🗣️ 20. Syntax overview

French builds a clause around a finite verb that almost always has an expressed subject, in the order subject – verb – object, with the pronoun objects and the negative particle ne packed in front of the verb as clitics, with most adjectives after the noun, and with prepositions and articles in front of every noun phrase. Case is not marked on nouns, so the order and the little words carry the grammar. Around this core, the spoken language uses a second system that moves phrases out of the clause (dislocation, clefts) and fills it with markers of speech. This chapter is the entry to the syntax part of the manual. It gives a map of the part, a one-page account of the French clause, the way a French sentence is written down as a Universal Dependencies (UD) tree, the French relation inventory with counts, the places where the French treebanks disagree, and what Mova's parser does with French today.

Contents:

1. The chapters of the syntax part
2. The French clause in one page
3. Dependency grammar and the CoNLL-U record
4. Word classes in brief
5. The French relation inventory
6. French subtypes
7. Design decisions that shape every French tree
8. The treebanks and how they differ
9. UD, SUD and the history of the GSD annotation
10. Order, projectivity and the clitics
11. Written and spoken French in the numbers
12. How Mova handles it
13. Sources

1. The chapters of the syntax part

The forms the syntax works on (pronouns, determiners, verbs, auxiliaries, participles) are in the morphology part; the meaning and use of the sentences in the pragmatics part. Each chapter of this part ends with a section How Mova handles it, which says what Mova's tools do with the point and where they stop. Mova's French is much smaller than Mova's English: there is no French tokenizer, lexicon or morphology, and the pragmatics reads English words. The sections say so, construction by construction.

2. The French clause in one page

The basic order. A declarative clause is subject – verb – object: Le chat mange la souris. The subject is nearly always expressed. French is not a pro-drop language: an empty subject is replaced by an empty il (Il pleut), and in the spoken language even a noun subject is usually repeated by a pronoun before the verb (Mon frère, il est parti). A noun subject comes after the verb only in a few set frames: in inserted speech (dit-il), after some adverbs, in relative clauses, and in the stylistic inversion of written narrative (chapter 29).

The verb group. The finite verb is preceded by a string of clitics in a fixed order — ne, subject pronoun, object pronouns, y, en — and followed by pas or another negative word. Compound tenses and the passive put an auxiliary before the participle; the negation surrounds the auxiliary (Je n'ai pas vu), and short adverbs sit between auxiliary and participle (Il a bien mangé). Modal and aspectual verbs take a bare infinitive (Il veut partir), others take à or de before it (Il commence à parler, Il décide de partir).

The noun phrase. Determiner – (short adjective) – noun – (adjective) – prepositional complement – relative clause. The determiner is obligatory for a singular count noun. Adjectives follow the noun unless they belong to a short list of frequent ones (grand, petit, bon, nouveau, premier …). A noun complement is introduced by a preposition (le livre de Paul); there is no stacking of nouns as in English.

Prepositions. The preposition is the head word of an oblique phrase in traditional analysis and a dependent in UD. À and de are the two most frequent words of the grammar after the article, and they merge with le and les into au, aux, du, des.

Subordination. A subordinate clause is introduced by que, by a relative pronoun, by si, or by a conjunction; the introducing word is never omitted. The mood of the subordinate verb depends on the governing word. A subject that is the same as the main subject is expressed by an infinitive instead of a finite clause (Je veux partir, not *Je veux que je parte).

Questions and negation. Questions are formed by intonation, by est-ce que, or by inversion of a subject pronoun; wh-words are placed in front or left in place. Negation has two parts around the verb, of which the first (ne) is mostly dropped in speech.

Information structure. Because a stressed word is not enough, French marks topic and focus by construction. A topic is placed outside the clause and taken up by a pronoun (Cette épreuve, tu la fais); a focus is isolated by a cleft (C'est Marie qui a pris le livre). A dislocation, a cleft or an il y a … qui presentative occurs in 14.6% of the sentences of the spoken ParisStories and 11.0% of Rhapsodie, against 1.8% in the written GSD (2026-11-06; chapter 29).

What French does not have. It has no case on nouns, no phrasal verbs, no dative shift (*Il donne sa sœur le livre), no stranded prepositions (*la femme que je parle à), no do-support and no free subject-verb inversion. None of this is a matter of tags: an English model, which expects an English order and English function words, cannot be used on French as it is, so Mova retrains its parser on French trees (§12).

3. Dependency grammar and the CoNLL-U record

UD represents a sentence as a tree of words. Each word depends on exactly one other word, its head, through a labelled relation, and one word, the root, depends on none. The unit is the syntactic word, which can differ from the written word. The tree is stored in the CoNLL-U format, one line per word with ten tab-separated columns: ID, FORM, LEMMA, UPOS, XPOS, FEATS, HEAD, DEPREL, DEPS, MISC.

Here is a sentence of UD French GSD with its record (the DEPS column is empty in every French treebank of the release, and XPOS is empty in all of them except PUD, ALTS and ParTUT, which carry a tag set of their own; the table leaves both out):

fr-ud-dev_00007 · UD French GSD Elle lutte pour échapper aux tueurs à ses trousses. she fights for escape-INF to-the killers at her heels — "She is fighting to escape the killers on her heels."

Four things in this record are typical of French.

Multiword tokens. The written words au, aux, du, des (and the rarer auquel, duquel, auxquels, desquelles) are each two syntactic words: a preposition and an article. Line 5-6 holds the written form, lines 5 and 6 the words. In UD French GSD (all three files, counted 2026-11-06) there are 11 022 multiword tokens: du 4 390, des 3 298, au 2 477, aux 808, and 49 others. Because des is also the plural indefinite article (des enfants), only the contracted des = de + les is split: the indefinite one stays a single DET.

Elision and hyphens. An elided word is a word of its own (l', qu', n', jusqu'), whose lemma is the full word (le, que, ne, jusque). In inversion the subject pronoun is cut off from its verb and keeps its hyphen (-t-il, -vous). In the written text there is no space between them; the tree has two words, with SpaceAfter=No on the first.

The lemma of personal pronouns. The lemma column of the GSD conflates the personal pronouns: il, elle and le all have the lemma lui (as in line 1 above, with the original form kept in wordform), je and me have moi, se has soi. Counts by lemma are therefore counts of groups. The forms are reliable.

Features. French nouns carry Gender and Number, adjectives and participles also Gender and Number (agreement), finite verbs Mood, Tense, Person, Number, articles Definite, PronType=Art, pronouns Case (nominative, accusative, dative: clitics only) and Emph (clitic or strong form). The MISC feature Exponence[Gender]=Inherent marks gender as a property of the noun; on agreeing words it is not marked. The Subject= attribute (§7) records how the subject of a non-finite verb is found.

4. Word classes in brief

UD uses 17 universal parts of speech. The figures are those of UD French GSD (400 385 words, counted 2026-11-06):

Three decisions of the French UD are worth knowing here. Auxiliaries are three verbs only: être and avoir in compound tenses and the passive, and faire in the causative (GSD changelog, version 2.1). Modals (pouvoir, devoir, vouloir) and the verbs of future and immediate past (aller, venir de) are full verbs with an xcomp. A copula is only être. Other linking verbs (devenir, rester, sembler, paraître) take an xcomp predicate. The article and the preposition are separate words, so a prepositional phrase has three layers: aux tueurs = case à + det les + the noun.

The spoken treebanks differ in one more point: in the written GSD the interjection class is almost empty (87 words), whereas the spoken Rhapsodie marks 207 sentences whose root is an interjection (oui, ouais, non, ah).

5. The French relation inventory

The UD relations fall into the core arguments (subject, object), the oblique and nominal dependents, the clausal dependents, the function words, and the special relations. The table gives, for each universal relation used, its function in French and the counts in UD French GSD (all three files, 400 385 words, 2026-11-06) and, per 10 000 words, in three other treebanks. Subtypes are folded into their universal relation in this table and listed in §6.

Values for aux, expl, flat, acl, obl and nsubj add the subtypes in §6.

These numbers show the shape of written French: a long noun phrase (det, case, nmod and amod are 44% of all relations), few clauses per sentence (24.5 words per sentence, and one clausal dependent — advcl, ccomp, acl, xcomp, csubj — for about every 22 words), and one root verb for three sentences in four. The spoken treebanks show the opposite: more subjects and more adverbs per word, almost no nmod, appos or flat:name, and a large layer of discourse, reparandum, expl:comp and dislocated:subj.

6. French subtypes

UD allows language-specific subtypes after a colon. French GSD uses the following; the counts are for the whole treebank (2026-11-06).

The ParisStories and Rhapsodie treebanks add dislocated:subj (302 and 203) and dislocated:obj (34 and 21), and parataxis:parenth and parataxis:obj for speech.

7. Design decisions that shape every French tree

The following decisions determine the shape of a French tree. They are those of UD in general, and where the French treebanks go their own way they are marked.

Content words are heads. The head of a clause is its lexical predicate; auxiliaries, the copula, prepositions, articles and subordinating conjunctions are leaves. Je n'ai pas compris: compris is the head; ai is aux:tense, n' and pas are advmod, Je is nsubj. Il est grand: grand is the head, est is cop. Dans la maison: maison is the head, dans is case, la is det.

Oblique phrases are split into argument and modifier. The French GSD distinguishes an obl:arg (the prepositional phrase is required by the verb or adjective: participer à un projet, parler de politique, naître à Paris) from an obl:mod (it is an adjunct: travailler à Paris, en 1995). The distinction was introduced in 2018 and applied by hand to preposition phrases with à, de, par, dans, avec first (GSD changelog); the rest of the treebank was converted by rule. The difference is a matter of valency and is treated in chapter 23. A nominal indirect object (donner un livre à Marie) is therefore obl:arg, never iobj.

The reflexive and expletive clitics are expl. The empty il (Il pleut, Il faut partir), the demonstrative ce of a cleft or of a copular sentence (C'est difficile), the se of a pronominal verb (Il se souvient), the se of a middle or passive (Cela se vend bien) and the y and en that are part of a verb idiom (il y a, s'en aller) are not arguments. They are expl with four subtypes (subj, pv, pass, comp), introduced in version 2.5 (2019).

Causative faire is an auxiliary and its infinitive is the head. In Pathé fait monter une usine (fr-ud-dev_00214) the head is monter; fait is aux:caus; usine is the obj of monter; and Pathé, the subject of fait, is attached to monter as nsubj:caus. The person made to act, when expressed, is obj:agent or iobj:agent (Il fait chanter les enfants, Il fait lire le livre aux enfants; chapter 23). Laisser and the verbs of perception are not auxiliaries: laisser filer le ballon has filer as an xcomp of laisser (fr-ud-dev_00051). The treebank is nearly consistent here: faire is aux:caus 250 times, and an infinitive is an xcomp of faire only 9 times (2026-11-06).

Subjects of non-finite verbs are recorded in MISC. French GSD has an attribute Subject= on the non-finite or participial verb that says where its subject comes from: SubjRaising (the subject of the governing verb), ObjRaising (its object), OblRaising (its prepositional object), Generic (an arbitrary subject), Instantiated (a subject expressed elsewhere in the sentence); ParisStories and Rhapsodie add NoRaising. It occurs 6 765 times in GSD, 3 892 of them SubjRaising (e.g. échapper in the table above), 343 ObjRaising and 251 OblRaising. The labels are the annotators' own; the values given here are our reading from examples. The attribute is the only record in the tree of what is called control and raising in chapter 23.

Multiword expressions. Frozen sequences are fixed (en effet, afin de, dès lors, c'est-à-dire), 3 243 words in GSD; light-verb constructions are obj:lvc (avoir lieu, mettre fin); idioms are marked in MISC (InIdiom, Idiom: 3 246 and 2 671 in GSD) without changing the tree.

Names and foreign phrases are flat. Marie Curie is flat:name with the first name as head; 7 005 in GSD, only 31 in the spoken ParisStories.

Ellipsis is orphan. In Il a le pitjantjatjara comme langue maternelle et l'anglais comme seconde langue (fr-ud-dev_00393) the second conjunct has no verb: anglais is promoted to conj and langue is its orphan. 208 in GSD; 201 under conj.

Non-projectivity. Crossing arcs are allowed. They are discussed in §10.

8. The treebanks and how they differ

The UD 2.18 release holds nine treebanks of French or of a regional language of France that the manual uses (counts from the files, 2026-11-06). They fall into the written ones (GSD, Sequoia, PUD, ParTUT, FQB), a treebank of sixteenth-century legal French (ALTS), a treebank of a regional language, Poitevin-Saintongeais (PoitevinDIVITAL, counted here only for completeness), and the spoken ones (ParisStories, Rhapsodie). The Old and Middle French treebanks of the release are not treated.

The disagreements are real and they matter for any tool trained on more than one treebank. We found these:

1. Subtypes. GSD, Sequoia, ParisStories and Rhapsodie subtype obl into obl:arg, obl:mod, obl:agent; PUD and ALTS do not (PUD has 1 422 plain obl, ALTS 5 521). ALTS has plain expl (623) and plain aux (977), keeping only aux:pass (846); it has no obj:lvc, and its passive subject is a plain nsubj (there is no nsubj:pass in it).

2. C'est. In GSD the c' of c'est is nsubj 350 times and expl:subj 187 times; in ParisStories it is nsubj 607 times and expl:subj only 33. The second treebank treats c'est as an ordinary copular clause with a demonstrative subject, the first reserves expl:subj for the cleft frame. The consequence is that a model trained on both learns a blurred boundary between a subject and an expletive.

3. Negative words. GSD marks jamais with Polarity=Neg (86 times); ParisStories and Rhapsodie mark it with PronType=Neg (24 and 26 times) and give it no Polarity. Pas is Polarity=Neg in all four.

4. Questions. FQB analyses the noun phrase after Qu'est-ce que as dislocated (139 times in 2 289 sentences) where GSD makes the sentence a cleft. See chapter 26.

5. The root. 73% of GSD sentences have a verb as root (11 969 of 16 342), 67% in ParisStories (1 869 of 2 776), 59% in Rhapsodie (1 881 of 3 209): the remainder are nominal sentences, adjective sentences and, in speech, interjections.

6. Multiword tokens. All the treebanks split au, du, des; GSD and ParisStories split -t-il and -ce from the verb, ParisStories adds Liaison features to its words.

The advice is the same as in the other manuals: keep one treebank's conventions when training, convert before mixing. Adding the four open spoken and written training sets to GSD did not change the LAS on GSD (§12), which is some evidence of the conflict.

9. UD, SUD and the history of the GSD annotation

The French treebanks have been maintained next to a second annotation scheme, SUD (Surface-Syntactic Universal Dependencies). SUD is near-isomorphic to UD; the main difference is that it takes functional heads: the auxiliary heads the participle, the copula heads the predicate, the preposition heads the noun (gerdes-2018). In SUD Je n'ai pas compris has ai as head and compris as its comp:aux; in UD it is the opposite. UD French GSD is maintained in SUD and converted automatically to UD since version 2.6 (2020, Grew rules, GSD README). A tool that reads the UD files therefore reads converted trees, and conversion artefacts exist (such as the Subject= attribute, which is a SUD-side annotation exported into MISC).

The history of GSD in six lines, from its changelog:

• 1.0–1.4 (2015–2016): initial conversion of the Google treebank (a content-head conversion of the 2013 universal treebank); lemmas, features and the completion of truncated sentences (Wikipedia sentences had been cut).
• 2.0 (2017): conversion to the new guidelines; manual annotation of the nmod / obl distinction in ambiguous cases.
• 2.1 (2017): only être, avoir, faire are AUX; harmonisation across French treebanks of the causative, the copula and the auxiliary; possessive determiners become det.
• 2.2–2.3 (2018): obl split into obl:agent, obl:mod, obl:arg (done by hand for 43.5% of the cases at first); subtyping of à.
• 2.5 (2019): expl:subj, expl:comp, expl:pass; aux → aux:tense; a systematic review of non-projective trees; improved ellipsis.
• 2.6–2.7 (2020): maintained in SUD, converted.

The effort to bring the French treebanks closer to the UD guidelines and to make them internally consistent is described in the paper that accompanies these changes (guillaume-2019): a conversion with rules, validated and measured on both consistency and closeness to the guidelines. The scheme as a whole, its history and its principles are in (demarneffe-2021). For spoken French, where the question of what a "sentence" and what "punctuation" are is harder, the guidelines used by Rhapsodie and the SUD spoken treebanks are given by (kahane-2021), and the recent format of Rhapsodie that joins the syntactic tree with a prosodic layer is in (garcia-2025).

10. Order, projectivity and the clitics

A dependency tree is projective if no two arcs cross when drawn above the sentence. A projective tree always corresponds to a continuous phrase-structure; a crossing arc is a discontinuity. French is almost projective. In the UD files (2026-11-06, arcs to punctuation ignored) the share of sentences with at least one crossing is:

A study of non-projective dependencies in the French treebanks (perrier-2021) finds four linguistic sources that together explain 97% of the crossings: clitic climbing (the object clitic stands before the auxiliary or the governing verb, not before the verb of which it is an argument: Il en a pris trois with en depending on trois, or Il lui a permis de partir), deep extraction (a relative or interrogative pronoun separated from the verb that governs it), minimisation of dependency length (heavy phrases moved to the end), and pairs of distant dependent words (for instance ne … que). The same study shows that the choice of scheme changes the count: in SUD, where the auxiliary heads the participle, a clitic before the auxiliary crosses the arcs of the subject and the adverbs, and there are two and a half times as many non-projective nodes as in UD (perrier-2021). In UD the same clitic does not cross, since both depend on the participle.

In practice this means that a left-to-right shift-reduce parser, which builds projective trees, loses the clitic arcs of the 4% of GSD sentences that have a crossing, and the whole of a ParisStories sentence in about 6%. Mova's parser is of that kind (§12).

Clitic climbing in history. The order of the Modern French clitics (all before the finite verb, none before the infinitive of a restructuring verb) is a late state. Old and Middle French had obligatory climbing and interpolation between the clitic and the verb; climbing was lost before the beginning of the 19th century (olivier-2023). The modern order, in which every clitic is verb-adjacent and proclitic, is discussed in the pronouns chapter; here it matters that it is a reason for the crossings above.

11. Written and spoken French in the numbers

The spoken treebanks and the written ones are, for syntax, two samples of the same language with different frequencies. These ratios (2026-11-06) show it.

The spoken French clause has almost no noun subject (32 per 10 000 words against 263 in the written corpus: 3.5% of the plain nsubj in ParisStories, 12% in Rhapsodie, 55% in GSD) and is made of pronouns; it has few nominal modifiers and many adverbs and discourse markers; it uses il y a about thirteen times as often per word as the written text (212 in ParisStories, 3.7 per 10 000 words in GSD against 49.5). The tools of the written language need to be rebuilt for it. The tables of the next chapters repeat these two columns for each construction.

12. How Mova handles it

What is done. French is read by the English tree parser (en) after the treebank's tags have been mapped to the tag set the parser knows (en xpos-fill, in en/src/main.rs, function xpos_of). The mapping gives each word a Penn tag from its UPOS and FEATS: NOUN → NN/NNS, ADP/SCONJ → IN, a verb form by VerbForm and Tense, PRON → PRP (relative or interrogative WP), DET → DT (possessive PRP$, interrogative WDT), ADV → RB (interrogative WRB), CCONJ → CC. Mood, case, gender and the difference between a clitic and a strong form are lost in the tag. Mova's relation enum (en/src/gram.rs) has only the English subtypes: a French label that is not in it falls back to its universal relation. So expl:subj, expl:pv, expl:pass, expl:comp are all expl; obl:arg and obl:mod are both obl; aux:tense is aux; nsubj:caus is nsubj; advcl:cleft is advcl; flat:name is flat; while nsubj:pass, obl:agent, aux:pass and acl:relcl are kept. The parser therefore cannot tell an expletive subject from a reflexive clitic, or an argument from an adjunct. Wherever this manual says that the parser learns a French relation, it learns the universal one.

Measured. On UD French GSD (UD 2.18, train → dev, 2026-10-22) the baseline is LAS 87.00 with the parser's own tags and 89.78 with the treebank's gold tags (UAS 91.08 and 92.59); the tagger reaches 96.47% on the dev set, 84.8% on words unseen in training (7.2% of tokens). Training on the four other open training sets as well (Sequoia, ALTS, Rhapsodie, ParisStories) gave 87.02 and 89.83, no gain on the written dev set. Both numbers are on the gold tokenization.

The tokenizer is the limit. The en tokenizer is the English finite automaton, with a small dictionary learned from the training file (69 abbreviations, 1 994 hyphenated words kept whole, 5 apostrophe forms). On raw French text (en tok-eval, train on GSD train, test on GSD dev, 2026-11-06) it matches the treebank's tokens exactly in 26.1% of the 1 476 sentences (token precision 92.19, recall 85.72, F1 88.84): it keeps the elided clitic joined to the next word (l'obscurité, n'ai, qu'il, c'est, jusqu'alors, vas-tu) and does not split au, aux, du, des into preposition and article. Parsing raw French therefore starts from a wrong token sequence in three sentences out of four, and the 87.00 does not apply.

What the parser does with raw French. The results below are from en explain, which trains the parser on the GSD training file (12 epochs, beam 8, CPU, 2026-11-06) and parses sentences given as raw text through the English tokenizer. It is a trace of the model, not a score, and the sentences were chosen by us to show the constructions of this part; the same trace is used in the following chapters. Labels are those Mova can produce (the French subtypes are folded, as explained above). The first block is parsed correctly or nearly so, the second shows where it fails.

Writing the elision with a space does not help (qu' il gives the three tokens qu, ', il), so the parser cannot be fed treebank-style tokens through the English tokenizer. The same traces are used in the How Mova handles it sections of the following chapters.

What Mova does not do for French. No French tokenizer (clitic splitting, au/du/des, hyphenated pronouns); no morphology or lemmatizer; no French subtypes in the relation set; no enhanced graph (the French treebanks have none: the DEPS column is empty in all of them); no handling of non-projective arcs (the parser is a projective shift-reduce parser, §10); no French pragmatics or frames (the frames reader in prag/src/frames.rs takes negation, tense and modality from English words). The first steps would be a French tokenizer lexicon built from the multiword tokens of the treebanks (11 022 in GSD, each a two-word split), the relation subtypes of §6 added to the enum, and a conversion of GSD to the shapes the frames reader expects.

13. Sources

• Universal Dependencies guidelines and the French treebank READMEs (UD 2.18): UD_French-GSD (changelog), ParisStories, Rhapsodie, Sequoia, PUD, FQB, ALTS, PoitevinDIVITAL, ParTUT, in ud-treebanks-v2.18; counts by the author with awk, 2026-11-06.
• de Marneffe et al., Universal Dependencies (2021): demarneffe-2021.
• Guillaume, de Marneffe & Perrier, Conversion et améliorations de corpus du français annotés en Universal Dependencies (2019): guillaume-2019.
• Gerdes, Guillaume, Kahane & Perrier, SUD or Surface-Syntactic Universal Dependencies (2018): gerdes-2018; and the 2019 update, gerdes-2019.
• Perrier, Étude des dépendances syntaxiques non projectives en français (TAL, 2021): perrier-2021.
• Kahane et al., Annotation guidelines of UD and SUD treebanks for spoken corpora (2021): kahane-2021; Botero-Garcia et al., An intonosyntactic treebank for spoken French: what is new with Rhapsodie? (2025): garcia-2025; Kahane et al., Status of morphosyntactic features (2025): kahane-2025.
• Olivier, Sevdali & Folli, Clitic climbing and restructuring in the history of French (Glossa, 2023): olivier-2023.
• Mova: en/src/main.rs (xpos_fill, xpos_of, tok_eval, parse_eval, explain), en/src/gram.rs (Rel), prag/src/frames.rs; Open parsers: witnesses and tests; baselines in What every language gets.
• The full list, with licences, is in Sources.

This page in Markdown → https://md.mova.elamur.ai/free/en/docs/languages/fr/20-syntax-overview/