🗣️ 38. Dialogue, discourse markers and reported speech

Conversation has its own order: turns, pairs of actions that call for each other, preferred and dispreferred responses, repair, and the steady work of showing that one has understood. Spoken French does this work with a large set of small words — bon, ben, bah, alors, donc, enfin, quoi, hein, voilà, tu vois, du coup, genre — with the filled pause euh, with the feedback tokens ouais, mh, d'accord, with vocatives, and, in writing, with the typography of dialogue (dashes, guillemets, the incise dit-il). This chapter describes these devices, how the French UD treebanks annotate them, how often they occur, and what Mova does with them, which for French is almost nothing yet. Speech acts and the answer particles oui, si, non are in 31. Speech acts; tu/vous, titles and politeness formulas in 32. Politeness and hedging; connectives that link propositions (parce que, donc, pourtant) in 36. Discourse relations and coherence; the spoken register as a whole in 37. Register and variation; the punctuation of dialogue in chapter 7. Every count is on UD 2.18, all splits, made on 2026-11-06.

Contents

1. The organisation of conversation
2. Recruitments and responses
3. Feedback and grounding
4. Discourse markers
5. The filled pause, interjections and laughter
6. Disfluency and reformulation
7. Vocatives and address
8. Reported speech
9. Dialogue in written narrative
10. Questions and answers in dialogue
11. French dialogue corpora and dialogue acts
12. How Mova handles it

1. The organisation of conversation

Conversation analysis describes talk as an orderly system of turns (Sacks, Schegloff & Jefferson 1974; not open — see the English chapter, 38. Dialogue, discourse markers and reported speech §1, for the units: turn-constructional unit, transition-relevance place, adjacency pair, preference, insertion sequence, repair, pre-sequence). The units are the same in French; what differs is the material that fills them.

Transition-relevance places. Listeners do not wait for silence to take the floor: they predict where the current turn can end. Uro et al. (2024, uro-2024-evaluation-perceptive-de-l-anticipation, CC BY 4.0) tested this on French television talk (the REPERE corpus). Fifty-three French speakers heard or read 64 inter-pausal units, cut by zero to three prosodic words before their end, and judged whether they could take the floor now, soon or not yet. With the audio, listeners anticipated a turn end up to three prosodic words before it came, and they were more reluctant to take the floor at non-terminal boundaries; with the transcript alone they anticipated less. The prosody of speech, not only its words and syntax, carries the turn-end cues. A second paper by the same group annotates transition-relevance places and interruptions in French conversation (uro-2024-annotation-transition-relevance-places-interruptions, CC BY-NC 4.0: it informs, it is not quoted).

For a text-only system the consequence is plain: in a transcript the turn boundaries are given by the transcriber, and the cues that told the listener where they were — falling pitch, lengthening, the final quoi or hein — survive only in part, as words.

Dispreferred responses in French have the shape described for English: a delay, a preface (ben, bah, euh, bon), an appreciation or an apology (c'est gentil, désolé), an account (c'est que…, c'est-à-dire que…), and a softened refusal (je vais pas pouvoir, ça va être compliqué). Ben and bah (§4) are the French counterparts of English well in this position.

2. Recruitments and responses

A recruitment gets another person to do something practical (Floyd, Rossi & Enfield (eds.) 2020, books/e/en/enfield-2020-getting-others-things-pragmatic, open; English chapter §2). French is not one of the volume's eight languages, so no comparable French count exists in an open source we hold. What the grammar offers is described in 31 §4 (the ladder from the imperative to Est-ce que vous pourriez…?) and 32 §6 (the conditional and the imperfect of politeness). Two French points matter for dialogue:

• The response to a request is often not a word but an action, with a token on top: Tiens. / Tenez. (handing something over; the imperative of tenir used as a presentative), Voilà. (marking that the thing asked for is done or given), Je vous en prie. (granting). Voilà is the most general completion marker of spoken French (§4).
• Refusing is done with an account more than with non: Je peux pas, j'ai du boulot. A bare Non. to a request is curt, as in English.

3. Feedback and grounding

Participants show each other that they have heard and understood (Clark & Schaefer 1989, papers/c/cl/clark-1989-contributing-discourse; Clark & Brennan 1991, papers/c/cl/clark-1991-grounding-communication; English chapter §3). The French feedback inventory is short and very frequent.

Prévot and colleagues studied feedback in the Corpus of Interactional Data (CID, about eight hours of conversation recorded in Aix-en-Provence) (prevot-2013-quantitative-view-feedback-lexical-markers, CC BY-NC-SA 3.0: informs only, not quoted). In our words: the core feedback items there are ah, bon, ben, euh, mh, ouais, oui, non, d'accord, OK and voilà; they combine freely (some two hundred distinct combinations, mostly repetitions such as ouais ouais ouais and base items preceded by a modifier such as ah or euh); and speakers differ a great deal — the heavy feedback producers are those who produce many light backchannels (mh, ouais), much of it in overlap. The SUMM-RE meeting corpus (Hunter et al. 2024, hunter-2024-summ-re-corpus-french-meeting, CC BY 4.0) counts backchannels per minute with a deliberately simple definition — single-word utterances among oui, ouais, hm, mh, non, ok, ah, ben, bien, eh, euh, voilà — and finds fewer backchannels and more filled pauses in video-call meetings than in meetings in person: grounding costs more when the channel is poorer, as Clark & Brennan predict.

In the UD treebanks, feedback is visible only where a listener's token is transcribed inside the speaker's unit or as its own unit. Rhapsodie (dialogues and monologues) carries a # speaker comment on every one of its 3 209 sentences; ParisStories is mostly monologue narratives told to a listener. Their feedback tokens are INTJ with discourse: ouais 104, oui 106, mh 43, voilà 74 occurrences as discourse in the two treebanks together (§4.3). Kahane et al. (2021, kahane-2021-annotation-guidelines-ud-sud-treebanks, CC BY 4.0) propose that a listener's filler such as mh, uttered during the speaker's utterance, be its own tree with features pointing to the utterance it attaches to; the UD 2.18 French treebanks do not have these features (no AttachTo in Rhapsodie or ParisStories, checked 2026-11-06).

4. Discourse markers

4.1 What they are

Discourse markers (French marqueurs discursifs, also particules énonciatives, petits mots du discours) are mostly invariable words and phrases that do not add to the truth conditions of the utterance and that show how it relates to what came before, to the speaker's state of mind, or to the management of the interaction. Herreño Castañeda & Sillaire (2025, castaneda-2025-annotation-de-marqueurs-discursifs-le, CC BY 4.0) distinguish three families of function: structuring the discourse logically and in time (the connectives, for which LEXCONN is the reference list; 36), expressing the speaker's attitudes and emotions, and managing the interaction. Most markers are polyfunctional, and most also have a non-marker use (quoi the pronoun, enfin "at last", genre "kind, gender", après "after"), so the first task is to tell the uses apart.

That task is hard even for experts. In the CODIM project, a corpus of about 110 million tokens of spoken, written and online French was pre-annotated with local grammars; for après, only 848 occurrences were sure markers, 75 151 sure non-markers and 70 059 left as candidates — more than half of the candidates in the spoken and online parts. Two experts then annotated candidates as contrastive marker (après, c'est pas grave "then again, it doesn't matter") or non-contrastive: a naive first protocol reached Cohen's κ = 0.54, a refined one κ = 0.69, with most disagreements in speech (the same paper). Sow et al. (2025, sow-2025-etat-de-l-art-sur, CC BY 4.0), reviewing discourse markers in NLP, note that LEXCONN lists 328 French connectives with their syntactic categories and relations, and that nearly all NLP work on markers is on written text, not on speech or its transcripts.

4.2 The main French markers

Two observations. First, several of these are grammaticalised clauses: tu vois, tu sais, je veux dire, je crois, je sais pas are verbs whose object is the host utterance and which cannot take their own tense or modifiers (*tu as vu in the same function) — Kahane et al. (2021) cite this as the reason to treat them as markers rather than parentheticals. Second, the written language uses only a few of them (bref, eh bien, enfin, quand même); in edited prose most appear only in dialogue.

Bally (1909, Traité de stylistique française, bally-1909-traite-de-stylistique-fran, public domain), one of the first to describe the langue parlée as a system rather than a corruption of writing, already set the expressive "affective" means of speech — interjections, exclamative words, intonation — beside its logical means (§§ 261 and following of his treatise); the discourse-marker literature since the 1980s has given that intuition a lexicon.

4.3 In the treebanks

How the UD French treebanks annotate markers depends on the treebank. The written treebanks (GSD, Sequoia, PUD) follow the UD rule that a marker keeps the relation of its word class: donc and alors are advmod (GSD: donc advmod 177, alors advmod 250 and mark 124), and discourse is kept for interjections. The spoken treebanks (Rhapsodie, ParisStories), converted from SUD, give discourse to markers of any word class: ADV (enfin, donc, alors), NOUN (coup in du coup, genre), the ADP head of en fait, and verbal markers (tu vois, tu sais, je veux dire, je crois). Kahane et al. (2021) explain the choice: the category-based distinction of UD was judged unwarranted, so SUD extends discourse to verbal markers (in UD, parataxis:discourse would be expected). In ParisStories donc is advmod 238 times and discourse 61 times: the line is drawn by use, not by word.

discourse per treebank (UD 2.18, all splits, syntactic words with punctuation as denominator, 2026-11-06):

The spoken treebanks have six hundred times the rate of GSD. The difference is partly the language and partly the annotation: if Rhapsodie and ParisStories followed the GSD convention, their ADV and verbal markers would not be discourse, but the interjections alone (euh, bah, ben, ouais, hein, mh) would still give some forty per thousand words.

The most frequent discourse lemmas:

Two spoken corpora, two profiles. ParisStories (young Parisian speakers telling stories, recorded 2019–2022) is the corpus of du coup, bah, genre and en fait; Rhapsodie (recorded earlier, more varied situations, including radio and public speech) has more hein, ben, alors and je veux dire, and du coup is not among its frequent markers. Genre as a marker (78 in ParisStories) does not occur as discourse in Rhapsodie at all. These are the generational markers that chapter 37 describes; the two treebanks are too small and too different in their situations to date the change, but they show its direction. In GSD, the written web texts (reviews, blogs) use bref to close a review (Bref, à déconseiller !, fr-ud-test_00215) and attention as a warning.

Short examples from the spoken treebanks (CC BY-SA 4.0; transcription as in the treebank):

(1) et du coup tu finis avec une vraie boule de pâte. (ParisStories_2019_cuisineApproximative_14) — du coup as upshot of the preceding step.

(2) et en gros euh, ensuite, ensuite, tu mets le beurre, genre des morceaux de beurre. (ParisStories_2019_cuisineApproximative_5) — euh filled pause, ensuite repeated (the first is reparandum), genre introducing an approximation.

(3) pas une semaine après, hein. (ParisStories_2020_concoursInstagram_41) — final hein, seeking agreement with a self-correction.

(4) du coup, on a dit, bah, voilà, quoi, c'est notre chien! (ParisStories_2021_adoptionMouts_21) — four markers before the reported words.

(5) mais, je veux dire, ils ont travaillé normalement. (Rhap_D0002-7) — je veux dire as discourse, its verb vouloir the head.

(6) euh, je pense que la chirurgie, c'est pareil, hein. (Rhap_D0006-100).

4.4 Positions

Where in the unit each marker stands (ParisStories and Rhapsodie together; first and last non-punctuation word of the sentence as segmented by the treebank; "medial" includes a marker after a conjunction or after another marker, as in et euh or bah ouais; 2026-11-06):

The table confirms the positional classes of §4.2: quoi and hein are closers (71% and 67% of their marker uses are the last word), ah and alors openers, bah, ben and bon turn-initial prefaces that usually follow et, mais or another marker rather than standing first. Euh is everywhere. Bref is split between opening a summary and closing a unit.

5. The filled pause, interjections and laughter

The filled pause euh (/ø/, often lengthened) is the most frequent discourse word of spoken French: 1 598 in ParisStories (37.3 per thousand words) and 994 in Rhapsodie (22.5). It is not an interjection in meaning: it holds the floor while the speaker plans, and it marks the place of a coming repair (§6). After UD 2.18, the French guidelines added the subtype discourse:filler for it, together with discourse:tag for tag questions (…, hein ?, …, non ?) and conj:reform for reformulations (Mova's notes on the UD guidelines: dialects/ud-next/seeds.md#fr-spoken-subtypes); the 2.18 treebanks counted here do not use them yet. Euh is spelled euh in French transcripts (heu is an older variant).

Interjections proper — ah, oh, eh, ouf, aïe, ouille, bof, hop, chut, beurk, waouh, oh là là, mince, zut, punaise (euphemistic), putain (vulgar, extremely frequent in some speech) — express emotion and take no dependents in the usual case. Bally (1909, §261) described the primary interjections ah! eh! oh! as former involuntary cries that have come to be used with a symbolic value for pure affective movements, beside "exclamative words" that are ordinary words used as exclamations (Ciel !, Dommage !). GSD has 87 INTJ tokens in 400 385 words (2026-11-06), Sequoia 5: written French keeps its interjections for quoted speech, greetings (bonjour, salut) and the web.

Some interjections are verbs or nouns frozen in the imperative or as a noun phrase: tiens, tenez ("here", "well, well"), allez (encouragement, closing: allez, salut), dis donc, dites donc (calling attention, mild reproach), voyons (reproach), attention, pardon, dommage. UD tags the frozen imperatives as VERB or INTJ depending on the treebank (ParisStories: tenir as discourse 7 times).

Laughter is transcribed by the spoken corpora as a non-speech event and is not part of the UD trees. It is a dialogue move with its own placement in turns (Ginzburg et al. 2020, ginzburg-2020-laughter-as-language); for French, Bigi and colleagues have studied laughter in the CID (bigi-2016-laughter-french-spontaneous-conversational-dialogs, CC BY-NC 4.0: informs only), and SUMM-RE marks laughs and coughs in its corrected transcripts (Hunter et al. 2024).

6. Disfluency and reformulation

Spoken French repeats and repairs constantly: le, le, le vacarme; j'ai euh, j'entends, j'écoutais pardon la radio (ParisStories_2019_devoirPhilosophie_11). The two spoken treebanks have more reparandum than any other relation of speech except discourse: 1 035 in ParisStories (24.2 per thousand words) and 1 205 in Rhapsodie (27.2); the written treebanks have none (one in ParTUT).

The annotation is not the standard UD one, and a reader of these trees must know it. Kahane et al. (2021) analyse repetitions, corrections and reformulations as paradigmatic lists — the same syntactic position filled several times — following Blanche-Benveniste's work on spoken French; in SUD they are a subtype of conj, conj:dicto, and nothing said is treated as cancelled. In the conversion to UD, conj:dicto is renamed reparandum, but the first conjunct remains the head: the repaired element depends on the reparandum, the reverse of the usual UD analysis where the reparandum depends on the repair. The authors also chose not to separate voluntary reformulation from disfluency, because the line between them cannot be drawn reliably. Truncated words are marked with a tilde (pl~, plein d'enfants, ParisStories_2021_couruLaVoir_11).

Reformulation markers introduce the repair or the second formulation: enfin (the most frequent), je veux dire, c'est-à-dire, pardon, ou plutôt, disons: on vit vraiment dans un petit micro~, enfin, vraiment dans un petit microcosme (Rhap_D0006-47). Grabar & Eshkol-Taravella have worked on detecting reformulations and predicting their pragmatic functions in spoken French (grabar-2016-prediction-automatique-de-fonctions-pragmatiques, CC BY 4.0; grabar-2016-detection-reformulations-spoken-french, CC BY-NC 4.0).

7. Vocatives and address

A vocative names or calls the addressee; UD gives it the relation vocative from the head of the clause. The choice of address term — first name, Monsieur/Madame, title, endearment, insult — is a politeness choice and is treated in 32 §5; here, the dialogue side.

French politeness expects the greeting to be completed by a term of address in formal service encounters (Bonjour Madame rather than a bare Bonjour); in familiar speech, generational vocatives play the role of English dude, mate: mec, meuf, frère, gros, ma poule. ParisStories has cauchemar, meuf, tu peux même pas savoir. (ParisStories_2022_11_copineEnCarton__33), with meuf as vocative.

The vocatives per treebank (2026-11-06): GSD 15 (NOUN 10, PROPN 5), Sequoia 53 — all NOUN, 37 monsieur and 16 madame, the opening formulas of the European Parliament debates in its Europarl part (Monsieur le Président, Madame la Présidente) —, Rhapsodie 41 (PROPN 25, NOUN 16: a host greeting guests by name, as in Rhap_M2006-1), ParisStories 16, ParTUT 71 (licence NC: counted, not quoted), PUD 1. The parliamentary formula is the single most frequent vocative of written French in these treebanks.

8. Reported speech

8.1 Forms

Tense agreement (concordance des temps) in indirect speech after a past reporting verb shifts present to imperfect, past to pluperfect, future to conditional: Il a dit : « Je viendrai. » → Il a dit qu'il viendrait. Spoken French often keeps the original tense when the content still holds (Il m'a dit qu'il vient demain); the shift is the rule of writing (chapter 13, chapter 24).

8.2 The incise

After or inside a direct quotation, French reports the speaker in an incise with the subject inverted: dit-il, répondit Marie, a-t-elle ajouté, s'écria le roi, précise le ministre. The inversion is obligatory in the standard language, whether the subject is a pronoun or a noun; the euphonic -t- appears between a vowel-final verb and il, elle, on (a-t-il dit, demanda-t-elle). Spoken familiar French has an uninverted qu'il dit, qu'elle me fait (Je viens pas, qu'il me dit), which grammars mark as popular; Tobler (1905, Mélanges de grammaire française, tobler-1905-melanges-de-grammaire-fran, public domain) traces the Old French background of these mixed constructions (ch. XXXIX: direct speech introduced by que, direct speech continuing indirect speech).

Kahane et al. (2021) distinguish the incise (parataxis:insert) from a parenthetical clause (parataxis:parenth): an incise is not saturated — its object is the quote — and cannot stand alone, it belongs to the formal registers (news, stories, sermons), and in French it is marked by subject inversion. The counts bear this out. In GSD (2026-11-06) there are 183 parataxis:insert, 182 of them headed by a verb; of these, 139 have the subject after the verb (76%) and 43 before it. The verbs are verbs of saying: dire 21, ajouter 19, expliquer 15, déclarer 15, indiquer 14, affirmer 13, préciser 12, souligner 8, annoncer 8, écrire 6, rapporter 5 — the reporting verbs of the press. Examples:

(7) "Il est désormais clair que l'incident est spécifique au moteur Trent 900", a ajouté Rolls-Royce. (fr-ud-train_00610) — root clair (the head of the quote), parataxis:insert(clair, ajouté), nsubj(ajouté, Rolls-Royce) after the verb.

(8) L'accord engendrera une perte financière considérable pour l'UE, a souligné la même source. (fr-ud-dev_00803) — a report without quotation marks, the same structure.

This is the reverse of the English EWT analysis, where the reporting verb is the root and the quote its ccomp (English chapter §7.1). A reader that looks for the speaker of a quote must, in French UD, go from the quote's root down to a parataxis:insert child and take its nsubj, wherever that subject stands. The spoken treebanks have almost no incises (Rhapsodie 15 parataxis:insert, mostly écrire in a read-aloud text; ParisStories none), but they have parenthetical tu sais, je sais pas as parataxis:parenth (ParisStories 27, Rhapsodie 39).

8.3 Free indirect speech

The style indirect libre — a character's words or thoughts given without a reporting verb, in the narrator's tense and person, but with the character's exclamations, questions, deictics and evaluations — is a classic of French narrative prose from La Fontaine to Flaubert and Zola. The name comes from Charles Bally's 1912 article (Le style indirect libre en français moderne, Germanisch-Romanische Monatsschrift 4; not in our library — wanted), written after Tobler had noted the mixture of direct and indirect report in older French. Its marks:

• imperfect and conditional in the narrator's frame (il viendrait demain, not il viendra);
• third person for the character (il, not je);
• the character's deictics: demain, ici, maintenant where the narrator would write le lendemain, là, alors;
• questions and exclamations with inversion or !: Qu'allait-il faire ?;
• the character's vocabulary and evaluation (ce maudit train).

Nothing in a UD tree marks it; it is a property of a stretch of text, recognised by the combination of narrative tense and the speaker's subjectivity.

9. Dialogue in written narrative

French typography gives dialogue its own layout (chapter 7):

« Vous partez déjà ? demanda-t-elle.
— Il le faut, répondit-il. Le train est à huit heures.
— Alors, bon voyage. »

• an opening guillemet before the first reply and a closing one after the last, in the traditional style; many modern novels drop the guillemets and use only dashes;
• a dash (tiret, an em dash or en dash with a space) at the head of each change of speaker;
• the incise inside the reply, set off by commas, with inversion (§8.2), and without closing the guillemets;
• narration between replies closes the guillemets, or, in the dash-only style, stands as an ordinary paragraph.

The treebanks barely contain literary dialogue: in GSD the opening guillemet « occurs 955 times, 101 of them sentence-initially, the straight quote " 952 times (63 initial), the closing » 894 times, and the long dash — twice (2026-11-06); the quotations are press quotations and titles, not dialogue. Sequoia's sentence-initial hyphens (312) are list items of its medical texts, not dialogue dashes. A French dialogue reader will therefore have to be built and checked on public-domain novels and tales rather than on the treebanks.

10. Questions and answers in dialogue

A question makes an answer relevant. French has a three-way answer system to polar questions: oui to a positive question, non to either, and si to contradict a negative question or statement (Tu ne viens pas ? — Si ! "Yes, I am"); a bare oui after a negative question is ambiguous and often corrected by the hearer (31 §6.1). Spoken answers are reinforced or softened: ben oui, bah non, ouais ouais, si si, mais si, ben si, pas du tout, absolument, tout à fait, carrément (young, emphatic).

Questions in dialogue are often not requests for information:

• tag questions seek confirmation: C'est loin, non ?, Tu viens, hein ?, C'est bien ça ?, n'est-ce pas ? (careful). The new subtype discourse:tag is for these (§5).
• rising declaratives (Tu pars ?) are the ordinary yes/no question of speech (chapter 26); with déjà (Tu pars déjà ?) they carry surprise.
• indirect answers (Tu viens ce soir ? — J'ai du travail.) leave the yes or no to inference. Müller & Plank (2024, muller-2024-indirectqa-understanding-indirect-answers-implicit, CC BY-NC 4.0: informs only) built a French and Spanish set of indirect answers to implicit polar questions; it is discussed in 33.
• Hein ? and Quoi ? alone are requests for repair (§3); Comment ça ? asks for justification.

11. French dialogue corpora and dialogue acts

Dialogue-act schemes — DAMSL, SWBD-DAMSL, the ISO 24617-2 standard — are described in the English chapter (§10) and in 31 §12; the discourse markers of §4 and the feedback tokens of §3 fill exactly the feedback, turn-management and discourse-structuring dimensions of ISO 24617-2. The French corpora with dialogue or conversation, and what their licences allow:

The Claire compilation is a useful map of the field: it gathers parliamentary debates, theatre (the public-domain Théâtre classique collection among them), interviews, meetings and free conversation from more than twenty French corpora, and its authors had to split their release by licence because so many of the spoken corpora are shared for research only (Louradour et al. 2024). The same constraint holds for Mova: most French conversation data can be used for measurement but not built into what is released (Data and model licences).

No French corpus we hold is annotated with dialogue acts in the ISO 24617-2 sense under an open licence; Prévot and colleagues annotated communicative functions of feedback in the CID (prevot-2015-annotation-classification-french-feedback-communicative, CC BY-NC-SA 3.0: informs only).

How Mova handles it

Nothing in Mova reads French dialogue today. Checked against the code on 2026-11-06:

• Say frames and quotes are English. The frame reader recognises speech by English verbs of saying (SAY and SAY_MORE in prag/src/frames.rs: say, tell, ask, answer, reply, shout, cry…, 43 lemmas, plus VerbNet communication classes) and expects the English structure, where the reporting verb is the root and the quote its ccomp or a parataxis. The French structure of GSD (§8.2) — the quote as root, the reporting verb as parataxis:insert with an inverted subject — is not read.
• Guillemets are not quotes for Mova. The quote tracker of the pragmatics labeller (quote_spans in prag/src/feats.rs) opens and closes direct speech on “ ” " ‘ ’ and ', and carries open quotes across the sentences of a paragraph. It does not know « », nor the dialogue dash at the head of a line. On a French novel, every reply would be read as narration (voice=narrator). When French text is passed through en xpos-fill, « and » are tagged as English opening and closing quotes (`` and ''), which is right for the tag but does not reach the quote tracker, which reads the raw text.
• The labels are language-neutral; the features are not. The label set of prag/src/schema.rs (form, act, indirect, hedge, polarity, voice) has no English in it: voice (narrator / character / none) and the acts (greet, thank, agree, disagree, ask, request…) apply to French dialogue as they stand. The features that predict them are English: the word cues (please, the speech verb), the tree from the English model, the quote tracker.
• The dialogue layer is English. world/src/talk.rs classifies a user's turn with English lists — greetings (hi, hello, hey), thanks (thanks, thank you), apologies (sorry), question starts (what, who, do, can…), request starts (please, stop, tell…), praise and blame words — and its politeness detectors are the English strategies of Danescu-Niculescu-Mizil et al. (2013). A French turn (Bonjour, tu peux me dire où on en est ?) gets none of them.
• Questions under discussion are English. world/src/qud.rs types questions by English wh-words and English place prepositions and resolves he/she/it/they; French qui, où, quand, pourquoi, comment, est-ce que and inversion are not there.
• Coreference by speaker (the speaker sieve in coref/src/sieve.rs) can use a # speaker comment, which Rhapsodie has on every sentence; but the sieve's pronoun lists are English (I, we, you), so French je, nous, tu, vous, on are not linked to the speaker (35).
• The trees. French reaches Mova only as UD trees, read by the English parser after en xpos-fill (LAS 87.00 with the parser's own tags on GSD, train → dev, 2026-10-22; README). The parser has been trained and measured on GSD, which has almost no discourse (49) and no reparandum; it has not been measured on Rhapsodie or ParisStories, where these relations are 4–8% of all words, and their head direction for reparandum (§6) is the opposite of the English one. The Mova dialect's decision to keep discourse for interjections only (dialects/mova/seeds.md#discourse-adverbs) matches GSD, not the spoken French treebanks; converting those would need a rule that turns ADV and verbal discourse into advmod and parataxis. The French spoken subtypes (discourse:filler, discourse:tag, conj:reform) are recorded as candidates and not used (dialects/mova/seeds.md#spoken-subtypes).

What is being built. The Eel, the worker that improves Mova's trees and pragmatics, has a round kind prag-fr for French pragmatics in its round schedule (tools/eel/src/schedule.rs, tools/eel/src/native.rs, in development on 2026-11-06): a native CPU tool of Mova's own is run on a pinned French dev sample, its candidate is installed only on a measured gain, and its proposals are queued for review, never admitted as gold. The French pragmatics tool that such a round would run is not yet written, and there is no French measurement. Above the per-language layers, a language-independent meta-pragmatics — the acts, voices, feedback and turn functions of this chapter described once, with each language supplying its own cues (euh and um, hein and right?, « » and “ ”, the inverted incise and the English said the Rabbit) — is planned; nothing of it is implemented. The language-neutral label set of prag/src/schema.rs is the natural place for it to start.

For dialogue, a French layer would need, in order of cost:

1. Counts from the trees (§§4–7 of this chapter): discourse, reparandum, vocative, parataxis:insert and parataxis:parenth per text, as register features; no lexicon needed.
2. Quotes: « » and the dialogue dash in the quote tracker; the GSD incise structure (quote root → parataxis:insert → nsubj) in the frame reader; the French reporting verbs (dire, demander, répondre, ajouter, expliquer, déclarer, affirmer, préciser, s'écrier, murmurer…).
3. A closed French marker lexicon with the functions and positions of §4.2, checked against the counts of §4.3–4.4; the feedback inventory of §3; the answer particles oui / si / non.
4. French lists for the dialogue layer: greetings, thanks and their answers, apologies, question words, request formulas (from 31 and 32).
5. A gold set: there is none under an open licence for French dialogue acts; a small one built on public-domain plays and novels and on the CC BY-SA spoken treebanks would be the yardstick.

Sources

• Bally, C. 1909. Traité de stylistique française, I–II. Heidelberg: Winter. bally-1909-traite-de-stylistique-fran (public domain)
• Bally, C. 1912. Le style indirect libre en français moderne. Germanisch-Romanische Monatsschrift 4. Not in the library — wanted.
• Tobler, A. 1905. Mélanges de grammaire française, trans. M. Kuttner and L. Sudre. Paris: Picard. Ch. XXXIX. tobler-1905-melanges-de-grammaire-fran (public domain)
• Uro, R., Rilliard, A., Doukhan, D., Tahon, M., Laurent, A. 2024. Évaluation perceptive de l'anticipation de la prise de parole lors d'interactions dialogiques en français. JEP 2024. uro-2024-evaluation-perceptive-de-l-anticipation (CC BY 4.0)
• Uro, R. et al. 2024. Annotation of Transition-Relevance Places and Interruptions for the Description of Turn-Taking in Conversations in French Media Content. LREC-COLING. uro-2024-annotation-transition-relevance-places-interruptions (CC BY-NC 4.0: informs only)
• Prévot, L., Bertrand, R. et al. 2013. A quantitative view of feedback lexical markers in conversational French. SIGDIAL. prevot-2013-quantitative-view-feedback-lexical-markers (CC BY-NC-SA 3.0: informs only)
• Prévot, L. et al. 2015. Annotation and classification of French feedback communicative functions. prevot-2015-annotation-classification-french-feedback-communicative (CC BY-NC-SA 3.0: informs only)
• Herreño Castañeda, P., Sillaire, M. 2025. Annotation de marqueurs discursifs : le cas de la désambiguïsation de après. RECITAL 2025. castaneda-2025-annotation-de-marqueurs-discursifs-le (CC BY 4.0)
• Sow, F. et al. 2025. État de l'art sur les marqueurs discursifs en traitement automatique des langues. sow-2025-etat-de-l-art-sur (CC BY 4.0)
• Kahane, S. et al. 2021. Annotation guidelines of UD and SUD treebanks for spoken corpora: a proposal. TLT/SyntaxFest. kahane-2021-annotation-guidelines-ud-sud-treebanks (CC BY 4.0)
• Botero-Garcia, M. P. et al. 2025. An intonosyntactic treebank for spoken French: What is new with Rhapsodie? garcia-2025-intonosyntactic-treebank-spoken-french-what (CC BY 4.0)
• Hunter, J. et al. 2024. SUMM-RE: A corpus of French meeting-style conversations. TALN 2024. hunter-2024-summ-re-corpus-french-meeting (CC BY 4.0)
• Louradour, J. et al. 2024. Claire: Large Language Models for Spontaneous French Dialogue. TALN 2024. louradour-2024-claire-large-language-models-spontaneous (CC BY 4.0)
• Rennard, V. et al. 2023. FREDSum: A Dialogue Summarization Corpus for French Political Debates. Findings of EMNLP. rennard-2023-fredsum-dialogue-summarization-corpus-french (CC BY 4.0)
• Grabar, N., Eshkol-Taravella, I. 2016. Prédiction automatique de fonctions pragmatiques dans les reformulations. grabar-2016-prediction-automatique-de-fonctions-pragmatiques (CC BY 4.0); Detection of Reformulations in Spoken French, LREC 2016, grabar-2016-detection-reformulations-spoken-french (CC BY-NC 4.0)
• Bigi, B., Priego-Valverde, R. et al. 2016. Laughter in French Spontaneous Conversational Dialogs. LREC. bigi-2016-laughter-french-spontaneous-conversational-dialogs (CC BY-NC 4.0: informs only)
• Gorisch, J. et al. 2014. Aix Map Task corpus. LREC. gorisch-2014-aix-map-task-corpus-french (CC BY-NC 4.0: informs only)
• Devillers, L. et al. 2004. The French MEDIA/EVALDA Project. LREC. devillers-2004-french-media-evalda-project-evaluation (CC BY-NC 4.0: informs only)
• Müller, C., Plank, B. 2024. IndirectQA. muller-2024-indirectqa-understanding-indirect-answers-implicit (CC BY-NC 4.0: informs only)
• Clark, H. H., Schaefer, E. F. 1989. Contributing to Discourse. papers/c/cl/clark-1989-contributing-discourse; Clark, H. H., Brennan, S. E. 1991. Grounding in Communication. papers/c/cl/clark-1991-grounding-communication
• Ginzburg, J., Mazzocconi, C., Tian, Y. 2020. Laughter as language. Glossa. ginzburg-2020-laughter-as-language
• Floyd, S., Rossi, G., Enfield, N. J. (eds.) 2020. Getting others to do things. Language Science Press. books/e/en/enfield-2020-getting-others-things-pragmatic
• Sacks, H., Schegloff, E. A., Jefferson, G. 1974. A simplest systematics for the organization of turn-taking for conversation. Language 50(4). Not open — wanted.
• Universal Dependencies French treebanks, UD 2.18: ParisStories, Rhapsodie, GSD (CC BY-SA 4.0), PUD (CC BY-SA 3.0), Sequoia (LGPL-LR), ParTUT (CC BY-NC-SA 4.0, counts only). Counts made on 2026-11-06 from UD_French-*, all splits together, one count per syntactic word, punctuation in the denominator; sentences quoted by sent_id.

This page in Markdown → https://md.mova.elamur.ai/free/en/docs/languages/fr/38-dialogue-and-discourse-markers/