A
nnu. Rev. Psychol. 2001. 52:223­47
Copyright c 2001 by Annual Reviews. All rights reserved
THINKING
Arthur B. Markman
Department of Psychology, University of Texas, Austin, Texas 78712;
e-mail: markman@psy.utexas.edu
Dedre Gentner
Department of Psychology, Northwestern University, Evanston, Illinois 60208;
e-mail: gentner@nwu.edu
Key Words reasoning, analogy, domain-specific processes,
mental models, learning
s Abstract Reasoning processes allow the human cognitive system to go beyond
the information readily available in the environment. This review focuses on the pro-
cesses of human thinking, including deductive reasoning, induction, mental simulation,
and analogy. We survey recent trends across several areas, including categorization,
mental models, cognitive development, and decision making. Our chief organizing
principle is the contrast between traditional approaches that focus on abstract logical
reasoning and a number of current approaches that posit domain-specific, knowledge-
intensive cognition. We suggest that some instances of domain-specific cognition result
from domain-general processes operating on domain-specific representations. Another
theme is the link between reasoning and learning. We suggest that learning typically
occurs as a byproduct of reasoning, rather than as an end in itself.
CONTENTS
INTRODUCTION . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 224
STRUCTURAL ALIGNMENT AND MAPPING. . . . . . . . . . . . . . . . . . . . . . . . . . 226
Learning and Reasoning: The Shift from Active Processing to Storage . . . . . . . . . 228
MENTAL MODELS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 228
Logical Mental Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 229
Causal Mental Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 229
Relating Mental Models to Theories . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 231
The Development of Theories and Mental Models . . . . . . . . . . . . . . . . . . . . . . . 233
CATEGORIZATION . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 235
Similarity and Categorization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 235
Inference from Categories . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 237
Feature Centrality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 238
Domain-Specific Approaches to Categorization . . . . . . . . . . . . . . . . . . . . . . . . . 239
SUMMARY AND CONCLUSIONS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 241
0066-4308/01/0201-0223$14.00 223

2
24 MARKMAN GENTNER
INTRODUCTION
We have called this chapter "Thinking" rather than using the more traditional title
"Reasoning" because we think it important to go beyond the traditional deductive
logic focus of the term reasoning and include other ways of arriving at new con-
clusions, including induction and analogy. In this paper we survey recent trends in
the field across several areas, including categorization, mental models, cognitive
development, and decision making. A theme that emerges frequently in recent
research is an interest in going beyond lab-oriented paradigms to investigations
of real-life cognitive activities. A related theme is the relation between abstract
logical reasoning and concrete domain-specific reasoning--or to put it another
way, between abstract and knowledge-based approaches to reasoning. The idea of
reasoning is often equated with the notion of purely logical processes that operate
independent of content. In popular imagery, icons such as Mr. Spock and HAL
draw their intellectual power from pure logic. Likewise, within psychology the
study of reasoning has focused largely on the use of content-independent logical
rules (e.g. Johnson-Laird & Byrne 1991, Rips 1994). However, other research
shows that the content being reasoned about influences people's reasoning ability,
even for tasks to which logical rules are applicable (e.g. Cheng & Holyoak 1985,
Cosmides 1989, Wason & Johnson-Laird 1972). Partly because of these findings,
there has been considerable recent interest in how people learn and use rich domain
representations such as theories and mental models.
As an example of the kinds of phenomena that need explanation, Bassok et al
(1998) asked college undergraduates to write addition and division word prob-
lems. Presumably, college students are experts at addition and division; it should
be straightforward for them to think abstractly about simple arithmetic. However,
the content of the problem influenced the form of the word problem written. When
asked to write a problem involving two members of the same category (e.g. apples
and pears), the students found it easier to write an addition problem than a divi-
sion problem. In contrast, when given members of thematically related categories
(e.g. apples and baskets), they found it easier to write a division problem than an
addition problem. Thus, even in the seemingly abstract domain of mathematics,
cognitive performance is affected by domain content.
Such effects have led some researchers to suggest that content and context
are fundamental to reasoning (Newell & Simon 1972). Some theorists assert that
human learning is conservative, with representations that are tied to the initial
learning situation (Gentner 1989, Medin & Ross 1989). Indeed, some have taken
the extreme opposite of the logicist view, arguing that there is no utility to a gen-
eral notion of representation or process. One such view is the situated cognition
approach that assumes all thinking is fundamentally context-governed (Suchman
1987). A related position is the embodied cognition view, according to which
cognitive processes are optimized to mesh with particular sensorimotor activi-
ties (Glenberg 1997, Pfeifer & Scheier 1999). For example, Glenberg reviewed
evidence that some spatial reasoning tasks are facilitated by specific motor
movements consistent with route-following.

T
HINKING 225
Another approach to domain-specific reasoning is modularity theory, which
assumes that human learning and development requires innate modules for cer-
tain domains such as physical causality or psychology (e.g. Hirschfeld & Gelman
1994). Finally, another approach that postulates domain-specific reasoning is evo-
lutionary psychology, which posits domain-specific modules selected by evolution
to solve complex reasoning tasks (Tooby & Cosmides 1989).
These views have had the salutary effect of pointing to phenomena neglected
by traditional approaches to cognition. However, in the extreme, the focus on
domain-specific modules can lead to abandoning the search for general processes
in favor of particularistic descriptions. Our reading of the findings supports a
more moderate conclusion. We suggest that both domain-general and domain-
specific reasoning have a place in psychology and suggest three lines of recon-
ciliation. First, we invoke a distinction proposed by Newell & Simon (1972)
between strong and weak methods of reasoning. Weak methods are general strate-
gies that can operate without special knowledge of a domain, such as means-ends
analysis or logical inference rules like modus ponens. Strong methods make in-
tensive use of represented knowledge, as in reasoning by example. Weak meth-
ods are valuable because of their generality; they provide a means of operat-
ing on novel or knowledge-poor domains. However, Newell and Simon asserted
that strong methods are typically superior when the appropriate knowledge is
present. Many recent systems have focused on architectures that permit both
general and specific knowledge to be used. For example, Sloman (1996) re-
viewed evidence that people apply specific knowledge to problems in order to
answer them quickly, and that this knowledge can even compete with rule-based
processes.
The second point of contact is that many instances of domain-specific
cognition result from domain-general processes operating on domain-specific rep-
resentations. Domain specificity can thus be captured without abandoning the
idea of domain-general cognitive processes. For example, in cognitive architec-
tures like Anderson's ACT system (Anderson 1993, Anderson & Lebiere 1998),
it is assumed that the structure of semantic memory can strongly influence the
performance of the model. A third point of connection is that the distinction be-
tween abstract and concrete cognition often behaves as a continuum rather than
as a dichotomy. Furthermore, the continuum is learning related. As discussed in
the next section, alignment and abstraction processes can result in a natural transi-
tion from domain-specific to domain-general reasoning (Cheng and Holyoak 1985,
Gentner & Medina 1998). This last trend is an instance of another basic theme that
emerges from our survey: In many natural contexts, learning occurs as a byprod-
uct of reasoning, rather than as an end in itself. For example, Ross (1997) showed
that people's representations of categories are influenced by the way they use the
categories in problem solving.
The plan of this chapter is as follows. We first lay out structure mapping, a
general process that can capture many domain-specific effects. We then cover re-
search in mental models and theories, traditionally an arena of rich domain-specific
cognition. Finally, we survey research in categorization and category-based

2
26 MARKMAN GENTNER
induction as a case study of the ways domain-general and domain-specific in-
formation interact in cognitive processing.
STRUCTURAL ALIGNMENT AND MAPPING
As noted above, we need accounts of reasoning processes that can deal with rich
domain-specific knowledge. In each of these cases, we need to find a way to allow
information about the new situation to be related to background knowledge.
Comparison is a domain-general process that allows the detection of com-
monalities and differences in a pair of situations or domains. Many models of
similarity have been proposed (e.g. Shepard 1964, Tversky 1977). We focus on
recent accounts of similarity that are structurally sensitive, and that allow for com-
plex mental representations and hierarchical relational structures (Falkenhainer
et al 1989, Gentner & Markman 1997, Goldstone 1994, Holyoak & Thagard 1989,
Hummel & Holyoak 1997, Medin et al 1993). In particular, structure-mapping
theory (Gentner 1983) has been used to model analogy, similarity, and metaphor
in perceptual and conceptual tasks (Gentner & Markman 1994, 1997; Markman &
Gentner 1993a,b, 1996, 1997).
Structure mapping involves two processes: structural alignment and inference
projection. The idea behind structural alignment is that comparing structured
representations requires a process that is sensitive to similarities, not only be-
tween the elements but also between the connections between elements. The struc-
tural alignment process matches semantically similar relations, but beyond this it
seeks maximal structurally consistent systems of matches. Structural consistency
involves two constraints: parallel connectivity and one-to-one correspondence.
Parallel connectivity requires that if a pair of predicates is placed in correspon-
dence, then their arguments must also be placed in correspondence, and one-to-one
correspondence requires that an element in one representation can match to at most
one element in the other representation.
For example, consider the following analogy:
(1) Fred loves Mary,
so Fred bought Mary cookies.
(2) Joyce loves Fred,
so Joyce bought Fred candy.
Fred let Joyce drive his car.
A process that was not sensitive to relational structure would arrive at the
correspondences Fred(1) Fred(2), cookies candy, and so on. Structural alignment
invites the arguments of matching relations to be placed in correspondence. Thus,
once the relational correspondence loves(1) loves(2) is found, Fred(1) is mapped
to Joyce and Mary is mapped to Fred(2). Further, by one-to-one correspondence,
once the correspondence Fred(1) Joyce is determined, Fred(1) cannot also match
with Fred(2).

T
HINKING 227
A further tenet of structure-mapping is the systematicity principle (Gentner
1983, 1989), which posits that matches between systems of relations connected
by higher-order constraining relations are preferred to matches between isolated
relations. Such relations include causality or means-end relations in conceptual
domains and symmetry or monotonicity in perceptual arenas. This claim is sup-
ported by studies that have demonstrated that pairs are seen as more similar when
they share systematic relational matches than when they do not (e.g. Gentner et al
1993, Lassaline 1996, Markman & Gentner 1993b) and that matches are seen as
more central to an analogy when they are connected to other matching assertions
than when they are not (Clement & Gentner 1991).
Systematicity is manifest not only in which commonalities emerge from a com-
parison but also in which differences are noticed (Markman & Gentner 1993a). In
particular, alignable differences tend to be highly salient in similarity comparisons.
These are differences that result when nonidentical elements are placed in corre-
spondence by virtue of their connections to the common structure. For example, in
the example analogy, cookies and candy will be placed in correspondence, because
they fill the same role. Thus, they are an alignable difference. In contrast, Fred's
car, which has no correspondence in the first set, is a nonalignable difference--a
difference not connected to the common system that has no correspondence in the
other domain.
Commonalities are more important in judgments of similarity than are differ-
ences (Sj¨oberg 1972, Tversky 1977). Structural alignment extends this observation
to alignable differences, assuming that alignable differences are more prominent
than nonalignable differences in similarity tasks. As evidence for this supposi-
tion, it has been shown that alignable differences influence similarity comparisons
more than do nonalignable differences (Markman & Gentner 1996), are more likely
to be listed than nonalignable difference in difference-listing tasks (Markman &
Gentner 1993a), are more fluently produced in speeded tasks (Gentner & Markman
1994), and serve as better recall cues for information presented in a comparison
than do nonalignable differences (Markman & Gentner 1997).
Systematicity also influences which new inferences are drawn from a com-
parison (Bowdle & Gentner 1997, Clement & Gentner 1991, Markman 1997,
Spellman & Holyoak 1996), thus providing a means of constraining possible in-
ferences. Information that is connected via higher-order constraining relations,
such as causality, to the common system is more likely to be projected to the tar-
get than is other knowledge about the base. These candidate inferences serve as
informed guesses.
Recent research has begun to investigate the real-time processes of compar-
ison. There is evidence for an initial alignment process that is symmetric be-
tween base and target, followed by a directional inference process (Falkenhainer
et al 1989, Wolff & Gentner 2000). The alignment between two representations
is computed via a local-to-global process that begins with individual element
matches and then imposes global consistency constraints (Goldstone & Medin
1994).

2
28 MARKMAN GENTNER
To summarize, comparison can be viewed as the alignment of structured
representations. This process begins with local matches and gradually arrives at
the maximal structurally consistent system of correspondences. The comparison
process can generate new conclusions in at least four ways: (a) it promotes a focus
on common connected relational systems that can serve as useful abstractions;
(b) it highlights alignable differences of a pair, which are likely to be relevant;
(c) it invites candidate inferences; and (d ) it sometimes leads to re-representations
of one or both domains to increase their similarity (Gentner & Wolff 2000). These
processes contribute to learning as well as to online reasoning.
Learning and Reasoning: The Shift
from Active Processing to Storage
It is often easier to retrieve something from memory than to derive it by reasoning
processes. This idea is evident in many theories of automaticity. For example,
Logan (1988, 1996) suggested that in the initial stages of learning a skill, peo-
ple obtain solutions by carrying out an algorithm. For example, they might learn
to add two numbers by starting with the larger one and counting on to add the
other. With sufficient practice, they come to be able to retrieve previously stored
answers. Siegler & Shipley (1995) observed this progression in children learning
to do arithmetic, who shifted from overtly counting on their fingers to retrieving
answers from memory.
A similar process of shifting from active computation to memory retrieval can
be seen in metaphor comprehension (Bowdle & Gentner 1999; Gentner & Wolff
1997, 2000). When a novel metaphor is presented (e.g. that lawyer is a boa con-
strictor), an interpretation (e.g. the lawyer is inexorable) is actively computed via
alignment processes. However, if the metaphor base is repeatedly used in the same
nonliteral way, its metaphorical interpretation may be stored as an alternate word
sense. Such stock metaphors can behave like conventional categories (Glucksberg
et al 1997) and do not need to be recomputed.
People's propensity to cache frequently encountered solutions can be used to
predict their future behavior. In consumer psychology, the single best predictor of
a person's purchases in a category is "brand loyalty"--the prior probability dis-
tribution of purchases of products in that category (e.g. Guidagni & Little 1983).
People tend to buy again what they bought in the past. Our pattern of storing the
results of frequently experienced processes is a form of cognitive economy, but
it sometimes works against us. In the classic Einstellung or "set" effect noted by
the Gestaltists, people's continued use of a past successful solution hampers their
ability to see a better solution to a problem.
MENTAL MODELS
A mental model is a representation of some domain or situation that supports
understanding, reasoning, and prediction. There are two main approaches to the
study of mental models. One approach seeks to characterize the knowledge and

T
HINKING 229
processes that support understanding and reasoning in knowledge-rich domains.
The other approach focuses on mental models as working-memory constructs that
support logical reasoning (Johnson-Laird 1983, Johnson-Laird & Byrne 1991).
Although we are primarily interested in causal mental models, we first briefly
discuss logical mental models.
Logical Mental Models
There is considerable research on how people perform logical reasoning tasks
(Evans & Over 1996; Johnson-Laird 1983, 1999; Rips 1994). One prominent
explanation for the way people perform logical reasoning is that they form a
mental model of the situation (Johnson-Laird 1983). In this view, when faced with
a logic problem, the solver imagines a set of tokens organized so as to capture
the relevant aspects of the premises. For example, given the premise "all archers
are bakers," the solver might imagine a group of people. All those with arrows
would also have cakes, but there would be others with cakes and no arrows. With
each additional premise, the model is adjusted, and if more than one model is
consistent with the premises, then alternate models are constructed. After all the
premises have been presented, the resulting model(s) are summarized to give
the conclusion. The difficulty of a given problem is determined by the number
of different mental models that have to be constructed. This approach has been
applied to a variety of logical reasoning situations, including classical syllogisms,
multiply quantified statements, and inductive arguments (Johnson-Laird 1983,
1999; Johnson-Laird & Byrne 1991). However, some researchers have argued
that logical mental models fail to capture aspects of human processing that are
captured in propositional models (Rips 1986, 1994), or that mental models are
actually isomorphic to other forms of diagrammatic reasoning in logic (Stenning &
Oberlander 1995).
Causal Mental Models
Causal mental models are mental representations that are used in reasoning and
that are based on long-term domain knowledge or theories (Gentner & Stevens
1983). They are used to explain reasoning about physical systems and mecha-
nisms (Hegarty & Just 1993, McCloskey 1983) such as spatial representation
(Tversky 1991), human-computer interaction (Norman 1988), ecology (Kempton
et al 1994), and the development of astronomical knowledge (Vosniadou & Brewer
1992).
Causal mental models differ from logical mental models in two further ways
besides their domain of application. First, the elements in a logical mental model
are tokens rather like algebraic symbols, whereas the tokens in a causal mental
model correspond to elements of causal systems. Second, logical mental models
are created on the spot and involve only information currently active in working
memory, whereas causal mental models, even those currently active in working
memory, are assumed to draw on long-term memory structures.

2
30 MARKMAN GENTNER
Causal mental models (hereafter called mental models) are related to several
otherkindsofrepresentationalstructures(Markman1999). Schemas(orschemata)
are general belief structures. Scripts are schemas summarizing event sequences
and are characterized by a chiefly linear temporal order with limited inferential
flexibility. Naive theories or folk theories are global systems of belief, typically
encompassing larger domains such as biology. The terms mental models and naive
or intuitive theories overlap in their application, though mental models are typically
more specific in their application than are theories.
The Use of Mental Models One way people employ their mental models is to
perform mental simulations--to imagine the future trajectory of a system given a
set of initial conditions. These predictions involve qualitative estimates and often
neglect some of the information relevant to the behavior of a system. For example,
Gilden & Proffitt (1989) showed people collisions between a moving ball and
a stationary ball and asked them which ball was heavier. Making this judgment
requires attending both to the velocity of the balls and their trajectories prior to
and following the collision. The results showed that people tended to use simple
heuristics such as "the ball moving faster after the collision is lighter," or "the ball
that ricochets is lighter." Whereas these heuristics are often correct, they do not
reflect an integration of the information necessary to answer the question correctly
on all trials. A similar analysis of people's predictions about the behavior of rolling
wheels suggests that people understand the physics of forward (i.e. translational)
motion of rolling wheels, but not the physics of rotational motion (Proffitt et al
1990).
An important aspect of these mental simulations is that they are qualitative
(Forbus 1984, Kuipers 1994). That is, people reason about relative properties of
physical systems such as direction of motion, relative speed, and relative mass.
People do not estimate values of specific quantities, nor do they carry out mathe-
matical simulations of the behavior of a system. These strategies are quite pow-
erful, and have been used as the basis of computational simulations of complex
physical systems (e.g. Forbus et al 1991), but they do not require the significant
computational machinery that would be necessary to carry out detailed quantitative
simulations.
Mental simulation often involves the use of imagery. Schwartz & Black (1996)
asked people to solve gear problems such as, "Imagine a pegboard with seven
gears on it arranged in a circle so that each gear meshes with one next to it on
each side. If you rotate one gear clockwise, in which direction will the gear next
to it move?" To answer such questions people often mentally simulate the gears'
motions, sometimes with accompanying hand gestures (Hegarty & Just 1993,
Metz 1985, Schwartz & J. Black 1996, Schwartz & T. Black 1999). Over time,
people gradually learn the parity rule--that every other gear in a sequence turns
in the same direction, whereas adjacent gears turn in the opposite direction--and
shift from using imagery to solving the problems by rule. (Use of the parity rule
reveals that the above problem is a trick. Circuits consisting of an odd number

T
HINKING 231
of gears will lock, because two adjacent gears will attempt to move in the same
direction.)
Mental simulations may also have a motor component. In one set of studies,
Schwartz & Black (1999) showed people a partially filled glass and asked them to
predict how far it would have to be tilted in order for the liquid to pour out. People
performed poorly, consistent with the Piagetian finding that people have difficulty
predicting the level of water in a tilted glass (e.g. Howard 1978). However, when
those same people were asked to hold a glass with a line drawn on it at the water
level, and to tilt it until the glass would pour (with their eyes closed), they were
significantly more accurate than in the explicit prediction task. Thus, it appears
that some aspects of mental simulations involve representations that are strongly
coupled with motor movements.
To summarize, recent research in mental models suggests that people can use
mental simulations to reason about physical events. These simulations incorporate
a small number of variables about the event and are often qualitative. Because the
simulation process is effortful, with increasing experience people shift to using
learned rules or cached results. More generally, this research reveals that people
typically possess multiple models of complex systems, some highly context-bound
and others more abstract.
Relating Mental Models to Theories
As noted above, there is a close relationship between mental models and intuitive
theories. Many researchers have suggested that people's knowledge can be char-
acterized as a theory about the world (Carey 1985, Gopnik & Meltzoff 1997, Keil
1989, Murphy & Medin 1985). The word theory has been used in a number of
different ways to describe knowledge. In its loosest form, the term theory is used
for the causal knowledge that people use to infer nonobvious properties and to
explain observed patterns in the world.
Gopnik & Meltzoff (1997) have explored the possibility that children's knowl-
edge can be characterized in the same terms as scientific theories: as theoretical
knowledge that is abstract and coherent, that has explicit causal knowledge, and
that makes specific commitments about the ontology of objects. They suggest
that children create and use theories much as scientists do to predict and interpret
events in the world. This does not, of course, mean that children's understanding
is as deep as that of scientists (or even as adults). Preschool children often focus
on object properties in their categorizations and explanations and fail to perceive
functional or relational commonalities (Gentner & Medina 1998, Halford 1992).
There are some points in favor of the claim that scientific reasoning provides a
good model for reasoning in children and adults. First, as Nersessian (1999) notes,
scientific reasoning encompasses processes that go beyond hypothesis testing, such
as mental simulation and analogy, that are also found in commonsense reasoning.
Second, both scientists and nonscientists appear more willing to accept evidence
that supports their beliefs than evidence against them. Kuhn (1997) showed this

2
32 MARKMAN GENTNER
pattern in nonscientists. She first ascertained their views as to which factors matter
in a situation and then showed them counterevidence. For example, people who
believed that humor increases TV show ratings were shown pairs that differed only
in humor, but received identical ratings. People did not change their belief in the
importance of humor, but instead constructed other explanations for the finding.
Koehler (1993) found a similar pattern among scientists. Even though sci-
entists state that their prior beliefs should not influence their evaluation of new
evidence, they often set a more stringent acceptance criterion for contrary ev-
idence than for consistent evidence. Thus, both scientists and nonscientists
appear to resist negative evidence, although possibly to different degrees. Inter-
estingly, there is some divergence of opinion as to the best policy on inconsisten-
cies. Koehler suggests that on Bayesian grounds it is reasonable for scientists to
be skeptical of evidence that contradicts their beliefs. However, Dunbar (1995)
carried out long-term observations of working microbiology laboratories and con-
cluded that attention to inconsistent findings is a major predictor of success in a
laboratory.
There are also differences between scientists and nonscientists. First, scien-
tists possess deeper causal theories than novices. Chi et al (1981) showed that
expert physicists sorted physics problems into principle-based categories such as
"momentum" and "conservation of energy," whereas novices sorted them into cat-
egories with similar diagrams. Second, scientists' knowledge in the domain of
expertise is likely to be more explicit than that of novices. This greater degree of
explicitness may arise in part from the need for scientists to communicate with
other members of the scientific community, and from the specific feedback that
scientists get about their ideas. Consistent with this proposal, Chi & VanLehn
(1991) found that students learning science benefit significantly from generating
explicit explanations for themselves as they study.
Third, although both scientists and nonscientists show resistance to data that
are inconsistent with their prior beliefs, scientists are trained to seek such data:
e.g. to include in their studies conditions that could disconfirm the hypotheses. In
contrast, nonscientists have a strong confirmation bias when testing hypotheses
(Evans & Over 1996, Klayman & Ha 1987). The Wason 246 task is a classic
demonstration of this confirmation bias. People are told that the sequence 2­4­6
was generated using a rule, and asked to determine the rule by producing further
elements. The typical strategy is to guess the rule (such as "increasing by 2")
and then test only sequences that satisfy the rule, thereby collecting confirma-
tory evidence. Most people fail to discover the correct rule ("any increasing se-
quence"), because they do not test sequences that would disconfirm their hypoth-
esized rule.
In summary, there are commonalities in reasoning styles between scientists and
nonscientists, and the differences are often of degree, not kind. However, the com-
bined force of these differences--such as greater explicitness of knowledge and
greater commitment to seeking disconfirming evidence--may lead to a substantial
overall difference in practice.

T
HINKING 233
The Development of Theories and Mental Models
One route to understanding domain-specific reasoning is to study its development.
Developmentalists have explored the way in which domain knowledge is acquired
and structured in key domains such as mathematics, physics, and biology. Much
of this research is motivated by the question of whether infants are born with
domain-specific modular capacities for learning and reasoning. Just as linguistic
development has been posited to be the result of an innate language acquisition
device, so too might other conceptual abilities reflect an innate endowment for
particular domains such as naive physics or a naive biology. For example, Leslie
(1991) has proposed a specific theory of mind module and posited that autism is
the result of a defective theory of mind module. Some theorists propose a combi-
nation of domain-specific and domain-general attentional capacities and learning
processes (Baillargeon 1994, Carey 1985). At the other extreme, some theorists
argue that purely general learning principles can account for the development of
domain-specific modules (Elman et al 1996, Thelen & Smith 1993). According to
these theorists, modules are the result of learning. At present, the issue is not re-
solved. Regardless of the outcome of the debate between nativists and empiricists,
however, this body of work is of immense value for understanding how domain
knowledge is acquired.
The development of biological knowledge is an arena of intense current
interest. Carey (1985) probed children's understanding of the properties of animals
(e.g. dogs or worms), plants, and inanimate objects. She asked about biological
properties (bones and a heart), biological behaviors (eating, sleeping, and having
babies) and psychological properties (thinking). The youngest children (4-year-
olds) seemed to take a "man is the measure of all things" approach. They deter-
mined the likelihood that an entity had a property based on its similarity to humans.
For example, they attributed biological properties to a mechanical monkey--but
not to other inanimate objects--over half the time. With increasing age, children
came to differentiate based on biological categories--e.g. between mammals and
nonmammals. By the age of 10, children correctly treated the mechanical monkey
as an inanimate object. Inagaki & Hatano (1987) suggested that the early "person-
ification" responses represent children's productive use of analogy to reason from
a well-understood base domain--human beings--to other less familiar species.
Carey's findings suggest that young children's biological theories may not dif-
ferentiate among animals, or even between animals and plants. Carey further sug-
gested that children's naive biology develops from their naive psychology, which
is initially used to predict the actions of the people around them. Other researchers
maintain that a naive theory of biology appears early in development. For example,
Coley (1995) found that kindergarteners attributed biological properties (e.g. "has
blood") according to taxonomic category membership, but attributed psychologi-
cal properties (e.g. "thinks") according to whether the animal was domesticated.
Coley concluded that a distinction between biological and psychological properties
is present in young children and becomes more pronounced over development.

2
34 MARKMAN GENTNER
Young children reason about animate and inanimate objects in very different
ways. Rosengren et al (1991) showed children pictures of animals labeled as "baby
animals." When the children were shown new pictures and asked to choose which
picture showed the baby as an adult, even preschool children consistently selected
the larger animal. They would even accept a shape transformation (e.g. from a
caterpillar to a butterfly) rather than choose a smaller "adult." Thus, children expect
that animals will get larger, and would rather assume that an animal changed shape
than that it got smaller. In contrast, when children were presented with artifacts,
they expected them to stay the same size (though possibly to become chipped and
broken). Hatano & Inagaki (1999) extended this method to plants and found that
children expected both plants and animals to grow rather than to stay the same
size. These findings suggest that young children distinguish between living and
nonliving things.
However, children's initial understanding of the living-nonliving distinction
may not go very deep. Simons & Keil (1995) showed children pairs of pictures
of animals or machines with cutaways showing the "insides"--either organs or
machinery (e.g. gears and chains)--and asked them to select the right insides for
each whole. Only half of the 3- and 4-year-olds associated organs with the in-
sides of animals and machinery with the insides of artifacts. Other children could
choose correctly for artifacts but not for animals. Not until 8 years of age could
children all select correctly. Simons & Keil suggest a developmental shift from
abstract comprehension to specific causal knowledge: Preschool children distin-
guish between animate and artifact categories, but lack deep causal understanding
of these domains. Although this claim that children possess abstract domain theo-
ries but lack causal understanding may seem rather tenuous, Keil and his colleagues
suggest that adults' mental models of physical systems are similarly incomplete
(Wilson & Keil 1998). They find that college students often believe that they know
how complex artifacts like car starters and toilets work, but are unable to explain
when probed for details. They argue that people are generally unaware of their
lack of specific causal information about how common objects work, and call this
effect the shallowness of explanation.
Finally, there has been research on how biological information is organized
in the brain. Some neuropsychological patients appear to have category-specific
impairments for animals; they can name pictures of artifacts but have trouble
naming animals (e.g. Caramazza & Shelton 1998, Warrington & Shallice 1984).
On the basis of these findings, Caramazza & Shelton (1998) suggest that evolution
might have selected for specific mechanisms that facilitate the representation of
animate and inanimate objects, and that the observed deficits reflect the breakdown
of these systems. Regardless of the resolution of this issue, these findings add
weight to the animate/inanimate distinction as an organizing principle for category
information.
To summarize, studies of biological knowledge suggest that preschool chil-
dren can make core distinctions such as the animate/inanimate distinction, but
that their understanding deepens with development. Both children and adults have

T
HINKING 235
significant gaps in their models and may be unaware of the extent of the missing
information. Much remains to be understood. For example, do the various facets of
theanimate/inanimatedistinction--living/nonliving, self-moving/nonself-moving,
and so on--reflect unified categories from the start, or are they acquired piece-
meal? Is naive biology acquired through general learning processes or through
domain-specific processes?
CATEGORIZATION
Categorization research includes both how people classify things into categories
and how people draw inferences from known categories. Most research on cat-
egorization has focused on domain-general issues such as the role of similarity
or the perception of feature centrality. We begin with these approaches and then
describe some domain-specific approaches.
Similarity and Categorization
Intuitively, similarity and categorization seem tightly linked. It is not surprising,
then, that many studies of categorization have focused on the kinds of similarity
relationships between a new exemplar and a stored category representation that
allow the new item to be classified as a member of a known category. Prototype
models assume that the stored category representation is a summary of the most
typical feature values for members of a category, and that new exemplars are
classified on the basis of their similarity to these prototypes (e.g. Hampton 1995,
Reed 1972, Rosch & Mervis 1975). Exemplar models assume that people store
away specific exemplars of the category and classify new items on the basis of
their similarity to the stored exemplars (e.g. Kruschke 1992, Medin & Schaffer
1978, Nosofsky 1986).
Most models of similarity are domain general: They assume that some mea-
sure of proximity can be calculated for any pair of representations. Two classic
influential approaches are the spatial view and the featural view. Spatial models of
similarity (e.g. Nosofsky 1986; Shepard 1962, 1964) model individual concepts
as points in a multidimensional space and assume that the similarity of a pair of
objects is inversely related to the distance between them. Feature-based models
of similarity (e.g. Tversky 1977) assume that objects are represented by sets of
features, and that the commonalities of a pair are the features in the intersection
of the sets representing each object, whereas the differences are just the features
not in the intersection. These models have been criticized because they typically
assume that a fixed set of dimensions or properties applies to a category (Schank
et al 1986), ignoring the role of feature learning during categorization (Schyns
et al 1998). Reflecting this perspective, classification experiments typically utilize
a fixed set of dimensions, each with a small number of values. Similarity-based
models of categorization have also been faulted for ignoring the role of theory in

2
36 MARKMAN GENTNER
determining the relevance of features to categories (Goldstone 1994, Murphy &
Medin 1985).
Current approaches to similarity like the structure-mapping approach described
above may better meet some of these requirements. As Goldstone (1994) points
out, structural alignment provides constraints on which commonalities among
items are relevant for categorizing them. In particular, systematic relational struc-
tures are likely to be important for categorization. This point helps to bridge the gap
between theory-based and similarity-based categorization. If similarity is com-
puted by a structural alignment process rather than a feature-matching process,
then common relational systems will be included in the similarity computation
(and even favored, by the systematicity principle); thus, common causal systems
will naturally be part of the category representation (Gentner & Medina 1998).
Structural alignment may also influence the way category representations are
learned (Lassaline & Murphy 1998). Markman & Wisniewski (1997) asked people
to list commonalities and differences of pairs of object categories at different
levels of abstraction. People could more easily list commonalities and alignable
differences for pairs of categories within a superordinate (e.g. pairs of vehicles
or pairs of weapons) than for pairs from different superordinates (e.g. a vehicle
and a weapon). This result suggests that object categories are organized around
taxonomic clusters whose numbers are readily comparable with other members of
the cluster, but not with members of different clusters.
These clusters may develop through reminding-based category learning (Ross
et al 1990, Spalding & Ross 1994). In this view, when a new item is presented, it
is given an initial representation that serves as a retrieval cue for items in memory.
The concepts retrieved serve as a template for constructing the representation of the
new item. This process helps create clusters of categories whose members are all
comparable. There is some evidence for this reminding-based process. Zhang &
Markman (1998) had people learn about a sequence of new brands of microwave
popcorn. People first saw a description of the first brand; in a second session, they
saw a description of the first brand again followed by descriptions of two more
brands whose properties overlapped with those of the first brand. People learned
properties of the later brands better if they were alignable differences with the
first brand than if they were nonalignable differences, suggesting that they learned
the later brands by aligning them to the initial one. These findings suggest that
structural alignment influences the representations of new categories.
The studies just described are part of a larger movement that explores the
influence of reasoning with categories on what is learned about categories--an
instance of the "learning as a byproduct of reasoning" theme mentioned at the
beginning of this paper (see Brooks 1978). Ross (1997, 1999) has demonstrated
that when features are relevant to a secondary task (either problem solving or a
secondary classification) their salience is increased in a primary classification task.
Yamauchi & Markman (1998) found that when categories are learned in the process
of making feature predictions, their representations focus more on relationships
among features than do those of categories learned by classifying new instances.

T
HINKING 237
Finally, the issue of whether language influences category structures
has returned to the research foreground. Recent work is exploring whether and
how possessing specific semantic categories affects conceptual processing
(e.g.Cabrera&Billman1996, Gentner&Loewenstein2000, Gumperz&Levinson
1996, Malt et al 1999), as well as how the act of communication affects people's
learning and reasoning (Clark 1996, Garrod & Anderson 1987, Markman & Makin
1998).
Inference from Categories
Once categories are established, people can use them to infer features of a new
situation. This kind of reasoning is studied using the category-based induction
task. In these tasks, people are told that one or more categories has some property
and asked how likely it is that some other category has the same property. For
example, people might be told that robins have property X and that bluebirds have
property X, and asked how likely it is that all birds have property X. The strength
of this argument can be compared to that of another argument, such as: "Robins
have X, Ostriches have X; how likely are all Birds to have X?"
Many of the central phenomena in category-based induction were mapped out
by Osherson et al (1990). According to their similarity-coverage model, the per-
ceived strength of a conclusion increases with the similarity between the premises
and the conclusion category. Thus, "robins have X, therefore bluebirds have X" is
judged to be a stronger argument than "ostriches have X, therefore bluebirds have
X." The similarity-coverage model also predicts a diversity effect in reasoning.
Premises from diverse categories lead to stronger arguments than premises that are
highly similar to each other. For example, college undergraduates find the argument
"robins have X, ostriches have X, therefore all birds have X" to be stronger than the
argument "robins have X, bluebirds have X, therefore all birds have X." A related
model of induction based on feature overlap was developed by Sloman (1993).
These views of category-based induction rely on overall similarity or feature
overlap as a predictor of argument strength. However, recent research suggests
that the comparisons involved here may be sensitive to causal alignments. For
example, Heit & Rubinstein (1994) showed people arguments in which the prop-
erties to be inferred were either behaviors or anatomical properties. People based
their judgments of argument strength not on overall similarity, but on similarity
with respect to the particular dimension: Thus, whales were considered likely to
share a new behavioral property of fish, but a new anatomical property of cattle.
Furthermore, it appears that the subjective strength of the induction depends on the
causal alignment between premises and conclusion categories (Lassaline 1996).
Category-based induction is also influenced by expertise in a domain. Medin,
Atran, and their colleagues (Coley et al 1999, Medin et al 1997) have undertaken
a cross-cultural study of biological reasoning. They contrasted category-based
induction performance of college undergraduates with that of Itzaj Mayans from
Guatemala (who have extensive experience with their native flora), with scientific

2
38 MARKMAN GENTNER
taxonomists (who study classification of trees), landscape workers (who decide
where to plant trees in residential settings), and park maintenance workers (who
take care of trees in local parks). One striking difference among these groups arises
with diversity-based reasoning. As discussed above, college students consider that
a diverse set of premise categories increases the strength of an argument. The
taxonomists and landscape workers also preferred diverse arguments. However,
both Itzaj Mayans and park maintenance workers showed the opposite effect,
preferring arguments with similar categories in the premises to arguments with
dissimilar categories in the premises. An analysis of the justifications suggests that
the Mayans and landscape workers based their judgments on ecological factors.
For example, landscape workers thought that a disease affecting white birches
and river birches would affect more trees than a disease affecting white pines and
weeping willows (Coley et al 1999). Although the second pair is more diverse
than the first, maintenance workers argued that the first pair (the birches) were
more likely to cause diseases of all trees, because this pair of trees is more widely
planted and more susceptible to disease than pine and willow. Thus, differences
in expertise influenced how people evaluated inductive arguments.
To summarize, recent research in category-based induction explores both gen-
eral processes and specific knowledge structures. People align specific causal
systems between the premises and conclusions to guide their inductions. With
increasing expertise, people are able to form more elaborate causal explanations.
Feature Centrality
A related area in which domain-general approaches have been applied to catego-
rization is in the determination of the centrality of a feature in a category. The
notion of centrality is that some properties of a category are more important than
others. For example, for a robin, the property of having wings seems to be more
central than the property of having a red breast. It is easier to imagine a robin with-
out a red breast than one without wings. A related point is that some categories
seem more cohesive than others. Research has focused on structural properties that
may determine cohesiveness and feature centrality.
Gentner(1981)exploredthedifferencesbetweennounsandverbsandsuggested
that nouns differ from verbs in the relational density of their representations. That
is, the semantic components of noun meanings are more strongly interconnected
than those of verbs. One consequence of this difference is that the meanings of
nouns seem less mutable than the meanings of verbs. For example, people asked
to paraphrase a sentence like "the lizard worshipped" are more likely to change
the verb's meaning (e.g. "the small gray reptile lay on a hot rock and stared at the
sun") than the noun's meaning (e.g. "a man with scaly skin prayed"). In this view,
the ease of altering a property of a concept is influenced by the degree to which
that property is interconnected with others.
This interconnectivity idea was explored at the level of individual properties by
Sloman, Love, and Ahn (Love & Sloman 1995, Sloman et al 1998). They found

T
HINKING 239
that the perceived mutability of a feature--that is, how willing people were to
accept changes in those features for members of the category--was influenced
by its interconnectivity with other features of the category. Features with few
interconnections were more mutable than features with many interconnections.
More specifically, there is evidence that an important determinant of feature
centrality is whether a property participates in a higher-order constraining structure
such as causality. This is consistent with the systematicity principle discussed
above. For example, Ahn and colleagues have demonstrated the importance of
causally relevant properties (Ahn 1999, Ahn et al 2000). Using sorting tasks, Ahn
(1999) demonstrated that people prefer to create categories based on items with
a common cause or a common effect, suggesting that they align causal structures
across items, and focus on the alignable information. Likewise, Rehder & Hastie
(1998) found that common causal relations led items to be placed in the same
category. Ramscar & Pain (1996) asked subjects to categorize a set of stories that
varied systematically in their similarity relations. Subjects classified the stories by
common causal structure rather than by common object features.
Ahn et al (2000) further suggested that causes may be more important than
effects. They found that when people categorize new items with missing features,
they are more likely to classify an item into a given category when the cause
is present and the effect is missing than when the reverse is true (provided the
causal relation is plausible). This suggests that the position of a property within a
relational structure (as the cause rather than as the effect) influences the importance
of that feature in a category.
Similar findings have been shown with temporal relations (Sloman et al 1998).
However, we would not expect that just any relation (e.g. taller or darker) or even
any relation between relations (e.g. conjunctions) would contribute to centrality.
The systematicity effect depends on higher-order constraining relations: relations
that bind their relational arguments in informative ways. Future work will have to
investigate the effects of content and structure further.
Domain-Specific Approaches to Categorization
Although categorization has traditionally been modeled using domain-general pro-
cesses, recent research has focused on domain-specific aspects of categorization.
First, research in the interface between psychology and anthropology has used
cross-cultural similarities and differences in categorization to study the relative
contributions of the objective structure of the world and the constraints of the
cognitive system to human conceptual structure. Second, as discussed above, re-
search in cognitive development has traced the development knowledge structures
that characterize specific domains. One major focus in this work is naive biology:
e.g. the distinction between animate and inanimate entities in category acquisition.
Malt (1995) reviewed studies exploring folk biological categories across cul-
tures to examine whether category structure is more strongly influenced by the
objective information available in the world or by the structure of the cognitive

2
40 MARKMAN GENTNER
system. One way to approach this is to compare folk systems with scientific
taxonomies, which presumably are designed to capture objective facts about the
world. Perhaps unsurprisingly, the evidence that bears on these questions shows a
mixture of influences. In general, folk biological (and folk botanical) categories
often do bear a strong resemblance to scientific categories, particularly at the genus
level (roughly similar to the basic level). This suggests that there are clusters of
features in the world, and that people are sensitive to these clusters (Atran 1990,
Berlin 1972). However, there are also clear influences of cultural factors on these
categories. For example, the folk botanical category "tree" does not pick out a
distinct class of plants, and many objects labeled as trees are more closely related
in scientific taxonomies to plants we would call bushes than they are to other trees.
Instead, categories like tree carry with them certain functional properties such as
providing shade.
A second way in which domain-specific information influences categorization
is that, as noted above, domain theories can affect the set of features people use to
classify an item. This phenomenon can be seen in children as their domain theories
develop. Many researchers have found that when children hear a new noun applied
to an object, they tend to generalize the word to other objects with the same shape
(e.g. Baldwin 1989, Landau et al 1988), but that with increasing knowledge there
is a shape-to-taxonomic shift (Imai et al 1994). Children come to use taxonomic
information to generalize word meanings (Markman 1989, Waxman 1990).
Consistent with our earlier discussion of the development of domain theories,
children appear to consider animacy important in categorization. There is evidence
that children are less likely to generalize a new label according to shape when they
believe that the object is animate. Jones et al (1991) found that preschool children
given labels for a set of novel objects generalized according to shape unless the
objects had eyes, in which case they also took into account the texture of the objects.
Another early dimension may be whether something is an artifact. Bloom (1998)
found that young children pay more attention to the intended function of an item
if they are told it is an artifact.
People's knowledge of the functions of categories can also constrain the features
they focus on (Heit 1997, Murphy & Allopenna 1994, Spalding & Murphy 1996,
Wisniewski & Medin 1994). Wisniewski (1995) had people learn about novel
functional categories whose features varied in frequency of association and in
functional relevance. For example, if the function of a particular category was
to kill bugs, then a functionally relevant feature might be "contains poison," and
an irrelevant feature might be "manufactured in Florida." People who were told
the functions of the categories prior to learning focused strongly on functionally
relevant features, even when these features occurred more rarely than functionally
irrelevant features.
These examples illustrate both the importance of specific content and the way
in which content interacts with higher-level beliefs. For example, in the devel-
opmental studies described here, the importance of a given feature is influenced
by children's domain understanding. In cross-cultural research, the content of

T
HINKING 241
categories is explored in order to ask fundamental questions about how perceptual
experience and cultural conventions interact in cognitive structure.
SUMMARY AND CONCLUSIONS
Human reasoning is characterized by a mix of domain-general and domain-specific
aspects. Domain-general processes such as logical deduction operate on the struc-
ture of representations, independent of content. Their value is in their wide ap-
plicability across domains. At the other extreme are reasoning processes that are
specific to particular modalities or content areas. Analogy--structural alignment
and mapping--operates at an intermediate level. It has a domain-specific aspect in
that it is sensitive to domain content. However, the constraints on these processes
(e.g. structural consistency and systematicity) are domain general. Our review
suggests that many domain-specific effects can be captured using alignment and
mapping processes.
The research reviewed here opens up a number of important avenues for future
research. One particularly important area is cross-cultural work. In the section on
categorization, we briefly touched on some research with specialized populations
including Itzaj Mayans, and groups of American tree experts. This work is part of
a growing trend in psychology to move beyond the college undergraduates who
typically participate in studies to examine how cultural difference and knowledge
difference influence behavior. For example, members of collectivist cultures (like
many Asian cultures) may be more willing to take risks than members of individ-
ualistic cultures (like the one in the United States), because of a belief that the
cultural influences are likely to become an important part of research on reasoning
processes.
The role of domain-specific knowledge in reasoning is also central to
evolutionary psychology. Cosmides and Tooby (1994) point out that the interests
in domain-specific modules in development that we discussed above intersects
with work by evolutionary psychologists on mechanisms for solving domain-
specific problems. Evolutionary psychology argues that these domain-specific
modules evolved to solve problems that were prominent in the environment of
early hominids. While it is too early to evaluate the evolutionary perspective, the
parallels between this work and research in cognitive development bear further
scrutiny.
Domain-specificity has also been explored in cognitive neuroscience. Earlier,
we mentioned the proposal by Caramazza and Shelton that there may be specific
brain mechanisms to process animate and inanimate objects. There is continued
exploration of the way the brain processes information about animate and inani-
mate objects (as well as faces) in both lesion studies (e.g. Tranel et al 1997) and
imaging studies (e.g. Perani et al 1999). At present it is difficult to draw firm
conclusions from this work, but significant new research is likely to be done in
this area as techniques for brain imaging improve.

2
42 MARKMAN GENTNER
The rise of research on imaging heightens the importance of behavioral research
on thinking that was the focus of this paper. Current brain imaging techniques are
only able to explore tasks that take place in a few seconds. Long complex tasks are
difficult to give to patients who may have significant cognitive impairments. Fur-
ther, complex tasks that take place over a period of minutes or even hours are
inappropriate for imaging techniques. However, a better understanding of rea-
soning behavior will lead to task decompositions that may be more amenable to
techniques from cognitive neuroscience.
We are entering an exciting period of research on reasoning. Significant progress
have been made on both domain-general and domain-specific processes. This work
has opened up new avenues for exploration. The next step in this progression of
research will require the development of models that combine domain-general and
domain-specific approaches into unified models of reasoning in cognition.
ACKNOWLEDGMENTS
This research was supported by ONR grant N00014-92-J-1098 and NSF-LIS grant
SBR-9511757 to DG and by NSF grant SBR-9905013 given to ABM. This chapter
was partially prepared while DG was a Fellow at the Center for Advanced Study
in the Behavioral Sciences. We are grateful for the financial support provided by
the William T. Grant Foundation, award #95167795
Visit the Annual Reviews home page at www.AnnualReviews.org
LITERATURE CITED
Ahn WK. 1999. Effect of causal structure on
category construction. Mem. Cogn. 27:1008­
23
Ahn WK, Kim NS, Lassaline ME, Dennis MJ.
2000. Causal status as a determinant of fea-
ture centrality. Cogn. Psychol. In press
Anderson JR. 1993. Rules of the Mind. Hills-
dale, NJ: Erlbaum
Anderson JR, Lebiere C, eds. 1998. The Atomic
Components of Thought. Mahwah, NJ: Erl-
baum
Atran S. 1990. Cognitive Foundations of Nat-
ural History: Towards an Anthropology of
Science. New York: Cambridge Univ. Press
BaillargeonR.1994.Howdoinfantslearnabout
the physical world? Curr. Dir. Psychol. Sci.
3(5):133­40
Baldwin DA. 1989. Priorities in children's ex-
pectations about object label reference: form
over color. Child Dev. 60:1289­306
Bassok M, Chase VM, Martin SA. 1998.
Adding apples and oranges: semantic con-
straints on application of formal rules. Cogn.
Psychol. 35:99­134
Berlin B. 1972. Speculations on the growth
of ethnobotanical nomenclature. Lang. Soc.
1:51­86
Bloom P. 1998. Theories of artifact categoriza-
tion. Cognition 66:87­93
Bowdle B, Gentner D. 1997. Informativity and
asymmetry in comparisons. Cogn. Psychol.
34(3):244­86
Bowdle B, Gentner D. 1999. Metaphor compre-
hension: from comparison to categorization.
In Proc. 21st Annu. Conf. Cogn. Sci. Soc., pp.
90­95. Vancouver, BC: Erlbaum
Brooks LR. 1978. Non analytic concept forma-
tion and memory for instances. In Cognition
and Categorization, ed. E Rosch, BB Lloyd,
pp. 169­211. Hillsdale, NJ: Erlbaum

T
HINKING 243
Cabrera A, Billman D. 1996. Language-driven
concept learning: deciphering Jabberwocky.
J. Exp. Psychol.: Learn. Mem. Cogn. 22(2):
539­55
Caramazza A, Shelton JR. 1998. Domain-
specific knowledge systems in the brain: the
animate-inanimate distinction. J. Cogn. Neu-
rosci. 10:1­34
Carey S. 1985. Conceptual Change in Child-
hood. Cambridge, MA: MIT Press
Cheng PW, Holyoak KJ. 1985. Pragmatic rea-
soning schemas. Cogn. Psychol. 17:391­416
Chi MTH, Feltovich PJ, Glaser R. 1981. Cat-
egorization and representation of physics
problems by experts and novices. Cogn. Sci.
5:121­52
Chi MTH, VanLehn KA. 1991. The content
of physics self-explanations. J. Learn. Sci.
1:69­105
Clark HH. 1996. Using Language. London:
Cambridge Univ. Press
Clement CA, Gentner D. 1991. Systematicity as
a selection constraint in analogical mapping.
Cogn. Sci. 15:89­132
Coley JD. 1995. Emerging differentiation of
folkbiology and folkpsychology: attribu-
tions of biological and psychological proper-
ties to living things. Child Dev. 66:1856­74
Coley JD, Medin DL, Proffitt JB, Lynch E,
Atran S. 1999. Inductive reasoning in folkbi-
ological thought. See Medin & Atran 1999,
pp. 205­32
Cosmides L. 1989. The logic of social ex-
change: Has natural selection shaped how
humans reason? Studies with the Wason se-
lection task. Cognition 31:187­276
Cosmides L, Tooby J. 1994. Origins of domain
specificity: the evolution of functional or-
ganization. In Mapping the Mind, ed. LA
Hirschfeld, SA Gelman, pp. 85­116. New
York: Cambridge Univ. Press
Dunbar K. 1995. How scientists really reason:
scientific reasoning in real-world laborato-
ries. In The Nature of Insight, ed. RJ Stern-
berg, JE Davidson, pp. 365­95. Cambridge,
MA: MIT Press
Elman JL, Bates EA, Johnson MH, Karmiloff-
Smith A, Parisi D, Plunkett K. 1996. Re-
thinking Innateness: A Connectionist Per-
spective on Development. Cambridge, MA:
MIT Press
Evans JS, Over DE. 1996. Rationality and Rea-
soning. East Sussex, UK: Psychology Press
Falkenhainer B, Forbus KD, Gentner D. 1989.
The structure-mapping engine: algorithm
and examples. Artif. Intell. 41:1­63
Forbus KD. 1984. Qualitative process theory.
Artif. Intell. 24:85­168
Forbus KD, Nielsen P, Faltings B. 1991.
Qualitative spatial reasoning: The CLOCK
project. Artif. Intell. 52:417­72
Garrod S, Anderson A. 1987. Saying what you
mean in dialogue: a study in conceptual and
semantic co-ordination. Cognition 27:181­
218
Gentner D. 1981. Some interesting differences
between nouns and verbs. Cogn. Brain The-
ory 4:161­78
Gentner D. 1983. Structure-mapping: a the-
oretical framework for analogy. Cogn. Sci.
7:155­70
Gentner D. 1989. The mechanism of analog-
ical learning. In Similarity and Analogical
Reasoning, ed. S Vosnindon, A Ortony, pp.
199­241. New York: Cambridge Univ. Press
GentnerD,LoewensteinJ.2000.Relationallan-
guage and relational thought. In Language,
Literacy, and Cognitive Development, ed. J
Byrnes, E Amsel. Hillsdale, NJ: Erlbaum. In
press
Gentner D, Markman AB. 1994. Structural
alignment in comparison: no difference
without similarity. Psychol. Sci. 5:152­
58
Gentner D, Markman AB. 1997. Structural
alignment in analogy and similarity. Am. Psy-
chol. 52:45­56
Gentner D, Medina J. 1998. Similarity and
the development of rules. Cognition 65:263­
97
Gentner D, Rattermann MJ, Forbus KD. 1993.
The roles of similarity in transfer: separat-
ing retrievability from inferential soundness.
Cogn. Psychol. 25:524­75

2
44 MARKMAN GENTNER
GentnerD,StevensAL,eds.1983.MentalMod-
els. Hillsdale, NJ: Erlbaum
Gentner D, Wolff P. 1997. Alignment in the pro-
cessing of metaphor. J. Mem. Lang. 37:331­
55
Gentner D, Wolff P. 2000. Metaphor and knowl-
edge change. In Cognitive Dynamics, ed. E
Dietrich, AB Markman, pp. 295­342. Mah-
wah, NJ: Erlbaum
Gilden DL, Proffitt DR. 1989. Understanding
collision dynamics. J. Exp. Psychol.: Hum.
Percep. Perform. 15:372­83
Glenberg AM. 1997. What memory is for. Be-
hav. Brain Sci. 20(1):1­55
GlucksbergS,McGloneMS,ManfrediD.1997.
Property attribution in metaphor comprehen-
sion. J. Mem. Lang. 36:50­67
Goldstone RL. 1994. The role of similarity
in categorization: providing a groundwork.
Cognition 52:125­57
Goldstone RL, Medin DL. 1994. The time
course of comparison. J. Exp. Psychol.:
Learn. Mem. Cogn. 20:29­50
GopnikA,MeltzoffAN.1997.Words,Thoughts
and Theories. Cambridge, MA: MIT Press
Guidagni PM, Little JDC. 1983. A logit model
of brand choice calibrated on scanner data.
Mark. Sci. 2:203­38
Gumperz JJ, Levinson SC. 1996. Introduction:
linguistic relativity re-examined. In Rethink-
ing Linguistic Relativity, ed. JJ Gumperz, SC
Levinson, pp. 1­18. Cambridge, MA: Cam-
bridge Univ. Press
Halford GS. 1992. Analogical reasoning and
conceptual complexity in cognitive develop-
ment. Hum. Dev. 35(4):193­218
Hampton JA. 1995. Testing the prototype the-
ory of concepts. J. Mem. Lang. 34:686­
708
Hatano G, Inagaki K. 1999. A developmental
perspective on informal biology. See Medin
& Atran 1999, pp. 321­54
Hegarty M, Just MA. 1993. Constructing men-
tal models of machines from text and dia-
grams. J. Mem. Lang. 32: 717­42
Heit E. 1997. Knowledge and concept learning.
In Knowledge, Concepts, and Categories, ed.
KLamberts, DShanks, pp.7­41.Cambridge,
MA: MIT Press
Heit E, Rubinstein J. 1994. Similarity and prop-
erty effects in inductive reasoning. J. Exp.
Psychol.: Learn. Mem. Cogn. 20:411­22
Hirschfeld LA, Gelman SA, eds. 1994. Map-
ping the Mind: Domain Specificity in Cog-
nition and Culture. New York: Cambridge
Univ. Press
Holyoak KJ, Thagard P. 1989. Analogical map-
ping by constraint satisfaction. Cogn. Sci.
13:295­355
Howard IP. 1978. Recognition and knowl-
edge of the water-level principle. Perception
7:151­60
Hummel JE, Holyoak KJ. 1997. Distributed
representations of structure: a theory of ana-
logical access and mapping. Psychol. Rev.
104:427­66
Imai M, Gentner D, Uchida N. 1994. Children's
theories of word meaning: the role of shape
similarity in early acquisition. Cogn. Dev.
9:45­75
Inagaki K, Hatano G. 1987. Young chil-
dren's spontaneous personification as anal-
ogy. Child Dev. 58:1013­20
Johnson-Laird PN. 1983. Mental Models.
Cambridge, MA: Harvard Univ. Press
Johnson-Laird PN. 1999. Deductive reasoning.
Annu. Rev. Psychol. 50:109­35
Johnson-Laird PN, Byrne RMJ. 1991. Deduc-
tion. Hillsdale, NJ: Erlbaum
Jones SS, Smith LB, Landau B. 1991. Object
properties and knowledge in early lexical
learning. Child Dev. 62:499­516
Keil FC. 1989. Concepts, Kinds and Cognitive
Development. Cambridge, MA: MIT Press
Kempton W, Boster JS, Hartley J. 1994. En-
vironmental Values in American Culture.
Cambridge, MA: MIT Press
Klayman J, Ha Y. 1987. Confirmation, discon-
firmation and information in hypothesis test-
ing. Psychol. Rev. 94:211­28
Koehler JJ. 1993. The influence of prior beliefs
on scientific judgments of evidence quality.
Organ. Behav. Hum. Decis. Process. 56:28­
55

T
HINKING 245
Kruschke JK. 1992. ALCOVE: an exemplar-
based connectionist model of category learn-
ing. Psychol. Rev. 99:22­44
Kuhn D. 1997. Is good thinking scientific think-
ing? In Modes of Thought: Explorations in
Culture and Cognition, ed. DR Olson, N Tor-
rance, pp. 261­81. New York: Cambridge
Univ. Press
Kuipers B. 1994. Qualitative Reasoning: Mod-
elingandSimulationwithIncompleteKnowl-
edge. Cambridge, MA: MIT Press
Landau B, Smith LB, Jones SS. 1988. The im-
portance of shape in early lexical learning.
Cogn. Dev. 3:299­321
Lassaline ME. 1996. Structural alignment in
induction and similarity. J. Exp. Psychol.:
Learn. Mem. Cogn. 22:754­70
Lassaline ME, Murphy GL. 1998. Alignment
and category learning. J. Exp. Psychol.:
Learn. Mem. Cogn. 24:144­60
Leslie AM. 1991. The theory of mind impair-
ment in autism: evidence for a modular
mechanism of development? In Natural The-
ories of Mind: Evolution, Development and
Simulation of Everyday Mindreading, ed. A
Whiten. Oxford/Cambridge, MA: Blackwell
Logan GD. 1988. Toward an instance theory of
automaticity. Psychol. Rev. 95:492­527
Logan GD. 1996. The CODE theory of vi-
sual attention: an integration of space-based
and object-based attention. Psychol. Rev.
103:603­49
Love BC, Sloman SA. 1995. Mutability and the
determinants of conceptual transformability.
In Proc. 17th Annu. Conf. Cogn. Sci. Soc.,
pp. 65­59. Pittsburgh, PA: Erlbaum
Malt BC. 1995. Category coherence in cross-
cultural perspective. Cogn. Psychol. 29:85­
148
Malt BC, Sloman SA, Gennari S, Shi M, Wang
Y. 1999. Knowing versus naming: similarity
and the linguistic categorization of artifacts.
J. Mem. Lang. 40:230­62
Markman AB. 1997. Constraints on analogical
inference. Cogn. Sci. 21:373­418
Markman AB. 1999. Knowledge Representa-
tion. Mahwah, NJ: Erlbaum
Markman AB, Gentner D. 1993a. Splitting the
differences: a structural alignment view of
similarity. J. Mem. Lang. 32:517­35
Markman AB, Gentner D. 1993b. Structural
alignment during similarity comparisons.
Cogn. Psychol. 25:431­67
Markman AB, Gentner D. 1996. Commonal-
ities and differences in similarity compar-
isons. Mem. Cogn. 24:235­49
Markman AB, Gentner D. 1997. The effects of
alignability on memory. Psychol. Sci. 8:363­
67
Markman AB, Makin VS. 1998. Referential
communication and category acquisition. J.
Exp. Psychol.: Gen. 127(4):331­54
Markman AB, Wisniewski EJ. 1997. Similar
and different: the differentiation of basic
level categories. J. Exp. Psychol.: Learn.
Mem. Cogn. 23:54­70
Markman EM. 1989. Categorization and Nam-
ing in Children. Cambridge, MA: MIT Press
McCloskey M. 1983. Naive theories of motion.
See Gentner & Stevens 1983, pp. 299­324
Medin DL, Atran S, eds. 1999. Folkbiology.
Cambridge, MA: MIT Press
Medin DL, Goldstone RL, Gentner D.
1993. Respects for similarity. Psychol. Rev.
100(2):254­78
Medin DL, Lynch EB, Coley JD, Atran S. 1997.
Categorization and reasoning among tree ex-
perts: Do all roads lead to Rome? Cogn.
Psychol. 32:49­96
Medin DL, Ross BH. 1989. The specific charac-
ter of abstract thought: categorization, prob-
lem solving and induction. In Advances in
the Psychology of Human Intelligence, ed.
RS Sternberg, 5:189­223. Hillsdale, NJ: Erl-
baum
Medin DL, Schaffer MM. 1978. Context theory
of classification. Psychol. Rev. 85:207­38
Metz KE. 1985. The development of children's
problem solving in a gears task: a problem
space perspective. Cogn. Sci. 9:431­71
Murphy GL, Allopenna PD. 1994. The locus
of knowledge effects in concept learning. J.
Exp. Psychol.: Learn. Mem. Cogn. 20:904­
19

2
46 MARKMAN GENTNER
Murphy GL, Medin DL. 1985. The role of the-
ories in conceptual coherence. Psychol. Rev.
92:289­315
Nersessian N. 1999. Model-based reasoning in
conceptual change. In Model-Based Reason-
ing in Scientific Discovery, ed. L Magroni, N
Nersessian, P Thagard, pp. 5­22. Dordrecht,
The Netherlands: Kluwer
Newell A, Simon HA. 1972. Human Problem
Solving. Englewood Cliffs, NJ: Prentice-Hall
Norman DA. 1988. The Psychology of Every-
day Things. New York: Basic Books
Nosofsky RM. 1986. Attention, similarity
and the identification-categorization rela-
tionship. J. Exp. Psychol.: Gen. 115:39­57
Osherson DN, Smith EE, Wilkie O, Lopez A,
Shafir E. 1990. Category based induction.
Psychol. Rev. 97:185­200
Perani D, Schnur T, Tettamanti M, Gorno-
Tempini M, Cappa S, Fazio F. 1999. Word
and picture matching: a PET study of se-
mantic category effects. Neuropsychologia
37:293­306
Pfeifer R, Scheier C. 1999. Understanding In-
telligence. Cambridge, MA: MIT Press
Proffitt DR, Kaiser MK, Whelan SM. 1990. Un-
derstanding wheel dynamics. Cogn. Psychol.
22:342­73
Ramscar M, Pain H. 1996. Can a real distinction
be made between cognitive theories of anal-
ogy and categorisation? Proc. 18th Annu.
Conf. Cogn. Sci. Soc., pp. 346­51. Hillsdale,
NJ: Erlbaum
Reed SK. 1972. Pattern recognition and catego-
rization. Cogn. Psychol. 3:382­407
Rehder B, Hastie R. 1998. The differential ef-
fects of causes on categorization and similar-
ity. Proc. 20th Annu. Conf. Cogn. Sci. Soc.,
pp. 893­98. Hillsdale, NJ: Erlbaum
Rips LJ. 1986. Mental muddles. In The Rep-
resentation of Knowledge and Belief, ed. H
Brand, RM Harnish. Tucson: Univ. Ariz.
Press
Rips LJ. 1994. The Psychology of Proof:
Deductive Reasoning in Human Thinking.
Cambridge, MA: MIT Press
Rosch E, Mervis CB. 1975. Family resem-
blances: studies in the internal structure of
categories. Cogn. Psychol. 7:573­605
Rosengren KS, Gelman SA, Kalish CW, Mc-
Cormick M. 1991. As time goes by: chil-
dren's early understanding of growth in ani-
mals. Child Dev. 62:1302­20
Ross BH. 1997. The use of categories affects
classification. J. Mem. Lang. 37(2):240­67
Ross BH. 1999. Postclassification category use:
the effects of learning to use categories after
learning to classify. J. Exp. Psychol.: Learn.
Mem. Cogn. 25(3):743­57
Ross BH, Perkins SJ, Tenpenny PL. 1990.
Reminding-based category learning. Cogn.
Psychol. 22:460­92
Schank RC, Collins GC, Hunter LE. 1986.
Transcending inductive category formation
in learning. Behav. Brain Sci. 9:639­86
Schwartz DL, Black JB. 1996. Shuttling be-
tween depictive models and abstract rules:
induction and fallback. Cogn. Sci. 20:457­
98
Schwartz DL, Black T. 1999. Inferences
through imagined actions: knowing by sim-
ulated doing. J. Exp. Psychol.: Learn. Mem.
Cogn. 25:116­36
Schyns PG, Goldstone RL, Thibaut JP. 1998.
The development of features in object con-
cepts. Behav. Brain Sci. 21:1­54
Shepard RN. 1962. The analysis of proximities:
multidimensional scaling with an unknown
distance function. I. Psychometrika 27:125­
40
Shepard RN. 1964. Attention and the metric
structure of the stimulus space. J. Math. Psy-
chol. 1:54­87
Siegler RS, Shipley C. 1995. Variation, selec-
tion, and cognitive change. In Developing
Cognitive Competence: New Approaches to
Process Modeling, ed. TJ Simon, GS Hal-
ford, pp. 31­76. Hillsdale, NJ: Erlbaum
Simons DJ, Keil FC. 1995. An abstract to con-
crete shift in the development of biological
thought: theinsidesstory.Cognition56:129­
63
Sj¨oberg L. 1972. A cognitive theory of similar-
ity. G¨oteborg Psychol. Rep. 2 (monograph)

T
HINKING 247
Sloman SA. 1993. Feature-based induction.
Cogn. Psychol. 25:231­80
Sloman SA. 1996. The empirical case for two
systems of reasoning. Psychol. Bull. 119:3­
22
Sloman SA, Love BC, Ahn WK. 1998. Feature
centrality and conceptual coherence. Cogn.
Sci. 22:189­228
Spalding TL, Murphy GL. 1996. Effects of
background knowledge on category con-
struction. J. Exp. Psychol.: Learn. Mem.
Cogn. 22:525­38
Spalding TL, Ross BH. 1994. Comparison-
based learning: effects of comparing in-
stances during category learning. J. Exp. Psy-
chol.: Learn. Mem. Cogn. 20:1251­63
Spellman BA, Holyoak KJ. 1996. Pragmatics in
analogical mapping. Cogn. Psychol. 31:307­
46
Stenning K, Oberlander J. 1995. A cognitive
theory of graphical and linguistic reasoning:
logic and implementation. Cogn. Sci. 19:97­
140
Suchman LA. 1987. Plans and Situated Ac-
tions: The Problem of Human-Machine
Communication. Cambridge, UK: Cam-
bridge Univ. Press
Thelen E, Smith LB. 1993. A Dynamic Systems
Approach to the Development of Cognition
and Action. Cambridge, MA: MIT Press
Tooby J, Cosmides L. 1989. Evolutionary psy-
chology and the generation of culture: I.
Theoretical considerations. Ethol. Sociobiol.
10(1­3):29­49
Tranel D, Damasio H, Damasio AR. 1997. A
neural basis for the retrieval of conceptual
knowledge. Neuropsychologia 35:1319­27
Tversky A. 1977. Features of similarity. Psy-
chol. Rev. 84:327­52
Tversky B. 1991. Distortions in memory for
visual displays. In Spatial Instruments and
Spatial Displays, ed. SR Ellis, M Kaiser, A
Grunewald, pp. 61­75. Hillsdale, NJ: Erl-
baum
Vosniadou S, Brewer WF. 1992. Mental models
of the earth: a study of conceptual change in
childhood. Cogn. Psychol. 24:535­85
Warrington EK, Shallice T. 1984. Category spe-
cific semantic impairments. Brain 107:829­
54
Wason PC, Johnson-Laird PN. 1972. Psychol-
ogy of Reasoning Structure and Content.
London: Routledge
Waxman SR. 1990. Linguistic biases and the
establishment of conceptual hierarchies: ev-
idence from preschool children. Cogn. Dev.
5:123­50
WilsonRA,KeilF.1998.Theshadowsandshal-
lows of explanation. Minds Mach. 8:137­
59
Wisniewski EJ. 1995. Prior knowledge and
functionally relevant features in concept
learning. J. Exp. Psychol.: Learn. Mem.
Cogn. 21:449­68
Wisniewski EJ, Medin DL. 1994. On the inter-
action of theory and data in concept learning.
Cogn. Sci. 18:221­82
Wolff P, Gentner D. 2000. Evidence for role-
neutral initial processing of metaphors. J.
Exp. Psychol.: Learn. Mem. Cogn. 26:512­
28
Yamauchi T, Markman AB. 1998. Category
learning by inference and classification. J.
Mem. Lang. 39:124­48
Zhang S, Markman AB. 1998. Overcoming the
early entrant advantage: the role of alignable
and nonalignable differences. J. Mark. Res.
35:413­26