Elamur: **Mova** · [Parse](https://mova.elamur.ai/free/en/parse/) · [Dictionary](https://md.dictionary.elamur.ai/free/en/) · [Translate](https://md.translate.elamur.ai/free/en/) · [Research](https://md.research.elamur.ai/free/en/) · [Groupants](https://md.groupants.elamur.ai/free/en/)

[Mova](https://md.mova.elamur.ai/free/en/) · [this page in HTML](https://mova.elamur.ai/free/en/docs/languages/fr/30-pragmatics-overview/)

[home](https://md.mova.elamur.ai/free/en/) / [docs](https://md.mova.elamur.ai/free/en/docs/) / [languages](https://mova.elamur.ai/free/en/docs/languages/) / [fr](https://md.mova.elamur.ai/free/en/docs/languages/fr/) / 30. Pragmatics: an overview

# 30. Pragmatics: an overview

**On this page**

1. [1. The chapters of the pragmatics part](#1-the-chapters-of-the-pragmatics-part)
2. [2. Layers of meaning](#2-layers-of-meaning)
3. [3. Context](#3-context)
4. [4. Where French grammar carries pragmatic meaning](#4-where-french-grammar-carries-pragmatic-meaning)
5. [5. From the tree to pragmatics: the cues in UD French](#5-from-the-tree-to-pragmatics-the-cues-in-ud-french)
6. [6. Sentence types in the treebanks](#6-sentence-types-in-the-treebanks)
7. [7. Resources for French pragmatics](#7-resources-for-french-pragmatics)
8. [8. What the open research says](#8-what-the-open-research-says)
9. [9. Pragmatics in Mova](#9-pragmatics-in-mova)
10. [Sources](#sources)

Pragmatics is the study of meaning in use: what a speaker does with a sentence, what they mean beyond what they say, what they take for granted, whom they are talking to, and how sentences hang together in a text or a conversation. This chapter opens the pragmatics part of the French manual. It sets out the layers of meaning that chapters 31–38 describe, the context they depend on, the places where French grammar itself carries pragmatic meaning (the *tu*/*vous* choice, the conditional of politeness, the clitics, the dislocated topic, the discourse particles), the cues that a Universal Dependencies tree gives for them, and what Mova does with French pragmatics: today nothing in its pragmatic layer reads French, and a French layer is being built for the Eel (§9). Every count is on the UD 2.18 French treebanks and was made on 2026-11-06.

Contents

1. [The chapters of the pragmatics part](#1-the-chapters-of-the-pragmatics-part)
2. [Layers of meaning](#2-layers-of-meaning)
3. [Context](#3-context)
4. [Where French grammar carries pragmatic meaning](#4-where-french-grammar-carries-pragmatic-meaning)
5. [From the tree to pragmatics: the cues in UD French](#5-from-the-tree-to-pragmatics-the-cues-in-ud-french)
6. [Sentence types in the treebanks](#6-sentence-types-in-the-treebanks)
7. [Resources for French pragmatics](#7-resources-for-french-pragmatics)
8. [What the open research says](#8-what-the-open-research-says)
9. [Pragmatics in Mova](#9-pragmatics-in-mova)

---

## 1. The chapters of the pragmatics part

| file | what it covers |
| --- | --- |
| 30-pragmatics-overview.md | this chapter: layers of meaning, context, French carriers, the tree cues, the sentence types, Mova's position |
| [31. Speech acts](https://md.mova.elamur.ai/free/en/docs/languages/fr/31-speech-acts/) | what sentences do: assertions, questions, requests, orders, offers, promises; the imperative, the conditional, *vouloir* and *pouvoir* |
| [32. Politeness and hedging](https://md.mova.elamur.ai/free/en/docs/languages/fr/32-politeness-and-hedging/) | *tu* and *vous*, address, formulas, the conditional of politeness, hedges, letters and e-mails |
| [33. Implicature and presupposition](https://md.mova.elamur.ai/free/en/docs/languages/fr/33-implicature-and-presupposition/) | what is meant but not said, and what is taken for granted |
| [34. Irony and figurative language](https://md.mova.elamur.ai/free/en/docs/languages/fr/34-irony-and-figurative-language/) | irony, litotes, euphemism, metaphor, idioms |
| [35. Reference and coreference](https://md.mova.elamur.ai/free/en/docs/languages/fr/35-reference-and-coreference/) | referring expressions, clitics, *y*/*en*, *on*, *ça*, *ce dernier*, associative anaphora, dislocation; Democrat and ANCOR; the `coref` crate on French |
| [36. Discourse relations and coherence](https://md.mova.elamur.ai/free/en/docs/languages/fr/36-discourse-relations/) | coherence relations and French connectives (*car*, *parce que*, *puisque*, *mais*, *or*, *pourtant*, *puis*); ANNODIS, FDTB, LEXCONN; tense in narrative |
| [37. Register and variation](https://md.mova.elamur.ai/free/en/docs/languages/fr/37-register-and-variation/) | the label of a text (language, dialect modifier, style of speech); *soutenu* to *argotique*, verlan, speech and writing, official, literary, web French, song lyrics, house styles and norms |
| [38. Dialogue, discourse markers and reported speech](https://md.mova.elamur.ai/free/en/docs/languages/fr/38-dialogue-and-discourse-markers/) | turn-taking, feedback, *bon*, *ben*, *bah*, *quoi*, *hein*, *genre*, *du coup*, *euh*; incises and free indirect discourse; dialogue corpora |

The syntax chapters (12 and 20–29) are the base. Every pragmatic reading in a UD-based system starts from a tree, and French grammar leaves a good deal of the work to word order, clitics and particles, so the syntax chapters are cited throughout. The varieties of French (Québec, Belgium, Switzerland, Africa) are in [chapter 8](https://md.mova.elamur.ai/free/en/docs/languages/fr/08-national-varieties/); chapter 37 treats the registers and the styles of speech.

## 2. Layers of meaning

One French utterance carries several layers at once. Take a request overheard at a counter, *Vous auriez l'heure, s'il vous plaît ?* ("Would you have the time, please?").

| layer | what it is | in the example |
| --- | --- | --- |
| **sentence meaning** | what the words and the structure say, independent of use | a yes/no question about whether the hearer possesses the time |
| **force** | the act performed (ch. 31) | a request for the time; the conditional *auriez* and *s'il vous plaît* soften it |
| **implicature** | what the speaker means by saying it (ch. 33) | "tell me what time it is", not "answer yes or no" |
| **presupposition** | what must already be taken for granted (ch. 33) | that there is a time to be known and that the hearer is a person addressed formally |
| **social meaning** | the relation and the stance the form signals (ch. 32, 37) | *vous* and the conditional: distance, deference, a stranger |
| **reference** | who and what is meant by each expression (ch. 35) | *vous* is the hearer; *l'heure* is the current time |
| **discourse** | how the utterance fits what came before (ch. 36, 38) | an opening move, after *Excusez-moi* |

The layers are separable. *Vous auriez l'heure ?* and *Quelle heure est-il ?* have almost the same force and different social meaning. *Tu parles !* can mean "you are speaking" or, depending on intonation and context, "you bet" (agreement) or "nonsense" (disagreement): the sentence meaning does not decide the force. French linguistics distinguishes the **énonciation** (the act of uttering, with its speaker, hearer, time and place) from the **énoncé** (the utterance produced), and French pragmatic description is often framed that way. The large modern reference grammars (Riegel et al., Grevisse and Goosse, the *Grande Grammaire du français*) are known to us only by their cards, so nothing in this part rests on them (see [Sources](https://md.mova.elamur.ai/free/en/docs/languages/fr/sources/)).

The older grammars already see the problem. Brunot (1922, `brunot-1922-la-pensee-et-la`) objects to the school habit of treating the imperative as the grammar of command: a command, a request or a plea is also expressed by futures and infinitives with the sense of orders, by bare words thrown out as commands (his examples: *Silence !*, *Dehors !*, *Arme sur l'épaule !*), by exclamations that demand, advise or beg, and by the conditionals and imperfects of timid requests, "all the subtleties of politeness" (our examples of the last: *Je voulais vous demander…*, *Pourriez-vous…*). All that, he says, cannot be tied to the poor mutilated form called the imperative. That is the point of view of the whole part: the same act is made by many forms, and the same form makes many acts.

## 3. Context

Meaning in use depends on what the speaker and the hearer share. The standard parts of context, with the French expressions that depend on each:

| part of context | what it holds | French expressions that depend on it |
| --- | --- | --- |
| **the speaker, the hearer and their relation** | who speaks to whom, their rank, age, intimacy | *tu*/*vous*, titles (*Madame*, *Maître*), first names, the imperative or the conditional, *s'il vous plaît* |
| **the time and place of the utterance** | the deictic centre | *maintenant*, *hier*, *demain*, *ici*, *là*, *ce soir*, the *passé composé* versus the *passé simple* |
| **the discourse so far** | what has been said and by whom | pronouns (*il*, *elle*, *ça*), definite descriptions, *ce dernier*, *celui-ci*, ellipsis, *oui*/*si* |
| **the common ground** | what both take as known | presupposition triggers (*encore*, *déjà*, *ne … plus*, *cesser de*), *bien sûr*, *tu sais* |
| **the setting and the genre** | where and why they talk: a shop, an exam, a letter, a forum, a lecture | the register (*soutenu*, *courant*, *familier*), the opening and closing formulas |
| **the channel** | speech or writing; face to face or distant | intonation and gesture in speech; punctuation, capitals and emoticons in writing |

French differs from English in several of these. It cannot avoid the relation: every utterance addressed to one person forces a choice between *tu* and *vous*, which English does not. Its gender marking makes the speaker's and the hearer's gender visible in the agreement of an adjective or a participle (*Je suis fatiguée*), so a text can reveal who is speaking. Its spoken and written registers are further apart than the English ones, to the point that Mova's own spoken French treebanks show a different grammar (*on* for *nous*, *ça* for *cela*, *ne* dropped, dislocation: [chapter 37](https://md.mova.elamur.ai/free/en/docs/languages/fr/37-register-and-variation/)). And the oral tradition is strong: a French pragmatics grounded in text alone loses what intonation and gesture do, which is a good deal in the cases of irony and of the discourse particles.

The hearer is never given, but is always assumed. Chapter 32 shows that **the choice of address** is the most reliable pragmatic signal French grammar offers; the open research on learners (Sperlich et al. 2016, `sperlich-2016-interaction-politeness-systems-korean-learners`) shows the choice of *tu* and *vous* in a request is a matter of both explicit rules and tacit knowledge of the situation: their Korean learners handled *tu* better than *vous*.

## 4. Where French grammar carries pragmatic meaning

The English manual's pragmatics part notes that pragmatic categories are signalled by grammatical form, but not one-to-one. In French the grammar does a good deal of the signalling, and it is worth listing before the chapters take the points one by one.

| French form | pragmatic effect | chapter |
| --- | --- | --- |
| *tu* / *vous* | the relation between speaker and hearer: intimacy, equality, or distance and deference | 32 |
| conditional (*je voudrais*, *pourriez-vous*) | softens requests and wishes; also marks reported or doubtful information (*le président serait malade*) | 31, 32, 33 |
| imperative | the plain form of an order, a request, an invitation or a piece of advice, depending on context | 31 |
| *s'il vous plaît* / *s'il te plaît*, *merci de*, *veuillez* | explicit politeness markers and formulaic requests | 31, 32 |
| inversion versus *est-ce que* versus intonation | the register of a question | 26, 31, 37 |
| *ne … pas* versus *pas* alone | the register: written and formal versus spoken | 27, 37 |
| *on* for *nous* | informal inclusive "we" (speech) versus *nous* (writing) | 35, 37 |
| clitic dislocation (*Moi, je … ; Lui, il …*) | topic, contrast, emphasis, in speech | 29, 35 |
| clefts (*C'est … qui*) and pseudo-clefts (*Ce que … c'est*) | focus, exhaustivity | 29, 33 |
| the subjunctive after *vouloir que*, *il faut que* | the mood of the dependent clause is conventional, but the choice of the main verb signals a directive | 24, 31 |
| *oui* / *si* / *non* | answers to affirmative and negative questions: *si* corrects a negative question | 31, 38 |
| *bon*, *ben*, *quoi*, *hein*, *voilà*, *du coup*, *en fait*, *enfin* | discourse markers: turn management, stance, repair | 36, 38 |
| *donc*, *alors*, *mais*, *pourtant*, *or* | connectives with pragmatic uses beyond their logical ones | 36 |
| gender agreement (*Je suis contente*, *On est allées*) | the speaker's or the referent's gender | 35 |
| literary tenses (*passé simple*) | register and genre: narration, formal writing | 13, 37 |

The same form is often used for more than one act: **the conditional of politeness** (*Je voudrais un café*) and **the conditional of hearsay** (*Il y aurait trois morts*) are one verb form and two unrelated pragmatic functions.

## 5. From the tree to pragmatics: the cues in UD French

A UD tree does not carry pragmatic labels, but it carries cues. The ones that matter for French:

| cue in the tree | what it can signal | example |
| --- | --- | --- |
| `Mood=Imp` on the root | a directive act: command, request, invitation | *Soyez concis, courtois et pertinents.* (fr-ud-dev_00959) |
| `Mood=Cnd`, with a first- or second-person subject | a softened request or wish | *je voudrais vous parler d'amour* (ParisStories_2019_devoirPhilosophie_14) |
| `Mood=Cnd` with a third-person subject and a news context | hearsay or uncertainty (see ch. 33) | *aurait été* |
| the relation `vocative` | a term of address: who is spoken to | *Madame*, *maître*, *mamie* |
| the relation `discourse` | an interjection, a particle or a marker | *euh*, *bon*, *ben*, *hein*, *enfin* |
| the relation `dislocated` (and `dislocated:subj`, `dislocated:obj`) | a topic detached from the clause | *ça, c'est des spécialités…* (Rhap_D0006-125) |
| the relation `parataxis` and `parataxis:parenth` | a parenthetical, a reporting clause | *voilà*, *je pense*, *tu vois* |
| the relation `reparandum` | a repaired or abandoned stretch, a disfluency | *il é~, il a pas mal réagi* |
| `Person=2` pronouns, with the `Polite` feature | the hearer; and, on *vous*, whether the address is polite | *vous* with `Polite=Yes` |
| `PronType=Int` and a final `?` | a question | *Pouvez-vous nous en dire plus ?* (fr-ud-train_02355) |
| `expl:subj` | an impersonal or an expletive subject: no reference | *Il y a…*, *C'est…* |

French treebanks do mark some of these more fully than English ones. The UD French guidelines use `Polite=Yes|No|Unknown` on *vous*, in the MISC column of the treebank (**not** in FEATS), and the Rhapsodie treebank annotates it on every *vous*. The values are: `Yes`, a singular *vous* of politeness; `No`, a real plural; `Unknown`, ambiguous. Counted on 2026-11-06 (tokens of the pronoun *vous*):

| treebank | *vous* (PRON) | with a `Polite` value | `Yes` | `No` | `Unknown` |
| --- | --- | --- | --- | --- | --- |
| GSD | 276 | 100 | 4 | 2 | 94 |
| ParisStories | 52 | 50 | 7 | 15 | 28 |
| Rhapsodie | 426 | 426 | 50 | 1 | 375 |
| Sequoia | 130 | 73 | 9 | 0 | 64 |

Two readings follow. First, the annotation is incomplete in GSD and Sequoia, where it covers about a third and a half of the tokens. Second, where it is made, **most tokens are `Unknown`**: one *vous* alone often cannot be classified as a polite singular or a plural without the situation. Only 11% of the annotated *vous* tokens in the four treebanks together (70 of 649) are marked as polite singular. The values are sparse labels, not a resource large enough to train an address classifier. Chapter 32 returns to them.

The dependency relations are counted in the treebanks below (all splits, syntactic words with punctuation as the denominator, 2026-11-06):

| relation | GSD (400 385) | ParisStories (42 789) | Rhapsodie (44 234) | Sequoia (70 545) | ALTS, 16th c. (68 088) | FQB (23 896) |
| --- | --- | --- | --- | --- | --- | --- |
| `discourse` | 49 | 3 246 | 1 919 | 3 | 15 | 0 |
| `vocative` | 15 | 16 | 41 | 53 | 14 | 0 |
| `reparandum` | 0 | 1 035 | 1 205 | 0 | 0 | 0 |
| `dislocated` and its subtypes | 75 | 354 | 250 | 17 | 53 | 139 |
| `parataxis:parenth` | 0 | 27 | 39 | 0 | 0 | 0 |
| `expl:subj` | 931 | 314 | 425 | 237 | 0 | 333 |
| `INTJ` words | 87 | 2 391 | 1 910 | 5 | 4 | 3 |

`discourse` and `reparandum` are almost absent from the written treebanks and make up a large part of the spoken ones: `discourse` is 759 per 10 000 tokens in ParisStories and 434 in Rhapsodie, against 1.2 in GSD. `dislocated` (with `dislocated:subj`, `dislocated:obj`, `dislocated:mod`) is 18 per 10 000 tokens in GSD and 83 and 57 in ParisStories and Rhapsodie respectively, and the spoken ones are nearly all subject dislocations (302 of 354, 203 of 250: *moi, je …*, *ça, c'est …*; [chapter 35](https://md.mova.elamur.ai/free/en/docs/languages/fr/35-reference-and-coreference/)). The FQB count is the question-bank construction (*Ces enfants, on ne les voit jamais ?*).

## 6. Sentence types in the treebanks

The shape of the sentence is a first approximation of the speech act. Counted from the final punctuation and the root of each sentence (UD 2.18, 2026-11-06):

| treebank | sentences | mean length (tokens) | final `?` | final `!` | root is a verb | root is not a verb |
| --- | --- | --- | --- | --- | --- | --- |
| GSD | 16 342 | 24.5 | 1.8% | 2.2% | 73.3% | 26.7% |
| ParisStories | 2 776 | 15.4 | 5.2% | 3.0% | 68.7% | 31.3% |
| Rhapsodie | 3 209 | 13.8 | 5.7% | 0.0% | 60.1% | 39.9% |
| Sequoia | 3 099 | 22.8 | 1.5% | 0.7% | 63.3% | 36.7% |
| PUD | 1 000 | 24.7 | 1.1% | 0.4% | 85.8% | 14.2% |
| ALTS (16th-century legal French) | 2 026 | 33.6 | 0.4% | 0.2% | 92.7% | 7.3% |
| FQB (questions) | 2 289 | 10.4 | 97.3% | 0.0% | 59.7% | 40.3% |

Four points. First, **questions** are rare in running text (about 2% in written French, about 5% in the spoken treebanks) and are the whole point of the FQB treebank (a question bank, 97% final question marks). Second, **exclamations** in `!` are about 2% in the web texts of GSD (reviews and blogs, where *Bravo !* and *Attention !* are frequent) and in the spoken treebanks depend on the transcription convention: Rhapsodie has a single sentence ending in `!`, so the absence of exclamation there is a convention of the transcription, not a fact of speech. Third, **verbless sentences** are common: from a quarter of the GSD sentences to two fifths of Rhapsodie. In French they are often titles, captions, fragments of reviews (*Très bon accueil.*), answers (*Oui.*, *Pas du tout.*) and, in speech, the stretches before a repair. Any pragmatic labelling that starts from a finite verb at the root misses a quarter to a third of the sentences. Fourth, the spoken sentences are about 40% shorter than the written ones (13.8 and 15.4 tokens against 24.5), but this reflects the segmentation of the transcription as much as the speech.

The punctuation of the spoken treebanks is the annotators' (Rhapsodie and ParisStories transcripts have no original punctuation); the question marks there are decisions of the transcribers, informed by intonation (Kahane et al. 2021, `kahane-2021-annotation-guidelines-ud-sud-treebanks`, discuss the principles of segmenting speech into "sentences" in UD and SUD treebanks).

**The mood of the root verb.** For the root verbs with a finite form, the moods are, in GSD: `Mood=Ind` 23 930 finite verbs and auxiliaries in the whole treebank (not only roots), `Mood=Cnd` 431, `Mood=Sub` 379, `Mood=Imp` 137. In the spoken treebanks the imperative is rare: 31 in ParisStories and 65 in Rhapsodie, where *dire* alone accounts for 18 of the 65 (*dites-moi*, *dites*). The conditional is 431 in GSD (1.1 per 1 000 tokens), 47 in ParisStories (1.1 per 1 000), 89 in Rhapsodie (2.0 per 1 000). By lemma, in GSD the conditional of *avoir*, *être*, *pouvoir* and *devoir* make up the large majority, with the 3rd person overwhelmingly dominant (111 *avoir*, 73 *être*, 66 *pouvoir*, 38 *devoir* in the 3rd person): the written conditional is mostly the conditional of hypothesis, report and attenuated assertion, not of requests. The first person is 6.5% of the conditionals in GSD (28 of 431), 28% in ParisStories (13 of 47), 43% in Rhapsodie (38 of 89) and 27% in Sequoia (40 of 146): in speech and in parliamentary speech the conditional is much more often the conditional of wishes and polite requests (*je voudrais*, *j'aimerais*).

## 7. Resources for French pragmatics

The corpora that exist, with their licences as we know them. For the licence rule the manual follows [Data and model licences](https://md.mova.elamur.ai/free/en/docs/licence/DATA/): share-alike and non-commercial data may be cited, counted and tested on; none is shipped.

| resource | what it holds | licence | use here |
| --- | --- | --- | --- |
| UD French GSD | 400 k tokens, written web French (reviews, blogs, news, Wikipedia) | CC BY-SA 4.0 | counts, short examples |
| UD French ParisStories | 42.8 k tokens, spoken stories told in Paris, with `Polite`, liaison and sound links | CC BY-SA 4.0 | counts, short examples |
| UD French Rhapsodie | 44.2 k tokens, 3 h 10 of spoken French (30 monologues and 27 dialogues), with prosodic annotation (Botero-Garcia et al. 2025, `garcia-2025-intonosyntactic-treebank-spoken-french-what`) | CC BY-SA 4.0 | counts, short examples |
| UD French Sequoia, FQB | news, Europarl, Wikipedia, medical (Sequoia); questions (FQB) | LGPL-LR | counts |
| UD French ParTUT | law, Wikipedia, tales | CC BY-NC-SA 4.0 | tests only, never quoted |
| ANCOR | about 418 000 words of transcribed spoken French annotated for coreference; mostly ESLO interviews and some higher-interaction material (Grobol 2019, `grobol-2019-neural-coreference-resolution-limited-lexical`) | non-commercial share-alike (the paper is CC BY-NC-SA 3.0; the corpus licence to be confirmed from its distribution) | counts only, never quoted, see chapter 35 |
| Democrat | 58 written texts, 689 000 words, 198 000 referring expressions, 20 000 chains with at least two mentions (Lopez 2023, `lopez-2023-approches-neuronales-pour-la-detection`) | CC BY-SA 4.0 (the CorefUD conversion, whose train and dev parts are held; counts in chapter 35) | counts, short quotes; see chapter 35 |
| LEX-CONN | the lexicon of French discourse connectives with their relations | open (to be confirmed) | see chapter 36 |
| SUMM-RE | about 95 hours of spontaneous meeting conversations, 25 h manually corrected (Hunter et al. 2024, `hunter-2024-summ-re-corpus-french-meeting`), 20 h discourse-annotated (Prévot et al. 2025, `prevot-2025-segmenting-large-french-meeting-corpus`) | the paper is CC BY 4.0 | see chapter 38 |
| Claire | French conversation and theatre transcripts for training language models (Louradour et al. 2024, `louradour-2024-claire-large-language-models-spontaneous`) | the paper is CC BY 4.0 and says that models, data and code are published under open licences; an earlier dataset paper (Hunter et al. 2023) is CC BY-NC-SA 4.0, so the licence is to be confirmed | see chapter 38 |
| TREMoLo-Tweets | 228 505 tweets (6 M words) labelled for casual, neutral and formal register (Mekki et al. 2021, `mekki-2021-tremolo-tweets-multi-label-corpus`) | the paper is CC BY 4.0 | see chapter 37 |
| IndirectQA | French and Spanish subtitles with indirect answers to polar questions (Müller & Plank 2024, `muller-2024-indirectqa-understanding-indirect-answers-implicit`) | the paper is CC BY-NC 4.0 | informs only, see chapter 33 |

We hold the texts of the open papers; of the corpora they describe we hold the UD treebanks and the CorefUD conversion of Democrat. Where a corpus licence is marked "to be confirmed", the paper is the source of the description and nothing is claimed about the licence beyond it. The list is a set of leads, not of holdings; [Sources](https://md.mova.elamur.ai/free/en/docs/languages/fr/sources/) keeps the table with the slugs.

## 8. What the open research says

A short orientation to the French papers on pragmatic topics that we hold with open licences, chapter by chapter; each chapter gives what the paper found.

- **Politeness and address**: Sperlich, Leem & Ahn 2016 on how Korean learners acquire *tu* and *vous* in requests; Lancien 2020 (`lancien-2020-variation-stylistique-en-francais-quebecois`) on how Quebec French speakers adapt their speech to the identity of the interlocutor. Chapter 32.
- **Modality and stance**: Colli & Battistelli 2025 (`colli-2025-exploration-de-la-modalite-en`) and Colli et al. 2024 (`colli-2024-modal-sense-classifier-french-modal`) on modal profiles of spoken and written French and on the senses of *pouvoir*; Deng 2023 (`deng-2023-variationist-analysis-two-french-attitude`) on *je pense* and *je crois*. Chapters 31, 32, 33.
- **Presupposition**: Garassino et al. 2024 (`garassino-2024-politicians-vs-chatgpt-study-presuppositions`). Chapter 33.
- **Idioms and wordplay**: Beauchemin et al. 2026 (`beauchemin-2026-idiom-understanding-as-tool-measure`), Noël et al. 2026 (`noel-2026-eiffel-novel-benchmark-measure-bias`), Bezançon et al. 2025 (`bezancon-2025-forbidden-fruit-is-sweetest-annotated`) and 2023 (`bezancon-2023-reconnaissance-de-defigements-dans-des`); Zhou et al. 2024 (`zhou-2024-une-comparaison-de-l-intonation`) on ironic intonation. Chapter 34.
- **Reference**: Grobol et al. 2017 (`grobol-2017-interoperable-annotation-co-references-democrat`), Grobol 2019, Oberle 2019 (`oberle-2019-detection-automatique-de-chaines-de`), Bourgois & Poibeau 2025 (`bourgois-2025-elephant-coreference-room-resolving-coreference`), Milintsevich 2025 (`milintsevich-2025-impact-asr-transcriptions-french-spoken`), Guenoune et al. 2020 (`guenoune-2020-dataset-anaphora-analysis-french-emails`), Amsili & Seminck 2017 (`amsili-2017-google-proof-collection-french-winograd`), Cunha & Heidinger 2026 (`cunha-2026-animacy-restrictions-without-animacy-features`). Chapter 35.
- **Discourse relations**: Sow et al. 2026 (`sow-2026-relex-common-denominator-connectives-discourse`), Laali & Kosseim 2017 (`laali-2017-automatic-mapping-french-discourse-connectives`), Rauh et al. 2023 (`rauh-2023-multilingual-connective-database-aligning-german`), Blochowiak et al. 2020 (`blochowiak-2020-what-type-subjectivity-lies-behind`), Pastor et al. 2024 (`pastor-2024-la-reconnaissance-automatique-des-relations`). Chapter 36.
- **Register and variation**: Mekki et al. 2021 and 2025, Lecorvé et al. 2018 (`lecorve-2018-construction-conjointe-d-un-corpus`), Kogkitsidou & Antoniadis 2016 (`kogkitsidou-2016-l-architecture-d-un-modele`). Chapter 37.
- **Dialogue and discourse markers**: Lee et al. 2020 (`lee-2020-tude-comparative-de-correlats-prosodiques`), Sow 2025 (`sow-2025-etat-de-l-art-sur`), Herreño Castañeda & Sillaire 2025 (`castaneda-2025-annotation-de-marqueurs-discursifs-le`), Uro et al. 2024 (`uro-2024-evaluation-perceptive-de-l-anticipation`), Dinkar et al. 2022 (`dinkar-2022-fillers-spoken-language-understanding-computational`). Chapter 38.

Two public-domain works give the historical perspective. Nyrop's *Grammaire historique de la langue française*, vol. 5 (`nyrop-1899-grammaire-historique-de-la`), has the fullest account we hold of address (*tu*, *vous*) and of the replacement of the first person by *on*. Brunot's *La pensée et la langue* (`brunot-1922-la-pensee-et-la`) is a notional grammar that starts from what the speaker wants to say (to order, to ask, to deny) rather than from the parts of speech, and is the nearest the public-domain literature gets to a pragmatic grammar of French.

## 9. Pragmatics in Mova

### 9.1 The English design, and what carries over

Mova's pragmatic layer is built for English and is described in the [English chapter 30](https://md.mova.elamur.ai/free/en/docs/languages/en/30-pragmatics-overview/). In brief: a UD tree is read into **frames** (who did what to whom, with links of cause and sequence); each sentence gets one of a closed set of **labels** (act, polarity, uncertainty, voice, and others); small rule modules read **implicatures and presuppositions** (`prag/src/infer.rs`), **irony candidates** (`prag/src/irony.rs`), **reference games** (`prag/src/rsa.rs`); a **world** keeps the discourse (questions under discussion in `world/src/qud.rs`, the dialogue acts of the family in `world/src/talk.rs`). Everything is deterministic and explained: a reading says which cue produced it. The labels are produced by a trained classifier, repaired by an agreer that follows the tree (`prag/src/agree.rs`: the sentence form comes from the tree, a root imperative with no subject becomes a request).

### 9.2 What is done for French

**Nothing in the pragmatic layer reads French.** The honest statement, checked against the code on 2026-11-06:

- The `prag`, `coref` and `world` crates take a tree from the `en` annotator and use English word lists: the verbs of saying and knowing, the scalar items, the polar words of the irony detector, the positive and negative situations, the pronoun lexicons, the connective lists. None has a French list.
- French is read only by the English tree parser, after `en xpos-fill` has mapped the treebank's tags to the tag set `en` knows ([README](https://md.mova.elamur.ai/free/en/docs/languages/fr/)). On UD French GSD (UD 2.18, train → dev, 2026-10-22) that gives LAS 87.00 with the parser's own tags and 89.78 with gold tags. The tree is therefore good enough to read the *cues* of section 5 (a `vocative`, a `discourse` marker, a `dislocated` topic, a root `Mood=Imp`) as the English reader reads its own, but nothing consumes them for French.
- **The mood is not in the parser's input.** `en xpos-fill` writes a Penn Treebank tag computed from the UPOS and the FEATS of the treebank (`xpos_of` in `en/src/main.rs`); the English tag set has no conditional, subjunctive or imperative, so a French conditional becomes `VBZ`, `VBP` or `VBD` ([chapter 13](https://md.mova.elamur.ai/free/en/docs/languages/fr/13-verbs/)). The FEATS stay in the file, but Mova's parser reads only word forms and tags, and Mova's own tagger predicts tags only, with no French FEATS and no French lemmas. On gold treebank trees a French reader can take `Mood=Cnd` or `Mood=Imp` from the FEATS; on raw French text parsed by Mova the mood is not available at all.
- Mova has no French lexicon of discourse markers, of address terms, of speech-act verbs or of connectives, no French register lexicon (the register lexicons are `global/data/register-*.tsv`, English) and no French coreference table. The `coref` crate reads the CorefUD format (`coref/src/corefud.rs`, written for the English GUM layout of the entity fields), so the French CorefUD collections (ANCOR and Democrat) could in principle be read once their entity layout is checked; the sieves, its gender, number and animacy lexicon and its pleonastic *it* test are English.

### 9.3 What French pragmatics in Mova would need

An order, from the cheapest to the dearest, built on what exists:

1. **Cues counted from the trees.** The relations of section 5 and the forms of chapters 31–38 can be counted per treebank as features of a text: share of `discourse` words, of `vocative`, of `Mood=Cnd` with a first- or second-person subject, of *tu* against *vous*, of dislocation, of *ne* dropped. This is the same way chapter 37 reads the register of a corpus. It needs no new lexicon, only the trees.
2. **Small closed lexicons**, written once and checked against the trees: the French discourse markers and their functions (chapter 38), the address terms (chapter 32), the speech-act and modal verbs (chapters 31–32), the connectives and their relations (chapter 36, with LEX-CONN as the model), the presupposition triggers (chapter 33).
3. **French rules in the existing modules.** The agreer's tree-form rule (a root imperative with no subject is a request) carries over with `Mood=Imp`; the QUD stack carries over with *est-ce que*, inversion and the final question mark; the scalar-implicature scales need French items (*certains*, *plusieurs*, *tous*; *peut-être*, *sûrement*, *certainement*).
4. **A French coreference model.** The sieves need a French pronoun paradigm (*il*, *elle*, *ils*, *elles*, *le*, *la*, *les*, *lui*, *leur*, *y*, *en*, *on*), gender and number agreement, the *ce dernier* and *celui-ci* rules, and a test for the expletive *il* and the event *ça* ([chapter 35](https://md.mova.elamur.ai/free/en/docs/languages/fr/35-reference-and-coreference/)).
5. **A baseline on French data**: the labels cannot be scored without a French gold set. The French pragmatic resources are few (§7), and the nearest to a gold set for the speech-act or politeness labels is none.

None of these steps has been taken in the code of the pragmatic layer itself. The chapters that follow say, topic by topic, what the sub-task needs, with the open resources that exist for it.

### 9.4 What is being built: the Eel's French layer and the meta-pragmatics

The Eel is the worker that improves Mova's trees and pragmatics from its own measurements. On 2026-11-06 its development branch gives it a schedule of round kinds (`tools/eel/src/schedule.rs`): beside the English parser and autoencoder rounds there are a French parser round (`parser-fr`), a French autoencoder round (`ae-fr`, conditioned on language, dialect and style; [chapter 37](https://md.mova.elamur.ai/free/en/docs/languages/fr/37-register-and-variation/)), a **French pragmatics round (`prag-fr`)**, an English–French cross-check (`cross`) and two translation rounds. `prag-fr` is a *gated native round* (`tools/eel/src/native.rs`): a pinned executable of Mova's own, on the CPU and without a teacher, measures the installed French pragmatics state and a candidate on a pinned dev sample; the Eel checks the executable's hash, that the dev sample is the one pinned on the first round, and that no training key is a dev or test key; it installs the candidate only on a dev-metric gain, or, in measurement mode, records the value only if the round's own negative control is beaten; proposals are queued for review and never become gold by themselves. The scheduler, the gates and their tests exist; **the French pragmatics executable that the round runs is still being written**, so there are no French labels, no French metric and no French measurement yet. What it will label, and from which cues, is what chapters 31–38 lay out.

Above the language layers, a **language-independent meta-pragmatics** is planned: one inventory of acts, stances, politeness, reference, discourse relations and situational features (the style of speech among them), into which the English layer and the French layer each map their own forms. The English labels are already closed codes with no English in their values (`prag/src/schema.rs`: form, act, indirect, hedge, polarity, voice), which is the natural seed of such an inventory; but no mapping from French forms to it exists, no code implements a shared schema across languages, and nothing has been measured. The division of labour the chapters assume is this: the meta level holds what a request, a presupposition or a contrast relation *is*; the French level holds how French marks it (the conditional of *vouloir*, *encore*, *pourtant*, *tu*/*vous*).

## Sources

- Brunot, F. 1922. *La pensée et la langue*. `brunot-1922-la-pensee-et-la` (public domain).
- Nyrop, K. 1899. *Grammaire historique de la langue française*, vol. 5 (syntax). `nyrop-1899-grammaire-historique-de-la` (public domain).
- Sperlich, D., Leem, J., Ahn, E.-J. 2016. The interaction of politeness systems in Korean learners of French. `sperlich-2016-interaction-politeness-systems-korean-learners` (CC BY 4.0).
- Kahane, S. et al. 2021. Annotation guidelines of UD and SUD treebanks for spoken corpora: a proposal. `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).
- Grobol, L. 2019. Neural coreference resolution with limited lexical context and explicit mention detection for oral French. `grobol-2019-neural-coreference-resolution-limited-lexical` (CC BY 4.0).
- Lopez, C. et al. 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).
- Hunter, J. et al. 2024. SUMM-RE. `hunter-2024-summ-re-corpus-french-meeting` (CC BY 4.0). Prévot, L. et al. 2025. `prevot-2025-segmenting-large-french-meeting-corpus` (CC BY 4.0).
- Louradour, J. et al. 2024. Claire. `louradour-2024-claire-large-language-models-spontaneous` (CC BY 4.0).
- Mekki, J. et al. 2021. TREMoLo-Tweets. `mekki-2021-tremolo-tweets-multi-label-corpus` (CC BY 4.0).
- Müller, C., Plank, B. 2024. IndirectQA. `muller-2024-indirectqa-understanding-indirect-answers-implicit` (CC BY-NC 4.0: informs only).
- The rest of the papers named in section 8 are listed with their licences in [Sources](https://md.mova.elamur.ai/free/en/docs/languages/fr/sources/).
- Universal Dependencies French treebanks, UD 2.18: GSD, ParisStories, Rhapsodie (CC BY-SA 4.0), Sequoia, FQB (LGPL-LR), ALTS (sixteenth-century legal French of Normandy and Guernsey, CC BY-SA 4.0; its counts are not modern French and are given for contrast only), PUD (CC BY-SA 3.0), ParTUT (CC BY-NC-SA 4.0). Counts made on 2026-11-06 from UD_French-*, all splits together, one count per syntactic word (multiword tokens counted as their parts), punctuation included in the denominator.
- [← previous 29. Ellipsis, word order and information structure](https://md.mova.elamur.ai/free/en/docs/languages/fr/29-ellipsis-word-order-information-structure/)
- [next → 31. Speech acts](https://md.mova.elamur.ai/free/en/docs/languages/fr/31-speech-acts/)
- [↑ back to Modern Standard French (fr) — a reference manual for Mova](https://md.mova.elamur.ai/free/en/docs/languages/fr/)