1. Introduction

A number of constructions in various languages display a different behavior in the scope of epistemic and deontic modals. For example, the German indefinite determiner irgendein gives rise to different inferences under the two kinds of modals (Aloni and PortReference Aloni, Port, Fainleib, LaCara and Park2010; Kratzer and ShimoyamaReference Kratzer, Shimoyama and Otsu2002). Furthermore, while the Romanian determiner vreun is licensed under epistemic modals, but not under deontic modals (Fălăuş Reference Fălăuş2009), concessive scalar particles like Slovenian magari are licensed under deontic modals but not under epistemic ones (e.g. Crnič¸, Reference Crnič, Fainleib, LaCara and Park2011b). Although seemingly a diverse set of observations, there is a curious commonality in the proposed explanations for the relevant data: all of the three case studies independently assume that deontic and epistemic modals have different free choice (FC) potential, in the sense that these modals differ in the way that they license so-called FC-inferences. We consider FC-inferences to be inferences of the form in (1b) and (2b) associated with either existentials or disjunctions under modals, as in (1a) and (2a) respectively:

While epistemic FC-inferences, i.e., FC-inferences associated with epistemic modals, mostly seem to be well-behaved pragmatic inferences outside of compositional semantics, deontic FC-inferences seem able to penetrate into the computation of semantic values much more freely. It is this Modal Variability Hypothesis, as we will call it, that is the main object of scrutiny in this chapter.

We argue that the hypothesized differences in FC-potential are not due to a sharp semantic contrast between modalities, but propose a pragmatic explanation for the Modal Variability Hypothesis. We acknowledge the possibility, explored in great depth by Chierchia (Reference Chierchia2006, to appear a) and others, of (quasi-)pragmatic inferences taking effect “locally” during the composition of semantic meaning, but we still argue for a genuine pragmatic perspective on the distribution of such local pragmatic effects. We suggest that whether a pragmatic inference reliably enters semantics proper is a matter of pragmatic fossilization: since deontic FC-inferences are of a different type of information than epistemic FC-inferences, the former are more readily adaptable to the computation of semantic values.

Our main argument for the pragmatic difference between deontic and epistemic information hinges on the contextual variability of universal FC-inferences, i.e., FC-inferences that apparently occur in the scope of universal quantifiers, as is the case when an utterance of (3a) conveys the strengthened meaning expressed in (3b).

We argue that the availability of universal FC-inferences depends, among other things, on the contextual relevance of the inference in question. Contrasting situations in which deontic and epistemic universal FC-inferences do or do not arise, we argue that deontic FC-inferences are more clearly relevant for practical purposes than their epistemic counterparts. This suggests that deontic information is more readily perceived as “base-level” information about what the actual state of affairs is than epistemic information. In turn, this makes plausible why, from a diachronic perspective, deontic FC would fossilize sooner and/or more readily than epistemic FC.

The proposed intuitive difference in relevance for “base-level” information can be made tangible in dynamic semantics. In standard dynamic accounts of epistemic modality (VeltmanReference Veltman1996; GilliesReference Gillies2004; YalcinReference Yalcin2007), epistemic modals are not eliminative updates, but tests on whether the currently accumulated information supports or is compatible with some piece of further information. In contrast, deontic modal statements that provide directly useful information about practically relevant permissions and obligations are most plausibly treated as base-level informative, on a par with propositional information. We show that the difference between deontic and epistemic FC-inferences clearly shows in a dynamic semantics in the way that the former are persistent, while the latter are not only not persistent but even anti-persistent. Persistent information survives information growth; anti-persistent information survives information loss.

The chapter is structured as follows. After reviewing the relevant data and theories on Romanian vreun, Slovenian magari (and related constructions), and German irgendein in Section 2, we formulate the Modal Variability Hypothesis in Section 2.4. Section 3 discusses but dismisses an obvious semantic explanation of modal variability implemented in an alternative-based framework (AloniReference Aloni2007b; Chierchia to appear a). Building on data concerning universal FC-inferences in Section 4, Section 5 states our pragmatic explanation of modal variability. Section 6, finally, tries to corroborate the informal picture by tracing the underlying intuitions in dynamic semantics. Section 7 concludes with some reflections on the conceptual consequences of our proposal.

2. Data in support of modal variability

In this section, we briefly review the main properties of Romanian vreun (Fălăuş Reference Fălăuş2009, Reference Fălăuş, Li and Lutz2011, Reference Fălăuş2012), Slovenian magari (and related constructions) (CrničReference Crnič, Fainleib, LaCara and Park2011b), and German irgendein (Aloni and PortReference Aloni, Port, Fainleib, LaCara and Park2010; Kratzer and ShimoyamaReference Kratzer, Shimoyama and Otsu2002), focusing on their behavior in modal contexts. We also sketch three of the most recent accounts of these data sets. Interestingly, all of these analyses assume a different FC-potential for deontic and epistemic modals. We therefore formulate the Modal Variability Hypothesis in Section 2.4.

3. Semantic explanations for modal variability

A natural way to implement the Modal Variability Hypothesis assumes a formal/sortal difference between deontic and epistemic modals: FC-inferences are derived as semantic entailments for disjunctions and existentials under the former, but not under the latter. We are not aware of any such account in the literature, so we attempt to develop one that is as viable as possible. This hypothetical analysis adopts Chierchia’s (to appear a) framework of polarity and free choice and attempts to accommodate the phenomena we discussed in Sections 2.1 and 2.3, namely the different behavior of vreunand irgendein under deontic and epistemic modals, by employing, within Chierchia’s framework, an explicit semantic analysis of deontic FC-inferences along the lines of Aloni (Reference Aloni2007b).

The main motivation for adopting Aloni’s (Reference Aloni2007b) analysis comes from the following considerations. A first potential difficulty for an entailment account of deontic FC-inferences is that it would predict total variation effects for all indefinites under deontic modals, even unmarked ones as in “John may/must marry someone,” which clearly does not entail that any person is a permissible marriage option for John. Aloni’s (Reference Aloni2007b) analysis provides a potential way out of this difficulty by allowing two different representations for indefinites – an alternative-inducing and a flat representation – and by letting modals be sensitive to the alternatives generated in their scope. Only alternative-inducing indefinites give rise to free choice entailments on this account. Unmarked indefinites can then be required to adopt flat representations.

According to Aloni (Reference Aloni2007b), alternative-inducing representations of indefinites generate a genuine set of propositional alternatives, but not flat representations:

Possibility modals and imperatives (but nothing prevents us from extending this analysis to necessity modals as well)⁶ are assumed to entail that all alternatives generated in their scope should be compatible with the relevant modal base:

On this analysis of modality, only alternative-inducing indefinites give rise to FC-inferences in modal environments. Unmarked indefinites like someone can be then required to employ flat representations. Marked indefinites like irgendein and vreun instead should use alternative-inducing representations. FC-effects would be generated solely for the latter.

Let’s adopt the analysis of modality in (32), but solely for priority modals. Since FC-inferences would be generated under priority modals but not under epistemic modals, the distribution of concessive scalar particles would be explained along the lines of Crnič (Reference Crnič, Fainleib, LaCara and Park2011b). As for epistemic indefinites, adopting Chierchia’s (to appear a) alternative-based account, we could assume that both irgendein and vreun generate singleton domain alternatives as in (30b). Under epistemic modals, via recursive exhaustification in the vein of Chierchia or Fox we can derive partial variation effects, as in example (11) from Fălăuş¸. Under deontic modals, total variation effects are independently derived because of (32). Vreun would then be predicted to be out under deontic modals, because of a clash between the deontic FC-entailment and the lexically encoded ban on total variation, as in example (12). The different behavior of irgendein under deontic and epistemic modals would also be accounted for: semantic total variation would be derived under the former (via (32)), and obligatory pragmatic partial variation effects would be generated under the latter (via recursive exhaustification). There are however a number of problems for this strategy.

First of all there are other examples of marked indefinites for which we want to derive partial variation under epistemic modals, but, as it seems, no total variation under deontic or other priority modals, like Spanish algún or Italian un qualche. Let us illustrate these facts for the Italian case. The compatibility of (33) with the HideAnd Seek Scenario in (23) shows that in epistemic contexts un qualche only generates partial variation effects (Aloni and PortReference Aloni, Port, Fainleib, LaCara and Park2010):

That un qualche does not give rise to total variation effects under priority modals is demonstrated by the following example from the web, where the continuation explicitly specifies that not any kind of basic skill is enough:

Why do these examples constitute a problem for the strategy we are investigating in this section? To generate epistemic partial variation in Chierchia’s approach we need to assume that un qualche activates singleton alternatives, but then, given (32), total variation under priority modals would be automatically generated as well (unless we assume that in Italian priority modals behave differently than in Romanian and German, which seems quite implausible).

A second more conceptual problem of this implementation is that there is no reason why the generation of deontic FC-inferences formalized in (32) should not automatically extend to epistemic modals. The difference in FC-potential between epistemic and deontic modals would be stipulated. Why would only the latter be sensitive to the alternatives generated in their scope?⁷

A final and quite general problem with a semantic explanation of (29) is that the to-be-explained difference in FC-potential does not seem to be a black-and-white matter, but rather calls for a more nuanced picture of why only some FC-inferences penetrate into compositional semantics. This is the topic of the following section.

4. Contextual-variability of universal free choice

In order to come to grips with the variable FC-potential of deontic and epistemic modals, this section looks more closely at the behavior of disjunction in the scope of these modals. FC-inference can arise for both deontic and epistemic modals if they embed a disjunction, as witnessed by the following examples:

But also other modalities, e.g., ability modals, can give rise to FC-inferences:

Moreover, as observed by Klinedinst (Reference Klinedinst2006), Eckardt (Reference Eckardt, Sauerland and Stateva2007), and Fox (Reference Fox, Sauerland and Stateva2007), FC-like inferences may even arise for disjunctions when embedded under non-modal existential constructions (38).

Under a standard semantic analysis of the components involved, all of these putative FC-inferences would not be considered semantic entailments. This could mean that these inferences are pragmatic inferences, but it could also mean that the standard semantics is wrong. The main argument why these inferences are probably not semantic entailments comes from the observation that FC-inferences do not seem to arise in all downward-entailing environments, such as in the scope of quantifier none:

So far the picture seems quite homogeneous, with different kinds of modals (and even non-modal constructions) giving rise to FC-inferences, with some evidence that these are pragmatic entailments. But the impression of peaceful homogeneity is spoiled, when we look at how readily FC-inferences associated with disjunctions and different modals embed. Indeed, in spite of the data in (39) and (40), there is not a total ban on FC-inferences possibly occurring in the scope of other logical operators. It has been observed, for instance, that FC-inferences associated with disjunction under deontic modals apparently can take scope under universal quantifiers, so-called universal free choice (henceforth: UFC) (cf. ChemlaReference Chemla2009b, for empirical evidence):

For our purposes here it is important to notice that UFC does not arise as readily for other sorts of modals. More concretely, the apparent strength of UFC seems to depend on the kind of modality. This was suggested first by Geurts and Pouscoulous (Reference Geurts and Pouscoulous2009) based on introspective judgments, but is also backed-up by empirical data (Tiel Reference Tiel2011). The study of Tiel provides clear evidence that deontic UFC, as in (41) is much more readily endorsed by the participants than epistemic UFC, as in (42).

On a coarse-grained picture, the different UFC-potential of deontic and epistemic modals seems to support the Modal Variability Hypothesis in (29): deontic FC-inferences infiltrate compositional semantics, epistemic FC-inferences do not, or at least not to the same degree. The problem is that this coarsed-grained picture is too coarse-grained. As soon as we zoom in, we realize that there is a devil in the quantitative details. The problem is that the Modal Variability Hypothesis in (29) would rather lead us to expect that while deontic UFC is readily attested, epistemic UFC should be impossible, or at least very inaccessible. But that is not supported by Tiel’s data: epistemic UFC is attested, but much less prominent than deontic UFC. In other words, although (29) does square well with the qualitative difference in UFC-potential of deontic and epistemic modals, it does not square well with the more nuanced quantitative picture.

Of course, there can be many factors figuring into the explanation why epistemic UFC is less readily attested. We could think that epistemic FC is more computationally costly than deontic FC (irrespective of which system computes these inferences), or less available for other reasons – reasons that lie outside of formal semantics proper. In essence, this is exactly the line of explanation that we would like to explore here. We would like to suggest for consideration the idea that the strength of an FC-inference is relative to the relevance of the inferred information.^{8, 9} To demonstrate this, we consider sentences that could give rise to UFC-inferences in two contexts each: one which makes the UFC-inference relevant for practical purposes and one which does not (cf. Malamud forthcoming, for a similar approach to the interpretation of plural definites). Our main argument is that contexts of the former type are not equally natural for different kinds of modals, suggesting that the availability of embedded FC-inferences depends on pragmatic factors of relevance of information in context.

Let’s consider a “neutral” case first, namely the UFC-potential of ability modals. The study of Tiel attests only very little UFC-potential to ability modals. And, indeed, it is also intuitive that the sentence in (43a) does not readily invite the UFC-inference in (43b).

However, the availability of (43b) as an inference from (43a) seems to depend crucially on the context of utterance and, in particular, on the relevance of the information in (43b). To see this consider two scenarios:

When uttered in a context like (45), the sentence (43a) clearly does not convey the strengthened meaning in (43b). In contrast, when uttered in a context like (44), it seems to us that (43a) does convey (43b). It seems to us that, when you get the information in (43a) in context (44), it would be justified for you to assign to either the violin or the trombone any member of the ILLC you choose. If we had not meant you to do that, we should have given you more specific information. In other words, the UFC-inference in (43b) is available in (44) but not (45) because the question whether the “ability predicate”

holds true of any arbitrary member of the ILLC and any arbitrary relevant instrument is of acute practical relevance in context (44), but not in context (45).

We suggest that the availability of UFC-inferences under deontic and epistemic modals similarly depends on matters of contextual relevance. Take the more critical case of epistemic UFC. We claim that whether an utterance of (47a) triggers the UFC-inference in (47b) depends on context, similarly to the previous case of ability modals.

To test this intuition, consider again two scenarios:

In context (49), the UFC-inference in (47b) clearly does not heavily invite itself, because it is not relevant in this context. But in the context (48), where it is relevant for your practical choice of action whether the “epistemic relation”

is true of each member of the ILLC and each victim in question, the UFC-inference does seem to arise. Of course, the example is quite contrived and set up deliberately so that Clouseau’s epistemic state about whether x is a suspect for case y matters for practical purposes. But the artificiality of this context is exactly what we are interested in. We hypothesize that, in general, it takes rather exceptional and involved situations in which an epistemic UFC-inference would arise because “epistemic relations” like (50) are seldom practically relevant.

This is different for deontic UFC in context. Let’s consider the question whether (51a) gives rise to the UFC-inference in (51b).

Parallel to the previous cases, here are two scenarios, one which makes the UFC inference practically relevant, and the other one which does not.

According to our intuitions, the UFC-inference in (51b) clearly arises in context (52), and not, or at least clearly less strongly, in context (53). The context (52) is one where it is a relevant issue whether the “deontic predicate”

is true of all relevant x and y. What is interesting for our argument is that the context (52) does not at all seem very contrived or artificially set up. If at all, then it’s rather (53) that appears less natural or common.

Of course, having provided only one pair of contexts for each case, we extrapolate from a weak inductive basis here, but we still feel that there is a general difference in the naturalness of contexts in which UFC-inferences for different types of modals would be (practically) relevant. We suggest that this is due to a pragmatic circumstance: whether information expressed by modalized predicates like (46), (50), or (54) are frequently relevant for concrete practical purposes.

5. Modal variability from pragmatic relevance

If the picture we just sketched is correct, then a pragmatic explanation for the Modal Variability Hypothesis pops into view, especially when we consider a diachronic perspective: if putative FC-inferences of different modal flavors differ in the frequency or strength of their contextual relevance, and so are available more or less frequently/strongly, then, over time, this will lead to differences in the extent to which these inferences could infiltrate compositional semantics. Essentially, this is a direct application of the well-established notion of pragmatic fossilization that yesterday’s pragmatics will gradually become tomorrow’s semantics.¹⁰

More concretely, suppose a community of language users has available two kinds of constructions, type-1 and type-2. Type-1 constructions are able to give rise to deontic FC-inferences; type-2 to epistemic FC-inferences. Suppose further that these FC-inferences are originally entirely pragmatic and do not enter the recursive computation of semantic values at all. But now let’s also assume, as we tried to make plausible above, that deontic FC-inferences are more readily available than epistemic FC-inferences if we look at a random sample of conversational contexts in which either type of construction is used. In that case, if pragmatics fossilizes, then it must be that deontic FC-inferences, simply due to a slight advance in contextual prominence, would be more prone to enter the compositional computation of semantics values. As the differences in informational relevance are presumably universal, the hypothesized asymmetry in the pattern of fossilization would likewise presumably be universal.¹¹ Consequently, the Modal Variability Hypothesis might be a true empirical generalization at the current stage of language evolution, not because of a strong sortal difference between modalities, but because of a tendentious pragmatic difference that is operative on a long time scale.

Essentially, this diachronic view endorses the general possibility of “local implicatures” in the spirit of Chierchia (Reference Chierchia2006, to appear a, to appear b) and Chierchia et al. (Reference Chierchia, Fox, Spector, Maienborn, von Heusinger and Portner2012), but also adds its own particular flavor to the overall picture. Chierchia’s approach to “grammaticalized implicatures” has obvious empirical advantages over purely pragmatic approaches when it comes to positive fit to the empirical data: allowing pragmatic inferences to systematically intrude into compositional semantics yields more fine-grained, but also seemingly better empirical predictions in many domains of application than standard “globalist” accounts of pragmatic inferences (cf. Chierchia et al. Reference Chierchia, Fox, Spector, Maienborn, von Heusinger and Portner2012; SauerlandReference Sauerland2012, for recent arguments). But allowing implicatures into compositional semantics runs the risk of generating problems with negative fit: the generating power of potential “local implicatures” is overwhelming and not borne out empirically. For instance, lacking special intonation that might trigger alternative pragmatic processes, FC-inferences do not arise in the scope of none as in (39). To cope with issues of “negative fit,” usually some variant of the strongest meaning hypothesis of Dalrympleet al. (Reference Dalrymple, Kanazawa, Kim, Mchombo and Peters1998) is employed (cf. Chierchia to appear b; Chierchia et al. Reference Chierchia, Fox, Spector, Maienborn, von Heusinger and Portner2012; Sauerland Reference Sauerland2012). Roughly speaking, the strongest meaning hypothesis states that if two alternative semantic interpretations are available, then ceteris paribus the semantically stronger one is to be preferred. Unfortunately, the case of modal variability cannot be explained by appeal to a strongest meaning hypothesis. Adding FC-inference to an existential modal sentence always creates a stronger meaning irrespective of the nature of the modal. Hence, the strongest meaning hypothesis cannot explain any differences in FC-potential. Consequently, the pattern of available local pragmatic effects must be explained in a different way. This is exactly where our present proposal adds to the picture. The proposed diachronic perspective vindicates the general possibility of local pragmatic effects, but both adds a general explanation for patterns of distribution of local pragmatic effects, and also yields, not a black-or-white picture, but the more nuanced quantitative perspective that, arguably, the empirical data demand.

To make these ideas at least somewhat more precise, we would like to conclude this book by sketching the outlines of a framework, couched in dynamic semantics, that shows how deontic, but not epistemic FC-inferences could be treated as embeddable pragmatic effects, because of a difference in the usual base-level relevance of the information conveyed by these modals in discourse (cf. Aloni Reference Aloni, Guevara, Chernilovskaya and Nouwen2012).

6. Deontic and epistemic FC in discourse

To account for the different discourse behavior of deontic and epistemic FC, we will adopt a dynamic account of epistemic modality (GilliesReference Gillies2004; VeltmanReference Veltman1996; YalcinReference Yalcin2007) combined with a more classical approach to deontic modality and propose a genuinely dynamic mechanism to generate and incorporate pragmatic inferences which is sensitive to the differences between the two modalities. We will assume that FC-inferences of a given form ϕ are implicatures that can be captured by specifying a set of states in which ϕ would optimally or rationally be uttered by a speaker (given the usual assumptions of cooperativity, mutual recognition of relevance of information etc.). Aloni (Reference Aloni, ten Cate and Zeevat2007a) and Franke (Reference Franke2011), among others, both spell out systems that give us exactly that, and we will stay close to these accounts, but nothing hinges on specifics here. Also, our exposition here focuses on FC-inferences derived from disjunction, but this too is for convenience only. The crucial bit is the integration of the information provided by deontic and epistemic FC-inferences into the discourse: only deontic FC-inferences are persistent and therefore can be fully integrated (via Aloni’s (Reference Aloni, Guevara, Chernilovskaya and Nouwen2012) +I-operator) into the dynamic process of interpretation in a non-trivial way. The reason why deontic and epistemic information is treated differently in dynamics semantics is, we suggest, exactly the difference in practical relevance of information of that kind, as argued above in Section 4.

7. Conclusion

Summing up, we have here reviewed evidence from three independent case-studies that the FC-inferences associated with deontic modals appear more “semanticized” than their epistemic counterparts. Taking the possibility of local pragmatic effects seriously, but noting that the usual explanation for the distribution of local pragmatic effects in terms of the strongest meaning hypothesis (cf. Chierchia to appear b; Chierchia et al. Reference Chierchia, Fox, Spector, Maienborn, von Heusinger and Portner2012; SauerlandReference Sauerland2012) does not apply in the case at hand, we tried to give a diachronic pragmatic explanation for the different degree to which deontic and epistemic FC-inferences penetrate into compositional semantics. Our proposal revolves around the notion of pragmatic fossilization, and we argued that deontic FC-inferences are practically relevant more often than epistemic FC-inferences. We discussed data on universal FC-inferences in favor of this claim and showed how this difference cashes out in dynamic semantics as a difference in persistence of the FC-information. In sum, according to our proposal, it is essentially basic features of contextual relevance that affect why some inferences penetrate compositional semantics more readily than others, even to the extent that these asymmetries in fossilization patterns inform grammaticality judgments (as in the case of vreun or *magari).

An important conceptual question is how this explanation of the Modal Variability Hypothesis relates to the “grammatical view” of scalar implicatures, as championed by Chierchia (Reference Chierchia2006, to appear a) and others. As mentioned before, our account acknowledges the possibility of local pragmatic effects, but, migrating beyond applications of the strongest meaning hypothesis, we have tried to give a diachronic pragmatic explanation for the distribution of these. This essentially commits us to a particular interpretation of the “grammatical view.” In a recent paper, Chierchia (to appear b) argues that the “grammatical view” of scalar implicature calculation can (but need not) be seen as a stepwise expansion of the more confined classical Gricean reasoning that has long been taken for gospel (Grice Reference Grice, Cole and Morgan1975). But it can also be considered a purely grammatical thing. What we suggested in this chapter commits us to consider the “grammatical view” in the first manner, as intrinsically pragmatic, even if integrated into formal semantic theory – that is, subject to general functionalist considerations about language use. This is opposed to a strictly grammatical view which would consider language an object of inquiry on its own, detached from issues of language use, that would then not admit functional considerations to play a role in explanations of local pragmatic effects.

Even if equivalent in terms of actual predictions, there is a conceptual difference between these two perspectives on the “grammatical view” that shows if we look at which auxiliary assumptions would be deemed feasible to adapt the formalism to new sets of empirical data. The data we discussed in Section 2 is such a set of novel empirical data that challenges the current formulation of the “grammatical view,” in particular, its sole reliance on the strongest meaning hypothesis to account for the distribution of embedded implicatures. To account for (something like) the Modal Variability Hypothesis, an intrinsically pragmatic grammatical view can appeal to differences in frequencies of contextual relevance. But a strict grammatical view cannot. Viable explanations for the strict grammatical view could involve stipulating (and motivating) formal differences in the way different modals interact with domain alternatives and the like. No knock-down argument has been given here that the Modal Variability Hypothesis cannot be explained in terms of a strict grammatical view. But the burden of proof that modal variability is amenable to a strictly non-pragmatic explanation lies with those who favor this view.

Acknowledgments

This research has been largely inspired by the work of Gennaro Chierchia: mostly by his recent research on free choice and local implicatures but also by his less recent contributions on dynamic semantics. Maria Aloni would further like to thank Gennaro Chierchia for supervising her 1996 Master thesis on “Dynamic semantics and epistemic modalities,” and for his contagious passion for linguistic research: as this article tries to demonstrate, in even the most intricate empirical data some logical pattern can be found. We also thank Stephanie Solt and four anonymous reviewers for helpful comments, as well as the financial support of the Netherlands Organization for Scientific Research (NWO).