🚀 go-pugleaf

The following is a second PSYCOLOQUY target article on reading and
inference, now open for commentary. It is posted in two parts. PART I
is self-contained; commentaries not dependent on the experimental
details may prefer to address only PART I. PART II, which appears in
the next posting, reports the experiments on salience. Note that
although posted separately, both parts will be archived jointly as a
single article in the PSYCOLOQUY permanent archive.

   PART I of Glenberg & Mathew: WHEN MINIMALISM IS NOT ENOUGH:
                                MENTAL MODELS IN READING COMPREHENSION
                                Paragraphs 1 - 30, Lines 1 - 492
                                [Tables and References will appear in
                                PART II]

-------------------------------------------------------------------------
psycoloquy.92.3.64.reading-inference-2.1.glenberg-mathew  Sunday Dec 13 1992
ISSN 1055-0143           (50 paragraphs, 12 references, 4 Tables, 911 lines)
PSYCOLOQUY is sponsored by the American Psychological Association (APA)
                Copyright 1992 Arthur M. Glenberg & Shashi Mathew

                WHEN MINIMALISM IS NOT ENOUGH:
                MENTAL MODELS IN READING COMPREHENSION

                Arthur M. Glenberg
                Department of Psychology
                University of Wisconsin - Madison
                Madison, WI 53706
                glenberg@macc.wisc.edu

                        and

                Shashi Mathew
                Department of Psychology
                Harvard University
                Cambridge, MA 02138
                mathew@isr.Harvard.edu

    0.0 ABSTRACT: McKoon and Ratcliff (1992) contrast a minimalist
    theory of reading with a constructionist or mental-model theory
    "that readers automatically construct a full representation of the
    real-life situation described by a text." We take issue with three
    aspects of McKoon and Ratcliff's arguments. First, the contrast
    between the minimalist and the mental-model account is unwarranted:
    The minimalist account applies to automatic inferences, whereas
    mental models are not often constructed automatically. Second,
    mental models are not full representations of real-life situations.
    Last, we report the results of two experiments demonstrating
    problems with their "salience" account of phenomena previously
    attributed to mental models.

    0.1 KEYWORDS: reading comprehension, mental models, inferences,
    minimalism, salience

1.0  PART I

1.1.0  INTRODUCTION

1.1.1  McKoon and Ratcliff (1992) have argued for a minimalist strategy
in the investigation of automatic inferences during comprehension.
Minimalism is intended to apply to reading in the absence of specific
goals such as learning new information from a text. According to
minimalism, when this condition is met, automatic inferences are made
in two situations only. First, inferences are made automatically to
preserve local coherence. Second, inferences are made automatically
when the requisite information is highly available. McKoon and Ratcliff
contrast minimalism with several other theories they interpret as
positing that inferences are made under less restricted conditions. We
agree with McKoon and Ratcliff that the minimalist position provides a
sound (although not the only) research strategy, namely, assume the
minimum until forced by the data to assume more. We do have three
specific disagreements, however, each of them involving the concept of
mental models. The first disagreement concerns the extent to which
mental models are formed automatically, the second the extent to which
mental models are "life-like,"  and the third has to do with McKoon and
Ratcliff's interpretation of their Experiment 5, in which they purport
to demonstrate that previous data supporting mental-model theory are
best viewed as demonstrating the operation of "salience."

1.1.2  Virtually all pertinent theories postulate that comprehension
results in the creation of some form of cognitive representation.
Debate focuses on the processes used to derive the representation, the
nature of the representation (e.g., propositional or imaginal), and the
information that is represented. Among mental-model theories of
comprehension, the greatest consensus concerns the last mentioned area.
That is, a mental model is a representation of what the text is about,
a representation of the events, objects, or processes described by the
text, rather than of the words, sentences, or structure of the text
itself. Thus, a mental model is likely to capture the spatial (Bryant,
Tversky, and Franklin, 1992; Glenberg, Meyer, & Lindem, 1987; Morrow,
Greenspan, & Bower, 1987) and temporal structure (Glenberg & Langston,
1992) of the events described in the text. Most mental-model theories
do not deny the possibility of other levels of representation which may
well be required for successful grammatical and referential processes.
The basic intuition, however, is that these other levels alone do not
fully capture the meaning of the term "comprehension."

1.2.0  ARE MENTAL MODELS FORMED AUTOMATICALLY?

1.2.1  McKoon and Ratcliff (1992) state that "The most important claim
of many `mental model' theories of text comprehension is that the
mental representation of a text automatically depicts the events
described by the text in a `life-like' way," (p. 456) and "readers
automatically construct a full representation of the real-life
situation described by a text" (p. 458). To what extent are these
statements correct? Note that there are really two components to be
addressed: the extent to which mental models are formed automatically
and the extent to which they should be considered a "full"
representation of the situation described by the text.

1.2.2  We do not think most mental modelers would hold that mental
models are constructed automatically (at least not all of the time).
Although there is no universally accepted definition of automatic
processing, some broadly agreed upon characteristics of automatic
processing are that it is accomplished without intention, that it is
unavoidable, and that it requires a minimum of effort. Because reading
for deep comprehension is unlikely to be automatic in this sense, and
because mental models are just what deep comprehension is all about,
few mental modelers would propose that the construction of a mental model
is automatic. Furthermore, many of the experimental demonstrations of
mental-model-like effects have used procedures that engender
nonautomatic processing. For example, Morrow et al.'s (1987) subjects
were required to memorize a spatial layout before reading the test
texts. Bryant et al.'s (1992) subjects were allowed to study the
initial portions of a text "for as long as they wished." Radvansky and
Zacks (1992) used a multi-trial study procedure. Glenberg and Langston
(1992) demonstrated that for some texts, the construction of mental
models seems to require that the text be accompanied by illustrations;
thus, mental model construction was certainly not automatic for the
unillustrated texts. Given this brief survey of methods for inducing
mental models we can see that it is not "an important claim" of
mental-model theory that mental models be constructed automatically.

1.2.3  In contrast to what McKoon and Ratcliff's have suggested,
Glenberg, Kruley, and Langston (in press) have proposed several
preconditions for the construction of mental models. For example,
readers must be attending to the words and grammatical relations and
must have the appropriate lexical and grammatical knowledge. They must
also have some domain knowledge to be able to construct a
representation of the events, for example, knowledge about the
characteristics of the entities producing the event, the types of
relations these entities can enter into, and how these relations can be
represented. In highly familiar domains, this knowledge may be readily
available, thereby facilitating the construction of mental models. This
seems to have been the case in the experiments reported by Glenberg et
al. (1987) and Sanford and Garrod (1981). When this knowledge is not
readily available, however, constructing mental models may be
cognitively laborious.

1.3.0  ARE MENTAL MODELS LIFE-LIKE?

the 1.3.1  Some mental model theorists do claim that mental models are
life-like, but in the particular sense that there are some analogical
correspondences between mental models and the situations that are
represented. The correspondences are between real-world numbers and
numbers in the representation as well as between real-world spatial
relations and spatial relations in the representation. Thus, when a
text is about a single actor, the corresponding mental model will have
an entity corresponding to that actor. In a text in which two actors
are foregrounded (e.g., in a conversation), the mental model will have
entities corresponding to both participants. At least some spatial
relations between entities also appear to be represented analogically
(see Bryant et al., 1992 and Morrow et al., 1987). These claims of
limited analogical correspondence contrast with McKoon and Ratcliff's
statement that mental models are "a full representation of the
real-life situation described by a text."

1.3.2  Part, at least, of the difference between McKoon and Ratcliff's
characterizations and what mental modelers actually propose may be
traced to differences in representational assumptions and the language
used to describe those assumptions. According to Palmer (1978), in a
propositional representation, all relations must be explicitly
encoded. For example, suppose that a text describes three sticks, A, B,
and C, and that the text mentions that A is taller than B and B is
taller than C. A propositional representation of the relations among
the sticks may consist of the propositions "taller than (A, B)" and
"taller than (B,C)."  The fact that stick A is taller than stick C is
not represented until it is explicitly encoded as "taller than (A,C)."
In a representation in which number and the taller-than relation are
represented analogically, the situation is different. That is, given
that it is (analogically) represented that A is taller than B and that
B is taller than C, the fact that A is taller than C is then also
represented. Thus, the analogical representation is "life-like" in the
limited sense that the analogical representation of "taller than"
preserves the property of transitivity. The "inference" that A is
taller than C is likewise made automatically, in that it is part of the
representation. Nonetheless, there is no claim that other aspects of
the situation need be represented fully or analogically. Thus, the
exact heights of the sticks, their thicknesses, surface
characteristics, densities, chemical compositions, etc., need not be
represented to qualify the representation as a mental model.

1.3.3  In light of these distinctions, consider a prediction attributed
by McKoon and Ratcliff to mental model theory: "[A] mental model for a
text such as the actress fell from the fourteenth story should include
the inference that she died. It would not be reasonable, from the
mental model point of view, to leave her suspended in mid air" (p.
457). If mental models were indeed a complete representation of a real
situation, it would perhaps be reasonable to infer from the
representation that the actress would die. Mental models are not
complete representations, however. Given the brief segment of text, a
mental model would probably include one representation corresponding to
the actress and one corresponding to the fourteenth floor. If the model
represented spatial relations analogically then the actress would be
(in the analogical representational medium) below the fourteenth
floor. However, there is nothing in mental-model theory to suggest that
readers will necessarily fill in the rest of the details and have the
actress plunge to her death. The text could in fact continue in a
variety of ways that would deny just that possibility (e.g., the
actress is saved by falling into a swimming pool). Thus, McKoon,
Ratcliff, and Seifert's (1989) data demonstrating that most people do
not completely infer the actress's death is consistent with both the
minimalist and the mental-model hypotheses.

1.3.4  If people do not automatically infer outcomes such as the death
of the actress, why do they bother to construct mental models?
Primarily because understanding a text consists of creating a
representation of what the text is about, that is, a mental model.
Also, mental models allow inferences (e.g., that the actress will die)
if one is asked to infer what is likely to happen. Thus, if the text
continues with "She died instantly," the mental model is useful in
understanding how that could have come about. It is important to note
that if the text continues with "A gust of wind blew her onto a ledge
at the thirteenth floor," the mental model is also useful in
understanding how this could have come about.

1.4.0  CAN "SALIENCE" EXPLAIN THE GLENBERG, MEYER, AND LINDEM (1987) RESULTS?

1.4.1  One piece of evidence consistent with a mental-model level of
representation was provided by Glenberg, Meyer, and Lindem (1987). The
subjects in their experiment read short texts describing ordinary
situations. An important feature of the texts was that their
propositional structure was quite different from that of the situation
described by them. Consider as an example a text which describes a main
actor, John, preparing for a marathon. The critical sentence in the
associated version of the text is "After doing a few warm-up exercises,
he put on his sweatshirt and went jogging."  Note that the associated
version of the text describes a situation in which the main actor,
John, is associated spatially with a target object, the sweatshirt.
The critical sentence in the dissociated version of the text describes
a situation in which the main actor and the target object are spatially
dissociated, "After doing a few warm-up exercises, he took off his
sweatshirt and went jogging." If subjects are tracking the structure of
the situation, that is, building a mental model from the text, then we
would expect to find evidence of a stronger connection between the
actor and the target object in the a condition than in the dissociated
condition. Furthermore, Glenberg et al. assumed that subjects would
keep the main actor foregrounded because the actor is referred to by a
pronoun (see, for example, Sidner, 1982). Because the target object
remains spatially close to the actor in the situation described by the
associated condition, it too should remain foregrounded and easily
accessible. In fact, after the last mention of the target object and a
one-sentence delay, subjects were faster in making a recognition
response to the name of the target object in the associated condition
than in the dissociated condition. There was little difference in
responding immediately after the last mention of the target object,
presumably because the target object was foregrounded (due to recent
mention) in both cases.

1.4.2  Note that these results are difficult to reconcile with the view
that comprehension results in a propositional representation of the
text. In the associated condition, the main actor and the target object
are conjoined in a single proposition, something like, "on (John,
sweatshirt)."  In the dissociated condition, the two are also conjoined
in a single proposition, something like, "off (John, sweatshirt)."
Glenberg et al. reasoned that if these propositions formed the only
representation then there would be no reason to predict differential
speed of responding to the target object.

1.4.3  McKoon and Ratcliff (1992) proposed that the Glenberg et al.
results do not reflect the construction of a mental model, but instead
are due to "salience." That is, in the associated condition, the target
object is viewed as more salient -- "more relevant to the topic of
[the] discourse" (p. 460) --  than in the dissociated condition. According
to McKoon and Ratcliff, it is this difference in salience that produces
the differential speed of responding, not a difference in the structure
of a mental model.

l.4.4  Before we introduce new data (PART II, see Sections 2.6 and 2.9,
Experiments 1 and 2 for complete details), there are already two
reasons to question the salience hypothesis. First, the very notion of
salience may well derive from a mental model. That is, what make an
object or event salient are its relations to the other objects and
events and our knowledge about what those relations imply for further
action. Yet for nonscript texts that knowledge may well be derived from
constructing a mental model to ascertain the relations. Furthermore,
once a model is constructed, it may be "run" (Gentner and Stevens,
1983) to determine future states or implications. Second, the salience
account has difficulty addressing an interaction found by Glenberg et
al. (1987). The associated/dissociated variable affected responding to
the probe (e.g., sweatshirt) when the probe was delayed by one or two
sentences. However, when the probe occurred immediately after the
sentence in which the target object was introduced, there was no effect
of the associated/dissociated variable. If differential salience of
objects is conferred by propositional relations (rather than by mental
models), differences in salience should be measurable immediately after
reading the sentence introducing the object because the relevant
propositions should be formed by this point. It should be noted,
however, that the force of this second point is reduced by possible
effects of verbatim representations.

1.4.5  These arguments notwithstanding, McKoon and Ratcliff conducted
an experiment (their Experiment 5) to test the salience account. They
modified the Glenberg, Meyer & Lindem (1987) texts so that each
critical sentence introduced two objects, the original target object
(which will be referred to as the GM&L target) and an initial location
for the GM&L target (the initial location will be referred to as the
M&R target). McKoon and Ratcliff write about the M&R target that,
"[T]he location was something that could not move with the main
character" (p. 460). An example of the initial portion of a modified
text in the associated condition is: "John was preparing for a marathon
in August. After doing a few warm-up exercises, he put on a sweatshirt
from the laundry and went jogging." In this example, "laundry" is the
M&R target

1.4.6  McKoon and Ratcliff reasoned that if subjects are building
mental models, the associated/dissociated variable should affect
responding to the GM&L target but not to the M&R target because the
latter "could not move with the main character." However, suppose the
associated/dissociated variable differentially confers salience on the
propositions encoding the GM&L target. Remember that the M&R target is
a location for the GM&L target; hence they are encoded in the same
proposition. Thus, if the associated/dissociated variable confers
salience on the proposition encoding the GM&L target, the salience of
the M&R target should also be affected by the associated/dissociated
variable. This is exactly what McKoon and Ratcliff found, supporting
the salience interpretation of both their own data and the original
Glenberg et al. data.

1.4.7  Although these data appear compelling, for the reasons mentioned
previously we questioned the salience hypothesis and conducted two
experiments to test it further. In the first experiment (see PART II,
paragraphs 2.l.0 to 2.8.4 for details), we asked people to judge the
salience of the GM&L and the M&R targets. The judgments were reliable,
but they were not strongly related to the associated/dissociated
variable. In fact, for the M&R targets, subjects judged the target to
be more salient in the dissociated condition than in the associated
condition, just the opposite of what is predicted by the salience
hypothesis.

1.4.8  If differential salience does not account for the results of
McKoon and Ratcliff's Experiment 5, what does? In particular, why
should recognition of the M&R target depend on the
associated/dissociated status of the GM&L target? One hypothesis is
derived from a simulation of mental models that we are developing
(Glenberg, Kruley, & Langston, in press). In this simulation, mental
models are constructed using pointers arrayed in a spatial working
memory. Distances between pointers are representationally meaningful in
that distance in the model has an analogical correspondence to
distances in the situations being described (or to other salient
dimensions in the text that are assigned to spatial dimensions for
purposes of constructing a mental model). By manipulating the pointers,
subjects can derive inferences that are not explicit in the text.
Furthermore, pointers in the mental model activate corresponding
propositions in memory. For example, as one continues to think about
John's sweatshirt by keeping a pointer in the mental model, other
information about the sweatshirt (encoded propositionally) remains
active too.

1.4.9  Given this background, consider what happens when a subject
keeps a GM&L target (such as sweatshirt) in the mental model in the
associated condition. The representation of the GM&L target is
activated, as is propositional information about the target. One such
proposition is that the GM&L target has an initial location, namely,
the M&R target. Thus, information about the M&R target is also
activated by virtue of the GM&L target remaining in the mental model.
Similarly, when the GM&L target is not in the mental model (in the
dissociated condition), there is a reduction in activation of the GM&L
target and of associated propositions, including the proposition
encoding the M&R target. Thus, although the associated/dissociated
variable is manipulated with reference to the GM&L target, it may well
have a mediated effect on the accessibility of the M&R target.

1.4.10  One would expect any mediated effect to be rather small.
However, those effects should be enhanced to the extent that other
components of the mental model also have a mediating effect on the
accessibility of the M&R target. One other source of mediating effects
is the association of the M&R target and the main actor. McKoon and
Ratcliff state that the M&R target was a location "that could not move
with the main character"; thus, as the main actor moves, the M&R target
should always be dissociated. However, in preparing the materials for
Experiment 1, we noticed several instances in which this did not appear
to be the case. For example, in one text about a fisherman in a
rowboat, the critical dissociated sentence was, "He had just finished
eating when the wind blew the empty bag out of his hand and into the
water."  The GM&L target is "bag," and the M&R target is "hand" (the
original location of the bag). Clearly, the M&R target is a location
that will move with the main actor. In a text about an athlete
preparing to play racquetball, the dissociated sentence was, "He
changed his shirt and took his watch off his wrist."  The GM&L target
is "watch," and the M&R target is "shirt."  The intended referent for
"shirt" was apparently the one the actor removed in the act of
changing. However, when the word "shirt" is used as a recognition
probe, it refers equally well (or better) to the new shirt that the
main actor is wearing. A third example comes from a text about a
businessman in a restaurant. The dissociated sentence is "Leaning back
in his chair, he glanced at a woman from the next table who was walking
out of the room." The GM&L target is "woman" and the M&R target is
"table." Although the GM&L target is dissociated from the main actor by
virtue of the woman leaving the room, the M&R target, "table," remains
close to the main actor.

1.4.11  In all, we identified six texts in which the M&R target is not
clearly dissociated from the main actor, so that the main actor can
provide a type of mediated priming. Removing the six texts has a
selective effect. In particular, for the M&R target, removing those
texts greatly reduces the associated/dissociated effect (from 51 msec
to 20 msec), while slightly enhancing the effect for the GM&L target
(from 62 sec to 81 msec).

1.4.12  Given these data, an attempt to replicate McKoon and Ratcliff's
Experiment 5 using rewritten texts was in order (see Section II,
paragraphs 2.9.0 to 2.11.4 for details). We started with the texts used
by McKoon and Ratcliff, but we rewrote freely to accomplish several
goals. First, the M&R target was always a location for the GM&L
target. Second, that location was always introduced in a prepositional
phrase immediately following the GM&L target. Third, the combination of
the critical sentence and the following sentence (the one immediately
preceding the recognition probe) always dissociated the main actor and
the M&R target. Of course, whether or not the main actor was
dissociated from the GM&L target was the major variable in the
experiment. Fourth, we ensured that there was only one main actor (in
some of the texts used by McKoon and Ratcliff there were several actors
referred to by pronouns, thus granting multiple actors "in focus"
status).

1.4.13  Consider the predictions for the recognition times for the M&R
and GM&L targets based on the mental-model account and the salience
account. The mental-model account predicts a large effect of the
associated/dissociated variable for the GM&L target (the one that is
associated or dissociated from the main actor), but a much smaller
(mediated) effect for the M&R target. The salience account is based on
the assumption that the associated/dissociated variable confers
salience on both the GM&L target and the M&R target, the original
location of the GM&L target. Thus the associated/dissociated variable
should affect both the GM&L target and the M&R target.

1.4.14  Although not all aspects of the data were as clean as one would
like, the main result was clear. With the rewritten texts, the effect
of the associated/dissociated variable (46 msec measured by texts) was
significant for the GM&L target whereas the effect (0 msec, measured by
texts) was not significant for the M&R target.

1.5.0  GENERAL DISCUSSION

1.5.1  The salience account of the associated/dissociated effect
appears to be incorrect in three respects. First, it cannot easily
handle the absence of an effect immediately after the target is
introduced (Glenberg et al., 1987). Second, reliable, direct judgments
of salience do not conform to the associated/dissociated effect for the
M&R targets (Experiment 1). Third, after the stimulus materials were
vetted for inappropriate texts, the effect of the variable on the M&R
target disappeared (Experiment 2 and reanalysis of McKoon and
Ratcliff's data).

1.5.2  The data are consistent with the mental-model theory offered by
Glenberg et al. (in press). While reading, subjects construct a mental
model that may represent space analogically. Significantly, when the
text describes central features (e.g., the main actor and a target
object) as spatially close, they are related more strongly in the
mental model than when the text describes the same features as farther
apart. Thus, when the main actor is foregrounded by virtue of
pronominal reference, the spatially close target object (in the
associated condition) tends to remain highly accessible; when the
target is described as spatially far from the target (in the
dissociated condition), it is less accessible.

1.5.3  In concluding PART I, we note that we have few disagreements
with McKoon and Ratcliff's proposals. The minimalist hypothesis is
about the conditions under which inferences are made automatically.
Although there is no universally accepted definition of automatic
processing, we assume McKoon and Ratcliff mean that these inferences
are drawn without intention, that they are unavoidable, and that they
require a minimum of effort. As we have noted elsewhere (Glenberg et
al., in press; Glenberg & Langston, 1992), we think the construction
and maintenance of a mental model need not be automatic in these
senses. Instead, forming mental models may well be strategic, and in
unfamiliar domains, it may well be quite effortful. Also, some research
demonstrating mental model effects during comprehension (Bryant et al.,
1992; Morrow et al., 1987) involves extended learning procedures that
would be difficult to view as automatic.

1.5.4  In addition, the minimalist hypothesis is meant to apply when
reading "In the absence of specific, goal directed, strategic
processes" (McKoon and Ratcliff, 1992, p. 440). Because the sort of
reading for good comprehension that requires mental model construction
is often goal directed and strategic, those situations that lead to
mental models are just those of little concern to tests of the
minimalist hypothesis.

1.5.5  Finally, we agree with McKoon and Ratcliff that readers
generally do not construct a "full" representation of the situation
described by the text. As we argued in the introduction, however, this
does not exclude the possibility that readers are constructing mental
models.

1.5.6  The one remaining disagreement seems to be settled by the
findings reported in PART II of this article. McKoon and Ratcliff argue
that the findings reported in Glenberg et al. (1987) reflect the
operation of "salience." In contrast, the evidence is now strong that
the effect of the associated/dissociated variable is due to the
construction of a mental model. In fact, it may well be that
representation in a mental model is one source of highly available
information that supports the judgment of salience and the process of
drawing minimal, automatic inferences.

-------------------------------------------------------------------------
[end of PART I; PART II will appear in next posting but in the archive
it will be stored as one single article]