Links without LocationsInformation Packaging: From Cards to BoxesHerman Hendriks1Utrecht Institute of Linguistics OTSUtrecht UniversityTrans 103512 JK UtrechtThe Netherlandsphone: +31-30-2536183fax: +31-30-2536000ILLC/Department of PhilosophyUniversity of AmsterdamNieuwe Doelenstraat 151012 CP AmsterdamThe Netherlandsphone: +31-20-5254509fax: +31-20-5254503e-mail: Herman.Hendriks@let.ruu.nlThis paper is organized as follows. First, in Section 1, an outline is given ofthe theory of information packaging|i.e., the structuring of propositionalcontent in function of the speaker's assumptions about the hearer's informa-tion state|as it is presented by Vallduv�� (1992, 1993, 1994), who identi�esthe informational primitives focus and ground, link and tail, adapted fromthe traditional pragmatic focus/ground and topic/comment approaches, and1. The present paper is a merged, updated and extended version of `Links without Lo-cations' and `Information Packaging: From Cards to Boxes', which appeared in P. Dekkerand M. Stokhof (eds.), Proceedings of the Tenth Amsterdam Colloquium, Institute of Lan-guage, Logic and Computation, University of Amsterdam, pp. 339{358, and in T. Gallowayand J. Spence (eds.), Proceedings of Semantics And Linguistic Theory VI . CLC Publica-tions, Ithaca, New York, pp. 75{92, respectively. Section 3 of the present paper is largelybased on joint work of Paul Dekker (ILLC/Department of Philosophy, University of Am-sterdam) and the author, who would like to thank Paul Dekker, Elisabet Engdahl, FritzHamm, Sieb Nooteboom, Tanya Reinhart and Enric Vallduv�� for cooperation, discussionand stimulation. 1

who concludes|as is explained in Section 2|that the exploitation of infor-mation states of hearers by the information-packaging strategies of speakersreveals that these states have at least the internal structure of a systemof �le cards along the lines of Heim (1982, 1983): links, which correspondto what are traditionally known as topics and which are typically markedby L+H* pitch accents in English, say where|on which �le card|the fo-cal information goes, and tails indicate how it �ts there. This conclusionis challenged in Section 3, where it is argued that it begs the question. If�le card systems are assumed, then the information-packaging strategies doseem to contribute to e�cient information exchange|however, the questionis whether this assumption itself is justi�ed. Moreover, it is shown that theidea that links specify a locus of update in information states that are sys-tems of �le cards is problematic for various reasons. Therefore, Section 4o�ers an alternative account in terms of the discourse representation struc-tures of Discourse Representation Theory (see Kamp 1981, Kamp and Reyle1993), which are ontologically less committed than the `dimensionally richer'�le card systems, since discourse representation structures do not come withlocations. The latter aspect raises the question what purpose links do serveif they do not serve to specify a locus of update: a di�erent perspective onthe function of links is required. According to the perspective o�ered in Sec-tion 4, linkhood|and hence L+H* pitch accent in English|serves to signalnon-monotonic anaphora: the discourse referent Y of a link is anaphoric toan antecedent discourse marker X such that X 6� Y. This hypothesis a�ectsa wide range of phenomena. In addition to its contribution to an analy-sis of (non-)association with focus, it is shown to subsume `non-identity'anaphora, contrastive stress, correction, pronoun referent resolution, andrestrictiveness of relatives and adjectives. In Section 5, �nally, it is pointedout that the account of links given here is consistent with and can actuallybe considered a partial execution of the intonational-informational researchprogram that is outlined in Pierrehumbert and Hirschberg (1990).1 Information PackagingThe notion of information packaging is introduced in Chafe (1976):[The phenomena at issue] have to do primarily with how themessage is sent and only secondarily with the message itself,just as the packaging of toothpaste can a�ect sales in partial 2

independence of the quality of the tooth paste inside. (Chafe1976: 28)The basic idea is that speakers do not present information in an unstructuredway, but that they provide a hearer with detailed instructions on how tomanipulate and integrate this information according to their beliefs aboutthe hearer's knowledge and attentional state:To ensure reasonably e�cient communication, [t]he speakertries, to the best of his ability, to make the structure of his ut-terances congruent with his knowledge of the listener's mentalworld. (Clark and Haviland 1977: 5)On all levels the crucial factor appears to be the tailoringof an utterance by a sender to meet the particular assumedneeds of the intended receiver. That is, information packagingin natural languge re ects the sender's hypotheses about thereceiver's assumptions and beliefs and strategies. (Prince 1981:224)For instance, sentences such as (1) and (2) are truth-conditionally equivalentin that they express the same proposition, but each of them `packages' thisproposition in a prosodically di�erent way:2The teacher loves ice cream(1) The teacher loves ice cream(2)Typically, speakers will use (1) if the hearer at the time of utterance knowsnothing about or is not attending to the teacher's relation to ice cream, whilethey will use (2) if the hearer at the time of utterance knows that there existsa relation between the teacher and ice cream, is attending to this relation,but does not know what it is. Apparently, speakers are sensitive to suchdi�erences in the hearer's knowledge and attentional state, and hearers relyon this:2. Italics are used for unaccented expressions; small caps for expressions that bear a(focal) H* pitch accent; and boldface for expressions that bear a L+H* pitch accent.This is the terminology of Pierrehumbert (1980). H* accent and L+H* accent are calledA accent and B accent, respectively, in Jackendo� (1972). 3

speakers not using this device systematically give their listen-ers a harder time. (Nooteboom and Terken 1982: 317)Truth-conditionally equivalent sentences which encode di�erent informationpackaging instructions are not mutually interchangeable salva felicitate in agiven context of utterance: for example, of the above sentences, only the �rstone is a felicitous answer to the question What does the teacher love? It isthis context-sensitivity that has traditionally placed information packagingwithin the realm of pragmatics, where two in uential approaches can be dis-tinguished, the `topic/comment' approach and the `focus/ground' approach.According to the focus/ground approach, sentences consist of a `focus'and a `ground'.3 The focus is the informative part of the sentence, the partthat (the speaker believes) makes some contribution to the hearer's mentalstate. The ground is the non-informative part of the sentence, the part thatanchors the sentence to what is already established or under discussion in(the speaker's picture of) the hearer's mental state. Although sentences maylack a ground altogether, sentences without focus do not exist.The topic/comment (theme/rheme) approach splits the set of subex-pressions of a sentence into a `topic', the|typically sentence-initial|partthat expresses what the sentence is about, and a `comment', the part thatexpresses what is said about the topic. Topics are points of departure forwhat the sentence conveys, they link it to previous discourse. Sentences maybe topicless: so-called `presentational' or `news' sentences consist entirely ofa comment.In Reinhart (1982), it is argued that the dimension of `old'/`new'information is irrelevant for the analysis of sentence topics. Instead, thenotion of `pragmatic aboutness' is de�ned in terms of the organization ofinformation. The set PPA(S) of Possible Pragmatic Assertions that can bemade with a sentence S expressing proposition ' is de�ned as follows:PPA(S) = f'g [ fha; 'i j a is the interpretation of an NP in Sg(3)A pragmatic assertion ha; 'i is assumed to be about a. The possibility foran NP interpretation a to serve as the topic of a pragmatic assertion ha; 'i3. The ground is also known as `background', as `presupposition' and as `open propo-sition'. In phonology, the term `focus' is often used for intonational prominence. Thatis, any constituent which bears pitch accent is said to be a focus. Although in general,(part of) the informational focus is marked by prosodic prominence, not every accentedconstituent is a focus in the informational sense. In particular, accented constituents mayalso be topics/links. 4

is subject to further syntactic and semantic restrictions, cf. footnote 8 belowNotice, by way of example (adopted from Dahl 1974), that the sen-tence The teacher loves ice cream gives rise to the parallel topic/commentand ground/focus partitions indicated in (4) if it answers the questionsWhatabout the teacher? What does he feel? , whereas it induces the partitionsspeci�ed by (5) in the interrogative context What about the teacher? Whatdoes he love?topic commentThe teacher loves ice creamground focus(4) topic commentThe teacher loves ice creamground focus(5)The fact that the two informational articulations correspond to di�erentpartitions in (5) shows that neither of them is by itself capable of capturingall the informational distinctions present in the sentence. Therefore, the twotraditional binomial articulations of focus/ground and topic/comment arecon ated into a single trinomial and hierarchical one in Vallduv��'s accountof information packaging (1992, 1993, 1994). The core distinction is theone between new information and anchoring, between focus and ground.In addition, the ground is further divided into the `link', which correspondsapproximately to the topic in the traditional topic/comment approach,4 andthe `tail'.5 In a picture:� topic commentlink tail focusground focus `aboutness'`old'/`new'(6)4. To the extent that links correspond to the topic in the traditional topic/commentdistinction, Vallduv��'s theory is quite similar to the analysis of sentence topics presentedin Reinhart (1982), where a pragmatic assertion of ' about a is formalized as ha;'i, inthat a functions as a kind of `locus of update' for ' (see below). The two approaches di�erin that Reinhart allows assertions without a `locus of update' (since also ' 2 PPA( S))and topics that express new information.5. The hierarchy does not imply constituency or (even) continuity. In particular, the twoparts (link and tail) of the ground may not constitute a linear unit at the surface. Moreover,sentences may have more than one link, and more than one element may constitute thetail. 5

Given this articulation, the answer The teacher loves ice cream to thequestions What about the teacher? What does he love? will receive thefollowing analysis:The teacher loves ice creamlink tail focusground focus(7)Roughly speaking, the di�erent parts|focus and ground, link and tail|ofa sentence S have the following informational functions in Vallduv��'s theory.The focus encodes IS, the information of S, which can be metaphor-ically described as �S, the proposition expressed by S, minus Kh, the infor-mation (the speaker presumes) already present in the hearer's informationstate. The ground performs an ushering role|it speci�es the way in whichIS �ts in the hearer's information state: links indicate where IS should goby denoting a location in the hearer's information state, and tails indicatehow IS �ts there by signaling a certain mode of information update.Of course, talking about ushering information to some location in thehearer's information state presupposes that this information state has somesort of internal structure. In this respect, Vallduv�� purports toagree with Heim that there has to be some additional internalstructure in the hearer's model of the common ground thatplays an important role in natural language interpretation,even if this internal structure is of tangential relevance in truthvalue computation. It is this internal structure of informationstates which is, in fact, crucially exploited by the di�erentinformation-packaging strategies used by speakers in pursuingcommunicative e�ciency. (Vallduv�� 1994: 7)2 Files in FocusIn fact, Vallduv�� takes the metaphor of Heim's �le change semantics (1982,1983) literally, in that he assumes that the information in the hearer's modelis organized in �les, i.e., collections of �le cards. Each �le card representsa discourse entity: its attributes and its links with other discourse entitiesare recorded on the card in the form of conditions. Such a discourse entitymay be known to the hearer but not salient at the time of utterance, it may 6

be salient at the time of utterance, it may be completely new to the hearer,it may be inferable from what the hearer knows, etc. Discourse entitiesmediate between referring expressions (noun phrases) and entities in thereal world: inde�nite noun phrases prompt hearers to create a new �le card,and de�nite noun phrases incite them to retrieve an already existing �le card.Both de�nites and pronouns denote already existing �le cards, but pronounsdenote salient �le cards, whereas (other) de�nites refer to non-salient ones.File change comprises the above aspects of �le card management,but it also involves content update, i.e., the incorporation of informationconveyed by a given sentence into records on novel and familiar �le cards,and this is where Vallduv�� lets information packaging come in.Links are associated with so-called goto instructions. In �le changesemantics, the target location of such a declaration is a �le card fc. Atail points at an information record|normally a (possibly underspeci�ed)condition|on such a �le card, record(fc), and indicates that it has to bemodi�ed (or further speci�ed) by the focus information IS of the sentence.The associated instruction type is called update-replace. In the absenceof a tail, the focus information IS of a sentence is simply added at the currentlocation. The associated instruction type is called update-add.Sentences may lack links and tails (recall that the focus is the onlynon-optional part of a sentence), so the following four sentence types can bedistinguished:a: link-focusb: focusc: focus-taild: link-focus-tail(8)The respective sentence types in (8) are associated with the (compound)instruction types in (9):a: goto(fc)(update-add(IS))b: update-add(IS)c: update-replace(IS,record(fc))d: goto(fc)(update-replace(IS,record(fc)))(9)The sentence and instruction types in (8) and (9) can be illustrated withthe following examples, where links, tails and foci are speci�ed by means of[ L: : :], [T: : :] and [F: : :] brackets, respectively, and accented expressions in 7

foci and links are|as above|written in small caps (representing H* pitchaccent) and boldface (for L+H* pitch accent), respectively:a: link-focus: [LThe boss][Fhates broccoli]goto(fc)(update-add(IS))b: focus: [FHe always eats beans]update-add(IS)c focus-tail: [FHe is not][Tdead]update-replace(IS,record(fc))d: link-focus-tail: [LThe boss][Fhates][Tbroccoli]goto(fc)(update-replace(IS,record(fc)))(10)As regards the �rst example, suppose that a newly appointed temp is or-dering dinner for the boss and asks the executive secretary whether there isanything that he should know about the boss' taste. The executive secretarygives the following answer:[LThe boss][Fhates broccoli](11)Example (11) is a link-focus construction, and as such it is associated with agoto(fc)(update-add(IS)) instruction. The link subject the boss speci�esa locus of update fc, viz., the card representing the boss|card #25, say.The focus verb phrase hates broccoli speci�es the information IS that has tobe added to this card. Suppose that broccoli is represented by card #136.Then, passing over some formal details, the update-add(IS) instructionassociated with the focus hates broccoli amounts to adding the condition`hates(25,136)' to the locus of update, i.e., the boss' card #25. Moreover,the record ` ; 25 ', a pointer to the locus of update, is added to card #136,rendering the condition `hates(25,136)' on card #25 `accessible' from card#136. Vallduv�� says that this linking mechanism, which designates a uniquelocation for content update, is `much more e�cient' than straightforwardmultiple recording of information on cards.25boss(25) 136broccoli(136) =) 25boss(25)hate(25,136) 136broccoli(136); 25[FHe always eats beans](12) 8

Example (12), an all-focus construction, is associated with a simple update-add(IS) instruction. Here, this instruction involves the addition of the focusinformation IS that the value of the current card always eats beans. Thatis: if it is interpreted immediately after example (11) and if its adverbiallymodi�ed transitive verb phrase is left unanalyzed for simplicity, it amountsto adding the condition `always eats beans(25)' to card #25.The presence of a tail in a sentence signals a mode of update di�er-ent from the straightforward update-add(IS) instruction. A tail indicatesthat a (possibly underspeci�ed) record on a �le card has to be replaced (orspeci�ed further). The material in the tail serves the purpose of determiningwhich record. Suppose, for example, that (13) is a reaction to the statementSince John is dead, we can now split his inheritance:I hate to spoil the fun, but [Fhe is not][Tdead](13)With this focus-tail example, the speaker instructs the hearer to replace therecord on the current locus of update|card #17, say, for John|expressingthat the value of card #17 is dead by one saying that he is not dead. Inshort, the tail serves to highlight a condition on �le card #17, the one sayingits value is dead. This condition is then modi�ed in the way speci�ed by thematerial in the focus.In addition to the option of replacing a record on a �le card, thereis the possibility of further specifying an underspeci�ed record, somethingwhich is assumed to be going on in the link-focus-tail example (14) givenbelow. Suppose now that the newly appointed temp asks the executive sec-retary whether it was a good idea to order broccoli for the boss, and thatthe executive secretary gives the following answer:[L The boss][Fhates][Tbroccoli](14)The idea is that the temp has an underspeci�ed record on his card for theboss, which says that the boss has some attitude towards broccoli. The lackof information about the nature of this attitude is re ected by the record`att', and it is this record which is replaced by `hate' after hearing theexecutive secretary's answer (14):9

25boss(25)att(25,136) 136broccoli(136) ) 25boss(25)hate(25,136) 136broccoli(136); 25Di�erent languages choose di�erent structural means to spell out the sameinformational interpretations. Vallduv�� studies the manifestation of informa-tion packaging in several languages, with an emphasis on Catalan and En-glish. Cross-language comparison shows that in expressing information pack-aging, languages exploit word order and prosody in various ways. Roughlyspeaking, English structurally realizes information packaging by means ofalternative intonational contours of identical strings, whereas Catalan has aconstant prosodic structure and e�ectuates information packaging by meansof string order permutations. In fact, Vallduv�� argues that languages suchas Catalan supply empirical support for the representation of informationpackaging sketched above, since these languages package their informationin a much more salient way than, for example, English. Thus, while infor-mational interpretations may be expressed exclusively by prosodic means inEnglish, information packaging instructions in Catalan are straightforwardlyre ected in syntax.In English, the focus is associated with a H* pitch accent (writtenin small caps), links are marked by a L+H* pitch accent (written in bold-face), and tails are structurally characterized by being deaccented. One andthe same string may be assigned di�erent intonational phrasings in order torealize di�erent informational interpretations. In particular, the focal pitchaccent may be realized on di�erent positions in the sentence. This is illus-trated by the sentences (16), (18) and (20), construed as answers to thequestions (15), (17) and (19), respectively:What did you �nd out about the company?(15) [FThe boss hates broccoli](16) What did you �nd out about the boss?(17) [LThe boss][Fhates broccoli](18) What does the boss feel about broccoli?(19) 10

[LThe boss][Fhates][Tbroccoli](20)In Catalan, the situation is as follows. Metaphorically speaking, one can saythat Catalan focal elements remain within a so-called `core clause', but thatground elements are `detached' to a clause-peripheral position. In particular,links are detached to the left, and non-link ground elements undergo right-detachment. As a result of detaching both links and tails, the core clause(CC) is left containing only the focus of the sentence:links [CC focus ] tails(21)Consider the Catalan counterparts (22), (23) and (24) of (16), (18) and(20), respectively. The all-focus sentence (22) displays the basic verb-object-subject word order. In (23) and (24), the link subject l'amo has been de-tached to the left. In (24), moreover, the tail direct object el br�oquil hasbeen detached to the right, leaving a clitic (l') in the focal core clause. Notethat intonational structure plays a part in Catalan too, albeit `a rather lameone' (Vallduv�� 1993: 33): a focal H* pitch acent is invariably associated withthe last item of the core clause.[FOdia el br�oquil l'amo](22) [LL'amo][Fodia el br�oquil](23) [LL'amo] [Fl'odia][Tel br�oquil](24)The above observations provide con�rmation that information packaginginvolves syntax as well as prosody; hence any attempt to reduce informa-tion packaging to either syntax (for Turkish, cf. Ho�man 1995) or prosody(for English, cf. Steedman 1991, 1992, 1993) is inadequate from a cross-linguistic point of view.6 Accordingly, Hendriks (1996) treats the range ofvariation in the structural realization of information packaging as displayedby Catalan and English by means of the sign-based categorial grammar for-malism of Hendriks (1994), which takes its inspiration from Oehrle's (1988,6. Note, moreover, that the structural realization of information packaging in Catalan in-volves both syntax and prosody. E.g., the informationally non-equivalent sentences [FOdiael br�oquil l'amo] and [FOdia el br�oquil] [Fl'amo] di�er only prosodically. The sameholds for English, where the structural realization of information packaging also may in-volve syntax: Reinhart (1982: 63) notes that a fronted NP such as Felix in the sentenceFelix, it's been ages since I've seen him must be a topic (i.e., link). 11

1993) work on generalized compositionality for multidimensional linguisticobjects and shares characteristics with HPSG (Head-Driven Phrase Struc-ture Grammar|see Pollard and Sag 1987, 1994). Basically, this formalismis a both intonationally/syntactically and semantically/informationally in-terpreted version of a double `dependency' variant (see Moortgat and Mor-rill 1991) of the non-associative Lambek (1961) calculus, enriched with theunary operators of Moortgat (1994). The signs, the grammatical resources ofthis formalism, are Saussurian form-meaning units which re ect the fact thatthe dimensions of linguistic form and meaning contribute to well-formednessin an essentially parallel way:intonational term / type . informational term(25)The treatment of information packaging in this formalism di�ers from manyof its predecessors (including other extensions of standard Lambek calculussuch as Oehrle 1991, Van der Linden 1991, and Moortgat 1993), in that itdoes not employ focusing operators, but, instead, makes use of `defocusing'operators that license the presence of links and tails. Acccording to mostapproaches, focused constituents are semantic functors which take the non-focused part of the sentence as their argument. This analysis is based onsuch assumptions as made in Szabolcsi (1981, 1983) and Svoboda and Ma-terna (1987), where focus is not only considered an information-packagingprimitive but also an implicit truth-conditional exhaustiveness operator, andon semantic studies of the phenomenon of `association with focus' as pro-vided by Jacobs (1983), Rooth (1985), Krifka (1991), and others who haveargued that the quanti�cational structure of so-called focus-sensitive opera-tors is crucially determined by the traditional pragmatic focus-ground par-tition. However, Vallduv�� argues convincingly that `the claim that focusedconstituents truth-conditionally entail exhaustiveness leads to extreme po-sitions' (1992: 170), and Vallduv�� and Zacharski (1993) show that `associa-tion with pragmatic focus' is not an inherent semantic property of `focus-sensitive' operators, which may express their semantics on partitions otherthan the focus-ground one|witness obvious cases of association with sub-segments of the informational focus, with links, and with other parts ofthe ground. This dissociation of the pragmatic focus-background distinctionfrom issues of exhaustiveness and focus-sensitivity dispels the need of ana-lyzing focused constituents as operators which semantically take scope overthe non-focused parts of the sentence, which can be considered an advan-tage. As sentences may lack links and tails, such analyses do not immediately 12

re ect the core status of the focus, which is the only non-optional part of asentence. In some sense, then, all-focus sentences constitute the basic case,and the cases where there is a ground are derived from such basic all-focusstructures.3 Cards and BoxesVallduv�� has it that[. . . ] a proper understanding of information packaging, i.e., ofthe actual strategies used by human agents in e�ecting infor-mation update by linguistic means, will help us gain furtherinsight into the structural properties of the cognitive statesthese dynamic strategies manipulate. (Vallduv�� 1994: 24)As we have seen, the basic idea of information packaging is that in discourse,speakers not only present information to their interlocutors, but also pro-vide them with detailed `instructions' on how to manipulate and integratethis information. With respect to the role of these instructions in the deter-mination of those aspects of the structure of information states which arerelevent to natural language interpretation, Vallduv�� claims the following:The use of these instructions reveals that speakers treat infor-mation states as highly structured objects and exploit theirstructure to make information update more e�cient for theirhearers. (Vallduv�� 1994: 3)More speci�cally, concerning `the internal structure of information stateswhich is, in fact, crucially exploited by the di�erent information-packagingstrategies used by speakers in pursuing communicative e�ciency' (1994: 7),it is argued that information packaging instructions contribute in two waysto the optimization of information update, since they provide means to� designate a �le card as the locus of information update and hencecircumvent the redundancy of multiple update; and� identify the information of the sentence and its relation to informationalready present in the hearer's model. 13

(Recall that the information of the sentence, IS, is expressed by the focus,and that the ground has an ushering role with respect to IS: links indicatewhere IS goes, and tails indicate how it �ts there.) So, summing up, Vallduv��concludes that information states constitute systems that have at least theinternal structure of a collection of �le cards connected by pointers.Though the presented arguments may appear to be intuitively quiteappealing, it can be argued that, strictly speaking, they are not as com-pelling as they seem. Somehow, Vallduv�� is begging the question: `talkingabout ushering IS to a location in the hearer's model Kh [. . . ] does not makemuch sense unless one assumes some sort of rich internal structure for Kh'(Vallduv�� 1994: 7). The relevant question, however, is whether this assump-tion of `some sort of rich internal structure' itself makes sense of anythingbesides the ushering function of links.If �le card systems are assumed, then the information-packaging in-struction types apparently do contribute to e�cient information exchange.And if this assumption is warranted, it may even serve as an explanation ofthe fact that we do appear to �nd these ways of packaging information in avariety of languages. Nonetheless, the more theoretical question is whetherthis assumption itself is warranted, and whether the organization of linguis-tic information exchange really presupposes such information states. Afterall, ushers can be very useful, but there are also halls that have unnum-bered seats! Maybe links really make no sense without �les, but, for thatmatter, maybe we simply fail to understand what links do. The notion of`ushering IS to a location' may be just as metaphorical as the notion of `�lecard collection'. For instance, �les are, as Vallduv�� puts it, `dimensionallyricher' than the card-less discourse representation structures of DiscourseRepresentation Theory (see Kamp 1981, Kamp and Reyle 1993), since each�le card introduces its own `representational space' where all its records areto be found while there is no sensible notion of location in discourse rep-resentation structures. Still, a hearer who employs discourse representationstructures has an easier job from a bookkeeping perspective than a hearerwhose information states are collections of �le cards connected by pointers.This can be illustrated as follows. Imagine an utterance made byIrene, a speaker who organizes her utterances on the basis of the assump-tion that her audience stores information using collections of �le cards con-nected by pointers, to Hans, a hearer who employs discourse representationstructures. Clearly, it would be inappropiate to say that Irene uses links tousher IS to a location in the hearer's model KS, since there is no sensiblenotion of location in Hans' discourse representation structures. Still, this 14

does not at all preclude Hans from updating his discourse representationwith the proposition that Irene attempts to get through. And worse, he haseven got considerably less to do than a hearer who uses collections of �lecards connected by pointers. Compare the following link-focus example:[LFrank5][F ew from Amsterdam9 to Oslo8 via Stuttgart2](26)Neglecting various details, if a �le clerk is to update her �le in order torepresent the information expressed by example (26) in the way sketchedabove, she has to carry out the following sequence of instructions:7goto(5)(update-add( ew(5,9,8,2)))goto(9)(update-add( ; 5 ))goto(8)(update-add( ; 5 ))goto(2)(update-add( ; 5 ))goto(5)(27)Hans, on the other hand, only has to carry out the following instruction:update-add( ew(5,9,8,2))(28)This example may serve as an indication that none of the data discussedabove precludes the use of, say, Kampian discourse representation structuresinstead of Heimian �les. Clearly, there may be evidence for assuming thereto be �les at work, and one of the last things this paper would like to claimis that people organize their information in simpler systems than collectionsof �le cards (or discourse representation structures, for that matter). Onthe contrary. The only point is that the use of �les does not appear to beimperative so far.7. Assuming that establishing links to the locus of update is done via packaginginstructions|of course, these links have to be established somehow. Note, by the way,that the �le clerk's task would not be made easier by structure sharing (something sug-gested by Enric Vallduv�� (personal communication)), because also the structure sharingwill itself have to be established somehow|in the following way, for example:goto(5)(update-add( 1 ew(5,9,8,2)))goto(9)(update-add( 1 ))goto(8)(update-add( 1 ))goto(2)(update-add( 1 ))goto(5) 15

Notice that Vallduv��'s conclusion is, in some sense, unfalsi�able. Discourserepresentation structures can model precisely the same information as �lecard systems, except for one small di�erence. The only thing that discourserepresentation structures lack is a marked discourse referent correspondingto the �le notion of `current locus of update', i.e., the location where the�le clerk happens to �nd herself. If we assume that discourse representationstructures have a way of marking such a discourse referent j|by a condition`clerk at(j)', say|, then the two systems di�er only in the way in whichthey display their information: in one big box, or on several cards connectedby pointers. But, moreover, one can show that given Vallduv��'s speci�c useof pointers to �le cards, there is actually a bijective correspondence betweenhis �les and the class of discourse representation structures with atomicconditions and one marked discourse referent for the current locus of up-date. For note that conditions `rel(i1; : : : ; in)' are invariably added on cardi1, inducing pointers ` ; i1 ' on the cards i2; : : : ; in. Hence the followingcorrespondence can be established:i1. . .rel(i1; : : : ; in). . . i2. . .; i1. . . : : : in. . .; i1. . . () : : : ; i1; : : : ; in; : : :. . .rel(i1; : : : ; in). . .j. . .eb6�� DD b. . . () : : : ; j; : : :. . .clerk at(j). . .The idea that links specify a locus of update in information states thatare collections of �le cards connected by pointers is problematic for variousreasons. First, it is unclear what locus of update must be associated withquanti�ed, negative and disjunctive links, or|more in general|where andhow quanti�ed, negative and disjunctive information has to be put. Second,the existence of sentences with more than one link is enigmatic. Third, thereplacement operation triggered by the presence of tails is complicated bythe use of �le cards. And fourth, the approach leads to the counterintu-itive conclusion that pronouns form part of the focus. These issues will beaddressed in the remainder of this section. 16

(a) Vallduv�� observes that �les are `dimensionally richer' than the dis-course representation structures (drss) of Discourse Representation Theory.Now, this is true to the extent that each �le card introduces its own `rep-resentational space' where all records concerning that �le card are to befound. In order to be actually richer, nonetheless, �les must be adapted tomodel more than merely atomic conditions|i.e., individuals having proper-ties and standing in relations at various spatio-temporal locations. Amongother things, they should be able to model quanti�ed, negative and disjunc-tive information. Discourse Representation Theory allows the constructionof complex conditions from sub-drss, and these conditions|by an appro-priate semantic interpretation procedure|model precisely such information.Heim, who explicitly speaks of �les and �le cards as metaphors (1982: 276and 302�.), spells out quanti�ed, negative and disjunctive information inpurely semantic terms, i.e., in terms of the domains and satisfaction sets of�les. However, it is not clear how such information must be expressed in thenon-metaphorical �le card set-up of Vallduv�� (1994).For one thing, what loci of update are speci�ed by the links of sen-tences such as (29), (30) and (31)?8 On what �le card(s)|if any|shouldthe information expressed by these sentences be put?[LEvery man][Fwalks](29) [LNo man][Fwalks](30) [LJohn or Mary][Fwalks](31)For another, how should this information be put? One might think of us-ing sub-�les, but then, where must these be put? Are they attached to amain �le, or must they be attached to a main �le's �le card? Which one?Interestingly, Heim raises similar questions in her 1983 paper:8. Though `links tend to be de�nite NPs' (1992: 77), Vallduv�� notes the `restricted exis-tence of inde�nite links' (1992: 46). `Sentences with quanti�er links are' claimed to be `lessnatural than others, causing raised eyebrows among some Catalan speakers. Sentences likeA tots els estudiantsi elsi donen un carnet ti \To all students they give an ID" or Atothomi no eli tracten ti igual \Everybody they don't treat the same" are extremelynatural, some other sentences sound odder. Most sentences, however, are felicitous oncethe right context is construed, although in some cases it may require some sophistica-tion' (Vallduv�� 1992: 153). Analogously, Reinhart notes that if they `can be interpreted(pragmatically) as denoting sets, universally quanti�ed NPs, as well as speci�c and genericinde�nite NPs, can serve as topics' (1982: 65{66). 17

Take a simple sentence [. . . ]: It is raining . In the context of the�le metaphor, one doesn't quite know how to deal with thissentence. As an informative sentence, it ought to call for anupdating of the �le somehow: but what exactly is the �le clerksupposed to do? The information that it is raining does notbelong on any particular �le card, it seems, since each �le cardis a description of an individual, but It is raining is not aboutany individual. Should the �le clerk perhaps write on somearbitrary card: `is such that it is raining'? Or should he writethat on all cards? And what if the �le so far doesn't containany cards yet? [. . . ] Quanti�ed and negated propositions aresimilarly puzzling if we are so ambitious as to want to saywhat exactly the �le clerk does in response to them. Under themodest aspect of domain and satisfaction set change, however,they pose no problem. (Heim 1983: 183{184)It should be noted here that such a `modest' position cannot be retained inthe set-up of Vallduv�� (1994), because there the entities to be updated mustbe �les, and not their domains and their satisfaction sets.(b) Vallduv�� (1992: 104) notes that there is no structural restrictionon the number of links in Catalan. `Sentences may have more than one link,as in the Catalan example (32).[LEl br�oquil] [La l'amo] [Fl'hi van regalar]the broccoli to the boss obj-iobj 3p-past giveApprox.: `The broccoli the boss (they) gave it to him (for free)'.(32)In these cases the speaker directs the hearer to go to two addresses and enterthe information under both.' (Vallduv�� 1992: 60, example number adapted).So, assuming that `they' have card #3 and that the boss and broccoli stillpossess their respective cards #25 and #136, this means that the sentenceis not associated with the instruction (33),9 but with an instruction alongthe lines of (34).*goto(136)(goto(25)(update-add(give(3,136,25)))(33) goto(136)(update-add(give(3,136,25))goto(25)(update-add(give(3,136,25))(34)9. Note, by the way, that the `goto(136)' constitutes a super uous detour in instruction(33). 18

But this raises questions. What is the current locus of update after (34)has been carried out? Is the �le clerk suddenly simultaneously working ontwo di�erent �le cards? If she isn't (suppose she is only working on card#25), does this then mean that (34) is equivalant to (35), the instructionassociated with the one-link sentence (36)?goto(25)(update-add(give(3,136,25))goto(136)(update-add( ; 25 ))goto(25)(35) [LA l'amo] [Fhi van regalar el br�oquil]](36)But if (34) and (35) are equivalent, then why does Catalan allow multiplelinks at all? And how could (34) and (35) be non-equivalent|what sensecould multiple loci of update make that pointers cannot?(c) Above an informal sketch was given of Vallduv��'s analysis of tail-containing sentences in terms of update-replace instructions. It can beexpected that various complications will arise when it comes to giving anexplicit formalization of the replacement instructions associated with tails.Any attempt at giving an appropriate and fully general de�nition of theseinstructions will have to confront a number of questions. Thus, how exactlydo you know which record has to be replaced or speci�ed further? Is thereguaranteed to be such a record? Is there a unique one, and what happensif there are more? Is it always one record that has to be replaced, or do wesometimes need to replace a group of records? What kind of match mustthere be between the material in a tail, and the material in the target record?Of course, these are tough nuts that have to be cracked when it comes tocoming to theories of belief revision.Here we will just present a simple example which illustrates thatthe replacement operation triggered by the presence of tails is speci�callycomplicated by the idea that information is organized in �le card systems.Suppose that Louis van Gaal utters (37), whereupon Johan Cruij� reactswith saying (38):[LAjax][Fwon](37) [FNo, Barcelona][Twon](38)Assume the �le cards #1 and #2 for Ajax and Barcelona, respectively. Now,clearly, Johan Crui� here instructs Louis van Gaal to replace his record 19

according to which Ajax won by one according to which Barcelona did.Presumably, this should not (only) be done on the card for Ajax. Insteadof the straightforwardly simple (39), the complex instruction given in (40)seems to be needed.*update-replace(won(2),won(1))(39) update-replace( ,won(1))goto(2)(update-add(won(2)))(40) (d) An example typical of the way in which Vallduv�� analyzes pro-nouns can be obtained by combining the above example sentences (11) and(12) into one text:[LThe boss][Fhates broccoli][FHe always eats beans](41)The �rst sentence is a link-focus construction, and therefore associated withan instruction to go to the �le card of the boss, thereby turning it into thecurrent locus of update, and to enrich that �le card with the informationthat the boss hates broccoli (and the broccoli �le card with a pointer tothe �le card of the boss). The second sentence is an all-focus construction,associated with the simple instruction to add the focus information that thevalue of the current locus of update always eats beans to the current locusof update. Hence if it is interpreted immediately after the �rst sentence, itamounts to adding the information that the boss always eats beans to thecard of the boss.Note that the pronoun he obviously does not induce replacementor shift the locus of update. Hence it cannot be a link or a tail, and thisinevitably leads to the conclusion that it forms part of the focus. This is acounterintuitive result, however, since it is also clear that the interpretationof the pronoun is provided by the value of the current locus of update|which does not constitute new information, but can be assumed to be alreadypresent in the hearer's information state.4 Non-Monotonic AnaphoraLet us wind up the discussion so far. It has been argued that the datadiscussed above do not enforce the conclusion that information states have 20

at least the structure of a collection of �le cards connected by pointers. Forthat matter, the phenomena can also be accounted for in terms of discourserepresentation structures, and it is very well possible that circumventing �lecards might lead to the avoidance of the complications that were outlinedin the previous section.In view of these considerations, a card-less alternative will be de-fended in the present section, according to which information states aremodeled by means of discourse representation structures, which are onto-logically less committed than the `dimensionally richer' �le card system, inthat discourse representation structures do not come with locations.But if, as was argued above, the use of �les does not appear to beimperative, then a question must be faced: what purpose do links serveif they do not serve to specify a locus of update by ushering to locations?What does `ushering to a location' mean if representations do not come withlocations? Thus a di�erent perspective on the function of links is required.Below a tentative answer will be suggested that carries less presuppositionsthan the �le metaphor.This alternative perspective has its heuristic starting point in Kampand Reyle (1993), who note that processing a plural pronoun does not alwaysinvolve equating the discourse referent it introduces with one introducedearlier through the processing of some other plural NP. Kamp and Reyleconsider the following example:John took Mary to Acapulco. They had a lousy time.(42)Here, the plural pronoun they does not have a single NP for its antecedent.Rather, the `antecedent' has to be `constructed' out of various parts of thepreceding text. Such examples, which are very common, seem to suggest thatplural pronouns can pick up any antecedent that can be obtained from an-tecedent information by logical deduction. However, the deductive principlesthat are permitted in this context turn out to be subject to restrictions.Eight of the ten balls are in the bag. They are under the sofa.(43)The pronoun they in (43) cannot be understood as referring to the two ballsthat are missing from the bag. Apparently, subtracting one set from anotheris not a permissible operation for the formation of pronominal antecedents.The permissible process of antecedent formation displayed by (42)is called Summation: a new discourse referent is introduced which repre-sents the `union' of individuals (John and Mary) and/or sets represented 21

by discourse referents that are already part of the discourse representa-tion structure. Other permissible processes are Abstraction, exempli�ed by(44), which allows the introduction of discourse referents for quanti�ed NPs(compare also footnote 8 above), and Kind Introduction, which introducesdiscourse referents for a certain `genus' explicitly mentioned in the text by a(simple or complex) noun. If they in (45) refers to the (few) men who joinedthe (conservative) party, we are dealing with Abstraction. The more naturalreading, where they refers to men in general (and the party is presumablynon-conservative), is a case of Kind Introduction.I found every book Bill needs. They are on his desk.(44) Few men joined the party. They are very conservative.(45)In their discussion of the inferential processes available for the constructionof antecedents for (plural) pronouns, Kamp and Reyle suggest the followinggeneralization:What sets the admissible inference processes of Summation,Abstraction and Kind Introduction apart from an inadmissibleinference pattern such as set subtraction is that the former are[. . . ] strictly positive (Kamp and Reyle 1993: 344),or `cumulative' in the following sense: the newly created discoursereferent represents an entity of which the discourse referentsused in the application of the rule represent (atomic or non-atomic) parts (Kamp and Reyle 1993: 394).Notice that, when this generalization is taken in conjunction with a principlethat anaphora invariably involves the addition of an equational condition `X= Y' for an anaphoric expression with discourse referent Y and a|possiblyinferentially created|antecedent discourse referent X (and such an equa-tional approach is standard practice in Dicourse Representation Theory),the necessary result will be that anaphora is always (upward) monotonic:if an expression with discourse referent Y is anaphorically dependent on anexpression with discourse referent X, then X � Y.1010. Let sets A and B be partially ordered by �A and �B, repectively. In mathematics, afunction f : A! B is called monotonic i� a �A b entails that f(a) �B f(b) for all a and bin A. Note that the same notion is involved here for (i) the function f : NP! DR which 22

The latter result, however, does not seem to be borne out by the facts.For example, Van Deemter (1992, 1994a) presents cases of `non-identityanaphora' along the lines of (46), as well as minimal pairs such as (47) and(48): Our neigbours are extremely nice people.He is a teacher, she is a housewife.(46) John fed the animals. The cats were hungry.(47) John fed the animals. The cats were hungry.(48)It can be observed that the pronouns he an she are anaphorically depen-dent on our neighbours in (46), but that the discourse referents of the pro-nouns represent entities which are proper subsets of the entity representedby the discourse referent of the antecedent: obvious cases of non-monotonicanaphora.Moreover, whereas the reading of (47) where the cats is anaphoricto the animals strongly and monotonically suggests that all animals fedby John were cats, the reading of (48) where the cats is anaphoric to theanimals does not. It even seems to imply that John fed at least one non-cat.11 Again, we are dealing with non-monotonic anaphora.Note that the texts (47) and (48) di�er only in the assignment ofL+H* accent to the noun phrase the cats, which is the distinguishing markof links in English. Hence our alternative hypothesis concerning links:Non-Monotonic Anaphora Hypothesis (NAH):Linkhood (marked by L+H* accent in English) serves to signalnon-monotonic anaphora. If an expression is a link, then itsdiscourse referent Y is anaphoric to an antecedent discoursereferent X such that X 6� Y.(49)This hypothesis a�ects a range of phenomena. In Hendriks (draft) it is ar-gued that a discourse-representational approach of information packagingassociates every occurrence of a noun phrase a with its discourse marker f(a); (ii) �NPsuch that a �NP b i� a is the antecedent of anaphor b; and (iii) �DR such that X �DR Yi� X � Y .11. `Strongly suggests' and `seems to imply' instead of `entails', since though the e�ectsare quite strong, they are of a pragmatic, rather than a logico-semantic, nature. See also(c), on pronoun referent resolution, below. 23

such as the one sketched here provides a sound background for an ade-quate theory of the phenomena commonly referred to as `association withfocus' (Jacobs 1983; Rooth 1985, 1992; Krifka 1991), thereby contributingto the integration of pragmatic theories of information packaging with con-temporary research in formal semantics. More in particular, the approach iscapable of handling the non-marginal cases of association of `focus-sensitive'operators such as only with non-focal parts of the sentence that have been at-tested in the literature (Vallduv�� 1992; Partee 1994; Vallduv�� and Zacharski1993). That is, not only the phenomenon of `second occurrence focus' isaccounted for, but also the link-sensitive behaviour of `focus-sensitive' op-erators is given an analysis of which the approach to links given here is anessential ingredient. Below it will be shown that, in addition to this, theNAH subsumes not only the so-called `non-identity' anaphora just exem-pli�ed and analyzed in Van Deemter (1992, 1994a), but also the cases ofcontrastive stress discussed in Rooth (1992) and Vallduv�� (1992, 1994), andthe corrections mentioned by Pierrehumbert and Hirschberg (1990). It con-tributes to an explanation of the e�ect of pitch accenting on pronoun referentresolution noted in Cahn (1995), Kameyama (1994), Vallduv�� (1994), amongmany others, and it sheds light on the distinction between restrictive andnon-restrictive relative clauses and adjectives (see Kamp and Reyle 1993).(a) The relationship between non-identity anaphora and linkhood canbe demonstrated even more saliently with relational nouns:Ten guys were playing basketball in the rain.The fathers were having fun.(50) Ten guys were playing basketball in the rain.The fathers were having fun.(51)Thus, whereas (50) has an `identity' reading where the fathers is anaphoricto ten guys which|monotonically|suggests that all ten guys playing bas-ketball in the rain were fathers who were having fun, the reading of (51)where the fathers is anaphoric to ten guys does not. This reading seemsto|non-monotonically|imply that the fathers who were having fun con-stitute a proper subset of the ten basketball-playing guys. Since father isa relational noun, there is, next to this `subsectional' reading, also a|non-monotonic|`relational' reading of (51) on which the fathers of the ten guysplaying basketball in the rain were having fun.Observe, by the way, that Kamp and Reyle's example (42) of Summa-tion, a case of monotonic non-identity anaphora in which the pronoun they 24

typically appears unaccented, shows that is not so much the `non-identity'as the `non-monotonicity' of the anaphora which is responsible for the L+H*accent (that is: the linkhood) of the anaphor.(b) According to Rooth, contrast is the cornerstone of the interpre-tation of focus phenomena: `Intonational focus has a semantic import re-lated to the intuitive notion of contrast within a set of alternative elements'(1992: 113), and Vallduv�� gives the following example of `contrastive' links(1993:14):Where can I �nd the cutlery?The forks are in the cupboard, butThe knives I left in the drawer.(52)Note, however, that contrast is not really necessary for L+H* accent:12Where can I �nd the cutlery?The forks are in the cupboard.(53)Mere non-monotonicity su�ces.(c) As Pierrehumbert and Hirschberg (1990) observe, L+H* accentoften arises in corrections. Thus, sentence (55) is a natural way of correcting(54). John was stung by mosquitos.(54) He was stung by bees.(55)Interestingly, it can be observed that non-monoticity plays a role here too.Notice that a sentence such as (56) is less naturally uttered in the context of(54), even though, semantically (i.e., truth-conditionally) speaking, this sen-tence is an impeccable expression of the proposition entertained by a speakerwho believes that John was actually stung by bees as well as mosquitos.He was stung by insects.(56)12. Nor is contrariety (as proposed in Van Deemter 1994b), witness:Where can I �nd the cutlery?The forks are in the cupboard, and the knives too. 25

In fact, a speaker of (54) who is `corrected' by someone's utterance of (56)might very well react by uttering (57):Mosquitos are insects.(57)Apparently, it is the fact that mosquitos are insects that thwarts the assign-ment of L+H* accent to insects in sentence (56).(d) Many authors have paid attention to the e�ect of pitch accentingon pronoun referent resolution. The examples below stem from Lako� (1971).Paul called Jim a Republican. Then he insulted him.(58) Paul called Jim a Republican. Then he insulted him.(59)For grammatical reasons (syntactic parallellism), the preferred antecedentsfor the unstressed pronouns he and him in (58) are Paul and Jim, respec-tively. The preferences are reverse for the stressed pronouns he and him in(59).13 In the theory of Kameyama (1994), this phenomenon is accountedfor in the following way:� A grammar subsystem represents the space of possibilities and a prag-matics subsystem represents the space of preferences;� Stressed and unstressed pronouns have the same denotational range|the same range of possible values;� Complementary Preference Hypothesis (CPH): A stressed pronountakes the complementary preference of the unstressed counterpart.However, the NAH formulated in (49) is capable of predicting the CPHe�ects: adding L+H* accent to pronouns means the addition of a pragmaticsignal that the anaphora involved is non-monotonic. In the case of singularantecedents with entity-representing discourse referents,14 this means thatthe anaphor does not corefer with its antecedent. Correspondingly, we havethat pronominal stress turns the pragmatically determined preference for a13. The fact that (59) insinuates that calling someone a Republican is an insult is essen-tially due to the de-accenting of insulted in the second sentence of (59).14. Or, equivalently, singleton-set-representing discourse referents. 26

certain grammatically possible antecedent into a pragmatically determinedpreference for non-coreference with that antecedent.15(e) The sentences (60) and (61) (taken from Kamp and Reyle 1993:255) illustrate the familiar rule of English orthography that non-restrictiveclauses are set apart from the surrounding text by commas, but that restric-tive clauses are not.The son who attended a boarding school was insu�erable.(60) The son, who attended a boarding school, was insu�erable.(61)Note that (60), in which the relative clause is used restrictively, suggests thatthere is more than one son, but only one who is boarding. In (61), wherethe relative clause is used non-restrictively, the suggestion is rather thatthere is only one son, of whom it is said not only that he was insu�erablebut also, parenthetically as it were, that he attended a boarding school. Ifthe prosody of these sentences is taken into account, it will be clear thatthis pragmatic di�erence is in keeping with the NAH as formulated in (49).Similar observations can be made with respect to the (non-)restrictivenessof the adjectives and nouns in (64) (Kamp and Reyle 1993: 372).The son who attended a boarding school was insufferable.(62) The son, who attended a boarding school, was insufferable.(63) John fed the animals. The young cats were hungry.John fed the animals. The young cats were hungry.John fed the animals. The young cats were hungry.John fed the animals. The young cats were hungry.(64)15. Thus let fpg; fjg; fxg; fyg constitute the respective discourse referents of the nounphrases Paul, Jim, he and him. Then the grammar subsystem speci�es ffpg; fjgg asthe range of possible values for fxg and fyg, and the pragmatics subsystem (building onsyntactic parallellism) speci�es fpg and fjg as the respective preferred antecedents for thepronominal discourse referents fxg and fyg. Since the pronouns are stressed, however, therespective preferences of fxg and fyg for fpg and fjg mean that fpg 6� fxg and fjg 6� fyg,that is: p 6= x and j 6= y.

27

5 ConclusionIn Pierrehumbert and Hirschberg (1990), `The Meaning of Intonational Con-tours in the Interpretation of Discourse', it is proposed that speakers use tuneto specify a particular relationschip between the `propositional content' re-alized in the intonational phrase over which the tune is employed and themutual beliefs of participants|speaker S and hearer H|in the current dis-course, where tune, or intonational contour, is taken to be a sequence of low(L) and high (H) tones, made up from pitch accents, phrase accents andboundary tones, and tune meaning is assumed to be built up composition-ally. Pitch accents mark the lexical items with which they are associated asprominent: accented items are salient, not only phonologically but also froman informational standpoint. Pierrehumbert and Hirschberg follow Beckmanand Pierrehumbert (1986) in distinguishing six pitch accents: two simpletones, H* and L*, and four complex ones, L*+H, L+H*, H*+L and H+L*,where the `*' indicates that the tone is aligned with a stressed syllable. Pitchaccents are believed to convey information about the status of the individualdiscourse referents, modi�ers, predicates and relationships speci�ed by thelexical items with which the accents are associated.(a) `In general, we believe that all accent types can be usedto convey information to H about how the propositionalcontent of the (perhaps partially) instantiated expressioncorresponding to the utterance is to be used to modifywhat H believes to be mutually believed' (1990: 289).With respect to the two L+H pitch accents (L*+H and L+H*), it is observedthat: (b) `[The L+H pitch accents] `are employed by S to conveythe salience of some scale (de�ned here [. . . ] as a partialordering) linking the accented item to other items salientin H's mutual beliefs' (1990: 294), and `S employs theL+H* accent to convey that the accented item|and notsome alternative related item|should be mutually be-lieved' (1990: 296).Finally, the most common use of L+H* in the data collected by Pierrehum-bert and Hirschberg 28

(c) `is to mark a correction or contrast. In such cases S sub-stitutes a new scalar value for one previously proposedby S or by H|or for some alternative value available inthe context' (1990: 296).By way of conclusion we may observe that the account of links detailed inthe present paper can actually be considered a partial execution|viz., forL+H* pitch accent|of the intonational-informational program outlined inPierrehumbert and Hirschberg (1990), for note that on the present accountthe following analogous claims are assumed to hold:(a) Pitch accent is an aspect of information packaging|i.e,the structuring of propositional content in function of thespeaker's assumptions about the hearer's knowledge andattentional state;(b) L+H* pitch accent does invoke a scale, viz., the partialordering `�' on discourse referents, and conveys that theantecedent the accented item �nds itself `linked to' is as-sociated with a discourse referent that is not a subset ofthe discourse referent of the accented item; and(c) Correction and contrast are among the manifestations ofthe non-monotonic anaphora signaled by L+H* pitch ac-cent, which were argued to involve (non-)association withfocus, `non-identity' anaphora, contrastive stress, correc-tion, pronoun referent resolution, and restrictiveness ofrelatives and adjectives.References[1] Beckman, M., and J. Pierrehumbert (1986). `Intonational Structure inJapanese and English'. Phonological Yearbook 3, 15{70.[2] Bosch, P., and R. van der Sandt (eds.) (1994). Focus and NaturalLanguage Processing. Proceedings of a Conference in Celebration ofthe 10th Anniversary of the Journal of Semantics. Working Papers 6(Vol. 1: Intonation and Syntax), 7 (Vol. 2: Semantics), and 8 (Vol. 3:Discourse) of the IBM Institute for Logic and Linguistics, Heidelberg.[3] Cahn, J. (1995). `The E�ect of Pitch Accenting on Pronoun ReferentResolution'. Manuscript. MIT, Cambridge (Mass.). 29

[4] Chafe, W.L., (1976). `Givenness, Contrastiveness, De�niteness, Sub-jects, Topics and Point of View'. In C.N. Li (ed.) (1976), Subject andTopic, 25{55. Associated Press, New York.[5] Clark, H.H., and S.E. Haviland (1977). `Comprehension and the Given-New Contract'. In R.O. Freedle (ed.) (1977), Discourse Production andComprehension, 1{40. Lawrence Erlbaum Associates, Hillsdale (NewJersey).[6] Dahl, �O. (1974). (1974). `Topic-Comment Structure Revisited'. In �O.Dahl (ed.) (1974), Topic and Comment, Contextual Boundedness andFocus. Papers in Text Linguistics 6. Helmut Buske, Hamburg.[7] Deemter, K. van (1992). `Towards a Generalization of Anaphora'.Journal of Semantics 9, 27{51.[8] Deemter, K. van (1994a). `What's New? A Semantic Perspective onSentence Accent'. Journal of Semantics 11, 1{31.[9] Deemter, K. van (1994b). `Contrastive Stress, Contrariety and Focus'.In P. Bosch and R. van der Sandt (eds.), 39{49.[10] Engdahl, E. (ed.) (1994). Integrating Information Structure intoConstraint-based and Categorial Approaches. ESPRIT Basic ResearchProject 6852, Dynamic Interpretation of Natural Language. DYANA-2Deliverable R1.3.B. ILLC, University of Amsterdam.[11] Heim, I. (1982). The Semantics of De�nite and Inde�nite Noun Phrases.Ph.D. Dissertation University of Massachusetts, Amherst. Published in1989 by Garland. New York.[12] Heim, I. (1983). `File Change Semantics and the Familiarity Theoryof De�niteness'. In R. B�auerle, C. Schwarze and A. von Stechow (eds.)(1983), Meaning, Use and Interpretation of Language. De Gruyter,Berlin, 164{189.[13] Hendriks, H. (1994). `Information Packaging in a Categorial Perspec-tive'. In Engdahl (ed.), 89{116.[14] Hendriks, H. (1996). `Intonation, Derivation, Information'. In C. Casa-dio (ed.), Proceedings of the Third Roma Workshop on Proofs andLinguistic Categories. University of Bologna.[15] Hendriks, H. (draft). `Information Packaging and \Association withFocus" '. Manuscript, Utrecht University.[16] Ho�man, B. (1995). `Integrating \Free" Word Order Syntax and In-formation Structure'. Manuscript, University of Pennsylvania.[17] Jackendo�, R. (1972). Semantic Interpretation in Generative Grammar.MIT Press, Cambridge (Mass.). 30

[18] Jacobs, J. (1983). Fokus und Skalen: Zur Syntax und Semantik vonGradpartikeln im Deutschen. Niemeyer, T�ubingen.[19] Kameyama, M. (1994). `Stressed and Unstressed Pronouns: Comple-mentary Preferences'. In P. Bosch and R. van der Sandt (eds.), 475{484.[20] Kamp, H. (1981). `A Theory of Truth and Semantic Representation'.In J. Groenendijk, T. Janssen and M. Stokhof (eds.) (1981), FormalMethods in the Study of Language. Mathematical Centre, Amsterdam.Reprinted in J. Groenendijk, T. Janssen and M. Stokhof (eds.) (1984),Truth, Interpretation and Information. Selected Papers from the ThirdAmsterdam Colloquium. Foris, Dordrecht.[21] Kamp, H., and U. Reyle (1993). From Discourse to Logic. Introductionto Modeltheoretic Semantics of Natural Language, Formal Logic andDiscourse Representation Theory . Kluwer, Dordrecht.[22] Krifka, M. (1991). `A Compositional Semantics for Multiple FocusConstructions'. Linguistische Berichte, Suppl. 4, 17{53.[23] Lako�, G., (1971). `On Generative Semantics'. In D. Steinberg andL. Jacobovitz (eds.) (1971), Semantics. Cambridge University Press,Cambridge, 232{296.[24] Lambek, J. (1961). `On the Calculus of Syntactic Types'. In R. Jakob-son (ed.) (1961), Structure of Language and its Mathematical Aspects.Providence.[25] Linden, E.-J. van der (1991). `Accent Placement and Focus in Cat-egorial Logic'. In S. Bird (ed.) (1991) Declarative Perspectives onPhonology . Edinburgh Working Papers in Cognitive Science. Eccs,Edinburgh.[26] Moortgat, M. (1993). `Generalized Quanti�cation and DiscontinuousType Constructors'. In W. Sijtsma and A. van Horck (eds.) (1993),Proceedings of the Tilburg Symposium on Discontinuous Dependencies.De Gruyter, Berlin.[27] Moortgat, M. (1994). `Residuation in Mixed Lambek Systems'. In M.Moortgat (ed.) (1994), Esprit Basic Research Project 6852, DynamicInterpretation of Natural Language, dyana-2 Deliverable R1.1.B. Illc,University of Amsterdam, and to appear in IGPL Bulletin.[28] Moortgat, M., and G. Morrill (1991). `Heads and Phrases. Type Calcu-lus for Dependency and Constituent Structure'. OTS Research Paper,University of Utrecht.[29] Nooteboom, S.G., and J.M.B. Terken (1982). `What Makes SpeakersOmit Pitch Accents?'. Phonetica 39, 317{336. 31

[30] Oehrle, R. (1988). `Multidimensional Compositional Functions as aBasis for Grammatical Analysis'. In R. Oehrle, E. Bach and D. Wheeler(eds.) Categorial Grammars and Natural Language Structures. Reidel,Dordrecht.[31] Oehrle, R. (1991). `Prosodic Constraints on Dynamic GrammaticalAnalysis'. In S. Bird (ed.) Declarative Perspectives on Phonology . Ed-inburgh Working Papers in Cognitive Science. Eccs, Edinburgh.[32] Oehrle, R. (1993). `String-based Categorial Type Systems'. Manuscript.Department of Linguistics, University of Arizona, Tucson.[33] Partee, B. (1994). `Focus, Quanti�cation, and Semantics-PragmaticsIssues. Preliminary Version'. In P. Bosch and R. van der Sandt (eds.),363{378.[34] Pierrehumbert, J. (1980). The Phonology and Phonetics of EnglishIntonation. Ph.D. Disertation. MIT, Cambridge (Mass.). Distributedby the IULC.[35] Pierrehumbert, J., and J. Hirschberg (1990). `The Meaning of Into-national Contours in the Interpretation of Discourse'. In P. Cohen,J. Morgan and M. Pollack (eds.) Intentions in Communication, MITPress, Cambridge.[36] Pollard, C., and I. Sag (1987). Information-Based Syntax and Seman-tics. Vol. 1: Fundamentals. CSLI, Stanford.[37] Pollard, C., and I. Sag (1994). Head-Driven Phrase Structure Gram-mar. University of Chicago Press, Chicago, and CSLI, Stanford.[38] Prince, E. (1981). `Toward a Taxonomy of Given-New Information'. InP. Cole, Radical Pragmatics. Academic Press, New York, 233{255.[39] Reinhart, T. (1982). `Pragmatics and Linguistics: An Analysis of Sen-tence Topics'. Philosophica 27, 53{94.[40] Rooth, M. (1985). Association with Focus. Ph.D. Dissertation Univer-sity of Massachusetts, Amherst.[41] Rooth, M. (1992). `A Theory of Focus Interpretation'. Natural Lan-guage Semantics 1, 75{116.[42] Steedman, M. (1991). `Structure and Intonation'. Language 67, 260{296.[43] Steedman, M. (1992). `Surface Structure, Intonation and \Focus" '. InE. Klein and F. Veltman (eds.) Natural Language and Speech. Sympo-sium Proceedings, Brussels, November 1991. Springer, Berlin.[44] Steedman, M. (1993). `The Grammar of Intonation and Focus'. In P.Dekker and M. Stokhof (eds.) (1993), Proceedings of the Ninth Ams- 32

terdam Colloquium, December 14{17, 1993, Part III. Illc, Universityof Amsterdam.[45] Svoboda, A., and P. Materna (1987). `Functional Sentence Perspectiveand Intensional Logic'. In R. Dirven and V. Fried (eds.) Functionalismin Linguistics. John Benjamins, Amsterdam.[46] Szabolcsi, A. (1981). `The Semantics of Topic-Focus Articulation'. In J.Groenendijk, T. Janssen and M. Stokhof (eds.) (1981), Formal Methodsin the Study of Language. Mathematical Centre, Amsterdam.[47] Szabolcsi, A. (1983). `Focussing Properties, or the Trap of First Order'.In Theoretical Linguistics 10, 125{145.[48] Vallduv��, E. (1992). The Informational Component. Garland, NewYork.[49] Vallduv��, E. (1993). `Information Packaging: A Survey'. Report pre-pared for Word Order, Prosody, and Information Structure. Centre forCognitive Science and Human Communication Research Centre, Uni-versity of Edinburgh.[50] Vallduv��, E. (1994). `The Dynamics of Information Packaging'. InEngdahl (ed.), 1{27.[51] Vallduv��, E., and R. Zacharski (1993). `Accenting Phenomena, As-sociation with Focus, and the Recursiveness of Focus-Ground'. In P.Dekker and M. Stokhof (eds.) (1993) Proceedings of the Ninth Amster-dam Colloquium, December 14{17, 1993, Part III. Illc, University ofAmsterdam.

33