Offprint from

CURRENT TRENDS
~··
·..

I~ IN LINGUISTICS
I
I Edited by

l THOMAS A. SEBEOK

VOLUME 12

Linguistics and
Adjacent Arts and Sciences

1974
MOUTON
THE HAGUE • PARIS

I
I
I
'.

SEMANTICS AND COMPREHENSION

HERBERT H. CLARK*

INTRODUCTION

Comprehension is fundamental to the study of language. In their investigations,

most linguists have relied heavily on what people understand a sentence to mean,
and their theories have been designed.in part to account for these so-called
'linguistic intuitions'. On the other hand, psycholinguists have more often been
interested in the difficulties people have in understanding sentences, and their
theories have usually been designed to account for these difficulties. More gener-
ally, while linguists have usually been concerned with· the end-product of compre-
hension, psychologists have been concerned with the process of comprehension.
The present chapter reports on an attempt to develop a general theory of compre-
hension that includes selected aspects of both the process and end-product of
comprehension. The purpose of section 1 is simply to outline the theory, its
assumptions, and its justification. The purpose of sections 2 through 5 is to apply
this theory quite specifically to four basic constructions in English - namely,
negatives, locatives, compar~tives, and actives and passives - and to demonstrate
that the theory is adequate to account for the major psychological (and linguistic)
evidence available on these constrnctions. Although the theory is certainly not
complete at the present time, it is useful in the following respects: (1) it makes
sense of the previous experimental work on comprehension in a unified framework;
(2) it points out the gaps in our knowledge and theory of comprehension; and (3)
it prowdes a unified framework for future empirical studies and theories in com-
prehension.
Comprehension will be approached mainly as a problem in semantics, in the
broadest sense of this word. The first problem of the theory is to characterize
what it is that a person knows of the meaning of a sentence once he has 'com-
prehended' the sentence. This is what I will call the representation problem, and it
will be our first concern in the application of the theory to the various English con-
structions. The second problem is to characterize the processes by which a person
uses the 'meaning' of a sentence to do other things - e.g. decide whether the
sentence is true or false, answer questions about that sentence, carry out commands
* I~ indebted to Eve V. Clark for her helpful comments on the manuscript of this chapter.
 1292 HERBERT H. CLARK

made by the sentence, and so on. This will involve us in the processing problem.
The processing problem, however, is very closely tied to the representation prob-
lem, for the method by which the semantic representation of a sentence is processed
depends critically on the form the representation itself is assumed to take in the
mind. The strategy in the present chapter is therefore to characterize the semantic
representation of a sentence as completely as possible and then to investigate the
subsequent processes that are required in using this representation in sentence
verification, question answering, instruction following, and the like.

!. REPRESENTATION

1.1 The Problem

The first order of business is to characterize what it is that a person has come to
know in reading or listening to a sentence. This information, normally .called the
meaning of the sentence, is to be contrasted with judgments one can make about
that meaning - e.g. is it true? is it silly? is it acceptable? is it nonsensical? and
so on. The semantic representation - or simply the representation - of a sentence
refers to a notation that characterizes the meaning of that sentence, what a person
knows of the content of that sentence.

1.11 The deep structure assumption

The semantic representation of a sentence, I assume, is equivalent to, or closely
related to, the linguistic deep structure of the sentence. By nsing the term 'deep
structure', I do not mean to pre-judge the current linguistic controversies as to
whetherlanguage..is best accounted for by a deep structure plus semantic inter-
pretation (as in Chomsky 1965), by a system of generative semantics (as in Lakoff
1971; McCawley 1968; Postal 1970), by a case granunar (as in Fillmore 1968;
Anderson 1971), or by something else. Rather, 'deep structure' is meant
to refer to an adequate linguistic representation of the underlying structure of a
.. !
sentence, whatever that will turn ont to be. In the present chapter, we will need to
refer only to linguistic facts and representations that onght to be acceptable to all
variants of current generative granunar.
The a priori justification for the deep structure assumption is quite straight-
forward. .One requirement of a general theory of comprehension is that semantic
representations conform to two constraints: (1) they must satisfy all linguistic
intuitions, or linguistic facts, about what a sentence means, whether it is acceptable,
semantically anomalous, ungrammatical, or whatever, what its accurate para"
phrasesare,· alid so on; (2) it must also be compatible with a process that accepts
a semantic representation and manipulates it appropriately to answer •questions,
follow directions, decide on issues of truth and falsity, and so on. Since the goals
of linguistic theory are simply those of (1), we automatically satisfy the first con-
 SEMANTICS AND COMPREHENSION 1293
dition by making the deep structure assumption - at least to the limit of present
day linguistic theory. In contrast, very little has been said previously about the
second condition: that is one of the main goals of this chapter. To put it another
way, the liuguistic facts at our disposal are both far more extensive and far more
systematic than our knowledge about the subsequent processes that make use of
semantic representations. Thus it seems prudent to begin with the best qualilied
semantic representations derived from liuguistic theory and to explore the possible
'mental processes' that can be built onto them to account for question answering,
sentence verification, instruction following, and the like. The snecess of the present
chapter to a large extent is due to the fact that this strategy has been followed as
closely as possible. As will be seen, the strategy enables us to make some 'very
strong generalizations about the mental processes that manipulate semantic repre-
sentations.
To put the justification in another way, the claim is being made that liuguistic
deep structme accurately represents what people know once they hav01 'compre-
hended' a sentence. In advance of many other concrete examples to be given in
subsequent sections, consider the negative sentence John isn't happy. Linguistically,
this sentence is composed roughly of two propositions, John is happy and It is
false, with the first embedded within the se.cond, as in that John is happy is false.
The semantic representation of this sentence might be coded as follows: ((John is
happy) is false). Such a representation characterizes much of what a person
understands the sentence to mean: (1) that happy applies to John and not to some-
thing else; (2) that false applies to the proposition that John is happy, not simply to
John or to something else; (3) that to show that the sentence is true, it is enough
to show that the embedded proposition (John is happy) is false; and so on. In
addition, it will be shown i)! section 2 that the two propositions, one embedded
within the other, are essential to the psychological model that accounts for the
latencies with which people are able to verify negative sentences like this one. In
short, there is evidence - both liuguistic and psychological - that the linguistic
deep structme does accurately characterize what people understand a sentence to
mea.J,l.
There are limitations, however, on how literally we should accept the deep
structure assumption, for it could be taken in either a strong or a weak form. In its
strongest form, the assumption is that a person who has comprehended a sentence
'knows' every detail represented in liuguistic deep structure. In a weaker form, the
assumption is that whatever it is that a person 'knows' of a sentence is consistent
with liuguistic deep structure, but he does not necessarily represent all or even
most of the details of deep structure each time he comprehends_ a sentence. It
seems highly unlikely that the strong form will be correct, though it is quite
plausible that some version of the weak form will. This implies that the linguistic
deep structure contains many details that a listener might well pass over and
leave unrepresented on any particular occasion; nevertheless, thos~ details that he
1294 HERBERT H. CLARK

does represent mnst be consistent with deep structure, since otherwise the listener
will be said to have misunderstood the sentence. This point is best made with an
example. Consider the sentence John thought to himself, 'Mary has just come into
the kitchen'. If the listener were interested only in knowing where Mary was, he
might well ignore the presupposition of come, that John must be in the kitchen
(cf. Fillmore 1967a), and represent come simply as 'move'. But if the listener
were interested in knowing where John was, he could not ignore that presupposi-
tion; he would have to represent come as 'move' plus its presupposition, i.e. as
'move toward John'. It is difficult to assert a priori that the listener will always
represent the presupposition of come, since it is easy to imagine a listener who is
quite unable to say where John was, even though the information was in the
original sentence the listener had read and 'understood' a moment before. The
assumption is, however, that whenever the listener does represent the presupposition
of come, he will do so in a way consistent with the linguistic representation of come
in its context.
1.2 The Processing Problem
The listener is assumed to constrnct a semantic representation for each sentence
he understands, and that representation is assumed to serve as the basis for all
further processes that require the comprehended information. ·This brings us to
the processing problem: just what processes does the listener use on semantic
representations in order to verify sentences, answer questions, and follow instrnc-
tions appropriately? For this purpose, I have formalized the complete representa-
tion and processing problems in a specific yet easily generalizable schema or format.
The schema is not an arbitrary one, for it has empirical consequences that can be
tested in any comprehension task that it is applied to. As we will see, the schema
itself is given considerable empirical support in the literature to be examined later.

1.21 The four-stage schema

The assumption is that most 'comprehension' tasks can be broken down into
four identifiable stages (cf. Clark 1969a; Clark and Chase 1972). Con-
sider the so-called verification task in which the experimental subject is given a
sentence and a picture and is asked whether the sentence is trne or false of the
picture. At Stage 1, the subject is presumed to encode the sentence- i.e. to
constrnct a semantic representation for the sentence. At Stage 2, the subject codes
the picture in a representation that is in the same format as the sentence. At
Stage 3, the subject carries out a series of meutal operations that compare his
representation of the sentence with that of the picture to determine whether they
are synonymous or not. And at Stage 4, the subject takes the outcome of the
Stage 3 comparison operations and converts that outcome - e.g. true -'-- into the
vocalization 'trne', into the press of a button marked 'trne' or into whatever re-
sponse is appropriate. So the process includes four stages - a sentence represen-
 SEMANTICS AND COMPREHENSION 1295
tation stage, a picture representation stage, a comparison stage, and a response
stage. The four stages must be serially ordered: the sentence and picture must be
represented before their representations can be compared; the comparison stage
has to be complete before the response stage can realize the outcome of the com-
parison stage; however, the two representation stages could be in either order,
depending on the task requirements.
The four-stage process just given is applicable to comprehension tasks of all
ldnds with minor alterations in the stages. First, let us examine the verification
task again in a little more detail. Assume that a subject is required to verify the
sentence A is above B against a picture of an A above a B. Hypothetically, the
process might appear as in:
1 Stage 1: Represent sentence as (A above B)
Stage 2: Represent picture as (A above B)
Stage 3: Compare (A above B) with (A above B) for synomymy
Stage 4: Produce the response 'true'
Although it is critical to know how the sentence and picture are represented at
Stages 1 and 2, the heart of the task is the Stage 3 comparison process. It must
be expanded into a set of mental operations that compare the Stage 1 and 2 repre-
sentations piece by piece, since in much more complex examples, the comparison
will not be as obvious or simple as in this example. Indeed, the main proposals of
the present chapter are concerned with the Stage 3 comparison operations and their
effects on the latency of the Stage 4 response. I will return to the Stage 3 mental
operations later.
In contrast, consider a question-answering task, such as If John killed Harry,
then who died? In this case, the four stages would appear as in:
2 Stage 1: Represent premise as (John caused (Harry die))
Stage 2: Represent question as (X die)
Stage 3: Compare (John caused (Harry die)) and (X die): find X
Stage 4: Produce the response 'Harry'
Stages 1 and 2 in 2 are identical in format to the same stages in the verification task;
if the same sentences had been used, the representations would be identical. Stage 4
is also the same except for the specific response. There is a difference, however,
at Stage 3. Whereas Stage 3 simply matched the two representations for identity
in the verification task, it must match representations and pick out an X in the
question answering task. It will become cleat that this is a minor difference, for the
matching operations are carried out in the same way in both tasks.
In the next task to be illustrated, the subject is given a page full of digits and
is required to carry out the instruction Cross out all even numbers. This task
might be represented as 3:
ri 3 Stage 1: Represent instruction as (you cross out X (if (X is even)))
i Stage 2: Represent the current digit (e.g. 8) as (8 is. even)
Stage 3: Compare representations from Stages 1 and 2
Stage 4: Produce response consistent with (you cross out 8)
1296 HERBERT H. CLARK

This task, in fact, contains a verification process within it. One can think of the
Stage 3 comparison as having two parts: in the first, the subject compares the if-
clause of the instruction with the (8 is even) exactly as in the verification task; in the
second, the subject 'attaches' the outcome (true) of this comparison to the main
clause of the instruction to produce the representation (true (you. cross out 8))
from which the response is made at Stage 4. So although this task is superficially
quite different from the verification task in 1, it can be conceptualized in ahnost
identical stages and comparison processes.
The four-stage schema, as applied to the verification, question-answering, and
instruction-following tasks, is more• than just a convenient conceptualization of
these tasks. Underlying the schema are three very important empirical claims.
The first claim is that a sentence is represented (understood) in the same way no
matter what the task is. Stages· 1 and 2 are in principle identical in form for the
verification, question-answering, and instruction-following tasks. The sentence
A is above B, for example, would he represented as (A above B) at Stage 1 no
matter what the task. Although this claim is open to empirical test, the evidence
available (to be presented later) suggests that this claim is indeed true. The second
empirical claim is that the representation stages {Stages 1 and 2) are separate from
the comparison stage (Stage 3), which in tnru is separate from the production of
the response (Stage 4). Furthermore, the representation stages, the comparison
stage, and the response stage must be carried out in this order. I will present
evidence later· that supports this empirical claim. The third and final claim is
that the comparison stages (Stage 3) in the three tasks all have something iu
common: they work from the semantic represent~tions of sentences, carry out
comparisons and manipulations on these underlying representations, and eventually
produce outcomes, (like true or false) that serve as the basis for Stage 4 response.
One part of this claim is that these comparison stages do not operate directly on
the surface structure of the sentences, on mental images, or on other 'uninter-
preted' representations. Evidence for this claim will be considered later. A second
part of this· claim is that comparisons are made by checking for the IDENTITY of
various parts of the Stage 1 and ·2 representations. These claims are part ofan
assumption I have previously called the PRINCIPLE oF CONGRUENCE (Clark 1969a).

1.:22 ' The prinCiple of congruence

Imagine that' a person has read and understood the sentence John killed Harry,
Now we ntight ask bin\, Who died?, a questiori that is simple enough for the listener
to answer. Our listener koows, roughly, that h<>must find a term which, when it
replaces 'the· who in Who died?, will make the corresponding declarative sentence
true. One possible way to do this would be to match surface structures - to seeif
died can be fouild in the original sentence and, if so, to t;lke the term immediately
preceding died as the correct answer. This, of course, will not work here, and it
would not work even in sentences that e:Xplicitly contain died, as in The person who
 SEMANTICS AND COMPREHENSION 1297
died was Harry, Harry- according to John -died with his boots on, etc. etc.
To answer Who died?, therefore, this listener must search not through surface struc-
ture, but through deep structure, seeking congruence between selected parts of the
semantic representations of the premise and the question. Who died? can be an-
swered once John killed Harry is represented as (John caused (Harry die)) and
Who died? as (X die). Then (X die) can be matched with (Harry die), X replaced
by Harry, and the question answered by 'Harry'.
This is one line of reasoning that led to the PRINCIPLE OF CONGRUENCE, which as-
serts simply that the Stage 3 comparison operations work to make underlying repre-
sentations completely congruent with each other. The basic operation, then, is the
match. (X die) is matched against (Harry die); the two representations are found
to be identical except for the mismatch of X and Harry; and so X is replaced by
Harry in accordance with the requirements of the task. Thus the answer 'Harry
died' (or in elliptical form 'Harry did' or simply 'Harry') is fully congruent with a
complete part of the semantic representation of the sentence John killed Harry. To
take another illustration, the sentence A is above B, represented as (A above B),
might be verified against a picture represented as (B below A). Since these two
representations are not congruent, an additional operation is required to transform
one of the two representations, e.g. (B below A) into (A above B), in order to make
them congruent. After this transformation, the two representations are completely
congruent, an identity match will succeed, and the sentence can be pronounced
'true' of the picture. '
The principle of congruence has several important theoretical and empirical con-
sequences. First, the principle provides the basis for constructing a set of Stage 3
comparison operations that will correctly solve thetask at hand. Without such a
principle, it is impossible to, specify just how Stage 3 would go about deciding, for
example, whether (A above B) and (B below A) are synonymous or not. In more
complex cases to be considered later, the Stage 3 operations would be quite ad hoc
without such a principle. Second, the principle implies that certain comparisons
should be more difficult than others. In particular, whenever there is incongruence,
an additional martipulation operation is required of Stage 3 in order to bring the
two representations into congruence. ,If it is assumed simply that each of these
additional operations takes time to perform, tasks with incongruent semantic repre-
sentations should take longer, or elicit more errors, than tasks with congruent
semantic representations. I will use this particular argument time and again to
make predictions about response latencies in verification, question-answering, and
instruction-following tasks. In particular, this second implication has been shown
to hold very closely for the verification of negative sentences (Clark and Chase
1972), a fact that will be discussed at length in section 2.
 1298 HERBERT H. CLARK

1.3 Recapitulation

In section 1, then, I have outlined a method for approaching comprehension tasks.

Psychologists have rarely studied comprehension alone, for comprehension is nor-
mally such a private experience that it gives psychologists very little to measure.
The usual methodological trick is to embed comprehension itself in other sorts of
tasks - tasks which range from simply asking the subject to indicate whether a
sentence is true or false to asking him to follow a difficult instruction as fast as
possible. In each of these cases, the subject has been tricked into comprehending a
sentence, since an understanding of the sentence is required before he can carry out
the task. Thus, instead of having to account for comprehension alone, the psy-
chologist has to account for comprehension pins all the subsequent activity that the
task requires - e.g. making decisions, responding, crossing ont digits, etc. This
has led to a formulation of what it is that all these tasks have in common.
The main proposal is that such comprehension tasks can be broken down into
four separable stages: (1) a sentence representation stage; (2) another sentence (ot
picture or knowledge) representation stage; (3) a stage in which the representations
of (1) and (2) are compared in some way; and (4) a response stage. These stages
help to define what the psychological theory must account for. In particular, it is
necessary to know: (a) the form in which sentences are represented at Stages 1 and
2; and (b) the mental operations by which these representations are compared at
Stage 3. The present theory therefore makes two strong assumptions. First, the
DEEP STRUCTURE ASSUMPTION asserts that a sentence is represented in memory in a
form that is equivalent to, or closely related to, its deep underlying linguistic repre-
sentation. The second assumption, the PRINCIPLE OF coNGRUENCE, nsed in conjunc-
tion with the first, \'Sserts that two such semantic representations are compared for
congruence, and in !he absence of congruence, there must be mental operations that
bring the two representations into congruence.
It is the deep structure assumption and the principle of congruence that lie at the
. base of the present theory. In the following sections, I will discuss how the deep
structure assumption is applied to four different linguistic constructions - the
• negative, the locative, the comparative, and the active and passive. Armed with
this assumption, and the principle of congruence, I will then predict the relative
difficulty of each of these constructions as used in a wide variety of comprehension
tasks. Then I will review the main results of the experimental literature for each
construction and show how the results bear out these predictions, thereby con-
firming the present theory of comprehension.

2. NEGATION

Negation is undoubtedly one of the most fundamental conceptual devices of lan-

guage. It is used in the explicit denial of facts (e.g. Katherine isn't present), in the
 SEMANTICS AND COMPREHENSION 1299
affirmation of negatively stated facts (Katherine is absent), and so on. From the
point of v·iew of syntax, negation is found in vastly different guises - in particles
(not, never, no), adjectives (absent, few), adverbs (hardly, scarcely, seldom), verbs
(forget, lack, fail), nouns (absence, lack, failure), prepositions (from, out, off), and
so on. In short, negation is an extremely heterogeneous phenomenon - even in
English. Nevertheless, negation appears to be a phenomenon that should be treated
as a single semantic device in language. Indeed, the purpose of section 2 is to
present a single nnified model of the comprehension of negation - one based on
the principles of section 1 - that is capable of accounting for the main facts about
all types of negation in the existing psychological literature.
Because negation was one of the first linguistic notions to be studied by psycholo-
gists in the last decade, the negation literature is really quite large. The pioneering
studies on negation were carried out by Wason (1959, 1961, 1965; Wason and Jones
1963), and this series was followed by investigations by Eifermaun (1961 ),
McMahon (1963), Gough (1965, 1966), Slobin (1966), and many others. The main
result of these studies can be summarized succinctly: negation is more difficult to
comprehend than affirmation. Although this result is hardly surprising, these
studies also uncovered a number of other important facts about negation. Unfor-
tunately, the secondary results varied considerably from study to study, so much so
that even two apparently similar studies often appear to have quite contradictory
results. For example, some studies found that true negative sentences were more
difficult to verify than those that were false, whereas other studies found just the
opposite. One of the main goals of this section is to reconcile these and other con-
flicting results on the comprehension of negation.
It was partly because of the inconsistency of the previous results that William
Chase and I carried out a series of experiments on an instance of negation
that is experimentally much. cleaner than any of those studied by previous investi-
gators (Clark and Chase 1972, 1974; Chase and Clark 1971, 1972;
Young and Chase 1971; Singer, Chase, Young, and Clark 1971). Indeed,
we have found what appears to be the paradigm case for the comprehension of
negation. I will refer to this case as the 'true' method or model of negation (the
meaning of 'true' will becpme clear later). In addition, Richard Young and
William Chase (1971) have shown that negatives can be comprehended by several
methods that consist of converting negatives directly into affirmatives and com-
prehending the resultant affirmatives. I will refer to these methods as the 'con-
version' methods or models of negation. One important thing about the 'conversion'
models is that they are really just special cases of the 'true' model, although
'conversion' methods produce results that are quite different from those of the
'true' method. The second important thing is that the conflicts in the previous results
can be explained by assuming that the subjects in the previous studies were using
predominantly one or the other method of negation.
Even more recently, several students and I have carried out a series of studies on
 1300 HERBERT H. CLARK

other types of negation - in particnlar, various types of 'implicit' negation, e.g.

absent, few, scarcely any, and so on. The importance of these studies is that they
show that implicit negation is comprehended in the same manner as explicit nega-
tion, except for several very important differences, which appear to be closely
related to the suppositions of the sentence.
The plan of this section is as follows. In section 2.1, I will present the 'true' and .
'conversion' models of negation and their rationale. In that section I will treat only
explicit sentence negation, as in the example, Katherine isn't at home. In section
2.2, I will review all of the previous literature on this type of negation in order to
show that the previous results can be accounted for by the 'true' or 'conversion'
models or by some combination of the two. In section 2.3, I will take up the lin-
. guistic problem of how to characterize the several dilferent types of negation,
especially in terms of snpposition; and in section 2.4, I will consider th.e stndies that
bear on the psychological properties of those types of negation. Jn section 2.5, I
will consider an instance of 'perceptual' negation, that is, :an instance iu which a
diagram is mentally represented in a negative code that behaves in all othedashions
like a linguistic negative. Finally, in section 2.6, I will summarize the main models
and findings in negation.

2.1 The 'True' and 'Conversion' Models of Negation

The 'true' and 'conversion' models of. negation are most ·easily described within the
context of a particular verification task that Chase andl put to our subjects. In it,
we presented the subject with a display that contained a sentence on the left and a
picture on the right. The subject's task was to read the sentence, then look at the
picture, and then press a 'true' or a 'false' button as quickly as possible. to indicate
whether the sentence was true or false of the picture. The subject was timed from
the moment the display appeared to the moment·he,pushed the button. ·.There were
eight possible sentences: plus is above star, stat is above plus, plus is· below star,
star is below plus, star isn't above plus, plus isn't.above· star; star- isk~t below. plus,
and plus isn't below star. And there were two possible· pictures: a star (a typed
asterisk) above a plus; and a plus above a star.. mall, then, ther<nvere 16 displays,
four of which contained Tiue.Affirmative sentences, four•False:Affirmative, four
True Negative; .and four False Negative. The goal we set for ourselves. was to
account for the variation in the times it took the subjects to decide: whether the
sentences were true or false.
As proposed in section 1.2, we assumed that the total process could·. be. broken
down into four stages. At Stage 1, the ~ubject encodes the sentence.in·a mental
representation of some sort. At Stage 2, he encodes the picture in the.: same repre-
sentational format. At Stage 3, the· subject compares the Tepresentations: he has
constructed for the sentence and picture. to see .whether or not they· match. And
 SEMANTICS AND COMPREHENSION 1301
at Stage 4, he takes the output of this comparison stage aud converts it into some
sort of response. The 'true' and 'conversion' models differ in how the subject is
assumed to have represented the sentence at Stage 1; otherwise, the two models are
identical.

2.11 The 'true' model of negation

2.111 The representation problem. The fundamental question here is, what is
the mental representation of a negative sentence? In accordance with the deep
structure assmnption I am making throughout, the negative sent<;nce is assumed to
be represented mentally as something very close to its deep structure. The sentence
A is above B, for example, is a simple proposition represented, perhaps, in terms
of its internal subject-predicate relations as ((A)NP(above B)vP)s. For now, the in-
ternal structure of A is above B is of little interest, so t will simp!)! represent its
deep structure as (A above B). But now, when sentence negation is attached to
A is aboveB, the resnlt is .A isn't above B, whose deep structure might look some-
thing like the following: (((A above B)s)NP(false)yp)s. This cocld be paraphrased
as that A is above B is false or it is false that A is above B or it isn't the case that
A is above B. For present prrrposes, i will symbolize the deep structure of this
negative simply as (false (A above B)). It c~uid have been symbolized by a.more
common notation, say Neg (A above B), but I prefer the present notation since it
makes the semantic nature of the negative explicit. The point of this notation is to
show that a negative sentence, like (false (A above B)),:is.simply a non-negative
proposition embedded in a higher sentence which asserts that that embedded propo-
sition is false. As we will see, there ate important consequences that follow from
the embedded nature of the negative.
Once this representational scheme has been adopted, the eight basic types of
sentences in the verificatio~ task can be represented very simply. Consider a
schematic picture of au A above a B aud the eight possible sentences that could
have been paired with that picture. The representations for the eight sentences are
shown in the' second column of Table I. These eight sentences· fall into four cate-
gories, True AffirniatiV'es, False Affirmatives, True Negatives; aud False Negatives,
•
; .i
.; I
with a sentence containing above and a sentence containing below in each category .
Ii The problem of how to represent the pictures at :stage 2 ·is just as critical as the
problem of how to represent the sentences at Stage l. As discussed in section 1.2,
however, the reqD.iremerit is that the pictures must be represented in the same
forinat as the sentencbs; aud this leaves: us with the dedsion:of whether au A above
f
a B is coded as (A above B), (B below A), (fidse (B above A)), or (!alse (A below
I B); all of~hicli are consistent with the picture: Although the experimental evidence
i
-.--j caimot be presented here (cf. Clark and Chase 1972; aud section 3), it shows
that for this partii:nlar experiment the picture is coded as (A: above :B) when the sen-
l tence contains the word above, but as(B belowA)'when the, sentence contains the
word' below. These representations ate shown in the . third colrimn of Table I.
i
1302 HERBERT H. CLARK

TABLE I. Sentence representations, picture representations, and latency·

components for the 'true' model of negation
Stage 1 Stage2
Sentence Picture
Sentence Type Representation Representation Latency Components

True Affirmative
(A above B) (A above B) to
(B below A) (B beloW A) t 0 +a
False Affirmative
(B above A) {A above B) to +c
(A below B) (B below A) to+a+c
True Negative
(false (B above A)) (A above B) to +c+(b+d)
(false (A below B)) (B below A) to+a+c+(b+d)
False Negative
(false (A above B))' (A above B) to +(b+d)
{false (B below A)) (B below A) to+ a + (b +d)

2.112 The Stage 3 comparison operations. Stage 3 is meant to decide whether

a particular Stage 1 sentence representation (in Table I) is synonymous with ito
Stage 2 counterpart, the picture representation. Stage 3's job then is to come up
with a judgment of 'true' or 'false' about the synonymy of the sentence and picture.
For this reason, Stage 3 is said to keep track of a TRUTH INDEX, whose value is
true or false. The goal of Stage 3 is to manipulate the truth index in such a way
that the index cmitains the correct value at the end of the process. It is assumed
that the truth index contains the value true at the start - that is, the sentence is
considered true of the picture until there is evidence to the contrary. Various Stage
3 (mental) operations then change the truth index from true to false, and even from
false back to true again, depending on the result of various comparisons ·that the
operations carry out. ··
The heart of the 'tnie' model of negation is to be found in the mental operations
that make up the Stage 3 comparison process. These mental operatiol)s could be
shown as a flow diagram, as a tree diagram, as a set of branching rules, or .other-
wise. For convenience, I will represent the mental operations as a set of four
branching rules, which are shown in Table II. I will illustrate the use of these rules
for the True Negative sentence B isn't above A, the fifth line down in Table I. The
sentence is represented as (false (B above A)) at Stage 1; the picture is represented
as (A above B) at Stage 2. Operation 1 compares the embedded strings (B above
A) and (A above B), and finds that they do not match. So. the process goes to
Operation 1a, which changes the presupposed truth value. true to false. Then Opera-
 SEMANTICS AND COMPREHENSION 1303
TABLE II. The comparison process (stage 3) of the 'true' model of negation

1. Compare the embedded strings of sentence and pictnre.

A. If they match, go to 2.
B. If they do not match, go to Ia.
1a. Change value of truth index into its opposite; then go to 2.
2. Compare the embedding strings of sentence and picture.
A. If they match, stop.
B. If they do not match, go to 2a.
2a. Change value of truth index into its opposite; then stop.

tion 2 compares the embedding strings, (false 0) and (0), and finds that they do
not match. So the process goes to Operation 2a, which changes the truth index
~set at false by Operation 1a - from false back to true again. The fioal value of
the truth iodex is therefore true, which is iodeed the correct value for this particular
True Negative sentence. Although the four different tyjj!:s of sentences take dif-
ferent paths through these Stage 3 mental operations, careful workiog through of
the four types will convioce the reader that the Stage 3 operations of Table II
produce the correct answer io every case. .
Stage 4 is simply a formal stage io which the outcome of the Stage 3 operations
-i.e. the'fioal value of the truth index- is converted ioto the appropriate re-
sponse; In Clark and Chase (1972), that response was a push of a button
but . it could have been the vocalization of the words 'true' or 'false'. In any
case; this stage is of little ioterest here, for the responses are direct and are assumed
not to affect the relative verification latencies in the problems we are considering.
2.113 Predictions of verification latency. In constructing this model, Chase and
I wanted to be able to predict the relative verification latencies for the eight sen-
tence types of Table I. For this purpose, we made several simple assumptions.
First, the four stages are carried out io succession. Second, the times taken at each
stage are constant from condition to condition. And third, e~ch mental operation
at Stage· 3 consumes a specific increment of time, and these separate increments
are additive. With these assumptions, it is possible to predict the latencies of the
eight sentence types of Table I.
· At Stage 1, the subject encodes a sentence that contains either above or below.
Forreas~ns that I will discuss io section 3, it was assumed that below takes longer
to encode at this stage than above. To be precise, below takes an iocrement of time
a longer to encode. Similarly, it seemed likely a priori that positive and negative
sentences would not be encoded io equal times. The assumption was that negatives
would take an increment of time b longer to encode than affirmatives. At Stage 3,
Operations 1, 1a, 2, and 2a were assumed to consume iocrements of time x, c, y,
and d, respectively. Note that the sentences differ io What Stage 3 operations they
require: True· Affirmatives require Operations 1 and 2; False Affirmatives require
1304 HERBERT H. CLARK

1, 1a, and 2; True Negatives require 1, 1a, 2, and 2a; and False Negatives require
1, 2, and 2a. All sentences therefore require Operations 1 and 2, and they differ
only in whether they require 1a, 2a, or both; that is,, they differ in whether they will
consume increments of time c, d, or both.
The predictions made by these assumptions about parameters a, b, c, and d are
sho\'l<n for each sentence iri Table I. The parameter to is simply a wastebasket
category, which contains all the time not taken up by the predicted parameters a,
b, c, and d. The True Negative sentence A isn't below B, for example, consumes
a and b at Stage 1 because it contains below and a negative, and it consumes c and
d at Stage 3 because it requires both Operations 1a and 2a. Note that b. iind d are
perfectly correlated iri this table, making it impossible to separate them ,out iri this
experiment. I have therefore listed (b + tf) as a sirigle parameter, which I will call
simply Negation Time. Similarly, I will call the parameter c Falsification Time.
We will need to refer to Negation Time and Falsification Time again and again- iri
our review of the literature.
2.114 Experimental evidence for the 'true'· model. Data from Clark andChase
(1972) and other related experiments confirm this model with an extra-
ordinary degree of accuracy. In one experiment, twelve subjects were run for 160
trials each iri the manner described previously. From the overall data, Below Time
+
a was estimated to be 93 msec, Negation Time (b tf) 685 msec; and Falsification
Time c 187 msec. Once these three values. were substituted into the forinulae of
Table I, the latencies predicted by these formulae were very close 'to the actual
overa:ll latencies for these eight types of sentence . .Statistically, these three. parameters
accounted for 99.6% <if the variance iri the .rctual latencies. The 'true' model of
negation received further support in at least five other, experiments; in the latter
variants of the iriitial experiment, the subjects were required to look at the picture
first,· to attend to particular parts of the picture, or to conform to other differences
in procedure. In other words, the 'true' model of negation was supported•.in a wide
variety of conditions, most often accounting for over 97% of the variance iri the
observed latencies.
So far I have talked .ofthe 'true' model of negation, particularly Stage. 3, only
iri terms of very abstract representations and comparison operations. Brit all the
stages of this model reflect quite accurately what ·subjects claim they .are doing.
The typical subject describes his intuitions somethirig like this:· 'Wherr I geFthe
sentence B is above A, I first look over at the top of the pictrire:' We presume,
for this reason, that he is encoding the picture .as (A above B), not as· (B below A).
'But ·theri I notice th.at the .B isn't on tOp as I expected, so I change my answer to
false and push the, "false" button.' . Here we presume that he finds· a mismatch· at
Operation 1 and so must go to·Operation la, which changes the truth value• from
true to false, which then serves as the basis for his button push. 'When,.J•get·the
sentence A isn't aboveB; again I look at the top of the picture', again presumably
encoding the picture as (A above B). 'But here I notice that the A is above the.B,
 SEMANTICS AND COMPREHENSION 1305
while the sentence says that A is not above B. Since this is contradictory, I change
my answer to false.' Here he has described carrying out Operation 2, finding a
mismatch, and then carrying out Operation 2a, which changes the truth value from
true to false. 'With the sentence B isn't above A, I look at the top, find that A is
above B, and note that the sentence and picture don't match; but I also remember
that the sentence is negative so I realize that even though there is a mismatch I
must say "true".' Here he is describing finding a mismatch at Operation 1, carrying
out Operation 1a, finding a mismatch at Operation 2, and carrying out Operation
2a. In summary, subjects report changing their answer from true to false and back
to true again on True Negatives. This is exactly what the model does as well.
In one sense, this model is not particularly new. Wason (1961), for example,
formulated an informal model similar to this one based on what his subjects said
that they were doing. A similar model of negation was also proposed by Gough
(1965) and Slobin (1966). Unfortunately, however, the models they proposed were
unable to account for their data in the direct and accurate way that they were able
to account for the data in Clark and Chase ·(1972). The reason, as I will
argue below, was because some of the subjects in Wason's and Gough's tasks
in particular were treating negatives quite differently from this model. For this
reason, Wason and Gough were unable to confirm that the normal process of veri·
fying negative sentences ·consists of two well-defined operations that consume fixed
amounts of time.

2.12 The 'conversion' methods for comprehending negatives

The subjects in Clark and Chase (1972) could have used a much more
direct method for verifying negative sentences - if they had wanted to. For
example, they could have converted the sentence A isn't above B immediately into
A is below B and then looked at the picture for verification. If they had done this
consistently for all the negatives, the False Negatives should have been 180 msec
slower, rather than 180 msec faster, than the True Negatives, just as the False
Affirmatives were 180 msec slower than the True Affiriiiatives. Evidence internal
to the data is therefore in complete agreement ·with the subjects' reports that they
had all used the 'tme' method of negation described abov~.
In principle, however, subjects can make conversions of this sort. Young and
Chase (1971) carried out a series of ingenious experiments on the assump-
tion that they could simply instruct subjects to make particular conversions, and if
subjects followed these instructions properly, their latencies would change radically
from those shown by the. subjects in the Clark and Chase experiments. Young and
Chase essentially repeated the Clark and Chase experiment just described, but
asked their subjects to 'convert' each sentence to another form before they looked
at the picture. Each subject proceeded through the task using one of four 'conver-
sion' rules: (1) delete the negative and· change the preposition, converting e.g.
A isn't above B into A is below B; (Z) delete the negative and reverse the two
1306 HERBERT H. CLARK

nouns, converting e.g. A isn't below B into B is below A; (3) changes below into
above and reverse the two nouns, converting e.g. A isn't below B into B isn't above
A; and (4) change below into above and change positive into negative or vice versa,
whichever is appropriate, converting e.g. A is below B into A isn't above B. Three
very reliable subjects worked under each of the four 'conversion' instructions in
four different sessions.
Young and Chase predicted that the latencies of their subjects' judgments could
be broken down into two parts. Fiist, the conversion itself would take an increment
of time k to perform. And second, the 'true' model of negation would work on the
semantic representation that results once the 'conversion' had been performed. As
illustration, consider the first conversion rule listed above. It changes each negative
into an affirmative with the opposite preposition, so that isn't above becomes below
and isn't below becomes above. Table III shows the sentences as the subjects saw

TABlE III. Sentences, sentence representations, and latency components for one
f conversion'· method for comprehending negatives

Stage 1 Stage 2
Sentence Picture Latency
Sentence Type Sentence Representation Representation Components

True Affirmative
A is above B (A above B) (A above B) to
B is below A (B below A) (B below A) to +a
False Affirmative
B is above A (B above A) (A above B) to +c
.. AisbelowB (A below B) (B below A) to +a+c
True Negative
B isn't above A (B below A) (Bbelow A) to +k +a
A ;isn't below B (A above B) C'\ above B)
to+k
False Negative
A isn't above B (Abelow B) (B below A) to+ k+a+c
B isn't below A (B above A) (A above B) to+k +c

them, the Stage 1 representations of the sentences after conversion, the picture repre-
sentations, and the appropriate latency components from the 'true' model of negation.
Note that the latency components contain k for each negative sentence converted, and
no longer contain (b +d), the parameters associaiedwith Negation Time in the 'true'
model. In short, the latency components are predicted from a straightfor-ward appli-
cation of the 'true' model of negation plus Conversion Time k. In the experiments of
Young and Chase, this model was supported again with a very high degree of
accuracy. With three parameters estimated for each- of the three subjects, the model
 SEMANTICS AND COMPREHENSION 1307
shown in Table III >recounted for over 95% of the variance in each of the subjects.
Since this kind of accuracy is mre in psychological experiments, these results add
considerably to the credibility of this particulfff 'conversion' model. Young and
Chase were just as successful in predicting the latencies under the other three 'con-
version' rules.

2.13 Comparisons of the 'true' and 'conversion' models of negation

The most important difference between the 'true' and 'conversion' models of
negation is that they predict different patterns of latencies fot True and False
Affirmatives and Negatives. The main comparisons can be seen in Table IV, which
lists the predictions of the two models (ignoring the above-below difference for the
moment). The pffflicular 'conversion' model illustrated here is for rule 1 of the
Young and Chase experiment, since that rule is the only one that has been used
spontaneously by subjects in other experiments in the literature. First, note that
False Affirmatives fffe-predicted to be an increment c slower than True Affirmatives
under both models. The reason for this, of course, is that the two models are
identical for affirmative sentences. For the negatives, however, True Negatives are
c slower than False Negatives under the 'true' model, but c FASTER than False
Negatives under the 'conversion' model. To put it another way, there is a true-false
by positive-negative interaction under the 'tnie' model, but no such interaction
under the 'conversion' model. The consequence of these effects is that, under ihe
'true' model, true sentences should be equally as fast overall as false sentences, but
under the 'conversion' model true sentences should be an increment c faster overall
than false sentences. It is -these consequences of the 'true' and 'conversion' models
that will allow us to make sense of the previous literature on negation.
TABLE IV. The latency 'Components for the ttrue' and 'conversion' models
of negation
Sentence Type 'True' Model 'Conversion' Model
True Affirmative to to
False Affirmative to+ c to+ c
True Negative to+c+(b+d) to +k
False Negative to +(b+d) to+c+k

The second important difference between the 'true' and 'conversion' methods is
that the 'ttoe' method of negation will work under all circumstances, whereas the
'conversion' method will not. When the subject uses a 'conversion' method, he is
normally cheating a bit. By changing A isn't above B, for example, into A is below
B, the subject has lost some of the information in the first sentence, since below
implies isn't above, but isn't above does not imply below. This pfffticular conver-
sion would result in an incorrect answer for a picture of an A and a B side by side:
the sentence A isn't above B would be judged true of this picture, while its 'con-
 1308 HERBERT H. CLARK

verted' form A is below B would incorrectly be judged false. The 'conversion'

methods, therefore, can be used correctly only under very special circumstances,
as in the Young and Chase experiment in which the subject knew he would never
encounter anything but pictures of an A and a B (i.e. a star and a plus) one above
the other. It is this fact about the 'conversion' method that justifies the term 'true'
in the 'true' method of negation. The 'true' method is the only method that is
always applicable and that retains all the information of the negative sentence.

2.2 Previous Studies of Negation

These two models of negation - the 'true' and 'conversion' models -can now be
profitably applied to the previous research on the comprehension of negative sen-
tences. Although the majority of these studies are verification tasks, they differ in
a number of ways. The most immediate difference is in the kinds of evidence that
sentences are to be verified against. In some studies, sentences are verified against
pictures, as in the Clark and Chase experiments; in others, they are verified against
previous knowledge, as in the verification of Nine isn't an even number; and in still
_,.:·-- other cases, they are verified against other sentences, as in If x precedes y, then y
isn't preceded by x. To be able to form a general theory about the results of these
experiments - and the results are indeed very similar - we must assume that all
three types of evidence - pictures, previous knowledge, and other sentences - are
represented at Stage 2 in the same format. This assumption is a very important one,
for without it, we would be forced to formulate a different theory for almost every
one of the experiments that follow; with it, the 'true' and 'conversion' models apply
accurately to a la<ge number of studies.
The first study to be examined is one of Wason's pioneering studies on negation
(Wason 1961) in which subjects were timed as they verified whether sentences like
Nine is not an even number were true or false. Among his sentences were True
Affirmatives (e.g. Nine is an odd number), False Affirmatives (Nine is an even
number), True Negatives (Nine is not an even number), and False Negatives (Nine
is not an odd number). This task is easily formulated in the four-stage model dis-
cussed previously. At Stage 1, the subject sets up a representation of the sentence,
say (false (9 is even)); at Stage 2, he sets up a representation of his a priori know-
ledge about the number in ·the sentence, in this case (9 is odd); at Stage 3, he carries
out the four-operation comparison process shown in Table II, coming up in this
instance with the correct. truth value true; .and at Stage 4, he says 'true'. Thus, if
Wason's subjects were actually using the 'true' method of negation, his results
should be predicted accurately by the 'true' model.
Wason, however, asked each one of his· subjects how he. carried out the task.
Significantly, he found two types of. subjects. The. first type - about half the
subjects- reported using a method equivalent to. the 'true' method of negation.
 SEMANTICS AND COMPREHENSION 1309
The second type - the remaining subjects - reported converting not even to odd
and not odd to even each time they encountered a negative sentence. Since odd
and even are contradictories, this is a perfectly good strategy. QActually, the use
of not instead of n't probably encouraged this kind of conversion, since it invites
contrastive stress - 'Nine is not an odd number' - which brings with it the con-
clusion, 'It is an even one'.) From these reports alone, we should expect the results
to show a mixture of the 'true' and 'conversion' models, and this is exactly what
Wason found. He found that True Affirmatives took 260 msec less time than False
Affirmatives, in agreement with both models. On the other hand, he found True
Negatives to be 30 msec faster than False Negatives. The 'true' model predicts that
True Negatives should be 260 msec slower than False Negatives, whereas the 'con-
version' model predicts the opposite. A mixtnre of the two types of subjects would
therefore approximately cancel out tlie difference between True and False Nega-
tives, just as Wason found. Wason also found that True Negatives elicited the most
errors. This would occur because some subjects in the 'true' method would have
failed to carry out both Operations 1a and 2a, thus producing the wrong answer;
it is ouly on True Negatives (in the 'true' method) that both comparison operations
are required.
In Wason and Jones (1963), a variation of the same experiment was repeated
with only slightly different results. Overall, True Affirmatives took 420 msec less
than False Affirmatives, but in this case True Negatives were SLOWER than False
Negatives by 160 msec. Thus, in this experiment, there must also have been a mix
of people using the 'true' and 'conversion' methods, but there should have been
slightly more people using the 'true' method. Unfortunately, Wason and Jones did
not report on the percentages of subjects using the two methods, although they
implied that both kinds of .subjects were present. And again, there were more
errors on True Negatives than on the other three sentence-types.
Eifermann (1961) repeated the Wason experiment on Hebrew speakers with two
types of Hebrew negation. .Again, she used sentences like Seven· is ·not an even
number, in its Hebrew form, and timed subjects' latencies to a correct verification.
For the Hebrew negative lo, True Affirmatives were verified 67 msec. faster than
False Affirmatives, and True Negatives 51 msec faster than False Negatives. For
the other negative eyno;· True Affirmatives were 232 msec faster than False Affir-
matives, and: True Negatives 24 msec slower than False Negatives. Apparently,
there was a mixture of people in both groups using the 'true' and:iconversion'
methods. In the error data, however, there was a clear tendency in both kinds of
negatives for more True Negative errors than False Negative ones. This foiiows
directly from the extra comparison operation required for the True Negative sen-
tences under the 'true' method of negation.
Gough (1965, 1966) used a task in which subjects were first read a sentence and
were then shown a picture that was either true or false with respeCt to the sentence.
The sentences affirmed or denied that a boy or a girl hit or kicked a boy or a girl,
1310 HERBERT H. CLARK

and the pictures depicted the same possible affirmative events. The subjects were
timed from the onset of the picture to their answer. In three separate experiments,
Gough found that True Affirmatives were verified faster than False Affirmatives by
125 msec to 165 msec. But in one experiment he found True and False Negatives
to have the same latency, in another True Negatives were 55 msec FASTER than
False Negatives, and in the final experiment True Negatives were 175 msec SLOWER
than False Negatives. That is, the third experiment fits the 'true' method, while the
first two do not. However, as Slobin (1966) pointed out, in the first two experi-
ments .it was sometimes possible to short-cut the verification of True Negatives
merely by noting that the picture was inapplicable to the sentence - e.g. it con-
tained. two girls, while the sentence talked only of boys. In the final experiment,
however, the sentences and pictures were both restricted to a boy and girl, one
hitting the other. In this instance, it is difficult to apply a 'conversion' method or
use any shortcuts, but it IS easy to apply the 'true' method, and the latter predicts
the data quite well.
Slobin (1966), using a method quite similar conceptually to the third experiment
of Gough's, also obtained data that fit the 'true' method quite well. For his subjects,
True Affirmatives were 130 msec faster than False Affirmatives, and True Nega-
tives were 230 msec slower than False Negatives. Although these two differences
(130 msec and 230 msec) should be the same, they were apparently not reliably
different from each other.
An experiment by Trabasso, Rollins, and Shaughnessy (1971) further illus-
trates the difference between the 'true' and 'conversion' methods of negation, but
in quite a different way. Basically, Trabasso et al. attempted to separate out the
Stage 1 encoding time for negatives from the Stage 3 comparison times. They were
able to do this by presenting a 'sentence' to. the subject, who, when he had read and
encoded it, pressed a button that presented the verifying picture; Trabasso et al.
measured both the encoding time and the verification time. In one experiment, they
presented •subjects with 'sentences' like luz green ('is green') rand kov orange ('isn't
orange') under one of two conditions: in one, the sentences and pictures contained
only two colors; in the other, they contained four colors. The subjects .in the first
condition, then, conld use the '(:onversion' method, since kov orange always implied
green and kov green always implied orange. And the results showed that the sub-
j~cts did carry out these conversions. The times at 'the encoding stage were longer
for the negatives (kov) than for the affirmatives, since subjects were changing the
negatives into the appropriate affirmatives during this preverification phase. But
the verification times were approximately the same for affirmatives and negatives.
The result is exactly what the 'conversion' method wonld predict, since at the be-
ginning of the verification stage all the sentences are affirmative and should there-
fore alltake about ·the same time. The subjects iri the ·second condition, however,
because they could not use the 'conversion' method, were forced to use the 'true'
method because kov green ('not green') did not imply any single color; therefore,
 SEMANTICS AND COMPREHENSION 1311
kov green could not be converted into a single positive color name at the time of en-
coding. Herrce, although the Stage 1 encoding times for negatives were still longer
than for affirmatives in this second condition, the Stage 3 verification times for
negatives were fully 255 msec longer than for affirmatives, just as the 'true' method
would predict.
In another experiment Trabasso et al. presented the picture before the 'sentence'.
Here they expected the subjects not to use the 'conversion' method, since the 'true'
method in this case would be far sinipler to carry out. Indeed, their results fit the
predictions of the 'true' method quite well. True Affirmatives were verified 95 msec
faster than False Affirmatives, whereas True Negatives were verified 57 msec
slower than False Negatives, and the difference between 95 msec and 57 msec was
not reliable.
In another recent experiment, Wales and Grieve (1969) collected data that
further confirm the 'true' method of negation. They asked subjects to verify sen-
- tences like Given 6 and 8, the next number is not 1, in which the numbers were
supposed to add up to 15. Note that in such problems as these, there is no possi-
bility for using a 'conversion' method. As expected, the results are in approximate
agreement with the 'true' method. In the group of subjects for whom the difficulty
of the addition task was not corrfounded with verification difficulty, True Affirma-
tives were verified 140 msec faster than False Affirmatives, and True Negatives
610 msec slower than False Negatives; however, it is not clear whether the differ-
ence between 140 and 610 is statistically reliable, and if so, why.
Finally, Greene (1970a, b) asked her subjects to judge whether two sentences
were the same or different in meaning. Although this procedure is different from
the previous ones, it can be conceptualized in exactly the same way. The process
in this case would consist of .setting up a mental representation of the first sentence,
setting up one of the second sentence, and then comparing the two representations
by the 'true' method of negation. Although Greene had her subjects compare the
meanings of several kinds of sentences, the conditions of interest here are those in
which subjects were given an active sentence, e.g. x exceeds y, followed by a nega-
tive sentence with the same meaning, y does not exceed x, or one with different
meaning, x does not exceed y. In the present nomenclature the two latter sentences
can be considered to be a True Negative and a False Negative, respectively, vis a
vis the first sentence. According to her results, subjects were able to verify that
a True Negative had .the same meaning as the first sentence 730 msec slower than
that a False Negative had a different meaning, exactly in agreement with the 'true'
method of negation; when there was an additional passive transformation involved
(e.g. when the True Negative was y is not exceeded by x), then True Negatives were
stlll 280 msec slower than False Negatives. So even in this quite different sort of
task, the 'true' method of negation is stlll able to account for the relative com-
prehension and comparison tinies for negative sentences.
In summary, all the studies of explicit negation we have exanrined are approxi-
 1312 HERBERT H. CLARK

mately consistent with the 'true' method, the 'conversion' method, or with some
combination of the two. I say 'approximately' because there are several results that
deviate slightly from the precise predictions of the two models, although these
deviations appeared in experiments with considerably more experimental error, with
'looser' timing methods, and so on than the confirming cases. Probably the most
precise techniques used were those in the Oark and Chase, Young and Chase, and
Trabasso et al. experiments, and they were the studies that showed the most precise
confirmation of these models. The 'true' and 'conversion' models therefore find
considerable support in the previous literature. This is not to say that these two
models explain everything. For one thing, previous studies show considerable varia-
tion in the relative difficulty of negative sentences. In Trabasso .et aL, for example,
negatives were approximately 200 msec slower than affirmatives, while in the Clark
and Chase experiment, negatives were 685 msec slower than affirmatives. Why
should Negation Time vary so much between these two studies? Part of the answer,
I will argue, is to be found in differences in the scope of negation. Jlut to discuss
this point and the other studies on negation, I must first describe, classify, and
represent the different types of negation found in English.

23 Types of Negation in English

As I pointed out earlier, English contains many types of negation other than the
simple sentence negation. Negation, for example, seems to be implicit in many
single lexical· items, like absent, forget, few, small, below, and so on. What are the
differences between these various types of negation? One important distinction to
be made is betwe~J,I explicit and implicit negation, between e.g. isn't present and
absent. Although this division can .be made on syntactic grounds, the division will
also be shown to be the result of differing presuppositions on the part of the
. speaker. A second important distinction to be made is between full and quantifier
negation, between e.g. not many and few. Although there are probably many other
possible distinctions.among negatives, these two are important because they can be
shown to be implicated directly in the way negatives are comprehended.

2.31 Explicit versus. implicit negation

The.sentence·Helen isn't home· denies that Helen is at.home. When would it be.
appropriate,for.,me to.make such a denial? If I ·thought that you had expected
Helen to .be home, or had said so, or had implied so in what you had just said, I
would be very likely to say, Helen isn't .at home. Note that it seems appropriate
to say I know you think Helen is at home, but she. isn't at home, but quite in-
appropriate to say I know you think He/en is at school, but she isn'tat home. These
examples hint at the fact that a speaker usually makes an assumption about the
·beliefs (or apparent beliefs) of his listener whenever he utters a denial. Specifically,
 SEMANTICS AND COMPREHENSION 1313
he normally supposes that the listener does or could well believe in the truth of
what is being denied. In saying Helen isn't at home, the speaker supposes, for any
one of a number of reasons, that the listener believes that Helen is or might be at
home. As Wason (1965) pointed out, The whole isn't a fish is a plausible negative
since it is reasonable to assume that a listener might think that whales are fish; but
The whole isn't a bird, though true, sounds highly incongruous, for surely no one
could believe that a whale is a bird. It appears, then, that a full representation of
what a listener understands of a negative sentence ought to contain something about
the sentence's (or speaker's) suppositions.
Denials of this sort, however, appear to be different from other types of negatives,
and this distinction coincides with Klima's (1964} distinction between affirmative
and negative sentences. Negatives can be identified, according to Klima, by whether
they can co-occur with any in the same clause and with either as a final tag when
the negative is in the second of two and-conjoined clauses. Consider the criterion
of any-acceptance first. None is negative since 4d is accep1Bble:
4 a) Ali of the men had some dinner.
b) *None of the men had some dinner.
c) *All of the men had any dinner.
d) None of the men had any dinner.
The criterion of either-conjunction leads to the same conclusion, since Sd is also
acceptable:
5 a) The men left and all of the women left too.
b) *The men stayed and none of the women left too.
c) *The men left and all of the women left either.
d) The men stayed and none of the women left either.
By these two criteria, there.area number of specific words in English that tum out
to be negative, e.g. scarcely, hardly, few, seldom, little, etc., besides the obvious
ones, e.g. not, no, none, never, etc.
The interesting point about these criteria is that they fail to identify such words
as absent, forget, except, and. without as negative; even though these words are
approximately synonymous with expressions that are considered negative, i.e. not
present, not remember, but not, and not with respectively. Thus, Klima's criteria
can be used to separate out explicit negatives (like not present) from implicit
negatives (like absent): the explicit negatives conform to his co-occurrence rules,
whereas the implicit negatives fail to do so. A quick perusal of English reveals
any number of explicit/implicit pairs: not have/lack, not succeed/fail, not success-
ful/unsuccessful, and so on.
The explicit/implicit distinction can be seen most clearly in a striking minimal
pair- few (an explicit negative) and a few (an implicit negative). They are distin-
guished by any-!'cceptance:
6 a) Few of the men ate any dinner.
b) *A few of the men ate any dinner.
1314 HERBERT H. CLARK

and by either-conjunction:
7 a) Many men stayed and few women left either.
b) *Many men stayed and a few women left either.
Although at first glance few and a few appear to mean the same thing, a closer look
shows that they differ in what they suppose of the listener's or speaker's expec-
tations. The sentence Few men left, for example, supposes that many or all of the
men were expected to leave. Thus, of the two sentences,
8 a) I expected all the men to leave, but few did.
b) *I expected none of the men to leave, bnt few did.
only Sa contains an acceptable prior expectation. On the other hand, A few men
left supposes that few or perhaps none of the men were expected to leave. Thus
a few has exactly the opposite effects on acceptability as few, as can be seen in 9:
9 a) *I expected all of the men to leave, but a few men did.
b) I expected none of the men to leave, but a few men did.
Note that 9a can be made acceptable by adding only before a few, since only is an
explicit negative that thereby makes the sentence acceptable (cf. Hom 1969); 9a
can also be made acceptable by placing contrastive stress on few, which has approx-
imately the same consequence as adding only.
In short, explicit negatives actually deny positive suppositions on the part of the
speaker or listener (No, it isn't true. Few men left.), while implicit negatives
merely affirm the already negative suppositions of the speaker or listener (Yes, it's
true. A few men left.). In this sense, the explicit negatives really do deny, while
the implicit negatives actually affirm.
The main problem is how to represent these subtle differences in the notation for
their semantic representation. The differences between lOa, lla, and 12a seem to
be appropriately captured in the representations of lOb, llb, and l2b, respectively,
whose paraphrases are shown in c in each case:
10 a) Many men aren't leaving.
b) (false (suppose (men (men are many) leave)))
c) . It is false to.suppose that many men are leaving.
11 a) · Few men are leaving.
b) (men (false. (suppose (men are many))) leave)
c) It is false to suppose that the men who are leaving are many.
12 a) A few men are leaving.
·b) (men (suppose (false (men are many))) leave)
c) It is correct to suppose that the men who are leaving are not many.
The main point of this notation is to show how these three types of negation differ
in their scope. In 10, the whole sentence is found within the scope of the false;
in 11, it is the supposition that is denied; and in 12, false lies within the scope of
the supposition. To put it still another way, Few men are leaving makes the as-
sumption that some men are leaving, but asserts that the number is not many,
whereas A few men are leaving makes the assumption. that the number of men
referred to is not many and it affirms that this number is leaving.
 SEMANTICS AND COMPREHENSION 1315
We now have come to the point where we can distinguish three different 'dimen-
sions': positive-negative, affirmation-deuial, and agreement-contradiction. First of
all, a sentence with a negative is not necessarily a denial. A deuial is specifically
a sentence that asserts that something is false, where that something is presupposed
to be possible. Some negatives are not denials, e.g. A few men left, since it is only
their suppositions that are negative. It is clear nnder these distinctions, then, that
Klima's criterion of either-conjunction is a criterion for denials, not for negation.
Note 13 and 14:
13 The men without hats caught cold, and the women who didn't wear
hats caught cold '
a) too.
b) *either,
14 The men who caught cold were without hats, and the women who caught
cold didn't wear hats
a) *too.
b) either.
In 13, there is clearly a negative (didn't) within the second main clause, but the
second clause itself is an affirmation, so it takes too, not either. On the other hand,
the second main clause of 14 is a denial, so it takes either, not too, The negative
in 13, since it is contained in a restrictive relative clause, is part of a p~esnpposi­
tion of the second clause: the second clause of 13 presupposes that there were women
who didn't wear hats (cf. Vendler 1967). If we think of a sentence as consisting of an
assertion plus perhaps some presuppositions, then a negative in the assertion makes
the sentence a denial, whereas a negative in the presupposition leaves the sentence
simply as an affirmation. Notice that this is just the difference I have tried to
capture between few, a deuial with a positive supposition, and a few, an affirmation
with a negative supposition. Finally, agreement-contradiction is distinguishable
from both positive-negative and affirmation-deuial. An agreement can be either
positive or negative, and either an affirmation or a deuial:
15 a) So Mary has been here all day? Indeed, she has.
b) So Mary hasn't been here all day? Indeed, she hasn't.
And so can a contradiction:
16 a) So Mary has been here all day? I'm sorry, she hasn't.
b) So Mary hasn't been here all day? I'm sorry, she has.
The matter is certainly more complicated than I have indicated here, but these
are the main distinctions to be made.
Another example of a minimal pair is not present (an explicit negative) and
absent (an implicit one), To say John isn't present and John is absent is clearly to
refer to the same objective situation. Again, the difference between them lies in
whether one wants to deny the supposition that John is present, as in John isn't
present, or to affirm the supposition that John isn't present as in John is absent.
I should add one caution here, I have pointed out that the supposition of the
1316 HERBERT H. CLARK

sentence normally refers to assumptions that the speaker has supposed that the
listener has made about the subject at hand. However, this does not appear to be
a hard and fast rule. When I say Helen isn't at home, what it appears that I am
saying is really this: 'Suppose that Helen is at home; well, that supposition is false.'
The supposition is like a temporary condition set up so that I can make some point
with reference to that condition. The temporary condition set up, of course, will
not be an arbitrary one; it will normally be. pertinent to the speaker's and listener's
beliefs at the moment. Furthermore, it appears that what I have been calling sup-
position is closely related, if not identical, to presupposition, as identified e.g. by
Fillmore (1970) or Lakoff (1970). A presupposition or' a sentence is a proposition
that must be true for the sentence to be 'felicitous'. For example, for John to say,
'You should stop beating your wife', he would have to believe that you have been
beating your wife; otherwise, the sentence would be infelicitous. The sentence
*I know you have never beaten your wife, but you should stop beating her is
infelicitous (and unacceptable). This type of nnacceptability seems to have its
parallel in negation; e.g. *I know you think that John is absent, and he isn't present.
The unacceptability is not quite as severe here possibly, but it seems to be similar
in origin. On the other hand, the usual test for whether a part ofthe meaning of a
sentence is a presupposition. or not is whether that part changes meaning nnder
negation. For example, Stop beating your wife and Don't stop beating your wife
both presuppose that you have been beating your wife, so the latter is a presupposi-
tion, not an implication of the first two sentences. In the same sense, it seems that
both John is absent and John isn't absent 'presuppose' an assumption (perhaps very
temporary) that John isn't present. But it is impossible to apply the same criterion
to John isn't present, which already contains an explicit sentence negative. I have,
therefore, tentatively called the phenomenon jnst discussed for negatives 'supposi"
tion', instead· of 'ptesupposition', but the hypothesis is that supposition and pre"
supposition are probably very closely related. For now, we can consider supposition
to be simply a temporary: condition or assumption set up for the sentence.

2.32 Full versus•quantifier negation

The second distinction to be made among the negatives is between full and quan-
tifier negatives. The distinction•is parallel to the difference between antonyms that
are contradictories and those that are contraries (cf. Lyons 1968). Present and
absent are contradictories, since not present implies absent and vice versa; ln con-
trast, many and few are ouly contraries, since not many does not necessarily imply
few, although jew always implies not many. To put this another way, present
and absent are contradictories since *John· is neither present nor absent is a con-
tradictory sentence; many and few are contraries since There were neither many
nor few people at the movie is not contradictory. Thus, l wish to call absent a fuU
negative, and few a quantifier negative.
This distinction is quite independent of the explicit"implicit distinction. There-
 SEMANTICS AND COMPREHENSION 1317
fore, there are four possible classes of negatives: (1) explicit full negatives, like
not present; (2) implicit full negatives, like absent; (3) explicit quantifier negatives,
like few; and (4) implicit quantifier negatives, like a few. Unfortunately, this brings
up a notational problem. The notation I have used for few is indistinguishable from
the one that would be used for not many. The false applied to many to produce
few is actually only applied to a feature - the feature +Polar - of the word
many; in contrast, the false applied to many to get not many applies somehow to
the whole lexical item. It is not immediately obvious how this difference is easily
represented in a simple notation, so I have ignored this problem in the present
section. The convention to be used here is that a false appli~d to a quantifier
produces a quantifier negative only (e.g. few), and not a full negative (e.g. not
many). Ignoring the difference between the two types of false has no consequences
, in the present chapter.
i
By this characterization, full and quantifier negatives differ slightly in scope.
A fnll negative will always contain complete propositions or lexical items within
its scope, as in e.g. not many. In contrast, a quantifier negative will contain only
a single feature such as +Polar within its scope, as in e.g. few. This difference
in scope will be important in a review of the studies that compare different types
of negation.

2.4 Previous Studies of Implicit Negation

Although there are few studies of implicit negation in the literature, the studies that
.exist show that the 'true' method of negation is just as applicable to implicit nega-
tion as to explicit negation, although with several important differences. These
studies also suggest a secorid hypothesis, something I will call the scope of the
negation hypothesis, which is this: the greater the scope of the negative the more
difficult the negative is to comprehend. More specifically, this hypothesis asserts
that Negation Time will increase for increases in scope of negation. In order to
examine these two hypotheses, I will ·review three series of experiments, one on
except, a second on not present and absent, and a third on a number of explicitly
or implicitly negative quantifiers. Next, I will collect together all the Negation and
Falsification Times found. in the literature and compare therri in the light of these
two hypotheses. Finally; I will discuss two previous studies aud how the notions
of supposition just discussed apply to them. ·

2.41 Except
i The first implicit negative to be examined is except. It has been carefully inves-
tigated by Sheila Jones (1966a, b; 1968) in an important series of experiments on
the nse of except in instructions. These experiments are particularly significant,
I
i
!
because they show that the 'true' method of negation is applicable to instructions
1318 HERBERT H. CLARK

as well as verification tasks and to implicit as well as explicit negatives, and because
they give the first evidence for the scope of negation hypothesis.
In her experiments, Jones gave her subjects a pencil, a sheet of paper filled with
the digits 1 through 8 in a random order, and asked them to cross out certain digits
according to instruction. She timed them for each sheet of digits they crossed out.
Some subjects were given the 'inclusive' instructions: 'Mark the numbers 3, 4, 7,
and 8', and others were given the 'exclusive' instructions: 'Mark all numbers except
1, 2, 5, and 6.' It should be noted that the two instructions are equivalent in two
important respects. Both require the subject to keep only four digits in mind, and
both require him to mark out digits 3, 4, 7, and 8. Th~ important difference is that
the inclusive instruction is positive, and the exclusive instruction implicitly negative.
Before examining the results, however, we should attempt to formulate how
Jones' subjects presumably wonld have represented the instructions mentally. The
inclusive instruction 17a might best be represented as a conditional, as in 17b,
which is paraphrased in 17c:
17 a) Mark the numbers 3, 4, 7, 8.
b) (you mark X (if (X is (3 or 4 or 7 or 8))))
c) Mark X (the number you are currently checking) if X is 3, 4, 7, or 8.
The corresponding exclusive instruction, its representation, and its paraphrase are
given in 18:
18 a) Mark all numbers except 1, 2, 5, 6.
b) (you mark X (if (false (X is (1 or 2 or 5 or 6)))))
c) Mark X if it isn't true that X is 1, 2, 5, or 6.
That is, these two instructions are identical in form, except that the inclusive in-
struction contains a positive conditional where the exclusive one contains a negative
one. Both instructions can be thought of as consisting of two parts: a test plus a
command. The test (if ...) is positive for the inclusive instruction and negative for
the exclusive instruction, whereas the command (Mark X) is positive under both
instructions ..
The effects of these instructions on crossing-out time are now easy to predict,
since each test (of each test-plus-command instruction) can be viewed simply as a
specific instance of the 'true' method of negation described earlier. ·At stage 1 of
that method, the subject codes the test as, for example, (false (X .is (1 or 2 or 5
or 6))); at Stage.2, he codes the number he is currently testing, say (X is 2); at
Stage 3, he compares the two codes by the four-operation comparison process
shown in Table II; and at Stage 4, he enters the answer into the command and
carries out the command. By this analysis, the inclusive instructions contain a test
that is a True Affirmative half the time and a False Affirmative the other half of
the time. In contrast, the exclusive instructions contain a test that is a True Nega-
tive half the time and a False Negative the ·other half of the time. Since affirmatives
require one less Stage 3 comparison operation, overall, than negatives, the inclusive
instructions should be faster than the exclusive instructions, even though both in-
 SEMANTICS AND COMPREHENSION 1319
structions make the subject carry out an equal number of tests and exactly the same
commands. In agreement with this prediction, Jones found that the inclusive in-
structions took about 67% as much time as the exclusive ones.
Jones (1966a) also carried out an interesting variation of the above instructions.
Instead of using the inclusive and exclusive instructions equivalent in their. test sets,
she changed the inclusive instruction to read 'Mark the numbers 1, 3, 4, 6, 7' and
the exclusive instruction to read 'Mark all numbers except 2, 5, 8'. For this in-
clusive instruction 5/8 of the tests will be True Affirmative, 3/8 will be False
Affirmative, but each test will take longer than before, since each X must be tested
against five, instead of four, members of the test set. For the exClusive instruction,
5/8 of the tests will be the more difficult True Negative, and 3/8 will be False
Negative, but here each negative test will take Jess time than before since the test
set consists of three, instead of four, digits. Jones found that under these instruc-
tions the difference between the two instructions was considerably reduced: the
inclusive instructions were carried out in about 95% of the time it took for the
exclusive instructions, although the difference was still statistically significant. This
agrees with intuition that the more digits there are to be marked out, the easier it
will be to follow the exclusive instructions. If the set of digits to be marked out
includes as many as seven of the eight digits, the exclusive instructions, which
specify the single digit not to be marked out, should be easier to follow than the
inclusive instructions, which specify the seven digits that are to be marked out,
just because the negative test set is so much smaller than the positive test set.
The errors subjects made in Jones' experiments demonstrate another phenome-
non, but it appears to be independent of the 'true' model of negation. On the
inclusive ·instructions in the second experiment of Jones (1966a), she found fewer
'false positive' errors (instances of digits incorrectly marked out) than 'omissions'
(failures to mark out a digit·that should have been marked out); the error percent-
ages were 2% and 6%, respectively. In contrast, the exclusive instructions showed
more 'false positive' errors than 'omissions', 8% to 4%. The same pattern of
errors is found in Jones 1966b, with the four conditions eliciting 1%, 6%, 8% and
6% errors, respectively. Such a pattern can be accounted for under the assumption
that people are more apt to miss a digit they are looking for than select a digit as
being in the test set when it is not. This could well be a perceptually motivated
phenomenon based on the way people search the visual features of a digit for a
match with the digits in memory. In any case, the pattern could not have been
produced by the 'true' model, since for the inclusive instructions it woUld predict
more False Affirmative errors ('false positive' errors) than True Affirmative errors
('omissions') and for the exclusive instructions more True Negative errors ('omis-
sions') than False Negative errors ('false positive' errors). The data were just the
reverse in both cases. In agreement with the 'true' model, however, there were more
errors on the negative than affirmative instructions in all the experiments. This
effect is independent of the possible perceptual scanning strategies.
1320 HERBERT H. CLARK

In another experiment, Jones (1968) gave her subjects either explicit or implicit
negations as instructions for the same type of task. Her instructions were the
following:
19 a) 'Mark all the numbers except 2, 5, 8'
b) 'Except for the numbers 2, 5, 8, mark all the rest'
20 a) 'Mark all the numbers, but not 2, 5, 8'
b) 'Do not mark the numbers 2, 5, 8; mark all the rest'
The instructions in 19 contain the implicit negative except, whereas those in 20
contain the explicit negative not. Her main finding was that the implicit instruc-
tions 19 were easier to carry out than the explicit ones 20. She also found that
when she asked .her subjects what they were doing both just before and just after
the task, they overwhelmingly tended to restate the instruction they were given as
an implicit negative 'self-instruction', especially as instruction 19a.
If the theory of comprehension we are examining is valid for this task, we must
be able to trace the difference between the implicit and explicit instructions to a
difference in the mental representations. At first glance, the two instructions
appear to be identical: they both specify only the set that is not to be marked. But on
closer examination, they differ in one important respect. If we again think of each
instruction as consisting of a test plus a command, the explicit negative instructions
place the negative on the command and test together, whereas the implicit negative
instructions place the negative on only the test, as we saw above. The represen-
tation for 19 and for 20 would therefore look something like 19c and 20c, respec-
tively:
19 c) (you mark X (if (false (X is (2 or 5 or 8)))))
20 c) (false (you mark X (if (X is (2 or 5 or 8)))))
So in the explicit instruction 20, the test itself will be either a True or False Af-
firmative, but the command - the consequent - will be either a True or False
Negative; the reverse holds for the implicit instruction 19. That 19 was easier to
carry out than 20 is therefore reflected in a real difference in their mental represen-
tations.
Here is the first evidence we have that scope of negation is important. Note that
the scope of negation in the easier instruction 19 is less than in 20, where the
negation includes everything in its scope. The difficulty in handling the wider
negative scope of 20 would also explain why Jones' subjects preferred to refor-
mulate it as the referentially equivalent, but psychologically easier, implicit in-
struction 19.

2.42 Present and absent

The next implicit negative to be examined is absent, a full negative like except.
Absent cannot take any (*John has been absent at any parties this month), whereas
its explicit negative counterpart not present can (John hasn't been present at any
parties this month). Absent also does not allow either-conjunction (*John was
 SEMANTICS AND COMPREHENSION 1321
gone and Mary was absent either), although not present does (John was gone and
Mary wasn't present either). The most important property of absent, however, is
that it is referentially identical with not present: the sentences John is absent and
John isn't present will dearly always describe the same referential situation. The
difference between them, as I pointed out before, is in their suppositions. John isn't
present denies the supposition that John is present, whereas John is absent affirms
a supposition that John isn't present. Concomitant with this difference in suppo-
sition is a difference in scope of negation: the scope of the negative in isn't present
is larger than the scope of the negative in absent. So here again we have a case in
which we can see whether the 'true' method of negation applies to implicit negatives
as well as whether the scope of negation hypothesis is confirmed or not. The rele-
vant experiments are reported in Clark and Young (in preparation).
The predictions of the 'true' method of negation are more clearly visible after the
assumed mental representations of the possible sentences have been made explicit.
The four types of sentences used in Clark and Young are shown in 21 through 24
followed by their representations and the approximate paraphrases of those re-
presentations:
21 a) A is present.
b) (true (suppose (A is present)))
c) It is correct to suppose that A is present.
22 a) A isn't present.
b) (false (suppose (A is present)))
c) It is false to suppose that A is present.
23 a) A is absent.
b) (true (suppose (false (A is present))))
c) It is correct to suppose that it is false that A is present.
24 a) A isn't absent·
b) (false (suppose (false (A is present))))
c) It is false to suppose that it is false that A is present.
The not present versus absent difference is now well defined, since A isn't ,present
and A is absent have a true and a false interchanged: 'It is false to suppose that it is
true that A is present' versus 'It is trueto suppose that it is false that A is present'.
In the actual task of Clark and Young (in preparation), subjects were asked to
verify a sentence describing whether a star or plus was or wasn't present or absent;
on the opposite side· of the display, the picture was always either a star (*) or a
plus (+). The subjects were timed from the moment they looked at the display to
the moment they pushed the 'true' or 'false' button:
The 'true' model must be recast slightly for present and absent, since the mode!
must now be able to accommodate the suppose of the sentence representations and
the double negative in isn't absent. At Stage 1, the subject would represent the
sentences as in 21 through 24. At Stage 2, he would represent the pictures either
as (star is present) or (plus is present), depending on which figure was there. It is
1322 HERBERT H. CLARK

at Stage 3 that a problem arises, since the four-operation process (in Table II) has
no devices for handling either suppose or double negatives. The solution I will
use temporarily is simply to ignore the embedding strings labeled suppose and to
add a third pair of operations (Operations 3 and 3a), which compare the next
embedding strings out and change the truth index on finding a mismatch. The result
of this proposal is that the 'true' model applies to not present and absent in the
identical way and that the false double negative isn't absent should be slowest to
verify since its truth index would have to be changed four times.
The results of Oark and Young confirm the predictions from the 'true' model of
negation. True present statements were verified about 180 msec faster than false
present statements, just as they should be for affirmative sentences. In contrast,
both true not present and true absent statements were verified 95 msec slower,
respectively, than their false counterparts, just as they should be for negative sen-
tences. The two increments of time here should be the same if subjects were always
going through the processes exactly as outlined above; the difference between the
two increments, however, suggests that some subjects might have been using a
method to be described in 2.5 below. Finally, tnie isn't absent statements were
verified more quickly than false ones; however, almost all the subjects reported
'converting' isn't absent into is present, making the predictions for the isn't absent
case inapplicable.
These results also support the scope of negation hypothesis. It was found that
absent statements were 371 msec slower, overall, than present statements, whereas
not present statements were fully 640 msec slower than present statements. That
is, isn't present took 270 msec longer than absent to verify. This difference cannot.
be accounted for by differences in the referential properties of isn't present and
absent, since there are none. It also cannot be accounted for by the extra reading
time of isn't present over absent, since careful measurements show that reading
could account for no more than 20 msec of this difference and it probably accounts
for less. The sole difference between isn't present and absent is therefore in the
semantic representations of thesetwo expressions: the negative of isn't present con-
tains the supposition, while the negative of is absent is contained in the supposition.
The 270 msec:difference - at least most of i t - must therefore be associated with
this difference in representation.
What is the source of the 270 msec difference? There. are two possibilities.
(1) Isn't present could take longer to represent at Stage l than absent. Or (2) isn't
present could consume more time in the Stage 3 operations than absent. Although
very little evidence is available on this issue, one of Jones' results already discussed
seems to suggest that (2) might be true. She found that except instructions were
easier than but not instructions, and we noted that this difference was related to a
difference in scope of negation in the. representations. of these instructions. The
important point, though, is. that her subjects presumably comprehended the instruc-
timi just once - at the beginning of the task - but used the instruction over and
 SEMANTICS AND COMPREHENSION 1323
over again in crossing ont the digits. In other words, Stage 1 occurred once, bnt
Stage 3 recurred throughout the experiment. The advantage of except over but not
must therefore be located at Stage 3. This suggests that the d increment, the time
taken by Operation 2a in the comparison process of Table II, is simply longer
when the false dominates more in its embedded string, as it does for both isn't
present (vs. absent) and but not (vs. except).
Obviously, the evidence available does not allow us to specify. the mechanisms
that underlie the scope of negation hypothesis. We must be satisfied here with a
statement of the hypothesis; we will have to account for it later. ·The actual mech-
anism might necessitate a slight change in the specific form of the Stage 3 opera-
tions listed in Table II, but the operations must nevertheless be approximately
correct, since they are necessary to account for the explicit negatives, for isn't
present, absent, except, and but not, and for several other examples that follow.
It is probably just as important as the hypothesis itself that we have identified an
important difference in comprehension that can only be specified as a difference in
semantic representations of sentences; in particular, the semantic difference is in the
form of the suppositions of the sentences. We will now examine another instance
where supposition is critical in accounting for the comprehension of certain im-
plicit negatives.

2.43 The Just and Carpenter study

Marcel Just and Patricia Carpenter (1971) have completed a comparative
study of several types ·of negation in English, In their task, they asked subjects to
verify sentences like Most of the dots are black and A small proportion of the
dots are black against pictures that contained either 14 black dots and 2 red dots
or the reverse. Jnst and Carpenter explicitly set out to investigate three categories
of negation. In Category 1, there were three exemplars of explicit negation:
(A) The dots are red vs. The dots aren't red; (B) There are red dots vs. There are
no red dots; and (C) All of the dots are red vs. None of the dots are red. (Unlike
the pictures for Categories 2 and 3, the pictures paired with the Category 1 sen-
tences contained either all red or all black dots,) Clearly, Exemplar A contains a
sentence negation, while in Exemplars B and C, the negative encompasses only the
quantifier and therefore has a slightly smaller scope. Category 2 contained exemplars
of explicit quantifier negation: (D) Most of the dots are red vs. Hardly any of. the
dots are red; (E) Lots of the dots are red vs. Scarcely any of the dots .are red; and
(F) Many of the dots are red vs, Few of the dots are red. Finally, Category 3 con-
tained three exemplars·of implicit quantifier negation: (G) The majority of the dots
are red vs. The minority of the dots are red; (H) A large proportion of the dots are
red vs. A small proportion of the dots are red; and (I) 14 out of 16 dots are red vs.
2 out of 16 dots are red.
The results of the Jnst and Carpenter study support the 'true' model of negation
in several very important ways. First, they found that negatives were more difficult
1324 HERBERT H. CLARK

than affirmatives in all three categories. Second, they found that in Categories 1
and 2, which contained all explicit negatives, True Positives were faster than False
Positives, and True Negatives were slower than False Negatives. For Category 1
this is a replication of previous results, but for Category 2 this is a result that shows
that the 'true' model of negation also holds for explicitly negative quantifiers, like
few, scarcely any, and hardly any. But third, they found, quite unexpectedly, that
in Category 3 True Positives anq True Negatives were both verified faster than
their false counterparts. Although this result appeared somewhat mysterious at
first, Just and Carpenter showed that this apparent inconsistency was nonetheless
in agreement with the 'true' model of negation.
Just and Carpenter accounted for the Category 3 inconsistency by the way the
pictures were coded for those exemplars. (I have recast Just and Carpenter's ex-
planation in terms of supposition to make it compatible with the present analysis
of negatives.) Consider many-few (a Category 2 pair of quantifiers) vs. many-a few
(substituting for a Category 3 pair of quantifiers). When verifying the true sentence
Many of the dots are black- i.e. (true (suppose (dots (dots black) are many)))-,
the subject can choose whether to code the majority of the dots in the picture (the
14 black), or the minority (the 2 red). In the former case, he would code the
picture as (dots (dots black) are many), but in the latter case, he would code it as
(false (dots (dots red) are many)). Note that the coding for the majority of the dots
will be congruent with the sentence, bnt the coding of the minority will not. Jnst
and Carpenter assumed that subjects coded the majority of dots for both many
and few (Category 2); this would produce the correct predictions from. the 'true'
model of negation. In· contrast; they were assumed to code the majority and mi-
nority, respectively, for many and a few (Category 3); this likewise allows the cor-
rect predictions from the 'true' model. Restated slightly, subjects were always
assumed to code the picture in terms of the supposition of the sentence: many and
few both contain suppositions about the majority ·of the dots; but mariy and a few
contain suppositions abouithe majority and minority of the dots, respectively. This
coding is hatnral for other .reasons. Mqny and few respectively affirm and deny
something about the majority of .the dots; in contrast, many and a few are both
affirt:nations, one about the majority of the dots and the other about the minority .
.. In order to justify this \':J<planationcof·the difference.between Categories 2 and 3,
Jnst·and C!itpeilter re-ran the.two ca(egories,·bvt.this. time they presented the
pictures a half-second before the sentence appeared. In. this way the coding of the
pictures ·could not be contingent on the type. of; negative in the sentence. In orie
condition, the subjectS were instructed to code the majority of the dots from the
picture, and in the other, the minority. ·Their results. here were in full agreement
with their previous explanation. When subjectS coded·the majority of the dots only,
Categories 2 and 3 were in complete agreement, in contrast with the results of. the
first experiment. And when subjects ·ooded the minority only, the two categories
were again in complete agreement, but in this case True Positives and False Nega-
 SEMANTICS AND COMPREHENSION 1325
tives were slower, instead of faster, than False Positives and True Negatives, re-
spectively. This result is also consistent with the 'true' model of negation, once the
codings of the pictures produced by the new instructions have been taken into
account. The conclusion is that when subjects are forced to code the pictures in
one particular way, the subjects then verify the exemplars of Categories 2 and 3 in
exactly the same way. In all, the Just and Carpenter results fit nicely into the
present unified explanation for negation.
The Just and Carpenter results also support the scope of negation hypothesis.
Their exemplars of negation can be approximately ordered for decreasing scope.
Exemplar A is sentence negation and should have roughly the greatest scope;
Exemplars B and C have less scope than A since they negate ouly a quantification;
Exemplars D, E, and F are all explicit quantifier negatives, so they are next; Exem-
plars G, H, and I have the least scope, since their negatives are included within the
supposition. Significantly, the Negation Time of these four groups of exemplars
decreased approximately with scope (cf. below) .•This evidence, however, is not as
direct as in the present-absent and except examples above, since so many other
things seem to be varying at the same time in Jnst and Carpenter's exemplars.

2.44 Other studies of implicit negation

Scattered throughout the psychological literature are a number of studies that
show that certain other negative adjectives and prepositions have properties con-
sonant with the 'true' model of negation. Although I report on these studies in
more detail in sections 3 and 4, I will summarize that evidence briefly here.
In Clark (1969a) I reviewed the previous literature on deductive reasoning that
used problems of the sort, 'If John is better than Pete, and Dick is better than
John, then who is best?' In.problems such as this, the comparative adjective could
be either a positive adjective or an hnplicit negative one - e.g. better or worse.
The previous literature on these problems shows that the positive adjectives in com-
parative form are easier to comprehend than the negative adjectives. Problems
with the comparatives of good, much, fast, far, tall, happy, warm, deep, old, and
high were solved consistently faster or with fewer errors than problems with their
respective implicitly negative opposites. This evidence, along with the results of
Just and Carpenter, leads to the tentative conclusion that all the lexically marked
adjectives like bad, little, slow, etc., are implicit (quantifier) negatives.
Prepositions can also be either hnplicit full negatives or hnplicit quantifier nega-
tives. Out, for example, is a full negative preposition, since John isn't in the room
and John is out of the room imply each other. The three pairs of full negative pre-
positions in English are to and from, in and out, and on and off. Carole Offir and I
recently studied to and from and into and out of in a verification task and found that
sentences with to and into were verified more quickly than sentences with from and
out of, respectively. Examples of preposition pairs that contain implicit quantifier
negatives are above and below, ahead of and behind, and in front of and in back of.
1326 HERBERT H. CLARK

They are quantifier pairs since, for example, A isn't above B does not imply A is
below B although the latter implies the former. Among pairs of this sort, a review of
the literature shows that above, on top of, in front of, ahead of, and before (the
positive members of the pairs) are comprehended more quickly than their negative
opposites below, under, in back of, behind, and after.
In summary, the evidence for the comprehension of certain implicitly negative
adjectives and prepositions is in general agreement with the 'true' model of negation.

2.45 Comparison of previous studies on negation

It is instructive at this point to collect together the results from a number of past
studies, for this evidence adds somewhat more support to the scope of negation
hypothesis as well as to the 'true' model of negation. In Tables V through VIII,
I have listed the Negation and Falsification Times from the studies that have
measured the latencies in the verification of positive and negative sentences from
the moment the sentence was presented to the moment the subject responded. In
each table, I have listed the investigators of the particular negative, the negative
studied (the positive counterpart of which is usually obvious), and the Negation and
Falsification Times for each negative. The four tables list, respectively, explicit full
negatives, explicit quantifier negatives (with two exceptions to be noted later),
implicit full negatives, and implicit quantifier negatives. The Falsification Times
were estimated in each case from the difference between false and true positive
sentences only, since in many studies the subjects were presumed to use a 'conver-
sion' method, thus making the negative sentences unusable for this purpose.
The first observation that one ·can make about these tables is that Negation Time

TABLE V. Negation times for explicit full negation

Negation Falsification
Source Example of Negative Sentence Used Time Time

Wason Seven is not an even number~ 550 260

Wason & Jones Seven iS nOt an even number. 440 420
Eifermann Seven is not even (in Hebrew with lo). 736 67
Seven is not even (in Hebrew witheyno), 600 232
Clark & Chase Star isn't above plus. 685 187
Star isn't above line. 7()9 155
Star isn't above plus. (picture first) 556 148
Star isn't above plus. (picture first) 660 145
Clark & Young The star isn't present. 640 187
Just & Carpenter The dots aren't red. 463 133
Wales & Grieve Given 6 and 8, the next number is not 1. 455 140
Median 600 155
 SEMANTICS AND COMPREHENSION 1327
TABLE VI. Negation times for explicit quantifier negation

Negation Falsification
Source Example of Negative Sentence Used Time Time
(I
Just & Carpenter None of the dots are red. 306 164
There are no red dots. 76 218
Hardly any of the dots are red. 385 167
Scarcely any of the dots are red. 384 161
Few of the dots are red. 210 263
Trabasso et a!. KOV green. 200 95
Median 258 166

TABLE vn. Negation times for implicit full negation

Negation Falsification
Source Example of Negative Sentence Used Time Time

Clark & Young The star is absent. 371 242

The hole is absent. 1851 291
Chase & Clark Absent. 145 219
Clark & Offir It is coming from town B. 184 155
Mary has just come out of the kitchen. 201 150
Median 185 219

TABLE VITI. Negation times for implicit quantifier negation

Negation Falsification
·.
Source Example of Negative Sentence Used Time Time
.- ~~--;' :, _.:. -_·:
93 187
·.'!
'·:-[ Clark &Chase Star is below plus,
135 155
I
! Star is below !ille.
Star is below plus. (picture first) 117 148
Star is below phis. (picture first) 84 145
I'
· Just & Carpenter A minority of the dots are red. 241 96
!~
,1;1 A small proportion of the dots are red. 214 240
jt;" 2 out of 16 of the dots are red. 31
72
!I Clark & Peterson Star is lower than plus. 251 180
I
Clark The pink one is in back of the blue one. 189 127
I Median 135 148
I
1328 HERBERT H. CLARK

is positive regardless of the negative, although Negation Time varies from about a
tenth of a second to almost three-quarters of a second. Falsification Time is also
positive in all these studies. However, it does not vary nearly as much as Negation
Time, but ranges between only about a thirtieth of a second to two-fifths of a second
with the majority of the Falsification Times below a fifth of a second. Most im-
portant, there are no exceptions to the proposal that Negation Time and Falsifi-
cation Time are truly positive increments.
The second observation is that Negation Times appear to vary systematically
across types of negation, whereas Falsification Times ·do not. The median Nega-
tion Times order themselves from longest to shortest as follows: explicit full nega-
tion (600 msec), explicit quantifier negation (258 msec), implicit full negation
(185 msec), and implicit quantifier negation (145 msec). That is, Negation Time
decreases approximately from explicit to implicit negation, and from full to quan-
tifier negation. In more detail, the explicit full negatives clearly have the longest
Negation Times; its exemplars, whlch vary from 440 msec to 736 msec, do not
overlap with those of any other category. Although the exemplars of the other three
categories have considerable overlap with one another, there is at least a hlut that
the medians might be reflecting real differences among the latter three categories.
In ·contrast, the median Falsification Times vary very little from category to
category, hovering around 160 msec for most exemplars in the four tables.
Tills second observation, then, supports the scope of negation hypothesis, since
Negation Times follow the scope of negation, whereas :Falsification Times do not.
Explicit negatives have greater scope of negation than implicit negatives, and full
negatives have greater scope than quantifier negatives, and these differences are
correlated with differences in Negation time. (It should be noted that the first two
exemplars in Table .VI are included there, not because they are truly explicit quan-
tifier negatives, but because their scope is less than the sentence negatives in Table
V and is approximately equal to the explicit quantifier negatives.) These data, then,
constitute support, albeit less than overwhelming support, for the scope of negation
hypothesis.
The second observation also supports the four 'Stage 3, comparison operations
shown in Table II. If Negation Time can vary independently of Falsification Time,
as these results show, then the mental operations that produce Negation Time must
be separate from the mental operations that produce Falsification Time. Signifi-
cantly, this is a property of the Stage 3 comparison operations of the 'true' model
of negation. Operations 1 and 1a .are quite separate from Operations 2 and 2a.
A third observation is that Negation Times and Falsification Times are relatively
homogeneous within each table despite the wide variety of exemplars there. Take,
for example, the explicit full negatives in Table V•• The exemplars in this table in-
clude locatives (Star isn't above plus), attributjves (The dots aren't retf), and predi•
cate nominatives (Seven is not· an even number). The exemplars also include
sentences in whlch only one word - other than the negative - is significant for
 SEMANTICS AND COMPREHENSION 1329
the verification (The dots aren't red), two words (Seven is not an even number),
or three words (Given 6 and 8, the next number is not 1). It appears that Negation
Time is around 600 msec, and Falsification Time is around 155 msec, regardless of
other incidental properties of the sentences: there does not seem to be any syste-
matic variation in Negation Time and Falsification Time with, for example, num-
ber of significant words. If these observations are supported by future research, it
constitutes evidence for the relative independence of Negation Time and Falsifi-
cation Time from certain other attributes of sentences.
Summary data like these, of course, should be interpreted With a great deal of
caution; First, the experimental methods varied considerably from study to study;
in some cases, the experimental error was large, but in others it was small. Second,
we have no idea how other differences in the negative exemplars - like the differ-
ence between from and out of - affect Negation Time and Falsification Time.
Furthermore, we have only separated out the negative exemplars according to gross
differences in scope of negation. Two exemplars within a single table might well
differ in scope of negation in a subtler way, and by hypothesis, this should have an
effect on Negation Time as well. Third, Negation Time can be significantly affected
by the range of objects (e.g. pictures) that the sentences are verified against. Con-
sider, for example, the three instances of absent listed in Table VII. The first
sentence was verified against a picture of a plus ('true') or a star ('false'), the second,
against a picture of an empty square ('true') or a circle within a square ('false'), and
the third, against a picture of a disk absent ('true') or present ('false'). The three
corresponding Negation Times, then, vary because of subtle interactions between
the negative sentences themselves and the type of evidence they are compared
against. Although I will discuss the reasons for the Negation Time differences in
this particular example later, I am quite uncertain as to which experimental varia-
tions are important in the other exemplars and which are not. Fourth, Negation
Times can be reduced with practice. Young and Chase (1971), and Singer,
Chase, Young, and Clark (1971), who ran subjects for 10 to 15 days,
found that although Negation Time began at 600 msec, it diminished to as little as
100 msec by the end of all that practice. Since the experiments listed in these four
tables could also well differ in how practiced the subjects were (either on negatives
in general or on the particular negatives of the experiment), some of the variation
in these tables might be due to practice effects.
In summary, the support for the scope of negation hypothesis, though quite
meager, is still suggestive. ~The most direct evidence is that Negation Time for
absent is considerably less than for not present (Clark and Young 1971),
and that the instruction 'Mark all the numbers except 2, 5, 8' is easier than the in-
struction 'Mark all the numbers but do not mark 2, 5, 8' (Jones 1968). Suggestive,
bnt not conclusive, is the combined evidence from the existing studies on negation.
They show that explicit full negation has longer Negation Time than explicit quan-
tifier negation, implicit full negation, or implicit quantifier negation, and that th~
1330 HERBERT H. CLARK

latter three categories might possibly be ordered in this way from longest to
shortest Negation Time. In contrast, Falsification Times do not differ much from
·category to category. As yet, there seem to be no counter-examples to the hypo-
thesis. Better confirmation of this hypothesis waits on studies, like that of Clark
and Young, in which scope of negation can be varied without concomitant changes
in verifying evidence, practice, and other confounding effects.

2.46 On the putative 'prohibitive connotation' of negatives

Other proposals have been made as to why one negative is more difficult than
another. Eifermann (1961) and Wason and Jones (1963), for example, proposed
that certain types of negatives differ in what they called a 'prohibitive connotation'.
The first proposal of this sort appears to have been made by Eifermann. She
claimed in her study that the Hebrew negative lo was more difficult than the
Hebrew negative eyno. And since lo is used in all negative contl)xts, including the
prohibitive one, whereas eyno is not, she suggested that the greater difficulty of lo
over eyno might be related. to lo's prohibitive connotation which somehow makes
it more difficult to process. Wason and Jones (1963), referring to Eifermann,_ then
tentatively suggested that the difference between the two kinds of negatives in their
study might also be attributed to 'an inhibition of response specifically associated
with the prohibitive connotations of the word "not'".
Unfortunately, a closer look at Eifermann's data reveals that Eifermann's con-
clusion of a difference between lo and eyno was quite unwarranted. In her study,
one group of subjects received lo and the other eyno, in True Negative and False
Negative sentences. But both groups received the identical True Affirmative and
False Affirmative sentences. Eifermann's data show that, for some unknown reason,
the group that received lo verified both Affirmative and Negative sentences an
average of 325 msec slower than the group that received eyno. The important point
here is that the former group was not ouly slower on the lo sentences, but also on
the affirmative sentences that were identical for both groups. In spite of this fact,
Eifermann concluded that lo was more difficult than eyno. The appropriate way
to compare lo and eyno is to use Negation Time, as we have done in the above
discussion. The Negation Time is 736 msec for lo and 600 msec.for eyno, and
from the statistics that Eifermann reports, the difference between the two is not
statistically significant. What reinforces the conclusion that lo and eyno are not
really different is·that on another task in which subjects were Tequired to construct
true or false affirmative or negative sentences there were no differences between
lo and eyno. One can ouly conclude that there is no empirical support in Biter-
mann's data for her proposal that 'prohibitive connotations' of certain types of
negatives make them more difficult to process.
Given the present framework, one .could also question Wason and Jones' con-
clusion that some sort of response inhibition could have. served. as an explanation
 SEMANTICS AND COMPREHENSION 1331
in their study. Their study went as follows: They gave their subjects either explicitly
negative sentences, like 7 is not an even number, or logically negative sentences, like
7 med an even number. The positive counterparts were 7 is an even number and
7 dax an even number, respectively. The subjects were to 'tick' true sentences and
'cross' false sentences, but for med and dax, the subjects had to figure out the cor-
rect system for themselves. Wason and Jones found that Negation Time for the
explicitly negative sentences was greater than Negation Time for the logically
defined negative sentences, which they took as evidence for the prohibitive con-
notations of explicit over logical negatives. But Wason and Jones also questioned
their subjects and found that those subjects given the· explicit sentence used either
the 'true' or 'conversion' method of negation in verifying the sentence. In contrast,
only a third of the subjects given the logical negative did so. Instead, most subjects
in the latter condition rigged up some kind of decision procedure, e.g. 'if an even
number and Dax, or an odd number and Med, then press "tick"; otherwise press
"cross"'. This, of course, is quite a different verification procedure from either
the 'true' or 'conversion' model of negation, never'fheless one that could be
modeled by a process of mental operations. There is no reason to believe that
the Negation Time should be slower or faster for the various logical methods. One
can only conclude that, yes, there are several possible ways of verifying negative
sentences, and the med and dax can lead subjects to use methods different from
the 'true' or 'conversion' methods of negation. But none of these methods leads to
the conclusion that the 'prohibitive connotations of the word "not"' are involved
with explicit negatives.
The inevitable conclusion appears to be that there is no direct evidence at ·all for
the hypothesis that the difficulties of some negatives are the result of a 'prohibitive
connotation' of those negatives.

2.47 On the 'contexts of plausible denial'

In section 2.3, I argued that sentence negation as in John isn't at home is best
thought of as consisting of a supposition - John is at home - plus a denial of that
supposition. Furthermore, I pointed out that the supposition is normally something
that is plausible to the speaker and listener; for example, John isn't at home would
be said if it was expected, believed, or supposed in that context that John was at
home. In other words, the supposition has to be plausible in context. By some-
what similar reasouing, Wason (1965) has argued in a very interesting study that
negatives can normally be used to describe only the exceptional item in a· set .of
otherwise homogeneous objects. From this standpoint, Wason proposed two hypo-
theses. The Exceptionality hypothesis can be illustrated as follows: In au athletic
team of ten people, only one of which (player number ten) is a: girl, it is appro-
priate to say Player number ten isn't male, but inappropriate to say Player num-
ber one isn't female. By this hypothesis, it is appropriate to select the exception for
negative comment. The Ratio hypothesis, ·which is perceptibly different· from the
1332 HERBERT H. CLARK

Exceptionality hypothesis, goes as follows: For the same athletic team, it is appro-
priate to say Exactly one player isn't male, but inappropriate to say Exactly nine
players aren't female. Although the present treatment of supposition and Wason's
treatment of exceptions are similar in spirit, they differ at a crucial point. Where
Wason would argue for both the Exceptionality and Ratio hypotheses, the sup-
positional treatment would predict that the Exceptionality hypothesis is valid, but
the Ratio hypothesis is not.
Let us examine what the suppositional treatment would predict for the Excep-
tionality and Ratio hypotheses. First, consider the two sentences under the Excep-
tionality hypothesis, their semantic representations, and their paraphrases:
25 a) Player number ten isn't niale.
b) (false (suppose (player number ten is male)))
c)· It is false to suppose that player number ten is male.
26 a) Player number one isn't female.
b) (false (suppose (player number one is female)))
c) It is false to suppose that player number one kfemale.
Thus, underlying 25 is the supposition that the particular player being referred
to is male, while underlying 26 is the supposition that the particular player being
referred to is female. Since most of the players are male, the supposition of 25 is
plausible a priori, but that of 26 is not. So by the suppositional analysis, 25 ought
to be more in line with a priori expectations - i.e. congruent with expectations -
than 26, hence easier than 26. Note that 26c - It is false to suppose that player
number one is female - evokes a sort of surprise reaction, 'But why do you think
I would suppose that?'
In contrast, the two sentences under the Ratio hypothesis, their semantic repre- ·
sentations, and their paraphrases are shown in 27 and 28:
27 a) Exactly one player isn't male.
b) (false (suppose (exactly one player is male)))
c) It is false to suppose that exactly one player is male.
28 a) Exactly nine players aren't female.
b) (false (suppose (exactly nine players are female)))
c) It is false to suppose that exactly nine players are female.
Thus, underlying 27 is the supposition that exactly one player is male, and under~
lying 28 is the supposition that exactly nine· players are female. But note. that both·
of thes.e suppositions are implausible in context. If most of the players are men;
the supposition in 27 that exactly one playeris male is. contradictory to that eJ<pec"
tation; exactly one has the effecLof saying no more or no less than one player .is
male, which·is not at aH congruent with the apriori expectation. As Just and
Carp.enter pointed out, 27 is in a sense a double negative, with an explicit negative
in isn't and an. implicit negative in exactly one (like their 2 out of 16). That is, 27
has the sense of the following analysis: (false (suppose (players (false (players are
many)):are male))), paraphrased as It is incorrect to suppose that there aren't many
 SEMANTICS AND COMPREHENSION 1333
players who are male; The supposition of 28 - that exactly nine players are
female- is just as implausible io this context. Thus, the suppositional analysis would
claim that the Ratio hypothesis is incorrect: both 27 and 28 should be difficult,
since neither of their suppositions is plausible withio the context of this particular
athletic team.
In agreement with the present suppositional analysis, Wason's experimental
results supported the Exceptionality hypothesis, but not the Ratio hypothesis. He
found that sentences like 25 were relatively easier than ones like 26, but that sen-
tences like 27 and 28 were equally difficult. In his experiment, the subjects viewed
and described a set of eight circles, seven of which were, say, blue and one of
which was red; immediately afterwards, they attempted to complete an iocomplete
sentence ('Circle number 7 is not .•.') with the correct color as quickly as possible
while they were timed. But I should add one note of caution about Wason's results.
He had one group of subjects always complete the Exceptionality-type sentences,
and another group always complete the Ratio-type sentences. Unfortunately, io
the period just before they were to complete the sentences, the Exceptionality group
tended to describe the set of circles with the exceptional item first ('Circle number 4
is blue and the rest are red'), whereas the Ratio group did just the opposite ('Seven
circles are red and one is blue'). This gives us strong reason to suspect that iden-
tical pictures were encoded io two different ways by the two different groups.
We know from previous experiments (cf. Clark and Chase 1972; Clark and
Young, io preparation) that the form the coding of a picture takes io memory can
have dramatic effects on the subsequent verification of that sentence. Although it
is difficult to guess just how the encoding should have affected Wason's experiment,
one should nevertheless be cautious io accepting his results as final.
Wason's results nonetheless appear to support the analysis of negative sentences
into a supposition and a negation. To explain Wason's results simply by appealing
to general properties of sets and their exceptions appears not to work. Instead, the
correct and niore general rule is: a negative sentence will be easy if the supposition
of that negative is plausible io that context. The more general rule, unlike Wason's
rules, allows us to bring together io one unified theory a number of results that
would otherwise not be related to each other. As an alternative to Wason's correct,
but fairly restricted Exceptionality hypothesis, the present suppositional treatment
can be extended to other types of negatives, particularly implicit negatives like
except, excluded, and without, which contain suppositions quite different from their
explicit negative counterparts. The suppositional analysis, of.course, would predict
quite different results for the explicit negatives that Wason used and for the
implicit negatives.

2.5 Negation in Cognition

As a final piece of evidence for the 'true' model of negation, I will now consider
1334 HERBERT H. CLARK

several examples from perception that appear to require the present model of nega-
tion. Although perception has traditionally been kept quite distinct from psycho-
linguistics, it is obvious that there must be a representational system common to per-
ception and language; otherwise, it would be impossible for people to talk about
pictures, to verify sentences against pictures, to imagine a scene as described by a
sentence, and so on. Indeed, as I argued in section 2.2, it is necessary to assume
that pictures are represented, at some level of processing, in semantic representa-
tions that are just like those of sentences. For example, a picture of an A above
a B is normally represented as (A above B), which i~ identical to the underlying
representation of the sentence A is above B.
The main point of perceptual representations is that they are normally positive
-that is, we code perceptual objects as positive entities. 'That is the Eiffel Tower',
we think, not 'That isn't Hoover Tower', or 'That isn't Coit Tower', or 'That isn't
my landlady'. The reason is clear. In most cases, there is only one possible iden-
tification of what a thing is, but many of what it is not. Yet there are instances
where it is important, and eveii: necessary, to think of something in negative terms.
Significantly, the instances of negation in perception that I have found follow the
same logic as the linguistic examples of negation discussed above; that is, negation
in perception fits into the 'true' model of negation, and it can be analyzed in terms
of the supposition the perceiver has of the situation he is observing. I will first
examine evidence for the normality of positive coding, and then discuss a study
in which subjects found it necessary to use negative perceptual representations.

2.51 Positive perceptual coding

In the Clark and Chase study, to be discussed more fully in section 3, it was
found that a picture of an A above a B can be viewed in at least two different ways.
The subject can Iiiok at the top figure (A) in the picture, and if he does he will
encode the picture as (A above B). Or he can look at the bottom figure (B), and
in this case he' will encode the picture as (B below A). When asked simply to view
the picture as a whole, the code will always be (A above B). In other words, an
A above a B is normally coded in a positive form - in terms of the implicitly
positive preposition above, although when the subject views the picture in a specific
way, he can code it in terms of the negative preposition below instead. Here is one
piece of evidence that pictures are encoded in positive representations, except under
unusual conditions.
A second finding of the same study is that subjects could not have been coding
the pictures as explicit negatives. A picture of an A above a B, for example, could
have been coded as (false (B below A)), which is a representation that is also true of
the picture. But the evidence showed that subjects did not use such negative codes.
Note that the use of this type of code would generally result in incorrect judgments
of truth value, for the above coding is consistent with sentences like A is beside B,
even though the picture is incorrectly described by such sentences. Negative repre-
 SEMANTICS AND COMPREHENSION 1335
sentations generally lose information that is contained in the normal positive code,
unless that negative code specifies so much that it is unwieldy.
A third piece of evidence for the normality of positive codes is found in the
present-absent study (Clark and Young, in preparation) reported above. There it
was assumed that the pictures were encoded in a positive form; for example, a plus
was assumed to be coded as (plus is present), not as (false (star is present)). The
alternative assumption, that a picture is coded negatively, leads to incorrect predic-
tions of the verification times in that study. So here are three simple examples of
how pictures are normally encoded in their positive forms.

2.52 Negative perceptual coding

Although subjects normally use positive codes, it is easy to contrive cases in
which they will be forced to use negative codes instead. Consider a picture of a
square with or without a smaller circle inside it. The square with a circle is meant
to represent the presence of a hole, while the square without a circle is meant to
represent its absence~ I have recently used these two pictures in a verification
experiment in which subjects were given sentences like The hole is absent together
with a picture of a square with or without a circle inside it and were asked to judge
whether the sentence was true o~ false of the picture. In this experiment, the
sentences describing the presence or absence of the hole were mixed in with sen-
tences describing the presence or absence of a star or plus; the pictures for the latter
were either a plus or a star.
The rationale of the study was that the pictorial presence of the circle would be
represented as (hole is present), while the pictorial absence could only be repre-
sented in a negative way, as (false (hole is present)). Thus the hole-sentences will
be verified against either a positive or a negative representation of a picture; in
contrast, the star-sentences (or plus-sentences) will always be verified against posi-
tive representations. The representations for the sentences describing the hole and
for the corresponding pictures are shown at the top of Table IX. They are to be
compared with the representations for the sentences describing the star or plus at
the bottom ofthe table. When the four-operation comparison process of Table II
- the 'true' method of negation - is applied to these. representations, the process
still produces the correct truth value in each instance, but the latency components
for the hofecsentences tum out to be different from those for the star-sentences.
In particular, c never appears in the ho/eCsentences, and b and d are not correlated
in the ,hole-sentences as they are in the star-sentences. In short, the 'true' model of
negation predicts quite different patterns of latency for the two kinds of sentences.
These predictions are nicely confirmed by the results of the experiment. For the
sentences describing the star and plus, the results were as usual: True Positives and
False Negatives were faster, respectively, than False Positives and True Negatives.
For the hole-sentences, however, True Positives and True Negatives were both
faster than their false counterparts: in this case there was no interaction between
1336 HERBERT H. CLARK

TABLE IX. Hypothetical representation of sentences and pictures under

two conditions
Sentence Type Sentence Code Picture Code Latency Components

True Affirmative (hole present) (hole present) to

False Affirmative (star present) . (false (hole present))· to +d
True Negative (false (hole present)) (false (hole present)) to +b
False Negative (false (hole present)) (hole present) to +b+d

True Affirmative (star present) (star present) to

False Affirmative (star present) (plus present) to+ c
True Negative (false (star present)) (plus present) to+c+b+d
False Negative (false (star present)) (star present) to +b+d

true-false aod positive-negative. This is just what the 'true' model would predict.
The experiment as I have just described it lacks ao importaot control. The hole
is absent was always confirmed by the ·physical or perceptual absence of a figure
in the picture (viz. by the absence of the circle within the square), whereas The star
is absent was always confirmed by the physical presence of a figure (the plus). One
might argue that the difference in latency patterns between the hole- and star-
sentences is attributable to the perceptual properties of the picture, not to the
semaotic representation of the picture itself, as I have been arguing. As a defense
against this kind of counterargument, a ·second carefully counterbalanced con-
dition was included in which hole was replaced· by lid, so that The lid is present
was true when the circle was absent from the square, aod The lid is absent was true
when the circle was present. In agreement with the semaotic representations of the
pictures, the lid-sentences had ·exactly the same pattern of verification latencies as
the hOle-sentences. The· conclusion is that it is the semaotic interpretation of the
picture that is critical in this verification task, not simplf the physicalc attributes of
the picturec
In suuunary, the study •with.the. hole~ aod ltd-sentences constitutes one example
of how a picture must·be interpreted:negatively aod:shows. that when it is, the ~true'
model of.negationis "till'·correct in predicting relative :verification latencies, rThe
negative :iilterpretations:in this study, furthermore, depended: crucially·on what' the
subject supposed the picture was representing.' For example, the pictrtte of a square
with a circle inside it was interpreted a8 the presence .. of a hole whert. the subject
supposed that the.picture was·meaot•torepresentholes, but as .the absence of a lid
when he·supposed that thepicture·was:tueaot to representlids,··In short, ortrpefc
ceptual system mitkes use of negation too; aod this negation is fundamentally. bf the
same sort that is found in laoguage: it cbnsists· of a proposition (e.g. the hole is
 SEMANTICS AND COMPREHENSION 1337
present) embedded within another predicate (it is false) that denies the truth of the
embedded proposition.

2.6 Summary

In this section, I have proposed a very general model for the comprehension of
negatives. Basic to this model is the assumption that sentences like Helen isn't at
home consist of two underlying propositions: the positive proposition Helen is at
home, aud the proposition it is false, into which the first proposi1;ion is embedded.
In verification tasks, for example, it was~proposed that Helen isn't at home is repre-
sented as (false (Helen at home)) at Stage 1, and its verifying evidence, say, that
Helen is at school is represented at Stage 2 as (Helen at school). The Stage 3 com-
parison process of the model consists of a. series of operations that compare the
two representations. First, (Helen at home) is compared with (Helen at school), the
two are found to mismatch, so a truth index is changed from true to false, Second,
the embedding string (false 0) is compared with the lack of one (0), the two are
found to mismatch, so the truth index is changed back from false to true. At Stage
4, a response is performed in agreement with the final value. of the truth index
true. This model was called the 'true' model of negation, and it was contrasted
with several 'conversion' models in which negative sentences are 'converted' into
positive ones before they are encoded at Stage 1. In a review of the psychological
literature on comprehending negatives, it was then shown that the 'true' and 'con-
version' models accounted for the main findings on comprehension time and com-
prehension errors in this literature.
Another important observation made about negatives in this section was that
sentences like Helen isn't present, which contain explicit negatives, contain positive
suppositions like Helen is present which are denied; in contrast, sentences like
Helen is absent, which contain implicit negatives, contain negative suppositions
like Helen isn't present which are affirmedc This suppositional analysis led to the
expectation that explicit and iuiplicit negatives would behave differently, and this
expectation was verified in a number of studies on negation. This literature, in
fact, suggested a hypothesis about scope of negation:. the larger the scope of a
negative within a sentence, the larger the Negation Time for that negative should
be. A review of the negation literature was consistent with this hypothesis, although
the evidence was not as complete as it should be. But the more general lesson was
that it is imperative to take suppositions into account in ci>nstructing a model of
negation, for suppositions were shown to be implicated in a number of studies on
negation;

3. LOCATIVES

Another basic construction all languages have in one form or another is the locative,
1338 HERBERT H. CLARK

whose function is to describe the spatial positions of objects. In English, locatives

normally consist of a nominal referring to the object whose position is being de-
scribed plus a prepositional phrase, as in The unicorn is in the garden. The position
of an object (the unicorn) is normally described only by specifying that position in
relation to another position (that of the garden). This section, then, will be con-
cerned mostly with the comprehension of sentences that contain prepositional
phrases.

3.1 Semantic Representations of Locatives

The semantic representations of locatives must first take account of the linguistic
asymmetries between A and B in sentences like (1) and (2):
(1) A is above B.
(2) B is below A.
Sentences (1) and (2) dearly refer to the same state of affairs, namely an A above
a B, but even though they have the same denotation, they are not in fact synony-
mous. In (1), the position of A is being described with respect to the position of B,
while in (2), the reverse is true. Why would one want to describe A with respect
to B, say, instead of the reverse? One significant reason is that the speaker pre-
sumes that the listener already knows the position of B but not the position of A;
therefore, the speaker will describe the position presumed to be unknown in
relation to the one presumed to be known. Linguistically, this can be put another
way: sentence (1) presupposes that the position of B is known. That this is true is
demonstrated by the fact that Where is A? is answered acceptably by (1), but not
(2), and Where is B? by (2), but not (1). The question Where is A? indicates that
the position of A ls unknown, and therefore, B is below A is unacceptable as an
answer, since it presupposes that the position of A is known, thereby violating the
implication of the question. Psychologically, the B of A is above B is a 'point of
reference', and the sentence merely asserts that A has a certain relation to that
point of reference.
Sentences (1) and (2) therefore cannot both be represented in the same form, say
as above (A, B). Rather, the proper representation must contain the information
that one position is being described with respect to the other. A fairly complete
notation might represent A is above B as ((position A) (above (suppose (known
(position B))))), glossed as 'the position of A is above the supposed known position
of B'. But this notation is too cumbersome to retain throughout this section, so I
will use simply (A above B) as a shorthand for this more complete notation.
Even this representation is incomplete, however, for it does not fully indicate
the relation between above and below. These .two prepositions are not just two
arbitrarily different prepositions in English but are converses: A is above B implies
and is implied by B is below A. This fact might be represented for the present time
 SEMANTICS AND COMPREHENSION 1339
by a featural notation, with above as [+Vertical [+Polar]], and below as [ + Ver-
tical [-Polar]]. The +Vertical feature represents all those verticality features
that above and below have in connnon, and the +Polar, a feature that specifies
whether the point of. reference is to be the lower or the highet of the two objects.
However, the analysis does not end here, since the features +Polar and-Polar
have a far broader nse than just the words above and below. Consider the four
adjectives English uses .exclusively for the description of verticality: high, low, tall
and short. It is well known that high and low, and tall and short, are asynnnetrical,
with high and tall the unmarked or semantically positive members of the pair (Bier-
wisch 1967). An important property of these four words, however, is that they all
refer to measurement from a reference point upward. High means 'of much height'
or 'of much distance upward', whereas low means 'of little height' or 'of little
distance upward'; although the point of reference for height is normally the ground,
it is always below what is being measured. The analogous statements can be made
for tall and short. Although English also contains the pair deep-shallow, which can
be used for the description of measurement downward, as in The ocean is deep, depth
is not exclusively a vertical measurement, as can be seen in The cupboard is deep
and Cushman went deep into the forest (cf. Bierwisch 1967). More generally, depth
means 'distance into an enclosed space from its surface' and is not related directly
to verticality. In short, English presupposes that vertical measurement is made in
an upward direction from a point of reference.
The property of upward measurement, of course, is a property of A is above B,
but not of B is below A. This suggests that A is above B is the normal mode of
description of two objects, and B is below A is the marked, or negative, case and
is used only when there is some special reason to choose the upper object as the
reference point. ·The notation that indicates that above contains +Polar and below
-Polar represents this fact only if it is remembered that a plus is the unmarked, or
. positive, value on the polarity feature.
Psychologically, the unmarked or normal character of above vis-a-vis below leads
to the following proposal. The representation for above is assumed to be set up in
essentially one step, for the feature +Vertical always, redundantly, briogs with it
the feature +Polar. This redundancy rule wonld explain the normality of above
over below and the fact that high, low, tall and short all presume measurement
upward. On the other hand, the representation for below is assumed to be set up
in two steps, first by setting up [+Vertical [+Polar]] and second by changing the
sign assigned to Polar by the redundancy rule from + to - . A model such as this
implies that below shonld take longer to represent initially than above, and evidence
will be presented later that is consistent with this proposal.

3.11 Properties of English prepositions

A more serious representation problem arises when one tries to represent all of
the English prepositions in a feature notation. At present, thi~ is clearly impossible
1340 HERBERT H. CLARK

to do, and ultimately, it might well turn oui to be the wrong approach. Never-
theless, it is instructive to consider certain generalizations that can be made about
some English spatial terms, particularly since some of the generalizations are im-
portant for the psychological experiments that will be discussed. I will first outline
a proposal that relates the spatial model underlying prepositions and adjectives in
English to certain human perceptual processes. To describe and give evidence for
this proposal in detail would go far beyond the scope of the present chapter, so a
simple outline of the argument will have to suffice. I will then point out that cer-
tain other prepositions are found in positive-negative·pairs.
3.111 Quantifier negation .. English contains considerable evidence (cf. also
Bierwisch. 1967; Teller 1969) for the proposal that the semantic properties under-
lying English spatial terms are derived ultimately from the way we perceive the
world around us. The main center of the perceptual world is the ego: positions are
perceived in relation to the ego, as far or near, in front or in back, up or down, left
or right, and so on. Normally, the positive visual field (~e part that is normally
visible)consists of the visible space that is in front of the ego and above the ground;
the boundaries for this field are the ego for front-back and the ground for up-down,
and the field is .symmetrical about left-right. This suggests that perceptually the
visual field consists of three fundamental coordinates- up-down, front-back, and
left-right ~ and that upwardness from the ground is positive, forwardness from
the ego is positive, and since left and right are approximately symmetrical, both
directions are positive.
These perceptual facts have their semantic 'COnsequences. As pointed out above,
verticality in English is normally expressed. as measurement upward: height means
'distance upward from the ground'. This was also taken to mean that above de-
scribes the normal case and below the semantically marked case. These semantic
facts, .by hypothesis, are more or less direct consequences of our perceptual organi-
zation of the world, since upward is positive in the visible field. Distance in English
is also normally expressed ·as measurement away from the ego. The sentence How
faris San Francisco? in the unspecified case would be taken to mean 'How far is it
to San Francisco from here?' The ego's being the reference point of the perceptual
space makes ·this.fact explicable· too. There is also some .indication in English that
in front:of_is unmarked or positive with respect to in back of (in the same sense that
above is unmarked:>with respect to below), since words like backwards and behind
have negative connotations ~ e.g., Little Eddie is backward and Mary is some-
what behind in her work- whereas their opposites do not. This again is explicable
from perception since. forwardness, .like upwardness, is positive in the perceptual
field.
One can think of these properties in terms of negation. Objects in the positive
'perceptual field' are expressed in positive terms, whereas those in the negative
field are expressed as implicitnegatives of the positive term. In the terminology
of section 2, above-below, up-down, in front of-in back of, ahead-behind, and so
 SEMANTICS AND COMPREHENSION 1341
on are positive-negative pairs in which each of the negatives is an implicit quan-
tifier negative. Thus the sentence A is below B is a 'weak' negative: it is not a
denial, but rather an affirmation of the presence of a negative value on the above-
below dimension.
In short, the hypothesis is that the semantic properties of the spatial words in
English- e.g. high, low, tall, short, deep, shallow, far, near, above, below, front,
back, ahead, and behind - are ultimately the result of how people perceive the
space immediately around them. Although I have hinted only at the relation be-
tween positivity in the perceptual field and positive-negative distinctions in seman-
tics, it can be shown that perception and semantics are very tightly interrelated in
other ways too.
3.112 Full negation. As Gruber (1965) has pointed out, there are also three
pairs of prepositions in English that contain implicit full negatives. Unlike above-
below, ahead-behind, and in front of-in back of, which contain implicit quantifier
negatives, to-from, in-out, and on-off contain full negatives. John is out of the
house, for example, is referentially identical with John isn't in the house, showing
that the relation between in and out is much like the relation between present and
absent, as discussed in section2. The positive and negative force of in and out can
even be seen in such constructions as Jane talked George inio staying and Jane
talked George out of staying. The first sentence implies, positively, that George
stayed, whereas the second implies, negatively, that George did not stay. To and from
are the basic directional prepositions in English, and their positive and negative force
can be seen in John went to the house and John went from the house. The first
implies, positively, that John is at the house, while the second implies, negatively,
that John is not now at the house. On and oft are similar to in and out. These
three pairs of prepositions,-- then, would be expected to behave very much like
present-absent, since absent is an implicit full negative too, but somewhat differ-
ently from above-below, since below is an implicit quantifier negative.

3.12 Time as a spatial metaphor

In English, as in many otherlanguages, time is normally expressed by locatives
used metaphorically. It has often been noted, for example, that the spatial and
temporal prepositions are almost completely co-extensive - e.g. at home and at
noon, in a house and in a moment, between now and then and between here and
there, before me and before noon, behind the house and behind the appointed time,
and so on. The questions When is the party? and Where is the party? have much the
same form and ·are answered in much the same way, e.g. at noon and at home. It
can be shown that the temporal prepositions and their interlocking meanings are
derivable from the spatiaL prepositions and their semantic field, so that what holds
for spatial prepositions should also hold for temporal prepositions. The full expli-
cation of this proposal, however, cannot be accomplished in the short space of the
present chapter, so this bare outline must suffice for now. The. only reason for
1342 HERBERT H. CLARK

bringing up time as a spatial metaphor here is to account for before and after as
temporal subordinating conjunctions and to predict how they should behave in
question answering tasks.
The two ·constructions of interest are 29a and b:
29 a) He sold his Cord before he bought a Citroen.
b) He bought a Citroen after he sold his Cord.
As Kuroda (1968), ·McKay (1968) and E. Clark (1969) have argued convincingly,
the conjunctions before and after in 29 are derived linguistically from prepositional
phrases in the. underlying structure. For example, 29a and 29b are derived from
something like 30a and 30b, respectively:
30 a) He sold his Cord before the time at which he bought a Citroen.
b) He bought a Citroen after the time at which he sold his Cord.
The sentences in 30 simply contain the prepositional phrases before a time and
after a time with relative clauses attached to each time. Therefore, the subordinate ·
conjunctions before and after reduce the prepositions and their subordinate clauses
to ordinary relative clauses. Since the conjunctions before and after, then, are
derived from temporal locatives, which in tum are 'derived' semantically from ·
spatial locatives, it is quite appropriate to symbolize 29a and 29b, respectively, as
the locatives in 31a and 31b:
31 a) (S1 before S,)
b) (S, after S1)
These representations are meant to be formally identical to the previous locative
notations, with S1 representing the first event in time and S2 the second event.
The representations of 31 are better justified than they may first appear. First
of all, if the sentences in 29 really are locatives as represented in 31, then they
should have the same asymmetries of presupposed known and unknown positions
as true locatives. And this is so. Note that He sold his Cord.before he bought a
Citroen really asserts something like 'The .time at which he sold his Cord was be-
fore the time at which he bought a Citroen', and presupposes both the sale of the
Cord and the purchase of the Gtroen. But more important, it presupposes that the
listener knows when the. purchase of the Citroen occurred. The time of the Citroen
purchase is the point of reference for locating the time of the Cord's sale, so this
is exactly analogous to 'space' locatives; Second, and closely related to the· first
point, the answer to when questions behave just like the answers to where ques-
tions. e.g. When did John buy the Citroen? is answered appropriately by He did it
after he sold his Cord, or more simply by After he sold his Cord or After the sale
of the Cord or After four; it would normally. be inappropriate to answer He sold
his Cord before he did it.
The temporal conjunctions before and after should therefore behave just like the
locative prepositions above and below. The fact that temporal before and after are
semantically derived from spatial before and after, analogous to in front of and in
back of, respectively, suggests that before is implicitly positive, and after implicitly
negative, and that before should therefore be encoded more quickly than after.
 SEMANTICS AND COMPREHENSION 1343
3.13 Summary
The semantic representation of a locative, then, is to be denoted simply as, e.g.
(A above B). This notation, however, is shorthand for a number of facts that must
also be represented in a more complete version of the notation. First, the position
of B is presupposed to be known, and the position of A is being described with
respect to B's position. Second, above and below are converses such that they
differ only on the single feature +Polar. Third, the representation for the positive
word above is· hypothesized to be formed in a single step, as opposed to that for
the implicit quantifier negative below, which is formed in two. Fourth, prepositions
like from, off, and out are implicit full negatives that should behave much like
absent, another such negative. And fifth, temporal conjunctions like before and
after are assumed to be semantically derived from their spatial counterparts and
therefore predicted to behave like the latter - i.e. as locatives. I now tum to
several studies of locative sentences in English in order to examine how the seman-
tic representations for locatives just proposed fit into the theory of comprehension
given in section 1. The tasks in these studies fall into three categories - sentence
verification tasks, question-answering tasks, and instmction following tasks. I will
discuss the three types of tasks in tum, showing how the theory applies equally to
each of them.

3.2 Sentence Verification

3.21 Above and below

One of the. most basic studies of locative sentences is the one previously discussed
(Clark and Chase 1972) in which subjects were asked to verify sentences like
The star isn't below the plus against a picture of, say, a star above a plus. That study
demonstrated important facts about locatives: (1) the meaning of above is repre-
sented faster at Stage 1 than the meaning of below; (2) the verifying picture can be
represented at Stage 2 in either of two ways; and (3) the picture is normally re-
presented in terms of above, not below. These three facts are just what should be
expected, given the linguistic features that were presumed to be part of the mental
representations of locatives.
3.211 Stage 1 coding. In what is now a large number of experiments
(Seymour 1969; Chase .and Clark 1971, 1972; Clark and Chase 1972, 1974;
Young and Chase 1971), sentences than contained above were always found
to be verified about 80 to 100 msec faster than those that contained
below. This difference appeared over all kinds of variations in experimental
method, e.g. when the picture was looked at before the sentence, when
the picture was looked at' after the sentence, when the top or bottom of
the picture had been partially obscured, when subjects were told to attend selec-
tively either to the top or bottom of the picture, and so on. In Young and Chase
(1971), subjects were required to change isn't above into below and isn't
1344 HERBERT H. CLARK

below into above. In accord with the other stndies, whenever the final coding of
the relation was in terms of above, that coding was verified more quickly than the
other coding. This last result demonstrates that it is not the printed word above
that is easier to perceive, etc., than the printed word below, but rather it is the
coded meaning of above that is easier to form than the coded meaning of below;
thus, neither reading time nor, say, greater familiarity with the printed word above
can explain the above-below difference. In short, these results support the thesis
that below takes longer to encode, to represent in semantic form at Stage 1, than
above.
The only other pair of prepositions that has been stndied is in front of and in
back of (Clark, unpublished data). Significantly, sentences that contained in front
of were verified about 190 msec faster than those that contained in back of, again
confirming that the positive member of the pair is represented faster.
3.212 Stage 2 coding. A critical factor in the process of verifying sentences like
A isn't above B against pictures is how one encodes the pictures. Consider an A
above a B. Whenever people look at pictures like this, they find it difficult not to
attend selectively to either the A or the B in the pictnre instead of to the picture as
a whole, even when the picture is quite small. Whenever they attend to the A, it
could be assumed that they· are implicitly trying to encode the position of the A,
not the B. In effect, they ask themselves, 'Where is the A?' the answer to which,
of course, is A is above B, which describes the position of A with respect to some
other known position in the vicinity of A. These considerations suggest the follow-
ing thesis: people encode pictures in terms of the figure they have attended to.
What should their code be when they are not specifically instructed" on how to
attend to the picture? It was observed" previously that in English the normal point
of reference is at the bottom of the measured dimension (as in tall, short, high, low).
This fact suggests that people normally consider the lower of two points as the
point of reference arid encode the upper point relative to this point.. In other words,
without any other constraints people should normally encode "an A above a B as
Ads above B; or rather as the semantic representation underlying this sentence.
Indeed, the experiments of Dark and Chase (1972) demonstrate these two
proposals about encoding pictures quite nicely. In one"'experiment; subjects were
asked to look at the pictnre before they read the sentence nnder one of the three
folltiWing instructionst·(1) fixate the top figure mthe pictnre; (2) fixate the· bottom
figure; (3) fixate the. picture as a whole. •Under instruction (1), the latencies of
sentence verification were consistent with the picture coding .(A above B), not (B
below A), whereas under instruction (2), the latencies were consistent with the
reverse. And finally, the pattern of latencies nnder instruction (3) was just like that
nnder instruction (1), indicating that under normal atteritional conditions, either
subjects fixate the top and therefore code the pictnre as (A above B), or they
simply look at the pictnre as a whole, taking the lower figure to be the point of
reference. It was also shown that minstances in which•subjects read the sentence
 SEMANTICS AND COMPREHENSION 1345
before they looked at the picture, they invariably looked at the top of the picture
whenever the sentence contained above, but at the bottom whenever it contained
below. In doing this, the subjects were therefore coding the picture in such a way
that the prepositions of the picture's code and the sentence's code were identical;
this meant that they had Jess manipulation to do in the later comparison process in
order to test whether the sentence and picture were in agreement or not.
Preliminary to another study (Clark and Chase 1974), subjects were
simply asked to describe a picture of an asterisk either above or below a plus in.
their own words. Their descriptions were overwhelmingly couched in terms of above,
or over, or the like. In the same experiment, subjects were also given pictures of
an asterisk either above or below a three-quarter inch long line. In this case the
asterisk above the line was described in terms of above, and the asterisk below the
line, in terms of below. The line proved to be a salient point of reference, and, as
such, it overrode the normal considerations of viewing the lower of the two figures
as the p9int of reference. In spite of this asymmetry between lines and stars, how-
ever, it was found in the subsequent experiments that subjects still coded the pic-
tures independently of these asymmetries when they read the sentence before they
looked at the picture; wh~n subjects looked at the picture first, in contrast, the
asymmetries and their associated codings had the appropriate consequences on the
verification process. This is accounted for by the fact that subjects attended to the
picture in the sentence-first condition contingent only on the preposition of the
sentence: look at the top if the preposition is above, but at the bottom if it is below,
regardless of whether the asterisk is above or below the line. In the condition in
which subjects viewed the picture first, many of them, when uninstructed about
where to look, simply always coded the position of the asterisk with respect to
the line.
3.213 The Stage 3 comparison process. We are now in a position to study the
verification process of locatives in detail. The main principle that will emerge here
again' is the principle of congruence: the verification process attempts to manipulate
the underlying repnisentation of the sentence and that of the picture so that the two
are· exactly congruent.
Corulider the sentence A is below B, which is to be verified against a picture of
an A above a B. The following verification process has been confirmed by the
results in · Clark and Chase (1972) whenever the subject reads the sen-
tence· before he looks' at the picture: At Stage 1, the subject codes the sentence as
(A below B); at Stage 2, he looks over at the bottom of the picture, which he
encodes as (B below A); at Stage 3, he merely checks to see if the subject nouns
of the two sentences are identical, and since they are not in this instance, he changes
.i the answer he presupposed to be true to false; and at Stage 4 he answers false .
The manipulation he carries out at Stage 3 tilkes about 150 msec, milking the non-
I negative false sentences 150 msec slower than true sentences.
Whenever the subject looks at the picture first, however, the process must neces-
1346 HERBERT H. CLARK

sarily be more complicated. First, he codes the picture as (A above B) and then he
codes the sentence as (A below B). The stage 3 comparison process consists of
two steps: first he checks the two subject nouns to see if they are identical; when
they are not, he transforms the picture from (A above B) into (B below A); then he
checks the two prepositions to see that they are identical; since in this in-
stance they are not, he changes the answer he presupposed to be true to false.
At Stage 4, he answer<: 'false'. What is important here is that whenever the picture
is presented first, false sentences that contain above are processed with two time
consuming operations, not just one: the subject noun-matching operation takes
approximately 200 msec, and the second true-to-false translation operation takes
150 msec, as it did in the sentence-first condition. The signilicance of this process
is perhaps not in the details of the comparison operations, but rather in the general
principle that underlies these operations - the principle of congruence.
In summary, the studies on above and below confirm several a priori constraints
on how locative sentences should be represented. The positive locatives above and
in front of take less time to encode than their negative counterparts; pictures are
normally encoded in terms of the positive preposition above; and the represen-
tations of sentences and pictures are compared under the constraints of the prin-
ciple of ·congruence.

3.22 To and from

In a recent series of studies (Clark and Offir, in preparation), Carole Offir and I
investigated the pairs to-from and into-out of, which contain the implicit full nega-
tive prepositions from and out of. In one experiment, subjects were required to
verify sentences like It is coming from Town B against schematic pictures of a sym-
bolic car (an arrow).pointing in the direction, e.g. to Town B and away from Town
A. Among other results of this experiment, we found that sentences containing to
were verified about 180 msec faster than. those containing from, This is consistent
with the hypothesis that from is the negative member of this pair of prepositions.
In a second experiment, subjects were given problems like John. thought 'Mary has
just come out of the house'; therefore, John could be in the.ho~Jse ('false') and.were
required to say whether the conclusion was true or false of tl!e premise. Agaijl, sen-
tences containing into were found to be verified about 200 ms.ec faster than those
containing out of. Furthermore, out of. was found to behave according to the 'true'
model of negation, since, in agreement wit!! predictions from the 'true' model, an
out of in the above premise followed J:>y a not in in the conclusion was ,verified
more quickly than out of followed by in. The second experime!lt, then, yields even
more convincing support for the thesis that out otis comprehended as an impljcit
full negative, since the results here are similar to those for. present and absent in
section 2.
 SEMANTICS AND COMPREHENSION 1347

3.3 Answering Questions

The psychological literature has produced only two studies that I know of on an-
swering questions raised about either explicitly or implicitly locative sentences, and
they are by Huttenlocher (1968) and Smith and McMahon (1970). Both of these
studies are important since they verify in quite some detail (1) the differences be-
tween implicitly positive and negative prepositions, {2) the role of the presupposed
point of reference, and {3) the requirement for the principle of congruence.~ Before
these studies are described further, it is necessary to determ!ne exactly what is
meant by congruence in a question-answering task concerning locatives.

3.31 Congruence
The two main questions that could be asked about the locative John is ahead of
Pete are (1) Where is John (or Pete)? and (2) Who is ahead (or behind)? Let us
represent A is ahead of B as in 32:
32 (A ahead of B)
and the questions Where is A? lll1d Who is ahead?, respectively, as 33 and 34:
33 (A at X)
. 34 (X ahead y)
That is, 33 and 34 are both simply locative sentences, but the location is being
sought in 33 and an object (X) at some location is being sought in 34. There is quite
appropriate justification for these two representations. The question Where is A?
is simply another form of the question A is at what place? (Katz and Postal 1964),
which has, according to our notation convention, just the representation that is
found in 33. The question Who is ahead? is a complete locative, except that it
lacks a prepositional object, represented here as y, simply because it is unknown to
the questioner.
To see the effects of congruence, I must first outline the process of answering a
simple question. In a task that will serve as an illustration, the subject is to read
the sentence A is ahead of B and then answer the question Where is A? The sub-
ject would set up (A ahead of B) for the sentence, (A at X) for the question, and
then try to match the two representations. In this case, the match is relatively easy,
for the two representations are congruent except for the unknown but questioned
location. The. .subject merely needs to replace ·at X in the second string by the
ahead of B of the first, since all else is congruent, and the answer is therefore 'A is
ahead of B' or more simply 'ahead of B'. What if the question had been Where is
B?, that is (B at X)? Now the two representations - namely (A ahead of B) and
(B at X) - are not congruent: the question asks about the positimi of B, but B is
presupposed to be in a known position in the first sentence. To answer·this ques-
tion, the subject must first perhaps reformulate the sentence's representation,
changing it from (A ahead of B) into (B behind A), then attempt to match it to the
question's representation again. This time, of course, the (B behind A) and (B at
1348 HERBERT H. CLARK

X) are congruent, except for the wanter answer; so at X is replaced by behind A,

and the subject can answer 'B is behind A' or simply 'behind A'.
The answers to questions in English are typically elliptical, but only in such a way
that the answer consists of the word or phrase that would properly replace the who,
where, why, etc. of the question and make a full sentence. Thus, an answer to Who
left just now? can be simply George, and this means George left just now; the an-
swer to Which table is Jane sitting at? is that one, meaning Jane is sitting at that
table; and so on. Note that for Where is Jane?, one must answer with a full prepo-
sitional phrase, like at that table, not just that table; the same can be said for ques-
tions containing when. The answers to how and why questions must be answered
by phrases more complex than a single noun phrase also, e.g. How did you do it?
by jumping up and down, and Why did you go? because I felt like it. The point is
that because the answer to where questions must be a prepositional phrase (except
for here and there), this indicates that the whole at X, a somewhat clumsy way of
representing where, must be replaced by a prepositional phrase, like ahead ofB, in
the process of answering Where is A?
Now we must look at congruence in questions like Who is ahead? Consider the
problem A is ahead of B; therefore, who is ahead? The sentence and question are
represented respectively, as (A ahead of B) and (X ahead of y); the two strings are
compared, found to be congruent except for replacing X by A, and y implicitly by
B, and the answer is simply 'A'. If the question had been who is behind?, of course,
the two representations would have been (A ahead of B) and (X behind y), which
are not at all congruent. The sentence's representation would perhaps be reformu-
lated as (B behind A), which could then be directly compared with the congruent
(X behind y), and the answer B is directly forthcoming.
The principle of congruence, therefore, is capable of predicting which questions
will be easy to answer and which will not. The four sentence-question sequences
in 35 are congruent, in the sense in which I have used the term above:
35 a) A is ahead of B. Where is A?
b) B is behind A. Wh.ere. is B?
c) A is ahead of B. Who is ahead?
d) B is behind A. Who is behind?
whereas thos.e in 36 are not:
36 .a) A is ahead of B. Where is B?
b) B is behind A. Where is A?
c) A is ahead of B. Who is behind?
d) B is behind A. Who is ahead?
If it is assumed simply that each one of the extra mental operations required to
make the sentences of 36 congruent with their questions takes time to carry out,
then the sentence-question sequences of 35 should take less time to answer than
those of 36. This principle is confirmed in the following studies.
 SEMANTICS AND COMPREHENSION 1349
3.32 Before and after
Smith and McMahon (1970) have recently reported on an important series of
experiments on the comprehension of sentences of the form He danced before he
sang, He sang after he danced, Before he sang, he danced, and After he danced he
sang. This series was apparently initiated in order to account for some previous
work (Clark and Clark 1968) that revealed that subjects make certain systematic
errors in recalling sentences of this form. In each of Smith and McMahon's ex-
periments, subjects were required to read such a sentence and then answer either
the question What happened first? or What happened second? The question was
presented before the sentence in one experiment, directly after the sentence in an-
other, and after a considerable delay in a third. In the latter two experiments, the
subject was given the sentence, allowed to 'understand' it for as long as he needed,
then required to press a button, at which time he was given the question or a delay
followed by the question. This procedure enabled Smith and McMahon to measure
sentence-encoding time separately from question-answering time. The pattern of
latencies observed in these experiments was very complex indeed. One kind of
sentence was consistently faster than another in one experiment, but consistently
slower than the other in a second. Smith and McMahon rightly observed that these
inconsistencies reflect the very flexible cognitive strategies subjects have for come
prehending sentences; whereas one strategy is useful in one task; another is more
useful in a second task. Smith and McMahon therefore proposed a model of com-
prehension expressed as a flow diagram with empty boxes that the subject could
fill in with ,particular operations for particular tasks. Smith and McMahon did not
explain what the possible strategies are or how the empty boxes are filled in for
their particular experiments.
It appears, however, that by concentrating on the apparent inconsistencies in
their results, Smith and McMahon overlooked several very important generali-
zations that could have been made. And these generalizations make good sense when
the task is characterized as follows. The subject's first job in each task is to s'ef up
a semantic representation for 'the sentence and question at hand. To do this, he
must note the particular surface structure of the sentence and use, probably, some
heuristic devices in" deciding what the representation should be. Just as Smith and
McMahon observed, these heuristic devices can vary considerably from experiment
to experiment. But what is probably constant from experiment to experiment is
what the final semantic representation of the sentence and question should look
like, and how long it takes to set• up the representations for before and after:
Specifically, before-sentences must still be represented as (S1 before S2), and after-
sentences as (S• after S1) no matter what the task; and the Stage 3 process of com-
paring; say, (S, before S2), against the underlying representation of What happened
first? should still follow the rules of congruence. Several of these generalizations
will, be exaniined.
First, Smith and McMahon consistently found that sentences containing before
1350 HERBERT H. CLARK

were answered more quickly than those containing after. In the experiments where
they were able to separate the initial inspection time of the sentence from the time
taken to read and answer the question, Smith and McMahon found that before
took less time than after during the initial encoding stage, but approximately the
same amount of time during the Stage 3 comparison process. That is, the before-
after difference can be attributed specifically to the differences in their Stage 1
encoding time. This result, then, is exactly comparable to the conclusion reached
in· Clark and Chase (1972), in which it was found that above took Jess time
than below to encode at Stage 1.
The before-after difference is further supported in 'Smith and McMahon's data
when they also consistently found that What happened first? was answered in less
time than What happened second? The representations for the two questions 37a
and 37b can be represented approximately as in 38a and 38b, respectively:
37 a) What happened first?
b) What happened second?
38 a) (X befcire (anything else happened))
b) (X after (one thing happened))
That is, implicitly 37a and 37b contain before and after, or rather the semantic
representations underlying before and after. Thus, if tbe semantic representation
of before .takes less time to set up than that of after, then 37a should be answered
more quickly overall than 37b. This is exactly what Smith and McMahon found.
Second, the results of Smith and McMahon folly support the principle of con-
gruence. Under the present scheme, Before he sang, he danced and He sang before
he danced would both be represented as. 39a, and both After he sang, he danced
and He sang after he danced would be represented as 39b:
39 a) (S1 before S.)
b) (S,•after s,)
in which S1 is he danced and S. is he sang. The two. questions What happened first?
and What happened second? are represented as 38a and 38b, respectively. The
sentence 39a is congruent with the question 38a, and 39b with 38b, while the other
two pairings are incongruent. According to hypothesis, it shouJd, take less time to
answer in the congruent cases than in the incongruent ones .. This is equivalent to
saying that it should be easier to answer with Si than with s, for the before-sen-
tences, but the reverse for the after-sentences; that is, it should be easier in all sen-
tences .to answer with the event specified· in the main, not the subordinate, clause
of.the sentence. This prediction is confirmed.in all three of Smith and McMahon's
experiments on before. and after.
The remaining results in the Smith and McMahon data are not directly appli-
cable to the underlying representations and their comparisons, but rather seem to
relate to leftcright heuristic strategies for 'scanning' and 'parsing' surface structure.
And it was these remruning results that were inconsistent from experiment to
experiment, For example, it was. easier to comprehend sentences with. the sub-
 SEMANTICS AND COMPREHENSION 1351
ordinate clause before the main clause in some instances, but harder in others; also
it was easier to comprehend sentences with S1 and S2 mentioned in their true chrono-
logical order in some instances, but harder in others. Neither of these properties of
sentences, however, is directly related to the final representations of the sentences .
and their comparisons, so these inconsistencies only serve to emphasize that the
actual construction of semantic representations is quite separate from the com-
parison of two such representations at a later stage. In sum, Smith and McMahon's
results appear to be quite compatible with the process of answering questions that
has been outlined in this chapter.

3.33 Ahead and behind

In another series of experiments reported by Smith and McMahon, subjects were
given sentences like John is leading Bill and John is preceded by Bill and were
asked either Who is ahead? or Who is behind? The sentences were either active
or passive and contained lead, follow, trail, or precede as _the transitive verb.
Although thes'e experiments were even more consistent than the before-after ex-
periments, Smith and McMahon still preferred the empty box model of comprehen-
sion. Here again it appears that stronger conclusions can be drawn once the task is
viewed as a process consistjng of separate encoding and comparison stages. Some
justification is needed, however, for bringing up a study on transitive verbs in a
section on locatives. The reason is that sentences like John is leading Bill contain
implicit locatives. The questions Who is ahead? and Who is behind?, of course, are
explicitly locative, indicating that what the subject must derive from his under-
standing of the sentences is their underlying locative information. The analysis I
will propose for the locative nature of lead, follow, precede, and trail, however,
should be considered only tentative until much more is known about the linguistic
relations between words like lead and genuine locatives.
Conceptually, the sentence John is leading Bill seems to contain two parts:
(1) the sense that John is directing Bill's movement or perhaps simply that John is
doing something in relation to Bill; and (2) the sense of the sentence John is ahead
of Bill. It does not matter here whether sense (2) is thought of as actually part of
the sentence or as simply implied by the meaning of lead: the point is that sense
(2) is Ji.ecessary.for answering the-question Who is ahead? or Who is behind? For
our purposes, then, lead and .precede 'contain' the meaning of come-ahead, and
trail and follow, the meaning of come"behind; Smith and McMahon implied as
much, too, when they referred to the former two as 'come-first verbs' and the latter
two as 'come-second verbs'. Thus A leads or precedes B contains 40, and B fol-
lows or trails A contains 41:
·40 (A comes ahead of B)
41 {B comes behind A)
Furthermore, the sentence A comes ahead of B in tum contains (or implies) the
sentence A is ahead of B, a simple locative; that is 40 implies 42, and 41
implies 43:
1352 HERBERT H. CLARK

42 (A ahead of B)
43 (B behind A)
With the simple locatives of 42 and 43, all the principles established previously
for locatives should hold. The principle of congruence, for example, predicts that
it will be easier to answer Who is ahead? for 42, hence for sentences that contain
lead or precede, whereas Who is behind? will be easier for 43, hence for sentences
that contain trail or follow.
Smith and McMahon's sentences also differed in voice, e.g. John is leading Bill
vs. Bill is led by John. Although linguists differ somewhat in their derivation of
passives, it is generally agreed that the same set of functional relations underlie
both the active and passive. (Section 5, below, is devoted entirely to active-passive
differences.) Thus, both of the above sentences have the same underlying proposi-
tional structure (John leads Bilf), so both contain (John comes ahead of Bilf) and
both are subject to the same rules of congruence. But passive sentences give us a
unique opportunity to compare surface and deep structure with respect to the notion
of point of reference.
Consider the question Who is ahead? and the possible answers in 44:
44 a) A precedes B.
b) B follows A.
c) B is preceded by A.
d) A is followed by B.
As expected from the analysis of precede as come-after and follows as come-before,
the question Who is ahead? is answered acceptably by 44a, but not 44b. ·But the
passives in 44c and 44d pose special problems, for they are both relatively un-
acceptable as answers to Who is ahead? As we will see in section 5, 44c might be
paraphrased as 'As for the position of B, it is in a place relative to A'; that is, the
passive makes the A the point of reference and this is in direct conflict with the
point of reference in the underlying proposition (A ahead B). Thus, Who is ahead?
and 44c are congruent in their underlying propositions - (X ahead y).and (A
ahead B), respectively - but not in their points of reference. A person trying to
answer Who is ahead? from 44c must therefore·make adjustments for the incon-
gruent points of reference and this should take time. It should be noted that 44c
with contrastive stress on B actrtally changes .the point of reference. from A to B
and thereby makes 44c quite acceptable as an answer to Who is ahead? 44d poses
an even more complex problem. Its natnral point of reference, as a passive; is B
and its underlying proposition is (B behind A). The person trying to answer Who
is ahead? must first get at the underlying proposition (B behind A); but that should
be difficult to do, since the point of reference of the passive (B) and of the under-
lying proposition (A) are incongruent. Next, the person must try to match (B
behind A) to (X ahead y) and, just as in 44b, this will require a transformation of
some sort, perhaps changing (B behind· A) to (A ahead B). To sum up; passives
should be more difficult than actives, since to derive the underlying propositions a
 SEMANTICS AND COMPREHENSION 1353
person must change the point of reference of the passive to match the point of
reference of the underlying proposition. Second, 44b and 44d should be more
difficult, respectively, than 44a and 44c, since the former require a transformation
that changes their underlying propositions into something congruent with the under-
lying propositions of the question.
Smith and McMahon's pattern of latencies supports this analysis. The principle
of congruence predicts that Who is ahead? should be answered quickly by the deep
structure subjects (the logical agents) of lead and precede, and Who is behind?, by
the deep structure subjects of trail and follow; the other four ques~ion-answer com-
binations should be slow. To put it another way, subjects should be able to answer
more quickly with the deep structure subject than with the deep structure object
for all four verbs. This is just what Smith and McMahon found. Furthermore,
since passives have points of reference in conflict with the points of reference of
their underlying propositions, they should be slower altogether than actives, and
this is also confirmed in 'Smith and McMahon's dat\}.
Ahead is positive with respect to behind in the s'ame way that above is positive
with respect to below, before to after, and in front of to in back of. It should be ex-
pected from the encoding hypothesis that Who is ahead? should be encoded more
quickly overall than Who is behind?, and this too is consistent with Smith and
McMahon's results.
If lead and precede are both come-ahead verbs and trail and follow both come-
behind verbs, however, then the same principle would predict that lead and precede
should be encoded faster than trail and follow, and this was not confirmed in Smith
and McMahon's data. The failure to confirm would be serious only if lead, precede,
trail, and follow could be completely characterized as come-ahead or come-behind.
Obviously they cannot. As pointed out above, lead, for example, seems to imply
a directive function of the leader in addition to its come-ahead meaning. Trail,
unlike the other verbs, is derived from the noun trail and seems to emphasize 'being
on the track of'. And precedes appears to mean simply 'be in front of'. It is·un-
kno\vn, of course, how the surplus meanings of lead, precede, trial, and follow affect
their encoding times, so there is no reason to expect the first two to be necessarily
faster than the second two. An additional problem is that Smith and McMahon
used precede (like the other three verbs) in the present progressive tense: e.g. John
is preceding Bill. For many people, precede is only a stative verb (cf. Lakoff 1966)
and cannot be used in the progressive tense: John precedes Bill is fine, but *John
is preceding Bill is unacceptable. Of the four verbs Smith and McMahon used,
precede was consistently the slowest, and it could well have been for this reason.
The two words they used that seem most comparable are lead and follow, for they
are converses in at least one of their senses. The two words in fact agree with the
hypothesis: lead was found to be faster overall than follow. T4e encoding time of
the verbs, then, could perhaps be construed as confirming the ahead-behind asym-
metry, although the data are not without serious problems.
1354 HERBERT H. CLARK

One of Huttenlocher's {1968) experiments on deductive reasmring funher sup.

ports this analysis of the transitive verbs leads and trails. In this experiment, sub-
jects were given problems like 'John is leading Panl; Paul is being trailed by Bill;
therefore, who is first?' These problems contained all possible combinations of is
leading, is trailing, is led by, and is trailed by and of the questions Who is first?
and Who is last? Each problem was solved in two steps. The subjects were first
read the first prenrise and were asked a question about it; then they were read the
second prenrise and question and were timed from the end of the question to their
answer. Unfortunately, since the latencies and errgrs of this experiment were
reported only in a very abbreviated form, it is not possible to confirm all the com-
parisons derivable from the present theory.
Only four gross variations in the data are reported. In such problems, the first
prenrise always delineates the ordering of, say, X and Y, as in X is leading Y. The
second prenrise then must describe a third object, call it A, with respect to either
X or Y, and thls conld be active or passive, and it could contain A as either the
deep structure subject or object of the verb. The data are reported in term.s of these
latter four categories, collapsed across the questions Who is first? and Who is last?
It should be ;,oted that the second prenrise must always contain the nriddle term of
the three term ordering - e.g. the term John in an ordering with Bill first, John
second, and Paul third. Thus, for one of the two questions, the answer will be the
A of the second prenrise, but for the otl)er, the answer will be a term out of the first
prenrise. If the answer is to be taken from the second prenrise at all, it will always
be A, and not the other term. By exactly the same reasoning as in the Snrith-
McMahon experiments, the principle of congruence predicts that when the answer
to a locative question- and Who is first? and Who is last? are locative questions-
is the deep structure subject, that answer will be congruent to the question and will
be easy to make; otherwise, the answer will be dlfficnlt. This is exactly the pattern
of Huttenlocher's results. For second prenrises in the active voice, the prenrises
with A as the subject had fewer errors - though no longer reaction times - than
those with A as the object of the verb. For second prenrises in the passive voice,
the same result held for both errors and latencies, with A easier as the deep struc-
ture subject - as the logical agent of the verb. Huttenlocher's results also verify
the prediction that passive sentences will take longer to answer questions of than
active sentences. Unfortunately, the difference between Who is first? and Who is
last? was not reported. The present analysis would predict that Who is first?, be-
cause. it wonld be represented approximately as (X ahead of (everyone else)),
should be encoded more quickly than Who is last?, i.e. (X behind (everyone else)).

3.4 Following Instructions

3.41 On top of and under
Imagine a child sitting in front of a ladder-like structure that contains three
 SEMANTICS AND COMPREHENSION 1355
compartments, one above the other, into which the child could place colored blocks.
And imagine that there was a blue block in the middle compartment. The child is
now handed a red block and is given instruction 4Sa or instruction 4Sb:
45 a) Make it so that the red block is under the blue block.
b) Make it so that the blue block is on top of the red block.
In either case, the child is supposed to put the red block into the empty compart-
men! immediately below the blue block. The theoretically interesting question here
is: Which instruction is easier to follow? Before looking at the results of tbis im-
portant study by Huttenlocher and Strauss (1968), we must examine how the present
theory would answer tbis question, since tbis task clearly demands of the child that
he comprehend a locative sentence in solving the problem.
This, and other instructional tasks of this kind, can be analyzed in two almost
equivalent ways. The first way is to see what constraints are imposed on the task
merely by the way the child presumably represents these two instructions. Ignoring
th\l Make it so that for the moment- and Hutteulocher and Strauss show that one
can - instructions 4Sa and 4Sb can be represented as 46a and 46b, respectively:
46 a) (Red block under blue block)
b) (Blue block on top of red block)
Instruction 46a; like all locatives, presupposes that the position of the blue block is
known, and it indicates the position ofthe red block with respect to the blue block;
exactly the reverse is true in instruction 46b. The task, however, is quite well
defined, with the blue block already fixed, its position known, and the red block
to be placed, its position unknown. Of the two instructions, only 46a is congruent
with the task so defined. To make instruction 46b congruent with the task at hand,
some manipulation must be performed. In short, 46a should be easier when the
fixed referent block is the poip.t of reference - the object of the preposition - in
the instruction.
The second way of viewing tbis task, and the one I will use from now on, is to
imagine th<tt the child is implicitly asking himself the question, 'Where does the red
block go?' which itself reduces to 'Where should the red block be?' Because of its
presuppositional constraints, this question is answered acceptably by 46a, but not
by 46b. From 46a and from the general process of answering questions outlined
previously, the child can form directly, but still implicitly, 'The red block should
be under the blue block' and hence 'The red block should go (or be placed) under
the blue block'. From 46b, the correct self-instruction can be formed ouly in a
more indirect manner.
The first and second ways of viewing tbis task are almost equivalent, and they
stem from the same presuppositional considerations. Both predict .that 46a will be
easier than 46b. The data of Huttenlocher and Strauss fully confirm these predic-
tions. Instructions with the red block (more generally, the moveable block) as the
subject of locatives were carried out more quickly and elicited fewer errors than
the ·other possible instructions.
1356 HERBERT H. CLARK

Unfortunately, Huttenlocher and Strauss' resnlts are of very little help in deter-
mining exactly what subjects do on incongruent instructions like 46b. The resnlts
of Bern (1970), however, are helpfnl. Bern found a group of four-year-old children
who simply could not follow incongruent instructions like 46b at all. These children
carried out the incongruent instruction each time either inconsistently or consis-
tently wrong. When the task is viewed as an implicit question-answering task - the
second of the two suggestions above --' then the children were initially unable to do
anything but try to match the question Where should the red block be? and the
instruction; they could not manipulate the situation further. So Bern tried to teach
them how ·to solve the problem. Her procedure was to present another model
ladder with the correct solution directly alongside the ladder the child was sup-
posed to fix up; Bern wonld then read the instruction, remove the model ladder,
and then urge the child to make his ladder look like the model as he remembered it.
After this training, according to Bern, the child was able to conceptualize the prob-
lem by imagining what the solution was supposed to look like and by then placing
the red block to fit this goal. In the end, all the children were able to respond
correctly most of the time, although the incongruent problems always took longer.
From the present viewpoint, Bern's observations can be cast in a slightly different
form. When the child is·being shown what the final result shonld look like, he is
probably restructuring how he shonld view the red and blue blocks: instead of
thinking of the blue block as the fixed block, he can now think of the red block as
fixed. He can then ask 'Where should the blue block be?' and although the child
is physically moving the red block, he is thinking of where the blue block shonld
be with respect to it. In support of this view, Bern found that when the child was
asked immediately after placing the block what the instruction was, he replied most
often with the instruction he was given, not with some transformation of that in-
struction. This view also coincides with an adult's intuitions when he carries out
such a task. It appears, then, that Betn's resnlts show that children do have to
restructure incongruent problems; they do this not by transforming the instruction,
but by changing the implicit point of reference in the .final display.
The present model of comprehension would also predict that'instructions con-
taining on top of, a positive preposition, shonld be easier than those containing
under; an implicit qiiantifier negative. Although Huttenlocher and Strauss' data
were apparentl{not sensitive enough to detect this difference, Bern's results (per-
sonal commutiication} confirm this'prediction. ·
The present view of the instruction task differs slightly from Huttenlocher and
Strauss' explanation. In the account they preferred, the .child is seen as treating the
subject of a locative instruction as an 'actor'; and since the block the child has in
his hands is also treated as an actor; the instruction is easier when the block in the
hand is described as the subject of tbe locative. But this account gives no reasoii
why the subject of a locative should be considered an 'actor'. It is not agentive,
as the deep structnre subject of a transitive verb is, and it clearly does not have to
 SEMANTICS AND COMPREHENSION 1357
be animate. The attributes Hutteulocher and Strauss apparently want to ascribe to
the subject of the locative are derivable ouly from the notions of (1) presupposed
point of refererrce and (2) congruence. The latter two notions are clearly more
general and more powerful than Hutteulocher's 'actor', for they enable us to view
the difficulties of incongruent instructions as just a special case of the general theory
of comprehending sentences, a theory that accounts for other facts about locatives,
as well as facts about negatives, comparatives, transitive verbs, and other con-
structions.

3.42 Push and pull

In Hutteulocher, Eisenberg, and Strauss (1968), children were asked to follow
instructions in a task identical in concept to the on top of and under study. The
child might hold a red truck in !tis hand and have a blue truck in view in front of
!tim. He .would then be given one of four instructions:
47 a) Make it so that the red truck is pushing the blue truck.
b) Make it so that the blue truck is pushing the red truck.
c) Make it so that the blue truck is being pushed by the red truck.
d) Make it so that the red truck is being pushed by the blue truck.
or the same four instructions with push replaced by pull.
Conceived of as a question-answering task, these push-pull problems fit quite
uicely into the analysis previously given to the Smith-McMahon experiments on
lead, trail, precede, and follow. The instructions of 47 - again excluding the
Make it so that phrase - would therefore be approximately represented, respec-
tively, by 48:
48 a) (red truck cause (blue truck move)) & (red truck behind blue truck)
b) (blue truck cause (red truck move)) & (blue truck behind red truck)
c) (red truck cause (blue truck move)) & (red truck behind blue truck)
d) (blue truck cause (red truck move)) & (blue truck behind red truck)
This notation, wltich views each instruction of 47 as the conjunction of a sentence
with a motion verb and a locative, is ouly meant to be approximate, for push and
pull have much more complex meanings than are represented here (see below). The
main point is that a locative is at least implied.by, if not directly contained in, such
sentences, since Where is A? can be answered acceptably by A is pushing B,
though not by A is moving B, wltich is not locative in such a specific sense. In any
case, instructions 47a and 47c both presuppose that the position.of the blue truck
is known, and so are directly congruent with the question Where should the red truck
be? The other two instructions presuppose the position of the red truck, hence they
are incongruent with this implicit question. The prediction is, then, that 47a and
47c, in which the movable truck is the deep structure agent, should be .easier than
the remaining two. Further, the point of reference considerations that applied to
lead, trail, precede, and follow apply here too, leading to the prediction that· actives
slibuld be easier than passives. The results of Huttenlocher et a!. and Bern (1970)
together corrfirm both of these predictions.
1358 HERBERT H. CLARK

Huttenlocher et al.'s data, however, appear to raise one problem for the present
model. The claim has been made here that push implies the locative behind or in
back of, and pull, ahead of or in front of. Under the assumption that ahead of is
represented more quickly than behind, pull should be represented more quickly
than push. But the results run exactly 'COUnter to this prediction. Other factors,
however, seem to be at work here. The locatives ahead and behind underlying pull
and push are really quite derivative, and seem to depend on our specific knowledge
of trucks; people know empirically that one truck pushes another from behind.
More generally, push means to 'move in a forward direction from the agent' (note
the· unacceptability of *John pushed it toward himself), whereas pull means to
'move in a backward (vis-a-vis the agent) direction towards the agent' (note the
uuacceptability of *John pulled it away from himself, at least without odd contriv-
ances - I am indebted to Patricia Carpenter for these observations). Push and
pull, therefore, contain the meauings move forward and move backward, respec-
tively: the fust contains a positive preposition (before or in front of) and the latter
a negative. The assumption that positive locatives should be comprehended faster
than negative ones, then, predicts that push should be faster than pull, in full agree-
ment with Huttenlocher et al.'s data. Quite speculatively, then, it could be claimed
that the most general meanings of push and pull do confirm the positive-negative
differences in direction prepositions, in spite of the fact that the specific locatives
derived from push and pull appear to indicate the opposite. Obviously much more
linguistic and psychological work is needed either to confirm or to disconfirm this
or any other explanation of the locative nature of push and pull.

3.5 Conclusion

In this section, it has been argued that locatives must be represented with at least
two properties. Consider Helen is in the garden. First, the sentence must contain
one object (Helen) - what the sentence is about- and its location (in the garden),
which is normally expressed in English by a prepositional phrase. The prepositional
phrase (in the garden) specifies a se'COnd· object (the garden) as point of reference
and a relation (in) that describes the position of the fust object (Helen) with respect
to the point of reference (the garden). Second; the mental representation of loca-
tives must indicate· that the position of the second object (the garden) is presupposed
by the speaker m be known to the listener; That is; the garden is of little value in
describing ·the location of Helen if the location of the garden is unknown.
Witli these two· considerations in mind, the principle of congruence enables us to
make several simple but powerful predictions about the comprehension of locatives.
In verifying locative sentences, the point of reference· in the sentence must be
identieal to the point of reference in the information being used to verify it; other-
wise; the point of reference of the verifying information must be changed; thereby
 SEMANTICS AND COMPREHENSION 1359
causing the incongruent case to take longer to verify. The same is true of answer-
ing questions. If the question presupposes the same point of reference as does the
locative sentence being queried, then the answer is easily constructed; if not, one
point of reference must be altered, and the answer takes more time to construct.
Following instructions is just like answering questions. If the objed that is to be
placed in response to a locative instruction is not the point of reference of the in-
struction, then the placement is relatively easy; otherwise, the placement must be
delayed while the perceived point of reference of the task is altered. In sum, con-
gruence is required of points of reference in locatives in order t<:J verify sentences,
answer questions, and follow instructions; incongruence leads to a restructuring of
the problem, a step that takes time and makes the incongruent instances more
difficult.
Finally, it was proposed that prepositions expressing upwardness and forward-
ness are positive with respect to their implicitly negative antonyms; this proposal is
based on the view that certain semantic features are ultimately derived from percep-
tion. As a consequence, the unmarked or positive prepositions should be encoded
more quickly at Stage 1 than their marked or negative counterparts. This proposal
was verified directly for above and below, in front of and in back of, before and
after, ahead of and behind, first and last, and with considerable speculation, also
for the pairs lead and follow, and push and pull. In addition, English has three
pairs of prepositions - to-from, in-out, and on-off - that contain implicit full
negatives, and here it was predicted that the negative prepositions should behave
like other implicit full negatives. This was confirmed for the pairs to and from, and
into and out of.

4. COMPARATIVES

English contains a number of constructions capable of making comparisons, The

principal ones are the strict comparative (hereafter called simply the comparative),
as in John is taller than Pete, and the equative as in John is as tall as Pete. Com-
parisons can also be made with prepositional locatives (e.g. John is above Pete in
intelligence, John is intelligent beside Pete, John has it over Pete in intelligence),
sentential complements (e.g. John is too intelligent to go to college, John is in-
telligent enough to go to college, John is so intelligent that he finished college early),
and comparative verbs (e.g. John exceeds Pete in intelligence, John's intelligence
has increased, John is becoming intelligent, John is getting intelligent). This list is
certainly not exhaustive. In spite of the variety of English comparison construc-
tions, psychologists have worked in detail with only the two principal constructions
- the comparative and the equative. In this section, I will therefore consider only
these two constructions and how they are comprehended. The treatment I will give ·
will not necessarily apply to the other comparison constructions, since there are
 1360 HERBERT H. CLARK

no a priori reasons to think that they are comprehended by exactly the same mental
processes as the two principal constructions.
Psychologists have studied comparatives from three points of view. First, many
have studied the more or less direct comprehension of the comparative, e.g. by
asking subjects to verify or answer questions about a comparative sentence. As we
will see, these studies are all fairly well accounted for in the present theory of com-
prehension with a few properties. Second, the comparative has been used to in-
struct subjects to do something, e.g. place an object with respect to another object.
'This type of task emphasizes a slightly different property of the comparative. And
third, other psychologists have investigated the young child's understanding (and
misunderstanding) of the comparative. The developmental studies, far from being
separate from the studies on adults, actually help to confirm the model of com-
prehension proposed for adults. I will discuss each of these experimental areas in
tum, attempting in each case to bring all the phenomena studied under one theo~
retical umbrella. Before discussing these thn)e areas- of comprehension studies, I
will take up the representation problem, the question of how the comparative
should be represented mentally.

4.1 The Representation Problem

The sentences John is better than Pete and Johlt·is as good as Pete have two very
obvious properties. Both sentences contain two nouns referring to objects that are
being compared, and the two objects in each case are being compared on a semantic
scale represented by the adjective in the sentence. In line with the proposals made
in Oark (1969a, b), I want to suggest that the comparative and equative sentences
are best thought of as the comparison of two sentences, not just of two nouns, and
that it is the two underlying sentences that make the comparisons either easy or
difficult to understand.

4.11 Underlying sentences

Most linguistic analyses of the comparative· demand that it be derived from two
underlying sentences (Lees 1961; Smith 1961; Huddleston 1967; Doherty and
SchwartZ 1967). John is better than Pete is derived, ulthnately, from something
like that which underlies John is so good and Pete is so. good, where John is so good
means something like 'John is good to an unspecified degree'. That is, the com-
parative is derived from ·two sentences ,in which the adjectives have a 'degree
marker' in. front of them. The two underlying sentences are then conjoined by a
series of transformations that insert more than (or as-as) giving, successively, John
is (more. than Pete is good) good, John is more good than Pete is good, .John is
m'!re good than Pete is, John is more good than Pete, John is better than Pete,
A similar derivation holds for John is' as good as Pete. Note that whenever the
 SEMANTICS AND COMPREHENSION 1361
adje·ctives of the two underlying sentences do not have the same sense, the second
adjective cannot be deleted, as in John is taller than Pete is wide. And if the tense
of the two underlying verbs differs, the second verb cannot be deleted either, as in
John is better than Pete used to be. Ross (1969) has pointed out, furthermore, that
the second sentence of the comparative may really be derived from a relative clause
of some sort, as in. John is good to an extent to which Pete isn't, in which the
degree markers, symboli~ed here as to an extent, are shown qplicitly. Although
the details of the linguistic analysis still remain somewhat of a mystery, it is safe
to say that the main outline of the analysis - that the comparative is derived from
two sentences - is well-founded and this is really all that we will need here.
The two sentences that underlie the comparative can be thought of as presupposi-
tions of the comparative. To say.Johwis better than Pete is to presuppose that we
can decide how good John is and how good Pete is, or that John is so good and
Pete is so good, or that John and Pete are somewhere on the goodness scale. John
is better than Pete can be paraphrased as The goodness of John is more than the
goodness of Pete, or How good John is is more than how good Pete is. In these two
paraphrases, the underlying pair of sentences is quite explicit: John's goodness is
derived ultimately from the same structure which underlies John is so good, and
John is how good (how goodlohn is) is he nominal clause corresponding to John
is so good. And like all presuppositions, those of John is better than Pete remain
unchanged under sentence negation (Fillmore 1970): John isn't better than Pete still
supposes that John and Pete are being compared on the goodness scale. This
presuppositional evidence, then, confirms even further the more syotactically
oriented evidence that John is better than Pete is derived from two primitive sen-
tences something like John is so good and Pete is so good.
The ouly linguistic analysis I know of that does not derive all comparatives from
two underlying sentences (Campbell and Wales 1969) is unsatisfactory, I think, on
grounds related to this presuppositional evidence. Campbell and Wales argue that
sentences like John is better thtinPete should be derived from one underlying sen-
te:nce, but sentences like I like John better than Pete from two. Their principal
reason for wanting to derive the two· differently is that 'in [a sentence like The
river is wider than the road] two objects are being. compared whereas in [a sentence
like The river is as deep as it is wide] it is two properties, namely the depth and
width of the river, which are being,compared'. This assertion seems false on the
face of it. The first sentence, <The river is wider than the road, is just as much a
comparison of two properties (namely, the width of the river and the width of the
road) as the second one is. This information is explicit in the paraphrase The
river's width is more than the road's width and is nicely captured in the two-sen-
tence analysis. So one major reason for objecting to the onecsentence analysis is
that its principal motivation is based·on the very dubious premise that simple and
more complex comparatives 'differ in what they compare. But there are still other
reasons for objecting to the Campbell and Wales one-sentence analysis of com-
1362 HERBERT H. CLARK

paratives. For example, according to Campbell and Wales, 'Putting it at its

weakest, given two analyses of comparatives of roughly equal value from a purely
linguistic point of view, we should like to see a willingness on the part of linguists
to accept psychological considerations as being relevant here' (p. 219). As we shall
see, the one-sentence analysis does not allow us to account for a large number of
important psychological results that we will examine, whereas the two-sentence
analysis does. So by Campbell and Wales' own criterion, we should reject their
single-sentence analysis in favor of the more powerful and more general two-sen-
tence analysis.
Now we come to the question of how we should represent the comparative and
equative constructions. A fairly complete representation of John is better than
Pete might be ((extent (John is gootl) (false (extent (Pete is gootf))))), which is to be
paraphrased John is good to an extent to which Pete is not. For our purposes,
however, it will be enough to represent 49a simply as 49b, and 50a as SOb:
49 a) John is better than Pete.
b) ((John is good+) (Pete is good))
50 a) John is as good as Pete.
b) ((John is good=) (Pete is good))
In these representations, the + is meant to indicate a degree or an extent greater
on the good dimension, whereas the = is meant to indicate equal degree or extent.
Sentence negation does odd things in this notation. The negative of 49a is shown
in Sla whose representation is in Slb; but since Sla is equivalent in meaning to
Pete is as good as John, the representation in Slb is really equivalent to the re-
presentation in Slc. The comparable relations hold for 52a, 52b, and 52c:
51 a) John isn't better than Pete.
b) (false ((John is good+) (Pete is good)))
c) ((John is good) (Pete is good=))
52 a) John isn't as good as Pete.
b) (false ((John is good=) (Pete is good)))
c) ((John is good) (Pete is good+))
The final sentence to be represented is John is less good than Pete, which is, shown
in 53a with its representation in 53b:
53 a) John is less good than Pete.
b) ((John is good-) (Pete is good)),
The - marker denotes a negative rather than a positive degree.

4.12 Underlying adjectives

The fnll semantic content of John is better than Pete is not completely specified
in the representation ((John is good+) (Pete is gootl)), for although this notation
shows the synonymy between John is better than Pete and Pete isn't as good as
John, it does not show the near-synonymy of John is better than Pete and Pete is
worse than John. To be able to relate the latter two sentences, English speakers
 SEMANTICS AND COMPREHENSION 1363
must know that the presuppositions of John is better than Pete- that John and
Pete are being measured for relative goodness - can be translated into the presup-
positions of Pete is worse than John - that they are being measured for relative
badness. Although it is tempting at first to say that relative goodness and relative
badness are the same thing, this conclusion can be shown to be incorrect. Onr job
for the moment is to specify how the goodness and badness scales are related and
to write this into the mental representation of the comparative.
Evidence reviewed previously by Clark (1969a, 1970b) shows clearly that good
and bad are asymmetrical. The asymmetry is that good has It neutral 'nominal'
sense besides its more usual 'contrastive' sense, whereas bad has only the 'contras-
tive' sense. There are three main examples that illustrate the non-committal 'nomi-
nal' sense of good. First, note that the name of the good-bad scales is goodness,
not badness. The term goodness normally spans the whole range of qualities from
good to bad, while badness spans only the range from neutral to bad; goodness,
however, can also be used in a second sense that spans the range only from the
neutral point to the good end of the scale. Second, the question How good was the
dinner? makes no presupposition as to the quality of the dinner: it could have been
good or bad. But the question How bad was the dinner? clearly presupposes that
the dinner was bad. The question How good was the dinner? with a stress on how,
however, does make the presupposition that the dinner was good, and like the how
bad? question, requests merely that the listener specify the extent of the quality
questioned. So again, good can be used either neutrally or with a contrastive
connotation. Third, although good and bad do not exhibit this quality, long, the
neutralizeable member of the long-short pair, is also the member that is able to
accept measure phrases. Thus, two feet long is acceptable, whereas two feet short
is not. And two feet long does not imply that what is being measured is actually
long; rather, it merely specifies that the dimension being measured is length, as
. opposed to depth or height or the like.
Putting this evidence together, we see that good, the positive term, is the only
term able to serve as the name of the scale; hence, this sense of good- say, good1
- might be called the 'nominal' sense of good. Bad, on the other hand, is capable
only of specifying that something is below the norm on the goodness scale; it only
can serve in a 'contrastive' sense. Bad's positive counterpart is good2 , the 'contras-
tive' sense of good, as in John is good . .The interrelations among good., good2 , and
bad might be diagrammed like this:
goodness1

goodness. badness

Goodness, is superordinate to both goodness2 and badness. Stated yet another way
to say that something is good2 or bad is to presuppose that it can be evaluated on
1364 HERBERT H. CLARK

the goodness1 scale; similarly, to say that somethlng is either long2 or short is to
presuppose that we can speak of its length. On the other hand, we can say that
somethlng has goodness, without ever committing ourselves to whether it is good,
or bad. In short, good1 should be thought of as the superordinate, or the seman-
tically prior sense, of the two contrastive senses, gootk and bad.
All this discussion is pertinent to the comparative, since it must be decided for
the sentence John is better than Pete whether the good underlying that sentence is
good1 or good2• In the evidence examined above, the ouly time good could take on
the sense of good1 was when good was modified by a degree marker, as in How good
was the dinner? Other evidence seems to show that this is also true in general.
But it was also noted that John is better than Pete is derived from the sentences
that 'COntain degree markers - as indicated by the so in John is so good. This
means that John is better than Pete might be interpreted in two ways, one as John
is better1 than Pete, which presupposes only that John and Pete are being evaluated
somewhere on the good-bad scale, and the other is John is bette,.. than Pete, which
presupposes that John and Pete are, in fact, on the good end as opposed to the bad
end of the scale. On the other hand, Pete is worse than John must always presup-
pose that John and Pete are being evaluated for badness.
Now we return to the original problem: how can we translate from presupposi-
tions about goodness to ones about badness, and vice versa? For us to be able to
judge that John is better than Pete is synonymous with Pete is worse than John, we
must know: (1) that better can be interpreted in the sense of good,; (2) that good,
is superordinate to or presupposed by bad; (3) therefore, more good1 (better) is
entailed by the expression less bad. It is clear that the relation between the two
sentences is not a simple one, for in translating from one to the other, we are forced
to choose good1 before the translation can work.
The representatiOn problem for better ancl worse can be solved, therefore, by
specifying the relation between good and bad in the notation. Good1 might be re-
presented as [Evaluative [Polar]], good• as [Evaluative I+ Polar]], and bad as
[Evaluative [-Polar]]. This notation expresses the differences between the three
senses quite accurately. Good1 is simpler than gooa. or bad, since good1 is not
specified for polarity. Also, gooa. and bad presuppose good, since the former con-
tain all the specifications of good1 plus the additional specification of polarity on
the good-bad scale. And furthermore, the translation rules for getting from John is
better than Pete to Pete is worse than John can be written in terms of the simple
addition or deletion of the sign in front of Polar. This notation and its conse-
quences constitute What I will call the PRINCIPLE OF LEXICAL MARKING. Although
this principle will be important in the following discussion, it will not be necessary
to keep this cumbersome notation throughout. So it must be kept in mind that the
good and bad underlying comparative sentences are not as entirely separate as the
notation ((John is good+) (Pete is good)) might suggest; rather, they are identical
except for the sign in front of the feature Polar.
 SEMANTICS AND COMPREHENSION 1365
4.13 Model for the comprehension of comparatives
A model for the comprehension of the comparative that has been proposed pre-
viously (Clark 1969a, b; 1970a) was designed to be general enough to account for a
wide variety of comprehension phenomena. That model will be shown here to
account for: (1) people's ability to answer questions like Who is best? asked of the
comparative; (2) the relative difficulty of comparatives whose presuppositions are
about badness (and other implicitly negative terms) as opposed to goodness (and
other positive terms); (3) the difficulties people have in solving such three-term
series problems as If John isn't as bad as Pete, and Dick isn't as good as Pete, then
who is best?; (4) people's judgments of what is an appropriate paraphrase of a com-
parative; (5) the latencies people show in verifying true and false comparatives;
(6) people's difficulties in following instructions, containing comparatives, on where
to place things; and (7) the odd mistakes that children make in comprehending
comparative sentences.
4.131 Stage I representations. The model simply asserts that the comparative in
54a and the negative equative in 55a are represented at Stage 1 as 54b and 55b,
respectively:
54 a) John is better than Pete.
b) ((John is good+) (Pete is good))
55 a) Pete isn't as good as John.
b) ((Pete is good) (John is good+))
Also, the less-comparative 56a is represented as in 56b:
56 a) Pete is less good than John.
b) ((Pete is good-) (John is good))
Another representation that will also be needed later is that for Who is best? Since
this question is approximately equivalent to Who is better than anyone else?, it
therefore has the same presuppositions as the latter paraphrase. For the present
purposes, I will represent 57a simply as 57b:
57 a) Who is best?
b) (X is good++)
The two + s simply denote that the superlative degree of goodness is being specified.
Another problem of Stage 1 is how to construct representations for the under-
lying adjectives good and bad. The model presumes that good and bad are re-
presented as particular feature complexes. The question is, should the two feature
complexes be equally easy to form? Intuitively, the answer is no, for it seems that the
unmarked or positive member of the pair (good) should be easier to represent at
Stage 1 than the negative (bad). This would agree with the evidence presented in
the previous sections on implicitly positive and negative lexical items: the positive
prepositions, for example, were consistently found to be represented more quickly
at Stage 1 than the implicitly negative onesc But if this is the case, how is this to be
accounted for? In a previous paper (Clark 1969a), I proposed one model that pre-
dicts .a difference in encoding time; in the present chapter, I would like to consider
1366 HERBERT H. CLARK

another possibility as well. I will call these two suggestions Proposal I and Proposal
IT, respectively.
Under Proposal I, comparatives with good are expected to be coded more quickly
at Stage 1 than those with bad since good is normally neutralized in meaning in
comparative constructions, The argument is as follows. By the principle of lexical
marking, good, is represented in a simpler form in the feature complexes than good2
or bad. The assumption is that since this is true, good1 should take less time to
represent than good2 or bad. This hypothetical difference in representation time
has two consequences. First, the good of comparatiye or equative constructions
will normally be understood in the sense of good, since that is the simpler and first
sense to be constructed. Second, and as a result of the first consequence, com-
parisons containing underlying good will be represented at Stage 1 more quickly
than those containing bad, since the former will normally be understood as the
simpler good, and the latter can only be understood as the more complex bad. In
short, tbis predicts that comparatives with good (or any other implicitly positive
adjective) will be encoded more quichly than those with bad (or any other
implicitly negative adjective).
Under Proposal II, comparatives with. good would be represented more quickly
than those with bad because both good1 and good2 ar.e represented m.ore quickly
than bad. The argument is similar to the one proposed to account for the difference
between above and below in section 3. In tbis case, whenever the feature Evalu-
ative is formed it could carry along no Polarity feature at all, and then itis inter-
preted simply as good, meaning 'extent of evaluation'. This, of course, should be
the easiest to form. But, the feature Evaluative can also be formed carrying along
with it the feature +Polar redundantly to form good2 , which is taken to mean
'extent of positive evaluation'. To form bad, therefore, it takes yet another step,
and that is to change the feature +Polar to -Polar. The meaning of bad is there-
by reached by forming the representation meaning 'extent of positive evaluation'
and changing it to one meaning 'extent of negative evaluation\ Under tbis proposal,
then, three levels of representation time are distinguished. First, good1 should be the
fastest, sinceit requires one less feature to be set up than either good2 or bad. And
second, .good2 should. be faster to represent than bad, since the feature +Polar is
set up redundantly with the feature Evaluative; and to form bad, the sign on that
feature must be reversed in a furtber step.
As is clear, Proposal II actually contains Proposal I as one of its parts. Although
the evidence neede<! to support either of these two proposals is very difficult to
obtain, some results will. be presented later that suggest that at least Pnipdsal I is
true. This evidence shows that an unmarked adjective that neutralize& is coded
more quickly than an adjective that does not. But there also seems to be evidence
that the positive adjective large, for example, is encoded more quickly tban the
implicit negative small even· when the adjectives are not used in the comparative
form and large cannot therefore be neutralized (Chase and Clark, unpublished data);
 SEMANTICS AND COMPREHENSION 1367
and there seems to be support for Proposal II in other unpublished data as well.
The bulk of the evidence, however, is consistent with both proposals and cannot
discriminate between them.
4.132 Stage 3 comparisons. In the present model of comprehension of compara-
tives, the Stage 3 comparison process is very similar to the corresponding processes
for negatives and locatives. The main requirement again is that two underlying
representations must be compared for congruence, and if they are not congruent,
certain extra· manipulation operations must be performed in an attempt to make
them congruent. The main point concerning congruence that will interest us will be
in the presuppositions of the comparative: if the comparative encoded at Stage 1
does not contain the same presuppositions as whatever is represented at Stage 2,
then the Stage 3 comparison operations must make these two sets of presupposi-
tions congruent before any other comparison operations can be performed.
Since specific applications of this model are easiest to describe in the context of
particular tasks, I now tum to the experimental evidence on the comprehension of the
comparative in four types of tasks: question-answering tasks, paraphrasing tasks,
verification tasks, and instruction-following tasks. Evidence from all of these
studies will be shown to support the main two proposals of this model of com-
prehension: the principle of congruence and the principle of lexical marking.

4.2 Answering Questions

4.21 The two-term series problems

One of the principal investigations of comparatives is found in a series of studies
(Clark 1969a, b; 1970a) in which subjects were required to answer questions about
comparative sentences. In the simplest of these experiments, the task is so easy that
it seems trivial. Subjects were asked to answer questions of the following type: If
John isn't as bad as Dick, then who is best? Here the if-clause (the 'proposition')
asserts certain information, and the question immediately asks about that infor-
mation. In these so-called two-term series problems the propositions contained is
bette.r than, is worse than, isn't as good as, or isn't as bad as as the relational term
comparing two common names, and the question was either Who is best? or Who
is worst? The subject was timed while he read and answered the question; his
comprehension time was taken as the reading-plus-answering time minus the read-
ing time alone.
This task can be viewed exactly as the other question-answering and verification
tasks for the negatives and locatives have been viewed - as a four-stage process.
Given If John isn't as bad as ·Pete, then who is best?, the process would go as fol-
lows. At Stage 1, the subject encodes the proposition as ((John is bad) (Pete is
bad+)). At Stage 2, he represents the question as (X is good+ + ); this notation
specifies that the question presupposes the goodness scale, normally the goodness1
1368 HERBERT H. CLARK

scale, and that an X is wanted that fulfills the appropriate description. At Stage 3,
he tries to match the question to the information in the proposition, and after a series
of manipulations succeeds in replacing X with John. And at Stage 4, he utters 'John',
the answer that has been produced by the Stage 3 manipulations. Clearly, it is again
the Stage 3 comparison and manipulation operations that are of most interest in
this process.
The main point about Stage 3 is that its operations obey the principle of con-
gruence. Note that ill comparing (X is good++).to ((John is bad) (Pete is bad+)),
the first string is congruent with neither of the underlying sentences of the second
string. The finding of congruent information, then, requires that the presuppositions
of either the first or second string be changed. Let us assume that it ;s the question
that is changed, from (X is·good+ +)to (X is bad-), that is, from 'who is best' to
'who is least bad? Now there is congruence of the presuppositions of the propo-
sition and question. By other minor manipulations, the question (X is bad-) can
be made fully congruent to the subsentence (John is bad) of the proposition, thus
making X= John, the correct answer. If instead the proposition is Jolin· isn't as
good as Pete, and the question is who is best?, then the presuppositions of the
proposition and question already match and there is no need for the first operation
that reformulates the question in terms of the presupposition for badness. There-
fore, Stage 3 requires one less operation - the question translation operation -
when the presuppositions of the proposition and question are congruent than when
they are not. As a result, the two-term series problems with presuppositional con-
gruence should be answered more quickly than those without.
The results of the experiment of two-term series problems nicely confirm the
predictions of this model. Table X lists the four types of propositions and opposite
them the solution time, in seconds, for the two questions who is best? and who is
worst? Proposition'! was answered more quickly with the question who is best?

TABLE X. Mean solution time for two-tetm series problerrzs

Form of question
Form of proposition Analysis Who is best? Who is worst? Mean

·Ais good
I A is better than B .61 .68 .64
B is good
A is bad
II B' is worse than A 1.00 .62 .81
B is bad
A is bad
I' A iSn't .as bad @.S B 1.73 1.58 1.66
Bis bad
B is good:
II' B isn't as good as A 1.17 1.47 1.32
A is good
 SEMANTICS AND COMPREHENSION 1369
and Proposition II, with the question who is worst? Similarly, Proposition I' was
answered more quickly with the question who is worst?, and Proposition II', with
the question who is best? In these four instances, the problem was solved faster
when the proposition and question had the same presuppositions than when they
did not. Furthermore, the surface properties of the problems could not, by them-
selves, aiiow us to make these predictions. For Propositions I and II, the answer
was given more quickly when it was the subject of the sentence; in contrast, for
Propositions I' and II', the answer was given more quickly whim it was the predicate
term. The main pattern of solution time, then, appears to be. accounted for by
the congruence or incongruence of presuppositions, as the model of comprehen-
sion predicts.
The comprehension model with its principle of lexical marking makes another
prediction for the results in Table X - namely, that propositions with underlying
good should be comprehended more quickly than those with underlying bad. This
prediction is borne out with the comparative sentences, 1.10 sees to 1.50 sees, as weii
as with the negative equative sentences, 1.64 sees to 1.80 sees. Again other attributes
of the sentences could not allow us to make this prediction. In the comparative
propositions, the easier one, Proposition I, has the 'better' term as its subject,
whereas in the negative equative propositions; the easier one is Proposition II',
which has the 'worse' term as its subject. So again, solution time seems to be
accounted for maiuly by properties of the presuppositions.

4.22 Deductive reasoning

The second kind of question-answering task to be examined is one that would
traditionaily be included under the rubric of deductive reasoning, for it requires
the subject to answer questions about the joint information from two premises. But
it is reaiiy just a simple extimsion of the previous two-term series problem caiied
the three-term series problem: e.g. If John isn't as good as Pete, and Dick isn't as
bad as Pete, then who is worst? Although this task has been investigated by a series
of psychologists (Burt 1919; Hunter 1957; Donaldson 1963; DeSoto, London and
Handel1965; Hanr!el, DeSoto and London 1968; Hutteulocher 1968), the previous
stodies have made only minor reference to the linguistic processes underlying the
solution of these problems. The model to be presented here, first proposed in Clark
(1969a, b), makes the claim that the major difficulties in the three-term series prob-
lems are to be accounted for by the processes that deal with the comprehension
and manipulation of the information in comparative sentences. That is, this task
can be considered merely another example of how comparatives are comprehended,
hence it should exhibit the same qualities found in the two,term series problems
and common to other tasks involving the comprehension of comparatives. The
difference between studies of three-term series proble)IIs and other comprehension
studies is maiuly one of complexity: there is a wide variety of three-term series
problems that require detailed analysis before predictions can be.made about their
1370 HERBERT H. CLARK

solution time. I will try merely. to outline the argument, giving evidence from Clark
(1969a, b) where appropriate. For a more detailed argument, the reader should
consult Clark (1969a, b).
In Clark's (1969a) task, subjects were given 32 types of three-term series prob-
lems in the following form:·
If John isn't as good as Pete,
And Dick isn't as bad as Pete,
Then who is worst?
Pete John Dick
The subject read the problem to himself and said the answer ont lond as soon as he
TABLE XI. Geometric mean solution times for three-term series problems
Form of question
Form of problem Analysis Who is best? Who is worst? Mean

A is better than B, A is good

I B is good 5.20 5.80 5.49
B is better than C. Cis good
C is worse than B, A is bad
li B is bad 6.10 5.74 5.91
B is worse than A. Cis bad
A is better than B, A is good
III B is good, bad 5.09 5.58 5.33
C is worse than B. Cis bad
B is worse than A, A is bad
IV B is bad, good 5.53 5.73 5.63
B is better than C. Cis good
A isn't as bad as B, A is bad
I' Bis bad 6.96 6.25 6.59
B isn't as bad as C. Cis bad
C isn't as good as B, A is good
II' B is good 5.84 6.61 6.22
B isn't as good as A. Cis good
A isn't as bad as B, A is bad
III' B is bad, good 6.53 6.50 6.52
C isn't as good as B. Cis good
B isn't as good as A, A is good
IV' B is good, bad 5.78. 6.63 6.19
B isn't as bad as C. Cis bad
 SEMANTICS AND COMPREHENSION 1371
could while he was timed. ·using the convention that the letters A, B, and C in
each problem stand for the best, middle, and worst person, Table XI lists the eight
basic combinations of propositions and their average solution times. Since the two
propositions of each pair can occur in either order, and the question can be either
Who is best? or Who is worst?, there are actually 32 possible three-term series
problems - excluding, for now, those whose answers are indeterminate. The
model proposed above makes several important predictions about the relative
solution time of these 32 problems. ,
The first prediction to be made is based on the principle of congruence. This
principle predicts that those problems for which the presuppositions of the ques-
tion are the same as those of the proposition that contains the answer should be
easy, and those problems for which this does not hold should be difficult. For
Problems I, II, I', and II', then, this principle predicts that Problems I and II',
which presuppose the goodness scale, should be easier with the question Who is
best? than with Who is worst?, whereas Problems II and 1', which presuppose the
badness scale, should be easier with the question Who is worst? than with Who is
best? As shown in Table XI, each of these individual predictions is confirmed,
with an average solution time of 4.50 for the congruent problems and an average
of 4.75 for the incongruent ones.
A second example of congruence is found in Problems III, IV, III', and IV'.
First consider Problem III. The answer to the question· Who is best? is A, which
is found in a proposition that presupposes goodness; likewise the answer to Who
is worst? is C, which is found in a proposition that presupposes badness. That is,
the question and the proposition that contains the answer are congruent in Prob-
lem III. This is also true of Problem IV'. But for Problems IV and III', this is
not the case. The answer, in both problems, for Who is best? is still A, but in
these problems A is found in a proposition that presupposes badness; similarly,
the answer to Who is worst? is C and is found in a proposition that presupp<;>ses
goodness. Problems IV and III' therefore contain questions and corresponding
propositions that are incongruent in their presuppositions. Solving Problems IV
and III', then, requires an extra step for translating the presupposition of the ques-
tion into one that is congruent with a proposition, and so Problems IV and III'
should be solved more slowly, respectively, than Problems III and IV'. The solu-
tion times in Table XI confirm this prediction.
Finally, the solution times in Table XI confirm the prediction made by the
principle of lexical marking that comparatives that presuppose goodness should be
solved faster than those that presuppose badness. Since Problems Ill, IV, III', and
IV' contain both a good- and a bad-proposition, they are irrelevant to this predic-
tion. But Problem I presupposes goodness and Problem II badness although they
are otherwise identical, so this hypothesis would predict Problem I to be solved
faster than Problem II. And this is confirmed. For the same reasons Problem II'
'hould be solved more quickly than Problem 1', and that prediction is also con-
 1372 HERBERT H. CLARK

firmed. The model of comprehension of the comparative, therefore, is fully con-

sistent with the results from this set of three-term series problems.
In a slightly different experiment (Clark 1969b), the same three-term series
problems were solved by another group of subjects who were given 10 seconds to
solve each problem. For this task the difficulty of each problem was measured by
the number of subjects who made an error on that problem by the end of the 10
second period. As one might expect, the results in this second task only confirmed
what was found in the first: those problems that were solved faster in the first task
were solved correctly more often in the second task. Thus, this second experiment
fully confirms all the predictions made for the first. But this second experiment is of
additional interest because subjects solved 32 indeterminate problems in addition
to the 32 determinate problems shown in Table XI. One such indeterminate prob-
lem islf John is better than Pete, and Dick is better than Pete, then who is best?
ln this example, both Johu and Dick are better than Pete, and there is no way of
determining who is best. The correct answer to this problem is 'can't tell', which
was an alternative the subjects were allowed to give. The interest in these prob-
lems is not so much in their difficulty as in the mistakes the subjects made on
them, for the model that we have been discussing makes quite specific predictions
about what those errors should be. ·

TABLE XII. Two types of indeterminate problems and their percentage errors
Problem Analysis Principal Errors

If G is better than J G is good G: 26%

VII And J is worse than H His bad H: 6%
Then who is best? J is good, bad J: 2%

If. G isn't as bad as J G is bad G: 12%

VII' And J isn't as good as H His good H: 30%
Theri who is best? J is good, bad J: 2%

The 'errors are best illustrated in the indeterminate problems containing one
prbposition that presupposes goodness and one that presupposes badness. Two
such problems, labeled VII and VII', are shoVI'n in Table XII. To see what errors
subjects should make on these two problems, consider the underlying presupposi-
tions oftheir premises; these presuppositions are shown in the second colinnn of
Table XII. In VII, we find G is good, H is bad, and J is both good and bad. In
VII', however, which is simply VII with is better than replaced by isn't as bad as,
and is worse than replaced by isn't tis good as, we find that the case is reversed;
now G is bad, H is good, and I is both good and bad. When the subject is asked
to search through the Stage, 1 representations of the premises, to find an answer
for Who is best? (i.e. (X is good++)), he will naturally search for an answer in a
 SEMANTICS AND COMPREHENSION 1373
premise with a presupposition that is congruent with the question (i.e. a premise
with presuppositions containing good). In VII, this strategy should lead the subject
to find G, since G is good is congruent with (X is good+ + ), whereas H is bad is
not. In Xll', the strategy would lead the subject to answer H, since H is good is
the congruent string while G is bad is not. The correct answer to both problems,
of course, is 'can't tell'. Significantly, this strategy predicts the errors the subjects
made quite accurately. As shown in Table XII, on VII subjects chose the incorrect
but congruent answer G 26% of the time, but chose the other two incorrect and
incongruent answers H and J a total of only 8% of the time. In, Vll', on the other
hand, subjects gave the incorrect but congruent answer H 30% of the time, but
the incorrect and incongruent answers G and J a total of only 14% of the time.
Thus, the strategy of searching through the Stage 1 semantic representations for con-
gruent representations strongly predicts the specific errors subjects made in those
two problems. The other problems with the same properties as VII and VII' con-
firmed the predictions in precisely the same way (cf. Clark 1969b: 211).
In the experiments on deductive reasoning discussed so far, the principle of lexi-
cal marking - that positive adjectives should be comprehended faster than nega-
tive ones - has had only one piece of confirming evidence, namely, that com-
paratives with good were comprehended faster than ones with bad. But in sifting
through the past literature on the three-term series problem, I have turned up many
other examples of positive-negative pairs of comparative adjectives and their rela-
tive difficulty of comprehension. In particular, the data of Burt (1919), Hunter
(1957), DeSoto eta!. (1965), Flores d'Arcais (1966), Handel eta!. (1968), Hutten-
lpcher (1968) and Clark (1969a) show that the positive adjectives better, faster,
warmer, higher, deeper, more, farther, taller, happier, and older are comprehended
more easily, respectively, tllan their negative counterparts worse, cooler, lower,
shallower, less, nearer, shorter, sadder, and younger. I have not found a single
exception to this generalization in the psychological literature or in my own un-
published studies with various adjectives.
Jones (1970), moreover, has made some specific tests of Proposal I of the
principle of lexical marking. As she pointed out, heavy and light are seman-
tically unmarked and marked, respectively, whereas dark and light are not asym-
metrical in this respect; note that neither darkness nor lightness is really the
proper scale name for the light-dark dimension, particularly when applied to hair
color, and neither dark nor light neutralizes in How dark? type questions. Similarly,
thick and thin are semantically unmarked and marked, respectively, whereas fat
and thin are not. If Proposal I is correct, therefore, heavier should be easier than
lighter, and thicker should be easier than thinner, whereas there should be little or
no difference between darker and lighter, and between fatter and thinner. These
predictions were confirmed by subjects who solved three-term series problems as
they were timed. So with the evidence from the previous literature and from Jones,
thereis very strong support for the principle of lexical marking.
 1374 HERBERT H. CLARK

4.23 Psychophysical judgments

Remotely related as they might seem, several studies on the psychophysical judg-
ments of the relative brightness of lights lend even more support to the present
model of the comparative. In the principal experiment, Audley and Wallis (1964)
simultaneously presented two very bright lights, with one slightly brighter than the
other, or two very dim lights, again with one slightly brighter than the other, and
asked subjects to judge either which was brighter or which was dimmer. The sub-
jects were timed from the presentation of the left and right light patches to the press
of the corresponding left or right button. Audley and, Wallis found that subjects
could judge which was brighter faster for the very bright patches than for the very
dim ones, but which was dimmer faster for the very dim patches than for the very
bright ones.
With only one added assumption, this resnlt fits quite nicely into the present
conception of the comparative. The assumption is that the pair of bright patches
are encoded as sometbing like ((left is bright+) (right is bright)) -thereby en-
coding the comparative brightness of the left and right patches - and 'that the
very dim patches are encoded as ((left is dim+) (right is dim)). The questions
Which is brighter? and Which is dimmer?, of course, are encoded, respectively, as
something like (X is bright+) and (X is dim+). The brighter-question, then, is
congruent in its presuppositions with the encoding of the pair of bright patches,
and the dimmer-question, with the dim patches, so the principle of congruence
applies and predicts the differences that Audley and Wallis found. It should be
noted that Wallis and Audley (1964) carried out almost the same experiment fot
high and low tones with the same resnlts, and Shipley, Norris, and Roberts (1946)
found similar resnlts for comparative judgments of good and bad esthetic objects.
Further manipulati~ns of the stimuli also produced data consonant with the prin-
ciple of congruence.

4.3 Paraphrasing

The second kind of study included in the broad .category of ·comprehending com-
paratives is illustrated by Flores d'Arcais's (1966) important study on paraphrases
of comparative sentences. Flores d' Arcais, too, noted that the interpretation of
comparative sentences was influenced by their presuppositions - although he did
not use the term presupposition. He noted, for example, that Lambs are less
ferocious than lions seems to make assumptions about the underlying dimension
that were different from those in Lambs are more gentle than lions bnt almost the
· same as those in Lambs aren't as ferocious as lions. In this paraphrase experiment,
then, he asked subjects which of several sentences was the best paraphrase of sen-
tences like Lambs are less ferocious than lions. His resnlts may be simply stated in
terms of the present model of the comparative. A paraphrase was judged more ac~
ceptable when it contained the same presuppositions as the target sentence. Lambs
 SEMANTICS AND COMPREHENSION 1375
are less ferocious than lions, for example, since it is represented mentally as ((lambs
are ferocious-) (lions are ferocious)), was judged more similar to Lambs aren't as
ferocious as lions, ((lambs are ferocious) (lions are ferocious+)), than to Lambs
are more gentle than.lions, ((lambs are gentle+) (lions are gentle)). Flores d'Ar-
cais's results re-affirm the conclusion that subjects derive and use an abstract re-
presentation of the comparative sentence that is something like ((lions are fero-
cious+) (lambs are ferocious)).
I

I 4.4 Sentence Verification

The comparative has been used in only one study of sentence verification that I
know of, and that is one by Flores d'Arcais (1966). He was interested in studying
the comparative adjectives more and less, since it appeared to bim that less was
psychologically more complex than more. Flores d'Arcais's task was an elegantly
simple one. He gave his subjects sentences llke A lion is less ferocious than a sheep
and asked them to indicate whether the sentences were true or false as fast as they
could by pressing one of two buttuns. Flores d' Arcais timed his subjects from the
presentation of the sentence to the response. Table XIII lists the four types of sen-
tences Flores d'Arcais used, an example of each type of sentence, and the verification
time for each type.

TABLE XIII. Geometric mean verification times for comparative sentences

(from Flores d'Arcais 1966)
Type of Sentence Example· of Sentence Verification Times

True More A lion is more ferocious than a sheep. 2.68 sec

False More A lion is more peaceful than a sheep. 2.90
True Less A lion is less peaceful than a sheep. 3.53
·False Less A lion is less ferocious than a sheep. 3.75

In suggesting one explanation for these. results, Flores d' Arcais hypothesized
'that the subject has to formulate the right assertion before reaching the decision.
For example, when given the statement "A lion is more peaceful than a sheep", the
subject would have to "formulate" one of the two following statements: "A lion is
more ferocious than a sheep"; or "A sheep -is more peaceful than a lion"~ that is, a
true statement, and match it with the one given, to reach the conclusion that the
statement is false' (p. 12). This suggestion is quite consistent with the present
theory of comprehension, although it is instructive to spell out this model in more
detail.
In the present theory, the sentence A lion is more peaceful than a sheep would
be represented at Stage 1 as in 58:
1376 HERBERT H. CLARK

58 ((lion is peaceful+) (sheep is peaceful))

The critical question, however, is how one's knowledge about lions and sheep are
represented at Stage 2. One possibility is that the subject accepts the presupposi-
tions given in the sentence and attempts to express his knowledge in accord with
those presuppositions, as in 59:
59 ((lion is peaceful -) (sheep is peaceful))
At Stage 3, these two representations would be compared, and since they are con-
gruent in their underlying presuppositions, this is easily done by simply comparing
the + and the - (or the equivalent). A mismatch is found, the subject changes a
presupposed truth index from true to false (just as with negatives - cf. section 2),
and the outcome, false thereby serves as a basis for a push of the 'false' button.
Under this model, false sentences should take longer than true ones, no matter
whether the sentence contains more or less. Flores ·d'Arcais's data are consistent
with this, showing that false sentences take 220 msec longer than true ones for both
more and less sentences. Furthermore, since more is positive and less implimtly
negative, more should take less time to code at Stage 1 than less, thus mill<ing the
more sentences take less time overall than less sentences. Flores d'Arcais's data are
consistent with this prediction too, showing a 850 msec difference between more and
less. Unfortunately, this 850 msec seems considerably longer than 200 to 300 msec
Negation Time shown for the other implicit negatives in section 2, and this suggests
that there might be something additional that differentiates the more and less state-
ments. What this might be remains to be seen.
Flores d'Arcais's data, however, allow us to reject certain other models as to how
this comparison process might proceed. The most interesting of the alternative but
incorrect models is one in which the subject sets up his Stage 2 representation
INDEPENDENTLY of,his Stage 1 representation. Consider the case in which the sub-
ject has set up representation 58 at Stage 1, and at Stage 2, he sets up represen-
tation 60, which is true statement about the relative ferocity of lions and sheep:
60 ((lion is ferocious+) (sheep is ferocious))
In this case, the underlying presuppositions of 60. and 58 are incongruent, so the
subject would have to carry out some manipulation to make them congruent; after
that manipulation, he would find a mismatch of the + and - in the resulting
pair of strings and would change his p.res11pposed truth value true to false, as usuaL
If carried through in detail, however, this analysis predicts that the difference be-
tween True. Less and False Less verification times should be smaller or even the
reverse of the difference b_etween True More and-False More verification times.
The data in Table Xlll are not at all consistent with.this modeL It should be noted
that the model just presented is similar to considering the Tru.e More, False More,
True L~ss, and False Less sentences to be equivalent to A .lion is ferocious, A lion
is peaceful, A lion isn't peaceful, and A lion isn't ferocious, respectively, that is, as
equivalent to True Positive, False Positive, True Negative, and False Negative sen-
tences. The 'true' model of negation, it will be recalled, predicts that True Nega-
 SEMANTICS AND COMPREHENSION 1377
lives should be slower than False Negatives, and this is inconsistent with the alter-
native interpretation just given of Flores d'Arcais's data.
This second incorrect model demonstrates very nicely how important the under-
lying presuppositions of the comparative are. In this model, it is assumed that
the subject at Stage 2 set up representation 60, which is a true description of the
relative ferocity of lions and sheep. But how did the subject know to represent their
relative ferocity rather than their relative laziness, their relative size, their relative
speed, or the like? To have done this the subject must already have taken into
account the presuppositions of peacefulness in the original sentence (in this case
58); he would have had to have known that presuppositions about ferocity can be
translated into presuppositions about peacefulness, whereas presuppositions about
laziness, size, speed, and the like cannot. In other words, it is simply not possible
for a subject to set up a Stage 2 representation independently of the presuppositions
of the Stage 1 representation and guarantee the possibility of deriving an answer.
This line of reasoning. argues that the most efficient Stage 2 representation to form
is one that agrees precisely with the Stage 1 representation in its presuppositions,
and this is jUst the model (presented previously) that best accounts for the Flores
d'Arcais data.
In smmnary, Flores d'Arcais's data can be accommodated within the present
theory of comprehension if it is assumed that subjects represent their previous
knowledge at Stage 2 in a way cmisistent with the presuppositions of the com-
parative sentence they are trying to verify. Although it was argued that this assump-
tion is an eminently plausible one, more experimental work will have to be carried
out to confirm this and the other assumptions of the model in full.

4.5 Following Instructions

The ·comparative has also been studied in tasks that require subjects to place all
objectin accordance with a description, as in 'Make it so that the red block is
higher than the black block'. Although the first study of this kind was Rutten-
locher's (1968) theoretical paper on deductive reasoning, this was explicitly based
on her previous work on locatives and actives and passives, as described in sec-
tion 3; The main thrust of the·1968 paper was to present ·an explanation for deduc-
tive reasoning based on the ·notion· that people treat the ·propositions of a three-
term ·series problem as an instruction to·arrange objeCts in a series- in this case
on an imaginary visual display. Hrittenlocher's paper has been followed up more
recently by tw(}; studies, Huttenlocher et al. (1970) mid· Clark and Peterson (in
preparation). In order to put all this information together, however, it is necessary
first to examitie the properties of the comparative as a: locative and to explore the
consequences ·of those properties.
1378 HERBERT H. CLARK

4.51 Point of reference in comparatives

Since the comparative sentences A is higher than B and A isn't as high as B both
describe locations, they might be expected to have properties similar to the loca-
tive, and they do. For locatives we saw that there were constraints on what could
be used to answer the question Where is A? The same constraints hold for com-
paratives, as can be seen in the question-answer sequences of 61:
61 a) Where is A? A is higher than B.
b) *Where is B? A is higher than B.
c) Where is A? A isn't as high as B.
d) *Where is B? A isn't as high as B.
Like locatives, a comparative or equative construction describes where its subject
is with respect to its predicate term. The term in the predicate - B in 61 - al-
ways serves as a reference point of position.
These facts can be stated in terms of presupposition. The sentences A is higher
than B and A isn't as high as B both presuppose that the height of B is known, as
indicated in a rough paraphrase of the former, A is at some height that is greater
than' the known height of B. It is this presupposition that accounts for the unac-
ceptability of 61b and 61d, for the question Where is B? implies that B's position
is not known whereas the two answers A is higher than B and A isn't as high as B
presuppose that B's position is known. In other words, the comparative A is higher
than B has presuppositional properties very similar to the locative A is above B,
and since this is true, the comparative and locative should share many properties
when they are used in carrying out instructions.
The point of reference is as important in the comparative as it was in the loca-
tive. Again, it is assumed that instructional tasks can be treated as implicit ques-
tion-answering tasb. When the subject has A in his hand and is told 'Make it so
that A is higher than B', the notion is that the subject implicitly asks 'Where should
A go {or be), given that I know the position of B?', and he then tries to match this
question with the instruction he was given to come up with the correct answer. So
at Stage 1, the subject encodes the instruction; at Stage 2, he encodes the implicit
question; at Stage 3, he compares the two representations to find the answer; and
at Stage 4, he carries out the instruction.
To examine this process in more detail, imagine that the subject was holding a
red block in his hand and imagine that he was looking at a display with a blue
block in it. The first instruction given to the subject is 'Make it so that the red block
is higher than the blue block'. This instruction and the implicit· question would be
represented at Stages 1 and 2, respectively as 62a and 62b:
62 a) {{red block is high+) {known {blue block is high)))
b) {(red block is X) {known {blue block is high)))
In this notation, I have explicitly represented the point of reference in the (known
(blue block is high)) type of notation; furthermore, I have represented the question
Where should the red block be with respect to the blue block? in a comparative
 SEMANTICS AND COMPREHENSION 1379
form, since this notation represents both the form the answer must be in and the
information that the position of the red block is being questioned with respect to
the blue block. The Stage 1 and Stage 2 representations in 62 are obviously fully
congruent, so the Stage 3 comparison process would find it necessary only to re-
place X by high+, forming the implicit instruction 'The red block should be higher
than the blue block'. At Stage 4, it is a simple matter for the subject to conform
to this instruction by placing the red block higher than the blue block.
In contrast to this simple problem, consider the subject holding a red block and
given the instruction, 'Make it so that the blue block is higher than the red block'.
The Stage 1 and Stage 2 representations for this problem are shown in 63a and
63b, respectively:
63 a) {(blue block is high+) {known {red block is high)))
b) ((red block is X) (known (blue block is high)))
The two strings 63a and 63b are now quite incompatible and require much mental
manipulation before they become congruent. We can only specnlate at this point
on exactly what these manipnlations are, but it is sufficient to assume simply that
they are time-consuming operations not reqnired by the problem represented in 62.
This analysis therefore predicts that the problem represented in 62 shonld be solved
more quickly than the problem represented in 63. In the same way, the instruc-
tion 'Make it so that the red block isn't as high as the blue block' is represented
in 64a, and 'Make it so that the blue block isn't as high as the red block' in 65a;
it is assumed for both instructions that the red block is in hand and the blue block
is fixed, giving the implicit questions 64b and 65b:
64 a) ((red block is high) (known (blue block is high+)))
b) ((red block is X) (known (blue block is high)))
65 a) ((blue block is high) (known (red block is high+)))
b) ((red block is X) (known (blue block is high)))
Again, 64 shows congruence, and 65 does not.
The relative difficulty of instructions 62 through 65 can be summarized quite
simply. In placing a red block in an array that already contains a blue block, the
subject must ask himself implicitly, 'Where is the red block to go, given that I know
the position of the blue block?'. Since this query contains an implicit where-ques-
tion, it is easily answered only by those instructions that also presuppose that the
position of the blue block is known. Thus, instructions 62 and 64 should be easy,
and 63 and 65 shonld be hard.

4.52 Experiments on the comparative as an instruction

The most direct study on the instructional properties of the cqmparative is found
in Clark and Peterson (in preparation). Rather than ask subjects to place blocks
in an array, as Huttenlocher and Strauss (1968) asked their very young subjects to
do, Clark and Peterson gave the subject a red pencil, a black dot on a piece of
paper, an instruction like 'The black dot isn't as far left as the red dot', and asked
1380 HERBERT H. CLARK

him to place a red dot in the appropriate place on the paper as quickly as he could,
while ·he was timed. Although this experiment was successful, it was somewhat
cumbersome mechanically and the timing was not very accurate. For this reason,
Clark and Peterson carried out a similar experiment in which subjects were pre-
sented printed instructions like 'The blue one isn't as high as the pink one'. With
a blue line off to the right of the sentence, the subject was to 'draw' in a pink line
in the appropriate place with respect to the blue line by pressing one of two verti-
cally arranged buttons; the upper ·button represented the drawing of a line above
the blue one, and the lower button the•drawing of a line below the blue one. The
displays were presented in a tachistoscope, and the subject was timed from the
presentation of the display to the press of the button. I will describe only the
second experiment since the first one, with the black dots and red pencil, produced
much the same results as the second.
The sentences Clark and Peterson used. had either the blue one or the pink one
as the subject, and is higher than,.islowerthan, isn't_ as high as, isn't as low as,
is better than, is worse than, isn't as good as, or isn't as bad as as the relational
term. For the evaluative relations, the subjects were told to think of the lines with
the better one on top and the worse on the· bottom; this is how subjects normally
place evaluative objects in a vertical array (cf. DeSoto et a!. 1965; Clark 1969a;
Jones 1970).
The results of this experiment are in fnli agreement with the model I have de-
scribed above. When subjects were asked to 'draw in' the pink line,. each of the
eight kinds of sentences was responded to faster when the pink one was in the
subject position than when it was in the predicate. The occasional errors subjects
made further support this conclusion•. There were fewer errors when the pink one
was in the subject position .titan' when it was in the predicate. These results also
constitute additional evidence for the principle of lexical marking, for those sen-
tences with high were answered 194 msec faster than those with low, and those with
good, 208 msec faster: thall those with bad.
It is instructive here, however, to contrast the results of the instructional task
with those of the two-term series problems discussed above. The present claim is
that the instructiomil task is equivalent to answering questions like those in 66:
66 a) If A is higher than B; then wherds A? [On top]
b) ·If A is higher than B, themwhere is B? [On bottom]
c) If A isn't as high as B, then were is A? [On' bottom}
d) If A isn't as high as B, then where is B? [On top]
These are very similar to the two"term series .problems in 67:
67 a) If A is higher than.B, then which is higher?· [A]
b) If A is higher thall B, then which: is lower? [B]
c) If A isn't as high as B; then which.is higher? [B]
d) If A isn't as high as B, then which is lower? [A]
Clearly, the comparatives, or conditionals,: of· 66 and 67 are identical. and it is
 SEMANTICS AND COMPREHENSION 1381
only the information requested of the comparatives that is not. But the important
congruence relation in 66 is that between the points of reference of the comparative
and its question, whereas the important congruence relation in 67 is that between
the. underlying presupposed semantic dimensions (height vs. lowness) of the com-
parative aud .question. Considerations of congruence predict that 66a and 66c
should be answered more quickly, respectively, than :66h aud 66d, and 67a aud
67c more quickly, respectively, than 67b and 67d. To state these predictions an-
other way, when the conditional is .a strict comparative. (e.g. A is higher than B),
it is always easier to answer questions about the subject of the sentence (A) than
about B, no matter whether the question is a where- or which'..type question. In
contrast, when the conditional is a negative equative (e;g .. A isn't as high as B),
this is not the case: it is easier to answer where-type questions ·about the subject
of the sentence (A), but easier to answer which-type questions about the term in
the predicate (B). Following instructions; therefore, will not have the same proper-
ties as answering Who iS: higher? type' questions, even though the properties. for
both are predictable from the principle of congruence.

4.53 Deductive reasoning

We now return to the deductive reasoning problems lil>e IfJohnis better than
Pete, and John is worse than Dick, then who is· worst? in order to examine two
important attempts (DeSoto et al. 1965; Huttenlocher 1968) to explain the relative
difficnlty of three-terin series problems. Both of these theories were founded on
the premise that people reason in their mind's eye by setting up and :manipulating
imaginary symbols that stand for the three terms of the problem.. I have included
these two theories under the rubric of 'following instructions', because both of them
(Huttenlocher quite explicitly) treat the.. comparatives of each problem as an in-
struction to construct ail iriiaginary display. What is important about these two
imagery theories, however, is that since• they both make predictions that are dif-
ferent from those of the linguistic. theory that I have. championed above, one can
use empirical results to test the validity of the theories~: In.fact, the available evi-
dence appears to support the linguistic theory •aud invalidate the two imagery
theories. But let us look more closely at both the theories· aud the evidence.
4.531 Theory of spatial paralogic. DeSoto eta!. (1965) were the first to suggest
that one conld explain the relative difficnlties .of three"tenno series. problems by
appealing tb .. a theory of 'spatial,paralogic'; .. First, th¢y noted that,people often
say that they actually do. set up imaginary arrays in their minills eye while solving
these problems.· Given A is bettimihan B, the subject will imagine au A above a B;
then given a secondproposition B.-is better than C, he will imagine· the fnll com-
plement, with A which is above B, which is above C.. 'When asked for Who is
best?, the subject merely 'looks' at the top·of the anay,·picksoutA, and says 'A'.
If this is what subjects actually do, DeSoto et a!. pointed out, then the difficulties
of the three-term series problem with better and/or worse: could be explained by
1382 HERBERT H. CLARK

two principles: (1) the principle of preferred direction (my nomenclature) states
that it is easier to construct or read off displays from top down than from bottom
up; and {2) the principle of end-anchoring states that it is easier to construct or
read off displays from the ends in than from the center ont. As it happens, these
two principles predict difficulty correctly for the comparative problems that DeSoto
et al. examined, but not for the negative equative problems that I examined {Clark
1969a, b).
The principle of directional preference predicts that a problem like 68:
68 A is better than B
And B is better than C
should be easier than 69:
· 69 C is worse than B
And B is worse than A
because the subject will always build 68 from the top down and 69 from the bot-
tom np. Subjects do, in fact, report that they visualize both problems with A on
top and C on the bottom. But this principle also predicts that 70:
70 A isn't as bad as B.
And B isn't as bad as C
should be easier than 71:
71 C isn't as good as B
And B isn't as good as A
for exactly the same reasons. And Jones (1970} has found that subjects generally
do set •up 70 and 71 with A on top and C on the bottom. However, the results
of Clark (1969a), shown in Table XI, go directly counter to this prediction, with
70 actually harder than 71. According to the .principle .of lexical marking, of
course, this result is quite explicable, since 70 has .the more difficult bad nuder-
lying it and 71, the easier good. Thus, the principle of lexical marking succeeds
where DeSoto et al.'s principle of directional preference fails.
Their second principle, end-anchoring, predicts that 72:
72 A is better than B
And C is worse than B
should be easier' than 73:
73 'B is worse than A
And B is better than C
since the, propositions in 72 both mention the.extieme terms {A and C) before they
mention the middle term B, but. the reverse is ctrue in 73. That is, the subjects
construct a visual Tepresentation from the ends in in the former problem, but from
the center out in the hitter. However, this principle also predicts, and for exactly
the same reasons; that 74:
74 A isn't as bad as B
And C isn't as good as B
should be easier than 75:
 SEMANTICS AND COMPREHENSION 1383
75 B isn't as good as A
And B isn't as bad as C.
This prediction is incorrect, as the results of Clark (1969a), shown in Table XI,
indicate. On the other hand, we saw that the principle of congruence correctly
predicts that 74 (Problem III' in Table XI) should be harder than 75 (Problem
IV'). So here again, the linguistic principle succeeds where the principle from the
theory of spatial paralogic fails.
4.532 Theory of constructing visual imagery. Hutteulocher has proposed a
theory of constructing visual imagery in order to account for many of the same
phenomena that DeSoto et al.'s theory was designed to account for. The main
difference between her theory and DeSoto et al.'s was that she attempted to con-
struct a more satisfying explanation for the end-anchoring effects found by DeSoto
et a!. Her thesis was this: People typically show certain difficulties in following in-
structions that are couched in terms of comparative sentences. For example, as we
saw in section 3 and earlier in this section, it is easy to place something expressed
as the subject of a locative or comparative sentence, but difficult otherwise. If
people construct visual images in the same way they construct actual physical
arrays, the tasks that require imagery should therefore show the same difficulties
as the tasks that require physical manipulation. It is important to see what predic-
tions this theory makes.
Consider the incomplete problem in 76 (analogous to Problem III in Table XI):
76 A is better than B
And C is worse than B.
After reading the first proposition, the subject would set up a visual image with A
above B. The important point in the process, however, comes in the placement of
the third object C. Since B is already in place, C will be easy to place with respect
to B if C is the subject or' ihe second proposition. This conditions hold for 76,
so 76 should be a relatively easy problem to solve. In contrast, consider 77
(analogous to Problem IV in Table XI):
77 B is lower than A,
And B is higher than C.
Reading the first proposition here would result in a visual image of A above B.
But in this case, the third term C, to be placed with respect to B, is in the predicate
of the second proposition and should therefore be difficult to place. The theory
predicts, then, that 76 should be easier than 77, a prediction that is verified in
Table XI.
But this theory makes incorrect predictions for the negative equative problems
analogous to 76 and 77. Consider 78 and 79 (Problems III' and IV', respectively,
in Table XI):
78 A isn't as bad as B,
And C isn't as good as B.
79 B isn't as good as A,
And B isn't as bad as C.
1384 HERBERT H. CLARK

78 has the same properties as 76, since C, the third term to be placed, is the subject
of the second proposition. By Huttenlocher's theory and because the subject term is
easier to place in negative equatives as well as comparatives (d. Clark and Peter-
son's results), 78 should be relatively easy. By contrast, 79 should be hard just as
77 was, since in 79 the third term to be placed (C) is in the predicate of the second
position. But the results in Table XI show that 78 is harder than 79, not easier,
as Huttenlocher's theory predicts. Thus this demonstration (and other data in
Clark 1969a, b) appears to disconfirm Huttenlocher's theory of constructing visual
images.
This argument invalidating Hutten!ocher's hnagery theory was presented in
Clark (1969a, b), but more recently, Huttenlocher, Higgins, Milligan, and Kauff-
man (1970) have replied to that argument, giving more evidence said to favor the
theory of constructing spatial hnages against the invalidating argument just pre-
sented. Although there are too many details to go into here (cf. Clark and Peter-
son, in preparation), I would like to show briefly how this additional evidence does
not save the theory of constructing spatial hnages from the disconfirmation just
made.
The invalidation argument presented in Clark (1969a, b) required only one
assumption in order to test the original Huttenlocher proposal, and that was that
C would be easier to manipulate physically when it was the subject of a negative
equative instruction, as in C isn't as good as B, than when it was in the predicate,
as in B isn't as good as C. Although unverified at that time, that assumption, as.
we saw, has since been ·confirmed by Clark and Peterson (in preparation), so the
invalidation argument of Clark (1969a, b) remains unchanged. In order to coun-
ter this argument, however, Huttenlocher et al. (1970) carried out two pairs of
parallel experhnents. In one experiment, for example, they had their subjects solve
three-term series problems in their heads, and in the parallel experiment, they had
different subjects solve the same problems visually on a felt-board, by placing felt
figures with names on them in a vertical array on the board. The subjects were
timed in both instances. Since. the results of the two experiments were virtually
identical, Huttenlocher et al. argued (1) that the subjects must have gone through
the same mental operations in the first experiment as they did in the second, and
(2) since the subjects were manipulating objects in the second experiment, they
must have been manipulating hnaginary objects in the first.. Both parts of this
conclusion, however, are logically invalid. One could just as easily.<atgue in the
second part, for example, that since the subjects were carrying out the !iilguistic
processes of the linguistic theory in the first experiment, they must have been ·doing
the same in the second. Without any additional evidence, therefore, these data could
just as well support the !iilguistic theory as the imagery theory.
But there is additional evidence, and it shows at least that the subjects do not
behave in a manner consistent with the imagery theory. The 'requisite evidence is
that the subject of negative instructions is easier to place than the predicate, as
 SEMANTICS AND COMPREHENSION 1385
demonstrated in Clark and Peterson (in preparation). The logic is as follows:
(1) Clark and Peterson found that the surface subjects of both the comparative and
the negative equative constructions are easier to place in instructional tasks than
the predicate. (2) Since (1) is true, Huttenlocher's (1968) imagery theory would
predict that problems like 78 should be easier than problems like 79. (3) But Hut-
tenlocher et a!. and Clark (1969a, b) found for both the reasoning and the felt-
board experiments that problems like 78 were harder, not easier, than problems
like 79. Therefore, (4) since the imagery theory makes incorrect predictions in
tlris instance, it must be wrong. In short, the existing data run counter to Rutten-
locher's particular imagery theory, yet are quite consistent with the present theory
of comprehension.

4.6 Comprehension of Comparatives by Children

One surprising source of support for the present analysis of the comparative is to
be found in the literature on comprehension in clrildren. In several previous studies,
children of various ages have been reqnired to understand comparatives in carry-
ing out several types of tasks. It was found that these children persisted in making
several types of mistakes that revealed deficiencies in their understanding of the
adjectives underlying the comparative as well as in their understanding of the
comparative construction itself. The most extensive studies of these phenomena
have been carried out by Donaldson (1963), Donaldson and Balfour (1968), and
Donaldson and Wales (1970). Although I have discussed the relevance of these
studies to adult comprehension in detail elsewhere (Clark 1969a; Clark 1970b), I
will summarize the arguments briefly here.
One of the main facts to be accounted for is that children about three and one-
half years old consistently misunderstand the word less as if it were the word more.
Donaldson and Balfour gave their young subjects two cardboard trees with cardboard
apples on them and asked them 'Wlrich tree has more apples on it?' and 'Wlrich
tree has less apples on it?'. All of the clrildren, except Jor one or so of the more
soplristicated ones, pointed to the tree with more apples on it in answer to both
questions. Similarly, Donaldson and Wales gave their subjects other graded sets
of objects and asked them questions like 'Give me one that's bigger than this one',
'Give me one that's wee-er than tlris one', 'Point to the biggest one', and 'Point to
the wee-est one' (the clrildren, obviously, were Scottish). The findings in tlris type
of questioning were consistent with the more~less results: clrildren correctly inter-
preted the positive adjectives (big, long, thick, high, tall, fat) more often than they
did the implicit negative adjectives (wee, short, thin, low, short, thin), and this was
true when the adjectives were in both the comparative -er and the superlative -est
forms. It is perhaps not too surprising that children should misinterpret just those
forms that adults find more difficult to comprehend, but the more basic problem
1386 HERBERT H. CLARK

is to account for both the adult's aud child's difficulties in a unified theory.
It appears that we cau identify several primitive stages in the comprehension of
comparatives by children. These stages are characterized by the fact that children
have incomplete knowledge of the adjectives aud presuppositions of the compara-
tive. I will outline a hypothetical series of stages that are (1) consistent with the
data that have been collected, yet (2) ordered from the hierarchically simplest in-
formation found in the comparative to the most complex.

4.61 The first stage

Consider Donaldson aud Balfour's comparative This tree has more apples than
that one. Although the children had the superficial appearance of correctly com-
prehending this sentence - they made correct choices when forced to choose the
tree with more apples - there is other evidence that they did not. When Donald-
son aud Balfour presented two equal trees aud asked their subjects 'Does one tree
have more (or less) apples on it thau the other?', some of the children replied with
expressions like 'both of them', 'both the trees', 'that one-does aud that qne', 'they
two ones', 'each tree', and 'these two have', answers which clearly do not make
sense from the adult point of view. These answers, in fact, almost seem to be ans-
wers to a question like 'Does this tree or that tree have SOME apples on it?'; the
answers affirm the fact that both trees have apples on them. ·
This observation leads to the hypothesis that This tree has more (or less) apples
on it than that tree is interpreted at the first stage approximately as ((this tree
has some apples) (that tree has some apples)), paraphrased roughly as 'This tree
has some apples on it and so does that tree'. This interpretation has two important
properties. First, it contains two primitive sentences, just like the comparative, aud
cau be developed into the full comparative at a later age. And second, the com-
parative adjective .more is taken to mean 'some' or 'a quantity of', which is just the
NOMINAL sense• that is superordinate to, or semantically presupposed by, the·adjec-
tives much and little, which themselves underlie more aud less. The first stage inter-
pretation of the comparative, therefbre, contains at least the essentials of the full
interpretation·: it specifies that the two compared objects are on a semantic dimen-
sion, aud it specifies, although only in part, what that dimension is .
.. This description, however, is still not enough to account for why the children in
DonaldSon and Balfour's study were able to interpret more correctly but not less.
Note that they were asked to choose the tree that had more (or less) apples on it.
If they interpreted this to meau 'choose the tree with some apples on it' they could
have chosen either tree; So we have to make the additional assumption that 'some'
or 'a quantity of' is best exemplified by the tree with more apples on it. This is a
reasonable assumption. 'Some' or 'a quantity of' refers to a positive amount along
the quantity dimension, aud the tree with more apples on it is not ouly au example
of such a positive amount, but it is also positive in amount with respect to the
other tree. The lesser amount would be taken as the reference point for the
 SEMANTICS AND COMPREHENSION 1387
quantity dimension perceptually, as certain other studies have shown.
In all, there is a surprising amount of evidence for this primitive first stage in
the acquisition of the comparative.
First, the fact that less is interpreted as if it meant more, and shorter as if it meant
longer, wee-er as if it meant bigger, etc., is consistent with the above characteriza-
tion of the first stage of acquisition. That is, since both more and less would mean
'some' or 'a quantity of', it would appear that less was being interpreted as more,
just as Donaldson and Balfour found.
Second, the elliptical answers 'both of t\lem', 'that one does ·and,that one', and the
like that Donaldson and Balfour observed constitute evidence for the proposal that
the interpretation consists of two underlying propositions. As pointed out above,
these answers could be considered shortened versions of, e.g, 'That one has some
and that one has some', answers that quite explicitly express the two underlying
. propositions.
· Third, Donaldson (1963) and Duthie (1963) found that some young children
interpreted Betty is older tlum May to be the same as May is older than Betty and·
even as Betty is younger than May. For these children, apparently, the sentences
were all synonymous and meant simply something like 'Betty is different in age
from May'. This example again indicates that the children were simply interpreting
the sentences in terms of the nominal sense ('of age') of the words younger and
older. Also, the children showed that they 'knew' the two underlying sentences of the
comparative by taking into account the age of both Betty and May. The one addi-
tional fact that this example shows is that Betty is younger than May did not mean
for these children simply 'Betty is so old and May is so old'; the interpretation had
the additional constraint that the ages had to be different, although it did not mat-
ter in which direction the difference lay.
Fourth, in his study of how children solved three-term series problems with
quantified comparative sentences (e.g. Tom is four years younger than Dick),
Duthie (1963) found that some of the youngest children treated the underlying
adjective in its nominal sense quite explicitly. One child, for example, was asked
in the middle of a problem how he knew that Tom was font; he replied, 'Because
it says that Tom is four years younger than Dicl<' (p. 237). Other children made
the same error with 'four years older' and other similar constructions. Both younger
and older appear to have taken on the nominal meaning 'of age' for these children.

4.62 The second stage

The second stage is a transitional oue in which the comparative Betty is younger
than May is interpreted approximately as ((Betty is young) (May is young)), which
can be paraphrased as 'Betty and May are young but (perhaps) different in age'.
The children at this stage will have passed from the point at which they interpreted
younger and older simply to mean 'of age' to the point at which they 'know' that
younger and older indicate something about the age of Betty and May relative to
 1388 HERBERT H. CLARK

some standard age; Although the evidence for this transitional stage is still less
adequate than one would like, it can be seen in the following illustrations.
First, Piaget (1921) studied the mistakes that nine- and ten-year-olds made in
solving the following three-term series problem (from Burt 1919): Edith is fairer
than Suzanne; Edith is darker than Lili. Which of the three has the darkest hair?
To quote Piaget (1928): 'It is as though [the child] reasoned as follows: Edith is
fairer than Suzanne so they are both fair, and Edith is darker than Lili so they are
both dark. Therefore Lili is dark, Suzanne is fair, and Edith is between the two.
1n other words, owing to the interplay 9f the relatiqns included in the test, the
child, by substituting the judgment of membership (Edith and Suzanne are "fair",
·etc.) for the judgment of relation (Edith is "fairer than" Suzanne), comes to a con-
clusion which is exactly opposite of ours.' (p. 87). Piaget's interpretation of these
mistakes is consistent with the proposal that children in this transitional stage
interpret Edith is fairer than Suzanne as ((Edith is fair) (Suzanne is fair)), a notation
that specifies 'judgment of membership' without indicating 'judgment of relation'.
Second, Donaldson (1963) studied the mistakes children made in solving the
problem: Dick is shorter than Tom. Dick is taller than John. Which of these three
boys is tallest? Like Piaget, Donaldson found that some children had great diffi-
culties putting the information from two premises together. One main problem
for Donaldson's subjects, however, was that the 'judgment of membership' - the
judgment that Dick is shorter than Tom implies simply that Dick and Tom are
short - induced them to think that there were four people instead of three. Since
Dick was short in the first premise, buttall in the second, and since these two facts
are incompatible, many children simply assumed that there must be two Dicks, a
tall one and a short one. For example, one. young girl's solution to the above prob-
lem was, 'This Dick [second premise] is tallest, John is next tallest, Tom is third
and then it's Dick [first premise]' (p. 131). In this example, the child could grasp
the comparative information within premises, but she could not put information
from the two premises together because ofthe disparity in the underlying adjec-
tives. The children in Donaldson's study made the two-Dick error in all types of
problems, but most often in those problems in which Dick appeared in two com-
paratives with different underlying adjectives (like the above problem).
Third, ·Some of the children in Donaldson's study verbalized the underlying pre-
suppositions of the problem directly. One .child, for example, seemed to be unable
to interpret comparative information one moment, but quite able to the next. This
child was quoted as saying, 'It says that DiCk is shorter than Tom, so Dick is short
and Tom is short too'. But in the next sentence she continued, 'And Dick is taller
than John so Dick is tall and John is short' (p. 131). In the first sentence the child
is expressing the underlying presuppositions of the sentence, and this is consistent
with the proposed representation of the comparative at this transitional second
stage; in the second sentence, the child is expressing the comparative information
in the only way she knows how, with the non-comparative adjectives tall and short.
 SEMANTICS AND COMPREHENSION 1389
It should be noted that children who are at this stage can appear to answer
questions like Which is taller? and Which is shorter? correctly without a full under-
standing of the comparative information. Since Which is taller?, for example,
presupposes tallness, and this is something that the child knows at this stage, the
child can simply pick out the tall object as opposed to the short one in answer
to the question; this action would be consonant with his primitive interpretation of
the comparative, and would also appear to be correct. Problems of the kind Piaget
referred to are required before one can see that the child has not grasped the full
meaning of the comparative. '

4.63 The third stage

In the third and final stage in the acquisition of the comparative, the child is able
to appreciate the complete meaning of the comparative, treating Betty is older than
May as ((Betty is old+) (May is old)), just as an adult would.
In the processing of comparatives by adults, example~ have been given to demon-
strate that subjects must manipulate the underlying presuppositions of the com-
parative and question into congruence before the problem can be solved. Some of
Donaldson's children, in fact, made this manipulation quite explicit, One boy was
given the problem: Tom is taller than Dick. Dick is taller than John. Which of
these three boys is shortest? The boy promptly explained, 'This means Dick is
shorter than Tom, John is shorter than Dick. So that gives the answer- it's John'
(p, 121). To make the premises congruent with the question Who is shortest?, he
explicitly rephrased both of the premises in terms of shorter. When asked why he
did this, he answered, 'I thought it would help'. Donaldson reports that many other
children did this too.
In summary, the acquisition literature suggests that children comprehend the
comparative in much the way that adults do. The children appear to treat the com-
parative as an amalgam of two sentences and take into account the underlying
adjectival dimension of the ,comparative. The difference between children and
adults is that children appear to have less cmnplete knowledge of the compara-
tive. The youngest children; for example, appear to interpret the underlying adjec-
.·\
tive in only its 'nominal' sense,a.ndwhile this sense identifies the semantic dimen-
sion on which the two compared objects are found, it· does not contain any' com-
.i parative information. The children .who. are slightly older appear to 'interpret the
iurderlying adjectives contrastively (e.g. young vs. old), but this is still not enough
i for true comparisons. This child must be fairly old (perhaps as much as ten years
I old) before he is able to grasp all the subtleties of the .adult interpretation of the
. I comparative. As a note of caution, however, it must be pointed out that before
I these (or any other) hypothetical stages can be taken seriously, much more sys-
tematic investigatibn is needed of the comprehension of comparatives by'children.
I
1390 HERBERT H. CLARK

4.7 Summary

In this section, I have presented a model of how people represent the comparative
(John is better than Pete) and equative (John is as good as Pete) constructions and
have tried to demonstrate how these representations would be used within a general
theory of comprehension to account for several phenomena in the comprehension
o:l' comparatives. The three principal properties of these representations were (1)
that they contain the two objects being compared, (2) that they specify the under-
lying semantic dimension on which the two objects were presupposed to be com-
pared, and (3) that they indicate that the second of these two compared objects is
the presupposed point of reference. The previous results on answering questions
were shown to be accounted for by assuming that the subject had to make the
premise and question congruent in their presuppositions before the question could
be answered. Similar constraints were required to account for the previous results
on paraphrasing and sentence verification. It was in the tasks that required subjects
to follow instructions that the point of reference of the ·comparative was important.
If the point of reference of the task to be carried out was different from the point
of reference described in the comparative instruction, then the instruction was
difficult to follow.
A second issue discussed in this section was the relative difficulty of implicitly
negative or marked comparative adjectives (e.g. worse) over their positive counter-
parts (e.g. better). A large number of such positive-negative pairs were discovered
in the comprehension literature and were shown to fit what I have called the prin-
ciple of lexical marking. Furthermore, the evidence from children's misinterpreta-
tions of comparative adjectives seems particularly. strong on this point: at first
children are correctly able to 'interpret' only the unmarked adjectives, and even
when older, they are more likely to err on marked than unmarked adjectives.

5. ACTIVE AND PASSNE SENTENCES

Among the earliest constructions to be studied in transformational-generative lin-

guistics were the simple active and passive (affirmative declarative) constructions.
In 1957, for example, Chomsky argued that William likes Furness..and Furness is
liked by William express the same underlying proposition and should therefore be
derived from the same underlying source. Chomsky's proposal was that active and
passive sentences are derived from identical kernel sentences; the derivation
of a passive, however, requires a Passive Transformation in addition to certain
obligatory transformations. Since 1957, there has been much linguistic debate on
the derivation of the passive (and its corresponding active, for that matter), and the
debate still continues. In this section, the goal is not to solve all the difficult lin-
guistic problems associated with the passive, but rather to use what is known to
suggest certain properties of the comprehension of actives and passives.
 SEMANTICS AND COMPREHENSION 1391
The active and passive constructions have also had a long history of controversy
within psychology. It all started with Miller (1962) who first proposed a cognitive
theory of comprehension that might now be described as follows: To understand a
sentence, a person must derive the deep structure of that sentence. And to do that,
he must consider the surface structure, work backwards through the transfor-
mations that would produce that surface structure from deep structure, and come
up with the proper deep structure. Thus, an active sentence should take less time
to comprehend than its corresponding passive, since a person can derive the deep
structure of the active relatively directly, but that of the passive only by 'detrans-
forming' the surface structure through the passive transformation. It also follows
I
i that negative passives will take more time than either negatives or passives alone,
since the former require at least one more transformation than either of the latter,
and so on. Miller's hypothesis eventually became known as the derivational com-
plexity hypothesis: sentences that are derivationally more complex should be more
difficult to understand than those that are less complex. The subsequent popular-
ity of the active and passive contrast has been immense: it has been used both by
those trying to confirm or disconfirm the derivational complexity hypothesis and by
those who simply wanted two constructions that apparently differed in difficulty.
Since Miller's original paper, a number of studies have given support to the
derivational complexity hypothesis (e.g. Miller and McKean 1964; Mehler 1963;
Savin and Perchonock 1965), but just as many or perhaps more have raised ques-
tions about it. For example, if active and passive sentences are remembered in the
same form except for a footnote for the passive indicating that it requires an extra
transformation, the pattern of errors in memory for the deep structure subject, verb,
and object of the active and passive ought to be similar; but empirically the patterns
differ considerably (Anderson 1963; Coleman 1964). Another prediction of this
hypothesis is that the deep propositions of the sentence are retained in memory,
and the footnotes specifying how to transform them into particular surface struc-
tures are lost independently of those deep propositions; evidence to the contrary
has been found by Fillenbanm (1966, 1968a, b), Sachs (1967), Clark and Clark
(1968), Clark and Stafford (1969), Clark and Card (1969), and others. In com-
prehension, Gough (1965, 1966), and Slobin (1966), for example, have argued that
the derivational complexity hypothesis cannot explain a number of the consistent
differences. in verification times that they found with active, passive, positive, and
negative sentences. More recently, Fodor and Garrett (1966) and Watt (1970)
have argued from very general evidence that the derivational complexity hypothesis
just cannot be correct. For example, Watt (1970) pointed out that this hypothesis
.i would predict For anyone to please Dee is hard should be easier to comprehend
! than Dee is hard to please, a prediction, like many many other similar ones, that
is surely false.
The particnlar point that many psychologists have worried about is that active
and passive sentences emphasize different things, and this seems contrary to .the
1392 HERBERT H. CLARK

derivational complexity hypothesis. Jeffrey sired Rhodri seems to 'emphasize'

Jeffrey, whereas Rhodri was sired by Jeffrey seems to emphasize Rhodri instead.
The difference has been called one of 'theme' (Clark 1965; Clark and Begun 1968),
'emphasis' (Johnson-Laird 1968a, b), 'salience' or 'focus of attention' (Turner and
Rommetveit 1968; Tannenbaum and Williams 1968), and so on. Although the
phenomenon alluded to by these various expressions has received little attention
from transformational linguists, they have in some excellent studies by Firbas
(1964), Halliday (1967), and other linguists. Yet a complete psychological account
of active and passive sentences and their comprehensiO!l will have to' deal with this
problem, particularly since it seems to take on innnediate importance in many
psychological tasks.
In spite of the differences between actives and passives, however, one fact r~
mains: people will' judge Jeffrey sired Rhodri and Rhodri was sired by Jeffrey
to be synonymous in some sense. Later in this section, we will examine several
studies that seem.to show that whatever else is wrong, with the derivational coii1•
plexity hypothesis, it is correct in claiming that the propositions expressed by an
active and its corresponding passive are the same. And this fact is in full accord
with_ intuition. The part of the derivational complexity hypothesis that fails to hold
up is the identification of comprehension difficulty with derivational, complexity -
with number of transformations intervening between surface and deep structure.
So the analysis that is about to be offered will retain the requirementthat the pro-
positional content of actives and passives be the same, while discarding the re-
quirement that the process of deriving deep structure be the inverse of a linguistic
derivation of the comprehended sentence.

5.1 The Representation Problem

The first requirement for the 'representations of actives and passives, then, is that
they indicate, the same propositional structure., Thus, we might represent the pro•
positi<inal content of Vivienne saves money and Money is saved by Vivienne by 80:
, 80 {Vivienne save money)
In a more elaborate notational- scheme,_ of course, we would want to represent the
actual functional relations or ca~es (a Ia Fillmore 1968) of the three- words, indi-
cating that Vivienne is the agent, money is the qbject, and save is the, verb. We
might also want to indicate that Vivienne is the subject and save, money is the
,predicate, although psychologically,this will be seen_ to--have less importance than
the case relations. So it shou,Id be remembered simply that the three ,positions
within the parentheses standfor Agent, Verb, and Object (or Recipient)•
English, however, contains very few true transitive verbs, and in fact save' is, not
one of them. Vivienne saves money is really a causative construction- (Chomsky
1965; Lyons 1968) that can be paraphrased approximately as Vivienne causes
money to be safe ~ that is, as- an embeddedsentence something like 81:' " ,
 SEMANTICS AND COMPREHENSION 1393
81 (Vivienne causes (money be safe))
English causatives will always look something like 81, with l! cause (actually ao
abstracLverb closely related to the word cause) as the embedding verb and with
one or aoother type of embedded sentence. Sets 82 through 88 all show how the
causatives in a, paraphrased approximately as b, contain the simpler sentence in c.
82 a) John mbved the car.
b) John caused the car to move.
c) The car moved.
83 a) Fats killed the roach.
·b) Fats caused the roach to die.
c) The roach died.
84 a) Tommy crated the apples•
. b) Tommy.caused the apples to be in a crate.
c) The apples were in a crate.
85 a) The Navy enlisted the men.
b) The Navy ·eal1sed the men to be on a list.
c) The men were on a list.
86 a) Helen enlarged the hole.
b) Helen·caused the hole to be larger.
.c) The hole was larger.
87 a) Charlotte cleaned the windows.
b) Charlotte ·caused the windows to be clean.
c) The windows were cleao.
88 a) Horace taught Rogers how to drive.
b) Horace caused Rogers to learn how to drive.
c) Rogers learned how to drive.
It is evident from these illustrations that causatives cao be of maoy types. The
embedded c sentences cao be simple intransitives, as in 82 and 83; locatives as in
84 .aod .85; simple attributives, i.e. ·sentences with predicate adjectives, as in 86
aod ·87; or even transitive sentences, as in 88. Furthermore, although the causa-
tive verbis often identical or similar phonetically to its embedded intraositive verb,
as with move-move, lay-lie, set-sit, etc., it sometimes is not,. as with kill-4ie in 83
aod teach-learn in 88. Once alLthe causative verbs have been removed from the list
of what appear to be traositive verbs i1;1 English, only a few 'true' •traositives remain,
e.g ... hit as •in· Melvin hit the ball;: But even these 'trne' traositives are suspect, for
they too' seem til'indicate ao. agentive or causative function of the subject; it is just
that trese verbs do.. not seem to have corresponding intransitives in English. When
Fillmore (1968) calls Melvin in Melvin hit the ball Agentive, he appears to be claim-
.ing. that all proper traositives have this causative fmrction in common. But what-
ever the linguistic theory turns out to 'look like,. traositive sentences (excluding
those with instrumental subjects, like The knife cut the butter; ·those with receptive
subjects like John saw the fight, aod certain others) cao be said to be causative in
nature.
1394 HERBERT H. CLARK ·

One reason for discussing the causative nature of most transitive sentences is to
give us a way of talking about certain differences between active and passive sen-
tences. Consider the sentence Fats killed the roach. This sentence can be divided
up approximately into a cause and an effect: the cause of the roach's dying is Fats
(or rather some action by Fats), and the effect of this cause was that the roach died.
The sentence Fats killed the roach, roughly, presumes that Fats acted as the cause
of something, and it asserts that the effect was that the roach died. But the corre-
sponding passive sentence The roach was killed by Fats presumes approximately
that something happened to the roach, and it asserts that what happened was that
the roach died and furthermore that Fats was the cause of that death. In other words,
the active and passive differ in what is presumed by the speaker to be known to
the listener: the cause or what was affected.
All this subjective analysis can be firmed up with a few linguistic examples. Con-
sider the question-answer sequences in 89:
89 a) What did Fats do? Fats killed the roach.
b) What did Fats do? The roach was. killed by Fats.
Clearly, 89a is acceptable whereas 89b without contrastive stress is not. On the
other hand, the pattern is reversed for the question-answer sequences in 90:
90 a) What happened to the roach? Fats killed the roach.
b) What happened to the roach? The roach was killed by Fats.
In this case 90b is acceptable, whereas 90a without contrastive stress is not. In
accord with the subjective analysis, the question What did Fats do? presupposes
that Fats effected something, whereas What happened to the roach? presupposes
that the roach was affected by something. The simpler question What happened?,
which presupposes only that something happened, is answered equally well by Fats
killed the roach and by The roach was killed by Fats, as it should be, since the
presupposition that something happened is implied by the respective presupposi-
tions of the active and passive sentences. It should be pointed· out here that What
happened to the roach? can be answered simply by The roach was killed without a
specification of the causative agent, even though such an agent is always implicit
in the passive. The passive can always be agentless. Similarly, the full effect need
not be specified in an answer to What did Fats do?, for the answer could be Fats
ate, a sentence in which the object of the verb is implicit. . Only a few English verbs,.
however, allow the deletion of the. object.
The linguistic evidence just presented; then, must·be taken into account in con-
sidering how people mentally represent the information in active and passive
sentences.. The semantic representations apparently should reflect two pieces of in-
formation: (1) that the propositions underlying the. two constructions are the same;
and (2) that what is asserted by the two is different. A notation that approximates
these two goals for sentences 9la and 92a is shown in 91b and 92b, along with
approximate paraphrases of those representations in 91c and 92c:
 SEMANTICS AND COMPREHENSION 1395
91 a) Fats killed the roach.
b) (Fats did (Fats cause (the roach die)))
c) As for what Fats did, he killed the roach.
92 a) The roach was killed by Fats.
b) ((Fats cause (the roach die)) happened to the roach)
c) As for what happened to the roach, it was killed by Fats.
Both 91b and 92b contain the same underlying proposition (Fats cause (the roach
die)), but the embedding clauses of 91b and 92b are different: the embedding
clause of 91b is meant to indicate that what Fats did is being !l&Serted, while that
of 92b is meant to indicate that what happened to the roach is being asserted.
Moreover, the representations in 91 and 92 express the facts illustrated in 89 and
90 quite ditectly. They also reflect the respective pseudo-cleft sentences for 91a
and 92a- namely, What Fats did was kill the roach and What happened to the
roach was that it was killed by Fats. The particular notation conventions used in
91 and 92 should not be taken as final, but rather as a temporary means for
expressing the important semantic facts that need representing. For the present,
the exact form of the notation is less important than the notions the representations
are meant to express.
Observations very similar to those just made can also be found in the work of
Firbas {1964) and Halliday (1967). Fitbas, for example, pointed out that sentences
tend to open with thematic elements and close with rhematic elements, where
'thematic elements are such as convey facts known from the verbal or situational
context, whereas rhematic elements are such as convey new, unknown facts'. Halli-
day (1967), in adopting this terminology, has spoken simply of theme and rheme,
where in the active sentence the theme is usually the agent, and in the passive the
theme is the object; the theme is the noun phrase presumed known from context.
But this view of theme does ·not seem quite complete enough to express the differ-
ences required for the question-answer sequences in 89 and 90. Under this view of
theme, Fats killed the roach asserts.something about Fats, not about what Fats did;
its paraphrase would be As for Fats, he killed the roach, instead of As for what
Fats did, he killed the roach. Likewise, under this view the passive would be
paraphrased as A·s for the roach, it was killed by Fats, instead of as the more com-
plete Asfor what happened to the roach, it was killed by Fats. Even though theme
does not seem· complete enough for present pmposes, the word 'theme' will never-
theless be a convenient term to use in referring to the surface subjects of the active
and passive sentences and to their presuppositional effects.
With the suggestive representations of 91 and 92 in mind, we can now proceed
to the psychological literature on active and passive sentences. In the sections that
follow, I will take up the topics of semantic judgments - that is, what people
understand the active and passive sentences to mean - sentence verification, and
question answering, followed by a discussion of certain universals of word order in
transitive sentences.
 1396 HERBERT H. CLARK

5.2 Semantic Judgments

One method for studying how people ·understand or represent active and passive
sentences is to ask people what snch sentences mean. Rather than ask subjects this
question directly, however, the psychologist would normally ask subjects to make a
comparative judgment of some sort - how similar are these two sentences in
meaning? how sensible is this sentence? which of these two situations is better
described by this sentence? which of these two sentences is more acceptable? and
so on. Some of these methods are designed to bring qut the gross similarities be-
tween two different constructions, and others, their subtle differences. So it shoulq
be kept in mind thaLall the studies to be examined do not reveal every facet of
meaning in active and passive sentences.

5.21 Judgments of similarity and anomaly

The gross similarity between actives and passives is nicely demonstrated in a
study of Clifton and Odom {1966). Thekinterest, among others, was to· measure
the judged similarity between pairs of constructions all derived from the same
underlying proposition - in particular, the simple active, simple passive, negative,
question, negative passive, passive question, negative question, and passive negative
question. The subject was given a proposition in· one syntactic form - e.g. The
house was not built by the carpenter - and was asked to judge its 'similarity' to
the same proposition in all the other seven syntactic ·forms;
The predictions the present analysis makes of these results are based on the two
separate aspects of the representations: first, their identical underlying propositions,
and second, their differing embedding strings. First, actives and their correspond-
ing passives should be judged highly similar, .since they contain the same under-
lying propositions. ·Actives. and passives should be more similar, for example, than
actives and their corresponding negatives, or passives and their corresponding
negatives, since in these latter cases; subjects will rightly .judge a proposition and its
denial to be quite dissimilar in meaning', ·Clifton and Odom's data confirm these
predictions: simple actives and,:passives were· judged more similar than any other
pair of constructions in their study._
The second type of prediction. is derived from .the· differences between. actives
and passives. Note, for example, that the active questionDid Melvin strike Eloise? .
asks for confirmation of what Melvin did,' not of what happened to· Eloise, while
the reverse is true of the passive question Was -Eloise struck by Melvin? That is,
the yes/no questions corresponding to actives and passives. do not query tlie under-
lying •propositions themselves, but rather the strings iri which they are embedded
- i.e. (Melvin did 0) and (0 happened to Eloise), respectively. The active and
passive questions might therefore be paraphrased as.93.and.94:
93 As for what Melvin did, did he strike Eloise?
94 As for what happened to Eloise, was she struck by Melvin?
 SEMANTICS AND COMPREHENSION 1397
The representations of 93 and 94 would be more congruent with positive state-
ments of the same voice rather than the opposite voice, since active statements
assert what active questions ask about, and passive statements assert what passive
questions ask about. In short, active and passive questions should be judged more
similar to simple affirmatives in the same voice than in the opposite voice. This
prediction is verified in Clifton and Odom's results. The analogous predictions,
based on parallel arguments, can be made· to show that actives and passives should
be more similar to negative statements in the same voice,-that active and passive
questions should be more similar to negative statements in the same voice, and so
on. All these predictions are upheld in Clifton and Odom's dat~,
A second sfudy,onjudgments.ofseinantic anomaly.(Clark and Begun 1968) sup-
ports the present analysis in quite· a different way. In that study, subjects were
asked to judge on a scale of 1 to. 7 how sensible or nonsensical certain sentences
were. The sentences were constructed by takit\g perfectly. good active sentences
and substituting words. from other random sentences into them as the agent, verb,
or object, or sometimes·in both the. agent and -verb position, and so on; the sub-
stituted word or words, however, were always. of the same form class as the word
they replaced. The resulting sentences, which ranged from very sensible to very
nonsensical, were judged by the subjects in both their active and passive forms.
If the judged 'sensibleness' of the active and passive sentences is merely a judg-
ment of the sensibleness of their underlying propositions, as we should expect,
then an anomalous active and its passive counterpart should be judged equally
anomalous regardless of voice. The results confirmed this prediction: judgments
of anomaly were made of the underlying proposition aud were not affected by
voice.
But a second part of the Clark and Begun study was more revealing. Consider
The calf ·lassoed the presideiit and The president was la8soed by the calf, two sen-
tences containing the same underlying proposition and judged .equally nonsensical.
In a second part of the study, subjects were asked simply to change one word (a
noun or verb) in each sentence to make the sentence .more sensible. Given the first
sentence, subjects were most likely to change lasso ot 'president, but not calf,
whereas given the second sentence, they were· most likely to.: change lasso or. calf,
but not president. So even tho.ugh both sentences were ·judged as equally none
sensical, subjects.saw·the anomaly as arising out <if different sources: the deep verb
or object contained the: anomaly .in the first sentence, but the deep verb o.r agent
contained tlie anomaly in the. second. Apparently, the subjects accepted the pre-
suppositicms.that·the,calf had done something in the active and that something had
happened to the president in the passive; they accepted the themes of the two
sentences asgiven or known:' As a consequence, the'ysaw the,imomaly arising out
of. what the calf had done and in what had happened to .the president, respectively,
and were loath to alter the themes. These results, then, further support the two
properties of active and passive representations.--'"' their common underlying pro-
positions, and their differing embedding propositions or themes.
1398 HERBERT H. CLARK

5.22 Judgments of 'emphasis' in active and passive sentences

Slightly less direct evidence for the proposed representations of active and pas-
sive sentences is to be found in two studies by Johnson-Laird (1968a, b). His
hypothesis was simply that active sentences lay emphasis on the agent in the under-
lying proposition, whereas passives lay emphasis on the object. In terms of the
representations in 91 and 92, the noun mentioned in the embedding sentence
-the theme of each sentence -is the ·one that is 'emphasized' in both sentences.
In one study, Johnson-Laird gave his subjects two sentences ~ say, Red follows
blue and Blue is followed by red-, two long rectangles on a piece of paper, and
a red and a blue crayon. The subjects were instructed to color in parts of the
rectangles corresponding to the agent and object of each sentence; itt such a way
that another person conld figure out which of the two sentences referred to which
of the two rectangles. The resnlts were just as Johnson-Laird expected. The sub-
jects tended to make the color of the surface subject (the theme) cover more of the
rectangular strip than the color of th!l remaitting term; and this effect was more
pronounced for passives than for actives. In a second study, Johnson-Laird gave
his subjects symmetrical or asymmetrical diagrams of the same kind and 'asked
them to judge the relative appropriateness of several active and passive sentences
for those diagrams. The resnlts in this study confirmed the conclusions drawn for
the first. In short, subjects tend. to emphasize the theme - the term presupposed
to be known - in both actives and passives,

5.23 Definiteness of the theme in active and passive sentences

In agreement with Firbas (1964) and Halliday (1967), the present analysis has
assumed that the theme or surface subject of an active or passive sentence is nor-
mally presupposed by the speaker to be known to the listener. Tills assumption
has several immeOfate implications. First, the theme should normally be a definite
noun phrase - e.g. the roach as opposed to a roach - , since definite noun
phrases require or presuppose previous mention of their referents, whereas indefi-
nite noun phrases imply· that their referents are being introduced for the first time.
Second, the theme shonld often be expressed as a pronoun, since pronouns are
shorthand for previously mentioned or situationally implied ·nouns. Tirlrd, the
theme shonld be brief, since the theme, as a noun phrase already known to the
listener, normally requires few ;jf any qualifying relative clauses or adjectives.
These tluee predictions are nicely demonstrated ·in Svartvik's (1966) statistical
examination of an exteiiSive corpus of-English prose. He found thatthe theme in
passives was most often definite, contained· about half as many words on the
average as the agent, was a pronoun far more ·often than the agent was, and so on.
Although Svartvik did not collect the comparable statistics on the active, it seems
likely that the ·theme ·of the active (normally the agent) wonld have comparable
properties.
Consider definiteness. alone in a ' proposition like baker watched candlestick-
 SEMANTICS AND COMPREHENSION 1399
riU!ker. If only one of the noun phrases is definite, it will normally be made the
theme of the sentence. So when baker is definite and candlestick-maker is indefi-
nite, 95a will seem appropriate, and 95b, though possibly acceptable, will cer-
tainly seem inappropriate:
95 a) The baker watched a candlestick-maker.
b) A candlestick-maker was watched by the baker.
But when candlestick-maker is definite and baker is indefinite, the reverse will be
true, with 96b more appropriate:
96 a) A baker w•.tched the candlestick-maker.
b) The candlestick-maker was watched by a baker. ·
The inappropriateness of 95b and 96a seems even more pronounced when the sen-
tences of 95 and 96 are preceded by contexts that mention the definite noun phrase
and truly motivate its definiteness, or when the definite noun phrases in each sen-
tence are replaced by the pronoun he. Thus, whichever of the agent or object is
definite will normally be made the theme, since it is the noun phrase that is pre-
supposed known from previous context.

5.24 The interpretation of quantifiers in active and passive sentences

Closely related to the properties of definiteness just discussed are certain pro-
perties of indefinite determiners as investigated in a very interesting pair of studies
on quantifiers by Johnson-Laird (1969a, b). These studies were apparently based
on Chomsky's (1957) observation that 'Everyone in the room knows at least two
languages may be true, while the corresponding passive At least two languages are
known by everyone in the room is false' (p. 101). In particular, Johns.on-Laird
noted that in 97,
97 a) Every philosopher has read some books.
b) Every book has been read by some philosopher.
the some is normally taken to mean 'some or other', whereas in 98,
98 a) Some books have been read by every philosopher.
b) Some philosopher has read every book.
the some is normally taken to mean 'some in particnlar'. Whether some is inter-
preted as 'some in particular' or as 'some in general' depends, as Johnson-Laird
noted, not on its function in the underlying proposition, but rather on whether it
is the theme of the sentence or not. To. give these informal observations more
empirical backing, Johnson-Laird devised a technique by which subjects could
indicate non-linguistically what they understood sentences like 97 and 98 to mean.
This experiment supported his informal observations as well as several subsidiary
predictions that need not concern us here. In a second study, Johnson-Laird asked
other subjects to decide whether inferences like the following were true or false:
Every philosopher has read some books; therefore, some books have been read
by every philosopher. Since some or other logically implies some in particular,
but some in particular does NOT imply some or other, the inference '97a therefore
 1400 HERBERT H. CLARK

983' Will normally be judged as true, whereas the inference '98a therefore 97a'
will be judged. as false. In sum, Johnson-Laird demonstrated with normative data
that some is usually interpreted as 'some in particular' when it is the theme of an
active or passive sentence, but as 'some or other' when it occurs as the object of
the active or the agent of the passive.
The property Johnson-Laird has investigated in quantifiers is what Fillmore
(l967b) has called 'specificity'. Indefinite. determiners, according to Fillmore, can
be marked as either +Specific ('some in particular') or --Specific ('some or other'),
One particular place where this feature is necessary is in sentence negation. Con-
sider Some philosopher has read every book. Whenever some is marked as + Spe-
cific, then to indicate that it is false that some·PARTICULAR philosopher· has read
. every book one would have to say Some philosopher hasn't read· every book. But
whenever some is marked as --Specific, then to indicate that iUs. false that some
philosopher OR OTHER has read every book one would have to say No philosopher
has read •every book.· Thus, the feature ±Specific has concrete syntactic con-
sequences besides the obvious differences. •
One very important consequence not discussed by Fillmore is that --Specific
quantifiers cannot· be antecedents for definite pronouns in subsequent clauses:
Consider the sentences in 99: ·
99 a) *Every philosopher has read some books, but· they aren't very
exciting books at alL .. "
b) Every philosopher has read Plato's Republic and Aristotle's Poetics,
but they aren't very exCiting books at all.
c) Some books have been.read..by every philosopher, but they aren't
very exciting books at all.
Note that 99a, in which the some is normally interpreted as --Specific, is de-
cidedly odd; the they in the second clause seems to have no well-defined ante-
cedent. The they is not wtong because it. refers to a noun phrase in the predicate
of the sentence, for 99b is perfectly .acceptable; and it is .not, wrong because it
refers to an indefinite noun phrase,.for 99c is acceptable.t<io. The difference be-
tween 99a and 99c is that the former contains a -Specific smne, while the latter
contains a +Specific some. The definite. pronoun they must apparently have a·
+Specific ·antecedent. This is. further demonstrated in 99d and 99e: ·
99: d) *Few ·books have been read by every philosopher, ·but they aren't
very exciting books at all.
e) A few ·books have been read by every philosopher; but they aren't
very exciting books at .all.
As.• Fillmore pointed mit, few and a fiiw are respectively --Specific and + Spe-
cific; no matter where they occur. If --Specific quantifiers. carniot be antecedeuts.
to definite pronouns, then 99d should be odd, but 99e. acceptable, ·and this agrees
with our judgments of the acceptability of these sentences:
These facts about specificity in the antecedents to. definite .pronouns .en.able us
 SEMANTICS AND COMPREHENSION 1401
to account for Johnson-Laird's results if we assume the semantic representations of
actives and passives given earlier. Consider the paraphrases for the semantic re-
presentations of 97a and 98a, respectively:
100 a) As for what every philosopher has done; he has read some books.
b) As for what has happened to some books, they have been read by
every philosopher.
100a contains notlring to force anything other than a ---Specific interpretation on
some books, whereas 100b does. Since some books in 100b is the antecedent for
the definite pronoun they, some books must be +Specific. These interpretations
of 100a and 100b are in agreement with facts that Johnson"Lair'd has pointed out.
Although the paraphrases in 100 are probably not completely accurate reflections
of the underlying semantic representations of 97a and 98a, they do demonstrate
two points: (1) the Surface subject of a sentence introduces a theme, and for the
rest of the sentence to refer back to tlris theme, it, must be a· wellodefined object;
(2} therefore, if the theme is indefinite, it must be marked as +Specific, for other-
wis0 no definite statement can be made about it. Apparently, indefinite noun
phrases are normally assumed to be non-specific; it is only when the indefioite
noun phrases must be +Specific for thematic or other syntactic reasons that the
indefinite will be marked as +Specific. This informal rule, then, would account
for why some isinterpreted as +Specific in the surface subject position, but not
necessarily so when in other positions.
This explanation, however, leaves two facts unaccounted for; First, people do
not invariably interpret some in the surface subject position as +Specific (cf. John-
soh-Laird's resnlts). The reason for tlris appears to be that people can change the
theme of a sentence ·by non-normal stress patterns. For •example, the some in
Some books have been1ead by every philosopher is normally taken to be +Spe"
cific; but When books is given contrastive stress; the some b'ecomes -Specific. In
the stressed version, some books is no longer what is 'given', so it can be taken as
---Specific. The second fact to be accounted for is why the few in Few books have
been read by everyiphilosopher can still be ---Specific although it is the theme of
the sentence. This appears to contradict the generalization that themes must be
+Specific. But it shonld ·be.remembered that few is a negative (cf. section 2), and it-
appears that negation itself makes few ---Specific.,_, Consider the ordinary negative
John isn't-at home. While its positive couhterpatt•John'•is at home is quite specific
about John's whereabouts, the negative· senteric'e is non-specific. The same ob-
servation can be made of few: The proper paraphrase of Few books have been
read by every Philosopher is probably .something like As·for. what has.happened
to·books,.few-of:them have been read··by every philosopher. In this case, them
refers to-specific books, and few-simply denies that the number of books is many.
Put another way, few is not part of the presupposed theme: it is part of the asser-
tion of the sentence and, as such, it does not need to be. +Specific; .This is in
agreement with. the facts about few and a few discussed in section 2.
1402 HERBERT H. CLARK

In summary, themes must be +Specific, and the semantic representations pro·

posed for the present analysis in some sense 'predict' this fact. It is this property
(that themes must be +Specific) that causes some in the subject position to be
normally interpreted as 'some in particular'. It also accounts for Chomsky's ob-
servation that At least two languages are known by everyone in the room and its
active counterpart are not synonymous, since the indefinite at least two languages
in the passive version is the theme and must be +Specific, whereas in the active
version it is not the theme and is normally interpreted as --specific. This analysis .
can also be extended to account for the greater acceptability of definite over in-
definite themes: definite themes are automatically +Specific and· so are always
acceptable, whereas indefinite themes will only be acceptable when they are inter-
preted as +Specific. The broader generalization appears to be that themes are
more acceptable the more.exactly they refer to some identifiable object; +Definite
themes fit this criterion better than -Definite themes, and +Specific better than
--specific. The present semantic representations for actives and passives encom-
pass these facts quite naturally.

5.3 Sentence Verification

In this section, I tum to two studies directly concerned with the underlying proces-
ses of comprehension in active and passive sentences. In both Gough (1965, 1966)
and Slobin (1966), subjects were required to listen to a sentence, view a picture,
and then judge as quickly as possible whether the sentence was true or false of the
picture. As usual, these verification tasks ·can be divided into four stages: a sen-
tence encoding stage, a picture encoding stage, a comparison stage in which the sen-
tence and picture representations are compared, and a response stage. And verifi-
cation· latencies can be predicted from a full specification of these four stages.
Unfortunately, without more data than are found in these two studies, it is impos-
sible to specify these stages as completely as necessary. So I will present two alter-
native ways of viewing the process, Scheme I and Scheme II, both of which are
consistent with .the data of• Gough and Slobin. More evidence will be required to
choose between the two or to reject these two for a third.
To understand Scheme I, consider Gough's (1965) task in which the subject
is presented, say, either The boy 1lit the girl or The girl was hit by the boy along
with a picture of a boy or girl either hitting or kicking a boy or a girl. At Stage 1,
it is assumed that The boy hit the girl is represented as {the boy did (the boy hit
the girl)), and The girl was hit by the boy as ((the boy hit the girl) happened to the
girl). At Stage 2, it is assumed that the subject sets up an active-like representation
of the picture - one containing (someone did 0) - when he has just read an
active sentence and a passive•like one - one containing (0 happened to some-
one)- when he has just read a passive. In other words, the subject knows that he
 SEMANTICS AND COMPREHENSION 1403
must encode the pictUie with respect to an agent in the former case, but with re"
spect to the recipient of the action in the latter. So, for example, the picture for a
true active sentence would be (the boy did (the boy hit the girT)), and that for a
true passive would be ((the boy hit the girT) happened to the girT). At Stage 3, the
subject would compare the representations of the sentence and picture, and at Stage
4, he would respond appropriately.
The Stage 3 comparison operations in Scheme I would be relatively simple and
almost identical to the first operations of the 'true' method of negation discussed
in section 2. The first comparison operation would check for the identity of the
embedded strings. For true sentences, there would always be a match, and for false
sentences, there would always be a mismatch. For false sentences, it is necessary
to have a second operation that would change the presupposed value of the truth
index true to a new value of false. Since this second operation would consume
time, the false sentences would take longer than the true ones.
Gough's (1965, 1966) latencies can be applied to Scheme I. They are easily
summarized: actives took less time than passives; true sentences took less time than
false ones; and these two effects were independent of each other. The second re-
sult ~ that true sentences should be faster than false ones - follows directly from
Scheme I, but the first result is not so obvious. To explain the active and passive
difference, it must be assumed that (1) actives take less time to represent at Stage 1
than passives; or (2) pictures take less time to represent in an active format than in
a passive format; or (3) both (1) and (2) are the case. A second study of Gough's
(1966) eliminates the first possibility. Whereas in the first study the picture was
presented simultaneously with the first consonant of the last word of the sentence,
in the second study it was presented 3 sec. after the initial consonant of the last
word. In the first study, then, it might perhaps be assumed that the representation
of the sentence was still being constructed after the picture had been presented;
but this is highly implausible in the second study. Nevertheless, actives took less
time than passives in both studies - 90 msec in the first and 115 msec in the
second. So the active-passive difference in Gough's tasks cannot be accounted
for by differences in representation time (although his results do not preclude such
a difference, since he measured latencies from the end of each sentence rather than
the begimring). In conclusion, Scheme I would have to attribute the active-passive
difference to the relative difficulty of representing pictUies in a passive-format.
Now consider Scheme II. It does not differ from Scheme I at Stage 1. But at
Stage 2, it is assumed that the picture is invariably represented in an active-format
· - for example, as (the boy did (the boy hit the girT)). Stage 3 must therefore con-
tain an additional set of comparison operations. The" first operation, say, compares
the embedding strings of the sentence and picture representations. For active sen-·
tences, of course, this. operation will find a match. But for passives, it will find
that (0 happened to the girT) and (the boy did 0) are not identical, so a second
operation will manipulate, say, the second string to make it identical to. the first, in
If
1404 HERBERT H. CLARK

this case by changing (the boy did 0) into a passive format (0 happened to the
gir£); The third and fonrth operations would repeat the comparisons of Scheme I,
checking the identity of the embedded strings and changing the truth value jnst in
case there is a mismatch. As for latencies, the second operation wonld be applied
only in the case of passive sentences, making passives take longer than actives; and
the fourth operation wonld apply in the case of false sentences, making false take
more. time than true. Thus; Scheme II wonld also accurately predict the latencies
of Gough's first two Studies; but in Scheme II; passives would take longer than
actives because of an extra Stage 3 operation.
.In their essentials, however, Schemes I and II both attribute the increased diffi-
culty· of passives to the same thing: the picture must be represented in a passive
format before it can be compared to the representation of a passive sentence, and
representing the picture in this way takes more time. Some possible reasons for
this difference will be taken np in section 5.5.

5.31 ·Encoding strategies

The study by Slobin (1966) contains an important innovation over Gough's
studies of active and passive sentences. Slobin distinguished what he called re-
versible sentences from non-reversible ones. The·boy hit the girl~ one of Gough's
sentences - is reversible since the sentence still makes sense with the agent and
object interchanged- e.g. The girl hit the boy-; but The girl picked the flower
is• non-reversible, since The flower picked the girl is anomalous. Slobin's basic
finding was that whereas actives were faster than passives when they were reversi-
ble, there was no difference between actives and passives when they were non-
reversible. Slobin suggested ·that this was perhaps due to the fact that the agent
and object in the .non-reversibk•sentences can be distinguished semantically and
need no help from the syntax,· hence both active and passive sentences are equally
easy.ito comprehend. But if' we. take this suggestion. to mean that the Stage 1
encoding of non-reversible actives and passives· is· now equivalent, this still does
not. explain· ·why Gough found active-passive differences in his 3-sec. delay
condition. :;; '"
· Slobfu's important resnlt, however;· might be'·ticcounted· for in the picture en-
coding stage. The pictures for reversible sentences will always contain two animate
objects·_,____ boys or girls; for example, in Gough's tasks. Since· this -is the case, the
subject cannot' discriminate-' agent from· recipient iri the pictures without judging
what 'the· action is and which of the two people instigated the action. By this
process· alone, the picture 'could well be coded in an active format more quickly
than. in a passive. format. But ·the picture for. reversible sentences will always con-
tain orie animate and one inanimate object, and so the subject can immediately
discriminate ·agent from recipient- Thus,df the sentence is active, he. can imme-
diately code the picture in terms ofwhanhe agent did, and if the sentence is pas-
sive; he can immediately code what happened to the recipient of the action. To
 SEMANTICS AND COMPREHENSION 1405
account for Slobin's results the two codes in this case would have to take the same
amount of time to construct.
Note that this explanation favors Scheme I over Scheme II. That is, the source
of the active-passive difficulty in Gough's tasks and in Slobin's reversible sentences
derives from the fact that a picture is more easily coded in an active format, by
first perceiving the instigator of the action and then encoding what he did. In the
case of non-reversible sentences, the normal process of encoding the picture can
be bypassed, since the instigator and recipient of the action are both readily per-
ceived and the correct coding can be made relatively quickly in· either case.

5.4 Answering Questions

The psychological literature· contains several excellent examples of question-

answering tasks for active and passive sentences. All these .examples appear to
support the present proposals for the semantic representations of active and passive
sentences.

5.41 The Wright study

The main study to be considered is Wright's (1969), which contains some of the
most convincing: evidence available that is counter to Miller's (1962) original thesis
that actives and . passives are simply represented in the identical form (except for
the footnote specifying voice). Briefly, Wright read her subjects 60 sentences one
at a time; 5 sec. after each sentence, she asked the subjects a question about that
sentence. Half the sentences were active, and half were passive, and half the ques-
tions were active, and half were passive. The questions asked for the agent, the. ob-
ject, or the verb by various means -e.g. Who: followed the girl? By whom was the
boy followed? What did· the boy do? etc. Her interest was in the number of .errors
subjects made on each type of question. Indeed, it is somewhat of a surprise that
subjects should make any errors at all on this task, but Wrightfound errors in about
13% of the answers. And·as she expected, subjects made more errors in answer
to some questions than to others. But before we exaniine these errors, it is profitc
able to see'just whatcthe present theory would predict.•.:'
'Let us consider the· schematic 'active and passive sentences A hit B and B was
hit by A· and the possible qiJestions that could be asked of them. ·These two sen-
tences would be represented· at Stage 1 as 101 and ,102, respectively:
101 (A did (A bit B))
102 ((A bit B) happened to B)
The first two active questions to be considered, Who hit B? and Whom did A hit~,
would be .represented at Stage 2 .by 103 and 104, respectively:
103 (X did (X hit B))
' 104 (A did (A bit X))
 1406 HERBERT H. CLARK

The two passive questions By whom was B hit? and Who was hit by A? are given
in lOS and 106:
lOS ((X hit B) happened to B)
106 ((A hit X) happened to X)
For purposes of the Stage 3 comparison process, it is obvious that the active ques-
tions 103 and 104 are congruent with 101, but incongruent with 102, whereas the
passive questions lOS and 106 are congruent with 102, but incongruent with 101.
So the first prediction is that a question in the same voice as the sentence should
be easier to answer than one in a different voice, since in the latter incongruent
cases, the representations of the sentence and question must be manipulated in
some (unknown) way to make them congruent.
But this simple principle - that same voice is easy and different voice is hard
- will not work for questions asked of the verb. The two active questions What
did A do? and What ·happened to B? can be represented as in 107 and 108:
107 (A did (X))
108 ((X) happened to B)
The two passive questions What was done by A? and What was done to B? are
perhaps best represented as 109 and 110:
109 (A did (X))
110 ((X) happened to B)
Representations 107 and 109 were made identical since it is difficult to differen-
tiate what is asked for in What did A do? and What was done by A?; the second
question, for example, does not seem to imply that there was necessarily a recipient
of the action (as the normal passive would imply), and both queStions are answered
acceptably by the simple sentence He ran. The questions What happened to B?
and What was done to B?, on the other hand, both necessarily imply a recipient of
an action and so are represented in the same way, as 108 and 110, even though the
first question allows a broader range of answers - e.g. He died, He fell, etc. -
which do not contain .agents. Although there is some uncertainty about represen-
hitions 107 through 110, they seem to be approximately correct.
The representations in 107 through 110, however, are either directly congruent
or directly incongruent with the sentence representations in 101 and 102: 107 and
109 are congruent with 101, the active sentence,'and 108 and 110 are congruent
with 102, the passive sentence. The Stage 3. .process of replacing the X by the
appropriate subcomponent in the sentence representations should therefore be easy
in the congruent cases, but difficult in the inco11gruent ones.
Wright's (1969) results are shown in Table XIV. The sentence-question se-
quence is given on the left, and the correct answer and percentage of .errors for
each sequence are shown on the right. (Wright, obviously, used a large variety of
full English sentences; A hit B is simply a schematic .example.) Her data confirm
all the predictions made above, and in considerable detail. First, when the voice
of the sentence and question were alike in the who questions, subjects made few
 SEMANTICS AND COMPREHENSION 1407
TABLE XN. Percentage errors from Wright (1969)
Sentence Question Correct Answer Errors

A hit B. Who hitB? A 3%

A hit B. Whom did A hit? B 6
A hit B. By whom was B hit? A 3
A hit B. Who was hit by A? B 19
B was hit by A. WhohitB? A 6
B was hit by A. Whom did A hit? B 6
B was hit by A. By whom was B hit? A. 1
B was hit by A. Who was hit by A? B 7
:1
A hit B. What did A do? hitB 7
A hit B. What happened to B\' hit by A 17
A hit B. What was done by A? hitB 8
A hit B. What was done to B? hit by A 10

B was hit by A. What did A do? hltB 14

B was hit by A. What happened to B? hit by A 9
B was hit by A. What was done by A? hitB 16
B was hit by A. What was done to B? hit by A 7

errors; otherwise they made many. To illustrate, the active questions Who hit B?
and Whom did A hit? each elicited fewer errors from A hit B than from B was hit
by A, whereas the reverse was true for each of the passive questions By whom
was B hit? and Who was hit by A? Second, voice made no difference in the what
questions (at the bottom of Table XIV), just as predicted; nevertheless, congruence
of underlying representations did. Thus, What did A do? and What was done by
A?, both of which presuppose that A caused something and ask for an effect, were
easier for A hit B, which presupposes that A caused something and asserts what
i
i that ~ffect was. In contrast, What happened to B? and What was done to B?, both
of which presuppose that B ~as affected by something and ask what the effect and
cause were, were easier for B was hit by A, which makes the same presupposition
~·F and asserts what is asked for. In short, the principle of congruence predicts the
main differences in these data quite nicely.
Wright's data show several other striking effects. Perhaps the most obvious one
is that subjects were able to answer with the agent mor.e accurately than with the
I
i
(deep structure) object, aiJ.d ,this effect is quite independent of the voice of the sen-
tence and question. This generalization, of course, cannot be made without refer-
' ence to the propositions that un.derlie the active and passive; because of this re-
1408 HERBERT H. CLARK

quirement, the correct representations for actives and passives must specify these
propositions in the identical form. The present proposal conforms to thls require-
ment, since (A hit B) underlies both the active and passive. Although there is no
ready explanation for the agent-object difference at present, it must ultimately be
accounted for by the Stage 3 comparison operations. For example, Stage 3 might
first contrive to make the (A did 0) and (0 happened to B) embedding strings of
the sentence and question congruent, followed by the next strings in, strings of the
form (A caused 0), followed by the most embedded strings, like (B received a
blow) or however the innermost string would be represented. In this scheme; the
agent would be processed sooner than the object, producing the desired result.
Without further data, however, such a Stage. 3 process is merely speculative. It
should also be noted here that a simpler answer will not do. In the what questions,
asked of the verbs, it was no easier, overall, to express what the agent did than
what happened to the object, so it is not simply that the agent and what he. did is
'foremost' in the subject's mind with respect to the object and what happened to it.
The explanation of the ,~gent-object effect will have to be closely tied to the form
of the questions.
A second striking result of Wright's experiments is that actives are no easier than
passives in a task of thls type. If anythlng, Wright's data show that pas.sives are
slightly easier. Most previous studies have shown that .actives are easi~r than pas-
sives, but it appears, from the vantage point of Wright's experimen}, that the pre-
vious studies have given unfair advantage to the active, Gough'~ tasks (1965;
1966), for example, made it difficult to differentiate the agent from the recipient in
the pictures; since humans were used as both agent and recipient, agentiveness was
perhaps prominent, giving advantage to the active. In Slobin (1966), this problem
was eliminated in non-reversible sentences, and passives were just as easy as ac-
tives. It is important to note ·that Wright's sentences were reversible, and yet she
still found that passives were rio harder than actives. This fact reinforces the pre-
vious suggestion that the difficulty of the passive over the active voice in reversible
sentences comes not at the sentence encoding stage in Gough's arid Slobin's expe,ri;;·
ments, but rather at the picture encoding stage. One could conclude that actives
and passives each have their own important place in language, and when the proper
conditions prevail, actives aie easier than passiVes, or P~sives ate easier- thru;t ·ac..:
lives. It is jnst that actives are probably appropriate in' a wider range or more
common set of contexts.

5.42 Transitives and locatives .

In section 2 we exanun.ed 'two studies on answering questions of actiye and pas-
sive sentences -namely, Smith·and McMahon's(1970) ruid Hntteillociier's (1968)
studies. Although the transitive sentences of these stUdies 'were analyzed as if iliey
contained implicit locatives, they coutd alscr be viewed' as true actives and passives.
What is important is that the locative and transitive analyses of these sentences
 SEMANTICS. AND COMPREHENSION 1409
yield the same predictions and conclusions. It is worthwhile to examine these sen-
tences as transitives and to compare. this analysis with the previous locative analysis.
As usual, A is leading B and B is led by A (as two examples) will be represented
at Stage 1 by 111 and 112, respectively:
111 {A did {A lead B))
112 {{A lead B) happened to B)
The problem is how to represent the questions Who is ahead? and Who is behind?
so that they could ever be made congruenCwith 111 and 112. .The most obvious
solution is simply to treat them as actual transitives comparable to Who is leading?
and Who is following? That is, Who is qhead? and Who is behind? would be re-
presented something like 113 and 114, respectively:
113 (X did (X lead y))
114 (X did (X follow y))
The words lead and follow should not be taken literally in 113 and 114; they are
rather surface representations of ·the notion of transitivity combined with the mean-
ings of ahead and behind, respectively.
Once this notation has been adopted, the congruence of sentence and question
becomes obvious. Let us examine the four possible premise-question sequences.
First, question 113 is completely congruent with premise 111, so the sequence 111
+ 113 should be the easiest one to answer. Second, question 113 is not congruent
with 112 in the embedding strings, so to answer sequence 112 .+ 113, the subject
would first have to change premise 112 from ({A lead B) happened to B) to (A did
(A lead B)), or perform some other equivalent alteration. This second sequence
should therefore be harder than the first. Question 114 is congruent with neither
111 nor 112. So third, to answer the sequence 111 + 114, the subject must simply
change around premise 111·from (A did (A lead B)) to (B did (B follow A)), or
something equivalent. And fourth, to answer the sequence 112 + 114, the subject
must change the premise even more radi'cally; for example, he might change 112
from ((A lead B) happened to B) to (A did (A lead B)) to (B did (B follow A)), or
something equivalent. That is, congruenee' is achieved only after a double trans-·
formation. In sum, the first sequente should be easiest, the second and third of
moderate difficulty, and the fourth the hardest. This analysis implies (1) that ac-
tives should take less time than passives, ·and· (2) that Who is ahead? should take
less time to answer than Who is behind? for premises with lead and precede, but
more time to answer for premises with trail and follow. In agreement with this
analysis, both of these predictions are consistent with Smith and •McMahon's .and
Hutteulocher's data.
To summarize, the transitive analysis makes the same predictions in. these sen-
tences as the previous locative analysis. And indeed it should, for these two analyses
are virtually identical. The main point of ·contact between the two is with the basic
assumption that the transitive (A lead B) and the locative (A ahead of B) are
closely related. The locative analysis assumes that the httter underlies the former,
 1410 HERBERT H. CLARK

while the transitive analysis assumes that the latter can be treated as if it were the
former. There is good evidence, besides that presented previously, that transitives
and locatives are related in just this way. Consider such pairs of sentences as:
A is across B and A crosses B; A is around B and A surrounds B; A is on top
of B and A tops B; etc. Such examples show that the agent of the transitive is
normally equivalent to the first term of the locative, and the recipient to the point
of reference of the locative. The assumption that locatives and transitives cor-
respond in this way seems well justified.

5.5 The Universality of the SO Order

In his discussion of universals in word order, Greenberg (1963) has pointed out
that the functional subject- the agent - normally precedes the object in surface
structure in all the languages of his diverse sample. Although the verb might come
before, between or after the subject and object, the order of the latter - S and
0 - is always the same. But why? The answer, as I will argue in this section, lies
in how people comprehend and produce active and passive sentences and in how
they per7eive the world about them.
Word order is forced on language because of the way people speak. People
cannot utter words in a two or three dimensional array: the array must be oue
dimensional. As a phenomenon intimately related to production, then, word order
should reflect the constraints placed on language merely because oue thing has to
be expressed before another. I will consider two simple constraints placed on word
order by our perceptual surroundings, and then show how these constraints could
be marshalled to explain why the SO order is the normal one in all languages.

5.51 Surface structure and chronological order

The first constraint placed on language is that events must normally be expressed
in chronological order whenever the chronology is not explicitly marked in another
way. Consider a speaker who is describing events as they occur. By necessity, he
must describe the events in chronological order, since the events being described
are by definition occurring in chronological order. But this property of present
tense or simultaneous descriptions appears. to be the basis for the rule that all se-
quences of sentences in the same tense are taken to mean that the events are being
described in chronological order uuless otherwise marked. Consider, for example,
the sequence in 115:
115 Melvin dropped the ball. Melvin ran across the field.
This sequence would normally be interpreted to mean that Melvin dropped the
ball before, not after, he ran. And this constraint also explains why and is not
interpreted simply as a symmetrical ampersand when it conjoins tw.o verbs- as
it normally is when it conjoins two noun phrases. In these instances, it takes on the
meaning 'and subsequently' (cf. Johnson-Laird 1967;1969c; Staal1968) as in 116:
 SEMANTICS AND COMPREHENSION 1411
116 Melvin dropped the ball and ran across the field.
And it is not that this constraint applies to adnlt speech alone. Children first de-
scribe events aronnd them as well as imaginary or past events in simple sentences
in perceived chronological order (E. Clark 1969). Sentence order is often the only
way very yonng children have of indicating temporal order, for their speech lacks
conjnnctions and temporal adverbs, and at first it even lacks tense. Furthermore,
children np to the age of about four take clause order to be the only indicant of tem-
poral order. For example, they interpret ALL the sentences She danced before she
sang, Before she danced she sang, She danced after she sang, and After she danced
she sang to mean that the dancing occnrred before the singing, in spite of the fact
that the second and third sentences are marked to the contrary (E. Clark 1971).
Since chronological order of sentences is therefore of primary importance
to the interpretation of sequential events, it is likely that the following two proper-
ties of English are nniversal: (1) the order of two otherwise unmarked sentences in
snrface strnctnre will normally be taken to be the chronological order of the events
they describe; and (2) the conjunction and will normally take on the meaning 'and
subsequently' - never 'and previously' - when it conjoins two clauses otherwise
unmarked for time sequence.

5.52 Surface structure and cause and effect

The second constraint placed on language is derived from the fact that cause and
effect always occur in that order - cause before effect. This is no place to discuss
just when two temporally related events are also interpreted as causally related, but
certain such sequences are. In these cases, two ordered sentences (again other-
wise unmarked for sequence) or two sentences conjoined by and take on not only
the meaning 'and subsequently', but also the meaning 'and consequently' (cf. Staal
1968). Note the difference l':ietween 117 and 118:
117 a) John fell off his bicycle. He broke his arm.
b) John fell off his bicycle and broke his arm.
118 a) John broke his arm. He fell off his bicycle.
b) John broke his arm and fell off his bicycle.
The interpretation of 117 would normally be that the fall caused the break, but that
of 118 wonld not. English, of course, has other ways of expressing cause and effect,
as in John broke his arm because he fell off the bicycle, but in these constructions
the cause and effect are always explicitly marked by a word like because, since, so,
consequently, etc. Thus, since cause and effect by definition always occur in this
order, the following are probably also linguistic universals: (1) the order of two
otherwise unmarked clauses will normally be taken to indicate cause and effect in
those cases where a causal interpretation is possible; and (2) the word and con-
joining two such clauses will be interpreted as 'and consequently', never as 'be-
cause'-.
1412 HERBERT H. CLARK

5.53 SO order in transitive sentences

Now consider the cause and effect sequences in 119:
119 a) Fats acted. The roach died.
b) Fats acted and .(as a consequence) the roach died.
Interpreted as cause-effeCt sequences, 119a and 119b are equivalent to the simple
transitive sentence in 120:
120 Fats killed the roach.
In other words, 120 is roughly an amalgamation of the two simple intransitive sen-
a
tences in 119a in cause-effect relationship. As pointed .out above, the sequence
of primitive sentences in 119a has one single properly that distinguishes it as a
cause and effect sequence, and that is the surface structure ·order of the two ,primi-
tive sentences Fats acted and The roach died. The most natural way to amalgamate
these two sentences, then, would be to preserve this order while making other alter-
ations. The most primitive way to amalgamate, of course, is simply to conjoin the
two sentences, as in 119b. More complete amalgamations, however, are shown
in 121:
121 a) Fats' actions caused the roach to die.
b) Fats' actions made the roach die.
c) Fats caused the roach to die.
Sentences 121a and 121b still contain the two primitive sentences Fats acted hlld
The roach died in quite explicit form;. sentence.121c drops explicit mention of,Fats'
actions, but retains explicit mention of the roach dying; finally .sentence 120 amal-
gamates cause to die in a single verb kill. In going from 119 through 121 to 120,
one is going from the most productive to the least productive expression of cause
and effect; that is, any cause and effect sequence can be expressed ·as in 119, but
the possibilities are vastly limited when~ one is forced to choose a transitive con-
struction as in 1211: The most important fact about this series of amalgamations,
however, is that the one distinctive property of 119 - the surface order of the
cause and effect - is preserved through .the series. In the final. amalgam 120, the
order of the cause (what Fats did) and the effect (whaLhapperied to the roach)
is preserved in the subject-object order. It is natural that all languages• should like-
wise evolve a special-transitive construction that pre8erves !his· cause-effect order.
And this, of course, would account for ·why·Greenberg. fourid·that the SO order is•
universal as the normal, unmarked order of the· subject and object in transitive
sentences. ·:J ,-
The universality of the SO. order can be viewed in a different, but equivalent
way. 'In perceiving cause-effect sequences; people will normally view• the 'causing
event before the caused event ~ because· that is· by definition the· order in which
they occur: Thus, a simultaneous description:of these events should•mention the
cause before the. effeet. Just as the surface order of tWo• otherWise umnarkedn;en"
tences reflects the chronological order of the two events described by them,. so
should the mention of the causative agent before the recipient of the action reflect
 SEMANTICS AND COMPREHENSION 1413
the chronological order in which the agent and recipient are perceived. This implies
again that the SO order is the normal order, whereas the OS order, as in the pas-
sive, should be .marked, or non-normal.
It is interesting to note that a number of other facts about English transitive sen-
tences might also be derived from the temporal properties of cause and effect This
argmnent, again, depends on the close relationship of time and space spoken of in
section 3. Note that from here to there implies a direction of movement, from the
source here to the goal there. Likewise, movement is also implied~ in temporal
descriptions as in from noon to midnight, which is always taken to mean the time
interval with noon as its beginning not as its end. That is, the fro'm-phrase indicates
the preceding time and the to-phrase, the following time. Now notice the preposi-
tions that are often nsed in transitive sentences. First, there is the to in Melvin
kissed Eloise and he also did it to Irene. The to in the second clause is merely to
indicate that Irene is the recipient of the action - the goal or the end-point in the
causeceffect sequence. Second, there is the out of and from in Julia left out of anger
(meaning Anger caused Julia to leave) and Constance was in great pain from the
accident (meaning The accident hurt Constance greatly). The out of and from in
these examples indicate the causal agent, the beginning of the cause-effect se-
quence. (For many other illustrations of to and from underlying transitive sen-
tences, see Gruber 1965.) These examples of to and from underlying transitive sen-
tences in English are quite consistent with the view that to and from reflect the
chronology of the cause and effect sequence underlying transitives.
We are now in a position to say why it should be easier to encode a picture in an
active format than a passive format (cf. section 5.3). People normally perceive a
cause and effect sequence in its natural chronological order. So a plausible way
to encode a picture that is alleged to show a cause and effect relation is first to
start at the beginning, finding the causal agent and encoding what he did, rather
than to start at the end, fioding the recipient of the action and asking what hap-
pened to him. The former strategy, of course, results in a representation of the
picture in an active format. For non-reversible sentences (like Margaret picked the
flower), however, the recipient can be determined in the picture without identifying
the action at all, so the picture can be encoded with respect to what happened to
the recipient. This might well be the reason why actives and passives do not differ
for non-reversible sentences, but do for reversible sentences.
·.On the one hand, E. Clark (1969, 1971) has shown that children rely
heavily. on order in surface structure as an indicator of chronological order; and
Clark and Clark(1968) have shown that even adults are strongly affected by clause
order in meuiory tasks. In another context Bever (1970) has argued that a primary
strategy for both children and adults is to try always to interpret a noun-verb-noun
sequence first as agent-verb-object, unless there is some indication to the contrary.
One of Bever:s pieces of evidence for this strategy is that children at certain ages
consistently interpret passives like The cow was kissed by the horse to mean The
1414 HERBERT H. CLARK

cow kissed the horse, while they rarely do the reverse. But if transitive sentences
are viewed as an implicit conjunction of cause and effect, and in that order, then
Bever's strategy, if correct, appears to be just a special case of the more general strat-
egy E. Clark (1971) has noted - namely, that children first take surface
order to be identical to chronological order. That is, children interpret noun-verb-
noun sequences as agent-action-object, not as object-action-agent, simply because
the agent and object are in chronological order in the former sequence, but not the
latter. Of course, this explanation is confounded with frequency of occurrence in
English, since actives with agent-action-object order are more common than pas-
sives. But one could argue that actives are more prevalent than passives for just
this reason: it is only the actives that are in agreement with the universal SO order
that is demanded by the two surface order constraints discussed previously.
In sum, it is not at all implausible that certain facts about actives and passives
can be accounted for ultimately by two very general {and probably universal} con-
straints on word order. First, a sequence of umnarked sentences is taken as de-
scribing its constituent events in chronological order. Second, such chronological
sequences can further be interpreted under the appropriate circumstances as cause
and effect. One consequence of these two constraints is that the cause should uni-
versally be described ~fore the effect innormal surface order, thus accounting for
the universality of the SO order in languages of the world. Furthermore, these two
constraints give some account for the strategies children use in interpreting se-
quences of sentences (or clauses) and transitive sentences. This line of reasoning
is just one example of how certain general properties of cognition might be thought
to affect linguistic structure - here specifically surface structure - in a signifi-
cant way.

5.6 Conclusion

In section 5, it has been argued that the comprehension of active and passive sen-
tences can also be viewed as a four stage process, with the principle of congruence
as the guiding rule for the comparison stage. The active and passive sentences Fats
killed the roach and Theroach was killed by Fats, it was seen, have the proposition
(Fats caused (the roach die)) in common, but the active asserts what Fats did and
the passive, what happened to the roach. The f)Ill representations of the active and
passive were therefore given as (Fats did(Fats caused (the roach die))) and ((Fats
caused (the roach die)) happened to the roach), respectively, in ari attempt to repre-
sent both the similarities and differences between the active and passive sentences
in the same format.
In a review of the psychological literature, we saw that people interpret senten-
ces in a :way that is consistent with these representations. Although actives and
passives are judged as similar in meaning, the themes of the two sentences ·are more
 SEMANTICS AND COMPREHENSION 1415
often treated as known, given emphasis, and interpreted as definite or specific.
Next, we saw that the comprehension of actives and passives was also consistent
with these representations. In verification tasks, actives were generally easier thao
passives. The most plausible reason for the difficulty of the passives, it was argued,
is that the verifying pictures - particularly for non-reversible sentences - are
easier to encode in an active rather than passive format. It was in a question-an-
swering task that the proposed representations of actives and passives received
their best support. Here we saw that active and passive questions were most easily
answered when what was queried were active and passive sentepces, respectively.
The exception to this generalization came in questions that asked about the verb;in
this case, questions asking about what the agent did were easier for actives than
for passives, whereas questions asking about what happened to the object were
easier for passives than for actives. So the experimental literature supports the
proposal thatactives and passives express the same proposition, but assert different
things about it.

6. CONCLUSIONS

In section 1 of this chapter, I outlined a general theory of comprehension which

suggested that people represent the meaning of a sentence in an abstract symbolic
form at Stage 1, represent other information in the same format at Stage 2, com-
pare these two representations by a series of match and manipulation operations
at Stage 3, and 'convert' the symbolic Stage 3 outcome into a response at Stage 4.
The most basic assumptions of the theory were (1) that sentences are represented
at Stage 1 in a format similar to their linguistic deep structures, and (2) that the
Stage 3 comparison process inust carry out its operations in conformity with the
principle of congruence. In sections 2 through 5, I examined empirical evidence
on the comprehensibn of negatives, locatives, comparatives, and actives and pas-
sives, and demonstrated how the theory could account for the main findings ·on
latencies and errors in the psychological data.
In this section, I will attempt to evaluate a few of the successes and failures of
this endeavor. On the· plus side, the most obvious success k that the theory can
account for the results of previous studies. Less obvious, however, is the fact that
the theory contains important empirical generalizations about the structure of
'comprehension" tasks and about the semantic coding of all types of .information.
These generalizations will be discussed in section 6.1. On the minus side, the
theory as presented fails to cover certain points, some of which are irrelevant to the
validity of this theory, others of which are more serious. Thes.e will be discussed iu
section 6.2. In the final section, section 6.3, I will suggest several directions the
· study of comprehension might take in the future.
 1416 HERBERT H. CLARK·

6.1 Two Empirical. Generalizations

6.11 Analysis of 'comprehension' tasks

The 'comprehension' tasks examined in this chapter have been of several dif-
ferent types - sentence verification, answering questions, following instructions,
and paraphrasing, to mention only the most prominent. The heuristic that has been
used to tie all o€. these tasks together is the four-stage schema, by which each task
was decomposed into two representation stages, a comparison stage, and· a re-
sponse stage. In spite of its simple and innocent appearance, this heuristic em-
bodies s~w:ral significant empirical claims about such 'comprehension' tasks. One
clai]n is.·that a sentence is represented at Stage 1 in the same way for all types of
tasks. This generalization is consistent With the fact that the results from a whole
gan>ut of tasks containing a particular construction are accounted for by the use of
· a single Stage 1 semantic representation. Another claim is that the Stage 3 com-
parison operations are guided by the principle of congruence in all types of tasks.
This claim was verified by the consistent finding that 'comprehension' tasks were
carried out more quickly and With fewer errors when the Stage 1 and 2 semantic
representations were congruent than when they were not. Yet another claim is that
reading an instruction and then following it is conceptually equivalent to reading
an instruction, .asking an implicit question about the instruction, answering the ques-
tion;. and then producing an appropriate answer. This claim also was consistent
With. the results of the relevant studies. Indeed, the main .claims of the present
theory ·of comprehension would not have been testable Without the four-stage
schema.

6,12 .Yhe semanticrepresentation of information

One generalization made possible ·by this four-stage schema cis that information
encoded .at Stage 2 is represented in the same format as the semantic representation
at Stage 1 regardless of where that information came from. Included among the
type.s ·of infortnation 'encoded at. Stage 2 were: pictures, previous knowledge re-
trieved from memory, other sentences· (in verification tasks), questions (in question~
answering tasks), and the physical situation (in instruction-following. tasks). The
evidence .for. this generalization was that a number of resnlts.t:ould not have been
accounted·for Without it. Consider the 'true' model of negation. When applied to
. verification tasks in which. pictures, previous knowledge, and. other sentences were
variously· used· as verifying. evidence, the model accounted equally well for each
type of evidence. Without the generalization that all three types of evide.nce are
coded in the same way, it would have ·been necessary to construct three different,
but: partly overlapping models of negation, rnlili.g out the 'true' modelof negation
in its. general.form. This is dearly an undesirable consequence, and the assumption
that all Stage 2 representations are ·semantic in form·is .a natural solution to this
problem.
 SEMANTICS AND COMPREHENSION 1417
It is important to point out that other generalizations could have been possible.
Whereas the Stage 1 and 2 representations were assumed to be 'semantic' or 'inter-
pretive' in the present theory, they could have been pictorial or acoustic in form
instead. Thus, it is possible to imagine a theory of comprehension in which the
sentence is stored in memory acoustically, and other information is first converted
into an acoustic image and compared piece by piece with the image of the sentence.
For most tasks, however, alternatives like thls can be ruled out as implausible a
priori. When a subject is asked to decide, for example, whether Star is above plus
is true or false of a picture of a star (*) above a plus ( + ), he is ,really being asked
whether his 'interpretation' of the sentence is equivalent in meaning to his 'inter-
pretation' of the picture. He could answer 'no' on a number of grounds: e.g. he
might think that star cannot be used to refer to the asterisk, that the picture is
upside down, or even that the sentence means 'the star is above the surface of the
paper on which the plus is printed'. Clearly, it is the meaning of both sentence and
picture that counts. So it is natural that the comparison process should operate
on the comparable 'interpretations' of sentence and picture, as it does in the present
theory. This generalization is discussed more fully in Clark and Chase (1972),
where it is further supported by empirical evidence.

6.2 Problems to be Solved

The present theory of comprehension, as proposed here, contains a number of

theoretical gaps. Some of these gaps are deliberate and relatively unimportant to
the validity of the present theory, while others reflect fundamental questions.

6.21 Deliberate omissions '·

6.211 The derivation of semantic representations. Some psychologists might
argue that the theory I have presented is not a theory of 'comprehension' at all, but
rather a theory of 'sentence verification', 'deductive reasoning', or the like. They
would point out that thls theory does not specify the process many people would
call 'comprehension' - Le. the process by which people 'derive' semantic represen-
tations from the words and phrases of surface structure. Obviously, I have not
attempted to account for this part of the comprehension process, nor have I in-
tended to. One strength of the present theory is that it can stand without a specifi-
cation of the process by which the semantic representations are derived. The
present theory begins with the hypothetical semantic representation of a sentence,
and the only assmnption required' is that such a representation can be derived from
·-.-·-. surface structure. How it is derived affects no part of the theory as presently stated.
The process of how people derive semantic representations from surface struc-
ture, though extremely important, is very difficult to study - witness a number
of current attemptsAo experiment on this aspect of comprehension. More signifi-
1418 HERBERT H. CLARK

cantly, however, the present theory could be of considerable help to these current
theories. The argument is as follows: Before one can develop a theory of deriving
semantic representations from surface structure, one must be able to specify in
some detail just what the semantic representations should look like. The latter
question is most easily answered by studying how sentences behave in tasks of the
kind reviewed in the present chapter. That is, the present theory specifies the form
the semantic representations of sentences take in working memory, and if correct,
the form of these representations should then place powerful constraints on pro-
cesses alleged to derive them from surface structure. , Indeed, the present theory
appears to conform quite well to several current proposals about how semantic
representations are derived from surface structure, proposals too extensive to re-
view here (cf. e.g. Watt 1970; Bever .1970).
As to the charge that the present theory is one of 'deductive reasoning' rather
than 'comprehension', it matters very little how one refers to the theory. 'Com-
prehension' is an ambiguous term: it can refer either to the process of comprehen-
sion - as in Pierre's comprehension of English is slow - or to the end product
of comprehension -.as in Pierre's comprehension of English is poor. Although
this chapter is also about 'deduction' in the most general sense of that word, the
basic concern of the chapter is with 'comprehension' in the latter sense.
6.212 The derivation of stage 2 representations. Another topic deliberately
ignored was the theory of how people derive Stage 2 semantic representations from
pictures, previous information, or the physical context of the task. Like the process
of deriving Stage 1 representations, the process of deriving Stage 2 represen-
tations does not have to be specified for the present theory to be valid. Here again,
one could argue that the most efficient strategy for studying this derivation process
is to learn as much as possible about the end product of this process - the Stage 2
representations -'and then work backwards to see what operations could derive
such a representation from the pictures, previous knowledge, or the like. The
present theory, if correct, puts constraints onthe form such derivational processes
may take.
This is not to say that there are no important problems to be solved with regard
to the derivation process. Consider a priori information, like the oddness and
evenness of numbers. The Wason (1961) experiments could be accounted for, it was
argued, if it was assumed that his subjects encoded the information about numbers,
e.g. Eight is even, at Stage 2 in a positive form. It seems reasonable that information
of this sort should be represented in a positive form, since in more general examples
it would take many negative propositions to specify the information of a single posi-
tive proposition. But it seems highly unlikely that Wason's subjects had stored
intact propositions like Eight is even in permanent memory. This information
might have been stored, for example, as a rule: if· the number is a multiple of two,
it is even, and otherwise it is odd. The retrieval of Eight is even is therefore almost
certainly a derivative process, even though the evidence seems to show that the
 SEMANTICS AND COMPREHENSION 1419
final representation of this information is in the form of the semantic representation
underlying Eight is even. In more complex cases - e.g. Flores d'Arcais's sentence
Lions are more ferocious than sheep - , the problem of what is stored and how it
is retrieved for representation at Stage 2 is compounded; the evidence requrred for
verification seems even more derivative here than in Wason's task. The problem
of how 'interpreted' information is stored in and retrieved from permanent memory,
it appears, is much farther from solution than the problems covered in the present
chapter. And it will not be solved without extensive, careful empirical investiga-
tions.
6.213 The semantic representations of lexical items. Another problem bypassed
in the present theory is the question of how most lexical items are represented. In
a few instances, specific proposals were made about the representations of words
- for example, in the feature notations for above and below and umnarked and
marked comparative adjectives (e.g. better and worse), and in the 'propositional'
decomposition of absent. In most instances, however, such a detailed specification
was not required. It did not matter, for instance, how star and plus were represented
in the Chase and Clark experiments as long as their Stage 1 representation times
were independent of the remaining latencies, as was found. The representations of
most lexical items are still too difficult to specify, and they add little to the main
processes discussed in the present theory.
There are, nevertheless, important exceptions to this last statement. The discus-
sion of the comprehension of sentences like John is absent gives just a hint of how
important the representation of a single lexical item can be. That absent can be
decomposed into not and present suggests that many other lexical items are actuaily
a complex amalgam of primitive propositions, each embedded within another in a
very specific structure. As !.will suggest below, one extension of the present theory
will be to study the semantic representations of such lexical items in much greater
detail.

6.22 Problems of central interest

In the present theory, there are also several fundamental issues that are more
centralto its validity. Although some of these have been hioted at before, I will
discuss each of them io more detail here. ·
6.221 The. semantic representations of sentences. In this chapter, I have made
use of a particular format to specify the semantic representations of the sentences.
In an important sense, this notation is central to the theory, for it reflects the prop-
erties the semantic representations are claimed to have. For the most part, how-
ever, the function of the notation has been more to suggest what these properties
are than to specify them exactly, and this has added a certain amount of impre-
cision to the present chapter. Although unfortunaie, the imprecision seems una-
voidable at the present time. Part of the problem lies in the dependence of this
notation on !inzuistics. Few linguists have tried to specify, for example, how single
1420 HERBERT H. CLARK

lexical items like absent ought to be represented in a 'propositional' notation {al-

though cf. McCawley 1968). Even Jess is known about how presuppositions should
be represented in such a notation {cf. Filhnore 1970). Unfortunately, this chapter
was heavily concerned at times both with the internal structure of lexical items
and with presuppositions; Therefore, before the sometimes makeshift notational
schemes of the present chapter can be improved, we will have to know much more
about (1) the linguistic deep structures of these constructions, and (2) the way
these structures relate to other parts of the theory of comprehension.
Closely tied to the notational problem is the question of how much of the deep
structure of a sentence a person, truly represents. In section 1, it was pointed out
that the deep structure assumption should probably be taken in a 'weak' sense.
People do not represent all the potential deep structure of a sentence each time
they comprehend it, but what they do, represent is consistent with deep structure.
But what do, incomplete or ,impoverished semantic representations look like? The
claim is that they ate 'consistent' with linguistic representations, where 'consistent'
means informally that the inferences a subject will draw from his understanding of
a sentenc,e are ones that would have been drawn from a complete linguistic represen-
tation. ,This requirement, if correct, implies that some possible representations are
'c,onsistent' and some are not, yet the properties of the representations specified by
this requirement have not been exanrined so far. Another related question is: How
is it possible to discover what is and what is not represented of a sentence in a partic-
ular instance? The present theory should be helpful in making decisions about this
point (cf. Clark 1969a: 396 for such an example) unless the semantic representa-
tions change wildly from instance to instance. The answers to both questions,
though qnlte out of reach at the present time, are critical to a complete understand-
ing of the process of comprehension.
6.222 Stage,3 comparison operations. Throughout this chapter, the Stage 3
comparison operations have been formulated subject to several constraints. The
main constraints were (1) that Stage 3 consist of an algorithmic series of mental
operations, {2) that the operations symbolically compare and manipulate the Stage
Land 2 representations, and (3). that the overall processes conform to the principle
of congruence ,,....;. thaNwo representations had to be identical for a deCision to be
made. Stage 3 was used for making predictions of latencies; generally;'the greater
the: number of: Stage 3 mental operations required, the longer the task. took. Al-
though this series of constraints enabled us to predict the direction of most differ-
ences in comprehension time, there are only a few instances - e.g. the, Clark and
Chase study - ,fonwhlch a model has been worked out in any detail. For many
other~ases; plausible models were constructed, leaving most particulars of, Stage 3
unspecified,, To be complete the ,theory will have to specify ,the Stage 3· operations
for:aUJasks., · ,
At first glance, there .seem to be several slight disparities among the specific
models of comt;>rehension that have been presented. First, consider the ordet· oJ.
 SEMANTICS. AND COMPREHENSION 1421
the comparison operations in the Clark and Chase model. Evidence indicated that
the first comparison made in the Stage 1 and 2 strings was of the embedded strings,
and this was followed by a comparison of the embedding strings. The order of
these comparisons might be called 'inside-out'. In contrast, one of the models
proposed for actives and passives contained the order 'outside-in'. Is this a signifi-
cant disparity? Should, the order of comparisons be the same for all types of con-
structions? If the answer to the last question were yes, we could add a powerful
constraint to the theory of comprehension, But there are many other significant
ways of constraining the order of comparison; for example, pre&uppositions might
always be compared before assertions, or more specific information might always
be compared before less specific information, and so on, The problem seems
insoluble with the present paucity of data. A second disparity among the separate
models of comprehension is that the Stage 3 manipulations were viewed as altering
the Stage 1 representation in some cases, but the Stage 2 representation in others.
Is this disparity a critical one? This question too is unanswerable at the present
time,
These disparities bring us to an important issue concerning Stage 3, and that is
how to build more constraints into the comparison operations. Accurate predic-
tions about specific latencies will not be possible without constraints on the order
of the comparison operations, on the possible manipulation operations, and the like.
These constraints will have to be discovered empirically.

6.3 The Future of Comprehension Theory

Despite its problems, the present theory of comprehension holds considerable prom-
ise for future studies in semantics. Two directions in which the theory ought to be
extended are towards the study of single lexical items and towards the study of
more complex constructions unexplored by psychologists,
How the present theory can be utilized to study single lexical items is best illus- .
!rated by a recent study, by Carole Offir and myself, on the presuppositional prop-
erties of the words come and go, Consider the sentence John thought, 'Mary has
just come into the kitchen', one of the sentences we studied, One presupposition
of this sentence is that John must be in the ldtchen. The same sentence with go in
place of come has just the opposite presupposition: John cannot be in the ldtchen.
(These presuppositions are spelled out in detail by Fillmore 1967a). The question
is whether these two presuppositions are represented in approximately this form
when one 'understands' and represents the sentence. If this were so, then the
. i semantic representations of the sentence with come should 'contain' 122, and that
of the sentence with go should 'contain' 123:
122 a) John is in the ldtchen.
"'' (John in ldtchen)
1422 HERBERT H. CLARK

123 a) John isn't in the kitchen.

b) (false (John in kitchen))
If these presuppositions are treated in comprehension exactly as if they were asser-
tions, then problems like John thought, 'Mary has just come into the kitchen',
therefore, John isn't in the kitchen should be verified as follows. At Stage 1, the
subject sets up the presupposition (along with the rest of the representation) (John
in kitchen); at Stage 2, he sets up the conclusion (false (John in kitchen)); at Stage
3, he compares the two representations by the 'true' method of negation; and at
Stage 4, he responds 'false'. With respect to this type, of conclusion, the sentence
with come is a positive, and the sentence with go is· a negative. One could imagine,
of course, other ways of representing (or deducing) the presuppositions of come
and go, although admittedly the model just presented is the simplest and most
appealing. The important point is, however, that the present theory of comprehen-
sion - in this case, the 'true' model of negation - is capable of testing which of
these.altematives is correct. Although the come-go study superficially has the same
appearance as any other investigation of comprehension in sentences, it is' basically
a study of the semantics of come and go as lexical items. As this example shows,
however, the presuppositions (and more generally, the meanings) of come and go
cannot be studied ·without placing the two words in sentential contexts. The pre-
suppositions of come and go must be stated relative to their agentive and locative
cases and relative to the speaker and listener of the utterance: none of these func-
tional 'arguments' is specified for come and go in isolation. The same lesson can
be extended to most, if not all, lexical items.
One does not have to look very far to find other examples of lexical items that
can be investigated by the present techniques. 1n section 2, it was argued empiri-
cally that absent was decomposible into not and present- that is, John is absent
can be represented'as (true (suppose (false (John is present)))). A similar argument
was applied to the quantifiers few, scarcely any, and their ilk, to the prepositions
out of, from, and their ilk, and in other sections to verbs (e.g. lead and follow),
subordinate conjunctions (before and after), and other lexicaLitems. In principle,
the present theory could be applied to words in any lexical category - even to
nouns. In short, the present theory holds much. promise for the study of single
lexical items - especially those which, like come and go, have a very complicated
set of underlying propositions.
The present theory can easily be extended to new construction types as well.
Potential candidates include: relative clauses (e.g. The man who hit John was
struck by Dick), complement structures (John refused, to go home), coordinate
constructions (John neither tripped nor fell), subordinate constructions (John fell
because he tripped), indirect questions (John asked Mary to leave the room), com-
mands (Raise your hand), questions (Who left the house early?), polite requests
(Will you please raise your hand?), and so on. A study of these constructions
should help us define the form semantic representations take in memory and the
 SEMANTICS AND COMPREaENSION 1423
properties intrinsic to the Stage 3 comparison process, as well as many other issues
raised in this chapter. But this future work seems particularly important, for it
could well hold a key that both psychologists and linguists could use to unlock the
secrets of language.

REFERENCES

ANDERSON, B. 1963. The short-term retention of active and r.assive sentences.

Unpublished doctoral dissertation, Johns Hopkins University.
ANDERSON, J. 1971. The granrmar of case: Towards a localist theory. London,
Cambridge University Press.
AUDLEY, R. J., and C. P. WALLIS. 1964. Response instructions and the speed of
relative judgments. I. Some experiments on brightness discrimination. BrJ-
Psych 55.59-73.
BEM, S. L. 1970. The role of comprehension in children's problem ·solving. Dev-
Psych 2.351-58.
BEVER, T. G. 1970. The cognitive basis for linguistic structures. Cognition and
the development of language, ed. by J. R. Hayes. New York, Wiley.
BIERWISCH, M. 1967. Some semantic universals of German adjeetivals. FL 3.1-36.
BuRT, C. .1919. The development of reasoning in school children. Journal of
Experimental Pedagogy 5.68-77, 121-27.
CAMPBELL, R.N., and R. J. WALES. 1969. Comparative structures in English. JL
5.215-51.
CHASE, W. G., and H. H. CLARK. 1971. Semantics in the perception of verticality.
BrJPsych 62.311-26.
- - . 1972. Mental operations in the comparison of sentences and pictures.
Cognition in learning and memory, ed. by L. Gregg. New York, Wiley.
CHOMSKY, N. 1957. Syntactic structUres. The Hague, Mouton.
- - . 1965. Aspects of the theory of syntax. Cambridge, Mass., M.I.T. Press.
CLARK, E. V. 1969. Language acquisition: The child's spontaneous description of
events in time. Unpublished doctoral dissertation. Department of Linguistics,
University of Edinburgh.
-.-. 1971. On the acquisition of the meaning of 'before' and 'after'. JVLVB
10.266-75.
CLARK, H. H. 1965. Some structural properties of shnple active and passive sen-
tences. JVLVB 4.365-70.
- ·-. 1969a. Linguistic processes in deductive reasoning. PsychRev 76.387-404.
- - . 1969b. The influence of language in solving tbree-term series problems.
JExPsych 82.205-15,
- - . 1970a. Comprehending comparatives. Advances in psycholinguistics, ed.
by G. Flores'd'Arcais and W. J. M. Levell. Amsterdam, North Holland Press.
1424 HERBERT H. CLARK
- - . 1970b. The primitive nature of children's relational concepts. Cogriition
and the development of language, ed. by J. R. Hayes. New York, Wiley.
CLARK, H. H., and J. S. BEGUN. 1968. The use of syntax in understanding sen-
tences. BrJPsych 59.219-29.
CLARK, H. H., and S. K. CARD. 1969. Role of semantics in remembering compa-
rative sentences. JExPsych 82.545-53.
CLARK, H. H., and W. G. CHASE. 1972. On the process of comparing sentences
against pictures.
1974. Perceptual coding strategies in the f!)rmation and verification of
descriptions. Memory and Cognition 2.101-11.
CLARK, H. H., and E. V. CLARK. 1968. Semantic distinctions and memory for
complex sentences. QJEP 20.129-38.
CLARK, H. H., and C. OFFIR. Forthcoming. On 'coming' and 'going'.
CLARK, H. H., and T. PETERSON. Forthcoming. Comprehension of the compara-
tive as an instruction. •
CLARK, H. H., and R. A. STAFFORD. 1969. Memory for semantic features in the
verb. JExPsych 80.326-34.
CLARK, H. H., and 0. YoUNG. Forthcoming. Explicit and iniplicit negation and
its comprehension.
CLIFTON, C., Jr., and P. 0DOM. 1966. Siniilarity relations among certain English
sentence constructions. Psychological Monographs 80 (5, Whole No. 613).
COLEMAN, E. B. 1964. The comprehensibility of seve<al grammatical transfor-
mations. Journal of Applied Psychology 48.186-90.
DESOTO, C.; M. LONDON, and S. HANDEL. 1965. Social reasoning and spatial
paralogic. Journal of Personality and Social Psychology 2.513-21.
DoHERTY, P., and A. ScHWARTZ. 1967. The syntax of the compared adjective in
English. Lg'43.903-36.
DoNALDSON, M. 1963. A study of children's thinking. London, Tavistock.
DONALDSON, M., and G. BALFOUR. 1968. Less is more: A study of language
comprehension in children. BrJPsych 59.461-71.
DoNALDSON, M., and R. J. WALES. 1970. On the acquisition of some relational
terms. Cognition and the development of language, ed. by J. R. Hayes. New
York, Wiley.
DvTHm, J. 1963, A further study of overlap error in three4erm series problems.
A study of children's thinking, ed. by M. Donaldson. London, Tavistock.
EIFERMANN, R. R .. 1961. Negation: A linguistic variable. Acta Psychologica
18.258-73.
F'ILLENBAUM, S. 1966. Memory for gist: Some relevant variables, L&S 9.217-'n.
- - . 1968a. Sentence similarity determined by a semantic relation: Learning of
'converses'. Proceedings of the 76th annual convention of the American Psy-
chological Association 3.9-10.
- - . 1968b. Recallfor answers to 'conducive' questions. L&S 11.46-53.
 SEMANTICS AND COMPREHENSION 1425
FILLMORE, C. J. 1967a. Deictic categories in the semantics of 'come'. FL 3.219-27.
- - . 1967b. The syntax of English preverbs. Glossa 1.91-125.
- - . 1968. The case for case. Universals in linguistic theory, ed. by E. Bach
and R. T. Harms. New York, Holt, Rinehart and Winston.
- - . 1970. Types of lexical information. Studies in syotax. FL supplementary
series, vol. 10, ed. by F. Kiefer. Dordrecht, Reidel.
FIRBAS, J. 1964. From comparative word-order studies. Bmo studies in English.
[Opera universitatis brunensis 93]. 4.111-26.
/ i FLoREs n'ARCAIS, G. B. 1966. On handling comparative sentences. Unpublished
manuscript. Harvard University, Center for Cognitive Studies.
FoDOR, J., and M. GARRETT. 1966. Some reflections on competence and per-
formance. Psycholinguistics papers, the proceedings of the 1966 Edinburgh
Conference, ed. by J. Lyons and R. J. Wales, pp. 135-53. Edinbirrgh Univer-
sity Press.
GouGH, P. B. 1965. GranJmatical transformations and speed of understanding.
JVLVB 5.107-11. •
- - . 1966. The verification of sentences: The effects of delay of evidence and
sentence length. JVLVB 5.492-6.
GREENBERG, J. H. 1963. Some universals of granJmar with particnlar reference
to order of meaningfnl elements. Universals of language, ed. by J. H. Green-
berg, pp. 58-90. Cambridge, Mass., M.I.T. Press.
GREENE, J. M. 1970a. Syotactic form and semantic function. QJEP 22.14-27.
- - . 1970b. The semantic function of negatives and passives. BrJPsych 61.17-22.
GRUBER,J. S. 1965. Studies in lexical relations. Unpublished doctoral disserta-
tion, M.I.T.
HALLIDAY, M.A. K. 1967. ·Notes on transitivity and theme in English: II. JL
3.199-244. .
HANDEL, S., C. DESOTO, and M. LoNDON. 1968. Reasoning and spatial repre-
sentation. JVLVB 7.351-7.
HoRN;L R. 1969. A presuppositional analysis of 'only' and 'evel:t'. Papers from
the Fifth Regional Meeting, Chicago Linguistic Society, pp. 98-107.
HUDDLESTON, R. D. 1967. More on the English comparative. JL 3.91-102.
HuNTER, I.M. L. 1957. The solving of three-term series problems. BtJPsych
48.286-98.
HUTIENLOCHER, J. 1968. Constructing spatial iniages: A strategy in ·reasoning.
PsychRev 75.550-60.
HUITENLOCHER, J., K. EISENBERG, and S. STRAuss. 1968. Comprehension: Rela-
tion between perceived actor and logical subject. NLVB 7.300-04.
HUITENLOCHER, J., E. T. HIGGINS, C. MILLIGAN, and B. KAUFFMAN. 1970. The
mystery of the 'negative equative' construction. JVLVB 9.334-41;.
HuTTBNLOCHER;J., and S. STRAuss. 1968. Comprehension mid a statement's re-
lation to the situation it describes. JVLVB 7.527'-30.
1426 HERBERT H. CLARK
JOHNSON-LAJRii, P. N. 1967. Katz on analyticity. JL 3.82.
- - . 1968a. The choice of the passive voice in a communicative task. BrJPsych
59.7-15.
- - . 1968b. The interpretation of the passive voice. QJEP 20.69-73.
--. 1969a. On understanding logically complex sentences. QJEP 21.1-13.
--. 1969b. Reasoning with ambiguous sentences. BrJPsych 60.17-23.
--. 1969c. '&'. JL 6.111-14.
JoNEs, S. 1966a. The effect of a negative qualifier in an instruction. JVLVB
5.495-501.
- - . 1966b. Decoding a deceptive instruction. BrJPsych 57.405-11.
- - . 1968. Instructions, self-instructions and performance. QJEP 20.74-78.
- ·-. 1970. Visual and verbal processes in problem-solving. Cognitive Psycho-
logy 1.201-14.
JusT, M. A., and P. A. CARPENTER. 1971. Comprehension of negation with
quantification. JVLVB 10.244-53.
KATZ, J. J;, and P.M. PosTAL. 1964. An integrated theory of linguistic ·descrip-
tions. Cambridge, Mass., M.I.T. Press.
KLIMA, E. S. 1964. Negation in English. The structure of language, ed. by J. A.
Fodor and J. J. Katz, pp. 232-46. Englewood Cliffs, N. J., Prentice-Hall.
KURODA, S. Y. 1968. English relativization and certain related problems. Lg
44.244-66.
LAKOFF, G. 1966. Stative adjectives and. verbs in English. Mathematical lin-
guistics and automatic translation. NSF-17.
- - . 1971. On generative semantics. Semantics, ed. by D. D. Steinberg and
L. A. Jakobovits. Cambridge, Cambridge University Press.
LEES, R. B. 1961. Grammatical analysis of the English comparative construction.
Word 17.171~85.
LYONs, J. 1968. Introduction to theoretical linguistics. Cambridge, Cambridge
University Press.
McCAWLEY, J.D. 1968. Lexical insertion in a transformational grammar with-
out deep structure. Papers from the Fourth Regional Meeting of the Chicago
Linguistic Society, pp. 71-80.
McKAy, J. C. 1968. Some generative rules for German thne adverbials. Lg
44.25-50.
McMAHoN, L. E. 1963. Grammatical analysis as part of understanding. a sen-
tence. Unpublished doctoral dissertation, Harvard University.
MEHLER,}. 1963.. Some effects of grammatical transformations on the recall of
English sentences. JVLVB 2.346-51.
MILLER, G. A. 1962. Some psychological· studies of: grammar. American Psy-
chologist 17.748-62.
MILLER, G. A., and K. McKEAN. 1964.. A chronometric study of some relations
between sentences. QJEP 16;297-308,
 SEMANTICS AND COMPREHENSION 1427
PIAGET, J. 1921. Une forme verbale de Ia comparaison chez l'enfant. ArchPs
18.141-72.
- - . 1928. Judgment and reasoning in the child. London, Paul Kegan.
PosTAL, P.M. 1970. Ou the surface verb 'remind'. Linguistic Inquiry 1.37-120.
Ross, J. R. 1969. A proposed rule of tree-pruning. Modem studies in English;
readings in transformational grammar, pp. 288-300, ed. by D. A. Reibel and
S. A. Schane. Englewood Cliffs, N. J., Prentice-Hall.
SACHS, J. 1967. Recognition memory for syntactic and semantic aspects of con-
nected discourse. Perception and Psychophysics 2.437-42.,
SAVIN, H. B., and E. PERCHONOCK. 1965. Grammatical structure and the imme-
diate recall of English sentences. JVLVB 4.348-53.
SEYMOUR, P. H. K. 1969. Response latencies in judgments of spatial location.
BrJPsych 60.31-9.
SHIPLEY, W. C., E. D. NoRRIS, and M. L. RoBERTS. 1946. The effect of changed
polarity of set on decision time of affective judgments. JExPsych 36.237-43.
SINGER, M., W. G. CHAsE, R. M. YoUNG, and H. H. CLARK. 1971. Practice
effects in the comparison of sentences and pictures. Paper presented at Mid-
western Psychological Association meetings, Chicago.
SLOBIN, D. I. 1966. Grammatical transformations and sentence comprehension
in childhood and adulthood. JVLVB 5.219-27.
SMITH, C. S. 1961. A class of complex modifiers. Lg 37.342--65.
SMITH, K. H., and L. E. McMAHON. 1970. Understanding order information in
sentences: Some recent work at Bell Laboratories. Advances in psycholinguis-
tics, ed. by G. Flores d'Arcais and W. J. M. Levelt. Amsterdam, North Hol-
land.
STAAL, J. F. 1968. 'Ancj' . .JL 4.79-81.
SvARTVIK, J. 1966. On voice in the English verb. The Hague, Mouton.
TANNENBAUM, P. H., and F. WILLIAMs. 1968. Generalization of active and pas-
sive sentences as a function of subject or object focus. JVLVB 7.246-50.
TELLER, P. 1969. Some discussion and extension of Manfred Bierwisch's work
on German adjectivals. FL 5.185-217.
TRABASSO, T., H. RoLLINS, and E. SHAUGNESSY. 1971. Storage and verification
stages in processing concepts. Cognitive Psychology 2.239-89.
TuRNER, E. A., and R. RoMMETVEIT. 1968. The effects of focus of attention on
storing and retrieving of active and passive voice sentences. JVLVB 7.243-8.

l'
VENDLER, Z. 1967. Linguistics in philosophy. Ithaca, Cornell University Press.
WALEs, R. J., and R. GRIEVE. 1969. What is so difficult about negation? Percep-
tion and Psychophysics 6.327-32.
L
r WALLIS, C. P., and R. J. AUDLEY. 1964. Response instructions and the speed of
relative judgments. II. Pitch discrimination. BrJPsych 55.133-42.
WASON, P. C. 1959. The processing of positive and negative information. QJEP
I 11.92-107.
I
I
1428 HERBERT H. CLARK

- - . 1961. Response to affirmative and negative binary statements .. BrJPsych

52.133-42.
-·-. 1965. The contexts of plausible denial. JVLVB 4.7-11.
WASON, P. C., and S. JoNEs. 1963. Negatives: Denotation and connotation: BrJ-
Psych 54.299-307.
WATT, W. C. 1970. On two hypotheses concerning psycholingnistics. Cognition
and the. development of language, ed. by J. R. Hayes. New York, Wiley.
WRIGHT, PA1;1UCIA. 1969. Transformations and the understanding ·Of sentences.
L&S 12.156--66.
YOUNG, R., and W. G. CHASE. 1971. Additive stages in the comparison. of sene
tences and pictures. Paper presented at .Midwestern Psychological Associa-
tion meetings, Chicago.

..

<' _,, ;
,·,;