Held C.&Co (Eds) Mental Models and The Mind

ADVANCES
IN
PSYCHOLOGY
138
Mental Models and the Mind

Current C
omments in Cognitive P
ool
Carsten Held Markus Knauti Gottfried Vosgerau Editors

MENTAL MENTAL MODELS MODELS AND AND THE THE MIND MIND
Current Current Developments Developments in in Cognitive Cognitive Psychology,
Psychology, Neuroscience, Neuroscience, and and Philosophy Philosophy of of Mind
Mind
ADVANCES ADVANCES IN IN PSYCHOLOGY
PSYCHOLOGY
138 138
Editor: Editor:
G. G. E. E. STELMACH STELMACH
ELSEVIER
Amsterdam Amsterdam – - Boston Boston - – Heidelberg Heidelberg - – London London – - New New York
York - – Oxford Oxford Paris Paris - – San San Diego Diego - – San San Francisco Francisco - – Singapore
Singapore – - Sydney Sydney - – Tokyo Tokyo
MENTAL MENTAL MODELS MODELS AND AND

THE THE MIND MIND
Current Current Developments Developments in in
Cognitive Cognitive Psychology, Psychology, Neuroscience,
Neuroscience, and and Philosophy Philosophy of of Mind

Mind
Edited Edited by by
Carsten Carsten HELD HELD Universität Universität Erfurt, Erfurt, Germany Germany
Markus Markus KNAUFF KNAUFF Max-Planck-Institute, Max-Planck-Institute, Germany Germany
Gottfried Gottfried VOSGERAU VOSGERAU Eberhard-Karls- Eberhard-Karls-Universität,
Universität, Germany Germany
ELSEVIER Amsterdam Amsterdam - – Boston Boston - – Heidelberg Heidelberg - – London London - –
New New York York - – Oxford Oxford Paris Paris - – San San Diego Diego - – San San Francisco Francisco
- – Singapore Singapore – - Sydney Sydney - – Tokyo Tokyo
ELSEVIER Inc. 525 B Street, Suite 1900 San Diego CA 92101-4495 USA ELSEVIERB.V. ELSEVIER B.V. ELSEVIER Inc. ELSEVIER ELSEVIER
Ltd. Ltd.
ELSEVIER ELSEVIER Ltd. Ltd. Radarweg Radarweg 29
29 525 B Street, Suite 1900 The The Boulevard Boulevard
84 84 Theobalds Theobalds Road Road P.O. P.O. Box Box 211 211
San Diego Langford Langford Lane, Lane, Kidlington Kidlington
London London 1000 The The 1000 Netherlands Netherlands
AE AE Amsterdam Amsterdam
CA 92101-4495 USA Oxford UK Oxford UK
OX5 OX5 1GB 1GB
WC1X WC1X 8RR 8RR UK UK
© © 2006 2006 Elsevier ElsevierB.V. B.V. All All rights rights reserved. reserved.
This This work work is is protected protected under under copyright copyright by by Elsevier Elsevier B.V., B.V., and and the the following following
terms terms and and conditions conditions apply apply to to its its use: use:
Photocopying Photocopying Single Single photocopies photocopies of of single single chapters chapters may may be be made made for for personal
personal use use as as allowed allowed by by national national copyright copyright laws. laws. Permission Permission of of the the Publisher Publisher and
and payment payment of of a a fee fee is is required required for for all all other other photocopying, photocopying, including including multiple multiple
or or systematic systematic copying, copying, copying copying for for advertising advertising or or promotional promotional purposes, purposes, resale,
resale, and and all all forms forms of of document document delivery. delivery. Special Special rates rates are are available available for for educational
educational institutions institutions that that wish wish to to make make photocopies photocopies for for non-profit non-profit educational educational
classroom classroom use. use.
Permissions fax fax homepage homepage Permissions (+44) (+44) 1865 1865 (http://www.elsevier.com/locate/permissions).
(http://www.elsevier.com/locate/permissions).
may may 853333, 853333, be be sought sought e-mail: e-mail: directly directly [email protected]. [email protected]. from from
Elsevier’s Elsevier's Rights Rights Requests Requests Department Department may may in in also also Oxford, Oxford, be be completed completed UK:
UK: phone phone on-line on-line (+44) (+44) via via 1865 1865 the the Elsevier 843830, Elsevier 843830,
In In Rosewood through through W1P rights rights Rosewood W1P the the 0LP, 0LP, agency agency USA, USA, the the UK; UK; Drive, Drive, Copyright
Copyright users users for for phone: phone: payments. payments. Danvers, Danvers, may may Licensing Licensing (+44) (+44) clear clear MA MA 20 20
permissions permissions 7631 7631 01923, 01923, Agency Agency 5555; 5555; USA; USA; Rapid Rapid and and fax: fax: phone: phone: make make
Clearance (+44) Clearance (+44) payments payments (+1) (+1) 20 20 7631 7631 (978) (978) Service Service 5500. 5500. through through 7508400,
7508400, (CLARCS), (CLARCS), Other Other the the fax: fax: countries countries Copyright Copyright (+1) (+1) 90 90 Tottenham Tottenham (978)
(978) may may Clearance Clearance have have 7504744, 7504744, Court Court a a local local Center, Center, and Road, and Road, reprographic
reprographic in in Inc., Inc., London London the the 222 222 UK UK
Derivative Tables Tables resale resale compilations Derivative compilations or or of of distribution distribution contents contents Works
Works
and and translations. translations.
may may of of be be such such reproduced reproduced material. material. Permission Permission for for internal internal of of circulation, circulation, the
the Publisher Publisher but but permission permission is is required required for for of of the the all all Publisher Publisher other other derivative derivative
is is required required works, works, for for including including external external
Electronic Permission Electronic Permission Storage Storage of of the the Publisher Publisher or or Usage Usage
is is required required to to store store or or use use electronically electronically any any material material contained contained in in this this work, work,
including including any any chapter chapter or or part part of of a a chapter. chapter.
Except Except as as outlined outlined above, above, no no part part of of this this work work may may be be reproduced, reproduced, stored stored in in a a
retrieval retrieval system system or or transmitted transmitted in in any permission permission Address any Address form form permissions permissions or
or of of by by the the any any Publisher. Publisher.
requests requests means, means, electronic, electronic, to: to: Elsevier’s Elsevier's mechanical, mechanical, Rights Rights Department, Department,
photocopying, photocopying, at at the the recording recording fax fax and and e-mail e-mail or or otherwise, otherwise, addresses addresses without
without noted noted above. above.
prior prior written written
Notice Notice No No products products ideas verification verification ideas responsibility responsibility contained contained liability, liability, of of
diagnoses diagnoses in in is is negligence negligence the the assumed assumed material material and and drug drug or by by or herein. herein. otherwise,
otherwise, the the dosages dosages Publisher Publisher Because Because should should or or from from for for of of be be rapid rapid any any any any
made. made.
injury injury advances advances use use or or and/or and/or operation operation in in the the damage damage medical medical of of any any to to sciences,
sciences, persons persons methods, methods, or or in in products, products, property property particular, particular, instructions instructions as as
independent independent a a matter matter of of or or
First First edition edition 2006 2006
Library A Library A catalog catalog of of record record Congress Congress is is available available Cataloging Cataloging from from in in the the
Publication Publication Library Library of of Data Data
Congress. Congress.
British British A A catalogue catalogue Library Library record record Cataloguing Cataloguing is is available available in in Publication Publication from
from the the British British Data Data
Library. Library.
ISBN-13: ISBN-13: 978-0-444-52079-1 978-0-444-52079-1 ISBN-10: ISBN-10: ISSN: ISSN: 0-444-52079-1 0166-4115 0166-4115 0-444-52079-1
(Series) (Series)
@ Printed Printed The The paper in paper in The The used used Netherlands. Netherlands.
in in this this publication publication meets meets the the requirements requirements of of ANSI/NISO ANSI/NISO Z39.48-1992 Z39.48-1992
(Permanence (Permanence of of Paper). Paper).
Working together to grow libraries in developing countries

www.elsevier.com | www.bookaid.org | www.sabre.org
ELSEVIER BOOK AID
Sabre Foundation
International
Contents
Preface 1
Contributors 3
General Introduction 5
Part I: Cognitive Psychology
Introduction 25
Mental Models, Sentential Reasoning, and Illusory Inferences 27 P.N. Johnson-Laird
Interaction of Knowledge and Working Memory in Reasoning About Relations 53 A. Vandierendonck, V.
Dierckx & H. Van der Beken
Mental Models in Learning Situations 85 N.M. Seel
Part II: Cognitive Neuroscience
Introduction 111
Resolving Valid Multiple Model Inferences Activates a Left Hemisphere Network 113 R.L. Waechter & V.
Goel
A Neuro-Cognitive Theory of Relational Reasoning with Mental Models and Visual Images 127 M. Knauff
v
vi Contents
vi Contents
Part III: Perception, Emotion, and Language
Introduction 155
Pictures, Perception, and Mental Models 157 K. Rehkdmper
Emotion, Decision, and Mental Models 173 M. Pauen
Language Processing: Construction of Mental Models or More? 189 B. Hemforth & L. Konieczny
Part IV: Philosophy of Mind
Introduction 207
Visual Imagery, Mental Models, and Reasoning 211 V. Gottschling
Mental Models as Objectual Representations 237 C. Held
The Perceptual Nature of Mental Models 255 G. Vosgerau
Index 277
Preface
The present book is the result of a workshop on "Mental Models and the Mind" that has been held at
the University of Freiburg in the summer 2003. The workshop brought together researchers from a
variety of disciplines: Cognitive psychologists reported their research on the representation and
processing of mental models in human memory. Cognitive neuroscientists demonstrated how visual
and spatial mental models are processed in the brain and which neural processes underlie visual and
spatial thinking. Philosophers talked about the role of mental models in relation to perception,
emotion, representation, and intentionality. Computer and education scientists reflected on the
importance of mental models, both theoretically and application-driven. As it is often the case after a
stimulating workshop, the idea of a book publication based on the contributions quickly arose. We
have asked all workshop participants for extended versions of their papers and have invited other
colleagues to contribute to the book. We owe special thanks to Phil Johnson-Laird for his support and
for an original contribution.We gratefully acknowledge financial support by several organizations for
the workshop and for our own research, including the writing and editing of this book. The Freiburg
workshop was supported by the Fritz- Thyssen- Stiftung within the Cross Section Area: Image and
Imagery. Our work has also been supported by the Deutsche Forschungsgemeinschaft (DFG) through
the Transregional Collaborative Research Center SFB/TR 8 Spa- tial Cognition and by the
VolkswagenStiftung through the Research Project Self-consciousness and concept formation in
humans. Markus Knauff is supported by a Heisenberg Award from the DFG. Gottfried Vosgerau has
been supported by a DAAD exchange fellowship held at NYU and a dissertation fellowship of the
Studienstiftung des deutschen Volkes. Every article in this book has been reviewed by another
contributor and by one external colleague. These colleagues provided authors and editors with
helpful comments and suggestions, in particular with respect to the multi-disciplinary audience we
hope to reach. We thus owe thanks to all our authors and the following external reviewers: Wolfgang
Huemer, Georg Jahn, Christoph Klauer, Albert Newen, Klaus Oberauer, Wolfgang Schnotz, Walter
Schaeken, Bernhard Schroder, Ralph Schumacher, Ger- hard Strube, Kai Vogeley, Lara Webber,
Stefan Wolfl, and Hubert Zimmer. Thanks also to Nadine Becker, Steffi von dem Fange, and Doreen
Schmidt for technical support, and, on the publisher's side, to Fiona Barron, Joyce Happee, and Simon
Pepping for a smooth and effective collaboration. Carsten Held, Markus Knauff, Gottfried Vosgerau
September 2005
1
This Page is Intentionally Left Blank
Contributors
Vicky Dierckx: Department of Experimental Psychology, Ghent University
Henri Dunantlaan 2, 9000 Ghent, Belgium
Vinod Goel: Department of Psychology, York University
Toronto, Ontario, Canada M3J 1P3 [email protected]
Verena Gottschling: Philosophisches Seminar, Johannes-Gutenberg-Universitat Jakob-Welder-Weg
18, 55099 Mainz, Germany [email protected]
Cars ten Held: Wissenschaftsphilosophie, Universitat Erfurt
99105 Erfurt, Germany [email protected]

Barbara Hemforth:
Laboratoire Parole et Langage, Universite de Provence 29, avenue Robert Schuman, 13621 Aix-en-
Provence Cedex 1, France [email protected]
Philip N. Johnson-Laird: Department of Psychology, Princeton University Princeton, NJ 08544, USA
[email protected]
Lars Konieczny:
Center for Cognitive Science, Albert-Ludwigs-Universitat Friedrichstr. 50, 79098 Freiburg, Germany
[email protected]
Markus Knauff: Center for Cognitive Science, Albert-Ludwigs-Universitat Friedrichstr. 50, 79098
Freiburg, Germany [email protected]

3
4 Contributors
4 Contributors
Michael Pauen:
Institut fur Philosophie, Otto-von-Guericke-Universitat Virchowstr. 24, 39016 Magdeburg, Germany
[email protected]
Klaus Rehkdmper: Institut fur Philosophie, Carl-von-Ossietzky-Universitat 26129 Oldenburg,
Germany [email protected]
Norbert M. Seel:
Institut fur Erziehungswissenschaft, Albert-Ludwigs-Universitat 79085 Freiburg, Germany seel@uni-
freiburg.de
Hannelore Van der Beken: Department of Experimental Psychology, Ghent University Henri
Dunantlaan 2, 9000 Ghent, Belgium
Andre Vandierendonck: Department of Experimental Psychology, Ghent University
Henri Dunantlaan 2, 9000 Ghent, Belgium Andre. Vandierendonck@UGent. be

Gottfried Vosgerau:
Philosophisches Seminar, Eberhard-Karls-Universitat Bursagassse 1, 72070 Tubingen, Germany
[email protected]
Randall L. Waechter: Department of Psychology, York University Toronto, Ontario, Canada M3J 1P3
Mental Models and the Mind Carsten Held, Markus Knauff, Gottfried Vosgerau © 2006 Elsevier B.V. All rights reserved
General Introduction: Current Developments in Cognitive

Psychology, Neuroscience, and the Philosophy of Mind
"Cognitive psychology," "cognitive neuroscience," and "philosophy of mind" are names for three
very different scientific fields, but they label aspects of the same scientific ambition: to understand the
nature of mental phenomena. Cognitive psychologists study mental processes as they are
indispensable for understanding human experience and behavior. They systematically observe such
behavior and then draw inferences from the observed data about unobservable mental processes. They
also apply their results to various domains of human life, including the design of new teach- ing
methods and the treatment of mental illness. Cognitive neuroscientists are concerned with the
connection between mental processes and the brain. They investigate how brain-events affect human
behavior. Philosophers of mind study the nature of mind, including consciousness, mental represen-
tation, and rationality. They ask questions such as: What is the relation between mind and brain, on the
one hand, and mind and world, on the other? Can machines think? How is the realm of beliefs and
knowledge con- nected to behavior? How come can I think about my own mental states? For many
decades, the three disciplines worked in relative isolation, but today they strongly overlap under the
roof of cognitive science. The goal of modern cognitive science, from our point of view, is to explain
how cognitive processes are related to and can be measured via behavior, how they are
computationally realized, and how these computations are biologically implemented in the brain. In all
sub-fields of cognitive science, the vast majority of researchers are familiar with the term "mental
model." Sometimes the expression is used as a synonym for "mental representation," but in most areas
it has a more
6 General Introduction
precise meaning. Building a bridge between the philosophy of mind and the empirical sciences of the
mind/brain, the present book develops a new perspective on the concept of mental models—from the
points of view of the mentioned disciplines: cognitive psychology, cognitive neuroscience, and
philosophy of mind. In the following, we provide a short introduction to the field. We initially sketch
some history of cognitive psychology and the newly emerging cognitive science as the background
against which the conception of a mental model has been invented (sec. 1-3). Then we describe the
conception in more detail (sec. 4) and outline the neuroscientific research it has inspired in recent
years (sec. 5). The last three sections draw a connection to the philosophy of mind. Mental models are
a special kind of mental representation. We sketch what philosophers think about mental
representation (sec. 6) and about how to make it scientifically accessible via a procedure called
"naturalization" (sec. 7). This will prepare the ground for a short outline of the special challenges
mental models produce for the philosophy of mind (sec. 8).

1. The decline of behaviorism and the cognitive turn
Today mental models play a key role in psychology and cognitive science, but that was not always
the case. Mental models basically ran through three different phases. In the early years of scientific
psychology, phenomena were described that we nowadays interpret as involving the use of mental
models, but they were described differently as the concept then did not exist. With the cognitive turn
(in the 1950) the conception developed very quickly and soon became one of the central notions of
cognitive psychology and cognitive science. In a third phase, the concept of mental models also
appeared in the cognitive neurosciences, where researchers are now searching for the neural correlates
of mental models. In the first part of the last century there was no room for mental models in
psychology. In his famous 1913 paper, Watson emphasized the study of observable behavior, rejecting
introspection—the direct observation of one's own inner life—and theories of the (un-)conscious as
unscientific approaches to psychology (Watson 1913). Following Watson's behaviorist proposal,
mentalistic descriptions were banned altogether from the psychological vocabulary in favor of
objective descriptions of behavior in depen- dence of stimuli. In psychological experiments, the stimuli
had to be con- trolled systematically and the ensuing behavior objectively measured. The method's aim
was to describe human (and animal) behavior as systematic regularities between input (stimuli) and
output (behavior). The cognitive system itself (the human brain) was viewed as a black box, the
internal
states of which are not amenable to scientific description or explanation. Therefore, the major concern
of behaviorists was conditioning (the learning from stimulus-response combinations). They attempted
to describe behavior of every kind, even as complex as linguistic behavior, entirely in terms of
stimulus-response patterns (cf. Skinner 1938, 1974). During the 1950s, more and more psychologists
began to challenge the behaviorist dogma of the cognitive system as a black box. A crucial step in this
development was a series of experiments with rats, described by Tolman in 1948. Though Tolman
started from behaviorist premises, his results turned out to be unexplainable without a concept of
mental representation. In one of the experiments, rats were trained to follow a path through a complex
maze in order to reach a food box. After the rats had performed perfectly (chosen the shortest way to
reach the goal), the trained path was blocked and the rats had to select another path from a variety of
alternatives. Astonishingly, most of the rats found a path that was close to the most direct connection
to the food box, whereas not a single rat erroneously tried to follow the original path on which they
had been trained. On the basis of these results, Tolman argued that the rats must have acquired an
internal representation (which Tolman did not call a "model") of the labyrinth. Today there is a large
body of evidence on how humans (and animals) explore their environment and mentally represent it.
Moreover, we now believe that such mental representations of spatial environments are constructed
even if we do not directly experience the environment when navigating through it, but also when we
just hear or read about it. Innu- merable studies in the field of text comprehension have shown that
mental models are routinely and immediately activated during word and sentence comprehension. If
individuals are asked to read texts, they regularly construct a mental model of the (possibly fictitious)
environment while reading (e.g. Glenberg 1997, Zwaan et al. 2002). As these results illustrate, today
most psychologists are convinced that more can be said about a cognitive system than just registering
input-output regularities. Its internal states, whatever they are, can be described in terms of the
functions they have for the whole system. In these terms, descriptions of the system's state can be given
that are, on the one hand, much more informative and detailed than a behaviorist would be willing to
grant, but that are, on the other hand, entirely independent of the concrete implementation. Progress in
the theory of computability was a major source for this new point of view (see next section).
The "cognitive turn", i.e. the switch from behaviorism to cognitive psychology, can be dated to the
appearance of Ulric Neisser's 1967 book Cogni- tive Psychology, which showed the application of the
new method to various areas in psychology. Soon, a new field called cognitive science arose from the
combination of methods from various disciplines. The leading method,

Table 1 The description of a chocolate vendor automaton with initial state 1
State
1122
Input
50 Cent
1 Euro
50 Cent
1 Euro
Output
none
chocolate
chocolate
chocolate, 50 Cent
Following State
2111
however, was imported from artificial intelligence research. As machines are programmed to solve
problems for which humans are said to require intelligence, humans themselves are viewed as such
problem-solving systems. The rationale of this comparison is the fact that, with reference to analo-
gous tasks, the states in both kinds of systems are functionally equivalent. Cognitive science today tries
to combine the experimental methods of cognitive psychology with the computational methods of
artificial intelligence in order to gain more insight into the functioning of the human mind. In regard of
the mentioned equivalence, the major premise of cognitive science is the "Physical Symbol System
Hypothesis" (Newell & Simon 1976). It states that the human brain is essentially a physical symbol
system. This implicates that cognition (what the human brain does) can be exhaustively described by
computational methods because a symbol system does nothing but computation. For this reason,
cognitive science is founded on conceptions of computational theory. It will be helpful to introduce
some of these conceptions.

2. The computational view of mind and the levels of description
The main concept of automaton theory is that of an abstract automaton. An automaton is defined by
states, in which the automaton can be, inputs, outputs, a function mapping every state and input to a
subsequent state and output, and an initial state. For simple automata, this can be written in a table—
see, e.g., the description of a chocolate vendor in table 1. The possible states of an automaton are
hence defined solely by the function mapping input and actual automaton state to output and
subsequent automaton state. Therefore, internal states of an automaton can be described functionally
(in terms of their "functional roles"). The view that cognitive
General Introduction 9 systems are automata and hence their internal l states can be described
functionally, is called functionalism (Fodor 1968, Putnam 2000). In mathematics, the intuitive notion
of an effectively calculable function has found several formalizations. One such famous formalization
involves the Universal Turing machine, introduced by Alan Turing 1936. This machine is a virtual
automaton and every function it can compute is called a Turing-machine computable function. Other
formalizations com- prise lambda-definable functions (Church 1932, Kleene 1935) and recursive
functions (Godel 1934, Herbrand 1932). It can be proved that all these formal analyses of the intuitive
notion of an effectively calculable function are equivalent. This result gave rise to the hypothesis now
known as the Church-Turing Thesis: Every function that is effectively calculable is Tur- ing machine
computable. The states of an automaton can be described in terms of functions. The functions
themselves can be viewed as effectively calculable in the sense of the Universal Turing machine.
Hence, every automaton can be modeled by a Universal Turing machine. However, the Universal
Turing machine is equivalent to other ways of modeling (lambda calculus, recursive functions, etc.).
Therefore, the functional description of an automaton is just as good (informative) as an algorithmic
description: it leads to a full description of an automaton in the sense defined above. There are many
equivalent ways to express the functions. Moreover, every algorithm can be implemented in various
ways. These considerations led to the characterization of three levels of description (Marr 1982): the
implementation level, the algorithmic level, and the computational (functional) level. An automaton
can be described on all three levels. However, the implementation level description does not offer
essentially new information compared to the functional level; on the contrary, details about algorithms
and their implementation are not interesting since there are many possible ways of implementing one
and the same (functionally described) automaton. Since the Universal Turing machine is a symbol
processing machine, computation can be identified with symbol processing. In cognitive science,
cognition is characterized as a form of computation, where internal states plus input are mapped to
internal states plus output. It follows that every form of cognition can be done by a physical symbol
processing machine. This hypothesis is exactly Newell and Simon's Physical Symbol System
Hypothesis. Adopting both the idea of different levels of description and the Phys- ical Symbol
System Hypothesis, human cognition can be fully described on a functional level. Because this level
can be described regardless of how

1 In the case of humans, internal states are often called mental states.
cognitive functions are implemented, human cognition can be implemented not only in the human
brain, but in any system equivalent to the Univer- sal Turing machine. Thus, also the functional
description of cognition that stems from behavioral experiments can be implemented on a computer.
The method of cognitive science is hence a combination of psychological methods and methods of
artificial intelligence. For each theory (functional description) of cognitive phenomena there should be
a possible implementation on a computer exhibiting the same phenomena. If both of these con-
straints (empirical accuracy and implementational possibility) are fulfilled, nothing of interest can be
added by looking at the original implementation (the brain).

3. The doctrine of mental logic
In the early years of the 20th century, the developmental psychologist Jean Piaget had studied how
children's ability to reason increases as they grow up. His research culminated in a theory of cognitive
development stating that children of different ages are equipped with (or have access to) different
inventories of inference rules as a basis for reasoning. Piaget's main assumption was that human
reasoning relies on a mental logic consisting of formal inference rules. More than fifty years later, the
computer metaphor of human cognition led to a renascence of the rule-based approach to reasoning.
These theories, especially prominent in the 1970s, state that mental representations have the form of
propositions, much like logical formulae. Reasoning is performed by applying syntactical rules to
transform propositions, like in logical proofs (Johnson-Laird 1975, Osherson 1975, Braine 1978, Rips
1983). Because this view was dominant at the time, Johnson- Laird (1983) calls it the doctrine of
mental logic. The view that humans basically perform syntactical transformations of propositions fits
very well with the program architecture known as rule- based systems. A rule based system has two
memories, one for propositions (the declarative memory) and one for syntactical rules (the procedu-
ral memory). If the system is given a set of new propositions it is able to select and apply rules from
the procedural memory. It will thereby generate new propositions (i.e. new information). This process
can, if necessary, include propositions from the declarative memory, for example axioms and
background knowledge. Therefore, theories of mental logic are very easy to program and they still are
popular among cognitive scientists. However, it is clear that humans, in many contexts, do not reason
logically sound. Especially if the context is poor and the reasoning task is abstract, humans fail to
generate correct conclusions. This fact is usually explained by limi-

tations of the working memory that must hold all the propositions needed for a certain inference. The
load on working memory, and consequently the difficulty of a task, increases with the number of rules
to be applied. Nev- ertheless, if the rules in the procedural memory are abstract logical rules, the
differences between reasoning with abstract and concrete material find no straightforward explanation.
A further question is, how these logical rules are learned and whether it is plausible to assume people
to have full abstract logical competence in that way. People with no background in psychology often
report that they do not use logical derivations but rather construct—before their mind's eye—an
integrated representation of the information given in the premises and then "read off" new
information, not explicitly given in the premises. This is the fundament of the theory of mental models.
4. Mental model theory
How do humans draw inferences? In contrast with the mental logic doctrine, a layperson will
quickly come up with the sensible idea that the content of all premises must be integrated into one
'picture.' Similarly, psychologists have conjectured that people integrate the information from the
premises into a single mental representation. In 1943, for instance, Kenneth Craik claimed that the
mind constructs "small-scale models" to anticipate events: If the organism carries a 'small-scale model' of
external reality and of its own possible actions within its head, it is able to try out various alternatives, conclude which is
the best of them, react to future situations before they arise, utilise the knowledge of past events in dealing with the present
and future, and in every way to react in a much fuller, safer, and more competent manner to the emergencies which face it.
(Craik 1943, 61) Craik's idea is the germ of what we today know as mental model theory. This theory
was first expounded by Philip Johnson-Laird in an article titled "Mental models in cognitive science"
(Johnson-Laird 1980) and, in full detail, in his book Mental models: towards a cognitive science of
language, inference and consciousness (Johnson-Laird 1983).2 The main purpose of this work was the
development of a theory of human thinking and reasoning that goes along with a critique of the
mentioned doctrine of mental

2 The editorial board of Cognitive Science, where Johnson-Laird's 1980 article appeared, has recently identified several
classic articles from the journal from the last couple of decades. The members of the committee rated Johnson-Laird's article
among the top ten because of its impact, innovation, and importance in promoting theoretical development in the field of
cognitive science.
logic in cognitive science. Johnson-Laird showed that human thinking and reasoning can be modeled
without propositional representations but, at the same time, saving the advantages of describing mental
processes in computational terms. He argues that mental models rather than formal logic underlie
syllogistic inferences, e.g., "All men are animals, the professor is a man, therefore the professor is an
animal." The article was the first to present a case for mental models as a computational theory of
human thought. Johnson-Laird extensively argued that different levels of description must be
distinguished in order to describe cognitive processes. Below the behavioral level lies the
computational level, and Johnson-Laird argued that a reasoning theory must be translated into a
computer language in order to be executable on a computer. The researcher must specify the
underlying representational format and the procedures that generate and manipulate this representation.
However this description must be in functional terms, rather then in terms of the bits and bytes that
move around in the computer hardware. Johnson-Laird thus endorses what we have called the
independence of functional level and implementation level. This is expressed, e.g., in the following
quotation:
We should not worry about the particular computer and its machine code, since the program could be executed on some
very different machines, and we do not want to make a different characterization for all these different sorts of computer.
(Johnson-Laird 1980, 100) According to mental model theory, human reasoning relies on the con-
struction of integrated mental representations of the information that is given in the reasoning
problem's premises. These integrated representations are the mental models. A mental model is a
mental representation that captures what is common to all the different ways in which the premises can
be interpreted. It represents in "small scale" how "reality" could be— according to what is stated in the
premises of a reasoning problem. Mental models, though, must not be confused with images. A mental
model often forms the basis of one or more visual images, but some of them represent situations that
cannot be visualized (Johnson-Laird 1998). Instead, mental models are often likened to diagrams since,
as with diagrams, their structure is analogous to the structure of the states of affairs they represent.
Prom the processing view, the model theory distinguishes between three different operations. In the
construction phase, reasoners construct the mental model that reflects the information from the
premises. In the inspection phase, this model is inspected to find new information that is not explicitly
given in the premises. In most variants of the model theory, the inspection process is conceptualized as
a spatial focus that scans the model to find new information not given in the premises (Bara et al. 2001,
Ragni et al. 2005, Schlieder & Berendt 1998). In the variation phase, rea-
soners try to construct alternative models from the premises that refute the putative conclusion. If no
such model is found, the putative conclusion is considered true. The theory's central idea is that of an
analogy of a mental model's structure with the structure of the situation modeled. In this respect,
mental models, much like the pictures of Wittgenstein's 1922 "picture theory", represent a certain
situation conceived in just one possible way. Sometimes a model captures what is common to all of the
different ways in which the possibility may occur. Then it is a perfect basis for reasoning. However,
sometimes—and in fact most frequently—reasoners are unable to survey the entire set of possibilities
and thus focus on a subset of possible models— often just a single model—which leads to incorrect
conclusions and illogical decisions. It is interesting that humans have preferences if a problem has
multiple solutions and that most people agree in their preferences. For a certain task, we tend to
construct almost the same single model—the preferred mental model—and to ignore others (Knauff
et al. 1998). The crucial difference with theories of mental logics is that no knowledge of logical rules
must be presupposed. The reasoner constructs and manip- ulates mental models not according to
abstract logical rules but according to the world which she represents. After having integrated all the
information of the premises in one (or more) consistent models, the conclusion can be directly "seen"
in the model (and eventually compared with conclusions from other models). In this way, logically
sound reasoning "emerges" from the format of representation. Failure of sound reasoning can be ex-
plained, as sketched above, from the fact that not all relevant models are constructed for many
problems. As illustrated in this book, cognitive psychologists have explored mental models from
very different points of view and carried out an extensive research program on how models engender
thoughts and inferences. Since the understanding of (linguistically presented) premises involves text
comprehension, the theory has been extended to provide a psycho-linguistic approach to semantic
processing. In the field of education, the role of mental model construction in learning has been
explored. The question whether mental model theory also contributes to the understanding of (visual)
perception is currently discussed in cognitive psychology and the philosophy of perception. These are
just a few examples for the immense influence of mental model theory across the borders of academic
disciplines. Today, this research effort is much more successful than the classical rule based
approaches of reasoning.
5. Mental models and the brain
In the last years, the position of the implementation-independency has been graded down because
most cognitive scientists now believe in the assumption that the understanding of brain-events can
provide insight into the computations they implement (cf. Gazzaniga et al. 2002, Johnson-Laird 1995).
That is what we call the third phase of mental models research. Still, behavioral methods are the via
regia to understand human thinking and reasoning with mental models. Today however,
neuroscientific research is adding important information about the characteristics of mental models.
Studies with brain-injured patients gave us an initial idea about which areas of the brain are involved
in thinking and reasoning and the availability of modern brain imaging methods currently contributes
enormously to our understanding of human thought and behavior. Researchers with a background in
mental model theory differ in some respects from other fields of cognitive neuroscience. They are still
cognitive psychologists with the goal to understand human experience and behavior and the
intervenient computational processes. They are not so much interested in what Uttal (2001) called the
new phrenology, namely the localization of cognitive processes including all the reductionistic
implications. Instead, they treat changes in cortical blood flow as a dependent variable, much as
response times or error rates. The background of this new turn in mental models research is that the
mentioned "independence of computational level" hypothesis makes some questionable assumptions,
after all. In particular, the supposition that each function computable by a Turing Machine can be
computed on all Turing-equivalent machines is not unqual- ifiedly true (Giunti 1997, Goel 1995).
Though it is true that computational processes can be realized in many different systems, it is not true
that they can be realized in all Turing-equivalent machines. The assumption of universal realizability
thus appears to be unwarranted (Goel 2004). A second reason for the new interests of mental model
researchers is that localization and dissociation can help to understand the cognitive processes
themselves. As Goel (2004) puts it, Gall & Spurzheim (1810-1819) was basically right and Lashley
(1929) wrong about the organization of the brain. Not all neural computations can be realized in all
brain areas. We know that there are highly specific brain regions dedicated to specific computations.
For instance, there are brain areas that exclusively process information in a verbal format, whereas
other areas only respond to incoming information in a visuospatial format. For the testing of
hypotheses it is essential that these cortical systems can be identified with rule-based and model-based
reasoning processes. Language-related brain areas are often identified with rule-based theories of
thinking, whereas activity in visuospatial brain ar-

eas during thinking is typically identified with mental models. A similar distinction is related to the
two brain hemispheres, where the right brain is related to the processing of mental models and the left
brain to more language-based abstract inference processes. Such localization can help us to test
different cognitive theories, since different theories—namely mental logic and mental model theories
of reasoning—make different predictions concerning the involved brain areas.

6. Mental representation as a philosophical problem
Mental models are mental representations of a certain type. The main problem in the philosophy of
mental representation is to characterize the relation between a mental representation and the
represented object. Naively speaking, a mental representation is an entity that 'stands for' another—the
represented object—, but here 'stands for' is just a metaphoric place-holder for 'represents', thus
requires further explanation. Obvious features of the representation relation can be isolated. First, it is
an asymmetric relation (if X represents Y, then Y usually does not represent X); and, second, there are
cases of misrepresentation, where, e.g., a cognitive system represents a horse as a cow. In the recent
literature on representation there are three main types of representation theories tackling these
problems: causal theories, similarity theories, and functional theories. The basic idea of the first type
of theories is that some mental entity represents another (non-mental or mental) entity because the
first is caused by the second. In this way, the asymmetry of the relation is straightforwardly explained.
However, the problem of misrepresentation is much harder to deal with. As sometimes Xs cause Y-
representations (misrepresentation), Y-representations should have a disjunctive meaning "X or Y"
according to the causal approach. To avoid this disjunction problem, Fodor (1987) introduces a nomic
relation between the Y-caused Y-representations and the X-caused Y-representations such that X-
caused Y-representations can occur only when there are Y-caused Y-representations, but not vice
versa. Similarity theories are based on the assumption that representations are similar to what they
represent. A general problem for these theories obviously is the explanation of asymmetry. 3 However,
the main problem seems to be to characterize the kind of similarity. Indeed, many different kinds have
been proposed (cf. Cummins 1989). Nowadays, the most attractive similarity relation seems to be
isomorphism, proposed in both philosophy (e.g. Cummins 1996, French 2002) and in psychology
(especially for mental

Although there are some similarity relations that are non-symmetric.
models; e.g. Palmer 1978, Gurr 1998). Still, it remains an open question what misrepresentation is: A
Y-representation is not similar to an X in the relevant sense (otherwise it would be an X-
representation), but still can be erroneously taken to represent X. The last group of theories imports
the notion of function from biology. Mental representations can be characterized by the functions they
have and eventually fulfill, just like organs or traits of an organism are characterized by their functions.
A heart is an entity that can be fully characterized by its function of pumping blood (independent of
whether it actually fulfills this function or not). In the very same way, a mental representation is an
entity that has the function to represent (e.g. Dretske 1994, Millikan 1984, 1989). The function of a
horse-representation, for example, is to stand for the horse within the functional architecture of a
cognitive system. Because of this function in the system it leads to certain states of the system con-
cerning the horse. A cow-representation, on the other hand, would be, in an obvious sense,
dysfunctional in this context. It does not lead to states concerning the horse (but to states concerning a
cow that is not there) and thus does not (indeed cannot) fulfill the function to stand for a horse.
7. Intentionality and its naturalization
The mentioned theories try to come to grips with asymmetry and misrepresentation. What about the
original feature that a mental representation 'stands for' the object it represents? This property
traditionally is an intentional property because it is possible that a mental representation stands for
something non-existent (as the cow-representation did in the previous example). Being a relation to
something possibly non-existent, arguably, is the mark of the intentional. More exactly, intentional
states are goal- directed, but these goals need not to exist. This non-existence implies that the states are
open-ended in the sense that they can have or not have a relatum (see, e.g. Chisholm 1957, 170).
Representations are intentional and open-ended states in this sense: they are directed toward the
entities they stand for and these entities need not exist even if the representations do. Obviously, the
properties of a mental representation's reference and intentionality are in urgent need of further
explanation, but philosophers are deeply divided about what form that explanation should take. One
large group thinks that such explanation should proceed in terms of relations of agents with their
environment that are more readily accessible to scientific treatment and direct observation. This
project is often described as "naturalizing the mind" (see, e.g. Loewer 1997). Others think that this
project is deeply misguided (see, e.g. Putnam 1982). The latter attitude,
well-founded as it may be, seems unhelpful in view of the fact that the cognitive sciences do already
investigate the mind with techniques that are both very successful and regarded as scientifically
acceptable. One way to summarize the qualms about the naturalization project is as follows. The
project is successful where it underpins empirical research on the mental as we see it done today,
because it rectifies suspicious philosophical vocabulary into scientifically acceptable terminology.
However, it fails where it aims to explain mental states in scientific terms by eliminating the
intentional vocabulary because the latter either will be tacitly reintro- duced to make naturalized
descriptions applicable (in the case of physical descriptions) or is entrenched in the scientific
terminology from the outset (in the case of biological descriptions). A good starting point to illustrate
this fact is, again, the "Physical Sym- bol System Hypothesis." In Simon's words, the hypothesis states
"that a physical symbol system [... ] has the necessary and sufficient means for general intelligent
action" (Simon 1996, 23). As we saw, the hypothesis is interpreted as saying that humans think and
represent by manipulating symbols and as such it has been a fruitful research hypothesis in cognitive
science. The ensuing hypothesis that humans are physical symbol systems is a helpful tool to make
several of their mental activities scientifically accessible in a new and fruitful way. Intentional
language is imprecise and thus, in comparison with, say, an algorithm for symbol manipulation, less
suited for describing such activities exactly. However, naturalization is at work here only in the sense
that the mental is made scientifically accessible by means of new tools. Only when physical symbol
systems, and thus humans, are interpreted as mere physical systems a more serious naturalization
project is initiated. Clearly, Simon himself aims in this direction. He introduces symbols as physical
patterns with the clear intention of interpreting physical symbol systems as purely physical systems.
The proposal is interesting, but ultimately doomed to fail—or so the skeptic will argue. Simon writes:
"Symbol structures can, and commonly do, serve as internal representations (e.g., 'mental images') of
the environment to which the symbol system is seeking to adapt." (Simon 1996, 22) It is here that the
skeptic will claim an ill-reflected re-introduction of intentional vocabulary ("serve," "seek," "adapt")
into a context that pretends to be pure physics. In cognitive science, the reliance on the program
advocated by Simons is history. As is reasonable, these sciences today make unrestricted use of
biological vocabulary. Do attempts to explain mental representation in this vocabulary count as
naturalizations? A key notion of biological explanations is the one of function. Indeed, we can
understand the above-mentioned functionalist theory of mental representation as a naturalization
project, if the mental representations having (and fulfilling or failing to fulfill) certain functions are
identified with organic states having these functions. Here,

it is the notion of function where the skeptic suspects a vicious circle. To explain intentional relations,
naturalizers utilize a teleological notion from biology, which is itself in need of explanation.
Philosophers of biology have long quarreled about the appropriate characterization of biological func-
tions (see, e.g. Sober 2000). Attempts at a coherent and satisfying characterization differ in many
important details, but nowadays they mainly appeal to natural selection as an explanation of a
biological subsystem's function for the including system. But it is specifically selection in animate
systems, thus in systems that exhibit biological activity, that is exploited for philosophical accounts of
biological functions, as, e.g., when the chance of reproduction under selective pressure is characterized
as the chance to survive and/or sexually reproduce. So, those who employ functional notions to
describe an entity tacitly refer to intentional notions via the goal- directedness of the very organism of
which the functional or dysfunctional element is an organic part or state. E.g., explaining a frog's
mental representation of a fly in terms of a state that functions to aid the frog in catching its food is to
utilize a notion that itself makes tacit appeal to the frog's activity of trying to eat, because that is what
that state, if it functions, contributes to. In turn, the fact that an equal state contributed, via
contribution to successful eating behavior, to the survival and sexual reproduction of the frog's
ancestors, explains the state's existence in the frog, but this evolutionary explanation of the functional
element presupposes a goal-directed activity of the frog and its ancestors. Without scientific
explanation the intentional phenomena remain mys- terious and it seems that such explanation must
take the course of naturalization. So far, however, only weak naturalization—making the mental
scientifically accessible while consciously preserving its description in the intentional vocabulary—is a
successful project. And it appears that the sciences don't need more naturalization.
8. Mental models and the philosophy of mind
The psychological insight that humans use mental models in many cognitive processes gives several
issues in the philosophical debate a new twist. The status of such models as mental entities stands and
falls with the one of mental representations in general. However, for those philosophers who want to
show that mental representations, as entities in their own right and with their own distinctive features,
do ultimately not exist, mental models raise the bar. Mental models have initially been proposed as
special representations that explain how humans reason, so the philosopher denying
the existence of mental representations faces the challenge: "Explain how people think!"—but
without mental models (see Johnson-Laird 1996, 90). Mental models offer still more challenges. After
all, cognitive psychologists distinguish them from other types of representation through their specific
functions (Johnson-Laird 1996, 91). Models are assumed to represent classes of situations as opposed
to images representing single situations (Johnson-Laird 1996, 120). This idea is based on the premise
that models contain abstract elements, which is in turn based on the assumption that they have the
function of aiding the execution or evaluation of syllogistic reasoning. Thus, it is essential to a
philosophical understanding of mental models to functionally differentiate them from other types of
representation.This differentiation involves several aspects: Firstly, it has to be described in what
respect the relation between a model and the represented situation differs from the relations between
other forms of representation and the objects they represent. Secondly, the question of what mental
models represent at all has to be answered. Thirdly, the status of the models itself within the cognitive
system has to be contrasted with the status of other forms of mental representation. This differentiation
involves not only a description of what a model is for the reasoner, but also an answer to the question
what kind of entities mental models are ontologically.

References
Bara, B., Bucciarelli, M. & Lombardo, V. (2001), 'Model theory of deduction: A
unified computational approach', Cognitive Science 25, 839-901. Braine, M. D. S. (1978), 'On the relation
between the natural logic of reasoning
and standard logic', Psychological Review 85, 1-21. Chisholm, R. (1957), Perceiving: A Philosophical Study,
Cornell University Press,
Ithaca. Church, A. (1932), 'A set of postulates for the foundation of logic', Annals of
Mathematics, second series 33, 346-366. Craik, K. (1943), The Nature of Explanation, Cambridge University
Press, Cam-
bridge. Cummins, R. (1989), Meaning and Mental Representation, The MIT Press, Cam- bridge, MA, London.
Cummins, R. (1996), Representations, Targets, and Attitudes, The MIT Press,
Cambridge, MA, London. Dretske, F. (1994), Misinterpretation, in S. Stich, ed., 'Mental Representation. A
Reader', Blackwell, Cambridge, MA, Oxford, pp. 157-173. Fodor, J. (1987), Psychosemantics, The MIT Press,
Cambridge, MA, London. Fodor, J. A. (1968), Psychological Explanation, Random House, New York.
French, S. (2002), 'A model-theoretic account of representation', Proceedings of
the PSA (Supplement). Gall, F. J. & Spurzheim, J. (1810-1819), Anatomie et Physiologie du Systeme
Nerveux en General et du Cerveau en Particulier, F. Schoell, Paris. Gazzaniga, M. S., Ivry, R. & Mangung, G.
R. (2002), Cognitive Neuroscience:
The Biology of the Mind, 2nd edn, W.W. Norten. Giunti, M. (1997), Computation, Dynamics and Cognition,
Oxford University
Press, New York. Glenberg, A. M. (1997), 'What memory is for', Behavioral and Brain Sciences
20, 1-55. Godel, K. (1934), 'On undecidable propositions of formal mathematical systems', Lecture notes taken
by Kleene and Rosser at the Institute for Advanced Study (Reprinted in M. Davis, ed., 1965, 'The Undecidable',
New York, Raven). Goel, V. (1995), Sketches of Thought, MIT Press, Cambridge, MA. Goel, V. (2004), Can
there be a cognitive neuroscience of central cognitive systems?, inD. Johnson & C. Erneling, eds, 'Mind as a
Scientific Object: Between Brain & Culture', Oxford University Press, Oxford. Gurr, C. A. (1998), On the
isomorphism, or lack of it, of representations, in K. Marriott & B. Meyer, eds, 'Visual Language Theory',
Springer, New York, Berlin, Heidelberg. Herbrand, J. (1932), 'Sur la non-contradiction de l'arithmetique',
Journal fur die
reine und angewandte Mathematik 166, 1-8. Johnson-Laird, P. (1996), Images, models, and propositional
representations, in M. de Vega, M. Intons-Peterson, P. Johnson-Laird, M. Denis & M. Marschark, eds, 'Models

of Visuospatial Cognition', Oxford University Press, New York, pp. 90-127. Johnson-Laird, P. N. (1975),
Models of deduction, in R. J. Falmagne, ed., 'Rea- soning: Representation and Process in Children and Adults',
Erlbaum, Hills- dale. Johnson-Laird, P. N. (1980), 'Mental models in cognitive science', Cognitive Sci-
ence 4, 71-115. Johnson-Laird, P. N. (1983), Mental Models, Harvard University Press, Cam-
bridge, MA. Johnson-Laird, P. N. (1995), Mental models, deductive reasoning, and the brain, in S. Gazzaniga,
ed., 'The Cognitive Neurosciences', MIT Press, Cambridge, MA, pp. 999-1008. Johnson-Laird, P. N. (1998),
Imagery, visualization, and thinking, in J. Hochberg, ed., 'Perception and Cognition at Century's End',
Academic Press, San Diego, CA, pp. 441-467. Kleene, S. C. (1935), 'A theory of positive integers in formal
logic', American Journal of Mathematics 57, 153-173, 219-244. Knauff, M., Rauh, R., Schlieder, C. & Strube,
G. (1998), Mental models in spatial reasoning, in C. Freksa, C. Habel & K. F. Wender, eds, 'Spatial Cognition
— An Interdisciplinary Approach to Representing and Processing Spatial Know- ledge', Springer, Berlin, pp.
267-291. Lashley, K. S. (1929), Brain Mechanisms and Intelligence: A Quantitative Study
of Injuries to the Brain, University of Chicago Press, Chicago. Loewer, B. (1997), A guide to naturalizing
semantics, inB. Hale &: C. Wright, eds,
'A Companion to the Philosophy of Language', Blackwell, Oxford, pp. 108-126. Marr, D. (1982), Vision: A
Computational Investigation in the Human Represen- tation of Visual Information, Freeman, San Francisco.
Millikan, R. G. (1984), Language, Thought, and Other Biological Categories, The MIT Press, Cambridge,
MA, London. Millikan, R. G. (1989), 'Biosemantics', The Journal of Philosophy 86(6), 281-297. Neisser, U.
(1967), Cognitive Psychology, Appleton-Century-Crofts, New York. Newell, A. & Simon, H. (1976),
'Computer science as empirical inquiry: Symbols and search', Communications of the Association for
Computing Machinery 19, 113-126. Osherson, D. N. (1975), Logic and models of logical thinking, in R. J.
Falmagne, ed., 'Reasoning: Representation and Process in Children and Adults', Erlbaum, Hillsdale. Palmer, S.
(1978), Fundamental aspects of cognitive representation, in E. Rosch & B. L. Lloyd, eds, 'Cognition and
categorization', Erlbaum, Hillsdale, NJ, pp. 259-302. Putnam, H. (1982), 'Why reason can't be naturalized?',
Synthese 52, 3-23. Putnam, H. (2000), The Threefold Cord, Columbia University Press. Ragni, M., Knauff, M.
fc Nebel, B. (2005), A computational model of human reasoning with spatial relations, in 'Proceedings of the
Twenty Seventh Annual Conference of the Cognitive Science Society', Erlbaum, Mahwah, NJ. Rips, L. J.
(1983), 'Cognitive processes in propositional reasoning', Psychological
Review 90(1), 38-71. Schlieder, C. & Berendt, B. (1998), Mental model construction in spatial reasoning: A
comparison of two computational theories, inU. Schmid, J. F. Krems & F. Wysotzki, eds, 'Mind Modelling: A
Cognitive Science Approach to Reason- ing, Learning, and Discovery', Pabst Science Publishers, Lengerich,
pp. 133- 162. Simon, H. (1996), The Sciences of the Artificial, 3rd edn, MIT Press, Cambridge,
MA. Skinner, B. F. (1938), Behavior of Organisms, Appleton-Century-Crofts, Inc., New York. Skinner, B. F.
(1974), About behaviorism, Knopf, New York. Sober, E. (2000), Philosophy of Biology, Westview Press,
Boulder, CO. Tolman, E. C. (1948), 'Cognitive maps in rats and men', Psychological Review
55, 189-208. Turing, A. M. (1936), 'On computable numbers, with an application to the
Entscheidungsproblem', Proceedings of the London Mathematical Society 42(2), 230-265. Uttal, W. R. (2001),
The New Phrenology, MIT Press, Cambridge, MA. Watson, J. B. (1913), 'Psychology as a behaviorist views
it', Psychological Review
20, 158-177. Wittgenstein, L. (1922), Tractatus Logico-Philosophicus, Routledge & Kegan
Paul, London.
Zwaan, R. A., Stanfield, R. A. & Yaxley, R. H. (2002), 'Do language comprehen-
ders routinely represent the shapes of objects?', Psychological Science 13, 168- 171.
Part I
Cognitive Psychology
Mental Models and the Mind 25 Carsten Held, Markus Knauff, Gottfried Vosgerau © 2006 Elsevier B.V. All rights
reserved
Introduction: Cognitive Psychology

Competence, difficulty, content—these three keywords stand for the sub- jects central to the study of
human thinking and reasoning. They mirror three research questions: By which mechanisms can
individuals reason? What factors cause reasoning difficulty? And: How do content and background
knowledge affect reasoning performance? In the last decades, reasoning research made much
progress in answering these questions. In specific cases, we think by applying mental rules, which are
similar to rules in computer programs. In most of the cases, however, we reason by constructing,
inspecting, and manipulating mental models. These models and the processes that manipulate them are
the basis of our competence to reason. In general, it is believed that humans have the competence to
perform such inferences error-free. Errors do occur, however, because reasoning performance is
limited by capacities of the cognitive system, misunderstanding of the premises, ambiguity of
problems, and motivational factors. More- over, background knowledge can significantly influence our
reasoning performance. This influence can either be facilitation or an impedance of the reasoning
process. Technically speaking, the abstract (logical) truth value of an inference can be the same as the
truth value of our prior knowledge— in this case the inference is supported. Or, the formal truth value
conflicts with the truth value of the prior knowledge—then the inference is more difficult, which
means it results in more errors or takes significantly longer. The first three chapters of this book are all
concerned with the mechanisms of reasoning, the causes for errors, or with the connection between
reasoning and prior knowledge. Johnson-Laird himself uses the model theory to explain how
individuals reason with sentential connectives, such as "if", "or", and "and" and why we commit errors
in such task. He shows that certain inferences yield systematic fallacies and that these fallacies can be
explained perfectly by the use of mental models. The chapter explains the models theory's predictions
and reports some studies corroborating the

26 Part I
occurrence of these "illusory" inferences. His research is impressive because no other theory is able to
explain the experimental findings. The chapter by Vandierendonck, Dierckx, and Van der Beken is
concerned with the connection between mental models and background knowledge. Vandierendonck
explains this connection in the field of relational reasoning and shows that this kind of inference is
based on an interaction of knowledge represented in semantic and episodic long-term memory on the
one hand and temporary information maintained in working memory on the other hand. His account is
very plausible and fits nicely with many experimental findings: Reasoners have a preference for visuo-
spatial representations, believability affects reasoning performance, and reasoning with transitive and
intransitive relations is related to different types of prior knowledge. Overall, the chapter shows that
reasoning (with relations) is based on a tight interplay of knowledge representations in long-term mem-
ory and temporary models in working memory. The chapter by Seel revolves around the function of
mental models in learning. The author is a pedagogue and thus interested in the potentials of mental
models to facilitate learning. For him, learning situations require the construction and manipulation of
mental models. His main argument is that models support the simplification and visualization of the
learning materials. The chapter reports on two empirical investigations that empha- size the facilitating
effects on models in multimedia learning and discovery learning. All three chapters of this part of the
book indicate the likely direction of future empirical research. Firstly, behavioural experiments will
continue to be the via regia to study human thinking and reasoning by means of mental models. They
will continue to be the most helpful means to understand the nature of human reasoning, in particular if
they are—as in the next chapters—combined with methods from cognitive neuroscience. Secondly,
mental models researchers will continue to suggest modifications and re- finements to explain new
experimental findings. And finally: The theory of mental models will find its way into applications.
The use of mental models in learning research is one example. Many other examples come from
computer science, especially from artificial intelligence, where the ortho- dox view that logic
representations together with forms of logical inference are sufficient to exhibit intelligent behavior is
complemented—and even rejected—by representations in the form of mental models.

Mental Models and the Mind 27 Carsten Held, Markus Knauff, Gottfried Vosgerau © 2006 Elsevier B.V. All rights
reserved
Mental Models, Sentential Reasoning, and Illusory Inferences

P.N. Johnson-Laird1
Department of Psychology, Princeton University 2
Abstract This chapter describes how individuals reason with sentential connectives, such as "if," "or," and
"and." They do not have a "truth functional" semantics for these connectives, but rather they construct models
of the possibilities compatible with sentences in which the connectives occur. Human working memory has a
limited processing capacity, and so individuals aim to construct only a single model at a time, and to represent
only those clauses in the premises that hold in each possibility. One unexpected consequence of the theory
emerged from its computer implementation. Certain inferences should yield systematic fallacies if reasoners
use mental models. The chapter explains this prediction and reports some studies corroborating the occurrence
of these "illusory" inferences. No one has yet devised an account of them on the basis of another theory.
Suppose that you are carrying out a test of system and you know that if the test is to continue then the
reactivity of the system must not have reached the critical level. You then observe that the reactivity has
reached the critical level. What should you do? It seems obvious that you should stop the test. The engineers in
charge at Chernobyl were in this position, but they continued the test (see Medvedev 1990). Why they
continued is puzzling, because the test was not only dangerous, but pointless. It led to
1 This research was supported by a grant from the National Science Foundation (BCS- 0076287) to study strategies in
reasoning. For their helpful advice, I thank Ruth Byrne, Vittorio Girotto, Geoff Goodwin, Uri Hasson, Karl Christoph
Klauer, Louis Lee, Markus Knauff, Walter Schroyens, Andre Vandierendonck, Clare Walsh, and Yingrui Yang. 2 E-mail:
[email protected]
28 P.N. Johnson-Laird
the disaster. One possibility is that the engineers failed to make a valid inference of the form: If A
then not B. B.Therefore, not A. where A stands for "the test is to continue" and B stands for "the
reactivity has reached the critical level." For several years, I have given groups of engineering
students a similar problem with an abstract content, such as: If there is a triangle on the board then
there is a circle on the board. There isn't a circle on the board. What, if anything, follows? Typically,
more than half of them respond that nothing follows from these premises. In fact, the premises yield
the conclusion: There is not a triangle on the board. This conclusion is valid: it must be true given that
the premises are true. But, the inference is quite difficult to make. The engineers are not reluctant to
make inferences, because with premises of this sort: If there is a triangle on the board then there is a
circle on the board. There is a triangle on the board, nearly all of them draw the valid conclusion:
There is a circle on the board. People do make mistakes, and the difference in difficulty between the
two previous inferences is one of the most robust effects in the psychology of reasoning (see, e.g.
Evans et al. 1993). Yet, reasoners are not always wrong. Psychologists therefore need to explain both
their logical ability and the cause of their mistakes. My aim in this chapter is to describe the mental
mechanisms underlying a major sort of reasoning, so-called "sentential reasoning", which is based on
negation and sentential connectives, such as "if," "or," and "and." The account is a development from
the theory of mental models (Johnson- Laird 1983, Johnson-Laird & Byrne 1991). The theory
postulates that the mind constructs models of the world that it uses to reason. It constructs them from
perception (Marr 1982), imagination (Metzler & Shepard 1982), knowledge (Gentner & Stevens
1983), and the comprehension of discourse (Stevenson 1993, Polk & Newell 1995, Oakhill &
Garnham 1996, Garnham 2001). A crucial distinction between models and other sorts of proposed
mental representation is that the structure of models corresponds to the structure of what they
represent: individuals are represented by individual tokens, properties by properties of these tokens,
and relations by relations among these tokens (see, e.g. Johnson-Laird 1983). In reasoning, a key step
is to establish a conclusion; its strength depends on whether any models of the premises refute it
(Johnson-Laird & Byrne

Mental Models, Sentential Reasoning, and Elusory Inferences 29
1991). The theory therefore provides a unified account of reasoning about what is necessary, probable, or
possible. A conclusion is necessary if it holds in all the models of the premises, it is probable if it holds in
most models of the premises (Johnson-Laird et al. 1999), and it is possible if it holds in at least one model of
the premises (Bell & Johnson-Laird 1998).
The model theory, as I refer to it, is based on a core principle that concerns the interpretation of connectives,
and that gives rise to systematic fallacies. These fallacies can be so compelling that they have an illusory
quality: it is hard to avoid succumbing to them even when you are on guard against them. You will understand
the principle more easily if I outline elementary logic. Hence, the chapter begins with such an account. It then
describes the interpretation of connectives in natural language, and illustrates the limitations of human working
memory. These limitations lead to the fundamental principle of the model theory: Mental models are
parsimonious. The chapter formulates the mechanisms that implement this principle in the construction of
mental models, which it contrasts with the reasoning of superhuman entities with unlimited working memories.
It reports some illustrative results of recent studies of the illusory inferences. These results corroborate the
theory.
1. Logic and truth-functional connectives
Logic treats sentences as expressing propositions; in everyday life, however, the proposition that a sentence
expresses almost always depends on its context. "I can hear you now"—an utterance all too common these
days—expresses different propositions depending on who says it, to whom it is addressed, and the time and
circumstances of the utterance. To keep matters simple, I will use sentences that depend as little as possible on
their context, and, where feasible, I will adopt the fiction that sentences are propositions.
Logic is the science of valid inferences. It is not concerned with how people make such inferences. Logicians
have formulated many different cal- culi for formalized languages. They can set up a calculus in two distinct
ways (see, e.g. Jeffrey 1981). The first way is formal, concerning patterns of symbols, but not their
interpretation. The sentential calculus concerns sentential connectives in their logical senses—a notion that I
will explain soon. Its formal specification depends on rules of inference, such as:
A or B, but not both. not-B Therefore, A.
Table 1 The truth table for an inclusive disjunction
There is a circle There is a triangle. There is circle on the board or
on the board. a triangle or both.
True
True
True True
False
True False
True
True False
False
False
The variables, A and B, can have as values any declarative sentences what- soever. The second way
to characterize a calculus is semantic. Consider an atomic sentence, i.e., one that contains neither
negations nor connectives: There is a circle on the board. Let's suppose that it is false. A compound
sentence is made from atoms by combining them with negation or sentential connectives. Here is a
negative compound: There is not a circle on the board. This assertion is true because, as I just told
you, the atom that it contains is false. Suppose that you also know another compound assertion, which
is a disjunction of two atoms: There is a triangle on the board or there is a circle, or both. This
disjunction is inclusive, because it allows that both atoms could be true. Hence, its meaning is
compatible with three possibilities: There is a triangle on the board and there is not a circle. There is
not a triangle on the board and there is a circle. There is a triangle on the board and there is a circle.
You already know that there is not a circle, and so you can eliminate all but the first possibility. It
follows that that there is a triangle. The formal rule above also allows you to make this inference, but
here you have made it on a semantic basis. Hence, in principle, human reasoning could be based on
formal procedures or semantic procedures, or both. The meaning of the preceding disjunction can be
laid out in the form of a truth table, which specifies the truth value of the disjunction for each of the
four possible contingencies—the three possibilities in which it is true, and the remaining possibility in
which it is false. Table 1 presents this truth table. Each row in the table shows a possible combination
of the truth values of the two atoms, and the resulting truth value of their inclusive disjunction. For
example, the first row is the possibility in which both atoms are true, and, as a result, the inclusive
disjunction is true too.

Mental Models, Sentential Reasoning, and Illusory Inferences 31
Truth tables were invented by the great American logician, Charles Sanders Peirce (see, e.g. Berry
1952), though Wittgenstein (1922) is often wrongly credited with their invention. A sentential
connective has a "logical" meaning when its interpretation can be summarized in a truth table. The
truth table shows how the truth value of a sentence containing the connective depends solely on the
truth values of its constituent propositions. Once you know their truth values, you can work out the
truth value of the sentence as a whole from the connective's truth table. Hence, an inclusive disjunction
in its logical sense is true or false solely as a function of the truth values of the constituent propositions.
As logicians say, a disjunction has a truth-functional meaning. This piece of jargon means: you feed in
truth values of the constituent propositions, and the truth table for "or" gives an output of a truth value.
In logic, a general recipe exists for interpreting compound sentences. You replace each atom with its
truth value—how you obtain such truth values is not part of the theory—and you progressively
simplify the compound according to the interpretation of each connective, until you arrive at a final
truth value for the sentence as a whole. This truth value depends only on the truth values of the atoms
and the truth-functional interpretations of the connectives. Here is an example of such an
interpretation. Consider the compound assertion in which "or else" is an exclusive disjunction, i.e.,
only one of the two clauses it connects is true: (A and not B) or else (C and D) and assume that all the
atoms are true: A: B, C, and D, are all true. The first step in the interpretation of the compound is to
replace its atoms with their truth values: (true and not true) or else (true and true) The next steps
simplify the expression according to the truth-functional meanings of negation and the connectives:
(true and false) or else (true and true) —according to the
meaning of not (false or else true) —according to the meaning of and true —according to the meaning
of or else Hence, the compound assertion is true given the values of its atoms. Logicians can use truth
tables to determine whether or not an inference is valid: It is valid if its conclusion must be true given
that its premises are true. One of the glories of twentieth century logic was Godel's discovery that there
are logics in which not all inferences that are valid in their semantic system can be proved using a
consistent formal system (see, e.g. Boolos & Jeffrey 1989). (The reason for the stipulation that the
system
is consistent is that an inconsistent system would allow any proposition including contradictions to be
proved.) The logic of sentential connectives, however, has the happy property that all inferences that
are valid on the basis of their truth-functional meanings are also provable in a consistent formal
system, and vice versa.
2. The interpretation of connectives in natural language
The psychology of reasoning would be simpler if all connectives in natural language were truth
functional. But, temporal connectives, such as "and then" or "before," are not truth functional. It is true
that Bush declared war on terrorism, and that terrorists attacked the USA, but the following assertion is
nevertheless false: Bush declared war on terrorism and then terrorists attacked the USA. The two
events occurred in the opposite order. In fact, the human interpretative system cannot be truth
functional, not even in the case of logical interpretations. As the example in the previous section
showed, a truth-functional interpretation starts and ends with truth values. It doesn't take into account
what individual atoms mean, what they refer to, or any temporal, spatial, or other such relation
between them: All it depends on are truth values. When you understand a sentence, however, you don't
end up with its truth value. Indeed, you may never know its truth value, which depends on the relation
between what it signifies and the state of the world. Comprehension starts with the construction of the
meaning of a sentence; it recovers its referents, their properties, and any relations among them—a
process that may depend on knowledge; and it ends with a representation of the possible situations to
which the sentence refers. In short, it starts with meanings and ends with models. The moral is clear.
No connectives in natural language are interpreted in a truth functional way (see Johnson-Laird &
Byrne 2002, Byrne 2005). Many uses of "if," "or," and "and" don't have a logical meaning. The
connective and can be interpreted to mean and then. The following disjunction: They played soccer
or they played some game seems innocuous. But, if you learn that the second atom is false, i.e.: They
didn't play any game you would not infer the truth of the first atom:
They played soccer. The formal rule I presented earlier would allow this inference to be made, but in
real life you wouldn't make it. You know that soccer is a game, and so you interpret the disjunction to
be compatible with only two possibili-

ties, and in both of them they played a game (see Johnson-Laird & Byrne 2002 for an account of such
"modulations" of interpretation). Hence, the disjunction no longer has a logical meaning.
3. The limitations of working memory
A mental model represents a possibility, or, to be precise, the structure and content of the model
capture what is common to the different ways in which the possibility could occur—a construal that I
owe to a logician, the late Jon Barwise (1993). When you are forced to try to hold in mind several
models of possibilities, the task is difficult. To experience this phenomenon of "memory overload" for
yourself, try the following problem:

June is in Wales or Charles is in Scotland, or both. Charles is in Scotland or Kate is in Ireland, or
both. What, if anything, follows? The disjunctions are inclusive, and so each premise is consistent
with three possibilities. The problem, of course, is to combine the two sets of possibilities. In fact,
they yield five possibilities, which support the valid conclusion: June is in Wales and Kate is in
Ireland, or Charles is in Scotland, or both. Five possibilities are too many to hold in mind at the same
time, and so, as the theory predicts, this inference is hard. My colleagues and I tested a sample of the
general population in an experiment, and only 6% of them drew a valid conclusion (Johnson-Laird et
al. 1992). The experiment also examined similar inferences based on exclusive disjunctions:
June is in Wales or Charles is in Scotland, but not both. Charles is in Scotland or Kate is in Ireland,
but not both. What, if anything, follows? These premises are compatible with only two possibilities,
and they yield the conclusion: Either June is in Wales and Kate is in Ireland or else Charles is in
Scotland. The problem was easier: 21% of the participants drew this conclusion or an equivalent to it.
Most people go wrong with both sorts of inference, and so you might wonder what conclusions they
draw. If they are trying to construct mental models of the various possibilities, then there are two
obvious predictions. The first is that if they grasp that there's more than one possibility but are unable
to discern what holds over all of them, then they should respond that there's no valid conclusion.
About a third of responses were of this sort. The second prediction is that if people overlook one or
more
possibilities, then their conclusions should correspond to only some of the possibilities compatible
with the premises. In fact, nearly all of the participants' erroneous conclusions were of this sort.
Indeed, the most frequent errors were conclusions based on just a single possibility compatible with the
premises. These errors cannot be attributed to blind guessing, because of the improbability of
guessing so many conclusions compatible with the premises. People prefer to reason on the basis of a
single model. Their erroneous conclusions are so hard to explain if they are relying on formal rules
that no-one has so far devised such an explanation (pace Rips 1994, Braine & O'Brien 1998). A simple
way in which to prevent reasoners from being swamped by possibilities is to give them an extra
premise that establishes the definite whereabouts of one of the persons, e.g.: June is in England. June
is in Wales or Charles is in Scotland, but not both. Charles is in Scotland or Kate is in Ireland, but not
both. What should then happen is that the interpretation of the first two premises yields only a single
possibility: June is in England Charles is in Scotland The combination of this possibility with those for
the third premise yields: June is in England Charles is in Scotland Kate is not in Ireland In this way,
the number of possibilities that have to be kept in mind at any one time is reduced to one. The
experiment included some problems of this sort, and they were easy. Diagrams can also improve
performance with disjunctive problems, but not just any diagrams. They need to make the task of
envisaging alternative possibilities easier (see Bauer & Johnson- Laird 1993). Your working memory
has a limited ability to hold models in mind. A superhuman intelligence, however, wouldn't be limited
in this way. Its working memory would not be a bottleneck, and so it could reason with much more
complex premises than you can. You don't realize your limitations because your social world is no
more complicated than your ability to think about it—it couldn't be—and your reasoning about the
physical world is good enough for you to survive.

4. The principle of parsimony
The model theory postulates that mental models are parsimonious. They represent what is possible,
but not what is impossible, according to assertions. This principle of parsimony minimizes the load
on working memory, and so it applies unless something exceptional occurs to overrule it. It
was introduced in Johnson-Laird & Savary (1999), who referred to it as the principle of "truth." This
name is slightly misleading, and so I have changed it here. Some critics have thought that the principle
means that mental models represent only those clauses mentioned in the premises. Such a view,
however, would imply wrongly that sentences have the same models regardless of the connectives that
occur in them. The principle of parsimony is subtle because it applies at two levels. At the first level,
mental models represent only what is possible. Consider, for example, how they represent the
exclusive disjunction: There is a circle or else there is a triangle but not both. Its mental models
represent the two possibilities: O A where "O" denotes a model of the circle, "A" denotes a model of
the triangle, and each horizontal line denotes a model of a separate possibility. Hence, the first row in
this diagram represents the possibility described in the first clause in the sentence, and the second row
represents the possibility described in the second clause. You will notice that two models of
possibilities are more parsimonious than the four rows of a truth table, which represent both what is
possible and what is impossible according to the premises. The second level at which the principle of
parsimony applies concerns individual models of possibilities: A mental model of a possibility repre-
sents a clause in the premises, whether it is affirmative or negative, only when the clause holds in that
possibility. This principle is exemplified in the mental models of the disjunction above. The first model
represents the possibility of a circle, but not the concurrent impossibility of a triangle. It contains no
explicit information about the triangle. Likewise, the second model represents the possibility of a
triangle, but not the concurrent impossibility of a circle. It contains no explicit information about the
circle. If you ask people to list what is possible given the preceding exclusive disjunction, they do
indeed list a circle as one possibility, and a triangle as another possibility, and they say nothing about
the status of the triangle in the first case or the status of the circle in the second case (Johnson-Laird &
Savary 1999). Yet, they have not entirely forgotten what is impossible in a possibility that they
represent. It is as though they made a mental footnote about it. But, the footnote is soon forgotten if
they have to carry out a taxing piece of reasoning or if sentences contain several connectives. Let's
consider a different sentential connective, the conditional, which joins together two clauses using "if"
and "then." Consider the assertion: If there is a circle then there is a triangle. You might ask: "And if
there isn't circle, what then?" The answer is that there may or may not be a triangle. The conditional in
its logical sense

is therefore compatible with three possibilities, which as usual I show on separate lines:
O A -.O A -IO -A where "-i" denotes negation. From adolescence or earlier, children list these
possibilities, as do adults, when they are asked what is possible given a conditional (see, e.g.
Barrouillet & Legas 1999, Barrouillet et al. 2000). However, because it's difficult to hold them all in
mind, when individuals reason from a conditional, they focus on the possibility in which both the "if"
clause, the antecedent, and the "then" clause, the consequent, occur. And so they construct the mental
model: O A But, if they were to construct only this model, then they would have represented a
conjunction: There is a circle and there is a triangle. They realize that the antecedent needn't occur:
There needn't be a circle. But, they de- fer the construction of an explicit model of this possibility.
They construct only a model that has no explicit content. It acts as a "place holder" to remind them that
there are other possibilities. The mental models of the conditional are accordingly: O A
where the ellipsis denotes the implicit model. Individuals should make a mental footnote that the
possibilities represented in the implicit model are those in which the antecedent doesn't occur, i.e.,
there isn't a circle. If they retain this footnote, then they can flesh out their mental models into fully
explicit models of the three possibilities. Now, you can understand why there is a difference in
difficulty between the two conditional inferences with which I began the chapter. The easy inference
follows at once from the mental models of the conditional, whereas the difficult inference does not.
One way to make the difficult inference is to flesh out the mental models into fully explicit models;
another way, which I will describe presently, is to make a supposition. Just as there are two sorts of
logical disjunction, inclusive and exclusive, so there are two sorts of logical conditional. You may
have understood the conditional above to mean that if, and only if, there's a circle then there's a
triangle. This interpretation is known as a biconditional, because it is equivalent to the assertion of
two conditionals: If there is a circle then there is a triangle, and if there isn't a circle then there isn't a
triangle. The biconditional is compatible with only two possibilities: O A

^O -. A
But, it has the same mental models as the regular conditional, except that the footnote states that the implicit
model represents the possibility in which both clauses fail to hold. If you retain the footnote, then you should
be able to flesh out your mental models into fully explicit models of the two possibilities. One reason that you
will try to do so is if you are unable to draw a conclusion from your mental models.
Table 2 summarizes the mental models and the fully explicit models of sentences based on the logical
meanings of the five principal sentential connectives. The ellipses represent implicit models, which serve as
place holders representing other possibilities that as yet have no explicit content and that are constrained by
mental footnotes. The fully explicit models flesh out mental models to represent all the clauses in the premises
in all the possibilities.
5. Truth tables versus models
You should now understand the difference between truth tables and models. Truth tables represent truth
values. Models represent possibilities. For example, the conjunction:
There is not a circle and there is a triangle is represented by a truth table with four rows, which represents
whether the atomic propositions are true or false. The only row for the conjunction that is true states in effect:
It is false that there is a circle and it is true that there is a triangle. In contrast, the conjunction has a single
mental model of a possibility:
-•O A Truth values are not possibilities, and the distinction matters in logic. When individuals refer to what is
"true" or "false," or mentally represent these terms, they are at risk of paradox, as in the famous example from
twentieth century logic:
This sentence is false. If this sentence is true then it is false; if it is false then it is true. Of course, the sentence
seems silly because it has no topic other than itself. Yet, logicians go to any lengths to avoid such paradoxes,
because they are a symptom of an inconsistent system (see, e.g. Barwise & Etchemendy 1987). No risk of
paradox occurs in referring to possibilities, e.g.:
This sentence is impossible. The sentence merely makes a false claim about the grammar of English: "true"
and "false" refer to the truth values of sentences, but "impossible" does not.
Table 2 The mental models and the fully explicit models for five sentential connectives
Connectives Mental models Fully Explicit models
Conjunction:
A and B: A B A B
Exclusive disjunction: A or B but not both: A A -> B B -.A B
Inclusive disjunction: A or B or both: AA
A
-.AA
-.BBB
Conditional:
If A then B:
Biconditional:
If and only if A then B:
BB
ABBB
-iBA
B -.B
Key: "-i" symbolizes negation, and "..." a wholly implicit model.
The difference between truth values and possibilities matters in psychology, because individuals
respond differently to questions about truth and falsity than to questions about possibility and
impossibility. For example, they tend to think that conditionals are true only in the case that both their
clauses are true, but they are happy to list as possible all three cases in Table 2, corresponding to fully
explicit models. Judgments of truth and falsity call for relating mental models to external possibilities
in order to derive truth values. When individuals list possibilities, however, they have only to
understand a sentence, and so they can flesh out their mental models into the three fully explicit
models of a conditional.
A
-.A
-•AA -.A
6. Mechanisms of model building
The model theory postulates that humans have a natural disposition to think of possibilities.
Alternative possibilities are represented as disjunctions of possibilities; and each model of a
possibility represents a conjunction of affirmative and negative propositions. The theory as it applies
to logical connectives therefore takes negation, conjunction, and inclusive disjunction, as
fundamental. In this second part of the chapter, I am going to describe the mechanisms that construct
models. These mechanisms have all been implemented in a computer program, and the program yields
a surprising consequence, which I'll get to by and by. But, I begin with negation, and then proceed to
connectives. Here is a problem that turns out to be harder than it seems at first sight (see Barres &
Johnson-Laird 2003). List the possibilities given the following assertion: It is not the case both that
there is a circle and that there is a triangle. Why isn't the task trivial? The answer is that you don't know
the answer, and so you have to infer it. You first have to work out what the unnegated sentence means:
There is a circle and there is a triangle. It allows just one possibility: O A The negative sentence rules
out this possibility to leave its complement, which is all the other possible models based on the same
two atoms and their negations. The first one that you're likely to think of is the mirror image of the
preceding possibility: -.O - A Some individuals go no further, but you will realize that there are two
other possibilities, in which one or other of the two shapes is missing: -.O A O - A In general, the way
to infer the correct interpretation of a negative sentence is to take its atoms, and to work out all the
possible combinations of them and their negations. You remove from these combinations those that are
compatible with the unnegated sentence, and what remains is the answer: the possibilities compatible
with the negative sentence. No wonder that people do not cope with the negation of compound
sentences well. They tend to be better at negating a disjunction than a conjunction, perhaps because the
former yields fewer models than the latter. Individuals represent a set of alternative possibilities as a
list of alternative models. Such a list corresponds to an inclusive disjunction. To combine two such
sets of models according to any logical relation between them, calls only for negation, which I've
described, and logical conjunction, which I'm

about to describe. When individuals interpret a set of premises, however, they construct a model of an
initial clause or premise, and then update this model from the remaining information in the premises.
Let's consider a pair of premises that illustrate the main principles of conjunction: If there is a triangle
then there is a diamond. There is a circle or else there is a triangle but not both. Before I tell you what
the resulting models are, you might like to think for yourself what possibilities are compatible with the
two premises. Most people think that there are two: a triangle and a diamond, or a circle. The mental
models of the first premise are: A 0

The core of the interpretative process is to update these models by forming a conjunction of them
with the models of the second premise. One possibility according to the second premise is that there
is a circle, and so the system conjoins: A 0 and O The triangle in the first model here occurs elsewhere
in the models containing the circle, and so the interpretative system takes the absence of the triangle
from the model containing the circle to mean that there is not a triangle. In effect, the conjunction
becomes: A 0 and O -• A Because there is now a contradiction—one model contains a triangle and the
other its negation—the result is a special null model (akin to the empty set), which represents
propositions that are contradictory. It represents what is impossible. The conjunction therefore yields
the null model: nil The system now conjoins the pair: A 0 and A The diamond doesn't occur elsewhere
in the set of models containing the model of the triangle alone, and so the two models are compatible
with one another. Their conjunction yields:
A 0 Similarly, the conjunction: and O yields O because the circle doesn't occur in the models
containing the implicit model. The final conjunction: and A yields nil because the triangle does occur
elsewhere in the models containing the implicit model, and so its absence in the implicit model is
treated as akin to its negation. The mental models of the conjunction of the premises are accordingly:
Table 3 The mechanisms for conjoining pairs of mental models and pairs of fully explicit models
1. If one model contains a representation of a proposition, A, which is not represented in the other model, then consider the
set of models of which this other model is a member. If A occurs in at least one of these models, then its absence in the
current model is treated as its negation (go to mechanism 2); otherwise its absence is treated as its affirmation (go to
mechanism 3). This mechanism applies only to mental models. 2. The conjunction of a pair of models containing
respectively a proposition and its
negation yield the null model, e.g.: A B and -iA B yield nil. 3. The conjunction of a pair of models that are not contradictory
yields a model
containing all the elements of both models, e.g.:
A B and B C yield A B C . 4. The conjunction of a null model with any model yields the null model, e.g.:
A B and nil yield nil.
A 0 O I have not shown the null models, because they do not represent possibilities. The two
models of possibilities yield the valid conclusion: There is a triangle and a diamond, or else there is a
circle. Table 3 summarizes the mechanisms for forming conjunctions of pairs of models. The same
mechanisms apply to the conjunction of fully explicit models. Here are the previous premises again: If
there is a triangle then there is a diamond. There is a circle or else there is a triangle but not both. Their
mental models can be fleshed out to be fully explicit by a mechanism that uses mental footnotes, but
I'll spare you the details. The fully explicit models of the conditional (see Table 1) are: A 0 -.A 0 -.A
-.0 Because the disjunction has two models, there are six pair-wise conjunctions, but three of them
are contradictions yielding the null model. The remaining pairs yield the following results: A 0 -.O - A
0O
-.A -.<> O The same conclusion follows as before:
There is a triangle and a diamond or else there is a circle.
But, reasoners who rely on mental models will fail to think about the second of these possibilities. They should
think that it is impossible to have the diamond and the circle. This prediction is typical of the model theory.
You can make suppositions when you reason, i.e., assumptions for the sake of argument (see, e.g. Byrne et al.
1995). Given a disjunction, such as:
There is a triangle on the board or there is a circle, or both, you can make the supposition that there isn't a
triangle on the board, and then infer as a consequence that in that case there is a circle on the board. You hold
in mind a possibility, which in this case corresponds to the negation of an atom in the premise, and then treat it
as though it was asserted categorically. You can then use the inferential mechanisms that I have already
described. If you are prudent, you remember that any conclusion depends on a supposition, and take this fact
into account in formulating a final conclusion. If a supposition leads to a contradiction (the null model), some
individuals appreciate that the supposition is impossible granted the truth of the premises. The procedure is
identical to the one that occurs in the construction of models of the following sort of conditional:
If A then both B and not B. The conjunction, B and not B, yields the null model. The interpretation of the
conditional calls for the conjunction of A and nil, which yields nil (see Table 3). What happens then depends
on whether individuals are relying on mental models or fully explicit models. With mental models, there
remains only the implicit model, which yields no conclusion. But, the fully explicit models of the conditional
are:
A nil -> A nil -• A -i nil The negation of nil in the third model yields the disjunction of the atoms that led to its
construction, and so this conjunction yields the conclusion:
not A. The corresponding principle in logic is known as reductio ad absurdum. In the model theory, it is a
consequence of a mechanism that makes suppositions, and of reasoning from fully explicit models.
In a review of theories of conditionals, Evans & Over (2004) claimed that the model theory makes no use of
suppositions, despite our several papers to the contrary (e.g. Byrne et al. 1995). They also argued that the
model theory is truth functional, despite the arguments that I have summarized above. Their review is
otherwise valuable. It is a pity that they mangle the model theory so badly, because it makes sense of
phenomena that are otherwise puzzling for them, e.g., the difference that I described earlier between judgments
of truth value and the listing of possibilities.
7. Superhuman reasoning
A computer program that I wrote to simulate the model theory can make inferences that are far
beyond the ability of human reasoners working without benefit of logic. Only a superhuman
intelligence, such as Hercule Poirot (Agatha Christie's famous fictional detective), could solve the
following problem without paper and pencil: Who helped to murder Mr. Ratchett on the Orient
Express? If Pierre helped if Dr. Constantine did then Greta helped too. If not both Harriet and Hector
helped then Dr. Constantine didn't help. Greta didn't help or Dr. Constantine did. Harriet didn't help or
the Princess Drago-Miroff, Mary, and Colonel Arbuthnot all helped. If Hector or Mary helped then
Pierre helped or Colonel Arbuthnot didn't help. So, who helped, who didn't help, and for whom is it
impossible to say? There are eight atomic propositions in the premises, and so their truth table has
256 rows. Likewise, there are multiple models, but if you build them up premise by premise, the final
result is a single model. It shows that all eight individuals helped to commit the murder. In many
other inferences, of course, the premises yield multiple models, but an algorithm exists for drawing
parsimonious conclusions that describe them (see Johnson-Laird & Byrne 1991, Ch. 9).
8. Some illustrative inferences
To illustrate the model theory and its predictions, I am going to consider some inferences that human
reasoners can make. The first inference is: Either Jane is kneeling by the fire and she is looking at the
TV or else Mark is standing at the window and he is peering into the garden. Jane is kneeling by the
fire. Does it follow that she is looking at the TV? Most people say: "yes" (Walsh & Johnson-Laird
2004). A second inference has the same initial premise, but it is followed instead by the categorical
denial:
Jane is not kneeling by the fire, and the question is: Does it follow that Mark is standing at the
window?
Again, most individuals say: "yes". Let's see what the theory predicts. The first premise in both
inferences is the same exclusive disjunction of two conjunctions. The theory predicts that individuals
should rely on mental models. Hence, they should interpret the first conjunction, Jane is kneeling by
the fire and she is looking at the TV, and build a model representing this possibility, which I will
abbreviate as follows: Jane: kneeling looking They should build an analogous model of the second
conjunction: Mark: standing peering These two models must now be combined according to an
exclusive disjunction. An exclusive disjunction has two mental models, which represent the two
conjunctions only in the possibilities in which they hold:
Jane: kneeling looking Mark: standing peering For the first inference, the conjunction of the
categorical premise: Jane is kneeling with the first model of the disjunction yields: Jane: kneeling
looking Its conjunction with the second model of the disjunction yields the null model. Hence, the
premises yield only the model:

Jane: kneeling looking and so individuals should respond: yes, Jane is looking at the TV. This
analysis may strike you as obvious. In fact, the inference is a fallacy. The principle of parsimony
postulates that individuals normally represent what is possible, but not what is impossible. When I
first wrote the computer program to simulate the theory, and inspected its output for a certain problem,
I thought that there was a bug in the program. I searched for the bug for half a day, before I realized
that the program was correct, and the error was in my thinking. What the program revealed is the
discrepancy between mental models and fully explicit models. The theory therefore predicted that
individuals should reason in a fallacious way for certain inferences. Indeed, the fallacies turn out to
be so compelling in some cases that they resemble cognitive illusions, and so my colleagues and I refer
to them as "illusory" inferences. If you succumbed to the illusion, then you are in the company of
Clare Walsh and myself. We studied these inferences, but it took us a couple of days to realize that
they were illusory, and that was after the discovery of other sorts of illusions. The fully explicit models
of the exclusive disjunction reveal the correct conclusion:
Jane: Jane: Jane: Jane:
kneeling kneeling kneeling -i kneeling

looking looking looking -i looking
Mark: Mark: Mark: Mark:
-i -i standing standing standing standing
-i peering peering -i peering peering

Jane: -> kneeling looking Mark: standing peering Jane: kneeling -i looking Mark: standing peering
When one conjunction is true, the other conjunction is false, and you will remember from my earlier
account that there are three ways in which a conjunction can be false. The categorical premise that
Jane is kneeling rules out the fourth and fifth possibilities. But, contrary to the illusory inference, it
leaves one possibility—the sixth one—in which Jane is kneeling but not looking at the TV. That is
why the illusory inference is invalid. Granted that Jane is kneeling, it does not follow that she is
looking at the TV. The second problem that I described has the categorical premise that Jane is not
kneeling by the fire, and poses the question of whether it follows that Mark is standing by the window.
Most people respond, "yes", which is a conclusion supported by the mental models shown above. The
fully explicit models show that this inference is valid. The categorical premise eliminates all but the
fourth and fifth models, and in both of them Mark is standing by the window. Our main experiment
examined a series of illusory inferences and control problems of this sort. The participants were much
more likely to respond correctly to the control problems (78% correct) than to the illusory problems
(10% correct): 34 of the 35 participants showed this difference. Illusory inferences occur in many
domains, including reasoning with quantifiers (Yang & Johnson-Laird 2000a, 6), deontic reasoning
(Bucciarelli & Johnson-Laird 2005), and assessing whether or not sets of assertions are consistent
(Johnson-Laird et al. 2004). I will describe two more examples. The first example (from Goldvarg &
Johnson-Laird 2000) calls for reasoning about what is possible: Only one of the following premises
is true about a particular hand of cards: There is a king in the hand or there is an ace, or both. There is
a queen in the hand or there is an ace, or both. There is a jack in the hand or there is a 10, or both. Is it
possible that there is an ace in the hand? The model theory postulates that individuals consider the
possibilities for each of the three premises. That is, they assume that the first premise is the one that is
true, and consider the consequences; then they assume that the second premise is the one that is true
and consider the consequences, and then they assume that the third premise is the one that is true and
consider the consequences. However, because the question asks only whether an ace is possible, they
can stop as soon as they find a premise that allows the presence of the ace in the hand. What is wrong
with this procedure? The answer is that when individuals consider the truth of one premise, they should
also consider the concurrent falsity of the other two premises. But, that is exactly what the principle of
parsimony predicts they will not do.

For the first premise, they accordingly consider three models, which each correspond to a possibility
given the truth of the premise: king ace king ace Two of the models show that an ace is possible.
Hence, on the basis of this premise alone individuals should respond, "yes." The second premise
supports the same conclusion. The third premise is compatible with it. In fact, it is an illusion of
possibility: reasoners infer wrongly that a card is possible. If there were an ace, then two of the
premises would be true, contrary to the rubric that only one of them is true. The same strategy,
however, yields a correct response to a control problem in which only one premise refers to an ace. A
problem to which reasoners should respond "no," and thereby succumb to an illusion of impossibility,
can be created by replacing the two occurrences of "there is an ace" in the premises above with,
"there is not an ace." Its control problem contains only one premise with the clause, "there is not an
ace." Figure 1 presents the results of an experiment in which we gave students 16 inferences, four of
each of the four sorts. Half of the illusions were based on disjunctive premises, and half were based on
conditionals. The participants' confidence in their conclusions did not differ reliably from one sort of
problem to another. As the Figure shows, they were very susceptible to the illusions but performed
well with the control problems, and the illusions of possibility were more telling than those of
impossibility. To infer that a situation is impossible calls for a check of every model, whereas to infer
that a situation is possible does not, and so reasoners are less likely to make the inference of
impossibility. This difference also occurs in problems that are not illusory (Bell & Johnson-Laird
1998). With hindsight, it is surprising that nearly everyone responded "yes" to the first of the problems
above, because it seems obvious that an ace renders two of the premises true. We therefore carried out
a replication with two groups of participants, and half way through the experiment, we told one group
to check whether their conclusions met the constraint that only one of the premises was true. This
procedure had the advantage that the participants did not have to envisage the circumstances in which
the premises did not hold. The group that received the special instruction was thereafter much less
likely to commit the fallacies (Goldvarg & Johnson- Laird 2000). If the preceding illusions result from
a failure to reason about what is false, then any manipulation that emphasizes falsity should reduce
them. The rubric, "Only one of the following two premises is false" did reduce their occurrence
(Tabossi et al. 1998), as did the prior production of false instances of the premises (Newsome &
Johnson-Laird 1996).
1OOn
Inferences of possibility
Inferences of impossibility
Fig. 1. The percentages of correct responses to fallacious inferences that are illusory and their control problems (based on
Goldvarg and Johnson-Laird, 2000)
The second example of an illusion is very compelling. The rubric, "one of these assertions is true and
one of them is false," is equivalent to an exclusive disjunction between two assertions. Consider this
problem, which is based on an exclusive disjunction: Suppose you know the following about a
particular hand of cards: If there is a jack in the hand then there is a king in the hand, or else if there
isn't a jack in the hand then there is a king in the hand.
There is a jack in the hand. What, if anything, follows? Nearly everyone—experts and novices alike
—infers that there is a king in the hand (Johnson-Laird & Savary 1999). Yet, it is a fallacy granted a
disjunction, exclusive or inclusive, between the two conditional assertions. The disjunction entails that
one or other of the two conditionals could be
false; and if one of them is false, then there may not be a king in the hand. Suppose, for instance, that
the first conditional is false. There could then be a jack but not a king—a judgment with which most
individuals concur (see, e.g. Oaksford & Stenning 1992). And so the inference that there is a king is
invalid: the conclusion could be false. An experiment examined the preceding problem and another
illusion, and compared them with two control problems in which the neglect of false cases should not
impair performance (Johnson-Laird & Savary 1999). The participants committed both fallacies in 100
percent of cases, and yet drew valid inferences for the control problems in 94 percent of cases. The
participants were again confident in both their illusory conclusions and their correct control
conclusions. Because so many expert psychologists have succumbed to illusory inferences, we have
accumulated many putative explanations for them. For example, the premises may be so complex,
ambiguous, or odd, that they confuse people, who, as a result, commit a fallacy. This hypothesis over-
looks the fact that the participants are very confident in their conclusions, and that the control
problems are equally complex. Likewise, when the illusions and controls are based on the same
premises, but different questions in the form of conjunctions, participants still commit the fallacies and
get the control problems correct (Goldvarg & Johnson-Laird 2000). Other putative explanations
concern the interpretation of conditionals. Individuals make the illusory inference with problems of
this sort: One of the following assertions is true and one of them is false: If there is a jack then there is
a king. If there isn't a jack then there is a king. This assertion is definitely true: There is a jack. Naive
individuals understand that a conditional, such as: If there is jack then there is a king. is false in the
case that there is jack but not a king. They also understand that the rubric to this problems mean that
one conditional is true and the other conditional is false. Hence, on their own account they should
refrain from inferring that there is a king. The analysis depends on nothing else. However, even if
some special factors exacerbate the illusions with conditionals, other illusions occur with problems
that do not contain conditionals, such as the problem with which I started this section of the chapter.
Many other robust phenomena in reasoning appear to arise from the principle of parsimony and the
resulting neglect of what is impossible or false. They include the results of Wason's "selection" task in
which individuals fail to grasp the relevance of an instance of a false consequent to testing the truth
or falsity of a conditional (see, e.g. Wason 1966, Wason & Johnson-Laird 1972).
9. Conclusions
This chapter has explained the mechanisms that construct models based on the logical interpretation
of connectives. These models do not represent truth values, but sets of possibilities. Individuals adopt a
variety of strategies to cope with reasoning problems, e.g., they may be guided by a given conclusion,
they may work forwards from the premises, they may make a supposition, and so on (Van der Henst et
al. 2002, Johnson-Laird & Hasson 2003). But, regardless of strategy, inferences that depend on a single
model are easier than those that depend on multiple models. Mental models abide by the principle of
parsimony: They represent only possibilities compatible with the premises, and they represent clauses
in the premises only when they hold in a possibility. Fully explicit models represent clauses when
they do not hold too. The advantage of mental models over fully explicit models is that they contain
less information, and so they are easier to work with. But they can lead reasoners astray. The
occurrence of these systematic and compelling fallacies is shocking. The model theory predicts them,
and they are a "litmus" test for mental models, because no other current theory predicts them. They
have so far resisted explanation by theories of reasoning based on formal rules of inference, because
these theories rely on valid rules. For several years, my former colleague Yingrui Yang has sought an
explanation based on a revised formal rule theory, but he has yet to succeed. To reason only about
what is possible is a sensible way to cope with limited processing capacity, but it does lead to illusions.
Yet, it does not imply that people are irredeemably irrational. The fallacies can be alleviated with
preventative methods. Otherwise, however, reasoners remain open to the illusion that they grasp what
is in fact beyond them.

References
Barres, P. & Johnson-Laird, P. (2003), 'On imagining what is true (and what is
false)', Thinking & Reasoning 9, 1-42. Barrouillet, P., Grosset, N. & Legas, J. F. (2000), 'Conditional reasoning
by mental models: chronometric and developmental evidence', Cognition 75, 237- 266. Barrouillet, P. &
Legas, J.-F. (1999), 'Mental models in conditional reasoning
and working memory', Thinking & Reasoning 5, 289-302. Barwise, J. (1993), 'Everyday reasoning and logical
inference', Behavioral and
Brain Sciences 16, 337-338. Barwise, J. & Etchemendy, J. (1987), The Liar: An Essay in Truth and Circu-
larity, Oxford University Press, New York.
Bauer, M. & Johnson-Laird, P. (1993), 'How diagrams can improve reasoning', Psychological Science 4, 372-
378. Bell, V. & Johnson-Laird, P. (1998), 'A model theory of modal reasoning', Cog-
nitive Science 22, 25-51. Berry, G. (1952), Peirce's contributions to the logic of statements and quantifiers, in
P. Wiener & F. Young, eds, 'Studies in the Philosophy of Charles S. Peirce', Harvard University Press,
Cambridge, MA. Boolos, G. &: Jeffrey, R. (1989), Computability and Logic, 3rd edn, Cambridge
University Press, Cambridge. Braine, M. & O'Brien, D., eds (1998), Mental Logic, Erlbaum, Mahwah, NJ.
Bucciarelli, M. & Johnson-Laird, P. (2005), 'Naive deontics: a theory of meaning,
representation, and reasoning', Cognitive Psychology 50, 159-193. Byrne, R. (2005), The Rational
Imagination: How People Create Alternative to
Reality, MIT Press, Cambridge, MA. Byrne, R., Handley, S. & Johnson-Laird, P. (1995), 'Reasoning from
suppositions',
Quarterly Journal of Experimental Psychology 48A, 915-944. Evans, J., Newstead, S. & Byrne, R. (1993),
Human Reasoning: The Psychology of Deduction, Erlbaum, Hillsdale, NJ. Evans, J. & Over, D. (2004), //,
Oxford University Press, Oxford. Garnham, A. (2001), Mental Models and the Representation of Anaphora,
Psych-
ology Press, Hove, East Sussex. Gentner, D. & Stevens, A., eds (1983), Mental Models, Erlbaum, Hillsdale,
NJ. Goldvarg, Y. & Johnson-Laird, P. (2000), 'Illusions in modal reasoning', Memory & Cognition 28, 282-
294. Jeffrey, R. (1981), Formal Logic: Its Scope and Limits, 2nd edn, McGraw-Hill,
New York. Johnson-Laird, P. (1983), Mental Models: Towards a Cognitive Science of Lan- guage, Inference,
and Consciousness, Cambridge University Press/Harvard University Press, Cambridge/Cambridge, MA.

Johnson-Laird, P. & Byrne, R. (1991), Deduction, Erlbaum, Hillsdale, NJ. Johnson-Laird, P. & Byrne, R.
(2002), 'Conditionals: A theory of meaning, prag-
matics, and inference', Psychological Review 109, 646-678. Johnson-Laird, P., Byrne, R. & Schaeken, W.
(1992), 'Propositional reasoning by model', Psychological Review 99, 418-439. Johnson-Laird, P., Girotto, V.
& Legrenzi, P. (2004), 'Reasoning from inconsis-
tency to consistency', Psychological Review 111, 640-661. Johnson-Laird, P. & Hasson, U. (2003),
'Counterexamples in sentential reason-
ing', Memory & Cognition 31, 1105-1113. Johnson-Laird, P., Legrenzi, P., Girotto, V., Legrenzi, M. &
Caverni, J.-P. (1999), 'Naive probability: a mental model theory of extensional reasoning', Psycho-
logical Review 106, 62-88. Johnson-Laird, P. & Savary, F. (1999), 'Illusory inferences: A novel class of erro-
neous deductions', Cognition 71, 191-229. Marr, D. (1982), Vision: A Computational Investigation into the
Human Repre- sentation and Processing of Visual Information, W.H. Freeman, San Francisco. Medvedev, Z.
A. (1990), The Legacy of Chernobyl, W.W. Norton, New York.
Metzler, J. & Shepard, R. (1982), Transformational studies of the internal representations of three-
dimensional objects, in R. Shepard & L. Cooper, eds, 'Men- tal Images and Their Transformations', MIT Press,
Cambridge, MA, pp. 25-71. Newsome, M. & Johnson-Laird, P. (1996), An antidote to illusory inferences, in
'Proceedings of the Eighteenth Annual Conference of the Cognitive Science Society', Erlbaum, Mahwah, NJ,
p. 820. Oakhill, J. & Garnham, A., eds (1996), Mental Models in Cognitive Science,
Psychology Press, Hove, Sussex. Oaksford, M. & Stenning, K. (1992), 'Reasoning with conditionals
containing
negated constituents', Journal of Experimental Psychology: Learning, Memory, and Cognition 18, 835-854.
Polk, T. & Newell, A. (1995), 'Deduction as verbal reasoning', Psychological Re- view 102, 533-566. Rips, L.
(1994), The Psychology of Proof, MIT Press, Cambridge, MA. Stevenson, R. (1993), Language, Thought and
Representation, Wiley, New York. Tabossi, P., Bell, V. & Johnson-Laird, P. (1998), Mental models in
deductive, modal, and probabilistic reasoning, in C. Habel & G. Rickheit, eds, 'Mental Models in Discourse
Processing and Reasoning', John Benjamins, Berlin. Van der Henst, J.-B., Yang, Y. & Johnson-Laird, P.
(2002), 'Strategies in senten-
tial reasoning', Cognitive Science 26, 425-468. Walsh, C. & Johnson-Laird, P. (2004), 'Co-reference and
reasoning', Memory &
Cognition 32, 96-106. Wason, P. (1966), Reasoning, in B. Foss, ed., 'New Horizons in Psychology',
Penguin, Harmondsworth, Middx. Wason, P. & Johnson-Laird, P. (1972), The Psychology of Deduction:
Structure and Content, Harvard University Press/Batsford, Cambridge, MA/London. Wittgenstein, L. (1922),
Tractatus Logico-Phiiosophicus, Routledge & Kegan
Paul, London. Yang, Y. & Johnson-Laird, P. (2000 a), 'How to eliminate illusions in quantified reasoning',
Memory & Cognition 28, 1050-1059. Yang, Y. & Johnson-Laird, P. (20006), 'Illusory inferences with
quantified assertions: How to make the impossible seem possible, and vice versa', Memory & Cognition 28,
452-465.
Mental Models and the Mind 53 Carsten Held, Markus Knauflf, Gottfried Vosgerau © 2006 Elsevier B.V. All rights
reserved
Interaction of Knowledge and Working Memory in Reasoning

About Relations
Andre Vandierendonck, Vicky Dierckx & Hannelore Van der Beken
Department of Experimental Psychology, Ghent University*
Abstract This chapter defends the thesis that relational reasoning is based on an interaction of knowledge
represented in semantic and episodic long-term memory on the one hand and temporary information
maintained in working memory on the other hand. Three lines of evidence relevant to this interaction are
considered. First, it is shown that reasoners seem to have a preference for a visuo-spatial representation of the
temporary mental models when this representational format is advantageous because there is a need to
represent structural information or because the relations are easily represented in this format. The second line
of evidence shows that apart from believability of the premises and believability of the conclusions, also
believability of the model(s) described by the premises plays a role. In a third part, transitive (order relations)
and intransitive relational problems (genealogical relations) are compared in order to clarify the role of
constraints on the relational representations on reasoning. The results obtained in these three lines of
investigation support the view that (relational) reasoning is based on a tight interplay of knowledge
representations in long-term memory and temporary models in working memory. Some discussion is devoted
to the possibility that this interaction is mediated by Baddeley's concept of episodic buffer.
E-mail: [email protected]
54 A. Vandierendonck, V. Dierckx & H. Van der Beken
1. Introduction
The present chapter develops the thesis that deductive reasoning about relations involves an interplay
of knowledge available in semantic and episodic long-term memory on the one hand and temporary
information maintained in working memory on the other hand. First, we specify the nature of
relational reasoning and show how the mental models theory of reasoning (Johnson-Laird 1983)
applies to this kind of reasoning. We also address the question how working memory and its different
components support relational reasoning. Next, attention will be focused on three specific ways in
which knowledge may affect reasoning performance. A first theme addresses the visuo-spatial basis of
relational reasoning. The second theme focuses on how activation of relevant background knowledge
available in cognitive schemata such as scripts may help or impair reasoning about temporal rela-
tions. Finally, we discuss reasoning about family relations. These constitute a set of relations which
have a rather rich semantic content imposing many constraints on the inferences that can be made.
1.1. RELATIONAL REASONING
From the information that Adam was born before Bob and that Bob was born before Carol, we can
easily infer that Adam was born before Carol. In this example, an inference must be made about the
relation between Adam and Carol and this is quite straightforward since "born before" implies a
temporal order relation which is transitive. Most of the research on relational reasoning has used such
transitive problems. It is not always the case, however, that the premises support a valid conclusion, as
can be seen in the following example. From Dedre was born before Eric and Dedre was born before
Fay, it is not possible to decide whether Eric was born before Fay or vice versa. These are
indeterminate problems and are more difficult to solve. Research based on problems with three terms
as in the examples given (three-term series), has shown that people solve such problems by integrating
all the relations into a single spatial array either with the largest values at the top and the smallest at
the bottom or in a left-to-right orientation with the largest values to the right (e.g. De Soto et al. 1965,
Huttenlocher 1968, Potts 1974). Although an alternative view was proposed and defended by Clark
(1969, 1971), the evidence collected over the years shows that the spatial array view yields the best
summary and explanation of the findings (see e.g. chapter 6 in Evans et al. 1993). The relational
content does not seem to matter much and the view applies equally well for determinate as for
indeterminate problems based on transitive relations (Barclay 1973, Foos et al. 1976, Huttenlocher
1968,
Interaction of Knowledge and Working Memory 55
Maybery et al. 1986, Mynatt & Smith 1977, Potts 1974, 1976, Sternberg 1980, 1981).
1.2. MENTAL MODELS
A spatial array is nothing else than a mental representation in which the "objects" are given a place along a
spatial line. It is a (mental) model representing the information given in the premises. In this sense, the spatial
array view is a special case of the mental models theory developed by Johnson-Laird and colleagues (Goodwin
& Johnson-Laird 2005, Johnson- Laird 1983, Johnson-Laird & Byrne 1989). The main thesis of this theory is
that deductive reasoning is a meaning-based, rather than a rule-based, process. Basically, reasoning follows a
three-stage course consisting of an interpretation of the premises, the formulation of a tentative conclusion and
a search for counter-examples.
Stages in reasoning. The first stage yields an interpretation of the information given in the premises. In
addition to a comprehension of each of the statements, this requires an understanding of the relationship
expressed in each premise. The premise information is then integrated into one or more models, whereby each
model is a representation of a possible situation in the world, given the information in the premises. With the
premises Glenda is smarter than Howard and Glenda is smarter than Igor, an initial model could be Glenda-
Howard-Igor.
In the next phase, the models are used to generate a tentative conclusion. From the model constructed in the
last example, a tentative conclusion would be that Howard is smarter than Igor. Because the conclusion is only
tentative, it is necessary to check whether the conclusion really holds. This is achieved in a third phase, by
searching for counter-examples. In our example, the reasoner might wonder whether it is not possible that Igor
is smarter than Howard. This corresponds to another model, namely Glenda-Igor-Howard which appears to be
consistent with the premises also. However, this second model is not consistent with the tentative conclusion.
Since there are now two models representing the premises and the tentative conclusion is consistent with only
one of the models, this conclusion cannot be valid. Actually, the reasoner is now able to recognize that the
premises allow to construct two models that are contradictions of each other. As a consequence, with the
premises given in the last example, the reasoner can conclude that no valid conclusion can be obtained.
Implicit models. A basic tenet of the mental models theory is that in the first stage an initial model is
constructed and that room is left for further elaboration of the representations by including an implicit model.
Whilst this is an important and central thesis in the theory which has
found support in studies of syllogistic and conditional reasoning (see e.g., Johnson-Laird 1999,
Johnson-Laird & Byrne 1989), in research on relational reasoning, on the contrary, the assumption
that the reasoner immediately constructs all relevant models has been prevalent (e.g., Evans et al.
1993, Johnson-Laird & Byrne 1989). However, Vandierendonck et al. (2004) have shown that the
latter assumption is probably incorrect. They used four-term series problems in which the premises
could be consistent with one, two or three different models. The findings supported the hypothesis
that reasoners construct an integrated representation of all the premise information and that in multi-
model problems an annotation is made which is further fleshed out when the need arises. This
elaboration of the annotation does not result in the construction of more than one model, however;
instead, the reasoners seem to construct a forked array which has been labeled "isomeric model"
(Schaeken et al. in press). For example, the premises John is larger than Kate, Kate is larger than Lee,
and John is larger than Meg, are consistent with three different models, namely John-Kate-Lee-Meg,
John-Kate-Meg-Lee, and John-Meg- Kate-Lee. Instead of constructing or elaborating all these models,
the reasoners start with constructing an annotated model such as John- (Meg) -Ka- te-Lee and
eventually work this out into a construction like John- KatM~Lea, which indicates that Meg may occur at
any position after John. Working memory. According to the mental models theory, a model is
constructed to integrate information extracted from the premises as a possible representation of what
is given to be true.2 By definition, a model is a temporary construction, needed for the derivation of a
tentative conclusion and for supporting the search for counterexamples. Once these processes come to
a closure, the need for maintaining the model(s) no longer exists. The designated medium for the
maintenance of temporary supporting information, such as models, is working memory. Indeed,
another assumption of the mental models theory is that models are constructed in working memory
and since working memory has a limited capacity, the more models that have to be maintained
simultaneously in working memory, the more likely it is that information will be lost and that errors
will be made and the more time the reasoner will need to arrive at a conclusion. Within the different
conceptualizations of working memory (see e.g., Miyake & Shah 1999, for a broad overview), the
model of Baddeley & Hitch (1974) provides a useful framework for the study of the interaction of
different cognitive tasks with working memory (see also, Baddeley 1986, 2000, Baddeley & Logie
1999). The strength of the model is that it

2This formulation may be perceived as strange. Nevertheless, it is related to a central tenet of the mental models theory,
namely that the reasoners only construct models on the basis of information that is given as being true.
conceptualizes working memory as a multicomponential system in which a number of automatically
operating processes are monitored by an executive controller. Originally, two subsidiary systems were
postulated, one for the maintenance of phonologically coded materials, the phonological loop, and one
for the maintenance of visuo-spatially coded materials, the visuo-spatial sketch pad. According to
recent developments, both the phonological loop (Baddeley 1986) and the visuo-spatial sketch pad
(Logie 1995) are considered to consist of a system for the passive maintenance of modality-specific
information and a looping device for refreshing the information kept in the store. Recently, a third
subsidiary system has been proposed, the episodic buffer, which has the task of integrating information
from both other subsidiary components with the contents of episodic memory (Baddeley 2000). All
these subsidiary systems are supervised by a central executive which is similar to the supervisory
attentional system described by Norman & Shal- lice (1986). Viewed within the context of this
general framework, there is little doubt that reasoning calls on the executive controller, at least to the
extent that reasoning involves the evaluation of tentative conclusions and the search for
counterexamples (see Evans 2000, for a discussion of this issue). Indeed, several studies using a dual-
task methodology have shown that reasoning is impaired when it is performed concurrently with a
demanding secondary task (see e.g., Klauer et al. 1997, Meiser et al. 2001, Vandierendonck & De
Vooght 1997). In a similar vein, there is evidence that the other working memory components also
play a role in deductive reasoning in providing storage for the maintenance of the models. However, as
will be explained later on in this chapter, it seems to depend on several factors such as the kind of
reasoning task, the number of models to be constructed, the size of the models, etc. whether and to
what extent reasoning calls on the phonological and the visuo-spatial subsystems (see e.g., Duyck et al.
2003, Gilhooly et al. 1993, 1999, Klauer et al. 1997, Meiser et al. 2001, Toms et al. 1993). Long-term
Memory. The construction of mental models, being a meaning- driven process, also heavily relies on
long-term memory. Because reasoning is assumed to rely on a semantic analysis of the premises,
reasoning is a process based also on accumulated knowledge. In fact, long-term memory support may
intrude in at least four different steps of the reasoning process. 1. The premises must be comprehended.
Apart from a process of language comprehension, this also involves recollection of contexts that are
relevant to the meaning of the premise sentences. Available information may provide "additional
premises" so that particular inferences are facilitated while others may be suppressed (e.g., Byrne
1989). In a similar vein, knowledge and beliefs may provide models or conclusions that are difficult to
ignore and that do affect the rep-

resentations reasoning is based on (e.g., belief biases, Revlin et al. 1980). 2. In the construction of
models, the information from several premises must be integrated. Again, available knowledge may
play a role to succeed in this task. This is one of the issues that will be further developed in the present
chapter. 3. General knowledge will support the process of conclusion generation. Knowledge
provides possible conclusions. It has long been known, that in the absence of knowledge, as in
reasoning problems with symbolic entities, reasoning is less biased. The present chapter also presents
evidence that knowledge embedded in schemata or scripts may provide alternative interpretations of
premises in relational reasoning. 4. In the search for counterexamples, the availability of relevant
and useful information retrieved from long-term memory may again affect the outcome of the
reasoning process. If a conclusion contradicts knowledge, an alternative is readily available and may
contribute to the observation of belief biases (e.g., Byrne et al. 2000, Newstead et al. 1992).
1.3. COUPLING OF WORKING MEMORY AND LONG-TERM MEMORY
For a long time, the role of long-term memory in reasoning has been a nuisance to reasoning
theorists, because the rule-based views on reasoning have always had difficulty to explain how the
application of logical rules could be intruded by knowledge (e.g., Henle 1962). For a meaning-based
theory, as the mental models theory is, these effects are part and parcel of reasoning. Interestingly, as
reasoning is supposed to be mediated by working memory, such a view also implies that at some
point working memory and long-term memory should interact on the way to obtain a solution to a
reasoning problem. Even though, thus far, the mental models theory does not provide a computational
view on deductive reasoning, this particular interaction enables a first approach to the specification of
a processing account of model-based reasoning. Before this can be achieved, however, we need more
knowledge about how this interaction affects reasoning. The present chapter is an attempt to bring
together such essential information along three lines, namely the visuo-spatial nature of relational
reasoning, the role of the believability of relational models and the specificity of relational
information. Each of these three aspects is elaborated next.

2. Visuo-spatial basis
In the introduction, we reviewed already the evidence from relational reasoning studies showing that
(a) from the start an integrated representation of the premise information is constructed, (b) that this
representation may be left implicit by adding an annotation that can be further elaborated, and (c) that
after elaboration a forked structure may result. This all very strongly suggests a spatial basis for the
representation of such models in line with the original spatial array theory. If this observation is
correct, one should expect that relational reasoning strongly relies on visuo-spatial working memory,
especially when the terms cannot be represented as a single ordered series. Indeed, if only one-model
problems are involved, the problems can be represented by a single order of tokens (terms, names, ...)
and this can be achieved by the phonological loop because all what is needed is a memory for a string
of tokens. This may be particularly easy when the terms can be represented by their first letter so that
a pseudo- word can be used to represent the model or when the terms are all short words that fall
easily within the storage capacity of the phonological loop (Dierckx et al. 2003). There may be at
least two reasons why with relational problems, reasoners would rely on the visuo-spatial sketch
pad rather than the phonological loop for constructing and maintaining the mental model(s). A first
reason is that the problem cannot be represented by a single string because some form of elementary
structure is present in the representation. A forked representation such as John-Kat°egLea thus would be
more easy to maintain in a visuo-spatial code. A second reason is that some relations are more easily
mapped on a spatial display than other ones. It is evident that spatial relations such as "left of,"
"above," "behind" are more easily represented in a spatial array than relations of other types.
However, temporal relations, such as "before," "at the same time," are also easy to map on a spatial ar-
ray, because we are used to this form of representation (time lines, clocks, etc.). For still other types of
relations, the mapping may be more difficult, but still possible. Whereas the first of these reasons for
using the visuo-spatial sketch pad to construct models mainly capitalizes on the structure and the
complexity of the models, the second reason is a more intrinsic one based on how easy the meaning of
the relation is translated into a spatial representation. In what follows, we will review empirical
evidence relevant for both aspects.
2.1. VlSUO-SPATIAL WORKING MEMORY IN SPATIAL AND TEMPORAL REASONING
To address the first issue, we can refer to a study of Vandierendonck & De Vooght (1997). These
authors investigated the involvement of three working memory components in four-term linear
reasoning tasks, namely the phonological loop, the visuo-spatial sketch pad and the central executive.
In the second experiment of their study, fourty-four participants were randomly assigned to four dual-
task conditions, namely control (reasoning only), articulatory suppression (reasoning while
continuously producing a fixed string of four digits at a fast rate), matrix tapping (reasoning while
continuously tapping the four corners of the numeric keypad at a fast rate) and random interval
repetition (reasoning while "shadowing" a series of randomly spaced time intervals by tapping a key 3 ).
In comparison to the participants in the single-task control condition, the participants in the articulatory
suppression condition did not have the possibility to verbally rehearse the premises. If model
construction uses the phonological loop, reasoning performance of these participants should be
drastically impaired, because it would be very difficult to maintain the information in working
memory. This inference is based on evidence showing that articulatory suppression interferes with
short-term memorization of verbal information (see e.g., Baddeley et al. 1984). Compared to the
participants in the control condition, those in the matrix tapping condition should show impaired
performance if reasoning is based on a visuo-spatial representation of the premise information. Again
previous research has shown that active and passive movements interfere with visuo-spatial working
memory (see e.g., Quinn 1994, Smyth & Scholey 1992, 1994). Finally, in comparison to the
participants in the control condition, those in the random interval repetition condition should be
impaired if reasoning relies on executive control processes. The premises were either presented at a
fixed speeded rate (3 s per premise) or they were presented in a self-paced way with registration of the
time taken to read and to process each premise. It was assumed that in the speeded condition, the
participants would not have enough time to integrate all the information in the premises in a model
and hence performance should be worse than in the self-paced reading condition. As expected, ac-
curacy was poorer in the speeded reading condition (51% correct) than in the self-paced condition
(61% correct), and reasoning performance was impaired in each of the three dual-task conditions.
3 Infact, this is a continuous simple reaction-time task with randomly selected inter- stimulus intervals.
I Time | Space
Control AS MT
DUAL-TASK CONDITIONS
RIR
Fig. 1. Average premise processing times for problems with spatial and temporal content as a function dual-task conditions.
The labels AS, MT and RIR refer to respectively articulatory suppression, matrix tapping and random interval repetition.
However, the most interesting aspect of these findings concerns the premise reading times in the
self-paced condition. Figure 1 shows the average premise reading times as a function of dual task
conditions and problem content (spatial or temporal relations). The figure clearly shows that com-
pared to the control condition, only the matrix tapping and the random interval repetition conditions
required more time to process the premises, while the articulatory suppression condition was not at all
slowed down. So, it is clear that even though the premises are presented verbally, no ver-
bal/phonological secondary task interference was observed. In other words, processing during premise
intake mainly relied on visuo-spatial and executive processes. Another interesting observation is that
the delay due to matrix tapping is larger than the delay due to random interval repetition. It is clear
from these data that this secondary task had a detrimental impact on the processing of the premise
information, while it is known from previous research that it interferes with visuo-spatial rehearsal
without requiring much attention (see e.g., Logie 1995, Smyth & Scholey 1992, 1994, 1996).
Together
with the lack of articulatory interference this indicates that a visuo-spatial representation is built and
maintained in visuo-spatial working memory.
2.2. THE RELATIONAL CONTENT
In the study just discussed, there are at least two reasons why the reasoners may have strongly relied
on their visuo-spatial working memory resources. The first reason is that the problems were based on
either spatial or temporal relations, which are easily mapped on a spatial representation, as already
pointed out. The second reason is that problems could be one- or three-model problems. This
variation in model complexity may have in- duced reasoners to rely more on a visuo-spatial
representation. It would be interesting to know, wether a visuo-spatial representation would be used
when such a representation is not so strongly favoured by the conditions. To clarify this situation, other
types of reasoning may be considered. Toms et al. (1993) worked with variations of conditional rea-
soning tasks. In such a problem, there are typically two premises. The first one expresses a conditional
relation between two terms, such as if it is a circle, then it is green; the second premise expresses a
simple state- ment, such as it is green. Over a series of studies in which Toms and colleagues varied
the contents and the secondary tasks used, they found that conditional reasoning was impaired under a
central executive load, but not with a concurrent visuo-spatial secondary task. Klauer et al. (1997)
used a more general variant of propositional reasoning and they found that when the problems
expressed a spatial relationship, reasoning was impaired by a concurrent visuo-spatial task. Given that
both in relational reasoning and in conditional reasoning based on spatial relations, visuo-spatial
working memory seems to be implied, the hypothesis can be put forward that visuo-spatial
representations are involved when- ever the relations express a spatial order or position. In order to
test this hypothesis Duyck et al. (2003) studied conditional reasoning about relations instead of
entities. Two kinds of relations were studied in conditions with and without a concurrent spatial task
(matrix tapping). One kind of relation was spatial, as in if Pete lives to the right of Paul, then Stan does
not live to the right of Kurt. The other kind of relation was based on joint activities without explicit
spatial order or location, as in if Pete plays tennis with Paul, then Stan does not play tennis with Kurt.
Sixteen problems of each of these two kinds were constructed, and these were made as similar as
possible except for the relation. These 16 problems were obtained as a factorial crossing of four
problem types (Modus ponens or MP: p —> q,p; Denial of the antecedent or DA: p —* q,~<p', Af-
firmation of the consequent or AC: p —> q, q; and Modus tollens or MT:
P —* 1i ~"l) x f°ur levels of negation in the first premise (both p and q positive, p negative and q
positive, p positive and q negative, and both p and q negative). Half of the 42 participants were
assigned to the condition with spatial problems and the other half solved the other (nonspatial)
problems. The reading times needed for each of the two premises and for the solution were registered.
The most interesting part of the latency data concerns the reading times for the first premise. As can
be seen in Figure 2, the concurrent spatial task did not slow down reading time for nonspatial problems
(left panel), while it did for spatial ones (right panel). This effect was not moderated by problem types
(MT, MP, ...), but it interacted with the presence of negations in the first premise. Actually the dual-
task effect was much larger in

I SinglQ-task Zl Dual-task
~25.0
i c/5200 8' 5 0 CLw 10.0

LJJCC 5.0 CL
0.0
I Single-task • Dual-task
25.0
2o.o
COLJJ 8CL CL
TII-
15.0
10.0
CC 5 . 0 CL mS 0.0
j
PROBLEM TYPES
PROBLEM TYPES
Fig. 2. Average premise processing times for the first premise on conditional reasoning problems as a function of the
structure of negations in the first premise and task condition for nonspatial relations (left panel) and spatial relations (right
panel).
problems with a negation than in those without negations. In these findings, two aspects deserve to be
discussed with respect to the role of the concurrent spatial task, namely its effect on spatial and
nonspatial problems, on the one hand, and its effect on the presence of negations in the first premise.
Concerning the first aspect, the present findings sug-

gest that the kind of relation, spatial rather than nonspatial, determines whether a problem will be
represented with the help of visuo-spatial resources. Since the terms of the relations were the same in
the two kinds of problems, only the kind of relation could affect the usage of visuo-spatial
representations. There is no a priori reason to assume that the terms will be represented differently in
the two kinds of problems. Interestingly, the findings are also consistent with those initially reported
by Toms et al. (1993), because they used only relations of the nonspatial kind. Further research could
show which other factors specifically determine the selection of the representational medium for the
the models; it would, for example, be interesting to know whether the terms themselves play any role.
With respect to the presence of negations, it could be argued that a premise such as if Pete lives to
the right of Paul, then Stain does not live to the right of Kurt will be translated into if Pete lives to
the right of Paul, then Stan lives to the left of Kurt. It may be expected that such a transformation
would cost time. However, the right panel of Figure 2 shows that, if anything, in the single-task
conditions the premises with a negation are comprehended more quickly than those without a
negation. What actually seems to happen is that in the presence of a concurrent spatial task, completing
a representation of the premise information is slowed when it contains negations. If the premises
would be converted as suggested above, there is no reason to expect that a spatial secondary task
would affect this process. It seems more likely that first a spatial representation of the premise is built
and that once this is completed a transformation is applied to this model when there is a negation.
Because these operations are performed on the spatial representation, they are expected to be slowed
down in the dual-task condition with a spatial load. For sure, this is not the end of the story. Future
research could build on this finding to bring further clarification. In summary, there seem to be two
aspects that determine whether a relational problem will be represented by visuo-spatial models. The
first is concerned with the structural complexity of the model(s). When the information cannot be
completely represented by a simple string of tokens, the visuo-spatial modality may be used because of
its power to represent structure. The second aspect relates to semantics. When the relation or the
relational content has more affinity to a spatial representation, the visuo- spatial modality seems to be
preferred, even though the reasoner remains in control of the choice actually made. This choice is not
unlimited, but with verbally stated reasoning problems, the reasoner has the choice to build a verbal-
phonological representation or a visuo-spatial representation. There are reasons to believe that the
reasoner has access to information about the gains and the costs of each alternative (cf. models of
strategical choice: Anderson 1990, Siegler & Shipley 1995).

3. Believability of relational models
In many studies of relational reasoning, the relations used are arbitrary ones, chosen in such a way
that the reasoner cannot rely on knowledge. The reason for this is that knowledge may affect
reasoning performance: a (tentative) conclusion which is at odds with our prior knowledge elicits more
efforts to find counterexamples than a conclusion that is consistent with what we know. The result is
that conclusions which are in agreement with our own beliefs are more often accepted than
conclusions violating our beliefs. This is known as the belief bias effect. Most of the research on this
effect has been performed in conditional and syllogistic reasoning (e.g., Evans et al. 2001, Klauer et
al. 2000, Morley et al. 2004, Newstead et al. 1992, Oakhill & Johnson-Laird 1985, Quayle & Ball
2000, Revlin et al. 1980, Thompson et al. 2003). Although relational reasoning offers many interesting
possibilities for such research because of the prominent reliance on visuo-spatial representations,
research with knowledge-related materials has been rather scarce. The studies by Roberts & Sykes
(2003) form a nice exception. They used premises with partly incorrect geographical or historical
information, such as The Pharaohs ruled after the Romans; The Pharaohs ruled before the Normans;
The Pyramids were built at the time of the Romans; At the time of the Normans, William the
Conqueror was king; therefore William the Conqueror was king after the Pyramids were built. With
this kind of problems, these investigators demonstrated a similar kind of belief bias effect as was
observed in categorical reasoning (Evans et al. 1983, 1994, Newstead et al. 1992), namely that
participants accepted believable conclusions more often than unbelievable ones and moreover within
indeterminate problems, the difference was larger for invalid than for valid conclusions. Considered
within the perspective of the mental models theory, not only the conclusions can match or contradict
general knowledge. It is also possible that the premises themselves describe situations (models) that
do or do not correspond with general knowledge. A context in which such effects may typically occur
is related to scripted activities. Within the context of schema theory, scripts are considered as a
particular kind of cognitive schema that can be described as a network consisting of nodes and
(mostly temporal) relations between the nodes. Some nodes are fixed, others are variable or optional
and can be filled in when the script is activated and for some of these variable and optional nodes the
script may provide defaults that can be overridden (see e.g., Abelson 1981, Bower et al. 1979, Brewer
& Treyens 1981, Graesser & Nakamura 1982, Nakamura et al. 1985). Typical scripts are a visit to a
restaurant, to the library, to a doctor, to a dentist, etc. Besides these temporally ordered scripts, there
exist also unordered
scripts, such as the circus script which specifies a number of acts typically performed in a circus, but
the order of the acts is not determinate. How scripts can be used to manipulate the believability of the
entire model given by the premises is illustrated by a study of Dierckx et al. (2004). They used
ordered scripts to create believable and unbelievable problems, which were then compared with neutral
problems based on unordered scripts. Neutral problems were included in order to allow inferences
on whether facilitation (better than neutral) or suppression (worse than neutral) was at the basis of the
observed differences between believable and unbelievable problems. For the construction of
believable and unbelievable problems an ordered script such as a visit to the dentist can be used. A
typical sequence of events for this script is: open mouth, localise toothache, anaesthetize, swallow
saliva, and plug tooth. Table 1 shows how these activities can be combined in a series of all
believable premises that taken together either add up to a believable (left panel) or an unbelievable
model (central panel). The table also shows how an unordered script (circus) is used to create neutral
problems. The most difficult part concerns the ordered scripts. In the believable problems, for
example, four of the five events mentioned form a strict or- dering; one event (swallow saliva) is a
typical script event that can occur anywhere in the sequence. The usage of such an event in each
script, makes it possible to create premises that are believable while the entire sequence may be
unbelievable because it violates the normal script order. The central panel of the table shows how
this is realized. Interestingly, the validity of the proposed conclusion can be varied independently from
the variation in believability, as is shown in the bottom panels of Table 1. Dierckx et al. (2004) used
problems like these to study model believability effects. For half of the participants a header
referring to the script was added. The authors found effects of model believability for invalid but not
for valid conclusions, as is shown in Figure 3. The solution latencies (left panel) of invalid
conclusions were faster in believable than in unbelievable problems, faster in believable than in neutral
problems, but not faster in neutral than in unbelievable problems. In comparison to the neutral prob-
lems, accuracy (right panel) was highest for the believable problems and lowest in the unbelievable
problems and both were different from the neutral. This shows that evocation of the script supports
reasoning when the model is believable and interferes when reasoning with the model is not
believable. It should be noted that this comparison is only based on conclusions that were not
contaminated by possible other effects: The same relations between the second and the fourth term was
tested in all conditions. Interestingly and as expected, the facilitating effect of believable problems
was not only present in solution time and accuracy; it was also observed already during premise
reading. Moreover, this study reveals an

Table 1 Examples of the construction of a believable, an unbelievable and a neutral model from sets of believable premises
based on script events in the study of Dierckx et al. (2004). The neutral problem is based on an "unordered" script and the
actions A-E are presented in any order.

Script actions A: open mouth
B: localise toothache C: anaesthetize
X: swallow saliva
D: plug tooth
Example 1 Believable problem
A before B B before C
C before X X before D
A: tame lions B: ride small bicycle C: throw cones
D: perform clown act E: train horses
Example 2
Example 3 Unbelievable problem
Neutral problem
A before B
A before B B before D
B before C
D before X
C before D X before C
D before E
Model representatior
l
A-B-C-X-D A-B-D-X-C
A-B-C-D-E
Valid conclusions B-X B-X
B-D
Invalid conclusions
X-B X-B
D-B
effect of believability (of the set of premises) while the believability of the individual premises and
the conclusion was held constant. This effect is distinct from the believability as studied thus far in the
literature, namely, the effect of believability of the premises themselves and of the believability of the
conclusion. By and large, these findings are consistent with the view that during premise presentation
a script is triggered and helps to maintain a consistent (believable) model, but seems to interfere for
the maintenance of (unbelievable) or inconsistent models. The main question is, whether these effects
are mediated by the interaction of long-term and working memory. As a further test of this hypothesis,
in an unpublished study, half of the participants were required to do some calculations after premise
presenta-
|Balievable • N e u l r a l I i Unbelievable
Valid Nonvalid
PROBLEM TYPE
I Believable i INsutral I I Unbelievable
1.0
a I0.9 8 0.8
| 0.7
§ 0.6
Valid NorwaNd
0.5
PROBLEM TYPE
Fig. 3. Average solution latency (left) and proportion of correct solutions (right) as a function of model believability and
conclusion validity in the one-model script-based problems of Dierckx et al. (2004).
tion and before conclusion verification. If working memory is used to build and maintain a temporary
model of the premise information, it may be expected that the extra task which requires both storage
and processing of information will compete with the working memory resources needed for the
maintenance of the model. Hence, the differences between believable and unbelievable models should
be enhanced. As expected, the additional calculation task had no effect on the verification accuracy
of valid conclusions; it also had no effect on the accuracy of invalid conclusions in believable and
neutral problems, but it dramati- cally lowered the accuracy of invalid unbelievable conclusions. This
effect is shown in Figure 4. Solution latencies were not affected much by the additional arithmetic
task, because in all cases, the model has been constructed by the end of the premise presentation phase.
It is important to point out that in this study, a working memory load was used which interfered with
the executive control processes, without itself requiring any storage. Since the neutral problems were
basically not affected by this load, it is clear that the effect is not due to competition for temporary
storage. To the contrary, the dramatic performance decrease in the unbelievable problems must be
accounted for by an increased competition between the model maintained in working memory and the
activated script in long-term memory. Be- cause of the load on executive control created by the
simple arithmetic task (see e.g., Deschuyteneer & Vandierendonck 2005), insufficient executive re-
sources, were left for an appropriate control of the interference between the conflicting working
memory and long-term memory representations.
69
Interaction |Be1ievable 0.70.5 I1I1
Knowledge and Working Memory •Neutral • Unbelievable
J-
_ Control Arilhmelic
0.90.So.s 1.0of
CONDITION
IBelievable •Neutral . |Unbelievable
Control Arithmetic
CONDITION
Fig. 4. Average accuracy of correct solutions in valid (left) and invalid (right) conclusion verifications as a function of
experimental condition and model believability in one-model script-based problems.
This set of findings with respect to model believability shows that prior knowledge may facilitate or
hamper reasoning performance. In particular, the activation of a model imbedded in knowledge (a
script) which is part of semantic memory, may help reasoning if this model matches the information
given in the premises. This is the case for premises resulting in a believable model. In contrast, when
the model imbedded in the script activated in semantic memory does not correspond exactly to the
information in the premises, then the model specified by the premises is unbelievable and this results
in both slower and less accurate reasoning. It must be stressed, though, that both the positive effect of
believable models and the negative effect of unbelievable models is not a general characteristic. In
fact, these effects do only seem to occur essentially when the conclusion to be drawn is invalid. For
valid conclusions, believability does not seem to matter so much. This could be an artefact, however,
because valid conclusions in unbelievable problems do not contradict the script (prior knowledge)
and so the effect of believability may not be playing any role in such conclusions. Another aspect that
deserves attention is that in all experiments discussed, attempts were made to include a neutral
condition. Although this is not easy to realize, the usage of unordered scripts seems to yield an ap-
propriate procedure for the construction of such neutral problems. There is much room for
strengthening and elaborating these findings. Nevertheless, with the data available it seems that prior
knowledge stored as scripts may be activated by the problem context and is then present while
the model derived from the premises is constructed. When the constructed model and the activated
model match, the task for the reasoner becomes easier because all what is needed is that the script is
kept active. On the contrary, when the constructed and the activated model differ, additional
(executive) control processes, such as conflict monitoring, will be needed to suppress the activated
model and to maintain the newly constructed model. Because this operation consumes more resources,
reasoning should become slower and more error-prone and this effect should be enhanced when the
resources are depleted by additional processes. This was confirmed in the study with an intervening
arithmetic task and it resulted in larger effects of model believability.

4. Specificity of relational information
Thus far, the present chapter focused on one particular kind of long- term memory effects accounted
for by contextual information. As shown, relevant background knowledge may help or impair the
reasoning performance. There are, however, other possibilities to study the interaction of long-term
and working memory. In genealogical reasoning, for example, the relation itself allows for a tight
coupling between working memory and long-term memory, because kinship information is given
directly via memory representations (Miller & Johnson-Laird 1976, sec. 5.2). Kinship relations are
abstract concepts; they cannot be characterized in terms of obvious external physical characteristics.
There is no way one can tell from perceptual evidence alone whether Paul is Billy's cousin. Kinship
terms have no "concrete" referents. In fact, they are purely cultural in content. Even a concept like
"father" may have different meanings in different cul- tures; in Western Europe, it may refer to the
biological relation (Albert is the biological father of Delphine) or to an adoptive parenthood (Mike is
the adoptive father of Lucia). Moreover, different societies have very different ways of categorizing
relatives; the Western European system is only one of many. For example, in Chinese (Mandarin)
older brother and younger brother are separate terms (i.e., "gege" and "didi") and in the Samoan
kinship terminology, siblings and cousins may be referred to by the same kin terms, e.g. the term
"tuafafine" may refer to the sister of a male anchor, as well as the female cousin of a male anchor
(Jonsson 1999). Kinship relations seem ideally suited to study the interaction of knowledge and
working memory in reasoning about relations. An additional advantage is that kinship terminology
represents a fairly com- pact, well-defined set of relations which, while being small enough to handle
easily, also varies considerably in complexity compared to the rather sim-

pie relations traditionally used in relational reasoning tasks (e.g., "taller than"), spatial relations (e.g.,
"on the left of") or temporal relations (e.g., "before"). Relations such as "father of," "uncle of," and
"grandchild of" are intransitive, while the "ancestor of" and "descendant of"-relations are transitive.
Family relations are often displayed in a spatial format, with verti- cally the descendency relations
(grandfather, father, grandmother, mother, grandparent, parent, son, grandson, daughter,
granddaughter, child, grandchild, ...) and with horizontally the sibling relations (brother, sister) and
the marriage relations. People have learned to use the very specific labels representing different types
of relationships as well as the spatial representation of the family tree (see e.g., Wood & Shotter
1973). On the basis of this, one may expect that people have acquired a number of logical rules to
make inferences from given family relations, but they will also have learned to build (spatial) models
to represent specific situations. A first question, then, concerns the issue whether people indeed have a
preference for the usage of mental models to represent given information as is apparent for other types
of relational reasoning. There are reasons to believe that this is indeed the case, but the present scope
prohibits development of these arguments (but see Dekeyser 1997, for more information on this
issue). Therefore, we shall assume that people indeed solve genealogical reasoning questions by
means of the representation of mental models. The second question, concerns the issue whether the
mental models constructed on the basis of genealogical relations are in any important way different
from the mental models constructed from other order relations. In a typical relational reasoning
problem, reasoners are given order information specifying for example that someone is older than
someone else. The spatial representation is sufficiently detailed when the terms (tokens) are given a
spatial position that corresponds to the order as expressed in the premises. If Ned is older than Olivia
and Olivia is older than Peter, then it does not matter whether Olivia is only a little bit or much older
than Peter. All what is needed to infer the correct conclusion that Ned is older than Peter is the order of
the terms. In genealogical reasoning, care must be taken that also the genealogical distance is rep-
resented when it is needed. If Roger is the grandfather of Stephen and Stephen is the father of Tom, the
inferred relationship Roger is the great-grandfather of Tom, must include the distance between the
terms. One possible solution would be to make explicit that Roger is the father of some x and that this
x is the father of Stephen and to build a representation that explicitly states this. In other words, the
two premises would result in the model "Roger - x - Stephen - Tom" from which it can be inferred that
Roger comes three steps before Tom, so that Roger is the great-grandfather of Tom. Making such
explicit repre-
sentations of the relations costs quite some effort in addition to the need for representational resources. Hence,
it may be expected that people will be reluctant to construct such explicit models. Nevertheless, the reasoner
must take care to develop a correct representation. This could be achieved in several ways. Firstly, it is
possible to place markers that indicate that the relationship can be fleshed out. Where the x-elaboration
discussed above gives a precise specification of the distance, it may also be possible to insert a marker that
there is a non-standard distance (e.g., Roger -()- Stephen), so that later on, if needed an elaboration is possible
by specifying Roger - x - Stephen for the grandfather relation or Roger - x - y - Stephen for expressing a
great-grandfather relation. Given some ideas present in the mental models theory and its developments, one
could expect that people would follow this strategy (e.g., annotations: Vandierendonck et al. 2004). Another
possibility is that reasoners represent the distance without adding any tokens. For the two premises in the last
example, this would result in a representation such as "Roger Stephen - Tom". A further possibility is that
reasoners will only represent the additional distance information if they expect to need it. In a context of
problems about who is who's ancestor, they will probably suffice with the typical relational model, as the
inference Roger is an ancestor of Tom does not call on the distance information at all.
4.1. STUDY 1: TRANSITIVE VERSUS INTRANSITIVE INFERENCE
In order to clarify how the relations are represented in different kinds of reasoning contexts, we compared
reasoning with order and genealogical relations (Van der Beken & Vandierendonck 2005). This was realized
in three different conditions. In a first condition, reasoners were given four premises based on the transitive
relation "older than" (for an example, see left panel of Table 2). After reading the premises, the reasoners were
asked to verify which one of two statements expressing the relation between the second and the fourth term
was correct. In the other conditions, similar premises were presented with the relation "father of" (see middle
and right panel of Table 2). In the second condition, the reasoners were asked to verify the "ancestor"
relationship between the second and the fourth terms. In the third condition, a genealogical relation had to be
verified between the two terms.
Figure 5 displays the main findings with respect to the solution latencies. A first observation is that the solution
times for the intransitive problems were longer than for the transitive problems, irrespective of whether the
latter were based on temporal or on genealogical premises. Secondly, in both conditions where transitive
conclusions were verified, the typical dis-
Table 2 Examples of the three types of problems used in the comparison of reasoning between transitive and genealogical
relations.
Condition 1
Relational
Transitive
Example 1
Believable problem
A before B B before C C before X X before D
Linear
Peter older than Roger Roger older than Klaus Klaus older than Steve
Steve older than Willy
Transitive
Roger older than Steve
Steve older than Roger
Condition 2
Temporal Genealogical
Example 2
Unbelievable problem
A before B B before D D before X X before C
Condition 3
Genealogical
Intransitive
Example 3
Neutral problem
A before B B before C C before D D before E
Genealogical
Peter father of Roger Roger father of Klaus Klaus father of Steve Steve father of Willy
Transitive
Roger ancestor of Steve
Steve ancestor of Roger
Intransitive
Roger grandfather of Steve Steve grandfather of Roger
Neither of both" a In the conditions with transitive relations, the alternative "neither of both" was
not used because this is not a possible alternative. In a control experiment, the
number of alternatives was equal across conditions, but that did not change the
findings.
tance effect was observed: representation the faster the the inference further was apart made.the 4
elements In the in the unified condition with verification of intransitive inferences, however, a
reversed distance effect

4 Figure 5 does not show this very clearly; the figure suggests no difference; nevertheless, the effect was statistically reliable.

Held C.&Co (Eds) Mental Models and The Mind

Uploaded by

Copyright:

Available Formats

Held C.&Co (Eds) Mental Models and The Mind

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Held C.&Co (Eds) Mental Models and The Mind

Uploaded by

Copyright:

Available Formats

ADVANCES

Mental Models and the Mind

Carsten Held Markus Knauti Gottfried Vosgerau Editors

MENTAL MENTAL MODELS MODELS AND AND

Cognitive Cognitive Psychology, Psychology, Neuroscience,

Neuroscience, and and Philosophy Philosophy of of Mind

Working together to grow libraries in developing countries

VolkswagenStiftung through the Research Project Self-consciousness and concept formation in

Toronto, Ontario, Canada M3J 1P3 [email protected]

Verena Gottschling: Philosophisches Seminar, Johannes-Gutenberg-Universitat Jakob-Welder-Weg

18, 55099 Mainz, Germany [email protected]

Cars ten Held: Wissenschaftsphilosophie, Universitat Erfurt

99105 Erfurt, Germany [email protected]

Provence Cedex 1, France [email protected]

Philip N. Johnson-Laird: Department of Psychology, Princeton University Princeton, NJ 08544, USA

Freiburg, Germany [email protected]

Institut fur Philosophie, Otto-von-Guericke-Universitat Virchowstr. 24, 39016 Magdeburg, Germany

Klaus Rehkdmper: Institut fur Philosophie, Carl-von-Ossietzky-Universitat 26129 Oldenburg,

Dunantlaan 2, 9000 Ghent, Belgium

Andre Vandierendonck: Department of Experimental Psychology, Ghent University

Henri Dunantlaan 2, 9000 Ghent, Belgium Andre. Vandierendonck@UGent. be

General Introduction: Current Developments in Cognitive

mental models produce for the philosophy of mind (sec. 8).

unscientific approaches to psychology (Watson 1913). Following Watson's behaviorist proposal,

combination of methods from various disciplines. The leading method,

cognitive science is founded on conceptions of computational theory. It will be helpful to introduce

some of these conceptions.

Therefore, the functional description of an automaton is just as good (informative) as an algorithmic

can be described regardless of how

added by looking at the original implementation (the brain).

generate correct conclusions. This fact is usually explained by limi-

mentioned doctrine of mental

thinking, whereas activity in visuospatial brain ar-

of reasoning—make different predictions concerning the involved brain areas.

representations (misrepresentation), Y-representations should have a disjunctive meaning "X or Y"

(especially for mental

Y-representation is not similar to an X in the relevant sense (otherwise it would be an X-

biological vocabulary. Do attempts to explain mental representation in this vocabulary count as

understand the above-mentioned functionalist theory of mental representation as a naturalization

identified with organic states having these functions. Here,

function of aiding the execution or evaluation of syllogistic reasoning. Thus, it is essential to a

what kind of entities mental models are ontologically.

representations, in M. de Vega, M. Intons-Peterson, P. Johnson-Laird, M. Denis & M. Marschark, eds, 'Models

Introduction: Cognitive Psychology

limited by capacities of the cognitive system, misunderstanding of the premises, ambiguity of

and reports some studies corroborating the

complemented—and even rejected—by representations in the form of mental models.

Mental Models, Sentential Reasoning, and Illusory Inferences

(Johnson-Laird & Byrne

disjunction is true too.

and the truth-functional interpretations of the connectives. Here is an example of such an

(true and false) or else (true and true) —according to the

be compatible with only two possibili-

yourself, try the following problem:

physical world is good enough for you to survive.

its logical sense

triangle. This interpretation is known as a biconditional, because it is equivalent to the assertion of

triangle. The biconditional is compatible with only two possibilities: O A

described, and logical conjunction, which I'm