Chapter 2
N-words and NPIs: Between syntax,
semantics, and experiments
Mojmír Dočekal
Masaryk University in Brno
In this paper I experimentally approach the following question: do strict negative
concord languages like Czech employ two strategies (syntactic and semantic) to
encode negative dependency between a verb and its argument(s) or not? And the
answer is: beside the default syntactic strategy (n-words), there is a class of negative
dependent expressions which are licensed by semantic rules.
Keywords: n-words, negative polarity items, experiments, neg-raising, Czech
1 Introduction
In this article I focus on a problem of dividing negative dependent expressions
into two classes: (i) n-woRds like Czech nikdo ‘nobody’ or Romanian nimeni
‘nobody’ (glossed as n-peRson) in (1a); (ii) negative polaRity items (NPIs) like
Czech sebemenší šance ‘slightest chance ’ or Romanian vreun ‘any’ in (1b). Despite
the long research traditions on both types of expressions (for NPIs see Heim 1984;
Ladusaw 1992; Kadmon & Landman 1993; Krifka 1995; Giannakidou 1997; Lahiri
1998; Gajewski 2011; Chierchia 2013; Crnič 2014 among many others; for n-words
see Laka 1990; Zeijlstra 2004; 2008 among others) there is still no consensus on
the relationship between the two classes of items.1
1Both n-words and NPIs are generally grammatical in sentences with a negated verb. There are
of course language-specific differences, e.g. English NPI any usually cannot appear in subject
position, Romanian vreun ‘any’ in (1b) behaves similarly but many languages allow NPIs to
freely occur in subject position – Błaszczak (2001) lists Hindi, Korean, Japanese among many
other languages where NPIs are licensed in any position of a sentence with a negated verb.
Slavic languages discussed in detail further belong to the set of languages allowing NPIs in
subject position, too.
Mojmír Dočekal. 2020. N-words and NPIs: Between syntax, semantics, and experiments.
In Franc Marušič, Petra Mišmaš & Rok Žaucer (eds.), Advances in formal Slavic
linguistics 2017, 15–49. Berlin: Language Science Press. DOI:10.5281/zenodo.3764845
Mojmír Dočekal
(1) a. Nimeni
n-peRson
nu
not
a
has
venit.
come
‘Nobody came.’
b. * Vreun
npi
student
student
nu
not
a
has
venit.
come
Intended: ‘No student did not come.’
(Romanian; Fălăuș & Nicolae 2016: 586, 591)
Nevertheless, it seems to be settled that the division between n-words and NPIs
correlates with the division between syntactic licensing and semantic licensing
along the following lines:
1) n-woRds are syntactically negative dependent expressions;2
2) negative polaRity items are semantically negative dependent expressions.
Some languages lexicalise the difference between NPIs and n-words, as shown
in the example (1) but sometimes the distinction manifests itself only via stress
(and usually consequently) focus marking. An example of the second strategy
is in (2) from Giannakidou & Zeijlstra (2017) where the non-focused expression
kanenan ‘anybody’ is (according to standard criteria) an NPI while the focused
expression Kanenan ‘n-peRson’ behaves as a n-word.
(2) a. Dhen
not
idhe
saw
kanenan
npi.person
o
the
Janis.
John
‘John didn’t see anybody.’
b. Dhen
not
idhe
saw
Kanenan
n-peRson
o
the
Janis.
John
‘John didn’t see anybody at all.’
(Greek; Giannakidou & Zeijlstra 2017: 17)
2This classification is of course very schematic and it can be a bit problematic to apply it to
a set of typologically diverse languages. Consider e.g. Romance languages where it seems to
be possible to use n-words in questions and in context without overt verbal negation (cases
of indirect negative verbs like doubt a.o.). Romance languages (and generally all non-strict
negative concord languages) allow moreover preverbal n-words in affirmative sentences (as a
rule in non-strict negative concord languages, preverbal n-words require positive verb, unless
the speaker wants to convey a double negation reading: see Laka (1990) for many examples and
further details). But even if the cross-linguistic scenery of n-words is more nuanced than the
distinction n-words=syntax, NPIs=semantics suggests, the classification is generally correct
and can be applied even to Romance (and generally non-strict negative concord languages),
once our syntactic toolbox is supplemented with phonologically null operators which license
n-words (see Zeijlstra 2004 a.o. for such a theory) – the licensing of such operators is of course
highly constrained (see again Zeijlstra 2004 and Zeijlstra 2008 for details).
16
2 N-words and NPIs: Between syntax, semantics, and experiments
Next to the classification of n-words as being basically licensed in syntax (either
via agreement or some other standard syntactic process) and NPIs as semantically
dependent expressions (occurring only in environments with specific
monotonicity properties) there is also an agreed-upon criterion of teasing apart
the two classes, one of its recent formalizations can be found in Giannakidou
& Zeijlstra (2017) – see (3), their example (16).3 The criterion is partially meaning
based and partially relies on context felicity of n-words. Its working will be
exemplified in the following sections.
(3) X qualifies as an n-word iff:
a. X can be used with structures with sentential negation or other X with
meaning equivalent to one ¬; and
b. X provides a negative fragment answer.
In this article I discuss mainly experimental evidence from Czech which allows
us to answer a research question: do languages like Czech (where the evidence
to differentiate between n-words and NPIs is very limited) distinguish between
n-words and NPIs (particularly the class of NPIs called strong NPIs)? Why is
Czech (and generally strict negative concord languages) a good data source for
finding differences between strong NPIs and n-words? Because even if the introduced
distinction between syntactically licensed n-words and semantically licensed
NPIs is supported by many researchers today (Zwarts 1998; Zeijlstra 2004
and Gajewski 2011 among others), there are very influential theories which subsume
n-words under NPIs (Ladusaw 1992) or observe close relationship of the
two classes (Laka 1990): in such theories the distinction between syntactic licensing
(n-words) and semantic licensing (NPIs) of course disappears. The question
of nature (if any … depending on the theory) of the distinction between n-words
and NPIs is theoretically still open and empirically is especially vexing in strict
negative concord languages because there the environment where a speaker can
get positive evidence about the distinction between n-words and NPIs boils down
to neg-raising contexts. This is the reason of centrality of neg-raising for NPI debate
– see further §2 and §3.1.
To foreshadow the experiments discussed in much bigger detail later, let us
consider the following set of Czech sentences (items from one of the experi-
3Beside n-words and their cross-linguistic variation with respect to the strictness of negative
concord, there is also a variation in NPIs: while generally NPIs are bad as negative fragment
answers, one particular subtype of them, minimizers provide felicitous fragment answers, see
Giannakidou (1998) and Błaszczak (2001) for further details. Thanks to an anonymous reviewer
for pressing these points about n-words and NPIs licensing variation.
17
Mojmír Dočekal
ments): if asked about grammaticality of such sentences, Czech native speakers
would consider (4a) ungrammatical, (4b) perfectly acceptable, (4c) good and (4d)
and (4e) bad to some extent. Such graded judgments of sentences containing
(what I will argue further to be) strong NPIs, in concreto graded acceptability of
strong NPIs depending on the presence of negation and/or the type of embedding
verb and some other factors was the original motivation for running the series
of experiments resulting in the current article. It is important to notice that there
is variation among speakers, variation caused by lexical items used in the tested
sentences, etc. This naturally calls for an experimental verification because relying
on a researcher’s intuitions in such cases can lead to totally conflicting claims:
e.g. Bošković & Gajewski (2011) state non-existence of neg-raising in Slavic languages,
while Dočekal & Dotlačil (2016a) defend limited existence of neg-raising
in Czech. The experimental data and their careful analysis – in the light of current
formal semantic theories – allow me to avoid such contradicting claims and
eventually isolate the relevant factors behind NPI licensing and an interaction of
the licensing with other syntax-semantics phenomena as neg-raising, etc.4
(4) a. * Zmizela
Lost
ani
not.even
jedna
one
knížka.
book
‘A single book is missing.’
b. Nezmizela
neg.lost
ani
not.even
jedna
one
knížka.
book
‘Not a single book is missing.’
c. Náš
our
nový
new
knihovník
librarian
nechce,
neg.wants
aby
comp
zmizela
lost
ani
not.even
jedna
one
knížka.
book
‘Our new librarian doesn’t want even one book to be missing.’
d. Náš
our
nový
new
knihovník
librarian
si
se
nepředstavuje,
neg.imagine
že
comp
zmizela
lost
ani
not.even
jedna
one
knížka.
book
‘Our new librarian doesn’t imagine that even one book is missing.’
e. Náš
our
nový
new
knihovník
librarian
neslyšel,
neg.heard
že
comp
zmizela
lost
ani
not.even
jedna
one
4Thanks to an anonymous reviewer for raising importance of this general background question
to me.
18
2 N-words and NPIs: Between syntax, semantics, and experiments
knížka.
book
‘Our new librarian didn’t hear that even one book was missing.’
The article is organized as follows: in the first, more theoretically based part (§2)
I will illustrate the empirical criteria distinguishing n-words and NPIs, then I will
tease apart so called weak NPIs from strong NPIs and lastly I will introduce the
basic observations about Czech and negative dependent expression. §3.1, §3.2,
and §3.3 will be more of the experimental linguistic character, they are heavily
based on the joint work with Jakub Dotlačil (partially reported in Dočekal &
Dotlačil 2016a,b; 2017). In concreto, I will report the experimental evidence for
distinguishing n-words from NPIs stemming from three classes of phenomena:
(i) neg-Raising contexts; (ii) fragment answers; (iii) likelihood manipulated contexts.
The nature of this article is more overview-like, the details about statistics,
design of the experiments, etc. can be found in Dočekal & Dotlačil (2016a,b; 2017);
Dočekal & Šafratová (2018).
2 NPIs vs. n-words: Theory
2.1 N-words
Let us start with introducing some important pieces of linguistic knowledge concerning
n-words, the expressions which are generally taken as syntactically dependent
on negation and which are different both from negative quantifiers on
the one hand and from NPIs on the other hand.
N-words crucially differ from Germanic negative quantifiers as the following
contrast in (5) shows: English verbal negation and a negative quantifier in (5a)
yield only a double negation reading while the word for word translation of (5a)
into Czech with the n-word nikoho ‘anybody’ and verbal negation in (5b) is interpretable
only with one negation scoping wide over the whole sentence as is
clear from the predicate logic formalization. Generally, n-words are syntactically
dependent expressions which occur only in languages where some form of negative
concord is attested.
(5) a. John didn’t see nobody. (English)
¬∃𝑥[peRson(𝑥) ∧ ¬see(John, 𝑥)]
b. John
John
nikoho
nobody
neviděl.
neg.saw
‘John didn’t see anybody.’ (Czech)
¬∃𝑥[peRson(𝑥) ∧ see(John, 𝑥)]
19
Mojmír Dočekal
The distinction between n-words and NPIs already mentioned in the criterion
in 3 is illustrated in (6): (6b) illustrates the unavailability of NPIs as fragment
answers versus the perfect acceptability of n-words in the same context in (6d)
– the Czech translation of the (6a) – (6b) mini-dialogue.
(6) NPIs ≠ n-words:
a. Whom did you talk to?
b. * Anybody. / Nobody.
c. S
with
kým
whom
jsi
aux
mluvil?
spoke?
‘With whom did you speak?’
d. S
with
nikým.
nobody
‘With nobody.’
There are at least two influential theories of n-words: the first one treats n-words
as non-negative indefinites (predicate of type ⟨𝑒, 𝑡⟩) which are required to be in
the scope of clause-mate negation (so called roofing requirement from Ladusaw
(1992), see Giannakidou (1997) for an historical overview). The second type of theory
compares n-words to agreement markers which nicely explains their locality
requirements, basically their need to be licensed syntactically by clause-mate
negation. The second type of approach is developed in Zeijlstra (2004) and Zeijlstra
(2008). In this article I will follow the syntactic agreement approach even
if nothing hinges too much on the particular framework as far as it constrains
the distribution of n-words to clauses with overt verbal negation. This locality
constraint is one of the usually mentioned contrasts between n-words and NPIs
since unlike NPIs which just need to be in a scope of negative element, n-words
need a local negation as the following contrast from Giannakidou & Zeijlstra
(2017) shows.
(7) Dhen
not
prodhosa
betrayed.1sg
mistika
secrets
[pu
that
eksethesan
exposed.3pl
[kanenan
anybody
/*Kanenan]].
n-body
‘I didn’t reveal secrets that exposed anybody.’
(Greek; Giannakidou & Zeijlstra 2017: 18)
It should be noted that the locality requirement of n-words varies across languages
but for n-words in Slavic languages the locality requirement is very strict
as observed already by Progovac (1993). So unlike in Spanish, Italian or Greek
20
2 N-words and NPIs: Between syntax, semantics, and experiments
where the licensing of n-words sometimes (especially in case of subjunctive embedding)
can span from a negation on the root verb to n-words in embedded
clauses, such licensing is ungrammatical in Slavic languages, see the following
examples from Czech.
(8) a. * Petr
Petr
neřekl,
neg.said
že
that
nikdo
n-body
přišel.
came
Intended: ‘Petr didn’t say that anybody came.’
b. Petr
Petr
řekl,
said
že
that
nikdo
n-body
nepřišel.
neg.came
‘Peter said that nobody came.’
c. * Petr
Petr
nechce,
neg.wants
aby
comp.sbjv
tu
here
nikdo
n-body
byl.
were
Intended: ‘Petr doesn’t want anybody to be here.’
d. Petr
Petr
chce,
wants
aby
comp.sbjv
tu
here
nikdo
n-body
nebyl.
neg.were
‘Peter wants nobody to be here.’
2.2 NPIs
A prototypical example of an NPI is the English expression any – see the seminal
work of Kadmon & Landman (1993) (and there for older references). If an NPI
occurs in a sentence without negation it results in an ungrammatical sentence –
(9). If it occurs in a negated sentence like in (10), the only interpretation is a scope
of any under negation: (10a) vs. the unavailable interpretation in (10b). In English
a quantifier few students (which shares with negation the relevant property of
downward entailingness – discussed shortly) licenses NPIs in the object: (11).
(9) * Peter visited anyone.
(10) Petr didn’t visit anyone.
a. Available: ¬∃𝑥[peRson(𝑥) ∧ visit(PeteR, 𝑥)]
b. Unavailable: ∃𝑥[peRson(𝑥) ∧ ¬visit(PeteR, 𝑥)]
(11) Few students visited anyone.
Next, negation is not the only expression licensing NPIs which (at least in the
case of so called weak NPIs) sets NPIs apart from n-words which are licensed
only by negation. Compare the following Czech paradigm in (12) where the
21
Mojmír Dočekal
NPI/minimizer sebemenší šance ‘slightest chance’ contrasts with the adjectival
n-word žádnou (glossed as n-adj). The NPI licensing expression in (12a) is the
quantifier málo studentů ‘few students’. The negation and other NPI licensing
expressions share the property of reversing the direction of entailment in their
argument. Notice how negation reverses entailment in Table 1: logical conjunction
entails logical disjunction in a positive case but negated logical disjunction
entails negated logical conjunction – notice the tautological status of both formulas
in Table 1. Because of the entailment reversion property of NPI licensors their
essential quality is called downward entailing (DE) and is generally accepted by
scholars as the most probable common denominator of NPI environments (since
Ladusaw 1992 at least).5
(12) a. Málo
few
studentů
students
mělo
had
sebemenší
slightest
šanci
chance
složit
to.pass
tu
the
zkoušku.
exam
‘Few students had the slightest chance to pass the exam.’
b. # Málo
few
studentů
students
mělo
had
žádnou
n-adj
šanci
chance
složit
to.pass
tu
the
zkoušku.
exam
Intended: ‘Few students had any chance to pass the exam.’
Table 1: Entailment properties of conjunction and disjunction
𝑝 𝑞 (𝑝 ∧ 𝑞) → (𝑝 ∨ 𝑞) ¬(𝑝 ∨ 𝑞) → ¬(𝑝 ∧ 𝑞)
1 1 1 1
0 1 1 1
1 0 1 1
0 0 1 1
In natural language the reasoning of monotonicity is frequently applied in relation
to sets, subsets and supersets. Notice the predicate logic implications in
5I will discuss in more detail the distinction between weak and strong NPIs. In the literature
there are various attempts to reclassify the landscape of NPIs, one of them Rullmann 1996,
following the work of Krifka 1995 and further elaborated in Lahiri 1998 points out that there is
a special class of NPIs – in Rullmann’s terms even-NPIs (ook maar-series in Dutch) which seem
to be an indefinite incorporated with the semantics of the scalar focus particle even. Even-NPIs
are inherently scalar and interact with focus. Czech ani-NPIs have precisely the characteristics
of even-NPIs as I will discuss later. In this respect they belong to the same class (even-NPIs) as
stressed English any and the ook maar-series in Dutch. Thanks to an anonymous reviewer for
pointing out the importance of this cross-linguistic comparison.
22
2 N-words and NPIs: Between syntax, semantics, and experiments
(13) which corresponds to the patterns from the propositional calculus. If there
is some x in the intersection of P and Q denotation then necessarily there is an
x in P and Q union (13a). And if there is no x in P and Q union, then there cannot
be any x in their intersection (13b). So in a affirmative sentence (in predicate
logic non-negated formula) the entailment goes from an subset (intersection) to
a superset (union) while in a negated sentence, the entailment is reversed and
proceeds from a superset (union) to its subset (intersection). A natural language
example is in (14): the denotation of NP red wine is a subset of the NP wine denotation
and in an affirmative sentence (14a) the entailment is from a subset to
a superset, not vice versa: (14b). In a negated sentence the entailment reverses:
(14c).
(13) a. ∃𝑥[𝑃(𝑥) ∧ 𝑄(𝑥)] → ∃𝑥[𝑃(𝑥) ∨ 𝑄(𝑥)]
b. ¬∃𝑥[𝑃(𝑥) ∨ 𝑄(𝑥)] → ¬∃𝑥[𝑃(𝑥) ∧ 𝑄(𝑥)]
(14) red wine → wine
a. John likes red wine. → John likes wine.
b. John likes wine. ↛ John likes red wine.
c. John does not like wine. → John does not like red wine.
The general condition stating that NPIs occur in downward entailing (DE) environments
can be stated like (15), from (von Fintel 1999: 100).6
(15) Fauconnier–Ladusaw’s Licensing Condition: An NPI is only grammatical
if it is in the scope of an 𝛼 such that 𝛼 is DE.
The downward monotonic and upward monotonic reasoning in case of quantifiers
works like this: upward monotonic quantifiers allow reasoning from subsets
to supersets while downward monotonic quantifiers from supersets to subsets:
(16). Natural language examples of upward, downward and non-monotonic quantification
are presented in (17).
(16) a. DET 𝐴 is upward entailing iff for any 𝐵, 𝐶 (𝐵 ⊆ 𝐶) DET 𝐴 𝐵 ⇒ DET 𝐴 𝐶
6The licensing condition has to be understood as necessary, not sufficient: there are cases of
intervention in NPIs licensing (see Linebarger 1987 for an early treatment and Homer 2008
for a more recent approach), then cases of NPIs being unacceptable even in simple negative
sentences (see Uribe-Echevarria 1994 and Błaszczak 2001). But as none of the experiments
reported further addresses such type of data, for the purposes of this paper I stick to (15), as a
working definition of NPI licensing.
23
Mojmír Dočekal
b. DET 𝐴 is downward entailing iff for any 𝐵, 𝐶 (𝐵 ⊆ 𝐶)
DET 𝐴 𝐶 ⇒ DET 𝐴 𝐵
c. if not upward or downward monotonic → non-monotonic
(17) Upward/Downward entailing and non-monotonic determiners:
a. some: Some toys are blue ⇒ Some toys are colored
b. few: Few toys are colored ⇒ Few toys are blue
c. exactly 𝑛: Exactly three toys are blue ⇎ Exactly three toys are colored
It is important to notice that monotonicity properties belong to a position in
a sentence and they are computed compositionally: so a position in a sentence
can be upward entailing even if it occurs in the scope of a downward entailing
quantifier. In (18) the object position is in the scope of two DE quantifiers
and consequently is upward monotonic, as the validity of the entailment pattern
shows.7
(18) a. [↓ At most three detectives arrested ↓[fewer than four ↑[criminals]]]
b. →[↓ At most three detectives arrested ↓[fewer than four ↑[humans]]]
2.3 Weak and strong NPIs
There is a class of NPIs, so called weaK NPIs with prototypical English examples
like any or ever. Weak NPIs occur in all downward entailing environments as
illustrated in (19).
(19) a. Bill didn’t ever say anything.
b. No student ever said anything.
c. Few students ever said anything.
d. At most 5 students ever said anything.
e. * Between 5 and 10 students ever said anything.
f. * {Some/all/most} students ever said anything.
The second class of NPIs instantiated by English expressions like in weeks, additive
either, and punctual until are so called stRong NPIs and as the name suggests,
7Early discussion of this compositionality which can lead to flip-flop effects in NPIs licensing
can be found in Baker (1970), a recent study incorporating some experimental findings is Geurts
& van Der Slik (2005).
24
2 N-words and NPIs: Between syntax, semantics, and experiments
they occur only in a subset of environments where weak NPIs are grammatical
as illustrated in (20).8
(20) a. Bill didn’t leave until his birthday.
b. No student left until his birthday.
c. * Few students left until their birthdays.
d. * At most 5 students left until their birthdays.
e. * Between 5 and 10 students left until their birthdays.
f. * {Some/most/all} students left until their birthdays.
The logical property which licensors of strong NPIs share (negation and no in
(20)) is a strengthened form of entailment reversal and usually is named antiadditivity.9
In using anti-additivity as the necessary condition for strong NPI acceptability
I follow seminal work of Zwarts (1998). There is a popular alternative
explanation of strong NPIs and their behavior in Gajewski (2011) which describes
their stricter distribution via downward entailing properties but checked both in
at-issue meaning and in the presupposition/implicature part of the meaning. I
will stick to the classic theory of anti-additivity here: the definition is in (21). (22)
illustrates the anti-aditivity (the quantifier no is anti-additive since negation is
always anti-additive as is clear from deMorgan’s law: ¬(𝑝∨𝑞) ↔ (¬𝑝∧¬𝑞)). But
DE quantifiers like few students in (23) are not anti-additive – imagine a scenario
8The interaction of strong NPIs and locality is a vast topic but notice the following pattern from
Romoli (2013: 317):
(i) a. John doesn’t think that Mary will arrive until tomorrow.
b. * John isn’t certain that Mary will arrive until tomorrow.
As the pattern shows, licensing of strong NPIs is always possible in case of negated neg-raising
predicates like think but results in ungrammaticality in cases of negated non-neg-raisers as the
predicate like be certain.
9The full hierarchy of negative strength is the following one: anti-morphicity > anti-additivity >
downward entailing (anti-morphicity defined after Krifka 1995: an operator 𝑂 is anti-morphic
iff: 𝑂(¬𝑋) = ¬𝑂(𝑋); negation is anti-morphic unlike English negative quantifier no as can
be seen from the following equivalence and non-equivalence John wasn’t happy = It’s not the
case that John was happy vs. No student wasn’t happy ≠ It’s not the case that no student was
happy). Being the strongest negative expression (like verbal negation) entails being classified
as anti-additive and downward entailing automatically. Strong NPIs are usually taken to be
licensed by operators of at least anti-additive strength – see the grammaticality of (20b). In
Slavic languages (strict negative concord) it is not that easy to tease apart anti-additivity and
anti-morphicity of negative NPs headed by no but it seems that strong NPIs in Slavic require
at least anti-additive licensors as well.
25
Mojmír Dočekal
with 10 students, three of them drinking and three of them smoking, then ∨ part
of (23) is false while ∧ part of (23) is true.10
(21) Anti-additive function: 𝐹 (𝑥 ∨ 𝑦) ↔ 𝐹 (𝑥) ∧ 𝐹 (𝑦)
(22) No student smokes or drinks ↔ No student smokes and no student
drinks.
(23) Few students smoke or drink ↮ Few students smoke and few students
drink
2.4 NPIs vs. n-words
Returning now to the broader question of distinguishing between NPIs (negative
dependent expressions licensed in semantics via notions like monotonicity
and/or anti-additivity) and n-words (negative dependent expressions licensed in
syntax via agreement), it is acknowledged that such a distinction corresponds
nicely with a well established modularity architecture of a grammar where usually
we distinguish between different forms of well-formedness such as syntactic
or semantic, corresponding to well-formedness which is located in different modules
of grammar. But the picture is not so clear when we consider recent theories
of NPI licensing where the logical properties correlate with syntactic acceptability
of NPIs. In concreto: if ungrammaticality of NPIs in upward entailing environments
is due to lack of the right monotonicity properties in them, then we
are in fact linking the domains of semantics with syntax. And in some theories
(Heim 1984; Crnič 2014) of NPIs licensing where the licensing of NPIs is postulated
via presupposition the linking goes even further: between the licensing in
pragmatics with syntactic acceptability. Recent theories of NPIs (Chierchia 2013)
and strong NPIs (Gajewski 2011) seem to point in the same direction.
Before we move to the experimental part of the article, let us have an outlook
of Czech data scrutinized in much more detail in the series of experiments I will
report. In Czech there are two candidates both at first sight reasonable for the NPI
10Corresponding to the full scale of negative strength – see the previous footnote – some researchers
like Krifka (1995) and Van der Wouden (2002) distinguish weak (licensed in downward
entailing contexts), strong (licensed in anti-additive contexts) and super-strong NPIs (licensed
in anti-morphic contexts). Due to the strict negative concord properties of Slavic languages
(discussed in the previous footnote too) I will stick to the basic dichotomy: weak/strong
NPIs where strong NPIs would subsume the strong and the super-strong NPIs from the more
nuanced classifications. Thanks to an anonymous reviewer for pointing out the importance of
this issue.
26
2 N-words and NPIs: Between syntax, semantics, and experiments
or n-word status: ani (jeden) ‘not even (one)’ and žádný ‘n-adj’. As the following
example demonstrates, both require clause-mate negation in basic cases, so both
can be thought of as either n-words or strong NPIs (the embedded clauses of
communicative verbs can be shown to be non-anti-additive: details to follow).
(24) a. Petr
Petr
neviděl
neg.saw
{ani
even
jednoho
one
/ žádného}
n-adj
studenta.
student
‘Petr didn’t see any student.’
b. {*Ani
even
jeden
one
/ *žádný}
n-adj
student
student
přišel.
came
‘Not even one/any student came.’
c. Petr
Petr
neslyšel,
neg.heard
že
that
{*ani
even
jeden
one
/ *žádný}
n-adj
student
student
přišel.
came
‘Petr didn’t hear that even one/any student came.’
So it is well conceivable that four logical possibilities of classifying ani jeden ‘not
even one’ and žádný ‘n-adj’ are reasonable. Czech tradition like Havránek et al.
(1960) can be interpreted as Table 2 suggests, so basically treating both types of
expressions as syntactically dependent on negation.
Table 2: Czech traditional grammar on ani vs. žádný
item/profile NPIs n-words
ani jeden  
žádný  
And as it is clear from the previous discussion, the division between n-words
and strong NPIs is subtle – the only other clause-mate environment (next to negation)
which passes the test of anti-additivity are prepositions like English without
(compare the equivalence of: John left the pub without paying and saying good bye
↔ John left the pub without paying or John left the pub without saying good bye).
So it is reasonable to ask a research question like (25). Neg-concord languages
like Czech (and generally all Slavic languages) do employ negative dependency
on negation via n-words, so is there a reason for a language to maintain a set
of expressions which does nearly the same job but is licensed in semantics? In
the rest of the article I will argue for the positive answer to the question: the experimental
evidence clearly shows that ani (jeden) ‘not even (one)’ expressions
pattern like strong NPIs, not like n-words, while žádný ‘n-adj’ are n-words.
27
Mojmír Dočekal
(25) Research question: do strict neg-concord languages even allow grammaticalization
of strong NPIs?
3 Experimental evidence
In the three following sections I will discuss the experimental evidence which allows
us to tease apart n-words from NPIs. First in §3.1 I will report evidence coming
from the behavior of NPIs in neg-raising constructions: neg-Raising (NR) is
a primarily interpretational phenomenon where a negation of verbs like think, believe
or want is most saliently understood as scoping over their embedded verb
(I don’t want to leave ≈ I want not to leave, compared with a lack of such interpretation
in case of non-NR predicates: I don’t say I will leave I am saying that I
will not leave). In §3.2 the evidence for distinguishing between n-words and NPIs
will come from their different acceptability as fragmentary answers to questions.
And in §3.3, the two classes will be shown to behave differently with respect to
their entailment and likelihood properties.
3.1 Neg-raising
Because NPIs are licensed in the semantic part of the grammar engine, they are
(ceteris paribus) expected to be able to be licensed at long distance. N-words
as syntactically dependent on negation have to obey strict locality conditions
unlike NPIs. And even more importantly, if the licensing of NPIs happens in
semantics, their licensing should be sensitive to properties of their embedding
verbs, in case of NR-predicates, NPIs should appear in the embedded clauses of
NR-predicates but are predicted to be unacceptable in the embedded clauses of
non-NR predicates (verbs of communication or causation). If we construe such
long distance licensing, the expected pattern should look like the one in Table 3:
n-words cannot be licensed across a clausal boundary, while NPIs can be licensed
from their embedding clause. Nevertheless in case of NR-predicates, we expect a
sharp difference between predicates like want or believe and non-NR predicates
like hear, say, or force.
The experimental results which bear on this issue are summarized in more
detail in Dočekal & Dotlačil (2016a). Let us call this experiment Experiment 1.
Experiment 1 consisted of 5 conditions demonstrated in (26), one of the items
of the experiment. The experiment tested acceptability of sentences containing
NPIs and focused on neg-raising.
28
2 N-words and NPIs: Between syntax, semantics, and experiments
Table 3: n-words vs. NPIs in Neg-raising and non-NR contexts
environment/status NPIs n-words
NR embedded  
non-NR embedded  
(26) a. Ztratila
Lost
se
se
ani
not.even
jedna
one
ovce.
sheep
‘A single sheep is missing.’
b. Neztratila
neg.lost
se
se
ani
not.even
jedna
one
ovce.
sheep
‘Not a single sheep is missing.’
c. Nový
new
bača
shepherd
v
in
Tatrách
Tatras
nechce,
neg.wants
aby
comp
se
se
ztratila
lost
ani
not.even
jedna
one
ovce.
sheep.
‘The new shepherd in Tatras doesn’t want even one sheep to be
missing.’
d. Nový
new
bača
shepherd
v
in
Tatrách
Tatras
si
se
nemyslí,
neg.think
že
comp
se
se
ztratila
lost
ani
not.even
jedna
one
ovce.
sheep
‘The new shepherd in Tatras doesn’t think that even one sheep is
missing.’
e. Nový
new
bača
shepherd
v
in
Tatrách
Tatras
neříká,
neg.say
že
comp
se
se
ztratila
lost
ani
not.even
jedna
one
ovce.
sheep
‘The new shepherd in Tatras doesn’t say that even one sheep is
missing.’
The sentences represent 5 environments listed below:
(A) an affirmative sentence
(B) a negative sentence
29
Mojmír Dočekal
(C) a clause embedded under negated NR predicates of intention and judgement/obligation
(e.g. want, advise)
(D) a clause embedded under negated NR predicates of opinion (believe)
(E) non-NR predicates
Experiment 1 tested only NPIs: ani jeden was one two NPIs in it, the second one
až do ‘until’ + time expression is not important for this article, n-words were
not tested. The descriptive statistics of Experiment 1 is visualized in Figure 1: the
x-axis represents the 5 conditions and the y-axis represents the 5-point Likert
scale (1 = the least acceptable, 5 = the most acceptable). The boxplots summarize
the acceptability in the usual manner. As is evident from the graph, Condition A
was the least acceptable, Condition B most accepted, all other conditions somewhere
in the interval between the two extremes. The most important difference
for this article is the one between the conditions C and D and E where E represents
non-NR predicates and was perceived as less acceptable by native speakers.
This seems to be result of unlicensed NPI in the embedded clauses of non-NR
predicates. I discuss the design of Experiment 1 in more detail here because the
following two experiments (viz §3.2 and §3.3) have an analogical design and statistical
modeling. When describing the following two experiments, I will be less
eloquent.
Figure 1: Results of Experiment 1
30
2 N-words and NPIs: Between syntax, semantics, and experiments
The results of the experiment can be theoretically explained in the scalar approach
to NR (Horn 1973; Romoli 2012; 2013). In the scalar theory of neg-raising
NR predicates (beside the assertion – (27a)) contribute the excluded middle (EM)
implicature to the semantic composition (27b). And finally the alternatives generated
by the implicature are exhaustified by EXH – (28).
(27) a. NR = 𝜆𝑝𝜆𝑥.□ 𝑥 [𝑝]
b. ALT( NR ) = {𝜆𝑝𝜆𝑥.□ 𝑥 [𝑝], 𝜆𝑝𝜆𝑥.[□ 𝑥 [𝑝] ∨ □ 𝑥 [¬𝑝]]}
(28) EXH(ALT(𝑝))(𝑝)(𝑤) = 𝑝(𝑤) ∧ ∀𝑞 ∈ EXCL(𝑝, ALT(𝑝))[¬𝑞(𝑤)]
I will illustrate the mechanics of the scalar theory of NR on an example item
from Experiment 1: (29). Formula in (30a) shows the alternatives generated by
the excluded middle implicature from (27b): it is the negated at-issue meaning
(¬want𝑠 [𝑝]) and the excluded middle part (¬(want𝑠 [𝑝] ∨want𝑠 [¬𝑝])). The excluded
middle in this case formalizes the involvement of the subject s: he either
wants the proposition p, or he wants the negation of p but he cannot be uninterested
with respect to p. The excluded middle for other classes of NR-predicates
has an analogous meaning: opinionatedness for know/believe, clear intentions for
plan, etc. Compare the lack of such an excluded middle meaning in predicates of
communication: a speaker can say p or neg p but he can be silent about p as well.
(30b) then shows the exhaustification of the alternatives: the at-issue meaning
remains the same but the excluded alternative is negated – the usual strenghtening
of the sentence meaning via negating its alternatives. The at-issue meaning
and double negated excluded middle alternative then (via deductive reasoning)
yield the semantic low scope of negation in the embedded proposition. So, as a
consequence of exhaustification of the NR predicate and its excluded middle implicature,
the negation is of the NR predicate is interpreted as having low scope
(semantically).
(29) ‘A new shepherd in Tatra mountains doesn’t want even one sheep to be
missing.’
¬want𝑠 [𝑝]
(30) a. ALT(¬want𝑠 [𝑝]) = {¬want𝑠 [𝑝], ¬(want𝑠 [𝑝] ∨ want𝑠 [¬𝑝])}
b. EXH(¬want𝑠 [𝑝]) = ¬want𝑠 [𝑝] ∧ ¬¬(want𝑠 [𝑝] ∨ want𝑠 [¬𝑝]) |=
want𝑠 [¬𝑝]
Let us recall that strong NPIs are licensed by anti-additive functions: functions
which obey deMorgan’s laws which naturally is true for negation: a natural
31
Mojmír Dočekal
language example is presented in (31a) and (31b) where the entailment is bidirectional
and in propositional logic in (31c) and (31d) where the same meaning
equivalence holds.
(31) a. It didn’t rain and it didn’t snow.
b. It didn’t rain or snow.
c. ¬𝑝 ∧ ¬𝑞
d. ¬[𝑝 ∨ 𝑞]
In the case of NR predicates like want in (32) the embedded clause qualifies as an
anti-additive environment due to the NR-transfer of negation: (32a) is equivalent
to (32b) – both require p and q being false in all possible worlds – see Table 4
with an example of two possible worlds. In such a model both logical formulas
in (32c) and (32d) are true.
(32) a. Susan does not want to sleep and she does not want to dance.
b. Susan does not want to sleep or dance.
c. □¬𝑝 ∧ □¬𝑞 ↔
d. □¬(𝑝 ∨ 𝑞)
Table 4: A fragment of possible worlds for (32)
world/proposition 𝑝 𝑞
𝑤1 0 0
𝑤2 0 0
But consider an example of non-NR predicates like say in (33a) and (33b). (33b)
does not follow from (33a) since non-NR predicates if negated allow only the
high scope of negation interpretation: (34a) – and such an interpretation is the
following: it requires there to be at least some possible worlds where the propositions
p and q are false. But (34a) is stronger: it requires both propositions p
and q to be false in all possible worlds. (34a) would be true in a valuation of
propositions across possible worlds in Table 5 but (34b) would be false in such
a model. In other words: non-NR predicates do not create anti-additive environment
in their embedded clauses. And since strong NPIs need anti-additivity, they
are unlicensed in the embedded clauses of non-NR predicates.
32
2 N-words and NPIs: Between syntax, semantics, and experiments
(33) a. Susan didn’t say that she will sleep and she didn’t say that she will
dance.
b. Susan didn’t say that she will sleep or dance.
(34) a. ¬□𝑝 ∧ ¬□𝑞 (true in the table)
b. ¬□[𝑝 ∨ 𝑞] (false in the table)
Table 5: A fragment of possible worlds for (33a)/(33b)
world/proposition 𝑝 𝑞
𝑤1 0 1
𝑤2 1 0
Returning now to the initial predictions: Experiment 1 confirmed the NPI status
of ani (jeden) – if ani (jeden) were an n-word, the contrast between NR predicates
(ani (jeden) licensed) and non-NR predicates (ani jeden not acceptable) would be
unexplained since syntactic licensing should not be sensitive to semantic distinctions
between anti-additive and non-anti-additive environments. So we can
conclude this section with a first clear experimental confirmation of classifying
ani (jeden) as a strong NPI. Moreover it was established that anti-additivity is a
necessary condition for licensing the strong NPI ani (jeden). Experiment 1 itself
did not establish contrast between strong NPIs (ani (jeden)) and n-words but its
results would be unexpected if ani (jeden) were not a strong NPI. Experiment 1
did not test intuitions for žádný, the reason for that is the following one: žádný
is perceived by native Czech native speakers to be grammatical only if it appears
in a sentence with local negation (26b) type of sentences. So unlike in case of ani
(jeden) where the judgments are much more graded, there is no need to experimentally
establish the acceptability of žádný.
Table 6: N-words vs. NPIs in Neg-raising environments
environment/status NPIs n-words
NR embedded  
non-NR embedded  
33
Mojmír Dočekal
3.2 Fragment answers
Another distinction mentioned already in criterion 3 is the distinction between
n-words and NPIs with respect to their ability to be fragmentary answers to questions.
Roughly, n-words are good fragmentary answers, while NPIs are generally
not acceptable as fragmentary answers. Similarly to the situation in NR contexts
reported in the last section, the acceptability of ani (jeden) as a fragmentary answer
seems to be more varied than in case of n-words which are always good
as fragment answers. I pre-experimentally noticed that especially in cases where
the question supplies more context, the NPIs seem more acceptable, following
the pattern in (35). The fragment answers were tested in two expriments; first in
Experiment 2 the fragment answers were tested against minimal context ques-
tions.
(35) a. Kdo
who
byl
was
dneska
today
večer
evening
na
on
náměstí?
square?
‘Who was today in the evening on the square?’
b. ? Ani
npi
jeden
one
člověk.
human
‘Not even one man.’
c. Kdo
who
tu
here
dneska
today
byl?
was
‘Who was here today?’
d. ⁇? Ani
npi
jeden
one
člověk.
human
‘Not even one man.’
In Experiment 2 (details can be found in Dočekal & Dotlačil 2017), there was
a negative interaction of ani and ellipsis in non-negative questions like (36). In
other words, as expected n-words were judged by speakers as better fragmentary
answers than NPIs. The statistical outcome is visualized in Figure 2 – the relevant
condition is ellipsis and blue bar for n-words, red for NPIs.
(36) Kdo
who
odešel
left
z
from
hospody?
pub?
‘Who left the pub?’
a. Žádný
n-adj
student.
student
‘No student.’
34
2 N-words and NPIs: Between syntax, semantics, and experiments
b. ⁇ Ani
npi
jeden
one
student.
student
‘Not even one student.’
Figure 2: Results of Experiment 2
The theoretical explanation of this known difference is usually provided via a
possible reconstruction of n-words and unavailability of reconstruction for NPIs.
Because NPIs are usually not able to reconstruct under a possible licensor in their
scope (De Swart 1998) like in the following example where NPI any student in
the cleft cannot reconstruct to its base object position under the quantifier no
professor which would license it.11
(37) *It is any student that no professor likes.
We further elaborated the fragment answer distinction in Experiment 3 (details
can be found in Dočekal & Dotlačil 2017) where we provided more contextual
11Again the ban on NPI reconstruction can be side-stepped with a carefully constructed example
as the following sentence from Uribe-Echevarria (1994: p.17) shows: A doctor who knew anything
about acupuncture was not available. It seems though that in such cases it is the whole
subject NP (containing the NPI) reconstruction which saves grammaticality of NPI and this
type of construction seems to be highly restricted. Nevertheless thanks to an anonymous reviewer
for pointing this out.
35
Mojmír Dočekal
informations like in the example item (38). In this experiment the correlation disappeared:
see Figure 3 – conditions FRagNPI vs. FRagNwoRd with no difference
in acceptability.
(38) Koho
whom
vyhodil
fired
profesor
prof
Palný
Palný
včera
yesterday
ze
from
zkoušky?
exam?
‘Who was fired by prof Palný during yesterday’s exam?’
a. Žádného
n-adj
studenta.
student
‘No student.’
b. Ani
npi
jednoho
one
studenta.
student
‘Not even one student.’
Figure 3: Results of Experiment 3
The ability of n-words to appear as fragmentary answers is usually taken as the
standard distinction of n-words against NPIs. But in a recent paper Fălăuș &
Nicolae (2016) observe a strikingly related phenomen: the authors claim (based
36
2 N-words and NPIs: Between syntax, semantics, and experiments
on data from many strict neg-concord languages) that in strict neg-concord languages
n-word answers to negative questions can have (surprisingly) a Double
Negation (DN) reading. This observation goes against the n-words vs. NPIs criterion
as it falsifies the meaning part of it: n-words and (reconstructed) negation
yield only one semantic negation. I checked Fălăuș & Nicolae’s (2016) claims with
10 native speakers of Czech and they seem to be valid – see example (39): there
seems to be even a preference (8/10) for the DN reading – (39a) but the negative
concord reading (39b) is considered to be possible (for 2 out of 10 speakers).
(39) Kdo
who
nepřečetl
neg.read
žádný
n-adj
článek?
article
‘Who didn’t read any article?’
a. Nikdo.
Nobody. (None of us read a single one.) NC (2/10):
¬∃𝑥, 𝑦[peRson(𝑥) ∧ aRticle(𝑦) ∧ Read(𝑥, 𝑦)]
≡ ∀𝑥[peRson(𝑥) → ¬∃𝑦[aRticle(𝑦) ∧ Read(𝑥, 𝑦)]]
b. Nikdo.
Nobody. (Each of us read an article) DN (8/10):
¬∃𝑥[peRson(𝑥) ∧ aRticle(𝑦) ∧ ¬Read(𝑥, 𝑦)]
≡ ∀𝑥[peRson(𝑥) → ∃𝑦[aRticle(𝑦) ∧ Read(𝑥, 𝑦)]]
Fălăuș & Nicolae (2016) solve the availability of DN reading of n-words via postulating
another (focus-related) position for covert negation (CN): in the left periphery
of a clause as in the tree in Figure 4. The position is according to Fălăuș
& Nicolae (2016) licensed via n-word movement to the left peripheral position
above TP. A negation in the left periphery is a second negation in a sentence,
next to the reconstructed negation from the question (surface negation, SN). So
the first negation in (39b) is the interpretation of covert negation, the second one
of the verbal negation. If we follow Fălăuș & Nicolae (2016), we can explain the
puzzling disappearance of contrast between n-words and NPIs (Experiment 3)
as a consequence of the covert negation – if such a negation appears in a clause,
the NPIs are licensed because they do not need to reconstruct under the scope
of verbal negation and then the contrast between n-words and NPIs disappears.
There are many questions raised by postulating such covert negation, especially
with respect to possible over-generation – at the end n-words in strict negative
concord languages cannot appear in sentences without negation but postulating
covert negation leaves this robust observation unexplained. Fălăuș & Nicolae
(2016) try to resolve such problems by restricting the covert negation only to
controllable set of cases, all somehow related to focus movement of n-words to
37
Mojmír Dočekal
CN
FOC
n-word
TP
NegP
SN VP
V n-word
Figure 4: Covert negation, syntax
the left periphery. I tried to verify their claims and conducted a small survey
again with the same 10 speakers of Czech and it seems that Fălăuș & Nicolae’s
general idea is confirmed with an interesting twist. Let us start with a basic case
– (40) is interpreted only with NC reading as is visible from the ranking in (40a)
and (40b) – a double negation reading is simply non-existent.
(40) Nikdo
n-peRson
ničemu
n-thing
nevěří.
neg.believes
‘Nobody believes anything.’
a. NC: ∀𝑥[peRson(𝑥) → ¬∃𝑦[entity(𝑦) ∧ believe(𝑥, 𝑦)]]
b. * DN: ∀𝑥[peRson(𝑥) → ∃𝑦[entity(𝑦) ∧ believe(𝑥, 𝑦)]]
But in case of information structure manipulation like in (41), which is even an
affirmative sentence, the double negation reading surprisingly emerges. A similar
pattern is observed in (42). The sentences moreover seem to have the double
negation reading only. This confirms Fălăuș & Nicolae’s hypothesis about focus
position of the CN: example (40), where there is no object movement to the left
periphery (unlike in (41) and (42)), has only the expected NC reading. In this article
it is not possible to explore more details of this interesting appearance of
double negation reading in a negative concord language like Czech but more importantly:
it seems to be reasonable to postulate another position for negation
in the left periphery of a clause, such a position (because it is somehow licensed
via focus) can then blur the picture of the fragmentary answer criterion and the
fluctuation of acceptability of NPIs as fragmentary answers observed in Experiment
3 is no longer a mystery, context manipulation can lead to a focus related
CN licensing of even strong NPIs as fragment answers.
38
2 N-words and NPIs: Between syntax, semantics, and experiments
(41) V
in
nic
n-thing
nikdo
n-peRson
nevěří.
believes
‘Nobody believes in anything.’
a. * NC (0/10): ∀𝑥[peRson(𝑥) → ¬∃𝑦[entity(𝑦) ∧ believe(𝑥, 𝑦)]]
b. DN (10/10): ∀𝑥[peRson(𝑥) → ∃𝑦[entity(𝑦) ∧ believe(𝑥, 𝑦)]]
(42) Nic
n-thing
při
at
té
the
zkoušce
exam
nikdo
n-peRson
nenapsal.
neg.wrote
‘Nobody wrote anything during the exam.’
a. * NC (0/10): ∀𝑥[peRson(𝑥) → ¬∃𝑦[entity(𝑦) ∧ wRite(𝑥, 𝑦)]]
b. DN (10/10): ∀𝑥[peRson(𝑥) → ∃𝑦[entity(𝑦) ∧ wRite(𝑥, 𝑦)]]
Summary of this section: there seems to be some evidence for classifying ani as
an NPI and žádný as an n-word which stems from the fragment answer experiments.
When the results diverge from the expected dichotomy, there seems to
be a reasonable explanation via postulation of a second covert negation in the
sentence.
3.3 Likelihood scenarios
The last environment discussed in this article concerns the semantic properties
of sentences where n-words vs. NPIs occur. The straightforward predictions are
the following:
1) n-words (licensed in syntax) should not be sensitive to logical properties
of their environment (they require just sentential/verbal negation)
2) NPIs are licensed in semantics and by definition are dependent on semantic
properties like DE, anti-additivity, etc.
I will pursue the line of distinguishing NPIs from n-words via the NPI sensitivity
to monotonicity and likelihood. And I will base my reasoning on a very
influential theory of NPI licensing, the so called simple even hypothesis of NPI
licensing (Heim 1984; Krifka 1995; Crnič 2014 – I will call the theory Heim/Crnič
theory further). The theory describes NPIs using the following three ingredients:
• NPIs associate with covert even – the formalization can be via a formal
[𝑒𝑣𝑒𝑛] feature carried by the NPIs, etc.
• NPIs (like focused element) generate sets of possible alternatives;
39
Mojmír Dočekal
• covert even associates with the alternatives and generates presupposition
of its prejacent being the least probable member of the set of alternatives
(entailing all the alternatives) – in case of association with even (some authors
suggest different covert licensors of NPIs too);
The immediate predictions of the Heim/Crnič theory is that NPIs should be sensitive
to probability and entailing properties. The first and the second one are
logically related: a proposition p cannot be more likely than a proposition q,
if p entails q: intuitive illustration – p being Rambo killed 100 enemies, q being
Rambo killed 99 enemies, p entails q and p is less likely than q; q does not entail
p and is more likely than q – see Crnič (2011) for details of relating entailing and
likelihood. The theoretical intricacies away, the prediction that NPIs should be
sensitive to logical properties like entailing or probability while n-word not is
uncontroversial, see Table 7 for a visualization of these predictions.
Table 7: N-words vs. NPIs in probability manipulated environments
property/item entailment/probability
n-words 
NPIs 
And exactly this prediction was tested in Experiment 2 and Experiment 3. In
both we found a strong correlation of ani and probability. As a side note: a corpus
survey (the biggest national Czech corpus, Křen et al. (2015)) confirms the likelihood
sensitivity of ani – a prototypical example in (43) shows that ani usually
associates with weak scalar items (ani jeden is the second most frequent collocation,
the first one another minimizer ani slovo ‘not a single word’). which via
scalar reasoning entails all other scalar alternatives (¬∃𝑋 [customeR(𝑋) ∧ #𝑋 =
1 ∧ enteR(𝑋)] → ¬∃𝑋 [customeR(𝑋) ∧ #𝑋 > 1 ∧ enteR(𝑋)]). And due to this
entailment the sentence with ani and a weak element associated with ani is the
least probable (entailing all other alternatives).
(43) tento
this
nyní
now
úspěšný
succesfull
podnikatel
businessman
[…] v
in
prvním
first
měsíci
month
neměl
neg.had
[ani
npi
jednoho
one
zákazníka]
customer
‘This currently succesfull businessman did not have even one customer in
the first month.’
40
2 N-words and NPIs: Between syntax, semantics, and experiments
In Experiment 2 the acceptability of ani with strong scalar items was tested
– example item in (44) where the scale of catholic hierarchy is most probably
⟨𝑝𝑟𝑖𝑒𝑠𝑡, 𝑏𝑖𝑠ℎ𝑜𝑝, 𝑐𝑎𝑟𝑑𝑖𝑛𝑎𝑙⟩ – cardinal being high scalar item in any case. The
scale entails contextual (not proper formal logical) entailment due to the facts of
world we know the following implicational hierarchy: ∃𝑥[become caRdinal(𝑥)
→ become bishop(𝑥) → become pRiest(𝑥)] and its reversal as invalid: ∃𝑥
[become pRiest(𝑥) ↛ become bishop(𝑥) ↛ become caRdinal(𝑥)]. To acquire
the grade of cardinal entails acquiring (ceteris paribus) acquiring all lower ranks
of catholic hierarchy but not the other way round. The scalar item cardinal is
the strongest (in the ad hoc scale), it entails all other items in the scale and is
consequently least likely (which fits the natural intuitions). If ani prefers weak
scalar items, it should be degraded with strong items, while n-words (as they are
not picky about semantic environments) should be more acceptable.
(44) […] nestal
neg.became
se
se
{ani
npi
/ žádným}
n-adj
kardinálem
cardinal
‘He didn’t become even a cardinal.’
And we found out that people overall preferred žádný (n-word) with strong scalar
items. The reason is that n-words do not have semantic requirements unlike NPIs:
ani prefers weak scalar items. The statistical results of Experiment 2 are in Figure
2, the pertinent condition liKelihood: ani (red) had mean acceptability very
much below the n-word’s mean acceptability (blue) (around 2.8 for n-words).
Experiment 3 was partially an elaboration of Experiment 2 – while Experiment
2 used an acceptability task, in Experiment 3 the truth value judgment task
was used in case of testing likelihood properties of ani. An example item is in
(45). Again it was tested how much worse is the acceptability of strong scalar
items with ani. In this scenario the scale is ⟨PhD, MA, BA⟩: here the scale is contextually
based on the likelihood of passing the exam (if the scale were based
simply on academic hierarchy, as in the acceptability testing in (44), it would
be ⟨PhD, MA, BA⟩ but in (45) the scale is reversed as passing the exam is prototypically
negatively correlated with the academic rank). The scale is (due to
the context) again based on contextual entailment: ∀𝑥[BA(𝑥) → pass(𝑥)] →
[∀𝑥[MA(𝑥) → pass(𝑥)] → ∀𝑥[PhD(𝑥) → pass(𝑥)]]. Therefore ani associates
again with the strongest scalar item (in the positive version of a tested sentence
entailing all its scalar alternatives). And as the statistical summary in Figure 3
shows (the relevant condition LiKeli_NPI s. LiKeli_NwoRd – blue color), speakers
again preferred n-words to ani NPIs. This again follows from ani’s semantic
requirements (it associates with weak items which in negative contexts become
41
Mojmír Dočekal
least likely among alternative scalar items) vs. n-words which do not have any
semantic sensitivity and are therefore more acceptable than ani.
(45) Scenario: prof. Novák yesterday examined an easy course which BA,
MA and PhD students attend. PhD students pass the exam always, MA in
most cases but BA only rarely. Včerejší
yesterday
zkoušku
exam
u
at
prof.
prof.
Nováka
Novák
nesložili
neg.passed
{ani
npi
/ žádní}
n-adj
bakaláři.
BA-students
‘No bachelors passed the yesterday’s exam by prof. Novák.’
Empirically both experiments strongly support the classification of ani as an NPI
which associates with weak scalar items and žádný as an n-word licensed in the
syntax (and consequently without any particular semantic sensitivity).
Table 8: Ani vs. žádný in probability manipulated environments
property/item probability/entailment
žádný 
ani 
The theoretical explanation of ani being an NPI which obligatorily selects
weak scalar items can be the following. The first thing to note is that the facts
observed in the experiments are only a piece of a bigger pattern where ani competes
in some environments with another scalar particle i ‘even’. In a recent experiment
(Dočekal & Šafratová 2018) it was confirmed that i obligatorily selects
strong scalar items, while ani weak items. Illustrated on a data pattern close to
the catholic hierarchy from Experiment 2 Czech native speakers are prone to the
following judgments (where * should be understood as total unacceptability in
experiments, ⁇ as in-between-acceptability and ✓ as nearly total acceptability –
statistic noise away – but of course only in case the judgments are related to the
set up scale, catholic hierarchy in (46).
(46) a. Upward entailing contexts:
i. Petr
Petr
se
se
nakonec
at-end
stal
became
{i
even
kardinálem
cardinal
/ ⁇i
even
knězem}.
priest
‘Petr in the end became { even a cardinal / even a priest }.’
42
2 N-words and NPIs: Between syntax, semantics, and experiments
ii. Petr
Petr
se
se
nakonec
at-the-end
stal
became
{*ani
not.even
kardinálem
cardinal
/ *ani
not.even
knězem}.
priest.
‘Petr in the end didn’t become even a cardinal.’
b. Downward entailing, non anti-additive contexts:
i. Jestli
If
se
se
Petr
Petr
stal
became
{i
even
kardinálem
cardinal
/ ⁇i
even
knězem},
priest
tak…
then
‘If Peter became even a cardinal, then …’
ii. Jestli
If
se
se
Petr
Petr
stal
became
*ani
not.even
kardinálem,
cardinal,
tak
then
…
‘If Peter didn’t become even a cardinal, then …’
c. Downward entailing, anti-additive contexts:
i. Petr
Petr
se
se
nakonec
at-the-end
nestal
neg.become
{*i
even
kardinálem
cardinal
/ *i
even
knězem}.
priest
‘Petr didn’t become even a cardinal at the end.’
ii. Petr
Petr
se
se
nakonec
at-the-end
nestal
neg.become
{⁇ani
not.even
kardinálem
cardinal
/
ani
not.even
knězem}.
priest
‘Petr didn’t become even a cardinal at the end.’
The pattern we observe is the following: i in upward entailing contexts and downward
entailing contexts prefers strong elements on a scale but it is unacceptable
with weak or strong scalar items in anti-additive contexts; ani prefers weak scalar
items in anti-additive contexts but it is unacceptable in upward entailing contexts
with both weak and strong scalar items (and in simple DE contexts). Such a pattern
is explainable (following the logic of argumentation in Crnič 2011) as i and
ani spelling out the following features:
(47) a. i … [even]
b. ani … [even,aa]
The feature [even] requires the association with covert even defined below in
(48) following Crnič (2014) among many others. The feature [aa] requires the
item to occur in an anti-additive environment. The items form a scale in (49) and
43
Mojmír Dočekal
compete for insertion via the usual Maximize presupposition principle which
requires the speaker to make her contribution presupposing as much as possible
(for the original formulation see Heim 1991).
(48) even 𝑤 (𝐶)(𝑝) is defined only if ∀𝑞 ∈ 𝐶[𝑞 ≠ 𝑝 → 𝑞 >liKely 𝑝]
(49) ⟨𝑖, 𝑎𝑛𝑖⟩
The observed distribution of i/ani and their strong/weak association is explainable
as follows:
1. Upward entailing environments: i is licit but only with strong scalar items
as then the even presupposition is satisfied, ani cannot be inserted as UE
environments clash with ani [aa] feature.
2. Downward entailing environments: i is licit with even scoping below the
DE operator: [ → [[even C] antecedent … i …] consequent ], ani cannot be
used due to the [aa] feature requirement.
3. Anti-additive environments: i cannot be inserted because Maximize Presupposition
dictates the insertion of the most specific item (ani in this case),
ani associates with weak scalar items: the scope [even C] [¬ …ani …].
4. The association of i/ani with ’wrong’ scalar items is perceived as bad (⁇)
but not totally ungrammatical – weak scalar item for i in upward entailing
contexts and strong scalar items for ani in anti-additive environments.
The last point seems to point to the existence of possible reversed scoping: [even
C][→ [antecedent …i …] consequent ] for i and [¬ [even C] …ani …] for ani which
would explain their allowed (even if not preferred) ’crossed’ association. But as it
was confirmed by Experiment 2 and Experiment 3 ani associates with weak items,
while i with strong scalar items (see Dočekal & Šafratová 2018 for details) by default.
This default scope exchange of i/ani which happens exactly in anti-additive
contexts (i prefers strong elements, ani weak elements but only in the scope of
negation – negation being the anti-additive licensor in 99%) reveals their unified
semantics where the flip-flop is a consequence of entailment/likelihood reversal
caused by the negation. The only difference between i and ani is the formal
feature [aa] which formalizes the morphological incorporation of negation into
ani. It would be possible to encode the scope differences via different features
([solo] of Crnič 2011 for the weak elements) but such a move would miss the
nice competition pattern which emerged from the data: namely i is in principle
44
2 N-words and NPIs: Between syntax, semantics, and experiments
expected in anti-additive environments but cannot be inserted as a consequence
of ani being more specific ([even,aa]).
Summary of this section: ani (jeden) ‘not even (one)’ behaves like a strong
NPI – this behavior was confirmed by Experiment 2 and Experiment 3 where
association with strong scalar items was sanctioned (against relatively acceptable
n-words modifying strong scalar items). Furthermore, ani competes with i – the
former prefers strong scalar items which was experimentally confirmed too. The
association with weak scalar items and competition with i would be unexpected
if ani were n-word.
3.4 Summary
Let us end this article by answering the question asked at the beginning: do nwords
and strong NPIs co-exist in natural language? And if yes (in some languages
like English they do co-exist for sure), does this distinction hold even in
strict neg-concord languages where the boundary between strong NPIs and nwords
is even more subtle? The experiments, their results and their theoretical
interpretation described in this article bring very strong support of the existence
of both classes of negatively dependent expressions even in a strict neg-concord
language like Czech. This result allows us to maintain the standard assumptions
concerning n-words (they are licensed syntactically) and NPIs (they are licensed
in semantics/pragmatics). More importantly, the data patterns of Czech NPIs
seem to strongly favor the NPI theories which base their licensing on concepts
like anti-additivity and likelihood (Zwarts 1998 in the first case, Heim 1984 and
Crnič 2014 in the second). Another issue touched in this article is unreliability
of our intuitions: it seems that distinguishing between n-words and strong NPIs
has to be based on such subtle data which can only be obtained by experimental
methods. The subtlety of judgments can explain differing stances on this distinction
in the previous literature where such opposing views as: n-words are a
subclass of NPIs (Ladusaw 1992, Fălăuș & Nicolae 2016 a.o.) versus n-words are a
separate class (Zeijlstra 2008 and Giannakidou & Zeijlstra 2017 a.o.) were maintained.
There is another pertinent question raised by the data: do all speakers
agree with respect to the distinction between n-words and strong NPIs? And if
no, is there a real dialectal variation or at least some correlation? The results of
the experiments in fact bear direct evidence on this fascinating question but the
space of this article is alas filled completely.
45
Mojmír Dočekal
Abbreviations
n-peRson n-word for persons
n-thing n-word for things
neg negation
npi negative polarity item
sbjv subjunctive
de downward entailing
pl plural number
sg singular number
comp complementizer
aux auxiliary verb
n-adj n-word for properties
aa anti-additive
se reflexive clitic
nR Neg-raising
Acknowledgements
I would like to thank two anonymous reviewers, the editors of the volume, the
audience at the FDSL 12.5 conference, especially Julie Goncharov and Lanko
Marušič, as well as Jakub Dotlačil, Iveta Šafratová, Anamaria Fălăuş and Hedde
Zeijlstra. All errors are, of course, my own. I gratefully acknowledge that the
research was supported by a Czech Science Foundation (GAČR) grant to the Department
of Linguistics and Baltic Languages at the Masaryk University in Brno
(GA17-16111S).
References
Baker, C Lee. 1970. Double negatives. Linguistic Inquiry 1(2). 169–186.
DOI:10.1080/08351816909389104
Błaszczak, Joanna. 2001. Investigation into the interaction between the indefinites
and negation. Berlin: Akademie Verlag.
Bošković, Željko & Jon R. Gajewski. 2011. Semantic correlates of the NP/DP parameter.
In Suzi Lima, Kevin Mullin & Brian Smith (eds.), NELS 39: Proceedings
of the 39th Annual Meeting of the North East Linguistic Society, 121–134.
Amherst, MA: GLSA Publications.
Chierchia, Gennaro. 2013. Logic in grammar: Polarity, free
choice, and intervention. Oxford: Oxford University Press.
DOI:10.1093/acprof:oso/9780199697977.001.0001
Crnič, Luka. 2011. Getting even. Cambridge, MA: MIT. (Doctoral dissertation).
Crnič, Luka. 2014. Against a dogma on NPI licensing. In The art and craft of semantics:
A festschrift for Irene Heim, vol. 1, 117–145. Cambridge, MA: MITWPL.
De Swart, Henriëtte. 1998. Licensing of negative polarity items under inverse
scope. Lingua 105(3-4). 175–200. DOI:10.1016/S0024-3841(98)00021-7
46
2 N-words and NPIs: Between syntax, semantics, and experiments
Dočekal, Mojmír & Jakub Dotlačil. 2016a. Experimental evidence for neg-raising
in Slavic. Linguistica 56(1). 93–109. DOI:10.4312/linguistica.56.1.93-109
Dočekal, Mojmír & Jakub Dotlačil. 2016b. When is not-believing believing that
not? In Robert Truswell, Chris Cummins, Caroline Heycock, Brian Rabern
& Hannah Rohde (eds.), Proceedings of Sinn und Bedeutung 21, 387–399.
Edinburgh: University of Edinburgh. https : / / sites . google . com / site /
sinnundbedeutung21/.
Dočekal, Mojmír & Jakub Dotlačil. 2017. Strong NPIs vs. n-words: Acceptability
experiment in Czech. Paper presented at Sinn und Bedeutung 22, Berlin. https:
//sinnundbedeutung22.wordpress.com/.
Dočekal, Mojmír & Iveta Šafratová. 2018. Even hypothesis of PPIs licensing:
An experimental study. In Joseph Emonds, Markéta Janebová & Ludmila
Veselovská (eds.), Language use and linguistic structure: Proceedings of the Olomouc
Linguistics Colloquium 2018, 253–273. Olomouc: Palacký University.
Fălăuș, Anamaria & Andreea Nicolae. 2016. Fragment answers and double negation
in strict negative concord languages. In Mary Moroney, Carol Rose Little,
Jacob Collard & Dan Burgdorf (eds.), SALT 26: Proceedings from the 26th Conference
on Semantics and Linguistic Theory, 584–600. Washington, DC: Linguistic
Society of America. DOI:10.3765/salt.v26i0.3813
Gajewski, Jon R. 2011. Licensing strong NPIs. Natural Language Semantics 19(2).
109–148. DOI:10.1007/s11050-010-9067-1
Geurts, Bart & Frans van Der Slik. 2005. Monotonicity and processing load. Journal
of Semantics 22(1). 97–117. DOI:10.1093/jos/ffh018
Giannakidou, Anastasia. 1997. The landscape of polarity items. Groningen: University
of Groningen. (Doctoral dissertation).
Giannakidou, Anastasia. 1998. Polarity sensitivity as (non)veridical dependency.
Amsterdam: John Benjamins.
Giannakidou, Anastasia & Hedde Zeijlstra. 2017. The landscape of negative
dependencies: Negative concord and n-words. In The Wiley Blackwell
companion to syntax, Second edition. New York: Wiley–Blackwell.
DOI:10.1002/9781118358733.wbsyncom102
Havránek, Bohuslav et al. 1960. Slovník spisovného jazyka českého. Praha: Nakl.
Československé akademie věd.
Heim, Irene. 1984. A note on negative polarity and downward entailingness. In
Charles Jones & Peter Sells (eds.), NELS 14: Proceedings of the 14th Annual Meeting
of the North East Linguistic Society, 98–107. Amherst, MA: GLSA Publica-
tions.
47
Mojmír Dočekal
Heim, Irene. 1991. Articles and definiteness. In Arnim von Stechow & Dieter Wunderlich
(eds.), 487–535. Berlin: de Gruyter. DOI:10.1515/9783110126969.7.487
Homer, Vincent. 2008. Disruption of NPI licensing: The case of presuppositions.
In Tova Friedman & Satoshi Ito (eds.), SALT 18: Proceedings of the
18th Semantics and Linguistic Theory, 429–446. Ithaca, NY: CLC Publications.
DOI:10.3765/salt.v18i0.2483
Horn, Laurence. 1973. On the semantic properties of logical operators in English.
Los Angeles, CA: University of California. (Doctoral dissertation).
Kadmon, Nirit & Fred Landman. 1993. Any. Linguistics and Philosophy 16(4). 353–
422. DOI:10.1007/BF00985272
Křen, Michal, Václav Cvrček, Tomáš Čapka, Anna Čermáková, Milena Hnátková,
Lucie Chlumská, Tomáš Jelínek, Dominika Kováříková, Vladimír Petkevič,
Pavel Procházka, Hana Skoumalová, Michal Škrabal, Petr Truneček, Pavel
Vondřička & Adrian Zasina. 2015. SYN2015: Reprezentativní korpus psané
češtiny. http://www.korpus.cz. Praha: Ústav Českého národního korpusu,
FF UK.
Krifka, Manfred. 1995. The semantics and pragmatics of polarity items. Linguistic
Analysis 25(3–4). 209–257. DOI:10.3765/salt.v0i0.2462
Ladusaw, William A. 1992. Expressing negation. In Chris Barker & David
Dowty (eds.), SALT 2: Proceedings from the 2nd Conference on Semantics
and Linguistic Theory, 237–260. Columbus, OH: Ohio State University Press.
DOI:10.3765/salt.v2i0.3030
Lahiri, Utpal. 1998. Focus and negative polarity in Hindi. Natural Language Semantics
6(1). 57–123. DOI:10.1023/A:1008211808250
Laka, Itziar. 1990. Negation in syntax: On the nature of functional categories and
projections. Cambridge, MA: MIT. (Doctoral dissertation).
Linebarger, Marcia C. 1987. Negative polarity and grammatical representation.
Linguistics and Philosophy 10(3). 325–387. DOI:10.1007/BF00584131
Progovac, Ljiljana. 1993. Negative polarity: Entailment and binding. Linguistics
and Philosophy 16(2). 149–180. DOI:10.1007/BF00985178
Romoli, Jacopo. 2012. Soft but strong. neg-raising, soft triggers, and exhaustification.
Cambridge, MA: Harvard University. (Doctoral dissertation).
Romoli, Jacopo. 2013. A scalar implicature-based approach to Neg-raising. Linguistics
and Philosophy 36(4). 291–353. DOI:10.1007/s10988-013-9136-2
Rullmann, Hotze. 1996. Two types of negative polarity items. In Kiyomi
Kusumoto (ed.), NELS 26: Proceedings of the 26th Annual Meeting of the North
East Linguistic Society, 335–351. Amherst, MA: GLSA Publications.
48
2 N-words and NPIs: Between syntax, semantics, and experiments
Uribe-Echevarria, Maria. 1994. Interface licensing conditions on negative polarity
items: A theory of polarity and tense interactions. Storrs, CT: University of Connecticut.
(Doctoral dissertation).
Van der Wouden, Ton. 2002. Negative contexts: Collocation, polarity and multiple
negation. London: Routledge.
von Fintel, Kai. 1999. NPI licensing, Strawson entailment, and context dependency.
Journal of Semantics 16(2). 97–148.
Zeijlstra, Hedde. 2004. Sentential negation and negative concord. Amsterdam: University
of Amsterdam. (Doctoral dissertation).
Zeijlstra, Hedde. 2008. Negative concord is syntactic agreement. Manuscript, University
of Amsterdam. https://ling.auf.net/lingbuzz/000645.
Zwarts, Frans. 1998. Three types of polarity. In Fritz Hamm & Erhard
Hinrichs (eds.), Plurality and quantification, 177–238. Dordrecht: Springer.
DOI:10.1007/978-94-017-2706-8_5
49