Sanders Spooren Noordman 1993 Coherence Relations in a Cognitive Theory of Discourse Representation

Coherence relations in a cognitive theory ofrepresentation

TED J. M. SANDERS, WILBERT P. M. SPOORENand LEO G. M. NOORDMAN

Abstract

Coherence relations such s CauseConsequence and ClaimArgumentestablish the link between two discourse segments. A review of existingaccounts of coherence relations shows that there is no consensus on thenumber of relations neededt nor on their exact nature.

We take coherence relations s cognitive entities that play a central rolein both discourse under-standing and production. We propose a concise theoryof coherence relations, wiih a limited number of cognitively basic concepts.It is hypothesized that readers use their knowledge ofthese basic concepts toinfer the coherence relations. As a result, the sei of relations is organizedint o a taxonomy in terms offour primitives that apply t o all relations, suchs the polar i ty of the relation.

To test the taxonomy, two experiments were carried out in which sentencepairs connected by coherence relations were presented to subjects. Theexperimental items were prototypical examples ofthe relations in the taxon-omy. In one experiment, subjects categorized experimental items on thebasis of the relation connecting the sentence pairs, in the other they labelledthe relations. The results indicate that the 12 classes distinguished in thetaxonomy are intuitively plausible and applicable and that the primitivesunderlying the taxonomy are psychologically salient.

In addition, the taxonomy is presented in the form of an outline of aprocess model that may provide a generally applicable systematic basis foran algorithm of coherence relation under standing. Consequences for suchvarying areas s linguistics, discourse processing, computational linguistics,and language acquisition are discussed.

1. Coherence and coherence relations

Discourse1 is more than a random set of sentences, because it coheres,or rather because people make a coherent representation of it. Discourse

Cognitive Linguistics 4-2 (1993), 93-133 0936-5907/93/0004-0093 $2.00 Walter de Gruyter

Brought to you by | Pontificia Universidad Catolica de ValpaAuthenticated | 158.251.32.82

Download Date | 9/5/12 9:29 PM

94 T.J.M. Sanders, W.P.M, Spooren and L.G.M. Noordman

coheres because among other reasons coherence relations holdbetween the segments (see, among others, Hobbs 1990). A coherencerelation is a means of combining elementary discourse segments which,for the sake of simplicity, we assume to be minimally clauses intomore complex ones. Consequently, a coherence relation is an aspect ofthe meaning of two or more discourse segments which cannot be describedin terms of the meaning of the segments in Isolation. In other words: itis because of this coherence relation that the meaning of two discoursesegments is more than the sum of its parts (see Hobbs 1979: 73 andMann and Thompson 1986: 58 for a similar view).

Hence, coherence relations are conceptual relations that establish thelink between two discourse segments. For example, the relation betweenthe two sentences in (1) is that of Claim-Argument: the second sentencefunctions s an argument for the claim that is expressed in the firstsentence.

(1) Maggie must be eager for apromotion. She's been working late threedays in a row.

Coherence is not a property of the discourse itself but of the representa-tion people have or make of it.2 Writers and Speakers link their ideaswhile producing a coherent discourse and readers and listeners constructa coherent representation of the Information in the discourse. Therefore,in the context of this article, when we say that discourse is coherent, thisphrase is shorthand for the idea that the discourse representation iscoherent. We take coherence relations s cognitive entities that play acentral role in both discourse understanding and discourse production.For the time being, our argument will be stated from a discourse under-standing perspective. Understanding a discourse means constructing acoherent representation of that discourse and a discourse representationis coherent if coherence relations such s Argument-Claim andCause-Consequence can be inferred between the discourse segments.

Note that this view is complementary to that of Halliday and Hasan(1976), who ascribe the texture solely to cohesive ties, that is, explicitlinguistic links such s connectives. Unlike cohesion, coherence does notdepend on the presence or absence of linguistic cues. Coherence relationsmay be either marked or unmarked. For instance, the Claim-Argumentrelation that exists in example (1) does not depend on the presence orabsence of the connective because between the segments. Nevertheless,connectives certainly play a role in guiding the Interpretation of therelation: the coherence relation that i s assumed between the segmentsmust be compatible with the meaning of the connective and with themeaning of the segments.



Coherence relations in discourse representation 95

2. Coherence relations in a theory of discourse representation

In this article we shall investigate the role of Coherence relations in atheory of discourse structure that aims at describing the link between thestructure of a discourse s a linguistic object and its cognitive representa-tion. In our view, coherence relations are an attractive basis for such atheory. As yet, there is no satisfying theoretical account of coherencerelations, although these relations (or very similar concepts) play a keyrole in interdisciplinary research on discourse.

We will illustrate both of these Claims in a concise review of thepertinent literature on the role of coherence relations in three differentfields: text understanding, text analysis, and computational linguistics.

2.1. Coherence relations in text understandingIn research on the effect oftext structure on text understanding, coherencerelations play an important role. There are two relevant experimentalfindings.

The first is that the meaning of the relation between the Segmentsaffects the processing of the segments. For expository text, Van E>ijk andKintsch (1983), Meyer and Freedle (1984) and Horowitz (1987) haveclaimed that differences in text structures cause differences in recall andunderstanding of the text. A good example is the work by Meyer andFreedle (1984), who assume that differences exist in the amount of organ-ization of different text types. One of the more organizing types i sCausation; a less organizing one is Collection. In an experiment, subjectslistened to text passages that differed only in the way in which (groupsof) sentences in the text were related to one another. In a free recall task,subjects were required to write down everything they could remember.One of the results was that recall of the Causation passage was superiorto recall of the Collection passage.3

The various text structures can be defined in terms of coherence rela-tions. For example, "Causation", "Collection" and "ProblemSolution"all resemble the rhetorical relations proposed by Mann and Thompson(1988) and the coherence relations presented by Hobbs (1990) and inSanders et al. (1992): CauseConsequence, List, ProblemSolution.4Despite the fact that some authors have explained these findings some-what differently (cf. Meyer 1985: 28), the Interpretation of these studiess providing evidence for the role of coherence relations in text under-standing is obvious. Understanding a discourse means constructing acoherent representation. As coherence relations play a crucial role in thisrepresentation, different relations will result in different representations.



96 T.J.M. Sanders, W.P.M. Spooren and L.G.M. Noordman

We have conducted reading experiments from which it can be concludedthat the processing of a particular discourse segment depends on the kindof coherence relation between the segment and the preceding discourse.This was demonstrated both in terms of the on-line understanding processand in off-line reproduction (Sanders 1992, eh. 4).

If the structure of a discourse afFects its processing, one may expectthat the linguistic marking of the structure influences the processing swell. This constitutes the second set of experimental evidence. Forinstance, Haberlandt (1982) showed that readers make use of the linguisticmarkers of coherence relations during processing. He used a reading timeparadigm, where he presented to readers sentence pairs in two conditions:the relation between the sentences was either explicitly marked orunmarked. Linguistic marking led to faster processing of the followingdiscourse segment. This is just what one might expect if coherence rela-tions are cognitively real.

2.2. Coherence relations in text analysisText analytic studies aim at descriptive adequacy in analyzing the struc-ture of different kinds of natural texts. In recent years, Mann andThompson have developed a theory of text structure that is closely relatedto the notion of coherence relations s defined above; cf. Mann andThompson (1986, 1988). Their Rhetorical Structure Theory is a descrip-tive framework for the organization of text. Among the relations incorpo-rated in this theory are Solutionhood, Sequence, Contrast and (Non-)Volitional Result. Mann and Thompson's success in analyzing differenttext types indicates that there is intuitive agreement among analystsconcerning the relations that can be identified between the Segments ina text.

Although this is a positive indication for the validity of coherencerelations in a theory of discourse structure, analytic studies do not aimat making explicit Claims concerning the role of coherence relations indiscourse understanding. The relations are considered s an analytic toolrather than s cognitive entities (see Grosz and Sidner 1986).

Nevertheless, s we have just seen, coherence relations play a role indiscourse understanding and therefore can be considered s cognitiveentities. If coherence relations are taken s cognitive entities, they canprovide a basis for a psychologically plausible theory of discourse repre-sentation. However, such a theory should at least generate plausiblehypotheses on the role of discourse structure in the construction of thecognitive representation. The basic problem with existing analytic propos-als is that they do not allow for plausible hypotheses about how a reader




arrives at the Interpretation of a particular coherence relation. The reasonis that it is unclear to what extent the analytic categories reflect humanknowledge: should the categories postulated by analysts be considered scognitive primitives? For example, it might be assumed that readersunderstand a discourse in which clauses are linked by a Claim-Argumentrelation because notions like Claim and Argument are cognitive primi-tives. This is problematic because the lists of coherence relations presentedin the literature are often unorganized and have no principled end, scan be concluded from the work by Hovy (1990), who collected morethan 350 relations proposed in the literature. Consequently, it is not clearhow many primitives must be assumed if there is a limit to this numberat all.

In our view, these problems can be overcome by the use of an organizedset of coherence relations. Such an organization will be worked out inthis article.

2.3. Coherence relations in computational linguisticsFor computational linguistics, coherence relations seem a suitable toolto construct or Interpret coherent discourse (see, e.g., Hovy 1988; Mannand Thompson 1988). However, there is a good deal of discussion aboutthe character and the number of relations needed (see also Hovy 1990and section 5 of this paper).

Some proposals (like that of Mann and Thompson 1988) present setsof 23 or so relations, others propose only two relations (Grosz andSidner, 1986). Although it has become clear that text planners need morethan, for example, two very general intentional relations (Hovy 1988), itis still not clear exactly what is needed. Like Hovy (1990), we claim thatcoherence relations are crucial for an adequate treatment of coherencein natural language generation and understanding and that the only wayout of the existing state of misunderstanding, confusion, and disagreementis to organize the set of possible relations. Hovy's (1990) taxonomy lacksa theoretically founded organization, s he states himself. In this articlewe will present an alternative taxonomy that is theoretically founded andempirically tested.

2.4. Conclusion: Organizing the set of coherence relationsThe conclusion from this overview is that coherence relations are anattractive starting point for a cognitively plausible theory of discourserepresentation provided that such a theory leads to:



98 T.J.M. Sanders, W.P.M. Spooren and L.G.Af. Noordmana) an organized set of coherence relations andb) explicit hypotheses about the way in which a reader understands a

coherence relation.The result should be a parsimonious theory of coherence relations,

with a limited number of cognitively basic concepts. We claim that readersuse their knowledge of these basic concepts to infer the coherence rela-tions. A relation such s ClaimArgument is taken to be composite: i tcan be analyzed in terms of a limited set of more elementary concepts,such s causality. These concepts, from now on called the cognitiveprimitives, are taken to be cognitively basic and apply also to othercoherence relations. The Interpretation of a coherence relation is consid-ered to be a process of checking the primitives. The result of this checkingis the Interpretation of the relation between the discourse Segments.

3. A taxonomy of coherence relations

Our taxonomy of coherence relations (which is described more extensivelyin Sanders et al. 1992) is based upon four primitives. These primitivesare properties of the coherence relations and consequently they are thecriteria for identifying the coherence relations. What distinguishes theseprimitives from other possible candidates is that they concern the rela-tional meaning of the relations. That is, they concern the informationalsurplus that the coherence relation adds to the Interpretation of thediscourse segments in Isolation.

3.1. Four cognitive primitivesIn the following, the primitives will be considered and illustrated byexamples. The examples stem from different types of naturally occuring(Dutch) texts: newspaper articles, advertisements, brochures, and non-specialist books. In each example, the context is explained in parentheses.The examples are in Dutch, followed by an English translation. Somerelations are marked linguistically, e.g., by a connective, others are not.This shows that coherence relations can be either implicit or explicit.

Basic Operation: Additive and causa! relationsThe first primitive concerns the Operation that is to be carried out on thediscourse segments. Two basic operations underly the coherence relations:the additive and the causal basic Operation. These two operations werepostulated because they justify the Intuition that discourse segments are




either weakly connected (addition) or strongly connected (causality)(Sanders et al. 1992).

An additive Operation exists if the relation between the t wo discoursesegments is simply that of logical conjunction (P & Q). A causal Operationexists if an implication relation (P>Q) can be deduced between the t wodiscourse segments. (2) is an example of an additive relationship, (3) ofa causal relationship.(2) (Centraal Beheer is een coperatie die zieh vooral met verzekeringen

bezig houdt.)De omzet bedraagt ongeveer 2,4 miljard glden. In 1988 steeg de winstvan 75 miljoen naar 103 miljoen glden.(Centraal Beheer is a Company dealing especially in insurance.)'The turnover is about 2.4 billion guilders. In 1988 the profitsincreased from 75 million to 103 million guilders.'

(3) (De hele middag en een groot deel van de avond waren de toegangs-weg naar de vertrekhal en een van de grote parkeerterreinenafgesloten.)Het laatste half nur werd ook de voorrijweg naar de aankomsthalafgesloten, zodat er in die tijd niemand via de gewone toegang hetaankomst- en vertrekgebouw in of uit kon.(The whole afternoon and a large part of the evening, the accessroads leading to the departure hall and one of the large parking lotswere closed.)'The last half hour the drive to the arrivals hall was closed, so thatduring that period nobody could leave or enter the terminal.'

In (2) several aspects of an insurance Company are listed. In (3) theconsequence of the fact that the road to the terminal was closed ismentioned in the second segment: nobody was able to enter or leave thebuilding.

Source of Coherence: Semantic andpragmatic relationsThe second primitive is called the Source of Coherence. The two valuesof this primitive are semantic and pragmatic. A relation is semantic ifthe discourse segments are related because of their propositional content,i.e., the locutionary meaning of the segments. For example, the sequencein (4) is coherent because it is part of our world knowledge that runningcauses fatigue. ([4] and [5] are constructed examples.)(4) Theo was exhausted because he had run to the university.(5) Theo was exhausted, because he was gasping for breath.



100 ./.. Sanders, W.P.M. Spooren and L.G.M. NoordmanA relation is pragmatic if the discourse Segments are related because ofthe illocutionary meaning of one or both of the Segments. In pragmaticrelations the coherence relation concerns the Speech act Status of thesegments. In the pragmatic relation (5) the state of affairs in the secondsegment is not the cause of the state of affairs in the first segment, butthe justification for making that utterance.

Note that the pragmatic relation in (5) is based on a "real world link"between a cause (being exhausted) and a consequence (gasping forbreath). This does not mean, however, that dependency on a real worldcausal link between the clauses is a general prerequisite of pragmaticcausal relations. See example (6), in which such a link is absent.(6) Theo was exhausted, because he told me so.Applying the semantic-pragmatic distinction to examples (2) and (3), itcan be concluded that the sequence in (3) is coherent because it is partof our world knowledge that when a road is closed, it cannot be used toget somewhere. In other words, (3) is a semantic relation: the writerdescribes something in the world that is coherent. The same goes for (2).

In the pragmatic relation expressed in (7) the state of affairs in thesecond segment is not the cause of the state of affairs in the first segment.The fact that Dylan* s music did not sound good is not the cause of hisnot being in top form. Rather, the finding that not a note sounded rightcauses saying the first segment. In other words: the second segmentpresents evidence for the claim expressed in the first. The Source ofCoherence is in the writer's communicative acts.(7) (De grote Dylan bestaat al lang niet meer. Spijtig genoeg is die

opmerking niet nieuw; De Meester heeft zijn hoogtepunt lang enbreed gehad.)Dylan kon ook nu de weg bergopwaarts weer niet vinden. Geen nootuit gitaar of mondharmonica klonk fatsoenlijk.(The great Dylan ceased to exist some time ago. Regrettably, thisremark is not news; the Master is definitely past his prime.)Once again, Dylan was far from being in top form. Not a singlenote from his guitar or mouth-organ sounded good.'

Order of the Segments: Basic andnon-basic orderThe third primitive is the Order of the Segments. Given the two basicoperations, two discourse segments can be connected in the basic or thenon-basic order. In (3) the first segment refers to the antecedent of thecausal basic Operation and the second segment refers to the consequent.




This coherence relation is said to display the basic order. The basicOperation underlying the relation in (8) links the antecedent "havingfeathers" with the consequent "being assigned to the Aves class". In (8)the second segment refers to the antecedent in the basic Operation, so therelation displays the non-basic order.(8) (De principes van de indeling zoals die hierboven is geschetst kunnen

het best worden geillustreerd aan de hand van een voorbeeld, zoalsde Kokmeeuw.)De Kokmeeuw is ingedeeld in de klasse van de Aves, omdat hij net alsalle andere vogels veren heeft.(The principles of the classification that was sketched above can beillustrated best by means of an example, such s the black-headedgull.)'The black-headed gull is assigned to the Aves class, because, likeall other birds, i t has feathers.'

As far s the linguistic marking of different relations is concerned, it isclear that, because of their syntactic valency, connectives put specificconstraints on the order in which segments can be realized. For instance,a subordinate conjunction like omdat 'because' can be used to realizeboth basic and non-basic order, whereas the use of the coordinate con-junction want Tor' demands non-basic order: first the claim, then theargument.

As additive relations are logically Symmetrie, the primitive Order ofthe Segments does not discriminate between different additive relations.Additive relations in which difFerences exist in the temporal order of theconnected segments, and which therefore turn out asymmetric, have beenleft aside (see Sanders et al. 1992, section 5.3).

Polarity: Positive and negative relationsThe last primitive with respect to which coherence relations can differ isPolarity, which can be positive or negative. A relation is positive if thet wo discourse segments function directly in the basic Operation. A relationis negative if not the discourse segments themselves but their negativecounterparts function in the basic Operation. The coherence relations in(2)-(8) are of a positive nature and (9) (a constructed example) is anegative counterpart of (8).(9) An ostrich is classified s a bird, although it cannot fly.(9) refers to the instantiation of the basic Operation linking the antecedent"not being able to fly" with the consequent "not being classified s a




bird". The first discourse segrnent expresses the negation of the conse-quent in the basic Operation.

In (10), a negative relation exists, apparently the instantiation of acausal basic Operation, between the antecedent "being very beautiful"and the consequent "being married".5

(10) (Zij was al een legende tijdens haar leven en haar mythe groeidedoor haar volstrekt gei'soleerde bestaan in een flat te New York.In 1951 werd zij Amerikaanse Staatsbrger, drie jaar later kreeg zijeen ere Oscar.)Hoewel Greta Garbo de maatstaf werd genoemd van schoonheid, iszij nooit geirouwd geweest.(She was already a legend during her life and her myth grewbecause of her completely isolated existence in a New Yorkapartment. In 1951 she became an American citizen, three yearslater she received an honorary Oscar.)Although Greta Garbo was called the yardstick of beauty, shenever married.'

The second segment expresses the negation of the consequent in the basicOperation. Positive relations are typically expressed by conjunctions likeand and because, negative relations by conjunctions like but and although.

The four cognitive primitives are combined to generate classes ofcoherence relations. The combination of the four primitives results in ataxonomy in which 12 classes of coherence relations are characterized.An overview of the taxonomy is given in Table l. The relations given foreach class in the taxonomy are taken to be prototypical examples of theclasses they represent. When more than one relation is presented for aclass, the relations within a class differ with regard to a segment-specificcriterion. It is important to stress here that for the moment we are aimingnot for descriptive adequacy but for the identification of the primitivesin terms of which the set of relations can be ordered; our goal is thesystematic classification of coherence relations. The taxonomy is intendeds a contribution to a psychologically plausible theory of discourse repre-sentation, rather than s a descriptively satisfying analytical Instrument.The classification of the relations will be discussed further in section 5,in which the taxonomy is presented s an outline of a flow chart forcoherence relation understanding.

Our central claim is that the coherence relations can indeed be classifiedsystematically in terms of the primitives that produce classes 1-12. Thishypothesis is tested in the rest of this article. As we do not primarily aimat descriptive adequacy, we do not claim that the list of relations inTable l is complete.6 We do claim, however, that the classes of relations




Table 1. Overview of the laxonomy and prolotypical relations

Basic Source of OrderOperation Coherence

Polarity Class Relation

Causa!

CausalCausa!

CausalCausal

CausalCausal

CausalAdditiveAdditive

AdditiveAdditive

Semantic

SemanticSemantic

SemanticPragmatic

PragmaticPragmatic

PragmaticSemanticSemantic

PragmaticPragmatic

Basic

BasicNon-basic

Non-basicBasic

BasicNon-basic

Non-basic

Positive

NegativePositive

NegativePositive

NegativePositive

NegativePositiveNegative

PositiveNegative

laIb23a3b45a5b67a7b89

lOalOb1112

Cause ConsequenceCondition ConsequenceContrastive Cause ConsequenceConsequence-CauseConsequence ConditionContrastive Consequence CauseArgument ClaimCondition ClaimContrastive Argument-ClaimClaim-ArgumentClaim-ConditionContrastive Claim-ArgumentListOppositionExceptionEnumerationConcession

can be the basis for a descriptively adequate set of relations comparableto that of Mann and Thompson (1988). To arrive at such a set, the classescan be further specified using segment-specific properties. One of thecandidate properties for further specification of the classes is "hypotheti-cality", referring to the hypothetical Status of one of the segments in aconditional relation s opposed to a non-hypothetical causal relation (seeLongacre 1983: section 3.4 for the same distinction between Causal andConditional relations). In fact, it is the segment-specific property ofhypotheticality that distinguishes between, for example,Cause-Consequence and Condition-Consequence s relations l a and Ib.The two relations belong to the same class because they do not differwith regard to the four primitives in the taxonomy. The relations in class10 are another example: Exception differs from the prototypicalOpposition relation in the segment-specific property of "specificity". Oneof the segments gives a general Statement and the other a specificStatement.

The examples presented above can be classified s follows in the classesof the taxonomy. (2): a List relation, in Class 9; (3): a Cause-Consequencerelation, in Class 1; (4) and (8): Consequence-Cause relations, in Class 3;(5), (6), and (7): Claim-Argument relations, in Class 7; (9): a ContrastiveConsequence-Cause relation, in Class 4; and (10): a ContrastiveCause-Consequence relation, in Class 2. The taxonomy presented here




differs somewhat from the original proposal (Sanders et al. 1992). First,the conditional relations are classified s examples of both semantic andpragmatic coherence relations, whereas originally they were only classifieds pragmatic. The fact that conditional relations appear s semanticrelations is illustrated by example (11), taken from a bank brochure. In(11) a Consequence-Condition relation (Class 3) is expressed. This is asemantic relation: having a credit balance (second segment) is the condi-tion for the possibility that the owner of the card can take out 500guilders (first segment). The writer describes a possible relationshipbetween two events in the world and the segments are linked because oftheir propositional meaning.(11) (De giromaat biedt rekeninghouders die in het bezit zijn van een

giromaatpas de mogelijkheid om lngs electronische weg geld opte nemen. Het grote voordeel is dat de giromaatpasbezitter inhet algemeen terecht kan buiten de openingstijden van postkantoorof postagentschap.)De giromaatpasbezitter kan per dag via de giromaat een bedrag vanmaximaal f 500, opnemen, mit s het saldo op de rekening vol-doende is.(The automatic teller offers account holders the opportunity totake out money electronically. The big advantage is that theGirocard owner can be served outside the opening hours of thepost office or post agency.)'The Girocard owner can take out up to a maximum of 500 guildersa day, provided that there is a credit balance.'

A second modification of our original proposal concernsGoalInstrument and ProblemSolution structures, which do not figurein the current version of the taxonomy. In a classification experiment(reported in Sanders et al. 1992) we found that Goal-Instrument relationswere often confused with other causal coherence relations likeArgument-Claim, although these relations differ in the Basic Operation.This led us to the conclusion that GoalInstrument relations have to beanalyzed s more complex structures than the other coherence relations,because two basic operations underlie them.7 Because of this property,specific to GoalInstrument and ProblemSolution structures, they areconsidered to be "discourse patterns" (cf. Hoey 1983) that are morecomplex than the coherence relations in the taxonomy (see Sanders inpreparation, for a more detailed analysis). Examples (12) and (13) (froma magazine and from an advertisement, respectively) show thatGoal-Instrument and Problem-Solution should indeed be considered shaving two causal basic operations.




(12) (Er krnen steeds meer voedingsmiddelen op de markt waarinsuiker is vervangen door zoetstoffen. ...)Veel rnensen willen meer weten over al die zoetstoffen. Daarom is ernu een Zoetmiddelen Informatiecentrum opgericht.(In an increasing number of food stuffs sugar has been replacedby sweeteners. ...)'Many people want to know more about all these sweeteners. Thatis why a Sweetener Information centre has been founded.'

The first causal basic Operation underlying the Goal-Instmment patternexpressed in (12) links the state of affairs ("the wish to know more aboutsweeteners") s the antecendent and the action undertaken to bring thatstate about ("an Information Centre has been founded") s the conse-quent. The second basic Operation is that between the action ("anInformation Centre has been founded") and the state of affairs thatresults from that action and that is positively evaluated ("to know moreabout sweeteners").

ProblemSolution patterns are very similar to GoalInstrument pat-terns. The main difference is in the linguistic realization of the "Problem"segment s opposed t o the "Goal" segment: in the first case the state ofaffairs is evaluated negatively, in the second case it is evaluated positively(Sanders in preparation). (13) is a typical example of a Problem-Solution pattern:(13) Een griepje hebben ze zo. Zorg daarom dat je een doosje Sinaspril

in huis hebt.(Dat helpt meteen bij hoofdpijn, kiespijn en pijn bij verkoudheiden griep.)'The flu is easily caught. Therefore, make sure you have a box ofSinaspril.*(It helps immediately in the case of headaches, toothache, andpains caused by colds and flu.)

The first basic Operation is between the negatively evaluated state ofaffairs ("The flu is easily caught") which is taken s negative becauseof flu being a disease and the writer's proposal for action to counteractthis state of affairs ("make sure to have a box of Sinaspril"). The secondbasic Operation can be identified between the proposed action ("havinga box of Sinaspril") and its intended result ("cures the flu").

The appendix to this article gives an Illustration of all the relations inthe current version of the taxonomy (the examples in the appendix wereused in Experiment 1; see Section 4).




3.2. Why Ms taxonomylThe four primitives in the taxonomy concern a property that we thinkcrucial for a viable account of coherence relations: the relational meaning.That is, the relations are classified on the basis of meaning aspects thatare attached to the meaning of the relation rather than to the meaningof the segments that are connected. A remaining question may still be:Why this taxonomy? Three answers to this question may be given:1. It is well-founded. All four primitives are important cognitive cate-gories, prominent in research on language and language behavior. As fars the Basic Operation is concerned, there is much evidence for the specialStatus of causal relations in discourse processing. For instance, Black andBern (1981), Trabasso and Van den Broek (1985) and Van den Broekand Trabasso (1986) have shown that the number of causal connectionsdetermines the importance of a discourse segment in a story.

Next, distinctions similar to our primitive Source of Coherence haveoften been proposed in discourse studies and (text) linguistics; see, forexample, Van Dijk (1979), Mann and Thompson (1988), Redeker (1990),Sweetser (1990), and Takahara (1990).8

The relevance of the order in which Information is presented has beenthe subject of psycholinguistic studies, from which it has become clearthat the iconic order in language in which the order of events corres-ponds to the order of mention facilitates discourse production andunderstanding more than non-iconic order (see Smith and McMahon1970; E. V. Clark 1971; Clark and Clark 1968). It should be noted thaticonic order which is investigated in the studies mentioned - is onlysynonymous to our basic order s far s semantic relations are concerned.Iconicity i s irrelevant for pragmatic relations, whereas Order of theSegments is relevant for both semantic and pragmatic relations. Linguisticevidence for the relevance of the primitive Order of the Segments comesfrom Mby Guarani, a South American Indian language. According t oDooley (1990), who analyzes the coherence relations in this language, thedifference between a pre- and postposed subordinate clause determinesthe type of coherence relation expressed by the Speaker.9

The last primitive, Polarity, is a well-known factor in psycholinguisticliterature: negative utterances are processed more slowly than their posi-tive counterparts (Wason and Johnson-Laird 1972; H. H. Clark 1974).2. Not only did it prove possible to identify four primitives that fulfillthe relational criterion the primitives are meaning aspects attached tothe relation rather than to the segments but the combination of theseprimitives also results in a productive system. After all, the combination




of the primitives generates 12 classes that make sense, i.e., that correspondt o our intuitions about coherence relations.3. There is experimental evidence in favor of the taxonomy. In ouroriginal proposal (Sanders et al. in press) we reported on two experimentsin which the predictions of the taxonomy were tested. It was shown thatrelations that are related in terms of the taxonomy were confused moreoften than relations that are not related. This was found both in aclassification task in which experts labelled the coherence relation con-necting two discourse segments and in a task in which naive subjectschose connectives to complete the segments. In the classification experi-ment, i t was shown that the 12 classes are intuitively plausible andapplicable. Text analysts were asked to label the relations between sen-tence pairs. Judges agreed with our classifications to a considerable extent.For example, of all the instances of what we took to be causal relations,83.5% of the subjects identified the relations s causal relations. Thisagreement was even higher for the values of the other primitives, exceptin the case of Source of Coherence: subjects agreed with our classificationof 65.2% of the cases in which the Source of Coherence was a semanticrelation and of 86. l % of the cases in which the Source of Coherence wasa pragmatic relation. A possible explanation for the relatively lowagreement scores on this primitive was that some of the items in theexperiment were not the clear representatives of the relations that wetook them to be.

In general, the results of the experiment confirmed the hypothesis thatin cases of disagreement the subjects' choices concern relations fromrelated classes rather than unrelated classes and therefore follow thegroupings of the taxonomy.In both experiments, subjects made very few misclassifications. This smallamount of confusion is a problem if one wants to find data on relationsbetween objects. It may be seen s a consequence of the task used in theexperiments: the subjects were not explicitly asked to make comparisonsbetween coherence relations. In the next section an experiment is reportedin which subjects were explicitly asked to make such direct comparisons.A second experiment addresses the question why the primitive Source ofCoherence was not s manifest s was expected.

4. Experimental evidence: Relations between coherence relations

4.1. Experiment 1: Similarities between coherence relationsIn Experiment l subjects compared the coherence relations connectingdiscourse segments. The experimental items were presented in a larger




context. The subjects in the experiment had to sort pairs of discoursesegments on the basis of the similarity of the coherence relations.

4.1.1. MethodThe material for the present experiment consisted of two sets of 17 pairsof discourse segments, embedded in a longer Stretch of discourse. Eachpair was taken to be a clear representative of one of the coherencerelations listed in Table 1. All items were in Dutch.

The first set is given in the Appendix. In the second set the contextwas that of a newspaper article on cases of alleged corruption in the localgovernment of Chicago. Where necessary, the context was changed mini-mally in order to highlight the intended coherence relation. An exampleof a context and a discourse segment pair (expressing a Consequence-Condition relation) is given in (14).(14) Context

In the second largest city of America, Chicago, the tradition ofgivinggifts to voters in return for their vote is still very much alive. At Msmoment a Democratic council member is on trial on the accusationof having given electrician licences and Jobs to voters.Discourse segment lHe can count on a long time in prison,Discourse segment 2if he is found guilty.

The experimental technique of card sorting (Miller 1969) was used. Thecontext and both discourse segments were presented to the subjects oncardboard cards. The discourse segments that were the target of theexperiment were indicated by letters and were printed in italics. Eachsubject saw Set l first and Set 2 second. The order of the items in eachset was randomized, but did not vary between subjects. The subjects wereinstructed to read each card carefully and to base their similarity judg-ments on the coherence relation connecting the two discourse segmentsand not on the content of the discourse segments. The instruction con-tained examples that had similar discourse segments but different coher-ence relations and examples that had different discourse segments butsimilar coherence relations. If the subjects judged two cards s similarwith respect to the coherence relation connecting the discourse segments,they were to put the two cards on the same pile and to write down ashort motivation for this choice. This motivation made it possible tocheck whether a choice had been based on the content of the segmentsor on the coherence relation connecting the segments. The subjects were




free to correct previous classifications. They were also free in the selectionof the number of piles. After the subjects had completed the classificationof the first set, the experimenter wrote down the classification and handedover the second set.

The subjects in the experiment were ten advanced students of discoursestudies. The reason for this choice is that the experimental task is of ameta-linguistic nature. It requires that subjects are used to giving judg-ments about texts and are more or less acquainted with an "analytic"view on text fragments, so that they can actually distinguish between thecontent of the Segments on the one hand and the coherence relation onthe other hand. Subjects did not receive any specific instruction of howto conceive of the "similarity between the coherence relations". Subjectsreceived payment for their participation.

4. l .2. Results and discussionThe data for each set were collected in a 17 by 17 similarity matrix. Therows and columns in the matrix correspond to the experimental items.Each cell in the matrix gives the number of subjects that put a particularitem on the same pile s the other item (minimum 0, maximum 10).Following Shepard (1972) a combined hierarchical cluster analysis (usingWard's method) and a multidimensional scaling analysis (Gemscal) werecarried out for each set. Both analyses give indications concerning therelationships among objects. In a cluster analysis objects that are closelyrelated are clustered at an early stage, whereas loosely related objects areclustered at a later stage. In a multidimensional scaling analysis theobjects are plotted in a geometrical space in such a manner that thedistances between the objects indicate the degree of relatedness: the morerelated the objects, the shorter the distance between them. The MDS-analysis of Set l resulted in a two-dimensional solution with a Kruskalstress of 0.17 (r2 = 0.85),10 The result of the combined analyses for Set lis plotted in Figure 1. There are four distinct clusters in Figure l:11positive causal relations (A, D, G, J), positive additive relations (M, P),negative relations (C, F, I, L, N, O, Q), and conditional relations (B, E,H, K). The two axes in Figure l correspond with t wo primitives in thetaxonomy. The horizontal axis clearly corresponds with the Polarityprimitive.

The vertical axis corresponds to the Basic Operation of the relation,although this Interpretation is somewhat less clear. Following the hypoth-eses, the axis would correspond to the Basic Operation if it distinguishesbetween causal and additive relations. This can indeed be observed bothfor the positive and for the negative relations (see Figure 1). The vertical




1.5

.5

0

-0.5

-l

-1.5

-2

-2 -1.5 -l 0 .5 l 1.5

A Cause-ConsequenceB Condition-ConsequenceC Contrastive-Cause-ConsequenceD Consequence-CauseE Consequence-ConditionF Contrastive Consequence-CauseG Argument-ClaimH Condition-ClaimI Contrastive Argument-ClaimJ Claim-ArgumentK Claim-ConditionL Contrastive Claim-ArgumentM ListN ExceptionO OppositionP EnumerationQ Concession

l 1 2 3 .3 -4 o5 *5 *6 .7 *7 *8 .9 10 010 o 12

Figure 1. Experiment l: Combined multi-dimensional and hierarchical representation ofcoherence relations for dato set l




dimension divides the negative relations into additive and causal relations.As far s the positive relations are concerned, the conditional relationsare on the one end of the dimension and the additive relations on theother. Since conditional relations belong to the causal relations inthe taxonomy, this is in accordance with the prediction. However, it mustalso be concluded that the non-conditional causal relations behave moreneutrally than expected.

The MDS-analysis of Set 2 resulted in a two-dimensional solution witha Kruskal stress coefficient of 0.17 (r2 = 0.88).12 The result of the com-bined analyses for Set 2 is plotted in Figure 2. The same four clustersthat were distinguished in the analysis of Set l show up in Figure 2: thefirst contains the negative relations (C, F, I, L, N, O, Q); the secondcontains the positive additive relations (P, M) plus Claim-Argument (J);the third cluster contains the positive causal relations (A, D, G); thefourth cluster contains the conditional relations (B, E, H, K). Again, thetwo dimensions in Figure 2 correspond with the primitives Polarity(horizontal) and Basic Operation (vertical).

Hence, Experiment l gives very similar results for the two sets of items:the groupings of the items in the two sets are almost identical. Thisprovides evidence for the existence of two primitives in the taxonomy,viz. Polarity and Basic Operation. However, there are also three unpre-dicted findings:1. There is no evidence whatsoever for distinctions on the basis of theprimitive Order of the Segments. In view of the fact that the role of thisprimitive was very evident in our previous experiments (see Sanders et al.1992), we assume that this negative finding does not argue against thisprimitive s a categorizing property. A rather speculative explanationmight be that the present task, in which both Segments are simultaneouslyavailable and have t o be compared with other pairs of segments, detractssubjects from a left-to-right processing of the relations.2. The conditional items B, E, H, and K behave somewhat anomalouslyin both figures. In Figure l conditional relations cluster with additiverelations before they cluster with causal relations. In Figure 2 causalrelations cluster with additive relations before they cluster with condi-tional relations. This deviant clustering is contrary to the prediction, sincethe taxonomy places conditional relations in the same class s causalrelations. Note, however, that conditional relations are on the same poleof the Basic Operation dimension s the causal relations, indicating thatthey are causal rather than additive in nature.3. There is only little evidence for distinctions on the basis of the primitiveSource of Coherence. If a distinction is made between semantic and




1.5

.5

-0.5

-l

-1.5

-2-2

-1.5 -.5 0 .5 1.5

A Cause-ConsequenceB Condition-ConsequenceC Contrastive-Cause-ConsequenceD Consequence-CauseE Consequence-ConditionF Contrastive Consequence-CauseG Argument-ClaimH Condition-ClaimI Contrastive Argument-ClaimJ Claim-ArgumentK Claim-ConditionL Contrastive Claim-ArgumentM ListN ExceptionO OppositionP EnumerationQ Concession

l 1 2 3 -3 -4 .5 *5 *6 .7 *7 *8 .9 10 010 o 12

Figure 2. Experiment 1: Combined multi-dimensional and hierarchical representation of thecoherence relations for data sei 2.

Note. Contrastive Argument-Claim (I) and Contrastive Claim-Argument (L) have identicalcoordinates and hence overlap completely.




pragmatic relations, it is invariably the weakest distinction that turns up.For instance, in both sets there is a cluster of semantic positive causalrelations A and D; in Set l there is a cluster H, K of pragmatic conditionalrelations and in Set 2 there is a cluster E, B of semantic conditional rela-tions. However, separations on the basis of a difference in Source ofCoherence are not found at all for negative relations.It is remarkable that of the four primitives in the taxonomy, the primitiveSource of Coherence is least agreed upon by the analysts. This can beseen both from the first experiment reported here and from the experi-ments reported in Sanders et al. (1992). Therefore, this primitive is furtherinvestigated in the next section.

4.2. Experiment 2: The Source of CoherenceThe lack of clear results for Source of Coherence in different experimentalstudies suggests a systematic confusion regarding this primitive. A pos-sible explanation for this confusion is that this primitive depends morestrongly on the context than the others. The items that were judged inthe experiments may not have been presented with enough linguisticcontext to enable judges to make clear distinctions between semantic andpragmatic relations. This explanation is supported by Redeker's (in prep.)finding that semantic and pragmatic relations are differentiated whenmore Information concerning the context of the related sentences is given.

To find out whether the semanticpragmatic distinction is indeed madein the presence of a communicativeiy clear context, an experiment wascarried out in which expert discourse analysts labelled coherence relationsbetween discourse segments. The aim was to find out whether otheranalysts agree with our classification of coherence relations in terms ofthe primitive Source of Coherence. This experimental task had been usedbefore (Sanders et al. 1992, experiment 1), but this time the discoursesegments were embedded in complete texts. In every text, there were t woexperimental items: one coherence relation that was semantic and onethat was pragmatic. Of each text two versions were constructed: a clearlyargumentative version and a clearly descriptive version. The experimentalitems were fully identical in both versions.

In this way the target segments were embedded in a large and communi-cativeiy clear context. Furthermore, each context had a bias for either asemantic or a pragmatic relation; semantic relations fit best into a descrip-tive context, whereas pragmatic relations fit best into an argumentativecontext.




Two hypotheses were tested:1. As Source of Coherence is a feature of coherence relations, the

difference between the semantic and pragmatic relations will berecognized by the analysts.

2. As the primitive Source of Coherence is recognized best by judgesin a context that corresponds to the character of the relation, theconfusion with regard to the Source of Coherence will be greaterwhen the intended relation differs from the context in which it isembedded. That is, confusion will be greatest when a semanticrelation occurs in an argumentative context or a pragmatic relationoccurs in a descriptive context.

Hence, the first hypothesis is that there is a main effect of the primitiveSource of Coherence. The second hypothesis states that there is aninteraction effect of Source of Coherence with context.

4.2. l. MethodThe material for the experiment consisted of four texts. Two versionswere made of each text: an argumentative and a descriptive version. Eachversion contained four sentence pairs between which the coherence rela-tion was to be judged. Two target pairs were identical in both versions.The two other pairs were filiers that varied with text versions.

The target relations in the experiment were positive additive relations(semantic: List; pragmatic: Enumeration) and positive causal relations(semantic: Cause-Consequence, Consequence-Cause; pragmatic: Claim-Argument, ArgumentClaim). The texts were based on non-specialistarticles and on articles and essays in newspapers and magazines. Examplesof target sentence pairs from experimental texts are given in (15) and(16), from a text about a migratory bird, the (European) crane. Thedescriptive version of this text was entitled "Crane Migration". In thisversion Information is given about the life of cranes, especially abouttheir behavior during and directly after their voyage from the north tothe south of Europe. The context directly preceding (15) is that the birdsarrive in the area where they will stay for the winter after a long journeythrough Europe. The title of the argumentative version of this text, whichwas written from the perspective of an ornithologist who has conductedresearch on the behavior of cranes, is "Misunderstandings about CraneMigration". The author argues that others are wrong in claiming that itis still unclear how migratory birds Orientale themselves and he presentsconclusive evidence for his own Claims, namely that cranes take theirorientation from the sun and the stars.




In (15) a semantic relation is expressed: Cause-Consequence. In (16)the pragmatic relation Argument-Claim is expressed:

(15) Doordat ze in krte tijd grote afstanden moeten aeggen,verkeren de kraanvogels bij aankomst in siechte conditie.'Because they have to fly long distances in a short period of time,the cranes are in a poor condition when they arrive.'

(16) Kraanvogels vliegen nooit als hei mistig is,Dus orienteren ze zieh op de zon en de sterren.'Cranes never fly when it is foggy.Therefore, they orientate themselves by the sun and the stars.*

The other three texts were constructed in a similar manner.There were four experimental conditions, constituted by two experi-

mental factors, each with two levels: type of context (argumentativedescriptive) and type of coherence relation (semanticpragmatic). Twosets of experimental texts were constructed. Each set consisted of fourtexts, two descriptive versions and two argumentative versions. Each setcontained only one version of each text. In each of the four conditionsin a set there was one experimental text. Each set was presented to tensubjects.

The texts were presented to the subjects twice. First, subjects wereasked to read the texts thoroughly. When they had finished reading thetexts, a two-page paper was handed to them. This paper contained thenames and (short) definitions of 17 relations in the taxonomy, s well san example of each relation. Subjects were asked to read through thispaper carefully.13

Next, the texts were presented to the subjects anew, in a different lay-out: the target sentence pairs were numbered and printed in bold type.Subjects read the text and judged each sentence pair printed in bold type.To that end they were to look through the list of relations and to choosean appropriate label for the relation connecting the sentence pair. Thesubjects were instructed to choose the most specific label fitting theexample. For instance, if they hesitated between "List" and "Opposition"relations, they were to choose "Opposition". After they had chosen alabel, the subjects wrote down the number and label of each item on ananswer sheet.

Twenty subjects took part in the experiment. They were paid for theirparticipation. As in Experiment l, subjects were advanced students ofDiscourse Studies at Tilburg University.



116 TJ.M. Sanders, W.P,M, Spooren and L.G.M. Noordman

4.2.2. Results and discussionThe replies were classified s corresponding or not corresponding withour classification. In general, there was a good deal of agreement: of 160choices by the subjects, 10 were not in agreement with our originalclassification (6.25%). Twelve subjects agreed on all eight items with ourclassification, six subjects disagreed on one item and two subjects dis-agreed on two items. The prediction following from the first hypothesisis that the observed number of responses corresponding to the originalclassification on this primitive is significantly higher than chance.

The data were analyzed in two ways: by subject and by item. In bothcases, the analysis concerned the amount of agreement on the primitiveSource of Coherence. The chance proportion was based on the binomialdistribution. According to this distribution, the probability of a t least 7out of 8 subjects correctly classifying the items and of at least 15 out of20 items correctly classified, is less than 5 percent. In the analysis bysubject, 18 subjects agreed more often than chance with the originalclassification (C/zi'2= 12.8, df l, /?


Experiment 2 showed that in an appropriate context the Source ofCoherence of a coherence relation is recognized.

The primitive Order of the Segments did not show up in the experi-ments. As explained earlier (section 4.1.3) we assume that this finding isa consequence of the task that was used and does not argue against thevalidity of the primitive s such.

5. The contribution of the taxonomy to a process model

Recent work in computational linguistics suggests that there is a directlink between the taxonomy presented earlier and the analysis of coherencerelations in natural language generation and understanding.

Hovy (1990) also reaches the conclusion that for a fruitful account ofcoherence relations an organized set is needed. He presents an insightfulstate-of-the-art discussion of computational research on discourse withregard to what he calls "discourse structure relations". He outlines thedilemma of having to choose between long lists of relations (e.g., Mannand Thompson 1988) and the extremely short lists of two intentionalrelations (e.g., Grosz and Sidner 1986). The problem with the first positionis that the lists are often unmotivated and unconstrained, and that theymay lead to analytic problems. The second position is inadequate fromthe point of view of text generation (Hovy 1990: 129), because textgenerators need more than only two intentional relations to produce acoherent text.

Hovy concludes that an approach is needed in which a descriptivelyadequate set of relations is presented that is neither unbounded nor adhoc. He then presents a hierarchical taxonomy of relations, in which 16relations figure at the second level (of three possible levels). Hovy makesa disclaimer by describing his taxonomy s unsatisfactory:I have not found any highly compelling top-level organization. Ideally, the toplevel should partition the relations into a few (say, three or four) major groupsthat share some rhetorical or semantic property. In the absence of a morecompelling Suggestion (for which I continue to search), I use here the top-leveltrifurcation of [Halliday 1985]... . (Hovy 1990: 130)It would seem to us that our taxonomy may play a role in filling the gapbetween Hovy's ideal and his present proposal.

Wu and Lytinen (1990), working along the lines of Hovy (1988), presentan algorithm for the Interpretation of the coherence relations presentedby Mann and Thompson (1988). Their proposal contains several interes-ting ideas concerning the Implementation of a text understanding System.




However, a major flaw is that the relations are treated in an ad hoc way:the prime classification of relations into classes of Make Adequate, Clarifyand Remind is arbitrary, and the relation between these classes is leftunspecified or underspecified. This leads to computational problems: theSystem can only work if it knows in advance what class of relations isappropriate. Another problem is that each relation is specified by anidiosyncratic set of properties/elements.

We want to suggest that much is to be gained by the use of a systemati-cally organized set such s the taxonomy that we have presented. Thecriteria for distinguishing between different relations can be standardizedusing the primitives in the taxonomy, and the procedure for coherencerelation Interpretation can be made explicit. In addition, the number oftests needed to arrive at the correct Interpretation can be radicallyreduced. This will be shown by a detailed comment on Wu and Lytinen's(1990) analysis of the Evidence relation, which is given in (17),

(17) Evidence P QIf1 (Men S P) and2 (Bei S Q) and3 (MB (Imp P Q)) and4 [(Men S (P - Q) or (Men S (P because Q))]Then5 (Evidence S Q P)6 (Ack H (Bei H P'))Glossary(Men S P): Speech action performed by the Speaker S who

states P.(Bei S P): P is the content of the belief held by Speaker S.(MB S H P): S is the holder of the mutual belief P and H is the

other partner who shares this belief. [By default Sand H are omitted.]

(Ack HP): H has an attitude towards a belief or command P.P': the declarative counterpart of P

1. It is unclear whether mentioning P (line 1) follows from mentioning(P Q). If so, line l is superfluous. If not, it is unclear why there isno rule (Men S Q).

2. The appeal to because in line 4 is inelegant, ,as there are Evidencerelations that are marked by connectives other than because(such s s, for, since, etc.). Moreover, it is unnecessary, since




(P because Q) can be considered s an instantiation of the moregeneral case (P Q).

3. The implication relation in 3 relates the antecedent P to the conse-quent Q. It is unclear what is meant by this line. If line 3 i s correct,one runs into problems in view of such examples s (18).

(18) He will not be prepared to do it, because he didn 't do it last week.

In this case "not being prepared to do something" is not the cause of"not having done something last week". Alternatively the authors mayhave intended to relate the antecedent Q to the consequent P. This wouldbe inadequate in view of Evidence examples like (5), (repeated here) inwhich "gasping for breath" does not cause the "being exhausted" butcauses the utterance concerning Theo's exhaustion.

(5) Theo was exhausted, because he was gasping for breath.An emendation of the Evidence algorithm might look s follows:

(19) Evidence P QIf1 (Men S (P-X-Q))2 (MB (Imp Q P'))3 (Bei S Q)Then4 (Evidence S Q P)

Glossary:X: Optional connectiveP': Uttering P

This procedure incorporates the four primitives of the taxonomy:1. Basic Operation. The implication relation mentioned in line 2 indi-

cates that the basic Operation is of a causal nature.2. Source of Coherence. The appeal to P' instead of P indicates that

the relation is of a pragmatic nature.3. Polarity. The segments that are mentioned (line 1) appear directly

in the implication relation (line 2), indicating that the relation is ofa positive nature.

4. Order of the Segments. The linear order of the segments mentionedis specified in line l. In view of the basic Operation in line 2, thisorder is non-basic.

5. The one property that is not determined by the four primitives and




hence is supposedly specific to the Evidence relation is given inline 3: S believes that Q is a firmly established fact.

In a similar vein, Wu and Lytinen's analysis of the Contrast relation (20)can be reanalyzed s the Opposition relation in (21).(20) Contrast P Q

If1 (Men S P)2 (Men S Q)3 (Men S P Q)4 (MB S (Assoc P Q: type OPP))Then5 (Men S (Contrast P Q))Glossary:OPP: Wu and Lytinen distinguish between three types of

Association relations, PartWhole, ClassInclusion andOpposite.

(21) Contrast P QIf1 (Men S (P-X-Q)) and2 (MB S (And P not-Q))Then3 (Opposition S P Q)Glossary:And: Our alternative to the Association relation of Wu and

Lytinen. Contrary to their proposal we see the Polarity ofa coherence relation s a property that is independent ofthe nature of the Operation underlying the coherencerelation.

In our reanalysis we have decomposed the Opposition-Associationrelation s a logical conjunction. The basic Operation is therefore additive.The Source of Coherence is semantic, since the Segments mentionedcorrespond to P and Q in the basic Operation and not to the uttering ofP and Q. The Polarity is negative (the segments that are mentioned,line l, do not occur directly in the basic Operation, line 2). The Order ofthe Segments is irrelevant since it is an additive relation. This means thatit does not matter whether P and Q figure in the basic Operation s theydo now, or in reverse order: (MB S (And not-Q P)).

Tentatively, we reformulate the taxonomy s a process model for




coherence relation understanding, in order to arrive at a systematicproposal. This process model is represented by the flow chart in Figure 3.The claim of the model is not to give a complete description of coherencerelation understanding. It is an outline of a process model incorporatingonly the primitives discussed in this article. The four primitives in thetaxonomy are presented s decision procedures and in this sense themodel represents hypotheses concerning the way in which readers under-stand coherence relations. The product of the flow chart is a filled in

Set reg i sie rOper 0

Sou 0Ord 0Pol 0

Figure 3. The taxonomy s a flow chart(Oper + / : causal/additive basic Operation; Sou + / : semantic/pragmatic source ofcoherence; Ord + / : basic/non-basic order; Pol + / : positive/negative polarity.)




register, which consists of the four primitives with their correspondingvalues. The values uniquely identify one class of coherence relations.

In the fiow chart, the System makes minimally two and maximally sixdecisions to arrive at the correct class of coherence relations. Dependingon the size of the class, a number of additional decisions may be necessaryfor the identification of one among several relations within a class. Itfollows from the diagram that it takes two Steps to arrive at a positivesemantic additive relation, where s it takes six Steps to arrive at anegative pragmatic causal relation.

By way of Illustration, the Interpretation of a Contrastive Claim -Argument relation is demonstrated. The initial assumption expressed inthe diagram is that the relation between the two segments is positive, andso the segments S1 and S2 are considered s the propositions P and Qin the basic Operation. They serve s the input for the subsequent tests.The first test carried out is whether the relation is semantic, that is,whether there is a relation between P and . This is not the case (thereis no positive semantic relation between P and Q) and the next test iswhether the relation is pragmatic. Again the test fails, because there isno relation between P and Q in terms of their illocutions (there is not apositive, but a negative pragmatic relation between the segments).15 It isconcluded that the relation between the two segments is negative, andthe process is rerun, now with the assumption that the relation betweenthe segments is negative. Therefore the input for the subsequent tests isa negative relation between S and S2 (i.e., a relation between 5 andnot-52 or S2 and not-S^). The test for a semantic relation fails, and thePragmatic test applies. The next test, for Causality, applies again, andthe test for Basic Order (does 5 correspond to P and does S2 correspondt ?) fails. Hence, in six Steps the correct (class of) relation(s) isarrived at.

There can be a short-cut in the route through the tree if there is aconnective marking the coherence relation. For instance, if the relationis marked by because it is no longer necessary to test the Polarity andthe Basic Operation of the relation, but it is still necessary to checkwhether 5 and S2 or their pragmatic counterparts are involved in therelation. Note that other languages have more explicit counterparts ofbecause cf. Dutch want (pragmatic), omdat (semantic) (see Spoorenand Jaspers, 1990); French car (pragmatic), parce que (semantic)(Bentolila 1986); Japanese kara (pragmatic), node (semantic) (Takahara,1990).

The process model implied by the flow chart incorporates severalassumptions concerning the processing of coherence relations. All ofthese are empirical Claims that are open to falsification. For instance, the




model predicts that negative relations take more decisions than positiveones, and hence are more complex. Similarly, additive relations are lesscomplex than causal ones.

Another type of claim is made by presenting the model s a flow chartrather than, say, s a decision matrix. This choice reflects the Intuitionthat the primitives may differ in their "weight" in the decision process.A different order of the primitives in the model would yield differentpredictions with respect to the number of Steps needed to identify aparticular relation and, consequently, may yield different predictions withrespect to possible confusions between relations. In terms of the numberof decisions to be made according to the model in Figure 3, the greatestdifference between (classes of) relations, is that between positive andnegative relations, the smallest that between basic and non-basic orderedrelations.

It should be stressed that, while the flow chart is a systematic overallproposal, it is at the same time still rather sketchy s a process model.The different tests incorporated in i t are in fact complete programmesthat need fr t her specification.

6. Discussion

We have argued that coherence relations are an attractive starting pointfor a cognitively plausible theory of discourse representation, providedthat such a theory generates plausible hypotheses about the constructionof a coherent discourse representation. By presenting an organization forthe set of coherence relations, this condition seems to be fulfilled. Theproposed taxonomy offers a plausible categorization of coherence rela-tions that accounts for the relationships among coherence relations. Wehave presented arguments from different fields for the primitives in thetaxonomy and we have given empirical evidence for their cognitiveplausibility.

In its form of a process model the taxonomy can be tested for itsadequacy: it should provide a systematic basis for an algorithm forcoherence relation understanding. Clearly, such an algorithm cannot beimplemented on the basis of the current proposal. If we consider thisalgorithm s an ultimate research tenet, then it becomes clear that thereare many questions left unanswered.

First, in order to arrive at a working algorithm for coherence relationInterpretation, the decisions representing the primitives will have to bespecified and operationalized further. Therefore, the taxonomy is in needof additional linguistic work on the further definition of the properties




of the primitives. In addition, a crucial issue for any algorithm fordiscourse production or understanding is the problem of knowledgerepresentation. Clearly, work has to be done on the amount ofknowledge an Interpretation System needs in making the decisions.Furthermore, our proposal could benefit greatly from linguistic work oncentral problems like the different realizations of the primitives in differentlanguages compare the remarks in the preceding section on the linguis-tic markers of positive causal relations in languages other than English.

Second, the end nodes in the flow chart are classes rather than indivi-dual relations. Additional criteria (which we claim to be segment-specific)are needed to arrive at the individual relations. Candidates for suchcriteria are Temporal specificity (distinguishing Sequences and Overlapsfrom mere Lists in Class 9), Hypotheticality (distinguishing conditionalsfrom other causal relations), Volitionality (distinguishing between Reasonand Explanation (Class 3), Specificity (distinguishing betweenSpecification, Generalization and Restatement within Class 9 andException and Opposition in Class 10), etc. Further elaboration of suchcriteria is needed (see Sanders 1992, eh. 6).

Other important questions remaining concern research on the cognitiveClaims in the process model. In our view, empirical evidence in favor ofsuch claims must come from psycholinguistic research in areas such sdiscourse processing and language acquisition. Some of these Claims seemmore than plausible. Consider, for instance, the results of psycholinguisticstudies concerning the salience of such classes s additive versus causaland negative versus positive relations (see section 3.2). The literature onlanguage acquisition also seems promising: it is a well established factthat children acquire negative relations later than positive relations andthat causal relations emerge later than additive relations (cf. Bloomet al. 1980).16

A last remark concerns phenomena that a cognitive theory of discourserepresentation should account for but that are missing in this article. Oneof the most important themes in discourse studies is the hierarchicalstructure of discourse (see Polanyi and Scha 1983; Polanyi 1988). Clearly,there are interesting interactions between the "relational meaning" prop-erties of coherence relations and their hierarchical properties. For exam-ple, causal relations all seem to be subordinating, whereas most additiverelations are coordinating. The relational and the hierarchical aspects ofcoherence should be regarded s complementary topics in research oncoherence, both of which need to be explored further.

In our view, interdisciplinary research from fields such s text anddiscourse analysis, linguistics, computational linguistics and cognitivepsychology leads to an exchange of results and a fruitful interaction of




theoretical insights and methodologies. It forms a suitable starting pointfor further work in the direction of a cognitively plausible theory of thecoherence and structure of discourse.

Appendix

Below, the relations in Table l are illustrated in a context [in brackets],in Dutch and in English. The context and the examples were used inExperiment l, reported in section 4.2.

Dutch[Elk najaar vertrekken veel vogels uit hun broedgebied in Noord-Europanaar het zuiden. Dat doen ze om de winter, waarin het zoeken naarvoedsel moeilijk wordt, voor te blijven. Een typische trekvogel is dekraanvogel (Grus Grus), die broedt in Noord-Scandinavie en overwintertin Spanje en Noord-Afrika.

Er wordt veel onderzoek gedaan naar de vogeltrek. Jaarlijks wordenenkele kraanvogels vlak voor het vertrek uit hun broedgebieden uitgebreidonderzocht door biologen. Vaak hebben de onderzoekers de grootstemoeite de dieren te vangen. In Noord-Spanje doet een Nederlandse"Kraanvogelploeg" elk najaar onderzoek naar de conditie van de vogelsals ze in hun overwinteringsgebied zijn aangekomen. Direct na aankomstworden enkele exemplaren gevangen, gewogen en nader onderzocht. Dereis dwars door Europa, die gemiddeld twee tot drie weken duurt, blijktde vogels erg veel energie te kosten.]

(1) a. CauseConsequenceDoordat ze in krte tijd grote afstanden moeten afleggen, ver-keren de kraanvogels bij aankomst in siechte conditie.

b. ConditionConsequenceAls ze binnen twee weken in Spanje zijn, verkeren de kraanvogelsbij aankomst in siechte conditie.

(2) Contrastive CauseConsequenceHoewel de kraanvogels uitstekende vliegers zijn, verkeren ze bijaankomst in siechte conditie.

(3) a. ConsequenceCauseDe kraanvogels verkeren bij aankomst in siechte conditie, door-dat ze net daarvoor de Pyreneeen zijn overgestoken.




b. ConsequenceConditionDe kraanvogels verkeren bij aankomst in siechte conditie, als zede Pyreneeen over krnen.(Sommigen overleven de tocht over die bergketen niet.)

(4) Contrastive ConsequenceCauseDe kraanvogels verkeren bij aankomst in siechte conditie, hoewel zeonderweg loch vaak en veel uitrusten.

(5) a. Argument-ClaimDe kraanvogels kunnen in Noord-Spanje gemakkelijk wordengevangen, dus verkeren ze bij aankomst in siechte conditie.

b. ConditionClaimMits ervan mag worden uitgegaan dat de onderzoeksgegevensbetrouwbaar zijn, verkeren de kraanvogels bij aankomst insiechte conditie.

(6) Contrastive ArgumentClaimAI wijzen niet alle onderzoeksgegevens in dezelfde richting, de kraan-vogels verkeren bij aankomst in siechte conditie.

(7) a. Claim-ArgumentZe verkeren bij aankomst in siechte conditie, want ze wegensoms nog maar de helft van wat ze wogen bij het vertrek uitScandinavie.

b. ClaimConditionZe verkeren bij aankomst in siechte conditie, als er tenminstevan mag worden uitgegaan dat gewicht een goede indicator isvoor hun conditie.(Bij aankomst wegen ze soms nog maar de helft van wat zewogen bij hun vertrek.)

(8) Contrastive ClaimArgumentDe kraanvogels blijken bij aankomst in siechte conditie, al wegen zeongeveer even veel als bij hun vertrek.

(9) ListIn groepen van gemiddeld 100 tot 300 exemplaren trekken de kraan-vogels Spanje binnen. Ze verkeren bij aankomst in siechte conditie.




(10) a. ExceptionDe kraanvogels verkeren bij aankomst in siechte conditie. Datgeldt niet voor de tweedejaars vogels.(Die blijken het sterkst).

b. OppositionDe kraanvogels verkeren bij aankomst in siechte conditie. Bijhun vertrek zijn ze daarentegen in optimale conditie.

(11) EnumerationDe kraanvogels verkeren bij aankomst in siechte conditie.Bovendien overleeft twintig procent van hen de tocht niet.

(12) ConcessionDe kraanvogels verkeren bij aankomst in siechte conditie, maarde meeste herstellen zieh snel.17

English[Every autumn many birds leave their breeding areas in Northern Europe,to travel south. They do that to avoid the winter, which makes if difficultfor them to find food. A typical migratory bird is the crane (Grus Grus),which breeds in North Scandinavia and stays in Spain and NorthernAfrica during the winter.

Much research is done on bird migration. Every year some cranes areexamined by biologists immediately before their departure frorn the breed-ing territories. Often, the researchers have trouble catching the animals.In the Northern part of Spain, a Dutch "crane team" examines thecondition of the birds every year on arrival in their winter territories.Directly after their arrival some birds are caught and examined further.The journey across Europe, which lasts two to three weeks on average,appears to cost the birds a lot of energy.]

(1) a. Cause-ConsequenceBecause they have to fly long distances in a short period oftime, the cranes are in poor condition on arrival.

b. ConditionConsequenceIf they make it to Spain in two weeks, the cranes are in poorcondition on arrival.

(2) Contrastive CauseConsequenceAlthough the cranes are good flyers, they are in poor conditionon arrival.




(3) a. ConsequenceCauseThe cranes are in poor condition on arrival, because they havejust crossed the Pyrenees.

b. ConsequenceConditionThe cranes are in poor condition on arrival, if they make itacross the Pyrenees.(Some do not survive the journey over these mountains.)

(4) Contrastive Consequence-CauseThe cranes are in poor condition on arrival, although they takelong and frequent rests during their journey.

(5) a. Argument-ClaimThe cranes can be caught easily in Northern Spain, so they arein poor condition on arrival.

b. Condition-ClaimProvided that it may be assumed that the research data can betrusted, the cranes are in poor condition on arrival.

(6) Contrastive ArgumentClaimAlthough not all research data point in the same direction, thecranes are in poor condition on arrival.

(7) a. Claim-ArgumentThey are in poor condition on arrival, for they have sometimeslost half of the weight they had when they left Scandinavia.

b. Claim-ConditionThey are in poor condition on arrival, at least if it may beassumed that weight is a good indicator for their condition.(Upon arrival they have sometimes lost half the weight theyhad when they left.)

(8) Contrastive Claim-ArgumentThe cranes appear to be in poor condition, although they weighabout s much s when they left.

(9) ListIn groups of 100 to 300 birds on average, the cranes enter Spain.They are in poor condition on arrival.




(10) a, ExceptionThe cranes are in poor condition on arrivah This does not holdfor the two-year-old birds.(They appear to be the strengest.)

b. OppositionThe cranes are in poor condition on arrivaL By contrast, theyare in a very good condition when they leave.

(11) EnumerationThe cranes are in poor condition on arrivaL Moreover^ morethan twenty percent of them do not survive the journey.

(12) ConcessionThe cranes are in poor condition on arrival, but most of themrecover quickly,

Received 28 June 1991 Tilburg UniversityRevision received 7 February 1992

Notes

* We would like to thank Sandy Thompson, Bob Dooley, several anonymous reviewersand our colleagues of the Discourse Studies Group at Tilburg University, especiallyLeonoor Oversteegen, for their cormnents on earlier versions of this article; MaartjeTel and Elles van Happen for experimental assistance; and Pieter Nieuwint for correct-ing our English. Needless to say, all remaining errors are our own.

1. In our view, there are no principled differences between spoken and written (mono-logic) discourse with respect to most phenomena we address in this article. Therefore,we prefer the neutral term "discourse" to the term "text", because the latter i s explicitlyrestricted to written language. However, "text" will be used with respect to researchthat i s specific for written language, also for conventional reasons (the field oftenknown s "discourse analysis" is different from that called "text analysis"). In addition,we will refer to experimental materials s "texts", to prevent misunderstanding aboutthe character of the Stimuli.

: 2. This idea has been put forward by many other authors, among them Charolles (1984).| 3. Sanders (1992, eh. 4) gives a critical review of these and other experimental findings.! 4. The fact that Meyer, like Mann and Thompson, is often more concerned with relations1 on a more global level is of course not a fundamental difference. Coherence relations

l exist both on a local level (between clauses) and on a global level (between paragraphsj or between chapters of a book).l 5- The many angry reactions to this sentence, which appeared in an obituary of Greta

Garbo (Volkskrant, April 17, 1990), clearly illustrate that the coherence relationexpressed is based on a causal Operation,

6. The completeness issue is discussed more extensively in our earlier proposal (Sanders




et al. 1992), where arguments are given for the absence of relations like Elaboration,Temporal Overlap and Temporal Sequence, and Background.

7. See Sanders et al. (1992, Section 3.4; notes 8,11).8. Although the distinctions made in the literature are similar to our semantic-pragmatic

distinction, there are also clear differences. It goes beyond the scope of this article t oelaborate on these differences.

9. "The decision to prepose or postpose a given subordinate clause is made largely on thebasis of the coherence relation (although ... discourse-pragmatic factors can changethe typical order). Therefore, in the case of switch-reference constructions, the relativeorder of clauses can provide a clue s to the relation" (Dooley 1990: 23).

10. The stress for the one-dimensional solution was 0,41 (r2 = 0.57); the stress for the three-dimensional solution was 0.13 (r2 = 0.91). Following Kruskal (1964: 16) we chose thetwo-dimensional solution, because the difference in stress between the one- and two-dimensional solution is much greater than the difference between the two- and three-dimensional solution. It is, however, possible to Interpret the three-dimensional solu-tion. The first dimension seperates three groups of relations. The positive causal rela-tions are at the one extreme, the negative relations are at the other extreme, and thepositive additive and conditional relations are in the middle. This dimension wouldthen reflect the distinction between negative and positive relations. At the one extremeof the second dimension are conditional relations; positive causal relations are at theother extreme, and negative and positive causal relations are in the middle. Thisdimension reflects the conditional-non-conditional distinction. The third

Documents

Sanders Spooren Noordman 1993 Coherence Relations in a Cognitive Theory of Discourse Representation