sequoia, PUD, FQB, ParisStories, Rhapsodie, ALTS, 2.18 (counts of 2026-11-06 by awk on the CoNLL-U files; see Sources for the licences).
• Mova: en/src/tok.rs, en/src/main.rs (xpos_of, tok_eval); en tok-eval GSD train → dev and test, 2026-11-06.

This page in Markdown → https://md.mova.elamur.ai/free/en/docs/languages/fr/07-punctuation-typography/