🗣️ 33. Implicature and presupposition

A sentence means more than it says. Certains étudiants ont réussi suggests that not all did. Il a arrêté de fumer takes for granted that he used to smoke. Il a réussi à ouvrir la porte suggests that it was hard. This chapter describes the two main kinds of inference of this sort in French: implicatures, which the hearer draws by reasoning about why the speaker said what they said, and presuppositions, which the form of the sentence itself requires to be taken for granted. It lists the French scales and triggers, the way inferences behave under negation and questions, the French tradition of présupposé and sous-entendu, the French conventional implicatures (mais, pourtant, même, d'ailleurs), the counts of the triggers in the open treebanks (UD 2.18, 2026-11-06), what the open research has found, and what Mova does today.

Contents

1. Grice and the maxims
2. Scalar implicatures: the French scales
3. Other conversational implicatures
4. Conventional implicatures
5. Presupposition and its triggers
6. Behaviour under negation, questions and embedding
7. Présupposé and sous-entendu: the French tradition
8. Metalinguistic negation and cancellation
9. Clefts and focus
10. Triggers in the treebanks
11. What the research has found
12. How Mova handles it

1. Grice and the maxims

Grice's account starts from a cooperative principle: make your contribution as is required, at the stage at which it occurs, by the accepted purpose of the exchange. Four groups of maxims follow (French: maximes de quantité, de qualité, de relation, de manière):

An implicature is what the hearer concludes on the assumption that the speaker follows the principle. It is calculable (derived from the sentence, its context and the maxims), cancellable (it can be denied without contradiction) and non-detachable (it goes with the content, not the wording). Certains étudiants ont réussi, voire tous cancels the "not all"; Certains étudiants ont réussi ; en fait, tous the same.

Generalised implicatures arise in nearly all contexts (the scalar ones); particularised ones depend on the particular context. The classic French example of the second kind: Il y a une station-service au coin de la rue in answer to Je suis en panne d'essence. It is relevant only if the station is open, and the hearer draws the inference that it is.

2. Scalar implicatures: the French scales

Horn's scale idea: items of a scale are ordered by informativeness, and to use a weaker item implicates that the stronger ones do not hold. French scales:

The exclusive reading of ou is a well-known case. French ou is inclusive in logic and exclusive by implicature in many contexts (Café ou thé ?); ou bien and soit … soit … make the exclusive reading explicit: Soit tu viens, soit tu restes. The numeral is "at least n" in meaning and "exactly n" by implicature: Elle a trois enfants — et même quatre is a cancellation that is not contradictory.

Cancellation and suspension are explicit: voire (Certains, voire tous les participants …), même (Certains, et même la plupart …), en fait (Certains. En fait, tous.), si ce n'est tous (Certains, si ce n'est tous, ont …), ou plutôt (Il est content, ou plutôt ravi). The scalar inference is also blocked in some contexts: under negation (Il n'a pas lu certains livres does not implicate that he has read all the others), in the antecedent of a conditional (S'il lit plusieurs de ces livres, il comprendra), and when the speaker's knowledge is limited (the inference "not all" becomes "I do not know that all").

The hearer's computation. The standard model, which many experiments confirm, treats the scalar inference as a step taken after the literal reading, at a cost. The Rational Speech Acts model (Frank and Goodman 2012) derives it as the choice by a listener of the reading that a rational speaker would most likely have meant by the sentence over the alternatives. Mova implements the model for small closed games in English (see below).

3. Other conversational implicatures

3.1 Relation: indirect answers

An answer that does not literally answer the question can be read as one by relevance: Tu viens à la soirée ? — J'ai un examen demain. means no. Tu aimes ce vin ? — Il est cher. means perhaps no, or perhaps yes with a qualification. A study of indirect answers to polar questions in French and Spanish film and series subtitles (Müller & Plank 2024, muller-2024-indirectqa-understanding-indirect-answers-implicit, a non-commercial paper that informs but is not quoted) notes that such answers are natural in dialogue and still hard for conversational systems, and that the questions need not have the form of a question (an implicit polar question is a statement that invites a yes or a no). The French words that signal the inference are the marks of a dispreferred answer (Eh bien …, Ben …, Euh …) and those of indirectness (Ça dépend, Je ne sais pas, Disons que …): see chapter 38.

3.2 Quality: irony and understatement

When the speaker says what they do not believe, or says too little, the hearer infers a different meaning: Quel beau temps ! in the rain, Ce n'est pas mal for something excellent, Il fait un peu frais in a gale. These are treated in chapter 34.

3.3 Manner: marked forms

A marked form implicates a marked meaning. Il a fait en sorte que la porte soit ouverte implicates something unusual about the way, by contrast with Il a ouvert la porte. Je ne dis pas non (litotes) implicates that the speaker is rather willing. Elle a réussi à trouver une place implicates, through the verb réussir à, that there was a difficulty, a conventional implicature or a presupposition according to the analysis. A double negative of the type Ce n'est pas impossible implicates "possible, but not very probable".

3.4 Questions and politeness

A question implicates that the speaker does not know the answer and wants to; a rhetorical question violates that and implicates the opposite assertion (Qui pourrait le croire ?). The politeness of a request lies in its implicatures (Tu as l'heure ?): the hearer is to infer that the speaker is asking for the time, not whether the hearer possesses one (chapter 31).

4. Conventional implicatures

A conventional implicature is a meaning attached by convention to a particular word, which does not affect the truth conditions of the sentence but is not cancellable like a conversational implicature. French has many, mostly in connectives and adverbs (the French theory of argumentation in the language of Ducrot and Anscombre makes them its centre; it is known to us by name and cards only).

Mais is the best studied. It marks a contrast of "argument" (the first segment would lead to a conclusion that the second one denies): Il est riche mais avare (the conclusion "he is generous" suggested by "rich" is denied). A mais with two statements neither of which opposes the other is odd (*Il est riche mais grand) unless the context supplies the argument. The conventions of donc, alors, mais, car and puisque as discourse connectives are in chapter 36.

5. Presupposition and its triggers

A presupposition is information that the sentence takes for granted: its content must be part of the common ground for the sentence to be appropriate. The speaker can use it to convey it as already known. A trigger is the word or construction that produces it.

The projection behaviour of the presuppositions is what separates them from entailments (below).

6. Behaviour under negation, questions and embedding

A presupposition survives when the sentence is negated, questioned, put in the antecedent of a conditional or under a modal: the family of sentences.

An entailment (Il a tué le chat entails Le chat est mort) does not survive negation. The test applies to all the triggers of the table of section 5: Il ne regrette pas que tu sois parti, Regrette-t-il que tu sois parti ?, S'il regrette que tu sois parti, … all presuppose that you left.

The presupposition can be filtered or blocked by the context:

• Si + the presupposition: Si Pierre a un frère, son frère est médecin — the presupposition of the possessive (that Pierre has a brother) is satisfied by the si clause.
• Conjunction: Pierre a un frère, et son frère est médecin.
• Disjunction: Soit Pierre n'a pas de frère, soit son frère est médecin.
• Attitude verbs shift the presupposition to the believer: Marie croit que Pierre a arrêté de fumer presupposes that Marie believes he smoked, not that he did.

When a presupposition is not in the common ground, the hearer usually accommodates it: Je suis arrivé en retard parce que ma voiture est en panne introduces the car as given even if the hearer did not know it. Accommodation is the way presupposition is used to carry new information as though it were old; it is also, as the next section shows, a device of persuasion and manipulation.

7. Présupposé and sous-entendu: the French tradition

French linguistics, following Ducrot (1972, Dire et ne pas dire), distinguishes two kinds of implicit content:

The French literature stresses the argumentative and interactional use of the presupposition, the point that Garassino and colleagues (2024, garassino-2024-politicians-vs-chatgpt-study-presuppositions) state with reference to Ducrot (1972; 1984): a presupposition is a linguistic act whose function is to restrain the speaker's possibilities of orienting the subsequent discourse; it sets the frame of the exchange and forces the hearer to accept it or to challenge it explicitly. To challenge a presupposition one must break the flow (Attendez, il n'a jamais fumé !), which is costly, while to challenge an assertion one only has to deny it. The presupposition thus passes information in a form that is hard to contest, which explains its use in advertising and politics.

8. Metalinguistic negation and cancellation

The standard negation, ne … pas, denies the proposition. Metalinguistic negation denies something else about the utterance: an implicature, a presupposition, a register, the pronunciation. The French examples:

The pattern is a negation followed by a correction (mais, ou plutôt, a comma, a pause) with a typical intonation. Blochowiak and Grisot (2018, blochowiak-2018-pragmatics-descriptive-metalinguistic-negation-experimental) report experiments with French speakers on the pragmatics of descriptive and metalinguistic negation: the long-debated question is whether the second is a separate (ambiguous) use of the negation or the same truth-functional negation applied after pragmatic enrichment. The debate bears on how a presupposition-aware system should treat ne … pas: the negation of a sentence can cancel its presupposition when it is metalinguistic, and a parser has only the correction (mais il n'y a pas de roi de France) to go by.

9. Clefts and focus

A cleft (C'est … qui/que) and a pseudo-cleft (Ce qui/que … c'est …) present part of the sentence as new (the focus) and the rest as given. C'est Pierre qui a cassé le vase presupposes that someone broke the vase and says that it was Pierre; it carries exhaustivity, the implication that only Pierre did it. The exhaustivity is cancellable (C'est Pierre qui l'a cassé, avec Paul), so it is an implicature or an inference of the structure, not an entailment; the debate about its status (assertion, presupposition, implicature) is wide. Subject focus is the subject of a study of French and Spanish (Cassara 2026, cassara-2026-subject-focus-french-spanish, CC BY 4.0, a book whose text we hold but did not read for this chapter).

Seul, seulement, ne … que, uniquement mark exhaustivity directly. Seul Pierre est venu. asserts that Pierre came and that no one else did (the second is the exclusive content), and presupposes (on one analysis) that Pierre came.

The structure of the French sentence leaves the focus to word order and intonation more than English does (chapter 29): the dislocation Pierre, il est venu presents Pierre as a topic, the cleft presents it as a focus, the dislocation and the cleft can be combined (Pierre, c'est lui qui l'a fait). The dislocated relations and the expl:subj of the clefts are the cues in the tree.

10. Triggers in the treebanks

String counts in the sentence texts of the UD 2.18 treebanks (2026-11-06; GSD, ParisStories, Rhapsodie, Sequoia; token totals 400 385, 42 789, 44 234, 70 545). The strings count words and phrases, not readings: toujours is a continuation word and also "always", même is "even" and "same", devenir contains venir-forms that are not iterative; the counts show what a word-list detector would see.

Three observations. First, the written corpus is rich in the scalar quantifiers (certains, plusieurs, quelques, la plupart: 661 in 400 k words of GSD, 1.65 per thousand tokens) and the spoken ones are not (23 in the 43 k of ParisStories and 29 in the 44 k of Rhapsodie, about 0.5 to 0.7 per thousand), where the scalar items are beaucoup, peu, tout and jamais. Second, clefts are far more frequent in speech: 149 c'est … qui/que strings in GSD (3.7 per 10 000 tokens), 54 in ParisStories (12.6) and 85 in Rhapsodie (19.2), and so are far more frequent in speech; the string includes many non-clefts (c'est une question que …), so the counts are an upper bound. Third, the conventional implicature of concession is divided by register: pourtant and d'ailleurs are written words (30 and 58 in GSD, 4 and 6 in ParisStories) while quand même is spoken (6 in GSD against 45 and 44 in the spoken ones).

An exact count of presupposition triggers needs the readings, and the French treebanks have no annotation for them. The nearest annotation is the dependency itself: a change-of-state verb is a verb with a given lemma, a definite description is a det of the type Definite=Def on a noun, a factive complement is a ccomp of a factive lemma, a cleft is an expl:subj ce with a acl:relcl or a relative clause introduced by qui/que (chapter 25). In the trees, 931 expl:subj occur in GSD, 314 in ParisStories and 425 in Rhapsodie, of which il is 611, 249 and 286 and ce 220, 34 and 83 (2026-11-06): the expletive ce of c'est constructions is 36% as frequent as expletive il in GSD, 14% in ParisStories and 29% in Rhapsodie.

11. What the research has found

Presupposition in political communication. Garassino, Masia, Brocca and Delorme Benites (2024, garassino-2024-politicians-vs-chatgpt-study-presuppositions, CC BY 4.0) compared the speeches of French and Italian politicians with texts that ChatGPT 3.5 generated for the same issues (immigration, the European Union), on a corpus (IMPAQTS) annotated for presupposition triggers and their discourse functions. Their results: on average the generated texts contain more questionable presuppositions (presuppositions that present as given a content that is not shared and not obviously true) than the politicians' texts; the two most frequent trigger types, change-of-state verbs and definite descriptions, together make up about 70% of all triggers in the politicians' texts and 80% in the generated ones; definite descriptions are the commonest trigger in the politicians' speeches and change-of-state verbs in the generated ones (a significant difference, χ² = 16.13, df = 1, p < .0001, Phi = .25); the other types (factives, relative clauses, adjectives, adverbs, clefts) are rare. The authors relate the tendency to a feature of the generated text: verbose and repetitive slogans made of change-of-state verbs (for example bâtir, which appears in the texts of all four politicians, and protéger) that presuppose a state of affairs the text does not argue for. The finding for a description of French is that the two triggers on which the whole of the political presupposition rests are the most easily detectable by word lists and dependency relations (a definite description, a change-of-state verb).

Metalinguistic negation (Blochowiak & Grisot 2018): see section 8.

Indirect answers (Müller & Plank 2024): see section 3.1.

Winograd schemas (Amsili & Seminck 2017; Wang, Seminck & Amsili 2022; amsili-2017-google-proof-collection-french-winograd, wang-2022-investigating-associative-switchable-negatable-winograd): pronoun-resolution problems that require world knowledge, of which section 4 of chapter 35 says more. They are relevant here because resolving the pronoun is an inference of the type of an implicature (a context-dependent conclusion by reasoning about what the speaker could mean).

Gaps. We hold no open French corpus annotated for scalar implicature, for presupposition outside the political speeches, or for conventional implicature. The French experimental literature on scalar implicature (for example on certains, ou, quelques) exists in journals that we do not hold in open form.

How Mova handles it

In English. Mova's inference module (prag/src/infer.rs; run with prag infer <en model> "<text>") computes, deterministically and with an explanation for each result:

• scalar implicatures from a table of scales (SCALES; after the Quantity axiom of Tatu and Moldovan 2012 and the Horn scales): a weaker item implicates that the stronger ones do not hold, except under negation, when a stronger item is also in the sentence, or when the speaker cancels it (if not all, in fact all, or even);
• presuppositions of triggers (presuppositions; after Cianflone et al. 2018 for the adverbs): again, still, also, too, stop, quit, continue, keep, know, realize, regret, forget that; a presupposition is licensed when the earlier text, up to 50 tokens back, has the same predicate, and accommodated otherwise, with the explanation saying which;
• Quality and Manner (maxims), with the question and the reordering axioms.

A separate module (prag/src/rsa.rs) runs the Rational Speech Acts recursion exactly, with integer weights, on small closed games supplied by hand.

In French, none of this runs. What is true on 2026-11-06:

• Every list is English: the scales in SCALES, the trigger lemmas in presuppositions, the cancellation phrases. A French sentence goes through the English tree parser after en xpos-fill and its lemmas (cesser, arrêter, encore, certains) match nothing.
• The trees carry what French needs where the treebanks have it: Definite=Def on determiners, the lemma of the verb, the expl:subj of clefts, ccomp of factives, advcl of avant que and après que, Mood=Sub on the embedded verb. After xpos-fill the mood is lost in the tag but not in the FEATS (chapter 13); a French reader of the subjunctive trigger must take it from the FEATS.
• The English module has no projection logic (holes, plugs, filters): a trigger fires wherever the parser shows it with the right relation. That limit carries over, and French adds the si filter of section 6 and the metalinguistic negation of section 8, which are the cases where the limit matters.
• The English module prints its inferences and does not store them in frames or in the world; the same would hold for a French version.

What is being built. A French pragmatics layer is being built for the Eel (round kind prag-fr in tools/eel/src/schedule.rs and tools/eel/src/native.rs, on the branch under development on 2026-11-06; chapter 30, §9.4). Its executable is still being written and has no inference component; nothing of this chapter is labelled or measured for French. The planned language-independent meta-pragmatics is where the kinds belong — scalar implicature over an ordered scale, the presupposition of iteration, of continuation, of factivity, of existence — while the French layer supplies the items (certains < tous, encore, ne … plus, regretter que, the definite article) and the French conditions (the si filter, the subjunctive after factive emotive verbs). No such schema exists in code yet.

What is needed, in order. (1) A French scale table (section 2) and a French trigger table (section 5) as data in the form SCALES and the trigger list already have, with the French cancellation phrases (voire, même, en fait, si ce n'est tous, ou plutôt). (2) The French scalar-item lemmas matched by lemma and part of speech, so that tous as a determiner and as a pronoun, or même as "even" and as "same", are separated; the counts of section 10 show how ambiguous the strings are. (3) The conventional implicatures of section 4 as a closed list of connectives with their relation, which is the content of the LEX-CONN-type resource of chapter 36. (4) A gold set: a small check suite in the style used for English can be written for French by hand; there is no open French gold for implicature or presupposition detection to measure against, and the political corpus of Garassino et al. is not distributed with the paper.

Sources

• Grice, H. P. 1975. Logic and conversation; Horn, L. R. 1972, 1989 (scales); Ducrot, O. 1972. Dire et ne pas dire; Frank, M. C., Goodman, N. D. 2012. Predicting pragmatic reasoning in language games. Cited as background and as cited in the papers below; not held.
• Garassino, D., Masia, V., Brocca, N., Delorme Benites, A. 2024. Politicians vs ChatGPT. A study of presuppositions in French and Italian political communication. AI-Linguistica 1(1). garassino-2024-politicians-vs-chatgpt-study-presuppositions (CC BY 4.0).
• Blochowiak, J., Grisot, C. 2018. The pragmatics of descriptive and metalinguistic negation: experimental data from French. Glossa 3(1). blochowiak-2018-pragmatics-descriptive-metalinguistic-negation-experimental (CC BY 4.0).
• Müller, C., Plank, B. 2024. IndirectQA. muller-2024-indirectqa-understanding-indirect-answers-implicit (CC BY-NC 4.0: informs only, not quoted).
• Cassara, M. 2026. Subject focus in French and Spanish (Language Science Press). cassara-2026-subject-focus-french-spanish (CC BY 4.0).
• Amsili, P., Seminck, O. 2017. A Google-proof collection of French Winograd schemas. amsili-2017-google-proof-collection-french-winograd; Wang, X., Seminck, O., Amsili, P. 2022. wang-2022-investigating-associative-switchable-negatable-winograd (both CC BY 4.0).
• UD French GSD, ParisStories, Rhapsodie (CC BY-SA 4.0), Sequoia (LGPL-LR); string counts of 2026-11-06 over the # text lines, all splits together; relation counts over the CoNLL-U files.
• Mova code: prag/src/infer.rs, prag/src/rsa.rs; the English chapter 33.

This page in Markdown → https://md.mova.elamur.ai/free/en/docs/languages/fr/33-implicature-and-presupposition/