Levels of representation in discourse relations - Semantic Scholar

Levels of representation in discourse relations ALISTAIR KNOTT, TED SANDERS, and JON OBERLANDER*

1. Cognitive linguistics and discourse Most schools of linguistics have their foundations in a study of the syntax and semantics of individual sentences. Cognitive linguistics is to some extent no exception: while it has always devoted considerable attention to the conceptual aspects of linguistic meaning and to the study of language in use, the focus of research has still traditionally been on single sentences or propositions rather than on larger units of text and discourse. However, within many schools of linguistics, a signi®cant body of work on discourse structure has developed over the last ®fteen years or so. Discourse-oriented studies in cognitive and functional linguistics include research on the discourse function of syntactic constructions such as subordinate clauses (Tomlin 1985; Matthiessen and Thompson 1987; Ramsay 1987) as well as on the need for functional and usage-based analyses of discourse (Langacker 2001; Barlow and Kemmer 2000). Discourse-oriented work in formal linguistics centers around the in¯uential dynamic semantic theories of Kamp and Heim (see, for example, Kamp and Reyle 1993), Heim (1982), and latterly Veltman (1996), which focus on issues relating to anaphora and presuppositions. In computational linguistics, there is a growing body of work on the interpretation and generation of extended texts; see, for instance, Lascarides and Asher (1993) or Strube and Hahn (1996) for discourse interpretation, and Hovy (1993) or Marcu (1997) for discourse generation. Interestingly, at the discourse level, the dividing line between cognitive linguistic approaches and traditional approaches seems less clear cut than at the sentence level. Many of the controversial proposals in cognitive linguistics concern the status of sentence syntax, in particular the question of whether syntax is an autonomous and purely formal level of representation (cf. Langacker 1986; Jackendo€ 1996). However, it is widely accepted that purely formal or syntactic principles play a far smaller role Cognitive Linguistics 12±3 (2001), 197±209

0936±5907/01/0012±0197 # Walter de Gruyter

198 A. Knott, T. Sanders, and J. Oberlander at the discourse level (see for example Halliday and Hasan 1976): it is hard, for instance, to make much sense of the idea of a structurally ``well-formed'' but semantically anomalous discourse. There is a consensus that the well-formedness of a piece of discourse is primarily to do with its meaningÐspeci®cally, to do with whether the meanings of its component sentences can be related together to form a coherent message. In formal and computational discourse linguistics, this point emerges in the emphasis on the role of world knowledge in discourse processing (see, e.g., Hobbs 1985; Hobbs et al. 1993; Lascarides and Asher 1993). In cognitive linguistics, a similar point emerges from the suggestion that a coherent text must emerge from, and lead to, a coherent cognitive representation (Sanders et al. 1992). In short, for research on discourse structure, there is considerable scope for the integration of work in cognitive linguistics with that from other traditions within linguistics. Purely from the perspective of cognitive linguistics, there are two further interesting reasons why the study of discourse is of particular relevance. Firstly, there is the issue of categorization, which has a long tradition in cognitive linguistics. Human beings categorize the world around them, usually unconsciously. The linguistic categories apparent in people's everyday language use have provided many interesting insights into the working of the mind (see, for instance, Lako€ 1987; Lako€ and Johnson 1999). In (text) linguistic approaches to discourse, the categorization of di€erent types of structural and semantic relations, and of the linguistic devices expressing them (connectives, lexical phrases) has played a prominent role over the last decade. For instance, the way in which speakers categorize related events by expressing them with one connective (because) rather than another (since) can be treated as an act of categorization that reveals language users' ways of thinking. On the basis of careful meaning-oriented analyses, it has been argued that the meaning and use of such connectives can be explained in terms of domains of use (Sweetser 1990) or di€erent sources of coherence (van Dijk 1977; Sanders et al. 1992; Knott and Dale 1994). This particular topic plays a major role in this special issue, because several contributions aim at rede®ning these existing distinctions, making use of constructs like subjectivity (following Langacker 1990) and perspective (Sanders and Spooren 1997). Secondly, one of the main concerns of cognitive linguistics is to develop integrated theories of language structure and language processes. Cognitive linguistics has already found ways to do just that: by assuming that structures and processes should be treated on a par, instead of separately (cf. Langacker 1987). The level of discourse readily presents itself as the ideal level in the study of language at which to further ¯esh

Levels of representation in discourse relations 199 out this claim (Sanders 1997b). A crucial characteristic of discourse is that it shows coherence. Relational coherence is one fundamental type of coherence: people connect discourse segments by inferring coherence relations of some kind, be it causal, contrastive or additive. Although coherence is generally considered a cognitive phenomenon, relatively independent of the exact linguistic realization in the discourse itself, both linguists and psycholinguists assume that connectives have the function of signaling relationships between discourse segments, thereby ``instructing'' interlocutors to construct a coherence relation between two clauses (see, among others, Gernsbacher and GivoÂn 1995; Noordman and Vonk 1997; Sanders and Spooren 2001). The study of linguistic markers of coherence from such a cognitive point of view leads to an account of language as a processing instructor. This holds not just for the study of connectives, as we will show, but also for other coherence markers, such as those of referential coherence. The grammar of referential coherence plays an important role in mental operations connecting incoming information to existing mental representations. For instance, on the basis of their linguistic realization, referent NPs are identi®ed as either those that will be important and topical, or as those that will be unimportant and nontopical. Hence, topical referents are persistent in the mental representation of subsequent discourse, whereas the nontopical ones are not (see Ariel 2001; Chafe 1994; GivoÂn 1995). As for relational coherence, there are many similar examplesÐbased on corpus studiesÐof this line of work in the functional linguistic literature, for instance on the tendency subordinated clauses exhibit for containing background information (see, for example, Tomlin 1985), or on the discourse function of if/when-clauses (Haiman 1992; Ramsay 1987) and purpose clauses (see Thompson 1985; Matthiessen and Thompson 1987 on in order to). The discourse function of purpose clauses appears to depend on their placement in relation to the main clause. In medial or ®nal position their role is a local elaboration, but in initial position their role becomes one of foregrounding information. They signal how to interpret the directly following clause and how to attach it to the preceding text. 2. Discourse relations In both cognitive linguistics and other traditions, an important building block for theories of discourse coherence is the notion of a discourse relation (also known in the literature as a coherence or rhetorical relation). Discourse relations are meaning relations which connect two text segments (minimally, single clauses). Examples are relations like CAUSE±CONSEQUENCE, LIST, and PROBLEM±SOLUTION. Such relations can be

200 A. Knott, T. Sanders, and J. Oberlander explicitly signaled by linguistic markers such as sentence or clause connectives; alternatively they can be left unsignaled for a reader or hearer to infer. Whether they are signaled or not, the key intuition in a theory of relations is that the coherence of a text can be attributed, in whole or in part, to the presence of appropriate relations between its constituent sentences and clauses. Of course, the proposal that a text is coherent in virtue of its discourse relations is vacuous until further information about the nature of these relations is provided. Firstly (and most obviously), a theory of relation semantics is required, which speci®es what kind of relations constitute discourse relations. There is a relation of some kind between every conceivable pair of sentences or clauses, even those whose juxtaposition does not result in a coherent segment of text: we are interested in identifying those relations whose presence between two juxtaposed sentences or clauses ensures or contributes to the coherence of the resulting complex text span. Secondly, a theory of span structure is needed, which indicates whereabouts in a coherent text discourse relations might be expected to be found. Must there be a discourse relation between every pair of adjacent sentences, for instance? How is coherence ensured between text units larger than sentences? Thirdly, a theory of relation signaling is needed, indicating how discourse relations manifest themselves in surface text, particularly in the form of sentence and clause conjunctions. Theories of relation signaling range from those which advocate a complete dissociation of discourse relations from conjunctive signals (cf. Mann and Thompson 1988) to proposals that the set of conjunctions tells us fairly directly about the set of discourse relations (cf. Knott 1996; Knott and Sanders 1998). Other researchers are interested in the semantics of conjunctions as a topic in its own right, and are not primarily concerned with the role they play in a theory of discourse coherence (cf. Halliday and Hasan 1976; Martin 1992). Much of the utility of the concept of a discourse relation derives from the way it sets up these three more precise research questions, and thus helps to structure the task of developing a theory of discourse coherence. The articles in this special issue focus predominantly on issues concerning relation semantics. Most of the texts studied in the articles presented in this issue are relatively small, typically pairs of sentences or clauses ( perhaps with some preceding or subsequent context), allowing the study of a single intersentential/interclausal relation. The assumption is that relations between adjacent sentences or clauses will constitute a key component of any theory of span structure, and that they are therefore worth studying in isolation. The articles presented here also take a similar perspective on relation signaling, focusing for the most

Levels of representation in discourse relations 201 part on discourse relations signaled explicitly by sentence or clause connectives, either because of an explicit interest in the semantics of connectives, or because of a decision to use connectives as an empirical window onto discourse relations. 3. A parameterization of discourse relations In the last decade, much research in relation semantics has focused on the question of how to taxonomize or classify the set of discourse relations (Hovy 1990; Knott and Dale 1994; Pander Maat 1998; Redeker 1990; Sanders 1997a). While work on the hierarchical classi®cation of discourse relations goes back at least as far as Ballard, Conrad, and Longacre (1971), Grimes (1975), and Halliday and Hasan (1976), the idea that a small number of reasonably orthogonal primitives is responsible for the differences amongst discourse relations is more recent, deriving from Hobbs (1985) and Sanders, Spooren, and Noordman (1992, 1993). The articles in this special issue all adopt this assumption to some degree. To get a feel for some of the primitives which underlie discourse relations, consider the following short texts: (1) (2) (3) (4)

Gareth gets on well with Betty. He loves disco music; she loves it too. Gareth and Edith have di€erent musical tastes. He loves disco music, while she despises it. Gareth grew up during the seventies, so he loves disco music. Edith grew up during the seventies, but she despises disco music.

The relations in examples (1) and (2) have something in common because they both make comparisons between two individuals. The relations in examples (3) and (4) can also be argued to have something in common, because they both rely on a piece of presupposed knowledge, namely that people who grew up during the seventies are likely to love disco music. Note also that the di€erence between examples (1) and (2) resembles the di€erence between examples (3) and (4). If you negate the last clause in examples (1) and (3), the resulting relations are the same as those in examples (2) and (4), respectively, and the conjunctions must be changed to restore coherence. We can hypothesize that the relations in these examples are composite constructs, de®ned in terms of two primitive bivalued parameters: one parameter might distinguish between COMPARATIVE and CAUSAL relations, and the other might distinguish between POSITIVE and NEGATIVE relations. Elaborating this simple account of relation semantics involves framing suitable de®nitions for these parameters, and for additional parameters, in a way that allows the de®nitions

202 A. Knott, T. Sanders, and J. Oberlander to interact so as to produce an appropriate range of composite relation de®nitions. A dominant distinction in existing classi®cation proposals is that between so-called SEMANTIC relations (also known, to some degree of approximation, as IDEATIONAL, EXTERNAL or CONTENT relations) and PRAGMATIC relations (also known as PRESENTATIONAL or INTERNAL relations). SEMANTIC relations, informally speaking, hold between the propositional content of the two related discourse segments, i.e. between their locutionary meanings; PRAGMATIC relations, on the other hand, involve the illocutionary meaning of one or both the related segments. To illustrate, consider the following examples: (5) (6)

The neighbors left for Paris last Friday. So they are not at home. The lights in the neighbors' living room are out. So they are not at home.

The relation in (5) can be interpreted as SEMANTIC because it connects two events in the world; our knowledge allows us to relate the segments as coherent in the world. A relation like that in example (5) could be paraphrased as ``the cause in the ®rst segment (S1) leads to the fact reported in the second segment (S2)''. In example (6), however, the two discourse segments are related because we understand the second part as a conclusion from evidence in the ®rst, and not because there is a causal relation between two states of a€airs in the world: it is not because the lights are out that the neighbors are not at home. The causal relation in example (6) could be paraphrased as ``the description in S1 gives rise to the conclusion or belief formulated in S2''. Hence the relation in this example is PRAGMATIC. The use of the connective so in both examples suggests that it can operate at di€erent levels of representation within a discourse, indicating a causal relation either at the level of propositional content or at the level of speaker intentions or illocutionary force. The distinction between SEMANTIC and PRAGMATIC levels of representation is noted by many students of discourse coherence; however, there is a great deal of discussion about the exact de®nition of the distinction (see, e.g., Bateman and Rondhuis 1997; Degand 1996, 1992; Hovy 1990; Knott 1996, 2001; Knott and Sanders 1998; Martin 1992; Moore and Pollack 1992; Oversteegen 1997; Pander Maat 1998; Sanders 1997a; Sanders and Spooren 1999). There is also considerable interest in the question of what work the distinction can do in other areas of natural language semantics. In particular, within cognitive linguistics, there are apparent connections with models of the semantics of modal verbs (cf. Sweetser 1990), perspective (cf. Sanders and Redeker 1996; Sanders and Spooren 1997), and mental spaces (cf. Fauconnier 1994; Fauconnier

Levels of representation in discourse relations 203 and Sweetser 1996; Dancygier and Sweetser 2000; Verhagen 2000). Alongside these discussions about SEMANTIC and PRAGMATIC relations, there are parallel discussions relating to the de®nitions of other discourse relation parameters, and to the relevance of these other parameters to other topics in natural language semantics (see for instance Pander Maat 1999). The articles in this special issue have their origins in papers presented at a workshop held at the University of Edinburgh in July 1999 on ``Levels of Representation in Discourse''. The workshop addressed the general question of how to derive a set of parameter de®nitions for discourse relations. A number of issues were subsumed under this general topic, in particular the following: ± ± ±

What are the relevant parameters in terms of which discourse relations can vary, and how should they be de®ned? How can empirical evidence be sought for a proposed parameter or set of parameters? What relation do the parameters of discourse relations have to other issues in linguistic semantics?

The workshop was interdisciplinary in its conception, with participation from cognitive linguists, computational linguists, formal logicians and psycholinguists. The articles in this issue re¯ect the issues which arose most prominently during the workshop. The articles by Pander Maat and Degand and by Pander Maat and Sanders squarely address the question of the de®nition of SEMANTIC and PRAGMATIC relations. The main proposal by Degand and Pander Maat is that SEMANTIC and PRAGMATIC causal relations should be seen as points on a continuum of relation types, rather than as alternative values for a discrete bivalued (or trivalued) parameter. The continuum is given by the degree of speaker involvement in the causal relation being expressed. Fleshing out this claim involves a reanalysis of existing de®nitions of SEMANTIC and PRAGMATIC. Evidence for the claim is sought using a corpus study of the distribution of Dutch and French connectives. Making use of further corpus data, Pander Maat and Sanders argue that the distribution of the Dutch connectives daarom `that's why' and dus `so' cannot be explained in terms of Sweetser's domains, or in terms of Sanders, Spooren, and Noordman's ``source of coherence'' parameter. Searching for the commonalities between volitional and epistemic relations, they argue for the relevance of the notions of subjectivity (Langacker 1990) and perspective (as used in, e.g., Sanders and Spooren 1997). Additional evidence is found in an experimental study of the connectives generated by subjects in a range of discourse contexts.

204 A. Knott, T. Sanders, and J. Oberlander An alternative reanalysis of the SEMANTIC/PRAGMATIC distinction is given in Jayez and Rossari. This article examines a range of conjunctive markers of causal discourse relations in French, from the perspective of formal semantics: accordingly, the relation de®nitions put forward in the article are motivated by their descriptive adequacy in a formal model of discourse, rather than by empirical experiment. The starting point for the article is the familiar fact that conjunctions do not always signal relations between the propositional contents of the clauses they link; its aim is to characterize the range of semantic objects underlying clauses (belief states, speech acts, etc.) between which relations can hold. To begin with, the article presents a formal model of belief states, and gives an analysis of the role of epistemic modals such as must in argumentative relations. This formal account allows a subtle analysis of the ®nding that some causal conjunctions (e.g., de ce fait, du coup), while holding between belief states, cannot be used to signal a conclusion reached by abductive reasoning from e€ect to cause, even though abductive reasoning, just like any reasoning, can be described in terms of ordinary causal relations between belief states. This analysis is then extended to account for relations involving speech acts. Again, the formal speci®cations of the semantics of conjunctions permit some subtle distinctions to be captured. The article by Louwerse is concerned not with a single parameter of variation for discourse relations, but with a set of parameters. The emphasis in the article is not on the SEMANTIC/PRAGMATIC distinction, but on distinctions between CAUSAL, ADDITIVE, and TEMPORAL relations, POSITIVE and NEGATIVE relations, and FORWARD, BACKWARD, and BIDIRECTIONAL relations. A novel source of information about the psychological validity of these parameters is used: studies of the eye movements of subjects reading texts containing relations signaled by sentence connectives. Finally, the article by Oberlander and Moore focuses on the interaction of two parameters: the order of presentation of the two spans in a discourse relation, and the placement of conjunctive signals in the ®rst or second span. This research is at the interface between corpus linguistics and psycholinguistics, but also draws on research in computational treatments of discourse. Empirical evidence about span order and cue placement is recruited from two sources: corpus analysis and psycholinguistic experiments involving reading times. Certain orders and placements are found to be harder to process; surprisingly, however, they are also relatively common, and often preferred over easierto-process alternatives. The explanation proposed for this ®nding concerns the interaction of information structure with span order and

Levels of representation in discourse relations 205 cue placement, and the issue of the level of hierarchy at which a relation appears in a text. 4. Directions for interdisciplinary work on discourse relations The articles in this special issue address an overlapping set of research questions about the semantics of discourse relation parameters, from a range of di€erent theoretical perspectives. There is a reasonable consensus across the articles about the interesting phenomena to be explained: in particular, the semantic/pragmatic distinction, or variants of it, and issues relating to the order of spans in discourse relations. However, there is not as yet any great consensus about how these phenomena should be modeled; it will be interesting to see whether further crossdisciplinary e€orts will strengthen this consensus. A number of concrete questions can be posed. For instance, can a formal account of the semantics of discourse relations such as that given by Jayez and Rossari deal with Pander Maat and his coauthors' suggestion that certain dimensions along which relations vary are continuous rather than discrete? Conversely, can the de®nitions of relation parameters used by psycholinguistic and cognitive linguistic researchers be formalized in a logical framework such as that given by Jayez and Rossari? To what extent are the de®nitions we propose for relations artefacts of the experimental techniques we use, or of the formal languages we adopt? Hopefully these questions will be addressed in future interdisciplinary work on discourse relations. Received 4 October 2001

Universities of Otago, Utrecht, and Edinburgh

Note *Author's e-mail addresses: [email protected]; Ted [email protected]; J.Oberlander@ ed.ac.uk.

References Ariel, Mira 2001

Accessibility theory: An overview. In Sanders, T., J. Schilperoord, and W. Spooren (eds.), Text Representation: Linguistic and Psycholinguistic Aspects. Amsterdam: Benjamins. Ballard, D. Lee, Robert Conrad, and Robert Longacre 1971 The deep and surface grammar of interclausal relations. Foundations of Language 4, 70±118.

206 A. Knott, T. Sanders, and J. Oberlander Barlow, Michael and Suzanne Kemmer 2000 Usage-based Models of Language. Stanford: Center for the Study of Language and Information. Bateman, John A. and Klaas Jan Rondhuis 1997 ``Coherence relations'': Towards a general speci®cation. Discourse Processes 24(1), 3±49. Chafe, Wallace 1994 Discourse, Consciousness and Time. The Flow and Displacement of Conscious Experience in Speaking and Writing. Chicago: Chicago University Press. Dancygier, Barbara and Eve Sweetser 2000 Constructions with if, since and because: Causality, epistemic stance and clause order. In Couper-Kuhlen, Elisabeth and Bernd Kortmann (eds.), Cause, Condition, Concession and Contrast. Berlin/ New York: Mouton de Gruyter, 111±142. Degand, Liesbeth 1992 Het ideationele gebruik van want en omdat. Nederlandse Taalkunde 4, 309±326. 1996 A situation-based approach to causation in Dutch with some implications for text generation. Unpublished Ph.D. dissertation, Universite Catholique de Louvain, Belgium. Fauconnier, Gilles 1994 Mental Spaces: Aspects of Meaning Construction in Natural Language. 2nd edition. Cambridge, MA: MIT Press. Fauconnier, Gilles and Eve Sweetser (eds.) 1996 Spaces, Worlds and Grammars. Chicago: University of Chicago Press. Gernsbacher, Morton and Talmy GivoÂn (eds.) 1995 Coherence in Spontaneous Text. Amsterdam/ Philadelphia: John Benjamins. GivoÂn, Talmy 1995 Coherence in Text vs. Coherence in Mind. Amsterdam/ Philadelphia: John Benjamins, 59±115. Grimes, Joseph E. 1975 The Thread of Discourse. The Hague: Mouton. Haiman, John 1992 Conditionals are topics. Language 54, 564 ±589. Halliday, Michael A. K. and Ruqaiya Hasan 1976 Cohesion in English. London: Longman. Heim, Irena 1982 The semantics of de®nite and inde®nite noun phrases. Ph.D. thesis, University of Massachusetts. Distributed by Graduate Linguistic Student Association. Hobbs, Jerry R. 1985 On the Coherence and Structure of Discourse (Technical Report CSLI-85-37). Stanford: Center for the Study of Language and Information. Hobbs, Jerry, Mark Stickel, Douglas Appelt, and Paul Martin 1993 Interpretation as abduction. Arti®cial Intelligence 63(12), 69±142. Hovy, Eduard E. 1990 Parsimonious and pro¯igate approaches to the question of discourse structure relations. In Hovy (ed.), Proceedings of the 5th International Workshop on Natural Language Generation. Pittsburgh, 128±135.

Levels of representation in discourse relations 207 1993

Automated discourse generation using discourse structure relations. Arti®cial Intelligence 63, 341±385. Jackendo€, Ray 1996 Conceptual semantics and cognitive linguistics. Cognitive Linguistics 7(1), 93±129. Kamp, Hans and Uwe Reyle 1993 From Discourse to Logic. Dordrecht: Kluwer. Knott, Alistair 1996 A data-driven methodology for motivating a set of coherence relations. Unpublished Ph.D. thesis, Department of Arti®cial Intelligence, University of Edinburgh. 2001 Semantic and pragmatic relations and their intended e€ects. In Sanders, Ted, Joost Schilperoord, and Wilbert Spooren (eds.), Text Representation: Linguistic and Psycholinguistic Aspects. Amsterdam: Benjamins, 127±152. Knott, Alistair and Robert Dale 1994 [1992] Using linguistic phenomena to motivate a set of coherence relations. Discourse Processes 18(1), 35±62. [Also available as Technical Report RP-34, Human Communication Research Centre, University of Edinburgh.] Knott, Alistair and Ted Sanders 1998 The classi®cation of coherence relations and their linguistic markers: An exploration of two languages. Journal of Pragmatics 30, 135 ±175. Lako€, George 1987 Women, Fire and Dangerous Things. Chicago/London: University of Chicago Press. Lako€, George and Mark Johnson 1999 Philosophy in the Flesh. The Embodied Mind and its Challenge to Western Thought. New York: Basic Books. Langacker, Ronald 1986 An introduction to cognitive grammar. Cognitive Science 10, 1± 40. 1987 Foundations of Cognitive Grammar 1: Theoretical Prerequisites. Stanford: Stanford University Press. 1990 Subjecti®cation. Cognitive Linguistics 1(1), 5±38. 2001 Discourse in cognitive grammar. Cognitive Linguistics 12(2), 143±188. Lascarides, Alex and Nicholas Asher 1993 Temporal interpretation, discourse relations and common sense entailment. Linguistics and Philosophy 16(5), 437± 493. Mann, William C. and Sandra A. Thompson 1988 Rhetorical structure theory: A theory of text organization. Text 8(3), 243±281. Marcu, Daniel 1997 The rhetorical parsing, summarisation and generation of natural language texts. Unpublished Ph.D. thesis, Department of Computer Science, University of Toronto. Martin, Jim R. 1992 English Text: System and Structure. Amsterdam: Benjamins. Matthiessen, Christian and Sandra Thompson 1987 The structure of discourse and ``subordination''. In Haiman, John and Sandra Thompson (eds.), Clause Combining in Discourse and Grammar. Amsterdam: Benjamins, 275±329.

208 A. Knott, T. Sanders, and J. Oberlander Moore, Johanna D. and Martha E. Pollack 1992 A problem for RST: The need for multi-level discourse analysis. Computational Linguistics 18, 537±544. Noordman, Leo and Wietske Vonk 1997 The di€erent functions of a conjunction in constructing a representation of the discourse. In Fayol, M. and J. Costermans (eds.), Processing Interclausal Relationships in Production and Comprehension of Text. Hillsdale, NJ: Erlbaum, 75 ±94. Oversteegen, Leonoor E. 1997 On the pragmatic nature of causal and contrastive connectives. Discourse Processes 24(1), 51±86. Pander Maat, Henk 1998 The classi®cation of negative coherence relations and connectives. Journal of Pragmatics 30, 177±204. 1999 The di€erential linguistic realization of comparative and additive coherence relations. Cognitive Linguistics 10, 147±184. Ramsay, Violeta 1987 The functional distribution of preposed and postposed ``if '' and ``when'' clauses in written discourse. In R. Tomlin (ed.), Coherence and Grounding in Discourse. Amsterdam: Benjamins, 383±409. Sanders, Jose and Gisela Redeker 1996 Perspective and the representation of speech and thought in narrative discourse. In Fauconnier, Gilles and Eve Sweetser (eds.), Spaces, Worlds and Grammars. Chicago: University of Chicago Press, 290 ±317. Sanders, Jose and Wilbert Spooren 1997 Perspective, subjectivity and modality from a cognitive linguistic point of view. In Liebert, W., G. Redeker, and L. Waugh (eds.), Discourse and Perspective in Cognitive Linguistics. Amsterdam: Benjamins, 85±112. Sanders, Ted 1997a Semantic and pragmatic sources of coherence. On the categorisation of coherence relations in context. Discourse Processes 24(1), 119±147. 1997b Psycholinguistics and the discourse level: Challenges for cognitive linguists. Cognitive Linguistics 8, 243±265. Sanders, Ted and Wilbert Spooren 1999 Communicative intentions and coherence relations. In Bublitz, W., U. Lenk, and E. Ventola (eds.), Coherence in Text and Discourse. Amsterdam: Benjamins, 235±250. 2001 Text representation as an interface between language and its users. In Sanders, Ted, Joost Schilperoord, and Wilbert Spooren (eds.), Text Representation: Linguistic and Psycholinguistic Aspects. Amsterdam: Benjamins, 1±25. Sanders, Ted J. M., Wilbert P. M. Spooren, and Leo G. M. Noordman 1992 Towards a taxonomy of coherence relations. Discourse Processes 15, 1±35. 1993 Coherence relations in a cognitive theory of discourse representation. Cognitive Linguistics 4, 93±133. Strube, Michael and Udo Hahn 1996 Functional centering. In Proceedings of the 34th Annual Meeting of the Association for Computational Linguistics. University of Santa Cruz, 270±277.

Levels of representation in discourse relations 209 Sweetser, Eve 1990 From Etymology to Pragmatics: Metaphorical and Cultural Aspects of Semantic Structure. Cambridge: Cambridge University Press. Thompson, Sandra 1985 Grammar in written discourse: Initial and ®nal purpose clauses in written English. Text 5, 55±84. Tomlin, Russell 1985 Foreground and background information and the syntax of subordination. Text 5, 85 ±122. van Dijk, Teun 1977 Text and Context. London: Longman. Veltman, Frank 1996 Defaults in update semantics. Journal of Philosophical Logic 25, 221±261. Verhagen, Arie 2000 Concession implies causality, though in some other space. In Couper-Kuhlen, Elisabeth, and Bernd Kortmann (eds.), Cause, Condition, Concession and Contrast: Cognitive and Discourse Perspectives. Berlin/ New York: Mouton de Gruyter, 361±380.