🗣️ 35. Reference and coreference

A French speaker who wants to talk about a person, a thing or an event chooses among names, full noun phrases, demonstrative phrases, clitic pronouns bound to the verb, strong pronouns, the pronominal adverbs y and en, the neuter ce, ça, cela and le, and, in a few places, nothing at all. The choice tells the hearer how easy the referent is to find. This chapter describes the French system of referring expressions and the way it differs from the English one (chapter 35 of the English manual): grammatical gender on every pronoun of the third person, clitics that cannot stand alone, dislocation with a resumptive clitic, the partitive en, the neuter pronouns that take up a whole clause, and a subject that cannot be dropped. It then gives the kinds of anaphora, the non-referring and generic uses, binding, reference in reported speech, the French coreference corpora and what research on them has found, the counts in the UD 2.18 French treebanks and in the CorefUD conversion of Democrat (all made on 2026-11-06), and what Mova does with French reference today, which is little.

Contents

1. Referring expressions in French
2. Givenness, accessibility and the reference chain
3. Kinds of anaphora
4. Non-referring and generic uses
5. Binding: se, soi, lui-même
6. Reference in reported speech and dialogue
7. French coreference corpora
8. What the research has found
9. Counts in the treebanks
10. How Mova handles it

1. Referring expressions in French

The forms of the personal pronouns, the order of the clitic cluster and the uses of y and en are in chapter 12; the points that matter for reference are these.

Gender is grammatical. Every pronoun of the third person singular and plural shows the gender of its antecedent, and the antecedent may be a thing: La table ? Elle est dans la cuisine; Le livre ? Il est sur la table. French has no neuter personal pronoun corresponding to English it: what English splits into he/she for people and it for things, French splits by the gender of the noun. Agreement in gender and number is therefore a far stronger filter on candidate antecedents than in English, for things as well as for people, and it is the first constraint every French resolver applies (Godbert & Favre 2017, godbert-2017-detection-de-coreferences-de-bout; Oberle 2019, oberle-2019-detection-automatique-de-chaines-de). Its limits are the epicene nouns (la victime, une personne, le professeur said of a woman), where the pronoun may follow the noun's gender or the person's, and the nouns that change gender between singular and plural use (gens).

Possessives agree with the possessed, not the possessor. Son livre and sa maison do not say whether the owner is a man or a woman: son and sa agree with livre and maison. Marie a vu Paul dans sa voiture leaves open whose car it is. English his/her gives the resolver a gender cue that French does not; French gives the number of the possessor only (son vs leur). The CorefUD Democrat data treats possessive determiners as mentions (5 333 mention heads with UPOS DET, 7.6% of all mentions, §7), following the French tradition that counts them as referring expressions because they vary in person and number with the possessor (Oberle 2019, citing Charolles 1988 and Schnedecker & Landragin 2014).

Clitics are not free words. A clitic cannot be stressed, coordinated or used alone. When a referent must be contrasted, coordinated or put after a preposition, the strong form appears: Lui, il reste ("he, for his part, stays"), Paul et elle, avec eux. With the verbs of the class penser à, tenir à, s'habituer à, a person is taken up by à + strong pronoun, not by the dative clitic (Je pense à elle, not *Je lui pense), and a thing by y (J'y pense) (chapter 12). The strong pronoun is thus both a marker of contrast and, with these verbs, the ordinary form.

Animacy is partly marked. The dative clitics lui and leur and the strong pronouns after a preposition are used for people in the standard; y and en are preferred for things (J'y pense, J'en parle). Godbert & Favre (2017) turn this into a rule: their system gives y and en the type inanimate and lui, elle, eux, elles in a prepositional or indirect position the type animate, and checks the candidate against it. In speech the boundary is loose (j'en ai parlé, de Paul is common), so the rule is a preference, not a constraint.

The subject is never dropped. French is not a null-subject language: a finite verb has a noun or a clitic subject, except in the imperative, in the infinitive and in coordination with a shared subject (Pierre boit et ∅ fume). Zero anaphora in French is thus limited to the coordination, which Oberle (2019) handles with a rule that copies the subject of the first conjunct, and to a few registers (notes, diaries, telegraphic style: Suis arrivé hier, Faut voir).

2. Givenness, accessibility and the reference chain

2.1 The scales

Gundel, Hedberg & Zacharski's givenness hierarchy and Ariel's accessibility scale (English chapter 35, §2) were stated for English, but their logic carries to French: the less information an expression carries, the more accessible the referent it signals. A French version of the accessibility scale, from most to least accessible, is:

Two French facts shape the scale. First, the clitic and the strong pronoun are two rungs where English has one (unstressed and stressed he). Second, French writing avoids repeating a noun more than English does; the school norm teaches the reprise by a synonym, a hypernym or an epithet (Victor Hugo… le poète… l'auteur des Misérables), so the rung of demonstrative and definite NPs is crowded with anaphores infidèles (§3.2). Oberle (2019) gives this avoidance of repetition as the reason to build lexical resources for them.

2.2 The reference chain

The French tradition, from Corblin and Charolles to Schnedecker, studies the chaîne de référence: the sequence of all expressions in a text that refer to one entity, together with its shape (how it opens, how often it is renewed by a full NP, where the pronouns cluster). The chain is the object the Democrat project annotated ("Description et modélisation des chaînes de référence", acknowledged in Oberle 2019) and the object that the Scolinter work on children's writing measures as continuité référentielle (Barletta 2024, barletta-2024-annotation-de-la-continuite-referentielle; Barletta & Ponton 2025, barletta-2025-analyse-de-la-continuite-referentielle): whether young writers keep a referent identifiable with pronouns and renewals, or lose it. The chain view differs from the pairwise view of most algorithms (pronoun → antecedent) in that it describes the whole sequence; a long chain in a novel can hold thousands of mentions spread over hundreds of pages (§8).

Boiteau, Mathet & Widlöcher (2025, boiteau-2025-integration-des-relations-inter-referents) argue that annotating chains is not enough: the relations between the entities themselves (a group and its members, a whole and its parts, an entity and its later state) form a second level, which they annotate with their own tool, OPERA. Their model is the clearest statement in the open French literature that coreference and bridging belong in one referential model rather than in two separate tasks.

2.3 Centering

Centering theory (Grosz, Joshi & Weinstein 1995, grosz-1995-centering-framework-modeling-local-coherence) ranks the entities of an utterance, in English by grammatical role. In French the subject clitic is the strongest continuation of the backward-looking centre, and dislocation adds a device English lacks in writing: a left-dislocated noun or strong pronoun announces the topic, and the clitic in the clause continues it (Mon frère, il est parti). A Continue transition in French writing is typically a clitic subject; a Shift is typically a full NP or a dislocation in speech. We have not found an open study that measures centering transitions on a French corpus.

3. Kinds of anaphora

3.1 Pronominal

The pronoun agrees with its antecedent in gender and number; person is third unless the antecedent is the speaker or the addressee. The clitic chosen depends on the function in the new clause, not in the old one: J'ai vu Marie. Je lui ai parlé (the antecedent was a direct object, the pronoun is a dative). The neuter le takes up a predicate or a clause rather than a noun: Elle est heureuse et elle ne l'a pas toujours été; Il me l'a dit (GSD fr-ud-test_00087).

(1) Democrat, CorefUD, articleswiki_zebre-p1-s1 to -p2-s3 (from a Wikipedia article, CC BY-SA 4.0): Zèbre est un nom vernaculaire […] pouvant désigner plusieurs espèces différentes d'herbivores de la famille des équidés […]. Ils se trouvent principalement en Afrique centrale et australe. Ces animaux se caractérisent par des bandes de rayures verticales noires et blanches.

The annotation links Ils to plusieurs espèces, and Ces animaux to Zèbre — two different chains; the next sentence takes the species up again with ce groupe. The pronoun ils is masculine plural, and the only masculine plural candidates are herbivores and équidés: the annotators chose espèces (feminine) on meaning, against agreement. The same sentence shows the limit of the agreement filter in the other direction: celle-ci in toutes les espèces descendantes de celle-ci is feminine singular and must take up une espèce commune, but the annotation links it to the chain of les équidés. A gender-and-number check would flag both links for review; it would also reject the first one, which is defensible. Agreement in French is a strong cue, not an oracle.

3.2 Nominal: anaphore fidèle and anaphore infidèle

The French terms, used in ANCOR and after it, separate a faithful reprise, with the same head noun (le chat… ce chat; ANCOR calls this a relation directe), from an unfaithful one, with a different noun: a synonym, a hypernym or an epithet (Mon chat… cet animal; La Seine… ce fleuve; Paris… la Ville Lumière) (Oberle 2019). The unfaithful reprise is frequent in written French and hard for a resolver because it needs lexical knowledge. Oberle builds two resources for it: a hypernym dictionary extracted from the first noun of each gloss of a parsed French Wiktionary (GLÀWI: chat → mammifère → animal, by transitivity), and a list of hypernyms and synonyms for named entities taken from YAGO (Zidane → footballeur, sportif, athlète). He notes that the relations depend on the domain: chat is an animal in ordinary use and an electronic conversation in computing.

The demonstrative determiner is the usual marker of an unfaithful reprise (cet animal, ce fleuve), because the demonstrative signals that the referent is already activated and the noun only classifies it. A definite NP with a new noun (l'animal) is possible but less explicit. In (1), Ces animaux is exactly this: a demonstrative with a hypernym of zèbre.

3.3 Associative anaphora (bridging)

A definite NP can refer to something not mentioned but inferable from what was: Nous arrivâmes dans un village. L'église était fermée. (the church of that village). French linguistics calls this anaphore associative; its study in French (Kleiber's work in the 1990s and 2000s) stresses that the relation is stereotypical — a village has a church, a car has a steering wheel — and that the anaphoric NP is definite because the associated object is unique relative to the antecedent. The classes are those of Clark's bridging (English chapter 35, §3.2): parts (une maison… le toit), roles (un procès… le juge), causes and consequences. The French works of Kleiber are not in the library; this paragraph rests on the general description and the English sources.

How frequent bridging is depends on the genre and on what the annotators chose to mark. In the 100 professional e-mail threads annotated by Guenoune et al. (2020, guenoune-2020-dataset-anaphora-analysis-french-emails), association (their name for bridging: any non-identity relation, part-of or "contextual association of ideas") is the second relation after identity: 252 of 1 856 annotated links, against 1 367 identity links; annotators disagreed most on these links. The CorefUD conversion of Democrat has no Bridge= marks at all (count of 2026-11-06), so the only French bridging data of size in Mova's reach is not annotated as such.

3.4 Demonstrative pronouns: celui-ci, ce dernier

When two candidates of the same gender and number compete, French writing chooses the most recent with celui-ci (celle-ci, ceux-ci, celles-ci) or ce dernier (cette dernière, ces derniers): Paul a écrit à son frère ; ce dernier ne lui a pas répondu (the brother, not Paul). A clitic would prefer the subject (Paul); the demonstrative pronoun overrides that preference and points to the last-mentioned. Celui-là in the pair celui-ci… celui-là traditionally points to the more distant, but this use is literary.

(2) GSD fr-ud-dev_00065: Il est le fils du Hall of Famer Ricky The Dragon Steamboat, et a été également entrainé par ce dernier. — ce dernier is the father, the last-mentioned man, not the subject il. In GSD, ce dernier is a DET ce plus an ADJ dernier that heads the nominal (here obl:agent), not a PRON: a resolver that looks for pronouns by UPOS misses it.

(3) Democrat sarrasine-p151-s1 (Balzac): - Mon cher, lui dit ce dernier, je suis chargé par notre ambassadeur de t'inviter à venir ce soir chez lui. — ce dernier names the speaker of the quote, and je in the quote is the same person (§6).

These forms are written French: GSD has 107 pronominal ce dernier (with dernier not amod) and 117 celui-ci, celle-ci, ceux-ci; the spoken treebanks have almost none (§9).

3.5 Abstract anaphora: ce, ça, cela, le

The neuter pronouns refer to situations, facts, propositions and stretches of discourse: Il est parti sans rien dire. Cela m'a surpris.; Ça, c'est beau (Rhapsodie Rhap_D2011-57: ça, c'est beau, hein.); Je le sais. Guenoune et al. (2020) call this the synthetic relation, "a summary of an idea previously described", after Lefeuvre's study of anaphores résomptives in c', cela, ça and ceci; it is the third relation in their e-mails (114 links). Godbert & Favre (2017) remove these pronouns from pronoun resolution altogether: their system first finds, from the dependencies, the pronouns of impersonal forms (il y a, il faut) and those that take up a sentence (– Louis est arrivé. – Oui, je le sais), and ignores them in what follows. That is the right first move for a resolver whose mentions are noun phrases: an abstract pronoun has no nominal antecedent and will be linked to a wrong one if it is not kept out.

Ce in c'est is ambiguous between a referring use (C'est grand, said of a thing in view; Ce sont des amis) and the expletive subject of a cleft (C'est lui qui recruta le successeur de Graham, GSD fr-ud-dev_00665), which UD marks expl:subj. In Democrat, ce and cela are the pronouns most often first mentions of their entity (375 of 428 ce heads and 144 of 165 cela heads, §7): they open new, abstract entities rather than take up old ones.

3.6 En and y

En takes up a phrase with de, and in particular a quantity of a type: J'ai trois frères → J'en ai trois; On en avait vue une dizaine (GSD fr-ud-dev_00067); Tsahal a commandé 130 Namers et en a reçu 10 (GSD fr-ud-train_06235). This partitive en is not identity coreference: the ten received are part of the 130 ordered, and en refers to the kind (Namers), with the numeral giving the new set. It is the French counterpart of English one-anaphora and of the partitive some of them; in UD GSD it is nmod of the numeral or the quantity noun (86 of 282 PRON en) or iobj/obj of the verb. Y takes up a place or a complement with à (On y mange correctement, GSD fr-ud-dev_00087). Both have idiomatic uses that refer to nothing (il y a, s'en aller, en vouloir à, s'y prendre), marked expl:comp or fixed in UD (§4).

3.7 Dislocation and the resumptive clitic

In speech, a referent is very often named in a dislocated position and taken up by a clitic in the clause: Mon frère, il est parti; Moi, je pense que…; Ce film, je l'ai vu; Il est parti, mon frère (right dislocation). The clitic and the dislocated phrase corefer within one sentence. UD French marks the dislocated phrase dislocated (with a subtype in the spoken treebanks: dislocated:subj, dislocated:obj, dislocated:mod, dislocated:obl) and the clitic with its ordinary relation.

(4) ParisStories, ParisStories_2020_maisonAbondonnee_27: donc, euh, sa maison, ben, elle avait été abandonnée. — sa maison is dislocated:subj, elle the subject.

(5) Rhapsodie, Rhap_D1002-14: ça vous dit quelque chose ça ? — a right-dislocated ça that repeats the subject ça.

The counts (§9) show how far this is a fact of speech: 354 dislocations in ParisStories (8.3 per 1 000 words) and 250 in Rhapsodie (5.7), against 75 in GSD (0.19) and 17 in Sequoia (0.24). In the spoken treebanks a dislocated subject almost always has a clitic subject on the same head (295 of 302 in ParisStories, 192 of 203 in Rhapsodie); a dislocated object has a clitic object about half the time (14 of 34, 14 of 21). The dislocated element is a strong pronoun in a third of the ParisStories cases (moi alone is 78 of 354) and ça in a fifth of those of Rhapsodie (56 of 250). For coreference this is the easiest link of all: the tree gives it. Godbert & Favre (2017) use exactly this for cataphora: Vous pouvez le décrire ce portefeuille — if the dependencies are right, the pronoun and the dislocated noun have the same governor and the link comes from the tree.

3.8 Cataphora

A pronoun can precede its antecedent. In French the common cases are the right dislocation (§3.7: Il est parti, mon frère; (5)), a fronted subordinate clause (Quand il est arrivé, Paul était fatigué), and the presentational voici/voilà and c'est + clause (C'est simple : il faut partir). As in English, a pronoun subject cannot take a noun in its own clause's complement as antecedent (*Il dit que Paul est fatigué, with il = Paul). In Democrat, only 39 entities whose first mention is an il exist (of 35 084 entities, §7): cataphora of the third person is rare in written narrative and news.

3.9 Split antecedents and groups

Sophie et Marie sortent ensemble. Elles sont très heureuses — elles refers to the group, which has its own gender and number (feminine plural) (Oberle 2019, who adds rules that build the group as a mention). French agreement with mixed groups is masculine plural (Paul et Marie… ils), so a masculine plural pronoun is compatible with any group containing a man, and a feminine plural pronoun with groups of women only. Guenoune et al. (2020) annotate split antecedents instead of building coordinated groups.

4. Non-referring and generic uses

Impersonal il. The impersonal subject is far more frequent than English expletive it because French has no there and builds its existential on il y a, its necessity on il faut, its weather on il pleut, and its extraposition on il est + adjective + de/que. In GSD 611 of 3 612 il (16.9%) are expl:subj; in the spoken treebanks the share rises to 38.4% in ParisStories (249 of 649) and 57.7% in Rhapsodie (286 of 496), and it is 55.4% in Sequoia (210 of 379), whose parliamentary and medical texts are full of il convient de, il est nécessaire que (§9). A resolver that treats every il as referring will therefore make an error on a sixth to a half of them; this is the French version of the English pleonastic-it problem, and larger. In speech the impersonal il is often dropped (faut voir, y a pas), which removes the problem and the token.

Generic and inclusive on. On is "people in general" (On dit que…) or, in speech, "we", which has almost replaced nous as a subject (chapter 12). In ParisStories on is 18.3 per 1 000 words against 2.5 for nous; in GSD 1.3 against 1.0 (§9). The two readings call for different treatment: generic on is not a mention of anyone; "we" on corefers with nous and with the speaker's earlier on. Democrat counts on as a mention and starts a new chain with a third of them (314 of 910 heads are first mentions). ANCOR, by contrast, does not link first and second person pronouns, which it treats as deictic (Oberle 2019).

Generic tu and vous. In conversation, tu (and vous among people who say vous) is used for anyone in a described situation: tu rajoutes un peu de sel, et tu rajoutes la farine (ParisStories_2019_cuisineApproximative_8, a recipe told to the hearer as a generic "you"). Whether it is the addressee or anyone can only be decided from the discourse.

5. Binding: se, soi, lui-même

The reflexive se (s') is used for all third persons and for on; the first and second persons use me, te, nous, vous. A reflexive clitic is bound by the subject of its own clause (principle A); a non-reflexive clitic of the third person is not bound by its clause's subject (principle B): Paul se lave (himself) vs Paul le lave (someone else). In GSD only a quarter of se/s' (587 of 2 337) are a true reflexive object (obj, qu'ils se sont consultés, fr-ud-train_06603) and 72 a dative (iobj); the rest are inherent reflexives (expl:pv, 986, Johnson s'appuyait sur) and reflexive passives or middles (expl:pass, 685), which refer to nothing (§9). The distinction is the annotators', not the form's (chapter 12).

The strong reflexive soi takes an indefinite or generic antecedent (Chacun pour soi, On est bien chez soi); with a definite antecedent the ordinary strong pronoun is used (Il pense à lui), so lui after a preposition may be bound by the subject: Paul ne pense qu'à lui (himself, or another man). French therefore lets a strong pronoun after a preposition corefer with the subject of its clause, against principle B as it is stated for clitics; lui-même removes the ambiguity (Paul ne pense qu'à lui-même). The forms in -même (lui-même, elle-même, soi-même) also have an intensive use without binding (l'autorité judiciaire elle-même, GSD fr-ud-dev_00594). GSD has 104 forms in -même and 11 soi (§9).

The reciprocal is also se (Ils se parlent), disambiguated if needed by l'un l'autre or entre eux.

6. Reference in reported speech and dialogue

Inside a direct quotation, je is the quoted speaker and tu/vous the quoted addressee; in the frame, the speaker is often named only after the quote by an inverted verb of saying (lui dit ce dernier in (3)). French typography marks direct speech with guillemets « » and, in novels, with a dash (—) at each change of speaker, often without closing guillemets; the speaker of a dash line is found by alternation and by the incises (dit-il, répondit Marie). Oberle (2019) writes rules for first and second person pronouns and possessives in quotations (Paul a dit : « Je suis étudiant. » → je = Paul); ANCOR leaves first and second persons unlinked; Democrat links them (2 257 heads je/j' are mentions). The automatic detection of direct speech in written French, which these rules need, is studied by Durandard et al. (2023, durandard-2023-automatic-annotation-direct-speech-written).

The choice of tu or vous is itself referential evidence: a vous addressed to one person picks out a different relation, and a switch from vous to tu in a dialogue marks the same addressee in a new relation, not a new addressee (chapter 32). A vous can also be plural, so its number is not known from the form: the verb agrees in the plural either way, and only a predicate adjective or participle shows it (Vous êtes prête vs prêtes).

In indirect speech the pronouns shift to the reporter's point of view: Il m'a dit : « Je viendrai » → Il m'a dit qu'il viendrait (je → il, coreferent with the subject of dire). This is the French case of the English pattern in which the subject of a verb of saying is the preferred antecedent of the subject pronoun of its complement clause.

7. French coreference corpora

The Democrat data in CorefUD. The conversion is in corefud-fr-democrat (train and dev; the README gives v1.1 of the conversion, 2023-02-24, and CC BY-SA 4.0). Its UD layer (lemmas, tags, features, relations) is automatic, which matters below. Counted on 2026-11-06 over train and dev, mention heads found from the head field of # global.Entity = eid-etype-head-other:

The pronoun heads, most frequent first: il 2 232 (and -il 75), elle 2 050 (and -elle 124), je and j' 2 257, relative qui 1 551, lui 1 040, nous 910, on 910, ils 605, vous 568, l' 562, me 525, le 494, ce 428 and c' 346, où 427, la 389, en 326, les 309, dont 299, moi 285, cela 165, tu 157. Three facts follow. Relative pronouns are mentions (qui, où, dont, que, lequel), unlike in GUM. The reflexive se almost never is (90 of 1 411 se tokens fall inside a mention, and 916 are expl in the automatic trees). And about a quarter of il are not mentions: 3 054 il tokens, 2 232 mention heads, so 822 (27%) are left out — the impersonal il of §4. The automatic trees mark only 2 of them expl, so a reader of this corpus cannot rely on its UD layer to find the impersonal il; the gold UD treebanks mark it (expl:subj, 611 in GSD), the parsed Democrat does not.

The chains are mostly pronominal in fiction and nominal in news: Bourgois & Poibeau (2025) find 74.95% pronominal mentions among the character mentions of their three novels, while Guenoune et al. (2020) give 74.8% pronominal anaphors in their e-mails and 41.1% in ANCOR. Democrat as a whole, mixing news, Wikipedia and fiction, has 28.3% pronoun heads among all mentions (including singletons, which are mostly nominal).

8. What the research has found

Rule-based systems. Godbert & Favre (2017) build an end-to-end French system for pronominal anaphora and faithful nominal reprises: a parser (MACAON), a mention detector trained on ANCOR (F 88.94 on predicted trees), semantic types for nouns and verbs from two lexicons of Dubois & Dubois-Charlier (the DEM and the LVF), and rules on gender, number, semantic class and distance. On the ANCOR test part (23 079 mentions) they report BLANC F 65.7. Their pronoun rules start by setting aside impersonal and sentence-taking pronouns (§3.5) and use the semantic type of the verb a pronoun depends on to choose an antecedent of the right class. Oberle (2019) builds a rule-based end-to-end system for written French that adds the unfaithful reprise (§3.2), groups (§3.9), zero subjects in coordination (§1) and first and second persons in quotes (§6). Evaluated on 3 250 tokens he re-annotated (a literary extract, news items, texts for learners of French), it reaches mention identification F 83.62 and MUC 58.1, B³ 64.53, CEAF 69.28, BLANC 53.08, against 50.2, 36, 31.9, 29.9 and 24.9 for the earlier rule-based RefGen. He concludes that machine learning is better for coreference in general and rules for specific phenomena, and proposes a hybrid.

Neural systems on speech. Grobol (2019, grobol-2019-neural-coreference-resolution-limited-lexical) adapts the end-to-end neural model of Lee et al. to ANCOR, with explicit mention detection (needed because ANCOR has singletons, which the end-to-end model ignores) and a span context of ten words on each side instead of the whole document, since the far context of transcribed speech is noisy. Mention detection reaches F 85.87; coreference on gold mentions a CoNLL average of 72.0 (against 75.3 for an earlier system that used gold features). Applied to Democrat, the same system does not carry well from spontaneous speech to writing (Grobol 2021, fr-nlp-2021-grobol-exploitation-corpus-democrat). Brassier et al. (2018, brassier-2018-classification-par-paires-de-mention; 2022, brassier-2022-valuation-comparative-de-systemes-neuronal) compare mention-pair classifiers and neural and statistical systems on the interactive part of ANCOR. Milintsevich (2025, milintsevich-2025-impact-asr-transcriptions-french-spoken) shows that models trained on human transcripts lose accuracy on automatic transcripts, which have punctuation, capitals, rephrasing and recognition errors; training on silver automatic transcripts as well as gold human ones helps, and models that favour recall degrade less than those that favour precision. Lopez (2023, lopez-2023-approches-neuronales-pour-la-detection) review the neural approaches to coreference chains for French.

Long documents. Bourgois & Poibeau (2025, bourgois-2025-elephant-coreference-room-resolving-coreference) annotate all character mentions of three full French novels from three centuries (285 176 tokens; inter-annotator agreement CoNLL F1 86.36). In their data 95% of pronouns are within seven mentions of their last antecedent, while proper and common nouns may be 270 mentions away, and the last 1% more than 1 700: pronouns, names and nouns need different search windows. Their pipeline (CamemBERT-based mention detection at F1 92–93; a mention-pair scorer with gender, person and distance features) reaches CoNLL F1 72.64 on gold mentions left to right and 74.32 when proper-noun mentions are clustered globally first; performance falls as documents grow, which is why systems trained on short texts disappoint on whole books.

World knowledge. The French Winograd schemas (Amsili & Seminck 2017, amsili-2017-google-proof-collection-french-winograd) are 107 pairs (214 questions) in which only knowledge of the world decides the antecedent (Nicolas n'a pas pu porter son fils parce qu'il était trop faible / lourd). Translating them from English was hard precisely because of gender: a schema works only if both candidates have the same gender and number as the pronoun, and French nouns often do not (cheveux is plural, siphon singular). Grammatical gender, which helps a resolver everywhere else, has to be neutralised to build a test of reasoning. Seminck, Segonne & Amsili (2019, seminck-2019-modeles-de-langue-appliques-aux) test language models on them.

9. Counts in the treebanks

Counted on 2026-11-06 from UD_French-{GSD,ParisStories,Rhapsodie,Sequoia}, all splits, syntactic words (multiword tokens counted as their parts), lowercased forms with UPOS PRON unless stated. GSD lemmatises il, elle and le as lui and se as soi (chapter 12), so the counts are by form, not by lemma.

Pronouns by type.

Speech has three times as many pronouns per word as writing, and the demonstratives (ça, c') are ten times as dense.

Selected forms.

Ça replaces cela almost entirely in speech (309 to 2 in ParisStories), on outnumbers nous seven to one there, and the written forms of re-identification (celui-ci, ce dernier) are absent from the spoken treebanks.

Y and en by relation (GSD). y: obl 222, expl:comp 154 (il y a), iobj 114; en: iobj 112, nmod 86 (the partitive with a numeral, §3.6), expl:comp 53, obj 15, obl 13. In the spoken treebanks y is mostly the y of il y a: expl:comp 213 of 262 in ParisStories and 202 of 258 in Rhapsodie.

Expletive and reflexive relations.

Reflexive se/s' (Reflex=Yes) in GSD: 2 337, of which expl:pv 986, expl:pass 685, obj 587, iobj 72; in ParisStories 241: expl:pv 91, obj 90, iobj 33, expl:pass 22.

Dislocation.

GSD and Sequoia mostly use the plain dislocated (50 and 11); the subtypes are the spoken treebanks'. The dislocated element, by UPOS and pronoun form: in ParisStories NOUN 151, moi 78, PROPN 25, ça 24, ce 15, lui 14, nous 13; in Rhapsodie NOUN 101, ça 56, moi 38, ce 15, PROPN 14.

How Mova handles it

Nothing in Mova resolves French reference. Checked against the code on 2026-11-06:

• The coref crate (coref — SLM coreference: a sieve of rules with a reason for every link) is a deterministic multi-pass sieve over UD trees, built and measured on English GUM. Its mentions (coref/src/mention.rs) are nominal heads from the tree, excluding expl, so on a gold French tree the impersonal il, the cleft ce, the idiomatic y/en and the expl:pv/expl:pass se would be kept out correctly. Person and number come from UD FEATS, so a French il would get third person singular. But gender and animacy come only from English closed lists: the pronoun table (pron_table: he, she, it…), the gendered nouns (gender_noun: mother, king…), the titles (Mr., Mrs.), and the gender lexicon of coref/src/lexicon.rs, learned from English noun–pronoun paths. The mention reader does not take Gender from FEATS. A French il and elle would both be third-person pronouns of unknown gender, compatible with any noun — the strongest French cue (§1) unused.
• Pleonastic and event pronouns (coref/src/it.rs) are recognised by English lemma lists (rain, seem, happen, true, mistake…). There is no list for falloir, pleuvoir, il est + adjective + de, ça/cela as event pronouns, or le taking up a clause.
• Quotations (coref/src/doc.rs, mark_quotes) are found from ", “ ” and the PTB quote tokens; French guillemets « » and the dialogue dash are not quote marks for it, so the speaker sieve would not see French direct speech.
• The CorefUD reader (coref/src/corefud.rs) reads Entity= brackets with the field order of # global.Entity; French-Democrat uses the same header Mova writes (eid-etype-head-other) and has no discontinuous mentions (which the reader rejects), so it should be readable. It has not been tried: there is no French coreference measurement in Mova.
• The frame reader's resolver (prag/src/frames.rs, ent_at, DUMMY_IT) works on English word forms (he, she, it, I, you) and English verb lists. It does not read French.
• The tree. French is parsed by the English parser after en xpos-fill (README): LAS 87.00 on GSD with its own tags, 89.78 with gold tags (UD 2.18, train → dev, 2026-10-22). The parser sees forms and PTB tags only (PRP for every personal pronoun, no case, no gender, no clitic/strong distinction), and the relations of this chapter (expl:*, dislocated*, iobj) have not been scored separately.

What is planned. French pragmatics is being built as a layer for the Eel, the worker that improves Mova's trees and pragmatics: on 2026-11-06 the Eel's schedule has a round kind prag-fr (a gated native round, CPU only, that measures a candidate on a pinned French dev sample and installs it only on a gain, with proposals kept for review; tools/eel/src/schedule.rs, tools/eel/src/native.rs, on the branch under development). The French tool that the round will run is not written yet, and it has no coreference component. Above the language layers, a language-independent meta-pragmatics — one label schema for reference, acts and relations shared by English, French and the other languages, with each language's forms mapped to it — is the design direction; nothing of it is implemented. For coreference that schema would hold the language-independent part of this chapter (identity, bridging, split antecedent, non-referring, generic, bound) and leave the French part (gender agreement on things, clitic versus strong form, the impersonal il, en of quantity, dislocation) to the French layer.

What a French resolver would need, from the cheapest, built on what exists:

1. Gender from FEATS in mention.rs for nouns and pronouns (GSD and Democrat have Gender on both), with a French pronoun table: il/le/lui masculine singular, elle/la feminine, ils/eux masculine plural, elles feminine plural, lui/leur dative and animate, y/en inanimate by preference (Godbert & Favre 2017), on and se without fixed features.
2. The non-referring uses from the tree (expl:*) on gold trees and, on Mova's own trees, from the French impersonal verbs and constructions (falloir, pleuvoir, il y a, il est + ADJ + de/que): §4 shows that a sixth to a half of il depend on it.
3. Dislocation links from the tree (§3.7): the clitic and the dislocated phrase on one head.
4. Ce dernier and celui-ci as most-recent-candidate pointers (§3.4), with ce dernier found as DET + ADJ.
5. Guillemets and the dialogue dash in mark_quotes, and the French verbs of saying in incises (dit-il).
6. A first measurement on French-Democrat dev with gold mentions, with the nearest-agreeing-pronoun baseline as the negative control, before any claim.

Sources

• Godbert, E., Favre, B. 2017. Détection de coréférences de bout en bout en français. TALN. godbert-2017-detection-de-coreferences-de-bout (CC BY 4.0).
• Oberle, B. 2019. Détection automatique de chaînes de coréférence pour le français écrit : règles et ressources adaptées au repérage de phénomènes linguistiques spécifiques. TALN-RECITAL. oberle-2019-detection-automatique-de-chaines-de (CC BY 4.0).
• Grobol, L. 2019. Neural Coreference Resolution with Limited Lexical Context and Explicit Mention Detection for Oral French. CRAC. grobol-2019-neural-coreference-resolution-limited-lexical (CC BY 4.0).
• Grobol, L. 2021. Exploitation du corpus Democrat par apprentissage artificiel. Langages 224. fr-nlp-2021-grobol-exploitation-corpus-democrat (CC BY 4.0, HAL).
• Brassier, M. et al. 2018; 2022. brassier-2018-classification-par-paires-de-mention, brassier-2022-valuation-comparative-de-systemes-neuronal (CC BY 4.0).
• Lopez, F. 2023. Approches neuronales pour la détection des chaînes de coréférences : un état de l'art. lopez-2023-approches-neuronales-pour-la-detection (CC BY 4.0).
• Milintsevich, K. 2025. Impact of ASR Transcriptions on French Spoken Coreference Resolution. CRAC. milintsevich-2025-impact-asr-transcriptions-french-spoken (CC BY 4.0).
• Bourgois, A., Poibeau, T. 2025. The Elephant in the Coreference Room: Resolving Coreference in Full-Length French Fiction Works. CRAC. bourgois-2025-elephant-coreference-room-resolving-coreference (CC BY 4.0).
• Guenoune, H. et al. 2020. A Dataset for Anaphora Analysis in French Emails. CRAC. guenoune-2020-dataset-anaphora-analysis-french-emails (CC BY 4.0).
• Amsili, P., Seminck, O. 2017. A Google-Proof Collection of French Winograd Schemas. CORBON. amsili-2017-google-proof-collection-french-winograd (CC BY 4.0). Seminck, O., Segonne, V., Amsili, P. 2019. seminck-2019-modeles-de-langue-appliques-aux (CC BY 4.0).
• Boiteau, A., Mathet, Y., Widlöcher, A. 2025. Intégration des relations inter-référents dans l'annotation de la coréférence. boiteau-2025-integration-des-relations-inter-referents (CC BY 4.0).
• Barletta, M. 2024; Barletta, M., Ponton, C. 2025. Continuité référentielle in the Scolinter corpus. barletta-2024-annotation-de-la-continuite-referentielle, barletta-2025-analyse-de-la-continuite-referentielle (CC BY 4.0).
• Durandard, N. et al. 2023. Automatic Annotation of Direct Speech in Written French Narratives. ACL. durandard-2023-automatic-annotation-direct-speech-written (CC BY 4.0).
• Grosz, B. J., Joshi, A. K., Weinstein, S. 1995. Centering. grosz-1995-centering-framework-modeling-local-coherence.
• Informing only (non-commercial licences, not quoted): Muzerelle et al. 2013, 2014 on ANCOR (muzerelle-2013-ancor-first-large-french-speaking, muzerelle-2014-ancor-centre-large-free-spoken); Wilkens et al. 2020 on Dem1921 (wilkens-2020-french-coreference-spoken-written-language); Nedoluzhko et al. 2022 on CorefUD (nedoluzhko-2022-corefud-1-0-coreference-meets); Désoyer et al. 2014, 2015 on ANCOR.
• Wanted: Kleiber, G. on anaphore associative; Corblin, F. 1995, Les formes de reprise dans le discours; Charolles, M.; Schnedecker, C. & Landragin, F. 2014 on reference chains.
• Data: CorefUD French-Democrat (v1.1 of the conversion, CC BY-SA 4.0), corefud-fr-democrat; UD 2.18 French GSD, ParisStories, Rhapsodie (CC BY-SA 4.0), Sequoia (LGPL-LR), ud-treebanks-v2.18. All counts made on 2026-11-06.
• The reference grammars of Sources for the descriptive part (chapter 12).

This page in Markdown → https://md.mova.elamur.ai/free/en/docs/languages/fr/35-reference-and-coreference/