Broaden Your Views: Implicatures of Domain Widening and the "Logicality" of Language Author(s): Gennaro Chierchia Reviewed work(s):

Source: Linguistic Inquiry, Vol. 37, No. 4 (Autumn, 2006), pp. 535-590 Published by: The MIT Press Stable URL: http://www.jstor.org/stable/4179384 . Accessed: 30/07/2012 14:13
Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at . http://www.jstor.org/page/info/about/policies/terms.jsp

.
JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms of scholarship. For more information about JSTOR, please contact support@jstor.org.

The MIT Press is collaborating with JSTOR to digitize, preserve and extend access to Linguistic Inquiry.

http://www.jstor.org

BroadenYour Views: Implicaturesof Domain Widening and the "Logicality" of Language

Gennaro Chierchia
This articlepresentsa unified theoryof polarity-sensitiveitems (PSIs) basedon the notionof domainwidening.PSIs includenegativepolarity items (like Italianmai 'ever'), universalfree choice items (like Italian andexistentialfree choice items (like Italqualunque'any/whatever'), ian uno qualunque'a whatever').The proposalis based on a "recursive," grammatically driven approach to scalar implicatures that breakswith the traditional view that scalarimplicaturesarise via postgrammatical pragmaticprocesses. The main claim is that scalaritems optionally activate scalar alternativesthat, when activated, are then recursivelyfactoredinto meaningvia an alternativesensitive operator similarto only. PSIs obligatorilyactivatedomain alternativesthat are factoredinto meaning in much the same way. Keywords:scalarimplicatures,negative polarity,free choice, alternative semantics

1 Introduction Overthe past few years, substantive of the semantics progresshas been madein our understanding 1 of negative polarityitems (NPIs). Therehas also been (quiterecently,in fact) important progress in the analysis of free choice items (FCIs).2As is well known, a stronglink exists between these two types of polarity-sensitive item (PSI). Robusttypologicalconsiderations point in thatdirection. According to Haspelmath(1997), roughly half of the approximately150 languages he surveyed
The basic outline of this articledates back to October2002, when it was presentedat the Universityof Massachusetts at Amherst. The first written version was finished in January2004 and was submittedto this journal. Extremely good comments by the referees and by many colleagues, more than I can acknowledgehere, have helped to give it its current form and to overcome, I hope, some of the initial difficulties. I know that problemsremain,and I have tried not to hide those I am aware of. Special thanks to Ivano Caponigro,Carlo Cecchetto, Veneeta Dayal, Danny Fox, Jon Gajewsky, Angelika Kratzer,Orin Percus, Luigi Rizzi, and Philippe Schlenker.I benefited also from presentationsof this work at the Universityof Potsdam(January 2003), StanfordUniversity(April 2003), the Ecole Normale Sup6rieure (May 2003), La Breteche (April 2004), and the University of Amsterdam(May 2004). After I completed the final version of this article, Menendez-Benito2005 became available. Furthermore, I co-taught a seminarwith Kai von Fintel, Danny Fox, and Sabine Iatridouon the topics covered in the article. As a result, it is more than ever an incomplete and partial "on the road"'report. lI have in mind, in particular,Kadmon and Landman 1993, Krifka 1995, and Lahiri 1998. For background,see referencestherein. 2 See especially Dayal 1998 and Kratzer and Shimoyama 2002. For relevant backgroundand alternatives,see referencestherein.
Linguistic Inquiry,Volume 37, Number4, Fall 2006 535-590 C) 2006 by the MassachusettsInstituteof Technology

535

536

GENNARO

CHIERCHIA

employ the same morphemesfor both negativepolarityandfree choice (FC) uses of PSIs, English being among them. The other half employ different series for the two uses, as is the case in Romance.If so many unrelatedlanguagesselect the same morphemesfor such seemingly diverse functions,the link between those functionscannotbe accidental.FCIs and NPIs must form grammatical classes that, although not identical, have a deep systematic relationshipto one another. However, the exact natureof this relationshipremains the object of an intense debate that has not yet reachedfirmconclusions(see, e.g., Horn 1999 for a criticaldiscussionof variouspositions). Here is, for example, an outstandingpuzzle. There are NPIs like English ever/Italianmai that (togetherwith minimizersand N(egative)-words;on the latter,see, e.g., Laka 1990) disallow FC uses, and there are FCIs like qualunquein Italianthat disallow negative polarityuses; in contrast with this, there are words like any (or irgendeinin German)that have both negative polarityand FC uses. Why? Let P1 be the propertythat characterizesNPIs that disallow FC uses (mai) and P2 the propertythat characterizesFCIs that disallow negative polarity uses (qualunque). Such propertiesmust be incompatible:having P1 (being an NPI like mai) must entail not having P2 (being an FCI like qualunque).Obviously, then, we cannot say that any has both P1 and P2, for such propertiesare incompatible.We could say thatany can have eitherproperty:it can be either an FCI or an NPI. This is tantamountto saying that any is ambiguous.But as we know from Haspelmath'ssurvey, roughly one languageout of two is like English: it has PSIs that do double duty. So the equivalentof any is lexically ambiguousin every second language.And which other lexical ambiguityworks that way? In the present article, I attemptto contributeto this ongoing debate by offering a precise hypothesisaboutthe semanticsand syntax of NPIs andFCls and theirrelationshipto one another. Among other things, such a hypothesis should explain why some morphemesallow only one of the two uses, why other morphemesallow both, and why items of the first kind so often mutate into items of the second kind. Building on the work cited in footnotes 1 and 2, I claim thatdomain widening, properlyconstrued,indeed constitutesa unifying basis for understanding PSIs. It also turns out that domain widening (throughthe role it plays in the grammarof polarity-sensitive relations)also constitutesan importantsource of insight into the relationshipbetween pragmatics and the computationalsystem of grammar. In the remainder of this introduction, I will informallyflesh out the main issues surrounding these questions and discuss in what ways they are of interest for the architectureof Universal Grammar. The domain-wideninghypothesis, since first proposed by Kadmon and Landman(1993), has been the main semantic insight aroundwhich investigationsof PSIs revolve. The intuition behindit is this. It is well known that as we communicate,we select domainsof discourseas our subject matter.NonreferentialDPs like every student,a student,and some studentare used with such domains in mind. For example, when we say, "Some studentdoesn't know me," we mean somethinglike 'some studentin D' (or 'some studentD', for short),where D is a set of individuals salient in the context of use (e.g., studentson this campus/inthis city/in this country/etc.).3 What
3 A standardreference in this connection is Westerstahl1988.

BROADEN

YOUR

VIEWS

537

Kadmonand Landmanpropose is that NPIs are indefinites(with a core semanticssimilarto that of some studentor a student),with the additionof an instruction to considerdomainsof individuals broaderthan what one would otherwise have considered. (1) a. a/some studentD b. any studentD + where D c D + If use of a plain indefinite alsome studentwould have naturallyled the hearerto focus on some salient domainD (say, the studentsaroundhere), use of any studentinvites the hearerto consider a set possibly largerthanD along some relevantdimension,with the inclusion of cases thatmight have otherwisebeen consideredmarginal(visiting students,studentson leave, or what have you). This rathersimple idea has the potentialfor explaining why NPIs like being in "negative" environments.Considera typical contrast: (2) a. *Thereis any studentD+ (in that building). b. There isn't any studentD+ (in that building). In a positive context, like (2a), widening the domain of an existential leads to a statementthat is weaker(i.e., less informative)than what we would obtain with a plain indefinite. Suppose, for example, that the set of new studentsis salient and that we would thereforebe thinkingof them in uttering"There is a studentin thatbuilding." Then, if our utteranceis in fact true, it remains so for any largerdomain (say, one that containsnew or old students).So what could be the point of widening the quantificational domain in such a case? If one is willing to accept an existential statementover some domainD, one shouldbe readyto accept it for any broaderdomain.Domain widening seems purposelessin positive contexts. Things are differentwithin the scope of negation.In such a case, considerationof a broader domain leads to a stronger (hence more informative)statement.For example, it may be used to convey that if you were focusing on new students,not only are there none of those around,but also there are no old studentsaround.In otherwords, there simply isn't any student(new or old) around.Broadeningour view is a sensible thing to do; in fact, it is a linguistic move we know we can make in more than one way (There wasn't a single student, There weren't students at all, etc.). So domain widening provides a natural"functional" basis for explaining the contrast in (2). The appeal of this line of explanationcan perhapsbe best appreciatedas follows. It was discovered in the 1970s that NPIs often like being in contexts that share a certainratherabstract propertywith negation-namely, downwardentailment:the capacity to license inferences from sets to subsets (John is not a smokerentails John is not a Marlborosmoker,etc.). Now we have a simple hypothesisaboutthe communicativefunctionof NPIs (i.e., domainwidening)thatmakes us readily see why such items would want to be used in downward-entailing (DE) contexts. Only theredo they seem to serve a reasonablecommunicativepractice:maximize informationcontent. This insight, of course, has to be turnedinto a "real" grammaticalconstraint:how does one go from basic "functionalistic" intuitionsbased on domain widening to actualgrammatical

538

GENNARO

CHIERCHIA

conditions, that is, pieces of the computationalsystem (that, say, rule sentences like (2a) out and sentences like (2b) in)? Thereis disagreementon how to accomplishthat. Kadmonand Landman (1993) stipulatea construction-specific constraint thatlimits domainwidening semantic/pragmatic to occurringonly in contexts where it leads to strengthening (in a sense, they try to make it part of the lexical meaningof any). Krifka(1995), instead,links domainwidening directlyto quantity implicatures.An NPI activates alternativeswith smallerdomains;this triggersan implicature,in accordancewith Griceanprinciples,that the alternativeselected is the strongestthe speakerhas evidence for. Finally, Lahiri (1998) proposes that the alternativesassociated with NPIs play a role similar to the one they play in focus semantics (cf. Rooth 1985, 1992); more specifically, NPIs have as part of their lexical meaning something that resembles the meaning of the focus particleeven. EvenJohndrankindicatesthatJohnwas the least likely personto drink.An indefinite with a widened domain does the same. There is(n't) any studentis interpreted roughly as 'There is(n't) even one student' (which makes sense only in DE contexts). The key issue thatarisesin this connectionis this:how does the pragmatics of communication interactwith specific lexical/grammatical conditions that license the presence of certainitems in certain structuresand not in others?Why do pragmaticallydriven conditions, which usually can be overridden,give rise in the case at hand to unsavablegrammaticality contrastssuch as those in (2)? By studying PSIs, we can hope to learn more about this fundamentalissue. Kratzerand Shimoyama(2002) have arguedthat domain widening may also play a role in the analysis of FCIs. They study in particularthe GermanFC indefinite irgendein. One of its canonical uses is illustratedin the following example: (3) Ich werde irgendeinen Doktor heiraten. I will a whatsoeverdoctor marry 'I will marryany doctor.' Intuitively, (3) indicates that I intend to marrya doctor, and that I am not at all choosy about who that might be: any doctor whatsoeveris possible. Kratzerand Shimoyamapropose that this too might be an implicaturetriggeredby domain widening. They argue that strengthening is not the only reason why one might want to widen a certaindomain. Extremeuncertaintyand hence reluctanceto rule out even the most far-fetchedpossibility might be anothersensible reason for exploiting domainwidening.By telling you thatthe indefiniterangesover a wide domain,I signal to you my intention not to rule any conceivable option out-whence the FC interpretation that any doctor is an option. This line of reasoninginsightfully extends the domain-wideningidea to FC uses. It also raises questions parallel to those we encounteredin our brief discussion of the grammarof "pure" NPIs. How can pragmatic,conversation-driven processes determinestrict morphosyntactic patterns?And what is the relation between two apparentlyvery differentuses of domain widening? Against this generalbackground, there are additionalspecific issues in the grammar of FCIs that stand out as particularly controversialand that may play an importantrole in advancingour One concerns their relation to modality. FCIs seem to be felicitous basically in understanding.

BROADEN

YOUR

VIEWS

539

the presence of (certainkinds of) modals, a point forcefully made by Dayal (1998). Even when to attaininterpretasuch modalsarenot overtlypresent,some kindof modalityseems to be required bility. Take for instance the following Germanexample: (4) Gestern hat irgendein Studentfur dich angerufen. yesterdayhas a whatever student for you called 'Yesterdaya student(I don't know/don't care who) called for you.' Even though this is clearly an episodic sentence (i.e., not modalizedby anythinglike an implicit generic), it indicatesthatthe speakerdoesn't know or doesn't care aboutthe identityof the caller, Consider, so it requiresthe presenceof a covertepistemicmodal of some sortfor its interpretation. by the same token, the following typical example of an FC use of English any: (5) YesterdayMary saw any studentthat wanted to see her. Sentence (5), like sentence (4), is episodic. Still, such a sentence seems to invite counterfactual conclusions:if, say, Joe had fancied seeing Mary, she would have seen him. This effect is subtly but robustly apparentmore with any than with its cousin and near synonym every (see Dayal 1998 for arguments).Wheredoes this implicit modality come from? Why does it patternin such peculiarways? force of FCIs. Germanirgendeinappearsto be A relatedissue concernsthe quantificational definitely existential. Sentence (3) indicates my willingness to marryone doctor, and sentence (4) indicates thatjust one studentcalled. In contrast,FC any, as exemplified by sentences like (5), appearsto be definitely universal. If one student wanted to see Mary and didn't, sentence (5) would be false.4 At the same time, even FC any (which is so clearly amenable to being understood universally)appearsto acquirean existentialflavorin certaincontexts.As Giannakidou (2001) observes, imperativesare one such context. (6) To continue, push any key. A sentence like (6) does not typically constitutean instructionto push all keys. Summing up, a host of intriguingopen questions surroundpolarity sensitivity. The main ones I intend to pursue are these: (7) a. Can domain widening constitutea semanticinsight capable of unifying all cases of polarity sensitivity (from NPIs to FCIs)?

4The universalcharacter of English any is arguedfor more extensively in Dayal 1998. Similarargumentshave also been developed for FCIs in Scandinavianby Seb0 (2001).

540

GENNARO

CHIERCHIA

b. Can domain widening, in particular,explain why different types of FCIs vary in force and in their link to modalities? their quantificational c. Domain-widening-basedaccounts are always pragmaticallydriven. What can we learn from domain widening about the relation between the computationalsystem and the pragmaticsof communication? The articleis organizedas follows. In section 2, I identify more explicitly the patternof FC constructionsin Italian, which bears out and justifies the claim that there are at least two types of FCls, an "existential" one and a "universal" one, with distinctscope properties.This pattern provides a rich testing ground for the hypothesis to be developed. In section 3, I present some This provides a general framebackgroundassumptionson the role of implicaturesin grammar. work for addressingthe role of pragmaticsin the grammarof polaritysensitivity. Hopefully, the functioningof PSIs will flow naturallyout of the grammarof implicatures.This idea is explored in the subsequentsections. In section 4, I discuss NPIs, and in sections 5 and 6, the two types of FCIs, "existential" and "universal." In section 6, I offer some tentativegeneralconclusions. Formal details are worked out in the appendix. 2 Some Italian Data: Two Types of Free Choice Items Italian(and, more generally,Romance)turnsout to be a good languagewith which to investigate the quantificational force of FC elements,for it has two relatedbutclearly differentsuch elements. The first is [un N qualunque/qualsiasi],which closely resembles Germanirgendein.The second is [qualunquelqualsiasi N], which moreclosely resemblesFC any. They clearlycontrastin quantificational force. Here is a minimal pair: (8) a. ?Sono uscito in strada e mi sono messo a bussarecome un matto like a madman (I) went out on the street and startedknocking ad una porta qualsiasi con i battentiin legno. at a door whatever with wooden shutters b. Sono uscito in strada e mi son messo a bussare come un matto like a madman (I) went out on the street and startedknocking a qualsiasi porta con i battentiin legno. at whateverdoor with wooden shutters Out of the blue, (8a) is slightly marginal;however, it can be interpreted if we imagine a context in which the agent goes out withoutknowing what to do and acts upon a door selected randomly. In such a (semimodalized)context, (8a) is interpreted existentially:I knocked on one door. The modifiercon i battentiin legno 'with wooden shutters'can readilybe construedin a nonrestrictive manner.Sentence (8b) is instead understooduniversally (I knocked on all doors with wooden shutters),andthe modifierhas to be construedrestrictively.The existence of differentconstructions (ultimately involving different lexical items) with different quantificationalforce clearly needs to be understoodbetter:if domain widening is systematicallyinvolved in FCIs, how can it give rise to such diverse effects?

BROADEN

YOUR

VIEWS

541

Schematically,this is the form of FCIs in Italian: (9) a. [INDEF NOUN FCI]5 dolce qualsiasi/qualunque un a sweet whatever due dolci qualsiasi/qualunque two sweets whatever b. [FC NOUN] dolce qualunque/qualsiasi whatever sweet The constructionsin (9a) and (9b) are probablysyntacticallyrelated, though I will not attempt any seriousanalysisof theirsyntacticstructure here.Froma semanticpointof view, these constructions have a common core (which I will try to bring out). However, as pointed out above, they also clearly differ in quantificational force, with (9a) interpreted existentiallyand (9b) interpreted much more universally (if one may say so). This was illustratedfor episodic contexts in (8). Under modals, we find a similar pattern(with one distinction,as we will see directly). (10) Future a. Domani interroghero qualsiasi studente tomorrow(I) will interrogatewhateverstudent che mi capiter'a a tiro. that I will lay my eyes on b. Domani interroghero uno studentequalsiasi. tomorrow(I) will interrogatea student whatever Imperative c. Prendiqualunquedolce. take any sweet d. Prendiun dolce qualunque. take a sweet whatever Modals of possibility e. Puoi prenderequalunquedolce. sweet (you) can take any V, 3

V, 3 3

V, 3 (?)

5The order [INDEF FC NOUN] is also found. uomo (i) un qualsiasi/qualunque a whatever man

In this order, however, the only possible realizationfor INDEF is the indefinite article. Numeralsare disallowed. (ii) *due qualsiasi uomini two whatever men I do not know why this is so.

542

GENNARO

CHIERCHIA

f. Puoi prendereun dolce qualunque. a sweet whatever (you) can take Modals of necessity prenderequalunquedolce con il liquore. g. Devi sweet with liquor any (you) must take [3-favoring context: If you go to Naples, you must go to Scaturchio] prendereun dolce qualunquecon il liquore. h. Devi a sweet whatever with liquor (you) must take

V, 3 (?)

admits a universal N] like (lOa). It uncontroversially Take a sentence with [qualunquelqualsiasi reading:(lOa) can readily be used to express my intentionto interrogateall students.However, within the scope of a modal (unlike what happens in episodic contexts such as (8)), [qualsiasi N] also seems to admit an existential reading;for example, I can also use (10a) to express my intentionto interrogate just one student.With some modalities(e.g., with imperatives),this ambiguity is very clear. In other cases (e.g., in (lOg)), the universalreading seems to be favored and a special context might be called for in orderto get the existential reading. So, there is a sharpand systematiccontrastbetween the [qualsiasi N] and [un N qualsiasi] The formeralways admitsa universalreading;however, in the scope of an overtmodal, structures. it also seems to be able to have an existentialreading(at least, often enough). The latteris always existential and is interpreted universally,if at all, in highly markedcircumstances. in (9a), andI will use universal Fromnow on, I will reserveexistentialFCIs for the structures FCIs for those in (9b). These are intendedas descriptivelabels (withoutprejudgingthe analysis). Another interesting difference between existential and universal FCIs concerns what has come to be known as the "subtrigging" effect,6 illustratedby the following paradigm: ho parlato con un qualsiasi filosofo. (11) a. ??Ieri yesterday (I) have spoken with a whateverphilosopher 'YesterdayI spoke with a philosopher(I don't know/don't care who).' ho che fosse b. ??Ieri parlatocon un qualsiasi filosofo yesterday (I) have spoken with a whateverphilosopherthat was.suBJuNc interessatoa parlarmi. interested in speaking with me 'YesterdayI spoke with a philosopher (I don't know/don't care who) that was interestedin speaking with me.' c. ??Ieri ho parlato con qualsiasi filosofo. philosopher yesterday (I) have spoken with any 'YesterdayI spoke with any philosopher.'

6 This terminologyis from LeGrand 1975.

BROADEN

YOUR

VIEWS

543

d.

leri ho parlato con qualsiasi filosofo che fosse yesterday(I) have spoken with any philosopherthat was.suBJuNc interessatoa parlarmi. interested in speaking with me 'YesterdayI spoke with any philosopherthatwas interestedin speakingwith me.'

Sentence (1 Ia), in which the existential FCI appearsunmodified, is marginal out of the blue (unless a special context is provided);if anything,the additionof a relative clause makes things worse (1 lb). Also, when unmodified,as in (1 Ic), the universalFCI is marginal(unless a special context is provided);however, the additionof a relative clause as in (ld) makes it completely for universalFCIs in episodic senacceptable.A modifier seems to restore full grammaticality tences. No similar effect is detectablewith existential FCIs. A final patternthat we will discuss involves the interactionof Italian FCIs with negation. This patternis potentially telling, as it reveals furtherscope differences between the two types of FCIs. A sentence like (12), for example, where negation has scope over a universal FCI, typically is acceptable only with the special intonationassociated with the so-called rhetorical reading. (12) Non leggero qualunquelibro. (I) won't read (just) any book Sentence (12) says that it is not the case that I will read every book (i.e., V) and suggests that I am going to read some special one. If we add a modifier and make the FCI more heavy (i.e., perhapsmore topical), things change. The rhetorical'notjust any old one' readingremainspossible. But next to it, a novel one appears. (13) Non leggero qualunquelibro che mi consiglieria Gianni. that Gianni will recommendto me (I) won't read any book Sentence (13) can also express that I simply won't read any book suggested by Gianni (i.e., a V 1 3 reading). So universal FCIs, at least in certain cases, display a scopal ambiguityvis-'a-visnegation. In contrast,an existential FCI embeddedundernegation has only the rhetoricalreading. (14) Non leggero un libro qualunque(che mi consiglierna Gianni). (I) won't read a book whatever (that Gianni will recommendto me) Sentence (14) can only mean thatI won't read any old book (recommended by Gianni).This fact is particularlyinteresting as it differs from what Kratzerand Shimoyama (2002) report about German irgendein (which is otherwise so similar to Italian uno qualunque). Under negation, Germanirgendeinis ambiguousbetween a rhetoricaland a nonrhetoricallNPI-like reading(as is the case with (13) in Italian).7Anyway, on top of this interestingcrosslinguisticcontrast,we see
7As Kratzerand Shimoyamaunderscore,the 'not just anyone' readingrequiresa special intonationor the presence of a focus particle.

544

GENNARO

CHIERCHIA

thatin Italian,universalandexistentialFCIsbehave differentlyundernegation,a differencewhose rationalewe would like to understand. Thus, ItalianFCIs display a ratherinterestingand in certainregardspuzzling pattern,which enables us to integrate the generalizationspresented so far in the literature.In particular,the in existentialityvs. universality) existence (in fact, coexistence) of two kinds of FCIs (contrasting with distinct scopal properties seems to be empirically supported.The interesting theoretical question is how exactly these two types of FCIs are related to each other and to other polarity phenomena. 3 Background: Pragmatics in Grammar As pointed out in section 1, the semanticallybased approachesto polaritysensitivity that we are considering all appeal to pragmatics,in some form or other. The problem that arises in this connectionis how pragmaticand morphosyntactic processes interactwith each otherin a modular system. With respect to this problem, I will be assuming that certainpragmaticprocesses (i.e., sysprocesses involving the speaker's intentionsand other aspects of the conceptual/intentional tem) are visible to (and accessed by) the computational system. More specifically, (some) implicatures are computedrecursivelyand compositionally,on a par with ordinarymeaningcomputation (and therefore are not part of a postgrammaticalprocess). The main motivation for such an assumption,in a nutshell, is twofold. First, NPI licensing can occur at any level of embedding. If implicaturesplay a role in such licensing, they must be computed at the relevant embedded site, on a par with compositionalsemanticprocesses and othercyclic (or phase-driven)syntactic processes. Second, scalar implicaturesplay a key role in deriving the propertiesof FCIs, as we will see. If so, then scalar implicaturecomputation must be partof (or accessible to) the computational system that determinesthe syntactic distributionof FCIs. An earlyapproach to pragmatics along these lines was developedby Gazdar(1979). Recently, similar ideas have been revived in work on "maximization" (Landman1998) and other scalar implicatures(Chierchia2004).8 Some generalconsequencesof these views for modernpragmatics are addressedby Recanati(2003). The approachto polarity sensitivity to be developed here has to rely on frameworksof this sort. For explicitness' sake, I will now outline a compositional system of scalar implicaturecalculation,as an example of "recursivepragmatics."I will do so in informalterms, leaving formaldetails to the appendix.The system I will presentis a slight (?) modificationof the one developed in Chierchia2004. It retainsmany featuresthatare speculative and approximate. 3.1 A RecursiveApproach to Pragmatics Each expression (or rather,its LF representation) in is associated with its meaning/denotation familiar ways. For example, (15a) is interpreted, say, as in (15b).
8 A bibliographicalremark.The basic ideas in Chierchia2004 were elaboratedin 1999 (and presentedat a series of workshopsand other venues); a writtenform essentially identical to the publishedversion has been circulatingsince 2001.

BROADEN

YOUR

VIEWS

545

(15) a. Many of your studentscomplained. b. Ilmany of your studentscomplainedll= manyD(of your students)(complained) I use logical formulas as stand-ins for the correspondingdenotations (see section Al of the appendix).The inferentialprocess throughwhich the (canonicalscalar)implicature arises, according to the familiarGriceanproposals,is often characterized along the following lines. First,when (iSa) is uttered,typically the set of alternativesin (16a-c) is being considered;this promptsthe inference in (16i-vi).9 (16) a. Some of your studentscomplained. b. Many of your studentscomplained. c. All of your studentscomplained. i. The speakerchose to utter(b) over (a) or (c), which would have also been relevant. ii. (c) entails (b), which entails (a) [the quantifiersform a scale]. iii. Given that (c) is strongerthan (b), if the speakerhad the informationthat (c) holds, she would have said so [quantity]. iv. The speakerhas no evidence that (c) holds. v. The speakeris well informedon the relevantfacts. Therefore vi. The speakerhas evidence that it is not the case that (c) holds. Notice thatthe last step, unlike the previousones, is not readilyjustifiable on the basis of Grice's maxims and (pure)logic. It seems to requirea "leap of faith" aboutthe informationstate of the speaker. Such a leap, called by Sauerland(2005) the epistemic step, is tantamountto a sort of neg-lowering,thatis, to pushingnegation across an epistemic modal (from not has evidence that to has evidence that not). This step is crucial in deriving scalar implicaturesand will also play a key role below in deriving the implicaturecharacteristic of FCIs. It is evident thatthe process in (16) does not consciously take place wheneveran implicature comes about. Rather,it seems to be automaticand unconscious in hearers/speakers just like so This suggeststhatit may be wrongto limit processes manyotheraspectsof semanticinterpretation. of this sortto root sentences.It is truethatthe reasoningin (16) concernsthe effects of utterances. But embeddedclauses are, after all, potentialutterances.And surely speakersdo routinelywork out the possible conversationaleffects of potential utterances.So it is conceivable that we run througha process like (16) in a cyclic manner,computingthe "utterancepotential" of embedded clauses compositionally.I will pursuethis idea here, by assuming that there are operationsthat "enrich" basic meanings and freely take place at scope sites. Such operations (together with certainassumptionsregardingfunctionalapplication)constitutethe core of recursivepragmatics. A crucial part of (16) is the observationthat a sentence is typically consideredagainst the of a set of alternatives. Once the alternative set (e.g., (16a-c)) is salientto illocutionary background sentencewill be per se informative.In this connection,Krifka(2003) agents,choosing a particular
9 The source is Grice 1989. Also see Horn 1989, Levinson 2000, and referencestherein.

546

GENNARO

CHIERCHIA

speaks of "motivatedinterpretation of alternatives,"typically guided by the awarenessthat one could have made weakeror strongerassertions.We can imagine a function11JALT that associates any item with its scalar alternatives.For example: (17) limanyof your studentscomplainedliALT { someD(of your students)(complained), manyD(of your students)(complained), everyD (of your students)(complained) }
-

We computesuch a set of alternatives using the sameoperationswe use to computeplainmeanings. In fact, this set can be computedin the same way as alternativesemanticsfor questions(Hamblin 1973) or focus (Rooth 1985, 1992). And something like (17) can, accordingly,be thoughtof as specifying one of the questions/issues under discussion, namely, the question "Roughly how many of your studentscomplained?" Alternativeskeep growing until they are factoredinto meaningby some operationthat produces pragmaticallyenrichedinterpretations. In the case of scalaralternatives,such an operation be can characterized rathersimply. The speakersuggests thatthe one she picks and its entailments is the only alternativeshe regardsas true. (18) flmanyof your studentscomplainedlls= a. manyD(of your students)(complained) A A p Vp pe limanyof your studentscomplainedlIALT (of your c students)(complained) p (where "c" stands for 'entails') manyD b. manyD(of your students)(complained) A allD (of your students)(complained) It is easy to see that (18a) is equivalent to (18b). The format in (18b) (adopted in Chierchia 2004, building on Krifka 1995) makes the scalar reinforcementtransparent. The formatin (18a) (proposedin Fox 2003, building on Groenendijkand Stokhof 1984) brings out the relationship between scalar enrichmentand adding a silent only to the basic meaning.In other words, it is as if scalar items bring to salience a question of the form "Roughly how many . .. ?" and the sentence winds up being taken as an exhaustiveanswer to such a question. Putting all this together, and adoptingthe abbreviationin (19a), we can define enrichment as in (19b). (19) a. Oc [q] - q A Vp [[pC A p] -> q c p] (O is a mnemonic for only: q and its entailmentare the only members of C that hold)10 b. 11iAts = Oc [OctILJ, where C = jj+||ALT At this point, the parallelwith focus semanticsbecomes hardto miss. The only differenceis that scalar alternativesare lexically driven and not necessarily activated by any special accentual pattern.
10I will assume that, for any p, Oc (p) is defined only if a suitable set of alternatives(in the case at hand, scalar ones) is available.

BROADEN

YOUR

VIEWS

547

A characteristic of scalarinferences is that they can be suspended.If we assume that scalar terms activate alternativesby default and that alternativesmust be factored into meaning, how is suspensionof implicaturespossible? We must assume thatthe defaultactivationof alternatives can be, in turn, suspended. A simple way to achieve suspension is to assume that each scalar is item comes in two variants(say, thanks to an abstractmorphologicalfeature [ ? c], where Oa mnemonic for strong): many[1i], or[ ,], and so on; a [+ F]item (e.g., many[?,]) has active - ]) has no active alternativesand must lead to enrichment,while a [- o] item (e.g., many[ alternatives and cannotlead to enrichment.Speakerschoose the featuresettingthatfits the context best.11[?+ ] correspondsto the [? F]-featurecommonly used in focus semantics;but whereas o is not (or not necessarily). F is phonologically interpreted, Actually, thereis a furtherdifferencebetween implicaturesand focus thatmakes things even more interesting(as it requiresthinkingof enrichmentrecursively).Under embedding,implicatures are sometimes preserved and sometimes "recalibrated."Let us see how this works by looking at an example. Consider a sentence like (20a). In principle, it can be enriched in two ways, representedby (20b-c) and (20d-e). (20) a. John believes that many of your studentscomplained. b. John believes that many of your studentscomplained and it is conceivable for all that John believes that not all did. c. Oc [believe (j, manyD(of your students)(complained))]= believe (j, manyD(of your students)(complained)) A believe (j, allD (of your students)(complained)) d. John believes that many, though not all, of your studentscomplained. e. believe (j, Oc [manyD(of your students)(complained)])= believe (j, manyD(of your students)(complained)) A allD (of your students)(complained) Working things out will reveal that enriching at the root level as in (20b) yields a ratherweak interpretation (compatiblewith John's believing that it is possible that all studentscomplained). Enrichingat the level of the embeddedclause as in (20d-e) results in something considerably stronger(it entails (20b)). I think that (20d), the version with the embeddedimplicature,is the and preferredreading.Whetherthis is right or not, (20d-e) is certainlya possible interpretation, to obtainit we must countenancethatbelieve applies to the enrichedinterpretation of its complement. That is, we countenancean applicationrule of the following form: (21) libelievethat Slis = Ilbelievells(dlthat Slls) Here, we see the recursiontaking shape.

1l This may look pretty close to an approachthat regardsscalar items as being ambiguousbetween a strong and a weak construal.But as we will see in discussing DE contexts, it is not so. The distributionof readings of scalar items in a DE context is clearly beyond the scope of any simple-mindedambiguityapproach.

548

GENNARO

CHIERCHIA

Thingschangeconsiderablyif we considera sentencelike (22a). Here,the embeddedimplicadoubtas -- believe). turecorrespondsto (22b-c), and the matrixone to (22d-e) (I am representing (22) a. John doubts that many of your studentscomplained. b. John doubts that many but not all of your studentscomplained. c. -- believe (j, OALT manyD (of your students)(complained)) d. John doesn't believe that many of your studentscomplainedbut believes that some did. e. -c believe (j, manyD(of your students)(complained))= -- believe (j, manyD(of your students)(complained)) A believe (j, someD (of your students)(complained)) is only available like (22b). This interpretation Sentence (22a) hardly ever has an interpretation stress on many); more normally, if (22a) in special contexts (and with the help of appropriate implicatesanything,it implicatessomethinglike (22d). Here, the original(embedded)implicature disappears-and a new one surfaces.'2 When (20) is comparedwith (22), it is immediatelyclear thatthe factor responsiblefor this patternmust be the monotonicity propertiesof doubt, which is a DE function (more or less assimilable to "not believing"). Roughly speaking, (canonical) implicatures(like those from many to many but not all) may well be preservedunder embedding within non-DE (i.e., nonwhen embeddedin DE functors "negation-like") functions,while typically they are recalibrated (a generalizationwe will refine shortly). This means that the semantics we use to compute the strong meaning in cases like (20)-(22) is this: that Slls = Oc Ildoubtil(llthat (23) fldoubt Sll) Putting (21) and (23) together,then, we get somethinglike this:'3

131s (24) lIaI

-

= Iotlls(Oc if cxis not DE} fl3fl), IloIls(1l13lls) otherwiseJ tOcIjk_II(IIj3I),

While this implementationis open to the allegation of being ad hoc, and one can surely try to improve on it, it embodies a ratherneat generalization. (25) In enrichinga meaning, accord preferenceto the strongestoption (if there is nothing in the context/commongroundthat preventsdoing so).'4

12 That negation affects implicaturecomputationwas alreadyobserved by Gazdar(1979). Horn (1989) generalized Gazdar's observationto all DE contexts. For detailed presentationand discussion of the relevant facts, see Chierchia 2004. For relevant discussion, also see Levinson 2000. The coming about of the implicaturein (22d) is what a simple-mindedambiguityapproachcannot account for. 13 See section A3 of the appendixfor a more precise formulation. 14 See Dalrympleet al. 1998 for use of a similar principle in a differentdomain.

BROADEN

YOUR

VIEWS

549

This principlepredicts the preferencefor the embeddedenrichmentin (20) and for the root one in (22), which seems in line with intuitionsand, if true, vividly exposes the "spontaneouslogicality" of language.In addingscalarimplicatures,speakersseek to optimize informationcontent contexts (like (= logical strength)in a way thatkeeps trackof the effect of entailment-reversing the DE ones). Notice thatthis reasoningcan applyiteratively(i.e., recursively).Forexample,we can embed a sentence like (22a) further;and if the embeddingfunction is not DE, then we can well get an embeddedimplicature.Considerthe following example: (26) a. I am sure that John doubts that all of your studentscomplained. b. I am sure that John disbelieves that all of your studentscomplainedbut he believes that some did. It is not hardto imagine a situationin which one would utter(26a) with the intentionof conveying something like (26b). of the notion of enrichedmeaning, the switch from So, in a compositionalcharacterization (20) to (22) can be obtained by a "clever" definition of functional application.This gives an idea of how the pragmaticsof scalar implicaturesmay be set up recursively. To complete the picture, I need to say something about multiple scales and implicaturesembeddedin the wrong place, as it were. I do this in the following two subsections.15 3.2 Multiple Scales Often enough, one particularsentence contains more than one scalar item; and if all such items have active alternatives,multiple implicaturesarise. (27) a. Someone[,,] smokes or[+,,] drinks. b. Someone (though not everyone) smokes or drinks(but not both). The strong meaning of (27a) is something like (27b). How can we obtain it? And how do we keep trackof multiple scales? The simplest way seems to me to allow multiplecyclic application of enrichmentat clausal nodes. So assuming an LF representation like (28a) for (27a), we want something like (28b) as its strong meaning. (28) a. someone[+ ,] [ti smokes or[+ ,] ti drinks] b. 0 [some[+,] (one) )xi 0 [smoke(xi) V[+,] drink(xi)]] Now, if we consider the scales of both some and or as part of the same set of alternativesto (27a), we get the following picture:

15"Globalistic" alternativesto this view can be found in Sauerland2004 and Spector 2003. See Chierchia2004 for argumentsagainst globalism. I should add, however, that it is technically feasible to adopt the algorithmsproposed by Sauerlandor Spector and use them in a cyclic manner,along the lines suggested here.

550

GENNARO

CHIERCHIA

(29)

someonei [ti smokes or t, drinks]

someonei [ti smokes and ti drinks]

everyonei[ti smokes or ti drinks]

everyonei[ti smokes and ti drinks] and arrowsindicatethe entailmentrelations.Whathappens,then, is that The spatialarrangement if we try to compute the implicatureat the root level in sentences like (27a), we will not find a uniquescale among the alternativesactivatedby the lexical entries.A naturalstipulationto make in this connection is that in such a situationwe would not know which scale to pick; hence, we would not know how to strengthen.On the other hand, if we apply strengtheningcyclically, handlingimplicaturetriggersin the orderin which they are introduced(as in (27b)), we deal with a uniquescale each time, which simplifies thingsgreatly.As this procedureseems naturalenough, I will adopt it: (30) a. To strengthenvia 0, the scale must be uniquely determined. = Oc (11411), where C is 4)'s scale in j4lJALT b. 1lflJs If thereis more than one scale for 4)in ALT, the definite description"4)'s scale in 4jj(lALT" fails. This forces us to choose the strengthening fails to be properand consequentlystrengthening representedin (27b), a welcome result. The full power of the presentsystem can be appreciated even more if we consider multiple occurrencesof scalar items within DE contexts. Here is a moderatelycomplicatedexample:'6 (31) a. No one who smokes and[+,,] drinkslives to 80[,+ ]. b. There are people who smoke or drink(but not both) and live to 80. c. There are people who smoke and drink and live to an age sufficiently close to 80. Assuming that the scalar terms and and 80 have active alternatives,(31a) implicates (31b) and (3 lc). This is indeed whatour definitionof application predicts;andit is perhapsworthunderscoring that the intended result cannot be obtained througha single applicationof the 0-operator. We have to use it twice, as follows: (32) a. 0 [O [no (Xxi one(xi) A smoke(x1)A drink(xi))](livesto 80)] b. 0 [no (Xx, one(x1)A smoke(xi) A drink(x1))] The squarebracketsin (32a) indicate the scope of 0. Considerin particular the most embedded occurrenceof 0, isolated in (32b). As the type of no one smokesand drinksis ((e, t), t), we have

16The following paragraph is not essential to understanding the basic workings of the system. See section A3 of the appendixfor details.

BROADEN

YOUR

VIEWS

551

to generalize 0 to this type (cf. Rooth 1985). So, in working (32b) out, we would be considering alternativesof this form: (33)

{ XP . no one who smokes and drinks P, XP . no one who smokes or drinksP }, P a

variableover properties

As usual, 0 says that the only alternativethat will hold is the one (that will be) uttered.So we get the following derivation: (34) 0 [no (Xx one(x) A smoke(x) A drink(x))] = XP [no (Xx one(x) A smoke(x) A drink(x))(P) A no (Xx one(x) A (smoke(x) v drink(x)))(P)] = XP [no (Xx one(x) A smoke(x) A drink(x))(P)A some (Xx one(x) A (smoke(x) v drink(x)))(P)] to the VP comes in, the second occurrenceof 0 takes its usual When the argumentcorresponding course and, at the end of the day, we get the intended strengthenedreading for (31a). The fact that the strengtheningof expressions headedby a DE functionrequiresthis stepwise, argumentby-argument, subclausalapplicationof 0 suggests thatmakingit partof the definitionof application itself is indeed correct(per definition (24)). Contraryto what happensfor non-DE contexts, the application of strengtheningto DE contexts cannot be readily accomplished via a clausal applicationof 0."7 Considerationof multiple scalarimplicaturesthus yields interestinglycomplex patternsthat I propose can be handledin systematicways, in spite of theircomplexity.The basic generalizations are (a) that enrichmenttakes place cyclically from the bottom up and (b) that when a function f is applied to an argumentA, if f is not DE, the argumentf(O [A]) is enriched;if f is DE, the result 0 [f(A)] is enriched("recalibration").In either case, additionof scalarimplicaturesleads to strengthening. While this seems to be generallycorrect,there are also cases of enrichmentthat do not lead to strengthening.Such cases too must somehow fit into the picture. 3.3 "Frozen" Implicatures Consideran example like this: (35) If many studentscomplained,we are in trouble. Within (35), the clause many students complained appears embedded in the antecedent of a of (35) is not conditional, a DE context. And in fact the (most salient) enriched interpretation something like (36a) but, if anything,something like (36b).
17

An even more complicatedcase (discussed in Chierchia2004; also see section A3 of the appendix)is this:

(i) Few people that smoke and drink live to 80. This has also the implicature "some do . . . ". Scalar enrichmentsof this complexity are not discussed in proposals alternativeto the present one, as far as I know.

552

GENNARO

CHIERCHIA

(36) a. If many but not all studentscomplained,we are in trouble. b. If many studentscomplained,we are in trouble,while if few studentscomplained, we are OK. If we express these options using the 0-operator, here is what we get: (37) a. if 0 [many studentscomplained],we are in trouble b. 0 [if many studentscomplained,we are in trouble] The scopes in (37a-b) correspondto the interpretations (36a-b), respectively.The preferencefor the interpretation represented by (37b) is in line with the preferencefor the strongestinterpretation (i.e., the option we have alreadyencounteredand discussed). However, there are cases in which a readingisomorphicto (37a) seems to emerge. Consider,for example, the following discourse: (38) If many studentscomplained,then we are betteroff than if all did. For (38) to make sense, the antecedenthas to be interpreted as follows: (39) If many though not all studentscomplained,then we are betteroff than if all did. This type of case (discussed in Levinson 2000) seems to involve an interpretation isomorphicto (37a). Now, as the readercan readily verify, interpretations of this sort are in fact weaker than the plain assertion.In general, it is easy to show that (40) [p -- q] C [O [p] -> q] So (37a) is an example of an enriched meaning that is not a strengthenedmeaning. The proposedsystem is designed to obtainstrengthened meanings.And, as it stands,it does not afford us interpretations like (37a). In Chierchia2004, I suggested that they are to be obtainedthrough something like domain selection. Here I wish to explore a differentpossibility, directly inspired by Fox (2003). We can imagineintroducing at LF somethinglike a "strongestmeaning" operator. So far, 0 has been used only in the semantic metalanguage;we might want to introduce an analogueof 0 at LF. Such an operator,call it "a," would be an abstractassertoricoperatorthat quite literally "freezes" or "locks in" the implicatures.cr [S] has as its (plain) meaning the (strongest)enrichedmeaningof S. Once u appliesto a constituent,the implicature of thatconstituent becomes partof its meaning and hence can no longer be removed or recalibrated. Formally: (41) IIJo Sll
IfSJS'8

In this connection, it is in fact tempting to adopt one of the familiar syntactic modes of projectingLF operators.For example, we might say that the feature [ ? cX] associatedwith scalar items is uninterpretable and needs to be checked by an (interpretable) abstractoperator(r (and vice versa: o" has to have a [ + (r] element in its scope). In the case underdiscussion, (35), such an operatorcan be attachedat different sites, as in (42).

18This is a simplification.See the appendixfor technical details.

BROADEN

YOUR

VIEWS

553

(42) a.

IP

IP

CP

IP

if many[+,] studentscomplained b. IP

CP

IP

if

IP

IP

many[+,, studentscomplained It should be observed,moreover,thata is not a direct syntacticprojectionof the enrichment operation0. For example, it can be shown that if p and q both contain scalar terms, then the following equivalence holds:
(43) a. kIa q[+1]]Il= 0 [p -> 0 q] [pr+?u] b. Example oa [if John drinksand[+,] drives, he gets two[+, ] months' probation] = 0 [if John drinksand drives -> 0 he gets two months' probation] = If John drinks and drives, he gets two months' (and no more) probation,while if he does only one of the two, he does not get two (or more) months' probation.

This is so because a single occurrence of o- can simultaneouslycheck several occurrencesof [ + c] (by analogy with wh-dependencies).The present approachalso rules out representations of the following sort, as cases of featuremismatch(where the second is a violation of minimality/ intervention,however one wants to implementit): (44) a. *v [John smokes or[-,] drinks] a'. Oc (smoke(j) v drink(j)) = undefined b. *vj [Johnis smoking or[-,0] grading some[+,] assignments] b'. 0 [smoke(j) vf-,, some[+,j (assignments)Xx grade(j, x)]

554

GENNARO

CHIERCHIA

as (44a'); this would be (44a) would be interpreted These are welcome results. LF representation semantically undefined (for the alternativesare not active). In (44b), the situation is different as (44b'), where the alternatives like (44b) would be interpreted (and worst).An LF representation associatedwith some are active, but those associatedwith or are not. As can readilybe computed, (44b') entails that John is not smoking, somethingwe clearly do not want as a possible meaning for a sentence like (44b) (for furtherdiscussion, see Chierchia2004, Fox 2003). I should also assumptionson featurecheckemphasizethatthese resultsareobtainedusing completely standard ing (or whateversubsumes its effects). which, as notedabove, has severalantecedents of a strongassertionoperator, The introduction in the literatureand is most directly inspiredby proposalsin Fox 2003, constitutes a departure from Chierchia2004. The link betweenthatproposalandthe presentone is, however, quitetrivial. notion of It boils down to definition (41): a' is defined in terms of the recursivelycharacterized enriched interpretation, 11Ils (which remains essentially the same as before). This is all quite sketchy (something only partiallyremedied in the appendix),but perhapssufficient for present purposes.WhatI have triedto do in this section is to set up a sufficientlyexplicit formalmachine (which can be providedwith some independentmotivation)in orderto formulatea (partlynew) theory of polarity phenomena building on the idea of domain widening. Let us now turn to formulatingsuch a theory. 4 Negative Polarity In this section, I will examine "pure" NPIs, namely, items like mailever that disallow FC uses presentedin section (minimizerslike lift afinger also fall into this category),using the framework 3. As I will show, such a frameworkallows us to readilyconceptualizethe role of the implicatures associated with PSIs. For convenience, I will illustratethe proposal mostly with English any, focusing on its negative polarity facet. The reader should bear in mind that a more adequate of items of the any type will have to wait until section 5. characterization 4.1 "Large" Domain-Alternatives Recursive pragmaticsenables us to systematize (and, in a sense, integrate) the proposals by Kadmon and Landman(1993), Krifka (1995), and Lahiri (1998) on NPIs. To show how, I will startout with a proposal close to Krifka's. Then I will modify it in ways that will bring out its connections to the others. Recall the basic idea: (negative polarity)any in English has the same meaning as an indefinite like some, plus domain widening. I will work toward my proposed implementationof this insight throughan example. Let us assume that every predicate carries a world variable, which is filled according to general principles (see Groenendijkand Stokhof 1984; for a more recent proposal, see Percus can be restrictedto contexlet us assumethatquantification (andabstraction) 2000). Furthermore, tually salient domains. Here is a simple example: (45) a. I saw a/some boy. b. Xw3xe D, [boy,(x) A saww(I,x)]

BROADEN

YOUR

VIEWS

555

Formula(45b) is the propositionexpressed by (45a).'9 I use set variables (D, D', etc.) to mark the (salient) quantificational domainassociatedwith DPs. D typically includes individualswhose existence we are sure about,along with individualswe may be less sure about.Take, for example, our neighborFred. For all we know, he might or might not have sons. So, dependingon specific aspects of the conversationaldynamics,D might include Fred's possible sons or not. Given a set D, Dw are those membersof D that actually exist in w. Fred's sons will be in Dw only if it turns out that in fact they exist in w. Adding (45b) to a common ground(the set of worlds that, for all the illocutionaryagents mutuallybelieve, might be actual) excludes from such common ground the worlds w' in which no memberof D existing in w' is a boy I saw.20Nothing new or particularly controversialso far (within a possible-worldssemantics). Now, the core meaningof a sentence involving any is just like (45b) plus domainwidening. I believe that domain widening takes place along two dimensions. First, we pick the largest possible quantificationaldomain among the reasonablecandidates.This means that all entities that for all we know might exist are factored in. Second, our uncertaintyabout quantificational domainsmay also have qualitativeaspects. Take Fred again, and considernow his nephew John. We are sure that John exists, but we may be uncertainwhetherhe is a man or still a boy. This means that in some worlds compatiblewith what we know, he is a boy; in others,he isn't. Using any boy, we might signal that our claim extends to him. How do we express this formally?Let us consider sheer domain size first. The only way to measure domain size is by comparison;this entails that the meaning of any must be inherently relational.It must involve comparisonamong D-alternatives.It is useful to visualize this with a toy example. (46) A system of "large" domains { a, b, c } D widest domain Dl I{ a, b } D2 = { b, c } { a, c } D3 Suppose D1-D3 are candidatedomains for what's aroundhere; then any would be associated with their union, D = Dl U D2 U D3. In choosing our quantificational domain in this way, we still have anchoringto a specific D, with the understanding that it is the largest one (among the alternativesat stake). Considernext the inclusionof "marginal"boys. This must amountto a kind of modalization: we take into considerationall those individuals that for all we know might be boys and might be in D.2' Puttingall this together,a sentencelike (47a) (if it was grammatical) would have (47b) as its meaning. This has to be consideredagainst the alternativesin (47c).
Here and throughoutI ignore the (important)differences between a and some. The main reference on the notion of common ground is Stalnaker 1978. The proposal in the text, which uses world-bound domains,can perhapsbe viewed as a way of representing"domain vagueness," which Dayal (1998) argues is characteristicof FCIs. Notice, in fact, the resemblanceto supervaluations(where each alternativecorrespondsto a partialinterpretation). For furtherdiscussion, see section 5. 21 This too can be viewed as generalizingwhat Dayal (1998) proposes for FC any.
19 20

556

GENNARO

CHIERCHIA

(47) a. *1 saw any boy. b. Meaning 3w'3xe Dw' [boyw(x) A saww(I,x)]22 c. Alternatives 3w'3xe D1,w' [boyw(x) A saww(I,x)], where 1 ? i c 3 is approprimustbe used to enrichplainmeaning.But whatkind of enrichment Active alternatives ate for any on the basis of the type of alternativethat by hypothesis it associates with? Given thatD-alternativesdo not form a scale, use of 0 (i.e., exhaustivization)seems inappropriate. Still, to the in choosing among alternatives,speakersdo tend go for strongestone they have evidence for. If this happensalso in the case of (47a), we wind up saying that even the most liberal (i.e., broad) choice of D makes the sentence true: in other words, the base meaning will acquire an Let us spell this implicature even-like flavor (as both Krifka(1995) and Lahiri(1998) propose).23 out: (48) Implicature 3w'3xEDw' [boyw'(x)A saww(I,x)] Cc 3w'3xe Di,w [boyw'(x) A saww(I, x)], where 1 ? i ' 3 and p Cc q = p is stronger (hence, less likely) than q relative to the common groundc However, given the way the domainsarechosen, (48) is logically false: all of the alternatives in (47c) are logically strongerthan the statement(47b); therefore,the latterstatementcannot be less likely thanits alternatives. Sentence(47a) enrichedby implicature (48) is inconsistent,whence its deviance. Contrastthis with what would happenin a negative (DE) context. (49) a. I didn't see anyDboy. b. Statement 3w'3xe D,' [boyw(x) A seew(I,x)] c. Implicature 3w'3xe Dw' [boyw(x) A seew(I,x)] Cc 3w'3xe Di,w [boyw(x) A seew(I, x)]
-

22 Formula(47b) ought to be relativized to an epistemic accessibility relation, which I am omitting for simplicity. Also, from now on, and when no confusion arises, I will omit Aw from formulas.So, for example, the formulain (47b) is to be understoodas a short form for (i).

(i) Xw3w'3xe Dw. [boy,.(x) A saww(I,x)] Actually, my proposalcorrespondsto what Krifkaproposes for what he calls "emphatic" any. For nonemphatic any, he proposes a purely scalar approach,accordingto which, assertinga sentence like (47a) leads to the simultaneous negationof all weakeralternatives(as in scalarreasoning).However, in positive contexts the result is contradictory, since it is impossible for an existential statementto be true in D without also being true in some of its subdomains.On the contrary,in negative contexts a sensible meaning results. Such an approachmakes wrong predictionsfor sentences like (i). (i) *There must be any studentin that building. The presence of a modal makes Krifka'sproposedimplicaturecoherent(somethingI must leave to the readerto verify). Consequently,(i) is predictedto be grammatical,contraryto fact. Krifka'sproposalfor emphaticany does not run into such a problem.
23

BROADEN

YOUR

VIEWS

557

The statement(49b) and the implicatureare consistent.This constitutesa green light to add them to our common ground.This additioninforms us that no matterwhat subset of D might turnout to be the actual domain, I saw nothing in that domain that could possibly be a boy. Domain widening yields its effects. The appealof this generalline of argumentshould be fairly clear. The even-like implicature flows from generalprinciplesof sensible use of alternatives(once one sees what the alternatives underconsiderationare). And it is also immediatelyclear that such an implicaturecannotbe met in positive contexts, which explains the distributionof NPIs. But with this, a potentialproblem comes readilyto mind:implicatures thatclash with the assertiondo not generallyyield ungrammaticality; they are simply removed (exploiting clashes of this sort is, in fact, the way implicatures are typically canceled). So why is a sentence like (47a) (an NPI-licensingviolation) ungrammatiof NPIs cal? Thereis an impasse here between the way domainwidening explainsthe distribution (using Griceanprinciples)and the way such principlesare typically taken to work. Evidently,while scalaralternativescan be deactivatedby the context, D-alternativescannot. Within "recursive" pragmatics,we have a possibly principledway of addressingthis issue. This approachto scalar implicatureshas led us to posit two variantsof scalar terms: a strong [ + u] variant,with active alternativesthat need to be used for enrichment;and a weak [ - a] variant with no active alternatives,for which no a is necessary (or possible). In this setup, it is indeed naturalto expect that there will be items associated with alternativesthat cannot be deactivated: [ + u] items with no weak variant.Any is [ + a] and lacks a weak variant(or, one might say, its weak variantis some[- ,], or some other weak indefinite). The effect of this is that such items will have to occur within the scope of u; their implicaturehas to be frozen in place, throughan abstractoperatoru. From a functionalisticstandpoint,this makes sense. If the role of domain widening is to induce an implicature,using an NPI in a context where such an implicaturecould not arise is self-defeating.Therefore,we can assume thatNPIs carryan (uninterpretable) feature that needs to be checked by an (specifically, a piece of possibly abstractnegative morphology)24 appropriate (interpretable) operator(namely, u). NPIs must be checked by a (i.e., one might say, enter an agreementrelation with u). In a way, the fact that NPIs need u provides independent evidence for it. 4.2 An Implementation Let me spell this out. I assume thatbesides 0 (whose definitionis given again in (50b)), another available mode of enrichmentis E (for even), defined as in (50a). (50) a. Ec (p) b. Oc (p)
=

A Vqe C [p cc q], where C = ALT p A Vqe C [q - p Cc q], where C = ALT

This format shows how close in meaning these two functors are. The choice between them is dictatedby the natureof the alternatives: if (and, ideally, only if) C contains a scale (unique for p), 0 is felicitous; if (and, ideally, only if) C contains partially orderedpropositions,like D24

Call it [ + - D-MAX], to differentiateit from the feature of scalar items.

558

GENNARO

CHIERCHIA

variants,E is felicitous.25In (5la), I specify an "official" lexical entry for any (but cf. section A2 of the appendix) and the alternativesit activates; in (51b), I spell out the specific form of pragmaticstrengtheningassociated with domain widening. (51) a. Lexical entryfor any i. ilanyDII = XPAQAw [3w'3xe Dw(Pw(x)) A QW(x)] = { XPXQXw ii. ALT(tlanyDII) [3w'3xe D','(Pw'(x) A QW(X))] D'c D A D' is large } iii. Any has an uninterpretable feature [ + a]. = Ec (lkkll), b. V1iAls where C = 11VPJALT As with 0, use of E shrinks the set of alternatives.Now, let us go back to the ungrammatical example (47a). In virtue of (5 laiii), it must occur in the scope of a. Here is what we get: (52) a. *I saw any boy. b. s [I saw any boy] c. Ec (]w'3xe Dw, [boyw(x) A seew(I, x)]) d. ]w']xe Dw, [boyw,(x)A seew(I,x)] cc ]w']xlxeDi,, [boyw(x) A seew(I,x)] Any carries a featurethat needs to be checked by u. As cxcan be adjoinedto clausal nodes, we do so in (52b) and the syntacticrequirement on any is duly met. However,o7 locks in the implicature.Thus, the interpretation of (52b) is (52c)-which is an unusablecontradiction (as the implicature it carries, (52b), is necessarily false). No way out. Contrastthis with what happens in a negative context (like (49), repeatedhere). (53) a. I didn't see any boy. b. - [I seeanyboy] c. Ec (-i 3w'3xe D, [boyw,(x)A see,(L, x)]) d. 3w'3xe D, [boyw(x) A seew(I,x)] Cc- w'ixe Di,, [boyw,(x)A see,(I, x)] In a sentence like (53a), we have an additionalsite at which the featureassociatedwith any can be checked, namely, after negation. The semantics we get this time is perfectly sensible, and domain widening comes happily to fruition (in the sense that it has led to something stronger than the availablealternatives).This generalizesto all DE contexts. We now see exactly how the computationalsystem forces NPIs to occur in DE contexts. Two observationsmay be appropriate. First, a can be thought of as what makes negation (and other DE heads) "strong" or "affective" (giving precise semanticcontent to this notion). Second, one might expect the special morphologythat induces checking or agreementwith the

25 Notice that if C contains scalar alternatives,Ec (p) yields a noncontradictory statementonly if p is the strongest member of the scale. Perhapswe might requirethat a form of enrichmentis felicitous only if it can yield meaningful results for all the alternativesat stake.

BROADEN

YOUR

VIEWS

559

implicature-freezing operatoru to be sometimes "visible." Cases of "negative concord" can be viewed in this light. (54) a. Non ho visto nessuno studenteparlarea nessun professore. professor student speak to no (I) not have seen no with 'I haven't seen any studentspeak any professor.' b. u [I saw any studentspeak with any professor] It is tempting (following the insights of Laka 1990 and Ladusaw 1992 embedded in a new framework)to explain negative concord along the following lines. N-words in languages like Italian have roughly the same semantics as (NPI) any. They are, therefore, domain-widening existentials.This forces checking by a, which can yield somethinginterpretable only in conjunction with negation and other negation-likeoperators.That is why negation must be present and can affect more than one N-word (without resulting in multiple negations). Moreover, since in the case of N-words, the NPI actually carriesa piece of overt negative morphology,the locality conditions on checking and the range of heads that can sustain cr and do the job may be more by languagenarrowlydefinedthanthose associatedwith any. This, in fact, seems to be supported internalevidence as well: nessuno, lit. 'no one', has a narrowerdistributionthan any (e.g., it is of ogni 'every');mai 'ever', which has no overtnegativemorphology, not licensed in the restriction instead has a distributionvery similar to that of any. There is obviously a lot of work to be done in this connection;nevertheless,the division of labor between syntax and semantics looks promising. A generalcriticismthat has been leveled againstthe domain-wideningidea is that widening evident with N-words. Just like does not seem to always have to take place. This is particularly its English translation,a sentence like (54a) can be used when the speakerhas a specific salient domain in mind, and it does not necessarily requireexpandingthis domain to include marginal cases. As it turns out, this is in fact consistent with the use of domain widening adopted here. The lexical entry for an NPI (see (51)) contains an implicit referenceto a specific domain,just like the entry for any other quantifier.So nessuno (or any) will be relativized to a specific, pragmaticallyset domain. However, alternativesare activated, and they automaticallygenerate the relevant implicature-which cannot be canceled. This mechanism sometimes reflects real aboutthe quantificational domain.But this doesn't have to always happen:sometimes uncertainty we merely have a formalrequirement. Power of grammaticization, one might say. Domain widening, as implementedhere, is a potential for domain widening. The main thrust of the present attemptis to place the semantics of any within a general multidimensionalsemantics.There are theory of implicatureprojection,in an alternative-based, just a few lexical options that Universal Grammar makes available:the kinds of alternativesan item activates (so far we have scalar alternativesand "large domain" alternatives)and whether or not weak variantsare available. Alternativesdeterminethe form of enrichment(O vs. E); the presence of weak variantsdetermineswhether such alternativescan be inert. The rest is set by the computationalsystem. Scalarand domain alternativesinteractin rich ways. We will see one effect of such interactionsbelow. The full scale of such interactionsis quite broad.For example,

560

GENNARO

CHIERCHIA

in Chierchia2004 it is arguedthatthe interventioneffect on NPIs is due precisely to such interaction. While the present frameworkdoes lend itself to checking the viability of such claims,26 exploringit fully is morethanI can do in this article,wherethe mainfocus remainsthe comparative grammarof PSIs. What I have done, in this connection, is to implementthe domain-widening idea in recursivepragmatics. 5 The Birth of Universal Readings In this section, I will examine FCIs of the any type (which allow negative polarityuses) and of the qualunquetype (which disallow negative polarity uses) and discuss where their properties andquantificational force come from.ThenI will come backto the relationbetweenthese elements and pure NPIs. 5.1 Antiexhaustiveness FC uses of elements like any is why they seem to switch One of the classic puzzles surrounding examples illustrate: force, as the following standard so naturallyto a universalor quasi-universal (55) a. Any cat meows. b. Yesterday,any studentthat was arounddroppedby. Dayal (1998) makes a convincing case that the universalforce of FCIs cannot be derived from a quantificationaladverb as a sort of quantificationalvariabilityeffect (see, e.g., Kadmon and Landman 1993, Giannakidou2001); rather,it must be endogenous to any itself. She proposes thatFCIs be analyzedas modalizeduniversals.I want to arguethatthe effects of Dayal's analysis can be derived as a furtherimplicatureof domain widening, elaboratingon an insight of Kratzer and Shimoyama (2002). In this section, I will extend (a variantof) Kratzerand Shimoyama's proposal for German(existential)FC irgendeinto universalFCIs (namely, Italian [qualsiasi N] and FC any), postponingdiscussion of existentials until the next section. Imagine that the alternativesunder considerationare not domains of approximatelyequal size, but ratherall of the possible choices (on a given maximaldomain).Imagine,in otherwords, that the structureof the alternativedomains is roughly this:

26

A referee points out interestingcases of this sort:

(i) Everyone who knows any motherof two children should tell her about this tax benefit. The interestof this example lies in the fact thatboth 0 and E must be at work andthatthe canonicalimplicature associated with two is absent (the sentence in fact implicates that it is not the case that people who know mothersof a single child should do anything).The LF representation and semantic interpretation of (i) should be as follows: (ii) a [everyone who knows any motherof two children should tell her about this tax benefit] (iii) 0 E (every (one who knows any motherof two children)(shouldtell her about this tax benefit)) The inverse order, E 0 (+), yields a contradiction(for only creates a nonmonotoniccontext, while E works only in a monotone setting). So, when we look at the structureof (i), where any motherc-commandstwo children, it seems that alternativesmust be factored into meaning as in a pushdown stack: the last one in is the first one out. This has rich consequences for interventioncontexts that, I think, corroboratethe general line explored in Chierchia 2004. A full discussion of the relevantissues must be deferredto anotheroccasion, however.

BROADEN

YOUR

VIEWS

561

(56) Dl = {a,b} D4 = { a }

D = {a, b, c } D2 = {b,c} D5 ={ b }

D3 = { a, c } D6 = { c }

Suppose that from among this finely structured range of alternativesyou were to pick one-say, = D3 { a, c } (by saying, for example, thatsomeone in D3 is the culprit).Whatwould thatconvey to your hearer?Clearly, that you are excluding other options; and, in particular,that you are excluding D5 (i.e., the complement of D3). The same holds for any other choice. Conversely, what would the choice of D, the maximaloption,convey? Plausibly,it would convey the opposite, namely, that you do not exclude any option whatsoever. This lays out the intuition.Now let us reconstructit formally. (57) a. (Yesterday)I saw any student(that wanted to see me). b. Assertion 3w'3xe Dw' [student,'(X)A seew(I, x)] Abbreviatedas: someD (student)(XxI saw x) c. Potential alternativeassertion I saw x), for any Di C D someDi(student)(Xx d. Strengthenedalternativeassertions 0 (someD0 (student)(XxI saw x)) = someDi (student)(XxI saw x) A c D-Di Clearly, asserting (57b) will induce the hearerto assume that the speakerhas no evidence that any strengthenedalternativelike (57d) holds. By the epistemic step, the hearer will therefore conclude that (57d) does not hold. (58) 0 (someDi(student)(XxI saw x)), for all Di

someD. (student)(XxI saw x), for any Dj

Now, let us work out (58) by putting the universalquantifiersover domains in the appropriate places. (59) a. VDi VDj [someD0 (student)(XxI saw x) A I saw x)]27 someDj(student)(Xx b. VDi VDj [someD0 (student)(XxI saw x) -> someDj(student)(XxI saw x)] In essence, this says that if I saw a studentis truein some domainD, it must be truein any other domain (containinga possible student).This, togetherwith the assertion,entails (60). (60) VD [someD(student)(XxI saw x)], where D contains possible students

27

someD,(student)(XxI saw x)] I use (59a) to enhance readability.The point in the text is not affected by this simplification.
A

To be precise, (59a) should be spelled out like this: (i) VD, VDjc D -D (student)(XxI saw x) [someD0

562

GENNARO

CHIERCHIA

A quasi-universalreading therebycomes about. The assertionby itself doesn't make it happen, and the implicatureby itself doesn't either. The universalforce stems from putting, as it were, two and two together(the assertionand the implicature).In doing so, we are using nothingmore thanplausibleGriceanprinciplesand domainwidening, on the assumptionthatthe D-alternatives form a "complete" lattice structureof the form in (56). So whatdoes the differencebetweenpureNPIs (like ever or ItalianN-words) andany amount to? We are playing here with two kinds of implicatures.The NPI implicatureis an even-like implicature(as suggested by Krifka(1995) and Lahiri(1998)); the FC implicatureis antiexhaustiveness (as suggestedby Kratzerand Shimoyama(2002)).28The lattercomes out of two factors: insisting on using domain widening in positive contexts (which excludes E as a possible enrichment operation)and activatingalternativesof any size (down to the smallest domains).It is thus plausible to maintainthat if a certainitem lexically activates alternativesof any size (including "small"ones), the form of enrichmentthat gets triggeredis antiexhaustiveness. Here is a possible implementationof all this. A reasonablecandidatefor the lexical entry of FC any might be this: (61) a. anyD = XPXQw'3xe Dw [Pw(x) A QW(x)] b. ALT (anyD) = { XPXQ3w'3xeD'w' [Pw'(x) A Qw(x)]:D' c D A D' n Xx3w'[Pw(x)] 0 0 } We have simply replacedthe conditionthat the domainsbe "large" with the one that alternative domains must standa chance (namely, contain things that might possibly satisfy the restriction). As a result, now even a D containing a single possible student(in the case of (57a)) will be in the alternativeset. And the strengtheningoperationthat naturallygoes with alternativesof this sort is antiexhaustiveness. (62) Antiexhaustiveness = 0-c l4XII, where C ik1I!s and 0-c 1kfItsALT

(p) =p

A Vq,q'e C [q -> q]29

It is plausibleto maintainthat0- can apply felicitously only when the alternativeset of domains forms a complete join semilattice as in the example in (56). Summingup so far,pureNPIs (like ItalianN-words) are associatedwith largeD-alternatives. This triggersan even-like enrichment,E. Use of E confines pure NPIs to DE contexts. FCIs like any are associatedwith alternativesof any size (includingsmall ones), which trigger0-. Everything else stays the same. Both NPIs and FCIs must be checked by the implicature-freezing operator.Here is a sample derivationinvolving FC any: (63) a. I saw any[+?,]student(that wanted to see me). b. u [I saw any[+ ] student]
28 Kratzer and Shimoyama(2002) actuallydiscuss the FC implicature only in the contextof what arecalled existential FCIs. Exploiting it to derive universalreadings is, as far as I am aware, an idea developed here for the first time. 29 A formulationlike this would make the antiexhaustivecharacter of 0 more transparent: (i) VqEC Oc, [q], where C' must be restrictedas in footnote 27 1 owe this idea to suggestions by Danny Fox and Jon Gajewsky.

BROADEN YOUR VIEWS

563

c. someD (student)Xx I saw x A VDi VDj [someDi(student)(XxI saw x) -> someDj(student)(XxI saw x)] d. Vae possible student n D [I saw a] (I continueto ignore,for simplicity,the modifierthatwantedto see me.) Formula(63d) constitutes a semiformalrenderingof the assertionand the implicaturetogether,which I will use from now on for convenience. Next, it is interestingto consider what happensto an FC element like any undernegation. In principle,a sentence like (64a) might have two scope options. The first is illustratedin (64b). (64) a. I didn't see any student(that wanted to see me). b. ox[I saw any student] c. Vae possible student n D [I saw a] Here negation has scope over the implicature-freezing operator.Accordingly, we first lock the implicaturein, then negate. The interpretation of (64b) is (roughly)as in (64c). This corresponds to the "rhetorical" reading (64a): the 'I didn't see just any student' type.30But there is also anotherpossibility, illustratedin (65). We can first negate, then "check" the implicature. (65) a. au [I saw any[ student] b. Statement someD (student)Xx I saw x c. Implicature I saw x) (student)(Xx VDi Dj [-, someD0

-i

someDj(student)(XxI saw x)]

Now notice that(65b) entails (65c). To see this, dropthe universalquantifier from(65c), instantiating it as an arbitrary Di in the alternativesets. (66) (student)(XxI saw x) -someD0
-.

someDj(student)(XxI saw x)

If D is (per our hypothesis) the largest domain, it is clearly impossible for (65b) to be true and (66) false, for (65b) entails both the antecedent and the consequent of (66). Conclusion: the implicatureis automaticallysatisfied in any situationwhere the statementis true. Just as with "pure" NPIs, in negative contexts we are left solely with domain widening;the FC implicature vanishes. The conclusion is simple and, arguably,compelling: a lexical item with an entry like (61) is predictedto have a quasi-universal force in positive contexts and to act like an NPI in negative contexts. Its (similarity to and) difference from pure NPIs is very explicitly laid out: it is a difference in the type of alternativesactivated.This explains why some languagesmight choose different lexical entries to signal association with different alternativesets, while others might opt to have one item covering both domains. It also explains why an item may start as a pure
30 It needs to be explained why the rhetoricalreadinggenerallyrequiresa special intonationalcontour.It would be desirableto derive this effect from the interactionof a principledproposalabout FCIs (such as the presentone arguably is) and the theory of focus. But this will have to wait for anotheroccasion.

564

GENNARO

CHIERCHIA

NPI and then turninto an FCI (by expandingits alternativesets) and vice versa. Finally, we also see that it is incorrect to think of any as "ambiguous" between NPI and FC interpretations: English any has a unitarymeaning, (61), which simultaneouslyaccountsfor its NPI uses (in DE contexts) and its FC uses (in non-DE contexts). 5.2 Subtrigging Dayal (1998) carefully lays out several of the key generalizationsabout FCIs like any. As we have seen, she concludes thatEnglish any is "inherently"modalized,universallyquantified,and domain vague. That insight seems to be basically correct.In fact, it fits with a view of polarity that is perhapsmore generalthan one could hope for. The "inherent"partof her proposalneeds to be qualified. The quantificational force of FC any is not writteninto its lexical entry. It stems from an implicature,triggeredby the domain alternativesactivatedby it. Dayal also proposes an account of subtrigging that, as far as I can make out, is the only one among those currently availablethat standsa chance of being right.This section is devoted to showing how her account extends to the proposalmade here. Consider sentence (67a) and its semantics, accordingto the presentproposal(67b). (67) a. *1 saw any student. b. VD 3w'3xe D,' [student,'(x) A see,(I, x)], where D contains at least a possible student What does (67b) actually say? In essence, that any possible studentis such that I saw her. This is an extremely strong statement-perhaps too strong to ever be true. I can only see actually existing students;I cannot see something that does not exist. Because of our liberal take, D is surely going to include some such nonexisting entities. But this makes (67b) much too strong to ever be true. Thereis a kind of presupposition clash here between the modalizedcharacter of the of the scope. It is as if we have gone too far with restrictionand the episodic/actualisticcharacter our domain widening, to the point of obtaining a restrictionunsuitableto be used in episodic statements. Now, consider an occurrenceof any "subtrigged" by a relative clause, as in (68a). What would the structureof the restrictorbe? While I do not have a full-fledged analysis, something like (68b) looks like a reasonableoutcome. (68) a. b. c. d. I saw any studentthat wanted to see me. VD 3w'3xe D, [student,'(x) A wantw(x,Xw`see, (x, me)) A seew(I,x)] D n \x3w' [studentw (x) A wantw(x,Xw`seew(x, me))] # 0 Xx [studentw (x) A wantw(x,Xw`seew(x, me))]

In (68b), we find three world variables.One is associated with the head noun student and gets bound (i.e., "modalized") by any (as is generallythe case). A second one is associatedwith the embedded infinitival clause [PRO see me]. The third one is associated with the relative clause (presumablythroughthe tense associated with the main verb want in the relative clause). This variableeventually gets associated with the actual world. The exact details of how this happens

BROADEN

YOUR

VIEWS

565

depend on specifics of the semantics of postnominalmodifiers and tense sequencing.However, the outcome of this process will, plausibly, give rise to a restrictionof the form shown in (68c) (obtainedthroughan intermediatestage that will look roughly like (68d)). Such a restrictionwill contain possible students who in fact wanted to see me (hence, they must be actual students). This resultsin a perfectlynaturalstatement,one thatcan be satisfied:every possible studentwho in fact wanted to see me (and hence must be actual) indeed saw me. Thus, subtriggingprovides the anchoringwe need to be able to use (modalized)FC items in episodic contexts. The general fact that FC items can be used in episodic contexts only subjectto specific restrictionstypically providedby a relative clause (but sometimes perhapsalso by informationpresentin the context) seems to receive a reasonableaccount.3 5.3 "Pure" Free Choice Items With this in place, we can now look at an interestingdifference between Italian FC qualsiasil qualunqueand English any. The differencelies primarilyin the behaviorof qualsiasilqualunque under negation. As noted in section 2, an unmodifiedqualunquewhen negated seems to have only the rhetorical'not just anyone' reading.For example: (69) a. (?)Non ho visto qualunquestudente. (I) not have seen whatever student 'I didn't see just any student.' cr [I saw any student] b. c. Vae possible student n D [I saw a] Out of the blue, (69a) is awkward,unless intonationand/orcontext warrantsa 'not just anyone' In presentterms, this means that (69a) admitsonly the LF representation in (69b), interpretation. in (69c). The otheroption, which is availablefor any (see (64a), which resultsin the interpretation repeatedhere as (70)), seems not to be available for qualunque. (70) a. a- [I saw any[ + ] student] b. Statement -- someD (student)Xx I saw x c. Implicature I saw x) VDi Dj H someDi(student)(Xx

I saw x)] someDj(student)(Xx

In commentingon (70), I observed that the implicaturein (70c) is entailed by the assertionand

31 Dayal (1998:460) furtherdiscusses the following interestingpatterninvolving FCIs with partitiverestrictions:

(i) You may pick any of the flowers. (ii) *You must pick any of the flowers. (iii) *Marypicked any of the flowers. FCIs of the any type seem incompatiblewith partitives,unless a possibility modal is present.This is what promptsDayal to introducethe concept of domain vagueness. I cannot address this patternwithin the limits of this article. Relevant discussion can be found in Menendez-Benito2005.

566

GENNARO

CHIERCHIA

hence "disappears."This makes any act like an NPI in negative contexts. In contrast,qualunque seems to be a "pure" FC element, which does not "double up" as an NPI, since it disallows reading (70a). The question then becomes, how are such construalsruled out? Addressing this question will help us pinpoint the exact differencebetween English and ItalianFCIs. A not unreasonable way to rule out (70) for qualunqueis by insisting that the strengthening must be proper; that is, adding the implicaturemust lead to something that is indeed stronger than the plain statementwithout implicature.In other words, the strengthenedstatementmust asymmetricallyentail the plain one. Clearly, this does not happen in (70). In contrast, if the results. implicatureis introducedin a positive context, properstrengthening qualunquestudente. (71) a. Vedro (I) will see whatever student b. Statement someD (student)Xx I will see x c. Strengthenedstatement i. LF a [I will see any student]
ii. Interpretation

someD (student)(XxI will see x) A VDi Dj [someDi(student)(XxI will see x) someDj(student)(XxI will see x)] = VaEpossible student n D [I will see a]
-

is factoredin (i.e., when we compute So (69) is admittedbecause at the time when the implicature oa [I saw any student]), we are in a positive context, where adding the implicatureproperly strengthensthe assertion.Negation comes in subsequently. be propercan be viewed as a presupposition that strengthening Technically,the requirement on the version of the freezing operatorselected by qualunque. undefinedotherwise. (72) licr Al= flu (rIl, if Il 4fl asymmetricallyentails j4+I; Boldface co is just like a with a presuppositiontacked on: cuyields a felicitous statementonly if the result of freezing the implicaturereturns something strictly stronger than the unenriched statement.We stipulatethat "pure" FC elements like qualunqueselect for uf (as opposed to U). As a consequence, the implicatureassociated with qualunquecan only be frozen successfully in positive contexts (the result can then, of course, be embeddedfurtheras in (69a)).32 Evidence for this analysis comes from the puzzling facts observed in (12)-(13), repeated here.

32 So cJ yields a sort of positive polarity effect, a notion that otherwise has no independentformal status on the presentapproach.

BROADEN

YOUR

VIEWS

567

(73) a. Non leggero qualunquelibro. (I) won't read whateverbook b. Non leggero qualunquelibro (I) won't read any book che mi consigliera Gianni. that Gianni will recommendto me

V (rhetorical) V (1 3); V

The factual generalizationis that while the rhetoricalreadingis the only option for unmodified FC qualunque,anotheroption becomes available when such items are modified (options that make such items start to act like NPIs). Now, we just saw how the rhetoricalreading for (a sentence like) (73a) is obtainedand why the NPI readingis absent.However, a furtheroption is expected, since, in principle,it shouldbe possible to scope the embeddedDP out. The correspondwould be (74b). ing LF representation (74) a. [qualunquelibro]i non leggero t, b. or [qualunquelibro]i non leggero ti c. VaEpossible book n D [I will read a] If we lock the implicaturein after having scoped the object out, as in (74b), the presupposition of the c-operator is met (i.e., we obtain something that asymmetricallyentails the unenriched interpretation of (74a)). However, the result constitutesa subtriggingviolation. Consequently,it will be ruledout by whateverrules out sentenceslike *I read any book.This immediatelypredicts that subtriggingwill rescue sentences like (74a), on the intended reading. This is indeed what of (73b) on the V -- readingseems to show. The relevantanalysis is given in the grammaticality (75). (75) a. [qualunquelibro che mi consiglieria Gianni]inon leggero t, b. or [qualunquelibro che mi consigliera Gianni]inon leggero ti c. Vae possible book that Gianni will recommendto me n D [I will read a] it gives the impressionthatqualunqueall of a sudden Since (75b) is equivalentto a 3 structure, takes on the behaviorof an NPI. As a matterof fact, however, this isn't so; and we now see why. An intricatepatternseems to fall into place in a ratherprincipledfashion. It is worth summarizingwhere we stand so far. The system of PSIs can be schematizedas follows: (76) The system of polarity-sensitiveitems (alcuno, mai, ever) u[D-MAX]: pure NPIs (any) u[D-MIN]: NPIs/FCIs (qualsiasi) cr[D-MIN]:pure FCIs What the elements in (76) have in common is that they (a) activatedomain (D-) alternativesand The latteris a device thatpreventsthe implicature (b) select for the implicature-freezing operator. (induced accordingto general Griceanprinciples)from being removed, a mechanismmotivated on the basis of scalar implicatureprojection.Where the items in (76) differ is (a) in the size of

568

GENNARO

CHIERCHIA

the D-alternatives(MIN/MAX) and (b) in the particularimplicature-freezing operatorselected. MAX-alternativesare "large" domains (expressing our agreement on core cases and doubts about marginalcases). Selection of MAX-alternatives triggersan even-like implicature.Such an implicaturecan be sustained only in DE environments(in non-DE environments,it results in include all possible domains,down to the smallest ones, thereby MIN-alternatives contradiction). indicatinga more radicaluncertainty.This results in a differentimplicature,antiexhaustiveness. Such an implicature,addedto the assertion,precipitatesa universalreading.Finally, implicature that the result is properly freezing can come aboutin two ways: with or withoutthe presumption one and strongerthan its input. That is, there are two variantsof v: a strong (presuppositional) a weak (presupposition-free) one. Lexical items freely select (throughagreement)eithervariant.33 If an item takes the "weak" option, it displays "double dealer" behavior:negative polarity in negative contexts, FC in positive contexts. If an item takes the "strong" option, it displays pure FC behavior. So it seems that systematicityperhapsraises its noble head. Several problemsremain,however. In particular, recall thatundercertaintypes of modalities(e.g., imperatives),the "universal" force of any seems to vanish: Push any button! Moreover, there is a whole class of FCIs for which a universalinterpretation is out of the question(Germanirgendein,Italianuno qualunque). What about them? 6 Existential Readings Strike Back In this section, I discuss existentialFCIs. The mainidea to be developed drawseven more directly from Kratzerand Shimoyama2002 than the one discussed in section 5. I will first presenttheir proposal,then discuss how it relates to the one I am making here. 6.1 CombinedEffects of Free Choice and IndefiniteMorphology Existential FCIs differ from universal ones in quantificationalforce. A furthercharacteristic of existentialFCIs, noted in section 1, is thattheirmarginalityin episodic contexts cannotbe rescued by subtrigging. ne ho discusso con un qualunquefilosofo (77) a. ??Ieri yesterday (I) of-it have discussed with a whateverphilosopher (che fosse disposto ad ascoltarmi). (that wanted to listen) b. leri ne ho discusso con qualunquefilosofo yesterday (I) of-it have discussed with whateverphilosopher che fosse disposto ad ascoltarmi. that wanted to listen

33 As readerscan check by themselves, if an item triggers the even-like implicature,the presuppositionof or can never be met; hence, pure NPIs can only select for presuppositionlesscr.

BROADEN

YOUR

VIEWS

569

c.

con un qualunquefilosofo. Avrei dovuto discuteme (I) should have discussed of-it with a whateverphilosopher

Out of the blue, (77a) is marginal,and the relative clause, if anything,makes things worse, in contrastto whathappenswith universalFCIs (see (77b)). An overt modalitycan rescue existential FCIs (as in (77c)). In fact, a way to rescue an existential FCI that is not overtly modalized, like the one in (77a), is to embed it (or imagine it embedded) in a context broadly construableas in the ab"modal." The generalizationthat emerges is that existential FCIs are ungrammatical sence of a modal of some sort, a modal that sometimes can be covertly supplied(perhapsin the form of an abstractassertoricmodality; on this, see also Kratzerand Shimoyama 2002). This of existentialFCIs (as Kratzerand Shimogeneralizationcould be directlybuilt into the grammar yama in fact do). We could simply state that an existential FCI must occur in the scope of a modal. But it would be more interestingif this link to modalitiescould be derived from what we have found out so far about FC in general and some other propertyof existential FCIs. Where should we be looking for such a property? An even superficialglance at the form of existential FCIs reveals that they are composed of the FC morphology (irgend in German,qualunquelqualsiasiin Italian)plus overt indefinite morphology(ein in German,any numeralin Italian).In the best of all possible worlds,the behavior of existential FCIs should follow from the grammarof FCIs (which we have, let us suppose, independentlyestablished) plus the standardcontributionof overt indefinite morphology. The (b) an "exactly" scalar two things:(a) existentialityand, importantly, latter,typically, contributes implicature.34 (78) a. A man walked in. b. Interpretation 3x [man(x) A walked in(x)] c. (Scalar) implicature twoD (man) Xx [x walk in] The existentialsemanticsin (78b) is alreadypartof the semanticsof universalFCIs, so thatcannot be what is specific to existential FCIs. That leaves us with the scalar implicature(78c)-which must thereforebe the culprit.Implausibleas this may appearprimafacie, it seems to follow that existentialFCIs must be characterized by threethings:(a) existentiality,(b) an antiexhaustiveness jointly These threeproperties over and a scalar (uniqueness)implicature. domains, (c) implicature should suffice to explain the special relationbetween existential FCIs and modals, as well as the other differences between existential and universal FCIs. As we will see, this is nearly on the mark.

34 I assume thatthe indefinitearticlehas roughlythe same semanticsas the first numeral one and thereforecompetes with numerals;I write [3nx ... ] for "there are at least n x's ... "

570

GENNARO

CHIERCHIA

To make things concrete, let us consider a hypotheticalexample. (79) a. ??Ho sposato un qualsiasi dottore. (I) have marrieda whatsoeverdoctor b. Basic assertion x) 3w'3ixEe Dw'(doctorw'(x) A I marryw c. Alternatives { 3w'31xeDw(doctorw (x) A I marryw x),... (x) A I marryw x), 3w'32xe Dw(doctorw
A I marryw A I marryw x),... x), 3w'32xeD1w'(doctor,'(x) 3w'31xeD1w'(doctor,'(x)

The basic meaningof an existentialFCI like (79a) is identicalto that of its universalFC counterpart, (79b). The alternatives,however, are different:an existential FCI is also a scalar term, so its alternativeswill contain both scalar (rows) and domain (columns) altematives, as shown in forms of enrichment(so that the (79c). These alternativesmust be used up throughappropriate requirementthat FC morphologybe checked by u can be duly met).35Accordingly, the scalar alternativesmust use 0, and the D-alternativesmust use 0-. The result is shown in (80). (80) Ila ho sposato un dottorequalsiasill = 0- (O (3w'31xe Dw'(doctorw'(x) A I marryw X)))36 Let us unpackthe effects of this complex form of enrichment,proceedingcompositionallyfrom the innermostoperator.As we know, 0 adds exhaustivityover the relevant scale. (81) a. 0- (3w'3lxeD,'(doctorw'(x)
marryw x))
A

I marryw x) A

3w'32xeD,'(doctorw'(x)

b. 0- (3w'3!1xeD,'(doctorw'(x) A I marryw x)) c. Alternatives { 3w'3!1xeDw'(doctorw'(x)A I marrywx), 3w'3! Ixe D'w(doctor,'(x) A I marryw x), A I marryw 3w'3! IxeDlwY(doctorw'(x) x), So the innerformulain (81a) says thatI marryexactly one doctorin the relevantdomain(abbreviated as in (81b)). At this point, the alternativeswill be the ones shown in (81c); that is, they will be all the D-variantsof (81b). Workingout the effects of 0- at this point gives (82a-b). (82) a. VD' VID [3w'3!1xeD1w(doctorw'(x)A I marryw x) A I marryw 3w'3! Ixe1YDw(doctorw'(x) x)] b. VD [3w'3!1xEEDw (x) A I marryw x)] (doctorw

35 See section A3 of the appendix for formal details. Also, as mentioned in footnote 26, the proper handling of complex alternativesets requiresmore work, especially in connection with intervention-type effects. What is provided in the text is thereforeparticularlytentativeinsofar as the handlingof multiple alternativesis concerned. 36 I believe that reversingthe scope of 0 and 0O leaves things unchanged.I will leave this for the readerto check.

BROADEN

YOUR

VIEWS

571

We see the usual effect of the FC implicature:formula(82b) must be true of every domain that contains a possible doctor.Now, if our alternativedomainscontainmore than one doctor (which they surely will, for otherwise there would not be D-altematives),then (82b) is inconsistent,for it says that the sentence I marryexactly one doctor must be true of every doctor. This seems to provide us with an account of why existential FCIs in plain episodic contexts are marginal(and cannot be rescued by subtrigging):the FC implicatureclashes with the scalar one. But now let's contrastthis with what happensin a modal context. We startby embedding (79a) underan (overt) modal and computingits interpretation. (83) a. Posso sposare un qualsiasi dottore. (I) can marry a whatsoeverdoctor b. Basic meaning A I marryw 3w R(wo, w) [3w'31Ixe D'w'(doctorw'(x) x)] 'There is an accessible world w, in which I marrya doctor.' The presence of the modal gives us a furthersite at which to compute the FC implicature.37 (84) a. 0- (3w R(wo, w) [O (3w'31xeD'w'(doctorw'(x)A I marryw x))]) b. 3w R(wo, w) [3w'3! lxe Dw'(doctorw'(x) A I marryw x)] A VD' [3w R(wo, w) [3w'3!Ixe D'w'(doctorw'(x) A I marryw x)]] It is not hardto see that (84b) is consistent.First, the assertionsays that there is some accessible world w in which somethingin D is a doctorI marry(and there are no two such things). Second, antiexhaustivenesssays that for every subdomainD' of D containinga doctor, there is a world in which I marrythat person. We obtain,in other words, a distributionof doctors across worlds: any possible doctor constitutesan option for me to marry.Here is the picture: (85) Worlds wI w2 wn Doctors I marry dl d2 dn

That is, the doctors must distributeover the worlds in such a way that in each world I marrya differentone, so that each possible doctorwinds up being the chosen one in some world or other (and,of course, uniquenesspreventstherebeingjust one worldin which I marryall of the doctors). This is a neat result. Without any stipulationwhatsoever,the interactionof modalities and the FC implicature (antiexhaustiveness) yields the rightkindof meaning.For(83a) plus its implica-

37 I am simplifying things considerably.For example, I am ignoringthe outermostscalaritem (i.e., the modalposso '(I) can' itself). Furthermore, there are other a prioriconceivable ordersof applicationof the relevantoperators-though I do not think they affect the main point. At any rate, as noted earlier, multiple alternativesets require principles of computationthat cannot be properlyexplored within the limits of this article.

572

GENNARO

CHIERCHIA

tures says that I must marryone doctor, and any conceivable doctor is a possible option. This is, in essence, Kratzerand Shimoyama'sinsight. What my proposaladds is that we do not have to worryaboutstipulatingthatexistentialFCIs must occur in the scope of a modal. For if it contains no modal, a sentence with an existential FCI is unusable.38 6.2 A Novel InterventionEffect We have concludedthatin nonmodalcontexts, existentialFCIs give rise to contradictory implicatures, rescuable by the insertion of a modal. This derives their meaning and distributionin a seemingly principled manner. However, potentially there is another way to rescue existential FCIs, namely, by inserting a quantifiedDP between the implicature-freezing operatorand the FCI. (86) a. (??)Un linguista ha sposato un qualunquedottore. a linguist has marrieda whatever doctor b. a [un linguistai [un qualunquedottorej[ti ha sposato tj]]] c. VD' [3y linguistw(y)A 3w'3! Ixe D'w'(doctorw'(x) A y marryw x)] d. For every doctor a, some linguist marriesa and only a Out of the blue, (86a) is odd (unless some "modalizing" context is provided).But it is not hard to see that(86c) is not contradictory in (86d) illustrates).So, if nothing (as the informalparaphrase is added, we would be predictingthat sentences like (86a) are grammatical, which is not correct. can a Only modals provide suitable environmentfor existential FCIs. We had hoped-indeed, claimed-that it was not necessary to build this in as a stipulationspecific to the grammarof FCIs, once the contributionof variousimplicaturesis properlydissected.39 Given facts like (86), though, it looks like we do have to stipulatesomething.We must stipulatethat no other DP can intervenebetween a and the DP thata associates with (i.e., the DP whose alternativesu operates on). A modal in the same position is fine. However, this type of restrictionhas a ring familiar from much work in syntax: it sounds like a minimalityeffect. If this is so, our hope not to have to say anythingconstructionspecific for existential FCIs would seem not to be misplaced after all. It is worth recalling some of the basic traits of the presentproposal.We are assuming that FC morphologyis semanticallyassociatedwith (minimal)D-alternatives,which must be factored into the meaning.We areimplementingthis by assumingthatthe FC morphemescarryan uninterpretable feature [ + oaD-MIN] that must be checked by the (interpretable) u-operator.This has the effect of locking the FC implicaturein place. Accordingly,the abstractstructure of a sentence like (86a) will be as in (87a), while the structureof a sentence with a modal will be as in (87b). (87) a. * [DP[-, D-MIN] [ ... DP[+u b. (r [can [ ... DP[+? D-MIN] ...
38

D-MIN] ***]]
11

This extends to universalmodals. See section A4 of the appendixfor a worked-outexample.

39 This point was made to me by Danny Fox, Jon Gajewsky, and Philippe Schlenker.

BROADEN

YOUR

VIEWS

573

Even thougha constraintlike (87a) is a stipulation,it indeed has a form familiarfrom much work andits associatedDP is disturbed of another, on locality. The relationbetweenou by the intervention DP. in not Modals the same do intervene. somehow unhomogeneous configuration Perhapsthis is for good reasons.For one thing, such modals cannotcarrythe feature[ + o D-MIN], since they are not associatedwith D-alternatives.Hence, they lack the relevantpropertythatmay interfere.40 Moreover, modals are heads, and here we are clearly dealing with XP (maximal projections) intervention.As Rizzi (1990, 2004) argues extensively, heads can interveneonly with respect to otherheads and maximalprojectionswith respectto othermaximalprojections.(See Rizzi's work for evidence and a detailed implementation.) that multiple FCIs can of course occur in the same sentence. Notice, furthermore, (88) a. Un qualsiasi cittadinopuo sollevare una qualsiasi questione. a whateverquestion a whatevercitizen can raise b. of [can [a citizen[+, D-MINJ [raise a questions[+ c D-MINI ]]] The LF representation of (88a) is somethinglike (88b);the modalpuo 'can' licenses the occurrence of both FCIs. And neither the modal itself nor the first FCI intervenesbetween the a-operator and the second FCI, as we would expect. Summing up, if the present approachis on the right track,it will provide (a) a reason why in plain nonmodalcontexts existentialFCIs are marginal(an implicatureclash), (b) a reasonwhy modals remove the interpretiveobstacle (distributionover worlds), and (c) a reason why DPs that could in principlealso remove the interpretiveobstacle fail to do so (intervention).Even if this approachturns out to be wrong or does not tell the whole story, it is still likely that the observationsin (86)-(88) point, at the very least at a descriptivelevel, in the directionof some sort of minimalityeffect. 6.3 Further Consequencesand Remarks The idea of a sort of "distributionacross worlds" is presentin differentforms in previous work on FCIs. One finds it, for example, in Dayal 1998, Giannakidou2001, and Seb0 2001 (with disagreementson the natureof the modalityinvolved). The first attemptto "deduce" this effect from Gricean principles was made by Kratzerand Shimoyama (2002). They, however, do not discuss the relation between existential and universal FCIs, nor do they derive the differences among them from the presenceversus absence of a scalarimplicature. Thereare otherdifferences as well between the present proposal and theirs. Kratzerand Shimoyama adopt an alternative semantics. Here, I stay within the boundariesof a multidimensionalsemantics (along the lines of Rooth's (1985, 1992) approachto focus or Krifka's (1995) proposal on NPIs, for example). Also, even though I would like to stay as neutral as I possibly can on details of implicature projection,the present proposal requiressomething like implicaturefreezing and, to the extent

40 Modals are of course associated with scalar alternatives;hence, they carry the feature [ + however, does not cause problems.

scal]. This feature,

574

GENNARO

CHIERCHIA

to which it is successful, providesevidence for it. The implicature-freezing bearsa family operator resemblanceto Rooth's (1992) -operator for focus; and it also resemblesFox's "abstract"only. As we saw, however, it has somewhat differentpropertiesfrom either of them. One consequenceof the presentapproachis that when an existentialFCI is not in the scope of an overt modal, if the resultingsentenceis somehow acceptable,the presenceof a covert modal operatorhas to be assumed;otherwise, the implicaturesassociated with the indefinite would be inconsistent.A sentence like (89a), then, must have an LF representation like the one in (89b). (89) a. Gianni e uscito di corsa e non sapendo che fare, Gianni ran out and not knowing what to do ha bussato ad una porta qualsiasi. knocked at a door whatsoever b. Elispeaker (u [Gianniknocked at a door] The abstractassertoricmodal in (89) could be interpreted as somethinglike 'it follows from what the speaker knows that Gianni knocked at a door'; the FC implicaturewould then be "it is consistentwith what the speakerknows thatany door might have been the one knockedat." This is a first approximation(more work needs to be done on the exact nature of the modalities involved), but it looks like a reasonablemove. Notice also that universalFCIs are not subjectto a similar requirement. They can be rescued by subtrigging(which does not work for existential FCIs). Evidently, a rescue strategythat employs overt lexical material(subtrigging)is preferred to one that employs null modals as in (89). Null modals must be a last resort. The presenttheory has a furtherconsequenceor, one might say, makes a furtherprediction. FCIs must be in the scope of the implicature-freezing operator,which, as we saw, comes in two variants:strong (presuppositional) and weak (nonpresuppositional). The presuppositionsof the can only be met in a positive context. The presuppositionlessversion can, strong (u-operator instead, function in both negative and positive contexts. We should thereforeexpect a difference between existentialFCIs parallelto the one found for universalFCls (betweenany andqualsiasi). This expectationseems correct.Italianand GermanexistentialFCIs seem to differ precisely along these lines (suggestingthatwe are probablydealing with a generalizedparametric variation between Romance and Germanic).Compare(90a) and (90b). (90) a. Niemand musste irgendjemand einladen. no one had to a person whateverinvite 'No one had to invite anybody.' b. Nessuno e costretto ad invitare una persona qualsiasi. no one had to invite a person whatever 'No one had to invite just anybody.' NPI reading/rhetorical

rhetorical

Kratzerand Shimoyamapoint out thatthe preferred interpretation of sentenceslike (90a), particularly if pronouncedwithoutspecial intonation,is a pureNPI-likereading.A second interpretation, the rhetorical 'not just anyone' reading, is also possible-for example, in the presence of a contrastiveintonationof some sort. The Italian counterpart of (90a), shown in (90b), has only

BROADEN

YOUR

VIEWS

575

the 'not just anyone' reading (consequently, (90b) requires contrastiveintonation or a special context of some sort). It does not have the NPI reading. This follows under present assumptions.At LF, the available options for German are as follows: (91) German a. LF 1 no one Xx (x MUST someD(person) Xy invite (x, y) b. Interpretation MUST [someD(person) Xy invite (x, y) A VD K someD (person) Xy invite (x, y)]41 c. LF2 a [no one Xx MUST someD (person) Xy invite (x, y)] d. Interpretation MUST [someD(person) Xy invite (x, y)] In German,implicaturefreezing can take place at two levels. The first is before the negative the negative operatorc-commandsu; thus, u applies operatorcomes in (i.e., in the final structure, to a positive assertion);the second is after negation (i.e., in the final structure,o c-commands is negation and thereforeapplies to a negative assertion).The first schematic LF representation given in (9 la); here, we first lock in the implicatureand then negate the result.The interpretation is roughly 'it is not the case that x must invite somebody and that anybody is an option', a The second possibility is given in (91c); in reasonablecandidatefor the rhetoricalinterpretation. this case, the FC implicatureis entailed by the assertionand thereforedisappears.An NPI-like behaviorresults. In contrastwith this, Italianselects for a. This choice is incompatiblewith the LF representation (91c), because it requiresthat the implicaturelead to properstrengthening,which can only happen if freezing applies to something positive (as in (91b)). Thus, Italian has only the LF representation correspondingto (9la) and, undernegation, allows only the rhetoricalreading. We are now also in a position to understand why even the most universalof the FCIs, like English any or Italianqualsiasi, suddenly acquiresan existentialreadingwhen embeddedunder certainmodals (a readingthat in fact sometimes emerges as the preferredone). (92) a. Taste any doughnut. b. Assaggia qualsiasi doughnut. The LF representation of (92) will clearly containthe modal operatorassociatedwith the imperative, whateverthat may be. This opens up the possibility of freezing the FC implicatureeither within the scope of the imperativeor at the top level (with scope over the imperative).Schematically:

41

For simplicity, I replace niemand 'no one' with plain negation.

576

GENNARO

CHIERCHIA

(93) a. b. c. d.

LIr you taste any doughnut

LIV you taste any doughnut crElyou taste any doughnut you taste any doughnutA V 0 you taste any doughnut EIIB

In (93a), we first freeze the implicature,obtaininga universalreading.Then the imperativecomes in. The result might be paraphrased as 'You must taste every possible doughnut',a possible (if disfavored)readingfor (92a-b). In (93c), firstthe imperativecomes in, thenwe freeze the implicature.The resultis fully equivalentto what we usually get with existentialFCIs (minus the uniqueness implicature).So the paraphrase is 'It is necessary that you taste a doughnutand for any particulardoughnut,it is possible for you to taste it'.42 On the whole, the patternof existential versus universal readings of FCIs is ratherintricate-yet it seems to be beginning to yield. 7 Concluding Remarks The PSIs we have discussed can be pulled togetheras follows: (94) The system of polarity-sensitiveitems r[MAX]: pure NPIs u[MIN]: NPIs/FCIs (universal) c[MIN]: pure FCIs (universal) u[MIN, SCAL]: NPIs/FCIs (existential) cr[MIN,SCAL]: pure FCIs (existential) (mai, ever) (any) (qualsiasi) (irgendein) (uno/due/tre/... NP qualsiasi)

Let us go throughthis chartand therebysummarizethe main points discussed above. What PSIs of the type studiedhere have in common is thatthey all involve domainwidening.Since widening is somethingwe only see by comparison,the form wideningmust take is the activationof a series of alternatives, out of which the largestgets selected.I implementthis in a bidimensionalsemantics in which next to the basic value, we compute a range of alternatives.Such alternativestrigger implicatures,accordingto general principles.The general point was made by Grice long ago: a conversational move is judged againsta background of othera prioriconceivablemoves. Selecting
42 A referee points out that the opposite patternnever arises. For example, the following Germanexample can only have the meaning in the gloss:

(i) Er konnte irgendwas tun. he could whatever-something do 'He could do something or other.' Sentence (i) cannot mean that he could do anything.On the present approach,this fact is derived as follows. Sentence (i) has two possible LF representations, namely, (ii) and (iii). (ii) [could uf [he do whatever-something]] (iii) aT[could [he do whatever-something]] In (ii), arapplies to a nonmodalizedsentence and the result is contradictory(see the discussion of (79) in the text). In (iii), a applies to a modalized sentence, which gives rise to the correct(existential)reading.No otheroption is available.

BROADEN

YOUR

VIEWS

577

can enrichcommunicaone move over anothercan be very telling. Riding on this, speaker-hearers tion in highly efficient ways, an opportunityexploited constantly and systematically.However, this does not take place just when an utteranceis completed, as one might think. It happens of meaning;implicatures can be factoredin a recursive,compositional the computation throughout manner. The elements in parenthesesin (94) arejust a mnemonicfor the alternativesassociatedwith the relevant entry. It is part and parcel of this general picture that implicaturesare determined by the natureof the alternatives.Much work still needs to be done in this domain in order to arriveat generalprinciples(which are not disguised "just so stories") on how implicaturescome about. Here is, however, the overall picture at this point. If the alternativesform a scale (i.e., a linearly orderedset), then choosing an element will naturallyindicate that all alternativesthat are not entailed are deemed not to hold. This closely resembles the behavior of only, and to capturethis fact we posit (following Fox 2003) a null operator0; I have arguedthat DE contexts requirespecial care in handling0 (essentially,0 has to be built into each step of functionalapplicationinvolving DE functors). If the alternativesdo not form a perfectscale, we seem to have at least two plausibleoptions. If we are considering possible domains of similar size, then we ought to choose the one that enables us to make the strongest (and hence least likely) statement;in this case, an even-like implicaturenaturallycomes about. I have formalizedthis via E. If, on the other hand, we are excluding no alternativeof any size, down to the smallest we thereforeought to choose the assertion possibility, then it sounds like we are really uncertain; thatexcludes the least, the one thatenables us to rule out fewer possibilities.Fromthis, the hearer will jump to the conclusion that we are trying to rule in most possibilities (hence, the existential statementbeing made is likely to hold of every alternative).This is formalizedvia 0-. The operatorsE, 0, and 0- are not syntacticallyprojected;they are only partof the semantic computation.However, at LF theremust be an implicature-freezing operatora, syntacticallyreal at least to the same extent as focus operators.This operator(which assigns to a sentence the strongest implicaturethat can be factored in without contradiction)is necessary to obtain the variousreadingsthat scalarscan give rise to. It is also crucialfor PSIs. It gives us a syntactically plausible way to state the requirement that implicaturestriggeredby PSIs cannot be removed. The system those proposals give rise to, though in many ways preliminary,unites formal explicitness with conceptual simplicity. Most of the similarities and differences among a fairly extended (and perhapstypologically significant)range of PSIs seem to fall into place. Appendix: The Formal Theory In this appendix,I will sketch a formallyexplicit characterization of the notion "(pragmatically) enriched meaning," building on Chierchia 2004. This characterization does not deal with all aspects of pragmaticenrichment.It takes the form of a recursive definition that to each wellformed LF representation ot associate its enrichedinterpretations The definition of Iliofls is klolIls.

578

GENNARO

CHIERCHIA

formulatedin terms of the standard definitionof (unenriched) which I take here for meaningIltlIl, granted.I assume that 1111 provides a mappingfrom LF into a partialversion of TY2 (i.e., a typed language with variablesover worlds and a semantics with truthvalue gaps; Gamut 1991). Since the numberof enrichedmeaningsis, in the general case, greaterthan one, lliks defines a set; that is to be thoughtof as a relation,ratherthanas a function.The notion of enrichedinterpretais, 11Ils the set of alternativesfor ox.In the general case, the set of tion I1otI[s exploits, in additionto l1o11, alternativesis defined for each expression cx,relative to one of its interpretations p (intuitively, the strongestone computedup to thatpoint);so we will be defining 's alternativesvia a function ( _, p)ALT, where p is an appropriate description (using, say, a logical form-that is, a TY2 representation)of &t'smeaning. The functions (cx,p)ALT and ltlls are defined by simultaneous recursion.Since typically p is the strongestenriched meaning of oc,I will abbreviate((x, p)ALT as ||4I5ALT. In section Al, I outline the main backgroundassumptions.In section A2, I discuss the key lexical entries.In section A3, the main section of this appendix,I providethe recursivedefinition. Each clause in the definition is matched by one or more examples that illustrateits workings. Finally, in section A4 I discuss the FC effect in the context of modals of necessity. Al Basics are representedby formulasof TY2. We assume that every predicateof TY2 that Interpretations representsa naturallanguage predicatecarries a world variable.Translationsare set up in such a way that the world variableof the main predicatein a clause is the one abstractedover under embedding (while the world variable associated with argumentscan be independentlyset; see Percus 2000). An example is providedin (95a).
A saww(I, x)) (95) a. I saw some student-- Xw3xe Dw(studentw(x) A saw,(I, x))Ilg b. Xg . jjXw3xeDw(studentw(x)

for functionsover assignmentsto variables Strictlyspeaking,formulassuch as (95a) are shorthand domains are the aspect (ultimately,contexts) of the form given in (95b). Since quantificational of context most directly relevantto our concerns, I will generallyrefer to (95) as functionsfrom domains into propositions.Once the properassignments are plugged in, formulas like (95) are used to incrementcommon grounds,understoodas sets of worlds (Stalnaker1978). Two formulas are D(omain)-variantsiff they are alphabeticvariantswith respect to some domain variable.Here is a semantic characterization of this notion: (96) D-variance a. q is a D-variantof p (in symbols, D-variant (p, q)), iff there are some i, j such that for every assignmentg, and every domain D, p(g[i/D]) = q(g[/D]). b. For any p, we designate the set of its D-variantsas D-variant (p). In the representationlanguage, we want to define both unrestrictedand restrictedquantification/abstraction. Let U be the domainof individualsand let the set of worlds W be a subset

BROADEN

YOUR

VIEWS

579

of U. Furthermore, let D be an arbitrary subset of U. For any world w, Dw is that subset of D containingall membersof U existing in w. (97) a. Unrestrictedquantification i. 113x4jw g = 1 if for some uE U,
g ii. 113x41w
=

jj4j1w,g[u/x]

0 if for all uE U,

jjfrjw,g[u/x]

1 = 0; undefined,otherwise
-

b. Restrictedquantification 1 i. 113xe DW+IJg = 1 if for some u such that ue IlDwllg, jj41g[u/x] ii. 113xeDw4Ijwg 0 if IlDwIlg # 0 and for all ue IlDwllg, IjjIw 9ru/x] 0; undefined, otherwise c. RestrictedA-abstraction then h(u) Dw.- 0g = h, where for every ue U, if IIDWIIg $ 0 and us IIDWIlg, IIXxe - 11(p11[x/u1; otherwise, h(u) is undefined
-

If 4~w is a formula whose "main" world variable is w, and R is an accessibility relation, then we express modalities as follows: (98) Modalities a. Vw' R(w, w') -* b. 3w' R(w, w') A

4'w' (abbreviatedas 11w4w )

'W

x w) (abbreviatedas Kw

Note that for Dw 4w to be true, + has to be undefinedor true in every world accessible to w; while for Kwfw to be true, + has to be true in some world accessible to w. We now turnto a characterization of the lexical entriesto be used in the recursivedefinition of strong meaning of a, for any expression ox. A2 Lexicon Here, we will consider two types of lexical entries that activate alternatives:scalar terms and polarity items. Let us start with scalar terms. For each lexical entry, we characterizeits basic and its alternativesALT(ot)by simply listing them. meaning IlkxIl A2.1 Scalar Items The lexical entries for scalar items are characterized as follows: some (Pw, Qw) = XPXQXw3x[PW(x)A Q,(x)] (99) jjsome[+,]Ij= XPXQXw ALT (some[+,]) = ALT (every[+,]) = ... { XPXQXw some (Pw, Qw), . . ., XPXQXw every (Pw, Qw) } { some', . . ., every' } (in abbreviatedform)
-

Let us turnto numerals.Following lonin and Matushansky(2005), I assume that the basic type of numeralsis ((e, t), (e, t)). Numerals also have a generalizedquantifiervariant,obtained from the basic type via existential closure. Here is the much simplified characterization of the basic version of numeralswe will adopt for presentpurposes:

580

GENNARO

CHIERCHIA

(100) Ilone[+,?II = XPXx\Xw[1(x) A P,(x)] ALT (one[,,,) = ALT (two[+,])

= ..P.

{ XPxXw[1(x) A PW(X)],
XPXxXw [2(x) A PW(X)],...

Here is the generalizedquantifierversion of numerals: (101) ilone?+,,Il= XPXQXw3x[1(x)

ALT (one[+,])
=

A PW(x) A Q,(x)]

{ XPXQXw3x[1(x) A PW(x)A QW(X)],

A QW(x)],... } XPXQXw3x [2(x) A PW(x)

ALT (two[+,,)

I assume the indefinite article a has the same meaning as one. A2.2 Polarity-SensitiveItems Any can be treatedin a manneranalogousto one, except that it does not impose any cardinality on its argument.It binds the world variableof its argument,and it activates "large" requirement subdomainalternatives. (102) a. Ilany +,III = XPXx3w' [xeDw' A Pw'(X)] A PW'(X)]: D' c D and D' is large } b. ALT (any[+,]) = { XPXx3w' [xeDDw The entry in (102) has a generalizedquantifiervariant,obtainedvia 3-closure.
= XPXQXw3w'3xxeD,' [Pw'(x)A QW(x)] (103) a. jIany[+f,jI b. ALT (any[+,]) = { XPXQXw3w'3xxeD'w' [Pw'(x)A QW(X)] D' c D and D' is large }

Keep in mind thatpure negative polarityany is a fiction. In fact, English any has the FC implicais actuallymore similarto that of Italianqualunquelqualsiasi, ture, so its interpretation sketched next. The FCI qualsiasi is an NP modifier.It binds the world variableof its argument; at the same time, it activates domain alternatives.Here is an example:
(104) a. Ilqualsiasi[ +,]I
= XPXx3w' [xe Dw A PW'(X)] b. ALT (qualsiasi[+,1) = {XPx3w' [xeD'W' A Pw'(x)]: D' n Ax3w [P,(x)] # 0 }

c D

For any lexical entry differentfrom the above, we assume that the set of its lexical alternatives is empty. (105) For any lexical entry oadifferentfrom the above, ALT(ot) = 0. A3 Simultaneous Recursive Characterization of jafllsand
lj(tfsALT

To provide a simultaneousrecursivecharacterization of Iloxtls and IkOxLIsALT, we first generalizethe definition of applicationto sets.

BROADEN

YOUR

VIEWS

581

(106) Generalizedapplication to the functions of a type appropriate If B is a set of functionsand A a set of arguments in B, then B(A)
=

{ (ox): 3eB,

ae

A}

Next, we introducethe following notationalconvention: (107) If p is a propositionand A a set of alternativesto p, then SCAL (p, A) = p's scale in A (if defined). Finally, we will use the following enrichmentoperations: (108) Enrichmentoperations a. Oc (P) = p A Vq [C(q)A q -> p c q] b. Ec (p) = p A Vq [C(q) -> p c q] c. ?-c (p) = p A Vq, q'e C [q -- q'], where the domain of q' is included in the complementof the domain of q

Opiis used to range over any of the above operations.These operationsare undefined if the context does not supply a suitablevalue for the variableC. Typically, such a value will take the form of a recursively specified set of alternatives. , are sets of objects of the same type. To Enrichedmeanings, Ilolis, and alternatives,IlxIlsALT enhance readability,I will use the following notationalconvention:
Enrichedmeanings llalls ={ 1... 2... } Alternatives ALT [[A, B, ... otilS A', B', . . . Each rule below is followed by a few examples that illustrateits workings.The examples consist of interpretations of LF structures. The LF structuresare simplified and use English words even arethe smallest sets of semantic and IlkxllsALT when the samplesentenceis Italian.Throughout, Ilo4ls type that satisfy the conditions that follow. values of the appropriate Here is the base of the recursion. (109) If ot is a lexical entry, then

Bs. kIlils= { kloll }

BALT-

BALT.kIllLT

||alS ALT

=
-

ALT if $ 0 (a), otherwise Ilals,

Example 1 lsposarells= { marry'} = [[ marry']] IsposarellsALT

} { marry'

582

GENNARO

CHIERCHIA

Example 2
ilqualsiasi[+, ]|ls =
l|qualsiasilJsALT

XPXx3w' [xe Dw' A P,'(x)]

A

[[ XPXx3w' [XeD W'A Pw (x)]: D' c D

D' n Ax3w [P,(x)] =#0]]

For an extension of this example, see example 1 under (114). Example 3 no' (Pw)(Qw)} = { no' } = { XPXQXw Jfno[+f]Jls InolISALT [[ not all', few', no' ]] The next step in the recursionis functionalapplication.However, I thinkthe system will be easier to grasp if we jump first to the clause-level enrichmentrules (pretendingthat functional applicationhas been properlydefined) and then come back to functionalapplication.(However, is defined,if they prefer.) readersarewelcome to look aheadat (114), wherefunctionalapplication (10) Scalar enrichment(SC) is of type t, then a. If JIa11
D { Oc (IlCX|): C = SCAL (IlotII, SCs. IkoCX|s IlolAls) }

({ oC (():
b. SCALT. (OC (1011))ALT
=

(E IIctiSALT, e is a D-variantof

Oc (Ilot) 1 J

and C = SCAL (t, { }, otherwise ItIkXI

}, if ilkxlIsALT)

not a singleton

Example 1 Some studentsmokes. Step 1 I1sometf,]studentsmokeslisD { some' (student')(smoke')} Note: This is derived by the functionalapplicationrule, yet to be specified. . . ., every' (student')(smoke')]] = [[some' (student')(smoke'), Ilsome[+, studentsmokeslIsALT Step 2

}= lsome[+,?] studentsmokeslls D { Oc (some' (student')(smoke')) { 1. some' (student')(smoke'), } by SCs 2. Oc (some' (student')(smoke')) { 1. some' (student')(smoke'), } by def of Oc 2. some' (student')(smoke') A many' (student')(smoke')
Step 3 (Oc Ilsome' (student')(smoke')lI)ALT = } by SCALT [[ some' (student')(smoke') ]] = { some' (student')(smoke') Note: The firstclause of SCALTwould yield { Oc (some' (student')(smoke')) }, which is a singleton; therefore,we revert to the plain meaning.

BROADEN

YOUR

VIEWS

583

Example 2 Few studentswith two papershave to read anythingfurther. Step 1 (derived by functional application) Ilfew[+,f]studentswith two papers[,,?] have to read anythinglls { 1. few (studentswith two papers)(haveto read someDthing), 2. few (studentswith two papers)(haveto read someDthing)A some (studentwith one paper)(hasto read someDthing), 3. Ec (few (studentswith two papers)(haveto read someDthing))} (Ilfew[ ? ] studentswith two papers[,?]Ijs(jlhave to read anythinglls))ALT [1not all' (studentswith two papers)(haveto read something) few' (studentswith two papers)(haveto read something) no' (studentswith two papers)(haveto read something) ]]
-

Step 2 have to read anythinglls +, studentswith two papers[ ? 1lfew[ { 1. few (studentswith two papers)(haveto read someDthing), 2. few (studentswith two papers)(haveto read someDthing)A some (studentwith one paper)(hasto read someDthing), 3. Ec (few (studentswith two papers)(haveto read someDthing)), 4. Oc (few (studentswith two papers)(haveto read someDthing))} by SCs { 1. few (studentswith two papers)(haveto read someDthing), 2. few (studentswith two papers)(haveto read someDthing)A some (studentwith one paper)(hasto read someDthing), 3. Ec (few (studentswith two papers)(haveto read someDthing)), 4. few (studentswith two papers)(haveto read someDthing)A
-

some (student with two papers)(has to read someDthing)

} reductions

Step 3 (Oc (Ilfewstudentswith two papershave to read someDthingl1))ALT } by SCALT I { lfew studentswith two papershave to read someDthingll I now turnto the enrichmentoperationtriggeredby maximaldomains (Max domainenrichment). (111) Max domain enrichment(Max) a. If Iloxl is of type t, then
Maxs. Ilotils 2 { EC (loxjl):C =
I
IItIIsALT

} n D-variant(11ox11)
and t has maximal domain

Ec (t): te

||CX||sALT

b. MaxALT. 10IISALT Example John didn't see anything.

and C = llxllsALT nl D-variant(g) }, if not a singleton { hloll }, otherwise

584

GENNARO

CHIERCHIA

Step I (fromprevious computations) (((John saw anything[+,,]ls) = { someD (thing)(Xxsaw(j, x)) } llnotlls (((John saw anything[,,] ls))ALT (linotlls [[ someD (thing)(Xxsaw(j, x)) someD'(thing)(Xxsaw(j, x)) someD" (thing)(Xxsaw(j, x))
*
.

Step 2 (notJohn saw anything[+? ]lS = { 1. someD(thing)(Xxsaw(j, x)) 2. Ec (-1 someD (thing)(Xxsaw(j, x))) } by Maxs { 1. someD (thing)(Xxsaw(j, x)) 2. someD (thing)(Xxsaw(j, x)) A Vpe C [-, someD (thing)(Xxsaw(j, x)) 5 p] } by def. of Ec
-

Step 3 (Ec ((lnotJohn saw anything[ ?111))AIT = { someD (thing)(Xxsaw(j, x)) } by MaXALT Note: Strictly speaking,Maxs gets to be applied only concomitantlywith FAs. See below. Here is the enrichmentoperationtriggeredby minimal domains (Min domain enrichment): (1 12) Min domain enrichment(Min) is of type t, then a. If Ilokll :C D { 0-c (Il(aI) Mins. Iloklls b. MinALT. (lt(ls=ALT

} D-variant((x((ll) { -C (t) : jel(ti(SALT and t has a maximal domain and C = (kU|(SALT nl D-variant(g) }, if not a singleton I { (loll}, otherwise J
= (lkollsALT n

Example 1 Gianni puo sposare un dottorequalsiasi. Gianni can marry a doctor whatever LF: [can [Gianni marrya doctor whatever]] Let us work out the innermostclause first. Step I D { Xw3w'3 ixe Dw(doctor,'(x) A marry,(G, x)) } by previous I([Gmarrya doctor whatever]lls computations I([Gmarrya doctor whatever]llsALT
[[ Xw3w'31xeD,'(doctor,'(x) A marry,(G, x)), A marryw(G, Xw3w'3 IxE D'w'(doctorw'(x) x)), A marryw(G, Xw3w'32xe Dw'(doctorw'(x) x)), ... A marryw(G, x)), ... Xw3w'32xeD'w'(doctorw'(x) A marryw(G, Xw3w'3ixe D"w'(doctorw'(x) A manryw(G, x)), x)), Xw3w'32xeD"w'(doctorw'(x)

In the above set, the rows are scales; the columns are D-variants.

BROADEN

YOUR

VIEWS

585

Step 2 fl[Gmarrya doctor whatever]lIs 2 { 1. Xw3w'3lxeDw'(doctorw'(x)A marryw(G, x)), 2. Oc (Xw3w'3lxe Dw'(doctorw (x) A marryw(G, x))) } by SCs { 1. Xw3w'3 lxe Dw(doctorw (x) A marryw(G, x)), 2. Xw3w'3ixe Dw'(doctor,'(x)A marryw(G, X)) A 3w'32xe Dw(doctorw (x) A marryw(G, x)) } by def. of Oc A marryw(G, { 1. Xw3w'3 Ixe Dw'(doctorw'(x) x)), 2. Xw3w'3! Ixe Dw(doctorw (x) A marryw(G, x)) } abbreviation (Oc II[G marrya doctor whatever]tI)ALT [[ O 0c (Xw3w'3lxeD,'(doctorw'(x) A marryw(G, x))), by SCALT A x))), Oc (Xw3w'3lxe D'tw(doctorw'(x) marryw(G, Oc (Xw3w'3lxe D",'(doctor,'(x) A marryw(G, x))),
*. ]] =

A marryw(G, 1[ Xw3w'3!ixe Dw'(doctorw'(x) x)), A marryw(G, Xw3w'3!ixe D'w'(doctorw'(X) x)), A marfyw(G,x)), Xw3w'3!lxe D",'(doctorw'(x) * 1] Let us turn,therefore, Going furtherat this stage (i.e., applying0-) would lead to contradiction. to the matrixclause. Step 3 D lican[G marrya doctor whatever]lls 1. A { ?w 3w"3lxxeD,(doctorw (x) marryw'(G, x)), 2. ?w 3w"3!1xc-Dw (doctorw (x) A marryw (G, x)) } by previous computations marrya doctor whatever]iis))ALT (Ijcanjls(II[G
= [[ ?w 3w"3!1xeDw,,(doctorw (x) A marry,'(G, x)), OW 3w"3!1xeDw (doctorw,,(x)A marryw(G, x)), ... A marryw A marryew(G, (G, x)), Ow 3w"3 Ixe D'w,'(doctorw,'(x) x)), ... ?w 3w"3! Ixe D'w (doctorw,"(x) 1 c XE A A 1x 3w"3! x)), 3w"3! (G, x)), ... (G, Ow OLw D"w "(doctorw,(x) marryw D"w,"(doctorw,(x) marryw

In the above set, the rows are scales; the columns are D-variants. Step 4 D Ilcan[G marrya doctor whatever]jls { 1. w 3w"3 lxe Dw (doctorw(x) A marryw (G, x)), 2. w 3w"3! lxe Dw (doctorw(x) A marryw (G, x)), 3. 0-c (Kw3w"3!1xeDw (doctorw,(x)A marryw'(G, x))) } by Mins { 1. Kw3w"31lxeDw (doctorw(x) A marryw (G, x)), 2. Kw3w"3! l xe Dw (doctorw (x) A marryw (G, x)), 3. VD Ow3w"3!lxeDw,'(doctorw,(x)Amarry'(G, x))} by def. of Oc (0-c (lican[Gmarrya doctor whatever]t1))ALT= [[VD Ow 3w"3! lxe Dw (doctorw (x) A marryw (G, x)), VD Lw 3w"3! lxe Dw (doctorw (x) A marryw (G, x))]]

586

GENNARO

CHIERCHIA

Step 5 D (lcan[G marrya doctor whatever]fls { 1. Kw3w"3 lxe Dw (doctorw(x) A marryw'(G, x)), 2. Ow 3w''3!lxeDw (doctorw(x) A marryw (G, x)), 3. VD Kw3w"3! ixe Dw (doctorw (x) A marryw'(G, x)), 4. Oc (VD 3\ 3w"3!1xeDw (doctorw (x) A marryw'(G, x))) } by SCs { 1. w 3w"3 lxe Dw (doctorw(x) A marryw (G, x)), 2. Kw3w'3! xe Dw (doctorw (x) A marryw'(G, x)), ixe Dw (doctorw(x) A marryw 3. VD Kw3w"3x! (G, x)), ixe Dw (doctorw 4. VD ?w 3w"3x! (x) A marryw(G, x)) A VD L]w3w"3!lixe Dw (doctorw (x) A marryw'(G, x)) } by def. of Oc
= (Oc (Ilcan[G marrya doctor whatever]lls))ALT

A marryw'(G, { ?w3w3lxwDxewD,(doctorw,(x) x)) } Here is the definition of the u-operatorin its two forms: (113) a. 11'4)j = nl 11tis a . I1I' lsALT { ((u J qIs } = n j(41js, C jkjIJj; if n jj4Jjs b. lic 4dII undefined,otherwise { 1k' b. 1k' (IALT PIIs } Example fjucan [G marrya doctor whatever]((= n (lcan[G marrya doctor whateverflls= A marryw'(G, ?w 3w"'3!lxExeDw(doctorw,"(x) x)) A VD Ow 3w"3! ixe Dw (doctorw (x) A marryw (G, x)) A VD Ew 3w'w3! ixe Dw,(doctorw (x) A marryw (G, x))
II((can [G marrya doctor whatever]ilsALTis the singleton of the above meaning.

Here is the definition of functionalapplication: (1 14) Functional application if 3 is not DE ilk'Ils(II1lls),

FAs. 11S PJls

(Il11(II131)) } | Iiils(llpli) U { oPci (111d(1l1311)).OPcn where Cl = |lo'll(lllIs ALT) and Cj+j = (OPCi (Iktjj(jIIjj)))ALT

I otherwise

The partof definition (114) thatpertainsto non-DE functions shouldbe self-explanatory. Insofar as non-DEfunctionsareconcerned,the idea is thatwe computeenrichedmeanings,freely applying enrichmentswith the initial value of C set to XIkl(OIIIsALT) until the set of alternativesbecomes a singleton(at which point, no form of enrichment can apply).The enrichments have to be applied by generalizing the relevant operationto the types that end in t (on the model of what Rooth (1985) does for only).
FAALT-

is ALT II=IIS ALT(i[YSALT), if |I@|I S { I13IIs(lhWls)AL otherwise llSjALT(((yjj),

not DE and jjyf(s is not a singleton J |

BROADEN

YOUR

VIEWS

587

Example 1 uno studentequalsiasi a student whatever

D
+ lIs)) + (]Is(lIwhateverIls + ,] studentwhatever[ + ((Istudentr Ila[ ,,IIs = IIa[ A 1(x) A QW(x)] } { XQXw3w'3x xe Dw' [studentw'(x) D { 1 studentin D } abbreviation, by FAs (non-DE) + ]JALT = luno studentequalsiasir [[I studentin D, 2 studentsin D, 3 studentsin D,... 1 studentin D', 2 studentsin D', 3 studentsin D',... 1 studentin D", 2 studentsin D", 3 studentsin D",...]]

by FAALT (non-DE)

The rows are scales; the columns are D-variants. Example 2 no studentwith two papers studentwith two papers[f+,]Ils 1jno[+?f]
-

{ 1. no (studentwith two papers),

2. Oc (no studentwith two papers)} by FAs (DE)

{ 1. no (studentwith two papers),

2. no (studentwith two papers)A some (studentwith one paper)} by def. of Oc

with two papers[+,] I S)ALT Ino[+,]fls(lstudent = [[ not all (studentswith two papers), few (studentswith two papers), no (studentswith two papers)]] by FAALT (DE) Example 3 No studentwith two papershas to read anythingfurther.
Ino[+? -] studentwith two papers[+,] has to read anythingjls

{ 1. no (studentwith two papers)(hasto read someDthing),

2. no (studentwith two papers)(hasto read someDthing)A some (studentwith one paper)(hasto read someDthing), 3. Ec (no (studentwith two papers)(hasto read someDthing))} [[ no (studentwith two papers)(hasto read someDthing), where C no (studentwith two papers)(hasto read someDthing), no (studentwith two papers)(hasto read someDWthing),...]] by FAs (DE) Itno[+,]studentwith two papers[+,] has to read anythinglISALT [[ not all (studentswith two papers)(haveto read someDthing), few (studentswith one paper)(haveto read someDthing), no (studentwith two papers)(hasto read someDthing)]] by FAALT (DE)
-

no[+,j studentwith two papers[+,] has to read anythinglls = { 1. no (studentwith two papers)(hasto read someDthing),

588

GENNARO CHIERCHIA

2. no (studentwith two papers)(hasto read someDthing)A some (studentwith one paper)(hasto read someDthing), 3. Ec (no (studentwith two papers)(hasto read someDthing)), 4. Oc' (no (studentwith two papers)(hasto read someDthing))} by SCs A4 Free Choice Effects with Universal Modalities In this section, I will show how the FC implicaturecomes aboutin the case of universalmodals. I will not considerthe derivationin every detail;instead, I will focus just on the relevantaspects. Consider: (1 15) Devo sposare un dottorequalunque. (I) must marry a doctor whatever (116) a. LF [must [I marrya doctor whatever]] b. Basic meaning A marry,'(I, x)) D, 3w"31IxeDw,,(doctorw,,(x) = Vw' [R(w, w') -> 3w"3 Ixe D,"(doctor,"(x)A marry,'(I, x))] c. Enrichedmeaning A marry,'(I, x))]) 0-c Oc' (Vw' [R(w, w') -- 3w"3 lxe Dv"(doctor,"(x) (1 17) Reductionsof enriched meaning a. 0-c Vw' [R(w, w') -> 3w"3! IxED,,(doctor,,"(x) A marry,'(I, x))] by def. of Oc b. VD' C D VD" c D-D' [Vw' [R(w, w') -A marryw'(I, 3w"3! IxE D'w,"(doctorw,"(x) x))] > [Vw' [R(w, w') -- 3w"3!ixe D`",(doctorw,(x)A marry,'(I, x))]] The assertion (116b) states that in every possible world, I marrya unique doctor selected from the widest domain D. The implicaturestates that if the assertionis true of a submaximaldomain D', it must also be true of every complementary domain. Suppose,for example, that the assertion is true of domain Dl = { a }; then it must also be true of every subset of D - { a }. But this clashes with the requirementthat the doctor I marrybe unique. Therefore,(1 17b) can only be vacuously true: it can be true only if the antecedentis false. We thus conclude that (1 16b) and ( 1l7b) entail ( 117c). (117) c.
3-,

D' C D Vw' [R(w, w') ->

(x) A w3!1xeD'w (doctorw

marryw (I, x))]

So, for (1 16a) to hold, given (117c), it must be the case that for every world there is a distinct memberof D that I marry. (117) d. Vw' [R(w, w') -> 3w"T! Ixe Dw,(doctorw,(x) Amarryw'(I, x))] This is the FC implicature.

BROADEN

YOUR

VIEWS

589

As a referee notes, it is crucial in the above proof that (117b) be restrictedto submaximal domains, to the exclusion of the maximal domainD itself. Otherwise,the statementitself would constitutean antecedentto (1 17b) and one would derive (118). (1 18) VD Vw' [R(w, w') -> 3w"3 Ixe Dw (doctorw(x) A marryw (I, x))] But (118) can only be satisfied in worlds where there is a unique doctor, a requirement that imposes exceedingly strong uniquenessconditions on the interpretation of (15). References
Chierchia,Gennaro.2004. Scalarimplicatures,polarityphenomena,and the syntax/pragmatics interface.In The cartographyof syntactic structures. Vol. 3, Structuresand beyond, ed. by Adriana Belletti. Oxford:Oxford University Press. Dalrymple,Mary,MakotoKanazawa,Yookyung Kim, Sam Mchombo,and StanleyPeters. 1998. Reciprocal expressions and the concept of reciprocity.Linguisticsand Philosophy 21:159-210. Dayal, Veneeta. 1998. Any as inherentlymodal. Linguisticsand Philosophy 21:433-476. von Fintel, Kai. 1999. NPI-licensing, Strawson-entailment, and context dependency.Journal of Semantics 16:97-148. Fox, Danny. 2003. The interpretation of scalarterms:Semanticsor pragmatics,or both? Paperpresentedat the University of Texas, Austin. Gamut, L. T. F. 1991. Logic, language, and meaning. Vol. 2, Intensional logic and logical grammar. Chicago: University of Chicago Press. Gazdar,Gerald. 1979. Pragmatics: Implicature,presupposition,and logical form. New York: Academic Press. Giannakidou,Anastasia.2001. The meaning of free choice. Linguisticsand Philosophy 24:659-735. Grice, Paul. 1989. Studies in the way of words. Cambridge,Mass.: HarvardUniversity Press. Groenendijk,Jeroen, and MartinStokhof. 1984. Studies in the semantics of questions and the pragmatics of answers. Akademischproefschrift,University of Amsterdam. Hamblin,Charles. 1973. Questions in MontagueEnglish. Foundationsof Language 10:41-53. Haspelmath,Martin. 1997. Indefinitepronouns. Oxford:Oxford University Press. Horn, LaurenceR. 1989. A natural history of negation. Chicago: University of Chicago Press. Horn,LaurenceR. 1999.Any and(-)ever:Freechoice andfree relatives.Paperpresentedat IsraelAssociation for TheoreticalLinguistics (IATL) 15, Haifa. lonin, Tania, and Ora Matushansky.2005. 1001 nights: The syntax and semantics of complex numerals. Ms., University of SouthernCalifornia,Los Angeles, and Universite Paris 8/CNRS. Kadmon,Nirit, and Fred Landman.1993. Any. Linguisticsand Philosophy 15:353-422. Kratzer,Angelika. 2003. Scalar implicatures:Are there any? Paper presentedat the workshop "Polarity, Scalar Phenomena,and Implicatures,"University of Milan - Bicocca, June 18-20. Kratzer,Angelika, and Junko Shimoyama. 2002. Indeterminatepronouns:The view from Japanese. In Proceedings of the ThirdTokyoConferenceon Psycholinguistics,ed. by Yukio Otsu, 1-25. Tokyo: Hituzi Syobo. Krifka,Manfred. 1995. The semantics and pragmaticsof polarity items. LinguisticAnalysis 25:209-257. Krifka, Manfred. 2003. Polarity items in questions. Paper presented at the workshop "Polarity, Scalar Phenomena,and Implicatures,"University of Milan - Bicocca, June 18-20. Krifka, Manfred,Francis Jeffry Pelletier, Gregory N. Carlson, Alice ter Meulen, GennaroChierchia,and GodehardLink. 1995. Genericity:An introduction. In Thegeneric book, ed. by GregoryN. Carlson and Jeffry Pelletier, 1-124. Chicago: University of Chicago Press.

590

GENNARO

CHIERCHIA

Ladusaw,William. 1992. Expressingnegation. In Proceedings of SALTII, ed. by Chris Barkerand David Dowty, 237-259. Columbus,Ohio: Ohio State University, Departmentof Linguistics. Lahiri,Utpal. 1998. Focus and negative polarity in Hindi. Natural Language Semantics6:57-125. Laka, Itziar. 1990. Negation in syntax: On the nature of functional categories and projections. Doctoral dissertation,MIT, Cambridge,Mass. Landman,Fred. 1998. Plurals and maximalization.In Events and grammar,ed. by Susan Rothstein.Dordrecht:Kluwer. Doctoraldissertation, Jean. 1975. 'Or'and 'any':The syntaxandsemanticsof two logical operators. LeGrand, University of Chicago, Chicago, Ill. Levinson, Stephen C. 2000. Presumptivemeanings. Cambridge,Mass.: MIT Press. of choice. Doctoraldissertation,Universityof Massachusetts, Menendez-Benito,Paula. 2005. The grammar Amherst. Percus,Orin.2000. Constraints on some othervariablesin syntax.NaturalLanguageSemantics8:175-229. Recanati,Franqois.2003. Embeddedimplicatures.Ms., InstitutJean Nicod, Paris. Rizzi, Luigi. 1990. Relativizedminimality.Cambridge,Mass.: MIT Press. Rizzi, Luigi. 2004. Locality and left periphery.In Thecartographyof syntacticstructures.Vol. 3, Structures and beyond, ed. by AdrianaBelletti, 223-252. Oxford:Oxford University Press. Rooth, Mats. 1985. Association with focus. Doctoral dissertation,University of Massachusetts,Amherst. Rooth, Mats. 1992. A theory of focus interpretation. Natural Language Semantics 1:75-117. Seb0, Kjell Johann.2001. The semanticsof Scandinavianfree choice items. Linguisticsand Philosophy24: 737-787. Sauerland,Uli. 2004. Scalar implicaturesin complex sentences. Linguisticsand Philosophy 27:367-391. Sauerland,Uli. 2005. The epistemic step. Paper presented at "ExperimentalPragmatics," University of Cambridge,April. Local or global? Paperpresentedat the workshop "Polarity, Spector,Benjamin.2003. Scalarimplicatures: Scalar Phenomena,and Implicatures,"University of Milan - Bicocca, June 18-20. Stalnaker,Robert. 1978. Assertion. In Syntax and semantics 9: Pragmatics, ed. by Peter Cole, 315-332. New York: Academic Press. Westerstahl,Dag. 1988. Quantifiersin formal and naturallanguages. In Handbookof philosophical logic, vol. 4, ed. by Dov M. Gabbayand Franz Guenthner,1-131. Dordrecht:Kluwer. Departmentof Linguistics Boylston Hall Harvard University Cambridge,Massachusetts02138 chierch@fas.harvard. edu Dipartimentodi Psicologia Universityof Milan - Bicocca Piazza dell'Ateneo Nuovo 1 201126 Milan Italy