Chapter 1 Pragmatic strategies - Paris Diderot Universityamsili/papers/van... · 2004-05-04 · Chapter 1 Pragmatic strategies Rob van der Sandt University of Nijmegen [email protected]

Chapter 1

Pragmatic strategies

Rob van der SandtUniversity of [email protected]

1.1 Introduction

In the seventies Gerald Gazdar developed a theory of formal pragmatics the core of whichwas his theory of presupposition projection.1,2 Though Gazdar’s theory covered a widerange of data and though, at its time, it was vastly superior over competing theories,its further development was hampered by some serious problems. One problem was thecounterintuitive notion of presupposition which did not capture the intuitive view of pre-supposition as being information which is in some sense given or taken for granted in adiscourse. A concomitant problem was his account of cancellation which was widely feltto be conceptually incoherent. In Gazdar’s view conversational implicatures constitute astronger kind of information than presupposition: presuppositional information may bedefeated by implicatures. This is at odds with our pretheoretical notion of presuppositionand conflicts with the standard way this notion is characterized. Intuitively, if a sentenceis felt to be presupposing, the presuppositional information is (in some sense to be madeexplicit) required for its utterance to be interpretable. The standard characterization ofconversational implicatures is a very different one. On the Gricean view a conversationalimplicature is not a semantic requirement for interpretation, but instead an inference thatis computed on the basis of the interpretation of a sentence, contextual information andconversational principles. The question then immediately arises how such inferences, whichare for their existence dependent on a previous interpretation of the utterance that invokesthem, can possibly defeat information which should already be part of the linguistic ornon-linguistic context for the inducing sentence to have an interpretation. This problem isthe more pressing since the received view tells us that conversational implicatures are theparadigm cases of defeasible inferences. Thus, if there is any direct interaction betweenpresupposition and conversational implicature, one would expect the former to override

1I would like to thank Bart Geurts, Emar Maier, Jennifer Spenader and the participants of the Kampworkshop on Presupposition for their helpful comments.

2Gazdar (1979), the published version, is a revision of his 1976 thesis which was completed under thesupervision of Hans Kamp.

1

2 CHAPTER 1. PRAGMATIC STRATEGIES

the latter and not the other way around.

Gazdar’s theory has a dynamic flavour. However, his dynamics is found only at the levelof sentential discourse processing and does not extend to the interpretation of subsententialconstituents. Moreover, presuppositional information is only computed and incremented inthe subsequent context after the propositional content has been computed and incremented.Current theories that are based on dynamic frameworks deviate in exactly this respectand give a very different picture of how presuppositions are evoked and resolved in acontext. In contrast to the Gazdarian picture they agree with Frege and Strawson thatpresuppositions are requirements for the interpretation of an utterance, they agree withStalnaker and Karttunen that presuppositional information is information that is somesense contextually given or taken for granted in a conversation. They agree with Lewisthat this information may not always be explicitly given in the context of utterance, butmay be reconstructed, and they thus hold that the felicity of a presupposing sentence maybe established by a process of accommodation. On current dynamic assumptions this meansthat accommodation is not simply a side product of contextual update, which only affectsthe output context, but should instead be reconstructed as an operation on input contextsso as to enable an utterance to be processed even if the original context didn’t admit it.And, finally, there is by now a consensus that an adequate account of presupposition andpresupposition projection is to be construed as a hybrid theory in which semantic andpragmatic principles interact. Existing views mainly differ with respect to the questionwhat exactly counts as contextual requirements for the interpretation of presupposingsentences. They differ moreover in their views on the type of mechanism and the status ofthe principles that govern accommodation. Heim, Beaver and other satisfaction theoristsbasically require that the context of utterance entails the presuppositional information.3

Current anaphoric accounts view presuppositions as identifiable objects that should findsome other identifiable object to serve as an antecedent for the presuppositional expression.4

And the latter are closer to the theory of Gazdar in accommodating this object in case itis not already there.5

In retrospect we may point to two major characteristics of Gazdar’s system that di-vorce his methodology from current theories, and that prevented analysts to develop hisaccount into a more adequate and conceptually appealing theory of presupposition. Thefirst characteristic to blame is what I will call his preKampian dynamism, the second hisvery unKampian view on the division of labour between semantics and pragmatics.6 Thesetwo problems are related. Gazdar’s system is confined to presupposition computation on apropositional level. In his simple model of discourse incrementation the dynamics does notextend below the level of full sentences, which means that it cannot handle quantificationaland anaphoric links beyond the sentence boundaries. As a consequence, presupposition,content and implicature expressions are incremented separately as entities in their ownright. Presuppositional or other pragmatic information will thus never enter into binding

3The satisfaction account originates in the work of Karttunen (1974) and Stalnaker (1974). See fordevelopments in a dynamic framework in particular Heim (1983) and Beaver (1997, 2001).

4See e.g. Van der Sandt (1992), Geurts (1995, 1999), Sæbø (1996), Krahmer (1998), Asher & Lascarides(1998), Bos (1999), Kamp (2001). See Zeevat (1992) for a hybrid account which incorporates ideas fromboth worlds

5See Geurts (1996) and Beaver (1997) for two different assessments of the state of the art.6See Kamp (1979) for a criticism of the classic division of labour between semantics and pragmatics

which antedates DRT two years.

1.2. GAZDAR’S DYNAMICS 3

relations with the semantic content of a sentence. Nor is presupposition treated in theFregean way as a definedness condition of the presupposing sentence. Instead presupposi-tion is treated as a purely pragmatic phenomenon which does not contribute to the semanticcontent of an utterance and thus cannot serve as a requirement for the interpretation ofthe inducing sentence.

The first problem, which has become known as the binding problem, can only be solvedwhen we adopt DRT or some other theory of dynamic interpretation. I will briefly commenton this issue in the next section and refer the reader to the extensive literature on dynamictreatments of presupposition that turned up in the nineties. The second problem, theinteraction between pragmatic inferencing and semantic interpretation, will be discussedwhen I turn to a discussion and reevaluation of Gazdar’s system (in particular section 1.5).I will then put forward some tentative remarks on the relation between presupposition andquestions in discourse.

The general setup of the paper is as follows. I will first show that even when werestrict ourselves to Gazdar’s conceptual frame, some oddities of his system can easily beremedied. This will set the stage for a reinterpretation of the major feature which governsthe behaviour of presuppositions in his system: the interaction between presuppositionsand his clausal implicatures. I will claim that these ‘implicatures’ should not be seen asordinary Gricean implicatures, but should instead be seen as constraints on the incomingcontext. In fact they can be reinterpreted as ‘issues’ or ‘questions’ in discourse. And, moreimportantly, I will point out that it is exactly the possibility of such a reinterpretation whichexplains why his original clausals did the work they were devised for. I will finally showhow such a reinterpretation, when implemented in current dynamic treatments, may givea new view on certain accommodation effects as giving rise to a special kind of bridginginference. I will illustrate this by means of a simple solution to a well-known probleminvolved in the analysis of disjunctive sentences.

1.2 Gazdar’s dynamics

Gazdar’s account is formulated in an incremental model of discourse processing in whicheach utterance changes the incoming context by adding the information conveyed to thenext context.7 As I pointed out in the previous section Gazdar’s dynamics does not extendbelow the sentence level and is not applicable when presuppositions and implicatures enterinto binding relations with quantified expressions. Gazdar’s presuppositions and implica-tures act as entities in their own right, but they may, in view of their content, constrainthe incrementation process of subsequent utterances.

Utterances are construed as context-sentence pairs.8 A discourse is an ordered sequence

7See Gazdar (1979) for the full exposition of his theory. Many detailed expositions are found in theliterature. See e.g. Soames (1982), Levinson (1983) , Van der Sandt (1988), Beaver (1997, 2001).

8Actually Gazdar construes them as indexed pairs of contexts and propositions, where the context istaken to be a set of propositions. Since one would like to conceive of ‘I am hungry’ said by me to youand ‘You are hungry’ said by you to me as different utterances expressing the same proposition, I willtake utterances as pairs of sentences and contexts where the context contains the parameters needed todetermine the proposition expressed, enriched with an information set which plays the role of Stalnaker’scontext set. This has no consequences for the discussion but simplifies the formulation of Gazdar’s system.Since Gazdar’s system is limited to modal propositional logic I will often use ‘context’ to refer to thisinformation set.


of utterances. The contextual information accumulated at a certain stage of a discourse is aset of propositions, which comprises all the information conveyed up to that point and thusincludes the contribution of the implicatures and presuppositions. Such an information setis interpreted as a commitment slate in the sense of Hamblin.9 After the utterance of asentence the speaker is thus not just committed to the truth-conditional content of his ut-terance, but also to all further information that has been conveyed by non-truth-conditionalmeans. The latter comprises in Gazdar’s theory not only implicatural information but alsothe presuppositional contribution.

In order to allow for the contextual defeasibility of presuppositions and implicaturesGazdar distinguishes between potential presuppositions (pre-suppositions in Gazdar’s ter-minology or proto-presuppositions as I will call them) and the actual presuppositions ofa sentence in a discourse. Similarly there is a distinction between potential implicatures(im-plicatures or proto-implicatures) and the actual implicatures conveyed by the utter-ance of a sentence. Both proto-implicatures and proto-presuppositions are generated ina purely mechanical way. Proto-presuppositions are read off from presuppositional con-structions such as definite NP’s, factive verbs, verbs of change and other items that arestandardly taken to be presupposition inducers. They are given by the lexicon and thesyntax of the language under consideration. These are the entities that may in the course ofthe incrementation process emerge as the actual presuppositions of an utterance. Whetherthey do so, will depend on the information contained in the context at the moment ofincrementation. The actual presuppositions of an utterance will thus always be a subsetof the set of proto-presuppositions.

Contextual incrementation is based on a hierarchical ordering of semantic content,proto-implicatures and proto-presuppositions, and is constrained by the requirement ofconsistency. Given a context, the proposition expressed by a sentence in this context,and given a previous determination of the proto-implicatures and proto-presuppositionsof this sentence, contextual incrementation proceeds in three stages. We first create anauxiliary context by adding the proposition expressed to the already established contextualinformation. This yields a context which consist of the union of the old context withthe truth conditional content of the sentence. In the next stage we add those consistentsets of proto-implicatures which retain consistency with the already established set ofpropositions, thus again establishing a new context which comprises both the content andthe implicatures associated with this sentence. Finally we add those consistent sets ofproto-presuppositions which are consistent with the set established by the incorporation ofthe implicatures. Only those proto-implicatures and proto-presuppositions which don’t giverise to contradictions in this staged process are thus established as the actual implicaturesand presuppositions of an utterance. This ordered process eliminates inter alia all proto-implicatures which are inconsistent with the propositional content, contextual informationor each other. It also eliminates all proto-presuppositions which are incompatible with theset resulting from the incorporation of the implicatures or which are inconsistent with oneanother. Only those proto-implicatures and proto-presuppositions which are added to thenext context are taken to be the actual implicatures and presuppositions of the sentenceprocessed. Figure 1 gives a pictorial representation of the incrementation procedure asGazdar conceives it.The following examples may illustrate the idea behind this general setup:

9See Hamblin (1971).

1.2. GAZDAR’S DYNAMICS 5

〈ϕ, c0〉〈χ, c0〉〈ψ, c0〉

c′

0c′′

0c′′′

0= c1 c′

1c′′

1c′′′

1= c2 c′

2c′′

2c′′′

2

c0ordered incrementation

sem. impl. pres. sem.impl. pres. sem.impl.pres.

Figure 1: Discourse

(1) a. Either John does not own a donkey or he keeps his donkey very quiet.b. Either John has an insulated stable, or he keeps his donkey very quiet.

The first disjunct of (1a) invokes inter alia the proto-implicature that for all the speakerknows John may not have a donkey. This proto-implicature is consistent with the contentof (1a). It will thus be incremented in the next context and be established as an actual im-plicature of an utterance of (1a). We then consider the contribution of the presuppositionsin the context that has thus been established by the incorporation of the implicature. Thesecond disjunct induces the proto-presupposition that John has a donkey. The assump-tion that John has a donkey conflicts with the information that has been established byadding the implicature that he may not have a donkey and is thus defeated.10 The proto-presupposition will thus not be inherited which correctly predicts that an utterance (1a)does not presuppose that John owns a donkey. In (1b) we observe a different situation. Dueto the fact that the first disjunct of (1b) is logically independent of the proto-presuppositionno conflicting implicature will be invoked in this case (the proto-implicatures that for allthe speaker knows John’s stable may or may not be insulated, are compatible with theassumption that he has a donkey). Hence the proto-presupposition will be inherited bythe next context and thus be established as an actual presupposition of the utterance.

Gazdar’s system captures the fact that survival of pragmatic information depends onthe context of utterance. It, moreover, correctly construes presupposition not as a propertyof the inducing sentence, but of its utterance in a context. However, in line with nearly allother theorists at his time he assumes semantic and pragmatic information are two differ-ent types of content that are computed and represented separately and will only be mergedafterwards in a more substantial proposition.11 On this view it is this more informativepropositional object which encodes the full message conveyed. Given a pragmatic view ofpresupposition this implies that we should be able to determine the proposition expressedby a sentence independently of the contribution of the presuppositional information. Thishas some unfortunate consequences. It yields a counterintuitive notion of propositionalcontent and, even worse, gives rise to binding problems when presuppositions and impli-catures entertain anaphoric relations with external antecedents. To illustrate the point letus have a look at (2):

10‘Conflicts’ in an intuitive sense. Logically there is no contradiction. See section 1.3.11Karttunen and Peters (1979) (though not Karttunen (1974)) is a particularly striking example. It

should be mentioned that Stalnaker’s work (1974, 1978), is an exception, though the Stalnaker inspiredtheory of Van der Sandt (1988) and the general criticism of the standard view in Van der Sandt (1992)may wrongly suggest the opposite.


(2) The mayor opened the exhibition.

Splitting up the presuppositional and content-contribution we obtain (3a) and (3b) for thepresuppositional information and the content respectively

(3) a. ∃x∀y[mayor(x) ↔ x = y]b. ∃x∀y[[mayor(x) ↔ x = y] ∧ open exhibition(x)]

that is, we are forced to encode the information that there is a mayor twice, once as part ofthe content expression and once as part of the presuppositional contribution. This capturesthe observation that presuppositional inferences of non-embedded sentences are preservedas an entailment of their inducing sentence. It also captures the fact that this informationcontributes to the proposition expressed when part of a complex structure. But note thatthe reason is not that ‘the presupposition’ is entailed by the subsentence that invokesit, as some pragmatic analyst have claimed. It only happens because the presuppositionalinformation is encoded twice, once as part of the content expression and once as a defeasibleproto-presupposition. It is the non-presuppositional content part that is preserved as a non-defeasible entailment. It is thus misleading to say that presuppositions cannot be defeatedin entailing environments. For even if the proto-presupposition is defeated or neutralized,its non-presuppositional counterpart in the content expression will still do its semanticduty. There is thus no principled reason why Gazdar’s system cannot be redefined in sucha way that the presupposition is defeated in such environments though the inference tothe purportedly presuppositional information goes through. The redundant encoding isnevertheless wrong for both technical and empirical reasons. Firstly, it does not capturethe fact that the non-presuppositional information conveyed by (2) is not that there is amayor who opened the exhibition, but that its utterance contributes to a context whichshould already provide a mayor and conveys there that he opened the exhibition. Thismeans that the non-presuppositional contribution should be given by an open sentencethat has to link up to the established presuppositional information. The procedure thusassumes a counterintuitive notion of propositional content. It moreover yields the wrongtruth-conditions when we consider non-extensional embeddings. Consider (4a) and notethat its utterance would normally presuppose that there is a mayor i.e. (4b) and convey(4c), that is the scalar implicature that it is not necessary that he opened the exhibition:

(4) a. It is possible that the mayor opened the exhibition.b. There is a mayor.c. It is not necessary that the mayor opened the exhibition.

When we encode the contribution of the presuppositional, content and implicature expres-sions separately, sentence (4a) is predicted to convey that there is a (unique) mayor in thisworld, that there is a (possibly) different one in the world where the exhibition was opened,and furthermore, by way of scalar implicature, that it does not hold for every world thatthere is a mayor who opened the exhibition. This is wrong and typical for all other theoriesthat adhere to a strict division of labour between the semantic and pragmatic component.12

A proper analysis should rather run along the lines of (5) and be analysed as sketchedin Figure (5):

12This also includes Van der Sandt’s (1988), though not his (1992) account.

1.3. EPISTEMIC MODIFICATION 7

(5) a. There is a mayor . . .it is possible . . .

but not necessary that he opened the exhibitionb. ∃x mayor(x) . . .

♦ open exhibition(x) . . .¬� open exhibition(x) . . .

SEMANTIC CONTENT∃x mayor(x) ♦ open exhibition(x) ¬� open exhibition(x)

(presupposed info (assertoric contribution (implicatural infocontextually required) dependent on the computed on the basis

presuppositional of contextual inforequirements) + content)

ACCOMMODATION INCREMENTATIONi.e. adjustment of the i.e. mapping of the adjustedcontext of utterance context into the next one

Figure 2: Types of information

We thus perceive an anaphoric process where the presuppositional expression should bind avariable in the (open) content expression and where the implicature is in turn anaphoricallydependent on both the presuppositional and the content expression. The content thusdepends on the presuppositional expression and takes priority over the implicature. Thispicture effectively reverses the order postulated by Gazdar.13

Gazdar’s theory is based on a standard static semantics. A solution of the bindingproblem had to wait until Kamp (1981) and Heim (1982) presented their versions of dy-namic semantics, recaptured meanings as relations between input and output contexts andshowed how to treat anaphoricity inter- and intrasententially in a uniform way. Variousproposals are found in the presupposition literature since then. I contend that the samedevelopment allows us to remove a second obstacle from Gazdar’s theory, the counterintu-itive and conceptually incoherent thesis that the contribution of implicatures takes priorityover the computation of presuppositions. But in order to do so I first have to discuss an-other feature of his theory that has been objected to by a variety of authors, the intuitivelyunreasonably strong epistemic modification of presuppositional, content and implicatureexpressions.

1.3 Epistemic modification

In the previous section I gave an informal sketch of Gazdar’s method of contextual incre-mentation as a kind of default inheritance mechanism. The basic constraint on updateis consistency. Only consistent subsets of information are inherited by the next context.However, in many cases intuitively conflicting information does not formally give rise toinconsistency between different types of pragmatic and semantic information. In order to

13See Geurts and Maier (ms.) for an extension of DRT which allows us to represent and interpret variouskinds of semantic and pragmatic information in a unified framework.


obtain inconsistency Gazdar takes recourse to a strong modal modification of the relevantexpressions.

The two modal operators figuring in Gazdar’s system are the K- and the P-operator ofHintikka’s (1962) epistemic logic. Kaϕ is to be read as ‘a knows that ϕ’ and Pϕ as ‘it ispossible, for all that a knows that ϕ’. These are the epistemic variants of the well-knownnecessity and possibility operators of standard modal logic and are thus interdefinable inone another:

Paϕ := ¬Ka¬ϕ

Gazdar limits himself to monologue. Hence he only needs indices for one speaker which maythus safely be skipped. And though his proofs are cast in terms of Hintikka’s epistemic logic,the modalities may be interpreted by normal Kripke semantics. The K- and P-operatorsmay be interpreted as the box and the diamond respectively. If we assume reflexivity andtransitivity for the accessibility relation the system semantically corresponds to KT4. Acentral property of knowledge is that anything that is known by someone is true. Reflexivityguarantees the validity of the scheme Kϕ → ϕ and thus that knowledge entails truth. Asecond property which will turn out to be crucial is positive introspection: if one knowssomething one knows that one knows that. Transitivity guarantees just that by validatingthe scheme Kϕ → KKϕ. We may add two further operators, B and C for belief. These‘doxastic’ operators are interdefinable just as their epistemic counterparts. Bϕ is to be readas the speaker believes that ϕ. And Cϕ means that it is compatible with the speaker’sbeliefs that ϕ. This weaker notion should not obey Bϕ → ϕ as there are false beliefs.It should however respect Bϕ → Cϕ (beliefs are consistent). We obtain this by skippingreflexivity and requiring just transitivity and seriality for belief.

According to Gazdar the assertoric utterance of a sentence ϕ commits the speaker toknowing that ϕ under Grice’s maxim of Quality. It is this commitment which gives rise toa so-called Quality implicature of the form Kϕ. Thus by the utterance of

(6) The cat is on the mat.

the speaker does not just commit himself to the cat being on the mat, but to knowingthat this is the case. This is too strong. Surely by an assertoric utterance of ϕ, a speakercommits himself to its truth, cannot admit disbelief in his statement and, when askedfor further justification, he may back up his claim by saying that he knows (or believes)it is true. That is, by the assertoric use of ϕ a speaker is committed to act as if ϕ istrue, but by his utterance he is neither saying nor implicating that he knows (or believes)that ϕ. His commitment is a consequence of uttering this sentence with assertoric force,but it is not part of the presuppositional, propositional or implicatural information andshould be handled at a meta level. Speaker commitment is already captured by Gazdar byinterpreting a context as a commitment slate and thus is already taken care of and handledas a fact about the context. Hence the epistemic commitment of the speaker to the truthof what he says need not be encoded by adding knowledge claims to the information setthat is taken to be his set of commitments.

Note that Grice already objected to construing epistemic commitment as some kind ofquality implicature.

1.3. EPISTEMIC MODIFICATION 9

On my account, it will not be true that when I say that p, I conversationallyimplicate that I believe that p; for to suppose that I believe that p [...] is just tosuppose that I am observing the first maxim of Quality [. . . ] it is not a naturaluse of language to describe one who has said that p as having, for example,‘implied,’ ‘indicated,’ or ‘suggested’ that he believes that p.

Grice (1989: 42 )

Proto-presuppositions are likewise modalized with the K-operator. The proto-presuppositionGazdar takes to be induced by (2) is not that there is a mayor but the stronger expres-sion that the speaker knows that there is a mayor. A typical utterance of (2) will thuspresuppose:

(7) K[there is a mayor]

The remarks just made with respect to the epistemic modification of assertoric utterancesapply here as well. When a speaker indicates that he presupposes that there is a mayor,he commits himself to the truth of what he presupposes, but just as in the case of plainassertions this is accounted for by interpreting a context as a commitment slate. There isthus no need to enter the description of a consequence of this commitment to the context.

Gazdarian implicatures derive from the maxim of Quantity and come in two varieties:clausals and scalars. Clausal implicatures are read off from the clauses of compound sen-tences. Given a compound sentence ϕ the mechanism yields for each non-entailed andnon-presuppositionally induced subsentence χ a set proto-implicatures Pχ and P¬χ. Fora disjunction of the form χ∨ψ this yields {Pχ, P¬χ,Pψ,P¬ψ}. The proto-implicatures of(1a) thus comprise P[John has a donkey] and P¬[John has a donkey]. Hence, by theutterance of this sentence the speaker will typically implicate that for all he knows it ispossible that John has a donkey and that for all he knows it is possible that John doesnot have a donkey, which might also be paraphrased more concisely as the speaker doesnot know whether John has a donkey. The informal explanation is that by the utteranceof a disjunction the speaker implicates that he does not take the truth of either disjunctfor granted. For if he had known that either disjunct were true, he would have been in theposition to make a much stronger statement. If he had known that John didn’t have a don-key, he should have said so directly, thereby producing a more informative statement. If hehad known that John had one, the plain assertion of the second disjunct would have beenshorter and more informative than the full disjunction. In both cases he would have beenrequired to choose the alternative expression by the maxim of Quantity. The implicaturefor the second disjunct is derived analogously. The hearer will thus conclude that since thespeaker asserted only the weaker statement, he was not taking the truth of either disjunctfor granted and thus conversationally implicated that he didn’t know whether John hada donkey or whether he kept his animal quiet. Clausals thus capture the suggestions ofepistemic uncertainty that typically go with the utterance of disjunctions, conditionals andepistemic possibility. I will argue in the next section that this inference should not be seenas a standard Gricean implicature and, moreover, that if it is properly construed, we don’tneed epistemic qualification in the Gazdarian way.14

14It should be remarked though that Gazdar’s definition overgenerates. This holds in particular forthose implications that are read off from the complements of verbs of propositional attitude. For example,Gazdar’s definition associates with (i) the set of proto-clausals (ii):


The second variety of implicatures found in Gazdar are scalars. These are derivedfrom quantitative scales in the sense of Horn (1972). Such scales are ordered n-tuples ofexpression alternatives the form 〈α1, α2, . . . , αn〉, where each αi is a linguistic expression ofa similar type. Given such a scale a sentence ϕ containing a weaker expression will invokethe implicature that the corresponding inference containing a stronger expression from thesame scale does not hold. So given a the scale 〈all, most, many, some, . . .〉 the utterance of(8a) will implicate that not all of the children were in the garden. Such scalars normallycome unmodalized. In Gazdar’s system they come modalized in a very strong way. Theproto-implicature assigned to (8a) is (8c). By a typical utterance of (8a) the speaker isthus not said to implicate that not all of the children were in the garden, but that he knowsthat not all of the children were in the garden.

(8) a. Some of the children were in the garden.b. All of the children were in the garden.c. K¬[all of the children were in the garden]

d. ¬K[all of the children were in the garden]

e. ¬[all of the children were in the garden]

This modification has generally be felt to be unreasonably strong and has been objected toby a number of authors. Soames (1982), Van der Sandt (1988) and Horn (1989) all suggestthat, if epistemic force is assigned, it should rather look like ¬Kϕ instead of K¬ϕ. For evenif we grant that a co-operative speaker will not utter (8a) when he knows that (8b) is true,it does not follow that (8a) can only be uttered co-operatively when the speaker knowsthat (8b) is false. It is perfectly acceptable for the speaker to utter (8a), when he does nothave information about the truth-value of (8b). Under normal circumstances a hearer willconclude from an utterance of (8a) that the speaker does not know whether all children arein the garden—not that the speaker knows that not all children were in the garden. Thehearer will draw the latter conclusion only if he assumes that the speaker has completeinformation about the situation at hand. If this does not obtain, (8c) is too strong. Theinference from (8d) to (8c) thus comes about as a result of the extra assumption thatthe speaker is well-informed about the truth-value of the corresponding statements whichcontain a higher ranking expression from the relevant scale.

Hirschberg (1991: 75-79) and Levinson (2000: 77-79) suggest a weaker epistemic modi-fication in terms of the belief thus distinguishing between the epistemic commitments thatcome about as a result from the speaker’s commitments in view of what is asserted versusthe commitments that arise from what he is merely implicating. In my view this strategyis misguided and based on a confusion between different levels of interpretation. Here Iconcur with Atlas as reported in Levinson (2000: 78). Just as in the case of presupposi-tional and assertoric information, the epistemic state of the agent should not be encoded aspart of the content of his utterance or of whatever is conveyed by his utterance. StandardGricean reasoning supports this view. When a speaker utters (8a) although he is in a

(i) John hopes that he has won the prize.(ii) {P[John has won the prize], P¬[John has won the prize]}These proto-clausals will survive a typical utterance of (i) unless the context provides information to thecontrary. However, intuitively these environments do not give rise to the inference that the speaker doesnot know whether he won the prize. This is easily remedied by excluding verbs of propositional attitudefrom the contexts that generate such clausals.

1.4. ELIMINATING EPISTEMIC MODIFIERS 11

position to utter the logically stronger (8b), he violates the maxim of Quantity. The factthat a speaker utters (8a) thus suggests—in light of the maxim of Quality ‘Say nothingfor which you don’t have adequate evidence’—that the speaker does not have completeinformation about the number of children being in the garden, and therefore avoids sayingmore than he knows. Only if the hearer knows that the speaker has full information in thisrespect, can he conclude that the speaker has chosen a logically weaker expression in orderto avoid a conflict with the submaxim ‘Don’t say what you believe to be false’. Thus itis only under those circumstances that he can infer (8c) The inference on the part of thehearer that the speaker does not know, believe or whatever, comes about as a result of areasoning process which involves inferences about the knowledge, beliefs and intentions ofthe speaker. These inferences as to the epistemic status of the speaker should however notbe made part of the content of what he conveys.

1.4 Eliminating epistemic modifiers

In the previous section I argued that the encoding of various degrees of epistemic commit-ment on the part of the speaker in the presuppositional, content and implicature expressionsis based on a confusion between what features of interpretation should be handled at thelevel of speech act theory, what should be handled by the dynamics of discourse interpre-tation and what at the level of content. And even when interpreted at a metalevel theepistemic force of knowing seems too strong. Nor does such a modification capture thedifferences between presuppositional, assertoric and various kinds of implicatural informa-tion.

When we have a closer look at the system, we see that the derivation of strong presup-positions, strongly modalized assertions and implicatures is an artifact of the system whichfor its empirical predictions depends crucially on two factors, (i) the hierarchical orderingof semantic content, proto-implicatures and proto-presuppositions and (ii) the requirementof consistency for the contextual update. Let us run though the example given before:

(9) a. Either John does not own a donkey or he keeps his donkey very quiet.b. K[John has a donkey]

c. P[John has a donkey],P¬[John has a donkey]

The occurrence of John’s donkey in the second disjunct of (9a) induces the proto-presup-position (9b). And, since the first disjunct is neither entailed nor (proto-)presupposed,this clause generates the proto-implicatures given in (9c). Contextual incrementation nowproceeds as follows. Assuming that the initial context is empty, K(9a) will be enteredinto an auxiliary context, thus establishing the input context for the incrementation ofthe implicatural and presuppositional information. Since the proto-clausals are consistentwith the context just created, this context will be updated by this information thus estab-lishing the proto-clausals as actual clausals of the utterance. The resulting context is theinput for the proto-presupposition. The latter won’t survive the incrementation procedurethough, since the context established by now comprises P¬[John has a donkey] whichis inconsistent with K[John has a donkey]. The context resulting from this ordered in-crementation procedure will thus comprise the (modalized) content of the sentence, theclausals, but not the proto-presupposition. The presupposition is said to be cancelled.


Note that this cancellation process depends both on the epistemic modification of the pre-suppositional material and on the order of incrementation. For, if the proto-presuppositionwere not modalized, it would wrongly survive, since the set {P¬[John has a donkey],[John has a donkey]} is consistent. If we would retain the modalization of the proto-presupposition but would switch the ordering around, i.e. if we would first incrementthe proto-presupposition and then the implicatural information, the proto-presuppositionwould equally survive, but now cancel the proto-implicature since the set {P¬ [John has

a donkey], K[John has a donkey]} is inconsistent. This would wrongly predict that(9a) presupposes (9b) and does not give rise to the inference that the speaker does notknow whether John has a donkey. The general setup of the system thus requires both amodalizing of the proto-presupposition and the stipulation that implicatural informationis stronger than presuppositional information.

I first give a solution to the problem of strong presuppositions and assertions anddiscuss the ordering of presuppositional and implicatural information in the next section.A simple solution to the first problem is obtained by simultaneously strengthening theconsistency requirement for contextual update to the requirement of doxastic defensibility,and eliminating the epistemic modalization of proto-presupposition and assertoric update.For a similar account I refer to Beaver (2001: 65-7).15

I noted above that the epistemic commitments need not be encoded as beliefs or knowl-edge of the speaker. Commitments are social consequences that derive from the utterancesa speaker actually produces. They may be reinforced, withdrawn, a speaker may stick tohis commitments or violate them. It is the utterances a speaker produces that give rise tohis commitments and this may be accounted for by entering the content of his utterances ina set which we then interpret as his commitment slate. Such a commitment slate can bestbe seen as an ‘edited summary of a dialogue’s preceding history’ Hamblin (1987: 229) andis close to Lewis’ (1979) notion of a conversational scoreboard. The fact that the informa-tion conveyed is part of such a commitment slate will subsequently put the speaker undera variety of obligations many of which are dictated by the Gricean maxims. He should beprepared to defend his claims, not redundantly repeat information etc. The basic pointis that the fact that this information is part of his commitment slate will put him undercertain obligations. His commitment slate thus should represent the content of what heis committed to, but not contain the description of his commitments. This is roughly theway in which Hamblin introduced the notion. It follows that the content of what a speakerconveys by making assertoric utterances will be entered as a part of the context, but thatthe effect of his being committed to knowing or believing what he says should be handledon a different level. Such an interpretation is in line with the framework for speech acttheory Gazdar proposes elsewhere (Gazdar 1981). Here speech acts are essentially viewedas instructions to change the context. An assertion, for example, will change a context inwhich the speaker is not committed to the truth of ϕ to one in which he is so committed.This is close to Stalnaker (1978) where we also perceive a principled division between thecontent of what is conveyed and the various effects this gives rise to.

15Beaver (2001) proposes to change the definition of consistency toCons(X) iff {Kϕ | ϕ ∈ X} 6� ⊥This comes down to requiring epistemic defensibility and yields the same results as presented here. Itis shown below that we may dispense with the K-operator altogether since doxastic defensibility alreadycaptures Gazdar’s data.


Gazdar (1979: 47) motivates the epistemic modalization of assertoric utterances byan appeal to Hintikka’s discussion of Moore’s paradox. Moore noted the oddity of thefollowing statement.

(10) The cat is on the mat, but I don’t believe it.

Clearly, what is expressed by this sentence is consistent. A speaker may say that the cat ison the mat without believing it. However, on no occasion of use his utterance of (10) willbe felicitous. Hintikka explains the oddity by pointing out that a speaker is expected tobelieve or at least to be in a position to conceivably believe what he says. This presumptionis violated in (10). In Gricean terms: by an utterance of (10) the speaker infringes themaxim of Quality. Hintikka’s explanation runs as follows. Though (11a) is consistent (11b)can be shown to be inconsistent, which means that (11a) is necessarily unbelievable.

(11) a. ϕ ∧ ¬Bϕ

b. B[ϕ ∧ ¬Bϕ]

The central notion is doxastic defensibility. Doxastic defensibility does not require consis-tency tout court, but it requires that it is consistently possible for a speaker to believe aset of sentences.

Doxastic defensibility A set of sentences {ϕ1, ϕ2 . . . ϕn} is doxastically defensible justin case B[ϕ1 ∧ ϕ2 ∧ . . . ϕn] is consistent.

Replacing the B- by the K-operator we get the corresponding notion of epistemic defen-sibility. We may further define the notion of doxastic implication in terms of doxasticindefensibility

Doxastic implication ϕ doxastically implies ψ just in case the set {ϕ,¬Bψ} is doxasti-cally indefensible.

Given the definition of defensibility it can easily be shown that {ϕ,¬Bϕ} or the correspond-ing {ϕ,C¬ϕ} are doxastically indefensible. The minimum requirements for the accessibilityrelation are seriality and transitivity. In order to interpret the knowledge-operator we addreflexivity.

Consider a serial and transitive standard model M = 〈W,Rd, Re, V 〉, where Rd, Re ⊆W × W . Rd is serial and transitive, Re reflexive, serial and transitive and Re ⊆ Rd

(knowledge implies belief). Assume for some arbitrary w in W , M �w B[ϕ ∧ C¬ϕ]. Sincethe necessity operator distributes over the conjunction this means

(i) M �w Bϕ and(ii)M �w BC¬ϕ.

Rd is serial, so according to (i) there will be a w′(wRw′) such thatM �w′ ϕ and according to(ii) M �w′ C¬ϕ. The latter means that there will be a w′′(w′Rw′′) such that M �w′′ ¬ϕ.But R is transitive. Thus wRw′′, and according to (i) we also have M �w′′ ϕ. ThusM �w′′ ⊥. Thus for no w,M: M �w B[ϕ ∧ C¬ϕ]. Hence {ϕ,C¬ϕ} is doxasticallyindefensible.Since Re ⊆ Rd we see that K[ϕ ∧ P¬ϕ] is also inconsistent. We thus have the same resultfor the knowledge operator. The set {ϕ,C¬ϕ} is thus doxastically and {ϕ,P¬ϕ} epistem-ically indefensible. We also have that the sets {ϕ,C¬ψ} and {ϕ,P¬ψ} are respectively


Bϕ,BC¬ϕ

w

C¬ϕ

w′

ϕ,¬ϕ

w′′

∀ ∃

Figure 3: Doxastic indefensibility

doxastically and epistemically indefensible whenever ϕ � ψ. For suppose that {ϕ,C¬ψ}were defensible in this situation. This would mean that for some w,M: M �w B[ϕ∧C¬ψ],but since ϕ � ψ we would also have M �w B[ψ∧C¬ψ] which as we just saw cannot obtain.

This is all we need to capture the cancellation predictions of Gazdar when strippingthe K- and P-operators from the presuppositional, content and implicature expression.

For the moment I will retain the epistemic modification of Gazdar’s clausals. I willdiscuss their status in the next section. But I will leave out the modal operators from thepresuppositional expressions and the scalars. Nor will I encode Gazdar’s Quality impli-cature as part of the assertoric update. By the utterance of ϕ the context will thus notbe updated with the information that the speaker knows ϕ, but just with the content-expression. I will, however, strengthen the requirement for contextual update. Instead ofrequiring simply consistency I will require that contextual incrementation does not giverise to indefensibility of the resulting context. This is in line with the previous discussion.The context should not contain a description of the speaker’s commitments, but simplythe content of what he is committed to, and the latter will only consist of informationwhich he can possibly believe or know. It turns out that redefined along these lines theresulting system yields the same predictions as Gazdar’s original formulation with respectto the behaviour of presuppositions and implicatures. It will turn out that this has someadditional benefits.

I will first run through the main cases where the interaction between contextual, im-plicatural and presuppositional information yields the empirical predictions Gazdar wantsto account for and conclude with a short summary of definitions. Throughout the re-vised version I will use doxastic operators for the encoding of proto-implicatures, proto-presuppositions and the contexual update. Whatever holds for them will apply a fortiorito the stronger epistemic variants.

(i) Context and content defeats clausal and presuppositional information

Consider first the effect of prefixing the K-operator to the assertoric update. A typicalexample of cancellation would occur in a modus ponens argument. Suppose we enterKp to the context. A subsequent utterance of p → q would add K(p → q) which inthis context entails Kq. The implication will moreover invoke the following set of proto-implicatures {Pp,P¬p,Pq,P¬q}. Since both {Kp,P¬p} and {Kq,P¬q} are inconsistent,the proto-implicatures P¬p and P¬q will be cancelled. Note that we get the same resultif we skip the K-operator and require defensibility for the result of the contextual update.The sets {p,C¬p} and {q,C¬q} are indefensible and the relevant clausals will equally get


cancelled.16 Cancellation of presuppositional material by content typically occurs whenthe (unmodalized) content conflicts with the presuppositional expression. So here Gazdar’ssystem gives the desired results without any epistemic modification. Since inconsistent setsof formulae are a fortiori indefensible, the original predictions are preserved.

(ii) The interaction between clausals and scalars

Scalars may be ‘suspended’ to use Horn’s terminology.17 An example is given in (12a).Here the consequent is said to invoke the scalar that not all the boys were at the party.Since no utterance of the full sentence will invoke this scalar it is cancelled in Gazdar’sterminology.18 Gazdar achieves this effect by giving clausals priority over scalars andmodalizing scalars in the strong way discussed above. In (12a) the antecedent invokesthe clausal P¬¬(12b) which yields a conflict with the proto-scalar K¬(12b). Again wemay leave out the modalization when requiring defensibility for the update in view of theindefensibility of {¬(12b),C¬¬(12b)}

(12) a. Some, if not all of the boys were at the party.b. All the boys were at the party

(iii) Clausals defeat presuppositions

The central predictions in Gazdar arise from the interaction between clausals and presup-positional expressions. We saw that K[John has a donkey], the proto-presupposition of(9a), was defeated by the scalar P¬[John has a donkey]. It holds generally in Gazdarthat a scalar P¬ψ defeats a proto-presupposition Kϕ whenever ϕ � ψ. Eliminating theK-operator from the presupposition expression and requiring doxastic defensibility for thecontextual update yields the same result. For we just saw that {ϕ,C¬ψ} is doxasticallyindefensible whenever ϕ � ψ.

(iv) Inconsistent presuppositions defeat each other.

Since inconsistent presuppositional expressions that come unmodalized already conflictthere is no need for further modalization.

It is a simple exercise to check that the arguments can be repeated in cases where incon-sistency in Gazdar is achieved by conjoining contextually established information with thecontent expression. It thus turns out that we may safely do without the epistemic modifi-cation in content, presuppositional and scalar expressions, provided that we simultaneously

16Note, by the way, that the positive clausals will pass Gazdar’s procedure unharmed. The resultingcontext will thus contain implicatures that are simultaneously entailed by the sentence that gave rise tothem. Though this is logically not a problem, it is strange and conceptually at odds with the notion ofimplicature.

17Note that the claim that subclauses give rise to the same scalars as their unembedded variants seemsto give rise to a projection problem for implicatures. This is wrong. Implicatures are by definition notconventionally attached to clauses but are computed on the basis of the content of a full utterance. Thissimply means that for implicatures there is no projection problem in the standard sense of the word.See Geurts (1999: 19-23) for an extensive discussion. I will ignore this point here and adopt Horn’s andGazdar’s assumptions for the moment.

18It is assumed that the whole sentence has the form of a conditional.


strengthen the consistency requirement to defensibility. The resulting system will have thesame predictive power. This does not require epistemic but just doxastic defensibility. Wethus may weaken the knowledge requirement to belief.

The following definitions assume an independent definition of the set of proto-presup-positions (PP (ϕ)) and and proto-implicatures (PI(ϕ)). They are close to Beaver’s (1997)reconstruction.19 In line with the previous discussion I assume that proto-presuppositionsand proto-scalars come unmodalized, and that the modality involved in the contextualupdate and in the encoding of the clausals is serial and transitive. I leave out the orderingbetween the clausals and scalars.

Defensible incrementation

(i) DefensibilityX is doxastically defensible (Defens(X)) iff{Bϕ | ϕ ∈ X} 2 ⊥

(ii) The defensible incrementation of X with Y (X ∪� Y )X ∪� Y :=X ∪ {ϕ ∈ Y | ∀Z ⊆ Y (Defens(X ∪ Z) → Defens(X ∪ Z ∪ {ϕ}))

Contextual update of information

(i) c+ ϕ = c ∪ {ϕ} propositional content

(ii) c++ϕ = (c + ϕ)∪� PI(ϕ) implicature

(iii) c+++ ϕ = (c++ϕ)∪� PP (ϕ) presupposition

Implicature and presupposition

Implicature

19See Beaver (1997: 962-3). Beaver’s definitions are closer to Gazdar in retaining the epistemic mod-ification of the content and presuppositional expressions. This could, however, easily be remedied byadopting Beaver’s 2001 suggestion (see note 15) and incorporating epistemic modification in the definitionof consistency. On the other hand, Beaver deviates from Gazdar and the definitions given here in definingthe presuppositions of an utterance ϕ as those proto-presuppositions that are in the final context but notentailed by ϕ. This has the advantage that an utterance of (i) does not presuppose (iii), but has thedisadvantage that the same holds for non-compound sentences like (ii), since these are supposed to entailthe presuppositional information.(i) John owns a donkey and keeps his donkey in hiding.(ii) John keeps his donkey in hiding.(iii) John owns a donkey.

Beaver 2001 (pp. 59-61) follows Gazdar in defining the presuppositions of an utterance as those proto-pre-suppositions that make it to the the output context, which gives exactly the opposite results. One would,however, want to retain the presupposition in (ii) but not in (i). A provisional solution is to adopt a two-dimensional approach (Karttunen and Peters (1979), Lerner and Zimmerman (1981)) and to distinguishbetween (a) the presuppositional information that is invoked in view of the presuppositional expression(Gazdar’s proto-presupposition), (b) its counterpart in the content expression, and then take care that theformer does not survive in (i). In the next section I will suggest a solution along these lines in a revisedGazdarian model (limited to a propositional language and neglecting the binding problem).

1.5. THE STATUS OF GAZDARIAN CLAUSALS 17

(i) ϕ implicates ψ in c iff

a. ψ ∈ PI(ϕ)b. ψ ∈ c+++ {ϕ}

Presupposition

(ii) ϕ presupposes ψ in c iff

a. ψ ∈ PP (ϕ)b. ψ ∈ c+++ {ϕ}

Before I discuss the status of Gazdarian clausals two points should be made. The firstconcerns the status of Gazdarian presuppositions, the second the status of implicaturalinformation. As to the first point, the definitions given above allow us to isolate thepresuppositional information that is conveyed by an utterance but that is not alreadypart of the incoming context. That is, in a more current terminology, the material thatis (to be) accommodated. Let me stress again that this does not capture the intuitivenotion of presupposition as information that is required to assign an interpretation to asentence. As to the second point, one would like to have a principled distinction betweenpresupposed and implicated material. Though presupposed material will normally be partof the incoming context, implicated material won’t. It clearly makes no sense to conveyinformation by way of implicature if this information is already taken for granted by theparticipants. However, as the system stands, neither of these distinctions can be made.The new information conveyed by ϕ in c is (c+++{ϕ})\c. But the presuppositions of ϕ inc are simply the proto-presuppositions of ϕ that are part of c+++{ϕ}, and the implicaturesof ϕ in c are those proto-implicatures of ϕ that satisfy the very same condition.

1.5 The status of Gazdarian clausals

In the previous section I noted that the strong epistemic modalization of the differentpieces of informational content in Gazdar’s original system is an artifact which is dictatedby the requirement of inconsistency between content, implicature and presuppositionalexpressions in order to account for cancellation effects. I showed that the same effect canbe achieved by eliminating epistemic modification and simultaneously strengthening therequirement on defeasible incrementation from mere consistency to doxastic defensibility.The reconstruction given above showed that this allows us to dispense with all epistemicmodification but for the clausals.

In this section I will discuss the status of Gazdar’s clausals, the counterintuitive stipu-lation of priority of clausal implicatures over presuppositionally induced information, andI will suggest a reinterpretation which does not conceive of clausals as conversational im-plicatures.

Gazdarian clausals are invoked by embedded sentences that are not entailed by theirmatrix. They thus do the work they are supposed to do in cases like (1a), since in herethe clausal that he may not have a donkey conflicts with the presuppositional expression.However, in a paradigmatic case like (13a) no clausals are invoked and the presupposi-tional expression thus passes unharmed. (13a) is thus on a par with (13b) which equallypresupposes that Jo


(13) a. John owns a donkey and keeps his donkey in hiding.b. John keeps his donkey in hiding.

By an utterance of (13b) a speaker will typically presuppose that John owns a donkey.When uttering (13a) he never will. A Stalnaker type explanation is straightforward. Whenuttering the first conjunct of (13a) we indicate that we don’t take the information that Johnowns a donkey for granted. When uttering the second conjunct we do. At the momentwe interpret the first conjunct, the context changes since we add asserted informationto the conversational background. By its very utterance its informational status thuschanges. It counts as presuppositional from now on and we may rely on the informationjust introduced in a presuppositional way. The crucial point is that given the assumptionthat presuppositional information should ideally be part of the incoming context, (13a) cannever be interpreted as being presupposing as this would defeat the point of the assertionof the first conjunct. Thinking along Gazdarian lines we thus would like to have a means todefeat the presupposition in (13a) but not (13b). Note that clausals as conceived by Gazdarwill not do the job. For firstly, the first conjunct of (13a) is entailed by the full sentencewhich means that no (proto-)clausal will be invoked and the incrementation procedure willthus leave presuppositional expression induced by the second conjunct of (13a) unharmed.And note, secondly, that given the interpretation of a context as a commitment slate thereis no way to redefine the system so as to invoke such clausals after all. A Gazdarian clausalcarries the information that both the relevant clause and its negation are compatible withthe knowledge of the speaker. This would clearly be an unfortunate condition to imposeon assertoric acts, and it would moreover violate Grice’s Quality maxim. Instead we wouldlike to say that an assertion is only felicitous in a context which is compatible with thetruth and the falsity of the relevant clause. But in order to say this we have to abandonthe interpretation of a context as the commitment slate of a speaker. If we view a contextnot in the Hamblin/Gazdar way as the information the speaker is committed to, butas a common ground in the Stalnaker/Karttunen sense, we can get the desired effect byadopting Stalnaker’s principle that an utterance must be compatible with, but not entailedby, the incoming context.20 If we furthermore associate with each embedded clause a localcontext in the sense of Karttunen and apply the principle with respect to the local contextof each subclause of a compound sentence (thus requiring that each subclause shouldbe compatible with but not entailed by its local context21) we simultaneously capturethe purported cancellation effects Gazdar captures by invoking clausals in non-entailingenvironments. Note that such a reinterpretation allows us to dispense with the epistemicmodification of the purported clausals.22

20See Stalnaker (1978)21Though the terminology might suggest a non-trivial similarity, this is in fact very different from

Karttunen 1974 and other satisfaction theorists, who require that the presuppositional expressions of eachsubclause are entailed by their local context

22This is the strategy followed in Van der Sandt (1988) who presents a projection mechanism that isinspired by Stalnaker. The same principles figure in a modified way in his (1992) anaphoric DRT-basedaccount.

1.5. THE STATUS OF GAZDARIAN CLAUSALS 19

Clausals are not implicatures

To understand why the ordering as stipulated by Gazdar does the work it is supposedto do, we have to see that Gazdarian clausals should not be construed as implicatures.Clausals differ in status from standard conversational implicatures in the Gricean sense.Conversational implicatures are typically computed on the basis of the incoming context,the content of the sentence to be processed and Gricean principles. They are typicallyincremented in the output context. Gazdarian clausals, I claim, differ in exactly thisrespect. They derive from pragmatic constraints on contexts, that is, constraints on inputcontexts.

To illustrate this point consider the difference between the implicatures invoked by sen-tences like (8a) and the suggestion of uncertainty that goes with disjunctions. Scalars andother types implicatures come with an utterance. They are, to use Levinson’s terminology,‘essentially gratuitous’.23 In this respect they differ both from presuppositional inferencesand clausals. The latter come about as inferences that result from contextual requirements,that is requirements the input context has to fulfill to accept an utterance as felicitous.The difference in cancellation behaviour illustrates this. In view of their very nature scalarsmay be easily defeated by conflicting or unsupportive assumptions as following examplefrom Levinson 2000 illustrates:

(14) A. Is there any evidence against them?B. Some of their identity documents are forgeries.

Here I would rather say that the purported implicature that not all of their documentsare forgeries is not even invoked. The suggestion of epistemic uncertainty that is typicalfor clausals is not so easily defeated, however. In fact, normal assertoric uses of sentencesinvoking clausals are never felicitous when the common ground conflicts with membersof the clausal set. Consider the disjunction (1a) and suppose that the truth of the firstdisjunct were taken for granted. This would not defeat the suggestion of uncertainty butinstead result in sheer uninterpretability. The underlying motivation is Gricean. For inthis situation the full disjunction would convey the same information as the mere utteranceof the second disjunct would have done. And the utterance of just the second disjunct isthus clearly a more efficient way to convey the desired information. The only examplesof purported cancellation of clausals I am aware of are examples of non-information ori-ented discourse, that is types of discourse in which the assumption that the conversationalmaxims are in force, is lifted. This happens for example in a modus ponens argument, orexam situations. However, such situations where conversational presumptions are lifted inadvance are very different from the cases of e.g. scalars where an implicature is computedon the basis of the asssumption that the maxims are in force.

Conceiving of a context as a common ground we may interpret a Gazdarian clausal set{Pϕ,P¬ϕ} as signalling that both Pϕ and P¬ϕ are compatible with but not entailed bycontextual information. Geurts (1995, 1999) pointed out that Gazdar’s clausals can be seenas unresolved issues in a discourse constraining the process of presupposition projection.24

23See the discussion in Levinson (2000: 60-63)24Just as Gazdar’s thesis Geurts’ thesis was written under the supervision of Hans Kamp. It was

completed in 1995. References are to his 1999 book which is an expanded and reworked version of thiswork.


Both in (1a) and in (15) the clausal set indicates that it is not known (believed, assumed)whether John owns a donkey.

(15) It is possible that John owns a donkey.

Instead of saying that disjunctions as in (1a) or epistemic modals as (15) invoke impli-catures we better say that they signal an issue in the ongoing discourse. Such an issuemay be addressed explicitly and be settled by the claim that John owns a donkey or bythe claim that he does not. Note that if an utterance signals in whatever way that thediscourse context contains some issue, it would be self-defeating if the utterance of the verysame sentence would settle this issue by presuppositional means. For example, in (1a) thiswould imply that the speaker both indicates that he does not know whether Johns ownsa donkey and assumes that the issue whether John owns one has already been settled.Standing issues in a discourse thus may prevent presuppositions from being projected tothe main context. This gives rise to a simple constraint on presupposition projection:

A presupposition may not be processed in such a way that the resulting struc-ture would settle an issue.

This view is supported by the now well-established account of information structure thatassumes a direct correlation between topics, questions and the acceptability of discourseunits.25 On this view a discourse owes its coherence to the fact that utterances can beviewed as answers to explicit or implicit questions. And the view is supported by the factthat there is a straightforward correlation between question/answer pairs and the infor-mation structure of a sentence. In the case of wh-questions this correlation is reflectedin focus/background division of a sentences. By lambda-abstracting over the non-focusedpart we obtain the focal background and by binding the relevant variable by a wh-operatorwe obtain the question. This question may be contextually given in which case the ut-terance that answers should be congruent with the question to be felicitous, that is thefocus of the answer should correspond with the wh-position in the question. If no ques-tion is found in the incoming context, the focus/background division provides the clue forthe interpreter to construct the question the utterance is supposed to answer. He will, touse to the presuppositional terminology, accommodate it. It is with respect to such anaccommodated question that the original utterance expresses an answer. The relevance ofthis fact for presupposition theory has already been observed by Strawson (1964). Withrespect to disjunctions Grice characterized the situation as follows:

A standard employment of ‘or’ is in the specification of possibilities [...] each ofwhich is relevant in the same way to a given topic. ‘A or B’ is characteristicallyemployed to give a partial [...] answer to some “W”-question to which eachdisjunct, if assertible, would give a fuller, more specific satisfactory answer.

Grice1989: 68

Simons (2001) develops this idea in a Hamblin-type theory of questions. She points out thatGrice’s topic can be characterized as the question the disjunction is supposed to partiallyanswer and she shows in detail that a disjunction can only provide an (informative) answerto a question in case each disjunct provides a (partial) answer to a question. Note that

25See for example van Kuppevelt (1995) or Ginzburg (1995).

1.6. BRIDGING TO QUESTIONS IN DISCOURSE 21

(1a), the disjunction we started with, is a special case. It is typically an answer to thefollowing yes/no-question:

(16) Does John own a donkey?

Note furthermore that in a partition analysis like Groenendijk and Stokhof’s26 the cellsof the partition induced by the interrogative correspond to the Gazdarian clausals if weinterpret these as signalling that both the information that John owns a donkey and thathe does not own a donkey are compatible with the context. And, note finally that suchan interpretation of clausals as contextually given issues or questions does not construethem as conversational implicatures, but typically interprets them as constraints on input-contexts. We will see that a proper appreciation of this fact suggests a simple solution tothe problem posed by the so-called bathroom sentences.

1.6 Bridging to questions in discourse

Evans (1977) pointed out that sentences like (17), that is the pronominalized version of(1a), are problematic for theories that treat pronouns as bound variables.27

(17) Either John does not own a donkey or he keeps it very quiet.

They are equally problematic for current theories of dynamic interpretation.28 Though theaccommodation mechanism of the anaphoric account of presupposition yields the correctresult for (1a) by predicting that the presuppositional anaphor his donkey will be locallyaccommodated in the representation of the second disjunct, it is unable to handle itspronominal counterpart. The reason is simple. On the anaphoric account presuppositionalconstructions allow accommodation since they normally carry enough descriptive materialto establish an antecedent. In the present case the information that John owns a donkeycan be retrieved from the full presuppositional construction. This material may be usedto create an antecedent at some accessible level of discourse structure in case the discoursedoes not already provide one. The explanation for (1a) now runs as follows. The firstconjunct is not accessible for anaphoric reference which precludes binding. In order to getan interpretation we thus have to accommodate. Global accommodation is not an option.Accommodation of the presuppositional information at the main level of discourse wouldviolate conditions on discourse acceptability: it would result in a structure that correspondswith the following infelicitous discourse

26See Groenendijk & Stokhof (1984)27The problem was first is pointed out and discussed by Geach (1962: 129-30). The example that has

become famous in the linguistic literature is Partee’s ‘Either there is no bathroom in this house or it is ina funny place.’ Hence the label ‘bathroom’-sentence.

28Krahmer & Muskens (1995) give a technical solution by reducing the problem posed by ‘bathroom’-sentences to the fact that double negation creates inaccessible contexts in DRT. They note that a disjunc-tion can be rendered as an implication which yields ¬¬ϕ1 → ϕ2 for ¬ϕ1 ∨ϕ2 and then present a semanticsfor DRT in which double negation acts as a ‘hole’ for anaphoric reference. This reduces the disjunctionproblem to the double negation problem. There are several problems with this proposal though. One isthat accessibility in such contexts does not always require that the negative element(s) are explicitly given,as pointed out by Geurts (1999: 74 ff and 128 ff) . A second problem is that it does not generalize to theepistemic lists of alternatives such as in (20a) as pointed out by Van Rooy (2001: 639).


(18) John owns a donkey. Either he does not own a donkey or he keeps it very quiet.

The plain assertion of ‘John owns a donkey and keeps it very quiet’ would have conveyedthe same information as (17) but in a less redundant way. The only accessible positionleft is the second disjunct and the material will thus be accommodated there, yielding aninterpretation which can be paraphrased as:

(19) Either John does not own a donkey or he owns a donkey and keeps it very quiet.

This correctly accounts for the interpretation of (1a). However, in conjunction with theclaim that presuppositional expressions are essentially anaphors this solution is problem-atic. If presuppositional expressions are treated on a par with pronouns it is imperative toprovide a unified treatment of both (1a) and (17), that is, we have to give a uniform expla-nation for the interpretation of the explicit presuppositional construction and its pronom-inalized counterpart. Since pronouns are devoid of descriptive content and thus lack thecapacity to accommodate, the explanation given for (1a) does not carry over to (17). Inthe latter the pronominal variable will not find an accessible antecedent and thus remainfree leaving this structure without an interpretation. The problem is aggravated when wesee that the phenomenon is not confined to disjunctions, but also occurs in conjunctionsas the following:29

(20) a. It is possible that John does not own a donkey, but it is also possible that hekeeps his donkey very quiet.

b. It is possible that John does not own a donkey, but it is also possible that hekeeps it very quiet.

As in the disjunctive case, an analysis by means of accommodation yields the right resultin (20a) but is not applicable in the case of (20b). And again we intuitively perceive a linkfrom the anaphoric expression in the second conjunct to a formally inaccessible positionin the first. We would thus, as in the disjunctive case, like to have an explanation whichholds both for the case where the anaphoric expression is given by a full description andfor its pronominalized counterpart. Note, moreover, that the modal examples are similarto the disjunctive counterparts in that the conjuncts give a specification of two alternativepossibilities which typically give an answer to the same yes/no-question. And note finallythat in both cases the specification of the possibility that John does not own a donkeyand the possibility that he owns one but keeps it in hiding, should be exhaustive withrespect to the knowledge of the speaker in order for him to know that the full sentence istrue. The parallelism between disjunctions on the one hand and modalized conjunctionsas in (20) on the other is made explicit in Zimmerman (2000) and Geurts (ms.) who arguethat disjunctions can be analyzed as (exhaustive) lists of epistemic possibilities. On thisaccount the logical form of ‘not ϕ or χ’ comes out as ♦¬ϕ ∧ ♦χ and (1a) may be thusparaphrased as (20b). Adopting this analysis allows us to give a uniform analysis of thedisjunctive case and such a list of epistemic conjuncts.

I want to suggest that the key to the solution is found in Kamp & Reyle (1993:187-90).They observe that a disjunction of the form ‘A or B’ can be paraphrased as ‘A or otherwiseB’. In such a paraphrase of (1a) the otherwise would refer to the other case, that is the case

29These examples originate in Liberman (1973), and have been extensively discussed in the earlierpresupposition literature.

1.6. BRIDGING TO QUESTIONS IN DISCOURSE 23

described by the first disjunct. According to Kamp and Reyle this licenses us to postulate arepresentation for what is referred to by otherwise and then extending that representationby incorporating it in the representation of [:he keeps it quiet]. Geurts (1999: 75)adopts this idea and proposes to insert the material into the second conjunct as an inferredantecedent. Both explanations are incomplete though. On Kamp and Reyle’s view thequestion arises where the material encoding ‘the other case’ comes from. It cannot simplybe constructed as the complement of the first disjunct for then we would end up with adouble negation and thus an inaccessible antecedent after all. We thus have to assumethat ‘the other case’ refers to the corresponding unnegated sentence and that the materialthus retrieved may be inserted. And on Geurts’ account we are faced with the question:inferred in what way? However, if we take the remarks by Grice (p. 20 above) seriouslyand once we see that the disjunction is to be interpreted as an answer to the (explicitlygiven or implicitly assumed) yes/no-question ‘Does John own a donkey?’, the answer isstraightforward. In order to interpret the disjunction the interpreter has to accommodatethe question. And adopting the standard Hamblin-type account of questions we find thealternative possibilities in the cells of the partition of the question thus accommodated.

To illustrate such an analysis let us first consider the non-disjunctive case:

(21) a. Does John own a donkey? *He keeps it very quiet.b. Does John own a donkey? Sure. But he keeps it very quiet.

The first discourse is out. The question does not imply that the speaker assumes thatthere is a donkey and the indefinite is not available for pronominal reference. The seconddiscourse is fine. The pronoun gets its value from the answer which in turn is one of thecells of the bipartition the question induces. This is uncontroversial.

The situation in the disjunctive case is basically the same. The second disjunct givesan answer to the question by being in the second cell of the partition. Together with theincoming context this cell figures as a provisional context for the interpretation of thisdisjunct. It follows that the use of the pronoun is licensed since John’s donkey inhabitseach of the worlds of this provisional context. The second disjunct thus gets a determinateinterpretation in its provisional context which entails that John owns a donkey. Theproposal I want to make thus agrees both with Kamp and Reyle that otherwise refers tothe other case and with Geurts that the interpretation is secured in view of a bridginginference. But unlike these authors I want to claim that ‘the other case’ is not constructedout of the first disjunct. Instead ‘the other case’ is given by one of the cells of the partitioninduced by the antecedently constructed question. Of course this does not imply thatthe indefinite in the question is accessible for anaphoric reference. In order to interpretthe disjunction we have to turn to the cells of the bipartition individually. The firstdisjunct corresponds with the cell which presents the negative answer. We then turn tothe alternative cell, thus deriving

(22) If John owns a donkey, he keeps it very quiet.

as the bridging inference.If this analysis is on the right track the relevant interpretation does not come about

by some anaphoric link across the disjuncts, nor is it due to straightforward insertion ofmaterial from the first disjunct into the second. The relation is indirect in that the fulldisjunction allows us to create some question by accommodation. It is the question thus


accommodated which figures as the context of interpretation of the disjunction and whichallows us to interpret the pronominal element.

1.7 Conclusion

In 1979 Hans Kamp gave the following formulation of the received view on the relationbetween semantics and pragmatics in a theory of meaning:

(i) The concept, or concepts, characterized by the recursive component of the theorybelongs to semantics.

(ii) Moreover no notion to which this component refers belongs to pragmaticsKamp 1979: 265

He went on to state and argue that there is ‘no good reason why semantics and pragmaticsshould be separate in the sense that [the principle stated above] tries to articulate’. Kamp’sargumentation was based on the analysis of non-assertoric speech acts. The same view,but now based on the analysis of generalized conversational implicatures, was recently putforward forcefully by Steve Levinson in Levinson (2000).

In this paper I discussed Gazdar’s pragmatic theory as a particulary elegant elaborationof the received view. Gazdar’s theory is based on the assumption that semantic content canbe determined without calling on pragmatic principles. A sentence in a context determinesa proposition in the classical sense. Pragmatic processing only enters the scene after thissemantic object has been established thus allowing us to determine a more substantial in-formational object, the information conveyed. Subsequent work in discourse representationtheory and, in particular, the developments in presupposition theory and related areas inthe nineties have shown that this picture is in need of a thorough revision. The idea thatthe semantic and pragmatic components of a theory of meaning can be seen as separatemodules in the sense that a sentence in a context yields a fully determined propositionwhich may then be ‘enriched’ or ‘strenghtened’ by pragmatic inferencing, thereby onlyaffecting the output context, turned out to be untenable. Instead, at the very moment anutterance is processed principles of a Gricean nature play a crucial role in determining thebackground with respect to which this utterance will get an interpretation. This is abun-dantly clear when we have a closer look at the all pervasive phenomenon of presuppositionaccommodation, and, in particular, when we look at the factors that constrain this process.

In this paper I went back in history, took Gazdar’s modular view as a starting point andshowed first that, even when we adhere to Gazdar’s original architecture, some obstaclescan easily be removed. I showed that there is no reason to stick to the strong epistemicmodalization of the various semantic and pragmatic components and, furthermore, thatthere is no need to retain the counterintuitive ordering between implicatures and presuppo-sitions. Adopting a dynamic view on meaning and reinstating presuppositional informationas information that is contextually required to establish an interpretation, I showed thatGazdarian clausals should not be conceived of as standard Gricean implicatures. Theyshould, instead, be seen as constraints on input contexts regulating the resolution processof inter alia presuppositional information. I suggested that they are better conceived ofas questions or unresolved issues in discourse. And I showed that it is exactly the possi-bility of such a reinterpretation that explains why Gazdar’s original clausals did the work

1.7. CONCLUSION 25

they were devised for. This illustrates why the semantic autonomy hypothesis cannot beupheld. In cases where presuppositional information is not yet there, we need to accom-modate. Accommodation is an operation on input contexts and this process is constrainedby factors which are uncontroversially pragmatic and Gricean in nature.


Bibliography

Asher, N. and Lascarides, A. (1998). The semantics and pragmatics of presupposition.Journal of Semantics, 15:239–300.

Beaver, D. (1995). Presupposition and Assertion in Dynamic Semantics. PhD thesis,University of Edinburgh.

Beaver, D. (1997). Presupposition. In van Benthem, J. and ter Meulen, A., editors,Handbook of Logic and Language, pages 939–1008. Elsevier, Amsterdam.

Beaver, D. (2001). Presupposition and Assertion in Dynamic Semantics. CSLI, Stanford.

Bos, J. (1999). The treatment of focus particles in underspecified discourse representa-tions. In Bosch, P. and van der Sandt, R., editors, Focus. Linguistic, Cognitive, andComputational Perpectives, pages 121–141. Cambridge University Press, Cambridge.

Evans, G. (1977). Pronouns, quantifiers and relative clauses I. Canadian Journal ofPhilosophy, 7:467–536.

Frege, G. (1892). ‘Uber Sinn und Bedeutung’. Zeitschrift fur Philosophie und philosophischeKritik, 100:25–50. Translation by Max Black in Geach and Black (1970): 56–78.

Gazdar, G. (1979). Pragmatics: Implicature, Presupposition, and Logical Form. AcademicPress, New York.

Gazdar, G. (1981). Speech act assignment. In A.K. Joshi, B. W. and Sag, I., editors,Elements of Discourse Understanding. Cambridge University Press, Cambridge.

Geach, P. T. (1962). Reference and Generality. An Examination of some Mediaeval andModern Theories. Cornell University Press, Ithaca.

Geurts, B. (1995). Presupposing. PhD thesis, University of Stuttgart, Stuttgart.

Geurts, B. (1996). Local satisfaction guaranteed: a presupposition theory and its problems.Linguistics and Philosophy, 19:259–294.

Geurts, B. (1999). Presuppositions and Pronouns. Elsevier, Amsterdam.

Geurts, B. (2003). Entertaining alternatives: disjunctions as modals. ms.

Geurts, B. and Maier, E. (2003). Layered DRT. ms.

Ginzburg, J. (1995). Resolving questions I and II. Linguistics and Philosophy, 18:459–527.

27

28 BIBLIOGRAPHY

Grice, P. (1989). Studies in the Ways of Words. Harvard University Press, Cambridge,Mass.

Groenendijk, J. and Stokhof, M. (1984). On the semantics of questions and the pragmaticsof answers. In Landman, F. and Veltman, F., editors, Varieties of Formal Semantics,pages 143–170. Foris, Dordrecht.

Hamblin, C. (1987). Imperatives. Blackwell, Oxford.

Hamblin, C. L. (1971). Mathematical models of dialogue. Theoria, 37:130–155.

Heim, I. (1982). The Semantics of Definite and Indefinite Noun Phrases. PhD thesis,University of Massachusetts, Amherst.

Heim, I. (1983). On the projection problem for presuppositions. In Barlow, M., Flickinger,D. P., and Wescoat, M. T., editors, Proceedings, Second West Coast Conference onFormal Linguistics, pages 114–125, Stanford, California. Stanford Linguistics Associ-ation.

Hintikka, J. (1962). Knowledge and Belief. Cornell University Press, Ithaca, New York.

Hirschberg, J. (1991). A Theory of Scalar Implicature. Garland, New York.

Horn, L. R. (1972). On the semantic properties of logical operators in English. PhD thesis,University of California, Los Angeles.

Horn, L. R. (1989). A Natural History of Negation. Chicago University Press, Chicago.

Kamp, H. (1979). Semantics versus pragmatics. In Guenthner, F. and Schmidt, S., edi-tors, Formal Semantics and Pragmatics of Natural Language, pages 255–287. Reidel,Dordrecht.

Kamp, H. (1981). A theory of truth and semantic representation. In Groenendijk, J. A.,Janssen, T., and Stokhof, M., editors, Formal Methods in the Study of Language. Foris,Dordrecht.

Kamp, H. (2001). Computation and justification of presuppositions. In Bras, M. and Vieu,L., editors, Semantics and Pragmatics of Discourse and Dialogue. Experimenting withcurrent theories. Elsevier, Amsterdan.

Kamp, H. and Reyle, U. (1993). From Discourse to Logic: Introduction to ModeltheoreticSemantics in Natural Language, Formal Logic and Discourse Representation Theory,Vol. 1. Kluwer Academic Publishers, Dordrecht.

Karttunen, L. (1974). Presuppositions and linguistic context. Theoretical Linguistics,1(181-194).

Karttunen, L. and Peters, S. (1979). Conventional implicature. In Oh, C. and Dinneen,D. A., editors, Syntax and Semantics II: Presupposition, pages 1–56. Academic Press,New York.

Krahmer, E. (1998). Presupposition and anaphora. CSLI Publications, Stanford.

BIBLIOGRAPHY 29

Krahmer, E. and Muskens, R. (1995). Negation and disjunction in discourse representationtheory. Journal of Semantics, 12:357–376.

Lerner, J.-Y. and Zimmermann, T. (1981). Mehrdimensionale Semantik. Die Presuppositio-nen und die Kontextabhangigkeik von ‘Nur’. Technical report, Universitat Konstanz,Sonderforschungbereich Linguistik, 99, Konstanz.

Levinson, S. C. (1983). Pragmatics. Cambridge University Press, Cambridge.

Levinson, S. C. (2000). Presumptive Meanings. The MIT Press, Cambridge, Massachusetts.

Lewis, D. K. (1979). Scorekeeping in a language game. Journal of Philosophical Logic,8(3):339–359.

Liberman, M. (1973). Alternatives. In Papers from the 9th Regional Meeting, pages 356–368, Chicago. Chicago Linguistic Society.

Sæbø, K. J. (1996). Anaphoric presuppositions and zero anaphora. Linguistics and Phi-losophy, 19:187–209.

Simons, M. (2001). Disjunctions and alternativeness. Linguistics and Philosophy, 24:597–619.

Soames, S. (1982). How presuppositions are inherited: a solution to the projection problem.Linguistic Inquiry, 13:483–545.

Stalnaker, R. C. (1974). Pragmatic presuppositions. In Munitz, M. K. and Unger, P.,editors, Semantics and Philosophy. Academic Press, New York.

Stalnaker, R. C. (1978). Assertion. In Cole, P., editor, Syntax and Semantics Vol. 9:Pragmatics. Academic Press, New York.

Strawson, P. (1964). Identifying reference and truth-values. Theoria, 30:96–118.

Strawson, P. F. (1950). On referring. Mind, 59:320–344.

van der Sandt, R. A. (1988). Context and Presupposition. Croom Helm, London.

van der Sandt, R. A. (1992). Presupposition projection as anaphora resolution. Journal ofSemantics, 9:333–377.

van Kuppevelt, J. (1995). Discourse structure, topicality and questioning. Journal ofLinguistics, 31:109–147.

van Rooy, R. (2001). Exhaustivity in dynamic semantics: referential and descriptive pro-nouns. Linguistics and Philosophy, 24:621–657.

Zeevat, H. (1992). Presupposition and accommodation in update semantics. Journal ofSemantics, 9:379–412.

Zimmerman, T. E. (2000). Free choice disjunction and epistemic possibility. NaturalLanguage Semantics, 8:255–290.

Documents

Chapter 1 Pragmatic strategies - Paris Diderot Universityamsili/papers/van... · 2004-05-04 · Chapter 1 Pragmatic strategies Rob van der Sandt University of Nijmegen [email protected]