🗣️ 34. Irony and figurative language

A speaker often means something other than the words say, in a way the hearer is meant to see: the reverse (irony), less (litotes, understatement), more (hyperbole), something softer or more decent (euphemism), something else that resembles it (metaphor), something close to it (metonymy), or something fixed by convention that has nothing to do with the parts (an idiom). This chapter describes these in French: the grammar and the vocabulary of irony, the figures of the French rhetorical tradition, euphemism, metaphor and metonymy, idioms and multiword expressions, and wordplay on fixed phrases (défigement). It gives what the open research has found, the counts in the French treebanks (UD 2.18, 2026-11-06) and what Mova does about it, which for French is nothing yet.

Contents

1. Literal and non-literal meaning
2. Irony
3. Understatement, litotes and hyperbole
4. Euphemism and dysphemism
5. Metaphor, simile and metonymy
6. Idioms and multiword expressions
7. Wordplay and défigement
8. Regional idioms and the dialect gap
9. The figures of the French rhetorical tradition
10. Counts in the treebanks
11. How Mova handles it

1. Literal and non-literal meaning

Non-literal language raises a question for a system that reads trees: the tree of a sentence is the same whether it is meant literally or not. Il a coûté les yeux de la tête and Il a coûté trois mille euros have the same structure; Quel beau temps ! has it under the sun and in a downpour. The choice between readings depends on the context, the speaker's attitude and, in speech, the intonation. The classical account is Gricean: the hearer sees that the literal reading violates a maxim (Quality, in irony and metaphor; Quantity, in litotes; Relation, in the odd answer) and looks for another one (chapter 33).

The older grammars treated the matter under the heading of the figures and of the evolution of meaning. Brunot (1922, brunot-1922-la-pensee-et-la) notes the habit of metaphor in the daily language: a dent has a racine, a family tree has branches and a dernier rejeton, a man is bon sous sa rude écorce; the vocabulary of everything abstract is borrowed, step by step, from the vocabulary of the concrete. Dauzat (1908, dauzat-1908-la-langue-fran-aise) shows how slang and the popular language, far from being a separate speech, are a laboratory of figures. For him French begins in Latin slang: tête comes from testa, properly a potsherd, which the slang used for the head where classical Latin had caput; joue was drawn "by metaphor" from a word for a bowl; peau first meant "fur" and boyau "black pudding"; and later fiacre was a slang name for the first hired carriages because the hiring office had been set up in 1640 at the hôtel Saint-Fiacre, a metonymy that the Academy admitted to its dictionary in 1718. This is the French case for a point that modern work on metaphor makes: figures do not leave the language; they become the language.

2. Irony

Irony says one thing and means another, with the hearer meant to see it. The French term is ironie, and the figure that says the opposite of what it means is the antiphrase. Sarcasm is irony with a target and a wounding intent. French uses both words; ironie is the wider one, and sarcasme is mostly the biting form. The theories:

2.1 What marks irony in French text

Irony has no grammatical form. It is signalled by context, by intonation in speech, and by a set of cues in text:

Merci bien ! and C'est ça, oui are good examples of conventionalised irony: said with a particular tone, they mean the opposite of what the words say. Ben voyons and Mais bien sûr convey "I do not believe it".

2.2 Intonation

The question whether ironic tone has its own acoustic pattern is debated, and the acoustic correlates are language-specific. Zhou, Al-Tamimi and Yoo (2024, zhou-2024-une-comparaison-de-l-intonation) elicited ironic and literal versions of the same utterance from 10 native speakers of French and 10 of Mandarin (12 scenarios of everyday conversation with acquaintances and friends) and compared eight acoustic correlates with random forests, then with mixed-effects models. They found that for French speakers the jitter was the most influential correlate, followed by the mean range of the fundamental frequency, the shimmer, the mean intensity and the harmonics-to-noise ratio; in the mixed-effects models the ironic utterances showed a significantly lower jitter and a significantly lower mean intensity (p = .050), a lower shimmer that is a trend (p = .060) and a lower F0 range that is not significant (a decrease of 0.236, p = .127). For Mandarin speakers the speech rate, the mean F0 and the mean intensity were the main correlates, and all three decreased in the ironic utterances (by 0.553, 0.268 and 0.423, all statistically significant). The conclusion is that the ironic tone of voice is language-specific, with different patterns of acoustic correlates, and the authors also found an effect of the speaker's gender on ironic utterances in French. The French result is thus a lowering of voice quality and intensity, not a pitch contour. The practical lesson for any text-only system is that what makes an utterance ironic in speech, the voice quality, is not in the transcript at all.

2.3 Irony in the French treebanks

No UD French treebank annotates irony. Some forms of the table above occur, and the counts show how little a word list alone can do (2026-11-06, string counts over the sentence texts; GSD / ParisStories / Rhapsodie / Sequoia):

Looking at the sentences gives the lesson. Of the ten occurrences of bravo in GSD, six are interjections and all six are sincere: Bravo pour cette magnifique maison d'édition ! (fr-ud-train_01614), Bravo Etienne pour ce beau travail ! (fr-ud-train_02559), Bravo, continuez, c'est ultra bon ! (fr-ud-train_09917), fr-ud-train_05719, fr-ud-train_09015 and fr-ud-train_12123 (reviews and forum posts); the other four are a surname (Christine Bravo) or a name (Kid Bravo). In Sequoia the seven occurrences are all the code name of an arms contract (le contrat Bravo). A system that took bravo as an irony marker would be wrong in every case in these two corpora, because the genre is reviews and encyclopaedic text. In the spoken treebanks, c'est ça is always an agreement or a confirmation (c'est ça, oui, ParisStories_2020_concoursInstagram_7; Rhap_D2010-142 c'est ça.; voilà c'est ça qui est incroyable effectivement, Rhap_D2011-8). In vous parlez of Rhap_D2002-41 (au moment où vous parlez d'étrange, de fantastique, …) the verb is the literal parler. The expected ironic forms are rare, since the spoken corpora consist of stories and interviews cooperatively told, and the written corpus of reviews, news and encyclopaedic articles: texts where irony is rare or covert.

The open research on French irony detection is on tweets, and its corpora are non-commercial (Karoui et al. 2015, karoui-2015-detection-automatique-de-l-ironie, CC BY-NC-SA 3.0; Bosco et al. 2016, bosco-2016-tweeting-being-ironic-debate-about, the French annotated corpus Twitter-MariagePourTous on the debate about the same-sex marriage law, CC BY-NC 4.0). They inform this chapter by title and abstract only, not by data. They show that the French irony data that exists is of one genre and one topic, and non-commercial.

3. Understatement, litotes and hyperbole

Litotes says less to mean more, typically by negating the contrary: Ce n'est pas mauvais (it is good), Il n'est pas bête (he is clever), Je ne te hais point (Corneille: I love you), Ce n'est pas rien (it is something), Ce n'est pas faute d'avoir essayé, Je ne dis pas non (I accept), Il n'est pas sans talent, Pas mal ! The Gricean account: the speaker chooses a marked, weaker form (Manner, Quantity) and the hearer enriches it to the strong reading. Understatement in general (meiosis) is the same effect without the negation: Il fait un peu frais in a gale, C'est un petit problème about a disaster.

Pas mal is a good example because it has three uses:

The counts are pas mal 3 in GSD, 15 in ParisStories, 3 in Rhapsodie, 0 in Sequoia (3.5 per 10 000 tokens in ParisStories against 0.7 in GSD): it is a spoken and informal word, and pas mal de is two thirds of the uses in speech (9 of 15). Litotes of the form pas très, pas trop, pas vraiment soften a negative judgment: La carte des menus n'est pas très fourni mais c'est souvent gage de qualité. (fr-ud-train_02564), Service beaucoup trop long et accueil pas très agréable. (fr-ud-train_03776). The tree gives these as advmod of the adjective with the negation advmod; the pragmatic reading is that pas très agréable means "unpleasant", a euphemistic downgrade, in a review.

Hyperbole says more: Je meurs de faim. Il y avait un monde fou. Je te l'ai dit mille fois. Un tas de gens. Des tonnes de documents. Ça m'a coûté les yeux de la tête. Je suis mort de rire. Un froid de canard. Il a un million de choses à faire. Colloquial French uses intensifiers freely (super, hyper, ultra, trop, carrément, grave), which lose their strength in use and need renewal (chapter 37). In the counts, des tonnes, des milliers, un tas de, mille fois, million occur 31 times in GSD and 1 in ParisStories, 0 in Rhapsodie and 3 in Sequoia, while super occurs 47, 38, 3 and 0 times: the exaggerated quantities are written, and the weakened intensifier is spoken.

Prétérition (paralipsis) says something by saying one will not say it: Je ne parlerai pas de son retard. Sans parler de … Passons sur le fait que … It is a figure of the French rhetorical tradition with a place in political speech.

4. Euphemism and dysphemism

An euphemism replaces a word felt as too harsh, rude or taboo by a gentler one. French is rich in them, and Brunot (1922) devotes a section to "the names that one avoids", noting that many things have been given disguised names "out of decency and politeness": during the First World War the administrative language replaced maîtresse by compagne in some contexts, and the name of the enema went through clystère, remède and lavage intestinal in two centuries; "French politeness", he writes, has at times carried these delicacies very far. The fields of euphemism are the same as in any language:

Dysphemism is the opposite: the harsher word chosen on purpose, as an insult or a mark of closeness (bouffer for manger, bagnole for voiture, clamser for mourir, crever). The popular register uses it freely (chapter 37). The euphemism treadmill is well known: the new word becomes tainted in turn (mourir → décéder → nous quitter) and a newer one replaces it. That is why no word list of euphemisms stays valid for long, and why the word is a cue to the register of the text: administrative French is euphemistic, the press uses a middle level, and speech varies by group.

5. Metaphor, simile and metonymy

5.1 Metaphor

A metaphor transfers the sense of a source term to a target term: Le temps, c'est de l'argent; Il est une vraie girouette; Cette réforme est un cheval de Troie; une pluie de critiques; noyé sous les dossiers; Je suis dans le brouillard. In the conceptual theory (Lakoff and Johnson), the metaphors are systematic: ARGUMENT IS WAR (défendre une position, attaquer un argument), TIME IS MONEY (perdre, gagner, économiser du temps), UP IS GOOD (au sommet de sa gloire, au plus bas). Pottiez, Hamon and Grabar (2025, pottiez-2025-detection-de-metaphores-dans-les), who study the metaphor in medical documents, give the standard definition (a figure that transfers the meaning of a source term to a target term, like TIME IS MONEY, and so identifies hidden similarities between two ideas) and note that the metaphor may embellish language, structure thought or explain complex notions, the last two especially in medical writing, where a disease is a battle (lutter contre un cancer) or a path (le parcours de soins).

Most metaphors in everyday French are conventional and do not feel like metaphors: le pied de la table, la tête du lit, un bras de fauteuil, le cours d'eau, grandir en maturité, saisir une idée. They are often registered in dictionaries as ordinary senses. The new or creative ones are what a reader notices.

Detection. The open research gives a sobering picture. Beligné et al. (2017, beligne-2017-detection-automatique-de-metaphores-dans) applied an automatic metaphor detection method based on latent topic models (LDA) to French geography articles and presented the first results with a discussion of the choices; Mpouli (2019, mpouli-2019-chronique-d-un-echec-identification) tried the same kind of LDA-based method (after Heintz et al. 2013) on texts of geography in French and English and concluded that the method was inadequate for the final aim, while it brought out the difficulties of such a project (the title is "Chronicle of a failure"). Pottiez et al. (2025) built a typology of metaphors and a corpus of clinical cases annotated with metaphorical uses, then tested a large generative model with several prompts: the best F-measure was 67.50, with a recall up to 74% for certain prompts, and 45.51% of the metaphors were typed correctly. These are field-specific results on specialised text, and none is a general detector for French.

5.2 Simile

A comparison (comparaison) uses comme, tel, pareil à, semblable à, on dirait, ressembler à, avoir l'air de, comme si: rapide comme l'éclair, têtu comme une mule, fort comme un Turc (the last two are stereotyped similes of long standing). The tree does not separate a literal comparison (Il est grand comme son père) from a figurative one (Il est grand comme une maison) either; the research has built a method to extract comparisons and pseudo-comparisons for the detection of figurative comparisons (Mpouli & Antoine 2015, mpouli-2015-extraction-et-analyse-automatique-des, CC BY-NC-SA 3.0, informs only). In the treebanks, comme si occurs 5 times in GSD, 5 in ParisStories, 2 in Rhapsodie and 1 in Sequoia, and en quelque sorte 2, 1, 5, 0; pour ainsi dire and si l'on peut dire occur 5 times in Rhapsodie only. These are the markers of a metaphor that the speaker is aware of: a hedge on a figure.

5.3 Metonymy and synecdoche

A metonymy names a thing by a related one: the container for the content (boire un verre), the author for the work (lire Zola), the place for the institution (l'Élysée a annoncé, Paris a décidé, Matignon, Bruxelles), the producer for the product (un Picasso), the cause for the effect, the sign for the thing (la couronne). A synecdoche names the whole by the part or the reverse (une voile à l'horizon, les bleus for the French team, une belle plume). Metonymy is the regular means by which French journalists name governments by capitals: the subject of a annoncé is a place name. Brun and colleagues' work on the resolution of metonymy for named entities (2008, 2009: brun-2008-resolution-de-metonymie-des-entites, brun-2009-resolution-de-metonymie-des-entites, CC BY-NC-SA 3.0, informs only) deals with this case, where a place name stands for an institution or a team, and is mentioned only to show that the problem has a tradition in French NLP.

In dependency trees, the metonymic subject is a nsubj as any other; only the semantics (a place cannot announce) shows the figure. This is the same case as in English, where Mova's absurdity matrix flags the odd role but does not separate metonymy from the other kinds of oddness (see below).

6. Idioms and multiword expressions

An idiom (locution figée, expression idiomatique) is an expression whose meaning is not the sum of the meanings of its parts, and whose form is fixed in whole or in part. French has several classes:

The last class is what UD French marks with the relation fixed: the dependents of a multiword function expression (a conjunction, a preposition, an adverb). In GSD there are 3 243 fixed dependents (about 0.8% of the tokens); the commonest fixed expressions, counted by head with its fixed dependents, are ainsi que (128), en effet (102), à partir (de) (100), alors que (80), un peu (75), par exemple (70), en tant que (59), d'ailleurs (58), tandis que (56), de plus (51), d'abord (49). In ParisStories they are en fait (153), parce que (140), un peu (60), quand même (45), parce qu' (40), et puis (39), et tout (39), pas mal (15); in Rhapsodie parce que (78), un peu (65), et puis (58), quand même (44), en fait (38), d'accord (26), bien sûr (18), à peu près (17). The lists are the connectives and the discourse markers of chapters 36 and 38. The idioms with a verb and a noun phrase (poser un lapin) are not marked fixed: UD treats them as ordinary verb–object phrases; the idiom lists tomber dans les pommes, poser un lapin, avoir le cafard, casser les pieds and coûter les yeux have no occurrence in the four treebanks (string counts of 2026-11-06), which is itself a measure of how sparse the idioms are in 560 000 tokens.

The PARSEME-FR annotation. The multilingual PARSEME shared task on verbal multiword expressions produced a French part, described by Candito et al. (2017, candito-2017-annotation-d-expressions-polylexicales-verbales). For French it covers idiomatic verbal expressions, inherently pronominal verbs and a generalisation of light-verb constructions, annotated on the French UD corpus and on Sequoia: 22 645 sentences with 4 962 annotated expressions, about one expression every 100 tokens, with a high rate of discontinuous expressions (40%). The figure of 40% discontinuity is the reason why a word-list detector does not work: Il a pris la décision and La décision qu'il a prise are the same expression. Pasquer et al. (2017, pasquer-2017-expressions-polylexicales-verbales-etude-de) study the variability of these verbal expressions in the corpus; Todirascu et al. (2019, todirascu-2019-polylexfle-une-base-de-donnees) built a database of multiword expressions for French as a foreign language; Zampieri et al. (2022, zampieri-2022-identification-des-expressions-polylexicales-dans) examine multiword expressions in tweets. Bagdasarov, Alves and Teich (2026, bagdasarov-2026-cheese-it-up-camembert-outperforms) compare a fine-tuned CamemBERT model with two open generative language models on the identification of French multiword expressions; the title states the result: the fine-tuned encoder wins. The PARSEME-FR annotation is the nearest thing to a French idiom gold standard; its own licence is to be confirmed from its distribution (the treebanks under it are CC BY-SA and LGPL-LR), and until then it is for counts and tests only.

Idioms and the tree. Because an idiom is non-compositional, the compositional reading (the sum of the parts) and the idiomatic one have the same tree. A system must recognise the idiom from the lexical items in a particular relation (the verb poser, the object lapin, the determiner un) and replace the compositional meaning by the idiomatic one. The syntactic flexibility varies (the passive of casser les pieds is odd, the passive of prendre une décision is normal), and the research on variability (Pasquer et al.) measures it.

7. Wordplay and défigement

French has a rich practice of wordplay: the calembour (pun), the contrepèterie (spoonerism), the jeu de mots on a fixed expression, the verlan, the mot-valise, the paronomasia. The specifically French concept is défigement ("unfreezing"): a fixed expression (figée) is altered lexically, syntactically, semantically or pragmatically so that a wordplay or a new meaning arises while the original is still recognisable. Bezançon et al. (2025, bezancon-2025-forbidden-fruit-is-sweetest-annotated) describe the phenomenon with the definition of Mejri (2013) that the underlying expression should always remain identifiable in the altered one, and call the altered forms puns in multiword expressions (PMWEs). They built the French Unfrozen Idioms in Tweets corpus (FRUIT): they queried the Twitter API with 216 conventional French expressions over three years (2020–2023), obtained 3 369 636 tweets, filtered them to 60 617 tweets that contain both the expressions and candidate PMWEs, and had three experts annotate a sample of 600. The agreement was low without guidelines (Krippendorff's α = 0.19 on 100 tweets), rose to 0.77 on a second sample with new guidelines (an easier sample), and was 0.67 on a third. The authors point out that the annotators disagreed on whether to give weight to the formal change of the expression or to the context. The result is a measure of how hard the notion is to apply: even experts do not agree on what counts as an unfrozen idiom without guidelines.

Bezançon & Lejeune (2023, bezancon-2023-reconnaissance-de-defigements-dans-des) propose a first method for the automatic recognition of défigements in French tweets, by aligning sequences on several layers of linguistic information (form, lemma, part of speech) to isolate possible unfreezings, then ranking the candidates by similarity. Dehouck (2023, dehouck-2023-generation-automatique-de-jeux-de) deals with the automatic generation of puns on first names (known to us by title). These works show that the figure is productive in social media, in headlines (Les Échos style plays on proverbs), in advertising and in political satire.

8. Regional idioms and the dialect gap

Idioms differ by variety. Beauchemin, Tremblay, Youssef and Khoury (2026, beauchemin-2026-idiom-understanding-as-tool-measure) built three benchmark datasets for testing the understanding of regional idioms: QFrCoRE, with 4 633 instances of idiomatic phrases of Quebec French, QFr-CoRT, with 171 regional idiomatic words, and MFrCoE, with 4 938 phrases of metropolitan French. They tested 111 large language models and found that while the models perform well on the metropolitan expressions, 65.77% of them perform significantly worse on the Quebec idioms, and only 9.0% favour the regional dialect. The finding for a grammar of French is the simple one that idioms are where varieties of French diverge most (dépanneur, magasiner, tabarnak for a Quebec speaker, against supérette, faire les courses, nothing, for a speaker in France; see chapter 8), and that coverage of "French" in language models means coverage of the metropolitan standard.

Noël et al. (2026, noel-2026-eiffel-novel-benchmark-measure-bias) built a benchmark of French idiomatic expressions in context, EIFFEL, in order to test the effect of the heavy English proportion in the training data of mainstream multilingual models. They compare multilingual models with models trained on higher proportions of French and train from scratch a series of 1B-parameter small models with different proportions of French and English data. Their conclusion is that EIFFEL brings out the benefit of a higher proportion of French data and shows the limits of standard benchmarks for measuring multilingual competence: a model can do well on standard French benchmarks without mastering French idioms. The benchmark was built with input from native French speakers and is described in a way that makes it reproducible for other languages; its annexes give the questions put to the native speakers on whether a French expression has an English idiom that is its word-for-word translation, or that results from a small change of its modifiers.

9. The figures of the French rhetorical tradition

French school grammar teaches the figures (figures de style) as a fixed list. A short table, with the French term, the definition and an example:

The chiasme example is the classical rhetorical proverb. The apostrophe of the figures is not the typographic apostrophe nor the vocative of the tree (though it often is a vocative: chapter 32). The term antiphrase is the French name for the central device of verbal irony.

10. Counts in the treebanks

Figurative language is sparse in the four treebanks, and the string counts for the most widely cited items are small (2026-11-06; GSD / ParisStories / Rhapsodie / Sequoia):

The véritable of GSD (36) is an intensifier of metaphor in the written corpus: une véritable zone franche de trafic de drogue (fr-ud-train_00574), une véritable brute psychopathe (fr-ud-train_00720), une véritable mascarade (fr-ud-train_02617). The adjective announces that the following noun is used figuratively but with literal force: it is a marker of metaphor, not a term for it. Littéralement is the same sort of marker, and often used for hyperbole: ça ne ressemble littéralement à rien (fr-ud-train_03423), where the word does not mean "literally" in the strict sense. The light-verb constructions of the table (prendre en compte, mettre en place, avoir lieu) are written and rare in speech; the construction il s'agit de is an almost exclusively written one (83 in GSD, 0 in ParisStories).

How Mova handles it

In English. Mova has a deterministic irony candidate detector (prag/src/irony.rs; after Joshi et al. 2015 and Riloff et al. 2013): it uses polar words of both signs from a small lexicon, counts sentiment flips and the longest same-polarity run, adds negative situations, emoticons, laughter and ironic phrases, and fires on a positive word governing or preceding a negative one unless a contrastive connector stands between them (the food was good but the service was bad is a sincere mixed judgment). The result is printed by prag infer. Idioms and register go through prag/src/lens.rs (prag lens; the expression base is closed and not in the free release), and the domains (humour, hidden meaning, slang, formal) and the absurdity trigger are in world/src/domains.rs. The agreer runs an irony model as a witness and only reports it (prag/src/agree.rs).

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

• There is no French list of polar words, negative situations, ironic phrases, idioms, proverbs or euphemisms in Mova. The English irony detector's lists (POSITIVE, NEGATIVE, SITUATIONS, PHRASES, CONTRAST) would need French counterparts, and the CONTRAST list (the connectors that make a mixed judgment sincere) would carry mais, cependant, pourtant, bien que, quoique, tout de même in French.
• The treebank counts of section 2.3 show a specific hazard for a French list: the common positive words (bravo, génial, formidable, magnifique) are, in the genres of the open French corpora, sincere. The English detector's rule needs a negative situation or a contrast to fire; a French version would too, and cannot use the positive word alone.
• The multiword expressions of PARSEME-FR could supply the idiom lexicon: 4 962 annotated verbal expressions on UD French and Sequoia (Candito et al. 2017), with labels for idiomatic expressions, inherently pronominal verbs and light-verb constructions; its licence is to be confirmed before it is used for more than tests. Mova does not load it.
• The fixed relation of UD French is already in the trees Mova reads, and the English parser, trained on the French treebank through en xpos-fill, predicts it; the multiword function expressions (ainsi que, en effet, alors que) are therefore handled at the tree level. The idioms of a verb and its object are not.
• Metaphor and metonymy are not detected in English or in French; the absurdity matrix flags odd roles, but does not separate metaphor from humour, and the nearest French result for detection, in a specialised domain, is 67.50 F for a large generative model (Pottiez et al. 2025). There is no prosody: ironic tone is out of reach in text.

What is being built. The Eel's French pragmatics round (prag-fr, in tools/eel/src/schedule.rs and tools/eel/src/native.rs on the branch under development, 2026-11-06; chapter 30, §9.4) is the place a French irony candidate and a French idiom lens would be measured. Its executable is still being written; no irony, idiom or figurative label exists for French, and none is measured. In the planned language-independent meta-pragmatics, irony and understatement are shared kinds (an evaluation whose polarity is reversed by the context; a weaker statement standing for a stronger one), while the French cues — Merci bien !, Ben voyons, pas mal, soi-disant, the French polar lemmas — belong to the French layer. Not implemented.

What is needed, in order of cost. (1) A French idiom and euphemism list built from the Wiktionary raw dump (CC BY-SA, for learning and tests only) and from PARSEME-FR, consulted by the lens. (2) A French version of the incongruity detector: a closed list of polar lemmas and negative situations, the French ironic phrases of section 2.1, and the contrast connectors; the explanation of each result printed as in English. (3) The marker words of section 10 (véritable, littéralement, pour ainsi dire, en quelque sorte, comme si) as cues of a figurative use. (4) A French sarcasm gold set: the open data we know is non-commercial and of one genre (tweets on one debate), so a small set of sentences labelled by hand, from the genres Mova reads, would be the first real measurement.

Sources

• Brunot, F. 1922. La pensée et la langue. brunot-1922-la-pensee-et-la (public domain).
• Dauzat, A. 1908. La langue française d'aujourd'hui. dauzat-1908-la-langue-fran-aise (public domain: author died 1955).
• Zhou, Z., Al-Tamimi, J., Yoo, H. 2024. Une comparaison de l'intonation ironique en français et en mandarin. JEP-TALN-RECITAL 2024. zhou-2024-une-comparaison-de-l-intonation (CC BY 4.0).
• Pottiez, C., Hamon, T., Grabar, N. 2025. Détection de métaphores dans les documents médicaux. pottiez-2025-detection-de-metaphores-dans-les (CC BY 4.0). Beligné, M. et al. 2017. beligne-2017-detection-automatique-de-metaphores-dans (CC BY 4.0). Mpouli, S. 2019. Chronique d'un échec. mpouli-2019-chronique-d-un-echec-identification (CC BY 4.0).
• Candito, M. et al. 2017. Annotation d'expressions polylexicales verbales en français. candito-2017-annotation-d-expressions-polylexicales-verbales (CC BY 4.0). Pasquer, C. et al. 2017. pasquer-2017-expressions-polylexicales-verbales-etude-de; Todirascu, A. et al. 2019. todirascu-2019-polylexfle-une-base-de-donnees; Zampieri, N. et al. 2022. zampieri-2022-identification-des-expressions-polylexicales-dans; Bagdasarov, S., Alves, D., Teich, E. 2026. bagdasarov-2026-cheese-it-up-camembert-outperforms (all CC BY 4.0).
• Bezançon, J. et al. 2025. Forbidden FRUIT is the sweetest. bezancon-2025-forbidden-fruit-is-sweetest-annotated; Bezançon, J., Lejeune, G. 2023. bezancon-2023-reconnaissance-de-defigements-dans-des; Dehouck, M. 2023. dehouck-2023-generation-automatique-de-jeux-de (title only) (all CC BY 4.0).
• Beauchemin, D. et al. 2026. Idiom understanding as a tool to measure the dialect gap. beauchemin-2026-idiom-understanding-as-tool-measure; Noël, C. et al. 2026. EIFFEL. noel-2026-eiffel-novel-benchmark-measure-bias (both CC BY 4.0).
• Karoui, J. et al. 2015. karoui-2015-detection-automatique-de-l-ironie (CC BY-NC-SA 3.0); Bosco, C. et al. 2016. bosco-2016-tweeting-being-ironic-debate-about (CC BY-NC 4.0); Mpouli, S., Antoine, J.-Y. 2015. mpouli-2015-extraction-et-analyse-automatique-des (CC BY-NC-SA 3.0); Brun, C. et al. 2008, 2009. brun-2008-resolution-de-metonymie-des-entites, brun-2009-resolution-de-metonymie-des-entites (CC BY-NC-SA 3.0): non-commercial, cited by title and abstract only, nothing quoted.
• Grice, H. P. 1975; Sperber, D., Wilson, D. 1981; Clark, H. H., Gerrig, R. 1984; Lakoff, G., Johnson, M. 1980: background only; not held.
• UD French GSD, ParisStories, Rhapsodie (CC BY-SA 4.0), Sequoia (LGPL-LR); counts of 2026-11-06. Sentences quoted by sent_id.
• Mova code: prag/src/irony.rs, prag/src/lens.rs, prag/src/agree.rs, world/src/domains.rs; the English chapter 34.

This page in Markdown → https://md.mova.elamur.ai/free/en/docs/languages/fr/34-irony-and-figurative-language/