Matthewson (2016) Modality

18
Modality
Lisa Matthewson
18.1 Introduction 525

18.2 Modal flavour 529
18.3 Modal force 542
18.4 Modal–temporal interactions 548
18.5 Typological and other remaining questions 555
18.1 Introduction
There are an infinite number of ways the world could have been. For exam-
ple, my first name happens to be Lisa, but my parents could have named me
Super Princess Lisa Superhero Easter Bunny.1 Another way to say this is there
are some possible worlds (Lewis, 1973) where my name is Super Princess
Lisa Superhero Easter Bunny. Natural language allows us to say all kinds of
things about these possible worlds – such as My parents should have named me
Super Princess Lisa Superhero Easter Bunny, or If my parents had named me Super
Princess Lisa Superhero Easter Bunny, I would have been angry at them. The ability
to talk about possible worlds is known as modal displacement (von Fintel
and Gillies, 2011, who follow Hockett’s (1960) discussion of temporal and
spatial displacement).
Modal displacement is, as far as we know, unique to human language,
universal to all languages, and acquired early. English-acquiring two-year-
olds freely talk about what is permitted, obligated, or compatible with some-
one’s abilities; they say things like you can’t get it till it’s Mother’s Day,2 he just
can borrow him,3 or you have to be quiet so she can sleep,4 and we expect that
1
Thanks to Madeleine Davis (p.c.) for the name idea.
2
www.youtube.com/watch?v=YQ9vv4CPzcY&feature=related.
3 4
www.youtube.com/watch?v=NDLK6dopfuc. www.youtube.com/watch?v=4Y5nbStIjeI.
526 L I S A M AT T H E W S O N
speakers of all languages (even those under-represented on YouTube) can

similarly talk about permission, obligation, and ability.
Interestingly, however, languages also vary in how they express and
categorize modal meanings. Unlike in English, in the Salish language
St’át’imcets (Lillooet) the same morpheme (the enclitic =ka) can express
either permission or obligation, as in (1). A different morpheme (the circumfix
ka- . . . -a) is used to express ability, as in (2).5
(1) wá7=ka s-lep’ i=k’ún7=a ku=pála7 máqa7

ipfv=deon stat-bury det.pl=fish.egg=exis det=one snow
i. ‘The eggs can stay in the ground for a year.’
ii. ‘The eggs have to stay in the ground for a year.’
(Rullmann et al., 2008, p. 329)
(2) Context: They also wrestled the young bulls.

wa7 xíl-em=wit ets7á kw=s=zwat-en-ítas swát=as
ipfv do-mid=3pl deic det=nmlz=know-dir-3pl.erg who=3sbjv
ku=wá7 ka-xilh-ts-tal’í-ha áti7 ku=xwém
det=ipfv circ-do-caus-top-circ deic det=fast
‘They did that to find out who could do it the fastest.’
(Matthewson, 2005, p. 89)
The goal of this chapter is to provide an introduction to three topics which

are central to understanding the semantics of modal elements: modal
flavour (Section 18.2), modal force (Section 18.3), and modal–temporal inter-
actions (Section 18.4). For each of these phenomena, I will outline its
treatment in a standard Kratzerian semantics (Kratzer, 1981, 1991b, 2012),
present some data on its expression cross-linguistically, and point the reader
to some ongoing debates and interesting unanswered questions. In the
remainder of this section, I give a basic introduction to the possible-worlds
approach to modality and to the three major topics we will be discussing.
Several introductions to modality are already available: see, for example,
von Fintel and Gillies (2011), Hacquard (2011), or Swanson (2008) for a more
philosophical perspective. Other relevant recent works include Portner’s
(2009) comprehensive monograph and Kratzer’s (2012) collection of updated
versions of her seminal papers. In this chapter I will be unashamedly Kratze-
rian, because Kratzer is where every student of modality needs to begin.
There are other approaches (see, e.g., Groenendijk et al., 1996; Ninan, 2005;
Yalcin, 2007; Lassiter, 2011b), and there is also of course no single Kratzerian
analysis, but rather a family of related analyses, which are constantly being
developed and updated in ongoing research.
5
Abbreviations in this chapter not covered by the Leipzig Glossing Rules: ii : series ii pronoun, circ :
circumstantial, cn : common noun connective, cust : customary aspect, deic : deictic, deon : deontic, dir :
directive transitivizer, emph : emphatic, epis : epistemic modal, exis : assertion of existence, infer : inferential
evidential, lex : lexical particle, mid : middle, mod : modal, necess : necessity modal, nonaff : non-affirmative,
pn : proper name, possib : possibility modal, prosp : prospective aspect, sens.non.vis : sensory non-visual
evidential, s.pl : plural subject, stat : stative, vai : intransitive animate verb, val : valency adjuster, vti :
transitive inanimate verb.
Modality 527
18.1.1 Possible worlds and modality

Propositions containing modal elements make contingent claims about the
actual world, by asserting something about a set of worlds in which certain
propositions are true. A simple example is given in (3).
(3) Context: The Youth Orchestra is holding auditions. At one point the
conductor says:
Madeleine can put her violin away now.
Can is a possibility modal, which introduces existential quantification over

worlds. So as a first pass, (3) asserts that there is at least one possible world
in which Madeleine puts her violin away. However, that first pass is far too
weak. There are an infinite number of worlds in which Madeleine puts her
violin away, and there are an infinite number of worlds in which she keeps
playing. (There are also other, less ‘normal’ worlds, in which she keeps her
violin out but starts using it as a coffee cup, or in which her violin sponta-
neously crumbles to dust.) As a second pass, (3) asserts that there is at least
one world in which Madeleine obeys the conductor’s actual-world rules for
the audition, and in which she puts her violin away.
Why did we say ‘the conductor’s actual-world rules’? This is what ensures
that (3) makes a contingent claim. In the actual world, the conductor’s rules
happen not to exclude Madeleine from putting her violin away. (Maybe the
rules include the proposition that each auditioner plays one piece and two
scales, and Madeleine has already played those things.) But it could easily
have been the case that the conductor demanded that she keep playing (if,
for example, the rules contained the proposition that auditioners play two
pieces and four scales). It is a general feature of modal elements that they
use some facts about the evaluation world to narrow down the set of worlds
being quantified over.
How did we know that (3) was a statement about rules (rather than about
abilities, or wishes, or what is possible given the available evidence)? And
given that it is about rules, how did we know that it expressed permis-
sion, rather than obligation? The former question concerns modal flavour,
and the latter concerns modal force. With English modal auxiliaries, modal
flavour questions are primarily (but not exclusively) settled pragmatically,
while modal force questions are primarily settled lexically.
The modal flavour of can in (3) is deontic. We know this because of the
discourse context, as the conductor is in a position of power and is likely to
make pronouncements about rules. In the alternative context in (4), can has
a different modal flavour: it imparts information about ability.
(4) Context: Madeleine is three and has recently started violin lessons. At first, she
wasn’t able to put her violin away in its case because she wasn’t strong enough
to loosen the bow. But she has grown stronger.
Madeleine can put her violin away now.
The modal flavour of can is not completely determined by context; it is also

lexically restricted. Can cannot have an epistemic flavour: it cannot assert
something about the worlds compatible with a body of evidence.6 Might, on

the other hand, can easily have this interpretation, as shown in (5b).
(5) Context: A homicide detective is summarizing the content of a report which

presents the findings of the investigative team.
a. #According to the report, Gerry can be the murderer.
b. According to the report, Gerry might be the murderer.
Non-auxiliary modal elements frequently lexically encode modal flavour.

For example, the attitude verb want encodes bouletic modality (to do with
desires), and the adverb maybe is unambiguously epistemic.
Turning to modal force, our example modal can is lexically specified to
existentially quantify over worlds. It contrasts with universally quantifying
modals like must or have to, as shown in (6).
(6) Context: A violin audition. The conductor says:

a. Madeleine can put her violin away now.
b. Madeleine must/has to put her violin away now.
Example (6b) expresses obligation, rather than permission; it says (roughly)

that in all worlds in which Madeleine obeys the conductor’s actual-world
rules, she puts her violin away.7 In other words, there is no way to obey the
actual rules without putting the violin away. But there is a pragmatic ele-
ment to modal force, too: notice that even (6a) probably leads, in the audi-
tion context, to Madeleine putting her violin away. Thus, in certain contexts
can-sentences can pragmatically function almost like obligations. However,
the obligation is cancellable in (6a), but not in (6b), as shown in (7). This
shows that must/have to semantically conveys obligation, but can does not.
(7) a. Madeleine can put her violin away now (but she can also keep
playing if she wants).
b. Madeleine must/has to put her violin away now (#but she can also
keep playing if she wants).
We have seen so far that the truth values of modal assertions depend on the
world in which they are uttered, because the set of worlds quantified over
is narrowed down by a set of propositions which are true in the evaluation
world. In a similar fashion, their truth values depend on the time at which
they are uttered. Suppose, for example, that the conductor’s actual-world
rules require that each auditioner play at least one piece and two scales. At
time t, Madeleine has played one piece and no scales, and five minutes later
at t , she has played two scales as well. Sentence (3) would be false at t, and
true at t , because at t Madeleine’s putting her violin away is not compatible
with the conductor’s actual rules, but at t , it is. Moreover, the propositions
6
Except when it appears under negation; see Palmer (2001, p. 103) and McCormack and Smith (2002,
p. 135), among others.
7
This is a simplification; see Section 18.2.1.
Modality 529
we use to narrow the worlds quantified over can change over time: the con-
ductor could change her rules from one minute to the next, or a person’s
abilities could change.
Just like modal flavour and modal force, modal–temporal interactions can
be lexically restricted. More often, however, they are influenced by tense or
aspect functional heads; this is especially easy to see in languages where
modals overtly inflect for these categories. The Dutch data in (8) show that
past marking on the modal auxiliary causes the modal to be evaluated at
some past time.
(8) a. we moeten winnen
we must.prs.1pl win.inf
‘We have to win.’
(In all worlds compatible with our obligations or goals at the
utterance time, we win.)
b. we moesten winnen
we must.pst.1pl win.inf
‘We had to win.’
(In all worlds compatible with our obligations or goals at some past
time, we win.)
c. we kunnen winnen
we can.prs.1pl win.inf
‘We are able to win.’
(In some world compatible with our abilities at the utterance time,
we win.)
d. we konden winnen
we can.pst.1pl win.inf
‘We were able to win.’
(In some world compatible with our abilities at some past time, we
win.)
(Hotze Rullmann, p.c.)
In the next section we turn to a more in-depth discussion of modal flavour.
18.2 Modal flavour
Our discussion so far has been completely informal and has also been sim-
plified. The first way in which it has been simplified is that we have bluntly
talked about modal flavour. However, one of Kratzer’s important proposals
is that modal flavour comprises two separate parameters – called ‘conversa-
tional backgrounds’ – a modal base and at least one ordering source. In Sec-
tion 18.2.1 we briefly review the motivation for this double relativity of
modals, and in Section 18.2.2 we discuss where the conversational back-
grounds come from. Section 18.2.3 presents some cross-linguistic data, and
Section 18.2.4 discusses ongoing debate about the status of the epistemic–
circumstantial division in conversational backgrounds.
18.2.1 Modal base and ordering source

What we have so far can be formalized as in (9).8 Modals are interpreted
relative to one conversational background, a function from worlds to
sets of propositions. By intersecting that set of propositions, we obtain
a set of worlds accessible from w, the worlds in which all the proposi-
tions are true.9 Necessity modals like must assert that the prejacent (the
core proposition embedded under the modal) is true in all the acces-
sible worlds. Possibility modals existentially quantify over the accessible
worlds.
(9) a. [[must]]w,g = λ f s, st,t .λp s,t .∀w [w ∈ ∩ f(w) → p(w ) = 1]

b. [[can]]w,g = λ f s, st,t .λp s,t .∃w [w ∈ ∩ f(w) & p(w ) = 1]
Kratzer shows conclusively that such an analysis is inadequate. The

first problem is that sometimes the propositions given by the conversa-
tional background are inconsistent, in which case all necessity claims come
out trivially true and all possibility claims come out trivially false. Take
example (10):
(10) Context: The conductor’s rules state that auditioners play at least one piece
and two scales, they play no more than five scales, and they play everything
from memory. When Jane auditions, she uses sheet music rather than playing
from memory. Now it’s Madeleine’s turn and she has played one piece and two
scales.
a. Madeleine must stop playing now.
b. Madeleine can stop playing now.
We want to predict that (10a) is false and (10b) is true. The simple system so
far predicts the reverse. The problem is that the conversational background
lumps together propositions describing the rules with propositions describ-
ing Jane’s actions, and the resulting set of propositions is inconsistent (there
are no worlds in which all the rules are obeyed and in which Jane disobeyed
the rules). Since there are no worlds to quantify over, the universal quantifi-
cation in (10a) is trivially true, and the existential quantification in (10b) is
false.
A second problem with the simple system is that it has no way to deal
with graded modality, as in (11).
(11) a. Michl is probably the murderer.

b. There is a good possibility that Michl is the murderer.
c. There is a slight possibility that Michl is the murderer.
d. Michl is more likely to be the murderer than Jakl.
(Kratzer, 1991b, p. 643)
8
I assume a basic familiarity with semantic formalism, approximately to the level of Heim and Kratzer (1998).
9
Propositions are sets of worlds (the worlds in which the proposition is true), so a set of propositions is a set
of sets of worlds. The intersection of those sets of worlds gives the set of worlds in which all the
propositions are true.
Modality 531
An analysis of these examples requires more than just checking whether

Michl is the murderer in at least one accessible world. We need to be able to
rank possibilities using a notion of normalcy or stereotypicality: given the
facts, some things are more expected than others.
A third problem is that the simple system predicts that (12a) asymmet-
rically entails (12b). If someone climbed Mount Toby in all worlds com-
patible with the actual-world evidence, then they climbed it in the actual
world (since the actual world has to be compatible with the actual-world
evidence).10
(12) a. She must have climbed Mount Toby.

b. She climbed Mount Toby.
(Kratzer, 1991b, p. 645)
The solution to all these problems is that modals are interpreted relative
to two conversational backgrounds: a modal base and at least one ordering
source. The modal base assigns to each world a set of propositions which are
true in that world, and the ordering source assigns to each world a (possibly
inconsistent) set of propositions representing, for example, norms, ideals,
laws, desires. The set of worlds in which all the modal base propositions are
true is then ordered according to how many of the ordering source propo-
sitions are true in each world, and the modal quantifies over only the most
ideal worlds as identified by the ordering source.
The ordering determined by a set of propositions A is defined in (13): w is
at least as close as z to the ideal determined by A if and only if all proposi-
tions in A that are true in z are also true in w.
(13) For all worlds w and z ∈ W : w ≤A z iff {p : p ∈ A and z ∈ p}

⊆ {p : p ∈ A and w ∈ p}
(Kratzer, 2012, p. 39)
If we assume for the purposes of simplicity that there is always a world (or
set of worlds) which comes closer to the ideal than any other worlds (i.e., we
adopt the Limit Assumption), we can define an operator which selects the
‘best’ worlds, and then have the modals quantify only over these. Examples
(14) and (15) draw on von Fintel and Gillies (2011, p. 61) and Portner (2009,
p. 67); f is the modal base and h is the ordering source. (See Lewis, 1973 and
Stalnaker, 1984 on the Limit Assumption, and for modal definitions which
do not adopt it, see Kratzer, 1981, p. 48 and Kratzer, 1991b, p. 644.)
(14) For a given order ≤A on worlds:

∀X ⊆ W[BESTA (X) = {w ∈ X : ¬∃w ∈ X[w ≤A w]}]
(15) a. [[must]]w,g = λ f s, st,t .

λh s, st,t .λp s,t .∀w ∈ BESTh(w) (∩ f(w)) : p(w) = 1
w,g
b. [[can]] = λ f s, st, t .
λh s, st,t .λp s,t .∃w ∈ BESTh(w) (∩ f(w))) : p(w) = 1
10
Von Fintel and Gillies (2010) argue that (12a) does entail (12b); see Section 18.3.1 below.
Here is an example of how the modal base and the ordering source interact
to give the right results.
(16) Context: As in (10), except that Madeleine has played one piece and five scales.
Madeleine must stop playing now.
Example (16) asserts that in all worlds which are compatible with the facts
about the audition and which are best according to the conductor’s actual-
world rules, Madeleine stops playing. This is the right result. The set of
worlds compatible with the facts includes only worlds where one of the con-
ductor’s rules has already been broken. But that’s okay – we just find the
best we can. Out of the worlds where Jane has already broken the rules, the
best ones are those where Madeleine obeys the rules. And in all of those, she
stops playing now.
As hinted already, all modal bases are realistic: they assign to any world
a set of propositions which are true in that world. In other words, for
any world w and modal base f, w is a member of ∩ f(w) (see, e.g., Kratzer,
2012, p. 55). It is often assumed that there are two types of realistic conver-
sational background: epistemic, and circumstantial/root. The former pro-
duces sets of propositions representing the available evidence; the latter
produces sets of propositions representing relevant facts about the circum-
stances. (In Section 18.2.4, I discuss this dichotomy further, and the difficul-
ties with formally distinguishing the two types.) If a conversational back-
ground is non-realistic, it must be an ordering source.11 Common examples
of flavours of ordering source are deontic (dealing with rules), bouletic (deal-
ing with desires), teleological (dealing with goals), and stereotypical (dealing
with normalcy). The first three of these operate on circumstantial modal
bases; Portner (2009) groups them under the term priority modals. See Palmer
(2001) on the different categories of modal flavour which are made use of
cross-linguistically.
18.2.2 Where do conversational backgrounds come from?

One of Kratzer’s important insights is that conversational backgrounds can
be restricted by things external to the semantics of the modal itself. For
a start, they can be restricted via overt adverbial phrases, as shown in
(17).
(17) a. According to the audition rules, Madeleine can put her violin
away now.
b. Physically, Madeleine can put her violin away now.
c. Given the information in the report, Gerry might be the
murderer.
11
This is because a non-realistic conversational background may contain an inconsistent set of propositions,
and using this as a modal base would lead to the triviality problems discussed above.
Modality 533
If there is no overt restriction on conversational background, it can be pro-

vided by the context of use.12 When uttered by a conductor during a violin
audition, can is likely to be interpreted deontically, while in a different con-
text, can can impart information about abilities. This context-dependency
can be modelled in various ways. In the system I adopted for concreteness
in (15), the modal base and ordering source are given by variables present
in the syntax, which receive their value from the contextually given assign-
ment function.
Given that modals can receive their conversational backgrounds from the
context of use, Kratzer argues that it is not necessary to postulate lexical
ambiguity for modals which allow varying flavours of conversational back-
ground. For example, we do not need can1 and can2 in the lexicon, where
can1 is deontic and can2 is an ability modal.
Importantly, Kratzer does not claim that conversational backgrounds are
restricted exclusively by context or by overt adverbial phrases. There is no
reason why conversational backgrounds should not also be linguistically
restricted by the modal itself, and Kratzer (1981, 1991b, 2012) gives examples
of German modals which are compatible only with certain types of conver-
sational background. We already saw that affirmative can may not be inter-
preted epistemically (see Portner, 2009, p. 55, for further details) and that
non-auxiliary modal elements routinely place lexical restrictions on conver-
sational background.
Interestingly, the fact that modals may have lexically restricted conversa-
tional backgrounds weakens the argument against an ambiguity analysis. It
is still conceptually preferable to avoid having can1 and can2 , if possible. But
if there are languages which have distinct lexical items for these two mean-
ings, there might in fact be some hidden ambiguities in languages which
use just one lexical item.
Recent work on the evidential properties of epistemic modals also sug-
gests that we may need to contemplate ambiguity for English modal aux-
iliaries. Von Fintel and Gillies (2010) and Kratzer (2009, 2012), among oth-
ers, argue that at least some epistemic modals, including must, are lexi-
cally restricted to relying on indirect evidence for the embedded proposi-
tion. One simple way to implement such a restriction only for epistemic uses
would involve ambiguity or polysemy between epistemic and non-epistemic
must.13,14
12
Even in the presence of an overt adverbial phrase, the conversational background can be given by the
context, as pointed out by Nauze (2008, p. 157). The modal in (i) has an epistemic interpretation, in spite
of the deontic adverbial.
Context: We are discussing the upcoming trial of the suspected criminal, Jockl. The trial has not yet
beg[u]n and we wonder what the outcome will be.
(i) In view of what the law provides, Jockl may be executed.
(Nauze, 2008, p. 157)
13
Nauze (2008, p. 153), for example, argues for a polysemy account.
14
An argument which is sometimes advanced against an ambiguity analysis is that there are as many
different particular conversational backgrounds as there are contexts and that this multiplicity of meanings
A final option for conversational backgrounds is that modals might

acquire some of their meaning restrictions through their position in the
syntactic structure. A recent proponent of this approach is Hacquard (2006,
2009, 2010), who spells out an analysis in which the structural height of
a modal indirectly restricts its conversational background. Hacquard pro-
poses that modals are relativized to events rather than worlds, and their
interpretation depends on features of the relevant event, such as its run-
time or participants. High-scoping modals in matrix clauses depend on the
speech event, and thus on the time of utterance and the speaker. Usually,
high modals receive epistemic interpretations, because epistemic modal
bases require an information state, which is usually provided by the speaker
of the speech event.15 Circumstantial modal bases, on the other hand,
depend on the circumstances of an event (e.g., its location, time, partici-
pants). As such they are VP-event-relative and will tend to scope low. This
provides a principled explanation for the pervasive finding that epistemic
modals scope above non-epistemic ones (see, e.g., Jackendoff, 1972; McDow-
ell, 1987; Picallo, 1990; Brennan, 1993; Cinque, 1999; Butler, 2003; Barbiers,
2006).
The goals of Hacquard’s proposal are a unified analysis which avoids lex-
ical ambiguity, and an explanation for the correlation between syntactic
height and interpretation. Her analysis insightfully addresses one of the
most puzzling issues in the study of modals, the tension between the desire
for a unified analysis, and an explanation for distinct readings. One interest-
ing question is whether the unified analysis is tenable cross-linguistically.
Hacquard states that a major motivation for rejecting an ambiguity analy-
sis of modals is the generalization that, cross-linguistically, modal elements
express a range of flavours. She writes that ‘this multiplicity of modal mean-
ings is common enough cross-linguistically, and in languages from differ-
ent families, so as to make a lexical ambiguity account unlikely: it is highly
improbable that the same lexical accident should be found in language after
language’ (Hacquard, 2011, p. 1489). She also argues that a problem with pro-
posals which assign separate lexical entries for roots and epistemics is ‘the
fact that, cross-linguistically, they are expressed by the same lexical items’
(Hacquard, 2006, p. 114).
However, it is not obviously true that the pervasive cross-linguistic pat-
tern is for single modal elements to express a range of modal flavours.
According to van der Auwera and Ammann (2011), slightly more than
half of their sample of 207 languages have no modals which allow both
would be impossible to deal with lexically (see, e.g., Hacquard, 2011, p. 11). However, this argument is
weakened by the fact that modals in many languages do lexically encode overarching modal flavours like
‘deontic’ or ‘epistemic’. An ambiguity account combined with contextual dependence is therefore tenable,
and would be parallel to Partee’s (1988a) analysis of the quantifiers many and few. Partee claims that
many and few are each ambiguous between cardinal and proportional readings, but the specific
cardinalities or proportions they require are dependent on context.
15
Or by the holder of the attitude event, in the case of a modal embedded under an attitude verb.
Modality 535
epistemic and non-epistemic interpretations (i.e., no instances of what they

call ‘overlap’ between modal flavours). Moreover, the percentage of lan-
guages which lexically encode conversational background may actually be
higher than this. As outlined by Davis et al. (2014) and Matthewson (2013b),
there are cases which van der Auwera and Ammann analyse as involving
overlap, but which really involve lexical distinctness. For example, van der
Auwera and Ammann argue that Kiowa has overlap between modal flavours
in half of its system (i.e., in either the possibility domain or the necessity
one). However, there is no evidence for overlap in the source cited (Watkins
and McKenzie, 1984, pp. 220–222), and Kiowa is actually a language where
epistemic and non-epistemic modality are expressed via distinct morpho-
syntactic strategies (Andrew McKenzie, p.c.). Epistemic modals are typically
adverbs, deontic possibility is expressed using the irrealis, and ability or
the lack of it is expressed by incorporating auxiliaries, inflection, or out-of-
control marking.16 Similarly, Mandarin is classified by van der Auwera and
Ammann as a language which standardly allows overlap, but this appears
to be incorrect: the overwhelming pattern in Mandarin is for modals to lex-
ically restrict modal flavour (Liu, 2013).
In short, our current cross-linguistic understanding allows us to conclude
only that some modals in some languages allow variable modal flavour,
while some or all modals in other languages lexically specify modal flavour.
This is just what the Kratzerian analysis leads us to expect. Languages
can view the common meaning components (quantification over possible
worlds) as criterial for lexical division, or they can focus on the distinction
between different modal flavours. As a final comment, a requirement for
extending Hacquard’s analysis to a universal proposal would be evidence
that the syntactic differences between different modal interpretations show
up in languages from different families. As shown by Hacquard and others,
the tests required to establish syntactic differences are often quite subtle,
and they have not been done for the majority of the world’s languages. This
is an area where further research is needed.
In the next section, we will get a taste of a language which lexically
encodes conversational background throughout its modal system.
18.2.3 Modal flavour in Gitksan

Gitksan is an endangered and understudied Tsimshianic language spoken in
northern British Columbia, Canada.17 Its core modal system is summarized
in Table 18.1 (Matthewson, 2013a; see also Peterson, 2010).
We see that there is no overlap between epistemic and circumstan-
tial interpretations. The former are strictly expressed by means of the
16
These different morpho-syntactic strategies may turn out to support the syntax–meaning correlations
argued for by Hacquard, but they would not support the claim that different modal flavours are produced
by the syntax in combination with a maximally general modal lexical entry.
17
ISO 639-3 code ‘git’. The term Gitksan is conventionally used to cover that part of the Nass-Gitksan dialect
continuum which stretches from Git-anyaaw (Kitwancool) to Ansbayaxw (Kispiox).
Table 18.1 Gitksan modal system
possibility (weak) necessity

plain da’akhlxw
circumstantial sgi
deontic anook
plain ima(’a)a
epistemic
reportative gat
a This element has various pronunciations, partly dependent on dialect: either ima, imaa,
or ima’.
second-position clitics ima(’a) and gat, the latter strictly by the clause-initial
verbs or predicative particles da’akhlxw, anook, and sgi. Examples are given
in (18)–(19). The reportative gat in (18b) is not necessarily translated into
English using a modal; however, Peterson (2010) argues that it has epistemic
modal semantics. A more revealing translation of (18b) would be ‘Given
what I heard, the berries might/must be ripe.’
(18) a. Context: There was a bad can of fish; everyone at the dinner got sick.
yugw=ima’=hl nee=dii am=hl hon=hl gup-diit
ipfv=epis=cn neg=cntr good=cn fish=cn eat-3pl.ii
‘The fish they ate must’ve been bad.’
(Matthewson, 2013a, p. 360, adapted from Peterson, 2010, p. 162)
b. Context: Your brother told you the berries are ripe now. Later, you tell me
the berries are ripe, based on what your brother told you.
hlaa mukw-t=gat=hl maa’y
incept ripe-3sg.ii=report=cn berries
‘The berries are ripe (I heard).’
(Matthewson, 2013a, p. 362)
(19) a. da’akhlxw-i-s Henry dim jam-t

circ.possib-tr-pn Henry prosp cook-3sg.ii
‘Henry is able to cook.’ / ‘Henry was able to cook.’
b. anook-xw(=hl) dim ha’w-s Savanna (k’yoots)
deon.possib-val(=cn) prosp go.home-pn Savanna (yesterday)
‘Savanna was allowed to go home (yesterday).’
c. sgi dim (ap) ha’w-s Lisa
circ.necess prosp (verum) go.home-pn Lisa
‘Lisa should/must go home.’ / ‘Lisa should (have) gone home.’
The split between epistemic and circumstantial readings is absolute in

Gitksan: the modals can be used with the flavours illustrated here, but they
cannot be used with the opposite ones. For example, in (20) the context
Modality 537
is (pure) circumstantial: the issue is not whether it is compatible with

the evidence that berries are growing here, but whether it is compatible
with the circumstances that berries could grow here (this is adapted from
Kratzer’s hydrangea example, 1991b, p. 646). We see that epistemic ima(’a) is
infelicitous.
(20) Context: You’re up in the Suskwa and notice a burnt patch of forest. You know
that huckleberries typically take seed in burnt alpine areas.
a. da’akhlxw=hl dim limxs=hl maa’y go’osun
circ=cn prosp grow.pl=cn berries loc.here
‘Berries might/can/are able to grow here.’
b. #limxs=ima=hl maa’y go’osun
grow.pl=epis=cn berries loc.here
‘Berries might be growing here.’
(Peterson, 2010, p. 158)
Lexical distinctions between modal flavours are pervasive in many lan-

guages, as mentioned in Section 18.2.2, and as shown recently in formal
work on St’át’imcets (Lillooet Salish: Matthewson et al., 2007; Rullmann
et al., 2008; Davis et al., 2009), Javanese (Austronesian: Vander Klok, 2008,
2012), Blackfoot (Algonquian: Reis Silva, 2009), Kwak’wala (Wakashan: Men-
zies, 2010), Nez Perce (Penutian: Deal, 2011), and Nsyilxcen (Okanagan
Salish: Menzies, 2013).
Even though many languages lexically distinguish epistemic from circum-
stantial modal bases, it turns out to be not so easy to distinguish the two
types theoretically. I turn to this issue in the next section.
18.2.4 Epistemic vs. circumstantial: the issues

Epistemic and circumstantial interpretations should be easy to tell apart.
Kratzer (2012, p. 52) gives the German pair in (21) to illustrate (English trans-
lations added). Example (21a) is circumstantial and asserts that Kathl has
the ability to make a pound of Quark out of this can of milk. Example (21b)
is epistemic and asserts that it is compatible with the evidence that Kathl
might actually do that. If Kathl can, but surely will not, make a pound of
Quark out of this milk, (a) is true and (b) is false.
(21) a. aus dieser Kanne Milch kann die Kathl ein Pfund
from this can of milk can the Kathl one pound of
Quark machen
cottage cheese make
‘Kathl can make a pound of Quark out of this can of milk.’
b. es kann sein, dass die Kathl aus dieser Kanne Milch ein Pfund
it may be that the Kathl from this can of milk one pound
Quark macht
cottage cheese makes
‘Kathl might make a pound of Quark out of this can of milk.’
The theoretical problem is that both types of modal base depend on facts
about the world of evaluation, and as such, there is no formal way to dis-
tinguish the two. As Nauze (2008, p. 155) points out, in many contexts the
very same set of propositions can serve as either an epistemic modal base or
a circumstantial one. Kratzer (2012, p. 24) concurs and writes that ‘[i]t now
seems to me a hopeless enterprise to try to characterize formal objects like
conversational backgrounds as “circumstantial” versus “epistemic”’.
To see the problem, imagine that the relevant circumstances for a partic-
ular modal assertion are identical to the relevant evidence we have (that is,
we know everything that is relevant). In this case, a circumstantial and an
epistemic modal base will contain the exact same set of propositions. Nauze
gives the example in (22), which involves similar modal flavours to (21) but
has a context in which the two modal bases (which we want to classify as
circumstantial and epistemic respectively) contain identical propositions.18
(22) Context: It is 2.50 pm on Saturday and we all know that France is playing
England at 3.00 pm in the Six Nations rugby tournament. Furthermore we
know that Fabrice is home (and has a television receiving the game).
a. Fabrice can watch the game.
b. Fabrice might watch the game.
(Nauze, 2008, pp. 154–155)
In recent work, Kratzer (2012, 2013) (building on work by Arregui, 2005,

2007, 2009; Hacquard, 2006, 2010) develops an alternative approach to
restricting the domains over which modals quantify. The new approach
includes a substantive rethinking of the epistemic–circumstantial divide.19
The major division Kratzer proposes is between factual and content modes
of projecting conversational backgrounds. Factual mode functions return a
set of worlds which all contain counterparts of some actual-world situation
or body of evidence.20 For example, (23) contains a factual modal. The claim
is that in all worlds in which the same rumour exists as in the actual world,
and in which the rumour bears the same relation to reality as it does in the
actual world, Roger was elected chief. The sentence commits the speaker to
the belief that Roger was elected in the actual world (modulo the wiggle
room provided by a stereotypical ordering source).
(23) Given the rumour, Roger must have been elected chief (#but he
actually wasn’t).
18
The problem cannot be solved by using different types of ordering sources. Nauze (2008, p. 155) also
shows that deontic and stereotypical ordering sources can contain identical sets of propositions in the
same context.
19
Another facet of the new approach is the idea that a modal’s domain is projected from an anchor. Anchors
may be individuals, events, situations, or bodies of evidence. Implicit domain-fixing functions map the
anchors to ordered sets of modal alternatives. Here I set the details of anchors aside, concentrating on the
issue of the epistemic–circumstantial divide.
20
On counterparts, see Lewis (1986).
Modality 539
Table 18.2 Traditional vs. new classifications of (modes of projection of)

conversational backgrounds
traditional classification circumstantial epistemic

new classification factual content
The content mode, on the other hand, provides functions which return
sets of worlds compatible with the propositional content of some actual-world
source of information. This is illustrated in (24). Here, the claim is that in
all worlds in which the content of the rumour is true, Roger was elected chief.
This allows the speaker to believe that the rumour was a lie.
(24) According to the rumour, Roger must have been elected chief
(but he actually wasn’t).
Notice that must behaves differently in (23) and (24), even though both
these uses would traditionally be classified as epistemic. The new classifi-
cation therefore splits the old class of epistemic modals. Content-mode epis-
temics are not based on the speaker’s knowledge, or necessarily on anyone’s
knowledge – the content of rumours and other similar sources of informa-
tion can be false.21
The traditional class of circumstantial modals, on the other hand, all use
the factual mode, as they all rely on some actual-world facts and quantify
over worlds in which counterparts of those facts hold. The relation between
the old and new classifications is schematized in Table 18.2.
The split between factual- and content-mode epistemics is lexicalized in
many languages. For example, the St’át’imcets reportative modal ku7 is fac-
tual, as shown by the fact that the St’át’imcets sentence on which (23)–(24)
are based does not allow the speaker to totally disbelieve the rumour (see
Matthewson et al., 2007). In contrast, the German reportative modal sollen is
a content-based modal and does allow the speaker to disbelieve the rumour
(Kratzer, 2012, p. 35).22 Other elements traditionally analysed as evidentials
also lexically encode the difference between factual- and content-mode con-
versational backgrounds. This is shown in (25)–(28) for St’át’imcets. The fac-
tual modal k’a requires the speaker to have indirect inferential evidence
for the prejacent, while the content modal lákw7a requires the speaker to
have sensory, non-visual evidence for the prejacent. Only lákw7a allows the
speaker to believe that the prejacent is false (Matthewson, 2011, 2012a).
21
Kratzer (2012) calls the two types of epistemic modals evidential vs. informational. This is a different use of
the term evidential from its traditional use, which picks out elements which encode information source
(Aikhenvald, 2004).
22
The Quechua and Cheyenne reportatives also allow the speaker to disbelieve the report (Faller, 2002;
Murray, 2009). Faller (2002) and Murray (2009) do not analyse their respective reportatives as having
modal semantics, but Faller (2011) analyses the Quechua one as a content-mode (‘informational’) modal,
similar to German sollen.
(25) #wá7=k’a ku=mám’teq láku7 áltsq7=a, t’u7 nílh=a

be=infer det=walk deic outside=exis but foc=a
cwílh=t’u7 ti=sk’éxem=a wa7 qan’ím-ens-an
after.all=just det=wind=exis ipfv hear-dir-1sg.erg
‘Someone might/must have been walking outside, but it was the
wind.’
(26) wa7 lákw7a ku=mám’teq láku7 áltsq7=a, t’u7 nílh=a
be sens.non.vis det=walk deic outside=exis but foc=a
cwílh=t’u7 ti=sk’éxem=a wa7 qan’ím-ens-an
after.all=just det=wind=exis ipfv hear-dir-1sg.erg
‘It sounded like someone was walking outside, but it was the wind.’
(27) #t’ec=k’a=t’u7 ku=páoy, t’u7 áoz=t’u7 kw=a=s áma
tasty=infer=just det=pie but neg=just det=ipfv=3poss good
#‘The pie might/must have been good, but it wasn’t good.’
(28) Context: It smelled as if the pie was good, but there was too much salt so it was
actually horrible.
t’éc=t’u7 lákw7a ku=páoy, t’u7 áoz=t’u7 kw=a=s
sweet=just sens.non.vis det=pie but neg=just det=ipfv=3poss
áma
good
‘The pie seemed good, but it wasn’t good.’
This analysis requires us to assume, following ideas found in McCready
(2010a), that the sensory evidence invoked by lákw7a has propositional con-
tent.23 So, the factual mode (27) might assert that the pie was good in all
worlds in which there was a counterpart of the actual-world smell, and in
which the smell bore the same relation to the taste of the pie as it did in the
actual world. (Anyone who asserts that cannot simultaneously assert that
the pie was bad.) The content mode (28), on the other hand, might assert
that the pie was good in all worlds compatible with the content of the
sensory-evidence proposition that the pie smelled good. Since my sensory
evidence might have been tricking me, it is possible that the pie was bad.
The fact that the St’át’imcets evidentials k’a and lákw7a lexicalize the split
between factual and content modes of conversational background projec-
tion has two consequences. The first is that one diagnostic which is some-
times used to argue for a non-modal analysis of evidentials collapses. The
diagnostic in question relies on whether an utterance containing the evi-
dential is compatible with the speaker’s knowing or believing that the pre-
jacent proposition is false (Faller, 2002; Murray, 2009). The story is that since
a modal asserts that its prejacent proposition is possibly or necessarily true,
23
Thanks to an anonymous reviewer for raising this issue. McCready uses the example of a bloody knife
hidden in a cupboard: the knife itself is just a knife, but the proposition that the knife is there is what serves
as evidence.
Modality 541
Table 18.3 A three-way categorization of conversational backgrounds
factual–circumstantial factual–evidential content–evidential

can might
English
must
ka (deon), k’a (infer), lákw7a
St’át’imcets
ka-. . . -a (circ) ku7 (report) (sens.non.vis)
an evidential which allows statements of the form ‘evidential-ϕ, but not

ϕ’ cannot have modal semantics. However, we just saw that content-mode
modals allow statements of this form, because they quantify over worlds
compatible with the content of some body of evidence, and do not place
any restriction on whether the prejacent is true in the actual world. So the
diagnostic of deniability is not sufficient to rule out a modal analysis of evi-
dential elements (Matthewson, 2012a).
The second consequence of the k’a/lákw7a situation is that it supports a
recently emerging idea about how to unify the old class of epistemic modals,
even under the new classification into factual and content modes. The idea
is that epistemics are modals which rely on indirect evidence about proposi-
tions and therefore that epistemic modals fall into the class of evidential
elements. As von Fintel and Gillies (2011, p. 113) put it, epistemic modals
have a ‘resistance to plain facts but sensitivity to stores of information’. The
‘stores of information’ can include ships’ logs, interview notes, or comput-
ers but always constitute indirect evidence of actual-world events. The idea
that epistemic modals always rely on indirect evidence is fleshed out in von
Fintel and Gillies (2007, 2010) and Kratzer (2012). In general, we are seeing
a convergence of research on epistemic modality in languages like English
with cross-linguistic research on evidentiality which, since Izvorski (1997),
has been exploring and debating the idea that at least some evidentials are
epistemic modals (Garrett, 2001; McCready and Asher, 2006; Matthewson
et al., 2007; McCready and Ogata, 2007; Speas, 2008; Peterson, 2010; Faller,
2011; Lee, 2011, among many others).
The picture I have been sketching leads to a classification as in Table 18.3,
where a three-way split is generated. The factual–content division is a differ-
ence in modes of projection (whether the modal quantifies over worlds con-
taining a counterpart of some situation or body of evidence, or over worlds
compatible with the propositional content of some body of evidence). The
circumstantial–evidential division is a difference in whether information
source is encoded.
We see that English lexically encodes the circumstantial–evidential split,
as well as having some modals which allow all three modes of projection.
St’át’imcets encodes the full three-way split. We also predict that languages
may have modals which are lexicalized to cover all kinds of factual inter-
pretations, including traditional circumstantial interpretations as well as
factual–evidential ones. Whether this prediction is upheld is a topic for
future research.
Before leaving the issue of modal flavour, I will mention one more issue of
ongoing debate in the area of epistemic modality. Suppose we could agree
that epistemic modals depend on stores of information which constitute
indirect evidence for actual-world events. There is then still debate about
whether epistemic modals necessarily rely on an agent who is the holder
of that store of information. If an agent is involved, how is that relativism
to an agent capturable? The literature on this topic is large and interesting,
and consensus has not yet been reached (for relevant discussion, see Garrett,
2001; Hacquard, 2006, 2010; von Fintel and Gillies, 2007, 2008, 2011; Yalcin,
2007; Stephenson, 2008; Waldie, 2012, among others).
18.3 Modal force
18.3.1 Issues in modal force

So far we have looked mainly at necessity and possibility modals, but a pic-
ture of modals as encoding either universal or existential quantification
over worlds is too simplistic in several important ways. First, there is the
existence of graded modality, as we see in (29) (repeated from (11)).
(29) a. Michl is probably the murderer.

b. There is a good possibility that Michl is the murderer.
c. There is a slight possibility that Michl is the murderer.
d. Michl is more likely to be the murderer than Jakl.
(Kratzer, 1991b, p. 643)
Kratzer (1981, 1991b) argues that the ordering source provides a way to
account for graded modal force. Kratzer defines various graded modal
notions such as ‘slight possibility’, ‘good possibility’, and ‘better possibility’;
examples are given in (30)–(31) (see discussion in Portner, 2009, pp. 69–71).
(30) p is a good possibility in w with respect to a modal base f and an

ordering source g iff:
∃u ∈ ∩ f(w)(∀v ∈ ∩ f(w)(v ≤g(w) u → v ∈ p))
(Kratzer, 1991b, p. 644)
(31) p is at least as good a possibility as q in w with respect to f and g iff:

¬∃u ∈ ∩ f(w)(u ∈ q − p & ∀v ∈ ∩ f(w)(v ∈ p − q → u ≤g(w) v ))
(Kratzer, 2012, p. 41)
A proposition p is a good possibility if there is some world u, such that

all higher-ranked worlds than u are p-worlds. To see whether p is at least
as good a possibility as q, we look at the worlds p and q do not have in
Modality 543
common (i.e., the worlds in q − p and in p − q). p is at least as good a possibil-

ity as q if there is no world in q − p that is ranked higher than all worlds in
p − q.
Recent work on graded modality has debated the extent to which the
ordering-source-based analysis is sufficient to capture all the relevant facts.
One issue is the existence of overtly quantifiable degrees of probability, as
in (32). These are not obviously handlable in a standard Kratzerian analysis
(although see Kratzer, 2012, pp. 42f.).
(32) a. It was twice as likely to rain as it was to snow.

(Swanson, 2008, p. 1202)
b. There is a 60% probability that it is raining.
(Portner, 2009, p. 73)
c. It is 95% certain that Jorge will win the race.
(Lassiter, 2011a, p. 198)
Researchers such as Halpern (1997, 2003), Swanson (2006, 2008), Port-

ner (2009), Yalcin (2007, 2010), Lassiter (2011a,b), Kratzer (2012) have
been exploring the idea that gradable modal elements like probable/
probably/probability should be analysed as probability operators, and/or that
their analysis should adopt insights from the study of gradable adjectives
(e.g., Kennedy and McNally, 2005a). A gradable adjectives analysis would
mean that these elements introduce a scale which is an ordered set of
degrees of possibility or probability.
A related issue to that of graded modality, which happens to be lexicalized
in English modal auxiliaries, is the distinction between strong and weak
necessity, illustrated in (33).
(33) a. After using the bathroom, everybody ought to wash their hands;
employees have to.
(von Fintel and Gillies, 2008, p. 116)
b. After using the bathroom, everybody should wash their hands;
employees must.
There are several different approaches to weak necessity modals (see

Portner, 2009, pp. 79ff. and Rubinstein, 2012, for recent summary and dis-
cussion). One family of approaches adapts the probability semantics just dis-
cussed for epistemic elements; it views weak necessity modals as assigning a
greater likelihood of success or utility to their prejacent than to other alter-
natives (see, e.g., Goble, 1996; Finlay, 2009, 2010; Lassiter, 2011b).
Another family of approaches to weak necessity involves domain restric-
tion and is based on the fact that universally quantifying over a smaller
set of worlds gives rise to a weaker claim.24 Following an intuition by
24
For a third set of approaches which rely on pragmatic differences between strong and weak necessity
modals (e.g., a presupposition on must that is missing on should, or evidential differences between the
two), see Ninan (2005), Copley (2006), Swanson (2008).
Sloman (1970), von Fintel and Gillies (2008) argue that ought universally
quantifies over a domain which is narrowed down by an additional restric-
tion, over and above the restrictions on the domain of must. In (34a), for
example, the claim is that in all the worlds where the relevant facts are the
same as in the actual world and in which you meet your goal of getting to
Ashfield, you take Route 2. There are no other options for meeting your goal.
In (34b), the claim is weaker: in all worlds in which you satisfy a secondary
goal (such as avoiding heavy traffic) you take Route 2.
(34) a. To go to Ashfield, you have to/must take Route 2.

b. To go to Ashfield, you ought to take Route 2.
Von Fintel and Iatridou argue that the secondary goal is due to a second
ordering source, whose presence can be signalled by counterfactual mor-
phology in a range of languages. This approach uses only tools provided by
a standard analysis of modals, but/and it raises interesting questions about
ordering sources. One issue discussed by Rubinstein (2012) is how we tell
which ordering source is which – which is the ‘primary’ and which the ‘sec-
ondary’? Rubinstein proposes that weak necessity modals rely on ordering
sources which involve a departure from the collective commitments of par-
ticipants in the discourse. For example, suppose there are four ways to get
to Amherst from Cambridge. A speaker who utters (35) conveys that while
taking Route 9 is not a necessity given the collectively agreed-upon goal of
getting to Amherst, it is a necessity given some assumption that the speaker
presupposes is not collectively shared with the hearer (such as seeing nice
scenery, or stopping for lunch at a particular place on the way).
(35) To go to Amherst, you ought to take Route 9.

(Rubinstein, 2012, p. 9)
Further questions raised by the ordering-source approach include the issue

of what prevents even a single ordering source from weakening a strong
necessity modal more than we would like. Similarly, what prevents a weak
necessity modal from becoming as weak as a possibility modal? And finally,
do epistemic strong necessity modals really even have one ordering source?
Both von Fintel and Iatridou (2008) and von Fintel and Gillies (2010) doubt
that they do. These authors disagree with one of the basic supposed empiri-
cal motivations for ordering sources, namely that (36a) is weaker than (36b)
because it involves a non-realistic, stereotypical ordering source.
(36) a. She must have climbed Mount Toby.

b. She climbed Mount Toby.
(Kratzer, 1991b, p. 645)
According to von Fintel and Gillies, epistemic must does not signal weaken-
ing with respect to the assertion of the plain prejacent. Instead, it signals
Modality 545
indirect evidentiality (as discussed in the previous section), and this leads to
a perceived weakening. One thing to note about this idea is that the evi-
dentiality claim is independent of the no-ordering-source claim. It could
conceivably be correct that must requires indirect evidence, but false that
it makes no use of an ordering source.
A final strand of current research in the area of modal force is modals
without duals. These are modals which do not come in necessity–possibility
pairs, but can be used in contexts supporting either necessity or possibility
claims. These have been most extensively discussed for languages other than
English (Rullmann et al., 2008; Davis et al., 2009; Peterson, 2010; Deal, 2011).
In the next subsection, I discuss two examples of such languages, Gitksan
and Nez Perce.
18.3.2 Modal force in Gitksan and Nez Perce

The Gitksan epistemic modals ima(’a) and gat, seen above in Section 18.2.3,
differ from English modal auxiliaries in that they do not appear to be spe-
cialized for a particular modal force. Instead of forming a paradigmatic
universal–existential pair, they contrast with each other only in information
source. Ima(’a) and gat are compatible with situations which license neces-
sity assertions, and with situations which license possibility assertions. Con-
sequently, acceptable translations into English differ widely, as shown in
(37).
(37) Context: You’re wondering where your friend is. You notice his rod and tackle
box are not in their usual place.
yugw=ima=hl dim ixw-t
ipfv=epis=cn prosp fish.with.line-3
‘He might be going fishing.’ / ‘He must be going fishing.’ / ‘He’s
probably going fishing.’ / ‘He’s likely going fishing.’ / ‘He could be
going fishing.’ / ‘Maybe/perhaps he’s going fishing.’
(Peterson, 2010, p. 161)
Note that in the example above the spelling of ixw has been corrected
(cf. Rigsby, 1986, p. 170).
The claim that ima(’a) is not semantically a necessity modal is supported
by data such as in (38). In this discourse context, must would be infelicitous
in English.
(38) Context: You thought your friend was fishing. But you see his rod and tackle
box are still at his house. You really don’t know if he’s fishing or not.
yugw=imaa=hl da’awhl ixw-t oo ligi nee=yimaa=dii ixw-t
ipfv=epis=cn then fish-311 or indef neg=epis=cntr fish-311
‘Maybe he’s fishing, maybe he’s not fishing.’
Peterson (2010) analyses ima(’a) as a possibility modal which can be strength-

ened via an ordering source. This idea is the inverse of the domain-
restriction approach to weak necessity discussed in the previous section,
and relies on the idea that while narrowing the domain of a universal quan-
tifier causes weakening, narrowing the domain of an existential quantifier
causes strengthening.
Superficially similar data, in the non-epistemic realm, are presented for
Nez Perce by Deal (2011). As shown in (39), the circumstantial modal o’qa
is translatable as, and at first glance apparently interpretable as, either a
possibility or a necessity modal.
(39) Context: I am watching people clean out a cooler and throw away various
things.
hi-wqíi-cix-∅ ’iléx̂ni hipt ke yox̂ hi-pá-ap-o’qa
3subj-throw.away-ipfv.pl-pres a.lot food rel dem 3sbj-S.pl-eat-mod
(i) ‘They are throwing away a lot of food that they could eat.’
(ii) ‘They are throwing away a lot of food that they should eat.’
(Deal, 2011, p. 574)
Deal shows, however, that in downward-entailing environments, o’qa

behaves only as a possibility modal would be expected to behave. This is
shown in (40), with the modal inside the downward-entailing first argument
of the universal quantifier ’óykala ‘all’.
(40) Context: As in (39).

hi-wqíi-cix-∅ ’óykala hipt ke yox̂ hi-pá-ap-o’qa
3subj-throw.away-ipfv.pl-pres all food rel dem 3sbj-S.pl-eat-mod
(i) ‘They are throwing away all the food that they could eat. They
are throwing away all their food.’
(ii) #‘They are throwing away all the food that they should eat (but
keeping some junk food).’
(Deal, 2011, p. 574)
Although the English translation in (40-i) entails (the first part of) (40-ii),
the Nez Perce sentence is not a viable way to express the specific negated-
necessity meaning of (40-ii). That is, the Nez Perce sentence cannot be under-
stood as conveying that they are throwing away all the food they should eat,
but keeping some food they could eat (the junk food).
Deal analyses o’qa as a possibility modal, which is acceptable in non-
downward-entailing necessity contexts because there is no contrasting
necessity modal to induce a scalar implicature. In other words, Nez Perce’s
circumstantial modal system parallels what English’s nominal quantifier
system would look like if it possessed the existential quantifier some, but no
universal quantifier like all or every. In such a version of English, (41a) would
fail to implicate that not all the guests brought presents and therefore
would be acceptable in a context where all of the guests brought presents.
Modality 547
This is parallel to the felicitous use of o’qa in the necessity context in (39-ii).
The downward-entailing (41b) would (just like in real-life English) be unable
to express the negated-universal meaning: (41b) would be unable to express
the claim that it is false that all of the guests brought presents (but true
that some of them did). This parallels the infelicity of o’qa in the downward-
entailing necessity context in (40-ii).
(41) a. Some of the guests brought presents.

b. It is false that some of the guests brought presents.
Peterson’s analysis of Gitksan epistemic ima(’a) and Deal’s analysis of Nez

Perce circumstantial o’qa make the same predictions for non-downward-
entailing environments (that the modals can be used in both possibility
and necessity contexts), but different predictions for downward-entailing
contexts (ima(’a), but not o’qa, is predicted to be acceptable in necessity
contexts). Unfortunately, the predictions for downward-entailing environ-
ments are difficult or impossible to test for ima(’a), owing both to indepen-
dent features of Gitksan, and to inherent difficulties with placing epistemic
modals in downward-entailing environments in the first place. See Matthew-
son (2013a) for Gitksan, and Deal (2011, p. 566), who writes about Nez Perce
that ‘Owing to difficulties in embedding certain epistemic expressions in
downward-entailing contexts, the intricacies of the epistemic system are not
yet fully understood.’
There are other approaches to modals without duals, for example a
domain-restricted necessity analysis by Rullmann et al. (2008), and a degree-
modal analysis by Kratzer (2012).25 There are also yet more empirical config-
urations, which raise further analytical questions. For example, Nsyilxcen
(Okanagan Salish) possesses two epistemic modals, mat and cmay. Mat is felic-
itous in both possibility and necessity contexts, but cmay is felicitous only in
possibility contexts. This is shown in (42)–(43).
(42) Context (possibility): You know that Mary loves to go running and often goes
on runs randomly. She could also be at the store or at school. I ask you, where
is Mary?
a. Mary mat ac-qíc-əlx
Mary mod cust-run-lex
‘Mary might be running.’
b. Mary cmay ac-qíc-əlx
(Menzies, 2013, p. 2)
25
One of Kratzer’s motivations for a degree-modal analysis is that Rullmann et al. (2008), Peterson (2010),
and Deal (2011) fail to explain the absence of duals in the respective languages (Kratzer, 2013, pp.
184–185). While a degree-modal analysis may turn out to be preferable, I am not sure that the absence of
duals is something which necessarily requires explanation. Some modals have duals, some do not. Attitude
verbs do not seem to have duals, for example.
(43) Context (necessity): Mary runs every day to train for a marathon. She usually
runs at 6 pm on Tuesdays. Today is Tuesday and it’s 6 pm. I ask you, where is
Mary?
a. Mary mat ac-qíc-əlx
‘Mary must be running.’
b. #Mary cmay ac-qíc-əlx
(Menzies, 2013, p. 2)
We thus see that within a single language’s epistemic system, one modal
behaves as if it lacked a dual, while one behaves like a strict possibility
modal. This raises interesting questions for the idea that possibility modals
in languages like English are weak because of scalar implicatures induced
by the contrasting necessity modal. See Section 18.5 for discussion of an
emerging formal typology of modal force.
18.4 Modal–temporal interactions
The denotations we have so far for English modal auxiliaries, repeated in

(44), are dependent on both a modal base and an ordering source, but con-
tain no sensitivity to time.
(44) a. [[must]]w,g = λ f s, st,t .

λh s, st,t .λp s,t .∀w ∈ besth(w) (∩ f(w)) : p(w) = 1
w,g
b. [[can]] = λ f s, st,t .
λh s, st,t .λp s,t .∃w ∈ besth(w) (∩ f(w))) : p(w) = 1
However, as pointed out above and as discussed by many authors includ-

ing Condoravdi (2002) and Ippolito (2003), the truth conditions of modal
statements depend not only on a world of evaluation but also on a time of
evaluation. Conversational backgrounds must therefore be functions which
take times as one of their arguments. A time-dependent denotation for must
is given in (45).
(45) [[must]]w,g = λ f s, st,t .λh s, st,t .λP i, s,t .λt.

∀w ∈ besth(w) (∩ f(w, t)) : P(t)(w ) = 1
Simply introducing a time of evaluation is of course not the full story. As

Portner (2009, p. 223) observes, there are at least three ways in which the
temporal interpretation of modal sentences can be restricted: first, by inde-
pendent tense or aspect operators and their scope properties; second, by
temporal restrictions in the lexical entries of the modals themselves, and
third, by ‘general semantic or pragmatic principles which help determine
temporal meaning, but which are not tied to any particular grammatical
Modality 549
element’. Portner (2009, Chapter 5) gives a comprehensive overview of the

issues in modal–temporal interactions and the major available approaches.
In this section I will first outline an approach to modal–temporal interac-
tions which draws on and generalizes Condoravdi’s (2002) work, and then
discuss the contentious issue of whether epistemic modals can scope under
past tense. In Section 18.4.3 I present some cross-linguistic evidence for the
proposed approach to modal–temporal interactions.
18.4.1 Sketch of an approach to modal–temporal interactions

The discussion in this subsection is presented with respect to English data
unless otherwise noted, but the architecture of the framework is intended
to be cross-linguistically applicable. Following much work in the literature, I
assume that modals appear syntactically below tense,26 but above viewpoint
aspect. The reference time interval which is provided by tense saturates the
time argument in the lexical denotation of the modal (see (45) above). The
tense thus provides the evaluation time for the modal base, called the tem-
poral perspective by Condoravdi (2002). Viewpoint aspect restricts the rela-
tion between the temporal perspective and the time of the event described
by the prejacent clause; this relation is called the temporal orientation of
the modal by Condoravdi. Temporal perspective and temporal orientation
are illustrated in (46).
(46) a. Merlin might win the game.

temporal perspective: present (based on available evidence at
utterance time)
temporal orientation: future (event follows temporal perspective)
b. Merlin might have won the game.
temporal perspective: present (based on available evidence at
utterance time)
temporal orientation: past (event precedes temporal perspective)
Simple versions of perfective, imperfective, perfect, and prospective view-

point aspects are given in (47)–(50). Exactly one of perfective or imperfective
applies in each clause, and either perfect or prospective may optionally add
between the first aspect and the modal.27
26
More generally (to take into account tenseless languages), what is crucial is that the modal is provided with
a reference time in whatever way the language independently provides reference times to non-modal
clauses.
27
Some analyses, including that of Condoravdi (2002), do not include a prospective viewpoint aspect but
instead place an inherent future semantics in the modal itself. I am using prospective here by analogy to
languages which overtly mark future temporal orientation with prospective aspect markers, such as Gitksan
(see Section 18.4.3). Kratzer (2011b) argues that English possesses a null prospective aspect marker which
co-occurs with modals, and its absence can lead to actuality entailments. (For other work on the important
topic of actuality entailments, which unfortunately I do not have space to go into here, see Bhatt, 1999;
Hacquard, 2006, 2009; Mari and Martin, 2007, among others.)
(47) [[pfv]]g,w = λP l,st λtλw.∃e[P(e)(w) & τ (e) ⊆ t] (Kratzer, 1998a)
(48) [[ipfv]]g,w = λP l,st λtλw.∃e[P(e)(w) & t ⊆ τ (e)] (Kratzer, 1998a)
(49) [[prf]]g,w = λP l,st λtλw.∃t [t < t & P(t )(w)]
(50) [[prosp]]g,w = λP l,st λtλw.∃t [t ≤ t & P(t )(w)]
The schemas in (51) show how the various temporal operators apply. In each
case the tense provides the temporal perspective (TP), and aspect restricts the
temporal orientation (TO). Note that prospective aspect merely induces non-
pastness; depending on the Aktionsart of the predicate and the perfective/
imperfective distinction, prospective can derive either strict future orienta-
tion (with an eventive perfective verb) or ambiguously present/future orien-
tation (with a stative predicate and/or an imperfective).
(51) a. He might dance:

pres (might (prosp (pfv (λe.he-dance(e)))))
TP: present TO: future
b. He might be dancing:
pres(might (prosp (ipfv (λe.he-dance(e)))))
TP: present TO: present or future
c. He might have danced:
pres(might (prf (pfv (λe.he-dance(e)))))
TP: present TO: past
d. He might have been dancing:
pres(might (prf (ipfv (λe.he-dance(e)))))
The schemas so far all have present tense, therefore present temporal
perspective. As is well known, English modal auxiliaries do not synchron-
ically inflect for tense, but in languages in which they do (such as Dutch
or German), past tense morphology on the modal gives past temporal per-
spective. This also happens with English semi-modals such as have/had to or
is/was able to. Otherwise, past TP in English is marked by the perfect aux-
iliary have. In (51c)–(51d), have introduces a low-scoping perfect aspect and
gives past TO, but have can also induce past TP when it combines with
a modal which contains historical past tense morphology (such as might
or could). These two different roles for have underlie the famous ambigu-
ity of might have-constructions, pointed out by many including Condoravdi
(2002), Huddleston and Pullum (2002), Ippolito (2003), and illustrated in
(52)–(53).28
28
The details of how have manages compositionally to contribute past TP have been debated in the literature
and go beyond what we can discuss here (see Condoravdi, 2002; Ippolito, 2003; Arregui, 2005; Hacquard,
2006; Laca, 2008; Rullmann and Matthewson, 2012, for discussion).
Modality 551
(52) Q: Why is John looking happy?

A: I’m not sure, but he might have won the game.
pres(might(prf(pfv(λe.he-win(e)))))
(53) Q: Why is John looking sad?

A: Because he might have won the game,
if he hadn’t fumbled that pass.
past(might(prosp(pfv(λe.he-win(e)))))
TP: past TO: future
The readings in (52) and (53) differ not only in temporal properties but
in conversational background: (52) is epistemic, while (53) is circumstantial
(more specifically, metaphysical, according to Condoravdi, 2002, although
see Abusch, 2012, for a dissenting view).29 How the unattested combina-
tions of temporal perspective, temporal orientation, and conversational
background are ruled out is a question that immediately arises. One sub-
part of the puzzle is the restriction of the circumstantial reading in (53)
to future orientation. If (53) does involve metaphysical modality, its future
orientation follows from Condoravdi’s (2002) Diversity Condition, or from
Werner’s (2003) Disparity Principle. These are general principles which rely
on the fact that metaphysical modals quantify over worlds which only vary
for times following the temporal perspective.
A second sub-part of the puzzle posed by (52)–(53) is the restriction of
the epistemic interpretation to a present temporal perspective. This can be
derived within Condoravdi’s system via a stipulation that epistemic modals
cannot scope under either past tense or the perfect auxiliary. This issue has
been the subject of much debate, and I discuss it in slightly more detail in
the next subsection.
18.4.2 Past epistemics

Can epistemic modals have past temporal perspectives? The issue is whether
sentences like (54a–c) can make an assertion about what was epistemically
possible or necessary at some past time.
(54) a. Jack’s wife couldn’t be rich. (Stowell, 2004, p. 625)

b. There had to be a hundred people there. (Stowell, 2004, p. 626)
c. There might have been ice cream in the freezer.
Many researchers have denied that epistemic modals can have past temporal
perspectives, at least in English. Hacquard (2011, p. 1495), for example, states
29
A metaphysical modal base is a circumstantial modal base which provides a set of worlds which are all
identical to the actual world up to and including the time at which the modal base is calculated
(Thomason, 2002).
that ‘epistemic modals are evaluated at the speech time’ (see also Groe-
nendijk and Stokhof, 1975; Cinque, 1999; Condoravdi, 2002; Stowell, 2004;
Hacquard, 2006; Borgonovo and Cummins, 2007; Demirdache and Uribe-
Etxebarria, 2008; Laca, 2008). A common belief is that the semantic restric-
tion on the temporal properties of epistemic modals derives from a syntac-
tic restriction, namely that epistemic modals scope over tense, while non-
epistemic modals scope under it. Hacquard (2011) (among others) adopts
Cinque’s (1999) hierarchy of functional heads, relevant portions of which
are given in (55):30
(55) Modepis > Tense > Aspect > Modvolitional > Moddeontic necessity
> Modability/deontic possibility
(Hacquard, 2011, p. 1496)
The view that epistemic modals always scope over tense has not gone unchal-
lenged. Epistemic modals have been shown to allow past temporal perspec-
tives in a number of languages; see Eide (2003, 2005) for Norwegian, Kratzer
(2009) for English and German, and Homer (2013), Mari (2010), and Mar-
tin (2011) for French. Within English and to a certain extent for other lan-
guages, the issue is still controversial. Von Fintel and Gillies (2008) argue
that (54c) does have a past-TP epistemic reading; the reading is facilitated
by a discourse context as in (56). Von Fintel and Gillies (2008, p. 87) state
that here ‘It is possible for [the speaker] to have said something true, even
though at the time of utterance she knows . . . there is no ice cream in the
freezer.’
(56) Context: Sophie is looking for some ice cream and checks the freezer. There is
none in there. Asked why she opened the freezer, she replies:
There might have been ice-cream in the freezer.
The debate about past epistemics goes beyond simple empirical disagree-
ment: among those who acknowledge that epistemic modals can have a
past temporal perspective, some deny that it reflects the ability of the epis-
temic modal to scope under a clause-mate past tense. It has, for example,
been proposed that the relevant readings involve an elided embedding atti-
tude verb or because-clause (Hacquard, 2006, 2011), or that they are felic-
itous due to contexts of (free) indirect discourse (Fagan, 2001; Boogaart,
2007; Hacquard, 2006, 2011). For arguments against some of these escape
hatches, see Homer (2013), Rullmann and Matthewson (2012), and see
Portner (2009, pp. 222–236) for summary and discussion.31
30
As discussed in Section 18.2.2 above, Hacquard does not derive the effects of syntactic position on
interpretation simply from a cartographic syntax. Her proposal is that the different interpretations arise
because modals interact locally with different elements which partially determine the modal’s
interpretation.
31
Portner (2009, p. 227) himself states that a past temporal perspective is rare but not impossible for
epistemic modals, and that the past readings may be limited to stative sentences. See also Iatridou (1990)
for relevant discussion, including the claim that some but not all of what we currently classify as epistemics
can scope under tense.
Modality 553
In a broader cross-linguistic context, Chen et al. (forthcoming) provide

evidence that twelve languages from seven families allow past-TP epis-
temics: English, Dutch, German, Mandarin, St’át’imcets, Northern Straits
Salish, Halkomelen (Salish), Gitksan, Blackfoot (Algonquian), Atayal (Aus-
tronesian), Javanese (Austronesian), and Ktunaxa (isolate). They argue that
cross-linguistically, modal flavour is independent of temporal perspective
and that all possibility modals can have past temporal perspectives, whether
they are interpreted epistemically or circumstantially. Some of the evidence
for this claim is presented in the next subsection.
18.4.3 Modal–temporal interactions across languages

In this section I present evidence for two proposals: first, that epistemic
modals can have a past temporal perspective in a range of languages, and
second, that viewpoint aspects within the prejacent clause help deter-
mine the temporal orientation. The first claim is supported by the data
in (57)–(59), from Blackfoot, Ktunaxa, and St’át’imcets respectively. Each of
these languages has lexical items dedicated to epistemic modality, so the
modal flavour is unambiguously epistemic. The past temporal perspective is
ensured by the discourse context and is not overtly marked because tense is
non-overt in all these languages.
(57) Context: Stacey bought a bone for Pat’s pet, thinking it might be a dog. Later,
she finds out the pet is a snake. When Pat asks her why she bought a bone, she
says:
matonni ni-maat-ssksini-’p-wa ot-aanist-a’pssi-wa
yesterday 1-neg-know.vti-loc:0-nonaff 3-manner-be.vai-3
piiksiksinaa-wa aahkam-omitaa-wa
snake-3 epis-dog-3
‘Yesterday, I didn’t know it was a snake, it might have been a dog.’
(Louie, 2012, elicited using the Feeding Fluffy storyboard, TFS
Working Group 2012)32
(58) Context: Your neighbour doesn’t show up for work and you know there’s been
a flu going around. You send your son to bring her hot soup. She actually took
the day off because her apartment flooded, so she asks why you sent her soup
in the middle of the day.
lin hin sa·nilxuʔ-ni
epis 2 sick-ind
‘You might have been sick.’
(Laturnus, 2012)
(59) Context: When you looked out of your window earlier today the ground was
wet, so it looked like it might have rained. But you find out later that
sprinklers had been watering the ground.
kwís=k’a=tu7
rain=epis=then
‘It might have rained.’
32
www.totemfieldstoryboards.org/stories/feeding_fluffy/.
Further evidence for past-TP epistemic modals is given in (60)–(62) from Gitk-
san; these data also support the second proposal of this section, namely
that viewpoint aspects control temporal orientation. Observe that (60)–(62)
are a minimal triplet containing the epistemic modal ima(’a), a uniformly
past temporal perspective (ensured by the context), and all three possible
temporal orientations, past, present, and future respectively. While in (60)
there is no overt marking of past orientation, in (62) the future orientation
is achieved by means of the overt prospective aspect marker dim.
(60) Context: When you looked out your window earlier today, the ground was wet,
so it looked like it might have rained. But you found out later that the
sprinklers had been watering the ground.
yugw=imaa=hl wis da’awhl
ipfv=epis=cn rain then
‘It might have rained.’ [based on my evidence earlier]
(61) Context: When you looked out your window earlier today, water was falling, so
it looked like it was raining. But you found out later it was the gutters leaking.
yugw=imaa=hl wis da’awhl
ipfv=epis=cn rain then
‘It might have been raining earlier.’
(62) Context: This morning you looked out your window and judging by the clouds,
it looked like it might have been going to rain, so you took your raincoat. Later
you’re explaining to me why you did that.
yugw=imaa=hl dim wis
ipfv=epis=cn prosp rain
‘It might have been going to rain.’
Prospective aspect marking is obligatory whenever a modal is future-

oriented in Gitksan (Matthewson, 2013a). This leads to at least three con-
sequences. First, English and Gitksan differ superficially, in that temporally
unmarked epistemic modals can be future-oriented in English but not in
Gitksan; this is shown in (63).
(63) yugw=imaa=hl wis

ipfv=epis=cn rain
‘It might have rained.’ / ‘It might be raining.’ / = ‘It might rain (in the
future).’
Context: You see puddles, and the flowers looking fresh and damp.
past TO
Context: You hear pattering on the roof. present TO
# Context: You hear thunder, so you think it might rain soon.
future TO
(Matthewson, 2013a, pp. 364–365)
Modality 555
Second, English and Gitksan are mirror images, since past orientation is
obligatorily marked in English (via have), but future orientation is obligato-
rily marked in Gitksan (via dim). And third, English and Gitksan can be anal-
ysed as abstractly parallel, if we are willing to postulate a null prospective
aspect in English and a null perfect in Gitksan (cf. Van de Vate, 2011).
In this section I have sketched a Condoravdi-inspired approach to modal–
temporal interactions, which makes the universal hypotheses that temporal
perspective is given by tense (or the language-internal functional equivalent)
and temporal orientation is restricted by aspect. Obviously, languages vary
independently in their temporal systems. An overarching goal is therefore
to establish the extent to which variation in modal–temporal interactions
reduces to those independent temporal differences. For example, we have
already seen that languages without overt past–present distinctions allow
past temporal perspectives without any overt marking. This is predicted by
the system presented. Whether other predictions are upheld is a matter for
future research.
18.5 Typological and other remaining questions
In this chapter we have discussed the basics in three major areas in modal-
ity research: modal flavour, modal force, and modal-temporal interactions.
Many empirical and analytical questions remain about all of these areas, as
well as about the interactions between them. Assuming that our ultimate
goal is a theory of universals and variation in human language, one impor-
tant task for the field is to gather information about modality in unfamiliar
and understudied languages. Formal research on such languages will allow
us to develop a formal typology of modality, which in turn will facilitate
greater theoretical understanding.
There is, of course, already a rich tradition of modality research in the
typological literature (see Bybee et al., 1994; de Haan, 1997; van der Auw-
era and Plungian, 1998; Palmer, 2001; Hengeveld, 2004; van der Auwera
and Ammann, 2011, among others). These works provide extensive infor-
mation about how modal notions are expressed across languages (e.g., as
verbs, affixes), what types of categories are encoded, and the grammatical-
ization paths of modal elements. Added to this is a recent growth of formal
research on modality, which is able (through hypothesis-driven investiga-
tion and the collection of negative evidence) to isolate precisely and accu-
rately the semantics of modal elements in a range of languages.33,34
33
For critiques of a non-hypothesis-driven typological approach to modality, see Matthewson (2013b) and
Davis et al. (2014). Along similar lines, Nauze (2008, p. 19) points out that a good typology of modality
must rely not only on general descriptive sources such as grammars but on ‘semantically motivated
descriptions’ of modal systems.
34
For a recent example of formal cross-linguistic research on modality, see Arregui et al. (2014), who provide
a modal analysis of imperfective aspects in several different languages.
One major set of questions which cross-linguistic study of modality needs

to address is what the lexicalization patterns are across languages for modal
flavour, modal force, temporal properties, and all combinations of these.
There is already evidence that flavour, force, and temporal properties are not
randomly combined. For example, there is a well-known flavour-orientation
correlation which we hinted at above, namely that circumstantial modals
usually – sometimes even obligatorily – have future orientation. As Kratzer
(2012, p. 54) writes, circumstantial modal bases target facts which are ‘exter-
nal or internal circumstances of people, things or places that determine
their possible futures’. In contrast, realistic epistemic modals target ‘evi-
dence of things implying or suggesting the presence of other facts in the
past, present, or future’.
The circumstantial/future-orientation correlation arises in different lan-
guage families and has been documented and analysed by, among others,
Coates (1995), Enç (1996), Condoravdi (2002), Stowell (2004), Copley (2006),
Werner (2003, 2006), Borgonovo and Cummins (2007), Van de Vate (2011),
Chen (2012), Kratzer (2012, 2013), Matthewson (2012b, 2013a), Thomas
(2013). There are also three-way correlations between force, flavour, and
temporal orientation, such as a restriction of epistemic necessity modals to
past orientation; see Werner (2006), Portner (2009) on English and Lekakou
and Nilsen (2008) on Greek.35 On the other hand, Tonhauser (2011) argues
that the Paraguayan Guaraní future marker -ta allows epistemic necessity
interpretations, and Giannakidou and Mari (2013) argue that the future
morpheme in both Greek and Italian conveys epistemic modality. Further
research is obviously required.
There has not yet been as much research into flavour–force correlations,
but an interesting question is whether one particular modal flavour might
be more likely to lack duals than another. Impressionistically, it seems that
epistemic modals (including elements which have traditionally been anal-
ysed as evidentials) may be more likely to lack duals. Gitksan is a case in
point; we saw above that this language encodes force distinctions in the
circumstantial domain, but not the epistemic. Another example is Niuean
(Polynesian), which possesses a general-purpose epistemic modal liga, usable
in contexts of both high and low certainty, but two circumstantial modals,
maeke and lata, which are specialized for force (possibility and necessity
respectively). Examples (64)–(66) show liga with various translations, which
correspond to the different modal forces the modal allows.
(64) liga kua fano tei

epis prf go prf
‘He/she/they might have left.’
(Matthewson et al., 2012, p. 224)
35
Abraham (1998, p. 233) argues more radically for modals of all forces that ‘reference to the future under
no circumstances gives an epistemic reading’.
Modality 557
(65) Context: Tom wasn’t fishing yesterday, and you were wondering about his
health. But today you see him fishing.
Hī ika a Tom he aho nei . . . liga malolo a ia
catch.fish fish abs Tom on day this epis strong abs 3sg
‘Tom is fishing today . . . he’s probably well.’
(Matthewson et al., 2012, p. 228)
(66) ne liga kua veli hifo e tama ke he pelapela

past epis prf fall down abs child to mud
‘The boy must have fallen in the mud.’
(Seiter, 1980, p. 13)
Example (67) shows the possibility modal maeke in one of its most frequent
uses, an ability reading, and Example (68) shows the necessity modal lata
with an obligation interpretation.
(67) kua maeke he tama ia ke taute pasikala afi

prf circ.possib at child that sbj fix bicycle fire
‘That child is able to fix motorbikes.’
(Seiter, 1980, p. 140)
(68) lata ke ō a tautolu he aho nei ki Queen Street

circ.necess sbj go.pl abs we.pl.incl on day this to Queen Street
‘We should go to Queen Street today.’
(Seiter, 1980, p. 133)
One of the most comprehensive contributions in the area of modal typology

is Nauze (2008). Nauze investigates a relatively small sample (six languages
from six families, as opposed to the 207 languages of van der Auwera and
Ammann, 2011). However, Nauze’s work has the advantage that it is based
not just on traditional descriptions such as grammars (which typically give
insufficient information about the semantics of modality) but also on field-
work, personal discussion with language experts, and targeted and/or for-
mal literature. Nauze advances the following (tentative) proposal:
Modal elements can only have more than one meaning along a unique
axis of the semantic space: they either vary on the horizontal axis and
thus are polyfunctional in the original sense of expressing different types
of modality or they vary on the vertical axis and can express possibility
and necessity, but they cannot vary on both axes. (Nauze, 2008, p. 222)
This proposed universal rules out a modal element which is polyfunc-

tional in terms of both modal flavour and modal force.
In very recent work, Vander Klok (2013) refines Nauze’s proposal. Vander
Klok observes that Nauze predicts that a language could possess one modal
expression which is polyfunctional along the modal force dimension, and
another expression which is polyfunctional along the modal flavour dimen-
sion, within the same domain (epistemic or non-epistemic). For example,
Table 18.4 Hypothetical modal system of the root domain: predicted to exist under
Nauze’s (2008) typology (Vander Klok, 2013, p. 18)
Modal flavour
Root domain
Deontic Pure circumstantial Teleological
Modal Necessity x y y
force Possibility x w z
a language should be able to contain one modal which is specified for a

particular non-epistemic flavour (e.g., deontic) and lacks a dual, covering
both possibility and necessity readings, and simultaneously contain other
circumstantial modals which are polyfunctional along the flavour dimen-
sion. A hypothetical system of this type is schematized in Table 18.4.
Based on languages like Gitksan and Paciran Javanese, Vander Klok
hypothesizes that a more restrictive universal might hold. She proposes
that languages allow for referential ambiguity along only one axis within
each modal domain (epistemic vs. non-epistemic). The system in Table 18.4
violates this restriction, because it involves referential ambiguity along
both the flavour and the force dimensions, within the same (circumstan-
tial) domain. The Gitksan system satisfies the more restricted universal:
although it possesses some modals which vary along a force dimension
(the epistemic modals ima(’a) and gat), and some modals which vary along
¯
a flavour dimension (the circumstantial necessity modal sgi, which allows
deontic, pure circumstantial, and teleological readings), the two types
of polyfunctionality are each confined to their own half of the system.
More cross-linguistic research is obviously required to test these typological
predictions.
18.5.1 Other areas of modal research

Modality is a very large topic, and any one paper on it must unfortunately
set aside many important sub-topics. One way in which the field of modality
extends beyond what I have discussed here is simply in terms of construc-
tion types. In this chapter the English data were drawn mainly from modal
auxiliaries (as is the tradition in much of the literature). However, the data
from other languages were not restricted to auxiliaries; we have looked at
modal verbs, affixes, second-position clitics, and adverbs. Modal semantics
also arises in conditionals, viewpoint aspects, tenses, moods, indefinite arti-
cles, adverbs, adjectives, infinitivals, and so on.36 The null hypothesis is that
the basic semantic concepts presented here will be equally applicable to
36
References for any one of these topics would be too numerous to cite.
Modality 559
other items involving modal semantics; any counter-examples to this pro-

vide interesting avenues of research.
Sometimes, the failure of a semantic analysis to extend to a new domain
is due to independent factors like syntax. For example, I mentioned above
that Chen et al. (forthcoming) found that epistemic modals can have past
temporal perspectives in eight languages. There are a few counter-examples
to the pattern in Chen et al.’s language sample, among them SENĆOTEN (the
Saanich dialect of Northern Straits Salish) and Hul’q’umi’num’ (the Island
dialect of Halkomelem Salish). According to Turner (2013), epistemic modals
in these two languages do not allow past temporal perspectives. However,
Turner suggests that this may be due to the fact that the epistemic modals in
question are of the wrong syntactic category. She draws a parallel between
the epistemic modals in SENĆOTEN and Hul’q’umi’num’ and the English
epistemic adverb maybe. Unlike the modal auxiliary might, maybe does not
scope under tense and cannot receive a past temporal perspective.
(69) Context: Sophie is looking for some ice cream and checks the freezer. There is
none in there. Asked why she opened the freezer, she replies:
There might have been ice cream in the freezer.
#Maybe there was ice cream in the freezer.
(adapted from von Fintel and Gillies, 2008, p. 87)
This is one simple example of how variation in modal semantics is reducible

to independent factors like the syntactic category of the modal elements. It
also illustrates how phenomena in unfamiliar languages have parallels or
counterparts within familiar languages when the latter are re-examined in
the light of the former.
Acknowledgements
I am very grateful to Maria Aloni, Henry Davis, and an anonymous reviewer
for helpful feedback on an earlier version of this paper. Many thanks to
St’át’imcets consultants Carl Alexander, Laura Thevarge, the late Gertrude
Ned, the late Beverley Frank, and the late Rose Agnes Whitley, to Gitksan
consultants Vincent Gogag and Barbara Sennott, and to Niuean consultant
co-author Lynsey Talagi. This research was supported in part by SSHRC grant
#410-2011-0431.

Matthewson (2016) Modality

Uploaded by

Copyright:

Available Formats

Matthewson (2016) Modality

Uploaded by

Document Information

Original Description:

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Matthewson (2016) Modality

Uploaded by

Copyright:

Available Formats

18

18.1 Introduction 525

speakers of all languages (even those under-represented on YouTube) can

(1) wá7=ka s-lep’ i=k’ún7=a ku=pála7 máqa7

(2) Context: They also wrestled the young bulls.

The goal of this chapter is to provide an introduction to three topics which

18.1.1 Possible worlds and modality

Can is a possibility modal, which introduces existential quantiﬁcation over

The modal ﬂavour of can is not completely determined by context; it is also

something about the worlds compatible with a body of evidence.6 Might, on

(5) Context: A homicide detective is summarizing the content of a report which

Non-auxiliary modal elements frequently lexically encode modal ﬂavour.

(6) Context: A violin audition. The conductor says:

Example (6b) expresses obligation, rather than permission; it says (roughly)

In the next section we turn to a more in-depth discussion of modal ﬂavour.

18.2 Modal ﬂavour

18.2.1 Modal base and ordering source

(9) a. [[must]]w,g = λ f s, st,t .λp s,t .∀w [w ∈ ∩ f(w) → p(w ) = 1]

Kratzer shows conclusively that such an analysis is inadequate. The

(11) a. Michl is probably the murderer.

An analysis of these examples requires more than just checking whether

(12) a. She must have climbed Mount Toby.

(13) For all worlds w and z ∈ W : w ≤A z iff {p : p ∈ A and z ∈ p}

(14) For a given order ≤A on worlds:

(15) a. [[must]]w,g = λ f s, st,t .

18.2.2 Where do conversational backgrounds come from?

If there is no overt restriction on conversational background, it can be pro-

A ﬁnal option for conversational backgrounds is that modals might

epistemic and non-epistemic interpretations (i.e., no instances of what they

18.2.3 Modal ﬂavour in Gitksan

Table 18.1 Gitksan modal system

possibility (weak) necessity

(19) a. da’akhlxw-i-s Henry dim jam-t

The split between epistemic and circumstantial readings is absolute in

is (pure) circumstantial: the issue is not whether it is compatible with

Lexical distinctions between modal ﬂavours are pervasive in many lan-

18.2.4 Epistemic vs. circumstantial: the issues

In recent work, Kratzer (2012, 2013) (building on work by Arregui, 2005,

Table 18.2 Traditional vs. new classiﬁcations of (modes of projection of)

traditional classiﬁcation circumstantial epistemic

(25) #wá7=k’a ku=mám’teq láku7 áltsq7=a, t’u7 nílh=a

Table 18.3 A three-way categorization of conversational backgrounds

factual–circumstantial factual–evidential content–evidential

an evidential which allows statements of the form ‘evidential-ϕ, but not

18.3 Modal force

18.3.1 Issues in modal force

(29) a. Michl is probably the murderer.

(30) p is a good possibility in w with respect to a modal base f and an

(31) p is at least as good a possibility as q in w with respect to f and g iff:

A proposition p is a good possibility if there is some world u, such that

common (i.e., the worlds in q − p and in p − q). p is at least as good a possibil-

(32) a. It was twice as likely to rain as it was to snow.

Researchers such as Halpern (1997, 2003), Swanson (2006, 2008), Port-

There are several different approaches to weak necessity modals (see

(34) a. To go to Ashﬁeld, you have to/must take Route 2.

(35) To go to Amherst, you ought to take Route 9.

Further questions raised by the ordering-source approach include the issue

(36) a. She must have climbed Mount Toby.

18.3.2 Modal force in Gitksan and Nez Perce