Do speakers and listeners observe the Gricean Maxim of Quantity?

Text-only Preview

Journal of
Memory and
Journal of Memory and Language 54 (2006) 554–573
Do speakers and listeners observe the Gricean Maxim
of Quantity?
Paul E. Engelhardt a,*, Karl G.D. Bailey b, Fernanda Ferreira a
a Department of Psychology, Michigan State University, East Lansing, MI 48824-1116, USA
b Andrews University, Berrien Springs, MI 49104, USA
Received 25 April 2005; revision received 12 December 2005
Available online 21 February 2006
The Gricean Maxim of Quantity is believed to govern linguistic performance. Speakers are assumed to provide as
much information as required for referent identification and no more, and listeners are believed to expect unambiguous
but concise descriptions. In three experiments we examined the extent to which naı¨ve participants are sensitive to the
Maxim of Quantity. The first was a production experiment which demonstrated that speakers over-describe almost one-
third of the time. The second experiment showed that listeners do not judge over-descriptions to be any worse than
concise expressions. The third experiment used the Visual World Paradigm to assess listeners’ moment-by-moment
interpretations of over-described utterances. This last experiment revealed that over-descriptions trigger eye movements
that can be interpreted as indicating confusion. The results provide support for the use of a simple heuristic such as
Minimal Attachment or Argument Saturation to create an initial parse. We conclude that people are only moderately
Ó 2006 Elsevier Inc. All rights reserved.
Keywords: Gricean Maxim of Quantity; Syntactic ambiguity resolution; Visual World Paradigm; Language comprehension; Eye
movements; Language production
A simple object such as an apple may be described
allow an addressee to identify an intended referent.
as the apple, the ripe apple, the ripe apple on the towel,
If there is only one apple in the relevant environment,
the ripe apple on the towel in the kitchen, and so on.
then the apple is sufficient. If there is more than one,
How does a speaker decide how much information
enough modification should be provided to allow the
to provide when referring to an object? Psycholinguists
apple to be uniquely identified. We will use the term
generally assume that speakers are cooperative and
under-description to refer to any expressions not con-
adhere to a Maxim of Quantity, which consists of
sistent with this principle. The second component of
two principles (Grice, 1975). The first is that speakers
the Maxim of Quantity is that speakers should not
should make their contributions as informative as
make contributions more informative than is necessary
required: enough information should be included to
for successful communication. If there is only one
apple, modifiers are not required and should not be
included. We will use the term over-description to refer
* Corresponding author. Fax: +1 517 353 3745.
to expressions that are inconsistent with this second
E-mail address: [email protected] (P.E. Engelhardt).
component of the Maxim of Quantity.
0749-596X/$ - see front matter Ó 2006 Elsevier Inc. All rights reserved.

P.E. Engelhardt et al. / Journal of Memory and Language 54 (2006) 554–573
It is easy to see why this Gricean principle has had
ings that speakers over-describe and that listeners per-
such an enduring appeal. Certainly, if a speaker wishes
form efficiently with over-descriptions are incompatible
to be understood, it makes sense that he or she would
with the Maxim of Quantity.
avoid producing expressions that do not contain enough
In the current studies we focus primarily on over-de-
information to allow a listener to identify the intended
scriptions, because evidence suggests they are more com-
referent. Similarly, it seems intuitively obvious that
mon than under-descriptions (Deutsch & Pechmann,
speakers would want to minimize their efforts and not
1982; Ferreira, Slevc, & Rogers, 2005; Olson, 1970;
include a potentially infinite number of modifiers that
Sedivy, 2003). In addition, the assumption that over-de-
are not likely to be informationally useful. Because it
scriptions are avoided has played a critical role in
seems plausible that speakers would adhere to the Max-
debates about the architecture of language processing.
im of Quantity, and because it seems reasonable to
These discussions have centered on restrictive post-nom-
assume that conversational partners are cooperative
inal modifiers, including prepositional phrases. A sen-
(Clark, 1992), listeners will have certain expectations
tence such as Mary put the frog on the towel into the
of speakers. Listeners expect them to provide enough
box contains a temporary syntactic ambiguity, because
information to allow referent identification, and will
on the towel could be either a modifier of frog (as it turns
therefore be bothered by under-descriptions. In addi-
out to be, in this case), or the location to which the frog
tion, they expect all information provided to be relevant,
is to be moved. The Garden-Path model of parsing
so if a modifier is included, it should be informative (in
assumes that syntactic representations are created serial-
the technical sense of ruling out states of affairs; Levi-
ly and that the system’s initial preference is for syntactic
son, 2000). Thus, over-descriptions are more than just
simplicity (Frazier, 1978). Moreover, the parser is an
a waste of time; they may actually disrupt comprehen-
informationally encapsulated module, consulting only
sion. For example, if a speaker uses the expression the
syntactic principles during the creation of the single ini-
large apple, the listener will assume that large is commu-
tial parse. If the initial parse is not viable (as is the case
nicatively relevant, and so she might assume that there is
in the above example), then the more complex structure
more than one apple in the discourse and that the mod-
will be built or reactivated. Thus, according to the Gar-
ifier is the information allowing the correct apple to be
den-Path model, listeners’ preference to interpret the
distinguished from any others. If in fact only one apple
phrase on the towel as a location rather than a modifier
exists, the listener has been misled about the existence of
is a consequence of the parser’s preference to avoid pos-
a set of objects, which may impair communication (cf.
tulating potentially unnecessary nodes.
Sedivy, Tanenhaus, Chambers, & Carlson, 1999).
In contrast, other models assume that the system
While these ideas are intuitively appealing, it is
considers all alternatives in parallel when there is an
important to know whether people adhere to these prin-
ambiguity, and that the parser consults all potentially
ciples of communication in practice. What is the evi-
relevant sources of information (MacDonald, Pearlmut-
dence that the Maxim of Quantity is followed, both by
ter, & Seidenberg, 1994; Spivey & Tanenhaus, 1998;
speakers and listeners? The experiments described below
Tanenhaus & Trueswell, 1995). If the sentence occurs
were designed to answer this question. Previous work by
in a linguistic or visual context in which there is only
Deutsch and Pechmann (1982) examined this issue by
one frog, the listener will apply his knowledge of Gri-
using contexts that contained multiple objects of the
cean principles and infer that on the towel is not a mod-
same type. Objects differed in size and color, and partic-
ifier, since it is not necessary. If there is more than one
ipants were asked to name one object that they would
frog, the putatively more complex structure will be
like to give as a present to someone else. In contexts with
favored from the outset, because the modifier interpreta-
multiple referents, Deutsch and Pechmann found that
tion allows the semantic/referential system to uniquely
adult speakers rarely produced ambiguous or under-de-
identify the appropriate referent. Thus, the assumption
scribed utterances. For example, in a display containing
that listeners assume a contrastive interpretation when
a red apple and a green apple, speakers almost never
they hear a complex description is critical in explaining
used the expression the apple. Instead, they produced
comprehenders’ tendency to experience a Garden-Path
enough modification to allow the intended referent to
for structures such as put the frog on the towel into the
be identified. Surprisingly, redundant or over-described
box when only one frog exists. (Other Garden-Path
utterances were produced fairly often, on over one-quar-
forms that turn on an ambiguity between a modifier
ter of trials. For example, in a display containing a single
and non-modifier interpretation are explained in a simi-
apple, participants would often refer to it as the red apple
lar way). If people do not have this expectation, then the
or the large apple, although the phrase the apple was suf-
tendency to prefer the non-modifier interpretation must
ficient for identification. Moreover, they argued that
have a different cause (e.g., the preference for syntactic
these over-described utterances actually led to more effi-
cient searches, indicating that comprehenders do not
One early model that assumed this application of the
suffer when they encounter over-descriptions. The find-
Maxim of Quantity (i.e. the principle of parsimony) to

P.E. Engelhardt et al. / Journal of Memory and Language 54 (2006) 554–573
the parsing of syntactically ambiguous sentences is Ref-
those involving ambiguous prepositional phrases (Spi-
erential Theory (Altmann & Steedman, 1988; Crain &
vey, Tanenhaus, Eberhard, & Sedivy, 2002; Tanenhaus,
Steedman, 1985). Crain and Steedman investigated
Spivey-Knowlton, Eberhard, & Sedivy, 1995; Trueswell,
structures such as the reduced relative (e.g., The neighbor
Sekerina, Hill, & Logrip, 1999). Tanenhaus et al. (1995)
warned about the criminal called the police), which have
investigated this type of ambiguity in an experimental
long been known to be difficult to understand (Bever,
task in which participants were required to move objects
1970). The Garden-Path model attributes this difficulty
in response to spoken instructions.
to the parser’s preference for the syntactically simple,
The displays contained either one or two relevant
main clause interpretation of warned about the criminal.
objects to be moved (in Figs. 1A and B, one apple or
But according to the Referential Model, the active struc-
two apples), and instructions to move the objects were
ture is built initially because the sentence occurs in a null
either syntactically ambiguous (1) or unambiguous (2).
discourse context and there is therefore no referential
support for the complex reduced relative form. Thus,
(1) Put the apple on the towel in the box. (ambiguous)
if the reduced relative occurred in a context that estab-
(2) Put the apple that’s on the towel in the box.
lished the existence of a set of neighbors, then the
sequence warned about the criminal would be interpreted
as a modifier right away, because the comprehender
When the visual context contained just one apple and
needs modification to identify the correct referent of
the instruction was ambiguous, participants fixated the
neighbor. In an offline judgment task, Crain and Steed-
incorrect location (the empty towel) shortly after hear-
man observed large effects of context on decision times
ing the prepositional phrase on the towel on 55% of tri-
als. With the unambiguous instruction, participants
almost never looked to the incorrect location. When
These results were later challenged in a series of
the visual context contained two apples, the empty towel
experiments with the same logic but employing eye-
was fixated much less often overall, and there was no dif-
movement monitoring to allow measurement of online
ference between ambiguous and unambiguous instruc-
processing (Ferreira & Clifton, 1986). Reduced relative
tions (15 and 12% of trials had fixations on the empty
sentences and sentences with temporarily-ambiguous
towel, respectively). This pattern is consistent with the
prepositional phrases (e.g., Mary put the book on the
predictions of the Referential Model, which assumes
table into her bag) were placed in contexts like the ones
that listeners expect speakers to follow the Maxim of
created by Crain and Steedman (1985). Ferreira and
Quantity. If the relevant visual world contains only
Clifton found that performance on a question–answer-
one apple, a modifier is unnecessary and the ambiguous
ing task was consistent with what Crain and Steedman
prepositional phrase is therefore initially interpreted as a
had observed, but online measures of processing (e.g.,
location, which results in looks to the empty towel. In
first-pass reading times) showed that the parser initially
contrast, with two apples, a modifier is required to allow
preferred the syntactically simpler structure. They con-
the correct referent to be identified, and the ambiguous
cluded that the referential principle applied during pro-
prepositional phrase is therefore immediately interpreted
cessing, but only on the output of the parser. The parser
as the contextually-necessary modifier.
initially selected the simpler structure, but when reanal-
These findings support the idea that over-descrip-
ysis was forced syntactically, contextual information
tions are problematic for listeners, but the evidence is
helped the comprehension system identify the correct
indirect. To directly test the Gricean Maxim of Quantity,
structure. (These issues were considered in further elab-
orations of the Referential Model; Altmann & Steed-
man, 1988; Clifton & Ferreira, 1989; Trueswell,
Tanenhaus, & Garnsey, 1994.)
The debate has been renewed with the introduction
of a new and powerful technique for studying language
processing, the Visual World Paradigm. In this para-
digm, spoken language comprehension is assessed by
monitoring eye movements as people view visual dis-
plays (either real-world or displayed on computer mon-
itor). Cooper (1974) demonstrated that people reliably
fixate pictures that are related to concurrent spoken lan-
One referent
Two referent
guage, and that eye movements are systematically time-
locked to related spoken words. During the past decade,
Fig. 1. (A and B) One referent (apple) and two referent (two
this paradigm has become popular for studying the pro-
apples) displays used in Tanenhaus et al. (1995) and Spivey
cessing of temporary syntactic ambiguities, particularly
et al. (2002).

P.E. Engelhardt et al. / Journal of Memory and Language 54 (2006) 554–573
we must examine whether speakers avoid producing
In the third experiment, we monitored the eye move-
over-descriptions, and we must assess to what extent
ments of participants as they moved objects in response
listeners are misled by them. We did this in the experi-
to those same instructions. If the Gricean Maxim of
ments reported below. First, a production study was
Quantity is used in the formulation of referential expres-
conducted in order to determine the types of utterances
sions, then speakers should not over-describe; for exam-
that participants generated when instructing another
ple, they should rarely refer to an apple in a display such
person to move the target object to either of the empty
as Fig. 2A as the apple on the towel. Listeners should
locations (see Figs. 2A and B). In the second experiment,
judge such an utterance/visual world combination to
participants rated the acceptability of four different
be inappropriate or unacceptable, because the modifier
instructions (see Table 1).
is informationally unmotivated. Listeners’ eye movements
Fig. 2. (A) Example stimulus for Experiment 1 (matching condition). (B) Example stimulus for Experiment 1 (different condition).

P.E. Engelhardt et al. / Journal of Memory and Language 54 (2006) 554–573
Table 1
required to give three instructions to another ‘‘subject’’
Instructions tested in Experiments 2 and 3
(in reality, a confederate) corresponding to the move
Instruction type
shown between each successive pair of pictures. Partici-
pants were told prior to the experiment that the ‘‘other
Location type
participant’’ would not necessarily have the objects in
phrase modifier
the same configuration as the one shown in the pictures.
(3) Put the apple on
This information was provided in order to prevent sub-
the towel
jects from using directional terms (i.e., ‘‘move the apple
(4) Put the apple
diagonally’’ or ‘‘move the apple to the right’’) in their
in the box
instructions. We constrained the instructions in this
(5) Put the apple on
the towel on the
way because our experiments attempt to determine
other towel
how much modification is usually provided and how
(6) Put the apple on
such modifiers are interpreted. It was therefore critical
the towel in the box
that participants use linguistic terms that did not depend
on objects being arranged in a 2 · 2 configuration.
should reflect their tendency to think that the ambiguous
The main purpose of this study was to determine
prepositional phrase is a location, and there should be
what type of instruction a naı¨ve participant would give
frequent looks to the empty towel. If the Gricean Maxim
to get a person to move a target object to one of the
of Quantity is not used by speakers and listeners or if
other available locations. In particular, we asked
other principles come into play and override it, then
whether the speaker would produce over-descriptions
we would expect the following: (1) Speakers should
such as put the apple on the towel in the box. In addition,
sometimes over-describe, (2) listeners should judge
we were interested in whether participants in the match-
over-described sentences as acceptable, and (3) eye
ing location condition would provide a modifier for the
movements should show little evidence that listeners
location because the object to be moved was already on
interpret the empty towel as a location, because listeners
an object of that same type. Thus, if the apple was on a
would not necessarily expect modifiers to be used to
towel and the new location was a different towel, they
implicate the existence of a set.
might tend to say put the apple on the other towel.
These three experiments together allow us to deter-
mine whether speakers avoid over-descriptions, whether
listeners find them infelicitous, and whether listeners
perform less efficiently when they encounter them. Alter-
natively, we might find dissociations between production
performance, explicit judgments of sentences, and
implicit comprehension performance.
Ten students from the undergraduate subject pool at
Michigan State University participated in the study for
course credit. Participants were native speakers of Eng-
Experiment 1: Production
lish and had normal or corrected-to-normal vision.
In Experiment 1, we examined the types of utterances
that participants produced in contexts such as those
shown in Figs. 1A and B. Our two referent context
Stimuli for this experiment consisted of four pictures
(Fig. 1B) was somewhat different, however, in that only
printed on a single sheet of paper. Pictures were labeled
one of the potential target objects was on a location
in the upper left corner as S, 1, 2, and 3, as shown in
object (for example, one apple was on a towel but the
Figs. 2A and B. The images were printed in grayscale
second apple was by itself). Participants instructed a
to prevent participants from using color to indicate
confederate to move real objects to both ‘‘matching’’
which object had to be moved. The difference between
and ‘‘different’’ locations. In both conditions, this object
each pair of pictures (i.e., S–1; 1–2; 2–3) consisted of a
to be moved was always on or in some other object (e.g.,
single movement of one object. We created a total of
an apple on a towel). In the ‘‘matching’’ location condi-
53 visual displays: five were for practice, 24 were exper-
tion, the instruction was to move the target object from
imental items, and 24 were filler displays. In experimen-
one location to another of the same type (e.g., the apple
tal trials, the target and distractor objects were always
on a towel was to be moved to another towel). In the
on the left side of the display and possible location
‘‘different’’ location condition, the target object was to
objects were on the right (as in Spivey et al., 2002). In
be moved to a different type of location (e.g., the apple
the practice and filler trials, target, distractor, and loca-
on a towel was to be moved to a box). Participants were
tion objects all had an equal probability of appearing
shown four pictures (see Figs. 2A and B) and were
in any of the four quadrants of the display, and no

P.E. Engelhardt et al. / Journal of Memory and Language 54 (2006) 554–573
constraints were placed on direction of movement.
in generating utterances (e.g., Now move that apple over
Practice trials consisted of three two-referent trials and
there). Participants were not given feedback concerning
two one-referent trials. Filler trials consisted of 16 one-
the quality of their instructions nor were they corrected
referent trials and eight two-referent trials. Eight of the
if an instruction was referentially ambiguous. Each par-
one-referent filler displays had no duplicated location
ticipant completed five practice trials followed by 24 crit-
objects (for example: an apple on a towel, a frog, an
ical trials (six trials in each condition) and 24 filler trials.
empty box, and an empty bowl). The other eight one-
The order of critical and filler trials was randomly
referent filler displays were like the one shown in
assigned for each participant. The utterances produced
Fig. 2A. The two-referent filler displays had two identi-
by each participant were recorded on audiotape, and
cal target objects (e.g., two apples); one of these objects
only the first instruction for each critical trial was tran-
was located on another object and the other was by
scribed. After completing all trials, the participant was
itself. The other two quadrants of the display were occu-
debriefed. The entire session lasted about 45 min. Partic-
pied by two distinct location objects so that no location
ipants were tested individually.
objects were duplicated in the display.
The experiment had a 2 · 2 within-participants
The confederate was placed in the waiting area of the
design: Visual contexts contained either one or two
laboratory and waited for the participant to arrive. Once
instances of the object to be moved (one-referent versus
he or she did, the experimenter addressed both in a sim-
two-referent condition), and the location type to which
ilar manner (for example, he always asked each of them
the object was to be moved either matched (towel to
to give their names). The confederate and the participant
towel) or was different (towel to box). Analyses were
both read the instructions and completed consent forms.
conducted with both participants (F1) and items (F2)
They were then shown all of the objects that would be
as random effects.
used in the experiment and were told their category label
Responses were categorized as follows: If the target
(e.g., apple, towel, and box; see, Table 2).
object (e.g., apple) was not modified, the utterance was
The participant and confederate were then seated on
labeled ‘‘bare target’’. This categorization applied even
opposite sides of a table facing one another. The exper-
if the location was modified by a word such as other. If
imental stimuli and a microphone were placed in front
the target object was modified, the utterance was
of the participant, and the objects corresponding to
labeled ‘‘modified target’’. This response category does
the stimuli were placed in front of the confederate. For
not distinguish between reduced and non-reduced
each trial, the participant gave three instructions, paus-
descriptions—i.e., put the apple on the towel in the
ing after each to allow the confederate to move an
box and put the apple that’s on the towel in the box
object. The confederate moved the real object corre-
were both categorized as modified target. The number
sponding to each instruction and responded with
of modified target responses was divided by the total
‘‘okay’’ after the move was completed. The experimenter
number of utterances produced for each condition.
added and removed objects between trials. A divider was
This proportion was the dependent variable for statis-
placed between the participant and confederate after the
tical analyses.
third practice trial to ensure that the participant would
Orthogonal to this categorization scheme, we also
not use physical cues such as eye gaze or hand location
tabulated whether a pre-nominal modifier such as other
was used to describe the location to which the object was
to be moved. A response either included other (or oppo-
Table 2
site, or another similar term) or it did not. In this
Objects used in Experiments 1–3
scheme, we ignored whether the noun phrase referring
Target & distracter objects
‘‘in’’ Locations
‘‘on’’ Locations
to the target object was modified or unmodified. This
analysis examined whether speaker would likely produce
the instruction put the apple on the towel when an object
is moved from one location to another location of the
same type.
One utterance was eliminated from the analysis
because the participant did not give a correct instruction
for the move shown.

P.E. Engelhardt et al. / Journal of Memory and Language 54 (2006) 554–573
Results are shown graphically in Fig. 3, where the
ent display; overall, almost a third of utterances included
proportion of modified target utterances is broken down
the unnecessary modifier. Again, recall that we refer to
by visual context (one vs. two-referent) and location
the modifier as unnecessary because the one-referent dis-
type (matching or different). Floating error bars show
plays contained just one apple (for example), so the
the 95% confidence interval for the pairwise comparison
expression the apple was sufficient to allow referent iden-
between adjacent condition means for the participants
tification. Of course, the modifier might be necessary or
data (Masson & Loftus, 2003). Let us consider first
at least compelling to the speaker for some other reason;
under what conditions people produced modified target
we will return to this possibility in the General
utterances. These should occur in the two-referent con-
dition but not in the one-referent condition, because
Now, let us consider whether people simply said put
modified target utterances in the one-referent condition
the apple . . . on the towel when the apple was to be
would constitute over-descriptions. Consistent with the
moved from one towel to another (i.e., the matching
Gricean Maxim of Quantity, we did indeed find that
condition). The answer is that they did not. Regardless
modified target utterances were more common in the
of whether there was one referent or two referents in
two-referent condition than in the one-referent condi-
the display (97% vs. 96%; both Fs < 1), people almost
F1 (1, 9) = 108.08,
p < .001,
always included a modifier when the location was
F2 (1, 5) = 257.83, p < .001, minF0 (1, 14) = 76.15, p <
matching (96.5%), and almost never included one when
.001. Of course, this is not surprising, as complex noun
phrases were essential for effective communication when
F1 (1, 9) = 644.78, p < .001, F2 (1, 5) = 198.24, p < .001,
two objects of the same type were present. Less compat-
minF0 (1, 8) = 152.62, p < .001 (see Table 3 for a break-
ible with the Maxim of Quantity is our finding that mod-
down of the types of modifiers produced in the matching
ified target instructions occurred on 30% of trials overall
condition). The interaction between these two variables
in the one referent condition. A one sample t-test
was not significant. Thus, it appears that if a speaker
revealed that the number of overdescriptions produced
intended to make a listener move an apple from one
in the one-referent condition was significantly greater
towel to another, he or she would modify the location
than 0, F1 (1, 9) = 11.0, p < .01, F2 (1, 5) = 102.96, p <
object so the resulting utterance would most likely be
.001, minF0 (1, 11) = 9.94, p < .01. (Note for this analysis
put the apple on the other towel. In other words, speakers
the t values were changed to F0s to calculate minF0.)
did not under-describe.
These over-descriptions occurred twice as often when
the location to which the object was to be moved was
matching rather than different (40% vs. 20%), but this
difference was not significant, being smaller than the
95% confidence interval of .30 for the comparison
These production data have yielded some unexpected
between condition means. There was no interaction
findings from the point of view of evaluating the Gricean
between number of referents and matching/different
Maxim of Quantity. Let’s begin with results that were
location, F1 (1, 9) = 3.56, p = .10, F2 (1, 5) = 3.01, p =
expected. First, consistent with the principle which states
.15, minF0 (1, 12) = 76.15, p = .23.
that speakers should provide as much information as nec-
The important finding from this analysis is that over-
essary, over 95% of the time speakers modified a noun
descriptions were surprisingly common in the one refer-
when there was more than one of the object that it denot-
ed. Thus, speakers avoided under-descriptions. However,
Matching location
they also included the modifier one-third of the time when
Utterance types produced
Delta value
Different location
it was not necessary (i.e., in the one referent display condi-
tions). Thus, it appears that speakers produce a surprising
number of over-descriptions—far more than would be
Table 3
Modifiers produced in matching condition in Experiment 1
tion of trials with modifier 0.3
(1) Other (i.e. the
other towel) (%)
(2) Opposite (i.e. the
opposite towel) (%)
Context type
(3) Directional (i.e. the
lower-right towel) (%)
Fig. 3. Results from Experiment 1, floating error bars show the
(4) Post-nominal (i.e. the
95% confidence interval for the pairwise comparison between
towel on the right) (%)
adjacent condition means.

P.E. Engelhardt et al. / Journal of Memory and Language 54 (2006) 554–573
expected if the Gricean Maxim of Quantity determined
and (6) correspond to ‘‘modified target’’ responses.
the form and content of people’s utterances. In addition,
However, one distinction that should be noted is that
speakers almost always included the word other (or oppo-
Experiment 2 examined temporarily ambiguous modi-
site or some semantic equivalent) in the matching location
fied target instructions, whereas the utterances produced
(towel-to-towel) condition. It was clearly pragmatically
in Experiment 1 were not necessarily ambiguous. The
relevant to them that the apple was to be moved from
bare target instructions should be judged as ‘‘bad’’ in
one towel to another, and so this was invariably marked
the two-referent conditions, regardless of whether the
in speakers’ utterances with the modifier other or opposite.
location is matching or different, because they are
In summary, speakers are fairly likely to over-de-
under-descriptions of the sort that do not permit unam-
scribe objects, but they avoid under-descriptions. Target
biguous identification of the intended referent. In addi-
objects were almost always described with a modifier
tion, instructions involving a move from one towel to
when there were two of the same type, and a destination
another should also be judged bad if the location object
was referred to as other or opposite when it was of the
is not modified, for example with the word other or
same type as the current location. In the next experi-
opposite, because, based on our production results, such
ment, we assess whether listeners judge under- and
an instruction is also an under-description (even though
over-described utterances to be infelicitous, as they
it ultimately does not produce any type of pragmatic
should according to the Maxim of Quantity.
confusion, since an object must be placed in a different
location to be correctly described as having been
‘‘moved’’). The two modified target instructions should
be rated high in the two-referent condition but low in
Experiment 2: Listener judgments
the one-referent condition, because modified instruc-
tions in the former condition are over-descriptions and
The purpose of this experiment was to assess whether
therefore violate the Gricean Maxim of Quantity. Again,
people find over- and under-described utterances infelic-
whether the location is matching or different should not
itous. Subjects judged spoken utterances by examining
matter, because the location is not relevant; all that mat-
two photographs presented side-by-side on a computer
ters is that with only one apple in the display, it is not
screen while they listened to a spoken instruction over
necessary to describe it as being on a towel.
headphones. The pictures were the same as those used
in Experiment 1 (see Figs. 2A and B, pictures S and
1). In the one referent condition, there was (for example)
an apple on a towel, a frog, an empty towel, and an emp-
ty box in the first picture. The pictures differed only in
that the relevant object had been moved either to the
matching location (towel to towel), or to a different loca-
Twenty-two people drawn from the same pool as for
tion (towel to box) in the second picture. In the two-ref-
Experiment 1 participated in this experiment. All were
erent condition, there were two relevant objects (e.g.,
native speakers of English, had normal or corrected-to-
two apples), an empty towel, and an empty box in the
normal vision, and none had participated in Experiment 1.
first picture. Again, the apple on the towel was moved
either to a matching or to a different location in the sec-
ond picture. The participants’ task was to judge whether
the instruction given was a good one to bring about the
The stimuli consisted of ‘‘before’’ and ‘‘after’’ pictures
change between the first and second pictures. Partici-
presented side by side on a computer screen. The pictures
pants rated the instructions on a scale of 1 to 5, where
were the same as pictures S and 1 from the previous exper-
1 was an incorrect instruction, 3 was an adequate
iment, except they were shown in color. Each participant
instruction, and 5 was a perfect instruction.
completed 10 practice trials and 72 regular session trials.
The instructions people received are shown in utter-
The latter consisted of 24 experimental (six in each condi-
ances (3–6) from Table 1, reprinted below:
tion, as in the previous experiment) and 48 filler trials. The
practice, filler, and experimental trials were designed so
(3) Put the apple on the towel. (matching)
that half of the instructions were ‘‘good’’ and half of the
(4) Put the apple in the box. (different)
instructions were ‘‘bad’’. A ‘‘good’’ instruction was one
(5) Put the apple on the towel on the other towel.
that unambiguously named the target object and goal,
and the instruction was correct for the movement shown
(6) Put the apple on the towel in the box. (different)
in the pictures. ‘‘Bad’’ instructions used either incorrect cat-
egory labels (target or location) or incorrect modification
Instructions (3) and (4) correspond to the ‘‘bare tar-
(for example, upper-left apple when the upper-right apple
get’’ responses from Experiment 1, and instructions (5)
was moved). Eighteen of the filler trials were ‘‘good’’ and

P.E. Engelhardt et al. / Journal of Memory and Language 54 (2006) 554–573
thirty were ‘‘bad’’, as all of the experimental instructions
participant responded. The order of trials was randomly
were ‘‘good’’ except for the bare target instructions in the
assigned for each participant. Participants were tested
two-referent condition. Sound files were recorded by a
individually, and the entire experimental session lasted
female native speaker of English, digitized using Comput-
25 min.
erized Speech Laboratory 4400 (Kay Elemetrics), and
converted to .wav format.
Results and discussion
Results are presented in Figs. 4A and B. Floating
The experiment was programmed using Superlab
error bars show the 95% confidence interval for the pair-
experimental software (Cedrus) and auditory stimuli
wise comparison between adjacent condition means for
were played through Sony MDR-CD60 headphones.
the participants data (Masson & Loftus, 2003). The
Participants completed the experiment on a Dell Opti-
result of a 2 ·2 · 2 repeated measures ANOVA revealed
plex GX 400 computer.
a significant interaction, F1 (1, 21) = 48.44, p < .001,
F2 (1, 5) = 42.25,
p < .01,
minF0 (1, 12) = 22.56,
p <
Design and procedure
.001. The following analyses are broken down by display
type because the Gricean Maxim of Quantity makes dis-
The experiment had a 2 · 2 · 2 design. Visual con-
tinct predictions for each type of context. The inferential
texts contained either one or two possible target refer-
statistics for the main effects and interactions are pre-
ents, the location type either matched the object that
sented in Table 4 (n.s. indicates not significant).
the target was currently on or was different, and the
instruction type either contained a prepositional phrase
One-referent context
modifier or did not (modified target vs. bare target).
All variables were manipulated within-subjects, but dis-
The main effects of location type and instruction type
play type was a between-items variable (the other two
were both significant. The different location conditions
variables were within-item).
produced higher ratings than the matching location con-
The bare target instructions in the different condition
ditions (4.61 vs. 3.93), and the modified target instruc-
were created by digitally removing the modifier (the first
tions were rated as better than the bare target
prepositional phrase) from the modified target instruc-
instructions (4.52 vs. 4.02). In addition, the loca-
tions. For the bare target instructions in the matching
tion · instruction
condition, a separate utterance like example (7) was cre-
Fig. 4A). Paired comparisons revealed four significant
ated, and then the first prepositional phrase was digitally
effects. First, the difference between the location condi-
removed. The creation of a separate utterance ensured
tions within the bare target instructions exceeded the
that the two bare target instructions had similar proso-
95% confidence interval of .31 for the difference between
dy. The resulting utterances sounded as natural as the
condition means. The different condition was rated high-
unaltered utterances (speech files can be obtained at
er than the matching condition (4.91 vs. 3.12). This
result is not surprising considering the disparity that
was observed in the production data (recall that the
(7) Put the apple in the box on the towel.
unmodified instruction was produced on 80% of trials
in the different condition and 60% of trials in the match-
Before the experiment began, participants were
ing condition). This result demonstrates that people
shown all of the objects that were used and were told
explicitly judge under-descriptions of the location to be
what label would be used for each (see, Table 2). Partic-
less felicitous than appropriate descriptions. It appears
ipants then read the instructions and signed a consent
that when there are two identical locations, a modifier
form, after which the experimenter asked if participants
(either pre- or post-nominal) is obligatory. The differ-
had any questions. Participants were instructed to rate
ence between the two location conditions with the mod-
the spoken instruction on a scale of 1–5, where a 1
was an incorrect instruction (i.e., the wrong category
confidence interval of .30 for the difference between
label or incorrect modification), a 3 indicated an ade-
condition means. The matching location was rated high-
quate instruction, and a 5 should be given to an instruc-
er than the different location (4.74 vs. 4.3). Again, this
tion that could not be better. Each trial began with a
result was predicted because the production data
message on the screen reading, ‘‘Press any key when
showed that speakers modified the target object on
ready’’, which was followed by a 1000 ms interstimulus
40% of trials in the matching condition. In addition,
interval. The visual display was shown for 2000 ms
speakers also modified the location object on 97% of
before the utterance was played out through head-
trials in the one-referent matching condition. It is there-
phones. The display remained on the screen until the
fore not surprising that comprehenders rated the

P.E. Engelhardt et al. / Journal of Memory and Language 54 (2006) 554–573
One-Referent Judgment Ratings
Matching location
Tw o-Re fe re nt Judg m e nt Ra tings
Matching location
Delta value
Delta value
Different location
Different location
ticipants ratings (1-5)

icipants ratings (1-5)


Bare Target
Modified Target
Bare Target
Modified Target
Instruction Type
Instruction Type
One-Referent Judgment Ratings
Matching location
Delta value
Matching location
Different location
Two-Referent Judgment Ratings
Delta value
Different location
ticipants ratings (1-5)

ticipants ratings (1-5)

Bare Target
Modified Target
Bare Target
Modified Target
Instruction Type
Instruction Type
Fig. 4. (A) Results from Experiment 2, one-referent condition, floating error bars show the 95% confidence interval for the pairwise
comparison between adjacent condition means. (B) Results from Experiment 2, two-referent condition, floating error bars show the
95% confidence interval for the pairwise comparison between adjacent condition means. (C) Results from Experiment 2b, one-referent
condition, floating error bars show the 95% confidence interval for the pairwise comparison between adjacent condition means. (D)
Results from Experiment 2b, two-referent condition, floating error bars show the 95% confidence interval for the pairwise comparison
between adjacent condition means.
modified target instruction higher, as it contained both
condition) when compared to a display that required
types of modification. Furthermore, the difference
modification of the target object.
between the two instruction conditions in matching loca-
tion displays exceeded the 95% confidence interval of .27
Two-referent context
for the comparison between condition means. The bare
target instruction was rated as worse than the modified
There was a main effect of instruction type (see
target instruction (3.12 vs. 4.74).
Fig. 4B). The modified target instructions were rated sig-
The difference between two different location instruc-
nificantly higher than the bare target instructions (4.35
tions also exceeded the 95% confidence interval of .38
vs. 2.8). This result is not surprising; under-descriptions
for the comparison between condition means. However,
can compromise communication, as an under-described
for this comparison the bare target instruction was rated
utterance will not permit a listener to identify the correct
higher than the modified target instruction (4.91 vs. 4.3),
referent from a set. A similar pattern was observed for
which is consistent with the Maxim of Quantity. We also
both the matching location and the different location.
observed a weak preference for utterances that modified
The location type · instruction type interaction in the
the target object inappropriately (i.e., in the one-referent
two-referent context was not significant. Ratings in the

Document Outline

  • Do speakers and listeners observe the Gricean Maxim of Quantity?
    • Experiment 1: Production
    • Methods
      • Participants
      • Materials
      • Procedure
      • Design
    • Results
    • Discussion
    • Experiment 2: Listener judgments
    • Method
      • Participants
      • Materials
      • Apparatus
      • Design and procedure
    • Results and discussion
      • One-referent context
      • Two-referent context
    • Experiment 3: Eye movements in the Visual World Paradigm
    • Method
      • Participants
      • Materials
      • Apparatus
      • Design and procedure
    • Results and discussion
    • General discussion
    • Acknowledgments
    • References