🗣️ 36. Discourse relations and coherence

A French text holds together when its parts are related: one sentence gives the cause of another (car, parce que, puisque), contrasts with it (mais, alors que, en revanche), concedes something against it (pourtant, bien que), follows it in time (puis, ensuite), or rephrases and sums it up (c'est-à-dire, bref, enfin). This chapter describes how French signals these relations: its connectives and their syntax in the UD trees, the order of clauses, the narrative tenses, and the questions a text answers. It also covers the frameworks and resources that annotate French discourse: SDRT in ANNODIS, a PDTB-style French Discourse Treebank, the LEXCONN lexicon of connectives, RST in translation, and DiscoGeM for implicit relations. It ends with what Mova's frame reader and QUD stack do with discourse structure: for English, a little; for French, nothing yet. Every count is on the UD 2.18 French treebanks and was made on 2026-11-06.

Contents

1. Cohesion and coherence
2. Coherence relations: the families
3. The frameworks and the French resources
4. LEXCONN: the relations of French connectives
5. French connectives and their syntax
6. Cause: car, parce que, puisque, comme
7. Contrast and concession
8. Time and narrative
9. Implicit relations
10. Questions under discussion
11. Segmentation and parsing for French
12. How Mova handles it

1. Cohesion and coherence

Cohesion (cohésion) is the set of surface ties between sentences: reference by pronouns and definite noun phrases (il, celle-ci, ce dernier, le directeur; chapter 35), ellipsis, connectives, and lexical repetition. Coherence (cohérence) is the relation a reader infers between the meanings. The two come apart. A text can be cohesive and still make no sense, and a text can be coherent with no tie at all: Ne sors pas. Il pleut. "Don't go out. It's raining." is read as a reason without any car. The English chapter gives the classic inventory, Halliday & Hasan's Cohesion in English (1976; not open). French linguistics follows the same split, usually in the terms of the French discourse tradition (the linguistique textuelle of the 1970s–1990s; not open, Wanted).

French expresses relations with a large closed class of connectives, and most of them are multiword. Brunot (1899, Précis de grammaire historique, brunot-1899, ch. on invariable words) notes that French kept few of the Latin particles. It rebuilt its adverbs, prepositions and conjunctions by putting words side by side: à cause de, désormais, toujours. The connectives show the same history (parce que, bien que, alors que, en effet, par conséquent, c'est-à-dire). Clédat (1894, Grammaire raisonnée, cledat-1894) gives the Latin source of car, quare "for which reason", a source it shares with no other French conjunction. The multiword form matters for parsing: a UD tree must decide whether en effet is one fixed word or a preposition with its noun (§5).

2. Coherence relations: the families

Every framework cuts the relations differently, but the same five families recur (Hobbs, Kehler, RST, PDTB and SDRT; the English chapter 36 §2 has the references):

The examples are from GSD (§5, §7). The last family holds the relations that link speech acts rather than facts. They are the "metatalk" relations of SDRT and LEXCONN (§4), and they are where puisque and de toute façon do most of their work.

3. The frameworks and the French resources

The three frameworks of the English chapter, RST, PDTB and SDRT, all have French applications. A fourth line of work starts from the connectives themselves.

SDRT: ANNODIS. SDRT (Asher & Lascarides 2003; not open) builds a graph of discourse units joined by relations that are either coordinating (Narration, Continuation, Contrast) or subordinating (Elaboration, Explanation, Background). New material may attach only on the right frontier. The main French corpus is ANNODIS (Afantenos et al. 2012, afantenos-2012-empirical-resource-discovering-cognitive-principles; Péry-Woodley et al. 2009, 2011, woodley-2009-annodis-une-approche-outillee-de, woodley-2011-le-corpus-annodis-un-corpus). It annotates texts of several genres in two ways: as a bottom-up SDRT structure over elementary discourse units (EDUs), and as a selective top-down annotation of larger structures such as enumerations. EXPLICADIS (Atallah 2015, atallah-2015-la-ressource-explicadis-un-corpus) adds a finer annotation of causal relations to ANNODIS-style texts.

PDTB: the French Discourse Treebank. The FDTB (Danlos et al. 2012, danlos-2012-vers-le-fdtb-french-discourse; Danlos, Colinet & Steinlin 2015, danlos-2015-fdtb1-premiere-etape-projet-french) applies the PDTB method to the French Treebank, a corpus of Le Monde. Its first stage, FDTB1, identified more than ten thousand connective occurrences and told them apart from non-discourse uses of the same words. The hard cases were ainsi as a manner adverb and prepositions that head complements: of 1 161 occurrences of pour + infinitive, 518 (44%) were purpose connectives.

RST in translation. No open native French RST treebank is in the research library. Pastor, Marino & Oostdijk (2024, pastor-2024-la-reconnaissance-automatique-des-relations) translated the English RST-DT test set (38 documents, 2 308 EDUs) into French one EDU at a time and ran the multilingual DMRST parser on it. Relation classification scored F1 0.54, against 0.65 on the English original, which is near the published scores for Spanish (0.63) and Portuguese (0.62). In their error analysis si was a reliable cue (73% of 41 conditions recognised), while et occurred in Joint, Temporal and other relations and misled the parser.

Lexicons and classifiers of connectives. LEXCONN (Roze, Danlos & Muller 2012, roze-2012-lexconn-french-lexicon-discourse-connectives) lists French connectives with their syntactic category and SDRT relations (§4). Laali & Kosseim (2017, laali-2017-automatic-mapping-french-discourse-connectives) mapped the 371 connectives of LEXCONN v2.1 (on average 1.3 relations each, 37 still unassigned) to PDTB senses. They parsed the English side of Europarl with a PDTB parser and carried the senses across the word alignment. The result, ConcoLeDisCo, holds 900 (connective, relation, probability) triples. Only Concession and Condition are defined alike in both inventories; on those, it recovered 36 of 44 LEXCONN connectives (81%). Rauh, Zaczynska & Bourgonje (2023, rauh-2023-multilingual-connective-database-aligning-german) aligned the French and German concessive connectives in the same way. Relex (Sow, Toussaint & Constant 2026, sow-2026-relex-common-denominator-connectives-discourse) marked connectives in 35 million tokens of French press and online debate with rule-based automata, which agreed with the FDTB at F1 0.76 once ambiguous candidates were set aside. It labelled each monosemous connective with its single LEXCONN relation and fine-tuned CamemBERT to predict 19 relations. On connectives unseen in training it reached F1 0.59. In its hand-checked test set, 19 of the 42 cases sent to the expert expressed two relations at once.

Implicit relations, causality, attribution. DiscoGeM 2.0 (Yung et al. 2024, yung-2024-discogem-2-0-parallel-corpus) has 12 834 implicit relations in English, German, French and Czech, each labelled by ten crowd workers who inserted a connective, so each relation is a distribution of readings (§9). FreCaDiS (Blochowiak et al. 2026, blochowiak-2026-argumentation-through-discourse-relations-subjectivity) annotates parce que and car in five genres, from SMS to fiction, as explanation or justification and as subjective or objective (§6). FRACAS (Richard et al. 2024, richard-2024-fracas-french-annotated-corpus-attribution) annotates quotations and their sources in 1 676 news texts: the attribution relation (chapter 38).

Papers under NC or ND licences inform this chapter but are not quoted. The open (CC BY) papers are paraphrased with their slugs.

4. LEXCONN: the relations of French connectives

LEXCONN 2012 lists 328 connectives with 428 usages: a connective with two relations has two usages (Roze, Danlos & Muller 2012). The tests for a connective include that it links two discourse segments, that its meaning is not compositional and that it contains no referential expression. The authors found each connective's relations by reading contexts in the Frantext corpus. They also used a substitution test: dès lors que, puisque and étant donné que can replace one another without changing the reading, so they share a usage.

The relations are fifteen of SDRT plus six the authors added because no SDRT relation fitted:

By share of usages, the largest classes are Opposition (9.5%), Result (8.1%), Concession and Continuation (7.4% each), Explanation (6.5%), Goal and Condition (5.8% each), Explanation* (5.6%) and Narration (5.4%). French thus has many connectives for opposing and conceding and for causes, and comparatively few for plain narration, where the order of telling does the work. The added Concession relation answers a real problem. ANNODIS had treated bien que and même si as markers of the coordinating Contrast. LEXCONN argues from the right frontier that concession is subordinating: a segment after the concessive can still attach to the segment before it.

The starred relations are where French grammar diverges most from a fact-only reading. In Il est malade, puisqu'il ne vient pas "He's ill, since he isn't coming", the absence is not the cause of the illness. It is the speaker's evidence for saying so (Explanation*). In Tu as fini ? Donc on peut partir. "Have you finished? So we can leave", donc introduces a conclusion about what to do, not a result of a fact (Result*). A reader that turns every puisque into a cause link gets these wrong (§12).

5. French connectives and their syntax

As in English, discourse connectives are a functional class, not a part of speech. In UD French they are coordinators, subordinators, adverbs, prepositional phrases and fixed multiword expressions. Their relation in the tree follows their syntax, not their discourse function.

The tree shows the difference between the two causal connectives at once (UD French GSD 2.18):

fr-ud-train_02191   Il purifie tout car il supprime les odeurs nauséabondes.
  purifie ─conj→ supprime ─cc→ car            (car: CCONJ, the cause is a conjunct)

fr-ud-train_03202   Le crédit était mort parce que les banques n'avaient plus de liquidités.
  mort ─advcl→ avaient ─mark→ parce ─fixed→ que   (parce que: the cause is an adverbial clause)

With car the cause is a coordinated main clause; with parce que it is a subordinate clause. This is the grammars' classic contrast: car cannot be clefted (*c'est car…) or answer pourquoi ?, while parce que can. A reader that looks for causes only under advcl misses every car.

How the connectives are annotated. All treebanks, all splits, one count per head word (the fixed parts of a multiword expression are folded into it):

Puis is a coordinator in UD French (cc 307 of 357), not an adverb. Donc, enfin and en fait are advmod in writing and often discourse in the spoken treebanks, which reflects the annotation guidelines as much as the language. The Mova dialect makes an adverb advmod, never discourse (dialects/mova/seeds.md#discourse-adverbs; chapter 38). Or is a trap: only 19 of its 90 occurrences are the connective; most of the rest are the noun or "gold".

The treebanks disagree on multiword connectives. GSD makes en effet one word (effet fixed, 102 times), while Sequoia analyses it as a preposition and a noun (effet obl:mod). The spoken treebanks do the same with par exemple. So a count of fixed en effet finds 102 in GSD and none in Sequoia, a difference of annotation, not of usage. Counting connectives across treebanks needs expressions matched on the words.

Frequency, written and spoken. Per 1 000 syntactic words, punctuation included, with the relation filters of the table above (car only as cc, si only as mark, après only as an adverb). Written is GSD, Sequoia, PUD and ParTUT (524 236 words). Spoken is ParisStories and Rhapsodie (87 023 words):

The spoken sample is small (87 000 words), so its rare items rest on single digits. Three points are still clear. Et is equally frequent in both, but speech builds its relations from a small set: mais, donc, parce que, alors, après, enfin, en fait. The concessive and oppositive adverbs of writing (cependant, toutefois, néanmoins, ainsi, bien que, tandis que) are absent from the spoken corpora, which use mais and quand même for concession and quand for time. And car and parce que swap places (§6).

At the start of the sentence. The share of sentences whose first word, punctuation aside, is a connective:

A quarter of the sentences of ParisStories, a corpus of spoken narratives, begin with et: the transcription cuts speech into short units, and et, donc and alors open each step of the story (§8). Written French rarely starts a sentence with a connective, and then mostly with mais (181 in GSD), cependant (71) or ainsi (64).

6. Cause: car, parce que, puisque, comme

French has four everyday causal connectives, and the grammars traditionally divide the labour among them as follows:

Car and parce que. The usual claim is that car gives a subjective, speaker-oriented reason and belongs to a high register, while parce que serves both objective and subjective causes. Blochowiak, Grisot & Degand (2020, blochowiak-2020-what-type-subjectivity-lies-behind, Glossa, CC BY 4.0) tested this on journalistic texts and a one-million-word SMS corpus, with annotation tasks. Their results do not support it. In SMS, car is about as frequent as parce que (1.09 against 1.52 per thousand words), so it is not confined to formal writing. Readers more often took car as presenting the information objectively, not subjectively. In SMS, car introduced events more often and parce que opinions more often. The authors propose two kinds of subjectivity: the type of causal relation (a cause in the world, or a reason for a belief or an act) and an evaluative, expressive subjectivity. Parce que carries the second more than car.

The treebanks show the register split more sharply. In writing car outnumbers parce que two to one (169 to 83; GSD alone 130 to 40). In the two spoken treebanks parce que outnumbers car twenty to one (274 to 13), and ParisStories has only 3 car. SMS, with its near-equal ratio, sits between speech and print. In GSD a car clause often justifies an evaluation or a piece of advice rather than giving the cause of a fact:

(1) GSD fr-ud-train_09346 Dommage pour eux car les plats étaient bon. "Too bad for them, because the dishes were good" (the writer's bon). The car clause justifies Dommage; it causes nothing.

(2) GSD fr-ud-test_00237 Il vaut mieux réserver car ce petit coin de Naples est bien connu. "Better book, because this little corner of Naples is well known": a reason for advice (a speech act).

In speech, parce que often stands alone. It opens a new transcription unit that explains, after the fact, what was just said:

(3) ParisStories ParisStories_2019_devoirPhilosophie_3–_4 donc je me suis dit, comment je vais faire pour rendre ce devoir ? / parce que euh, je suis pas très douée en philosophie à la base. "so I said to myself, how am I going to hand in this assignment? / because uh, I'm not very good at philosophy to start with." This parce que justifies the worry, not the assignment. It behaves like car, which spoken French barely uses.

Sentence-initial parce que opens 40 sentences in ParisStories and 39 in Rhapsodie, against 5 in GSD. Such a fragment must attach to the previous sentence or turn, because its own tree has no main clause (§10, §12).

Puisque. Puisque presents its clause as already accepted, so the relation it marks is often between the clause and the speaker's right to say the other one:

(4) GSD fr-ud-dev_00667 Choix payant puisque le milieu de terrain inscrit le 3e but de son club dix secondes après être entré en jeu ! "A choice that paid off, since the midfielder scored his club's third goal ten seconds after coming on!" The goal is the evidence for the evaluation payant.

(5) Rhapsodie Rhap_D2004-26 alors, en fait, vous êtes, euh, liée à la France, puisque vous êtes française. "so, actually, you are, uh, linked to France, since you are French." The speaker states the obvious and uses puisque to mark it as obvious.

LEXCONN counts puisque among the connectives of Explanation* (§4). Unlike car, it is not rare in speech (60 written, 33 spoken).

Result: donc, alors, ainsi, par conséquent. Donc is the commonest result connective, but in speech it also resumes a narrative (donc euh j'ai fait la conduite accompagnée au lycée, ParisStories ParisStories_2019_peripitiesVoiture_2) or brings back a question. In writing it is often a mid-clause adverb of inference: C'est un affluent de l'Aube en rive gauche, donc un sous-affluent de la Seine. (GSD fr-ud-train_07771). Par conséquent (10 in GSD, none in speech) belongs to writing. Ainsi is either result ("thus") or manner ("in this way"), and the FDTB had to separate the two (§3).

7. Contrast and concession

Mais is the second most frequent connective and covers both contrast (X, but Y differs) and concession (X, but Y despite X). The argumentative tradition of French linguistics (Anscombre & Ducrot; not open, Wanted) analyses p mais q this way: p argues for a conclusion r, q argues for not-r, and q is stronger. The analysis explains why mais can link sentences that share no topic, as in Il fait beau, mais j'ai du travail "It's fine, but I have work", where the conclusion is "let's go out" or not. Mais is cc in 1 518 of its 1 564 occurrences. When it opens a sentence, it is still cc of the root of that sentence, so in UD it attaches inside the sentence and does not link the two sentences. The link to the previous sentence is not in the tree.

Or introduces a new premise that turns the argument. It is the "minor premise" connective of the syllogism, typical of argument and legal prose:

(6) GSD fr-ud-dev_01069 Or, la mère de Leslie était invalide depuis sa naissance. "Now, Leslie's mother had been an invalid since birth." Or is cc of the root invalide, again inside its own sentence.

Or has no single English equivalent (now, yet, however each capture part of it), and it is rare: 17 written, 2 spoken.

The concessive and oppositive adverbs cependant, pourtant, toutefois, néanmoins and en revanche / par contre are written French: together about 0.7 per 1 000 written words, and almost none in the spoken corpora (§5). LEXCONN separates Opposition (cependant, malgré tout: the second segment goes against an expectation) from Concession (même si, bien que: the first segment is granted and does not prevent the second). En revanche and par contre mark plain contrast between two parallel facts: Si ses publications ont été peu nombreuses, la portée de son enseignement, en revanche, a été considérable. (GSD fr-ud-train_03288). Par contre was long criticised by purists, but it is in common use: GSD has 21 par contre against 27 en revanche, the former in reviews such as Par contre l'accueil est impeccable… (GSD fr-ud-test_00273).

Subordinating contrast and concession. Alors que and tandis que are temporal in origin ("while") and now mostly contrastive. In GSD, alors que 123 and tandis que 63 are mark. Many of these clauses contrast two parallel facts, as in 84,0% des logements sont possédés alors que 16,0% sont loués. (GSD fr-ud-train_08136). Others keep the temporal frame, as in Mais un jour, alors qu'il survole l'Amazonie, leur avion s'écrase dans la forêt. (GSD fr-ud-train_01965). Relex's authors give tandis que as their example of a connective whose relation depends on context, Contrast or Background (§3). The concessive bien que takes the subjunctive and même si the indicative (chapter 13, chapter 24). Bien que occurs 85 times in writing and never in the spoken treebanks, where même si and mais take its place.

Quand même "all the same, still" is the concessive adverb of speech: 89 occurrences in speech against 11 in writing, the strongest spoken skew in the table of §5. It also works as an intensifier (c'est quand même bizarre "it's really weird though"), so not every quand même marks a relation (chapter 38).

8. Time and narrative

8.1 Temporal connectives

Puis and ensuite mark the next event; alors marks the next event or the consequence; après is a preposition, a subordinator (après que) or, in speech, an adverb "afterwards; then again". Bras, Le Draoulec & Vieu (2001, bras-2001-temporal-information-discourse-relations-narratives, NC: read only) compared puis with un peu plus tard "a bit later" in SDRT. They argue that puis marks Narration proper and blocks a Result reading. Un peu plus tard blocks Narration and allows only a weaker, purely temporal succession, and it is compatible with Result. Two connectives with nearly the same temporal content thus differ in the discourse relations they allow.

In written narration puis chains events and also items in a description:

(7) GSD fr-ud-train_14479 Les Drangiens furent soumis par les Mèdes, puis par Cyrus le Grand. "The Drangians were subjugated by the Medes, then by Cyrus the Great." Here puis is cc on the second agent, Cyrus, a conj of Mèdes: the succession is between two events that share one verb.

(8) GSD fr-ud-train_00686 Arbuste, parfois densément ramifié, avec des rameaux d'abord laineux puis glabres. "A shrub, sometimes densely branched, with branches woolly at first, then hairless." This is the succession of two states in a botanical description, not a narrative sequence.

Après in speech. Herreño Castañeda & Sillaire (2025, castaneda-2025-annotation-de-marqueurs-discursifs-le) studied après in the CODIM corpus, about 110 million tokens of spoken, written and online French. A rule-based pass marked 848 occurrences as discourse marker and 75 151 as not, and left 70 059 undecided, more than half of them in the oral and online data. Linking discourse units, après can be additive (listing), consecutive, contrastive ("that said, mind you") or meta-discursive. To find the contrastive use, their protocol first excludes fixed phrases (après-guerre) and après next to a duration, an event noun or a place. It then looks for neighbouring markers (bon après, mais après) and finally asks whether the two situations are opposed. With it, agreement rose from κ 0.54 to 0.69. In the press, contrastive après hardly occurred. The treebanks agree: ParisStories has 82 adverbial après against 9 prepositional ones, GSD 40 against 317.

8.2 Tense and discourse

French narrative tenses carry discourse relations of their own. In the classic analysis of written narrative (Kamp & Rohrer 1983 on French tenses in DRT; Weinrich, Le Temps; both not open, Wanted), the passé simple moves the story forward: each new event follows the last (Narration). The imparfait sets a background state or an ongoing frame around the current event (Background). The plus-que-parfait moves back (Flashback). In speech and in most modern writing the passé composé takes over the role of the passé simple, and the narrative present is common.

Finite indicative verbs of the narrative tenses (lexical verbs and copulas, not auxiliaries). The passé composé is counted as an auxiliary in the present with a participle:

ParisStories is a corpus of personal narratives, and it has no passé simple at all. Its stories run on the passé composé for events, the imparfait for background (more imparfaits than passé composés), and connectives for the steps: et (26% of sentence starts), donc (7%), alors, après, et puis. The written passé simple of GSD sits in encyclopaedic history (Les Drangiens furent soumis…, ex. 7). So a reader that infers Narration from the passé simple has a cue in written history and fiction (chapter 13) but none in speech, where the cue is the connective and the passé composé.

Narrative order by connectives alone also misleads. Alors after an imparfait frame marks the event that breaks in (il lisait; alors on a sonné). After an event it marks the consequence. Alors que with an imparfait or present keeps a simultaneous background (ex. above, alors qu'il survole l'Amazonie). These are the interactions of tense, aspect and connective that Grisot (2018; see the English chapter §8) measures for English and French.

9. Implicit relations

Most relations in a narrative, and many in argument, have no connective. The FDTB authors estimated, from the share of explicit relations found in other languages (about half), that the ANNODIS texts contain some 1 600 connectives (Danlos et al. 2015). DiscoGeM 2.0 gives the cross-language picture. When crowd workers inserted connectives into implicit relations of parallel texts, the French readings did not always match the English or German ones. The translation sometimes made explicit, or changed, the relation the original left implicit (Yung et al. 2024). Translation into or out of French therefore cannot carry implicit relations across unchanged.

The last two rows are French-specific constructions whose relation is underspecified. En travaillant can be the means, the cause, or a simultaneous action (chapter 13). A detached participle (La production étant déficitaire, on importa alors des olives d'Algérie, GSD fr-ud-train_05152) is usually a cause or a background.

Juxtaposition in the trees. UD marks two juxtaposed clauses in one sentence as parataxis:

The spoken treebanks have fewer paratactic links than the written ones, though speech is more paratactic. Their transcriptions cut juxtaposed clauses into separate sentences, so the juxtaposition falls between trees, where UD records nothing. Most of what remains in speech is parenthetical (:parenth). ParTUT's near-zero is a convention of its conversion. Read across treebanks, parataxis measures annotation practice as much as the language.

10. Questions under discussion

The QUD view (Roberts 1996/2012; English chapter §9) treats a discourse as answers to a stack of open questions. Speakers often pose the question themselves: the Pourquoi ? Parce que… of argument, or the opening of an oral recipe, donc comment je fais les gougères ? (ParisStories ParisStories_2019_cuisineApproximative_1), which the rest of the story answers. Cleft and pseudo-cleft sentences (Ce qui m'étonne, c'est que…, C'est là que…) state the question in the sentence itself. Their information structure is in chapter 29, and their presupposition in chapter 33.

The French answer fragment Parce que… is the common case of an answer without a main clause (§6). A QUD reader must accept a parce que fragment as an answer to pourquoi, and a pour + infinitive fragment as an answer to pour quoi faire ? The English QUD module does not yet do even the English equivalent (§12). The FQB treebank, 2 289 questions with 172 quand and 19 pourquoi among their wh-words, is a source of French questions but has no answers.

11. Segmentation and parsing for French

Segmenting conversation. Hunter, Prévot & Muller (2023, prevot-2023-comparing-methods-segmenting-elementary-discourse) segmented the Corpus of Interactional Data (CID), conversational French, in two ways: by fine-tuning an existing segmenter on the manual EDUs, and by data programming, where hand-written labelling rules train a weakly supervised model. The two reached similar F-scores, and the rules needed far less annotation. Conversation is made of short contributions and one-word non-sentential units (answers, feedback), which segmenters trained on written text handle badly. Prévot, Bertrand & Hunter (2025, prevot-2025-segmenting-large-french-meeting-corpus) added 20 hours of manually segmented French meetings among three or four people, and 80 hours segmented automatically. A model fine-tuned on it scored F ≈ 0.86, still below written texts.

Weak supervision for relations. Maachou, Braud & Muller (2025, maachou-2025-supervision-faible-pour-la-classification) combine heuristics and pretrained models as weak labellers for relation classification across mismatched label sets, and come close to full supervision with a small part of the data. Badene et al. (2019, badene-2019-apprentissage-faiblement-supervise-de-la) use data programming for SDRT attachment. Both work on English corpora (RST-DT and PDTB; the chat corpus STAC). Their methods suit French, where gold data is scarce, but neither reports a French result.

What the papers say for Mova. Relex and ConcoLeDisCo show that the relation of a connective can be learned from cheap signals: monosemous connectives, or alignment with English. Pastor et al. show that si is a reliable cue and et is not. The après study shows that a marker's discourse use can be ambiguous even for expert annotators (κ 0.69 at best), and that genre decides much of the answer. A French reader should start from the connectives whose relation is stable: si, parce que, bien que, pour que, puis. It should treat et, donc, alors, après and enfin as uncertain.

12. How Mova handles it

Checked against the code on 2026-11-06.

12.1 What exists, for English

Mova keeps discourse relations as links between frames, not as a separate tree. The frame reader (prag/src/frames.rs, set_cause and set_next) sets them from the UD relation of a clause and its marker word. The English chapter 36 §11 shows a run.

The QUD stack (world/src/qud.rs) recognises answers to why through the list REASON = because, since, so, and the wh-types through English question words. The Manner inference "and = in that order" (prag/src/infer.rs, maxims) uses English lists of stative verbs (STATIVE) and reordering words (REORDER).

12.2 What happens to French

Nothing in this layer reads French. No pipeline passes a French tree to the frame reader. The reader does not check the language: prag roles <in.conllu> accepts any CoNLL-U file. But every rule except the last tests an English lemma, so on a French tree only the structural rule would fire. Read from the code (not run on French), the outcome would be this. Il purifie tout car il supprime les odeurs (car cc on a conj) would get next, a sequence where French means a cause. Le crédit était mort parce que… (parce mark) would get no link. Puis (cc + conj) would get the right next, by accident. A French reader needs:

1. A French connective table keyed on the word sequence and the UD relation, not on fixed (§5). LEXCONN is NC-ND and cannot be carried into a release, so the table must be written in Mova's own words from the open papers and checked against the trees; the counts of §5 give its first entries.
2. Cause on cc + conj for car (and donc as cc), which the English rules never needed.
3. Metatalk links for puisque, car and donc on a speech act (Explanation*, Result*, §4). A plain cause link would say that a goal caused an evaluation (ex. 4).
4. Cross-sentence attachment of sentence-initial parce que, mais, donc, or, which UD attaches inside their own sentence. In the spoken treebanks a quarter (Rhapsodie) to nearly half (ParisStories) of the sentences start with a connective.
5. Tense from FEATS. The passé simple / imparfait contrast (§8.2) is in the treebank's Tense feature. Mova's parser sees only forms and PTB tags, which have neither tense (chapter 30 §9.2).

12.3 The French pragmatics layer for the Eel, and the meta-pragmatics

The Eel, the learner that improves Mova's trees and pragmatics, schedules its rounds by kind. In the code in development on 2026-11-06, one kind is prag-fr, a gated round for French pragmatics (tools/eel/src/schedule.rs, KINDS; tools/eel/src/native.rs). Such a round runs a pinned native executable on the CPU. The Eel itself checks the pinned dev sample, the disjointness of train and dev keys and the reported metric. It installs a candidate only on a dev gain, and it queues proposals for review only. The French pragmatics executable is not written yet. There is no French connective table and no measurement. The schedule gives the work a slot and a gate, nothing more.

The meta-pragmatics is the planned language-independent layer: one schema of relations and acts that every language's reader writes to, with each language's connectives mapped into it. It is a direction, not code. What the readers share today is the UD relations they start from and the English frame fields (cause, next, quote, cond). §3 shows why the schema cannot simply borrow one framework's labels. LEXCONN's SDRT relations and the PDTB senses coincide only on Concession and Condition (Laali & Kosseim 2017), and implicit relations shift across translation (DiscoGeM 2.0). The mappings must stay explicit and revisable.

12.4 Not yet

• No French connective list, no French link rule, no French QUD reasons (parce que, car, puisque, comme).
• No metatalk relations (Explanation*, Result*) in any language; no Contrast, Concession, Opposition or Elaboration fields.
• No cross-sentence attachment of sentence-initial connectives.
• No discourse tree or graph and no EDU segmentation, for English or French.
• None of the French discourse resources of §3 is held in $MOVA_DATA. The open ones (ConcoLeDisCo, Relex) are not used even for measurement.
• The prag-fr round has a schedule slot and a gate, but no executable.

In Mova

• Links: prag/src/frames.rs (set_cause, set_next, quote); inference: prag/src/infer.rs (maxims, STATIVE, REORDER); QUD: world/src/qud.rs (REASON).
• Reading a CoNLL-U file: prag roles (prag/src/main.rs).
• The Eel's rounds: tools/eel/src/schedule.rs, tools/eel/src/native.rs (in development, 2026-11-06).
• The Mova dialect on discourse: dialects/mova/seeds.md#discourse-adverbs.
• English counterpart: 36. Discourse relations and coherence.

Sources

• Brunot, F. 1899. Précis de grammaire historique de la langue française. brunot-1899 (public domain).
• Clédat, L. 1894. Grammaire raisonnée de la langue française. cledat-1894 (public domain).
• Afantenos, S. et al. 2012. An empirical resource for discovering cognitive principles of discourse organisation: the ANNODIS corpus. LREC. afantenos-2012-empirical-resource-discovering-cognitive-principles (CC BY-NC 4.0: informs only).
• Péry-Woodley, M.-P. et al. 2009, 2011. ANNODIS. woodley-2009-annodis-une-approche-outillee-de, woodley-2011-le-corpus-annodis-un-corpus (CC BY-NC-SA 3.0: informs only).
• Atallah, C. 2015. La ressource EXPLICADIS. atallah-2015-la-ressource-explicadis-un-corpus (CC BY-NC-SA 3.0: informs only).
• Danlos, L. et al. 2012. Vers le FDTB : French Discourse Tree Bank. danlos-2012-vers-le-fdtb-french-discourse (CC BY-NC-SA 3.0: informs only).
• Danlos, L., Colinet, M., Steinlin, J. 2015. FDTB1, première étape du projet « French Discourse Treebank ». Discours 17. danlos-2015-fdtb1-premiere-etape-projet-french (CC BY-NC-ND 4.0: informs only).
• Roze, C., Danlos, L., Muller, P. 2012. LEXCONN: A French Lexicon of Discourse Connectives. Discours 10. roze-2012-lexconn-french-lexicon-discourse-connectives (CC BY-NC-ND 4.0: informs only).
• Laali, M., Kosseim, L. 2017. Automatic Mapping of French Discourse Connectives to PDTB Discourse Relations. SIGdial. laali-2017-automatic-mapping-french-discourse-connectives (CC BY 4.0).
• Rauh, S., Zaczynska, K., Bourgonje, P. 2023. Toward a Multilingual Connective Database: Aligning German/French Concessive Connectives. KONVENS. rauh-2023-multilingual-connective-database-aligning-german (CC BY 4.0).
• Sow, F., Toussaint, Y., Constant, M. 2026. Relex: A Common Denominator for Connectives and Discourse Relations in French. SIGdial. sow-2026-relex-common-denominator-connectives-discourse (CC BY 4.0).
• Pastor, M., Marino, E. B., Oostdijk, N. 2024. La reconnaissance automatique des relations de cohérence RST en français. TALN. pastor-2024-la-reconnaissance-automatique-des-relations (CC BY 4.0).
• Yung, F. et al. 2024. DiscoGeM 2.0: A Parallel Corpus of English, German, French and Czech Implicit Discourse Relations. LREC-COLING. yung-2024-discogem-2-0-parallel-corpus (CC BY-NC 4.0: informs only).
• Blochowiak, J., Grisot, C., Degand, L. 2020. What type of subjectivity lies behind French causal connectives? Glossa 5(1): 50. blochowiak-2020-what-type-subjectivity-lies-behind (CC BY 4.0).
• Blochowiak, J. et al. 2026. Argumentation through Discourse Relations and Subjectivity: FreCaDiS. LREC. blochowiak-2026-argumentation-through-discourse-relations-subjectivity (CC BY-NC 4.0: informs only).
• Richard, A. et al. 2024. FRACAS: a FRench Annotated Corpus of Attribution relations in newS. LREC-COLING. richard-2024-fracas-french-annotated-corpus-attribution (CC BY-NC 4.0: informs only).
• Bras, M., Le Draoulec, A., Vieu, L. 2001. Temporal Information and Discourse Relations in Narratives: the role of French connectives puis and un peu plus tard. ACL workshop. bras-2001-temporal-information-discourse-relations-narratives (CC BY-NC-SA 3.0: informs only).
• Herreño Castañeda, P., Sillaire, M. 2025. Annotation de marqueurs discursifs : le cas de la désambiguïsation de après. RECITAL. castaneda-2025-annotation-de-marqueurs-discursifs-le (CC BY 4.0).
• Hunter, J., Prévot, L., Muller, P. 2023. Comparing Methods for Segmenting Elementary Discourse Units in a French Conversational Corpus. NoDaLiDa. prevot-2023-comparing-methods-segmenting-elementary-discourse (CC BY 4.0).
• Prévot, L., Bertrand, R., Hunter, J. 2025. Segmenting a Large French Meeting Corpus into Elementary Discourse Units. SIGdial. prevot-2025-segmenting-large-french-meeting-corpus (CC BY 4.0).
• Maachou, K., Braud, C., Muller, P. 2025. Supervision faible pour la classification des relations discursives. TALN. maachou-2025-supervision-faible-pour-la-classification (CC BY 4.0).
• Badene, S. et al. 2019. Apprentissage faiblement supervisé de la structure discursive. TALN. badene-2019-apprentissage-faiblement-supervise-de-la (CC BY 4.0).
• Wanted (not open): Halliday & Hasan 1976; Asher & Lascarides 2003; Anscombre & Ducrot, L'argumentation dans la langue (1983); Kamp & Rohrer 1983; Weinrich, Le Temps (1973).
• Universal Dependencies French treebanks, UD 2.18: GSD, ParisStories, Rhapsodie (CC BY-SA 4.0; short quotations with sent_id), Sequoia, FQB (LGPL-LR), PUD (CC BY-SA 3.0), ParTUT (CC BY-NC-SA 4.0: counts only). Counts made on 2026-11-06 from UD_French-*, all splits together, one count per syntactic word, punctuation in the denominator; multiword connectives matched on the head word plus its fixed dependents (qu' normalised to que).

This page in Markdown → https://md.mova.elamur.ai/free/en/docs/languages/fr/36-discourse-relations/