🗣️ 09. Old, Middle and Classical French, and the literary tenses

The French in the rest of this manual is the French of about 1800 to today. The texts that Mova will meet also include French of earlier centuries: medieval epics and romances, the legal prose of Normandy, the Bible of the sixteenth century, Molière and Racine, the novels of the eighteenth century. They are read in schools, quoted in newspapers and held in every library of the Francophone world. This chapter says what the earlier states of the language were like, where they differ from the modern language in sound, spelling, morphology and syntax, what the literary tenses are (the passé simple and the imperfect subjunctive, which are archaic in speech and alive in writing), and how Mova handles them today. The short answer for Mova is that it handles none of it: the parser is trained on modern French, and §10 measures what happens when it meets an older text.

The treebanks of the earlier states are used for counts only where the licence allows quoting (§9); the PROFITEROLE treebanks are CC BY-NC-SA and are used for measurements only, never quoted, so the examples taken from them are constructed or taken from the public-domain editions (the Chanson de Roland, the Strasbourg Oaths).

1. The periods

The boundaries are conventional. Nyrop's Grammaire historique (nyrop-1899) and Brunot's Histoire de la langue française (brunot-1906) use "ancien français" for the period up to the end of the 13th century and "moyen français" for the 14th and 15th, with the 16th as a transition; Brunot treats the 16th century as the decisive period of the fixation of the language, in which the king's ordinance of Villers-Cotterêts (15 August 1539), which in its articles 110 and 111 prescribed that all judicial acts should be made in French, is the landmark of the official language (brunot-1906). The dates in the treebanks' own names follow the same division: PROFITEROLE says "from the 9th to the 13th century" for Old French and "from the early 14th to the late 15th" for Middle French.

The language of the Old French period was not one language. It was a group of dialects of the langue d'oïl (Francien of the Île-de-France, Norman, Picard, Champenois, Burgundian, Lorrain, Anglo-Norman, Walloon and others), with a written scripta in each. Francien became the standard slowly, with the political centre of the kingdom (Dauzat, dauzat-1908 and dauzat-1922, on the spread of the language of the capital and the retreat of the dialects). The Old French treebank mixes dialects: Yvain is in the dialect of Champagne, the Roland in the Anglo-Norman manuscript of Oxford, and Beroul's Tristan in a Norman-Anglo-Norman dialect.

2. Sounds, in brief

The manual describes modern sounds in the chapters 01–04. The history is shorter to state than to list.

• Vowels. Old French had many diphthongs and triphthongs that disappeared by the 16th century: ai, ei, oi, ie, ou/eu, uo → ue and a triphthong eau [eaw] (Nyrop's summary of the phonetics of Francien: nyrop-1899, §21). The vowel of mur was [y] and not [u] as in Latin; veir [vɛir] became voir. The nasal vowels were not yet independent; the nasal consonant was still pronounced after the vowel; later the nasal vowels arose and became lower.
• Consonants. The affricates [ts] and [dz] of cent and doze became [s] and [z]; [ʃ] and [ʒ] remained. The s before a consonant was pronounced (feste, beste) and was lost, leaving a long vowel or an accent in the spelling (fête, bête, forêt: the circumflex of 05); the final consonants were pronounced (chantet, murs, vint); the l in front of a consonant was vocalised (albe → aube, altre → autre, chevals → chevaux).
• The final -e. Unstressed vowels other than a disappeared in the passage from Latin to Old French (servire → servir, minus → meins), and the Old French -e was a weak vowel [ə] that could count as a syllable. In classical verse it counts as a syllable before a consonant and is elided before a vowel; modern prose speech has lost it (03. The syllable, stress, rhythm and intonation).
• Stress was on the same syllable as in Latin; with the reduction of the end of the word it came to fall on the last full syllable (03).

The sound changes explain the oddity of French spelling (05. The writing system): many silent letters were pronounced in the Old and Middle French period and the spelling was carried on.

3. Spelling and the problem of the text

Medieval spelling is not fixed. The manuscripts of one text show several spellings of one word (chose, cose, chouse), the scribe's dialect colours the forms, abbreviations are numerous (the tilde for m and n, the ampersand for et), and a word is not always written as a unit. For printing, the editors normalise: they expand abbreviations, fix word division, add accents and apostrophes. A treebank text such as the Old French ALTM is digitised from an edition (the Atiremens et jugiés d'eschequiers of 1314, edited by Génestal and Tardif in 1921) with the abbreviations expanded.

The 16th century has its own difficulty: printed texts with unfixed spelling (ledit jour vendredy 28 octobrix 1547, en l'Evesché). Three recent papers study the automatic normalisation of such texts:

• Gabay and colleagues (gabay-2020) frame the normalisation of 17th-century French as machine translation and compare a statistical and a neural system; they stress the philological work behind the training corpus, and aim at a tool for philologists.
• Rubino and colleagues (rubino-2024) normalise the orthography of 16th-century Middle French archives (a parallel corpus of manuscript and normalised text) while keeping the historical word forms: they do not modernise (their example normalises Ledictz jour, vendredy 28 octobrix 1547, en l'Evesché to Ledit jour vendredi 28 octobris 1547 en l'Évêché, where octobris keeps its period form), because spelling reduction without modernisation is what a diachronic study needs. Pre-trained causal language models and knowledge distillation improved on a strong baseline.
• Solfrini and colleagues (solfrini-2025) present a dataset of about 40 000 lines of 16th-century printed texts with "semi-diplomatic" normalisation rules, and a first model, as one step of a chain from OCR to machine-actionable data.

Normalisation is the first thing any tool for earlier French has to do, and it is a neural-translation task in each of these three papers, not a rule system.

4. Old French in outline

4.1 Nouns, articles and the two-case system

Old French had two cases for nouns: the cas sujet (subject, nominative) and the cas régime (oblique: object, prepositional complement, possessor). Masculine nouns show it by an -s in the singular subject and none in the plural:

(Nyrop gives the table with murs/mur and says that the old subject case "succumbs" before the oblique: nyrop-1899, vol. I §§21–27.) The declension carried the grammatical relation and so freed the word order. Nyrop's example is li chiens mort le cerf, with the same meaning as le cerf mort li chiens, the case marking showing who bites whom; another: mes amis plore sa seror ("my friend weeps for his sister") against ma suer plore son ami. The oblique also served as the possessive (li fiz le rei "the king's son", without a preposition). In Middle French the case system is lost, and the oblique case survives as the modern form: mur, murs, so that -s becomes the mark of the plural. The authors of the fifteenth century still used the old subject forms but had lost their exact value; Nyrop notes that Villon, trying to write a ballad "in old French" in the middle of the century, produced only a jumble of old forms (nyrop-1899, vol. I §§26–27).

Feminine nouns in -e had no subject s. Pronouns kept case longer (je/me/moi, il/le/lui, li for the dative). The definite article was optional where the modern language requires it: il prent espee "he takes (a) sword". The marker of the nominative masculine singular and plural article li (li rois "the king", li chevalier "the knights") is the most frequent trace of the case system in the texts: in the Old French PROFITEROLE treebank li occurs 4 995 times in 240 335 tokens; in Middle French PROFITEROLE 25 times in 119 001; in the 16th-century ALTS never; in GSD once, which is a name (counts of 2026-11-06).

4.2 Verbs

Old French verbs have more stem alternation and more endings than modern ones. The present of amer had the forms aim, aimes, aimet, amons, amez, aiment, with the stress-based alternation of the stem (aim- under the stress, am- without it), as Nyrop describes it (nyrop-1899, vol. I); the analogical levelling made the modern aime, aimes, aime, aimons, aimez, aiment. The first person singular had no ending in the present (je chant, je part), the -e being added later by analogy. The third person was -t (aimet), which survives in the modern -t of il sert, il vient.

The simple past (the ancestor of the passé simple) was a lively tense of narrative, with a rich set of strong forms, many of which survive: fis < feci, mis < misi, dis < dixi, vis < vidi, vint < venit (nyrop-1899: the -s of the first and second singular was extended by analogy, and a "u" form arose for -ui perfects: oi, ot > eus, eut). The compound tenses (j'ai chanté) were in the process of becoming tenses: the participle agreed with the object (constructed: il a prises les armes, with the participle placed before the object and agreeing with it).

Verbs of Old French had the auxiliary estre "to be" for most intransitive verbs, as the modern language still does for verbs of motion, and the future and conditional are the synthetic forms of Latin infinitive + habere (chanterai, chanterais), already formed.

4.3 Syntax

Old French syntax differs from the modern one in four ways that matter for parsing.

The verb tends to be in second position (V2). Any constituent, not only the subject, may stand first, and the finite verb then follows it immediately, with the subject after it if the first element is not the subject: Lors vint li rois "then came the king" (constructed). The construction declines in Middle French and is formulaic by the end of the period. Three recent open studies measure it:

• Larrivée (larrivee-2019), on a homogeneous corpus of Norman prose legal texts from 1150 to 1475, shows that the preverbal complement of a V2 clause is a discourse-old element (already mentioned or inferable) in the period 1150–1236, and that this categorical information value disappears before V2 itself ceases to be productive; the construction becomes formulaic.
• Wolfe (wolfe-2020) shows from a corpus of medieval texts that the relative frequencies of null, preverbal and postverbal subjects are connected with the mapping between syntax and information structure, which changes during the period.
• Larrivée, Pinzin, Poletto, Romanova and Touré (larrivee-2025) tested on a treebank the correlation between V2 and three "micro-cues" suggested in the literature (cues that a learner can read off the data), and found the expected statistical correlation, which supports the micro-cue model of acquisition.

The subject pronoun is often absent. Old French was a null-subject language in the main clause (Si prist l'espee "so he took the sword", constructed), and the pronoun is more a topic marker than a required subject. In the Old French PROFITEROLE treebank, of 15 691 finite verbs that are the root of their sentence, 6 748 (43%) have an overt subject (nsubj or a subtype); of those, 2 516 (37%) have the subject after the verb. In modern GSD, 7 193 of 7 594 (95%) have an overt subject and only 143 (2%) have it after the verb (counts of 2026-11-06, all files). The Old French ALTM (a legal text of 1314) has 289 of 329 (88%) with an overt subject, 45 of them postverbal, closer to the modern language; Middle French PROFITEROLE has 1 882 of 3 437 (55%), 525 postverbal; the 16th-century ALTS has 1 002 of 1 517 (66%), 129 postverbal. The share of clauses with an overt subject is 43% in Old French PROFITEROLE, 55% in Middle French PROFITEROLE, 66% in ALTS and 95% in GSD: the direction is the decline of the null subject and of subject inversion. The corpora differ in genre (verse, prose, law; the 1314 legal text is an exception at 88%), so the figures are indications, not a rate of change.

Negation. Old French negated with ne alone (Il ne vient), which is as in the modern literary language. A reinforcer (pas "step", point "dot", mie "crumb", goutte, mot) could be added but was not required. In the Old French PROFITEROLE treebank the negator forms ne, n', nen, nel, nes, non occur 5 367 times and the reinforcers pas, point, mie 719 times (a ratio of 7.5 to 1); in Middle French PROFITEROLE 1 232 and 365 (3.4 to 1); in ALTS 752 and 170 (4.4 to 1); in GSD, the written modern language, 1 773 and 1 142 (1.6 to 1; the figure counts ne/non and pas/point everywhere, not only in negation: 27. Negation gives the exact modern rates). Old French negation is thus mostly ne alone, and the two-part ne … pas is a development of Middle French.

Clitic placement. The object pronoun could precede or follow an infinitive and could be separated from its verb; a pronoun object at the head of a clause was bound to a following verb (the "Tobler–Mussafia" pattern). Olivier and colleagues (olivier-2023) study clitic climbing in the history of French (12. Determiners, pronouns and numerals).

4.4 The Chanson de Roland and the Strasbourg Oaths

The Strasbourg Oaths of 842 are the oldest text in a French-like language: Pro Deo amur et pro christian poblo et nostro commun salvament, d'ist di en avant, in quant Deus savir et podir me dunat, si salvarai eo cist meon fradre Karlo. The Chanson de Roland (c. 1100, in laisses of decasyllables with assonance) begins: Carles li reis, nostre emperere magnes, / Set anz tuz pleins ad estet en Espaigne. The first line already shows the subject case (li reis with -s, the article li), the optional article before the name, and the verb ad estet in its Anglo-Norman form. These are public-domain texts; the PROFITEROLE treebank annotates them and is used here only for counts.

5. Middle French in outline

Middle French is a period of transition.

• The case system is lost: roi for both subject and object. Only a few fixed forms keep traces of the old subject case.
• The article becomes obligatory and the definite and indefinite articles complete.
• Negation ne … pas spreads and ne alone goes out of use in the main clause, except in particular constructions (il ne cesse de pleuvoir, je ne saurais).
• The compound tenses (j'ai fait) spread at the expense of the simple past, in speech, and the subjunctive is more used.
• Word order moves to SVO. V2 declines (larrivee-2019; larrivee-2025).
• Latinising vocabulary. Many learned words from Latin and Greek enter, doublets appear (frêle/fragile, hôtel/hôpital), and the spelling is Latinised (doubter, escripre, sçavoir, temps from tens), which accounts for the silent consonants of modern spelling.
• The language of the law. Villers-Cotterêts (1539) made French the language of justice and administration, against Latin (brunot-1906).

The treebank record is meagre for this period. UD Middle French PROFITEROLE has 5 971 sentences and 119 001 words; its README still lists three extracts (the Chroniques of Jean II and Charles V of 1381, Jehan de Paris of 1494 and Book 1 of Commynes's Mémoires of 1497, 12 025 tokens), but the files of 2.18 also hold Le Jouvencel (1461) and a text whose sentence ids begin anglure (the pilgrimage account of the lord of Anglure), so the README undercounts its own data; UD Middle French ALTM has 217 sentences and 7 084 tokens of a Norman legal text of 1425; and the "French" treebank ALTS has two sixteenth-century texts, trial accounts from the Guernsey court (1563–1569) and an extract from Terrien's Norman law book (1578), 2 026 sentences and 68 088 tokens. Short quotations (CC BY-SA, with their ids) show the style of the law:

1425_Style_1_3_1_1_8 · UD Middle French ALTM D' un cas pour ordinaire peult bien faire le sergent en l' extraordinaire l' adjournement, se partie le requiert; of a case for ordinary can well make the sergeant in the extraordinary the summons, if party it requires — "For a case of the ordinary [session], the sergeant may well make the summons in the extraordinary one, if the party asks for it."

1563-1569_Guern_1_1_95_1_4 · UD French ALTS Et aprés il aperceult incontinent estre sain et guery. and after he perceived immediately be-INF healthy and cured — "And afterwards he saw at once that he was healthy and cured." (spelling aperceult, guery; accusative-with-infinitive after apercevoir)

1563-1569_Guern_2_1_1_8_2 · UD French ALTS Et dit que, dempuys qu'ylz partirent de Emontey, ilz ne furent en partie nulle juscque ad ce que ilz vienissent en Hyrlande par temps impetueux. and says that since they-left of Emontey, they not were in place no until that they came-SUBJ in Ireland by weather stormy — "And says that since they left Emontey they were nowhere until they came to Ireland in stormy weather."

The sentences show, in the second, the simple past (partirent, furent), the imperfect subjunctive after jusque … que (vienissent) and the double negation (ne … nulle); in all, a spelling in y for i (dempuys, ylz), double letters, the u of aperceult, and the Latinising spelling ad for à in jusque ad ce que.

6. Classical French (17th–18th century)

6.1 The grammarians and the norm

Classical French is the language that the court, the Académie française and the grammarians fixed in the 17th century. Vaugelas' Remarques sur la langue françoise (1647) is the manual of the good usage, defined by him: "the way of speaking of the soundest part of the Court, in conformity with the way of writing of the soundest part of the authors of the time" (vaugelas-1880, citing the first remark). He adds that the people are the master only of the bad usage, and that the usage of the capital outweighs the provinces. The edition of 1880 that the library holds includes the notes of Thomas Corneille and of the Académie that qualify the rules by the later usage. The Académie's dictionary (the first edition of 1694, paris-1694) is the other pillar.

Vaugelas and his commentators decide points of usage; examples from the text we hold:

• Gascon forms are condemned. Sortez ce cheval or tirez ce cheval for faites sortir ce cheval is "very badly said", though it had become common at Court and in all the provinces; the Gascons are blamed for it, because they turn intransitive verbs into transitive ones (tomber, entrer), and Vaugelas adds that no error establishes itself so easily as making a transitive verb of an intransitive one, since it shortens the expression (vaugelas-1880).
• Relative pronoun. Qui in the oblique cases is reserved for persons. The Académie's note confirms the rule "after much examination" but admits exceptions of usage: it condemns un cheval de qui j'ai reconnu les défauts (because dont is available) and accepts un cheval à qui j'ai fait faire de longues traites and un chien à qui elle fait mille caresses, concluding that usage often allows à qui beyond persons, above all for domestic animals (vaugelas-1880).
• Agreement of the participle with the following infinitive (je les ai fait venir or je les ai faits venir) is discussed at length; modern usage keeps fait invariable before an infinitive (and laissé since 1990: 06 §6).
• The sequence of tenses. After a verb in the past, the subjunctive is in the imperfect (je voulais qu'il vînt), and this literary system is the origin of the modern "imparfait du subjonctif" (13. Verbs).

6.2 The language of the period

The following features distinguish classical French from the modern language. Most are visible in the plays of Corneille, Molière and Racine.

Classical French has a very strong spoken-writing standard: the passé simple is the tense of narration and the passé composé the tense of conversation. The Port-Royal grammar states the rule that the passé simple (its "aorist") is said only of a time at least a day away from the moment of speaking: j'écrivis hier, but not j'écrivis ce matin or cette nuit, for which one must say j'ai écrit (arnauld-1660). Usage never followed it strictly, and the passé simple was already giving way in speech in the following century.

6.3 Verse

Classical verse (Corneille, Racine, Molière, La Fontaine) has fixed rules, many of them linked to the sounds of §2 (Kastner, kastner-1903):

• Syllable counting: the line is defined by its number of syllables. The alexandrine has twelve, divided in the middle by a caesura (6 + 6).
• The mute e counts as a syllable before a consonant (la bonne femme has four syllables in verse: la bon-ne fem-me, with the two e counted), but is elided before a vowel (vie hu-mai-ne: the e of vie does not count) and does not count at the end of the line (the feminine rhyme). Malherbe ruled that voyent, with the ending -ent after a vowel, must not stand inside the line since it is pronounced as a syllable (Malherbe in Kastner, kastner-1903: "'voyent' is pronounced in one syllable, therefore it must not be put in the line").
• Hiatus is forbidden since Malherbe: no word ending in a sounded vowel before a word that begins with a vowel; Kastner lists the few admitted exceptions, among them the hiatus left when the e after a stressed vowel is elided (kastner-1903).
• Alternation of rhymes: a masculine rhyme must not be followed by a different masculine rhyme, nor a feminine by a different feminine. The rule is "an acknowledged principle in French versification since the sixteenth century" (kastner-1903), for which Ronsard and Du Bellay gave the reason that poetry was still thought of as sung, and the mute e of feminine endings was "audible like any other". The rule has been followed by classicists and Romantics alike; the Symbolists were the first to ignore it when it suited them (kastner-1903).
• Rhyme is a rhyme for the eye as much as the ear (the final s and t are counted, the -ent rhymes with -ent only when the sound is the same).
• The diaeresis (tréma) and synaeresis (e.g. li-on in two or one syllable) fix the count of vowels in a word in verse (nation in two syllables, religion in three).

The mute e of verse is the strongest evidence that the e was a sounded vowel in the classical reading; the modern spoken language has lost it, but the reading of verse keeps it (03).

7. The literary tenses

The literary tenses are the passé simple, the passé antérieur (il eut chanté), and the two subjunctives of the past: the imparfait du subjonctif (qu'il chantât) and the plus-que-parfait du subjonctif (qu'il eût chanté). In spoken French they have been replaced by the passé composé and the présent du subjonctif; in writing they are alive, in narrative prose, in formal letters, and in some standard expressions.

7.1 The passé simple: forms

Four types, according to the vowel of the endings: -a- (all verbs in -er and aller), -i- (verbs in -ir of the second group, verbs in -re like vendre, dire (dis), prendre (pris), faire (fis), mettre (mis)), -u- (avoir, être, vouloir (voulus), pouvoir (pus), savoir (sus), devoir, lire (lus), boire (bus), vivre (vécus), croire (crus), connaître (connus)) and -in- (venir, tenir and their compounds). The accent circonflexe of the first and second plural, and of the third singular of the imperfect subjunctive (qu'il fût), is kept by the 1990 rectifications (05; cslf-1990, rule 4). The -ai of the first singular is pronounced /e/ and is homophonous with the infinitive and the participle (chantai, chanter, chanté): 05 §6.

7.2 The passé simple: use

It expresses a completed event in the past that is cut off from the present, in a narrative, and advances the story: Il entra, regarda autour de lui et sortit. The imperfect gives the background (Il pleuvait), and the pair imperfect + passé simple is the frame of the novel. In the third person it is common, in the first and second persons very rare, and in the plural of the first and second persons nearly extinct outside of fiction in the first person.

It has gone out of the spoken language of France (the Acadian French of 08 §6.2 is described as keeping a simple past in speech). It has gone from journalism in the 20th century, but it is alive in:

• fiction (the récit of the novel, in the third person), in historical writing and in biography;
• the language of the encyclopaedia for the dates of the dead (Napoléon mourut en 1821), and the "historic present" competes with it;
• children's tales and the translations of the tales of other languages;

Counts (UD 2.18, finite indicative forms marked Tense=Past, 2026-11-06). The annotation of GSD and Sequoia treats the passé simple as Tense=Past|Mood=Ind, and the imperfect as Tense=Imp.

In the written treebank 6.4% of the finite verbs are in the passé simple (the most frequent forms in GSD are fut 415, furent 84 and fit 40; GSD includes Wikipedia articles), in the spoken corpora 0 of 4 472 and 2 of 4 074. The imperfect subjunctive occurs 10 times in GSD, once in PUD. The passé simple is therefore the tense of a register (written narrative), not of a period.

7.3 The imperfect and pluperfect subjunctive

The imperfect subjunctive is built on the second singular of the passé simple: chanter → chantas → que je chantasse, que tu chantasses, qu'il chantât, que nous chantassions, que vous chantassiez, qu'ils chantassent; finir → finisse, finît; venir → vinsse, vînt; être → fusse, fût; avoir → eusse, eût. The pluperfect subjunctive is the auxiliary in the imperfect subjunctive plus a participle (qu'il eût chanté, qu'elle fût venue). The third singular in -ât, -ît, -ût is the only form that survives in modern literary writing, in a fixed use: the hypothetical ("eût-il voulu, il n'aurait pu" "even had he wished"), the concessive subjunctive (fût-ce "be it", dût-il "even if he had to") and the sequence of tenses after a main verb in the past ("Il fallait qu'il vînt"). The pluperfect subjunctive in -ût stands in as a literary conditional: Il eût fallu for Il aurait fallu. In modern writing the sequence of tenses is dropped; a main verb in the past is followed by the present subjunctive (il fallait qu'il vienne), as in speech.

8. Reading earlier French: what changes for the reader

• Spelling: old consonants (ay, oy, y for i, -ct-, -pt-, -ss-), accents missing (the Académie's printing of é and è is later), u/v and i/j not distinguished before the 17th century, long s ſ in print before about 1790.
• Words: archaic words with no modern successor (ainçois "rather", moult "much", ains, or "now", si as a particle of continuity); words that survive with another sense (gent "pleasant", ire "wrath", chef "head").
• Constructions with no modern equivalent: the V2 order, the object pronoun before the finite verb group, ne alone, the infinitif de narration (Et grenouilles de sauter), the absolute participle construction.
• Gender, agreement of participles and articles are not yet fixed (Vaugelas discusses them case by case).

9. The data

The annotation of PROFITEROLE uses its own tag set in the XPOS column (VERcjg for a conjugated verb, NOMcom, PROper, PONfbl, …). The treebanks use the UD relations with some conventions of their own for Old French (the Profiterole guidelines). The lemmas are modern French where possible, or those of the Dictionnaire du Moyen Français.

Resources for processing earlier French outside Mova: a corpus and neural models for the lemmatisation and part-of-speech tagging of Old French, with a global accuracy of about 97.5% for both lemma and part of speech (camps-2021; the companion paper on classical theatre, camps-2020, is CC BY-SA and is only mentioned); a morphological and syntactic lexicon of Old French built from dictionaries (sagot-2019); a symbolic grammar of French (FRMG) adapted to Old and Middle French in order to extend the SRCMF treebank with more texts (regnault-2019); a note that Old and Middle French forms had been wrongly included in the French UniMorph data and made the French inflection task look strangely hard (kezerian-2024), which is a warning that a lexicon built from "French" sources must be checked for period. The comparative study of Grobol and colleagues (grobol-2021) trains the same parser on the Old French SRCMF treebank and on GSD; its conclusion is that parsing Old French is indeed harder than parsing contemporary French, at least with the amount and heterogeneity of the treebanks available, in the word-level metrics and in the exact match of the main constituents; the qualitative analysis agrees; errors are not spread evenly over word orders, and while the least common orders have the most errors, the most common orders are not necessarily the best handled (provisional conclusions, in the authors' words).

10. How Mova handles earlier French

What Mova has. Nothing specific. There is no Old, Middle or Classical French lexicon, no normaliser, no list of the literary tenses and no treatment of the two-case system. The French parser is the English en parser, with the tags of a treebank mapped (en xpos-fill).

What happens when it meets earlier French: measured (2026-11-06). The measurements below use en xpos-fill followed by en parse-eval (the commands of PACKAGE.md); each of the four runs was made twice and gave the same numbers both times.

Three results follow. (1) A parser trained on in-domain earlier French does well (LAS 82.01 with gold tags and 76.79 with its own tags on Old French, UAS 86.46 and 82.83): the English parser can learn Old French syntax from 190 000 tokens, and the gap to modern GSD (LAS 87.00 own tags, 89.78 gold tags, 2026-10-22) is the order of 10 points with own tags and 8 with gold tags. The numbers are not comparable with those of Grobol's parser (a different system and split), but the direction agrees with their conclusion that Old French is harder. (2) A parser trained on modern French and run on Old French loses about 20 points with gold tags (62.01 against 82.01), and its own tagger collapses (own tags: LAS 30.10): the tagger meets forms it has never seen, with different spellings, and assigns wrong tags, which sends the parser far off. (3) Middle French is closer: the modern parser reaches 66.05 with gold tags and 47.84 with its own, which is 17 and 32 points below the in-domain parser.

These numbers measure only what the parser does with the tokens as given; they do not include tokenization of raw text, where the English tokenizer already loses about one token in nine on modern French (07. Punctuation, typography, capitals, abbreviations and numbers §12) and where earlier French adds clitics written with an apostrophe or a space (l' autre, as in the ALTM texts, which have a space after the apostrophe) and elisions that follow other rules.

What would be needed. The parse of earlier French needs, in order: (1) a normaliser from the manuscript or print spelling to a regular spelling (§3, with the neural models of gabay-2020, rubino-2024, solfrini-2025; a rule list is not enough because spelling varies by scribe); (2) a lexicon with lemmas and forms of the period (sagot-2019; with care about periods: kezerian-2024); (3) training on the period's treebanks, which Mova can do for each of the periods in §9 (and which the in-domain row of the table shows works at LAS 82–83 with gold tags); (4) a tagger on the period's tag set (camps-2021 reports 97.5% for Old French); (5) the literary tenses for modern French and Classical French, which a modern-French lexicon needs in any case: the forms of the passé simple and the imperfect subjunctive are regular and can be generated from the paradigm of §7, and the lemmatiser (when there is one) should map fut, vint, eût to être, venir, avoir.

What is not planned. Prosodic reading of verse (syllable counting with the mute e, §6.3) needs a French phonology that Mova does not have (03); the treatment of Old French dialects (Anglo-Norman, Picard, …) as separate languages, as the PROFITEROLE projects label the texts; the rendering of earlier French in modern French (a machine-translation task).

Sources for this chapter

• Nyrop, Kristoffer. 1899–1930. Grammaire historique de la langue française, vols. I–VI. Copenhague: Gyldendal. Public domain (nyrop-1899). Library: books/n/ny/nyrop-1899-grammaire-historique-de-la.
• Brunot, Ferdinand. 1905–1943. Histoire de la langue française des origines à 1900. Paris: A. Colin. Public domain for the volumes held (brunot-1906). Library: books/b/br/brunot-1906-histoire-de-la-langue.
• Vaugelas, Claude Favre de. 1647; ed. A. Chassang 1880. Remarques sur la langue françoise. Public domain (vaugelas-1880). Library: books/v/va/vaugelas-1880-remarques-sur-la-langue.
• Arnauld, Antoine & Claude Lancelot. 1660. Grammaire générale et raisonnée (Port-Royal). Public domain (arnauld-1660). The rule of the aorist.
• Kastner, Leon E. 1903. A History of French Versification. Oxford: Clarendon. Public domain (kastner-1903). Library: books/k/ka/kastner-1903-history-french-versification.
• Le dictionnaire de l'Académie françoise, 1st ed., 1694. Public domain (paris-1694). Library: books/p/pa/paris-1694-le-dictionnaire-de-lacademie.
• Dauzat, Albert. 1908. La langue française d'aujourd'hui. 1922. La géographie linguistique. Public domain (dauzat-1908, dauzat-1922).
• Conseil supérieur de la langue française. 1990. Les rectifications de l'orthographe (cslf-1990).
• Grobol, Loïc, Sophie Prévost & Benoît Crabbé. 2021. Is Old French tougher to parse? TLT 2021. CC BY 4.0 (grobol-2021-is-old-french-tougher-parse).
• Camps, Jean-Baptiste et al. 2021. Corpus and models for lemmatisation and POS-tagging of Old French. arXiv:2109.11442. CC BY 4.0 (camps-2021-corpus-models-lemmatisation-pos-tagging). The classical-theatre companion (camps-2020-corpus-models-lemmatisation-pos-tagging) is CC BY-SA and is only mentioned.
• Rubino, Raphael et al. 2024. Normalizing without modernizing: keeping historical wordforms of Middle French while reducing spelling variants. Findings of NAACL 2024. CC BY 4.0 (rubino-2024-normalizing-without-modernizing-keeping-historical).
• Solfrini, Sonia et al. 2025. Normaliser le moyen français: du graphématique au semi-diplomatique. RECITAL 2025. CC BY 4.0 (solfrini-2025-normaliser-le-moyen-francais-du).
• Gabay, Simon et al. 2020. Traduction automatique pour la normalisation du français du XVIIe siècle. CC BY 4.0 (gabay-2020-traduction-automatique-pour-la-normalisation).
• Larrivée, Pierre. 2019. To be or not to be informational: preverbal complements in Medieval French V2 configurations. Glossa 4(1). CC BY 4.0 (larrivee-2019-be-or-not-be-informational). Larrivée, Pierre et al. 2025. The correlation of micro-cues: the case of medieval French verb second. Glossa 10(1). CC BY 4.0 (larrivee-2025-correlation-micro-cues-case-medieval).
• Wolfe, Sam. 2020. Reconsidering variation and change in the Medieval French subject system. Glossa 5(1). CC BY 4.0 (wolfe-2020-reconsidering-variation-change-medieval-french).
• Regnault, Mathilde et al. 2019. Challenges of language change and variation: towards an extended treebank of Medieval French. CC BY 4.0 (regnault-2019-challenges-language-change-variation-extended).
• Sagot, Benoît et al. 2019. Développement d'un lexique morphologique et syntaxique de l'ancien français. CC BY 4.0 (sagot-2019-developpement-d-un-lexique-morphologique).
• Kezerian, Dayna et al. 2024. Ye olde French: effect of Old and Middle French on SIGMORPHON-UniMorph shared task data. CC BY 4.0 (kezerian-2024-ye-olde-french-effect-old).
• Olivier, Marc, Christina Sevdali & Raffaella Folli. 2023. Clitic climbing and restructuring in the history of French. Glossa. CC BY 4.0 (olivier-2023).
• UD Old French PROFITEROLE, Old French ALTM, Middle French PROFITEROLE, Middle French ALTM, French ALTS, French GSD, Sequoia, PUD, FQB, ParisStories, Rhapsodie, 2.18 (counts of 2026-11-06; licences in Sources).
• Mova: en xpos-fill, en parse-eval, en/src/main.rs; the runs of 2026-11-06 described in §10.

This page in Markdown → https://md.mova.elamur.ai/free/en/docs/languages/fr/09-old-middle-classical-french/