Scope without Syntax: A Game Theoretic

Approach
Luke M. Smith
August 30, 2018

Abstract
Here I argue that the commonly (and uncommonly) known facts about
the availability of quantifier scope interpretations fall out cleanly from
communicative constraints which Speakers and Hearers tactically navi-
gate to converge on the intended meaning of an utterance. This allows
a relatively complete and motivated theory of quantifier scope ambiguity
wholly without the need to resort to syntactic structure per se for the
main data. I model this theory game theoretically (von Neumann and
Morgenstern 1944), in an extensive game (Hart 1992) where speakers re-
ceive a payoff for successful communication, and decrements to payoffs
for the use of marked constructions. These assumptions are sufficient to
account for classical scope ambiguity data, but also newly compiled data
I present which argues crucially that word order rigidity, across languages
and constructions is the cause of scope ambiguity.

Contents
1 The Uniqueness of Quantifier Scope 2

2 Assumptions 3
2.1 Preference for Surface Scope . . . . . . . . . . . . . . . . . . . . . 3
2.2 Transformations are “marked” . . . . . . . . . . . . . . . . . . . . 4
2.3 Scrambling is not costly . . . . . . . . . . . . . . . . . . . . . . . 4

3 Basic English Data and Passives 4

4 Model 5
4.1 The Game Theoretic Core . . . . . . . . . . . . . . . . . . . . . . 7
4.2 Scrambling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
4.2.1 Scope in Scrambling Languages . . . . . . . . . . . . . . . 11
4.2.2 An Account of Scramblible Scope . . . . . . . . . . . . . . 11
4.3 The Generalization . . . . . . . . . . . . . . . . . . . . . . . . . . 13

1
5 The Generalization Applied 14
5.1 Flexibility of Negation . . . . . . . . . . . . . . . . . . . . . . . . 14
5.2 Construction-specific Rigidity . . . . . . . . . . . . . . . . . . . . 16
5.3 Why passivize when you can scramble? . . . . . . . . . . . . . . . 18

6 Theoretical Issues Solved and Opened 18

6.1 The Gambit of Linear Order . . . . . . . . . . . . . . . . . . . . . 18
6.2 Scope Interpretations are Not Licensed, but Pruned . . . . . . . 19

7 Towards a general, game theoretic theory of quantifier scope 20

7.1 Universal vs. existential quantifiers . . . . . . . . . . . . . . . . . 20
7.2 some vs. a vs. one; and other scope preferences . . . . . . . . . . 21
7.3 Empirical extensions . . . . . . . . . . . . . . . . . . . . . . . . . 21

8 Closing 22

List of Figures
1 The sequence of player choices . . . . . . . . . . . . . . . . . . . 6
2 The strategies for each player and their abbreviations . . . . . . 7
3 The extensive game in English . . . . . . . . . . . . . . . . . . . 8
4 A primer on extensive game theory notation as in Figure 3 . . . 8
5 Scramble dominates P assive as a strategy for Speaker . . . . . . 12
6 The extensive game in a scrambling language . . . . . . . . . . . 12
7 Empirical generalizations: Rigidity → Ambiguity . . . . . . . . . 17

1 The Uniqueness of Quantifier Scope

Quantifier scope ambiguity is often treated as a syntactic phenomenon, or at
least a syntactic epiphenomenon, but it bears some fairly marked traits that
make a complete theory of quantifier scope incompatible with a rigorous formal
approach to syntax.
Firstly, quantifier scope is known to be highly sensitive to linear order. This
is a generalization true generally, but especially true of languages with free word
order (Pafel 2004). Karimi (2003, p. xix) in fact notes that one of the “specific
properties” of scrambling languages is that “scope is usually determined by sur-
face (derived) positions of quantifiers.” However in a narrowly Minimal inter-
pretation of language, the phonological representation of a sentence, including
the linear order of words, should be incidental, but in any case not a factor that
contributes to or limits a sentence’s meaning.
At that, while it’s very common for syntacticians to evoke scope judgments
in arguments about syntactic features, there is no systematic metric for what
syntactic operations trigger scope judgments or vice versa. Refer to the litera-
ture ensuing after Han, Lidz, and Musolino (2007)’s interesting study in which
it was argued that Korean speakers internalize two different grammars (due to

2
poverty of stimulus) and as such exhibit different quantifier scope judgments.
One population of speakers allowed for an additional type of ambiguity that
the other population consistently rebuffed. That said, the debate arose as to
which population was the one with the extra reading: did low verbs create am-
biguity, or did verbs in T? While an empirical difference had been noticed, and
formal tools existed to account for some distinction, the lack of systematic use
of quantifier scope in syntactic theory makes references to scope only incidental.
At that scope is simply empirically problematic if we assume it is an output of
the syntax. While some sentences are acceptable and some clearly unacceptable,
scopal readings are notoriously unclear, context-dependent and generally victims
of whim.
For these reasons, it’s sensible to look at scope as a phenomenon conditioned
not so much by syntax, but by pragmatics and context. This is my goal. Using
Game Theory (Hart 1992; von Neumann and Morgenstern 1944) I will attempt
to craft a model of what communicative factors are sufficient to account for
different possible quantifier scope readings.

2 Assumptions
Before proceeding, I’ll make some a priori assumptions about scope interpre-
tations. We will see that much of the diversity of scope can be accounted for
merely taking these as assumptions interacting with each other.

1. Speakers and listeners prefer for quantifiers to be in “surface scope” order.

2. “Transformations” classically named (e.g. passives) are “costly” or “dis-
preferred” in some sense.
3. “Scrambling” in languages which exhibit it, is not similarly costly.

Before explaining the model, it’s at least worth justifying all three of these on
functional grounds. These are simply just so justifications of these assumptions,
none of which are fundamental to our argument here. It is not important why
these assumptions are true, only that they are decent priors for our model here.

2.1 Preference for Surface Scope

As noted above, languages have a general preference for surface scope interpre-
tation, meaning that ceteris paribus, if there are two quantifiers in a sentence,
the first will tend to scope over the other. This can be justified in various ways,
but perhaps most sensible is the idea that sentences are processed more or less
linearly, and as words are interpreted, the language faculty or cognition gener-
ally proceeds to generate a mental schema or image for the sentence (similar to
the ideas in Langacker (1987)).
A sentence with surface scope will make this process easier, as otherwise one
would have to revise one’s mental image when the truly wide scope is accessed
later in the sentence.

3
2.2 Transformations are “marked”
With respect to transformations, a longstanding assumption of transformational
grammar was that those utterances which are “transformed” are in some ways,
more marked or at least are derived from simplex expressions.

(1) a. Caesar crossed the Rubicon.

b. It was Caesar who crossed the Rubicon.
c. The Rubicon was crossed by Caesar.

That is, while all sentences in (1) share a semantically equivalent kernel, (1a)
is in someway more basic than the cleft (1b) or the passive (1c). Earlier ideas in
Generative Grammar assumed that this was because (1b) and (1c) were literally
derived from (1a), thus leading to the psycholinguistic thesis of the Derivational
Theory of Complexity, hypothesizing that the latter two sentences were marked
as they are later formations of the first.
While theories such as this have fallen out of favor in mainstream Genera-
tive Grammar, data from acquisition does indeed show that sentences such as
(1b) and (1c) are acquired and employed at a distinctly later stage of language
development.
Still, for our purposes, we could say something as simple as (1b) and (1c)
simply contain more morphemes or words than (1a). We should be clear that
the particular nature of the cost of “transformations” isn’t important for us,
only the general assumption that they are marked or dispreferrable.

2.3 Scrambling is not costly

This may fall out for free depending on one’s account of the costliness of transfor-
mations. For example, if we say that the passive is generally disprefered because
it consists in adding additional morphemes and words to a clause, we implicitly
say that scrambled sentences, since they require no additional morphology or
periphrasis, are not similarly costly. One could make similar arguements based
on frequency of construction or other discourse factors.

3 Basic English Data and Passives

With the background established, we can move to address data and spelling
out this model. First note some rudimentary scope facts in English represented
below:

(2) a. Every man saw a girl. (∀ > ∃; ∃ > ∀)

b. Everyone speaks two languages. (∀ > 2; 2 > ∀)
(3) a. A girl was seen by every man. (∃ > ∀; *∀ > ∃)
b. Two languages are spoken by everyone. (2 > ∀; *∀ > 2)

4
 Generally, unmarked active “kernel” sentences like (2a) and (2b) demon-
strate fairly robust scope ambiguity. Thus, (2a) can mean either that there
is one particular girl that every man saw (∃ > ∀) or that each man saw a
(potentially different) girl (∀ > ∃).
In the “transformed” passive equivalents of the two sentences, however, syn-
tactic ambiguity becomes unavailable and surface scope is typically the only
sensible reading, as below.

(4) a. A man ate every watermelon. (∃ > ∀; ∀ > ∃)

b. Every watermelon was eaten by a man. (∀ > ∃; ??∃ > ∀)
(5) a. Everyone loves someone. (∀ > ∃; ∃ > ∀)
b. Someone is loved by everyone. (∃ > ∀; ??∀ > ∃)

Thus (3a) is only true if there was only one girl in question, while (3b) is
only true if all of the people speak the same two particular languages. We see
that this scopal alternation generally holds across most active/passive sentence
pairs.1
It should be said that while (4b) and (5b) are labeled as requiring surface
scope, there are indeed situations where the inverse readings are possible or
required. While when some traditional accounts of scope categorically rule these
sentences out, for me, what is important is that the inverse passive readings are
simply highly dispreferred. As our analysis will show, there is nothing formally
syntactic that makes these sentences essentially bad, but merely the results of
the game theoretic analysis. The important fact here is merely that English
passives, without other context strongly imply only surface scope.

4 Model
It’s important to be clear about the game theoretics of communication. Re-
solving scope ambiguities like that in (2) is a kind of coordination game with
imperfect information. That is, two interlocutors must converge on the same in-
terpretation of a sentence which has been decided by circumstance; the Speaker
knows the required interpretation and in speaking attempts to signal it to the
Hearer. Both players “win” if the Speaker is succesful in leading the Hearer to
the correct interpretation. If the speaker intends for a universal quantifier, such
as every, to scope over an existential one, such as some, he ideally must word
his sentence in such a way to communicate this.
To be more specific, we can say this is a kind of 3-player coordination game
(where the sequence is indicated in Figure 1). The non-human Player 0, who
we can term Nature dictates what the needed scope interpretation should be,
that is, whether, based on circumstances, a subject quantifier must scope over
an object quantifier, or vice versa or any other combination of quantificational
1 Some exceptions will be discussed later.

5
 Player 0: “Nature”
Determines desired quantifier scope interpretation s

Player 1: “Speaker”
Determines a best wording w to communicate s

Player 2: “Hearer”
Must guess s based w

Figure 1: The sequence of player choices

elements.2 Nature represents the needed discourse environment which neither

the Speaker or Hearer determines themselves. We can think of the choice of
Player 0, Nature, as a random or given element of the model. Player 1, the
Speaker, knows what the needed scope given by Nature is, and has to encode a
linguistic message to communicate it to Player 2: the Hearer who is ignorant
of it (illustrated in Figure 3). In each language, the Speaker might have different
strategies usable to communicate this message, depending on the structure of a
language.3
If the Hearer’s interpreted scope and the scope selected by Nature match,
both human players get a payoff (represented as c for communication), while
if they do not match, there has been a miscommunication and this there is
no payoff. Similarly, we can formalize our assumptions in the Section 2. If
transformations, specifically passives, are costly or dispreferred in some way, we
can model them as saying that they reduce the payoff to the speaker to some
degree (designated by −p). Similarly, instances of free word order and so-called
“scrambling” do not reduce any payoff; this will be addressed in Section 4.2.
Lastly, surface scope should be universally preferred, so in situations where a
speaker and hearer settle on an inverse scope reading of a sentence, their payoffs
are reduced by some degree (symbolized by −i).
We should also assume that c > (p + i), meaning that it is always preferable
for the two interlocutors to understand each other even if transformations and
inverse scope may dig into that payoff. Also, p and i are not necessarily larger
or smaller than one another, and may vary from situation to situation, meaning
that in some situations, it may be preferable for a speaker to vie for inverse
2 We will only be including subject and object quantification for sake of simplicity to outline

some general principles.

3 Thus, our intent here is not to argue how or why a given sentence in a language is

grammatical or acceptable, but why Speakers choose to use a given sentence to communicate
a particular scope reading.

6
 Abrv. Strategy name
A Active Voice (Speaker)
P Passive Voice (Speaker)
S Interpret Surface scope (Hearer)
I Interpret Inverse scope (Hearer)
Sub > Obj Demand the subject scope over the object (Nature)
Obj > Sub Demand the object scope over the subject (Nature)

Figure 2: The strategies for each player and their abbreviations

scope rather than performing a transformation, while in others, it opposite may

be true.
The particulars of an instantiation of the game are based in the structure
of whatever language is being spoken. For any given language, there will be a
different set of syntactically valid utterances that the Speaker can use to signal
to the Hearer what the intended scope is. Given the constraints posited above,
we will assume that surface scope is the ideal, but different languages differ in
their abilities to use different transformations or scrambling.

4.1 The Game Theoretic Core

Now, given our constraints and model above, we can derive these facts of the
difference between actives and passives from the interaction of strategic inter-
pretation on the part of the two interlocutors. Assume the three player game
above (of Nature, a Speaker and a Hearer), dealing with the kernel sentence in
(5a) “Everyone loves someone,” depicted in the decision tree in Figure 3.
First, the Nature player determines whether the intended scope of the ut-
terance should be where the universal quantifier takes wide scope (∀ > ∃, or
Sub > Obj) or where the existential does (∃ > ∀, or Obj > Sub). Then the
Speaker, aware of Nature’s choice, takes his turn choosing either to word the
sentence as the active “Everyone loves someone” or the passive “Someone is
loved by everyone”. Lastly, the Hearer, unaware of Nature’s original decision,
chooses whether to interpret the sentences with surface scope or inverse scope.
To repeat, if the Hearer guesses the correct scope as defined by Nature, both
the Speaker and Hearer receive a payoff of c. If the players fail to engineer this,
both will receive no payoff. Additionally, because a passive transformation is
“costly,” the payoff of the Speaker will be deducted by p whenever he chooses to
produce a passive. Lastly, if the dispreferred inverse scope has to be employed,
both speakers will have a penalty of i.
Since this decision tree involves three players with substantive choices, it
helps to narrow down the decision to find Nash Equilibria, Schelling Points
or notable strategies. Let’s put ourselves in the position of the Hearer. The
Hearer is the one dealing with the informational asymmetry guessing the choice
of Nature given the Speaker’s utterance. Given the aforementioned decision
tree in Figure 3, the Hearer can make two hypotheses about Nature, that it

7
 Nature
Sub > Obj Obj > Sub

Speaker Speaker

A P A P

Hearer Hearer Hearer Hearer

S I S I S I S I
( ) ( ) ( ) ( ) ( ) ( )( ) ( )
c −i −p c−p−i 0 c−i c−p −p − i
c −i 0 c−i 0 c−i c −i

Figure 3: The extensive game in English

Notes on the Game Tree

Each branch represents a player choice. Payoffs to the Speaker and the Hearer are at
each terminal node, the Speaker’s being on top and the Hearer’s below.

The dotted lines represent the equivalence classes for the Hearer. That is, they unite
the nodes that the Hearer can not distinguish. To be clear, if the Speaker chooses the
“Active” strategy, the Hearer knows he must be on the nexus either first or third from
left, but can’t distinguish between the two since he is unaware of Nature’s original
choice.

Figure 4: A primer on extensive game theory notation as in Figure 3

8
Game Theoretic Terms
Imperfect information – When at least one player is not perfectly aware of the
actions of another. In our game, the Hearer is not aware of the decision of Nature,
ergo this is a game with imperfect information.
Incomplete information – When at least one player is not aware of the payoffs for
a player. Although this concept is frequently confused with imperfect information, we
do not have a incomplete information game.
Signalling – When a player voluntarily undergoes a costs to communicate his tactics
or strength. For example, a zebra, noticing a stalking lion may “irrationally” jump up
and down in place instead of running to show his spriness and tell the lion he isn’t an
easy target.
Nash Equilibrium – A point in a game where no player can improve his position
by changing strategies given what he knows. While Nash Equilibria are usually the
MacGuffin of game theoretic analysis, our game has no proper Nash Equilibrium (not
assuming iteration). The concept is usually attributed to Nash (1950) (hence the
name), but was originally formulated at least in Cournot (1838)’s theory of economic
duopoly.
Schelling Point – Sometimes called a Focal Point. A point which is not a Nash
Equilibrium, but due to some meta-game reasoning is a particularly marked. Originally
formulated in Schelling (1960).
Equivalence Classes – In an Imperfect Information game (like this one) nodes in a
decision tree that a player cannot distinguish due to his imperfect information.

chose to demand that the subject/agent scope over the object/patient, or that
it demands that the object should scope over the subject.
The key to the strategy is the cost of the passive p. Let’s take the situation
when Nature selects Sub > Obj. In that case, the hypothetical active form
“Everybody loves someone” already has the correct surface scope order. While
it is not immediately sure that the Hearer would determine that this active
clause is indeed the required order, it costs the Speaker neither decrements of p
or i.
If the Speaker were to passivize the sentence to “Someone is loved by ev-
erybody”, not only would he be incurring the loss of p for the passive trans-
formation, but if the Hearer did guess correctly that the sentences should have
inverse scope in this reading, both players would additionally be losing i.
The Speaker therefore is in a position of two theoretically uncertain out-
comes, one that can yield him c, while the other can yield him only c − p − i.
All things considered, c is preferable, and therefore using the active sentence
to express Sub > Obj should be preferable. While this is not a proper Nash
Equilibrium, since we are dealing with a non-simultaenous game, this decision
can act as a signal to the second player, the Hearer.
The Hearer, knowing that there is this Schelling Point for choosing the active
sentence when given Sub > Obj can therefore conclude by deduction that if the
Speaker for some reason chooses to word his sentence in the passive, it is nearly

9
certain that Nature meant the other alternative: Obj > Sub. Or put more
generally in (6).
(6) A speaker will not engage in a costly transformation which yields an
undesired scope order.
Or to spell (6) out more specifically in our context, see (7).
(7) The use of a costly reordering transformation, ceteris paribus, entails
that the underlying object should take wide scope over the subject. Or
put another way, scopal ambiguity dissappears in favor of surface scope
after a costly transformation.
To put it in more intuitive terms, if the subject does something costly like
passivization to a sentence, he is doing it for a reason, specifically here to avoid
the other loss of i. Passivizing only to also lose i is not a good Schelling Point
strategy. For this reason, in most pragmatic circumstances, passivized sentences
appear as unambiguous, seeing that we conclude that they are motivated to
avoid the cost of the inverse scope.
In the situation where Nature chooses Obj > Sub, the situation is less clear.
This is because the Speaker has two possible winning payoffs: c − i and c − p,
neither of which is necessarily preferable since we have not established whether
p > i, nor do I think one is always larger than the other. In this situation,
a Speaker could passivize and to avoid inverse scope order, or bite the bullet
and take inverse scope without the passivization, both with uncertainty. The
end result is that the active sentence “Everyone loves someone” does not clearly
communicate whether Nature choose Sub > Obj or Obj > Sub since there is no
obvious Schelling Point to rule out one of the strategies. Therefore, while the
English passive is unambiguous due to the presence of a Schelling Point, the
English active is not.

4.2 Scrambling
But how should scope ambiguities work where there are “costless” ways of re-
ordering quantified nominals? Scrambling languages present ways of reordering
nominals without a marked transformation, as we have assumed in Section 2.3.
In our model, Speakers in languages like this, such as German, Persian, Korean
and Japanese, have access to another strategy aside from producing an active
or passive clause. They may also scramble the object such that it appears to
the left of the subject.
First a theoretical note. The “scrambling” tendencies of each of these lan-
guages may be different: German “scrambling” is quite different syntactically
than Korean’s, etc. This is not so important to us here. We only need to know
if there is a valid reordering strategy in a language which is not marked in the
way that passives are. Why German or Persian or other languages vary with
respect to syntactic flexibility is not germane for us here, only the effects of
these traits on scopal possibilities. Our account here is not a theory of syntax,
only of quantifier scope.

10
4.2.1 Scope in Scrambling Languages
First the empirical generalizations. German, a scrambling language shows a very
different paradigm of scope availabilities than does English. Even in “kernel”
sentences like (8), surface scope is the only plausible interpretation. The same
is true in the scrambled sentence (9), where the object has been scrambled left
of the subject.

(8) dass eine Frau jeden liebt

that a woman everybody loves
“…that some woman loves everyone” (some > every; ??every > some)
(9) dass jeden eine Frau liebt
that everybody a woman loves
“…that a woman loves everyone” (every > some; ??some > every)

This universal surface scope is well mirrored in other languages. To repeat,

as Karimi (2003) and Pafel (2004) among others note, the general tendency is
for scrambling languages We can see similar patterns in Persian in (10).

(10) a. Har dāneshjui bāyad ye mas’ala-ro hal bo-kon-e.

each student must a problem-rā solution SUBJ-do-3sg
“Each student must solve a problem.” (∀ > ∃, ??∃ > ∀)
b. Ye mas’ala-ro har dāneshjui bayad hal bo-kon-e.
a problem-rā every student must solution SUB-do-3sg
“There’s a problem that every student must solve.” (∃ > ∀, ?∀ > ∃)

It should be noted that Karimi (2003) treats the latter sentence as ambigu-
ous, a judgment I could not replicate with my consultants in context. Still the
different judgment cited in Karimi (ibid.) is understandable given that the sur-
face scope order (∃ > ∀) logically entails the other (∀ > ∃), and the use of har
“every”, a quantifier with strong pragmatic preference for “wide scope”. We will
discuss this as a more general tendency in section 7.2.
Regardless, this generalization about scope effects is fairly well-accepted with
respect to scrambling languages, but it can also be shown to fall out from our
already mentioned assumptions.

4.2.2 An Account of Scramblible Scope

In our game theoretic framework, we can say that the Scramble strategy, which
consists of moving the object left of the subject, achieves the linear order of
passivization without the cost of p to the Speaker (Section 2.3). Because of
this, Scramble in a scrambling language is always preferable to the dominated
strategy Passive for the Speaker; the payoffs are shown in Figure 5. Therefore, a
rational Speaker need not consider Passive as an option for achieving the correct
scope interpretation if the alternative strategy Scramble is available. We will
discuss some typological implications of this in Section 5.3.

11
 Sub, S Sub, I Obj, S Obj, I
Active c −i 0 c−i
Passive −p c−p−i c−p −p − i
Scramble 0 c−i c −i

For every pair of decisions made by other speakers, each payoff for the Speaker is
greater if he chooses Scramble than if he chooses P assive, due to the −p penalty.

Figure 5: Scramble dominates P assive as a strategy for Speaker

Nature
Sub > Obj Obj > Sub

Speaker Speaker

A Sc A Sc

Hearer Hearer Hearer Hearer

S I S I S I S I
( ) ( ) ( ) ( ) ( ) ( ) ( ) ( )
c −i 0 c−i 0 c−i c −i
c −i 0 c−i 0 c−i c −i

Figure 6: The extensive game in a scrambling language

12
 Thus if we disregard the possibility of passivization as disprefered as a scope
technique, once again, both human players have two possible choices. We should
see that there is no straight-forward dominant strategy for either player, but
a very obvious signalling opportunity arises in the meta-game, or a Schelling
Point.
Specifically, independent of the Speaker and Nature’s choices, the Hearer will
want to avoid choosing the Inverse strategy. Given there is no Nash Equilibrium
in the simple game, there is no clear dominant strategy for the Speaker, but
even if a Hearer interprets his behavior as totally random, choosing the Surface
strategy increases the average payoff for the Hearer. The Speaker realizes this
(and shares the desire to avoid −i) and can strategically select his strategy based
on what will require the Hearer to not select the Inverse strategy. This acts
as a signal to the Hearer, or more generally a positioning strategy to move the
Hearer in to a place to get the best mutual payoff of c.
In the meta-game, the Speaker acts so that both players can be applicable
for the highest possible payoff of c. And a Hearer totally blind to the Speaker’s
actions should have a bias to the Surface scope interpretation strategy.
If the Hearer or Speaker violate this meta-strategy, they would be subject
to a decrease in expected returns over time, independent of the other players
actions. Therefore in a language with the free movement of nominals, we should
predict that Hearers should only try to interpret sentences in surface order in
normal situations, and that speakers should scramble or not depending on which
produces a sentence which gives the correct scope interpretations with a surface
scope reading.
All in all, the free availability of a costless movement makes avoiding −i
the only possible constraint, meaning that all of the choices the Speaker makes
should be assumed to avoid −i. This simply means that surface scope, as the
empirical judgments have shown, should be preferred at all times in scrambling
languages.

4.3 The Generalization

We can sum up the generalization of this game theoretic analysis of both English-
like and scrambling languages below in (11).

(11) Wherever there is free and costless word order, scope ambigu-
ities need not arise, but where word order is inflexible, scope
ambiguities occur.

This generalization simply falls out from the analysis we have outlined, and
we can widen the scope and look at other kids of scope ambiguities to see similar
effects.
Before that, just a restatement of the intuitions in intuitive terms. Hearers
assume that sentences with free word-order are always surface scope because,
due to the free word-order, the speaker could’ve put the words in another ideal
surface scope reading if such reading had been intended. On the other hand, in

13
English-like languages, costly transformations are unambiguous because hearers
assume that speakers would not have engaged in costly transformations unless
they intended the sentence to be in a special surface scope order.
But ambiguity arises in English-like languages when a sentence like “Every-
one loves someone” is produced. This is because hearers can say, “Ah, that may
just be the desired reading in surface scope, or perhaps it is a suboptimal order,
and the speaker didn’t want to undergo a costly transformation”.
Now our initial assumptions have accounted for much variation between
different languages, but there is scopal differences inside of languages between
different constructions that is worth outlining and accounting for in this novel
way.

5 The Generalization Applied

In the Generative Program, part of the common idea of quantifier scope differ-
ences between languages has been that there are parametric differences between
languages that not only cause syntactic differences, but also these scopal differ-
ences.
One language, due to a syntactic paramter, may have ubiquitous ambiguity
due to some parameter setting affecting “Logical Form”, one might have the
reverse.
I’ll argue that this conception is untenable, not just because of the better
account we can get from this type of game theoretic and pragmatic model, but
also because there are many examples of “local rigidity” which, in the same way
that English syntactic rigidity produces ambiguity, produce ambiguity only in
particular constructions in languages. Quantifier scope availability, therefore,
cannot be a language parameter setting, and must be grounded in the very
specific context of a construction, as I will show our theory here is.

5.1 Flexibility of Negation

We can take the generalization in (11) and compare it to the flexibility or rigidity
of non-nominal quantificational elements as well.
English expresses sentential negation in the element not. As a descriptive
generalization, not may occur only after a modal or another auxiliary. In normal
discoursive situations, it may not occur after main verbs or before a modal.
Many attempts have been made to describe and justify the specifics of these
facts. We will not address them here, but assume the empirical facts as given
syntactic constraints and proceed.
On to the scopal facts. Notice first that an English sentence with one modal
and one negation produce ambiguity.

(12) Billy can not go. (¬ > can; can > ¬)

(12) is ambiguous. Negation can take wide scope (which is inverse) such that
Billy is unable to go, or the modal can take wide (surface) scope, where Billy is

14
able not to go, if he so pleases.
In keeping with our assumptions, we can say that ambiguity arises because
the following order in (13) is syntactically invalid for other reasons in English.
(13) * Billy not can go.
Since (13) is syntactically ill-formed, we cannot, by normal syntactic means
force negation to linearly scope over the modal, thus its parallel sentence (12)
can be assumed to be a suboptimal enunciation of the meaning of an intended
(13). If we imagine a hypothetical “negation scrambling” language where the
equivalent of (13) is available, (12) should be unambiguously can > ¬.
Now that is the situation of modal and negation scope with one non-main
verb. However as inferred previously, where there are multiple auxiliaries, not
may freely occur after any one. Syntactic flexibility should reduce or eliminate
the possibility of ambiguity. This is the case as below.
(14) Billy could not have gone before we arrived.
(15) Billy could have not gone before we arrived.
Notice as there is flexibility of negation position with non-modal auxiliaries
in English, neither (14) nor (15) are ambiguous. In (14), we express the fact
that Billy was unable to go before our arrival. In (15), we express the possibility
Billy was able to not go, but in a world where Billy did go, (15) may still be
true. No one would utter (15) to mean that the negation takes wider scope
when (14), which places the negation more leftward, is a possible alternative.
Thus even in a single language our generalization holds. Syntactic rigidity
allows for ambiguity, while free flexibility creates situations where ambiguity
is ruled out due to the assumption that speakers have that surface scope is
universally preferred.
This is not just true from language to language or construction to construc-
tion, but in English, even when specifically addressing negation, any highly local
syntactic rigidity causes ambiguity and any highly local syntactic flexibility dis-
ambiguates.
And as expected, languages that can syntactically bear negation before all
modals/verbs, such as Chinese do not create the ambiguity in the rigid English
example.
(16) Shujuan keyi bu gen Guorong tiao wu. (Ernst 1998, p. 109)
S. may not with G. dance
“Shujuan may not dance with Guorong.” (may > not; *not > may)
(17) Shujuan bu keyi gen Guorong tiao wu.
S. not may with Guorong dance
“Shujuan may not dance with Guorong.” (not > may; *may > not)
The Persian situation is particularly interesting. In most situations, while
noun scrambling is mostly free, scrambling of the verb and its negation is more
marked. This manifests in that inverse scope is very possible in positions in-
volving a negation interfacing with another quantifier.

15
 (18) Yek dāneshjui ān ketāb-i-ro na-xānd.
one student that book-ACC not-read
“A student didn’t read that book.”

As we would predict, (18) is ambiguous. It can mean either a certain student

didn’t read the book (∃ > ¬) or that not one student read it (¬ > ∃). This
ambiguity arises because the movement of the verb is more marked.
In other situations, particularly in movement verbs, the Persian main verb
becomes more flexible. SVO order, where the negation is still a pre-verbal
clitic, is common with some movement verbs, and as expected, the ambiguity
evaporates in (19) and (20).

(19) Billy na-raft hame shahr-i.

B. not-went all city-IND
“Billy didn’t go to every city.” (¬ > ∀; *∀ > ¬)
(20) Billy be hame shahr-i na-raft.
B. to all city-IND not-went.
“Billy didn’t go to any city.” (∀ > ¬; *¬ > ∀)

Thus Persian movement verbs, which are uniquely more syntactically mobile
show precisely the same scopal effects we predict they should.

5.2 Construction-specific Rigidity

Since my statement here is that scope ambiguity is merely the result of lin-
ear rigidity in syntax, not of some language-wide parameter, we should see
the unambiguous surface scope of scrambling languages disappear in particular
constructions where normally scramblible nominals are tied in position.
Chinese, usually a very stablely scrambling or discourse configurational lan-
guage generally allows the low cost movement of nominals as illustrated in (21).
These sentences, as we should expect are unambiguous and force surface scope.
In (21a), everyone arrests different women, while in (21b), only one woman,
who apparently is a prolific criminal, is arrested.

(21) a. Meigeren dou zhuazou yige nüren.

everyone all arrest a woman
“Everyone arrested a woman.” (∀ > ∃; ∗∃ > ∀)
b. (You) yige nüren meigeren dou zhuazou.
(have) a woman everyone all arrest.
“A woman was arrested by everyone.” (∃ > ∀; ∗∀ > ∃)

However Chinese bei pseudo-passives require a particular word order. They

are rigid. The semantic object is promoted as the initial nominal, while the
agent follows the preverbal co-verb “bei” as shown in (22a). As (22b) shows,
however, the quasi-prepositional bei + agent constituent may not be fronted or
topicalized.

16
 Rigid Constructions Flexible Constructions
English main clauses Main clauses in scrambling languages
Persian negation Chinese negation
Typical English negation English negation around auxes
Chinese passives English passives*
All of these are ambiguous All of these are non-ambiguous

Figure 7: Empirical generalizations: Rigidity → Ambiguity

(22) a. Meigeren dou bei yige nüren zhuazou.

everyone all PASS a woman arrest
(Aoun and Li 1989, p. 142)

“Everyone was arrested by a woman.” (∀ > ∃; ∃ > ∀)

b. * Bei yige nüren meigeren dou zhuazou.
PASS a woman everyone all arrest

The scopal possibilities follow the predictions perfectly. Since flexible word
order is unavailable, (22a) is ambiguous: there can either be one woman arrest-
ing everyone, or each person can be arrested by a different woman.
The intuitions of our game theoretic model in essense propose that (22a) is
an ambiguous sentence because its alternative sentence (22b) is not available.
On the other hand, because (21a) does have a sensible counterpart (21b) with
a different scope order there is no reason for it to be ambiguous.
Interestingly enough, the scope possibilities in Chinese in normal clauses and
the bei pseudo-passive are precisely the opposite of English, again this falls out
from the fact that nominal movement is generally free in Chinese (meaning un-
ambiguous sentences normally) and the additional fact that bei passives are not
precisely equivalent to their active counterparts, but add additional meaning.
We will discuss this in the next subsection.
For now it should be noted that the data we’ve covered from a variety of lan-
guages move lockstep together: syntactically rigid constructions exhibit quan-
tifier scope ambiguity, while syntactically flexible constructions do not. This
is not a generalization about the differences between languages, but inside lan-
guages as well.
An approach to scope based in traditional syntax would be able account for
the differences between languages, saying that the different scope interpretations
are built into different syntactic parameters, but as we’ve shown here, there are
scope distinctions between individual constructions within a language.
This along with the fact that each of these differences correlate so well syn-
tactic rigidity or flexibility point to a deeper relationship between scope and
syntactic flexibility, a relationship which we have modeled causally above.

17
5.3 Why passivize when you can scramble?
We noted in Section 4.2.2 that our formal analysis treats the Speaker strategy
Passive strategy as a dominated strategy to Scramble in languages where Scram-
ble is available. A decent question for asking is, given our theory, why indeed
should there exist a passive is one can easily scramble words to a more desired
position? Should typical passives exist at all in “scrambling” languages?
First, it’s worth saying that this analysis only predicts that scrambling is a
superior strategy over passivization for the simple purpose of engineering surface
scope order. There may indeed be other reasons to use a passive in language.
That said, I think there is a case to be made that indeed scrambling languages
rely less on passivization and other transformations. Or at least, in scrambling
languages where passives exist, passives find a distinct semantic or pragmatic
niche.
Take the earlier example of Chinese, a scrambling language4 . Chinese indeed
have a construction sometimes called a “passive”, and that is the bei construction
we mentioned in example (22). Like English passives, Chinese passives promote
the object and demote the subject for a different argument structure schema,
and a different order.
However, while the two constructions are similar in form, the Chinese passive
has an added meaning: they imply that the action the patient is undergoing
is particularly undesired or unfortunate. That is, while the examples in (22)
show the word for “arrest”, it would be extremely strange to replace that with
a non-negative word such as “praise”, etc.
While it is not out objective here, it could certainly be argued that while
passives may not be the best strategies for getting a desired scope order, these
constructions may be put to other purposes or develop special meanings in
circumstance. This could feasibly be done in evolutionary game theory, but
that is beyond our purview here.
Still, implicitly, with the approach to language developed here, languages
are bundles of communicative strategies employed for communication.

6 Theoretical Issues Solved and Opened

6.1 The Gambit of Linear Order
It should be noted that the data of scope present an existential problem for
the general interpretation of syntax from a “logical form” perspective. On one
hand, the assumption has been that scope interpretations are read from quan-
tificational elements which interface with the hierarchical structure of language.
This structure (from the Chomskyan perspective) is construed as irrelevant to
the linear order of a sentence, which is a later realization of the expression in
phonological form.
4 Or at least a discourse-configurational language. Regardless, a language with relatively

free word order. What is important for us is not the cause of the multiple constructions, but
the fact that syntactic flexibility is an option.

18
 But the overwhelming reality of scope as a feature of natural language is
that it is manifestly and abundantly tied to linear order, nearly all of the data
presented here, along with that in the literature testify to this.
I think a proper understand of scope would be that all possible scope readings
of all sentences are theoretically possible at all times. In normal discoursive
situations, however, most possible readings as pruned out as implausible, based
on pragmatic circumstances or world-knowledge. This also would attest well
the conundrum of every syntax class, where graduate students sit around long
enough looking at sentences without context and start seeing all of the scopal
readings after long enough. My analysis here has endeavored to show why some
readings are ruled out in certain situations, although this is no be-all-end-all
solution to scope, precisely because it is a pragmatic, and perhaps extralinguistic
portion of language.
Such a framework would be able to maintain the statement that human
language, at its syntactic core, should be independent of linear order, as the
linear order effects are part of the pragmatic traits of language use and discourse.

6.2 Scope Interpretations are Not Licensed, but Pruned

Again, I have not crafted a universal account of scope ambiguities, and have
deliberately avoid some contradictory examples that I think explicable on other
grounds. Take the sentence pair below.

(23) Every boy ate an apple.

(24) An apple was eaten by every boy.

(23) follows the generalizations we’ve sketched here in, that it is ambiguous
(∀ > ∃, ∃ > ∀). (24), as a passive, is unambiguous, but not in the way we’ve
predicted here, but only inverse scope is allowed (∀ > ∃, *∃ > ∀), or at least,
inverse scope is highly preferred.
What rules out the surface scope interpretation of (24) is not the pragmatics
of passivization per se, but the interface of general world knowledge with the
inherent telicity of the verb eat with a count noun object. The predicate “ate an
apple” implies that the subject totally consumed an apple, but if the universal
quantifier is thought to scope over the existential “an apple,” this would have
to mean that every boy totally ate the same apple as every other boy, which is
logically impossible.
This makes the otherwise disfavored ∀ > ∃ interpretation the only logically
consistent option. If we rejigger the sentence to remove the telicity, as in (25),
we see that the expected scope possibilities return, even when the sentence is
still somewhat strange by that interpretation.

(25) Some of an apple was eaten by every boy. (∃ > ∀, ∀ > ∃)

Note also that if we imagine (24) in a discourse environment, we’re most

likely to think of contrastive focus or something else: “An apple was eaten by
every boy, a banana by every woman, a pineapple by every man…”

19
 I do not consider this a contradiction, but evidence in favor of the wider
point. Scope ambiguities are trimmed away by pragmatic factors. In (24), it is
world knowledge, in most of the other examples here, it’s economy of derivation.

7 Towards a general, game theoretic theory of

quantifier scope
Our empirical domain established here is quite robust in the data problems
we’ve addressed. Specifically, we can make predictions about what kinds of
languages and constructions make ambiguous scope readings available based on
independent factors. We have looked a number of constructions in a number of
languages, but the theory would predict these generalizations will continue to
be vindicated.
Our game theoretic model accounts fairly well for these scopal differences
motivated by the availability of different strategies. I think that this model can
be refined substantially in the future, and I plan to do so, but there’s some sense
in which the intuition is fundamentally “correct”.
That said, there are some domains of scope which I have not endeavored to
account for due to the more focused scope of the my data here. Two of these
domains are the most prominent:

1. The scope differences between universal and existential quantifiers.

2. The scope differences between particular quantifiers of the same “type”,
but generally taking different scope readings.

It’s worth discussing how these problems can be addressed, and how they
can be integrated into a general theory of game theoretic quantifier scope we’ve
begun here.

7.1 Universal vs. existential quantifiers

There have been game theoretic accounts to address these problems, notably
in Clark (2012). Clark models the difference between universal and existential
scopes as being an abstract game between falsifier and verifier algorithms. The
scopal differences between universal and existential quantifiers come about from
how these two “players” proceed to attempt to find a contradiction or vindication
of the truth of a sentence in a Model Theoretic framework.
Based on the linear order of a sentence, we may find a proof or disproof of
a sentence in a Model Theoretic world at different times for either existential
or universal quantifiers. Clark’s model, although a “mere” example in another
more general argument, is simple and highly effective at disentangling the scopal
tendencies of universal and existential quantifiers.
There is a possibility of combining this approach with ours into a linearly
processed incremental game, where the Hearer weeds out possible interpreta-

20
tions with world knowledge accounted for model theoretically, partially based
on cognitive coherence (Langacker 1987).

7.2 some vs. a vs. one; and other scope preferences

Another domain on which our current theory here is insolvent or at least agnostic
is the generally acknowledged tendency for some quantifiers to prefer to take
higher or lower scope (Feiman and Snedeker 2016). Take the pair of sentences
below.
(26) a. Every girl loves a man.
b. Every girl loves some man.
Either of these sentences can be interpreted as being ambiguous (∀ > ∃
or ∃ > ∀), but (26b) with the quantifier some seems to predispose one to be
somewhat closer to the ∃ > ∀ interpretation, where every girl loves one particular
man, say, Billy.
We can see a similar effect with universal quantifiers:
(27) a. Three postmen visited each house.
b. Three postmen visited every house.
(27a) seems to prefer the (∃ > 3) interpretation where at every house there
were three, possibly different postmen. (27a) prefers the reading (3 > ∃) where
exactly three postmen went to every single house.
While the account I have built here does quite well to predict the effects of
transformations and scrambling on scope interpretations, as well as predicting
typological facts, I do not think these narrow differences between the tendencies
of particular quantifiers can be captured in our current formalism.
However, an evolutionary game theoretic account (Maynard-Smith and Price
1973) could do so quite well. In fact, an evolutionary account could also answer
the general question of why languages indeed do have “synonymous” quantifiers
in the first place.
Specifically, given what we’ve modeled here, ambiguity is still a sizeable
problem for communication’s sake. To disambiguate sentences which maintain
ambiguity, language systems gradually evolve a meta-game or, really conven-
tions as to particular quantifiers having particular scope readings.
That is, in English, for example, some and a may both be existential quan-
tifiers, as every and each are both universal quantifiers, but over time, Schelling
Points arise where particular quantifiers are conventionalized as preferring wide
or narrow scope. In English, it seems that both some and each tend to prefer
wide scope while a and every prefer narrow scope.

7.3 Empirical extensions

Even more than formal modeling, I think that our generalization that syntactic
rigidity produces scopal ambiguity is an sensible find that could probably be
replicated on a much larger scale with many more constructions.

21
8 Closing
In closing, much of the confusion about scope can be alleviated by understanding
that scope availabilities are determined by pragmatic factors and implicatures
that can be modelled game theoretically. We’ve seen here that the facts about
the scope availabilities of most languages fall out quite effortlessly from assump-
tions about the cost of transformations, the costlessness of scrambling and the
wider syntactic capacities of a language.
I feel that much more work can be note to resolve questions in scope using
pragmatic facts, particularly in the areas of telicity and world knowledge. Ad-
ditionally, the tools of game theory can prove extremely effective at accounting
for scope and other quasi-pragmatic aspects of languages that have tradition-
ally for better or for worse been modeled as parts of or outputs of the formal
syntactic engine. Regardless, there seems to be good circumstantial evidence to
lend credence to the idea that scope is not a component of narrow syntax, but
a set of extra-UG implicatures we make about language use.
Additionally, other factors of grammar, such as binding in the classical sense
are in need of new life, once insurmountable problems were brought to tra-
ditional syntactic analyses of the data. It may be that these other factors,
binding, negative polarity items and cross-over effects may actually be deriv-
able on pragmatic grounds, and thus would eliminate such of the theoretical
mess and greatly economize and minimize the core language faculty.

References
Aoun, Joseph and Audrey Li (1989). “Scope and Consituency”. In: Linguistic
Inquiry 20, pp. 141–172.
Clark, Robin (2012). Meaningful Games: Exploring Language with Game The-
ory. MIT Press.
Cournot, Antoine Augustin (1838). Researches sur les principes mathématiques
de la théorie des richesses. L. Hachette.
Ernst, Thomas (1998). “Case and parameterization of scope ambiguities”. In:
Natural Language & Linguistic Theory 16.1, pp. 101–148.
Feiman, Roman and Jesse Snedeker (2016). “The logic in language: How all
quantifiers are alike, but each quantifier is different”. In: Cognitive Psychology
87, pp. 29–52.
Han, Chung-hye, Jeffrey Lidz, and Julien Musolino (2007). “V-raising and gram-
mar competition in Korean: Evidence from negation and quantifier scope”.
In: Linguistic Inquiry 38.1.
Hart, Sergiu (1992). “Games in extensive and strategic forms”. In: Handbook
of Game Theory with Economic Applications. Ed. by Robert Aumann and
Sergiu Hart. Elsevier.
Karimi, Simin (2003). Word Order and Scrambling. Blackwell Publishing Ltd.
Langacker, Ronald W. (1987). Foundations of Cognitive Grammar. Stanford
University Press.

22
Maynard-Smith, John and George R. Price (1973). “The Logic of Animal Con-
flict”. In: Nature 246, pp. 15–18.
Nash, John Forbes (1950). “Non-Cooperative Games”. PhD thesis. Princeton.
Pafel, Jürgen (2004). Quantifier Scope in German. John Benjamins Publishing
Company.
Schelling, Thomas C. (1960). The strategy of conflict. Harvard University Press.
von Neumann, John and Oskar Morgenstern (1944). Theory of Games and Eco-
nomic Behavior. Princeton University Press.

23