French adverb d'abord and Discourse Structure Myriam Bras1 Université de Toulouse, CLLE-ERSS, UMR 5263 CNRS & Université Toulouse-Le Mirail 1. Introduction This work is part of a broader study of linguistic units that have an impact on Discourse Structure, such as puis (Bras et. al 2001) or and (Gómez-Txurruka 2003), both markers of coordinating relations. To pay homage to Isabel Gómez-Txurruka, we have chosen the adverb d’abord which has an affinity with subordinating relations. Our first idea is that d’abord could be a subordinator, just as and is a coordinator, as Isabel showed in her very stimulating work. Among the adverbs of French, d’abord is described as a conjunctive sentence adverb, because its function is to establish a link between two discourse units (Guimier 1996, p.125). It is part of the sub-class of enumerative adverbs, sometimes also called ordering adverbs. Their function is to organize and structure discourse (Molinier and Lévrier 2000, p.57). Therefore, d’abord is often mentioned in the literature devoted to surface markers of enumeration. It is analyzed as a ‘marker of linear integration’ by Turco and Coltier (1988), as a ‘temporal organizer’ by Adam and Revaz (1989), and taken into account in fine descriptions of enumerative structures (Péry-Woodley 2000), (Luc 2001), (Porhiel 2007). These studies on enumerations are also useful for data mining applications (Jackiewicz 2002), (Jackiewicz and Minel 2003). These approaches take a macroscopic point of view on texts and analyze their organization on the basis of surface markers. D’abord is also analyzed in contrastive studies (Dalmas 1998), (Aoki 1999) that better reveal its general semantic content. We propose to analyze the role of d’abord in interaction with other elements at the semantic level, in a more semantic, bottom-up fashion. For 1 Many thanks to Laurent Prévot for the discussions on the formal proposal. I’m glad to associate him to this homage to Isabel. 77 that purpose, we chose the theoretical framework of SDRT that is specially designed to handle semantic and pragmatic information and their impact on discourse structure. 2. Theoretical framework to account for Discourse Structure 2.1. General principles of SDRT Segmented Discourse Representation Theory (Asher 1993, Asher and Lascarides 2003) is a formal theory of discourse. It follows on from Discourse Representation Theory (Kamp and Reyle 1993) to which it added the crucial notion of Discourse Structure by taking Discourse Relations into account. SDRT makes use of semantic information, world knowledge and pragmatic principles. It is actually a theory of the semantics-pragmatics interface. SDRT representations for discourses are called Segmented Discourse Representation Structures. An SDRS is a recursive structure consisting of labeled elementary DRSs (i.e., Discourse Representation Structures from DRT) representing a single clause and labeled sub-SDRSs linked together by Discourse Relations, such as Narration, Elaboration, Background, Continuation, Result, Contrast, Explanation… These elementary DRSs and the sub-SDRSs corresponding to complex discourse segments are the constituents of the SDRS representing the discourse. The elementary constituents describe eventualities, i.e. events or states. Each constituent is associated to a label. Labels are used to distinguish different occurrences of constituents : a given proposition or semantic content may have many different uses in different discourse contexts, and each occurrence of a constituent in a discourse structure will be affected differently, at least in principle, by the discourse context. Another way to think of these labels is as discourse referents for speech acts. To keep the whole theory within a first order setting, discourse relations take labels as arguments. We’ll use Greek letters (α, β,…) as variables to represent labels, which serve as markers for speech acts, Kα to represent the constituent with the label α and eα (eβ,…) to represent the main eventuality they describe in our formulation of the axioms for inferring and interpreting the various discourse relations linking constituents. To build an SDRS for a discourse, SDRT offers a “Glue Logic” and an “Update Function” that together determine a new SDRS for a given SDRS K0 representing the context (the discourse already processed) and a new 78 constituent α representing the information to be integrated into that context. The Glue Logic exploits the framework of `Commonsense Entailment' (Asher and Morreau 1991), a logic that exploits both monotonic (→) and non-monotonic (>) conditionals. In Commonsense Entailment (CE), φ > ψ means “if φ then normally ψ”. From φ > ψ and φ, CE entails ψ “by default”, that is, defeasibly, in the absence of further information regarding the truth value of ψ2. Asher and Lascarides (2003) develop the glue logic within the context of a description logic that makes extensive use of underspecification. While we will not go into details here, we will adopt their convention where predicates like discourse relation predicates that customarily would be understood as denoting two place relations will in the glue logic be represented as three place predicates, the third argument being reserved for the label of the smallest SDRS containing the formula involving the discourse relation. As far as the inferential tasks are concerned, the Glue Logic contains: (i) definitions characterizing which constituents in the contextually given SDRS are open for attaching β and (ii) axioms detailing what discourse relations may be inferred, on the basis of a variety of linguistic and common knowledge clues, in order to actualize the attachment of α to some open constituent of the contextually given SDRS. SDRT also contains axioms specifying the semantic effects of those discourse relations, which can be considered as meaning postulates. The Update Function is in charge of the proper hierarchization of the structure and of the resolution of the possibly existing underspecifications (e.g., anaphora and ellipses). 2.2. Discourse Relations SDRT distinguishes coordinating relations from subordinating ones. Narration and Continuation are coordinating relations, while Explanation and Elaboration are subordinating ones. Only subordinating relations may introduce complex SDRSs, in other words, the Update Function may gather several SDRSs into a new complex SDRS only if these constituents are attached to the same site with the same subordinating relation. Asher and Vieu (2005) provide criteria within SDRT for coordinating and 2 From φ > ψ, φ and ┐ψ, CE does not entails ψ, but ┐ψ. From φ > ψ, ζ > ┐ψ, φ → ζ, φ and ζ, ψ (and not ┐ψ) is inferred (Penguin principle). From φ > ψ, ζ > ┐ψ, φ and ζ, if φ and ζ are logically independent, CE cannot conclude ψ nor ┐ψ (Nixon Diamond). 79 subordinating relations. They also postulate that some relations like Result, though in most cases a coordinating discourse relation, may in certain contexts be subordinating instead. In SDRT several discourse relations may simultaneously link the same two constituents, thus distinguishing SDRT from other discourse approaches, notably RST (Mann and Thompson 1988). SDRT also allows for multiple ‘’superordinate’’ parents of a constituent, which means that SDRSs must be graphs and cannot be represented completely faithfully as trees. We now describe the semantics of one discourse relation of each kind: a coordinating one, Narration, and a subordinating one, Elaboration. Narration Narration encodes the principle ‘textual order matches real order’, that is to say the Gricean pragmatic maxim of manner “be orderly”. When two clauses are linked by Narration, they describe in sequence two successive events `of the same story'. Narration can be non-monotonically inferred if the predicate Occasion holds between the clauses to be related. This is what the following axiom expresses (λ represents the smallest constituent that will end up containing the formula that links α and β once the SDRS for the discourse is constructed and fully specified): Narration (? (α, β, λ) ∧ Occasion(α, β)) > Narration (α, β, λ) Occasion holds if the two clauses contain clues indicating that their main eventualities are of types that may belong to ‘the same story’. Occasion is a predicate of SDRT’s glue logic whose semantics involves those linguistic elements available in the logical forms of the discourse constituents that are relevant to inferring Narration. It exploits lexical semantics and shared knowledge in terms of scripts connecting certain event types in sequences in which one event ‘naturally’ leads to the next. For instance, (1) is an example of Narration where Occasion holds, since there is clearly in the shared knowledge a script in which, before entering, people knock at the door. (1) Pierre frappa à la porte. Il entra. (Peter knocked at the door. He entered.) We now turn to another way of inferring Narration, with a monotonic inference this time. This is when a specific discourse marker is present. In (Bras et. al 2001), it was shown that French adverb puis is such a marker. Its role in SDRT is described by the following axiom: 80 Narration_Puis (? (α, β, λ) ∧ puis(β)) → Narration (α, β, λ) Last, Narration can be inferred between α and β from information relevant to subsequent constituents, one of which is linked to β. An example motivating this sort of rule occurs with a simple discourse like (2): (2) Marie a ouvert la boite. Elle a pris un bonbon. Puis elle a refermé la boite sans faire de bruit. (Mary opened the box. She took a sweet. Then she closed the box again without a sound.) The last two clauses are forcibly linked by Narration because of the discourse connective puis. But this discursive link leads us to interpret the link between the first and the second clause as one of Narration. It is this sort of situation that the following axiom on Subsequent Relations, is designed to address: Subsequent Relations (? (α, β, λ) ∧ R(β, γ , λ’) > R(α, β, λ) where R is any discourse relation used in SDRT. Let us now examine the two major semantics effects of Narration on discourse content. The first one aims at capturing the fact that narratives must cohere in the sense that the events linked together by Narration must fit consistently and without significant spatio-temporal gaps as expressed in (Asher et al 1995). This doesn't mean that there should be no interval of time between the two events eα and eβ, but rather that no relevant event, i.e. no event interfering with eα or eβ, can occur during this interval. This constraint is formalized in (Bras et al. 2001) by the following axiom rewritten in the notation of Asher and Lascarides (2003): Spatio-temporal consequence of Narration ΦNarration(α, β) ⇒ eα ⊃⊂ (post(eα) ∩ pre(eβ)) ⊃⊂ eβ This axiom expresses a meaning constraint on Narration. It uses not the language of glue logic itself but the base logic in which the semantics of SDRSs is given. This language is much richer than the glue logic and can appeal to notions that are used traditionally in semantics, including the relation of abutment (⊃⊂) used in DRT and the functions poststate and prestate – the poststate of eα, post(eα), is a state that begins at the end of and persists indefinitely into the future, while the prestate of eα is the state that terminates at the beginning of eα and extends indefinitely far back into the past. 81 The right hand side of the axiom expresses that eα abuts post(eα) ∩ pre(eβ) which in turn abuts eβ. It entails in addition that there is no intervening event whose propositional content interferes with neither that of post(eα) nor that of pre(eβ), i.e., no `relevant' event that ends post(eα) before eβ starts or prevents pre(eβ) from holding right after eα has ended. From this axiom and uncontroversial ordering assumptions on events and their pre- and post-states (Event(e) → pre(e) ⊃⊂ e ⊃⊂ post(e)), we can deduce a relation of temporal precedence between the events eα and eβ: eα < eβ, which was the original temporal effect of Narration in (Lascarides and Asher 1993). Narration has a second semantic effect. It is motivated by the intuition that the elements of a Narration must belong to the `same story', i.e., they must have some common subject matter. To this effect, the axiom below expresses that the constituents connected together by Narration must have a common Discourse Topic. A Discourse Topic is a simple constituent which is contingent (i.e., not vacuous, not contradictory, not tautologic), and subsumes the constituents of a sub-SDRS, in this case, the constituents linked by Narration. Topic constraint on Narration ΦNarration(α, β) ⇒ ¬ (Kα Kβ) where is a merge operation defined in Asher (1993) for defining topics. The axiom states that the topic of α and β cannot be vacuous, which we formalize in SDRT as a necessary truth. If the Discourse Topic is not already present in the context, the introduction of a new Discourse Topic – actually linguistically implicit-- is required. It has to be added to the SDRS during the update. This will be ensured by the update function of SDRT, and more precisely by the following constraint (Asher and Lascarides 2003, p.219) where ⇓ denotes a ‘topic’ relation and it’s subordinating: Narration (α, β, λ) → ∃ δ ((δ = α β) ∧ ∃ γ (⇓ (δ, λ, γ))) Actually, “Topic constraint on Narration” above and the rules of glue logic imply that Narration can be non-monotonically inferred only if such a topic exists or can be built. To illustrate these structural effects, let us consider the simple example (1) again. This small text clearly tells the story of Peter arriving at someone's home or office. Such a Discourse Topic is inferred only because the two sentences are textually linked together and in this order. 82 Elaboration Elaboration is a subordinating relation. It introduces a complex constituent, that is, a SDRS that embeds other SDRSs. Let us illustrate this discourse relation with a self-made example inspired by the well-known Cervin example of (Kamp and Rohrer 1983) that we will analyse later on: (3) (a) Mixel a escaladé le Vignemale hier matin. (b) Il est parti du refuge des Oulettes au lever du jour. (c) Ensuite il a passé la Hourquette d’Ossoue, (d) puis il est arrivé au sommet vers midi. (e) Il était très heureux d’avoir enfin gravi ce sommet pyrénéen mythique. In (3), the label for the first constituent πa representing sentence (a) is elaborated by the labels of the constituents πb, πc, πd representing, respectively, sentences (b), (c) and (d). In the graphical representations of the Discourse Structure of (3) in Figure 1, πA is the label for the complex constituent embedding the elaborating constituents. We’ll adopt the representation on the right, which is the more standard one (the left one actually mixes DRS like notations and trees representing the discourse structure): Figure 1: Graphical representations of the SDRS for (3). As Elaboration is a subordinating relation, Elaboration(πa, πA) entails ⇓(πa, πA) which gives to πa the status of an explicit Discourse Topic. We now explain how this SDRS is built, because it will be useful for the analysis of examples with d’abord. Elaboration can be non-monotonically inferred if there is a subsumption relation between the types of the eventualities involved: ( ?(α, β, λ) ∧ Top(σ , α ) ∧ SubtypeD(σ, β,α)) > Elaboration3 Elaboration(α, β, λ) 3 We simplify the axiom of Asher and Lascarides (2003) as we will not consider different aspectual classes. 83 SubtypeD means that the type of the second eventuality is a subtype of the first one in the lexical semantics of the predicates or by some piece of shared knowledge. For (3), thanks to world knowledge about mountain climbing in the Pyrénées and lexical semantics of motion verbs escalader and partir we can infer SubtypeD. So the Elaboration Axiom applies and leads us to infer Elaboration(πa, πb) and build the first SDRS whose structure is graphically represented in Figure 2 (see the first step on the left handside). π0 is the local structure in which the Elaboration(πa, πb) condition will be embedded and corresponds to the third argument of Elaboration (λ) in the axiom above. Next, the label for the third constituent of this small discourse, πc, has to be attached, πb is an available attachment site, and we are in presence of ensuite that triggers Narration4, following an axiom similar to Narration_Puis above. So we infer Narration(πb, πc). Then we apply the axiom of Distributivity, meaning “that if you attach something to a constituent that is part of an Elaboration, what you attach should also be part of the Elaboration” (Asher and Lascarides 2003 p.207). The axiom Complex Constituents (ibid. p. 162) will lead us to build the complex constituent πA which as a whole elaborates πa : Elaboration(πa, πA, π0), as the middle graph in Figure 2 shows. Finally, we apply Narration_Puis to infer Narration(πc, πd) and the axioms Distributivity and Complex Constituents to build the discourse structure represented on the right of Figure 2. Figure 2: Steps of the construction of the SDRS for (3). 4 We have not proved it as we did for puis in (Bras et al. 2001), but we can reasonably assume it. 84 Let us analyze another typical example of Elaboration from (Kamp and Rohrer 1983) in (4): (4) (a) L’été de cette année-là vit plusieurs changements dans la vie de nos héros. (b) François épousa Adèle, (c) Jean-Louis partit pour le Brésil et (d) Paul s’acheta une maison à la campagne. As for (3), the label for the first constituent πa representing sentence (a) is elaborated by the labels of the constituents πb, πc, πd representing, respectively, sentences (b), (c) and (d). The final Discourse Structure is the same than the one represented above but the lower nodes are connected with Continuation relations instead of Narration. First, by lexical semantics that tells us that marrying events, leaving events and buying events are changes, we infer SubtypeD and Elaboration (πa, πb) and Elaboration (πa, πc). The axiom of Distributivity says that if we have inferred Elaboration(πa, πb, π0) we may attach πc to πb only if we can also infer Elaboration(πa, πc, π0), which is the case. The discourse relation that holds between πb and πc is Continuation, a coordinating relation with no temporal consequences. The following theorem adapted from (Prévot 2004, p. 60) states this explicitly: (?(β, δ, λ) ∧ Elaboration(α, β, λ) ∧ Elaboration(α, δ, λ)) → Continuation(β, δ, λ) Finally, Complex Constituents will lead us to build the complex constituent πA which as a whole elaborates πa: Elaboration(πa, πA, π0). Other ways of inferring Elaboration We now turn to the original Cervin example of (Kamp and Rohrer 1983) to illustrate an other way of inferring Elaboration5. (5) (a) L’année dernière Jean escalada le Cervin. (b) Le premier jour, (c) il monta jusqu’à la cabane H. (d) Il y passa la nuit. (e) Ensuite il attaqua la face nord. (f) Douze heures plus tard, il arriva au sommet. 5 There is another way of inferring Elaboration. For sentences like John promised Mary that he would phone her or John made a promise to Mary that he would phone her, where the promise is elaborated, by an axiom called Post Head Sentential Complements (Asher and Lascarides 2003, p. 285). We don’t describe it here because it is not useful for the analysis of d’abord. 85 In (Vieu et. al, 2005), we analyze Temporal Locating Adverbials like le premier jour or douze heures plus tard when they are dislocated to the left of the sentence, i.e. when they are IP-Adj, as they stand in (5). They play an important part in structuring discourse, by opening a temporal frame in which not only the current sentence but also subsequent sentences are gathered and their eventualities located, after the hypothesis of discourse framing of M. Charolles (1997). We show that this role could be expressed in SDRT by the introduction of a new topic. This new topic, equivalent to “what happened on the first day” for (b) Le premier jour, dominates all the constituents for the sentences in the scope of the frame introducer, namely πc (il monta jusqu’à la cabane H.) and πd (Il y passa la nuit). The temporal location contributed to this New Topic by the adverbial distributes over each constituent subordinated to it, therefore locating the going up event and the spending the night event in the first day. We thus have Elaboration (πb’, πc) and Elaboration (πb’, πd). Moreover, the semantic content of πd is a reliable clue indicating that the temporal setting represented by πb has to be closed (the night ends up this first day). This is the reason why the attachment of πe to πb’ is preferred. This treatment of Temporal Locating Adverbials in IP-Adj position as New Topic Introducers yields the following structure for (5). Figure 3: SDRS for (5). Spatio-temporal consequences of Elaboration Elaboration has a semantic effect on discourse content, it implies that the main eventuality of each elaborating constituent is part of that of the 86 elaborated clause. This implies that there is a temporal inclusion between them: ΦElaboration(α, β) ⇒ Part-of(eβ,eα ) Part-of(eβ,eα) ⇒ eβ ⊆ eα 3. Preliminary observations on d’abord As mentioned in section 1, previous work on d’abord describes its role as a conjunctive, which links its constituent with an antecedent constituent in the previous discourse. The first intuition is that d’abord has something to do with subordinate relations as it is a marker of the first step in a whole process, at least in some of its uses. To understand the role of d’abord in Elaborations, we take example (4) and we introduce d’abord in initial position: (6) (a) L’été de cette année-là vit plusieurs changements dans la vie de nos héros. (b) D’abord François épousa Adèle, (c) Ensuite Jean-Louis partit pour le Brésil, (d) puis Paul s’acheta une maison à la campagne. We could already infer Elaboration without d’abord, ensuite, puis in (4), but here we have Narration instead of Continuation between the elaborating constituents. The temporal order is not due to d’abord but to the others temporal connectives ensuite and puis. They are necessary to have a coherent discourse in presence of d’abord, on the other hand d’abord could be removed. When d’abord is internal as in (7) below, it is not useful to infer Elaboration either, and it can perfectly be removed, but if it is present, the second marker, here puis, cannot be removed. This observation may lead us to the conclusion that, on top of its own constituent, d’abord waits for a second constituent to be coordinated. (7) Laure et Maïlis se sont bien amusées cet après midi à Trento. Elles ont d’abord fait les boutiques, puis elles sont allées manger des glaces. Example (8) below is the Cervin example of (5) with d’abord in (b): (8) (a) L’année dernière Jean escalada le Cervin. (b) D’abord, (c) il monta jusqu’à la cabane H. (d) Il y passa la nuit. (e) Ensuite il attaqua la face nord. (f) Douze heures plus tard, il arriva au sommet. Left dislocated d’abord in (8) has a framing role, i.e. a New Topic Introducer one: the natural interpretation is to group together the going up 87 to the hut event and the spending the night event, thus getting a structure like the one in Figure 3. But what is the semantic content of this New Topic? D’abord does not convey a temporal location as location adverbials do. Nevertheless, it seems to us that it says: “I’m going to explain what happened first, be it one or more events”. It would not be the case with an internal d’abord in (9), for which we have a linear flat structure under πA: (9) (a) L’année dernière Jean escalada le Cervin. (c) Il monta d’abord jusqu’à la cabane H. (d) Il y passa la nuit. (e) Ensuite il attaqua la face nord. (f) Douze heures plus tard, il arriva au sommet. This left dislocated position of d’abord is worth examining, in order to understand better how discourse structure is revealed by linguistic clues. In the following, we want to restrict ourselves to initial and left dislocated d’abord. In the cases where the new topic introduced by a left dislocated d’abord dominates only one constituent, we’ll consider that initial d’abord and left dislocated d’abord are equivalent from the point of view of the discourse structure. A first study of the French adverb d’abord was done in the framework of a project focusing on the discourse relation of Elaboration and its linguistic clues (Bras and Le Draoulec 2007). We analyzed the uses of d’abord in a corpus of short newspapers texts in order to determine whether d’abord (without any restriction of position) was a reliable marker of Elaboration. This led us to distinguish prototypical cases, in which a discourse constituent is elaborated by the constituent introduced by d’abord, from other cases, in which the d’abord constituent elaborates a discourse topic that is not described explicitly by the preceding part of discourse, or which is subordinated by another subordinating relation. In this paper, we apply this first distinction to another corpus, taken from the Frantext textual database. On the basis of the data description, we work out a formal proposal in the SDRT framework. 4. D’abord under Explicit Discourse Topic 4.1. D’abord in temporal Elaborations By ‘temporal Elaborations’, we mean ‘temporally ordered Elaborations’ as the ones we have in (5), (8), (9), (7) or in (3), where an event is elaborated by its sub-events, and where textual order matches temporal order. In our corpus, that kind of example was not so frequent, maybe because of the restriction to initial or left dislocated d’abord. 88 Here is a typical example of temporal Elaboration found in our corpus: (10) (a) Une campagne de dénigrement, trop systématique pour n'être pas organisée, se déchaîna contre la personne même de Sélim. (b) Nos ennemis de l'ombre, armés de la calomnie corrosive, faisaient feu de tout bois. (c) D’abord on répandit le bruit absurde que Sélim était alcoolique, alors qu'il était la sobriété même. (d) Puis on prétendit partout que Sélim était stérile, c'est-à-dire maudit. (Michel de Grèce, La nuit du sérail, 1982, p.373) The sentence featuring d’abord elaborates the explicit topic described in (a). We will ignore (b), which gives a background on (a). The first part of (c) describes the spreading of an absurd rumor about Sélim, while (a) describes the development of the smear campaign. We can reasonably assume that subtypeD holds, then we can infer Elaboration(πa, πc). Next, thanks to Puis, we infer Narration(πc, πd) and a structure sketched in Figure 4, thanks to Distributivity and Complex Constituents (ignoring the complex structure under πc due to the Contrast relation introduced by alors que). Figure 4: Discourse Structure for (10). Example (10) does not provide any evidence of the role of d’abord: we could infer the same discourse structure in the absence of d’abord. Things go differently for example (11), where the clauses (b) and (c) are incomplete. (11) (a) Nous marchâmes longtemps. (b) D’abord, en traversant la ville, (c) ensuite, dans la steppe. (Andreï Makine, Le Testament français, 1995, p. 230) It is impossible to remove d’abord from (11), and it seems that d’abord is necessary to understand that the crossing the town event in (b) is a proper part of the walking event in (a) that will be followed by another proper part, that is foretold, so to say. We will not go into details to explain 89 how the incomplete segments are completed and how d’abord and ensuite take a part in that process, but we know that, to be able to solve the ellipses, we need to attach the constituents πb and πc to πa. We hypothesize that d’abord triggers the inference of the minimal subordinating relation: ⇓. Then, once the attachment is done and the failing material recovered, we can rely on a subtype relation between respectively, a walking/crossing the town event and a walking event, and a walking in the steppe event and a walking event to infer Elaboration(πa, πb) and Elaboration(πa, πc) and then get the same structure than for the example above. Other cases of temporally ordered elaborations are illustrated by example (12), where the explicit topic in (a) is an iterative event – getting used to going out – that gets elaborated by two plural events in (b) and (c). Here again, it is not the presence of d’abord but the presence of subtypeD that leads us to infer Elaboration(πa, πb). But the discourse is better with d’abord. (12) Alors qu'à U. Bridge, pendant dix ans, Lol était si peu sortie, si peu que son mari, parfois, l'y obligeait pour sa santé, (a) à S. Tahla elle prit cette habitude d'elle-même. (b) D'abord, elle sortit de temps en temps, pour faire des achats. (c) Puis elle sortit sans prétexte, régulièrement, chaque jour. (d) Ces promenades lui devinrent très vite indispensables. (Marguerite Duras, Le ravissement de Lol V. Stein, 1964, p.35) To conclude concerning cases of temporally ordered elaborations, it is not the presence of d’abord that leads to infer Elaboration, but a subtype relation between the eventualities. Nevertheless, example (11) shows a first evidence of the role of d’abord: it triggers, at least, the inference of ⇓ to recover the elaborating event. 4.2. D’abord in temporal Enumerations We now turn to the cases of enumerations. The use of d’abord in enumerations is very frequent, whatever the genre of the text, and this is probably the most described use of d’abord. As we saw in section 1, there is a rather extensive literature on enumerations and enumerative structures. Adverbs like d’abord, puis, ensuite, premièrement, deuxièmement, d’une part, d’autre part, etc., generally described as discourse organizers, play the role of specific markers of the co-items of enumerative structures. These structures start with specific starting-up segments, which are identified not only by punctuational (:) and layout markers, but also by their syntactic structure and their semantic features. For example, 90 quantified NP, with abstract nouns such as chose, élément, critère, aspect, but also fait, cause, raison, etc… referring to plural entities may constitute the starting-up segment of an enumerative structure (Péry-Woodley 2000, Porhiel 2007). In these structures, d’abord marks the first item of the enumeration list. The relation between the sentence featuring d’abord and the starting-up segment may be a topic relation, or an Elaboration relation. Let us have a closer look at this. We start with enumerative structures whose starting-up includes an event noun or a temporal noun. In (13) for example, we have a deverbal event noun deux décisions in the starting-up (a). Then each decision is described in (b) and (c) respectively, and textual order matches temporal order: (13) (a) Tout en roulant vers l'île Saint-Louis, il prit deux décisions. (b) D’abord, il allait persuader sa mère de quitter Paris au moins jusqu'au procès et de descendre dans le Midi, peut-être même en Italie où elle ne risquerait rien. (c) Et puis, de son côté, il abandonnerait la rue de Longchamp et louerait un studio meublé. (Michel Droit, Le Retour, 1964, p. 298) It is not clear whether we understand that the content of (b), deciding his mother to leave Paris and so on, is a decision rather from lexical and world knowledge or rather from the enumerative structure. It seems that it is the second solution. If it is the case, it means that subtypeD doesn’t hold, and that we cannot infer Elaboration by subtypeD. So we may want to infer Elaboration(πa, πb) by means of the enumerative structure, but let’s examine other cases, as (14) or (15). (14) (a) Deux séries de découvertes ont joué, de ce point de vue, un rôle particulièrement important. (b) D’abord, les procédés de conservation consistant à enfermer des légumes, du poisson, des fruits, de la viande, dans des boîtes de métal stérilisées (Nicolas Appert, 1809). (c) Ensuite les techniques du froid (Charles Tellier, 1878) : en 1878, un navire frigorifique a pu apporter d'Argentine de la viande […] (J.A. Lesourd, C. Gérard, Histoire économique XIXe et XXe siècle, 1966, p. 357) In (14), the starting-up includes both a temporal noun of period série and an event noun découverte. The two series of discoveries are made explicit by (b) the preservation processes of fresh food in tins and (c) the refrigeration techniques. Human interpreters understand that both (b) and (c) don’t describe the technique itself but its discovering: is it because the enumerative structure makes it clear or because they know that a technique has necessarily been discovered at a certain point? Probably both, but the 91 surface enumerative structure seems to be the stronger one, and here again, it seems that we may want to infer Elaboration(πa, πb) and Elaboration(πa, πc) thanks to the presence of the enumerative structure. Then if we have this two step elaboration structure as for example (10), there seems to be a difference between (10) and (14) considering the propositional content of the Topic constituent. The topic of (14) is lighter than the one of (10), and we think it should need to be enriched, as the New Topics for frame adverbials do. Lastly, we must underline that, as for (13), the textual order of the enumeration list matches the temporal order: ensuite marks Narration and the semantic content, i.e. the dates in (14) confirms that discursive temporal ordering. It is the same for example (15) below: (15) J'ai senti d'un coup toutes sortes d'implications dormantes se manifester, élever la voix, prendre le dessus. Or cela s'est fait en deux temps. D'abord, marche arrière, retour à Rennes, remise de mes pas dans leurs traces enfantines, adolescentes, etc. Cela s'appelle communément reculer pour mieux sauter. Ensuite identification brutale à celui de mes deux frères qui était le plus éloigné de moi, qui était au monde l'homme auquel je me croyais le plus étranger. (Michel Tournier, Les Météores, 1975, p. 37) The starting-up segment includes a temporal noun, temps, meaning phase here. Thus, with the help of d’abord we are inclined to interpret the enumeration list as a temporally ordered one and ensuite can play its role of Narration marker. In the next section, we examine the difference with the cases where the noun in the starting-up is more underspecified and has no temporal content. We’ll see that unlike the cases we’ve just examined, the second ones don’t imply a temporal order between the eventualities described in the enumeration list. But let us first draw a preliminary conclusion on the temporal enumerations: i. it seems useful to detect the presence of enumerative structures, relying on the form of the starting-up (quantified NP, etc.. ) and the presence of a item introducer such as d’abord, ii. it seems useful to detect that an enumerative structure is temporal, relying on the semantic type of the N and its link with the verb6 in 6 Let us mention that some verbal constructions with support or “light” verbs and quantified predicative nouns such as prendre deux décisions can be replaced by a semantically equivalent construction with a “heavy” verb and a “light” noun such as décider deux choses. Of course we should be able to identify both of them as temporal starting-ups. 92 iii. iv. 4.3. the starting-up: in those cases, the coordinating relation under the topic is Narration. we need to infer a subordinating relation between the starting-up and the enumeration items, but we’ve not chosen yet whether this relation should be Elaboration or a weaker subordinating relation such as ⇓ as suggested in the previous section, the topic in the starting-up constituent has to be completed. D’abord in non temporal Enumerations In what we call non temporal enumerations, the starting-up segments do not convey any temporal content, that is to say, no temporal nouns or event nouns, nor events at all7, the nouns still are abstracts and much lighter, in that they have a weak semantic content: critère, élément, chose, aspect, point, etc. as in (16), (17) or (18): (16) (a) La phrase s'y trouve définie selon deux critères. (b) D’abord, elle est l'ensemble de mots (coïncidant éventuellement avec un seul) que l'usager de naissance accepte comme complet, c'est-à-dire se suffisant à lui-même et n'exigeant pas d'addition pour être grammaticalement correct et sémantiquement interprétable. (c) Le second critère est formel : un certain contour intonationnel indique […] (Claude Hagège, L'Homme de paroles : contribution linguistique aux sciences humaines, 1985, p. 207) (17) (a) L'hygiène dentaire […] se fonde sur plusieurs éléments. (b) D’abord, il importe d'assurer un bon nettoyage des dents. (c) Puis il faut savoir faire appel au dentiste dès que l’on constate quelque chose d'anormal. (Encyclopédie médicale Quillet : nouvelle encyclopédie pratique de médecine et d'hygiène, 1965, p. 179) (18) Ce que j'aime dans ce jardin, c'est deux choses. D'abord, il y a la statue; c'est dans un coin à l'écart, elle est toute pleine de vert-degris […] Après la statue, il y a l'allée sur le côté avec les branches qui se rejoignent au-dessus des têtes, […] (Patrick Cauvin, Monsieur Papa, 1976, p. 91) These three examples are very different from the typical cases of Elaboration presented in sections 2, 3, 4.1 and from the temporal 7 See previous note. 93 enumerations described in 4.2. Here, we have no temporal ordering among the eventualities described in the subordinated constituents. In (16) and (18) we have descriptions involving states, while in (17) we have a prescription involving processes (plural iterative events) that may be concomitant, and co-extensive with the state described in the topic. So we could assume that enumerative structures trigger Elaboration and that the temporal consequence of Elaboration holds, just because we can exploit the possibility that the consequence of the part_of relation (eβ ⊆ eα) includes the co-extensive case (eβ ≡t eα). Or we could assume that we have a weaker subordinating relation, with no temporal consequences. But, in both cases, we must account for the fact that the connective puis in (17) does not trigger Narration, as the Narration_Puis axioms in section 2 predicts it. In (Bras et. al. 2001) we had left enumerative puis out of the scope of our study, that’s the reason why Narration_Puis only accounts for narrative cases. Here the data show that the enumerative use of puis comes when we are in a non temporal enumerative context. Let us examine a last example, (19), for which the starting-up is not that of a temporal enumeration: an aspect of a situation is not one of its part, but only a point of view on it. (19) [Il] veut que la situation de son pays soit connue en Amérique d'où peut venir l'aide militaire décisive. (a) Cette situation, il la dépeint sous deux aspects. (b) D’abord, il raconte l'histoire de juin 1940, du Massilia, de son arrestation, de son procès. Non pour se disculper, car cela, dit-il, ne concerne que lui et ses juges. Mais pour montrer au monde, et d'abord aux Alliés américains, ce qu'est devenue la justice française sous la férule de Vichy et de l'Allemagne nazie. (c) Ensuite, à travers le récit de sa vie clandestine dans la France occupée, il témoigne des sentiments profonds des français. (Pierre Mendès-France, Oeuvres complètes, 1, S’engager, p. 341) A first rapid reading may lead a human interpreter to infer that the protagonist raconte (tells, in (b)) and témoigne (brings evidence, in (c)) at the same time, by the same picture of the situation described in (a) and the previous discourse. But if he goes into the details of the semantic content of each segment (b) and (c), it becomes clear that what is told in (b) is something that happen before the facts through which he brings evidences in (c): the events of the second world war in France in june 1940 are anterior to the occupation of France. So here, it is the semantic content, together with world knowledge, that helps the interpreter to reorganize the eventualities in a temporal sequence, not ensuite as a temporal connective. 94 To conclude, we can echo the item (ii) in the conclusion of 4.2 and say: i. ii. iii. it seems useful to detect that an enumerative structure is not temporal, in those cases the coordinating relation under the topic is Continuation, and we must block the temporal role of connectives such as puis or ensuite, non temporal enumerations do not provide more clues than the temporal ones to help us determine the nature of the subordinating relation that has to hold between the starting-up and the enumeration items, but the minimalist assumption of ⇓ made in section 4.1 seems to hold water, for both kind of enumerations, d’abord is the first item in a series of markers (d’abord, puis, ensuite …) and at least one other constituent is expected: so d’abord requires a coordinate relation on the right with an underspecified constituent. Let us now come back to the question of whether the presence of an enumerative structure should systematically trigger the inference of Elaboration. It seems that a more exhaustive study of enumerative structures should be necessary to conclude and we propose to leave that for further investigations. This question of the limits of Elaboration in SDRT, and of what exactly this discourse relation covers on top of the typical cases of temporally ordered elaborations is discussed in the literature (see Prévot 2004, 191-201) and this further investigation should ground on this discussions. For now, we think that it is worth taking into account the presence of a relation at the textual level, which does not systematically goes with a relation at the semantic level. Asher and Lascarides (2003) distinguish between content-level relations as Narration or Elaboration and Text Structuring relations such as Contrast and Parallel. We propose the introduction of a new Text Structuring Relation called Enumeration. The inference of this relation is triggered by the identification of the beginning of an enumerative structure, i.e. a starting-up segment and an item introducer such as d’abord. Constraints on Enumeration should help us control that the number of steps announced in the starting-up are well realized in the following discourse. 4.4. Proposal All the partial conclusions we’ve drawn up to now in sections 3 and 4 lead to state that: 95 - - d’abord triggers, at least, the inference of ⇓ between an explicit discourse topic (be it a single event that can be decomposed in parts, a plural event, a plural abstract entity, an enumeration starting-up), d’abord waits for a second constituent to be coordinated, on top of its own constituent. So d’abord requires both subordination – above him in the discourse structure (with a segment that lies before him in the text) – and coordination – on its right at the same level in the structure (with a segment that will come after it in the text). We express that by the following axiom: D’abord ( ?(α, β, λ) ∧ d’abord(β)) > ( ⇓(α,β,λ) ∧ ∃ γ ∃ R ∈ Coord R (β,γ,λ)) In words, if β is to be attached to α in the local structure λ, and β features d’abord, then we infer non monotonically that the relation between α and β is a minimal subordinating relation ⇓ and that there exists a constituent γ to be attached to β with a coordinating relation. As Isabel Gómez-Txurruka showed in (Gómez-Txurruka 2003) for the connective and, and requires a coordinating relation and the requirement of a coordinating relation between two constituents implies the construction of a Coordinated Discourse Topic. Our analysis of d’abord was much inspired by her work on and. It is precisely the construction of this coordinate discourse topic that will help us build a substantial topic for our enumerations. We’ll have to examine precisely under which conditions the Maximise Discourse Coherence principle of SDRT will be able to keep the richer topic from this new CDT and α. Now, for the enumerative structures, we can formulate the following axiom: Enumeration (?(α, β, λ) ∧ Enum-Starting-up(α) ∧ d’abord(β)) > Enumeration(α,β,λ) It expresses that in presence of an enumerative structure, signaled by a starting-up expression (quantified NP, etc.. ) and the presence of an introducer item such as d’abord, we infer non monotonically the text structuring relation of Enumeration, which will also be triggered by other markers of enumeration. Enumeration is subordinating, therefore it entails ⇓(α,β,λ). We still have to account for the blocking of the temporal role of 96 connectives such as puis or ensuite in non temporal enumerations. But this problem lays outside the scope of this paper devoted to d’abord, so we leave it for further work. With these axioms, for all the data analysed so far, where (a) describes an Explicit Discourse Topic (be it an Enumeration starting-up or not) and (b) is the constituent featuring d’abord, we infer either Elaboration(πa, πb) when subtypeD holds, either Enumeration(πa, πb) when Enum-Startingup(πa) holds, either both relations. In the next section we analyze new data bringing other subordinating discourse relations into play. 5. D’abord requires a Discourse Topic In the examples analyzed so far, the topic was always explicit. So we could have inferred the right discourse relations with: ( ?(α, β, λ) ∧ d’abord(β)) > ⇓(α,β,λ) We explained that the condition “∃γ∃R∈Coord R(β,γ,λ)” was there to trigger the construction of a coordinated discourse topic and thus would help us enrich the explicit topic. The usefulness of this condition will come clearer in this section. 5.1. D’abord with Result Let us examine an example where the Discourse Topic is not explicit at first sight: (20) Elle a commencé à lui rendre visite le jour, puis elle s'est mise à découcher. Elle dormait dans la chambre de bonne, dans un lit à une place et tandis que la malheureuse luttait contre le sommeil, Simon délirait sur la lutte des classes. Il lui demandait de lui raconter son enfance et son adolescence en détail. Ça avait l'air de l'intéresser beaucoup. (a)Une nuit, (b) il lui a demandé de faire la même chose pour moi, (c) les récits d'enfance prolétarienne ne le lassaient jamais. (d) D'abord, Mariella a fait une scène de jalousie, (e) puis elle a répondu qu'il fallait qu'elle me demande l'autorisation de dévoiler mon passé. (Evane Hanska, Les Amants foudroyés, 1984, p. 62) If we examine the attachment of (d) Mariella exploded in a fit of jealousy to (b) he asked her to do the same thing for me [=to tell him my childhood], we are in the case of a Response relation along the terms of 97 Sandström (1993): the exploding in a fit of jealousy is a reaction or a response to the asking in (b). In SDRT theses cases are treated like Result cases as causeD(πb, πd) holds. Nevertheless, it is not the all story, in (d) d’abord tells us that there is going to be a second reaction of the asking, and this is what comes with (e), which can attach to (d) with Narration and also to (b) with Result (from lexical semantics of the verb, we have causeD(πb, πe). What we want to do intuitively is to group the two steps of the reaction and say that this whole thing is linked to (b) by Result, therefore we need to build this Implicit Discourse Topic. That was our idea in (Bras and Le Draoulec 2007). With the proposal we do in this paper, this grouping will occur under (b), and Result will be made a subordinating relation along the lines of (Asher and Vieu 2005, p. 605). We get the following structure, after simplification by MDC: Figure 5: Discourse Structure for (20). 5.2. D’abord with Explanation We now turn to another subordinating relation: Explanation. The use of d’abord to introduce the first argument of an explanation is very frequent. (21) illustrates such a use: (21) (a) Tu devrais épouser Laura, Jean-Pierre. (b) D’abord, elle est adorable. (c) Ensuite, c'est une des plus riches héritières du Brésil. (Romain Gary (Emile Ajar), Au-delà de cette limite votre ticket n'est plus valable, 1975, p. 253) Here the predicate causeD(πb, πa) holds and we infer Explanation(πa, πb). We could also have (22) with an explicit topic grouping together the two parts of the explanation: 98 (22) (a) Tu devrais épouser Laura, Jean-Pierre, pour au moins deux raisons. (b) D’abord, elle est adorable. (c) Ensuite, c'est une des plus riches héritières du Brésil. Here the quantified NP au moins deux raisons plays the role of a starting-up of enumeration, the semantics of the noun raison is a clue for Explanation. It is also a clue for a non temporal enumeration. This explains why we should block the temporal role of ensuite in (c). It is possible for (22), but not for (21). We have Explanation(πa, πb) and Explanation(πa, πc) and the structure is similar to the one in Figure 5 above with Explanation instead of Result. The examples of d’abord introducing the first explanation of a two or three fold explanation are very frequent. D’abord often co-occurs with an explicit marker of Explanation parce que as in (24) or in (25): (24) Je passe seulement mon temps à enfiler des robes et à les quitter pour en enfiler d'autres. Je suis mannequin... Il ne comprit pas le sens de ce mot: mannequin. D’abord, parce qu'étant du masculin, il ne pouvait s'appliquer à une femme. Ensuite, parce qu'il évoquait pour lui quelque chose d'inerte […] (Robert Sabatier, Les Allumettes suédoises, 1969, p. 220) (25) La discipline essentielle est alors le latin, d’abord parce que le latin est la langue de l' église, parce qu'aussi c'est à Rome plus qu'à Athènes qu'on cherche ses maîtres à penser. (Encyclopédie pratique de l'éducation en France, 1960, p. 125) In most of the cases the eventualities in the subordinating explaining constituents are states, and there is no temporal ordering between them. In some cases, however, we have events, as example (26) shows. We’ll account for that reading of ensuite in further work. (26) Marie a beaucoup maigri. D’abord parce qu’elle a fait un régime. Ensuite parce qu’elle a été malade. 6. Conclusion Our first idea was that d’abord could just be a subordinator as and is a coordinator, but it proves to be also a coordinator: we showed that d’abord requires both subordination with a constituent above him in the discourse structure and coordination with a constituent on its right at the same level in the discourse structure. 99 Analyzing the uses of d’abord led us to have a closer look at enumerations, and to sketch how the semantic content of the starting-up segments could give us precious clues for discourse structure, although not considered as discourse particles. For example, nouns such as raison or cause in the quantified NP included in the starting-up segment that we represent as a discourse topic, could be a sign of Explanation as well as the connective parce que does. We saw that the temporal substance of these nouns could help us understand whether the enumeration is temporally ordered or not. This seems crucial for the interpretation of some discourse connectives as puis and ensuite, whose temporal role proved to be blocked in non-temporal enumerations and explanations. This preliminary work opens motivating perspectives: - - - extend the analysis we’ve done here to all kind of enumerations, starting from the works we mentioned in section 1, and to analyze how the text structuring relation of Enumeration we’ve introduced interacts with content level subordinating relations such as Elaboration, Explanation, Result, examine the interactions between the connectives whose discourse roles were already formalized in SDRT : d’abord, puis, ensuite, et : account for their enumerative uses, account for the role of et when it follows one of the other connectives, account for other uses of d’abord, in particular the argumentative ones, within the same formal proposal. References Aoki, S. (1999). « Adverbes d'ordination, discours et énonciation : d'abord en français et mazu en japonais », in Combettes, B. ; Schnedecker, C. & Theissen, A. (éds), Ordre et distinction dans la langue et le discours. Actes du Colloque international de Metz, mars 1999, Paris, Champion, 5-16. Asher, N. (1993). Reference to abstract objects in Discourse. Dordrecht: Kluwer Academic Press. Asher, N., Aurnague, M. Bras, M., Sablayrolles, P., Vieu, L., (1995). « De l’espace-temps dans l’analyse du discours », Sémiotiques, 9, 11-62. Asher, N., Lascarides, A. (2003). Logics of Conversation. Cambridge: Cambridge University Press. 100 Asher, N., Morreau, M. (1991). “Commonsense Entailment: A Modal Theory of Nonmonotonic Reasoning.” In Mylopoulos, J. & Reiter, R. (eds), Proceedings of the Twelfth IJCAI: Morgan Kaufman, Los Altos, CA. 387-392. Asher, N., Vieu, L. (2005). “Subordinating and Coordinating Discourse Relations”, Lingua, 115(4), 591-610. Adam, J.M., Revaz, F. (1989). « Aspects de la structuration du texte descriptif : les marqueurs d’énumération et de reformulation », Langue Française, 81, 59-98. Bras, M., Le Draoulec, A., Vieu, L. (2001). « French Adverbial Puis between Temporal Structure and Discourse Structure”, in Bras, M., Vieu, L. (eds.), Semantic and Pragmatic Issues in Discourse and Dialogue: Experimenting with Current Theories, CRiSPI series, vol. 9. Amsterdam: Elsevier, 109-146. Bras, M., Le Draoulec, A. (2007). « D’abord : marqueur ou indice de la relation d’Élaboration ? », Journée scientifique du projet « Relations de Cohérence et Fonctionnement des Anaphores », funded by the Institut de la Langue Française, 29 juin 2007. Charolles, M. (1997). « L’encadrement du discours : univers, champs, domaines et espaces. », Cahier de Recherche Linguistique, 6, 1-73, Landisco, Nancy. Dalmas, M. (1998). “D’abord, et après ? Le marqueur d’intégration francais et ses cousins germains” In : Les corrélats anaphoriques, C. Schnedecker (ed.) Recherches linguistiques Vol. XXII, Université de Metz, Klincksieck, Paris, 75-95. Gómez-Txurruka, I. (2003). « The natural language conjunction and”, Linguistics and Philosophy, 26, 255-285. Guimier, C. (1996). Les adverbes du français, Ophrys, Gap/Paris. Jackiewicz, A. (2002), « Repérage et délimitation des cadres organisationnels pour la segmentation automatique des textes », in actes de CIFT’02, Hammamet, Tunisie, 95-107. Jackiewicz, A., Minel, J.-L. (2003). L'identification des structures discursives engendrées par les cadres organisationnels in Actes de TALN 2003, Batz-sur-Mer. Kamp, H., Reyle, U. (1993). From Discourse to Logic. Kluwer, Dordrecht. 101 Kamp, H., Rohrer, C. (1983). “Tense in texts”, in Bauerle, R., Schwarze, C., von Stechow, A. (eds), Meaning, use and the interpretation of language, Berlin : De Gruyter. Lascarides, A., Asher, N. (1993). “Temporal Interpretation, Discourse Relations, and Commonsense Entailment”, Linguistics and Philosophy 16(5): 437-493. Luc, C. (2001). “Une typologie des structures énumératives basée sur les structures rhétoriques et architecturales du texte », in Actes de TALN 2001, 263-272. Mann, W.C., Thompson, S., (1988). “Rhetorical Structure Theory: towards a Functional Theory of Text Organization”, Text, 8-3, 243-281. Molinier, C., Lévrier, F. (2000). Grammaire des adverbes, Droz : Genève/Paris. Péry-Woodley, MP (2000), Une pragmatique à fleur de texte, mémoire d’Habilitation à diriger des Recherches, Université Toulouse Le Mirail. Porhiel, S. (2007). « Les structures énumératives à deux temps », Revue Romane, 42-1, 103-135. Prévot, L. (2004). Structures sémantiques et pragmatiques pour la modélisation de la cohérence dans des dialogues finalisés, PhD Thesis/Thèse de doctorat, Université Paul Sabatier, Toulouse. Sandström, G. (1993). When-clauses and the Temporal Interpretation of Narrative Discourse. PhD Thesis, University of Umeå, Department of General Linguistics. Schnedecker, C. (1998). « Les corrélats anaphoriques : une entrée en matière », Recherches Linguistiques, 22, 3-36. Turco, G., Coltier, D. (1988). « Des agents doubles de l’organisation textuelle : les marqueurs d’intégration linéaire », Pratiques, 57, 40-60. Vieu, L., Bras, M., Asher, N., Aurnague, M., (2005). « locating adverbials in discourse », Journal of French Language Studies, 15, 173-193. 102
© Copyright 2026 Paperzz