<?xml version="1.0" standalone="yes"?>
<Paper uid="W01-1307">
  <Title>Temporal Information and Discourse Relations in Narratives: the role of French connectives puis and un peu plus tard</Title>
  <Section position="4" start_page="0" end_page="0" type="metho">
    <SectionTitle>
1 Introduction
</SectionTitle>
    <Paragraph position="0"> It is well known that tense, grammatical and predicational aspect are very important clues in the process of temporal interpretation of sentences, and of texts (see (Kamp and Rohrer, 1983) among others). As far as discourse level is concerned, it has been shown that world knowledge and knowledge about discourse structure play a role in determining the text global temporal structure (Lascarides and Asher, 1993; Caenepeel and Moens, 1994). In this paper we want to focus on the role of connectives and their intersentential semantics.</Paragraph>
    <Paragraph position="1"> We want to give an account of the way some temporal connectives in French interact with discourse relations. We will examine differences in the behavior of two connectives, namely puis (then, afterwards) and un peu plus tard (a bit later), when different discourse relations hold. At first sight, the temporal contribution of these two connectives should show little or no difference.</Paragraph>
    <Paragraph position="2"> We will see that matters are not so simple.</Paragraph>
    <Paragraph position="3"> This direction of work is inspired by a study of puis showing differences in behaviour according to the discourse relation at stake (Bras et al., 2001). It is also grounded on previous studies of adverbials of temporal location and adverbials of spatial location and of their role to build the spatio-temporal structure of discourse (Asher et al., 1994; Asher et al., 1995a; Asher et al., 1995b; Asher et al., 2001). We showed that, in the context of trajectories, relational adverbials like un peu plus tard and its spatial counterpart un peu plus loin (a bit further) had symetric roles in the process of locating the eventualities in space and time, and could both be given the same spatio-temporal interpretation.</Paragraph>
    <Paragraph position="4"> We will tackle a comparative analysis of connectives within the framework of a theory of discourse structure and discourse relations. We have chosen Segmented Discourse Representation Theory (SDRT) for its effective methodology in investigating and formalizing how different linguistic clues interact at the semantic/pragmatic interface to recover the discourse structure of a text. We will first briefly present SDRT (section 2), and the Discourse Relations (section 3) that we will use for our contrastive analysis of puis and un peu plus tard with respect to Discourse Structure (section 4).</Paragraph>
  </Section>
  <Section position="5" start_page="0" end_page="0" type="metho">
    <SectionTitle>
2 SDRT
</SectionTitle>
    <Paragraph position="0"> We first present the theorical framework that we chose to account for this interaction, namely SDRT. SDRT is a non trivial extension of DRT that takes discourse structure into account and offers a theory of the semantics/pragmatics interface. We give a brief outline of SDRT (see (Asher, 1993; Lascarides and Asher, 1993; Asher, 1996; Busquets et al., 2001) for a thorough presentation).</Paragraph>
    <Paragraph position="1"> In SDRT a discourse is represented by an SDRS.</Paragraph>
    <Paragraph position="2"> An SDRS is a recursive structure consisting of elementary DRSs (i.e., DRSs representing a single clause) and 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. We use Greek letters as variables for constituents, and a0a2a1 as constants.</Paragraph>
    <Paragraph position="3"> SDRSs are built up incrementally. SDRT defines a &amp;quot;Glue Logic&amp;quot; and an &amp;quot;Update Function&amp;quot; that together determine a new SDRS for a given SDRS a3 representing the context (the discourse already processed), and a new constituent a4 representing the information to be integrated into that context.</Paragraph>
    <Paragraph position="4"> The Glue Logic is embedded in the framework of &amp;quot;Commonsense Entailment&amp;quot; (Asher and Morreau, 1991), a logic that exploits both monotonic (a5 ) and non-monotonic (a6 )1 conditionals. The Glue Logic is specified by: a7 definitions characterizing which constituents in a3 are open for attaching a4 , a7 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 a4 to some open constituent of a3 , a7 axioms specifying the semantic effects of those discourse relations.</Paragraph>
    <Paragraph position="5"> We will see some of these two kinds of axioms below. The Update Function is in charge of the 1a8a10a9a12a11 means &amp;quot;if a8 then normally a11 &amp;quot;. From a8a13a9a14a11 and a8 , Commonsense Entailment entails a11 &amp;quot;by default&amp;quot;, that is, defeasibly, in the absence of further information regarding the truth value of a11 . From a8a15a9a12a11 , a8 and a16 a11 , Commonsense Entailment no longer entails a11 , but a16 a11 .</Paragraph>
    <Paragraph position="6"> proper hierachization of the structure and of the resolution of the possibly existing underspecifications (e.g., anaphora and ellipses).</Paragraph>
  </Section>
  <Section position="6" start_page="0" end_page="0" type="metho">
    <SectionTitle>
3 Discourse Relations
</SectionTitle>
    <Paragraph position="0"> In the following, we only present the discourse relations used in SDRT that will be helpful to analyse our data: Narration and Result. One important aspect of SDRT that is worth emphasizing at this point is that several discourse relations may simultaneously link the same two constituents.</Paragraph>
    <Section position="1" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
3.1 Narration
</SectionTitle>
      <Paragraph position="0"> Narration is a relation which is based on the Gricean pragmatic constraint of orderliness.</Paragraph>
      <Paragraph position="1"> When two clauses are linked by Narration, they describe in sequence two successive events &amp;quot;of the same story&amp;quot;.</Paragraph>
      <Paragraph position="2"> Let us first see more precisely what the semantics effects of Narration on discourse content are. Building on the previous analyses of (Lascarides and Asher, 1993), we showed in (Asher et al., 1995a) that a relation of Narration between two constituents a4 and a17 entails a temporal overlap between the resulting state of the main eventuality of a4 (noted a18a20a19 2), a21a23a22a25a24a27a26a29a28a30a18a31a19a33a32 , and the preceding state of a18a35a34 , a21a23a36a25a18a37a28a38a18a35a34a37a32 , in the absence of locating adverbials. This is what axiom (A1) expresses: A 1 Narrationa28a30a4a40a39a41a17a42a32a43a5 posta28a38a18 a19 a32a45a44 prea28a30a18a35a34a33a32 (A1) 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, 1996) and observed in earlier work on temporal order in narratives (see for example (Caenepeel, 1989)). This doesn't mean that there should be no interval of time between the two events a18 a19 and a18a27a34 , but rather that no relevant event can occur during this interval. From (A1) and uncontroversial ordering assumptions on events and their pre- and post-states3, we can deduce a relation of temporal precedence between 2As regards the ontology of eventualities, SDRT keeps building on DRT, and assumes the same Davidsonian approach. Two basic aspectual classes are distinguished among eventualities: events and states.</Paragraph>
      <Paragraph position="3"> 3a46a48a47a50a49a52a51a54a53a38a55a56a49a52a57a59a58 a60a54a61a62a49a35a55a56a49a52a57a64a63a66a65a67a49a68a63a66a65a69a60a25a70a29a71a72a53a30a55a56a49a52a57 , where a63a54a65 denotes temporal abutment, as used in DRT, or the &amp;quot;meets&amp;quot; relation as used in Allen's theory (Allen, ).</Paragraph>
      <Paragraph position="4"> the events a18 a19 and a18a27a34 : a18 a19a74a73 a18a27a34 . Actually, to fully capture the &amp;quot;no-significant-gap&amp;quot; constraint, a temporal overlap is too weak. As we will see in Section 4.1, we should guarantee that a21a23a22a25a24a27a26a29a28a30a18 a19 a32 persists (at least) up to the beginning of a18 a34 , and, conversely, that a21a23a36a25a18a37a28a38a18a35a34a33a32 starts when (or before) a18 a19 ends. So we propose here an improved version of (A1), where a75 denotes the &amp;quot;intersection&amp;quot; operator4: A 2 Narrationa28a30a4a40a39a41a17a43a32a48a5 a18 a19a69a76a78a77 a28 posta28a30a18 a19 a32a79a75 prea28a30a18a27a34a37a32a41a32 a76a78a77 a18a35a34 Narration has another semantic effect. It is motivated by the intuition that the elements of a Narration must belong to the &amp;quot;same story&amp;quot;, i.e., they must have some common subject matter. To this effect, axiom (A3) expresses that the constituents connected together by Narration must have a common Topic. A 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. If not already present in the context, it has to be added to the SDRS during the update. (A3) and the rules of the underlying logic actually imply that Narration can be non-monotonically inferred only if such a topic exists or can be built.</Paragraph>
      <Paragraph position="5"> A 3 a28a62a80 a3 a39a50a4a81a39a72a17a42a82 a83 Narrationa28a30a4a40a39a41a17a42a32a41a32 a5</Paragraph>
      <Paragraph position="7"> where a80 a3 a39a52a4a40a39a41a17a43a82 means that a17 is to be attached to a4 in the SDRS a3 , and a87 is a subordinating discourse relation whose semantics essentially involves subsumption between the topic and the elements of the narrative it summarizes.</Paragraph>
      <Paragraph position="8"> Now, how do we infer Narration? Since (Lascarides and Asher, 1993), the triggering axioms for Narration have changed to reflect the fact that Narration is not always a default in narratives (Asher, 1996). It is only a default if no other relation can be inferred, that is, if in the discourse there are no clues that other axioms could exploit to infer other discourse relations: A 4 a28a62a80 a3 a39a50a4a81a39a72a17a42a82 a83 a90 Clues_ a91a13a92a20a28 a3 a39a52a4a40a39a41a17a43a32 a83</Paragraph>
      <Paragraph position="10"> 4We assume that a71a50a99a101a100a10a71a103a102 , when applicable, yields a new state lasting the maximum interval of time during which both a71a50a99 and a71a103a102 hold, and whose propositional contents is the conjunction of a71 a99 and a71 a102 's propositional contents.</Paragraph>
      <Paragraph position="12"> where a91a13a92a45a96a20a96a20a96a106a91a64a104 are all the discourse relations used in SDRT5 but Narration, and Clues_ a91 a1 a28 a3 a39a50a4a81a39a72a17a42a32 holds whenever the propositional content of a4 and a17 and the discourse structure of a3 contain clues that could be exploited for inferring a91 a1 a28a30a4a40a39a41a17a42a32 .</Paragraph>
      <Paragraph position="13"> In addition, Narration can be non-monotonically inferred if the predicate Occasion holds between the clauses to be related: A 5 a28a103a80 a3 a39a52a4a40a39a41a17a43a82a95a83 Occasiona28a38a4a81a39a72a17a42a32a41a32a40a6 Narrationa28a30a4a81a39a72a17a42a32 Occasion holds if the two clauses contain clues indicating that their main eventualities are of types that may belong to &amp;quot;the same story&amp;quot;. In other words, Occasion exploits lexical semantics and shared knowledge in terms of scripts connecting certain event types in sequences in which one  event &amp;quot;naturally&amp;quot; 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.</Paragraph>
      <Paragraph position="14"> (1) Paul frappa a la porte. Il entra. (Paul  knocked at the door. He entered.) We can assume this kind of script to be encoded in the following axiom: A 6 a28a103a107 knocka28a38a18 a19 a39a52a108a109a39a52a110a86a32 a83 door-ofa28a30a110a23a39a50a111a54a32a106a112a113a4 6 a83 a107 entera28a30a18a35a34a23a39a50a108a45a39a50a111a54a32a106a112a114a17a42a32a43a5 Occasiona28a30a4a81a39a72a17a42a32 In this case, the sequence of events is only typical, i.e., &amp;quot;natural&amp;quot;: knocking isn't a necessary precondition to enter, and it doesn't cause the entering. Obviously, stronger dependence links between event types like precondition and cause also give rise to Occasion (and by non-monotonic inference, Narration), but some of them are also exploited to infer more specific discourse relations, like the next one to be presented here, Result. null 5SDRT explicit states that there should be a finite number of discourse relations, even though what these are precisely is not a settled matter yet. For the purposes of the present work, we will consider that these relations are: Narration, Background, Elaboration, Continuation, a115 (i.e., &amp;quot;Topic&amp;quot;), Result, Explanation, Contrast and Parallel, for which an SDRT account can be found in the literature.</Paragraph>
      <Paragraph position="15"> 6a116a8a118a117a120a119 means that the condition a8 appears in the constituent a119 .</Paragraph>
    </Section>
    <Section position="2" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
3.2 Result
</SectionTitle>
      <Paragraph position="0"> The Result relation has the semantic effect of implying a causal link between the main eventualities of the constituents it relates:  Result may be monotonically inferred on the basis of the presence in a17 of an explicit marker of causation as e.g., the conjunct donc (therefore) or the verb to result: A 9 a28a62a80 a3 a39a50a4a81a39a72a17a42a82a95a83a122a107 donca112a114a17a42a32a48a5 Resulta28a30a4a40a39a41a17a42a32 Result can also be non-monotonically inferred on the basis of lexical semantics or of some shared knowledge on the types of eventualities in a4 and a17 , as in the following two famous examples:  (2) Max poussa John. Il tomba. (Max pushed John. He fell.) (3) Paul eteignit la lumiere. Il faisait nuit  noire autour de lui. (Paul turned off the light. It was pitch dark around him.) In (2) and (3), it is again generic script-like information on pushing and falling event types, and on switching off the light event types and being dark state types that enables the reader to recover the causal links that the narrator most likely wanted to express. The presence of such clues indicating a possible causal link is expressed by the predicate D-Permissible-Cause. For instance, we assume that the following axiom encodes a plausible piece of shared knowledge:</Paragraph>
    </Section>
  </Section>
class="xml-element"></Paper>