🗣️ 18. Compounding, conversion and the minor processes

Besides affixation (16, 17), French makes words by compounding (chou-fleur, porte-monnaie, sourd-muet, pomme de terre), by conversion (le rire, un rouge, parler fort), and by a set of minor processes: clipping (prof, ciné), sigles and acronyms (SNCF, ovni), verlan and largonji (meuf, loufoque), blends (courriel, infox), verbal nouns (un appel, un saut), reduplication (dodo, chouchou), ellipsis (un vapeur for un bateau à vapeur) and eponyms (poubelle, boycotter). Compounds raise two spelling questions, hyphen or welding (§3) and the plural (§4). §12 says how the UD French treebanks annotate the results and what Mova does with them.

The historical analysis comes from three public-domain works: Brunot's Précis (1899, §§114–121, brunot-1899), Darmesteter's treatise on compounds (2nd ed. 1894, darmesteter-1894) and his study of new words (1877, darmesteter-1877), with Nyrop's volume on word formation (1908, nyrop-1899 t. III). The spelling rules are those of the official 1990 report (cslf-1990), held in full in the library. The counts are ours, made on the UD 2.18 French treebanks on 2026-11-06. Examples without a sent_id are our own.

1. What a compound is

A compound is a combination of words or stems that works as one word with a meaning of its own: un chou-fleur is a vegetable, not a cabbage that is a flower. French stress falls on the last syllable of the phrase, whether it is a word or a group (03), so stress gives no test. The usual tests are these:

None of these is enough alone, and the spelling is the weakest. Pomme de terre, chemin de fer, salle de bains, compte rendu are compounds by meaning and behaviour but are written with spaces, while peut-être and c'est-à-dire are hyphenated function words. The 1990 report describes a scale: a hyphenated compound competes with welding on one side (portemanteau, betterave) and with frozen expressions written with spaces on the other (pomme de terre, compte rendu). The further one goes from the frozen expression through the hyphenated compound to the welded word, the more clearly one has a single word (cslf-1990, analysis I.1).

Brunot's classes. Brunot (§116) separates two kinds of compound by whether a relation is left unsaid. A compound proper implies an ellipsis: arrière-cour stands for "a yard that is at the back", timbre-poste for timbre de la poste, couvre-chef for "what covers the head". A juxtaposition puts words together according to ordinary syntax, with nothing left unsaid: bienheureux, sergent de ville, pomme de terre. It differs from a free group only in calling up a single image: sergent de ville names a uniformed policeman, not a sergeant plus a town. The fusion is gradual. Pomme de terre has long given one image, but blanc de céruse still suggests white made from ceruse. A third, very large class is composition by particles, the prefixed words (surplis, défaire), treated in 16.

Compounds and juxtapositions tend to become simple words (§117). Sometimes the spelling shows it (plafond, piédestal are now one word), and the stress of the first part fades (licou < lie cou). Once welded, the word follows the ordinary sound laws: comes stabuli gave conestable, then connétable.

Order. A compound has a determined and a determining term (§115). Latin put the determining term first (silvicola "forest-dweller"), and that order survives in learned words. The analytic Romance tendency is to put the determined term first. Old French could express the complement without a preposition, as Latin did, which gave Hôtel-Dieu, fête-Dieu, Bourg-la-Reine (§121), and the same head-first order runs through chou-fleur, timbre-poste, wagon-lit. Head-last order comes back with neoclassical compounds (autoroute, hydroélectrique), with bound prefixes (mini-jupe) and with English loans (best-seller, hot-dog, week-end).

2. Types of compounds

2.1 Noun + noun

The head is the first noun. The second names a resemblance, a purpose, a measure or an origin:

The pattern is open (une ville-musée, un bébé-éprouvette, un dîner-débat). The gender is the head's: une année-lumière, un mot-clé. When the referent is a person, agreement may follow the referent rather than the head: porte-parole (a V + N compound, §2.3) is masculine as a word, but GSD has 13 masculine and 3 feminine tokens of it.

2.2 Noun + preposition + noun

The most frequent way to name things with two nouns is a phrase with de or à, the French counterpart of the English noun + noun compound (bathroom, windmill; compare en 18):

The preposition is chosen item by item, not by rule: un sac à main but un sac de voyage; une tasse à café is a cup for coffee, une tasse de café a cupful. The second noun has no article, and its number expresses the idea, not a quantity: une salle de bains, une confiture de fraises, du sirop d'érable.

2.3 Verb + noun: the porte-monnaie type

This is the most characteristic French compound. A verb form is followed by its object, and the whole names an instrument, a person or a place by what it does:

The verb form. The verb could be a bare stem, an imperative or a third-person present. Darmesteter argued at length for the second-person singular imperative, which governs the noun as in a command (boute en train, "put [them] in a good mood"). His evidence was medieval charters, where such names are rendered by a Latin imperative, Old French surnames such as Pais mouche and Boi l'awe, and doubled forms such as tournevire. Gaston Paris's preface to the 1894 edition endorses the analysis and notes that some compounds, mostly proper names, have an indicative instead (darmesteter-1894, preface and the chapter on verbal compounds). Synchronically the verb is frozen: nobody inflects it (*gratterons-ciel). The 1990 report states the consequence: in le ramasse-miettes the first element "is no longer a verb", the whole names one object and not an act, so it should take no plural mark in the singular (cslf-1990, analysis I.2).

Gender. The compound is masculine whatever the noun, because it names an object or an agent: un porte-plume, un garde-fou, un perce-neige (also feminine in usage). La garde-robe has the noun garde as its first element. Person nouns take the referent's sex: un/une porte-parole, un/une rabat-joie.

The type is productive for instruments: un porte-clés, un ouvre-bouteille, un chauffe-eau, un presse-purée, un essuie-glace, un lance-pierre. Some old members have parts that nobody analyses any more (rez-de-chaussée, croc-en-jambe, à vau-l'eau), which the report gives as one reason for welding.

2.4 Particle + noun

A preposition or adverb with a noun gives nouns of position, time and status:

The line with prefixation depends on whether the first element is also a free word. Après, sous, contre, sans, avant are prepositions; vice-, ex-, mini- are bound and belong to 16.

2.5 Adjective and noun

The last type puzzled grammarians until Darmesteter explained it as synecdoche: the bird is named after its red throat (darmesteter-1894, Gaston Paris's preface). Brunot gives two further facts. Pied-bot does not stand for "a man who has a club foot", since nothing is left unsaid: the part becomes the name of the whole, and such names became surnames (Jacques Bon Dos). And the gender may change: rouge-gorge was once feminine like gorge and is now masculine (§120).

2.6 Compound adjectives

The nationality series is fully productive in journalism. The colour rules are in 15.

2.7 Verb + verb, phrases and function words

• V + V and V + et + V: un laissez-passer, un savoir-faire, un savoir-vivre, un va-et-vient, un pousse-pousse. They are invariable.
• Clauses as nouns: un on-dit, un m'as-tu-vu, le qu'en-dira-t-on, un je-ne-sais-quoi, un sauve-qui-peut. They are invariable, and the hyphens mark the clause as one word. The report cites laissez-passer, sauve-qui-peut, qu'en-dira-t-on as the case where a sequence of words changes its grammatical class (analysis I.1).
• Compound verbs are rare. Maintenir (Latin manu tenere) is old; sous-estimer, contre-indiquer, s'entredévorer are prefixations. New verbs come from the -er conjugation, not from compounding (§5.3).
• Function words written as phrases or hyphenated: peut-être, c'est-à-dire, au-dessus, vis-à-vis, tout à coup, à cause de, parce que. They belong to the fixed expressions of 19.
• Numerals (dix-sept, vingt et un) are in 12.

3. Hyphen or welding

Dictionaries disagree on the hyphen, and usage hesitates (cslf-1990, analysis I.1). The 1990 report answers with a policy of extending welding (soudure or agglutination), "in keeping with the tradition of the Académie française". Its principles are these.

1. Weld what is felt as one unit. Welded words of old formation are no longer analysed: vinaigre, pissenlit, chienlit, portefeuille, passeport, marchepied, hautbois, plafond. The hyphen remains where an element has aged and is no longer understood (rez-de-chaussée, croc-en-jambe). It also remains where the sequence changes class (à propos → l'à-propos) or where the meaning differs from the sum of the words (rouge-gorge).

2. Do not weld where reading would suffer. The hyphen stays when two vowels meet that could be read as one unit (o-i, a-i, o-u, a-u: extra-utérin, génito-urinaire). Science sometimes uses the diaeresis instead (radioïsotope). The Académie also keeps contre-attaque, or contrattaque with elision, and allows the hyphen in learned compounds of more than two parts.

3. Learned elements weld. A learned element that is not a free word (narco-, poly-) is usually welded (narcothérapie), less often hyphenated (narco-dollar). When every element is learned, welding is obligatory (narcolepsie).

4. Particles weld. Contre- and entre- used as prefixes weld, "on the model of en, sur, supra": contrechant, à contrecourant, contrefeu, contrespionnage, entraide, entreligne, s'entredévorer. Welding removes the apostrophe (contrappel like contrordre). So do the Latin prefixes extra, intra, ultra, infra (extraconjugal, ultrafiltration, infrasonore) (recommendation 2).

5. Lists of welded compounds.

   • List A has a verbal first element: arrachepied (d'), boutentrain, brisetout, chaussetrappe, croquemitaine, croquemonsieur, faitout, fourretout, mangetout, passepartout, piquenique, porteclé, portemonnaie, portevoix, poussepousse, tirebouchon, tirebouchonner, tournedos, vanupied and others.
   • List B has nominal and adjectival elements: arcboutant, autostop, bassecour, branlebas, chauvesouris, hautparleur, lieudit, millefeuille, millepatte, platebande, potpourri, prudhomme, sagefemme, terreplein and others.
   • List C has onomatopoeia and expressive words: blabla, bouiboui, coincoin, froufrou, grigri, kifkif, mélimélo, pêlemêle, pingpong, prêchiprêcha, tamtam, tohubohu, traintrain, troutrou, tsétsé.
   • List F has loans "well established and not used as quotations": apriori, exvoto, statuquo, vadémécum; baseball, bluejean, cowboy, covergirl, fairplay, globetrotteur, hotdog, pipeline, weekend and others.

   The report adds that changing several thousand compounds at once was out of the question and that "usage can do it gradually" (list A, preamble).

The rectified spelling is a recommendation: both spellings are correct. The newspapers and web texts of GSD keep the traditional one. The lemma week-end has 15 tokens and weekend 4. Portemonnaie, aprèsmidi and hautparleur do not occur (2026-11-06). A reader of French meets both spellings, and a program that maps one onto the other makes no error.

The hyphen also marks clitics. Inverted subjects (a-t-il, dit-elle), imperatives with clitics (donne-le-moi), the demonstrative particles (celui-ci, ce jour-là) and -même (lui-même) are hyphenated, though none of them is a compound. A hyphen can therefore mark a compound (arc-en-ciel), a clitic attachment (donne-le-moi) or a numeral (dix-sept). A tokenizer has to tell the three apart (§12).

4. The plural of compounds

The traditional rules follow the analysis of each compound:

The V + N rule asked the writer to decide whether the noun means one thing or several. The report lists the resulting muddle: dictionaries distinguish des gardes-meubles (men) from des garde-meubles (places), "an erroneous analysis already denounced by Littré", and un porte-montre (for one watch) from un porte-montres (for several). They also contradict each other and themselves: un cure-dent but un cure-ongles; des après-midi but des après-dîners (analysis I.2).

The 1990 rule (rule II.2) treats V + N and prep + N nouns as ordinary nouns. They take no plural mark in the singular and a final -s in the plural: un pèse-lettre, des pèse-lettres; un cure-dent, des cure-dents; un garde-meuble, des garde-meubles (men or places alike); un après-midi, des après-midis; un sans-abri, des sans-abris. There are two exceptions. A noun with a capital stays as it is (des prie-Dieu), and so does a noun preceded by a singular article (des trompe-l'œil, des trompe-la-mort). Welded compounds follow the simple rule (des portemonnaies, des weekends). Loans take a regular -s (rule II.7; 19).

What the texts do. Hyphenated plural NOUN tokens in GSD (2026-11-06) include demi-finales 9, beaux-arts 5, sages-femmes, belles-lettres, années-lumière, sous-marins 3 each, and timbres-poste, sapeurs-pompiers, mots-clés, libres-penseurs, francs-maçons, avant-postes 2 each. All follow the traditional rules. The corpus has no plural V + N compound at all (gratte-ciel occurs three times, singular), so it says nothing about the 1990 rule. The form mot-valises (twice, lemma mot-valise; fr-ud-train_03645) follows neither system: an N + N compound of the mot-clé type has mots-valises.

5. Conversion

Conversion changes the class of a word without changing its form. Darmesteter calls it dérivation impropre, derivation without a suffix (darmesteter-1877, part I). French has less of it than English. A class carries visible marks (articles, endings, clitics), so a change of class is always marked by the context and often requires a change of form.

5.1 To noun

The processes do not have the same vitality. Darmesteter, surveying the new words of the nineteenth century, found nouns from adjectives "very much alive" and participial adjectives productive. Verbal nouns (la chauffe) and participial nouns (la retombée) were alive in popular speech but receding in the common and literary language before Latin derivation in -age, -ment, -tion. The infinitive as a noun was a dead process: le rire, le devoir are old, and new ones are not made (darmesteter-1877, conclusion).

A converted noun takes articles, adjectives and possessives (son beau sourire) and not degree adverbs (*le très rire). UD tags such a word by its use, so one form gets two tags. In GSD the form jeunes is ADJ 39 times and NOUN 23, pauvres ADJ 9 and NOUN 2, riches ADJ 12 and NOUN 1 (2026-11-06). Headless uses (les jeunes) are tagged both ways, and the choice is hard to automate.

5.2 To adjective and to adverb

• Noun as invariable adjective: une robe cerise, des yeux noisette, des chaussures marron, un air chic; also clipped words (un mec sympa). The colour rules are in 15.
• Adjective as adverb: parler fort, coûter cher, voir clair, chanter faux, travailler dur, sentir bon. The form is invariable, and the -ment adverb often means something else (fortement "strongly").
• Intensifiers of the familiar register: super, hyper, méga, grave, trop (c'est super bon, trop fort) — prefixes, adjectives and adverbs that have become degree words (15).

5.3 To verb: the -er default

French does not use a noun as a verb unchanged. A new verb takes an infinitive ending, almost always -er, the one productive conjugation (13): skier, téléphoner, texter, cliquer (from clic), and from English liker, zapper, booker, scroller. Strictly this is suffixation, not conversion. English to email corresponds to a suffixed French verb, so a French noun and its verb differ in form. Tutoyer, vouvoyer (from tu, vous) take the suffix -oyer.

5.4 Ellipsis

Nyrop describes a process close to conversion: a compound or phrase loses its head, and the modifier takes over the meaning (nyrop-1899 t. III, §571). Une ville capitale gives une capitale, un bateau à vapeur gives un vapeur, une poire de doyenné gives un doyenné. Modern French continues the pattern with un portable (téléphone portable), un mobile, la télécommande. Gender and number then come from the lost head, as in the eponyms of §11 (un bordeaux: un vin de Bordeaux).

6. Clipping (troncation)

Clipping shortens a word. It is very productive in the familiar register, and the result is a new word with its own gender and plural. Nyrop already discusses "the sometimes very violent shortenings that words undergo, especially in the argots" (t. III, §5).

Apocope cuts the end and usually keeps two syllables. The result often ends in a vowel, and -o is frequently added:

A clipped word keeps the gender of the full word (la fac, la pub, le resto). With person nouns the referent can decide (la prof).

Register in the treebanks (form counts, 2026-11-06). In written GSD (400,385 words) the familiar clippings are rare: sympa 19, resto 10, bac 4, télé 3, fac 1, appart 1, prof 1. The neutral ones are ordinary words: photo 13 and photos 14, métro 14, vélo 9. Spoken ParisStories is a tenth of GSD's size (42,789 words), yet it has fac 16, bac 6, prof 5 and maths 5. Per 10,000 words that is fac 3.7 against 0.02 in GSD. An example: bah la fac c'était un peu la même chose. (ParisStories_2019_experienceFac_13). GSD's single fac is in a summary of a television series: celui pour qui elle avait craqué en fac de médecine (fr-ud-train_08243).

Familiar suffixes on the stump: -o (apéro, prolo, dico, hosto, écolo, facho), -oche (cinoche, valoche, fastoche). The suffixes are in 17.

Aphaeresis cuts the beginning and is rare: (Amé)ricain → ricain, and bus from (omni)bus, (auto)bus. Given names are clipped both ways and reduplicated: Dominique → Domi, Nicolas → Nico, Francine → Fanfan.

7. Sigles and acronyms

French separates sigles, read letter by letter (SNCF, RATP, TVA, PDG, SMS, DVD, UE), from acronymes, read as words (ONU /ɔ.ny/, OTAN /ɔ.tɑ̃/, UNESCO, ovni). The line is blurred: SMIC is read /smik/. An acronym that becomes an ordinary noun is written in lower case and takes -s: un ovni, des ovnis (objet volant non identifié), le sida, and the English laser, radar.

Nyrop, writing in 1908, found the formation of words from initials "relatively rare in French". He quotes Paul Bourget calling it a habit "borrowed from Anglo-Saxon customs" and contrasts the frequency of the process in English (t. III, §5). The flood of sigles is a twentieth-century fact.

In the treebanks. GSD has 1,630 all-capital tokens of two to six letters (2026-11-06). 213 of them are Roman numerals tagged NUM (II 78, III 25, IV 23). Among the sigles the most frequent are ONU 34, UE 21, FC 14, UEFA 11, PS 11 (all PROPN), and DVD 17, CD 16, ONG 10, PIB 9, ADN 9 (NOUN). The tag follows what the sigle stands for: a proper noun for organisations, a common noun for things. None of the French treebanks uses the feature Abbr=Yes, so a sigle can only be found by its shape, and its expansion is not given.

Expansion is a problem of its own. Koptient (2020, koptient-2020) works on medical abbreviations for text simplification. Abbreviations are opaque, she notes, because only the first letters of the words are visible, and many are ambiguous. DP can be dialyse péritonéale, dilatation pneumatique or dysménorrhée primaire, and PC is a personal computer in general language. Classifiers trained on the context reached an F-measure of 0.888 in cross-validation on the training data (MLP and decision tree), but only 0.773 on a separate test corpus (decision tree). There, the baseline that always chooses the majority sense did better than every classifier. A sigle needs a list of its senses and a strong prior on the commonest one.

8. Verlan, largonji and javanais

French has a tradition of disguised word forms, used by groups of speakers (butchers, the criminal underworld, young people of the suburbs) to mark identity or to keep outsiders out. Darmesteter excluded true argot from his study of new words because it is "a language of convention, an artificial language" created by a class with good reasons to escape indiscreet ears. He kept the popular language, which follows the natural processes (darmesteter-1877, introduction). The output is a register (19 §6).

8.1 Verlan

Verlan is l'envers "the reverse", itself reversed. The syllables of the word are inverted, on the spoken form:

The eu of meuf, keuf, teuf, keum is the added schwa, pronounced as a full vowel /œ/ in the stressed syllable. Verlan words inflect like ordinary nouns and adjectives (une meuf, des meufs). Several have entered the general familiar register and lost their disguise: ouf "crazy", meuf "woman, girlfriend", relou "annoying", chelou "shady".

ParisStories (2026-11-06) has ouf 8, meuf 4 and the base mec 4; GSD has none of them. In c'est un truc de ouf, insupportable. (ParisStories_2021_loulouLeChat_32), ouf is a noun after de, so the verlan adjective has been converted. Another example: enfin, c'était la première meuf, euh, avec qui je parlais dans la, dans la classe (ParisStories_2022_19_contenteDeOuf__18).

8.2 Largonji and javanais

Largonji (from jargon) moves the first consonant to the end of the word, puts l- in its place and adds a suffix (-em, -ji, -oque, -é): jargon → largonji, boucher → loucherbem (the butchers' slang, also louchébem), fou → loufoque, en douce → en loucedé. Loufoque and en loucedé have entered ordinary familiar French. Javanais inserts av after the consonant of each syllable: bonjour → bavonjavour. It has left nothing in the standard language.

9. Blends, reduplication, onomatopoeia

9.1 Blends (mots-valises)

A blend fuses two words and drops part of each. French calls it a mot-valise, a calque of Lewis Carroll's portmanteau. Nyrop treats it as contamination and gives an old example, éclabousser from éclater and Old French esbousser (t. III, §6). A GSD sentence explains the Pokémon names Doduo and Dodrio as two blends: il s'agit de deux mot-valises composés de « dodo » volatile disparu et, respectivement de « duo » et de « trio » (fr-ud-train_03645).

Blends are most productive where terminology work coins replacements for English words, above all in Québec (courriel, clavardage, baladodiffusion).

9.2 Reduplication and onomatopoeia

The 1990 list C welds many of these (blabla, froufrou, pêlemêle, pingpong, tohubohu; §3). Onomatopoeia gives verbs and nouns: miauler, glouglouter, cliquer, chuchoter, un plouf.

10. Verbal nouns and back-formation

True back-formation (English edit from editor) is rare in French. The common process that resembles it is the verbal noun (substantif verbal, déverbal): a noun made from the verb with no suffix.

What the base is. Egger took the base to be the infinitive "with its ending cut off", and Brunot's bibliography lists his study under that title (brunot-1899, bibliography). Darmesteter followed Diez instead. The base is the stem of the singular present indicative, which explains maintien, soutien, relief, whose vowel is that of the stressed stem (il maintient), not of the infinitive (darmesteter-1877, part I, §5). In 1877 the process was alive in popular speech (la balade, la boulange) and receding elsewhere (§5.1).

These nouns compete with the suffixed action nouns in -age, -ment, -ion, -ure, -ance, -ade, -aison, a central topic of French derivational research. VerNom (Missud, Amsili & Villoing 2020, missud-2020) collects 25,857 verb–noun pairs for those seven suffixes from frCOW, a nine-billion-word web corpus of 2016, with their frequencies. The authors compare it with earlier resources (736 event nouns in Nomage; 8,848 verb–action-noun pairs in the first Démonette) and note that the suffixless conversion pairs (marcher → marche, découvrir → découverte, arriver → arrivée) are not yet in VerNom. Démonette2 (Namer et al. 2019, namer-2019) is to describe a derivational network of at least 366,000 entries, conversion among them.

Real back-formations exist in the learned layer, where a verb is made from a noun in -ion (télédiffusion → télédiffuser). They are few.

11. Eponyms and trade names

Proper names give common nouns, usually with a lower-case letter, a gender from the thing named and a regular plural:

• persons: une poubelle (the prefect Eugène Poubelle, who made rubbish bins compulsory in Paris in 1884), une guillotine (Dr Guillotin), une silhouette (Étienne de Silhouette), une mansarde (François Mansart), la nicotine (Jean Nicot), le braille, une béchamel, un sandwich; the units un watt, un volt, un ampère, un newton;
• places, mostly by ellipsis (§5.4): un bordeaux, un champagne, un camembert, le tulle, le jean (Gênes, Genoa), une macédoine;
• characters: un harpagon, un tartuffe, un don juan, un gavroche;
• verbs: boycotter (Captain Boycott), pasteuriser;
• trade names: un frigidaire, un kleenex, du scotch, un caddie, un post-it.

The reverse direction, common nouns as surnames (Boulanger, Lefèvre), is in 11.

12. How Mova handles it

What Mova does today. Mova has no French tokenizer of its own, no French lexicon and no French morphology. French treebanks are read by the English parser en after en xpos-fill has derived Penn-style tags from UPOS and FEATS (en/src/main.rs, xpos_fill, xpos_of). On UD French GSD (UD 2.18, train → dev, 2026-10-22) the tagger reaches 96.47%. 7.2% of the dev tokens are unknown to it, and it tags those at 84.8%. The parser's LAS is 87.00 with its own tags and 89.78 with gold tags. None of the processes of this chapter is implemented for French. There is no compound analyser, no clipping, verlan or sigle table, no hyphen or welding normaliser, and no link from a verb to its verbal noun. A clipped word, a verlan form or a sigle unseen in training is tagged from its letters and its neighbours, like any unknown word.

What the treebanks annotate (UD 2.18, all files of each treebank, 2026-11-06):

Hyphenated words in GSD are single tokens with their own lemma. By UPOS they are PROPN 1,558 (place and person names such as Saint-Denis), NOUN 937, PRON 325 (-t-il 59, lui-même 56, celui-ci 54), ADJ 217, ADV 125 (au-dessus 30, peut-être 24, au-delà 24), X 54, CCONJ 52 (all c'est-à-dire), NUM 44 and VERB 25. The commonest hyphenated noun lemmas are the compass points (sud-ouest 24, sud-est 24, nord-ouest 20, nord-est 18), then porte-parole 16, demi-finale 16, week-end 15, rendez-vous 13, chef-lieu 13, vice-président 12 and après-midi 11. A compound written with spaces is an ordinary phrase in the tree. In chips de pomme de terre (fr-ud-train_04284), pomme is an nmod of chips and terre an nmod of pomme, so nothing marks the compound. compound is nearly unused: GSD's six cases are on English material (sports, fan, pom-pom, test). French UD analyses its N de N names as syntax, where English UD uses compound thousands of times. flat:name (7,005 in GSD) links the parts of names, and flat:foreign (1,075) the words of foreign phrases. The Foreign=Yes forms are mostly English words in titles (the 91, of 84, and 28), not loanwords: integrated loans such as week-end or parking carry no feature.

The tokenizer. On raw text, en's tokenizer (en/src/tok.rs) learns from the training treebank which hyphenated words to keep whole and which to split (Lexicon::learn, the fields hyphen_words, hyphen_split, hyphen_prefix). On GSD it would learn French compounds. When it splits, though, it emits the hyphen as a token of its own (a · - · t · - · il), while GSD attaches the hyphen to the clitic (a · -t-il). Its clitic list is English only (n't, 's, 're, 've, 'll, 'd, 'm), and nothing splits French elisions (l'ami, qu'il) or the contractions du, au. The multiword-token lines are skipped when CoNLL-U is read (en/src/conllu.rs), so the parser sees de le, never du. No French tokenization accuracy has been measured: en tok-eval on GSD would give it. All the scores above are on gold tokens.

Not done, in order of what it would buy.

1. A French hyphen and elision policy for the tokenizer: keep compounds, attach inverted clitics as UD does, split elided articles and pronouns, and expand du, au, des, aux into syntactic words.
2. A French lexicon in which compounds carry their plural (porte-monnaie, invariable or -s by spelling system), and clippings, verlan forms and sigles link to their bases (fac → faculté, meuf → femme, SNCF → Société nationale des chemins de fer français). The candidates are in 19 §8.
3. A normaliser between the traditional and the 1990 spellings (week-end/weekend, des après-midi/après-midis), so that both analyse alike.
4. A derivational layer linking verbs and their nouns (appeler → appel), taken from Démonette2 or VerNom once their data licences are checked. The papers are CC BY 4.0, but that says nothing about the data.

Sources for this chapter

• Ferdinand Brunot, Précis de grammaire historique de la langue française, 1899, §§114–121 (compound and juxtaposition, order, reduction to simple words, rouge-gorge, Hôtel-Dieu), bibliography (Egger on verbal nouns). Public domain. brunot-1899.
• Arsène Darmesteter, Traité de la formation des mots composés, 2nd ed., 1894, with Gaston Paris's preface (the imperative in porte-feuille compounds, synecdoche in rouge-gorge). Public domain. darmesteter-1894.
• Arsène Darmesteter, De la création actuelle de mots nouveaux, 1877 (dérivation impropre, the verbal noun from the present stem, the vitality of the processes, argot). Public domain. darmesteter-1877.
• Kristoffer Nyrop, Grammaire historique de la langue française, t. III, Formation des mots, 1908, §§5–6 (abbreviation, initials, contamination), §571 (ellipsis). Public domain. nyrop-1899.
• Conseil supérieur de la langue française, Les rectifications de l'orthographe, Journal officiel, 6 December 1990: analysis I.1–I.2 and I.6, rules II.2 and II.7, recommendation IV.2, lists A, B, C, F. Official text, freely reusable. cslf-1990.
• A. Koptient, La désambiguïsation des abréviations du domaine médical, TALN-RÉCITAL 2020. CC BY 4.0. koptient-2020.
• A. Missud, P. Amsili & F. Villoing, VerNom : une base de paires morphologiques acquise sur très gros corpus, TALN 2020. CC BY 4.0. missud-2020.
• F. Namer, L. Barque, O. Bonami, P. Haas, N. Hathout & D. Tribout, Demonette2 : une base de données dérivationnelle du français à grande échelle, TALN 2019. CC BY 4.0. namer-2019.
• UD 2.18 French treebanks GSD, Sequoia, ParisStories, Rhapsodie, FQB, PUD, ALTS: counts of 2026-11-06 over all files of each. Quotations from GSD and ParisStories (CC BY-SA 4.0), short, with sent_id; Sequoia and FQB (LGPL-LR) for counts only.
• The English manual, chapter 18.
• Mova: en/src/main.rs (xpos_fill, xpos_of), en/src/tok.rs, en/src/conllu.rs; the baseline in What every language gets and the worklog of 2026-10-22.

This page in Markdown → https://md.mova.elamur.ai/free/en/docs/languages/fr/18-compounding-conversion-minor/