🗣️ 10. Word classes

A word class (part of speech, partie du discours, classe de mots, catégorie grammaticale) is a set of words that share inflection and distribution: they take the same endings and occur in the same positions. French has far more inflection than English, so a good part of its classes can be told by form: nouns have a fixed gender and a number, adjectives and determiners agree, a regular verb has forty-five simple finite forms. But the written form hides much of it (the plural of the noun is silent, the feminine of many adjectives is silent, see 11. Nouns §1), and a large share of the common words belongs to more than one class (que, le, en, tout, plus, comme, si). This chapter lists the classes as the traditional grammars and as Universal Dependencies define them, gives the tests that tell them apart, counts them on the French treebanks, sets out the boundaries that are hard to draw, and says how Mova's tagger decides. The chapters that follow describe each class in turn: nouns (11), determiners, pronouns and numerals (12), verbs (13, 14), adjectives and adverbs (15).

1. Variable and invariable words; open and closed classes

The French school tradition, already in the nineteenth-century grammars (bescherelle-1860), divides the words first into mots variables (those that change form: noun, determiner, adjective, pronoun, verb) and mots invariables (adverb, preposition, conjunction, interjection). The division is useful because it names what a learner must memorise, but it is not exact. Adverbs such as tout ("quite") agree in some uses (elle est toute contente); many pronouns are variable for gender and number (celui, celle, ceux, celles) while others are not (on, rien, qui); numerals vary only in some words (quatre-vingts, deux cents).

Another division is by openness.

In UD French GSD (UD 2.18, all three files, 400 385 words after the splitting of contractions, counted 2026-11-06) the closed classes DET, ADP, PRON, AUX, CCONJ and SCONJ together make 169 026 tokens, 42.2% of the words; the single most frequent lemma is the article le (43 301 tokens as a determiner), followed by the preposition de (31 302). The open classes NOUN, PROPN, VERB, ADJ, ADV total 173 113, 43.2%; punctuation is 11.0%.

2. The traditional classes and the UD classes

2.1 The traditional list

The French grammars of the nineteenth century follow the classical tradition and count ten "kinds of words" (espèces de mots); the Grammaire nationale of the Bescherelles lists them in this order (bescherelle-1860, in its opening chapter on the kinds of words): the noun (nom), the article, the adjective, the pronoun, the verb, the participle, the adverb, the preposition, the conjunction and the interjection. The participle was a class of its own because it is "verb and adjective". The school grammar of the late twentieth century reduces the list to nine classes by dropping the participle (it is a verb form, or an adjective) and replacing the article by the larger class of the déterminant (article, demonstrative, possessive, indefinite, numeral, interrogative and exclamative determiners). The modern scholarly grammars follow a similar list, with finer subclasses (the Grammaire méthodique of Riegel, Pellat and Rioul, riegel-2021, which we know by its card only).

2.2 The UD list

Universal Dependencies fixes seventeen universal parts of speech (UPOS) for all languages, and each language says how it uses them (demarneffe-2021; guillaume-2019 describes the French case). GSD uses sixteen of them (there is no PART). The table gives the mapping and the count in UD French GSD (UD 2.18, 2026-11-06).

The same mapping applies to the other open French treebanks, with different proportions that show the genre. The spoken ones (ParisStories, 42 789 words; Rhapsodie, 44 234 words; counted 2026-11-06) have a much richer function-word and discourse profile than the written GSD.

Where the written French has the nominal profile (nouns, determiners, prepositions: 50% of the words), speech has three times the pronouns, a seventh of the proper names and an interjection class that is a twentieth of all words (ben, bah, euh, ouais, hein, bon). A tagger trained on GSD alone has seen 87 interjections; speech needs its own treatment, in 37. Register and variation and 38. Dialogue, discourse markers and reported speech.

2.3 Features

UD adds features to the part of speech, and French uses the usual ones. In GSD, of the 75 330 NOUN tokens 30 689 are Gender=Masc|Number=Sing, 25 654 Fem|Sing, 10 961 Masc|Plur, 7 574 Fem|Plur; 4 + 17 + 134 + 92 + 205 tokens (the epicene, the unmarked and the foreign) lack one or both. Of the 23 820 ADJ, 12 726 are masculine and 11 094 feminine. The feature PronType has the values Prs (personal, 11 354 pronoun tokens), Rel (3 103), Dem (1 633), Ind (1 416), Neg (114), Int (67). The old tag set in the Penn style (the XPOS column) is empty in GSD (all 400 385 tokens carry _), so Mova must make its own tags from the UPOS and the features (§9).

3. The classes one by one

3.1 Nouns

Form: a fixed gender; a plural, silent in most words, with the exceptions of 11 §5. Distribution: head of a noun phrase; preceded by a determiner (a bare noun is rare: chercher fortune, avoir faim, par avion); modified by an adjective, a prepositional phrase (la maison de ma sœur), a relative clause. Subclasses: common/proper; count/mass (des pommes / du pain); animate/inanimate; the epicene nouns with one form for both sexes (un/une élève, un/une enfant, un/une collègue) and the nouns with one gender for both sexes (une personne, une victime, un témoin); the count/mass contrast is in 11. The tests that separate the nouns from other classes are: takes the definite article without changing sense (le manger and le boire are nominalised infinitives, see §4.5); takes a plural; can be the subject of a verb.

3.2 Determiners

The determiner is the class that the nineteenth-century grammars did not have as such. Its members are closed and each carries gender and number agreement with the noun (plusieurs has a plural only and chaque a singular only; both are invariable in gender): the definite le/la/l'/les, the indefinite un/une/des, the partitive du/de la/de l'/des, the demonstrative ce/cet/cette/ces, the possessive mon, ma, mes, ton, son, notre, votre, leur…, the interrogative-exclamative quel, the indefinites plusieurs, chaque, certain, aucun, nul, tout, quelques, différents, divers, the numerals when used before a noun (trois livres). Test: it stands before a noun and excludes another determiner (*le mon livre; Italian and Portuguese permit it). In GSD DET is 61 098 tokens, of which the lemma le is 43 301, un 10 065, son 4 348, ce 2 214, plusieurs 252, quelques 150, tout 136, chaque 133, certain 132, aucun 108 (UD lemmas, 2026-11-06). Details in 12.

3.3 Pronouns

The pronoun replaces a noun phrase and never takes a determiner of its own: it is a full noun phrase (il, lui, celui-ci, le mien, qui, rien, quelqu'un). French splits the personal pronouns into weak or clitic forms, which cannot stand alone and attach to the verb (je, tu, il, me, te, le, lui, y, en…), and strong or disjoint forms (moi, toi, lui, eux, soi), which can stand alone (Qui est là ? — Moi.). The clitics are unstressed, ordered in a fixed cluster before the verb (12 §9) and are written as separate words in the traditional spelling (je le lui dis), joined to the verb by a hyphen only in the imperative and in inversion (dis-le-lui). Subclasses by UD feature PronType: Prs personal and reflexive (11 354 in GSD, 64% of the class), Rel (qui, que, dont, lequel, 3 103), Dem (ce, celui, ceci, cela, 1 633), Ind (on, rien, personne, tout, plusieurs, 1 416), Neg (114), Int (67). A trap for counting: GSD gives the third-person pronouns one lemma, lui, whatever the form (il, elle, ils, elles, le, la, les, lui, leur, eux: 6 631 PRON tokens), as it gives moi to je, me, moi and soi to se; a count "by lemma" of the personal pronouns is therefore a count of persons, and the forms must be counted by form (12).

The three forms with several values deserve a mention. Qui and que are pronouns, relative or interrogative, in GSD; dont is a pronoun (283 tokens) and où an adverb (402). On ("one", "we") is a pronoun with no gender and no number of its own (the verb is singular); its value "we" is a register matter (37. Register and variation). Y and en (536 and 282 tokens as pronouns in GSD) stand for a prepositional phrase (j'y vais, j'en viens); they are the "pro-adverbs" of the older grammars and are PRON in UD because they can stand for a noun phrase.

3.4 Verbs and auxiliaries

Form: tense, mood, person and number (the "finite" forms, with the person mostly clear from the subject clitic alone in speech), and the non-finite forms: infinitive, present participle (chantant, the gérondif with en), past participle (chanté, which agrees in gender and number in several constructions). Distribution: the predicate of the clause; takes a subject (even the impersonal il); negated by ne … pas around the finite verb. UD: all verb forms are VERB, but the three verbs that form compound tenses and the passive are AUX: être (9 460 in GSD), avoir (3 340), and faire (271) in the causative and the sense "have done". The list is short by decision: the earlier GSD releases also counted modal and aspectual auxiliaries and some verbs of state as copulas, and the revision reduced the auxiliaries to être, avoir, faire and the copula to être (guillaume-2019, §4.2.1). Aller is therefore VERB even in the periphrastic future (il va pleuvoir): the lemma has 261 VERB tokens and 3 NOUN (un aller, "outward journey"). Pouvoir, devoir, vouloir, savoir are VERB too (modals are not a UD class). The periphrases and the modal-like verbs are described in 13. Verbs and 14. Irregular verbs. In GSD the participles are VERB (12 676 tokens with VerbForm=Part) when they head a clause or form a tense, and ADJ when they are used as an attributive adjective (une porte fermée, see §4.4).

3.5 Adjectives

Form: agrees in gender and number with the noun it qualifies (grand, grande, grands, grandes), with a silent or audible feminine. Distribution: modifier of a noun, after it by default (un livre intéressant), before it for a short closed list of frequent adjectives (un petit livre, une belle maison); predicate after être, sembler, devenir; modified by très, plus, assez, trop. Subclasses: qualifying (grand, rouge), relational, derived from a noun (présidentiel, solaire) and so not gradable, ordinal numerals (premier, deuxième), the adjectives of nationality (français; the noun Français has a capital). Details of agreement, position and comparison in 15. In GSD ADJ is 5.9%: premier, nouveau, public, national, grand, many of them in the plural.

3.6 Adverbs

The adverb is invariable and modifies a verb, an adjective, another adverb or a whole clause. It is the least homogeneous of the classes. Subclasses: manner (vite, bien, doucement); degree (très, trop, assez, plus); time (hier, déjà, toujours); place (ici, là, partout); negation (ne, pas, jamais, plus, guère); interrogation (comment, pourquoi, quand, où, combien); sentence adverbs (heureusement, peut-être, apparemment); connectives (donc, ainsi, pourtant, alors, d'ailleurs). The suffix -ment makes manner adverbs from adjectives (15 §10, 17). In GSD the most frequent ADV lemmas are ne (1 636), plus (1 339), pas (1 037), très (508), aussi (409), où (402), alors (376), ainsi (376), bien (349), également (321); ne and pas alone are over a sixth of all adverbs, because the two-part negation puts both in the class (27. Negation). Ne is tagged ADV with Polarity=Neg. Its frequency is a register marker: GSD has 1 636 ne against 1 037 pas (ne also pairs with plus, jamais, rien), ParisStories has 78 ne against 416 pas, Rhapsodie 220 against 374 (counted 2026-11-06): in spoken French ne is mostly dropped.

3.7 Prepositions

Prepositions introduce a noun phrase, an infinitive or a pronoun and form with it a phrase that is an adjunct or a complement: à, de, en, dans, pour, par, sur, avec, sans, sous, chez, entre, vers, depuis, pendant, selon. Two of them, de and à, are the most grammaticalised: de marks the genitive relation, the partitive, the agent, the infinitive complement (essayer de partir), and à the indirect object and the infinitive (apprendre à nager). In GSD, de alone is 31 302 of the 63 543 ADP tokens (49%), à 10 630 (17%), en 5 860, dans 3 137, pour 2 531, par 2 467. Complex prepositions are fixed expressions with the same function (à cause de, grâce à, auprès de, en face de, afin de, par rapport à); UD treats them as multiword expressions with fixed. Contractions with the definite article: à + le = au, à + les = aux, de + le = du, de + les = des; in GSD these are multiword tokens (11 022 of them, §7) split into ADP + DET: du → de + le.

3.8 Conjunctions

Coordinating (CCONJ): et, ou, mais, car, donc, or, ni, and the correlatives (soit … soit, ni … ni). The classical mnemonic "mais, ou, et, donc, or, ni, car" lists the seven. In GSD: et 8 132, ou 890, mais 833, puis 266, car 127. Subordinating (SCONJ): que (the complementizer, 2 260 by lemma if counted with the contracted qu'), si, comme, lorsque, quand, puisque, parce que, bien que, quoique, afin que, pour que, avant que, with their compound forms in que. A subordinator may bring a mood choice: bien que and pour que require the subjunctive (24. Coordination and subordination). Comme is a conjunction (comme il pleut), a preposition (comme un chien), an adverb of exclamation (comme c'est beau !); in GSD its 530 ADP tokens exceed its SCONJ ones (161).

3.9 Interjections and discourse words

Interjections are sentence-like words that do not enter the syntax: exclamations (ah, oh, hélas, zut, bravo), greetings and politeness words (salut, merci, bonjour, pardon), onomatopoeia (boum, ouf, aïe), and in speech the discourse markers ben, bah, bon, quoi, hein, voilà, tiens, allez, écoute. GSD has 87; the lemmas are merci 18, bref 13, oui 10, attention 9, bravo 7, dommage 6. ParisStories has 2 391 and Rhapsodie 1 910 INTJ tokens, 5.6% and 4.3% of their words; the discourse-marker side belongs to the pragmatics chapters.

3.10 Numerals

The numeral is not a class of its own in the traditional French list but a subset of determiners, adjectives, nouns and pronouns. UD gives it a part of speech NUM for the cardinal numbers in any use, determiner-like or not (deux 636, trois 243, quatre 125; the digits 2, 3, 5 and years like 2010: 159, 134, 131, 123 in GSD). The word un is NUM only in the counting sense (60 tokens), DET as the article (4 153) and PRON in l'un … l'autre (187). Ordinal numerals are ADJ (premier, deuxième, dernier). The number system itself is in 12 §7.

4. Where the boundaries are hard

4.1 Determiner or pronoun: le, la, les, un, leur, tout

The forms le, la, les are the definite article before a noun and the object clitic pronoun before a verb. In GSD: le is DET 15 425, PRON 283; la DET 10 956, PRON 109; les DET 9 738, PRON 129 (counts of the form, lowercased; 2026-11-06). The position is a good test (the article before a noun or its pre-nominal modifiers, the pronoun before a finite verb or infinitive) but not a perfect one: the pronoun follows the verb in the affirmative imperative (prends-le), stands before the presentatives voici, voilà (les voici), and the article stands before a nominalised infinitive (le savoir). Leur is a possessive determiner (leur maison, 455 in GSD, leurs 211) and a pronoun in je leur dis (51); the determiner takes the plural of the possessed thing (leurs maisons), the pronoun is invariable. The form un is the indefinite article (4 153), the pronoun in l'un des deux (187), the numeral (60). The form tout is an adjective (86), an adverb (140), a pronoun (112) and a determiner (48). GSD's choices for the whole lemma (all four forms) show the convention: the predeterminer of tous ces enfants, toute la journée is ADJ with the relation amod (455 of the lemma's tokens are ADJ amod); the word before a noun without an article is DET (tout homme, toute personne, 136); the pronoun is tout est perdu, ils sont tous là (112); the adverb is tout contents, tout à fait (143).

4.2 Que in four values

Que is the most multifunctional word in French. In GSD (form que, 1 962 tokens before elision): SCONJ 1 578 (je pense que…), PRON 231 (relative le livre que j'ai lu; interrogative que veux-tu ?), ADV 153 (restrictive ne … que, exclamative que c'est beau !, comparative plus grand que). The comparative que is often annotated SCONJ as well; the treebank's choice is reported rather than the grammarian's. A learner reads the four from the context: after ne it is "only"; after a comparative it is "than"; after a noun phrase with a missing object it is the relative; otherwise a complementizer.

4.3 En and y: preposition or pronoun

En is a preposition (en France, en dormant, 5 860 tokens in GSD) and a pronoun (j'en veux, 282). The two do not overlap by position, since the pronoun stands before the verb or its auxiliary, but a single word form must be told apart by the tagger. Y is a pronoun only (536).

4.4 Participle: verb or adjective

The past participle belongs to both classes. In the passive and compound tenses it is part of the verb; in une porte fermée it is an adjective, and in la porte est fermée either (state or passive, see 13 §13). The test is whether the participle can be modified by très (très fermé: adjective; *très fermée par le gardien) or has an agent phrase in par (verb). A rough upper bound for how often the choice arises: 3 566 GSD ADJ tokens end in a past-participle ending (-é, -ée, -és, -ées, -i, -ie, -is, -u, -ue, -us), a count that also includes true adjectives such as joli and vrai (counted 2026-11-06). The present participle (chantant) is VERB; the adjectif verbal (une histoire amusante) is ADJ and agrees. The gérondif (en chantant) is a verb form with en. Compare the English manual's treatment of -ing forms.

4.5 Substantivised words

Any word can be used as a noun when it takes a determiner: the infinitive (le boire et le manger, le savoir), the adjective (le beau, les riches, un Français), the adverb (le pourquoi, le mieux), the preposition (le pour et le contre), a conjunction (les si et les mais), a whole quotation. The UD guidelines let a converted word keep its class when it is only used as a head (an adjective riche in les riches stays ADJ; GSD has 37 ADJ and 1 NOUN tokens of the lemma riche, counted 2026-11-06), and give NOUN to words whose noun use is lexicalised (le déjeuner, le dîner, le pouvoir, le devoir). Where exactly the line falls in the treebank is the annotators' choice and is not uniform.

4.6 Adverb or preposition or adjective

Many words work as preposition and adverb, depending on whether they have a complement: avant, après, depuis, derrière, devant, dessus, dessous, autour, contre, pour. Il est arrivé avant (adverb), avant la fin (preposition). The adjectival adverbs (parler fort, chanter juste, coûter cher, voir clair, travailler dur) are invariable and keep the adjective's form; they are ADV in the treebanks. Debout and ensemble are adverbs; même is an adverb (même lui), an adjective (le même jour) and a pronoun (le même); tel is an adjective; autre an adjective or a pronoun; plus an adverb in all but the noun use (un plus, 2 tokens in GSD).

4.7 Invariable adjectives and nouns used as adjectives

Colour nouns as adjectives (orange, marron, cerise, citron) are invariable (des yeux marron), unlike rose, mauve, pourpre, écarlate, which agree. Nouns as adjectives (une jupe bleu marine, une personne cool, des tarifs maison) are ADJ if they can be modified (très cool), NOUN if they behave as appositions. The question is largely the one of the English manual on noun modifiers; French writes them without a hyphen and invariable.

4.8 Proper nouns

PROPN is 6.9% of GSD words (27 693) and in Wikipedia-like texts names run in sequences linked by flat:name: 7 005 GSD tokens have that relation. A name may contain a common noun (la Banque de France), a determiner, a preposition (Charles de Gaulle), and still be one name; the UPOS of the inner words follows their own class when the name is transparent (Banque NOUN) and PROPN when it is not (Gaulle). The capital letter is the writing marker, with the caveats of 07. Punctuation, typography, capitals, abbreviations and numbers.

5. The tag sets of the French treebanks

5.1 UD

The nine UD French treebanks that Mova holds are annotated in UD 2.18 with the UPOS set and a common feature set. The conversion from the older annotation scheme into UD is described by guillaume-2019: of the roughly two hundred rewriting rules that transform the Sequoia and French Treebank format into UD, a large part are simple changes of part-of-speech label (the paper's own example: N becomes NOUN) and of the names of features (g=m becomes Gender=Masc). The authors also note that the conversion treats the copula and the dependent relations in the order that makes the rule set simplest. The features in use in GSD (tokens carrying each, counted 2026-11-06) are Number 223 951, Gender 168 176, PronType 79 368, Definite 53 470, VerbForm 44 816, Person 39 299, Tense and Mood 24 877 each, Voice 11 160, Emph 10 298 (the strong-pronoun and -même marker), Case 8 065, ExtPos 4 583 (the class of a word inside a fixed expression), Poss 4 359, Polarity 3 240, Reflex 2 490, Foreign 1 946, Typo 936, NumType 216. The annotation of GSD has been repeatedly corrected since its first release, in a process that fixes similar errors in similar contexts everywhere at once (guillaume-2019).

5.2 The French Treebank (FTB) scheme

The first big treebank of French, the French Treebank of Abeillé, Clément and Toussenel (2003), annotated the newspaper Le Monde with a tag set that is often compared with the Penn Treebank's. The corpus has eleven lexical categories, plus ET (foreign word) and PONCT (punctuation); the categories carry subcategories (conjunction: subordinating or coordinating; noun: common or proper) and give 41 tags, and 218 once the inflectional information is added, against 36 tags of the Penn Treebank (abeille-2019). The tagger used in the original annotation worked with a reduced set of 103 tags, estimated at an 8% error rate at the time. Two choices differ from the UD ones: weak pronouns are a separate category of clitics (CL) and the strong pronouns another one (PRO); and the words au and du are not split (the corpus keeps one token). The corpus also records the internal categories of compound words (bien que is a compound subordinator whose parts are an adverb and a conjunction), and in the author's retrospective the numerals ought to have had one tag instead of four (determiner, adjective, noun, pronoun). The impersonal il and the passive past participle had no separate label in the first version; later versions added them (the paper cites Ribeyre et al. 2014).

The exact inventory of the 41 tags is in the corpus guide (Abeillé and Clément 1999), which we have not read; the statistical parsing experiments on the FTB use reduced variants of it.

5.3 What is common, what differs

5.4 Quality of the tagging

On edited written French the taggers are accurate enough that the evaluation itself becomes the question: Wisniewski (wisniewski-2019) shows on French part-of-speech tagging that much smaller test sets than the usual ones give the same ranking of systems, and that the examples saved can go into training, which helps most in domain adaptation. The FTB's own retrospective measured the agreement of a new expert with the corpus at κ = 0.99 on the main categories and 0.97 on the full tags; the disagreements were the ones of §4: adjective or past participle, adverb or preposition, the indefinites l'un, l'autre as determiner, adjective or pronoun, and de as preposition or determiner (abeille-2019, §4.1). Text outside the norm is another matter. On a small corpus of administrative and political texts in inclusive French (177 inclusive tokens), TreeTagger and spaCy made 3 to 7 points more tagging errors than on the same texts rewritten in standard French; the median-dot forms (doctorant·e·s) were the worst, mistagged on 8 and 12 of 17 tokens, while the coordination of both forms (toutes et tous) caused no extra errors, and TreeTagger gave no lemma at all for the new forms (grouin-2022). The forms are in 11.

6. Word classes and word formation

The class of a derived word is given by its suffix (17): -tion, -té, -ité, -ment, -age, -eur, -isme make nouns; -able, -ible, -eux, -al, -if, -ique, -aire make adjectives; -iser, -ifier, -er make verbs; -ment makes adverbs. The prefixes keep the class (16). Conversion (a word moved from one class to another without a mark: le boire, un rien, le bleu, un plus) is in 18. The derivational database VerNom (missud-2020) gives verb–noun pairs acquired from a very large corpus; the 2019 families-and-paradigms study of French derivation (sanacore-2019) shows the semantic relations between members of a derivational family.

7. Multiword tokens, compounds and words that are not words

The tokenisation determines the classes: the UD treebanks segment the text into syntactic words, not graphic words. The graphic word du is two syntactic words (de + le); au is à + le; des is de + les (preposition + article) but also the plain indefinite plural article des, which stays one token (DET); duquel is de + lequel. Elided forms are separate tokens after the apostrophe split: l'ami → l' ami; qu'il → qu' il; aujourd'hui is one word. In inversion the verb and the clitic are split at the first hyphen and the hyphen stays on the clitic: a-t-il is a + -t-il (PRON, 59 tokens in GSD), est-ce is est + -ce (27), voulez-vous is voulez + -vous; lexicalised hyphenated words stay one token (peut-être ADV 24, aujourd'hui ADV 142). In GSD there are 11 022 multiword-token ranges (counted 2026-11-06, lowercased: du 4 390, des 3 298, au 2 477, aux 808, and 49 of auquel, duquel, auxquels, desquelles, auxquelles), 2.75% of the number of words; every one of them is a preposition fused with the article.

8. Foreign words, symbols, abbreviations

GSD marks 2 957 tokens X (0.7%): mostly English words in names and titles (of 83, The 53, the 37, and 27, in 22), the placeholder _, etc. (28) and some French fragments. Symbols are SYM (718): the percent sign (296), the slash (123), the euro (45), the degree sign (33), the ampersand (26). Abbreviations keep the class of the full word (M. is PROPN or NOUN, km NOUN); units are NOUN (h 89, km 76, kg 22 tokens of the form). In the written text the elements of an acronym are one token (SNCF, PROPN). The abbreviations and typography are in 07. Punctuation, typography, capitals, abbreviations and numbers.

9. How Mova handles word classes

There is no French tagger in Mova. The English parser en reads a French treebank after its tags have been brought to the Penn-style tag set it knows, by the command en xpos-fill IN.conllu OUT.conllu (function xpos_fill and xpos_of in en/src/main.rs). It fills the XPOS column of every token whose XPOS is _ or is not an English tag (which is the case for all of GSD) with a tag computed from UPOS and FEATS alone:

The mapping drops the French distinctions that English does not draw: gender, the three persons (the English VBZ/VBP contrast only separates the third person singular from the rest), the mood (the subjunctive, the conditional and the imperative become VBZ/VBP like the present indicative; the passé simple and the imperfect both become VBD), the clitic/strong contrast, the partitive. One branch does not do what its name says on GSD: the treebank keeps the tense of a participle in the MISC column (Tense[denom]=Pres or Past), not in FEATS, and has no VerbForm=Ger, so neither VBG branch fires and all 13 769 participles of GSD (VERB and AUX), the 1 705 present participles (Tense[denom]=Pres) included, become VBN (counted 2026-11-06; see 13). What survives is the contrast infinitive / participle / finite verb, with the finite verb split by VBD and by person, which is enough for the parser to find the verbal heads and their auxiliaries.

Measured. On UD French GSD (UD 2.18, train → dev, 2026-10-22) en parse-eval gives LAS 87.00 with the parser's own tags and 89.78 with the treebank's gold tags; the difference of 2.78 points is what the tagger's own errors cost. The tagger's accuracy on unknown words is 84.8%; 10.07% of the dev NOUN tokens have a form not in the training noun forms (684 of 6 792), and 8.48% of the non-punctuation dev words (2 706 of 31 919).

Not done.

• Gender, the French verb morphology, the clitic class, the choice between homographs from French lexical knowledge: none of this is in the tagger; it learns from the context in the training treebank, so a form that is ambiguous in GSD (que, le, en, tout; of the 42 626 distinct lowercased forms of GSD, 3 107 occur with more than one UPOS, and the tokens of those forms are 52.7% of the corpus, counted 2026-11-06) is decided from the left and right neighbours.
• There is no French lexicon: no list of determiners, no list of auxiliaries, no lemmas other than those in the treebank.
• The classes of the closed lists in this chapter (the determiners, the pronouns, the prepositions, the subordinators) could be derived from the treebanks as a first closed lexicon, since the closed classes are covered almost completely by a few hundred lemmas; no tool does it today.
• The FTB tag set (the clitic and strong-pronoun categories, the unsplit contractions) is not read.

Possible next steps that this chapter suggests: a table of homograph pairs (§4) with the disambiguating context, to be used as seeds in the way the English manual's seeds are used; a check of the UPOS+FEATS combination against the closed lists above (a noun with Gender missing, a DET with PronType=Rel); a spoken-French check of the INTJ and discourse-marker class, since an English-trained tagger has a very different interjection profile.

Sources for this chapter

• Bescherelle, L.-N. et al., Grammaire nationale (1860 ed., bescherelle-1860, public domain): the ten kinds of words, variable and invariable.
• Abeillé, A., Clément, L. & Toussenel, F. 2003, French Treebank; the retrospective presentation: Abeillé, A., Clément, L. & Liégeois, L. 2019, Un corpus arboré pour le français : le French Treebank (abeille-2019, CC BY 4.0): categories, tag counts, clitics, compound words.
• Guillaume, B., de Marneffe, M.-C. & Perrier, G. 2019, Conversion et améliorations de corpus du français annotés en Universal Dependencies (guillaume-2019, CC BY 4.0): conversion rules, the 17 UPOS, the copula rule.
• de Marneffe, M.-C., Manning, C., Nivre, J. & Zeman, D. 2021, Universal Dependencies (demarneffe-2021, CC BY 4.0).
• Wisniewski, G. 2019, Combien d'exemples de tests sont-ils nécessaires à une évaluation fiable ? (wisniewski-2019, CC BY 4.0).
• Grouin, C. 2022, Impact du français inclusif sur les outils du TAL (grouin-2022, CC BY 4.0).
• Missud et al. 2020, VerNom (missud-2020); Sanacore et al. 2019 (sanacore-2019).
• Counts: UD 2.18 French GSD (fr_gsd-ud-{train,dev,test}.conllu), ParisStories, Rhapsodie, Sequoia, ud-treebanks-v2.18, counted 2026-11-06 with awk.
• Mova: en/src/main.rs (xpos_fill, xpos_of), baselines in What every language gets.

This page in Markdown → https://md.mova.elamur.ai/free/en/docs/languages/fr/10-word-classes/