Anastasia Giannakidou, Polarity Sensitivity as (Non) Veridical Dependency.
(Linguistik aktuell; Bd. 23). Amsterdam: John Benjamins, 1998.
281 pp., ISBN 1556199074 (alk. paper)
Reviewed by Mandy Simons, Dept. of Philosophy, Carnegie Mellon University
Greek is one of many languages in which are found Polarity Items,
expressions whose distribution is limited to a subset of the
environments in which other expressions of the same syntactic
category can occur. Well-known examples are English "any" and
idioms such as "give a damn". In this book, Giannakidou attempts
to answer two central questions about Polarity Items (PIs), based
on the rich variety of such items in Greek. The first question
is what she calls the licensor question: What property of the
linguistic environment renders PIs grammatical or ungrammatical?
Her answer is that it is the property of veridicality, which she
discusses and defines in Chapter Three. The second question is
why PIs should show the particular distribution that they do.
Giannakidou attempts to answer this question by "link[ing] the
lexical semantics of PIs to their limited distribution, and
[showing] how the former determines the latter" (p.20).
The book has four chapters. The bulk of Chapter One is a brief
presentation of some earlier approaches to the licensor question.
The chapter ends with an outline of a dynamic semantic
framework whose assumptions Giannakidou relies on in her account.
Some of this presentation is redundant, as several of the notions
defined are not utilized in the account. The specific framework
adopted is that of Groenendijk, Stokhof and Veltman (1996)
In Chapter Two, Giannakidou presents the data on Greek PIs. The
discussion here is detailed and informative. She argues that there are
four distinct categories of PI in Greek. First, she distinguishes a
class of Affective Polarity Items (APIs), which occur in the scope of
negation and other negative expressions, but also in non-negative
environments including the antecedents of conditionals and in the
scope of modals, all contexts which in Chapter Three are characterized
as non-veridical. The class of Negative Polarity Items proper (NPIs)
are limited to true negative environments, and differ from APIs in
both their syntactic and semantic behavior. The particular
interest of the API/NPI distinction lies in the fact that the same
words function in both roles in Greek. When these words receive
emphatic stress, they show one set of properties; when they do
not, they show another. Giannakidou thus argues that emphatic stress
marks a lexical distinction. In addition to APIs and NPIs, Giannakidou
distinguishes two further types of Polarity Item: Free Choice Items
(FCIs) and Subjunctive Relative Clauses (SRs). These classes,
too, are shown to have distinct distributional patterns.
Giannakidou then goes on to present what she calls the sensitivity
semantics of the items in each class, that is, the lexical semantic
properties to which she attributes the sensitivity of these
items to the veridicality or non-veridicality of the context. This
part of the discussion is far more problematic. In making her central
claims as to the semantics of PIs,Giannakidou invokes several
novel notions which I found to be inadequately defined. To illustrate
some of the difficulties, consider the presentation of the sensitivity
semantics of APIs. Giannakidou proposes that APIs are expressions of
existential quantification, but of a special kind. She calls them
"dependent existential quantifiers". These are initially defined as
in (1) (p.70, Sec.2.3.5). (The definition is altered slightly as
I cannot reproduce in this format the symbols used in the original.)
(1) An existential quantifier is dependent iff the variable x which it
contributes does not introduce a discourse referent in
[the actual world].
One problem is how to understand the term "discourse referent".
Giannakidou is assuming here GSV's semantic framework, but GSV
do not themselves use this term. They distinguish variables,
which are syntactic elements; pegs, which are "formal objects"
with which variables are associated; and the objects in the
domain of discourse. In her presentation of GSV's framework,
Giannakidou says that "since pegs serve no purpose other than
being 'intermediaries' between variables and discourse referents,
they can be dispensed with" (p.28). I infer from this that
Giannakidou is using the term "discourse referents" to refer to
members of the domain. (This in itself is a confusing use
Accordingly, I am uncertain how to understand the definition in (1).
In GSV, the use of a quantifier adds the associated variable to the
variables that are in active use, introduces a peg, and associates the
variable with a peg. The members of the domain are not affected by
linguistic items. Perhaps then we should set aside Giannakidou's
decision to dispense with pegs, and replace "discourse referent" in
definition (1) by "peg". However, it remains unclear how to interpret the
definition, as we cannot talk of pegs which belong to one world or another,
pegs being formal objects in the theory.
In the prose, Giannakidou says that "dependent existential quantifiers,
unlike regular existentials, cannot assert existence in the actual world."
However, in Chapter Three, Giannakidou modifies her definition of dependent
existential quantifiers to accord with the notion of relativized
veridicality which she introduces there (see below). The revised
definition, given in Sec.3.3.6 (p.140), is as follows:
(2) An existential quantifier is dependent iff the variable which it
contributes does not introduce a discourse referent in [any]
individual's epistemic model M(x).
An individual's epistemic model is a set of worlds all of which are
compatible with some aspect of the individual's epistemic state.
(I discuss this further below.) Again, the definition is problematic
as no explanation of what is meant by introducing a discourse referent
into a set of worlds is given. It is apparent, however, that the issue
is no longer existence in the actual world, but existence in an
Giannakidou claims that the distribution of APIs follows from their
being dependent existential quantifiers. However, as the definition
of this notion is not adequate, it is impossible to evaluate this claim.
A plausible way to interpret what she says in Chapter Two is that APIs
are expressions of existential quantification whose occurrence is
restricted to environments in which existence of an object satisfying
the associated description is not entailed. But this is merely to repeat
the proposed licensing condition, and does not achieve Giannakidou's
desired goal of deriving the licensing condition from the lexical
semantics of the items.
As further illustration of the difficulties posed by Giannakidou's
definitions, consider her proposal for the sensitivity feature of
subjunctive relatives (p.91, Sec.2.5.2). (I here use E for the
(3) [Op DP+SubjRel VP] has a truth value iff it is not known whether the
following is true: Ex [NP(x) & SR(x)]
The principal difficulty here is the phrase "is not known", which fails to
specify who is required to not know whether the existential statement is
true. Giannakidou perhaps intends to treat the requirement as a
presupposition; this is consistent with the proposal that failure to
satisfy the condition results in a truth value gap. However, this case is
unlike the normal case of presupposition, which requires some proposition
to be part of the common ground. Here we seem to have an
anti-presupposition, which requires that the common ground *not* contain
the existential proposition. It would be interesting to see how this
could be spelled out. Giannakidou's formulation, however, does not provide
the necessary details.
In Chapter Three, Giannakidou turns to the licensor question. First, she
defines a notion of veridicality, which I will discuss in some detail below.
Then, she argues that it is possible to state for each class of PIs a
veridicality-based licensing condition which accounts for the distribution
of the items in that class. Crucially, each class requires a different
relationship to the licensor. APIs, she argues, are licensed by
non-veridicality: they are grammatical in any nonveridical environment.
NPIs are subject to the more stringent requirement that they occur in an
antiveridical environment. Non-affective PIs (FCIs, and SRs) are subject
to what Giannakidou calls an anti-licensing condition: they are
ungrammatical in a veridical environment, but their grammaticality is not
guaranteed by non-veridicality. Following Linebarger (1980, 1987),
Giannakidou also allows that APIs may be licensed indirectly, via a negative
implicature generated by the sentence in which the PI occurs. Giannakidou
argues further that the licensing condition of each class is predictable,
given the sensitivity semantics of that class, although she does not
provide an account of why some PIs may be licensed indirectly and some
Chapter Four focuses on the syntactic licensing of APIs. Giannakidou
begins by relating her discussion to the analysis of Negative Concord
(NC). She takes it to be the case that the structures in which Greek
NPIs (but not other APIs) occur are instances of NC. Hence, an account
of the former should extend to NC structures generally. After providing
a typology of NC languages, Giannakidou discusses a number of existing
accounts of NC. She then gives her own account, according to which the
negative words which occur in NC structures -- in Greek, the NPIs -- are
polarity sensitive universal quantifiers. The polarity sensitivity of NPIs
requires them to occur in antiveridical contexts; but as quantifiers, they
must undergo QR, and take scope over negation at LF. This results in the
typical interpretation of NC structures. The central claim here is that
NPIs are required to escape the syntactic scope of the semantic licensor
at LF. This requirement, though, is not essentially syntactic, but results
from the required topic-hood of the items. In contrast, APIs are
subject to a positive syntactic licensing condition, namely, the
requirement that they occur in the c-command domain of the licensor at LF.
(Here, Giannakidou departs from earlier proposals, which took the relevant
level to be S-structure.) Thus, the differences in the semantic licensing
of the different types of PI are paralleled by differences in their
In this work, Giannakidou has taken an interesting approach to the issue of
polarity sensitivity, and has addressed important issues. First is the
question of the homogeneity of the expressions that show polarity
sensitivity. Her detailed exploration of Greek PIs leads to a
convincing argument for a categorization of these items into distinct
classes. However, she goes beyond the stage of categorization, and seeks
a feature which unifies the classes. This, she identifies as sensitivity
to veridicality. Giannakidou then takes the important next step of
attempting an explanation of why these items should have the sensitivity
she identifies. Each aspect of this project has its own strengths and
weaknesses. I have discussed the first two aspects above. I conclude with
a discussion of the notion of veridicality, which Giannakidou invokes as
the unifying property of polarity sensitivity.
Giannakidou first considers what she calls an "absolute notion" of
veridicality (pp.106-110, Sec.3.1.3). There are again some
difficulties with definitions:Giannakidou defines veridicality
only for monadic propositional operators and dyadic truth
functional connectives, although she is interested in characterizing
the veridicality of predicates such as "manage", "believe" and
"start", which do not belong to either of these classes. She also
wishes to characterize non-embedded simple clauses such as "Theodora
saw something" as veridical environments, although their
representation does not involve a propositional operator to which
veridicality can be attributed. However, her discussion indicates
that she has in mind a fairly intuitive notion of veridicality which
distinguishes, for instance, between "know", "believe" and negation.
"Know" (or alternatively, a clause embedded under "know") is
veridical, by virtue of the fact that "a knows that p" entails
that p. "Believe", on the other hand, is nonveridical, as the truth
of "a believes that p" entails neither that p nor that not p.
Negation is an antiveridical operator, as the truth of a negated
sentence (trivially) entails the falsity of the clause embedded
However, this notion of veridicality is not adequate for Giannakidou's
purposes. Clauses embedded under "believe", "say" and "dream" are
nonveridical by the above criteria; so are clauses embedded under "want",
"suggest" and "ask". However, APIs are ungrammatical in the former but
grammatical in the latter. Absolute veridicality thus fails to draw the
desired distinction between environments which license APIs and those
which don't. Giannakidou therefore replaces it with a notion of
relativized veridicality, which is defined using her notion of an
individual's model and a relativistic definition of truth.
An individual's model is a set of worlds associated with a particular
individual by some epistemic relation. What Giannakidou calls the belief
model of an individual is the set of worlds compatible with what that
individual believes at the actual world. Similarly, the dream model of
an individual is the set of worlds compatible with what that individual
dreams in the actual world, and the model of reported conversation of
an individual is the set of worlds compatible with what the individual
takes the reported conversation to be. (This last case is rather unclear.
I take it that what is meant is the set of worlds compatible with what
the individual believes to have been asserted in the conversation.)
Giannakidou takes a linguistic context to include a set of such models,
along with specification of the conversational common ground, the speaker,
the hearer, and the time, place and world of utterance.
The revised notion of veridicality is more or less as follows (p.112,
Definition 6). I have paraphrased slightly for simplicity.
(4) Relativized veridicality for propositional operators
If Op is a monadic propositional operator then:
(i) Op is veridical just in case the truth of [Op p] in a context c entails
that p is true at every world in some epistemic model in c.
Otherwise, Op is non-veridical.
(ii) A nonveridical operator Op is antiveridical just in case the truth of
[Op p] in a context c entails that NOTp is true at every world in some
epistemic model in c.
(iii) Epistemic models are: belief models, dream models, and models of
Let us set aside here the fact that the new notion is defined only for
propositional operators, and assume some appropriate extension applicable
to other embedding verbs and also to unembedded clauses.
The definition relies on Giannakidou's notion of truth-in-a-context. She
asserts that "sentences are not true or false in isolation, but they are
true or false with respect to an individual's epistemic state" (p.33,
Sec.1.3.3). Based on this, she gives the following truth definition for
(5) Truth for unembedded assertions
(i) p is true in context c iff the belief model of the speaker is such that
for all worlds w in that model, p is true at w
(ii) p is false in context c iff the belief model of the speaker is such
that for all worlds w in that model, p is false at w.
I am uncertain as to whether this is a useful notion of truth. The upshot
of the definition is that a proposition uttered in a context is true in
the context just in case the speaker believes it, and false in the context
just in case the speaker does not believe it. However, speakers can
unwittingly utter falsehoods, and can also accidentally speak the truth
while intending to lie. Giannakidou apparently intends to distinguish
between truth in a context and truth at a world, and so perhaps
would say that a proposition can be false with respect to the actual world
but true with respect to the context (to account for unwitting utterance of
falsehoods) and vice versa (to account for the opposite case). But some
clarification of how truth in a context relates to truth at a world is
needed. This relativistic notion of truth raises a multitude of further
questions, and has many far-reaching consequences. Unfortunately,
Giannakidou provides rather little justification for the view, and does
not discuss its consequences beyond its application in her account.
Perhaps, though, we could sidestep this problematic issue. Recall that
the move to relativized veridicality is required only because "believe",
"dream" and "say", which are non-veridical in the absolute sense, do not
license APIs. Suppose, then, that we assume a disjunctive notion of
veridicality: an expression will be veridical just in case it is veridical
in either the absolute sense or in the relativized sense. We will assume
standard definitions of truth and entailment, according to which unembedded
clauses will be veridical in the absolute sense. And we will follow
Giannakidou in assuming semantics for "believe", "dream" and "say"
according to which these verbs will be veridical in the relativized sense.
But I remain unconvinced that relativized veridicality is a robust notion.
Its definition relies on the assumptions of possible world semantics, and
indeed on particular approaches within that framework to the semantics of
the relevant verbs. If one attempts to recast the notion in theory-
neutral terms, the result is worrisome. Consider, for instance, the case
of "believe", as in:
(6) Theodora believes that it is raining
"Believe", or the clause embedded under it, is taken to be veridical in the
relativized sense because, assuming a particular possible world semantics
for belief, the truth of sentence (6) entails that "It is raining" is true
at every world in Theodora's belief model. But to say that "It is raining"
is true at every world in Theodorass belief model is merely to express in
the terminology of possible world semantics that Theodora believes that it
is raining. Put in these terms, it appears that we are to take "believe"
to be veridical because the sentence "Theodora believes that it is raining"
entails that Theodora believes that it is raining.
The final clause of the definition of relativized veridicality is a
further cause for concern. It stipulates that only three types of model
are relevant to determining veridicality: belief models, dream models,
and models of reported conversation. This ensures that if there were an
additional operator which satisfied clause (i) by virtue of some other
epistemic model (for instance, a desire model), that operator would not
count as veridical. No justification is given, though, for why these
models and no others should count, or why these three form a natural class.
The motivation for selecting these models seems to be the observation that
APIs are licensed in the scope of "believe", "dream" and "say". But then
we cannot use this definition of veridicality to account for the licensing
of APIs in these environments without incurring a circularity.
Giannakidou's presentation of the data and her informal discussion
convince me that there is a connection between the occurrence of PIs
and non-veridical environments, understood in the absolute sense first
discussed above. Given this, the puzzling question is why some
environments which are non-veridical in that sense fail to license PIs.
Giannakidou attempts to solve this puzzle by her notion of relativized
veridicality. Although I do not think that this attempt is successful,
Giannakidou has certainly drawn attention to a host of interesting
linguistic facts about polarity sensitivity and its relation to
veridicality in the absolute sense.
Groenendijk, J., M. Stokhof and F. Veltman (1996). "Coreference and
Modality". In S. Lapin (ed), The Handbook of Contemporary Semantic
Theory. Cambridge, Mass.: Blackwell.
Linebarger, M. (1980). The Grammar of Negative Polarity. PhD Dissertation,
--------- (1987). "Negative Polarity and Grammatical Representation."
Linguistics and Philosophy 10: 325-387.
Mandy Simons is an Assistant Professor in the Department of Philosophy at
Carnegie Mellon University. Her research is in formal semantics and
pragmatics, with a particular focus on topics which lie at the boundary
of the two. To date, her work has dealt with anaphora and presupposition,
and the semantics and pragmatics of disjunction.