🗣️ 17. Derivation II: suffixes

A suffix is a bound element added after a base. Unlike a prefix (chapter 16) it usually changes the word class, and it fixes the gender of the noun it makes: laver (V) → lavage (N, m), lavable (A), laveur (N, m), laveuse (N, f), laverie (N, f). Nyrop's volume on word formation (nyrop-1899, t. III) gives the suffixes most of its pages: nominal suffixes sorted by origin (Latin of popular formation, Latin of learned formation, foreign, French-made, of doubtful origin), the changes of suffix over time, and the verbal suffixes.

The chapter covers the general properties and the changes at the boundary (§1), nouns by meaning and the gender each suffix gives (§2), adjectives (§3), verbs (§4), adverbs (§5), popular and learned pairs (§6), what research has measured (§7), counts on the treebank (§8) and how Mova handles suffixes (§9). Compounding and conversion are in chapter 18; inflection (gender, number, verb endings) in chapters 11, 13 and 15.

1. General properties

1.1 What a suffix does

1.2 What happens at the boundary

Nyrop's chapter on "the fate of the primitive word" (nyrop-1899, t. III, §45) starts from the same observation: most suffixes attach to the word unchanged (fainéant → fainéantise, journal → journalisme), and the alternations are the exceptions. The last row of the table is a form of suppletion: the family has two stems (eau and aqua-: aquatique, aquarium). Démonette 2 treats such pairs as relations made opaque by a learned stem, with Saint-Étienne / stéphanois as its example (namer-2019). Before -iser and -ifier the learned stem is common too: électrique → électrifier.

1.3 Popular, learned and borrowed suffixes

• Popular suffixes continue Latin suffixes through the spoken language and were worn down by sound change: -eur (< -orem), -ier (< -arium), -aison (< -ationem), -té (< -itatem), -ure (< -atura), -ée (< -ata).
• Learned suffixes were taken from written Latin and Greek with their shape: -ation, -ité, -aire, -ique, -iste, -isme, -if, -al. According to Nyrop (t. III, §311), -ation first came in borrowed words (adoration, circulation), and on their model new action nouns were made from French verb stems (association, autorisation); he also notes that the learned form of a suffix can replace the popular one, as -ation took over from -aison (t. III, ch. IV).
• Borrowed suffixes are few (t. III, §350): Germanic -ard (vieillard, bavard) and -aud (lourdaud) from the oldest layer; -ade and -esque from the south and Italy after the Renaissance (promenade, romanesque); and recently English -ing (parking, camping; chapter 19).

2. Nouns

2.1 Action and result: from verbs

Competition. One verb can have two or three nouns with partly different senses: montage / montée, lavage / lavement ("enema"), abattage / abattement, éclairage / éclairement. Broad tendencies: -ation is the productive choice after -iser and -ifier (modernisation, vérification, mondialisation) and for learned verbs; -age leans to technical operations and processes (usinage, filtrage, piratage); -ment to states and results (changement, engagement); -ure and -ée take the participial stem and name a result or an extent. These are tendencies, not rules, and each pair is a fact of the lexicon.

What a large count shows. VerNom (Missud, Amsili & Villoing 2020, missud-2020) pairs verbs with nouns in seven suffixes, extracted by rules from frCOW, a web corpus of 9 billion words, and cleaned in several steps. It holds 25 857 pairs:

The first three make 85% of the pairs. The authors keep rare words and neologisms on purpose (pipolisation; the base even pairs kikouloler with kikoulolage), because earlier resources were too small to study rival suffixes: 8 848 verb/action-noun pairs in Démonette, 736 event nouns for 679 verbs in Nomage. Its quality was checked on four random samples of 100 pairs.

2.2 Agent, instrument and place

Two -eur series. The popular -eur attaches to the French verb stem and has the feminine -euse: chanter → chanteur / chanteuse, vendre → vendeur / vendeuse. The learned -teur continues Latin -tor on a Latin participle stem, with the Latin feminine -trice: acteur / actrice, directeur / directrice, lecteur / lectrice, instituteur / institutrice; -ateur / -atrice is the same on -er verbs (animateur / animatrice). A few old agent nouns have -eresse (vengeur → vengeresse, enchanteur → enchanteresse; demandeur → demanderesse survives in law), and ambassadeur has ambassadrice. Agent and instrument share the suffix (un batteur is a drummer or a whisk), and some -eur nouns are only instruments (ordinateur).

Lexeur (Wauquier 2018, wauquier-2018) is a hand-annotated resource of 5 974 agent nouns in -eur, taken from the TLFi and the web, each linked to its feminine in -euse and/or -rice, to its base (verbal or nominal) and to its action nouns. The feminine in -euse is far more frequent than in -rice (4 542 against 1 514); 78% of the entries have a verbal base; 22% of the families have no action noun, and the others have 1.47 on average (1 to 8). Building a mean "prototype" vector of the agent class in word embeddings, the study finds that its nearest neighbours depend on the corpus (Wikipedia, a newspaper, the web) and that words drift with lexicalisation (échangeur is now a motorway interchange more than a person who exchanges). The prototype of the feminine agent nouns has its own neighbourhood, trade names, words of topics culturally associated with women and women's names, which the author links to the negative connotation reported in earlier work: the feminine is not just the masculine with a gender change.

Feminine titles. For person nouns in -eur the feminine can be -euse, -rice, -eure, or the masculine with a feminine determiner, and the official guides of France, Belgium, Switzerland and Québec differ: une autrice / une auteure, une professeure, une ingénieure, une écrivaine, une cheffe. The full rules of the feminine are in chapter 11. In GSD (UD 2.18, all files, 2026-11-06) the feminines have lemmas of their own and are rare: 24 noun lemmas in -euse (60 tokens) and 21 in -trice (72 tokens; matrice among them is no agent noun), against 408 in -eur (2 462 tokens). The commonest are actrice 24, chanteuse 19, joueuse 9, impératrice 7: l'actrice et chanteuse Brigitte Nielsen (fr-ud-train_01436); en tant qu'actrice (fr-ud-train_01250).

2.3 Quality and state: from adjectives

-ité is the productive member: it attaches to learned adjectives and above all to -able/-ible and -al/-el (durabilité, lisibilité, mobilité; nationalité, réalité). -té is the popular form of the oldest layer (bonté, beauté). -esse (human qualities) and -eur (dimension and sensation: grandeur, chaleur) make few new words. The quality nouns in -eur are feminine, unlike the agent nouns: la hauteur, la longueur but le chanteur. Medieval French hesitated freely between these suffixes: Nyrop cites laideur, laidesse, laideté side by side (t. III, ch. IV); the modern language has kept one form per adjective, and which one is a fact of the lexicon.

2.4 Collective, place, tree, quantity

Une dizaine, une centaine mean "about ten, about a hundred" (une douzaine also "a dozen"), and la trentaine is an age (12. Determiners, pronouns and numerals). A shop in -erie has its keeper in -ier/-ière (boulangerie / boulanger, épicerie / épicier). In GSD the commonest nouns in -erie are military and institutional: infanterie 28, batterie 22, artillerie 18, galerie 18 (no derivation), cavalerie 12, gendarmerie 10 (2026-11-06).

2.5 Doctrines and their people: -isme, -iste, -ien

-isme names a doctrine, movement or attitude (socialisme, féminisme, capitalisme, réalisme, tourisme), a figure of speech (euphémisme) or a word from another language (anglicisme, belgicisme). The person is in -iste (socialiste, alpiniste), which also makes adjectives and is epicene (un/une journaliste), or in -ien for a thinker (kantien, cartésien). Both are productive (complotisme / complotiste, populisme), and in GSD -isme has the highest hapax rate of the large noun suffixes (§8). The English -ism, -ist are the same Greek-Latin suffixes, so these pairs translate predictably.

2.6 Diminutives and expressive suffixes

The diminutive often expresses affection or contempt rather than size, and many diminutives have lost the link to their base (une assiette is a plate). French diminutives are less free on common nouns than Italian -ino or Spanish -ito; first-name hypocoristics are the clearest living use (Jeannette, Pierrot). In GSD the string is a poor guide: 81 noun lemmas end in -ette (208 tokens) and 105 in -et (867), but the commonest are not diminutives (dette 20, recette 12, toilette 11, assiette 10; effet 156, juillet 131, projet 127, objet 60).

2.7 The gender of the derived noun

The suffix fixes the gender, with few exceptions. The table counts the Gender feature of GSD noun tokens whose lemma ends in the string and is longer than it (UD 2.18, all files, 2026-11-06). The match is by letters, so unsuffixed words are included; they explain most of the minority:

GSD marks Gender on 75 175 of its 75 330 noun tokens, so the counts are nearly complete. These 17 endings cover a large share of derived nouns and predict the gender almost perfectly except where an unsuffixed word shares the ending (-age, -eur, -té) or the noun is epicene (-iste). The research agrees: Sahai & Sharma (sahai-2021) trained classifiers on 124 803 French nouns from Dbnary (a Wiktionary-based database) and reached 92.5% with the spelling alone, 77.2% with the dictionary definition alone and 94.0% with both; adding pronunciation helped little. Their "textbook" baseline (masculine unless the word ends in -tion, -sion, -té, -son or -e, but not -age, -me, -ège) is the weak rule they beat.

3. Adjectives

-able / -ible. -able makes adjectives meaning "that can be V-ed" from verbs (lavable, mangeable, utilisable; on the stem of -ir verbs: finissable), often with in- (incontournable, impensable, inoubliable), and some older ones mean a quality (aimable, capable, confortable). -ible is found in borrowed Latin words (possible, lisible, visible, compréhensible, éligible) and does not make new adjectives, though its nouns in -ibilité do. Before -able a g keeps its sound with e (mangeable, échangeable). GSD has 141 adjective lemmas in -able (574 tokens) and 40 in -ible (268); the commonest, possible 61, faible 54, agréable 45, véritable 36, disponible 34, capable 34, responsable 33, are old lexicalised words (faible, véritable not even V + -able).

-al / -el / -ique. Relational adjectives do much of the work that English gives to noun modifiers: la vie culturelle "cultural life", le système scolaire "the school system", la politique économique. French has noun–noun compounds too (wagon-restaurant, mot clé; chapter 18), but the relational adjective is the ordinary choice. -al is the general form; -el is regular in some contexts, -uel (annuel, ponctuel), -iel (officiel, essentiel), -onnel (rationnel, professionnel) and after -tur- (culturel, naturel), and occurs in old words (mortel, cruel, éternel). Masculine plural -aux (nationaux, médicaux), with the exceptions banals, bancals, fatals, finals (also finaux), natals, navals (chapter 15). -ique belongs to Greek and Latin words and to the sciences (cubique, lyrique, atomique).

-eux. On popular bases (boueux, pluvieux, honteux) and learned ones (audacieux, ambitieux, religieux); the noun has -osité with the alternation /ø/ ~ /oz/ (nervosité, curiosité, générosité) and the adverb -eusement. GSD: 111 adjective lemmas in -eux, 674 tokens.

Gentilics. Names of inhabitants and languages come from place names with suffixes that cannot be predicted from the name: Paris → parisien, Lyon → lyonnais, Marseille → marseillais, Rome → romain, Danemark → danois, Bretagne → breton, Maroc → marocain, sometimes on a learned stem (Saint-Étienne → stéphanois). The adjective is lower case (un film français), the inhabitant capitalised (un Français), the language a lower-case masculine noun (le français) (07. Punctuation, typography, capitals, abbreviations and numbers). GSD has 160 adjective lemmas in -ien (842 tokens), 58 in -ais (777) and 40 in -ois (135), not all of them gentilics. A system needs a list.

4. Verbs

Only two models. Nyrop states it as a rule (t. III, §425): French verb derivation follows only the first conjugation (mur → murer) or the second (lot → lotir), directly or through a suffix (poète → poétiser, fumer → fumoter), and no French verb has been made on the model of the verbs in -oir or -re. Today nearly every new verb is first-group, borrowings included (scanner, liker, zapper, tweeter); the second group's one living pattern is the change of state from an adjective; the third group is closed (13. Verbs).

-iser and -ifier. -iser (Greek -izein through Latin) is the main suffix of verbs from adjectives, nouns and names (américaniser, mondialiser, médiatiser, hospitaliser), especially on -al, -el, -ique (nationaliser, officialiser, dramatiser); -ifier (Latin -ificare) takes learned bases and makes fewer new verbs. Their nouns are -isation and -ification. GSD has 152 verb lemmas in -iser (984 tokens) and 34 in -ifier (235): utiliser 241, réaliser 144, organiser 99, signifier 49, qualifier 43, préciser 36, modifier 28 (viser 37 ends in the letters without the suffix). First-group verbs make 1 828 of the 2 198 verb lemmas (83%) and 19 866 of the 31 769 verb tokens (63%).

Participles turned adjectives and nouns. Present participles in -ant and past participles are inflected forms (13. Verbs), but many are lexicalised as adjectives (brillant, charmant, fatigué, connu) and nouns (un étudiant, un employé, une mariée, un perdant). An agent noun in -ant (un habitant) is a converted participle, a different formation from the derived -eur noun (un chanteur).

5. Adverbs

The productive adverbial suffix is -ment, added to the feminine of the adjective; the rules and exceptions are in chapter 15. In brief: lent → lentement, heureux → heureusement, doux → doucement, fou → follement; adjectives that end in a vowel in the masculine take -ment on that form (vrai → vraiment, poli → poliment, absolu → absolument); -ant/-ent adjectives give -amment/-emment, both pronounced /amɑ̃/ (constamment, évidemment, récemment); a few take -ément (énormément, précisément, profondément); and gentil → gentiment. The 1990 rectifications drop the circumflex of assidûment, crûment, dûment, goulûment and of the other adverbs of this type, giving ambigument, assidument, congrument, continument, crument, dument, goulument, incongrument, indument, nument, like the long-established absolument (cslf-1990).

The adverb -ment and the noun -ment have different origins: the adverb continues the Latin ablative mente "with a mind" (clara mente → clairement), the noun continues Latin -mentum. In GSD (all files, 2026-11-06) word forms ending in -ment are 4 268 tokens: 2 497 ADV (58.5%), 1 677 NOUN (39.3%), 57 VERB (third-plural forms of verbs in -mer: forment, affirment, estiment), 32 PROPN. By distinct forms the adverbs are 54.2%. The ending alone is therefore nearly a coin toss between adverb and noun, and on forms unseen in training it is an exact one (of the 31 dev tokens in -ment whose form is not in the training file, 14 are adverbs and 14 nouns). What separates them is the base: an adverb in -ment is built on a feminine adjective (lente → lentement) or on -ant/-ent (-amment, -emment), while the nouns are built on verb stems (gouvern-er → gouvernement) or are unanalysable (moment).

GSD has 364 adverb lemmas in -ment (2 498 tokens) of 547 adverb lemmas (14 501 tokens): two thirds of the adverb types and 17% of the tokens. Comment (51 tokens) is among them but is no derivation. The commonest are également 321, notamment 190, seulement 74, vraiment 69, particulièrement 57, finalement 52, rapidement 51; in -amment/-emment: notamment 190, récemment 21, évidemment 10, constamment 9. Other adverb-forming endings are not productive (à reculons, à tâtons); most other adverbs are simple, phrasal (à peine, sans cesse) or adjectives used as adverbs (parler bas, coûter cher).

6. Popular and learned forms of one Latin suffix

As with the prefixes (chapter 16 §9), a Latin suffix often reached French twice. Brunot (brunot-1899, §165) describes the general mechanism of doublets: the vocabulary has a popular, a learned and a foreign layer, and one Latin word can appear in two of them with distinct senses (hôtel and hôpital from hospitale; façon and, through English, fashion from factionem). Suffixes show the same pattern inside words:

In productive use the learned form wins (-ation, -ité, -aire); the popular forms stay in the old words. The practical consequence is that a French suffix often has two shapes, each going with its own kind of stem: popular with the French stem (raison, bonté), learned with the Latin one (ration, bénignité).

7. What the research shows

• Event nouns. VerNom (missud-2020, §2.1) is designed for the quantitative study of the seven rival suffixes, with the frequencies of verb and noun in frCOW and the source of each pair.
• Agent nouns. Lexeur and Wauquier's distributional study (wauquier-2018, §2.2): the semantic content of -eur depends on the corpus and on lexicalisation, and the feminine agent nouns have a distinct distributional profile.
• Families and paradigms. Sanacore, Hathout & Namer (sanacore-2019, sanacore-2020) start from a family such as laver, lavage, lavoir, laverie, laveur, laveuse, lavette, lavable, lavement and align families on the same relations: imprimer, souder, laver, nettoyer each have a masculine agent noun (imprimeur, soudeur, laveur, nettoyeur), a modal adjective (imprimable, soudable, lavable, nettoyable) and an action noun (impression, soudage, lavage, nettoyage). The relation is the same though the suffix of the action noun changes. The 2020 paper describes the meanings within a family as "morphosemantic frames" in which each lexeme is typed and defined by its relation to the others.
• Paradigms are semantic. Démonette 2 (namer-2019) annotates each relation with a semantic type and an abstract semantic schema, and groups families into semantic paradigms. Its example: enseigner, enseignant, enseignement and chanter, chanteur, chant fall in one paradigm centred on a dynamic situation, while admirer, admirateur, admiration, whose verb is stative, does not, although the suffixes look alike.
• Polysemy of a suffix. Andreou & Petitjean (andreou-2017), modelling English deverbal nouns in -al (rental, disbursal) with frame semantics in XMG, argue that an affix's meaning cannot always be reduced to one underspecified meaning, and that type constraints decide which argument of the base verb the derived noun can refer to. The French -eur (agent, instrument, place) and -age (action, result, collection) raise the same question; the paper does not study French.

None of these gives Mova rules directly; they give data (VerNom, Démonette 2, both described in CC BY papers; the data licences must be checked before use) and a principle, families aligned in paradigms, that a French lexicon could follow.

8. Counts in the treebank

UD French GSD, UD 2.18, train + dev + test, 2026-11-06 (356 380 non-punctuation words). A lemma counts under a suffix when it ends in the letters, is longer than them, and has the stated UPOS; this is a string count that includes unsuffixed words. Hapax are lemmas with one token; P is hapax divided by tokens, Baayen's measure of productivity. On a corpus of 0.36 million words P is only a rough guide.

Class totals for comparison: nouns 9 466 lemmas / 75 330 tokens; adjectives 3 683 / 23 820; verbs 2 198 / 31 769; adverbs 547 / 14 501.

Reading it. The large suffixes (-tion, -ment, -eur, -ité, -al) have P of 0.05–0.07: well established, each new text brings few new members. -isme, -iste, -erie, -ette, -able have the highest P. For -isme (73% of its lemmas are hapax) the encyclopedic content inflates the figure, with one article per religion, movement or doctrine; for -ette the lexicalised words (§2.6) make the string count mean little. A balanced corpus would rank them differently.

How well the ending predicts the class (word forms, lowercased, 2026-11-06):

A suffix is a strong class cue for -ation, -iser, -able, and a weak one for -ique and -ment, where the context or the base must decide (§5).

9. How Mova handles suffixes

There is no French derivational morphology in Mova: no suffix list, no boundary rules, no families. French is read by the English parser en after en xpos-fill has given each treebank word a Penn-style tag computed from its UPOS and features (xpos_fill, xpos_of in en/src/main.rs; GSD baseline, UD 2.18, train → dev, 2026-10-22: LAS 87.00 with the parser's own tags, 89.78 with gold tags, PACKAGE.md). What bears on suffixes:

• Suffix features of the tagger. The tagger used by parse-eval is the averaged perceptron of en/src/ptag.rs; its features include the last one to four letters of the word (and the first one to three). For words unseen in training it adds a prior from the trigram tagger en/src/tag.rs, whose suffix model uses up to the last ten letters, learnt from the rare words of the training file, separately for capitalised and lowercase words. This is, in effect, all the suffix knowledge Mova uses on French: learnt letter strings, not suffixes. It works where the ending is a strong cue (-ation, -iser, -able, §8) and must rely on the neighbours for -ment and -ique. Accuracy on words unseen in training, GSD dev: 84.8% (2026-10-22).
• English derivation. en/src/derive.rs analyses an English word unknown to the dictionary as a dictionary word plus productive English affixes, keeping an analysis only if it composes back to the form; its rules are in the English word-formation seeds (en/seeds/morph/). Both taggers use it, with the English dictionary, as a hint for unknown forms. It knows no French suffix; on French it gives no hint for most words and a spurious one for some (an analysis with English affixes); this has not been measured.
• The tags. xpos_of maps NOUN to NN/NNS (by Number), ADJ to JJ, ADV to RB/WRB, VERB to VB/VBN/VBD/VBZ/VBP. Gender is not in the tag set, so the gender evidence of §2.7 is not used and agreement is not checked. Lemmas are the treebank's; GSD gives a feminine agent noun its own lemma (actrice, chanteuse), and Mova has no lemmatiser for unseen words.

What it does not do. No segmentation into base and suffix; no link between organiser and organisation (so no help for coreference such as organiser … l'organisation); no gender prediction for unseen nouns; no generation of feminine agent nouns or -ment adverbs; no boundary rules (§1.2) or learned stems (eau / aquatique); nothing for the new coinages (-isme/-iste nouns, -iser verbs), which are exactly the unknown words a tagger meets. For dependency parsing the cost is moderate, since a sentence's tree does not depend on how its words are built; for tagging unknown words and for any lexical task it is complete.

A first step. A French counterpart of derive.rs: a short rule table for the endings with a clear class and gender (§§2.7, 8), with -ment read as an adverb only when what precedes it is a known feminine adjective or an -amment/-emment form of an -ant/-ent adjective, standing on a French lexicon of bases (chapter 19); VerNom-style verb–noun pairs as links once their data licence is checked. It must be measured on GSD dev unknown words against the present tagger as negative control; a rule that does not beat it is not kept.

Sources for this chapter

• Kristoffer Nyrop, Grammaire historique de la langue française, t. III, Formation des mots (1908): §45 (the base unchanged), ch. IV (changes of suffix: laideur / laidesse / laideté, -aison replaced by -ation), §311 (-ation), §350 (borrowed suffixes), §425 (verbal derivation only on the first and second conjugations) (nyrop-1899, public domain).
• Ferdinand Brunot, Précis de grammaire historique de la langue française (1899), §165, doublets (brunot-1899, public domain).
• Conseil supérieur de la langue française, Les rectifications de l'orthographe (1990): the adverbs in -ument (cslf-1990, official text).
• Alice Missud, Pascal Amsili & Florence Villoing, VerNom: une base de paires morphologiques acquise sur très gros corpus (TALN 2020) (missud-2020, CC BY 4.0).
• Marine Wauquier, Analyse des noms agentifs dans les espaces vectoriels distributionnels (TALN-RECITAL 2018) (wauquier-2018, CC BY 4.0).
• Daniele Sanacore, Nabil Hathout & Fiammetta Namer, Semantic descriptions of French derivational relations in a families-and-paradigms framework (DeriMo 2019) (sanacore-2019, CC BY 4.0); Représentation sémantique des familles dérivationnelles au moyen de frames morphosémantiques (TALN 2020) (sanacore-2020, CC BY 4.0).
• Fiammetta Namer et al., Demonette2: une base de données dérivationnelle du français à grande échelle (TALN 2019) (namer-2019, CC BY 4.0).
• Marios Andreou & Simon Petitjean, Describing derivational polysemy with XMG (TALN 2017; English data) (andreou-2017, CC BY 4.0).
• Saumya Sahai & Dravyansh Sharma, Predicting and explaining French grammatical gender (2021) (sahai-2021, CC BY 4.0).
• UD French GSD, UD 2.18 (CC BY-SA 4.0): counts of §§2, 5, 8 (2026-11-06) and quoted sentences with their sent_id.
• Mova: en/src/main.rs (xpos_fill, xpos_of, parse-eval), en/src/ptag.rs, en/src/tag.rs, en/src/derive.rs; baselines in PACKAGE.md.

This page in Markdown → https://md.mova.elamur.ai/free/en/docs/languages/fr/17-derivation-suffixes/