The Psychology of Learning and Motivation, Volume 54: Advances in Research and Theory

V O LU M E F I F T Y- F O U R THE PSYCHOLOGY OF LEARNING AND MOTIVATION Advances in Research and Theory Series Edit...

Author: Brian H. Ross

44 downloads 1006 Views 2MB Size Report

This content was uploaded by our users and we assume good faith they have the permission to share this book. If you own the copyright to this book and it is wrongfully on our website, we offer a simple DMCA procedure to remove your content from our site. Start by pressing the button below!
Report copyright / DMCA form

DOWNLOAD PDF

V O LU M E

F I F T Y- F O U R

THE PSYCHOLOGY OF LEARNING AND MOTIVATION Advances in Research and Theory

Series Editor Brian H. Ross Beckman Institute and Department of Psychology University of Illinois at Urbana-Champaign Urbana, Illinois

V O LU M E

F I F T Y- F O U R

THE PSYCHOLOGY OF LEARNING AND MOTIVATION Advances in Research and Theory EDITED BY

BRIAN H. ROSS Beckman Institute and Department of Psychology University of Illinois at Urbana-Champaign Urbana, Illinois

AMSTERDAM • BOSTON • HEIDELBERG • LONDON NEW YORK • OXFORD • PARIS • SAN DIEGO SAN FRANCISCO • SINGAPORE • SYDNEY • TOKYO Academic Press is an imprint of Elsevier

Academic Press is an imprint of Elsevier 525 B Street, Suite 1900, San Diego, CA 92101-4495, USA 30 Corporate Drive, Suite 400, Burlington, MA 01803, USA 32 Jamestown Road, London, NW1 7BY, UK Radarweg 29, PO Box 211, 1000 AE Amsterdam, The Netherlands

Copyright # 2011, Elsevier Inc. All rights reserved. No part of this publication may be reproduced, stored in a retrieval system or transmitted in any form or by any means electronic, mechanical, photocopying, recording or otherwise without the prior written permission of the publisher Permissions may be sought directly from Elsevier’s Science & Technology Rights Department in Oxford, UK: phone (+44) (0) 1865 843830; fax (+44) (0) 1865 853333; email: [email protected]. Alternatively you can submit your request online by visiting the Elsevier web site at http://elsevier.com/locate/permissions, and selecting Obtaining permission to use Elsevier material Notice No responsibility is assumed by the publisher for any injury and/or damage to persons or property as a matter of products liability, negligence or otherwise, or from any use or operation of any methods, products, instructions or ideas contained in the material herein. Because of rapid advances in the medical sciences, in particular, independent verification of diagnoses and drug dosages should be made ISBN: 978-0-12-385527-5 ISSN: 0079-7421

For information on all Academic Press publications visit our website at elsevierdirect.com

Printed and bound in USA 11 12 13 10 9 8 7 6 5 4 3 2 1

CONTENTS

Contributors

ix

1. Hierarchical Control of Cognitive Processes: The Case for Skilled Typewriting

1

Gordon D. Logan and Matthew J. C. Crump 1. Introduction 2. What is Hierarchical Control? 3. The Two-Loop Theory of Typewriting 4. Distinguishing the Outer Loop and the Inner Loop 5. Words as the Interface Between Outer and Inner Loops 6. The Inner Loop is Informationally Encapsulated 7. The Outer Loop and the Inner Loop Rely on Different Feedback 8. Beyond Typewriting Acknowledgments References

2 3 6 7 9 12 15 18 23 23

2. Cognitive Distraction While Multitasking in the Automobile

29

David L. Strayer, Jason M. Watson, and Frank A. Drews 1. A Framework for Understanding the Sources of Driver Distraction 2. Do Cell-Phone Conversations Increase the Crash Risk? 3. Why Does Talking on a Cell Phone Impair Driving? 4. Are All Conversations Harmful to Driving? 5. Can the Interference Be Practiced Away? 6. Is Everyone Impaired by Using a Cell Phone While Driving? 7. Conclusions and Future Directions References

3. Psychological Research on Joint Action: Theory and Data

30 33 42 47 49 50 55 56

59

¨nther Knoblich, Stephen Butterfill, and Natalie Sebanz Gu 1. 2. 3. 4.

Introduction Emergent and Planned Coordination Evidence Discussion

60 62 66 91

v

vi

Contents

Acknowledgments References

4. Self-Regulated Learning and the Allocation of Study Time

93 93

103

John Dunlosky and Robert Ariel 1. Introduction 2. Self-Regulated Learning 3. Allocation of Study Time 4. Agenda-Based Regulation Framework 5. Comparing Accounts of Study-Time Allocation 6. Concluding Remarks and Some Directions for Future Research Acknowledgments References

5. The Development of Categorization

104 105 108 120 131 135 136 136

141

Vladimir M. Sloutsky and Anna V. Fisher 1. Introduction 2. Categorization and Selective Attention 3. The Role of Labels in the Development of Categorization 4. Early Categorization: What Develops? Acknowledgments References

6. Systems of Category Learning: Fact or Fantasy?

142 147 153 162 163 163

167

Ben R. Newell, John C. Dunn, and Michael Kalish 1. Introduction 2. Review and Critique of the Evidence I: Probabilistic Category Learning 3. Review and Critique of the Evidence II: Deterministic Category Learning 4. Reexamining Some Fundamental Assumptions 5. The Contribution of Mathematical Modeling 6. Discussion and Conclusions Acknowledgment References

7. Abstract Concepts: Sensory-Motor Grounding, Metaphors, and Beyond

168 170 179 188 199 206 210 210

217

Diane Pecher, Inge Boot, and Saskia Van Dantzig 1. Grounded Cognition 2. Representing Abstract Concepts: Some Evidence for Grounding 3. Explanations of Abstract Concepts 4. Discussion Acknowledgments References

218 224 228 240 242 243

Contents

8. Thematic Thinking: The Apprehension and Consequences of Thematic Relations

vii

249

Zachary Estes, Sabrina Golonka, and Lara L. Jones 1. Introduction 2. Definition and Differentiation 3. Dissociating Thematic Relations from Taxonomic (Categorical) Relations 4. Apprehension of Thematic Relations 5. Consequences of Thematic Relations for Cognition 6. Individual Differences and Cultural Effects 7. Conclusion References Subject Index Contents of Recent Volumes

250 251 256 261 268 280 285 287 295 301

This page intentionally left blank

CONTRIBUTORS

Robert Ariel Psychology Department, Kent State University, Kent, OH, USA Inge Boot Psychology Department, Erasmus University, Rotterdam, The Netherlands Stephen Butterfill Department of Philosophy, University of Warwick, Warwick, United Kingdom Matthew J. C. Crump Department of Psychology, Vanderbilt University, Nashville, TN, USA Frank A. Drews Department of Psychology, University of Utah, Salt Lake City, UT, USA John Dunlosky Psychology Department, Kent State University, Kent, OH, USA John C. Dunn School of Psychology, University of Adelaide, Adelaide, Australia Zachary Estes Department of Psychology, University of Warwick, Warwick, United Kingdom Anna V. Fisher Department of Psychology, Carnegie Mellon University, Pittsburgh, PA, USA Sabrina Golonka Department of Psychology, Leeds Metropolitan University, Leeds, United Kingdom Lara L. Jones Department of Psychology, Wayne State University, Detroit, MI, USA Michael Kalish Institute of Cognitive Science, University of Louisiana, Lafayette, LA, USA ¨nther Knoblich Gu Centre for Cognition, Donders Institute for Brain, Cognition, & Behaviour, Radboud University Nijmegen, The Netherlands Gordon D. Logan Department of Psychology, Vanderbilt University, Nashville, TN, USA

ix

x

Contributors

Ben R. Newell School of Psychology, University of New South Wales, Sydney, Australia Diane Pecher Psychology Department, Erasmus University Rotterdam, Rotterdam, The Netherlands Natalie Sebanz Centre for Cognition, Donders Institute for Brain, Cognition, & Behaviour, Radboud University Nijmegen, The Netherlands Vladimir M. Sloutsky Department of Psychology and Center for Cognitive Science, The Ohio State University, Columbus, OH, USA David L. Strayer Department of Psychology, University of Utah, Salt Lake City, UT, USA Saskia Van Dantzig Department of Psychology, Leiden University, Leiden, The Netherlands Jason M. Watson Department of Psychology, and The Brain Institute, University of Utah, Salt Lake City, UT, USA

C H A P T E R

O N E

Hierarchical Control of Cognitive Processes: The Case for Skilled Typewriting Gordon D. Logan and Matthew J. C. Crump Contents 1. Introduction 2. What is Hierarchical Control? 2.1. Hierarchy 2.2. Control 2.3. Hierarchical Control 2.4. The Case for Hierarchical Control, 0.0 3. The Two-Loop Theory of Typewriting 4. Distinguishing the Outer Loop and the Inner Loop 4.1. Distinguishing the Loops by Selective Influence 4.2. The Case for Hierarchical Control, 1.0 5. Words as the Interface Between Outer and Inner Loops 5.1. Words Prime Constituent Letters in Parallel 5.2. Words Activate Spatial Locations of Constituent Letters 5.3. Words Activate Motor Representations of Constituent Letters 5.4. The Case for Hierarchical Control, 2.0 6. The Inner Loop is Informationally Encapsulated 6.1. The Outer Loop Does Not Know Which Hand Types Which Letter 6.2. The Outer Loop Does Not Know Where Letters Are on the Keyboard 6.3. The Case for Hierarchical Control, 3.0 7. The Outer Loop and the Inner Loop Rely on Different Feedback 7.1. The Inner Loop Relies on the Feel of the Keyboard 7.2. The Outer Loop Relies on the Appearance of the Screen 7.3. The Case for Hierarchical Control, 4.0 8. Beyond Typewriting 8.1. Hierarchical Control in Other Skills 8.2. Hierarchical Control, Automaticity, Procedural Memory, and Implicit Knowledge

Psychology of Learning and Motivation, Volume 54 ISSN 0079-7421, DOI: 10.1016/B978-0-12-385527-5.00001-2

#

2 3 3 3 4 4 6 7 8 9 9 10 10 11 11 12 12 13 15 15 15 17 18 18 18 19

2011 Elsevier Inc. All rights reserved.

1

2

Gordon D. Logan and Matthew J. C. Crump

8.3. The Development of Hierarchical Control 8.4. Nested Control Loops in Everyday Cognition Acknowledgments References

20 22 23 23

Abstract The idea that cognition is controlled hierarchically is appealing to many but is difficult to demonstrate empirically. Often, nonhierarchical theories can account for the data as well as hierarchical ones do. The purpose of this chapter is to document the case for hierarchical control in skilled typing and present it as an example of a strategy for demonstrating hierarchical control in other cognitive acts. We propose that typing is controlled by two nested feedback loops that can be distinguished in terms of the factors that affect them, that communicate through intermediate representations (words), that know little about how each other work, and rely on different kinds of feedback. We discuss hierarchical control in other skills; the relation between hierarchical control and familiar concepts like automaticity, procedural memory, and implicit knowledge; and the development of hierarchical skills. We end with speculations about the role of hierarchical control in everyday cognition and the search for a meaningful life.

1. Introduction The idea that cognition is controlled hierarchically is ubiquitous but enigmatic. Hierarchical control was a critical issue in the cognitive revolution against the behaviorists in the 1950s (Lashley, 1951; Miller, Galanter, & Pribram, 1960) and it remains a common feature in modern theories of executive control in cognitive science and cognitive neuroscience (Badre, 2008; Cooper & Shallice, 2000; Logan & Gordon, 2001; Miller & Cohen, 2001; Norman & Shallice, 1986). Nevertheless, the case for hierarchical control is weaker than it ought to be, with much of cognitive psychology focused on simple tasks that can be explained readily in terms of the stimulus–response associations that the behaviorists favored (Logan, 1988; Shiffrin & Schneider, 1977). Many of the more complex tasks can be explained by nonhierarchical control as well as by hierarchical control (see Botvinick & Plaut, 2004, 2006 vs. Cooper & Shallice, 2000, 2006a, 2006b). In this chapter, we present the case for hierarchical control in skilled typewriting. We present a theory that claims there are two hierarchically nested feedback loops in skilled typewriting and we present evidence that the two loops can be distinguished by factors that selectively influence them, that words are the interface between the two loops, that the two loops share little knowledge beyond words, and that the two loops rely on different feedback. We offer this analysis as a method for making the case for

Hierarchical Control of Cognitive Processes: The Case for Skilled Typewriting

3

hierarchical control in other instances of cognitive control, and we draw implications from our analysis of typewriting to other issues and topics in cognitive control of thought and action.

2. What is Hierarchical Control? 2.1. Hierarchy The case for hierarchical control must begin with definitions of hierarchy and control so we know what we are looking for when we examine typewriting. A hierarchy is a representation that has at least two levels with a one-to-many mapping of elements in the higher level to elements in the lower level (Markman, 1999; Novick & Hurley, 2001). The texts that typists type satisfy this definition: Texts are made of paragraphs, paragraphs are made of sentences, sentences are made of words, and words are made of letters. Typists’ psychological representations of the texts must reflect this structure, so typing is driven by hierarchical representations. However, the debate about hierarchical control is not about hierarchical representation, but rather, about whether the processes that operate on the representations are also hierarchical. Hierarchical representation does not imply hierarchical processing. The same processes could operate at different levels of a hierarchical representation. For example, Schneider and Logan (2006, 2007) had subjects perform a sequence of tasks and argued that the sequence and the tasks were represented at different levels of a hierarchy but suggested that the same memory retrieval processes operated on both levels (also see Botvinick & Plaut, 2004). The case for hierarchical control requires more than demonstrating hierarchical representations.

2.2. Control A process is controlled if it is willfully directed toward the fulfillment of a goal (Logan, 1988; Miller et al., 1960). “Willfully” is a difficult component of this definition. Some argue that a process is willful if it is chosen voluntarily (Arrington & Logan, 2004). Others argue that a process is willful if it can be interrupted on demand (Logan, 1982; Logan & Cowan, 1984). However, there is no consensus on conceptual or operational definitions of will (Wegner, 2002). “Goal directed” is an easier component and we will focus on it in this chapter. We assume that typing is willful. People do not type by accident. Miller et al. (1960) expressed this definition of control in a generic feedback loop called a TOTE unit. TOTE stands for “test, operate, test, exit,” which involves a comparison of the current state with the goal state (test), followed by the execution of an operation intended to reduce the difference between the current state and the goal state (operate), followed by another comparison of the current state with the goal state (test). If the current

4


state matches the goal state, the task is completed, and the system goes on to pursue other goals (exit). If the current state does not match the goal state, the task is not yet complete, and the operate phase is engaged again, iterating until the goal state is attained. For example, the finger movements in typing the letter “T” can be described by a TOTE that compares the goal state (“is the index finger above the T key?”) with the current state of the finger (“the index finger is above the F key”), resulting in successive operations (movements toward the T key) until the goal state is satisfied (the index finger is above the T key and ready to strike it). TOTE units embody control because the operations are directed toward the attainment of a goal.

2.3. Hierarchical Control Miller et al. (1960) offered TOTE units as a cognitive or cybernetic alternative to the generic stimulus–response bond in behaviorist psychology. TOTE theory was a conceptual advance over stimulus–response bonds because TOTEs have more structure (see Chomsky, 1959). They can be concatenated to create extended chains of complex behavior, which Miller et al. called plans. For example, the plan for typing the letter “T” can be described by a sequence of two TOTEs, in which the first moves the finger from its current location to the location above the T key, as described in the earlier example, and the second depresses the key (the test is “is the key depressed?” and the operation is “push the index finger down”). More importantly for our purposes, TOTEs can be nested hierarchically. The operate phase of one TOTE can be replaced by another TOTE or a series of TOTEs that describe the details of the operation. Thus, the two-TOTE plan for moving the finger and depressing the key can be viewed as the operate phase of a superordinate TOTE that types the letter. The operations in the subordinate TOTEs need not resemble the operations in the superordinate TOTE, so hierarchical processes (nested operations) can be distinguished from hierarchical representations (nested goals). Moreover, TOTEs can be modular and informationally encapsulated (Fodor, 1983): a superordinate TOTE need not be aware of the operations in its subordinate TOTEs. The superordinate TOTE only needs to know that the subordinate TOTEs moved the current state toward the superordinate goal state. These features are important in typewriting, so we will adopt the idea of nested TOTEs—nested feedback loops—as our generic definition of hierarchical control.

2.4. The Case for Hierarchical Control, 0.0 The case for hierarchical control in any domain requires several demonstrations. First, there must be at least two levels of processing that can be distinguished by manipulations of experimental factors. Selective influence


5

by itself is not strong evidence for hierarchical processing because it is also found for separate processes that are not arranged hierarchically (e.g., Sternberg, 1969). Nevertheless, processes that cannot be distinguished from one another cannot be said to be organized hierarchically, so this is a necessary but not sufficient step in developing the case. To make this part of the case, we review studies of selective influence in typewriting. Second, the two levels of processing must operate on different kinds of information that exist at different levels of an informational hierarchy, with the higher level operating on higher-level information than the lower level. One way to make this part of the case is to demonstrate an informational hierarchy and show that information at some intermediate level serves as the interface between levels of processing. The higher level deals with units at this intermediate level and higher; the lower level deals with units at this intermediate level and lower. To make this part of the case, we review studies that show the role of words as the interface between language processes and motor control in typewriting. Third, the two levels of processing must divide the intellectual labor required to perform the task in a way that is consistent with hierarchical processing. The higher level must deal with larger structures and broader goals than the lower level. Moreover, the higher level should not know much about what the lower level is doing. The higher level should issue commands and determine whether they are executed without knowing the details of how the lower level executes the commands. The lower level should be informationally encapsulated, so the details of its processing are not available to the higher level (Fodor, 1983). To make this part of the case, we review studies that show that skilled typists know little about how they type. Fourth, the two levels of processing must utilize different kinds of feedback, appropriate to the goals they address. The higher level should process feedback about higher-level goals and the lower level should process feedback about lower-level goals. Demonstrating sensitivity to different kinds of feedback requires identifying the goals that drive each level of processing and identifying the states of the cognitive system or states of the environment that signal progress toward those goals. To make this part of the case, we review studies that identify different levels of feedback that are utilized in skilled typewriting. Finally, the levels of processing must be integrated in an overarching theory of the computations involved in successful performance. The first four steps mostly address whether two or more levels can be distinguished. The computational analysis specifies the relation between the levels, situating them in a larger system whose parts are individually necessary and jointly sufficient to perform the task at hand. This part of the case has already been made in typewriting: several computational theories have been proposed and tested ( John, 1996; Rumelhart & Norman, 1982; Salthouse, 1986;

6


Wu & Liu, 2008). Our analysis of typewriting abstracts two hierarchically nested levels of processing from these previous theories. Our efforts at establishing the first four steps in the case are grounded in these more complex theories, which specify the hierarchical relation between the two levels in computational terms.

3. The Two-Loop Theory of Typewriting Typewriting is a recent skill in human history. The QWERTY keyboard on which English speakers type was patented on July 14, 1868 by Christopher Latham Scholes. In 1876, the first book was created on a typewriter (Tom Sawyer by Mark Twain). By 1900, typewriters were in common use and Scholes was heralded as an emancipator for bringing women into the workplace as typists. From 1900 to 1980, typists were a small group of trained professionals who spent most of their time transcribing text—copy typing. The proliferation of personal computers in the 1980s brought typing to the masses. Now, typing is ubiquitous. Most American homes, businesses, and schools have computers, and most college-age people have strong typing skills. In 800 college-age typists we tested, mean typing speed was 68 words per minute (SD ¼ 18; range ¼ 21– 126), similar to yesterday’s professional typists. In 246 typists we surveyed, the average age at which they started typing was 10 years. They had typed for 10.8 years and currently spent 4.4 h per day on their computers. Seventy-eight percent had formal training in typing, averaging 44 weeks in duration. The nature of their typing was different from yesterday’s professional typists: Only 9% of their typing involved transcription, 51% involved composition (e-mail, essays), and 40% involved other activities (instant messages, search engines, user ID, etc.). In addition, they averaged 40 text messages per day (range ¼ 0–500). Compared to most tasks studied in cognitive science and cognitive neuroscience, typewriting is a complex activity. It involves many successive responses that are strongly constrained. The input is constrained by language processing and the output is constrained by the QWERTY keyboard. The act of typing itself is constrained by the need for speed and accuracy. All of the characters in each word have to be typed as quickly as possible in the correct order (Lashley, 1951). These constraints engage a wide variety of processes, ranging from language comprehension and generation to hand and finger movements ( Johns, 1996; Rumelhart & Norman, 1982; Salthouse, 1986; Shaffer, 1976; Wu & Liu, 2008). Our research has been guided by a simple model of typewriting that divides the many processes into two nested feedback loops or TOTEs: an outer loop that begins with language comprehension or generation and ends


7

with a series of words to be typed, and an inner loop that begins with a word to be typed and ends with a series of keystrokes (Logan & Crump, 2009). It is said that science carves nature at the joints. Our outer-loop inner-loop theory is intended to carve typewriting at a major joint, like severing an arm at the shoulder. Other theories address smaller joints, distinguishing more components and focusing on finer details of performance such as the timing of keystrokes (Gentner, 1987; Soechting & Flanders, 1992; Sternberg, Monsell, Knoll, & Wright, 1978; Terzuolo & Viviani, 1980; Viviani & Laissard, 1996) and the kinematics of finger movements (Flanders & Soechting, 1992; Gordon, Casabona, & Soechting, 1994; Soechting & Flanders, 1992). We do not deny the importance of these smaller joints. Instead, we argue that the distinction between the outer loop and the inner loop is important and general. It encompasses theories that address finer details and it leads to testable hypotheses and useful extensions to other instances of cognitive control. In recent years, we have amassed evidence that makes a strong case for hierarchical control in typewriting. We have shown that the outer loop and inner loop are affected by different factors, that they communicate at the level of words rather than sentences or keystrokes, that the outer loop knows little about the workings of the inner loop, and that the two loops rely on different kinds of feedback. In the remaining pages, we describe the experiments that support these claims and suggest how our experimental procedures might be used to generalize these claims to other instances of cognitive control.

4. Distinguishing the Outer Loop and the Inner Loop A straightforward way to distinguish the outer loop and the inner loop is to examine response times and interkeystroke intervals in discrete typewriting tasks, in which a single word is presented and subjects are instructed to type it as quickly as possible. Response time is the interval between the onset of the word and the registration of the first keystroke. In our theory, response time measures the duration of outer-loop processes that identify the word on the screen and pass it to the inner loop. Response time must also include the duration of inner-loop processes that prepare and execute the first keystroke. Thus, response time measures the duration of both loops. Interkeystroke interval is the interval between successive keystrokes. It is tempting to think of interkeystroke interval as a response time, but we should not yield to that temptation (Lashley, 1951). High-speed films of typists typewriting show that finger movements often occur in parallel, and the finger movement for one keystroke often begins before the finger

8


movement for the preceding keystroke ends (Flanders & Soechting, 1992). Thus, it is better to think of interkeystroke interval as measuring the finishing times of concurrent, temporally overlapping processes. In our theory, interkeystroke interval measures processing in the inner loop. The inner loop prepares and executes successive keystrokes, and interkeystroke interval measures the rate at which this occurs.

4.1. Distinguishing the Loops by Selective Influence In our theory, outer and inner loops can be distinguished by experimental factors that selectively influence response time and interkeystroke interval (Sternberg, 1969). For example, Logan and Zbrodoff (1998) ran a Stroop (1935) task with typewritten responses and found that the congruency of the color and the word affected response time but not interkeystroke interval. This suggests that congruency affects the choice of the word to type, which is the business of the outer loop, but not the execution of the keystrokes in the word, which is the business of the inner loop. Similarly, Logan (2003) had typists type words presented on the left or right side of a central fixation point and found Simon-type interference for words typed entirely with the left or right hand. The congruency of stimulus and hand locations affected response time but not interkeystroke interval, suggesting that congruency affects the choice of the word to type but not the execution of keystrokes. Some of the factors that are important in typewriting do not selectively influence a single stage of processing, and this limits the utility of selective influence in distinguishing the outer loop from the inner loop. For example, Crump and Logan (2010a) examined repetition priming in typewriting, presenting words several times and comparing performance on repeated words with performance on new words that had not been presented in the experiment. We found that repetition reduced both response time and interkeystroke interval, suggesting that repetition affects only inner-loop processing, which contributes to both measures. However, repetition could affect both inner- and outer-loop processing. Other data suggest that repetition priming facilitates perceptual and conceptual processing (Logan, 1990), which is part of the outer loop, so we can reject the hypothesis that repetition affects only inner-loop processing, but the data themselves cannot distinguish between the hypotheses. The problem is that repetition priming does not selectively influence outer- and inner-loop processing. More generally, factors that do not selectively influence the outer and inner loops cannot be used to distinguish between them (Sternberg, 1969). This limits the utility of the contrast between response times and interkeystroke intervals as a way to distinguish between outer and inner loops. Even more generally, the problem of selective influence shows the limits of defining processes in terms of the factors that affect them (Garner, Hake, & Eriksen, 1956; Logan, Coles, & Kramer, 1996; Sternberg, 1969).


9

We prefer to define processes in terms of the computations they require. Finally, the method of selective influence works only with discrete typing tasks in which response time can be defined meaningfully. It does not work with continuous typing tasks, which only provide information about interkeystroke intervals.

4.2. The Case for Hierarchical Control, 1.0 Studies of selective influence suggest that the outer and inner loops are affected by different factors, but they also indicate that the two loops are sometimes affected by the same factors. Thus, studies of selective influence suggest that two separate processes underlie typewriting but do not provide strong evidence for hierarchical control.

5. Words as the Interface Between Outer and Inner Loops There is abundant evidence that words are important units in typewriting. Manipulations of units larger than the word have little effect on typewriting: scrambled sentences are typed as quickly as intact ones (Fendrick, 1937; Gentner, Larochelle, & Grudin, 1988; Shaffer & Hardwick, 1968). However, manipulations of units smaller than the word have a strong effect: scrambled words and random letter strings are typed more slowly than intact words (Fendrick, 1937; Gentner et al., 1988; Shaffer & Hardwick, 1968). Studies that manipulated preview of the text to be typed found that increasing preview from 1 to 8 characters (approximately one word) increased typing speed, but further increases in preview up to 40 characters produced no further increase (Hershman & Hillix, 1965; Shaffer, 1973). More generally, reading (250–350 words/min) and speaking (120–200 words/min) are much faster than typing (50–100 words per minute; Rayner & Clifton, 2009), so the effects of units larger than the word may be absorbed while the outer loop waits for the inner loop to finish the current word. We interpret these effects as evidence for the proposition that words are the interface between the outer and inner loops. The outer loop generates a series of words to be typed, through language generation or comprehension, and passes them one by one to the inner loop. The inner loop takes each word and translates it into a series of letters to be typed, translates the letters into a series of keystrokes, and executes them one by one on the keyboard. The hierarchical relationship between words and letters is mirrored in the hierarchical relationship between the outer and inner loops: One word in the outer loop corresponds to several letters in the inner loop.

10


This perspective predicts that words will activate their constituent letters and the corresponding keystrokes in parallel. We tested this prediction in three ways.

5.1. Words Prime Constituent Letters in Parallel Crump and Logan (2010b) developed a priming technique, in which typists were given five- or seven-letter words as primes. On some trials, the prime was followed by another copy of itself, and typists typed the word. On other trials, the prime was followed by a single-letter probe, and typists typed the letter. The letter was either the first, middle, or last letter in the prime word or a randomly chosen letter that did not appear in the prime. We found that typists typed the single-letter probes faster if they appeared in the prime word than if they did not, suggesting that the prime activated all of its constituent letters. Priming was greater for the first letter than for the middle and last letters, but there was no difference in priming for the middle and last letters. This suggests that there is an advantage to the first letter, perhaps because it must be typed first, but the middle and last letters were activated similarly, which is consistent with the hypothesis that all letters in the word are activated in parallel. The priming words were presented visually, so it is possible that the priming effects were perceptual rather than motoric: seeing the letter in a word may have sped up perceptual processing of single-letter probes. We addressed this possibility in two ways. First, we presented auditory primes rather than visual ones and found the same effects: response time to singleletter probes was faster when the letters were part of the prime word than when they were not. Priming was greater for the first letter than for middle and last letters, but middle and last letters were primed equally. This shows that visual presentation of the primes is not necessary to produce withinword priming effects. Second, we presented strings of consonants as primes. The strings of consonants should be represented as several units in the outer loop, and we assumed that only the first unit would be passed to the inner loop. Consistent with this assumption, we found priming for the first letter of the string but no priming for the middle or last letter. This shows that visual presentation of the primes is not sufficient to produce within-word priming effects.

5.2. Words Activate Spatial Locations of Constituent Letters Logan (2003) had typists type single words presented to the left or right of a central fixation point. There were three types of words: LEFT/right words were constructed such that all letters were typed entirely in the left or right hand (e.g., rest), LEght/riFT words were constructed such that the first two letters were typed with one hand and the remaining letters were typed with


11

the other (e.g., swim), and Light/rEFT words were constructed such that the first letter was typed in one hand and the remaining letters were typed in the other hand (e.g., dump). If the words activated their constituent keystrokes in parallel, then there should be a Simon-type effect (Simon & Small, 1969), in which words presented on the same side of the screen as their constituent keystrokes should be typed faster than words presented on the opposite side of the screen. Importantly, the Simon effect should be stronger when more letters were typed in one hand; distributing letters across the hands should weaken the effect. Consistent with this prediction, the Simon effect was stronger with LEFT/right and LEght/riFT words than with Light/rEFT words. The effect of letters beyond the first suggests that the spatial locations of all the letters were activated in parallel.

5.3. Words Activate Motor Representations of Constituent Letters Logan, Miller, and Strayer (2011) presented LEFT/right, LEght/riFT, and Light/rEFT words centrally and measured the lateralized readiness potential in the electroencephalogram while typists typed them. The lateralized readiness potential is the difference in electrical potential between electrodes located over the left and right motor cortex (C3 and C4 in the international 10–20 system; Jasper, 1958). It reflects the process of response selection, measuring the growth in activation of motor representations of responses that are about to be executed (Coles, 1989). Logan et al. focused on the early part of the lateralized readiness potential time-locked to the first keystroke to measure activation of motor representations while the first response was selected. If the constituent keystrokes of a word are activated in parallel, then the amplitude of the lateralized readiness potential for the first keystroke should decrease systematically as progressively more keystrokes are activated in the opposite hand. Thus, the lateralized readiness potential should be greater for LEFT/right words than for LEght/riFT words, and greater for LEght/riFT words than for Light/rEFT words. However, if the constituent keystrokes are activated in series, then there should be no effect of subsequent letters on the lateralized readiness potential for the first letters of LEFT/right, LEght/riFT, and Light/rEFT words. The results were consistent with parallel activation: the amplitude of the lateralized readiness potential for the first keystroke decreased monotonically from LEFT/right to LEght/riFT to Light/rEFT words.

5.4. The Case for Hierarchical Control, 2.0 The idea that words are the interface between the outer loop and the inner loop provides strong evidence for hierarchical control. Outer-loop processes operate on structures larger than the word in comprehending and

12


generating language, and that processing results in a series of words that are passed one at a time to the inner loop. Inner-loop processes translate the word into letters, motor plans addressed to keyboard locations, and ultimately, keystrokes. Three lines of evidence provide strong support for the notion that outer-loop words translate to inner-loop motor plans in parallel, reflecting the one-to-many mapping that is characteristic of a hierarchy. Together with the evidence that the two loops may be influenced selectively by experimental factors, this evidence strengthens the case for hierarchical control of typewriting.

6. The Inner Loop is Informationally Encapsulated Our outer-loop inner-loop theory assumes that the intelligence required to type text is divided between the loops. The outer loop is concerned with language generation and comprehension, and its job is to produce a string of words to be typed. The inner loop is concerned with translating words into letters, finger movements, and keystrokes, and its job is to produce a series of keystrokes on the keyboard. With this division of labor, the outer loop does not need to know what the inner loop does. It only needs to provide the inner loop with words to be typed, one at a time. We believe that the outer loop does not know what it does not need to know. The inner loop is informationally encapsulated (Fodor, 1983), so the outer loop does not know how the inner loop does what it does. In this respect, typing is like other skills that exhibit paradoxical dissociations between explicit knowledge and implicit knowledge: practitioners of high-level skills know how to perform very well but do not know much about how they do it (Beilock & Carr, 2001; Beilock, Carr, MacMahon, & Starkes, 2002; Terzuolo & Viviani, 1980).

6.1. The Outer Loop Does Not Know Which Hand Types Which Letter The outer loop knows which words must be typed and it is able to spell the words, but it usually does not break words down into letters before passing them to the inner loop. The previous section summarized the evidence suggesting that the outer loop passes whole words to the inner loop, and the inner loop breaks the words down into letters and assigns the letters to particular hands and particular keyboard locations. This division of labor suggests that the outer loop does not know which hand types which letter but the inner loop does (it must because it types letters correctly).


13

Logan and Crump (2009) showed that the outer loop does not know which hand types which letter by having typists type only the letters assigned to one hand. This was very disruptive, as you can confirm for yourself by typing only the right-hand letters in this sentence. In one experiment, we had typists type whole paragraphs and told them to type only the left-hand letters (or only the right-hand letters). With these instructions, typists typed 14 words per minute and made errors on 33% of the words. When the same typists typed the same texts under instructions to type normally (i.e., to type all letters), their typing speed was 80 words per minute and they made errors on only 6% of the words. Similar results were found when typists typed single words preceded by a cue that told them whether to type the words in one hand (LEFT or RIGHT) or to type all the letters (WHOLE). The requirement to type only the letters in one hand increased the response time by 454 ms (a 54% increase), interkeystroke interval by 153 ms/keystroke (a 104% increase), and error rate by 16% (a 304% increase). A control experiment in which the letters to be typed were cued by color (“type only the red letters”) produced no disruption, suggesting that the difficulty lay in discovering which hand typed which letter. These gargantuan disruptions are paradoxical: They suggest that skilled typists do not know which hand types which letters, yet they choose the correct hand 5–6 times/s in normal typing. Our two-loop hypothesis resolves the paradox by proposing that the inner loop is informationally encapsulated. In order to discover which hand types which letter, the outer loop must observe the inner loop’s output. To type letters from only one hand, the outer loop must slow the inner-loop’s cycle time so that it has time to observe which hand was selected and inhibit the keystroke if necessary. Other investigators have shown similar disruptions from drawing attention to the details of performance in other skills (Beilock & Carr, 2001; Beilock et al., 2002). Our research provides one explanation for the disruption: performance must slow down so the outer loop can observe the details, and that disrupts timing and the fluency of performance.

6.2. The Outer Loop Does Not Know Where Letters Are on the Keyboard Another paradox in typing skill concerns knowledge of where the letters are located on the keyboard. Our intuitions as typists tell us we have little explicit knowledge of letter location, yet our fingers find the correct locations five to six times per second. We suggest that this is also a consequence of the division of labor between outer and inner loops and another example of encapsulated inner-loop knowledge. In this case, there may be stronger motivation for encapsulating knowledge about letter location in the inner loop: the locations of letters in words rarely correspond to the locations of letters on the keyboard (e.g., the letters in pout are in opposite left-to-right

14


order on the screen and on the keyboard). Encapsulating knowledge about letter location and communicating information about letters through the intermediary representations of words may reduce the costs of stimulus– response incompatibility. Liu, Crump, and Logan (2010) had skilled typists make explicit judgments of the relative locations of keys on the keyboard, using standard procedures from the literature on spatial memory (McNamara, 1986; Stevens & Coupe, 1978). We asked typists to imagine that they were standing on a key on the keyboard (e.g., F) facing a particular direction (e.g., the space bar) and then point to the location of another letter (e.g., W), indicating the direction with a mouse. When typists judged the relative direction with no keyboard in the testing room and so could only rely on explicit (outer loop) knowledge of keyboard locations, the absolute angular error (the unsigned difference between the actual angle and the judged angle) was 47 . A control group who made the same judgments when it could see the keyboard in the room had an absolute angular error of 28 . Another control group who made the same judgments after typing the letters on a keyboard covered by a box to prevent it from seeing the keys had an absolute angular error of 23 . In a second experiment, Liu et al. (2010) had skilled typists use a mouse to drag a depiction of a key to its location relative to another key. The absolute angular error was smaller overall, but it was still greater for typists who had to imagine the keyboard (29 ) than for typists who could look at the keyboard (17 ) or type the letters on a keyboard covered by a box to block vision (14 ). Absolute distance error—the unsigned distance between the correct location and the location to which they dragged the key—was 81 mm for typists who imagined the keyboard, 58 mm for typists who saw the keyboard, and 54 mm for typists who typed the letters on a keyboard covered by a box. This experiment allowed us to compare the relative precision of explicit and implicit knowledge of letters on the keyboard. In the explicit judgments, the standard deviation of the distance between the correct and judged location (signed distance error) was 28 mm. To estimate the standard deviation of the distance between correct and judged location in implicit judgments, we assumed that location was represented implicitly as a bivariate normal distribution centered on the key and that the percentage of correct responses reflected the proportion of the distribution that fell within the boundaries of the key. Typists who imagined the keyboard typed 93% of the keystrokes correctly in a typing test, so we assumed that 93% of the distribution fell within the boundaries of the key. We estimated a z score for the radius of the bivariate normal distribution by taking the square root of the 93rd quantile of a chi-square distribution with 2 degrees of freedom. The quantile was 5.2. The square root corresponds to a z score of 2.28, which corresponds to a standard deviation of 4.2 mm for implicit


15

knowledge of key location. Thus, explicit knowledge of keyboard location is (28/4.2¼) 6.7 times less precise than implicit knowledge of keyboard location. This analysis underestimates the difference in precision because it assumes that all typing errors are misplaced keystrokes, and that is not the case. Insertions, deletions, and transpositions are more common (Lessenberry, 1928; Logan, 1999). The standard deviation of implicit knowledge of key location may be much smaller than 4.2 mm.

6.3. The Case for Hierarchical Control, 3.0 Two lines of research provide strong evidence that the outer loop does not know what the inner loop is doing. This informational encapsulation is characteristic of the division of labor in hierarchical control, in which the higher level issues commands and notes that they are executed but does not know the details of how the commands are executed (Beilock & Carr, 2001; Beilock et al., 2002; Fodor, 1983). Together with the evidence for selective influence and limited word-level communication between loops, the evidence for informational encapsulation makes the case for hierarchical control even stronger.

7. The Outer Loop and the Inner Loop Rely on Different Feedback Feedback loops are defined in terms of the goals they are intended to achieve, the operations they carry out in order to achieve them, and the feedback they evaluate to determine whether or not the operations were successful. TOTE units evaluate feedback in the test phase by comparing the goal state and the current state. Thus, feedback can be identified by discovering the goal state and the current states—mental or physical—to which the TOTE is sensitive. Different TOTEs should be sensitive to different feedback. Nested TOTEs should be sensitive to a finer grain of feedback than the superordinate TOTEs in which they are nested. This suggests that the outer and inner loops should be sensitive to different kinds of feedback, and the feedback for the inner loop should be finer-grained than the feedback for the outer loop. We have two lines of evidence supporting this proposition.

7.1. The Inner Loop Relies on the Feel of the Keyboard Crump and Logan (2010c) asked whether the inner loop relied on different feedback than the outer loop, assessing the role of the “feel” of the keyboard in supporting skilled typing (also see Gordon & Soechting, 1995; Terzuolo

16


& Viviani, 1980). Our research was motivated in part by dueling intuitions from our own experience as typists about the role of the keyboard in skilled typing. On the one hand, we find it very difficult to type in the air or on a tabletop without a keyboard to support our typing. This suggests that the feel of the keyboard is essential. On the other hand, we believe that typing is a general skill that can be transferred readily to new keyboards. Otherwise, we would be reluctant to buy new computers or switch between keyboards on desktops and laptops. However, commercial keyboards are very similar, with keys of similar sizes at similar distances in similar layouts. Transfer outside of these familiar parameters may be difficult. Ultimately, the question is empirical, so we designed an experiment to test it. Crump and Logan (2010c) had typists type words on a regular keyboard and on “deconstructed” keyboards that successively removed familiar tactual and proprioceptive cues. First, we removed the keys from a regular keyboard and had typists type on the rubber buttons underneath them. This removes the usual tactual feedback while still providing the resistance or “give” of the regular keyboard, which the rubber buttons provide. Relative to the regular keyboard, the response time slowed by 144 ms (21%), interkeystroke interval slowed by 117 ms/keystroke (75%), and error probability increased by 0.08 (62%). Then, we removed the buttons and had typists type on the flat plastic panel underneath the buttons, in which the circuitry is embedded. This removes the resistance of the keyboard as well as the feel of the keys. Relative to the regular keyboard, the response time slowed by 296 ms (43%), interkeystroke interval slowed by 321 ms/keystroke (207%), and error probability increased by 0.23 (175%). Finally, we tested the typists on a commercially available laser projection keyboard that projected a life-size image of the keyboard on a tabletop. Like the flat keyboard, the laser keyboard removes the feel of the keys and the resistance of the buttons. Relative to the regular keyboard, the response time increased by 323 ms (47%), interkeystroke interval increased by 160 ms/ keystroke (103%), and error probability increased by 0.28 (207%). Similar effects were found in typing paragraphs. Typing speeds were 76, 52, 30, and 43 words per minute on the regular, button, flat, and laser keyboards, respectively. These large disruptions suggest that the feel of the keyboard is an important source of feedback that supports inner-loop performance. As a further test of the importance of feedback from the keyboard, we had 61 typists place their fingers on a blank piece of paper as they would if they were resting on the home row, and we traced the outline of their fingertips. The outline was curved, following the natural contour of the fingertips, and not straight, as it would be if the fingers were resting on the keyboard. The mean discrepancy from a straight line was 12.5 mm, which is about two thirds of the distance between the home row and the top or bottom row. This suggests that the feel of the keyboard is important in maintaining the proper alignment of the fingers on the keys.


17

7.2. The Outer Loop Relies on the Appearance of the Screen An important function of a feedback loop is to detect errors in performance. The processes that implement the operate phase of a TOTE may not always reduce the discrepancy between the current state and the goal state. Sometimes they increase it. In tasks like typewriting, operations that increase the discrepancy between the current state and the goal state produce errors such as typing the wrong letter, omitting a letter that ought to be typed, or typing the right letters in the wrong order (Lessenberry, 1928; F. A. Logan, 1999). Typists must detect these errors and correct them (Long, 1976; Rabbitt, 1978). The two-loop theory of typewriting claims that the outer loop and inner loop process different kinds of feedback, which implies that the two loops detect errors in different ways. The outer loop generates a series of words to be typed and so should monitor the accuracy with which words are typed. We suggest that it monitors the computer screen for the appearance of the intended word. If the intended word appears as it should, the outer loop assumes that it was typed correctly and moves on to the next word. If the intended word does not appear as it should, the outer loop assumes it was an error and asks the inner loop to make the screen look right. The inner loop generates a series of keystrokes and monitors proprioceptive and kinesthetic feedback to ensure that the right fingers moved to the right locations and struck the right keys. If the movements match intentions, typing should remain fast and fluent. If there is a mismatch, typing should slow down or stop. To test these claims, Logan and Crump (2010) had typists type single words and created mismatches between what typists actually typed and what appeared on the screen. We corrected errors that typists made, so the screen matched their intentions but their motor behavior did not, and we inserted errors that typists did not make, so their motor behavior matched their intentions but the screen did not. To measure outer-loop error detection, we had typists report whether or not they typed each word correctly. We assumed that the outer loop monitors the appearance of the screen and so would report corrected errors as correct responses and inserted errors as actual errors, showing cognitive illusions of authorship. Typists confirmed this prediction in an experiment with two alternative posterror responses (correct, error), calling corrected errors “correct” on more than 80% of the trials and inserted errors “error” on more than 70% of the trials, claiming authorship for the appearance of the screen even though it contradicted their motor behavior. In another experiment, we told typists we would correct some errors and insert some errors, and we gave them four alternative posterror responses (correct, error, corrected error, inserted error). We found cognitive illusions of authorship for corrected errors: typists were as likely to call them correct responses as corrected errors. We found no such

18


illusion for inserted errors: typists called inserted errors “inserted errors” as often as they called actual errors “errors.” The inner loop was not susceptible to these cognitive illusions of authorship. We measured inner-loop error detection by assessing posterror slowing. People often slow down on trials after an error in choice response time tasks (Laming, 1968; Rabbitt, 1966) and typists slow down after erroneous keystrokes (Gordon & Soechting, 1995). We found posterror slowing for actual errors and corrected errors but no posterror slowing after inserted errors. Thus, the inner loop knew the truth behind the illusion of authorship. The contrast between explicit error reports and posterror slowing suggests a dissociation between outer-loop and inner-loop error detection. Explicit reports of correct responses occurred both with actual correct responses, which produced no posterror slowing (since there were no errors), and with corrected errors, which produced posterror slowing. Explicit reports of erroneous responses occurred both with actual errors that exhibited posterror slowing and with inserted errors that exhibited no posterror slowing. This dissociation provides further support for our distinction between the outer loop and the inner loop.

7.3. The Case for Hierarchical Control, 4.0 Two lines of research provide strong evidence that the outer loop and inner loop rely on different kinds of feedback. Reliance on different feedback is strong evidence that the two loops engage in different computations, which supports the claim that they are different processes. Together with the evidence for selective influence, limited word-level communication, and informational encapsulation, the evidence for reliance on different feedback makes a strong case for hierarchical control in typewriting. Coupled with more detailed analyses of the computations involved inside the two loops ( John, 1996; Rumelhart & Norman, 1982; Salthouse, 1986; Wu & Liu, 2008), our case for hierarchical control of typewriting is complete and compelling.

8. Beyond Typewriting 8.1. Hierarchical Control in Other Skills Can the case for hierarchical control in typewriting be generalized to other skills? The answer depends on how unique typewriting is in the range of skills humans are capable of performing. Typing is like other skills in that proficiency is attained only after extensive practice. Our typists had 11 years of practice, logging in the 10,000 h necessary for truly expert performance (Ericsson, Krampe, & Tesch-Ro¨mer, 1993). Large amounts of practice may be necessary to develop the autonomy and modularity we see in inner-loop


19

processing. Hierarchical control may be seen in other skills that involve similar amounts of practice. Typewriting may be different from other skills in that it is grafted onto preexisting skills that are already well developed. Children learn to speak around 2 years of age and learn to read around 5 or 6 years of age. Their reading skills are grafted onto well-developed language skills, providing a new input modality. Our survey of typists suggests that children learn to type around 10 years of age, when language skills are quite sophisticated and reading skills are well developed. Typing skill is grafted onto these preexisting skills, providing a new output modality. We believe that this developmental history invites the development of modular typing skill, grafting a new inner loop onto a preexisting outer loop. Hierarchical control may be seen in other skills that graft new inputs or outputs onto preexisting skills. Grafting new skills onto old ones may be sufficient to develop hierarchical control but it may not be necessary. Speech production is controlled hierarchically, although it develops at the same time as language comprehension (Dell, 1986; Levelt et al., 1991). Typewriting may be different from other skills in that it is the result of performance that matters and not the performance itself. The end product of typing is an external text that conveys an intended meaning. The effort involved in producing the product does not matter much as long as the product looks as it should. Thus, the outer-loop processes that generate the intended meaning need not be concerned with the inner-loop processes that translate it into keystrokes. Skill at playing music is different from typewriting in that the performance itself matters more than the plan that generates it. The expressive aspect of music results directly from the nuances of the physical interaction of the musician’s effectors with the instrument. Guitar players evoke emotion with timing, vibrato, and bending and sliding notes ( Juslin, Karlsson, Lindstro¨m, Friberg, & Schoonderwaldt, 2006). Piano players evoke emotion by varying timing and striking the keys gently or robustly (Repp & Knoblich, 2004; Shaffer, Clarke, & Todd, 1985). The outer loop is directly concerned with inner-loop processes to be sure they convey the intended emotion. Thus, musicians may be more aware of what their fingers are doing than typists are. A strong case can be made for hierarchical control in skilled musical performance (Palmer, 1997; Shaffer, 1982), but the inner loop may not be as modular and informationally encapsulated as it is in typewriting.

8.2. Hierarchical Control, Automaticity, Procedural Memory, and Implicit Knowledge For better or for worse, a large amount of psychology is built around binary distinctions (Newell, 1973; Platt, 1964). Indeed, our distinction between the outer loop and the inner loop offers another one. Historically, three binary distinctions have been important in the psychology of skill: automatic and controlled processing (Shiffrin & Schneider, 1977), declarative and procedural

20


memory (Cohen & Squire, 1980), and explicit and implicit knowledge (Roediger, 1990). How do these ideas relate to the idea of hierarchical control? It is tempting to say that the outer loop is controlled and the inner loop is automatic. Indeed, the outer loop controls the inner loop and the inner loop is relatively autonomous. However, we believe there are automatic and controlled components in both loops. Our theory shifts the emphasis from whether processes are controlled to how processes are controlled. Both loops are controlled because they are willfully directed toward goals, but they control different things (relying on different feedback; Crump & Logan, 2010c; Logan & Crump, 2010). Both loops have automatic processes (e.g., lexical activation: Levelt, 1989; finger movements: Gordon et al., 1994), but the processes are controlled in that they serve larger goals (Bargh & Ferguson, 2000) and can be interrupted easily (Logan, 1982; Long, 1976; Rabbitt, 1978; Salthouse & Saults, 1987). Similar considerations apply to the distinctions between declarative and procedural memory, and between explicit and implicit knowledge. The inner loop is a paradigm case of procedural memory, but the language comprehension and generation processes in the outer loop also involve procedural memory (Levelt, 1989). Much of the knowledge in the inner loop is implicit, but it can be made explicit easily (although typists may have to slow down to do so; Logan & Crump, 2009), and much of the knowledge in the outer loop is implicit as well. Indeed, skilled performers often have more explicit knowledge about their skill than novices do (Beilock & Carr, 2001), although that knowledge may not be used directly to control performance. As with controlled and automatic processing, the key question in skilled performance is not whether memory is declarative or procedural or whether knowledge is explicit or implicit, but rather, how declarative and procedural memory, and explicit and implicit knowledge support performance. All the three binary distinctions were developed to address simpler tasks than typewriting, which typically involved single responses to single stimuli. We should not expect distinctions developed for simple tasks to generalize transparently to complex tasks like typewriting. The distinctions may apply to simple components of complex skills, but the components must be organized and coordinated to produce complex behavior, and that may require new concepts and new distinctions like TOTE theory (Miller et al., 1960), two-loop theory (Logan & Crump, 2009), and computational theories of typewriting ( John, 1996; Rumelhart & Norman, 1982; Salthouse, 1986; Wu & Liu, 2008). Complex tasks may require more complex explanations.

8.3. The Development of Hierarchical Control Skills like typewriting are acquired, so the hierarchical control they entail must develop during skill acquisition. It is not clear how this happens. Bryan and Harter (1899) documented plateaus in learning skill at sending and


21

receiving Morse code in telegraphers and argued that other complex skills should show similar plateaus (see LaBerge & Samuels, 1974). Improvements in the speed of sending and receiving messages occurred in steps as telegraphers learned letters, then words, and then phrases. Bryan and Harter interpreted the plateaus as indicating a hierarchy of habits, which implies hierarchical control, so we might see similar plateaus when other hierarchical skills are acquired. However, Keller (1958) reviewed skill acquisition studies that had been published in the years following Bryan and Harter and found little objective evidence for plateaus in the learning curves. Plateaus may be a subjective phenomenon. Newell and Rosenbloom (1981) proposed a theory of skill acquisition for hierarchically structured tasks. They argued that lower-level chunks were smaller than higher-level chunks and would repeat more often in the task environment. Consequently, lower-level chunks should be learned faster than higher-level chunks. MacKay (1982) proposed a theory of skill acquisition based on strengthening connections that made the opposite prediction. He argued that learning was proportional to the difference between current strength and the maximum strength, and that this difference was greater for higher-level structures than for lower-level structures. Thus, there should be greater learning at higher levels of the hierarchy. It is not clear what these theories would predict for typewriting. As we noted earlier, typing skill is grafted onto preexisting language skills, so the higher-level chunks and structures may already be learned before lowerlevel hand and finger skills are acquired. In learning one’s first musical instrument, higher-level structures involving musical phrases, scales, and keys may develop at the same time as lower-level skills that connect one’s effectors to the instrument. Anderson (1982, 1987) proposed a theory of skill acquisition based on the idea of collapsing a series of steps into a single step, which he called “composition.” In his theory, composition can operate at a single level, in which the steps in the series simply drop out (e.g., instead of counting to determine that 2 þ 3 ¼ 5, we simply remember the answer). It can also describe hierarchical skills, in which control of a sequence of actions is passed down from the cognitive system to the motor system. Anderson offers the example of dialing a phone number. As it becomes familiar, we may no longer think of the individual numbers; instead, we think of calling the person and let our fingers take care of the numbers. This proposal is unsatisfactory for understanding skills like typewriting, in which the important thing to explain is the way the motor system manages to execute a sequence of responses. At present, we know for sure that hierarchical skills can be acquired with practice, we are reasonably sure that extensive practice is necessary to develop them, and we have several theories of skill acquisition that may be helpful in understanding how hierarchical control develops. However,

22


we do not yet have a satisfactory explanation. The acquisition of typing skill may be harder to understand nowadays because people begin to acquire typing skill when they are children, so changes in skill are confounded with large changes in development. One strategy is to degrade skill by degrading the input (having skilled typists type nonwords; Crump & Logan, 2010b) or by degrading the output (having skilled typists type on flat keyboards; Crump & Logan, 2010a, 2010c). These degradations may push skilled typists back toward the beginning of the learning curve to a point at which their typing is no longer controlled hierarchically. If so, then training with degraded inputs and outputs may allow us to observe the development of hierarchical control in adult subjects (for a similar strategy for studying the acquisition of skill at arithmetic, see Zbrodoff, 1995).

8.4. Nested Control Loops in Everyday Cognition Nested control loops may be at work in many kinds of cognition. The inner loops take care of immediate goals, while the outer loops ensure that broader goals are satisfied. Our analysis has focused on skills like typewriting that require years of practice to attain proficiency, but nested control loops may be created ad hoc when they are required. A central tenet in theories of cognitive control is that executive processes are flexible, allowing us to address new situations coherently and solve novel problems expeditiously (Logan, 1985; Miller & Cohen, 2001; Monsell, 1996; Shiffrin & Schneider, 1977). The ability to create ad hoc control loops fits well with this tenet. Our use of language provides examples of ad hoc nested control loops. In a conversation, the inner loop may ensure that sentences express the intended meaning, while the outer loop ensures that the point of the conversation gets across. When meeting a new person, the inner loop lets us talk coherently, while the outer loop lets us make a good impression. In negotiation, the inner loop may raise issues and debate points made by the other party, while the outer loop addresses a hidden agenda. The specific goals may be ones we have never addressed before (Chomsky, 1959), yet we accomplish them with some degree of satisfaction. An interesting possibility is that we create outer loops as we need them. Most of the time, we may be “middle management,” carrying out tasks without much regard for higher goals. But when something unusual occurs, like an error, an interruption, or a crisis, we may step back and consider how our activity relates to higher goals (Vallacher & Wegner, 1987). We may not think much about the route home, but an accident on the road ahead may prompt us to create a new plan. We may not think much about survival while writing a chapter, but a fire alarm may revise our priorities. The outer loop may be dormant until we need it, or it may not even exist until we create it on the fly.


23

The idea of nested control loops may even provide some insight into the quest for the meaning of life. We may feel our lives are meaningful when we are pursuing intermediate goals so actively that we do not have time to consider whether we are fulfilling higher-level goals. We may be happier as middle managers (Frankl, 1959). When we reflect on higherlevel goals, we may find nothing of value (“Daddy, I don’t want to go to Europe.” “Shut up and keep swimming.”) or we may invent ad hoc goals to justify our existence. For now, we are happy to work toward the goal of finishing this chapter. What then?!!

ACKNOWLEDGMENTS This research was supported by grants BCS 0646588 and BCS 0957074 from the National Science Foundation.

REFERENCES Anderson, J. R. (1982). Acquisition of cognitive skill. Psychological Review, 89, 369–406. Anderson, J. R. (1987). Skill acquisition: Compilation of weak-method problem solutions. Psychological Review, 94, 192–210. Arrington, C. M., & Logan, G. D. (2004). The cost of a voluntary task switch. Psychological Science, 15, 610–615. Badre, D. (2008). Cognitive control, hierarchy, and the rostro-caudal organization of prefrontal cortex. Trends in Cognitive Sciences, 12, 193–200. Bargh, J. A., & Ferguson, M. J. (2000). Beyond behaviorism: On the automaticity of higher mental processes. Psychological Bulletin, 126, 925–945. Beilock, S. L., & Carr, T. H. (2001). On the fragility of skilled performance: What governs choking under pressure? Journal of Experimental Psychology: General, 130, 701–725. Beilock, S. L., Carr, T. H., MacMahon, C., & Starkes, J. L. (2002). When paying attention becomes counterproductive: Impact of divided versus skill-focused attention on novice and experienced performance of sensorimotor skills. Journal of Experimental Psychology: Applied, 8, 6–16. Botvinick, M., & Plaut, D. C. (2004). Doing without schema hierarchies: A recurrent connectionist approach to normal and routine sequential action. Psychological Review, 111, 395–429. Botvinick, M., & Plaut, D. C. (2006). Such stuff as habits are made on: A reply to Cooper and Shallice (2006). Psychological Review, 113, 917–927. Bryan, W. L., & Harter, N. (1899). Studies of the telegraphic language. The acquisition of a hierarchy of habits. Psychological Review, 6, 345–375. Chomsky, N. (1959). Review of Skinner’s Verbal Behavior. Language, 35, 26–58. Cohen, N. J., & Squire, L. R. (1980). Preserved learning and retention of pattern-analyzing skill in amnesia: Dissociation of knowing how and knowing that. Science, 210, 207–210. Coles, M. G. H. (1989). Modern mind reading: Psychophysiology, physiology, and cognition. Psychophysiology, 26, 251–269. Cooper, R., & Shallice, T. (2000). Contention scheduling and the control of routine activities. Cognitive Neuropsychology, 17, 297–338.

24


Cooper, R. P., & Shallice, T. (2006a). Hierarchical schemas and goals in the control of sequential behavior. Psychological Review, 113, 887–916. Cooper, R. P., & Shallice, T. (2006b). Structured representations in the control of behavior cannot be so easily dismissed: A reply to Botvinick and Plaut (2006). Psychological Review, 113, 929–931. Crump, M. J. C., & Logan, G. D. (2010a). Episodic contributions to sequential control: Learning from a typist’s touch. Journal of Experimental Psychology: Human Perception and Performance, 36, 662–672. Crump, M. J. C., & Logan, G. D. (2010b). Hierarchical control and skilled typing: Evidence for word level control over the execution of individual keystrokes. Journal of Experimental Psychology: Learning, Memory and Cognition, 36, 1369–1380. Crump, M. J. C., & Logan, G. D. (2010c). Warning, this keyboard will de-construct: The role of the keyboard in skilled typewriting. Psychonomic Bulletin & Review, 17, 394–399. Dell, G. S. (1986). A spreading-activation theory of retrieval in sentence production. Psychological Review, 93, 283–321. Ericsson, K. A., Krampe, R. Th., & Tesch-Ro¨mer, C. (1993). The role of deliberate practice in the acquisition of expert performance. Psychological Review, 100, 363–406. Fendrick, P. (1937). Hierarchical skills in typewriting. Journal of Educational Psychology, 28, 609–620. Flanders, M., & Soechting, J. F. (1992). Kinematics of typing: Parallel control of the two hands. Journal of Neurophysiology, 67, 1264–1274. Fodor, J. A. (1983). The modularity of mind. Cambridge, MA: MIT Press. Frankl, V. E. (1959). Man’s search for meaning: An introduction to logotherapy. Boston, MA: Beacon Press. Garner, W. R., Hake, H., & Eriksen, C. W. (1956). Operationism and the concept of perception. Psychological Review, 63, 149–159. Gentner, D. R. (1987). Timing of skilled motor performance: Tests of the proportional duration model. Psychological Review, 94, 255–276. Gentner, D. R., Larochelle, S., & Grudin, J. (1988). Lexical, sublexical, and peripheral effects in skilled typewriting. Cognitive Psychology, 20, 524–548. Gordon, A. M., Casabona, A., & Soechting, J. F. (1994). The learning of novel finger movement sequences. Journal of Neurophysiology, 72, 1596–1610. Gordon, A. M., & Soechting, J. F. (1995). Use of tactile afferent information in sequential finger movements. Experimental Brain Research, 107, 281–292. Hershman, R. L., & Hillix, W. A. (1965). Data processing in typing. Human Factors, 7, 483–492. Jasper, H. H. (1958). The ten-twenty electrode system of the International Federation. Electroencephalography and Clinical Neurophysiology, 10, 371–375. John, B. (1996). TYPIST: A theory of performance in skilled typing. Human-Computer Interaction, 11, 321–355. Juslin, P. N., Karlsson, J., Lindstro¨m, E., Friberg, A., & Schoonderwaldt, E. (2006). Play it again with feeling: Computer feedback in musical communication of emotions. Journal of Experimental Psychology: Applied, 12, 79–95. Keller, F. S. (1958). The phantom plateau. Journal of the Experimental Analysis of Behavior, 1, 1–13. LaBerge, D., & Samuels, S. J. (1974). Toward a theory of automatic information processing in reading. Cognitive Psychology, 6, 293–323. Laming, D. (1968). Information theory of choice-reaction times. London: Academic Press. Lashley, K. S. (1951). The problem of serial order in behavior. In L. A. Jeffress (Ed.), Cerebral mechanisms in behavior (pp. 112–136). New York: Wiley. Lessenberry, D. D. (1928). Analysis of errors. Syracuse, NY: Smith and Corona Typewriters, School Department.


25

Levelt, W. J. M. (1989). Speaking: From intention to articulation. Cambridge, MA: MIT Press. Levelt, W. J. M., Schriefers, H., Vorberg, D., Meyer, A. S., Pechmann, T., & Havinga, J. (1991). The time course of lexical access in speech production: A study of picture naming. Psychological Review, 98, 122–142. Liu, X., Crump, M. J. C., & Logan, G. D. (2010). Do you know where your fingers have been? Explicit knowledge of the spatial layout of the keyboard in skilled typists. Memory & Cognition, 38, 474–484, 138. Logan, G. D. (1982). On the ability to inhibit complex actions: A stop-signal study of typewriting. Journal of Experimental Psychology: Human Perception and Performance, 8, 778–792. Logan, G. D. (1985). Executive control of thought and action. Acta Psychologica, 60, 193–210. Logan, G. D. (1988). Toward an instance theory of automatization. Psychological Review, 95, 492–527. Logan, G. D. (1990). Repetition priming and automaticity: Common underlying mechanisms? Cognitive Psychology, 22, 1–35. Logan, F. A. (1999). Errors in copy typewriting. Journal of Experimental Psychology: Human Perception and Performance, 25, 1760–1773. Logan, G. D. (2003). Simon-type effects: Chronometric evidence for keypress schemata in typewriting. Journal of Experimental Psychology: Human Perception and Performance, 29, 741–757. Logan, G. D., Coles, M. G. H., & Kramer, A. F. (1996). Converging operations in the study of visual selective attention: Introduction. In A. F. Kramer, M. G. H. Coles, & G. D. Logan (Eds.), Converging operations in the study of visual selective attention (pp. xv–xxv). Washington, DC: American Psychological Association. Logan, G. D., & Cowan, W. B. (1984). On the ability to inhibit thought and action: A theory of an act of control. Psychological Review, 91, 295–327. Logan, G. D., & Crump, M. J. C. (2009). The left hand doesn’t know what the right hand is doing: The disruptive effects of attention to the hands in skilled typewriting. Psychological Science, 10, 1296–1300. Logan, G. D., & Crump, M. J. C. (2010). Cognitive illusions of authorship reveal hierarchical error detection in skilled typists. Science, 330, 683–686. Logan, G. D., & Gordon, R. D. (2001). Executive control of visual attention in dual-task situations. Psychological Review, 108, 393–434. Logan, G. D., Miller, A. E., & Strayer, D. L. (2011). Electrophysiological evidence for parallel response selection in skilled typists. Psychological Science, 22, 54–56. Logan, G. D., & Zbrodoff, N. J. (1998). Stroop type interference: Congruity effects in color naming with typewritten responses. Journal of Experimental Psychology: Human Perception and Performance, 24, 978–992. Long, J. (1976). Visual feedback and skilled keying: Differential effects of masking the printed copy and the keyboard. Ergonomics, 19, 93–110. MacKay, D. G. (1982). The problem of flexibility, fluency, and speed-accuracy tradeoff in skilled behavior. Psychological Review, 89, 483–506. Markman, A. B. (1999). Knowledge representation. Mahwah, NJ: Lawrence Erlbaum Associates. McNamara, T. P. (1986). Mental representations of spatial relations. Cognitive Psychology, 18, 87–121. Miller, E. K., & Cohen, J. D. (2001). An integrative theory of prefrontal cortex function. Annual Review of Neuroscience, 24, 167–202. Miller, G. A., Galanter, E., & Pribram, K. (1960). Plans and the structure of behavior. New York: Holt.

26


Monsell, S. (1996). Control of mental processes. In V. Bruce (Ed.), Unsolved mysteries of the mind (pp. 93–148). Hove, England: Erlbaum. Newell, A. (1973). You can’t play 20 questions with nature and win: Projective comments on the papers of this symposium. In W. G. Chase (Ed.), Visual information processing (pp. 283–308). New York: Academic Press. Newell, A., & Rosenbloom, P. S. (1981). Mechanisms of skill acquisition and the law of practice. In J. R. Anderson (Ed.), Cognitive skills and their acquisition (pp. 1–55). Hillsdale, NJ: Erlbaum. Norman, D. A., & Shallice, T. (1986). Attention to action: Willed and automatic control of behavior. In R. J. Davidson, G. E. Schwartz, & D. Shapiro (Eds.), Consciousness and selfregulation: Advances in research and theory (pp. 1–18). New York: Plenum. Novick, L. R., & Hurley, S. M. (2001). To matrix, network, or hierarchy: That is the question. Cognitive Psychology, 42, 158–216. Palmer, C. (1997). Music performance. Annual Review of Psychology, 48, 115–138. Platt, J. R. (1964). Strong inference. Science, 146, 347–353. Rabbitt, P. M. A. (1966). Errors and error correction in choice-response tasks. Journal of Experimental Psychology, 71, 264–272. Rabbitt, P. M. A. (1978). Detection of errors by skilled typists. Ergonomics, 21, 945–958. Rayner, K., & Clifton, C. (2009). Language processing in reading and speech perception is fast and incremental: Implications for event-related potential research. Biological Psychology, 80, 4–9. Repp, B. H., & Knoblich, G. (2004). Perceiving action identity: How pianists recognize their own performances. Psychological Science, 15, 604–609. Roediger, H. L. (1990). Implicit memory: Retention without remembering. American Psychologist, 45, 1043–1056. Rumelhart, D. E., & Norman, D. A. (1982). Simulating a skilled typist: A study of skilled cognitive-motor performance. Cognitive Science, 6, 1–36. Salthouse, T. A. (1986). Perceptual, cognitive, and motoric aspects of transcription typing. Psychological Bulletin, 99, 303–319. Salthouse, T. A., & Saults, J. S. (1987). Multiple spans in transcription typing. Journal of Applied Psychology, 72, 187–196. Schneider, D. W., & Logan, G. D. (2006). Hierarchical control of cognitive processes: Switching tasks in sequences. Journal of Experimental Psychology: General, 135, 623–640. Schneider, D. W., & Logan, G. D. (2007). Retrieving information from a hierarchical plan. Journal of Experimental Psychology: Learning, Memory and Cognition, 33, 1076–1091. Shaffer, L. H. (1973). Latency mechanisms in transcription. In S. Kornblum (Ed.), Attention and performance IV (pp. 435–448). New York: Academic Press. Shaffer, L. H. (1976). Intention and performance. Psychological Review, 83, 375–393. Shaffer, L. H. (1982). Rhythm and timing in skill. Psychological Review, 89, 109–122. Shaffer, L. H., Clarke, E. F., & Todd, N. P. (1985). Metre and rhythm in piano playing. Cognition, 20, 61–77. Shaffer, L. H., & Hardwick, J. (1968). Typing performance as a function of text. Quarterly Journal of Experimental Psychology, 20, 360–369. Shiffrin, R. M., & Schneider, W. (1977). Controlled and automatic human information processing: II. Perceptual learning, automatic attending, and a general theory. Psychological Review, 84, 127–190. Simon, J. R., & Small, A. M., Jr. (1969). Processing auditory information: Interference from an irrelevant cue. Journal of Applied Psychology, 53, 433–435. Soechting, J. F., & Flanders, M. (1992). Organization of sequential typing movements. Journal of Neurophysiology, 67, 1275–1290.


27

Sternberg, S. (1969). The discovery of processing stages: Extensions of Donders’ methods. In W. G. Koster (Ed.), Attention and performance II (pp. 276–315). Amsterdam: North Holland. Sternberg, S., Monsell, S., Knoll, R. L., & Wright, C. E. (1978). The latency and duration of speech and typewriting. In G. E. Stelmach (Ed.), Information processing in motor control and learning. New York: Academic Press. Stevens, A., & Coupe, P. (1978). Distortions in judged spatial relations. Cognitive Psychology, 10, 422–437. Stroop, J. R. (1935). Studies of interference in serial verbal reactions. Journal of Experimental Psychology, 18, 643–662. Terzuolo, C. A., & Viviani, P. (1980). Determinants and characteristics of motor patterns used for typing. Neuroscience, 5, 1085–1103. Vallacher, R. R., & Wegner, D. M. (1987). What do people think they’re doing? Action identification and human behavior. Psychological Review, 94, 3–15. Viviani, P., & Laissard, G. (1996). Motor templates in typing. Journal of Experimental Psychology: Human Perception and Performance, 22, 417–445. Wegner, D. M. (2002). The illusion of conscious will. Cambridge, MA: MIT Press. Wu, C., & Liu, Y. (2008). Cuing network modeling of transcription typing. ACM Transactions on Computer-Human Interaction, 15, 1–45, (Article 6). Zbrodoff, N. J. (1995). Why is 9 þ 7 harder than 2 þ 3? Strength and interference as explanations of the problem-size effect. Memory & Cognition, 23, 689–700.


C H A P T E R

T W O

Cognitive Distraction While Multitasking in the Automobile David L. Strayer, Jason M. Watson, and Frank A. Drews Contents 1. A Framework for Understanding the Sources of Driver Distraction 2. Do Cell-Phone Conversations Increase the Crash Risk? 3. Why Does Talking on a Cell Phone Impair Driving? 4. Are All Conversations Harmful to Driving? 5. Can the Interference Be Practiced Away? 6. Is Everyone Impaired by Using a Cell Phone While Driving? 7. Conclusions and Future Directions References

30 33 42 47 49 50 55 56

Abstract Driver distraction is a significant source of motor-vehicle accidents. This chapter begins by presenting a framework for conceptualizing the different sources of driver distraction associated with multitasking. Thereafter, the primary focus is on cognitive sources of distraction stemming from the use of a cell phone while driving. We present converging evidence establishing that concurrent cell phone use significantly increases the risk of a motor-vehicle accident. Next, we show that using a cell phone induces a form of inattention blindness, where drivers fail to notice information directly in their line of sight. Whereas cell-phone use increases the crash risk, we show that passenger conversations do not. We also show that real-world cell-phone interference cannot be practiced away and conclude by considering individual differences in multitasking ability. Although the vast majority of individuals cannot perform this dual-task combination without impairment, a small group of “supertaskers” can, and we discuss the neural regions that support this ability.

Most of the time we take driving for granted. But operating an automobile is the riskiest activity that most readers of this chapter engage in on a regular basis. In fact, motor-vehicle crashes were the leading cause of accidental deaths in the US in 2008 and are the leading cause of all deaths for people between the age of 1 and 35. The National Safety Council White Paper (2010) recently noted that driver distraction had joined alcohol and Psychology of Learning and Motivation, Volume 54 ISSN 0079-7421, DOI: 10.1016/B978-0-12-385527-5.00002-4

#


29

30

David L. Strayer et al.

speeding as leading factors in fatal and serious injury crashes. In this chapter, we will focus on driver distraction and some of the underlying causes that contribute to driving impairment. There are indeed many sources of driver distraction. Some “old standards” include talking to passengers, eating, drinking, lighting a cigarette, shaving, applying makeup, and listening to the radio (Stutts et al., 2003). However, the last decade has seen an explosion of new wireless “nomadic” devices that have made their way into the automobile, enabling a host of new sources of driver distraction (e.g., sending and receiving e-mail or text messages, communicating via cellular device, watching video movies, using the internet, etc.). It is likely that these new sources of distraction are more impairing than the old standards because they are more cognitively engaging and are often performed over more sustained periods of time. The primary focus of this chapter is on how driving is impacted by cellular communication (i.e., talking on a cell phone), because this is one of the most prevalent exemplars of this new class of multitasking activity. In fact, in 2010, the NSC estimated that the 28% of all crashes on the roadway were caused by the use of a cell phone to talk, dial, or text while driving (National Safety Council White Paper, 2010). Our chapter begins with a theoretical framework for understanding the different sources of driver distraction. Thereafter, our main focus will be on cognitive sources of distraction, with cell-phone use as the primary exemplar of this type of interference. Next, we will review evidence from our laboratory and elsewhere that establishes that driving is impaired with the concurrent use of a cell phone. Understanding why cell phones impair driving is important, and we will show that the use of a cell phone induces a form of inattention blindness, causing the drivers to fail to see critical information in their field of view. We also consider whether all forms of verbal communication impair driving and whether a driver can become sufficiently skilled at using a cell phone that they are no longer impaired by this activity (the answer to both questions is “NO”). Finally, we examine individual differences in this multitasking behavior. We will show that the majority of individuals suffer significant impairment when they use a cell phone while driving. However, there is a small percentage of individuals who have extraordinary multitasking ability and do not exhibit interference in the cell phone/driving dual-task combination. We show that these “supertaskers” exhibit a generalizable ability to multitask and present neuroimaging data that establish that frontal brain regions support this extraordinary ability.

1. A Framework for Understanding the Sources of Driver Distraction Figure 1 presents a framework for discussing the sources of driver distraction. Impairments to driving can arise from a competition for visual processing, wherein the driver takes their eyes off the road to interact with a

31

Multi-Tasking

Visual

High Moderate

Cognitive

Low

Manual

Figure 1 A framework for conceptualizing the sources of driver distraction.

device. Impairments can also arise from manual interference, as in cases where drivers take their hands off the steering wheel to manipulate a device. Finally, cognitive sources of distraction occur when attention is withdrawn from the processing of information necessary for the safe operation of a motor vehicle. These three sources of distraction can operate independently; that is, interacting with different devices can result in competition from one, two, or all the three sources. Figure 1 illustrates three hypothetical multitasking situations. The small blue inner circle represents a situation in which the driver engages in a concurrent activity that places low levels of demand on the visual, manual, or cognitive resources. An activity such as listening to a preprogrammed radio station at normal volume would be an example of low demand, in that it places little or no demand on visual, manual, or cognitive processing resources. The middle circle represents a situation in which the driver engages in a concurrent task that places moderate levels of demand on visual, manual, and cognitive resources. The outer circle represents situations in which the driver engages in a concurrent task that places high levels of demand on visual, manual, and cognitive resources. An example of this high level of interference might involve a driver using a touchscreen device to access information on the internet (e.g., a recent case we reviewed involved a younger driver who was killed when his vehicle collided into a semitractor trailer while he was manipulating information on his Facebook page using his cell phone). This interaction placed heavy demands on visual, manual, and cognitive resources, and activities such as these will inevitably end in a bad outcome. Holding other factors constant, the crash risk is higher for multitasking activities in the outer circle than for multitasking activities in the inner circle. There are two additional factors that are important to consider in discussions concerning driver distraction and crash risk. The first factor is

32


the duration of an activity that is concurrently performed while driving. In many instances, drivers attempt to multitask when they perceive the demands of driving to be low (e.g., while stopped at a traffic light). But, as the duration of interaction with a device increases, the ability of a driver to accurately predict the driving demands decreases. For example, changing a radio station may place demands on visual and manual resources, but the duration of that impairment is relatively short (e.g., 5 s or so). By contrast, a cell-phone conversation may extend for several minutes, and the conditions that were in effect at the beginning of a call may change considerably over this interval. In general, dual-tasking activities that tie up resources for longer periods of time will create greater cumulative impairments than activities with shorter durations. The second factor to consider is the exposure rate of an activity. The more drivers that engage in a distracting activity, the greater the impact to public safety. For example, below we will demonstrate that the risk of being in a motor-vehicle accident increases by a factor of 4 when drivers are talking on a cell phone. What compounds the risk to public safety is that at any daylight hour it is estimated that over 10% of drivers on US roadways talk on their cell phone (Glassbrenner, 2005). While there are many activities engaged in while driving that are associated with an equal or higher crash risk, few if any have the same exposure as using a cell phone. For the remainder of this chapter, we will examine cognitive sources of distraction, with a particular focus on the role that cell phones play in driver distraction. The cell phone is a relatively modern invention that has been in common use for less than 20 years. Over this period, use has skyrocketed, and as of 2010, more than 90% of the US population now carries a cell phone. Using a cell phone while driving has become commonplace, with 85% of drivers reporting that they use a cell phone while concurrently operating a motor vehicle (National Highway Transportation Safety Administration, 2006). And, as mentioned above, current estimates suggest that at any time during the day, more than 10% of drivers on the roadway talk on their cell phone. Even more alarming is that 2 out of 10 drivers who use a cell phone report that they have bumped into a person or object because they were distracted (Pew Internet and American Life Project, 2010). From a theoretical perspective, understanding the mechanisms underlying dual-task performance has been an important endeavor in psychology for over 60 years, and certain patterns of interference may prove useful for evaluating cognitive theory. In fact, several of the findings we discuss below prove challenging for current theories of attention and dual-task processing (e.g., see Strayer & Drews, 2007; Watson & Strayer, 2010). From an applied perspective, this issue is important as legislators attempt to craft legislation that addresses the safety concerns associated with multitasking. For example, at least six US States now have regulations that prohibit the use of hand-held cell phones while driving but permit the use of hands-free devices

33

Multi-Tasking

(IIHS, 2010). Implicit in this regulation is the assumption that a major source of the interference stems from the manual manipulation of the phone (i.e., holding the phone to listen and talk). We will see that this assumption is not supported by the empirical research.

2. Do Cell-Phone Conversations Increase the Crash Risk? There are several methodologies that have been used to address this question. Each methodology has strengths and weaknesses. Converging evidence from the different techniques provides a definitive answer to the question (“YES”). The simplest method uses naturalistic observations to see how their driving behavior is altered with the concurrent use of a cell phone to dial, talk, or text. In one such study, we observed over 1700 drivers as they approached a residential intersection with four-way stop signs. We determined through observation whether the drivers were or were not using their cell phone as they approached the intersection and whether they came to a complete stop (as required by law) before proceeding through the intersection.1 The resulting data are presented in Table 1. For drivers not using a cell phone, the majority stopped in accordance with traffic laws. By contrast, for the drivers who were observed talking on their cell phone as they approached the intersection, the majority failed to stop in accordance with traffic laws. For drivers not using a cell phone, the odds ratio for failing to stop was 0.27, whereas the odds ratio for failing to stop for drivers who were using their cell phone was 2.93. This 10-fold increase in failing to stop was significant (w2(1) ¼ 129.8, p < 0.01). Table 1 Frequency Totals for the 2 (Cell Phone in Use Vs. Cell Phone Not in Use) 2 (Stopping Violation Vs. No Violation) Observational Study of Four-Way Stop Sign Compliance.

On cell Not on cell

1

Stopping violation

No violation

82 352 434

28 1286 1314

110 1638 1748

This simple observational study has now become a standard used in research method courses at the University of Utah. It is a sure-fired way to ensure that students get significant and meaningful data that can be used for their class writing assignments (see www.psych.utah.edu/cellphonestudy/).

34


Observational studies have a high validity. After all, it is real driving and if a cell phone is in use, it is a real conversation. But one important limitation of the observational approach is that it cannot establish a causal link between the use of a cell phone and impaired driving. For example, it is possible that those drivers who regularly use a cell phone are willing to engage in more risky activities and that this increase in risk taking also leads drivers to engage in more risky driving behaviors such as running stop signs. Epidemiological studies provide another method for assessing the crash risk associated with using a cell phone while driving. Redelmeier and Tibshirani (1997) obtained the cell phone records of 699 drivers who were involved in a noninjury motor-vehicle collision. They used a casecrossover design in which the same driver was evaluated to see whether they were using a cell phone at several comparison intervals (e.g., same day of the week). The authors found that the odds of a crash were over four times higher when drivers were using their cell phone. McEvoy et al. (2005) replicated this procedure, but instead used crashes that required the driver to be transported to a hospital for medical care. Similar to Redelmeier and Tibshirani (1997), the odds of crashing were over four times higher when drivers were using their cell phone. As with observational studies, epidemiological studies have high face validity and establish a real-world association between use of a cell phone and crashes. However, like observational studies, this method does not establish a causal link between cell-phone use and crashes. Note that establishing a causal link between driving impairment and the concurrent use of a cell-phone is important if the research is to advance our theoretical understanding of driver distraction. The final method that we consider in detail involves the use of highfidelity driving simulators to establish a causal relationship between the use of a cell phone and driving impairment. Figure 2 shows a participant using our driving simulator. The simulator is composed of five networked microprocessors and three high-resolution displays providing a 180 field of view. It incorporates proprietary vehicle dynamics, traffic scenario, and road surface software to provide realistic scenes and traffic conditions. The dashboard instrumentation, steering wheel, gas, and brake pedal were taken from a Ford Crown VictoriaÒ sedan with an automatic transmission. For the majority of our studies, the simulator used a freeway road database simulating a 24-mile multilane highway with on- and off-ramps, overpasses, and two- and three-lane traffic in each direction. Our first simulator study used a car-following paradigm to determine how driving performance is altered by conversations over a cell phone. The participant’s task was to follow a periodically braking pace car that was driving in the right-hand lane of the highway. When the participant stepped on the brake pedal in response to the braking pace car, the pace car released its brake and accelerated to normal highway speed. If the participant failed

Multi-Tasking

35

Figure 2 A participant driving in the Patrol-Sim driving simulator.

to depress the brake, they would eventually collide with the pace car. That is, like real highway stop and go traffic, the participant was required to react in a timely and appropriate manner to vehicles slowing in front of them. Car following is an important requirement for the safe operation of a motor vehicle. In fact, failures in car following account for 30% of policereported accidents (e.g., National Highway Transportation Safety Administration, 2001). In our study, the performance of a nondistracted driver was contrasted with the performance of that same driver when they were conversing on either a hand-held or hands-free cell phone. We were particularly interested in examining the differences in driving performance of the hand-held cell-phone driver with that of the hands-free cell-phone driver, because six US States currently prohibit the former while allowing the latter form of cellular communication. To preview, our analyses will show that the performance of drivers engaged in a cell-phone conversation differs significantly from that of the nondistracted driver and that there is no safety advantage for hands-free over hand-held cell phones. Figure 3 presents a typical sequence of events in the car-following paradigm. Initially, both the participant’s car (solid line) and the pace car (long-dashed line) were driving at about 62 MPH with a following distance of 40 m (dotted line). At some point in the sequence, the pace car’s brake lights illuminated for 750 ms (short-dashed line) and the pace car began to decelerate at a steady rate. As the pace car decelerated, following distance decreased. At a later point in time, the participant responded to the decelerating pace car by pressing the brake pedal. The time interval between the onset of the pace car’s brake lights and the onset of the participant’s brake response defines the brake reaction time. Once the participant depressed the brake, the pace car began to accelerate at which point the participant removed his foot from the brake and applied pressure to the gas pedal.

36


100

80 Pace car’s speed (MPH)

60 Driver’s speed (MPH) 40 Following distance (m) 20 Pace car’s brake lights

Driver’s brake response

Time (s)

Figure 3

A typical sequence of events in the car-following paradigm.

Note that in this example, following distance decreased by about 50% during the braking event. Here, we report three parameters associated with the participant’s reaction to the braking pace car. Brake reaction time is the time interval between the onset of the pace car’s brake lights and the onset of the participant’s braking response (i.e., a 1% depression of the brake pedal). Following distance is the distance between the rear bumper of the pace car and the front bumper of the participant’s car. Speed is the average driving speed of the participant’s vehicle. Figure 4 presents the brake reaction time Vincentized cumulative distribution functions (CFFs) as participants reacted to the pace car’s brake lights. In Figure 4, the reaction time at each decile of the distribution is plotted, and it is evident that the functions for the hand-held and hands-free cellphone conditions are displaced to the right, indicating slower reactions, compared to the single-task condition. Analysis indicated that RT in each of the dual-task conditions differed significantly from the single-task condition at each decile of the distribution, whereas the distributions for hand-held and hands-free conditions did not differ significantly across the deciles. A companion analysis of median brake reaction time found that braking reactions were significantly slower in dual-task conditions than in singletask conditions, F(2,78) ¼ 13.0, p < 0.01. Subsidiary pair-wise t-tests indicated that the single-task condition differed significantly from the hand-held

37

Multi-Tasking

Cumulative distribution functions for car following task

Distribution decile

100

80

60

40 Baseline (no distraction) Hand-held cell phone Hands free cell phone

20

0 500

600

700

800 900 1000 Reaction time (ms)

1100

1200

1300

Figure 4 RT Cumulative Frequency Distributions (CDFs) for the single-task baseline condition and the hand-held and hands-free dual-task cell-phone conditions.

and hands-free cell-phone conditions, and the difference between handheld and hands-free conditions was not significant. In order to better understand the changes in driving performance with cell-phone use, we examined driver performance profiles in response to the braking pace car. Driving profiles were created by extracting 10 s epochs of driving performance that were time-locked to the onset of the pace car’s brake lights. That is, each time that the pace car’s brake lights were illuminated, the data for the ensuing 10 s were extracted and entered into a 32 300 data matrix (i.e., on the jth occasion that the pace car brake lights were illuminated, data from the 1st, 2nd, 3rd, . . ., and 300th observations following the onset of the pace car’s brake lights were entered into the matrix X[ j,1], X[ j,2], X[ j,3], . . ., X[ j,300]; where j ranges from 1 to 32 reflecting the 32 occasions in which the participant reacted to the braking pace car). Each driving profile was created by averaging across j for each of the 300 time points. Figure 5 presents the average driving speed profile, time-locked to the onset of the pace car’s brake lights, for the three conditions in the study. Over the 10-s epoch, participants in the single-task condition drove at a faster rate of speed than when they were conversing on a cell phone, F(2,78) ¼ 3.3, p < 0.05; however, vehicle speed during the prebraking interval did not differ significantly between conditions. Driving speed

38


Speed profile 58

Speed (MPH)

56

54

52

50

Baseline Hand-held cell phone Hands free cell phone

48 0

1

2

3

4

5 6 Time (s)

7

8

9

10

Figure 5 The driving speed profile plotted as a function of time. The single-task baseline condition is presented with the hand-held and hands-free dual-task cellphone conditions.

reached the nadir between 2 and 3 s after the onset of the pace car’s brake lights whereupon the participant’s vehicle reaccelerated toward prebraking speed. The difference in overall speed was primarily determined by the time it took participants to recover the speed lost during braking. In particular, the time that it took participants to recover 50% of the speed lost during the braking episode was significantly shorter in the single-task condition than the hand-held or the hands-free cell-phone conditions, F(2,78) ¼ 4.4, p < 0.01. Subsidiary pair-wise t-tests indicated that single-task recovery was significantly faster than either the hand-held or the hands-free cellphone conditions and that the rate of recovery time did not differ for the two cell-phone conditions. This sluggish behavior appears to be a key characteristic of the driver distracted by a cell-phone conversation, and such a pattern of driving is likely to have an adverse impact on the overall flow of dense highway traffic (see Cooper, Vladisavljevic, Medeiros-Ward, Martin, & Strayer, 2009). Figure 6 cross-plots driving speed and following distance to illustrate the relationship between these two variables over the braking episode. In the figure, the beginning of the epoch is indicated by a left-pointing arrow, and the relevant symbol (circle, triangle, or square) is plotted every third of a second in the time series. The distance between the symbols provides an indication of how each function changes over time (i.e., on a

39

Multi-Tasking

Speed × following distance

56

Speed (MPH)

54

52

50 Baseline (no distraction) Hand-held cell phone Hands free cell phone

48 24

25

26 27 28 Following distance (m)

29

30

Figure 6 A cross-plot of driving speed and following distance plotted as a function of time. The single-task baseline condition is presented with the hand-held and hands-free dual-task cell-phone conditions.

given function, symbols closer together indicate a slower change over time than symbols farther apart). The figure clearly illustrates that the relationship between driving speed and following distance is virtually identical for the driver distracted by either a hand-held or hands-free cell phone. By contrast, the performance of the participant in single-task conditions provides a qualitatively different pattern than what is seen in the dual-task conditions. In particular, the functions representing the dual-task conditions are displaced toward the lower right quadrant, indicative of a driver operating the vehicle more conservatively (i.e., somewhat slower and with a greater following distance from the pace car) than in single-task conditions. Figure 6 also illustrates the dynamic stability of driving performance following a braking episode. From a dynamic systems perspective, driving performance in single- and dual-task conditions can be characterized as operating in different speed-following distance basins of attraction with performance returning to equilibrium following each braking perturbation. Note also that the curves in Figure 6 for the nondistracted driver and the driver conversing on a cell phone did not intersect. This suggests that the basin of attraction created with either the hand-held or hands-free cellphone conversations was sufficiently “deep” that participants returned to their respective prebraking set points after a braking episode had perturbed their position in the speed/following-distance space.

40


Taken together, the data demonstrate that conversing on a cell phone impaired driving performance and that the distracting effects of cell-phone conversations were equivalent for hand-held and hands-free devices. Compared to single-task conditions, cell-phone drivers’ brake reaction times were slower and they took longer to recover the speed that was lost following braking. The cross-plot of speed and following distance showed that drivers conversing on a cell phone tended to have a more cautious driving profile, which may be indicative of a compensatory strategy to counteract the delayed brake reaction time. Elsewhere, Brown, Lee, & McGehee (2001) found that the sluggish brake reactions, such as the ones described herein, can increase the likelihood and severity of motor-vehicle collisions. Another way to evaluate these risks is by comparison with other activities commonly engaged in while driving (e.g., listening to the radio, talking to a passenger in the car, etc.). The benchmark that we used in our second study was driving while intoxicated from ethanol at the legal limit (0.08 wt/vol). We selected this benchmark because the epidemiological study by Redelmeier and Tibshirani (1997) noted that “the relative risk [of being in a traffic accident while using a cell phone] is similar to the hazard associated with driving with a blood alcohol level at the legal limit” (p. 465). If this claim can be substantiated in a controlled laboratory experiment, then these data would be of immense importance for public safety. In particular, the World Health Organization recommended that the behavioral effects of an activity should be compared to alcohol under the assumption that performance should be no worse than when operating a motor vehicle at the legal limit (Willette & Walsh, 1983). How does conversing on a cell phone compare with the drunk-driving benchmark? Here, we directly compared the performance of 40 drivers who were conversing on a cell phone with the performance of these same drivers who were legally intoxicated with ethanol. Three counterbalanced conditions were studied: single-task driving (baseline condition), driving while conversing on a cell phone (cell-phone condition), and driving with a blood alcohol concentration of 0.08 wt/vol (alcohol condition, verified using an Intoxilyzer 5000). Table 2 presents nine performance variables that were measured to determine how participants reacted to the vehicle braking in front of them. Three of the variables (brake reaction time, speed, and following distance) were used in our first study. We also added several new variables to provide a more fine-grained comparison between drunk driving and cell-phone conditions.2 Braking force is the maximum force that the participant applied to the brake pedal in response to the braking pace car. SD following distance is 2

These additional parameters did not differ between the hand-held and hands-free cell phone conditions in the first study.

41

Multi-Tasking

Table 2 Driving Performance Measures Obtained in the Alcohol, Baseline, and Cell-Phone Driving Conditions.

Total accidents Brake reaction time (ms) Speed (MPH) Following distance (m) Maximum braking force percentage of max SD following distance (m) Time to collision (s) Time to collision < 4 s Half-recovery time (s)

Alcohol

Baseline

Cell phone

0 779 (33) 52.8 (2.0) 26.0 (1.7) 69.8 (3.7) 10.3 (0.6) 8.0 (0.4) 3.0 (0.7) 5.4 (0.3)

0 777 (33) 55.5 (0.7) 27.4 (1.3) 56.7 (2.6) 9.5 (0.5) 8.5 (0.3) 1.5 (0.3) 5.3 (0.3)

3 849 (36) 53.8 (1.3) 28.4 (1.7) 55.5 (3.0) 11.8 (0.8) 8.1 (0.4) 1.9 (0.5) 6.3 (0.4)

the standard deviation of following distance. Time to collision (TTC), measured at the onset of the participant’s braking response, is the time that remains until a collision between the participant’s vehicle and the pace car if the course and speed were maintained (i.e., had the participant failed to brake). Also reported is the frequency of trials with TTC values below 4 s, a level found to discriminate between cases where the drivers find themselves in dangerous situations from cases where the driver remains in control of the vehicle (e.g., Hirst & Graham, 1997). Half-recovery time is the time for participants to recover 50% of the speed that was lost during braking. Also shown in the table is the total number of collisions in each phase of the study. We used a multivariate analysis of variance (MANOVA) followed by planned contrasts to provide an overall assessment of driver performance in each of the experimental conditions. MANOVAs indicated that both cell phone and alcohol conditions differed significantly from single-task baseline (F(8,32) ¼ 6.26, p < 0.01 and F(8,32) ¼ 2.73, p < 0.05, respectively). When drivers were conversing on a cell phone, they were involved in more rear-end collisions, their initial reaction to vehicles braking in front of them was slowed, and the variability in following distance increased. In addition, compared to the single-task baseline, it took participants who were talking on a cell phone longer to recover the speed that was lost during braking. By contrast, when participants were intoxicated, neither accident rates nor reaction time to vehicles braking in front of the participant nor recovery of lost speed following braking differed significantly from single-task baseline. Overall, drivers in the alcohol condition exhibited a more aggressive driving style. They followed closer to the pace vehicle and braked with more force than in the single-task baseline condition. Unexpectedly, our study found that accident rates in the alcohol condition did not differ

42


from baseline; however, the increase in hard braking is predictive of increased accident rates over the long run (e.g., Brown et al., 2001; Hirst & Graham, 1997). The MANOVA also indicated that the cell-phone and alcohol conditions differed significantly from each other, F(8,32) ¼ 4.06, p < 0.01. When drivers were conversing on a cell phone, they were involved in more rear-end collisions and took longer to recover the speed that they had lost during braking than when they were intoxicated. Drivers in the alcohol condition also applied greater braking pressure than drivers in the cellphone condition. Finally, the accident data indicated that there were significantly more accidents when participants were conversing on a cell phone than in the single-task baseline or alcohol conditions. w2(2) ¼ 6.15, p < 0.05. Taken together, we found that both intoxicated drivers and cell-phone drivers performed differently from the single-task baseline and that the driving profiles of these two conditions differed. Drivers using a cell phone exhibited a delay in their response to events in the driving scenario and were more likely to be involved in a traffic accident. Drivers in the alcohol condition exhibited a more aggressive driving style, following closer to the vehicle immediately in front of them, necessitating braking with greater force. With respect to traffic safety, the data suggest that when controlling for driving conditions and time on task, the impairments associated with cell-phone drivers may be as great as those commonly observed with intoxicated drivers.

3. Why Does Talking on a Cell Phone Impair Driving? The epidemiological studies establish that talking on a cell phone while driving increases the crash risk by a factor of 4. Moreover, several lines of evidence suggest that the crash risk is the same for hand-held and hands-free cell phones. For example, simulator-based studies reviewed above found that hands-free cell phones had the same impairment profile as that of hand-held devices. In addition, a recent analysis from the Highway Loss Data Institute compared US States that imposed a ban on driving while using a hand-held cell phone with comparable States that did not institute a ban and found no safety advantage for prohibiting hand-held cell phones (HLDI, 2009). Given that hands-free cell phones produce the same level of impairment as held-held units, it suggests that the source of interference is cognitive in nature. This follows because hands-free cell phones allow drivers to have their eyes on the road (i.e., little or no visual interference) and their hands

43

Multi-Tasking

Conditional probability of recognition

on the wheel (i.e., little or no manual interference). We have suggested that using a cell phone induces a form of inattention blindness whereby the cellphone conversation diverts attention from processing the information necessary to safely operate a motor vehicle (Strayer & Drews, 2007; Strayer, Drews, & Johnston, 2003; Strayer & Johnston, 2001). To test the inattention blindness hypothesis, we examined how cellphone conversations affect the driver’s attention to objects that are encountered while driving. The study used an incidental recognition memory paradigm to assess what information in the driving scene participants attended while driving. The procedure required participants to perform the driving task without the foreknowledge that their memory for objects in the driving scene would be tested. Later, the participant was given a surprise recognition memory task in which they were shown objects that were presented while they were driving and were asked to discriminate these objects from foils that were not in the driving scene. Differences in recognition memory between single- and dual-task conditions provide an estimate of the degree to which attention to visual information in the driving environment is distracted by cell-phone conversations. In this study, we also monitored eye fixation using an Applied Science Laboratories mobile 501 eye-tracker that allowed a free range of head and eye movements, thereby affording naturalistic viewing conditions for the participants as they negotiated the driving environment. Figure 7 presents the conditional probability of recognizing an object in the driving scene given that participants fixated on it while driving. This analysis specifically tests for memory of objects that were presented where the driver’s eyes were directed. That is, based on the eye tracking data, we

0.6 0.5 0.4 0.3 0.2 0.1 0.0 Dual task Single task Driving condition

Figure 7

Recognition memory in the single-task and dual-task conditions.

44


know that the driver’s eyes were on the road (directed at objects in the driving environment). Moreover, because we used a hands-free cell phone and the call was initiated before driving began, there was no manual interference when drivers were talking on the phone. Thus, any interference that is observed can be attributed entirely to cognitive interference. We restricted our analysis to objects that were fixated upon during the drive. In addition, we used hierarchical linear regression to statistically control for any differences in fixation duration. The analysis revealed that participants were more than twice as likely to recognize objects encountered in the single-task condition than in the dual-task condition, t(19) ¼ 4.53, p < 0.01. That is, when we ensured that participants fixated on objects in the driving scene, significant differences in recognition memory between single- and dual-task conditions were found. Even when the participant’s eyes were directed at objects in the driving environment for the same duration, they were less likely to remember them if they were conversing on a cellular phone. In a follow-up study, we asked participants to rate the objects in the driving scene in terms of their relevance to safe driving using a 10-point scale (participants were given an example in which a child playing near the road might receive a rating of 9 or 10, whereas a sign documenting that a volunteer group cleans a particular section of the highway might receive a rating of 1). Safety relevance ratings ranged from 1.5 to 8, with an average of 4.1. A series of regression analyses found that traffic relevance had absolutely no effect on the difference in recognition memory between single-task and dual-task conditions. This finding is important because it establishes that drivers do not strategically reallocate attention from the processing of less relevant information in the driving scene to the cell-phone conversation while continuing to give highest priority to the processing of task-relevant information in the driving scene. Figure 8 illustrates how the driving environment might be perceived by a driver who is not talking on a cell phone (panel A) and for that same driver when they are talking on a cell phone (panel B). In this example, the encoding of important objects (e.g., the flagman and the bicyclist) is impaired by the use of a cell phone. In fact, we have reviewed several real-world crashes where drivers report failing to see critical information such as stop signs and pedestrians that result in motor-vehicle accidents. Thus far, our studies assessing inattention blindness have relied on explicit memory measures to test the hypothesis that cell-phone conversations interfere with the initial encoding of information in the driving scene. However, an alternative possibility is that there are no differences in the initial encoding, but rather differences in the retrieval of the information during subsequent memory tests. This distinction is important because the former has direct implications for traffic safety, whereas the latter does not. To differentiate between encoding and retrieval deficits, we recorded

Multi-Tasking

45

Figure 8 A representation of what a driver might perceive when they are not talking on the phone (left panel) and when they are talking on a hands-free cell phone (right panel).

on-line measures of brain activity elicited by braking events in the driving environment. Prior research has found that the amplitude of the P300 component of the event-related brain potential (ERP) is sensitive to initial encoding conditions and that memory performance is superior for objects eliciting larger amplitude P300s during initial encoding (e.g., Fabiani, Karis, & Donchin, 1986; Otton & Donchin, 2000). We asked participants to follow a pace car that would brake at random intervals, and ERPs were time-locked to the onset of the pace car brake lights in both single- and dual-task conditions. The dual-task condition involved talking to a confederate on a hands-free cell phone. If the impairments in memory performance are due to differences in the initial encoding of objects in the driving scene, then P300 amplitude should be smaller in the dual-task condition than in the single-task condition. By contrast, if the memory differences are due to impaired retrieval of information at the time of the recognition memory test but not at the time of encoding, then we would not expect to find differences in P300 amplitude between the single-task and the dual-task conditions. The average ERPs recorded at the parietal electrode site are presented in Figure 9. Visual inspection reveals a large positive potential between 250 and 750 ms (the P300 component of the ERP). Statistical analysis indicated that the P300 component of the ERPs was significantly larger in the singletask than in the dual-task condition, t(15) ¼ 4.41, p < 0.01. In fact, P300 amplitude was reduced by 50% when the drivers were talking on the cell phone. These ERP data provide strong evidence for the inattention-blindness hypothesis. In particular, the brain activity associated with processing the information necessary for the safe operation of a motor vehicle is suppressed when drivers talk on a cell phone. Thus, drivers using a cell phone fail to see information in the driving scene because they do not encode it as well as

46

Amplitude (µV)


_ 5 µV Single task Dual task

+

–200

0

200

400

600

800

1000

1200

Time (ms)

Figure 9 Event-related brain potentials elicited by the onset of brake lights in singletask and dual-task conditions.

they do when they are not distracted by the cell-phone conversation. In situations where the driver is required to react with alacrity, these data suggest that those drivers using a cell phone will be less able to do so because of the diversion of attention from driving to the phone conversation. It is important to note that the demonstrations of inattention blindness described herein provide a pure measure of cognitive interference, because the participant’s eyes were on the road and they were not manually manipulating the phone in dual-task conditions. The studies assessing the inattention-blindness hypothesis tested memory for objects that were at fixation, ensuring that participants actually looked at objects in the driving scene. However, cell phones can also induce a form of tunnel vision, whereby drivers tend to direct their gaze directly ahead and tend to look less often in the periphery. The consequence of this tendency to fixate centrally is that drivers talking on a cell phone are less likely to see objects in the periphery (pedestrians, cars, roadside hazards) and make fewer glances at traffic signals at intersections (Harlbluk, Noy, Trbovich, & Eizenman, 2007). Alarmingly, some drivers talking on a cell phone do not even look at the traffic signals! In an unpublished study, Noy (2009) recorded eye movements in an instrumented vehicle when drivers were and were not talking on a handsfree cell phone. Figure 10 provides a visual illustration of the areas scanned by the driver as they operated a motor vehicle. The left panel illustrates visual scanning under normal conditions and the right panel illustrates visual scanning when drivers were talking on a hands-free cell phone. In this example, the driver talking on a cell phone would fail to see the bicyclist

Multi-Tasking

47

Figure 10 An illustration of how visual scanning is disrupted when drivers are talking on a hands-free cell phone. The left panel represents the scanning pattern of an undistracted driver and the right panel represents the scanning pattern when the driver is talking on a hands-free cell phone.

until it was too late to react. Note also that a driver talking on a cell phone suffers from both impaired visual scanning and inattention blindness, which helps to explain the high-crash rates associated with this activity. In sum, cell-phone conversations compete for attention with driving. The result is that visual processing is substantially impaired when drivers are talking on a cell phone (either hand-held or hands-free). This is seen both in the visual scanning of the driving environment (leading to tunnel vision) and in the extraction of information that is at fixation (leading to inattention blindness).

4. Are All Conversations Harmful to Driving? The preceding sections document that cell-phone conversations impair driving. But what about other conversations engaged in while driving? In particular, do in-vehicle conversations impair driving to the same extent as cell-phone conversations? One way to examine this issue is to compare the crash risk while conversing on a cell phone (established above as a fourfold increase) with the crash risk when there is another adult in the vehicle. Epidemiological evidence (Reuda-Domingo et al., 2004; Vollrath, Meilinger, & Kruger, 2002) indicates that the crash rate drops below 1.0 when there is an adult passenger in the vehicle (i.e., there is a slight safety advantage for having another adult passenger in the vehicle). Given that in many instances the passenger and the driver are conversing, these findings would seem to be at odds with the suggestion that any conversation task diverts attention from driving. However, there are also

48


situations where the passenger and the driver are not engaged in conversation, so a more precise analysis is needed. To provide a more formal comparison of the differences between passenger and cell-phone conversations, my colleagues and I returned to the driving simulator (Drews, Pasupathi, & Strayer, 2008). We recruited pairs of participants who knew each other before the study and randomly assigned one participant as the driver and the other as an interlocutor (a) on a cell phone or (b) as a passenger seated next to the driver in the vehicle. In single-task conditions, the driver was asked to drive down a multilane highway and take an exit at a rest stop located approximately 8 miles down the road. In dual-task conditions, the driver was asked to perform the same task while they were also engaged in a conversation with their friend. In all cases, the driver’s task was to exit the highway at the rest stop and park the vehicle. Drivers in single-task conditions had no trouble complying with this task, with a successful completion rate of 96%. However, there was a striking difference between cell-phone and passenger conversations in dual-task conditions. Passenger conversations (with a successful completion rate of 88%) did not significantly differ from single-task conditions, whereas 50% of the drivers engaged in a cell-phone conversation failed to take their exit. The difference between these two conversation conditions was significant, w2(1) ¼ 7.9, p < 0.05, providing clear evidence that the impairments to driving are not the same for all forms of conversation. We also examined the ability of drivers to maintain their lane position as they drove down the highway. We used an RMS error measure to determine variations in lane position. Single-task conditions did not differ from dual-task conditions involving an in-vehicle conversation (RMSe ¼ 0.45 vs. 0.40, respectively), whereas cell-phone conversations resulted in significantly greater lane deviation than passenger conversations (RMSe ¼ 1.0 vs. 0.4, respectively), t(39) ¼ 2.1, p < 0.01. To understand why passenger conversations differ from cell-phone conversations, we performed a detailed analysis of the conversations. Video analysis revealed that with in-vehicle conversations, the passenger often actively engaged in supporting the driver by pointing out hazards, helping to navigate, and reminding the driver of the task (i.e., exiting at the rest stop). In other cases, the conversation was temporally halted during a difficult section of driving and then resumed when driving became easier. These real-time adjustments to the conversation based on the demands of driving were not evident in cell-phone conversations. In effect, the passenger acted as another set of eyes that helped the driver control the vehicle, and this sort of activity is not afforded by cell-phone conversations. Another factor differing between passenger and cell-phone conversation is the content of the conversation. For example, a content analysis of the conversation revealed that there were significantly more references to traffic with passenger conversations (3.8) than with cell-phone conversations (2.1),

Multi-Tasking

49

t(46) ¼ 3.0, p < 0.01. This finding suggests that both the driver and the passenger share an awareness of the driving conditions, something that was significantly less likely with cell-phone conversations. Taken together, the epidemiological and simulator studies establish that not all conversations in the vehicle lead to impairments in driving. In particular, because the driver and an adult passenger adjust their conversation based upon the real-time demands of driving, in-vehicle conversations do not increase the odds of an accident. However, if that same conversation is performed over a cell phone, the conversation diverts the driver’s attention from the road and drivers are significantly more likely to be involved in a crash.

5. Can the Interference Be Practiced Away? Practice improves performance in some, but not in all contexts. A necessary condition for improvement is a consistency in the environment that can be capitalized upon with practice (Schneider & Shiffrin, 1977). If performance in the cell-phone–driving combination improves with practice, then it is possible that the impairments would diminish over time and the issues of cell phone-based driver distraction would abate as more and more drivers became proficient with this dual-task skill. However, an important aspect of driving involves reacting to unexpected events (e.g., a child running across the street, a deer darting across the road, road construction, a novel driving route, etc.), making it unlikely that driving can become automatic. Moreover, cell-phone conversations, by their very nature, vary from call to call. As a consequence, the consistency necessary to become an “expert” in talking on a cell phone while driving would appear to be missing. We tested to see if drivers could become expert cell-phone drivers with practice. The procedure involved identifying 30 individuals who used the cell phone regularly while driving (i.e., the experts who reported using the phone on 41% of their trips) and 30 drivers who did not use their phone while driving (novices). We tested these drivers in both single-task and dual-task conditions in both city driving and highway driving scenarios (Cooper & Strayer, 2008). We found no differences between the experts and novices (F < 1); both groups exhibited significant (and equivalent) impairment in dual-task conditions F(3,55) ¼ 10.7, p < 0.01. Thus, realworld experience using a cell phone while driving did not make the so-called experts any better at multitasking than the novices. We also used the driving simulator to test a “Groundhog Day”3 variation in which participants drove a scenario with the same event sequences 3

In the 1993 movie “Groundhog Day,” the actor Bill Murray plays weatherman Phil Connors who finds himself living the same day over and over.

50


for 4 days in a row (e.g., a pedestrian stepping out at a particular location). Our rationale was that the unexpected events would become more predictable and that the impairments to driving while using a cell phone would diminish with practice. This is exactly what happened. With practice, the number of collisions diminished from the first day (41 collisions) to the fourth day (18 collisions) of training, w2(1) ¼ 9.94, p < 0.01; however, even on the fourth day of practice, there were still twice as many collisions in dual-task conditions (12 vs. 6 for dual-task and single-task, respectively). To see if the improvements from Day 1 to 4 reflected a generalizable improvement in the ability to talk on a cell phone while driving, the participants were then transferred to a novel driving scenario. In the transfer phase, we observed significantly more crashes in dual-task conditions (26 collisions) than in single-task conditions (10 collisions), w2(1) ¼ 6.35, p < 0.05. In fact, the collision rates at transfer did not differ significantly from that observed on the first day of training. What the transfer analyses tell us is that the improvements observed with “Groundhog Day” training were specific to the training sequences, and when drivers were exposed to novel events at transfer, the pattern of dual-task impairment returned to the levels observed on the first day of training. Neither real-world practice nor simulator training made drivers perform better in novel dual-task conditions. There was no evidence that drivers became experts at the dual-task combination of talking on a cell phone while driving. We suggest that the dynamic nature of both driving and conversing on a cell phone precludes the possibility of practicing away the dual-task costs of this concurrent activity.

6. Is Everyone Impaired by Using a Cell Phone While Driving? A final issue to which we turn examines individual differences in the ability to concurrently talk on a cell phone while driving. We have provided clear evidence based upon group averages that using a cell phone while driving impairs performance. In fact, the evidence indicates that the interference is bidirectional, that is, not only does cell-phone use impair driving performance, but driving also interferes with the quality of the cell-phone conversation. But are there individual differences in the ability to multitask while driving? And, more importantly, are there “supertaskers” in our midst, individuals who can drive while simultaneously conversing on a cell phone without noticeable impairment? If so, what allows them to exhibit behavior that seemingly violates cognitive scientists’ current understanding of attention and dual-task control?

51

Multi-Tasking

To identify individuals with extraordinary multitasking ability, we paired the task of driving with an auditory version of the Operation Span (OSPAN) task. The OSPAN task involves maintaining the task goal of memorizing items and recalling them in the correct serial order while concurrently performing distracting math problems. Individual differences in OSPAN performance have been shown to predict behavior on a wide range of cognitive tasks thought to require frontal executive attention. Two hundred participants performed the driving and OSPAN tasks in combination and also performed each of the tasks separately. We predicted that most individuals would show substantial performance declines in driving and OSPAN when performed together compared to the singletask baseline measures. By contrast, individuals with extraordinary multitasking ability, if they exist, might be able to perform these two tasks in combination without impairment. The group-level data are presented in Figure 11. Dual-task performance was inferior to single-task performance for brake reaction time, F(1,199) ¼ 51.3, p < 0.01, following distance, F(1,199) ¼ 10.2, p < 0.01, OSPAN memory performance, F(1,199) ¼ 66.4, p < 0.01, and OSPAN math performance F(1,199) ¼ 30.6, p < 0.01. This pattern of performance is consistent with the well-established pattern of dual-task performance decrements associated with limited capacity attention.

Driving measures

OSPAN measures

1300 1200 1100

56 Single task Dual task

Memory performance

1400

30 25

Figure 11

Math performance

35

52 48 44 40

70

40 Following distance (m)

Break reaction time (ms)

1500

68 66 64 62

The group-level data for single-task and dual-task conditions.

52


Moreover, the data indicate bidirectional interference such that both driving and OSPAN measures suffered in dual-task conditions. Further scrutiny revealed a small subset of participants (N ¼ 5; 3 males and 2 females) scoring in the upper quartile of the OSPAN memory task (i.e., “high spans”) and showing no performance decline from single-task to dual-task across all the dependent measures. We used a stringent set of criteria for classifying participants as a “supertasker.” The first requirement was that performance on each of the four dependent measures was in the top 25% of the single-task scores for that variable, ensuring that the absence of dual-task costs could not be attributed to “sandbagging” in single-task conditions. The second requirement was that dual-task performance could not differ from single-task levels by more than the single-task standard error of the mean for that measure. Participants received a score ranging from 0 to 4, reflecting the number of measures in which they showed no dual-task decrement. Participants who earned a score of 3 (N ¼ 4) or 4 (N¼1) were classified as supertaskers (i.e., participants who performed both tasks at the same time with high levels of proficiency on each task) and those earning a score of 0–2 were classified as controls. Note that a score of 2 or lower indicates that one or both of the tasks were not performed as well in dualtask conditions as in single-task conditions. As illustrated in Figure 12, the dual-task cost for these supertaskers was zero; they performed as well, if not better, in the dual-task condition than

Driving measures

OSPAN measures

1400

65 Group average Supertaskers

Memory performance

Break reaction time (ms)

1500

1300 1200 1100

55 50 45

72 Math performance

40 Following distance (m)

60

35 30 25

71 70 69 68 67 66

20 Single task

Figure 12

Dual task

Single task

Dual task

A comparison of control and supertasker performance.

Multi-Tasking

53

they did in the single-task conditions. Independent sample t-tests comparing the difference between single-task and dual-task conditions indicated significantly smaller costs for supertaskers than for controls in brake reaction time, t(198) ¼ 5.0, p < 0.01; following distance, t(198) ¼ 3.1, p < 0.01; OSPAN memory performance, t(198) ¼ 4.6, p < 0.01, but OSPAN math performance did not differ (p > 0.10). We also compared the performance of supertaskers with the subset of participants who scored in the top quartile of the OSPAN task (i.e., high spans). Independent sample t-tests comparing the difference between single-task and dual-task revealed significantly smaller costs for supertaskers in brake reaction time t(49) ¼ 3.5, p < 0.01. and OSPAN memory performance t(49) ¼ 4.8, p < 0.01. There was also a trend for smaller costs in following distance for supertaskers t(49) ¼ 1.9, p < 0.06, whereas the costs in OSPAN math performance did not differ (p > 0.20). Note also that the supertaskers began in singletask conditions in the upper quartile of the distribution and became an even more extreme outlier in dual-task conditions. To ensure that this pattern of data did not arise by chance alone, we performed a Monte Carlo simulation in which randomly selected single– dual-task pairs of variables from the existing data set were obtained for each of the four dependent measures and then subjected to the same algorithm that was used to classify the supertaskers. The Monte Carlo procedure simulated 100,000 participants, and we found that by chance alone, 0.16% of the cases resulted in performance criteria that matched those of the supertaskers (compared to the obtained 2.5% of cases; a 15-fold difference). Logistic regression found that the frequency of supertaskers was significantly greater than chance w2(1) ¼ 17.9, p < 0.01. Given that this pattern cannot be attributed to chance, it suggests that an important individual difference variable underlies the effect. We have suggested that this individual difference is associated with differences in executive attention as mediated, at least in part, by the frontal cortex (Watson & Strayer, 2010). To test the hypothesis that the extraordinary multitasking behavior of supertaskers is mediated by differences in the frontal cortex, we invited our supertaskers plus three individuals who met the supertasker criteria in subsequent studies (making a total of eight supertaskers) and a control group matched on working memory capacity (assessed using the OSPAN task), age, handedness, and gender back for additional testing. This testing took place at least a month after the initial screening and involved having the participants perform a challenging N-back task while their brains were scanned using functional magnetic resonance imaging (fMRI). Participants were also retested on a single-task variant of the OSPAN task. In the dual N-back task, participants were instructed to respond when the letter and/or position of the square matched the stimuli N-times back (i.e., 1 time back in the 1-back condition, 2 times back in the 2-back condition, and so on). The N-back was administered as a dual task in that

54


visual/spatial and auditory/verbal stimuli were presented simultaneously, requiring participants to process both modalities independently ( Jaeggi et al., 2007). Participants completed the dual N-back task in two separate fMRI sessions in a Siemens 3T Trio MR scanner with a standard head coil. With the accuracy data, there was a significant effect of N-back load, with accuracy decreasing as load increased, F(3,30) ¼ 4.06, p < 0.01. More importantly, there was also a significant effect of group, F(1,10) ¼ 10.67, p < 0.01. The latter effect indicates that the supertaskers performed the dual N-back task more accurately than the controls. In addition, the test–retest reliability of the OSPAN task was higher for the supertaskers than for the controls, indicating a high level of stability for supertaskers. The stability of the OSPAN performance across several months reflects a reliable ability difference, and the superior performance in the dual N-back task suggests that the ability of supertaskers generalizes beyond the driving/OSPAN dual-task combination used for classification by Watson and Strayer (2010). That is, the supertasker classification reflects a reliable and generalizable ability difference. The fMRI analyses found several brain regions that differed for supertaskers and controls as they performed the dual N-back task. Of these, three frontal areas were of particular importance because they have been implicated in prior research on multitasking: frontopolar prefrontal cortex (FP-PFC), dorsolateral prefrontal cortex (DL-PFC), and anterior cingulate cortex (ACC). In all cases, the supertaskers had less activity at higher levels of load than controls. These neuroimaging findings provide an important biobehavioral marker of supertaskers’ performance and suggest that they are more efficient, achieving higher levels of accuracy in the dual N-back task with less metabolic activity (i.e., fewer resources). Note, however, that in terms of working memory capacity, supertaskers and controls did not differ, that is, there is something specific about multitasking that makes supertaskers unique. In other words, the dissociative pattern indicates that supertaskers excel at multitasking, but it is not the case that supertaskers are necessarily “smarter” across the board. Supertaskers have a remarkable ability to successfully perform two attention-demanding tasks that over 97% of the population cannot perform without incurring substantial costs in performance. Paradoxically, a recent study examining multitasking ability found that individuals who report multitasking more frequently do so less well than those who are less frequent multitaskers (Ophir, Nass, & Wagner, 2009). Indeed, our studies over the last decade have found that a great many people have the belief that the laws of attention do not apply to them (e.g., they have seen other drivers who are impaired while multitasking, but they believe that they are the exception to the rule), which is consistent with the general overconfidence of beliefs about one’s ability. In fact, some readers may also be wondering whether they too are supertaskers; however, we suggest that the odds of this are

Multi-Tasking

55

against them. The illusion that people harbor about their superior multitasking ability is likely to be driven by inattention blindness, whereby attention is diverted from sources of evidence that would indicate that their driving behavior is impaired. The discussion of supertaskers begs an interesting question: Why are we all not supertaskers? We suggest two possibilities. First, there may be some cost associated with being a supertasker. People are often faced with a stability/plasticity dilemma in which they must strike a delicate balance between being overly rigid and overly flexible in their processing style. Indeed, many clinical disorders are associated with an imbalance, being either overly rigid or overly flexible (DSM-IV, 1994). It may be that supertaskers excel at multitasking at the expense of other processing abilities. Second, there may be few costs (and possibly benefits) associated with being a supertasker, but the environmental and technological demands that favor this ability are relatively new, and any selective advantage for being a supertasker has yet to propagate throughout the population. Indeed, it has only been in the last few generations that technology has placed high value on multitasking ability. This time-scale is too short for any selective advantage to spread through the population. Together, these individual differences in multitasking behavior provide clear evidence for cognitive distraction (for the majority of us) and help to localize the areas of the brain (i.e., frontal cortex) that become overloaded when drivers attempt to talk on a cell phone while driving. In particular, these findings help to bridge the gap between applied cognition and cognitive neuroscience. Ultimately, we believe that the differences between supertaskers and controls can be leveraged to provide theoretical insight into why cognition does (or does not) break down for dual-task combinations beyond cell phones and driving.

7. Conclusions and Future Directions This chapter took an applied cognitive neuroscience approach to driver distraction, integrating methods and theories from cognitive science and cognitive neuroscience into the study of driving. Considering the ubiquity of driving and the fact that motor-vehicle crashes are the leading cause of accidental deaths in the United States, we believe that this work can have a significant impact. We focused on cognitive distraction and showed that for the most prevalent exemplar, driving while conversing on a cell phone, impairments can be as profound as operating a motor vehicle at the legal limit of alcohol. We showed that using a cell phone induces a form of inattention blindness and provided evidence using eye tracking and ERP methodologies of this impairment. We also used state-of-the-art

56


neuroimaging methods (fMRI) to identify several regions of the frontal cortex that are overloaded in multitasking situations. We also showed that talking on a cell phone differs in important ways from other forms of verbal communication (e.g., talking to an adult passenger in the vehicle). Translational research is often used to help guide public policy, and this has been the case with the research described herein. Members of our research team have participated in two National Distracted Driving Summits and briefed members of both the US House and Senate on the science of driver distraction. Given the explosion of new technologies that are making their way into the vehicle, the issues of driver distraction are likely to get much worse in the coming years. Unfortunately, there will be thousands of additional lives lost because a driver was multitasking instead of paying full attention to the road. We suggest two important directions for further research. First, a theoretically sound and methodologically rigorous technique should be developed to determine the distraction potential of a device before it is used while driving (and this is particularly true if the device is installed by the auto manufacturer). We suggest that it is unwise and unethical to integrate a device into the vehicle without first proving that it does not cause harm. By comparison, a drug company cannot market a drug unless it has gone through a rigorous set of evaluations to ensure that it causes no harm. This research need not be atheoretical. That is, not only will this research help to improve safety on the roads, but also it has the potential for helping to refine cognitive theory (as was the case for the research on supertaskers; for other examples, see Strayer et al., 2003). Second, it is important to understand why people continue to engage in a distracting activity when they acknowledge that it is risky. For example, surveys indicate that large segments of the driving public support legislation restricting or prohibiting the use of cell phones to talk or text. Yet, these same surveys also indicate that 85% of adult drivers talk on their cell while driving and 47% of adults report text messaging while driving. There is clearly a disconnect in that people support legislation that would restrict the activities in which they regularly engage. Understanding the bases for this disconnect is likely to be important both theoretically and in the process of helping to better translate our scientific understanding of driver distraction into good public policy.

REFERENCES Brown, T. L., Lee, J. D., & McGehee, D. V. (2001). Human performance models and rearend collision avoidance algorithms. Human Factors, 43, 462–482. Cooper, J. M., & Strayer, D. L. (2008). Effects of simulator practiced and real-world experience on cell-phone related driver distraction. Human Factors, 50, 893–902.

Multi-Tasking

57

Cooper, J. M., Vladisavljevic, I., Medeiros-Ward, N., Martin, P. T., & Strayer, D. L. (2009). Near the tipping point of traffic stability: An investigation of driving while conversing on a cell phone in simulated highway traffic of varying densities. Human Factors, 51, 261–268. Drews, F. A., Pasupathi, M., & Strayer, D. L. (2008). Passenger and cell-phone conversation during simulated driving. Journal of Experimental Psychology: Applied, 14, 392–400. DSM-IV. (1994). The diagnostic and statistical manual of mental disorders (4th ed.). Washington, DC: American Psychiatric Association. Fabiani, M., Karis, D., & Donchin, E. (1986). P300 and recall in an incidental memory paradigm. Psychophysiology, 23, 298–308. Glassbrenner, D. (2005). Traffic safety facts research note: Driver cell phone use in 2005—Overall results. DOT HS 809 967. Washington, DC: National Center for Statistics and Analysis, National Highway Traffic Safety Administration. Harlbluk, J. L., Noy, Y. I., Trbovich, P. L., & Eizenman, M. (2007). An on-road assessment of cognitive distraction: Impacts on drivers’ visual behavior and braking performance. Accident Analysis and Behavior, 39, 372–378. Hirst, S., & Graham, R. (1997). The format and presentation of collision warnings. In Y.I. Noy (Ed.), Ergonomics and safety of intelligent driver interfaces, (pp. 203–319). Hillsdale, NJ: Lawrence Erlbaum. HLDI. (2009). Hand-held cell phone laws and collision claims frequencies. IIHS. http://www.iihs.org/laws/cellphonelaws.aspx. Jaeggi, S. M., Buschkuehl, M., Etienne, A., Ozdoba, C., Perrig, W. J., & Nirkko, A. C. (2007). On how high performers keep cool brains in situations of cognitive overload. Cognitive, Affective & Behavioral Neuroscience, 7(2), 75–89. McEvoy, S. P., Stevenson, M. R., McCartt, A. T., Woodward, M., Haworth, C., Palamara, P., et al. (2005). Role of mobile phones in motor vehicle crashes resulting in hospital attendance: A case-crossover study. British Medical Journal, 331, 428–433. National Highway Transportation Safety Administration. (2001). Traffic safety facts—2001. Washington, DC: U.S. Department of Transportation (Rep. DOT 809 484). National Highway Transportation Safety Administration. (2006). The impact of driver inattention on near-crash/crash risk: An analysis using the 100-car Naturalistic Driving Study Data (DOT HS 810 594). Washington, DC: US Department of Transportation. National Safety Council White Paper. (2010). Understanding the distracted brain: Why driving while using hands-free cell phones is risky behavior. (on-line publication). Noy, Y. I. (2009). Human factors issues related to driver distraction from in-vehicle systems. (unpublished PowerPoint presentation). Ophir, E., Nass, C. I., & Wagner, A. D. (2009). Cognitive control in media multitaskers. Proceedings of the National Academy of Sciences, 106, 15583–15587. Otton, L. J., & Donchin, E. (2000). Relationship between P300 amplitude and subsequent recall for distinctive events: Dependence on type of distinctiveness attribute. Psychophysiology, 37, 644–661. Pew Internet and American Life Project. http://pewinternet.org/Reports/2010/CellPhones-and-American-Adults.aspx, Available from. Redelmeier, D. A., & Tibshirani, R. J. (1997). Association between cellular-telephone calls and motor vehicle collisions. The New England Journal of Medicine, 336, 453–458. Reuda-Domingo, T., Lardelli-Claret, P., Luna-del-Castillo Jde, D., Jimenez-Moleon, J. J., Garcıa-Martın, M., & Bueno-Cavanillas, A. (2004). The influence of passengers on the risk of the driver causing a car collision in Spain: Analysis of collisions from 1990 to 1999. Accident Analysis and Prevention, 36(481–489), 229–246. Schneider, W., & Shiffrin, R. M. (1977). Controlled and automatic human information processing: 1. Detection, search, and attention. Psychological Review, 84, 1–66.

58


Strayer, D. L., & Drews, F. A. (2007). Cell-phone induced inattention blindness. Current Directions in Psychological Science, 16, 128–131. Strayer, D. L., Drews, F. A., & Johnston, W. A. (2003). Cell phone induced failures of visual attention during simulated driving. Journal of Experimental Psychology: Applied, 9, 23–52. Strayer, D. L., & Johnston, W. A. (2001). Driven to distraction: Dual-task studies of simulated driving and conversing on a cellular phone. Psychological Science, 12, 462–466. Stutts, J., Feaganes, J., Rodman, E., Hamlet, C., Meadows, T., Rinfurt, D., et al. (2003). Distractions in everyday driving. AAA Foundation for Traffic Safety, (on-line publication). Vollrath, M., Meilinger, T., & Kruger, H. P. (2002). How the presence of passengers influences the risk of a collision with another vehicle. Accident Analysis and Prevention, 34, 649–654. Watson, J. M., & Strayer, D. L. (2010). Supertaskers: Profiles in extraordinary multitasking ability. Psychonomic Buelletin & Review, 17, 479–485. Willette, R. E., & Walsh, J. M. (1983). Drugs, driving, and traffic safety. Geneva: World Health Organization Publication no. 78.

C H A P T E R

T H R E E

Psychological Research on Joint Action: Theory and Data ¨nther Knoblich, Stephen Butterfill, and Natalie Sebanz Gu Contents 60 62 62 64 66 66 66 72 74 77

1. Introduction 2. Emergent and Planned Coordination 2.1. Emergent Coordination 2.2. Planned Coordination 2.3. Summary 3. Evidence 3.1. Emergent Coordination 3.2. Emergent Coordination During Joint Action 3.3. Consequences of Emergent Coordination 3.4. Planned Coordination 3.5. The Synergy of Planned and Emergent Coordination in Enabling Effective Joint Action 4. Discussion Acknowledgments References

85 91 93 93

Abstract When two or more people coordinate their actions in space and time to produce a joint outcome, they perform a joint action. The perceptual, cognitive, and motor processes that enable individuals to coordinate their actions with others have been receiving increasing attention during the last decade, complementing earlier work on shared intentionality and discourse. This chapter reviews current theoretical concepts and empirical findings in order to provide a structured overview of the state of the art in joint action research. We distinguish between planned and emergent coordination. In planned coordination, agents’ behavior is driven by representations that specify the desired outcomes of joint action and the agent’s own part in achieving these outcomes. In emergent coordination, coordinated behavior occurs due to perception–action couplings that make multiple individuals act in similar ways, independently of joint plans. We review evidence for the two types of coordination and discuss potential synergies between them. Psychology of Learning and Motivation, Volume 54 ISSN 0079-7421, DOI: 10.1016/B978-0-12-385527-5.00003-6

#


59

60

¨nther Knoblich et al. Gu

1. Introduction Human life is full of joint actions ranging from a handshake to the performance of a symphony (H. H. Clark, 1996). As Woodworth (1939, p. 823) pointed out, in many or all cases of joint action, it is not possible to fully understand individuals’ actions in isolation from each other: “Two boys, between them, lift and carry a log which neither could move alone. You cannot speak of either boy as carrying half the log [. . .]. Nor can you speak of either boy as half carrying the log [. . .]. The two boys, coordinating their efforts upon the log, perform a joint action and achieve a result which is not divisible between the component members of this elementary group.” How, then, can the basic processes enabling people to perform actions together be studied through psychological experiments? What are the perceptual, cognitive, and motor processes that enable individuals to coordinate their actions with others, and how can the seemingly irreducible components of joint actions (Hutchins, 1995) be characterized? This chapter provides an overview of current theories and experiments in psychology that have substantially enhanced our understanding of joint action. Generally, a joint action is a social interaction whereby two or more individuals coordinate their actions in space and time to bring about a change in the environment (Sebanz, Bekkering, & Knoblich, 2006). Coordinating one’s actions with others to achieve a joint outcome, such as lifting a basket together and placing it on a table, seems to require some kind of interlocking of individuals’ behaviors, motor commands, action plans, perceptions, or intentions. Early approaches to joint action originate in philosophers’ interest in the nature of joint intentionality. These approaches specify representational systems that enable the planning of joint actions. Philosophers generally agree that joint actions are actions done with shared intentions: what distinguishes joint actions from individual actions is that the joint ones involve a shared intention and shared intentions are essential for understanding coordination in joint action. This conceals deep disagreement on what shared intentions are. Some hold that shared intentions differ from individual intentions with respect to the attitude involved (Kutz, 2000; Searle, 1990 [2002]). Others have explored the notion that shared intentions differ with respect to their subjects, which are plural (Gilbert, 1992), or that they differ from individual intentions in the way they arise, namely, through team reasoning (Gold & Sugden, 2007), or that shared intentions involve distinctive obligations or commitments to others (Gilbert, 1992; Roth, 2004). Opposing all such views, Bratman (1992, 2009) argues that shared intentions can be realized by multiple ordinary individual intentions and other attitudes whose contents interlock in a distinctive way (see further Tollefsen, 2005).

Joint Action

61

The philosophical work on joint intentionality has guided research on language use where language is conceived of as a form of joint action (Brennan & Hanna, 2009; H. H. Clark, 1996). Focusing on common perceptions, common knowledge, and communicative signals, this approach situates joint planning in particular environments and particular interaction histories. For instance, the analysis of joint actions such as assembling furniture together or playing a piano duet has revealed how speech is used to prespecify who will do what and to agree on the specifics of the joint performance (H. H. Clark, 2005). Studies addressing how people solve spatial coordination problems have demonstrated that humans readily invent new symbol systems to coordinate their actions if conventional communication is not an option (Galantucci, 2009). The philosophical work on joint intentionality has also inspired groundbreaking research on the phylogenetic and ontogenetic roots of joint action and social understanding (Call, 2009; Carpenter, 2009; Tomasello, 2009). Melis, Hare, and Tomasello (2006) found that chimpanzees understand when they need to elicit the help of a conspecific to retrieve food and select the best collaborators to support their actions. This indicates that humans are not the only species to possess a representational system to support the planning of joint actions. However, it seems that humans are especially prone (“have a special motivation”, Tomasello, 2009) to engage in joint action and to help others to achieve their goals (Brownell, Ramani, & Zerwas, 2006). For instance, 1-year-old infants perform actions to help adults attain their goals (Warneken & Tomasello, 2007) and gesture helpfully to provide relevant information (Liszkowski, Carpenter, & Tomasello, 2008). By 3 years, children understand that joint action implies commitment of the individual partners (Gra¨fenhain, Behne, Carpenter, & Tomasello, 2009). Research on perception, action, and cognitive control has focused on the nuts and bolts of joint action, addressing the perceptual, cognitive, and motor mechanisms of planning and coordination. Ecological psychologists have studied rhythmic joint actions in order to determine whether dynamical principles of intrapersonal coordination scale up to the interpersonal case (Marsh, Richardson, & Schmidt, 2009). This research has shown that in many cases, the movement of limbs belonging to different people follows the same mathematical principles as the movement of an individual’s limbs (e.g., Schmidt, Carello, & Turvey, 1990). Cognitive psychologists have studied how coactors represent each other’s tasks and how the ability to predict each other’s actions supports coordination in real time (Sebanz, Bekkering, et al., 2006). The results of this research suggest that specific perceptual, motor, and cognitive processes support joint action (Knoblich & Sebanz, 2008; Semin & Smith, 2008) and that the needs of joint action shape individual perception, action, and cognition (Knoblich & Sebanz, 2006; Tsai, Sebanz, & Knoblich, in press).

62


This chapter provides a review of recent joint action research with a focus on the nuts and bolts of joint action. We begin by outlining a set of processes of emergent and planned coordination that support interpersonal coordination during joint action. We then review studies that have addressed particular processes of emergent coordination and planned coordination. In the last part of the chapter, we discuss evidence that could lead to an improved understanding of the interplay between planned and emergent coordination in enabling effective joint action.

2. Emergent and Planned Coordination We distinguish between two types of coordination that can occur during joint action, planned coordination, and emergent coordination. In planned coordination, agents’ behavior is driven by representations that specify the desired outcomes of joint action and the agent’s own part in achieving these outcomes. How much is specified about other agents’ tasks, perceptions, and knowledge may vary greatly. An agent may consider others’ motives, thoughts, or perspectives or simply wait for a particular action to happen (Vesper, Butterfill, Knoblich, & Sebanz, 2010). In emergent coordination, coordinated behavior occurs due to perception– action couplings that make multiple individuals act in similar ways; it is independent of any joint plans or common knowledge (which may be altogether absent). Rather, agents may process perceptual and motor cues in the same way as each other. Two separate agents may start to act as a single coordinated entity (Marsh et al., 2009; Spivey, 2007) because common processes in the individual agents are driven by the same cues and motor routines.

2.1. Emergent Coordination Emergent coordination can occur spontaneously between individuals who have no plan to perform actions together as well as during planned joint actions. For instance, pedestrians often fall into the same walking patterns (Van Ulzen, Lamoth, Daffertshofer, Semin, & Beek, 2008) and people engaged in conversation synchronize their body sway (Shockley, Santana, & Fowler, 2003) and mimic one another’s mannerisms (Chartrand & Bargh, 1999). In all of these instances of emergent coordination, similar behaviors occur spontaneously in two agents. Because these similarities do not seem instrumental for either individual goals or joint goals, emergent coordination has sometimes been portrayed as a single process (Semin & Cacioppo, 2008). However, if (as we believe) emergent coordination is a key facilitator of joint action, then it is essential to distinguish different sources of emergent coordination. We will distinguish between four such

Joint Action

63

sources, (1) entrainment, (2) common affordances, (3) perception–action matching, and (4) action simulation. 2.1.1. Entrainment Entrainment is perhaps the most widely studied social motor coordination process (Schmidt, Fitzpatrick, Caron, & Mergeche, in press). For instance, two people in rocking chairs involuntarily synchronize their rocking frequencies (Richardson, Marsh, Isenhower, Goodman, & Schmidt, 2007), and audiences in theaters tend to clap in unison (Neda, Ravasz, Brechte, Vicsek, & Barabasi, 2000). Entrainment is a process that leads to temporal coordination of two actors’ behavior, in particular, synchronization, even in the absence of a direct mechanical coupling. In dynamical systems research interpersonal entrainment is often considered as a particular instance of the coupling of rhythmic oscillators (Schmidt & Richardson, 2008) that is frequently observed in mechanical as well as biological systems. 2.1.2. Affordances Whereas entrainment occurs in the direct interaction between agents, common object affordances provide the basis for a further dynamical process of emergent coordination. Object affordances (Gibson, 1977), previously discussed as the “funktionale Toenung” of objects (von Uexkuell, 1920), specify the action opportunities that an object provides for an agent with a particular action repertoire. For instance, a chair “invites” sitting down on it. When two agents have similar action repertoires and perceive the same object, they are likely to engage in similar actions because the object affords the same action for both of them. This is a type of affordance that we will call common affordance because it leads to emergent coordination when agents perceive the same objects at the same time. Examples of objects with a common affordance that may induce emergent coordination include the arrival of a bus, an apple falling from a tree, and a shelter in the park. Another case of affordance, which we call joint affordance, is where objects have an affordance for two or more people collectively which is not necessarily an affordance for any of them individually. For example, a long two-handled saw affords cutting for two people acting together but not for either of them acting individually. 2.1.3. Perception–Action Matching: Common Action Representations A third process that can lead to emergent coordination is the matching of observed actions onto the observer’s own action repertoire. Such a matching can lead to mimicry of observed actions because perceiving a particular action activates corresponding representations that also guide the actions of the observer. Common representations in perception and action have been postulated in extensions (Hommel, Muesseler, Aschersleben, & Prinz, 2001;

64


Jeannerod, 1999; Prinz, 1997) of ideomotor theories of voluntary action control (James, 1890) and have received neurophysiological support from single-cell studies in monkeys and brain imaging studies in humans (Rizzolatti & Sinigaglia, 2010). In monkeys and humans, the matching is based on the similarity in actor–object relations. For instance, seeing someone grasp a grape activates grasping actions directed at small, round objects. In humans, the matching can also be based on similarity in intransitive movements that are not directed at objects. For instance, observing someone dancing will activate corresponding action representations if one knows how to dance (Calvo-Merino, Glaser, Gre`zes, Passingham and Haggard, 2005; Cross, Hamilton, & Grafton, 2006). The perception–action match can lead to emergent coordination because it induces the same action tendencies in different agents who observe one another’s actions (Knoblich & Sebanz, 2008). 2.1.4. Action Simulation: Common Predictive Models The fourth process of emergent coordination is closely related to the perception–action matching described above. Once a match between observed and performed actions is established, it enables the observer to apply predictive models in his or her motor system to accurately predict the timing and outcomes of observed actions. These processes are often referred to as action simulation (Sebanz & Knoblich, 2009) because they use internal models guiding an agent’s own actions to predict other agents’ actions in real time (Wolpert, Doya, & Kawato, 2003). To illustrate, a basketball player observing a shot will be able to accurately predict whether the shot will be a hit or a miss (Aglioti, Cesari, Romani, & Urgesi, 2008). Action simulation can lead to emergent coordination because it induces the same expectations about the unfolding of actions in different actors and thus induces similar action tendencies for future actions (Knoblich & Sebanz, 2008). This concludes our preliminary outline of four sources of emergent coordination. In the next main section, we present evidence for the existence of emergent coordination generally and for its occurrence in the context of joint action more specifically, and we discuss hypotheses about the positive consequences of emergent coordination for joint action. First, we turn to planned coordination which, unlike emergent coordination, depends on representing the outcomes of joint actions and individuals’ contributions to them.

2.2. Planned Coordination In order to perform joint actions, such as playing a piano duet or lifting a heavy log, planned coordination is usually required. In planned coordination, agents plan their own actions in relation to joint action outcomes or in relation to others’ actions, whereas planning is absent or confined to the

Joint Action

65

agent’s own actions in emergent coordination. The extent to which other agents’ tasks, perceptions, and knowledge are taken into account during planning of joint actions may vary greatly. Minimally, planned coordination requires a plan that specifies the joint action outcome, one’s own part in a joint action, and some awareness that the outcome can only be brought about with the support of another agent or force (X). For the minimal joint action plan, the identity of X and its part in the joint action can remain unspecified as captured by the formula “ME þ X” (Vesper et al., 2010). Starting with minimal representational requirements (A. Clark, 1997), allows one to address a wide range of joint actions that do not involve the detailed representation of other agents or their plans that have been postulated in philosophical approaches to joint action (Bratman, 1992; Tomasello, Carpenter, Call, Behne, & Moll, 2005). Given our focus on the nuts and bolts of joint action, assuming such detailed representations seems unnecessarily restrictive. Among the many processes contributing to planned coordination, we will focus on shared task representations and joint perceptions. 2.2.1. Shared Task Representations In the minimal cases of joint action, actors represent an outcome that they are not going to achieve alone and the task they need to perform themselves. Very often, though, joint action involves representations of the other agents who are actually and potentially involved. For instance, a chimpanzee who can only get food from a tray with the help of a conspecific may select one among several potential helpers according to how useful each is likely to be (Melis et al., 2006). This chimpanzee needs to represent the goal of obtaining food and their own task of pulling a rope but need not have detailed representations of the conspecific’s actions. Often, however, representations of others’ tasks are more detailed, specifying the actions others are going to perform. This is demonstrated by people’s proneness to represent the specifics of others’ actions and tasks (Sebanz, Knoblich, & Prinz, 2005). Shared task representations provide control structures that allow agents to flexibly engage in joint action. Shared task representations not only specify in advance the individual parts each agent (me and you in the simplest case) is going to perform but they also govern monitoring and prediction processes that enable interpersonal coordination in real time (Knoblich & Jordan, 2002; Pacherie & Dokic, 2006). For instance, two soccer players of one team, where one player is specialized on crosses and the other is specialized on headers, will monitor and predict each other’s running paths in the light of their individual tasks. 2.2.2. Joint Perceptions Planned coordination can be improved by including another’s perceptions into one’s own representation of the other’s task. This can consist in taking the other’s perspective in situations where coactors’ perspectives on a jointly

66


perceived environment differ such as when two actors sit face to face looking at objects to be assembled. Or it can consist in inferring what a coactor can or cannot perceive in situations where perceptual access to objects in the environment differs between coactors (Brennan & Hanna, 2009; Shintel & Keysar, 2009). Although it is debated how prone agents are to corepresenting each other’s perceptions, there is evidence that at least some aspects of another’s perspective are computed even when doing so hinders one’s own performance (Samson, Apperly, Braithwaite, Andrews, & Bodely Scott, in press). Corepresented perceptions might be highly useful for planned coordination in helping to establish perceptual common ground between actors (H. H. Clark, 1996), in enabling one to adapt one’s own task, and in facilitating monitoring of the other’s task.

2.3. Summary This section distinguished between two types of coordination, one—emergent coordination—involving multiple individuals acting in similar ways, thanks to common perception–action couplings, and another—planned coordination—involving representations of a joint action goal and contributory tasks to be performed in pursuit of it. Whereas emergent coordination involves processes such as entrainment and perception–action matching, planned coordination is supported by shared task representations and joint perceptions. In what follows, we review evidence for the existence of the various processes and structures we have linked to each type of coordination. Because much of this evidence is found outside of joint action, we also examine how these processes and structures facilitate joint action. In the final subsection, we also explore the synergy of emergent and planned coordination.

3. Evidence 3.1. Emergent Coordination 3.1.1. Entrainment For a long time, psychologists (Condon & Ogston, 1966; Trevarthen, 1979) have recognized the importance of rhythmic behavior in social interaction. Building on this earlier work, psychologists subscribing to a dynamical systems view now propose that entrainment is best understood as a selforganizing process that occurs in coupled oscillators (Haken, Kelso, & Bunz, 1985). The claim is that just as two clocks hanging on the same wall tend to synchronize because they are mechanically coupled (Huygens, 1673/1986), individuals may become automatically coupled through perceiving the same visual, auditory, or haptic information. This hypothesis was tested in experiments that determine whether people fall into

Joint Action

67

synchrony even though they try to keep their own speed (Schmidt & O’Brien, 1997). Such experiments provide converging evidence that people cannot resist falling into synchronous behavior with others. Schmidt and O’Brien (1997) were perhaps the first to study explicitly whether interpersonal entrainment between two people would occur even when both try to resist entrainment and try to maintain their own speed. In their experiment, two persons sitting side by side moved a hand-held pendulum. One person used her left hand and the other person used her right hand so that the two pendulums were located in between the two persons. On each trial, both individuals started out moving their own pendulum in a speed that was comfortable to them. Importantly, they were asked to look straight and thus did not see each other or the pendulums during this phase. For the second half of each trial, both individuals were asked to “maintain their preferred tempo from the first half of the trial while looking at the other participant’s moving pendulum” (Schmidt & O’Brien). Two results showed that participants could not resist falling into synchrony with each other. First, cross-spectral coherence, a sensitive measure of the correlation between the timing of the two individual movements, was higher during the second half of the trial than during the first half of the trial. Second, the relative phase between the two movements was much more frequently close to 0 and 180 during the second half of the trial than during the first half of the trial. Especially, the later result shows that the two individuals could not resist falling into synchrony, either in-phase (0 , same synchronized turning points for the pendulum) or antiphase (180 , different synchronized turning points for the pendulum). M. J. Richardson and colleagues (2007b, Experiment 2) obtained further evidence to support this claim. They asked two individuals to rock in a rocking chair at their preferred tempos either while looking at each other or while looking away from each other. The results demonstrated that the individuals could not resist interpersonal entrainment even if the “natural rocking frequencies” (eigenfrequencies) of the two rocking chairs they were rocking in differed. Unlike in the pendulum studies, participants were only drawn into in-phase coordination (same synchronized turning front and back) and not into antiphase coordination, suggesting that interpersonal entrainment varies depending on the specific body parts and the specific objects being moved. A further recent study (Oullier, de Guzman, Jantzen, Lagarde & Kelso, 2008) investigated whether the effects of unintended coordination occur for tapping movements. Two individuals were instructed to tap at a comfortable tempo with a finger. As in the previously described experiments, they were either looking at each other’s tapping movements or had their eyes closed. Auditory signals indicated when participants should open or close their eyes. Again, the analysis of relative phase revealed that participants strongly

68


tended to fall into synchrony (in-phase only). Surprisingly, two participants stayed entrained with each other even when they closed their eyes again after having seen each other’s movements. This finding conflicts with the view that interpersonal entrainment can be reduced to a coupling between oscillators, because this would predict that each individual should return to their preferred tempo. Accordingly, Oullier et al. (2008) propose that a “social memory” keeps participants synchronized when the visual input supporting the coupling is absent. Tognoli, Lagarde, De Guzman, and Kelso (2007) adapted the tapping task described above to investigate whether there are specific neural markers of interpersonal entrainment in the human electroencephalogram (EEG). In particular, they simultaneously recorded EEG from two people who were looking or not looking at each other’s tapping (see above). The results demonstrated that two oscillatory EEG components in the range between 10 and 11 Hz, Phi1 and Phi2, specifically occur during interpersonal entrainment. Whereas Phi1 activation decreased with increasing coordination, Phi2 activation increased with increasing coordination. It remains to be seen whether Phi1 and Phi2 can be established as a general marker of interpersonal entrainment across different experimental settings. Issartel, Marin, & Cadopi (2007) demonstrated that behavioral effects of interpersonal entrainment can even be obtained when participants are asked to freely move their forearms while explicitly instructed to ignore each other’s peripherally observed movements. Although under this instruction, participants did not engage in the joint rhythmic movements characterizing the studies above, their individual motor signatures (preferred movement frequencies) became more similar when they peripherally observed each other’s movements. This demonstrates that individuals cannot resist subtle interpersonal entrainment effects for “freely chosen” movements that look random and independent to an observer. Harrison & Richardson (2009) investigated whether the same principles govern the rhythmic movement of four limbs within and across organisms ( Jeka, Kelso, & Kiemel, 1993). Two participants were asked to walk around at a certain distance from each other. The participant walking behind could either see the other participant or was mechanically connected to the participant by a big foam cube, or both. The results showed that when the two participants were only visually or mechanically coupled, they fell into a coordinated walking pattern that very much resembled a horse pace. When they were visually and mechanically coupled, they fell into a walking pattern that very much resembled a horse trot. These findings suggest that the same stable multilimb coordination patterns can emerge within and across organisms (cf. Mechsner & Knoblich, 2004). Finally, the mechanisms of interpersonal entrainment have also been investigated in situations that involve more than two persons. One famous

Joint Action

69

example concerns the dynamics of the transformation of tumultuous applause into orderly and rhythmic clapping studied by Ne´da and colleagues (2000). They demonstrated that applauding audiences fall into synchrony by slowing down their own spontaneous clapping to roughly half its initial speed. Interestingly, the slower tempo required for synchronization of large groups considerably reduces the loudness of the applause so that synchronization disappears again to increase noise levels. 3.1.2. Affordance Whereas the studies described above provide ample evidence for interpersonal entrainment, the role of affordances as a mechanism for emergent coordination has so far not been addressed (Knoblich & Sebanz, 2008). Although object affordances have been studied extensively in experiments on individual perception ( Jones 2003; Tucker & Ellis, 1998, Yoon, Humphreys, & Riddoch, in press), we are not aware of psychological experiments addressing the role of affordances in emergent coordination. Such experiments would need to establish that similar action affordances induced by “usable” objects help actors to coordinate their actions. Such benefits should be particularly strong when actors have the same experience with the particular use of objects, because coordination should profit from the increased similarity in actor–object relations that results from frequently using objects in the same way. Some researchers have started to explore how the presence of another person creates affordances for acting together (Richardson, Marsh, & Baron, 2007). This reflects an interaction between planned and emergent coordination and will be discussed below. 3.1.3. Perception–Action Matching Perception–action matching is a further mechanism that can lead to emergent coordination. Whereas processes of entrainment can explain why two actors’ rhythmic actions get aligned, perception–action matching can explain why individuals tend to perform similar actions (Brass & Heyes, 2005) or actions that lead to similar perceptual consequences (Hommel et al., 2001; Prinz, 1997) while observing each other. Accordingly, studies on entrainment tend to address situations where people perform the same or very similar movements (but see Richardson, Campbell, & Schmidt, in press). However, visual and auditory entrainment should occur regardless of action similarity if two actions are performed at the same frequencies. The studies on perception–action matching, mimicry, and action simulation described below tend to exclusively focus on the similarity between observed actions and performed actions and neglect the role of timing. Several studies have demonstrated that observing a particular movement in another person leads to an automatic activation of the same movement in the observer (Brass, Bekkering, & Prinz, 2001; Bertenthal, Longo, & Kosobud, 2006). In Brass and colleagues’ (2001) experiment, participants

70


observed a video of either a lifting movement or a tapping movement on the computer screen. In one block, participants were instructed to respond to any movement on the screen with a tapping movement. In a second block, participants were instructed to respond to any movement on the screen with a lifting movement. Although participants knew exactly what to do in each trial, they were faster performing a lifting movement when they observed a lifting movement and faster performing a tapping movement when they observed a tapping movement. Stu¨rmer, Aschersleben and Prinz (2000) found similar results for manual movements. Participants observed videos of a hand that performed a spreading movement or a grasping movement from a neutral position. They were instructed to react to a color patch that occurred at the same time as the movement onset or shortly after movement onset. For one color, the response consisted in a spreading movement, and for the other color, the response consisted in a grasping movement. Thus, the observed hand movement was irrelevant for the response. Nevertheless, responses were faster when they corresponded to the observed movement, providing further evidence for the assumption that observing a movement activates the same movement in the observer’s motor repertoire. In another study, Kilner, Paulignan, and Blakemore (2003) asked participants to perform vertical or horizontal arm movements while observing vertical or horizontal movements of a human actor or of a robot. They found that the participants’ arm movements became more variable if they did not correspond with the observed human movement than when they corresponded with the observed human movement. The correspondence effect was not present when robot movements were observed. This finding suggests that perception–action matching occurs only if the kinematics of an observed movement is similar to the kinematics the observer would produce. Richardson, Campbell, & Schmidt (2009) have proposed an alternative explanation for this finding that is based on entrainment mechanisms. The studies described so far all involved simple movements that were not directed at objects. However, in animal research, the paradigmatic case for a close perception–action match consists in movements that serve to manipulate objects (Rizzolatti & Sinigaglia, 2010). Thus, several studies have investigated whether a perception–action match occurs when an observer perceives another person performing object-directed actions. Castiello, Lusher, Mari, Edwards, and Humphreys (2002) performed a study where participants observed a person grasping a small or a large object and were subsequently asked to grasp a small or large object themselves. If participants performed the same action they had observed before, they were faster in initiating their action and more effective in optimizing motor parameters such as a grip aperture (see also Edwards, Humphreys, & Castiello, 2003). Similarly, Bach and Tipper (2007) found that observing a person kicking

Joint Action

71

a ball facilitated foot responses, whereas observing a person typing on a keyboard facilitated finger responses. Griffiths and Tipper (2009) demonstrated that an observer does not only match the type of action observed but also specific kinematic parameters an observed actor adopts to avoid obstacles to reaching ( Jax & Rosenbaum, 2007; Van Der Wel, Fleckenstein, Jax, & Rosenbaum, 2007). 3.1.4. Action simulation Further studies have demonstrated that perception–action matching can induce motor predictions in the observer (Hamilton, Wolpert, & Frith, 2004; Wilson & Knoblich, 2005; Wolpert et al., 2003). This has been tested varying the similarity between observed actions and the observer’s action repertoire. Assuming that this similarity is higher when people perceive their own previous actions than when people perceive somebody else’s previous actions, Knoblich and Flach (2001) hypothesized that people should be better able to predict the landing position of a dart when observing their own throwing movement than when observing somebody else’s movement. The results confirmed the prediction and supported the assumption that perception–action matching can trigger motor predictions in the observer (see Knoblich, Seigerschmidt, Flach, & Prinz, 2002, for similar results in the handwriting domain). Converging evidence for this conclusion has been obtained in a study that compared professional basketball players’ and basketball reporters’ ability to predict the outcome of basketball shots (Aglioti et al., 2008). The hypothesis in this study was that basketball players’ expertise would allow them to more accurately predict whether a particular throwing movement would be a hit or miss, and this hypothesis was confirmed. Finally, it has been demonstrated that perception–action matching can influence attention. A study by Flanagan and Johansson (2003) demonstrates that perception–action matching can result in predictive eye movements, implying that an observer allocates attention to location or objects that the observed actor is expected to manipulate next. Flanagan and Johansson recorded eye movements of a person who moved a stack of objects from one location to another and compared them to the eye movements of people who observed a video recording of these actions. The results showed that the gaze behavior of participants observing the performance was highly similar to the gaze behavior of the person who had carried out the original action. These results suggest that perception–action matching does not only activate corresponding hand actions in the observer but also mimics processes of eye-hand coordination in the observed actor, in particular, the well-known temporal order “eye precedes hand”. Findings on inhibition of return across people show that perception– action matching can induce inhibition of return mechanisms for locations an observed actor attended to (Welsh, Elliott, et al., 2005). Inhibition of

72


return refers to the phenomenon that it takes individuals more time to detect a target when it appears in the same location as another stimulus presented shortly before the target. Welsh, Elliott, et al. (2005) and Welsh, Higgins, Ray, and Weeks (2007) demonstrated that observing a person respond to a target in a particular location slowed down an observer’s response to a target appearing at the same location. This between-person inhibition of return effect suggests that inhibitory attention processes can be triggered by mere action observation. The results of a recent study by Frischen, Loach, and Tipper (2009) also support this conclusion. This study demonstrated that observing another’s actions triggered inhibitory attention processes of negative priming. Interestingly, in the observation condition, the inhibitory processes followed the observed actor’s spatial reference frame and not the observer’s spatial reference frame. Overall, then, there is a rich body of evidence for three sources of emergent coordination and an open question about a fourth source, common and joint affordances. Note that the evidence we have reviewed so far mainly concerns nonjoint action situations where participants were not instructed to act together and where coordination among agents was not beneficial to performing the task and in some cases may have degraded performance. This raises two questions. First, what evidence is there that emergent coordination occurs when agents are performing a joint action? Second, how (if at all) could emergent coordination facilitate joint action? We now address these questions in turn.

3.2. Emergent Coordination During Joint Action In the studies reviewed in the previous section, emergent coordination occurred despite the fact that individuals were instructed to ignore each other’s actions or at least were not given reason to attend to each other. This section reviews studies where emergent coordination was studied in the context of joint action, including conversation. In all of these studies, two individuals showed emergent coordination of behavior that was apparently not necessary for achieving the goal of the joint action. This includes emergent coordination of movements such as body sway (Fowler, Richardson, Marsh, & Shockley, 2008) as well as emergent coordination of eye movements between the speaker and the listener in a conversation (Richardson, Dale, & Shockley, 2008). 3.2.1. Entrainment Although temporal coordination of speech patterns and body movements during conversation has been the subject of many observational studies (Condon, 1976; Kendon, 1970; Wachsmuth, Lenzen, & Knoblich, 2008), the systematic experimental study of interpersonal entrainment during joint action and conversation is quite new. Richardson, Marsh, and Schmidt

Joint Action

73

(2005) investigated interpersonal coordination of pendulum swinging while participants solved a puzzle task. Participants were asked to swing hand-held pendulums while jointly solving the puzzle. Two factors were varied: Participants either saw or did not see each other and they either talked or did not talk to each other. Interpersonal entrainment occurred only when participants perceived each other’s movements, implying that verbal interaction alone was not sufficient to produce a coupling between the individuals. However, the lack of interpersonal entrainment in the verbal interaction condition may be due to the dual task character of the study. Rhythmic pendulum movements and verbal rhythms may not have been sufficiently related to produce an interpersonal entrainment of manual movements through speech. Studies on interpersonal entrainment of body sway during conversation suggest that talking to each other can indeed be sufficient to produce interpersonal entrainment of body sway (Fowler et al., 2008), which consists in automatic movements that serve to keep a stable body posture. Shockley and colleagues (2003) asked two individuals to find subtle differences between two cartoon pictures either of which could only be seen by one of them. Participants were either facing each other or looking away from each other. The surprising finding was that talking to each other was sufficient to create interpersonal entrainment of body sway, as evidenced by a higher rate of recurrence in a cross-recurrence analysis (this analysis allows one to discover similarities in temporal patterns across different time series; Shockley, Butwill, Zbilut, and Webber, 2002). In a recent study, Shockley, Baker, Richardson, and Fowler (2007) extended these results by showing that particular properties of the conversation such as dyadic speaking rate and similarity in stress patterns give rise to acoustically mediated entrainment of body sway. Stoffregen, Giveans, Villard, Yank, and Shockley (2009) have identified further factors that modulate the entrainment of body sway such as the rigidity of the surface people are standing on. However, body sway is not the only type of movement that gets entrained during conversation. Two studies demonstrated that there is also acoustically mediated emergent coordination between the eye movements of speakers and listeners. Richardson and Dale (2005) recorded eye movements from speakers describing stories from an American sitcom they were highly familiar with while looking at the main characters. The verbal utterances were replayed to listeners (new participants) who were also familiar with the same sitcom while their eye movements on the same display of the main characters were recorded. Cross-recurrence analysis was used to determine overlap in the temporal patterns of the speaker and the listener. This analysis showed that verbal utterances were sufficient to produce emergent coordination between the eye movements of the speaker and the listener. In other words, verbal communication led to an overlap in the temporal rhythm between the speaker and the listener, thereby aligning attention. Similar

74


results were obtained in a setting where the speaker and the listener were engaged in a real-time dialogue (Richardson, Dale, & Kirkham, 2007). 3.2.2. Perception–Action Matching Studies on nonconscious mimicry during dialogue have also revealed emergent coordination based on perception–action matching. These studies demonstrate that observing the actions and mannerisms of a conversation partner leads individuals to perform the same movements without being aware of mimicking their partner. Chartrand and Bargh (1999) provided a demonstration of this “chameleon effect” by asking participants to take turns with another participant sitting next to them at describing photographs. The other participant was actually a confederate who engaged in particular mannerisms such as shaking her foot or rubbing her face. Video analyses showed that participants mimicked the confederate, rubbing their face more often when the confederate rubbed her face and shaking their foot when they observed their partner shaking her foot. Participants were not aware of their partner’s mannerisms and did not deliberately try to mimic them. This suggests that perceiving an action triggers corresponding action representations in the observer, which can lead to overt mimicry in the context of conversation. The extent to which people mimic others’ actions depends on individual characteristics, including the tendency to take others’ perspective (Chartrand & Bargh), and the tendency to rely on contextual information and to feel close to others (for a review, see van Baaren, Janssen, Chartrand, & Dijksterhuis, 2009).

3.3. Consequences of Emergent Coordination We have just seen that emergent coordination occurs in joint action contexts, for example, when two people are solving a puzzle together. But in the studies cited so far, emergent coordination appears to occur independently of participants’ individual and shared goals. As noted, emergent coordination may sometimes even interfere with individual action planning as when people cannot help falling into the same rhythm or mimicking observed actions. This may seem puzzling if, as we have proposed, emergent coordination can facilitate some joint actions. In fact, several recent studies suggest that emergent coordination has various effects that may serve a number of different psychological functions. These effects include increased affiliation and liking of a partner, increased willingness to cooperate with the partner, and increased understanding of the meanings a partner intends to convey in a conversation. 3.3.1. Entrainment It has long been suggested that the rapport between individuals is reflected in the synchrony of their body movements (e.g., Bernieri, 1988). More recently, Miles, Nind, and Macrae (2009) demonstrated that people judge

Joint Action

75

the connectedness of individuals in a dyad based on the perceived synchrony of their movements. Participants saw or heard footsteps of pairs of walkers walking in a more or less synchronized manner and rated their degree of rapport. The results showed that participants attributed the highest levels of rapport to those pairs of walkers that displayed in-phase or antiphase coordination, and assigned the lowest levels of rapport to walkers displaying phase relationships that were far from in-phase or antiphase. Thus, the most stable patterns of entrainment were clearly perceived as reflecting a close connection between individuals, regardless of whether information about the walkers’ synchrony was conveyed through visual or auditory information. That observers take synchrony to indicate rapport does not establish that these are causally related. Evidence for a causal link has been provided by Hove and Risen (2009). In their study, participants performed a tapping task next to another person, synchronizing their finger movement with a visual or auditory signal. Each of the two individuals responded to separate signals so that the target tempo for their tapping could be more or less similar. Even though participants knew that the signals determined to which extent they and their task partner were synchronized, those who had been more in synchrony with their partner subsequently reported liking her more. Entrainment also seems to boost people’s willingness to cooperate with group members (Wiltermuth & Heath, 2009). In a coordination game, participants who had walked in step in groups of three made more cooperative choices than participants who had not walked in step. Those who had engaged in synchronized walking also reported feeling more connected and trusting each other more. The same was true for groups of three singing in synchrony. Of particular interest was the further finding, using multiple rounds of a public-goods game, that following synchronous group actions, the level of participants’ contributions to the public good did not significantly fall as time went by, whereas the level of such contributions did decline over time in groups that had not engaged in synchronous behavior. These findings suggest that by increasing group cohesion, synchronous group action serves to increase altruistic behavior. Improvements in joint action performance following entrainment, as well as gains in understanding due to entrainment during conversation, provide further demonstrations of the benefits of entrainment. In a study by Valdesolo, Ouyang, and DeSteno (2010), two groups of participants rocked in rocking chairs. One group rocked next to each other, which allowed them to entrain, while the other group rocked back to back to avoid entrainment. Participants who had rocked in synchrony were subsequently better at an individually performed perceptual sensitivity task that required judging the speed of an occluded object, compared to participants who had rocked back to back. Interestingly, the increased perceptual sensitivity induced by the synchronized rocking may explain why

76


synchronized dyads also performed better in a subsequent joint action task that involved steering a ball through a labyrinth together. These findings provide a first indication that entrainment may have effects on the quality of subsequent joint action performance. The study by Richardson and Dale (2005) discussed above provides evidence that emergent coordination between the eye movements of a speaker and a listener can aid understanding. In their experiments, a speaker monologued while looking at an array of six characters; listeners then saw the same display (but not the speaker) while hearing the monologue. The degree to which the gaze between speakers and listeners overlapped predicted how many comprehension questions listeners subsequently answered correctly. A second experiment provided evidence that gaze coordination and comprehension are causally connected. While looking at an array of characters, listeners’ attention was drawn to particular locations at particular times by having the pictures of the characters flash. This made it possible to make listeners’ gaze pattern more or less similar to the speaker’s. Compared to a condition where the flashes appeared at shuffled times, participants indeed responded to comprehension questions more readily when their gaze had been drawn to the locations coinciding with the speaker’s gaze fixations. 3.3.2. Perception–Action Matching Increased liking seems to result not only from entrainment, but also from nonconscious mimicry, the tendency to perform the same actions as one’s interaction partner without being aware of doing so. The classic study on the “chameleon effect” (Chartrand & Bargh, 1999) included an experiment where the confederate either mimicked participants’ postures, movements, and mannerisms without them being aware of it, or simply sat next to them in a relaxed position. Participants who had been mimicked reported liking the confederate more and judged the interaction as smoother, suggesting that nonconscious mimicry may act as a kind of “social glue” (Lakin, Jefferis, Cheng, & Chartrand, 2003). Many studies have since confirmed and extended this finding (for a review, see Van Baaren et al., 2009). In particular, being mimicked does not only lead to increased liking of the person who did the mimicking, but seems to increase people’s prosocial orientation in general (van Baaren, Holland, Kawakami, & van Knippenberg, 2004). Participants whose postures had been mimicked were more likely to help pick up pens dropped by a stranger and donated more money to a charity. A further study by van Baaren and colleagues also suggests that mimicry increases the tendency to share resources. They demonstrated that when a waitress mimicked her customers, they gave her significantly larger tips (van Baaren, Holland, Steenaert, & van Knippenberg, 2003).

Joint Action

77

So far in this section, we have surveyed basic evidence for four sources of emergent coordination and reviewed evidence that emergent coordination occurs in the context of joint action. As we saw, there is evidence that two sources of emergent coordination, entrainment and perception–action matching, occur when people are acting together and, in particular, when they are engaged in conversation. The mere fact that emergent coordination occurs in joint action does not show, of course, that it plays any role in facilitating it. In fact, we saw that in some cases, emergent coordination may make performing joint actions harder than it would otherwise be. A crucial question, then, is how emergent coordination facilitates joint action. We have already seen part of the answer: emergent coordination promotes rapport and willingness to contribute to a group, which may indirectly benefit joint action; more directly, emergent coordination in the form of spatiotemporally coincident gaze appears to facilitate understanding. This evidence is consistent with a range of possible views on the significance of emergent coordination for joint action. Our own conjecture, supported below, is that emergent coordination cannot be fully understood in isolation from planned coordination, for many of the ways in which emergent coordination enables effective joint action depend on its functioning in combination with planned coordination. Before developing this theme, we first consider evidence for planned coordination at length.

3.4. Planned Coordination 3.4.1. Shared Task Representations Basic findings: In prototypical cases of planned coordination, agents represent an outcome to be achieved, their own task, and some aspects of other agents’ tasks in achieving that outcome. Psychological experiments performed with the aim of investigating how individual task performance is modulated by coactors’ tasks have shed light on the question of when and how others’ tasks are represented. Although representing a coactor’s task may not always be necessary, the findings of these experiments consistently suggest that humans form task representations that specify not only their own part, but also the part to be performed by the coactor. Moreover, the findings suggest that task representations entailing a specification of the individual tasks each agent is going to perform govern stimulus processing (Heed, Habets, Sebanz, & Knoblich, 2010), action monitoring (Schuch & Tipper, 2007), control (Sebanz, Knoblich, Prinz, & Wascher, 2006; Tsai, Kuo, Jing, Hung, & Tzeng, 2006), and prediction (Ramnani & Miall, 2004) processes during the ensuing interaction. A first study (Sebanz, Knoblich, & Prinz, 2003) investigated whether a response selection conflict between two action alternatives (a right and a left button press) that is known to occur within individuals is also observed across individuals in a social setting.

78


Participants responded to pictures of a red or green ring presented on an index finger pointing left or right. When participants performed the twochoice task alone, they responded to red stimuli by pressing a left button and to green stimuli by pressing a right button. Although the pointing direction of the index finger was irrelevant, participants responded faster to red stimuli when the finger pointed left than when it pointed right, and vice versa for green stimuli. This spatial compatibility effect demonstrates that the irrelevant spatial information of the stimulus elicited a response conflict when the finger pointed to the side opposite to the button that had to be pressed. The social version of this task tested whether the same response conflict would occur across individuals where neither individual’s task required taking the coactor’s actions or task into account. One participant responded to red stimuli by pressing a single button in front of her. Next to this participant sat another participant responding only to green stimuli with her own button. Thus, each task could be performed without taking the coactor’s task into account. Nevertheless, a response selection conflict was observed, with participants responding more slowly when the finger pointed at their coactor. A control condition showed that this interference did not occur when another person merely sat next to an individual participant. The findings of this first study suggest that participants did not ignore their coactor. Instead, they represented the action to be executed by the coactor so that a similar conflict in action selection occurred regardless of whether they were in charge of both actions or whether they performed the task together. These findings were replicated in a study where participants responded to odd and even numbers with left or right key presses (Atmaca, Sebanz, Prinz, & Knoblich, 2008). The numbers ranged from 2 to 9 and number magnitude was always irrelevant. It is well established that when individuals perform the parity task alone, as a two-choice task, left key presses are faster in response to small numbers and right key presses are faster for large numbers. This effect of number magnitude on parity judgments (the socalled “SNARC” effect) has been explained by the assumption that the perception of numbers automatically activates a magnitude representation on a mental number line going from the left to the right (Dehaene, 1997). Atmaca and colleagues showed that the same effect occurs when two people sitting next to each other perform the task together, so that one responds only to even numbers and the other only to odd numbers. Participants sitting on the left were faster when responding to small numbers, and participants sitting on the right were faster when responding to larger numbers. This suggests that, as in the compatibility task described above, participants represented their own action alternative in relation to the coactor’s actions. Cognitive mechanisms: These findings on shared task representations have raised many questions. In particular, there has been considerable debate

Joint Action

79

regarding the mechanisms underlying the observed corepresentation effects. Guagnano, Rusconi, and Umilta` (2010) tested whether corepresentation effects only occur when two actors perform complementary tasks, taking turns in responding, or whether corepresentation effects occur even when the two agents’ tasks are completely independent. In their study, the two participants in a pair performed independent detection tasks, one responding to red stimuli and the other to blue stimuli. The stimuli appeared either on the side on which the response should be made (compatible) or on the opposite side (incompatible). On 80% of the trials, the stimuli required a response from both participants to avoid turn taking and to make the two tasks maximally independent. A compatibility effect was observed when the two participants were sitting close to each other within arm’s reach (in the so-called peripersonal space). However, the compatibility effect vanished when the coactors were sitting outside of each other’s peripersonal space. Based on these findings, Guagnano and colleagues suggested that if a coactor is sufficiently close, the coactor provides a spatial reference point for coding the location of one’s own action. Instead of representing the specifics of the other’s task, then, coactors might simply use the other as a spatial reference. However, it remains to be specified how exactly such a spatial reference is established. Welsh (2009) reported similar compatibility effects when participants sitting next to each other crossed their hands and when they performed the same tasks with hands uncrossed. This finding suggests that if spatial coding is taking place, it can be flexibly based on the position of one’s body relative to the other’s body, or on the position of one’s hand relative to the other’s hand. A recent study (Heed, Habets, Sebanz, & Knoblich, 2010) demonstrates that both the spatial relationship between coactors and the task representations specifying the coactor’s part play a role. In this study, participants held cubes that emitted tactile stimulation on top (index finger) or at the bottom (thumb). Their task was to indicate via a foot response at which location the tactile stimulation had occurred. A light appeared on top or at the bottom of the cube (congruent or incongruent with the tactile stimulation) and was irrelevant for the task. When participants performed this task alone, responses to tactile stimuli were faster when the irrelevant light appeared in the same location (e.g., touch and light at the bottom) compared to when the light and the tactile stimulation appeared in opposite locations (e.g., touch at the bottom, light on top, Spence, Pavani, Maravita, & Holmes, 2004). Heed and colleagues tested whether this cross-modal congruency effect is modulated when a coactor performs a task involving the lights. Based on earlier findings on shared task representations, one might predict stronger cross-modal interference because the irrelevant lights are relevant for one’s task partner. However, representing the other’s task could

80


also facilitate task performance given that, unlike in previous studies, stimuli from two different sensory modalities were distributed between two coactors. The results indeed showed that the person responding to tactile stimuli could ignore the irrelevant light much better when her coactor responded to the location of the light. This effect only occurred, however, when the person responding to lights was sitting in the peripersonal space of the person responding to tactile stimuli, and when she responded to all lights. The reduction in cross-modal congruency was not observed when the person in charge of lights responded only to one of two different colored lights. This finding indicates that the weight assigned to the visual modality was changed when the partner’s task covered all visual events in peripersonal space. Thus, a representation of the other’s task modulated stimulus processing provided that the coactors were in a particular spatial relation to each other. Stimulating further debate about the mechanisms underlying shared task representations, some studies indicate that for others’ actions to be included in one’s own action plan, they must be visible (Welsh, Higgins, et al., 2007) and of a biological nature (Tsai & Brass, 2007), whereas other findings suggest that shared task representations occur even when people merely believe that they are acting together with another person (Ruys & Aarts, 2010; Tsai, Kuo, Hung, & Tzeng, 2008). Using the social compatibility task described at the beginning of this section, Tsai and colleagues told participants that they were going to perform the task (e.g., responding to red stimuli) together with a person in another room (responding to green stimuli) or with a computer program (the computer taking care of green stimuli). They found a compatibility effect when people believed that they were performing the task together with another person but not when they believed that they were performing the task with the computer. This indicates that the actual task performance is constrained by task representations formed in advance. Importantly, in the studies that found corepresentation effects with invisible coactors, participants constantly received (mock) feedback about the other’s actions (Ruys & Aarts, 2010; Tsai et al., 2008). This feedback may be necessary to maintain a representation of the other’s task. Neural mechanisms: Electrophysiological and brain imaging methods have been used specifically to investigate the processes occurring when a coactor does not need to act herself and awaits the other’s response. These studies have revealed two main findings. First, individuals seem to generate predictions of the other’s actions based on their representation of the other’s task (Ramnani & Miall, 2004). For instance, if two participants have been instructed to perform particular actions in response to certain color cues, seeing a color cue that specifies the action to be performed by a coactor elicits activation in brain areas associated with mental state attribution, which may reflect an ongoing prediction process.

Joint Action

81

Second, electrophysiological evidence suggests that acting together requires the recruitment of control processes to ensure that one does not act when it is the other’s turn. A positive event-related potential occurring 300–500 ms poststimulus was significantly more pronounced when participants needed to inhibit an action because it was their coactor’s turn compared to when they needed to inhibit an action because it was nobody’s turn to act (De Bruijn, Miedl, & Bekkering, 2008; Sebanz, Knoblich, et al., 2006; Tsai et al., 2006). Effects of shared task representations on action control have also been demonstrated by studies investigating the observation of errors (Bates, Patel, & Liddle, 2005; Schuch & Tipper, 2007; van Schie, Mars, Coles, & Bekkering, 2004). To investigate whether similar inhibitory processes occur in the person trying to stop an action and in an observer watching her coactor stop an action, Schuch and Tipper asked participants to respond to targets as quickly as possible, but to stop if a stop signal was presented shortly after the target. It is well known that participants respond more slowly on the trial following a stop signal, both if they have successfully stopped and if they have made an error. The results showed that participants were not only slower after they had stopped or made an error themselves, but also after their coactor had done so. This indicates that control processes governing one’s own actions are also active during a coactor’s performance, even if the coactor’s performance is irrelevant for one’s own task. The study by Ramnani and Miall (2004) mentioned above is also important in that it shows that completely arbitrary task rules are corepresented. In many other corepresentation experiments, the stimuli had a spatial dimension, leading to an overlap in perceptual features of the responses to be made by the two coactors and perceptual features of the stimuli (e.g., Sebanz et al., 2003). In contrast, in the study by Ramnani and Miall, the stimuli did not refer to the coactors in any such way. This allows one to conclude that coactors anticipated each other’s actions based on their task representation only. Converging results were obtained in a study where participants responded to a spatial stimulus feature next to a coactor responding to a certain stimulus color (Sebanz, Knoblich, & Prinz, 2005). Whereas some of the stimuli required a response from one participant only, others required a response from both participants. The results showed that participants responding to the spatial stimulus feature were slower when the stimulus color indicated that it was also the other’s turn to respond. This suggests that the (arbitrary) task rule specifying responses to color was represented by the individuals responding to the spatial stimulus feature. Learning: Recently, researchers have begun to investigate whether and how jointly practiced task rules modulate subsequent performance of another joint task (Milanese, Iani, & Rubichi, 2010). It is known from studies of individual performance that when participants respond to stimuli on the left with a right key press and to stimuli on the right with a left key

82


press, they subsequently show a reduced or even reversed spatial compatibility effect in a task where they need to respond to color and have to ignore the spatial position of the stimuli. That is, whereas participants would normally find it easier to make a left key press when a stimulus appears on the left, after the practice, participants find it easier to make a left key press when a stimulus appears on the right. Milanese and colleagues used this transfer effect to demonstrate that joint practice modulates subsequent joint performance of a compatibility task in just the same way. Interestingly, transfer effects also occurred when participants first performed the practice alone followed by the joint compatibility task, but not when they first performed the practice together, and then performed the compatibility task alone. This indicates that the representations guiding joint task performance may in fact be quite different from the representations guiding individual performance, with transfer occurring more easily from the individual to the joint case. Transfer studies may thus constitute a useful new way to study the nature of shared task representations. Social modulations: A related line of research has investigated how social factors, such as characteristics of the coactors and the nature of the interaction context, modulate the tendency to take each other’s tasks into account. To investigate possible links between impairments in mental state attribution and shared task representations, individuals with autism were asked to perform variants of the spatial compatibility task described above (Sebanz, Knoblich, Stumpf, & Prinz, 2005). They showed similar corepresentation effects as a matched control group of typical adolescents and adults, indicating that deficits in understanding particular mental states such as beliefs do not necessarily affect the tendency to represent the rules specifying a coactor’s task. However, recent findings provide a more nuanced view (Ruys & Aarts, 2010). Using an auditory version of the joint compatibility paradigm where participants believed that they were interacting with another person, Ruys and Aarts found that individuals who were good at inferring others’ mental states took the coactor’s task into account regardless of the interaction context, whereas individuals who were less able to infer others’ mental states showed signs of shared task representations only in a competitive context. The ability to infer others’ mental states was measured using the mind in the eyes test which requires selecting one of the four terms that best describes the emotional state of different pairs of eyes (Baron-Cohen, Wheelwright, Hill, Raste, & Plumb, 2001). These results indicate that limitations in the ability to infer others’ mental states, which may come with less of a propensity to do so, may also come with a decreased tendency to corepresent others’ tasks. The comparison between competitive and cooperative interaction contexts is of general interest in that it may reveal effects of particular prior intentions on task performance. Such effects have been demonstrated by studies comparing the performance of a grasping movement in a

Joint Action

83

collaborative context, where participants reach for and place a wooden block on a table to build a tower together with a coactor, and in a competitive context, where participants intend to place the block down sooner than their partner (Georgiou, Becchio, Glover, & Castiello, 2006; Becchio, Sartori, Bulgheroni, & Castiello, 2008). The kinematics differed already during the initial reach-to-grasp action, with longer movement duration, a higher movement path of the wrist, and a later time of opening the hand to grasp (maximum grip aperture) during cooperation than during competition. Control conditions where participants act alone under different instructions suggest that these effects are not simply due to movement speed, but instead reflect the intention underlying the grasping movement. Finally, affect also seems to play a role in task corepresentation. Using the joint compatibility task, Hommel, Colzato, & van den Wildenberg (2009) found corepresentation effects when participants acted together with a confederate who was friendly and cooperative, but not when they acted with a confederate who was intimidating and competitive. This result suggests that shared task representations only occur in positive relationships. However, it is also possible that mood is a key factor. When participants were presented with movies to induce a positive, negative, or neutral affective state before performing the joint compatibility task, corepresentation effects occurred only following positive and neutral mood induction (Kuhbandner, Pekrun, & Maier, in press). Summary: In short, behavioral, electrophysiological, and brain imaging evidence shows that humans represent not only their own tasks but also those of their partners and even those of people who they do not need to coordinate with. Much progress has already been made on questions about when agents represent coactors’ tasks. As we saw, whether agents represent others’ tasks does not appear to depend on whether doing so is necessary for performing their own tasks effectively, nor always on directly perceiving their coactors; but it does depend on believing that the other task is being performed by an agent rather than an algorithm, and in some cases, it depends on whether agents are acting in each other’s peripersonal space. While it is difficult, at present, to fully specify a detailed model of how shared task representations arise, there is much evidence on the related question of their effects. Representing a coactor’s task means needing to inhibit oneself from performing her actions and having one’s motor system become sensitive to her errors. Thus, shared task representations influence how agents monitor and plan their actions. In addition, shared task representations may also influence how the external world is perceived. This makes it easy to see how, in general terms, shared task representations could facilitate joint action. By representing their coactor’s tasks, agents are able to coordinate their actions and predict their joint outcome because they are each monitoring and planning both sets of actions.

84


3.4.2. Joint Perceptions Actors can adjust their actions to facilitate coordination if they are able to assess what their partner perceives at a particular moment in time. This may involve directing one’s own attention depending on where the other is looking, taking the other’s perspective, or inferring what the coactor can or cannot perceive in situations where perceptual access to objects in the environment differs between coactors. A study by Brennan and colleagues demonstrates that coactors are able to distribute a common search space by directing their attention depending on where the other is looking (Brennan, Chen, Dickinson, Neider, & Zelinsky, 2007). The task was to find the letter “O” among lots of Qs on a computer screen. Participants were instructed to indicate the presence or absence of the O as quickly as possible by pressing one of the two buttons. The two participants in a pair sat in different rooms and wore a headmounted eye tracker each. This made it possible to indicate the current gaze position of a searcher to her partner, who saw it displayed as a cursor on her screen. Both participants thus could see where their partner was looking. Joint search performance was much better than individual performance. Interestingly, joint performance was best when participants were not allowed to talk to each other. These findings suggest that being able to see where their partner was looking allowed people to divide the search space in an efficient manner. However, coactors do not always have access to the same visual input. When they are in different spatial locations, they may have different perspectives on the same scene, and only some objects but not others may be visible to both. There has been considerable debate as to how such differences could be overcome in communication. On the one hand, it has been proposed that people make partner-specific adaptations based on what they assume to be common knowledge (Brennan & Hanna, 2009). On the other hand, it has been argued that mental state inferences play only a limited role because they are time consuming and cognitively demanding, whereas processes of emergent coordination may be highly useful for achieving coordination (Shintel & Keysar, 2009). In the light of this debate, recent findings by Samson and colleagues are particularly relevant (Samson et al., in press). Participants were asked to judge their own or another’s visual perspective in situations where the two perspectives were the same or different. They found that even when taking the other’s perspective interfered with their own task performance because the two perspectives differed, participants computed the other’s perspective. This parallels the findings on task corepresentation discussed above. A study on effects of a coactor’s perspective on mental rotation (Boeckler, Knoblich, & Sebanz, in press) provides further evidence that people take another’s perspective into account even when this is not

Joint Action

85

required. Participants in a pair sat opposite each other and took turns in performing a mental rotation task with pictures of hands. While one participant performed the task, the participant opposite her either closed her eyes or looked at the pictures. When the coactor looked at the stimuli to be rotated, participants were slowed down when small rotations were required (large rotations from the other’s perspective) and were speeded up when larger rotations were required (smaller rotations from the other’s perspective). These findings indicate that the other’s perspective could not be ignored. Joint attention triggered a switch to processing the stimuli within the other’s, allocentric, reference frame.

3.5. The Synergy of Planned and Emergent Coordination in Enabling Effective Joint Action Although recent research has enhanced our understanding of the component mechanisms of emergent and planned coordination, it has still not been well understood how planned coordination and emergent coordination work together in order to enable effective joint action. However, numerous studies indicate that planning joint actions taps into several different mechanisms of emergent coordination recruiting the functionality of these fast and parallel mechanisms. 3.5.1. Synergy of Planning and Entrainment Entrainment is not only observed in situations where individuals do not plan to coordinate with each other but also in situations where individuals plan to coordinate their movements in order to obtain the joint goal of producing a particular movement pattern. Accordingly, the proponents of a dynamical systems approach to cognition have stressed the importance of entrainment in planned coordination. In particular, they have studied whether movement coordination across individuals follows the same principles that govern the coordination of limbs within individuals. Schmidt et al. (1990) asked two people sitting side by side to rhythmically swing their outer legs (the left leg of the person sitting on the left and the right leg of the person sitting on the right) at the same pace as a metronome that indicated different tempos. In the symmetric (in-phase) condition, participants performed synchronous forward and backward movements with their legs, flexing and extending them at the same time. In the parallel (antiphase) condition, participants performed synchronous leg movements, but now one participant extended the leg while the other flexed the leg and vice versa. The main finding was that the dynamical interpersonal coupling between the movements followed several predictions of the HKB equation (Haken et al., 1985) that was originally developed as a quantitative model of interlimb coordination within a person. In particular, participants found it

86


easier to perform symmetric movements than parallel movements and tended to switch from parallel into symmetric mode, especially at high movement speeds. These results show that the same quantitative relationship holds in planned within-person and across-person coordination of simple rhythmic movements. Further studies where participants were asked to swing hand-held pendulums at different speeds (Schmidt & Turvey, 1994; Schmidt, Bienvenu, Fitzpatrick, & Amazeen, 1998) also support this general conclusion. However, Schmidt and colleagues noted that planned interpersonal coupling of rhythmic movements was weaker and broke down more easily than acrossperson coupling of rhythmic movements. This could be an indication that across-person coordination involves mechanisms that are different from the within-person case. One such difference has been revealed in a recent study on joint tapping where two coactors overcompensated for each other’s timing errors when trying to tap in synchrony with each other (Konvalinka, Vuust, Roepstorff, & Frith, 2010). Further evidence that the interpersonal case is special comes from a developmental experiment that investigated drumming in young children. Children aged around 2.5 years deviated more from their preferred drumming tempo when they drummed with an interaction partner than when they drummed with a mechanical device producing the same rhythmic intervals as the interaction partner (Kirschner & Tomasello, 2009). In fact, their drumming in the social interaction condition was in a timing range not spontaneously performed by the children. This indicates that children (and adults) entrain more when they are engaged in social interactions. Thus, top–down influences of joint planning may modulate the extent to which entrainment occurs. Another interesting question with regard to the relation between planning and entrainment is how musicians manage to coordinate rhythmic performances. Goebl and Palmer (2009) investigated which cues pianists use to coordinate their performances. Not surprisingly, visual and auditory feedback was important for synchronization, indicating a crucial role for entrainment and online error correction mechanisms (Keller & Repp, 2004). When auditory feedback was absent, pianists produced ostensive cues for one another by finger movements. This indicates that, at crucial points of the musical performance, “communicative” coordination mechanisms ensure that performers’ joint plans stay aligned. Studies on the entrainment of eye movements between the speaker and the listener provide further evidence for top–down modulation of the entrainment processes through common knowledge and joint plans. Richardson, Dale, et al. (2007) recorded eye movements from two conversants engaged in a real-time dialogue about a Dali painting. Before starting the conversation, they either received the same or different information about Dali’s art. Crossrecurrence analysis revealed that the eye movements of two speakers who

Joint Action

87

shared common knowledge were more tightly temporally coordinated. In a further study, Richardson, Dale, and Tomlinson (2009) demonstrated that the coordination of eye gaze is not only modulated through common knowledge but also by sharing or not sharing a visual scene or believing that the conversation partner has or does not have access to the visual scene. 3.5.2. Synergy of Planning and Affordance A further interaction between planned and emergent coordination combines joint planning and dynamic actor–object relations (affordances). Perceiving the affordance of an object for oneself (would I be able to lift this object myself?), its affordance for another person (could this person lift this object?), or its affordance for the group (could we lift this object together, given the affordance the object has for me and the other) may provide the basis for deciding whether one should plan an individual action or a joint action. Richardson, Marsh, et al. (2007, Experiment 4) investigated this hypothesis in an experiment where they asked two individuals to lift planks of different length from a conveyor belt. Participants were free to decide to lift particular planks alone or together and were required to make their decision on the fly as the plank passed by on the conveyor belt. The results demonstrated that the decision to engage in joint action or individual action systematically depended on the ratio between plank length and the groups’ joint arm span. Moreover, the transition from individual action to joint action followed the same dynamic principles as the transition from unimanual to bimanual action in individual plan lifting (Richardson et al., 2007, Experiment 2). Importantly, participants with a longer arm span took into account the shorter arm span of their partner by choosing joint action more frequently than predicted by their individual arm span. Thus, the results provide clear evidence that affordances play an important role in deciding whether to perform a joint action or an individual action with an object. A study by Mottet, Guiard, Ferrand, and Bootsma (2001, Experiment 2) provides a further indication that joint action capabilities determine how individuals act in particular task environments. Mottet and colleagues asked participants to jointly perform rhythmical movements as fast as possible. One person moved a pointer in order to move between two targets that varied in size and were separated by different distances. The other person could move the targets, making the task easier for the person moving the pointer. The results showed that the combined movements of both persons followed Fitts’s law just as when one person performed the whole task bimanually. Fitts’s law predicts quantitatively the extent to which increases in target size and decreases in movement amplitude (distance between targets) allow for faster movements between two targets. Thus, Mottet and colleagues’ study provides evidence that actors can jointly optimize performance to particular object sizes and particular distances between objects.

88


3.5.3. Synergy of Planning and Perception–Action Matching Like entrainment, perception–action matching might appear as a low-level process operating largely independently of planned coordination. However, the control of perception–action matching processes is crucial for planned coordination where actors need to perform different actions. There is substantial evidence that higher-level planning processes modulate the matching of perceived actions onto the observer’s action repertoire. Attributing particular intentions to an actor can extinguish the tendency to mimic perceived movements and can trigger the activation of compensatory or complementary movements. Liepelt, von Cramon, and Brass (2008) showed that the intentions observers attribute to an actor modulate perception–action matching. Participants were instructed to lift either the index finger or the middle finger in response to a number that was presented between the index and the middle finger of a picture of a hand. The index or the middle finger of the perceived hand moved up as the number appeared, resulting in a congruency effect (slower responses when the finger to be raised in response to the number did not correspond to the perceived finger movement). The key manipulation was whether very small micromovements of the fingers occurred when a metal clamp restricted the fingers of the perceived hand, or without the clamp, so that the actor was free to move but only performed tiny finger movements anyway. The same movement kinematics led to a larger congruency effect when the fingers of the perceived hand were clamped, giving the impression that the actor was trying hard to move despite her fingers being restricted. Thus, the effect of action perception on action execution changed as a function of the intention attributed to the actor. The role of intention attribution is also demonstrated clearly by the finding that the same kinematics creates more or less interference with action execution depending on whether people believe that the movement of a dot reflects human motion or is generated by a computer (Stanley, Gowen, & Miall, 2007). In an adapted version of the paradigm developed by Kilner et al. (2003), participants performed horizontal or vertical arm movements in time with a dot moving horizontally or vertically. The perceived dot motion interfered with participants’ movements when they were told that a person generated the dot motion, but not when they were told that a computer generated the dot motion, regardless of whether the dot actually moved in a biological or nonbiological way. Experiments on ideomotor movements demonstrate that instead of mimicking perceived actions, people tend to make involuntary compensatory movements when they observe actions that are not in line with their own or an observed actor’s intentions (De Maeght & Prinz, 2004; Haeberle, Schuetz-Bosbach, Laboissiere, & Prinz, 2008; Knuf, Aschersleben, & Prinz, 2001; Sebanz & Shiffrar, 2007). For instance, participants tracking a ball

Joint Action

89

moving toward a goal on a computer screen moved left when the ball steered too far to the right (De Maeght & Prinz, 2004), even though they had no control over the ball movement. Sebanz and Shiffrar (2007) measured participants’ body tilt as they watched someone balancing along a wobbly foam roller with outstretched arms. When the actor shared the same spatial orientation as the participants, they tilted their upper body to the left when the actor was close to falling off the right side, and vice versa when the actor tilted too far left. These findings demonstrate that the intentions ascribed to actors can overrule the tendency to mimic perceived movements and induce compensatory movements. In the context of planned coordination, the tendency to perform complementary movements may prevail over the tendency to mimic the actions of one’s coactor (Van Schie, Waterschoot, & Bekkering, 2008). Participants were asked to grasp an object in an imitative context or in a complementary action context. The object could be grasped either on top by making a precision grip or at the bottom by making a power grip. In the imitative context, participants imitated the grasp of a coactor displayed on a computer screen, whereas in the complementary context, they acted as if they were taking over the object, performing a complementary grasp. On certain trials, a color cue instructed participants to perform a particular grip, regardless of the interaction context. If the interaction context played no role, participants should always be faster at executing corresponding grips. However, the results showed that in the complementary action context, participants were faster at making a complementary grasping movement, whereas in the imitative context, they were faster at making an imitative grasping movement. This demonstrates that planning to perform a joint action involving complementary action can override the tendency to mimic the coactor’s movements and, in fact, induces a tendency to perform the complementary movement. 3.5.4. Synergy of Planning and Action Simulation In the context of planned coordination, the matching between perceived and performed actions enables coactors to apply predictive models in their motor system to accurately predict the upcoming actions of their coactor, and to predict joint action outcomes. So far, only a few studies have directly addressed the role of action simulation in planned coordination. Kourtis and colleagues studied action simulation processes in a triadic social interaction where participants passed an object back and forth with an interaction partner or lifted it alone and a third actor always acted alone (Kourtis, Sebanz, & Knoblich, 2010). A cue instructed the actors about which actions, if any, they should perform, and a second later they were prompted to act. The crucial comparison was between trials where participants did not have to act themselves, but expected that either their interaction partner would lift the object alone or that the “loner” would lift the object alone. A neural marker of action simulation reflecting anticipatory

90


motor activation was more pronounced when people anticipated the action of their interaction partner (Kilner, Vargas, Duval, Blakemore, & Sirigu, 2004) than when they anticipated the same action to be performed by the loner. Given that the actions of the partner and the loner were identical in all respects, this indicates that action simulation is constrained by the relation between participants and their interaction partners. Planning to perform a joint action may also involve simulations of the coactor’s actions that lead to adjustments in individual action performance. Becchio et al. (2008) found that the movement kinematics of a reaching movement performed to grasp an object differed depending on whether the actor reached for the object to place it on a hand-shaped pad or to place it on another person’s palm at exactly the same location. The authors suggest that the smaller grip aperture and the lower speed at which the object was grasped in the joint action context reflect the need to handle the object in a way that makes it easy for the receiving person to grasp it. This may be taken as an indication that a simulation of the action to be performed by the partner guides individual action planning and control. Action simulation likely plays a key role in joint actions that require close temporal coordination of different individual actions, such as playing a piano duet. Findings from studies of temporal coordination suggest two different ways in which action simulation may support planned coordination (Sebanz & Knoblich, 2009). On the one hand, actors may run multiple parallel action simulations to predict the timing of other coactors’ actions (Keller, Knoblich, & Repp, 2007). In support of this assumption, Keller and colleagues found that pianists playing one part of a duet together with a recording of the other part of the duet were better synchronized when playing together with a recording of their own earlier performance than when trying to synchronize with another pianist’s performance. This may indicate that they used internal models in their motor system to predict the performance of both parts of the duet (their own and the one they synchronized with), which led to the best result when the actions to which they applied the models were their own earlier actions. However, action simulation can also support temporal coordination if the target of the prediction is the timing of jointly produced events (Knoblich & Jordan, 2003). Rather than generating separate predictions for their own and a coactor’s performance, agents might generate predictions regarding the temporal consequences of their combined efforts. Such predictions about joint action outcomes can only be made, however, after agents have had the opportunity to learn about regularities between their own actions, others’ actions, and the resulting effects. This was demonstrated in a study where participants were instructed to keep a circle on top of a target moving horizontally along the computer screen, using an “acceleration” and a “deceleration” key. Participants performed the task alone, controlling both keys, or in pairs, controlling one of the keys each. After

Joint Action

91

considerable practice, shared task performance was as good as individual performance, but only when participants received auditory feedback about the timing of each other’s actions. This suggests that accurate predictions about the timing of joint action outcomes can be made if agents have had the opportunity to trace back the consequences of their combined actions to their individual contributions.

4. Discussion The evidence reviewed above shows that emergent coordination and planned coordination each supports joint action. Emergent coordination can occur spontaneously between individuals who have no plan to perform actions together and relies on perception–action couplings that make multiple individuals act in similar ways. In planned coordination, agents plan their own actions in relation to joint action outcomes or in relation to others’ actions. Shared task representations and joint perceptions support these planning processes. Most forms of joint action likely require both emergent and planned coordination because there are complementary limits on what each can achieve. On the one hand, planning alone does not make people act at the right time, fall into synchrony, or predict others’ upcoming actions based on their own action repertoire. Although planning can prepare actors to perform their individual parts of a joint action, it does not guarantee successful implementation. Emergent coordination is likely the key to dealing with the real-time aspects of joint action. On the other hand, emergent coordination alone is limited in that it does not allow people to distribute different parts of a task among themselves, nor to adjust their actions to others so as to flexibly achieve joint outcomes. These aspects of joint action require planned coordination. The complementary limits of emergent and planned coordination suggest that it is the synergy of emergent and planned coordination that allows people to make music together, play team sports, or build a house. This synergy is partly a matter of how planned coordination modulates mechanisms of emergent coordination: the examples discussed above include greater entrainment in planned social interactions, the activation of action simulations for coactors but not independent third-parties, and, under the heading of perception–action matching, the possibility of performing actions which complement rather than match observed actions depending on either the nature of one’s own task or one’s representation of the observed agent’s task. But the synergy also involves modulation of planned coordination by emergent coordination, as where perception of joint affordances causes participants to switch from individual action to joint

92


action and where action simulation of a partner’s next action affects one’s own action planning. One big challenge for future research on joint action is to specify in more detail how emergent coordination and planned coordination work together. How can shared task representations tap into mechanisms of entrainment, perception–action matching, and predictive action simulation? Which perceptions need to be shared so that mechanisms of planned and emergent coordination will act in combination? Does emergent coordination have a role in how joint action plans are set up and how roles are distributed between individual actors? What is the role of emergent coordination in generating joint perceptions? A further challenge for joint action research is to discover interfaces that allow agents to integrate the more basic processes of emergent and planned coordination with the higher-level representations and processes postulated in theory of mind research such as common knowledge and mental state attribution. It is plausible that many cases of joint action, particularly those involving many distinct steps such as putting up a large tent on a wet and windy hill, depend on interlocking intentions and commitments in addition to emergent and planned coordination. How do attributions of intention and knowledge in the pursuit of joint action goals interact with the mechanisms of emergent and planned coordination? Some of the studies reviewed above indicate the possibility that what agents believe, the mood they are in, and their social relations with one another modulate the processes that are at the heart of performing joint actions. For instance, we saw that shared task representations can depend on beliefs about the status of a partner as an agent. To what extent can shared task representations also be modulated by explicit beliefs about the partner’s task, or by beliefs about the partner’s beliefs, or intentions about one’s own task? We have also seen that both planned and emergent coordination may sometimes conflict with the avowed intentions of agents; certainly, neither form of coordination appears to depend on agents making the attributions of mental states required for sharing intentions in any elaborate sense (e.g., Bratman, 1992). This raises the possibility that mental state attribution may sometimes be integrated only indirectly with emergent and planned coordination. To illustrate, recall that sometimes how close agents are to one another in space may affect their shared task representations. Consequently, mental state attribution might lead people to position themselves in ways that affect their shared task representations. Further questions concern whether and how emergent or planned coordination modulates attribution of mental states for joint action. Here, the studies linking rapport with synchronized behavior provide one possible model. Coordination cues may allow agents to draw conclusions about the chances of successful joint action with another agent. Furthermore, agents may be sensitive to

Joint Action

93

coordination cues that indicate whether their desires or beliefs are incompatible with their partner’s. Discovering interfaces between higher-level representations of minds and actions on the one hand and planned and emergent coordination on the other may provide psychologists studying human perception, action, and cognition with the opportunity to have a major impact on the design of robots that are built to engage in action with humans (e.g., Braun, Ortega, & Wolpert, 2009; Breazeal, 2002; Wachsmuth & Knoblich, 2008). For instance, engineers designing these robots face the problem of effectively distributing the workspace between man and machine (Vesper, Soutschek, & Schuboe, 2009) and enabling haptic interactions such as joint object manipulation (Bosga & Meulenbroek, 2007; Reed et al., 2006; van der Wel, Knoblich, & Sebanz, in press), or jointly carrying objects while walking (Streuber, 2008). Finally, psychological research on joint action may also lead to a fruitful exchange between experimental psychology and different disciplines in the humanities specialized in the use of discursive, observational, and phenomenological methods (DeJaegher, DiPaolo, & Gallagher, in press), especially musicology, anthropology, and philosophy. Joint actions are central in music history and music performance (Clayton, Sager, & Will, 2004; Keller, 2008) and play a key role in most worldly and religious rituals (Vogeley & Roepstorff, 2009). Thus, joint action can serve as a platform for planned and emergent coordination across disciplines.

ACKNOWLEDGMENTS We thank Brian Ross for many helpful comments and for providing us with the opportunity to discuss our views on joint action in detail. This research has been supported, in part, by the European Science Foundation (ESF) through a European Young Investigator (EURYI) award to Natalie Sebanz.

REFERENCES Aglioti, S. M., Cesari, P., Romani, M., & Urgesi, C. (2008). Action anticipation and motor resonance in elite basketball players. Nature Neuroscience, 11(9), 1109–1116. Atmaca, S., Sebanz, N., Prinz, W., & Knoblich, G. (2008). Action co-representation: The joint SNARC effect. Social Neuroscience, 3, 410–420. Bach, P., & Tipper, S. P. (2007). Implicit action encoding influences personal-trait judgments. Cognition, 102, 151–178. Baron-Cohen, S., Wheelwright, S., Hill, J., Rast, Y., & Plumb, I. (2001). The ‘Reading the Mind in the eyes’ test revised version: A study with normal adults, and adults with Asperger Syndrome or high-functioning autism. Journal of Child Psychology and Psychiatry, 42, 241–252.

94


Bates, A. T., Patel, T. P., & Liddle, P. F. (2005). External behavior monitoring mirrors internal behavior monitoring: Error-related negativity for observed errors. Journal of Psychophysiology, 19(4), 281–288. Becchio, C., Sartori, L., Bulgheroni, M., & Castiello, U. (2008). Both your intention and mine are reflected in the kinematics of my reach to grasp movement. Cognition, 106, 894–912. Bernieri, F. (1988). Coordinated movement and rapport in teacher–student interactions. Journal of Nonverbal Behavior, 24, 120–138. Bertenthal, B. I., Longo, M. R., & Kosobud, A. (2006). Imitative response tendencies following observation of intransitive actions. Journal of Experimental Psychology: Human Perception and Performance, 32, 210–225. Boeckler, A., Knoblich, G., & Sebanz, N. (submitted). Giving a helping hand: Effects of joint attention on mental rotation of body parts. Experimental Brain Research. Bosga, J., & Meulenbroek, R. G. J. (2007). Joint-action coordination of redundant force contributions in a virtual lifting task. Motor Control, 11, 235–258. Brass, M., Bekkering, H., & Prinz, W. (2001). Movement observation affects movement execution in a simple response task. Acta Psychologica, 106, 3–22. Brass, M., & Heyes, C. M. (2005). Imitation: Is cognitive neuroscience solving the correspondence problem? Trends in Cognitive Sciences, 9, 489–495. Bratman, M. (1992). Shared cooperative activity. The Philosophical Review, 101(2), 327–341. Bratman, M. (2009). Modest sociality and the distinctiveness of intention. Philosophical Studies, 144(1), 149–165. Braun, D. A., Ortega, P. A., & Wolpert, D. M. (2009). Nash equilibria in muti-agent motor interactions. PLoS Computational Biology, 5(8), e1000468. Breazeal, C. (2002). Designing sociable robots. Cambridge, MA: MIT Press. Brennan, S. E., Chen, X., Dickinson, C., Neider, M., & Zelinsky, G. (2007). Coordinating cognition: The costs and benefits of shared gaze during collaborative search. Cognition, 106, 1465–1477. Brennan, S. E., & Hanna, J. E. (2009). Partner-specific adaptation in dialogue. Topics in Cognitive Science, 1, 274–291. Brownell, C. A., Ramani, G. B., & Zerwas, S. (2006). Becoming a social partner with peers: Cooperation and social understanding in one- and two-year-olds. Child Development, 77, 803–821. Call, J. (2009). Contrasting the social cognition of human and nonhuman apes: The shared intentionality hypothesis. Topics in Cognitive Science, 1(2), 368–379. Calvo-Merino, B., Glaser, D., Gre`zes, J., Passingham, R., & Haggard, P. (2005). Action observation and acquired motor skills: An fMRI study with expert dancers. Cerebral Cortex, 15, 1243–1249. Carpenter, M. (2009). Just how joint is joint action in infancy? Topics in Cognitive Science, 1(2), 380–392. Castiello, U., Lusher, D., Mari, M., Edwards, M., & Humphreys, G. W. (2002). Observing a human or a robotic hand grasping an object: Differential motor priming effects. In Common mechanisms in perception and action. Series: attention and performance XIX (pp. 315–333). Oxford: Oxford University Press. Chartrand, T. L., & Bargh, J. A. (1999). The chameleon effect: The perception-behavior link and social interaction. Journal of Personality & Social Psychology, 76(6), 893–910. Clark, A. (1997). Being there: Putting brain, body, and world together again. Cambridge, MA: MIT Press. Clark, H. H. (1996). Using language. Cambridge, UK: Cambridge University Press. Clark, H. H. (2005). Coordinating with each other in a material world. Discourse Studies, 7, 507–525.

Joint Action

95

Clayton, M., Sager, R., & Will, U. (2004). In time with the music: The concept of entrainment and its significance for ethnomusicology. ESEM CounterPoint, 11, 96–99. Condon, W. (1976). An analysis of behavioral organization. Sign Language Studies, 13, 285–318. Condon, W. S., & Ogston, M. B. (1966). Sound film analysis of normal and pathological behavior patterns. The Journal of Nervous and Mental Disease, 143(4), 338–347. Cross, E. S., Hamilton, A. F., & Grafton, S. T. (2006). Building a motor simulation de novo: Observation of dance by dancers. Neuroimage, 31, 1257–1267. De Bruijn, E. R. A., Miedl, S. F., & Bekkering, H. (2008). Fast responders have blinders on: ERP correlates of response inhibition in competition. Cortex, 44(5), 580–586. De Maeght, S., & Prinz, W. (2004). Action induction through action observation. Psychological Research, 68(2–3), 97–114. Dehaene, S. (1997). The number sense. New York: Oxford University Press. DeJaegher, H., DiPaolo, E., & Gallagher, S. (in press). Can social interaction constitute social cognition? Trends in Cognitive Sciences. Edwards, M. G., Humphreys, G. W., & Castiello, U. (2003). Motor facilitation following action observation: A behavioural study in prehensile action. Brain and Cognition, 53, 495–502. Flanagan, J. R., & Johansson, R. S. (2003). Action plans used in action observation. Nature, 424, 769–771. Fowler, C. A., Richardson, M. J., Marsh, K. L., & Shockley, K. D. (2008). Language use, coordination, and the emergence of cooperative action. In A. Fuchs & V. Jirsa (Eds.), Coordination: neural, behavioral and social dynamics. Springer. Frischen, A., Loach, D., & Tipper, S. P. (2009). Seeing the world through another person’s eyes: Simulating selective attention via action observation. Cognition, 111(2), 212–218. Galantucci, B. (2009). Experimental semiotics: A new approach for studying communication as a form of joint action. Topics in Cognitive Science, 1(2), 393–410. Georgiou, I., Becchio, C., Glover, S., & Castiello, U. (2006). Different action patterns for cooperative and competitive behaviour. Cognition, 102, 415–433. Gibson, J. J. (1977). The theory of affordances. In R. Shaw & J. Bransford (Eds.), Perceiving, acting, and knowing: Toward an ecological psychology (pp. 67–82). Hillsdale, NJ: Erlbaum. Gilbert, M. (1992). On social facts. Princeton, NJ: Princeton University Press. Goebl, W., & Palmer, C. (2009). Synchronization of timing and motion among performing musicians. Music Perception, 26, 427–438. Gold, N., & Sugden, R. (2007). Collective intentions and team agency. Journal of Philosophy, 104(3), 109–137. Gra¨fenhain, M., Behne, T., Carpenter, M., & Tomasello, M. (2009). Young children’s understanding of joint commitments. Developmental Psychology, 45, 1430–1443. Griffiths, D., & Tipper, S. P. (2009). Priming of reach trajectory when observing actions: Hand-centred effects. Quarterly Journal of Experimental Psychology, 62(12), 2450–2470. Guagnano, D., Rusconi, E., & Umilta`, C. (2010). Sharing a task or sharing space? On the effect of a confederate in action coding. Cognition, 114, 348–355. Haeberle, A., Schuetz-Bosbach, S., Laboissiere, R., & Prinz, W. (2008). Ideomotor action in cooperative and competitive settings. Social Neuroscience, 3(1), 26–36. Haken, H., Kelso, J. A. S., & Bunz, H. (1985). A theoretical model of phase transitions in human hand movements. Biological Cybernetics, 51(5), 347–356. Hamilton, A., Wolpert, D., & Frith, U. (2004). Your own action influences how you perceive another person. Current Biology, 14, 493–498. Harrison, S. J., & Richardson, M. J. (2009). Horsing around: Spontaneous four-legged coordination. Journal of Motor Behaviour, 41, 519–524. Heed, T., Habets, B., Sebanz, N., & Knoblich, G. (2010). Others’ actions reduce crossmodal integration in peripersonal space. Current Biology, 20, 1345–1349.

96


Hommel, B., Colzato, L. S., & van den Wildenberg, W. P. M. (2009). How social are task representations? Psychological Science, 20, 794–798. Hommel, B., Muesseler, J., Aschersleben, G., & Prinz, W. (2001). The theory of event coding (TEC). Behavioral and Brain Sciences, 24, 849–937. Hove, M. J., & Risen, J. L. (2009). It’s all in the timing: Interpersonal synchrony increases affiliation. Social Cognition, 27(6), 949–961. Hutchins, E. (1995). How a cockpit remembers its speeds. Cognitive Science, 19, 265–288. Huygens, C. (1673/1986). The pendulum clock or geometrical demonstrations concerning the motion of pendula as applied to clocks. Ames: Iowa State University Press (R. J. Blackwell, Trans.). Issartel, J., Marin, L., & Cadopi, M. (2007). Unintended interpersonal co-ordination: Can we march to the beat of our own drum? Neuroscience Letters, 441(3), 174–179. James, W. (1890). The principles of psychology. New York, NY: Holt. Jax, S. A., & Rosenbaum, D. A. (2007). Hand path priming in manual obstacle avoidance: Evidence that the dorsal stream does not only control visually guided actions in real time. Journal of Experimental Psychology: Human Perception and Performance, 33, 425–441. Jeannerod, M. (1999). The 25th Bartlett Lecture. To act or not to act: Perspectives on the representation of actions. Quarterly Journal of Experimental Psychology, 52A, 1–29. Jeka, J. J., Kelso, J. A. S., & Kiemel, T. (1993). Spontaneous transitions and symmetry: Pattern dynamics in human four-limb coordination. Human Movement Science, 12, 627–651. Jones, K. S. (2003). What is an affordance? Ecological Psychology, 15, 107–114. Keller, P. E. (2008). Joint action in music performance. In F. Morganti, A. Carassa, & G. Riva (Eds.), Enacting intersubjectivity: A cognitive and social perspective to the study of interactions (pp. 205–221). Amsterdam: IOS Press. Keller, P. E., Knoblich, G., & Repp, B. H. (2007). Pianists duet better when they play with themselves: On the possible role of action simulation in synchronization. Consciousness and Cognition, 16(1), 102–111. Keller, P. E., & Repp, B. H. (2004). When two limbs are weaker than one: Sensorimotor syncopation with alternating hands. Quarterly Journal of Experimental Psychology, 57A(6), 1085–1101. Kendon, A. (1970). Movement coordination in social interaction. Acta Psychologica, 32, 1–25. Kilner, J. M., Paulignan, Y., & Blakemore, S. J. (2003). An interference effect of observed biological movement on action. Current Biology, 13, 522–525. Kilner, J. M., Vargas, C., Duval, S., Blakemore, S.-J., & Sirigu, A. (2004). Motor activation prior to observation of a predicted movement. Nature Neuroscience, 7(12), 1299–1301. Kirschner, S., & Tomasello, M. (2009). Joint drumming: Social context facilitates synchronization in preschool children. Journal of Experimental Child Psychology, 102(3), 299–314. Knoblich, G., & Flach, R. (2001). Predicting the effects of actions: Interactions of perception and action. Psychological Science, 12, 467–472. Knoblich, G., & Jordan, S. (2002). The mirror system and joint action. In M. I. Stamenov & V. Gallese (Eds.), Mirror neurons and the evolution of brain and language (pp. 115–124). Amsterdam: John Benjamins. Knoblich, G., & Jordan, S. (2003). Action coordination in groups and individuals: Learning anticipatory control. Journal of Experimental Psychology: Learning, Memory, & Cognition, 29, 1006–1016. Knoblich, G., & Sebanz, N. (2006). The social nature of perception and action. Current Directions in Psychological Science, 15, 99–104. Knoblich, G., & Sebanz, N. (2008). Evolving intentions for social interaction: From entrainment to joint action. Philosophical Transactions of the Royal Society B, 363, 2021–2031.

Joint Action

97

Knoblich, G., Seigerschmidt, E., Flach, R., & Prinz, W. (2002). Authorship effects in the prediction of handwriting strokes. Quarterly Journal of Experimental Psychology, 55A, 1027–1046. Knuf, L., Aschersleben, G., & Prinz, W. (2001). An analysis of ideomotor action. Journal of Experimental Psychology: General, 130, 779–798. Konvalinka, I., Vuust, P., Roepstorff, A., & Frith, C. D. (2010). Follow you, follow me: Continuous mutual prediction and adaptation in joint tapping. Quarterly Journal of Experimental Psychology, 63, 2220–2230. Kourtis, D., Sebanz, N., & Knoblich, G. (2010). Favouritism in the motor system: social interaction modulates action simulation. Biology Letters, 6, 758–761. Kuhbandner, C., Pekrun, R., & Maier, M. A. (in press). The role of positive and negative affect in the ‘mirroring’ of other persons’ actions. Cognition & Emotion. Kutz, C. (2000). Acting together. Philosophy and Phenomenological Research, 61(1), 1–31. Lakin, J. L., Jefferis, V. E., Cheng, C. M., & Chartrand, T. L. (2003). The Chameleon Effect as social glue: Evidence for the evolutionary significance of nonconscious mimicry. Journal of Nonverbal Behavior, 27, 145–162. Liepelt, R., von Cramon, D. Y., & Brass, M. (2008). What is matched in direct matching? Intention attribution modulates motor priming. Journal of Experimental Psychology: Human Perception and Performance, 34(3), 578–591. Liszkowski, U., Carpenter, M., & Tomasello, M. (2008). Twelve-month-olds communicate helpfully and appropriately for knowledgeable and ignorant partners. Cognition, 108(3), 732–739. Marsh, K. L., Richardson, M. J., & Schmidt, R. C. (2009). Social connection through joint action and interpersonal coordination. Topics in Cognitive Science, 1(2), 320–339. Mechsner, F., & Knoblich, G. (2004). Do muscles matter in bimanual coordination? Journal of Experimental Psychology: Human Perception and Performance, 30, 490–503. Melis, A. P., Hare, B., & Tomasello, M. (2006). Chimpanzees recruit the best collaborators. Science, 311(5765), 1297–1300. Milanese, N., Iani, C., & Rubichi, S. (2010). Shared learning shapes human performance: Transfer effects in task sharing. Cognition, 116(1), 15–22. Miles, L. K., Nind, L. K., & Macrae, C. N. (2009). The rhythm of rapport: Interpersonal synchrony and social perception. Journal of Experimental Social Psychology, 45(3), 585–589. Mottet, D., Guiard, Y., Ferrand, T., & Bootsma, R. J. (2001). Two-handed performance of a rhythmical Fitts task by individuals and dyads. Journal of Experimental Psychology-Human Perception and Performance, 27(6), 1275–1286. Neda, Z., Ravasz, E., Brechte, Y., Vicsek, T., & Barabasi, A.-L. (2000). The sound of many hands clapping. Nature, 403, 849–850. Oullier, O., de Guzman, G. C., Jantzen, K. J., Lagarde, J., & Kelso, J. A. S. (2008). Social coordination dynamics: Measuring human bonding. Social Neuroscience, 3(2), 178–192. Pacherie, E., & Dokic, J. (2006). From mirror neurons to joint actions. Cognitive Systems Research, 7, 101–112. Prinz, W. (1997). Perception and action planning. European Journal of Cognitive Psychology, 9, 129–154. Ramnani, N., & Miall, R. C. (2004). A system in the human brain for predicting the actions of others. Nature Neuroscience, 7(1), 85–90. Reed, K., Peshkin, M., Hartmann, M. J., Grabowecky, M., Patton, J., & Vishton, P. M. (2006). Haptically linked dyads: Are two motor-control systems better than one? Psychological Science, 17(5), 365–366. Richardson, D. C., & Dale, R. (2005). Looking to understand: The coupling between speakers’ and listeners’ eye movements and its relationship to discourse comprehension. Cognitive Science, 29, 1046–1060.

98


Richardson, D. C., Dale, R., & Kirkham, N. Z. (2007). The art of conversation is coordination: Common ground and the coupling of eye movements during dialogue. Psychological Science, 18(5), 407–413. Richardson, D. C., Dale, R., & Shockley, K. (2008). Synchrony and swing in conversation: Coordination, temporal dynamics and communication. In I. Wachsmuth, M. Lenzen, & G. Knoblich (Eds.), Embodied communication. Oxford, UK: Oxford University Press. Richardson, D. C., Dale, R., & Tomlinson, J. M., Jr.. (2009). Conversation, gaze coordination and beliefs about visual context. Cognitive Science, 33, 1468–1482. Richardson, M. J., Campbell, W. L., & Schmidt, R. C. (2009). Movement interference during action observation as emergent coordination. Neuroscience Letters, 449(2), 117–122. Richardson, M. J., Marsh, K. L., & Baron, R. M. (2007a). Judging and actualizing intrapersonal and interpersonal affordances. Journal of Experimental Psychology: Human Perception and Performance, 33, 845–859. Richardson, M. J., Marsh, K. L., Isenhower, R., Goodman, J., & Schmidt, R. C. (2007b). Rocking together: Dynamics of intentional and unintentional interpersonal coordination. Human Movement Science, 26, 867–891. Richardson, M. J., Marsh, K. L., & Schmidt, R. C. (2005). Effects of visual and verbal interaction on unintentional interpersonal coordination. Journal of Experimental Psychology. Human Perception and Performance, 31(1), 62–79. Rizzolatti, G., & Sinigaglia, C. (2010). The functional role of the parieto-frontal mirror circuit: Interpretations and misinterpretations. Nature Reviews Neuroscience, 11, 264–274. Roth, A. S. (2004). Shared agency and contralateral commitments. The Philosophical Review, 113(3), 359–410. Ruys, K. I., & Aarts, H. (2010). When competition merges people’s behavior. When competition merges people’s behavior: Interdependency activates shared action representations, Journal of Experimental Social Psychology, 46, 1130–1133. Samson, D., Apperly, I. A., Braithwaite, J. J., Andrews, B. J., & Bodely Scott, S. E. (in press). Seeing it their way: Evidence for rapid an involuntary computation of what other people see. Journal of Experimental Psychology: Human Perception and Performance. Schmidt, R. C., Bienvenu, M., Fitzpatrick, P. A., & Amazeen, P. G. (1998). A comparison of within- and between-person coordination: Coordination breakdowns and coupling strength. Journal of Experimental Psychology: Human Perception and Performance, 24, 884–900. Schmidt, R. C., Carello, C., & Turvey, M. T. (1990). Phase transitions and critical fluctuations in the visual coordination of rhythmic movements between people. Journal of Experimental Psychology: Human Perception and Performance, 16, 227–247. Schmidt, R. C., Fitzpatrick, P., Caron, R., & Mergeche, J. (in press). Understanding social motor coordination. Human Movement Science. Schmidt, R. C., & O’Brien, B. (1997). Evaluating the dynamics of unintended interpersonal coordination. Ecological Psychology, 9, 189–206. Schmidt, R. C., & Richardson, M. J. (2008). Dynamics of interpersonal coordination. In A. Fuchs & V. Jirsa (Eds.), Coordination: Neural, behavioral and social dynamics. Springer. Schmidt, R. C., & Turvey, M. T. (1994). Phase-entrainment dynamics of visually coupled rhythmic movements. Biological Cybernetics, 70, 369–376. Schuch, S., & Tipper, S. P. (2007). On observing another person’s actions: Influences of observed inhibition and errors. Perception & Psychophysics, 69, 828–837. Searle, J. R. (1990). Collective intentions and actions. In Consciousness and language (pp. 90– 105). Cambridge: Cambridge University Press. Sebanz, N., Bekkering, H., & Knoblich, G. (2006). Joint action: Bodies and minds moving together. Trends in Cognitive Science, 10(2), 70–76.

Joint Action

99

Sebanz, N., & Knoblich, G. (2009). Prediction in joint action: What, when, and where. Topics in Cognitive Science, 1, 353–367. Sebanz, N., Knoblich, G., & Prinz, W. (2003). Representing others’ actions: Just like one’s own? Cognition, 88, B11–B21. Sebanz, N., Knoblich, G., & Prinz, W. (2005). How two share a task: Corepresenting Stimulus–Response mappings. Journal of Experimental Psychology: Human Perception and Performance, 31, 1234–1246. Sebanz, N., Knoblich, G., Prinz, W., & Wascher, E. (2006). Twin Peaks: An ERP study of action planning and control in co-acting individuals. Journal of Cognitive Neuroscience, 18, 859–870. Sebanz, N., Knoblich, G., Stumpf, L., & Prinz, W. (2005). Far from action blind: Representation of others´ actions in individuals with autism. Cognitive Neuropsychology, 22, 433–454. Sebanz, N., & Shiffrar, M. (2007). Bodily bonds: Effects of social context on ideomotor movements. In Y. Rossetti, M. Kawato, & Patrick Haggard (Eds.), Sensorimotor foundations of higher cognition (attention and performance, XXII). Oxford, UK: Oxford University Press. Semin, G. R., & Cacioppo, J. T. (2008). Grounding social cognition: Synchronization, entrainment, and coordination. In G. R. Semin & E. R. Smith (Eds.), Embodied grounding: Social, cognitive, affective, and neuroscientific approaches (pp. 119–147). New York: Cambridge University Press. Semin, G. R., & Smith, E. R. (Eds.), (2008). Embodied grounding: Social, cognitive, affective, and neuroscientific approaches. New York: Cambridge University Press. Shintel, H., & Keysar, B. (2009). Less is more: A minimalist account of joint action in communication. Topics in Cognitive Science, 1, 260–273. Shockley, K. D., Baker, A. A., Richardson, M. J., & Fowler, C. A. (2007). Articulatory constraints on interpersonal postural coordination. Journal of Experimental Psychology: Human Perception and Performance, 33, 201–208. Shockley, K., Butwill, M., Zbilut, J., & Webber, C. (2002). Cross recurrence quantification of coupled oscillators. Physics Letters A, 305, 59–69. Shockley, K., Santana, M. V., & Fowler, C. A. (2003). Mutual interpersonal postural constraints are involved in cooperative conversation. Journal of Experimental Psychology: Human Perception and Performance, 29(2), 326–332. Spence, C., Pavani, F., Maravita, A., & Holmes, N. (2004). Multisensory contributions to the 3-D representation of visuotactile peripersonal space in humans: Evidence from the crossmodal congruency task. Journal of Physiology, 98, 171–189. Spivey, J. M. (2007). The continuity of mind. New York, NY: Oxford University Press. Stanley, J., Gowen, E., & Miall, R. C. (2007). Interference in performed movement during observation of a moving dot stimulus. Journal of Experimental Psychology: Human Perception and Performance, 33, 915–926. Stoffregen, T. A., Giveans, M. R., Villard, S., Yank, J. R., & Shockley, K. (2009). Interpersonal postural coordination on rigid and non-rigid surfaces. Motor Control, 13, 471–483. Streuber, S. (2008). Embodied in cyberspace: Coordination processes in a joint action transportation task. (Unpublished thesis). Stu¨rmer, B., Aschersleben, G., & Prinz, W. (2000). Correspondence effects with manual gestures and postures: A study of imitation. Journal of Experimental Psychology: Human Perception and Performance, 26(6), 1746–1759. Tognoli, E., Lagarde, J., De Guzman, G. C., & Kelso, J. A. S. (2007). From the cover: The phi-complex as a neuromarker of human social coordination. Proceedings of the National Academy of Sciences, 104, 8190–8195.

100


Tollefsen, D. (2005). Let’s pretend! Children and joint action. Philosophy of the Social Sciences, 35, 75–97. Tomasello, M. (2009). Why we cooperate. Cambridge, MA: MIT Press. Tomasello, M., Carpenter, M., Call, J., Behne, T., & Moll, H. (2005). Understanding and sharing intentions: The origins of cultural cognition. Behavioral and Brain Sciences, 28, 675–691. Trevarthen, C. (1979). Communication and cooperation in early infancy: A description of primary intersubjectivity. In M. Bullowa (Ed.), Before speech (pp. 321–343). Cambridge, UK: Cambridge University Press. Tsai, C.-C., & Brass, M. (2007). Does the human motor system simulate Pinocchio’s actions? Psychological Science, 18, 1058–1062. Tsai, C.-C., Kuo, W.-J., Hung, D. L., & Tzeng, O. J.-L. (2008). Action co-representation is tuned to other humans. Journal of Cognitive Neuroscience, 20, 2015–2024. Tsai, C.-C., Kuo, W.-J., Jing, J.-T., Hung, D. L., & Tzeng, O. J.-L. (2006). A common coding framework in self-other interaction: Evidence from joint action task. Experimental Brain Research, 175, 353–362. Tsai, C., Sebanz, N., & Knoblich, G. (in press). The GROOP effect: Groups mimic group actions. Cognition. Tucker, M., & Ellis, R. (1998). On the relations between seen objects and components of potential actions. Journal of Experimental Psychology: Human Perception and Performance, 24(3), 830–846. Valdesolo, P., Ouyang, J., & DeSteno, D. A. (2010). The rhythm of joint action: Synchrony promotes cooperative ability. Journal of Experimental Social Psychology, 46, 693–695. Van Baaren, R. B., Holland, R. W., Kawakami, K., & van Knippenberg, A. (2004). Mimicry and pro-social behavior. Psychological Science, 15, 71–74. Van Baaren, R. B., Holland, R. W., Steenaert, B., & van Knippenberg, A. (2003). Mimicry for money: Behavioral consequences of imitation. Journal of Experimental Social Psychology, 39, 393–398. Van Baaren, R., Janssen, L., Chartrand, T., & Dijksterhuis, A. (2009). Where is the love? The social consequences and moderators of mimicry in humans. Philosophical Transactions of the Royal Society B, 2381–2389. Van Der Wel, R. P., Fleckenstein, R., Jax, S., & Rosenbaum, D. A. (2007). Hand path priming in manual obstacle avoidance: Evidence for abstract spatio-temporal forms in human motor control. Journal of Experimental Psychology: Human Perception and Performance, 33, 1117–1126. Van der Wel, R. P. R. D., Knoblich, G., & Sebanz, N. (in press). Let the force be with us: Dyads exploit haptic coupling for coordination. Journal of Experimental Psychology: Human Perception and Performance. Van Schie, H. T., Mars, R. B., Coles, M. G. H., & Bekkering, H. (2004). Modulation of activity in medial frontal and motor cortices during error observation. Nature Neuroscience, 7(5), 549–554. Van Schie, H. T., Waterschoot, B. M., & Bekkering, H. (2008). Understanding action beyond imitation: Reversed compatibility effects of action observation in imitation and joint action. Journal of Experimental Psychology: Human Perception and Performance, 34(6), 1493–1500. Van Ulzen, N. R., Lamoth, C. J., Daffertshofer, A., Semin, G. R., & Beek, P. J. (2008). Characteristics of instructed and uninstructed interpersonal coordination while walking in pairs. Neuroscience Letters, 432(2), 88–93. Vesper, C., Butterfill, S., Knoblich, G., & Sebanz, N. (2010). A minimal architecture for joint action. Neural Networks, in press.

Joint Action

101

Vesper, C., Soutschek, A., & Schuboe, A. (2009). Motion coordination affects movement parameters in a joint pick-and-place task. Quarterly Journal of Experimental Psychology, 62 (12), 2418–2432. Vogeley, K., & Roepstorff, A. (2009). Contextualizing culture and social cognition. Trends in Cognitive Sciences, 13, 511–516. Von Uexkuell, J. (1920). Theoretische Biologie [Theoretical biology]. Berlin, Germany: Springer. Wachsmuth, I., & Knoblich, G. (Eds.), (2008). Modelling communication with robots and virtual humans. Berlin: Springer. Wachsmuth, I., Lenzen, M., & Knoblich, G. (Eds.), (2008). Embodied communication. Oxford: Oxford University Press. Warneken, F., & Tomasello, M. (2007). Helping and cooperation at 14 months of age. Infancy, 11(3), 271–294. Welsh, T. N. (2009). When 1 þ 1 ¼ 1: The unification of independent actors revealed through joint Simon effects in crossed and uncrossed effector conditions. Human Movement Science, 28(6), 726–737. Welsh, T. N., Elliott, D., Anson, G. G., Dhilion, V., Weeks, D. J., Lyons, J. L., & Chua, Romeo (2005). Does Joe influence Fred’s action? Inhibition of return across different nervous systems. Neuroscience Letters, 385(2), 99–104. Welsh, T. N., Higgins, L., Ray, M., & Weeks, D. J. (2007). Seeing vs. believing: Is believing sufficient to activate the processes of response co-representation? Human Movement Science, 26, 853–866. Wilson, M., & Knoblich, G. (2005). The case for motor involvement in perceiving conspecifics. Psychological Bulletin, 131, 460–473. Wiltermuth, S. S., & Heath, C. (2009). Synchrony and cooperation. Psychological Science, 20, 1–5. Wolpert, D. M., Doya, K., & Kawato, M. (2003). A unifying computational framework for motor control and interaction. Philosophical Transactions of the Royal Society of London B, 358, 593–602. Woodworth, R. S. (1939). Individual and group behaviour. The American Journal of Sociology, 44(6), 823–828. Yoon, E. Y., Humphreys, G. W., & Riddoch, M. J. (in press). The paired-object affordance effect. Journal of Experimental Psychology: Human Perception and Performance.


C H A P T E R

F O U R

Self-Regulated Learning and the Allocation of Study Time John Dunlosky and Robert Ariel Contents 1. Introduction 2. Self-Regulated Learning 2.1. Definition and Examples 2.2. A Metacognitive Approach to SRL 3. Allocation of Study Time 3.1. Modal Method, Dependent Measures, and a Typical Outcome 3.2. Reversals of the Modal Outcome and an Updated Literature Review 4. Agenda-Based Regulation Framework 4.1. Framework Description 4.2. Accounting for Item Selection 4.3. Accounting for Self-Paced Study Times 4.4. Summary 5. Comparing Accounts of Study-Time Allocation 5.1. Hierarchical Model of Self-Paced Study 5.2. RPL Framework 5.3. COPES Model 6. Concluding Remarks and Some Directions for Future Research Acknowledgments References

104 105 105 107 108 108 110 120 120 125 129 131 131 132 132 134 135 136 136

Abstract We explore how students allocate their study time—both with regard to which materials they select to study and how long they study them. A literature review revealed multiple outcomes that any theory of study-time allocation will need to explain: students often spend more time studying difficult to learn items (vs. less difficult ones), unless (a) little time is available for study or (b) the reward for correct recall is higher for the less difficult items. In the latter contexts, this shift to focusing on easier items is an effective strategy, and notably, students do not always make this shift when doing so would be effective. To account for these and other effects, we introduce the agendabased-regulation (ABR) framework, which assumes that students use agendas Psychology of Learning and Motivation, Volume 54 ISSN 0079-7421, DOI: 10.1016/B978-0-12-385527-5.00004-8

#


103

104

John Dunlosky and Robert Ariel

to allocate time to efficiently achieve their goals. We compare the ABR framework to other theories of study-time allocation and discuss avenues for future research.

1. Introduction Every person—every living organism—self-regulates, whether that involves the unintelligent (but adaptive) behavior of single-cell organisms honing in on a nutrient source or the calculated behavior of a student who is attempting to master calculus (for a general analysis, see Mithaug, 1993). Self-regulation itself is triggered by the detection of a discrepancy between the current state of a system and a goal state, and subsequent behavior is directed toward reducing this discrepancy. Consider the Escherichia coli bacterium, which is a single-cell organism that shows remarkable regulatory skills as it hones in on food. When nutrients are depleted, the bacterium tumbles randomly until it senses nutrient, and once it does, the tumbles decrease in frequency and it moves longer on the current path toward a potentially even greater concentration of nutrient (an observation of biologist Max Perutz, from Calvin, 1990). Although E. coli’s behavior may seem purposeful, its self-regulatory feat is made possible by “some inherited simple abilities such as swimming, tumbling, and sensing increasing yield” (Calvin, 1990, p. 32). In terms of self-regulated learning (SRL), students enrolled in a course are confronted with discrepancy once they receive the course syllabus. Discrepancy arises from each student’s perceived state of mastery of the course content and his or her expectations for achievement in the course. In a beginning college course for learning a new language, students may differ in their a priori knowledge about the language and may have different expectations for achievement, so the perceived discrepancies will differ and yield different patterns of self-regulation. Nevertheless, such discrepancy is necessary for regulation—students who can already speak French fluently would not study (or even take a course to learn French), unless other goals produced a discrepancy that studying would resolve. Perhaps their goal is to make extra money tutoring other students, so they decide to study the course examples to achieve this goal. Even in this case, regulation is triggered by discrepancy between the current state (no extra money, except for room and board) and a goal state (obtain enough money to pay for extracurricular activities). These examples stimulate numerous questions about human selfregulation: How do people attempt to regulate themselves to reduce perceived discrepancies? Is their regulation optimal? That is, would more regulation produce no extra gains and would less regulation produce fewer

Self-Regulated Learning and the Allocation of Study Time

105

gains (Mithaug, 1993)? If people’s regulation is suboptimal, what factors limit their optimality, and are some people better at regulating their behavior (and hence more likely to achieve their sought-after goals) than others? To address these questions in this chapter, we first define SRL in general and describe our current approach to answering these questions, which is informed by theory and data concerning the metacognitive control of learning. After our general overview, we provide a comprehensive review and theoretical analysis of one important aspect of SRL—students’ allocation of study time.

2. Self-Regulated Learning 2.1. Definition and Examples An act of SRL is any student behavior or cognition that is directed toward reducing a discrepancy between a current perceived state and a goal relevant to performance or learning.1 This definition for SRL was developed from system theories of self-regulation (Powers, 1973; for other perspectives, see Zimmerman & Schunk, 2001) and has several important implications. The first is that SRL is goal-oriented, because regulation involves reducing a discrepancy between internal states: a perceived state and a goal state. For instance, a student may want to recall the central theories of attention for an introductory course on cognition (a goal) but may be having difficulties recalling them (a perceived state). For a student, a discrepancy is not sufficient to stimulate self-regulation, because they often have many unmet goals, which cannot all be pursued at the same time. But, once a student takes on a given task, it is the student’s current perceived state and goal for it that will trigger and drive self-regulation. Second, the discrepancies that confront students often require planning to minimize; put differently, students will usually need to plan to achieve their goals, unless of course their goals are to merely pass an easy class (but even for this goal, self-regulation is required to attend class for exams). Self-regulation does not necessitate planning, but the kind of SRL that students often engage in will be most effective when guided by planning. Preparing for an exam will require some plan of attack, even if it means planning an all-night study session on the evening before the exam. Unfortunately, this plan is a popular one (Taraban, Maki, & Rynearson, 1999), and in this single cram session, SRL can be pushed to the maximum as a student decides how much to study

1

A psychological distinction exists between learning and performance goals (i.e., students who focus on the process of learning versus those who attempt to achieve a particular level of performance). For brevity, however, we used the term learning goal, unless we were discussing test performance per se.

106


various topics (e.g., those that the teacher emphasized), how much time should be spent reading versus taking practice tests, and so forth. Thus, planning is a fundamental aspect of effective SRL, just as it is for any other form of problem solving (e.g., Friedman, Scholnick, & Cocking, 1987; Miller, Galanter, & Pribram, 1960; Newell & Simon, 1972). Third, by constraining what counts as SRL, the definition itself reveals the breadth of the concept, namely, SRL includes any activity directed toward achieving some learning-relevant goal, whether it be mastering a concept in an independent study or achieving a particular grade in a required course. A few of the activities that students can use to achieve their course-related goals are listed in Table 1. Each activity involves subgoals (and hence discrepancies) that require regulating different behaviors: for successfully attending class, busy students may need to schedule courses at convenient times and rearrange work schedules. For completing homework assignments, students will need to plan when to complete each one and develop goals for completing them. Despite these differences, all the activities on this list can be engaged in the pursuit of learning. Viewed in this manner, achieving a more general goal (e.g., earning an A in a chemistry class) requires the coordination of many interrelated

Table 1

Activities That Involve Self-Regulated Learning.

Activities

Prestudy activities Goal setting Attending class Taking notes Completing in-class activities Time management Study activities Using comprehension and memory strategies Collaborative studying Allocating study time Self-testing Asking questions Seeking help from tutors/ teacher Completing homework assignments

Core readings into the relevant literature

Morisano, Hirsh, Peterson, Pihl, and Shore (2010) Clump, Bauer, and Whiteleather (2003) Pressley, Van Etten, Yokoi, Freebern, and Van Meter (1998) Blood and Neel (2008) Britton and Tesser (1991) McNamara (2010), Richardson (1998) Jeong and Chi (2007) Dickinson and O’Connell (1990) Crede´ and Kuncel (2009) Graesser, McMahen, and Johnson (1994) Azevedo, Moos, Greene, Winters, and Cromley (2008) Dettmers, Trautwein, Ludtke, Kunter, and Baumert (2010)


107

activities. Success at one activity often is necessary for successfully engaging in other activities. Students who fail to attend class cannot take notes, and without adequate notes, students may not be able to devise appropriate questions about their misunderstanding or may have difficulties prioritizing materials for study. Despite the interplay of these activities, each one is often investigated in isolation, and even some vital activities of SRL—for example, time management and collaborative study—have received little attention in the field. Admittedly, our own research on SRL suffers from isolationism in that we have almost exclusively investigated study-time allocation; nevertheless, the resulting agenda-based regulation (ABR) framework, which is introduced below, integrates metacognitive, cognitive, and social constructs in a manner that could be modified to investigate other SRL activities (Table 1) and their interplay. Finally, the definition of SRL and its implications can be used to generate interrelated and testable hypotheses about why students fail to achieve their goals. In general, students are expected to underachieve when they (a) fail to set goals, (b) do not remember their goals, (c) do not develop plans to achieve their goals, (d) develop ineffective plans, or (e) do not execute their effective plans properly. Concerning hypothesis testing, when students do not regulate effectively, one can understand this dysregulation by systematically investigating the contribution of the sources above. For instance, Dunlosky and Thiede (2004) observed that many students regulated their study time suboptimally, and then demonstrated that this dysregulation arose from failures to plan and from difficulties in executing plans when they were developed. Many other factors can contribute to poor SRL, such as a lack of motivation, inadequate declarative or procedural knowledge, and deficiencies in fluid intelligence and workingmemory capacity. Thus, as we discuss in further detail below, we advocate an approach that centers SRL within an information-processing system, because joint investigation of the contribution of these factors will be essential for a full understanding of student success.

2.2. A Metacognitive Approach to SRL Metacognitive processes—which include people’s monitoring and control of their cognitions—are central to our approach for understanding SRL and allocation of study time. Even though self-regulation itself does not require that regulators are conscious of their ongoing efforts (recall the E. coli example above), we assume that much of students’ SRL involves selfawareness and reflection—and hence, explicit metacognitive processes. In fact, although our definition of SRL is deliberately general, at least some scientists have included “metacognitive” in their definitions of selfregulation. Zimmerman (2001) notes that “students are self-regulated to

108


the degree that they are metacognitively, motivationally, and behaviorally active participants in their own learning processes” (p. 5). Our general definition does not demand that students are metacognitively active, but the emphasis on “reducing discrepancies between perceived states and goals” does suggest the utility of conscious monitoring and control processes. After all, students can explicitly monitor their ongoing progress to form a perception—or evaluation—of their current state, and then make explicit control decisions to continue reducing the discrepancy. Control processes may involve any of the activities listed in Table 1. Students may realize that they are making no progress understanding a concept in a physics class, and subsequently control their learning by completing more homework, seeking help from tutors, or even asking classmates questions about the concept. These activities will also involve an interplay between monitoring and control processes as students complete the activity on their way to achieving a larger learning goal. Successful SRL will depend on the coordination of metacognitive and cognitive processes along with the motivational orientation of a given student. Accordingly, in our model for the allocation of study time described below, we embed metacognitive processes within an information-processes system that describes how selective attention is constrained by the structure of memory (cf. Winne, 2001). Our ABR framework was developed to account for students’ allocation of study time, but the framework can be adapted to investigate any activity relevant to SRL. Before we introduce this framework, we first describe what inspired it: the modal method used to investigate students’ allocation of study time and some key outcomes that it has supported.

3. Allocation of Study Time 3.1. Modal Method, Dependent Measures, and a Typical Outcome The following method (and minor adaptations to it) has been most often used to investigate the allocation of study time. To-be-remembered items—usually paired associates (e.g., dog - spoon)—are presented individually at a fixed presentation rate. This initial study trial familiarizes the participants with the list of items and is often used to collect monitoring judgments. Typically, a participant will make judgment of learning (JOL) immediately after studying each item. This JOL is a judgment of the likelihood that an item will be later remembered on the test, and it measures participants’ perceived state of learning. After this initial trial, each item is presented again, and the participant either selects each item for restudy (called item selection) or studies each item as long as wanted


109

Table 2 Common Measures of Study-Time Allocation.

Overall allocation Number of items selected Amount of time studying per/item studied Across-item allocation Correlationa between initial item difficultyb and item selection Correlation between initial item difficulty and self-paced study Order that items are selected for studyc Measures involving test performance Mean test performance as a function of item selection Correlation between initial item difficulty and test performance Correlation between initial item selection and test performance Correlation between initial self-paced study times and test performance a b

c

All correlations are first computed across each individual participant’s values, and then the mean across correlations is reported. Item difficulty has been estimated in multiple ways—normative learning across items, individual participant’s judgments of learning (as described in the text), or learning on a test trial that occurs immediately after the initial (experimenter-paced) familiarization trial. The order of item selection is often compared to item difficulty, such as which level of item difficulty (easy or difficult) was more often chosen first for study (designated as “FC” in the column “Allocated study time to” in Table 3).

(called self-paced study).2 If participants are selecting items for study, then the selected items are usually presented again for self-paced study. Finally, a criterion test is administered. This modal method generates a rich data set, and some measures used to explore the allocation of study time are presented in Table 2. Although each of these measures is relatively common in any one article, only a subset of them (which are relevant to a focal question) are presented. For instance, if the question concerns whether reward value influences self-paced study time and test performance, then one would manipulate reward across items and compute mean time studying (and learning) as a function of reward. Another common question is, how do learners use monitoring to allocate study time? To answer this question, researchers have examined the correlational measures computed across items (middle, Table 2), such as correlating people’s JOLs made on the initial trial with subsequent item selection or self-paced study time. Given that each measure describes a somewhat 2

An exciting new wave of studies is exploring the kinds of practice schedule that learners use while studying, such as whether they prefer spacing or massing their study (Benjamin & Bird, 2006; Pyc & Dunlosky, 2010; Son, 2004; Toppino & Cohen, in press) and whether they prefer to reschedule practice test trials or restudy trials (Karpicke, 2009). In this chapter, we are mainly interested in item selection and self-paced study, but the ABR framework can be adapted to understand these other aspects of study-time allocation. In fact, Toppino and Cohen (2010) recently concluded that students use agendas when deciding whether to mass or space restudy trials.

110


different aspect of the relations among monitoring, control, and learning, the measures should not be conflated. Of all the measures listed in Table 2, those involving item difficulty (under “across-item allocation”) currently have had the largest impact on theory development. In a review of the early literature, Son and Metcalfe (2000) identified 46 experimental conditions that supported an analysis of the relationship between item difficulty and self-allocated study. Thirty-five of these conditions revealed a negative relationship between item difficulty and allocation. That is, as measures of item difficulty increased (i.e., items were more difficult to learn), people either were more likely to select items for study or spent more time studying them. This negative relationship in the early literature provided the empirical support for the monitoring-affectscontrol hypothesis (Nelson & Leonesio, 1988), which is that people monitor their ongoing learning and use this monitoring to decide when to stop studying. A specific instantiation of this hypothesis is that people set a single goal for learning and then continue studying a given item until the discrepancy between the current perceived state and the goal has been entirely reduced. This discrepancy-reduction model of study-time allocation offers a straightforward explanation. For example, assuming that difficult items (vs. less difficult ones) will require more study time to reach the set goal, then a prediction from this model is that people will allocate more study time to the difficult than less difficult items.

3.2. Reversals of the Modal Outcome and an Updated Literature Review Soon after this discrepancy-reduction model was introduced (Dunlosky & Hertzog, 1998), its inability to provide a complete account of study-time allocation became apparent. The disconfirming evidence was discovered using a straightforward adaptation of the modal method. Thiede and Dunlosky (1999) had participants study 30 paired-associate items. After an initial study and judgment phase, some participants were instructed to learn at least 24 pairs; that is, they had a high-learning goal, which was the kind of goal instruction that was (either implicitly or explicitly) used in the research cited in Son and Metcalfe (2000). Other participants were instructed that they needed to learn 6 of the 30 items. Immediately after they read the instructions, the stimulus (e.g., dog -?) of each pair was presented simultaneously in an array. That is, during selection, a simultaneous format displays the to-be-learned items together and hence participants can inspect them all during selection; by contrast, under a sequential format, each item is presented individually, and participants must decide whether to restudy each item as it appears. As we will discuss in detail later, the presentation format (i.e., simultaneous vs. sequential) moderates some study-time outcomes. For the present investigation (Thiede & Dunlosky, 1999), the simultaneous


111

format was used, and participants could select any item that they wanted to restudy. Participants with a high-learning goal demonstrated a negative relationship between JOLs and item selection (M correlation 0.37), which was consistent with the bulk of previous outcomes. By contrast, participants with a low-learning goal showed the opposite pattern: They selected more of the easier than difficult items for restudy (M þ0.38). Under both instructions, learners used their monitoring to allocate study time, but they used it in a qualitatively different manner. The shift in studytime allocation—from focusing on the difficult items to focusing more on the easier items—is called the shift-to-easier-materials (STEM) effect. In this chapter, STEM effect is used more generally to refer to any condition in which learners allocate more time to the easier-to-learn materials than the more difficult ones. STEM effects cannot be readily explained by the discrepancy-reduction model (Dunlosky & Hertzog, 1998), which predicts that people will always spend the most time with the more difficult (than less difficult) materials.3 Moreover, disconfirmation of this discrepancy-reduction model as the sole contributor to study-time allocation and further exploration of STEM effects have led to new and more comprehensive theories of study-time allocation (e.g., Ariel, Dunlosky, & Bailey, 2009; Metcalfe, 2009). Given that the presence and absence of STEM effects have had the largest impact on theory development, we reviewed relevant outcomes since Son and Metcalfe (2000). Our review (Table 3) was modeled from Table 1 in Son and Metcalfe (2000), so by combining both tables, interested readers can obtain a bird’s-eye view of the literature since Rose Zacks’s (1969) seminal paper. In contrast to Son and Metcalfe (2000), we separated outcomes relevant to item selection (top half of Table 3) and self-paced study time (bottom half), because these two aspects of study-time allocation may recruit a different subset of mechanisms. Also, unless otherwise noted, all participants were either implicitly or explicitly instructed to obtain a high-learning goal—that is, participants were expected to do their best to allocate study time to learn as many items as possible. Finally, our focus in this review—as in Son and Metcalfe—was to highlight conditions that led 3

The discrepancy reduction model of study-time allocation proposed by Thiede and Dunlosky (1999; see also Dunlosky & Hertzog, 1998) is a constrained version of general discrepancy reduction theory, which is derived from cybernetic theory. A key assumption of general discrepancy reduction theory is that a system attempts to reduce the discrepancy between its current state and a goal state. The discrepancy reduction model of study-time allocation is constrained in that the stopping criterion for the system is static. The system continues to operate (in this case, study) until the discrepancy between the current state and the (static and predetermined) goal state is completely reduced. Though many systems have this type of stopping criterion (e.g., thermostat regulating temperature and cruise control regulating speed in a car), it is not a necessary component of more complex systems that are concerned with optimizing gain while minimizing effort. Detecting discrepancy between goal states and current states is still important to trigger regulation, but these systems can terminate without reaching the goal state, such as when limited progress is being made toward the goal. Self-regulated learning in general is better characterized by this less constrained type of discrepancy reducing system.

Table 3 A Review of the Study-Time Allocation Literature Since Son and Metcalfe (2000).

Article

Participantsa

Item selection Ariel et al. (2009) Experiment 1

Adults

Condition/group

Formatb

Materials

Determinant of difficultyc

Allocated study time tod

High reward to easy items Sim E–E pairse Concrete/abstract Easy High reward to difficult items Difficult Same reward to all items Null Experiment 2 Adults High reward to easy items Sim Concrete/abstract Easy High reward to difficult items Difficult Same reward to all items Null Experiment 3 Adults High reward to easy items Sim Related/unrelated Easy High reward to difficult items Difficult Same reward to all items Null Experiment 4 Adults High reward to easy items Sim Concrete/abstract Easyf High reward to difficult items Difficult High reward to easy items Seq Easy High reward to difficult items Difficult Dunlosky and Thiede (2004): In all experiments, participants received a low-learning goal (recall at least 6 of the 30 items) Experiment 1 Adults Given plan to select 6 easy items Sim E–E pairs Concrete/abstract Easy Not given plan Easy Given plan to select 6 easy items Seq Null Not given plan Difficult Experiment 2 Adults Not given plan Sim Easy Experiment 3 Adults High working-memory span Seq Easy Low working-memory span Null

Finn (2008) Experiment 2

Adults

Experiment 3

Adults

JOLs framed about remembering JOLs framed about forgetting JOLs framed about remembering JOLs framed about forgetting

Seq

E–E pairs

JOLs

Difficult

Seq

E–E pairs

JOLs

Difficult Difficult

Kornell and Metcalfe (2006) JOLs Experiment 3a Adults Seq E–S pairse Experiment 3b Sim E–S pairs JOLs Kratzig and Arbuthnott (2009): The JOL-selection correlation was stronger for the youngest cohort Experiment 2 18–28 Trials 1 and 2 Seq E–E pairs JOLs 67–74 Trials 1 and 2 75–90 Trials 1 and 2 Metcalfe and Finn (2008) Experiment 1 Adults 1–3 condition Seq E–E pairs JOLs 3–1 condition Seq E–E pairs JOLs Experiment 2 Adults Immediate JOL 1–3 condition Seq E–E pairs JOLs Immediate JOL 3–1 condition Delayed JOL 1–3 condition Delayed JOL 3–1 condition Experiment 3 Adults Forward associations Seq Backward associations

Difficult Easyf Easy Difficult Difficult Difficult Difficult Difficult Difficult Difficult Difficult Difficult Difficult Difficult (continued)

Table 3 Article

(continued) Participantsa

Condition/group

Formatb

Materials


Metcalfe and Kornell (2003): Item selection was collapsed across study-time conditions by the authors Experiment 1 Adults Sim E–S pairs Norms Metcalfe and Kornell (2005): Only items not recalled during initial trial included for selection Experiment 2 Adults Seq Trivia Norms questions Experiment 5 Adults Sim E–S pairs Norms Experiment 6 Adults Seq E–S pairs Norms Price, Hertzog, and Dunlosky (2009) Experiments 1 & 2 Adults Sim E–S pairs Norms Older adults Rhodes and Castel (2009) Experiment 2 Adults Auditory presentation of words Seq Nouns JOLs in high or low volume Wasylkiw, Tomes, and Smith (2008) Experiment 3 Adults Five items tested and selected Not Cognitive JOLs reported science 10 items tested and selected Definitions 15 items tested and selected Self-paced study Ariel et al. (2009) Experiment 1 Adults High reward to easy items Simsub E–E pairs Concrete/abstract High reward to difficult items Same reward to all items


Easy FCg,h Difficult Easy FCh Easy Moderateh Moderateh Difficult

Difficult Difficult Difficult

Easy Difficult Null

Experiment 2

Adults

Experiment 3

Adults

High reward to easy items High reward to difficult items Same reward to all items High reward to easy items High reward to difficult items Same reward to all items

Hines, Touron, and Hertzog (2009) Experiment 1 Adults Older adults Koriat (2008): Four study-test trials Experiment 3 Adults Trial 1 Trial 2 Trial 3 Trial 4 Koriat and Ackerman (2010) Experiment 1 Self-paced groups Experiment 2 Koriat, Ackerman, Lockl, and Schneider (2008) Experiment 1 Grade school Grade 1 Grade 2 Grade 3 Grade 5 Grade 6

Simsub

Concrete/abstract

Easy Difficult Null Related/unrelated Null Difficult Null

Simsub

Seqall

E–E pairs

Prior recognition learning

Difficult Difficult

Seqall

H–H pairs

JOLs

Difficult Difficult Difficult Difficult

Seqall

E–E pairs E–E pairs

JOLs

Difficult Difficult

Seqall

H–H pairs

JOLs, norms

Null Null Difficult Difficult Difficult (continued)

Table 3 Article

(continued) Participantsa

Condition/group

Formatb

Koriat and Ma’ayan (2005) Experiment 1 Adults All delay conditions Seqall Koriat and Nussinson (2009): All study occurred under a time pressure Experiment 2 Adults Mental effort condition Seqall Control condition Koriat et al. (2006) Experiment 1 Adults Seqall Experiment 2 Self-paced group Experiment 3 Immediate JOLs Delayed JOLs Experiment 4 Four self-paced presentations Experiment 5 Differential incentive Constant incentive Experiment 6 Differential incentive—time constraint Constant incentive—time constraint Lockl and Schneider (2004) Experiment 1 First graders Emphasize speed Seqall Emphasize accuracy Third graders Emphasize speed Emphasize accuracy

Materials



H–H pairs

JOLs

Difficult

H itemse

Related/unrelated Easy Easy

H–H pairs

JOLs, norms

H items

JOLs

Difficult Difficult Difficult Difficult Difficult Difficult Difficult Easy

JOLs

Easy

E–E pairs

Related/unrelated Null Null Null Null

Metcalfe (2002) Experiment 1

Adults

5 s for study of three-item array Simsub

S–E pairs

Norms

15 s allowed for study Unlimited time for study Experiment 2

Spanish experts All time conditions

Experiment 3

Sixth graders

Experiment 4

Sixth grade experts

5 s for study of three-item array 15 s allowed for study Unlimited time for study 5 s for study of three-item array

Difficulth

15 s and unlimited time Experiment 5: 5 s study time with four study-test trials Adults Trial 1 Trials 2–4 Metcalfe and Kornell (2003) Experiment 1 Adults

Metcalfe and Kornell (2005) Experiment 7 Adults

Easy, moderateh Moderateh Moderate, difficulth Moderate, difficulth Easy, moderateh Nullh Nullh Nullh

Nullh Easy, moderateh

5 s for study of three-item array Sim 15 allowed for study Unlimited time for study

S–E pairs

JOLs

Moderateh Moderateh Moderate, difficulth

Learners with long study times Seqall Learners with short study times

S–E pairs

JOLs

Difficult Difficult (continued)

Table 3

(continued)

Article

Price et al. (2009) Experiment 1 Experiment 2

Participantsa

Condition/group

Adults Older adults Adults Older adults

Thomas and McDaniel (2007) Experiment 2 Adults Tiede and Leboe (2009) Experiment 1 Adults

All encoding tasks

Formatb

Materials



Simsub

S–E pairs

Norms

Simsub

S–E pairs

Norms

Difficulth Difficulth Difficulth Moderate, Difficulth

Seqall

Expository texts

JOLs

Difficult

Seqall

E–E pairs

Related/unrelated Difficult

JOLs, judgments of learning. Note. Any empty cell for a given article/experiment takes the value of the filled cell immediately above it. a Adult refers to college students. b Presentation format of items during item selection or self-paced study: sim(ultaneous), items presented simultaneous in an array; seq(uential), items presented individually. For self-paced study, (a) SIM refers to the fact that all items were presented in an array for selection, and then pacing occurred for selected items and (b) the outcomes are based either on study times for a subset of all items (i.e., usually those that were selected for study), which is indicated by the superscript “sub,” or for all items, which is indicated by the superscript “all.” c Refers to how item difficulty was measured for a given set of materials. d Refers to preference of allocation: easy refers to more time being spent on easy than difficult items (i.e., a STEM effect), difficult refers to more time being spent on the more difficult items (i.e., the modal outcome from the pre-2000 literature; Son & Metcalfe, 2000), moderate refers to preferring items of moderate difficulty, and null refers to no preference for allocating items to a specific level of difficulty. e E–E pairs, English–English stimulus–response pairs; E–S pairs, English–Spanish translation equivalents; H–H pairs, Hebrew–Hebrew pairs. f STEM effect was larger under the simultaneous than sequential format—see text for details (cf. Figure 1). g FC indicates that the measure of allocation was the difficulty level of the item that were most often the learners’ first choice (FC). h In these conditions, item difficulty covaried with item order (easiest items were in the prominent position of arrays), so the relationship between difficulty and allocation may in part be due to item ordering (see Section 4.4 for details).

119


learners to prefer easy (or difficult) items during selection and self-paced study. Some manipulations in these studies were relevant to investigating a different issue, so we do not present them in the table; the target article must be consulted for details. These reviews (from Son and Metcalfe and Table 3) include some notable results. Consider item-selection effects, which are presented in the top half of Table 3. First, when items are presented under a simultaneous format for selection (Sim, under the “Format” column), a STEM effect (an “Easy” in the column labeled “Allocated study time to”) occurs when the reward for correctly recalling items is higher for easy than difficult items. This STEM effect is illustrated by the two left-most bars in Figure 1. Also, under a simultaneous format, a STEM effect is more likely to occur when people have a limited time for study (e.g., when they do not have enough time to study all items) than when they have unlimited time (Thiede & Dunlosky, 1999). Second, all else equal, a STEM effect is more likely to occur when items are presented simultaneously in an array (Sim) than sequentially (Seq; e.g., Ariel et al., 2009; Dunlosky & Thiede, 2004; a nonsignificant trend occurred in Kornell & Metcalfe, 2006). When a STEM effect does occur under a sequential format, it is typically weaker than the corresponding STEM effect under a simultaneous format. This interaction with presentation format is illustrated in Figure 1: When easy items were slated to return more reward, learners selected fewer easier items (and more of the difficult items) for study under a sequential format than under a simultaneous format (Ariel et al., 2009).

Proportion of items selected

Easy

1 0.9

Simultaneous format

Difficult

Sequential format

1 0.9

0.8

0.8

0.7

0.7

0.6

0.6

0.5

0.5

0.4

0.4

0.3

0.3

0.2

0.2

0.1

0.1

0

High reward easy

High reward difficult

High reward easy

High reward difficult

0

Figure 1 Proportion of items selected when items are presented simultaneously for selection (simultaneous format) and one item at a time (sequential format) for selection. High reward easy indicates that the easier items were slated to return a higher reward if correctly recalled and difficult items were slated to return a lower reward. High reward difficult indicates that the difficult items were slated to return the higher reward. From Ariel et al. (2009).

120


As compared to item selection, self-paced study manifests fewer STEM effects, even under the same conditions. Put differently, when pacing study, students usually spend more time studying the more difficult-to-learn items than the less difficult items. This empirical generalization has boundary conditions, which include (a) conditions in which easy items are slated with a high reward and difficult items are slated with a low reward (Ariel et al., 2009) and (b) when the total time allowed for study is relatively brief given the total number of items that could be studied (Koriat, Ma’ayan, & Nussinson, 2006; Koriat & Nussinson, 2009; Metcalfe, 2002). Finally, remember that Table 3 just includes the outcomes that are relevant to the relationship between item difficulty and study-time allocation. Other effects have also informed theory development. For instance, the effects of incentives on study-time allocation have been explored in several studies (beginning with Le Ny, Denhie`re, & Le Taillanter, 1972) and have implications for how people pace their study. In Sections 4.2 and 4.3, we discuss other effects as they become relevant to theories of study-time allocation.

4. Agenda-Based Regulation Framework 4.1. Framework Description We now describe the ABR framework in detail, which was originally developed to account for the presence and absence of STEM effects and the overall flexibility of students’ study-time allocation. As a framework, the constructs and principles described next can be used post hoc to develop competitive hypotheses that can account for any allocation phenomena. The framework constrains these hypotheses to focus on the potential contribution of a subset of causal constructs (e.g., agendas, working-memory capacity, etc.), which according to the framework, are vital for effective study-time allocation. Thus, the fertility of the ABR framework will lie in its ability to guide the development and evaluation of more specific hypotheses that can offer a priori explanations for and predictions about study-time allocation. According to the ABR framework, “learners develop an agenda on how to allocate time to various study items and use this agenda when selecting items for study. Like many other theories of regulation (e.g., Benjamin, 2007; Carver & Scheier, 2000; Pintrich, 2000), the ABR model assumes that study regulation is goal-oriented” (Ariel et al., 2009, p. 433). A critical assumption is that when learners develop an agenda, they do so to maximize the likelihood of efficiently obtaining their goals (Thiede & Dunlosky, 1999). An agenda—or plan—describes the criteria used to decide which items to select for study or how long to study a given item, and the criteria


121

are chosen based on any number of learner characteristics and task constraints. Consider two cases. A student may be cramming for an upcoming exam and realize that enough time is not available to obtain an A. Thus, the student sets a lower goal (score at least a B) and chooses criteria that are likely to meet this goal given the limited amount of study time. In this case, the criterion for choosing what to study may include sections (a) that are most likely to be tested and (b) that can potentially be learned in the time remaining. By contrast, another student may have scheduled multiple study sessions before the exam, so as to obtain a very high grade. During an initial study session, this student’s criteria for selection may include all the sections that could be tested, and for subsequent study sessions, the criterion may select only those sections that have not been currently mastered. In these examples, note that the first student would demonstrate a STEM effect (focusing on the easier items), whereas the latter would spend the most time studying the sections that were more difficult to master. These two cases illustrate the flexibility of students’ allocation of study time; depending on their abilities, goals, and task constraints, any number of agendas could be constructed. By emphasizing the importance of ABR, this framework not only indicates how students may optimize their study-time allocation (i.e., by developing agendas in which study criteria match learner abilities and task constraints), but also indicates why students’ allocation may not be optimal. In particular, suboptimal allocation will occur when students do not construct agendas on how to use their study time, when they construct agendas that are suboptimal, or when they construct optimal agendas but fail to execute them properly. To account for how learners allocate their study time, the ABR framework embeds agenda construction and execution within Cowan’s (1988, 1995) information-processing model (Figure 2). This model describes how selective attention is constrained by the nature of the memory system and by the limited processing capacity of the central executive. Like other information-processing models, Cowan’s model describes the flow of information beginning with input to a sensory store. If attended, representations of these stimuli in the long-term store can be activated in the short-term store (STS), and the central executive can select a subset of these activated memories to remain in the focus of attention. The focus of attention is limited in capacity—a person can be aware of only a limited amount of information at any moment. Importantly, attention can be captured by goal-irrelevant stimuli, which can either be external (e.g., loud noise) or internal (e.g., irrelevant thoughts) to a learner. When goal-irrelevant stimuli capture attention, the central executive must inhibit the irrelevant stimuli and reactivate goal-relevant information, or the current learning goal may not be met. As depicted at the top of Figure 2, the exchange of information between the central executive and memory defines two major components of

122


Central executive Monitoring

Control

Long-term store

Agenda construction Focus of attention

Voluntarily controlled action

Habitual responses

Item selection and self-paced study

Activated memory Interest (STS) Self-efficacy Goal orientation Task and domain knowledge

Learning goals Time available Reward Font size of items Item ordering To-be-studied materials

Sensory store

Figure 2 The agenda-based-regulation framework embeds agenda construction within Cowan’s (1988) information-processing system. This figure was adapted from Dunlosky et al. (in press). See text for details.

metamemory—monitoring and control. In particular, the central executive receives information from the memory system (monitoring), and it can use this information to change the state of this system (control), such as by focusing attention on currently unactivated information in the environment, by using the criteria to make decisions about how long (or what) to study, and so forth. In this manner, the central executive and memory system function as the meta level and object level (respectively) in Nelson and Narens’s (1990) model of metacognition (for details, see Nelson & Narens, 1994; Van Overschelde, 2008). This interplay between monitoring and control also lends to the flexibility of allocation. Even if learners begin a study session with a particular goal, it does not mean that they will tenaciously pursue this goal if feedback from monitoring indicates that not enough progress is being made or that the given goal cannot be achieved. In either case, learners may stop executing the current agenda and devise a new one. According to the ABR framework (Figures 2 and 3), study-time allocation will be driven by multiple sources: agenda construction, agenda execution, habitual responding, and online monitoring and control. In the remainder of this section, we briefly discuss each source in turn. When learners are constructing an agenda, the central executive monitors the external and internal environment for task-relevant information. Some potentially relevant information is listed in the bottom-left corner of

123


Central executive Control

Monitoring

Long-term store Agenda execution

Voluntarily controlled action

Agenda criterion Focus of attention

Habitual responses

Item selection and self-paced study

Activated memory Interest (STS) Self-efficacy Goal orientation Task and domain knowledge

Learning goals Time available Reward Font size of items Item ordering To-be-studied materials

Sensory store

Figure 3 Agenda execution portrayed within Cowan’s (1988) information-processing system. Within the focus of attention, learners compare agenda criteria—developed during agenda construction—to the various to-be-studied materials, so as to allocate study time. Habitual (prepotent) responding to other environmental stimuli (not within the focus of attention) can also override agenda-based regulation and drive study-time allocation. See text for details.

Figure 2—such as learning goals, time available, reward, and so forth. Although these are characterized as being external to the learner, some of this information would be generated internally or from interactions between external stimuli and other information stored in long-term memory. For instance, while preparing for a final exam, a student may develop a learning goal from an analysis of previous test scores and her beliefs about how well she can perform on this exam. She may have earned Bs up to the final exam, but she may also have little self-efficacy for achieving a B on the final, perhaps because she waited until the night before to begin studying. This learner may set a low-learning goal (e.g., earn at least a C), and as illustrated in Figure 2, this goal would be activated in the STS as the learner constructed an agenda to achieve it. To achieve this goal efficiently, the learner may develop an agenda to focus on the easiest material that is most likely to be tested. Of course, a low-learning goal could likely be obtained using a variety of agendas, and the learner’s success at achieving the goal would in part depend on the quality of the agenda and how well it is executed. To execute any particular agenda (Figure 3), learners would at least need to maintain the activation of the agenda criteria (e.g., “select the easiest

124


material that will likely be tested”) in the focus of attention while they voluntarily control action toward selecting the appropriate materials for study. In Figure 3, the learner is maintaining attention on the agenda criteria and the to-be-studied materials, and using this information to control item selection. If the agenda includes criteria for how long to study each unit of material (e.g., master materials, the teacher said, would be tested and just read over the other materials), then the learner will also need to maintain activation of these criteria during self-paced study. Given that both agenda construction and execution occur within the focus of attention, capacity limitations can influence effective agenda use. Learners who have deficient central-executive processes may have difficulties constructing and executing agendas. Such difficulties would arise when learners do not inhibit task-irrelevant thoughts or when the number of relevant task constraints and agenda criteria exceeds their capacity. In either case, agenda-relevant information would not be maintained in the focus of attention, which in turn would compromise goal attainment (Kane & Engle, 2003). Even a relatively simple agenda (e.g., “select the easiest materials for study”) may be difficult to execute when goal-irrelevant information arises internally (e.g., mind wandering; Smallwood & Schooler, 2006) or from the environment. When the limits of the central executive have been exceeded, habitual responses may gain control of study-time allocation (Dunlosky & Thiede, 2004). Habitual responses are triggered by the stimulus environment and are not voluntarily controlled. Similarly, Koriat and Nussinson (2009) proposed that “study time is data driven rather than goal driven; it is mainly determined ad hoc by the item itself” (p. 1338). Thus, “a normatively difficult item may receive more study time because the item itself takes longer for a learner to process—not because the learner necessarily developed an agenda to voluntarily study difficult items longer than easier ones” (Dunlosky, Ariel, & Thiede, in press). Although an item may drive allocation, the ABR framework assumes that habitual responses to other environmental stimuli may also influence study-time allocation. Although speculative, learners may spend more time studying materials that are perceptually degraded, spend more time studying when environmental noise undermines selective attention to studying, or may merely study materials in the order in which they appear in notes or in a textbook without regard to differential importance of the materials for attaining a goal. Learners can develop an agenda that includes these factors: If their goal is to master all the material presented in class, they may decide the best way to achieve this goal is to start from the beginning. Even so, the ABR framework predicts that habitual responses will influence study-time allocation and can undermine ABR, and these habitual responses are expected to have a larger influence when the capacity of the central executive is exceeded. In this case, learners are expected to forget their agenda or learning goal and revert to habitual responding.


125

Finally, concerning online monitoring and control, learners can monitor their state of learning in service of controlling item selection and self-paced study. A learner monitors so as to evaluate whether the current goal(s) have been met; if they have been met, the learner will proceed to the next to-belearned material or to the next task. If they have not been met, the learner can make any number of control decisions: continue studying, change the strategy that was being used to study the current item, change the current goal, and so forth. Because monitoring is in part used to control study-time allocation, the accuracy of monitoring is related to the effectiveness of control (Dunlosky, Hertzog, Kennedy, & Thiede, 2005): learners who more accurately monitor their learning more effectively allocate study time and achieve higher levels of test performance (Thiede, 1999) and techniques that boost monitoring accuracy also lead to better allocation and learning (Thiede, Anderson, & Therriault, 2003). The ABR framework assumes that learners can use monitoring in different ways to control item selection and self-paced study. For instance, when pacing study, learners may use monitoring to assess a current state of learning or the rate of learning (for further discussion, see Section 4.3).

4.2. Accounting for Item Selection The ABR framework is composed of attentional and memory components (Figures 2 and 3) and assumes that learners attempt to construct and execute agendas to efficiently maximize goal attainment. The latter assumption provides (a) an obvious explanation for why reward influences item selection (Figure 1), because selecting items with a higher reward (or higher likelihood of being tested) will often maximize one’s expected outcome, and (b) a plausible explanation for the presence of STEM effects in item selection. For example, when people have unlimited study time and a highlearning goal, they tend not to select already recallable items and instead focus the bulk of their study time on the more difficult (unlearned) items. By contrast, with limited study time, they shift to spending more time studying the easier items (e.g., Thiede & Dunlosky, 1999). According to the framework, this shift arises because learners are attempting to optimize their learning given the task constraints. Put differently, if enough time is not available to learn the most difficult items, then focusing on the easiest items is a reasonable agenda to maximize performance. Consider another STEM effect from Metcalfe and Kornell (2003, Experiment 1), who had college students attempt to learn English–Spanish translation equivalents that differed in difficulty: a third were easier pairs (e.g., captain—capitan), a third were moderately difficult (prophecy— vaticinio), and a third were difficult (e.g., insecticide—garrapaticida). During each trial, three English cues (one each from the three difficulty levels) were presented simultaneously in a three-item array; item difficulty

126


(easy, moderate, difficult) was displayed over each English cue word. Participants could select to study the pairs in any order (and in most conditions, a time constraint for overall study time was employed); when a cue was selected, the Spanish translation equivalent was presented for study. The participants’ task was to study the items so that when they were later shown the English cue word on the test, they could correctly type the Spanish translation equivalent. The measure of item selection was the order in which participants selected items for study [denoted by “FC” (first choice) under the column “Allocated study time to” in Table 3]. They tended to select the easiest items first for study, and then moved on to the more difficult items if they had enough time to do so. Why do people prefer to study the easier items first? We have used these same pairs in some of our studies (e.g., Dunlosky & Ariel, in press), and even with unlimited study time, college students correctly recall almost none of the difficult items. It makes no sense to select the difficult items—students with no experience with the Spanish language just do not have a chance to learn them during a single restudy trial. Thus, similar to students with a time pressure (Thiede & Dunlosky, 1999), a reasonable approach to maximizing performance here would be to focus on the easier items, which potentially could be learned with another restudy attempt. Metcalfe (2002) aptly named this agenda as the region of proximal learning (RPL), because learners focus first on the easier unlearned items that by definition are in the students’ RPL. This RPL agenda can optimize behavior in some contexts (Atkinson, 1972), but other agendas are more optimal under other task conditions (Son & Sethi, 2006), so learners will need to use many agendas to efficiently maximize their goal attainment across different situations. In both cases (Metcalfe & Kornell, 2003; Thiede & Dunlosky, 1999), it is plausible that learners are constructing an agenda to select the easier items for study, and many students report choosing the easier items for study when doing so would be most effective (Dunlosky & Thiede, 2004). Nevertheless, even when focusing on easier items would be relatively effective, learners do not always allocate the most time to the easier items, such as (a) when items are presented in a sequential format for selection or (b) when items are presented simultaneously, but the easy items are placed in the least prominent positions of the array. We will consider each of these cases in turn. Concerning presentation format, Thiede and Dunlosky (1999, Experiment 6) used the modal method, and right before item selection, participants were instructed that their goal was to recall only 6 of the 30 items. As expected, when the items were presented simultaneously for selection, the participants selected 8.6 items for restudy and the correlation between JOLs and item selection was over þ0.50. These participants tended to select the easier items for study. By contrast, with the sequential presentation, participants selected almost 20 items for restudy and the corresponding correlation was lower than 0.40!

127


Concerning the placement of items within an array (i.e., under a simultaneous format), Dunlosky and Ariel (in press) used three item arrays—each array included an easier, moderately difficult, and difficult item (as in Metcalfe & Kornell, 2003). The easy items were presented either in the left-most position of the array or in the right-most position of the array. The right-most position was considered least prominent for these college students from Kent State University (OH, USA), because people from Western cultures tend to process information from left to right. These students often studied the easier item first when it was presented on the left. By contrast, they were much less likely to study the easy (and learnable) item first when it was presented on the right and instead studied the difficult (and unlearnable) items first. Put more generally, the position of items within an array in part influenced the order in which items were selected for study. In both cases described above, focusing on the easier items was the best strategy, so, why did learners fail to focus on the easier items? The ABR framework offers many answers to this question, because effective regulation can be undermined by deficits in many of its components, some of which are listed in Table 4. As noted above, learners may dysregulate if they do not develop an agenda, construct poor agendas, or poorly execute their agendas. These deficits in agenda construction and execution may arise from the more basic components of the framework. To illustrate, without any knowledge about optimal agendas for a given task, learners may not construct effective agendas, habitual responding could capture selection and result in poor agenda execution, and capacity limitations and inaccurate monitoring of learning could result in poor agenda execution. Given that so many components (and interactions among them) contribute to study-time allocation, the ABR framework can provide a post hoc explanation for any outcome. Although this flexibility could be viewed as a weakness, post hoc explanations from the framework yield testable predictions to guide further enquiry into why learners succeed (and fail) to attain their learning goals. In our own work, we have identified components from the ABR framework that explain why learners do not prefer to study easy items when doing so would comprise the best strategy. Let us first consider why Table 4 Components from the Agenda-Based Regulation Framework that Undermine Effective Study-time Allocation. Basic cognitive and metacognitive

Agenda-relevant activity

Attentional deficit Capacity limits Knowledge deficit Inaccurate monitoring Habitual responding

No agenda construction Poor agenda construction Poor agenda execution

128


the STEM effect does not occur under the sequential format (Thiede & Dunlosky, 1999). Dunlosky and Thiede (2004) gave participants a lowlearning goal (learn 6 of the 30 items) and a sequential format for item selection. Some participants were instructed to select any items that they wanted to study so as to meet the goal, whereas other learners were instructed to “select six of the easiest-to-learn items for restudy.” If learners were not constructing an agenda to select the easier items under the sequential format, then merely instructing them to use the agenda should produce a STEM effect. The mean number of items selected for restudy was smaller for the instructed (14.4) than for the uninstructed group (21.3), and the correlational measure (JOLs and item selection) was larger for the instructed (þ0.07) than the uninstructed group (0.46). These differences indicate that at least some of the uninstructed learners were not developing an agenda to study the easier items. However, the instructed group did not show a full-blown STEM effect (comparable values from an instructed group with the simultaneous format were 8.6 items selected and a JOLselection correlation of þ0.44). Why? One possibility was that these learners were having difficulties executing the agenda under the sequential format. To obtain a low goal under the sequential format, learners must keep track of how many items have already been selected, how many of the easiest items are left to be selected, and so forth. We speculated that keeping this information activated in the focus of attention while executing the agenda would exceed some learners’ capacity, which in turn would disrupt the maintenance of agenda-relevant information. To evaluate this idea, high and low working-memory span participants were instructed to obtain a low-learning goal and were given the agenda to do so (“select six of the easiest items”). As expected, under the sequential format, high spans demonstrated a STEM effect (correlation ¼ þ 0.39), whereas low spans did not (0.12). This evidence suggests that when items are presented sequentially for selection, learners have difficulties because they do not always construct an effective agenda; and, even when they are instructed to use an effective agenda, some learners’ capacity limitations disrupt agenda execution. The same explanation holds for why a sequential format undermines regulation when easy items are slated to receive a higher reward (see right-hand panel, Figure 1). For example, for the sequential format, when participants were separated into high and low spans, the former showed a large STEM effect when easier items were slated with the higher reward (similar to when items were presented simultaneously, left-hand panel of Figure 1), whereas low-span participants demonstrated no STEM effect whatsoever (Ariel et al., 2009). Now consider why people do not consistently prioritize the easiest item for study when it is presented in the least prominent position of a three-item array (i.e., in the right-most position). Dunlosky and Ariel (in press) attributed this outcome to a habitual reading response. Many learners may not


129

consider developing an agenda to locate the easiest item first for study. Instead, these learners’ behavior may be captured by a habitual reading response, which in Western culture is to proceed from left to right. Recently, we evaluated this idea by conducting the same experiment, but with people from Oman, who read Arabic from right to left (Ariel, Al-Harthy, Was, & Dunlosky, 2010). As expected, the outcomes were the opposite of those collected from Kent State students, who habitually read from left to right! Most importantly, our point here is that the ABR framework offers numerous explanations (Table 4) for the success and failure of study-time allocation. With the creative adaptation of the modal method, these explanations can be used to develop specific hypotheses that can be tested to further reveal the nature of study-time allocation.

4.3. Accounting for Self-Paced Study Times Similar to item selection, study time can be influenced by many components of the ABR framework: the dynamic interplay between online monitoring and decisions to terminate study, agendas that indicate how long to persist studying a given item, and habitual responding triggered by the external environment. The evidence to date has implicated the role of the first two sources of influence, whereas the influence of habitual responding has been unexplored. Accordingly, we first discuss hypotheses relevant to how monitoring serves to control study time and then consider the potential influence of agendas. In contrast to item selection, study times usually do not exhibit a STEM effect (for a few exceptions, see Table 3). Learners tend to spend the most time studying the more difficult items, and then use increasingly less time as item difficulty decreases. Thiede and Dunlosky (1999) argued that once an item was selected for study, a learner implicitly has set a high-learning goal for that particular item (even if they have a low-learning goal across all tobe-learned items); that is, learners select an item for study because they want to learn it. If so, it makes sense that item difficulty and self-paced study time would be inversely related, because it would take longer to learn the more difficult items. This discrepancy-reduction mechanism was introduced earlier, and since it was proposed, its lack of explanatory power has been repeatedly demonstrated (but for contexts in which it may control study decisions, see Benjamin & Bird, 2006; Toppino & Cohen, in press). For instance, even when people are instructed to achieve mastery, they often terminate study before mastering each item and even report terminating study before mastery (Nelson & Leonesio, 1988; Metcalfe & Kornell, 2005, Experiment 8). One difficulty with this discrepancy-reduction mechanism is that it assumes online monitoring is used only to compare the current state of learning to the learner’s static goal. The idea here is that a learner sets a static

130


goal for all items, and then allocation moves toward that goal mindlessly; this mechanism is inflexible. A variant of this monitoring–control relationship was proposed by Dunlosky and Thiede (1998): “Instead of conceptualizing the stopping rule as a function of one’s perception of a specific state of learning, the stopping rule may be a function of perceived change in learning. Termination of study will occur for an item when a person perceives no change in learning for a set amount of time, t. That is, a person will continue studying as long as he or she infers that progress is being made within a set amount of time” (p. 54). The idea that people monitor change is not new (Carver & Scheier, 1990) and has been recently called jROL (judging the rate of learning) by Metcalfe and Kornell (2005).4 This stopping rule is consistent with a main assumption of the ABR framework, because if one persists studying after progress has ceased, then study time is not being allocated efficiently. As important, the threshold, t, allows for more flexibility in study times, because under some task conditions, people may persist even when they perceive no change. The latter prediction is borne out by results from Nelson and Leonesio (1988), who instructed some students to continue studying each item until they were absolutely sure it was mastered (called accuracy-emphasized instructions) and other students to spend only as much time as was needed to learn each item (called the speed-emphasized instructions). Self-paced study times were substantially longer after accuracy- than speed-emphasized instructions; even so, this extra labor was in vain, because criterion test performance did not significantly differ between the groups. Our proposal is that learners develop a simple agenda to spend more time studying higher valued items (i.e., set a higher value of t; cf. Castel, 2007; Castel, Farb, & Craik, 2007; Metcalfe & Jacobs, 2010), and continue studying (even when progress is not being made) until the threshold is met. This conclusion is also supported by outcomes from other studies that have investigated the influence of incentive on study time. In these studies, incentive is typically manipulated by slating some items with a high reward for correct recall on a later criterion test and other items with a lower reward (Ariel et al., 2009; Dunlosky & Thiede, 1998; Koriat et al., 2006; Le Ny et al., 1972; Lockl & Schneider, 2004) or by manipulating the likelihood an item will appear on the criterion test (Ariel et al., 2009, Experiment 1; Dunlosky & Thiede, 1998, Experiment 3). For example, Ariel et al. (2009) assigned either a higher reward for learning easy items compared to difficult 4

At the risk of creating confusion, note that the jROL hypothesis is also based on a discrepancy-reduction mechanism: learners continue to study until the discrepancy between the current state (“making progress”) matches the goal state (“no progress being made”). Thus, although “the discrepancy-reduction mechanism” (often called, DRM) has been used to refer to our original hypothesis (in which learners continue studying till they meet a static goal; Dunlosky & Hertzog, 1998), all the current hypotheses for how monitoring is used to control study time are driven by discrepancy reduction.


131

items (five points for easy items vs. one point for difficult items), a higher reward for learning difficult items compared to easy items (five points for difficult items vs. one point for easy items), or a constant reward for both (one point for either easy or difficult items). Among the items participants selected for study, they persisted longer in studying higher reward items than lower reward items, regardless of item difficulty. Dunlosky and Thiede (1998) and Koriat et al. (2006) also manipulated point value for learning items, and participants in these experiments allocated more study time to the items slated with a higher than lower reward. These findings suggest that incentive may influence learners to adopt a higher threshold for terminating study than they would otherwise adopt for a given item.

4.4. Summary The ABR framework assumes that learners attempt to construct and execute agendas to efficiently maximize goal attainment, and that aspects of the cognitive architecture can influence the success of agenda use and the allocation of study time. The framework can be used to devise multiple explanations for any effect relevant to the allocation of study time, which in turn can drive empirical research aimed at isolating the causes of the focal effect. As discussed above, previous research has already begun to isolate the causes that are responsible for the presence and absence of STEM effects in item selection. Most research on self-paced study has focused on the influence of basic monitoring-and-control mechanisms (e.g., jROL), although recent research suggests that agendas also can influence study time (e.g., Ariel et al., 2009). Accordingly, the ABR framework provides a set of unified constructs (cognitive and metacognitive) that shows promise for guiding exploration of students’ allocation of study time and SRL.

5. Comparing Accounts of Study-Time Allocation The ABR framework was developed to explain how learners allocate their study time and why they succeed (or fail) to attain their learning goals. Three other accounts are available. Two of these accounts were designed to explain study-time allocation (hierarchical model of study time and the regional of proximal learning framework), and the COPES model by Winne and Hadwin (1998) was designed to explain SRL more generally. In Sections 5.1–5.3, we briefly describe these other accounts and compare each to the ABR framework.

132


5.1. Hierarchical Model of Self-Paced Study The hierarchical model of self-paced study (Thiede & Dunlosky, 1999) is a precursor to the ABR framework. Both emphasize that study-time allocation is goal-oriented and that learners develop agendas in an attempt to obtain their learning goals efficiently. The hierarchical model includes a superordinate level in which agenda construction occurs; when it has been completed, control is passed to the subordinate level in which the study time of selected items is driven solely by a discrepancy-reduction mechanism based on a static goal (for details, see Footnotes 3 and 4). Thus, agenda use and discrepancy reduction are inextricably linked, which is a major limitation of the hierarchical model, especially given that self-paced study time is not completely driven by this kind of discrepancy-reduction mechanism (i.e., continue studying until a static goal is met). By contrast, the ABR framework is more flexible, because it does not presuppose that monitoring is used in the same fashion in every learning context. Depending on the task constraints, monitoring of learning may be used to prioritize easier items for self-paced study, or it may be used to evaluate whether progress toward a learning goal is still being made. Given the explanatory limitations of the hierarchical model and the greater flexibility of the ABR framework, the latter offers a more viable platform to guide future research on study-time allocation (for a more detailed comparison of these accounts, see Ariel et al., 2009).

5.2. RPL Framework Allocating study time to the RPL means that learners allocate their study time first to the subset of unlearned items that are easiest to learn, and then they study items that are progressively more difficult to learn (Metcalfe, 2002). More recently, Metcalfe (2009; see also Metcalfe & Kornell, 2005) included two other components to form the RPL framework: (1) when selecting items for study, learners select unlearned items and not the learned ones, which can already be retrieved from memory, and (2) when pacing their study, learners monitor their progress and terminate study when progress is not being made. Next, we separately discuss the three components of the RPL framework. The use of an RPL to guide study provides a straightforward explanation for some STEM effects, because in many contexts, one’s RPL will include the easiest-to-learn items. In terms of the ABR framework, devoting time to one’s RPL is just one of the many agendas that a learner can construct in an attempt to efficiently maximize their learning. In contrast to the RPL, the ABR framework can also explain learners’ failure to use an RPL-like agenda when it should benefit learning, such as when learners allocate time to items presented in a sequence or when their capacity is exceeded.


133

The RPL framework assumes that learners first “attempt to eliminate materials that are already well learned,” (Metcalfe, 2009, p. 159), which typically includes materials that have been previously recalled from memory. This assumption is consistent with the prevailing evidence: Learners overwhelming select unlearned over learned (or previously recalled) items for study, and they use more time studying unlearned items (for a review, see Metcalfe & Kornell, 2005). Learners do select to restudy some previously recalled items (e.g., Karpicke, 2009, Experiment 2), but even here, their decision appears to be based on the belief that these previously recalled items have not yet been learned well. Learners’ bias to study less-welllearned items is also readily explained by the ABR framework, because under many conditions, it is not efficient to study items that have been previously recalled. Most critical for the current comparison, the ABR framework predicts that in some conditions, learners will select already learned items for restudy. This prediction would hold under any condition in which learners’ goal is to retain previously learned information over a long period of time. For instance, sometimes learners may place a premium on retaining some information and hence develop an agenda to overlearn it. We suspect learners commonly construct agendas to overlearn information (“overlearn highly valued information”) and hence will select already known items for restudy, but we admit that this speculation has not yet been empirically evaluated. The RPL framework includes the jROL mechanism, which implicates a particular stopping rule for how monitoring is used to control study time. According to jROL, “people stop studying an item when their perceived rate of learning approaches zero” (Metcalfe, 2009, p. 161). The ABR framework does not presuppose any particular stopping rule for self-paced study—anything goes. Learners can adopt many strategies to terminate study: They may terminate study once they believe no progress is being made (e.g., when time is limited), may continue persisting even when they are unsure of ongoing progress (e.g., when learning all items is highly valued and plenty of time is available), and may even decide to use information other than judgments of learning in deciding when to terminate study (Ariel et al., 2009). For the latter, they may decide to devote the most time to highly valued items that will most likely be included on the upcoming test. Given that learners could possibly terminate study in any of these ways, an intriguing avenue for research will be to answer questions like, why do learners choose to use one stopping rule instead of another? And, what is the effectiveness of a given stopping rule for yielding optimal gains? In general, we view the ABR and RPL frameworks as nested. The former is more general and emphasizes that learners can use many agendas and stopping rules to allocate study time. The latter embraces one agenda and one stopping rule, so it focuses on a subset of the ABR

134


components. Importantly, given that using an RPL agenda to guide study can yield optimal study-time allocation in some conditions (Atkinson, 1972; but see Son & Sethi, 2006), further exploration of this agenda promises to provide insight into the optimality of study-time allocation (for current directions on exploring learner’s use of an RPL agenda, see Metcalfe, 2009).

5.3. COPES Model Although the ABR framework was developed to account for study-time allocation and Winne and Hadwin’s COPES model was developed to account for SRL more generally, the two models are similar in many respects—both emphasize the importance of goal setting, planning, and metacognitive processes in self-regulation. The COPES model is too complex to describe here in any detail (for an excellent review, see Greene & Azevedo, 2007). In brief, SRL includes four stages: task definition, goal setting and planning, enactment, and adaptations. Students decide the task for studying, set goals, and plan how to approach it. These stages include processes relevant to agenda construction (Figure 2) in the ABR framework. Similarly, enactment pertains to agenda execution (Figure 3) and adaptations occur when students use feedback from task experience to subsequently control their learning. For the COPES model, students can C-O-P-E with any of the stages by noting the prevailing conditions and performing operations (strategies) to attain goals; these operations yield products that students evaluate by comparing them to standards. For the ABR framework, these functions are largely controlled by the central executive and often involve basic monitoring-and-control processes that can be relied on to achieve goals. Perhaps the largest difference between the COPES model and the ABR framework is that by embedding ABR in Cowan’s (1988) informationprocessing system, the ABR framework emphasizes the roles of system limitations (e.g., capacity limits of the central executive) and habitual responding, and in particular, how they could undermine the effectiveness of SRL. Preliminary evidence suggests that their influence on study-time allocation is likely to be meaningful. Even here, however, the two frameworks appear to differ in their emphasis, because the COPES model is based on an information-processing platform (Winne, 2001) and hence these factors (system limits and habitual responding) could be incorporated. Thus, the COPES model and the ABR framework are not incompatible, and either could be used to guide research on study-time allocation or any other SRL activity.


135

6. Concluding Remarks and Some Directions for Future Research In Section 1, we introduced a series of questions about self-regulation. To conclude, we answer each one with respect to how students allocate their study time, and in doing so, we consider the limits of our understanding and directions for future research. How do people attempt to regulate themselves to reduce perceived discrepancies? Answer: they construct and execute agendas that will allow them to efficiently achieve their goals. It is intuitively plausible that students use agendas to meet their goals, and agenda use can readily explain much of the evidence on study-time allocation. Nevertheless, the bulk of the evidence is indirect, so developing methods to directly measure ABR presents a major challenge for future research. Such methods will also be vital for exploring answers to other questions relevant to agenda use, such as “how do students set agendas for a given task?” And, after an agenda is set, “what factors lead to changes to the original agenda?” Is people’s regulation optimal? Not always. In the review above, we described several examples where learners did not effectively allocate their study time. In some cases, however, learners do make relatively effective decisions (e.g., Kornell & Metcalfe, 2006; Nelson, Dunlosky, Graf, & Narens, 1994). These and other studies have shown that when learners’ decisions on how to allocate study time are not honored, their test performance suffers. It is not clear, however, whether learners’ allocation in these cases is optimal—that is, whether they obtained the learning goal most efficiently. The evidence does indicate that people could have made worse decisions, but it does not rule out the possibility they could have made better ones. To rule out this possibility, a general theory of study-time optimization is needed that specifies what constitutes optimal regulation (achieving a goal while minimizing time invested) for a given student, learning goal, and to-be-learned materials. Comparing expectations from these theories to learners’ behavior will provide crucial evidence as to whether learners allocate their time optimally. Atkinson (1972) and Son and Sethi (2006) developed formal models of optimality for learning simple materials (although Son and Sethi’s model may generalize to other kinds of material as well). Most importantly, their approaches to modeling optimality (as a function of item difficulty and learning rates) are innovative and offer excellent starting points for developing a more general theory. Finally, if people’s regulation is suboptimal, what factors limit their optimality, and are some people better at regulating their behavior than others? Suboptimal study-time allocation can result from failures in agenda

136


use, cognition, or metacognition (Table 4). Failures in each one do contribute to suboptimal allocation: learners sometimes fail to construct agendas, and even when they do construct effective ones, limited workingmemory capacity can undermine effective agenda execution (Dunlosky & Thiede, 2004) and so can inaccurate monitoring of learning (Thiede, 1999). These are not the only factors that can cause suboptimal allocation of study time; other possible candidates are presented in Table 4, and still other factors (e.g., goal orientation, self-efficacy, need for cognition, etc.) have been examined (Zimmerman & Schunk, 2001). Although each of these factors can influence the success of students’ SRL, how they are coordinated as students attempt to achieve their long-term learning goals is largely unknown. The ABR framework offers insight into some ways that these factors can be coordinated, and in doing so, it can guide investigations aimed at discovering the foibles of self-regulated learning and how to fix them.

ACKNOWLEDGMENTS This study was supported by a James S. McDonnell Foundation 21st Century Science Initiative in Bridging Brain, Mind, and Behavior Collaborative Award.

REFERENCES Ariel, R., Al-Harthy, I., Was, C., & Dunlosky, J. (2010). Habitual reading biases in the allocation of study time. Unpublished data. Ariel, R., Dunlosky, J., & Bailey, H. (2009). Agenda-based regulation of study-time allocation: When agendas override item-based monitoring. Journal of Experimental Psychology. General, 138, 432–447. Atkinson, R. C. (1972). Optimizing the learning of a second-language vocabulary. Journal of Experimental Psychology, 96, 124–129. Azevedo, R., Moos, D. C., Greene, J. A., Winters, F. I., & Cromley, J. G. (2008). Why is externally-facilitated regulated learning more effective than self-regulated learning with hypermedia? Educational Technology Research and Development, 56, 45–72. Benjamin, A. S. (2007). Memory is more than just remembering: Strategic control of encoding, accessing memory, and making decisions. In A. S. Benjamin & B. H. Ross (Eds.), The psychology of learning and motivation: Skill and strategy in memory use, Vol. 48 (pp. 175–223). London, UK: Academic Press. Benjamin, A. S., & Bird, R. D. (2006). Metacognitive control of the spacing of study repetitions. Journal of Memory and Language, 55, 126–137. Blood, E., & Neel, R. (2008). Using student response systems in lecture-based instruction: Does it change student engagement and learning? Journal of Technology and Teacher Education, 16, 375–383. Britton, B. K., & Tesser, A. (1991). Effects of time-management practices on college grades. Journal of Educational Psychology, 83, 405–410. Calvin, W. C. (1990). The cerebral symphony. New York, NY: A Bantam Book.


137

Carver, C. S., & Scheier, M. F. (1990). Origins and functions of positive and negative affect: A control-process view. Psychological Review, 97, 19–35. Carver, C. S., & Scheier, M. F. (2000). Control processes and self-organization as complimentary principles underlying behavior. Personality and Social Psychology Review, 6, 304–315. Castel, A. D. (2007). The adaptive and strategic use of memory by older adults: Evaluative processing and value-directed remembering. In A. S. Benjamin & B. H. Ross (Eds.), The psychology of learning and motivation: Skill and strategy in memory use, Vol. 48 (pp. 225–270). London: Academic Press. Castel, A. D., Farb, N. A. S., & Craik, F. I. M. (2007). Memory for general and specific value information in younger and older adults: Measuring the limits of strategic control. Memory and Cognition, 35, 689–700. Clump, M. A., Bauer, H., & Whiteleather, A. (2003). To attend or not to attend: Is that a good question? Journal of Instructional Psychology, 30, 220–224. Cowan, N. (1988). Evolving conceptions of memory storage, selective attention, and their mutual constraints within the human information-processing system. Psychological Bulletin, 104, 163–191. Cowan, N. (1995). Attention and memory: An integrated framework. Oxford University Press Oxford Psychology Series, No. 26. Crede´, M., & Kuncel, N. R. (2009). Study habits, skills, and attitudes: The third pillar supporting collegiate academic learning. Perspectives on Psychological Science, 3, 425–453. Dettmers, S., Trautwein, U., Ludtke, O., Kunter, M., & Baumert, J. (2010). Homework works if homework quality is high: Using multilevel modeling to predict the development of achievement in mathematics. Journal of Educational Psychology, 102, 467–482. Dickinson, D. J., & O’Connell, D. Q. (1990). Effect of quality and quantity of study on student grades. The Journal of Educational Research, 83, 227–231. Dunlosky, J., & Ariel, R. (in press). The influence of agenda-based and habitual processes on item selection during study. Journal of Experimental Psychology: Learning, Memory & Cognition. Dunlosky, J., Ariel, R., & Thiede, K. W. (in press). Agenda-based regulation of study-time allocation. In P. A. Higham & J. P. Leboe (Eds.), Constructions of remembering and metacognition: Essays in honour of Bruce Whittlesea (pp. xx–xx). Basingstoke: Palgrave MacMillan. Dunlosky, J., & Hertzog, C. (1998). Training programs to improve learning in later adulthood: Helping older adults educate themselves. In D. Hacker, J. Dunlosky, & A. Graesser (Eds.), Metacognition in educational theory and practice (pp. 249–275). Hillsdale, NJ: LEA. Dunlosky, J., Hertzog, C., Kennedy, M., & Thiede, K. (2005). The self-monitoring approach for effective learning. Cognitive Technology, 10, 4–11. Dunlosky, J., & Thiede, K. W. (1998). What makes people study more? An evaluation of factors that affect people’s self-paced study and yield “labor-and-gain” effects. Acta Psychologica, 98, 37–56. Dunlosky, J., & Thiede, K. W. (2004). Causes and constraints of the shift-to-easier-materials effect in the control study. Memory & Cognition, 32, 779–788. Finn, B. (2008). Framing effects on metacognitive monitoring and control. Memory & Cognition, 36, 813–821. Friedman, S. L., Scholnick, E. K., & Cocking, R. R. (Eds.), (1987). Blueprints for thinking: The role of planning in cognitive development. New York: Cambridge University Press. Graesser, A. C., McMahen, C. L., & Johnson, B. K. (1994). Question asking and answering. In M. A. Gernsbacher (Ed.), Handbook of psycholinguistics (pp. 517–538). San Diego: Academic Press.

138


Greene, J. A., & Azevedo, R. (2007). A theoretical review of Winne and Hadwin’s model of self-regulated learning: New perspectives and directions. Review of Educational Research, 77, 334–372. Hines, J. C., Touron, D. R., & Hertzog, C. (2009). Metacognitive influences on study time allocation in an associative recognition task: An analysis of adult age differences. Psychology and Aging, 24, 462–475. Jeong, H., & Chi, M. T. H. (2007). Knowledge convergence and collaborative learning. Instructional Science, 35, 287–315. Kane, M. J., & Engle, R. W. (2003). Working-memory capacity and the control of attention: The contributions of goal neglect, response competition, and task set to stroop interference. Journal of Experimental Psychology. General, 132, 47–70. Karpicke, J. D. (2009). Metacognitive control and strategy selection: Deciding to practice retrieval during learning. Journal of Experimental Psychology. General, 138, 469–486. Koriat, A. (2008). Easy comes, easy goes? The link between learning and remembering and its exploitation in metacognition. Memory & Cognition, 36, 416–428. Koriat, A., & Ackerman, R. (2010). Metacognition and mindreading: Judgments of learning for self and other during self-paced study. Consciousness and Cognition, 19, 251–264. Koriat, A., Ackerman, R., Lockl, K., & Schneider, W. (2008). The memorizing effort heuristic in judgments of learning: A developmental perspective. Journal of Experimental Child Psychology, 102, 265–279. Koriat, A., & Ma’ayan, H. (2005). The effects of encoding fluency and retrieval fluency on judgments of learning. Journal of Memory and Language, 52, 478–492. Koriat, A., Ma’ayan, H., & Nussinson, R. (2006). The intricate relationships between monitoring and control in metacognition: Lessons for the cause-and-effect relation between subjective experience and behavior. Journal of Experimental Psychology. General, 135, 36–69. Koriat, A., & Nussinson, R. (2009). Attributing study effort to data-driven and goal driven effects: Implications for metacognitive judgments. Journal of Experimental Psychology: Learning, Memory & Cognition, 35, 1338–1343. Kornell, N., & Metcalfe, J. (2006). Study efficacy and the region of proximal learning framework. Journal of Experimental Psychology: Learning, Memory & Cognition, 32, 609–622. Kratzig, G. P., & Arbuthnott, K. D. (2009). Metacognitive learning: The effect of itemspecific experience and age on metamemory calibration and planning. Metacognition Learning, 4, 125–144. Le Ny, J.-F., Denhie`re, G., & Le Taillanter, D. (1972). Study-time of sentences as a function of their specificity and of semantic exploration. Acta Psychologica, 37, 43–53. Lockl, K., & Schneider, W. (2004). The effects of incentives and instructions on children’s allocation of study time. European Journal of Developmental Psychology, 1, 153–169. McNamara, D. S. (2010). Strategies to read and learning: Overcoming learning by consumption. Medical Education, 44, 340–346. Metcalfe, J. (2002). Is study time allocated selectively to a region of proximal learning? Journal of Experimental Psychology. General, 131, 349–363. Metcalfe, J. (2009). Metacognitive judgments and control of study. Current Directions in Psychology Science, 18, 159–163. Metcalfe, J., & Finn, B. (2008). Evidence that judgments of learning are causally related to study choice. Psychonomic Bulletin & Review, 15, 174–179. Metcalfe, J., & Jacobs, W. J. (2010). People’s study time allocation and its relation to animal foraging. Behavioural Processes, 83, 213–221. Metcalfe, J., & Kornell, N. (2003). The dynamics of learning and allocation of study time to a region of proximal learning. Journal of Experimental Psychology. General, 132, 530–542.


139

Metcalfe, J., & Kornell, N. (2005). A region of proximal learning model of study time allocation. Journal of Memory and Language, 52, 463–477. Miller, G. A., Galanter, E., & Pribram, K. H. (1960). Plans and the structure of behavior. New York: Henry Holt and Company. Mithaug, D. E. (1993). Self-regulation theory: How optimal adjustment maximizes gain. Westport, CT: Praeger. Morisano, D., Hirsh, J. B., Peterson, J. B., Pihl, R. O., & Shore, B. M. (2010). Setting, elaborating, and reflecting on personal goals improves academic learning. Journal of Applied Psychology, 95, 255–264. Nelson, T. O., Dunlosky, J., Graf, E. A., & Narens, L. (1994). Utilization of metacognitive judgments in the allocation of study during multitrial learning. Psychological Science, 5, 207–213. Nelson, T. O., & Leonesio, R. J. (1988). Allocation of self-paced study time and the “laborin-vain effect. Journal of Experimental Psychology: Learning, Memory, and Cognition, 14, 676–686. Nelson, T. O., & Narens, L. (1990). The psychology of learning and motivation (Vol. 26, pp. 125–141). New York: Academic Press. Nelson, T. O., & Narens, L. (1994). Why investigate metacognition? In J. Metcalfe & A. J. Shimamura (Eds.), Metacognition: Knowing about knowing (pp. 1–26). Cambridge, MA: MIT Press. Newell, A., & Simon, H. A. (1972). Human problem solving. Englewood Cliffs, NJ: PrenticeHall. Pintrich, P. R. (2000). The role of goal orientation in self-regulated learning. In M. Boekaerts, P. R. Pintrich, & M. Zeidner (Eds.), Handbook of selfregulation (pp. 451–502). New York: Academic Press. Powers, W. T. (1973). Behavior: The control of human perception. Chicago, IL: Aldine Publishing Company. Pressley, M., Van Etten, S., Yokoi, L., Freebern, G., & Van Meter, P. (1998). The metacognition of college studentship: A ground theory approach. In D. Hacker, J. Dunlosky, & A. Graesser (Eds.), Metacognition in educational theory and practice (pp. 347–366). Hillsdale, NJ: LEA. Price, J., Hertzog, C., & Dunlosky, J. (2009). Self-regulated learning in younger and older adults: Does aging affect metacognitive control? Aging, neuropsychology, and cognition, 17, 329–359. Pyc, M. A., & Dunlosky, J. (2010). Toward an understanding of students’ allocation of study time: Why do they decide to mass or space their practice? Memory & Cognition, 38, 431–440. Rhodes, M. G., & Castel, A. D. (2009). Metacognitive illusions for auditory information: Effects on monitoring and control. Psychonomic Bulletin & Review, 16, 550–554. Richardson, J. T. E. (1998). The availability and effectiveness of reported mediators in associative learning: A historical review and an experimental investigation. Psychonomic Bulletin & Review, 5, 597–614. Smallwood, J., & Schooler, J. W. (2006). The restless mind. Psychological Bulletin, 132, 946–958. Son, L. K. (2004). Spacing one’s study: Evidence for a metacognitive control strategy. Journal of Experimental Psychology: Learning, Memory, and Cognition, 30, 601–604. Son, L. K., & Metcalfe, J. (2000). Metacognitive and control strategies in study-time allocation. Journal of Experimental Psychology: Learning, Memory and Cognition, 26, 204–221. Son, L. K., & Sethi, R. (2006). Metacognitive control and optimal learning. Cognitive Science, 30, 759–774.

140


Taraban, R., Maki, W. S., & Rynearson, K. (1999). Measuring study time distributions: Implications for designing computer-based courses. Behavior Research Methods, Instruments & Computers, 31, 263–269. Thiede, K. W. (1999). The importance of monitoring and self-regulation during multi-trial learning. Psychonomic Bulletin & Review, 6, 662–667. Thiede, K. W., Anderson, M. C. M., & Therriault, D. (2003). Accuracy of metacognitive monitoring affects learning of texts. Journal of Educational Psychology, 95(1), 66–73. Thiede, K. W., & Dunlosky, J. (1999). Toward a general model of self-regulated study: An analysis of selection of items for study and self-paced study time. Journal of Experimental Psychology: Learning, Memory, and Cognition, 25, 1024–1037. Thomas, A. K., & McDaniel, M. A. (2007). The negative cascade of incongruent generative study-test processing in memory and metacomprehension. Memory & Cognition, 35, 668–678. Tiede, H. L., & Leboe, J. P. (2009). Illusions of competence for phonetically, orthographically, and semantically similar word pairs. Canadian Journal of Experimental Psychology, 63, 294–302. Toppino, T. C., & Cohen, M. S. (2010). Metacognitive control and spaced practice: Clarifying what people do and why. Journal of Experimental Psychology: Learning, Memory and Cognition, 36, 1480–1492. Van Overschelde, J. P. (2008). Metacognition: Knowing about knowing. In D. Hacker, J. Dunlosky, & A. Graesser (Eds.), Handbook of metacognition in education (pp. 47–71). New York: Psychology Press. Wasylkiw, L., Tomes, J. L., & Smith, F. (2008). Subset testing: Prevalence and implications for study behaviors. The Journal of Experimental Education, 76, 243–257. Winne, P. H. (2001). Self-regulated learning viewed from models of information processing. In B. J. Zimmerman & D. H. Schunk (Eds.), Self-regulated learning and academic achievement: Theoretical perspectives (pp. 153–189). (2nd ed.). Mahwah, NJ: Erlbaum. Winne, P. H., & Hadwin, A. F. (1998). Studying as self-regulated learning. In D. J. Hacker, J. Dunlosky, & A. C. Graesser (Eds.), Metacognition in educational theory and practice (pp. 277–304). Hillsdale, NJ: LEA. Zacks, R. T. (1969). Invariance of total learning time under different conditions of practice. Journal of Experimental Psychology, 82, 441–447. Zimmerman, B. J. (2001). Theories of self-regulated learning and academic achievement: An overview and analysis. In B. J. Zimmerman & D. H. Schunk (Eds.), Self-regulated learning and academic achievement: Theoretical perspectives (pp. 1–37). (2nd ed.). Mahwah, NJ: Erlbaum. Zimmerman, B. J., & Schunk, D. H. (2001). Reflections on theories of self-regulated learning and academic achievement. In B. J. Zimmerman & D. H. Schunk (Eds.), Selfregulated learning and academic achievement: Theoretical perspectives (pp. 289–308). (2nd ed.). Mahwah, NJ: Erlbaum.

C H A P T E R

F I V E

The Development of Categorization Vladimir M. Sloutsky and Anna V. Fisher Contents 142 143 147

1. Introduction 1.1. Theoretical Approaches to Category Learning 2. Categorization and Selective Attention 2.1. Spontaneous Learning of Categories Based on Multiple Overlapping Features 2.2. Salience and Categorization 3. The Role of Labels in the Development of Categorization 3.1. Labels as Features Contributing to Similarity 3.2. Labels Overshadow Visual Input 3.3. Effect of Semantic Similarity and Co-Occurrence Probability of Labels 3.4. Flexible Adjustment of Attention Allocated to Labels and Appearances 4. Early Categorization: What Develops? Acknowledgments References

147 151 153 153 156 157 159 162 163 163

Abstract The ability to categorize is a ubiquitous property of human cognition. However, despite much progress in understanding how children learn to categorize, several important issues remain highly debated. The central issues in this debate concern the role of perceptual input and the role of linguistic labels in categorization and category learning. This chapter reviews recent evidence bearing on these key issues. Based on this evidence, we suggest that (1) categorization begins with extracting structure from the input based on overlapping perceptual features and attentional weights and (2) development of the ability to selectively attend to some features while ignoring others underlies the learning of abstract concepts and categories.

Psychology of Learning and Motivation, Volume 54 ISSN 0079-7421, DOI: 10.1016/B978-0-12-385527-5.00005-X

#


141

142

Vladimir M. Sloutsky and Anna V. Fisher

1. Introduction Categorization is not a matter to be taken lightly. There is nothing more basic than categorization to our thought, perception, action, and speech. Every time we see something as a kind of thing, for example, a tree, we are categorizing (Lakoff, 1987, p. 139).

The ability to categorize is a ubiquitous property of cognition: it enables treating appreciably different entities (e.g., different dogs, different tokens of the same word, different situations, or different problems) as variants of the same thing. This critically important property of intelligence promotes cognitive economy (as different entities are represented the same way) and application of learned knowledge to novel situations (e.g., learning that Dobermans bark would lead one to believe that German Shepherds bark as well). In many ways, categorization in humans is remarkably similar to that in other species. For instance, human and nonhuman primates, monkeys, rats, birds, fish, and crickets can learn a variety of visual and auditory categories based on the distribution of features in the input (Chase, 2001; Pearce, 1994; Ribar, Oakes, & Spalding, 2004; Wyttenbach, May, & Hoy, 1996; Zentall, Wasserman, Lazareva, Thompson, & Rattermann, 2008). In the case of human infants, evidence for such perceptual groupings can be exhibited as early as 3 months of age (Bomba & Siqueland, 1983; Quinn, Eimas, & Rosenkrantz, 1993). In some other ways, categorization in humans is remarkably different from that in other species. For example, there is little evidence to suggest that nonhuman species have the ability to (1) organize categories into multiple levels of abstraction (such as Dalmatian ! Dog ! Mammal ! Animal ! Living Thing), (2) create ad hoc categories (i.e., categories that are not established in memory but derived to achieve a novel goal, such as “activities to do on a vacation in Japan”; Barsalou, 1991), or engage in abstract category-based inductive reasoning (i.e., inferring hidden properties on the basis of common category rather than perceptual similarity). Therefore, it is not surprising that questions of how people acquire these remarkable abilities in the course of ontogenesis and how category learning interacts with language learning have fascinated generations of researchers. Attempts to answer these questions generated multiple theoretical proposals and a large body of evidence, with surprisingly little agreement with respect to these fundamental questions. The goal of this chapter is to summarize the debate on the development of categorization and category learning and to review some recent evidence bearing on this debate.

The Development of Categorization

143

In what follows, we first summarize two approaches to human category learning that dominate the debate in the current literature. A comprehensive historical overview of the theoretical developments and associated findings—while highly relevant to the issues discussed below—are outside the scope of this chapter. Excellent overviews of these issues are available elsewhere (Murphy, 2002). We then discuss the mechanism of early categorization and possible changes in this mechanism in the course of development.

1.1. Theoretical Approaches to Category Learning There are a number of theoretical approaches aiming at explaining category learning and the development of mechanisms of categorization. Most researchers agree that even early in development, people can learn categories that are based on multiple overlapping perceptual features (Quinn et al., 1993; Ribar et al., 2004). Most researchers also agree that adults demonstrate the ability to go beyond perceptual input and to form categories based on abstract unobservable properties. There is little agreement, however, on when and how this ability emerges. Some researchers argue that the ability to form categories on the basis of unobservable features emerges from the ability to form perceptual categories (Rakison & PoulinDubois, 2001; Rogers & McClelland, 2008; Sloutsky & Fisher, 2004). Others argue that this is not the case and the two abilities are relatively independent (Booth & Waxman, 2002; Gelman, 2003; Keil, Smith, Simons, & Levin, 1998). Based on certain common theoretical assumptions, it is possible to classify these theories into two broad kinds: the knowledge-based approach and the coherence-based approach. These approaches offer divergent perspectives on the starting point, the mechanism and the development of categorization. 1.1.1. The Knowledge-Based Approach: Poverty of the Stimulus and Representational Constraints The basic assumption behind this approach is that input is underconstrained and therefore structured knowledge cannot be recovered from input alone. This assumption is based on the Poverty of the Stimulus argument, which was originally proposed for the case of language acquisition (Chomsky, 1980) and was later extended to many other aspects of cognitive development, including category learning. According to this argument, if knowledge cannot be recovered from data alone, then some innate constraints are necessary to explain its acquisition. It has been suggested that such constraints may come in the form of skeletal principles (R. Gelman, 1990), core knowledge (e.g., Carey, 2009; Spelke & Kinzler, 2007), or a variety of conceptual assumptions and naı¨ve theories (Gelman, 2003). While there is

144


no explicit agreed-upon definition of conceptual knowledge, the term “conceptual” is commonly used in reference to top-down knowledge that is used to structure input information. “. . .[C]hildren’s categories are theory based: They are constructed not merely on the basis of perceptual characteristics and regularities but on the basis of children’s beliefs and assumptions about the world and the way language works” ( Jaswal, 2004, p. 1872; see also Booth, Waxman, & Huang, 2005; Gelman, 2003; Markman, 1991). This meaning of conceptual biases and knowledge as distinctly different from learned associations is assumed in this chapter. The constraints described above are what we will refer to as representational constraints (to be contrasted below with other kinds of constraints)— domain-specific knowledge that guides learner’s attention toward a small set of features relevant for solving a specific problem in a given domain. The domain specificity is of critical importance because principles constraining learning in the domain of number (e.g., knowledge that integers never end; cf. R. Gelman, 1990) are not useful in the domain of space. This argument for representational constraints is based on the assumptions that we can successfully determine which constraints are required to solve a problem and which are available in the input. However, this assumption could be wrong. First, determining the required constraints is a highly nontrivial problem. For example, it has been traditionally accepted that the task of language learning requires the induction of a syntax that has an arbitrary large number of constraints. However, Chater and Christiansen (2010) have recently suggested this construal could be wrong. Specifically, they offered a critical distinction between “natural” and “cultural” induction (i.e., N-induction and C-Induction). N-induction involves the ability to understand the natural world, whereas C-induction involves the ability to coordinate with other people. While the traditional construal of the problem of language acquisition has conceptualized it as an extremely difficult N-induction problem (i.e., the discovery of abstract syntax), Chater and Christiansen suggested that the problem should be construed as a much easier problem of C-induction. Instead of solving a problem requiring an arbitrary set of constraints (i.e., the problem of N-induction), individuals simply have to make the same guesses as everyone else. In sum, according to the former construal, the number of required constrains is very high, whereas according to the latter construal, this number is substantially lower. Therefore, different conceptualizations of the same learning problem may result in vastly different estimates of the required constraints. Second, the number of constraints available in the input is also a nontrivial problem that still awaits a theoretical and empirical solution. Finally, there is much evidence that people possess powerful learning mechanisms allowing them to detect statistical regularities in the input (French, Mareschal, Mermillod, & Quinn, 2004; Go´mez & Maye, 2005; Pullum & Scholz, 2002; Saffran, Aslin, & Newport, 1996; Sloutsky & Fisher,


145

2008; Smith & Yu, 2008). In short, if the number of required constrained is not known, if the poverty of the input is questionable, and if some structure can be recovered from data, the assumption of innate representational constraints does not seem justified. 1.1.2. The Coherence-Based Approach: Similarity, Statistics, and Attentional Weights According to a different theoretical perspective, category learning early in development is driven by the statistical regularities in the input. The central idea is that items belonging to the same category have multiple overlapping features, which jointly guide category learning. For instance, members of the category “bird” have wings, feathers, and beaks; are referred to as “birds”; lay eggs; and possess a set of typical behaviors (such as flying and nesting). These features cohere and covary, providing a learner with richly structured input. Powerful learning mechanisms, such as perceptual and attentional learning, allow children to take advantage of this coherent covariation in the input (French et al., 2004; Hall, 1991; Rogers & McClelland, 2008; Sloutsky, 2003, 2010; Sloutsky & Fisher, 2004, 2008; Smith, 1989; Smith, Jones, & Landau, 1996). A common but hardly accurate assumption about this approach is that it is constraint-free and any correlation present in the input gets encoded and learned. If true, this would have made the problem of learning computationally intractable. There are three reasons why this does not happen. First, information that is attended to and learned is limited, in part, by the basic information processing properties of the organism—what we will refer to as processing constraints. It is well known that processing resources available to infants and young children are more modest compared to those available to older children and adults in terms of processing speed, processing capacity, attention capacity, memory, etc. (Baddeley, 1986; Demetriou, Christou, Spanoudis, & Platsidou, 2002). A case has been made that these limitations play a positive role early in development. As Elman (1993, pp. 17–18) put it, “Limited capacity acts like a protective veil, shielding the infant from the stimuli which may either be irrelevant or require prior learning to be interpreted” (for similar arguments, see Newport, 1990). Second, features available in the input differ in salience with salient features being more likely to get through the “protective veil” of limited capacity. In other words, more salient cues have a higher probability of being detected and learned than less salient cues. This learning in turn results in increased attention to these cues in the future, compared to other potentially salient cues (Hall, 1991; Kruschke & Blair, 2000; Nosofsky, 1986; Sloutsky & Spino, 2004). And third, many features in real input are “bundled together”— members of most natural kind categories have much in common and many of those common features distinguish these members from members

146


of other categories. These bundled features may mutually reinforce each other thus resulting in nonadditive effects on learning (see Rogers & McClelland, 2004; Smith, Colunga, & Yoshida, 2010; Sloutsky, 2010 for reviews). Many studies have documented that infants and young children are capable of sophisticated generalizations. For example, studies using a label extension task (e.g., “this is a dax, show me another dax”) have demonstrated that 2- and 3-year-olds generalize the novel name to the sameshaped item when the exemplar is a solid object, but to the same material item when the exemplar is a nonsolid item (Soja, Carey, & Spelke, 1991). One possible explanation of these findings is that at the start of word learning, children have knowledge of ontological distinction between objects and substances and that this knowledge manifests itself in this task (e.g., Soja et al., 1991). Another explanation is that the distinction between solids and nonsolids is supported by multiple syntactic and perceptual cues systematically covarying with the distinction. This massive covariation cues children’s attention to the distinction and results in smart generalization behaviors (e.g., Smith et al., 2010). Therefore, associative and attentional learning may give rise to smart behaviors. Another example of a smart behavior stemming from mundane mechanisms comes from work of Sloutsky and Fisher (2008) and we describe this work in a greater detail in a section below. To summarize, according to the knowledge-based approach, early generalization is driven by a few theoretically relevant features. Importantly, it is impossible to learn from input what the relevant features are—this knowledge comes a priori in the form of core knowledge, conceptual assumptions, and naı¨ve theories. According to the coherence-based approach, category learning is subserved by powerful learning mechanisms (i.e., perceptual and attentional learning) enabling extraction of statistical regularities in the input. If the end-point of development is a full-fledged system of concepts, it might be beneficial for an organism to start with more abstract ontological kinds and focus on few theoretically relevant features distinguishing among these kinds (Carey, 2009; Gelman, 1990; Spelke & Kinzler, 2007). However, there are several arguments as to why this is a highly implausible starting point. First, reliance on few relevant dimensions requires selective attention, whereas, as we argue in the next section, selectivity is a later developing ability. Therefore, selective attention may be insufficiently developed in infants and young children to support conceptual development. The necessity for selectivity, however, could be bypassed if (a) categories are based on multiple correlated features (something that is advocated by the coherence-based approach), (b) if category-relevant features are highly salient and capture attention automatically, or (c) if category labels serve as supervisory signals attracting attention to important


147

within-category commonalities and thus guiding category learning. However, as we argue in the Sections 2.2 and 3, respectively, there is very little empirical support for either (b) or (c).

2. Categorization and Selective Attention Learning of categories based on a few relevant dimensions requires selectivity, whereas learning of categories based on multiple overlapping features could be accomplished without selective attention. At the same time, young children may have difficulty focusing on few relevant features and ignoring multiple irrelevant ones, especially when irrelevant features vary independently of the relevant features. As a result, spontaneous learning of the categories based on a few relevant dimensions could be more challenging than spontaneous learning of the categories based on multiple overlapping features. In what follows, we review two lines of research, one demonstrating that under conditions of spontaneous learning, children are more likely to learn categories based on multiple overlapping features than on few ones. The second line presents evidence explaining the first one and indicating that young children have difficulty focusing on relevant information and ignoring irrelevant.

2.1. Spontaneous Learning of Categories Based on Multiple Overlapping Features In a set of studies, Kloos and Sloutsky (2008) examined the role of category structure in category learning. The measure of structure used in these studies was category density (i.e., the ratio of within-category variability to between-category variability). In one study, 5-year-olds and adults were presented with a category learning task where they learned either dense or sparse categories. These categories consisted of artificial bug-like creatures that had a number of dimensions of variation: the sizes of tail, wings, and fingers; the shadings of body, antenna, and buttons; and the numbers of fingers and buttons. For dense categories, multiple dimensions covaried and they were jointly predictive of category membership, whereas for sparse categories there were few category-relevant dimensions, with the rest of the dimensions varying randomly within- and between-category. Category learning was administered under either an unsupervised, spontaneous learning condition (i.e., participants were merely shown the items) or under a supervised, deliberate learning condition (i.e., participants were told the category inclusion rule). Critical data from this study are presented in Figure 1. The figure presents categorization accuracy (i.e., the proportion of hits, or correct identification of category members minus the proportion of

148


Accuracy (Hits-FA)

A

1 0.75 Learning condition 0.5

Unsupervised Supervised

0.25 0 Dense

Sparse

Category structure

Accuracy (Hits-FA)

B

1 Learning condition

0.75

Unsupervised Supervised

0.5 0.25 0 Dense

Sparse

Category structure

Figure 1 Mean accuracy scores by category type and learning condition in adults (A) and children (B). Error bars represent standard errors of the mean. After Kloos and Sloutsky (2008).

false alarms, or confusion of nonmembers for members) after the category learning phase. Data presented in Figure 1 clearly indicate that for both children and adults, sparse categories were learned better under the explicit, supervised condition, whereas dense categories were learned better under the implicit, unsupervised condition. Also note that adults learned the sparse category even in the unsupervised condition, whereas young children exhibited no evidence of learning. In addition, data from Kloos and Sloutsky (2008) indicate that while both children and adults exhibited able unsupervised, spontaneous learning of dense categories, there were marked developmental differences in unsupervised learning of sparse categories. Categorization accuracy in the unsupervised condition by category density and age are presented in Figure 2. Two aspects of these data are worth noting. First, there was no developmental difference in spontaneous learning of the very dense category,

149


Categorization accuracy

1.2 Adults Children

1 0.8 0.6 0.4 0.2 0 D=1

D = 0.39

D = 0.17

Category density

Figure 2 Unsupervised category learning by density and age group in Kloos and Sloutsky (2008).

whereas there were substantial developmental differences in spontaneous learning of sparser categories, with children exhibiting less evidence for learning than adults. Why are denser categories easier for young children? And why does learning of sparse but not dense categories undergo developmental progression? One possibility that has been discussed recently (e.g., Sloutsky, 2010) is that due to immaturities in the prefrontal cortex, young children have difficulty selectively focusing on relevant features, while ignoring irrelevant ones. At the same time, this ability seems to be critical for learning sparse categories. As prefrontal cortex matures, the ability to focus on relevant and ignore irrelevant improves, and so does the ability to learn sparse categories. This issue has been addressed in a series of recent yet unpublished studies by Yao and Sloutsky. These researchers presented participants with a series of tasks: simple matching, simple generalization, and complex generalization. Each of the tasks was a variant of the match-to-sample task consisting of a target and two test items, with one of the test items being a match for the target. There were two critical modifications to the standard task. First, each item had a less salient component that had to be used in matching and a more salient distracter component that had to be ignored (see Figure 3), with participants instructed to match the items on the basis of the less salient component. There were three within-subject conditions: supportive (Figure 3A), conflict (Figure 3B), and neutral (Figure 3C). In the supportive condition, the test item that matched the target on the less salient component, also matched the target on the more salient component. In contrast, in the conflict condition, the test item that matched the target on the less salient component mismatched it on the more salient component, whereas the mismatching test item matched the target on the more salient component. Finally, in the neutral condition, the more salient component was

150


B

A

C

Figure 3 Example of stimuli used by Yao and Sloutsky (2010). Small circles with symbols inside are task relevant, whereas the large colorful stimuli are task irrelevant. The item below is the target, whereas two items above are test stimuli. Participants are asked to match small circles of the target item with the small circles of one of the test items. (A) Supportive condition, (B) conflict condition, (C) neutral condition.

fixed across the target and test items, so that the task could not be performed on the basis of this component. Therefore, the Neutral condition served as the baseline for task performance. If participants ignore task-irrelevant information, their performance should not differ across the conditions. Alternatively, if participants cannot ignore the more salient (yet task-irrelevant) component, their performance should be above the baseline in the supportive condition and below the baseline in the conflict condition. In the simple matching task, participants were asked to match the target and one of the test items on the basis of the less salient component. Whereas 4-to 5-yearolds performed equally well across the conditions, 3- to 4-year-olds exhibited significant decrease in matching accuracy in the conflict condition. Furthermore, increase in task demands resulted in 3- to 4-year-olds and 4- to 5-year-olds failing the conflict condition, while succeeding in the neutral and supportive conditions. These findings support the idea that


151

young children have difficulty selectively focusing on some features, while ignoring other features, especially when these other features are more salient than the target features. While having difficulty attending selectively to few category-relevant features, young children can ably learn multiple dense categories without supervision, and this learning is often implicit. For example, in a recent study, Sloutsky and Fisher (2008) presented 4- and 5-year-olds with a contingency learning task. On each trial, participants were presented with triads of items. Sometimes, the triads appeared in Context 1 (i.e., they were presented on a yellow background at the top-right corner of a computer screen), and sometimes the triads appeared in Context 2 (i.e., they were presented on a green background at the bottom-left corner of a computer screen). When the triads appeared in Context 1, shape was predictive of an outcome, whereas when they appeared in Context 2, color was predictive of the outcome. Even though no feedback was provided, children quickly learned the dimension-context contingencies and generalized by shape in Context 1 and by color in Context 2. This finding is remarkable given that children exhibited little awareness of what they had learned (Experiment 2) and failed to exhibit this flexibility when they were explicitly told to focus on a given predictor (Experiment 3). These finding suggests that the learned flexibility was achieved by implicit learning, which is characteristic of the compression-based system. In sum, young children have no difficulty spontaneously learning dense categories, whereas they have greater difficulty learning categories based on few relevant features. We argue that this difficulty may stem from immaturities of selective attention, with young children having difficulty focusing on few relevant features while ignoring multiple irrelevant features.

2.2. Salience and Categorization As discussed above, learning categories bound by multiple correlated features can be accomplished in the absence of mature attentional selectivity. Similarly, such selectivity is likely unnecessary for learning a category defined by a single but highly salient feature that captures attention automatically. Could it be that conceptual knowledge can be construed as such a feature? This issue has been explored in recent studies that used a task-switch paradigm. Taskswitch paradigms are typically used to investigate executive control by presenting participants with a task in which performance demands change in the middle of the task, such as the Wisconsin Card Sort test (Berg, 1948). In a child-friendly adaptation of this test, the dimensional change card sort (DCCS) task children are presented with a set of cards depicting familiar objects that differ on two dimensions, such as color and shape (e.g., red and blue flowers, and red and blue boats; Zelazo, Frye, & Rapus, 1996). Children are first asked to sort cards based on one dimension and then to switch sorting

152


based on the other dimension. Despite understanding the instructions and reminders of the sorting rule given on every trial, children younger than 4 years of age often perseverate in sorting by the original dimension. Importantly, the pattern of perseveration is symmetric: it does not matter whether children start sorting by color or by shape—postswitch accuracy is reduced regardless of the starting sorting dimension. Perseveration errors described above are one type of switch costs—robust decrease in performance after task switch. The magnitude of switch costs is influenced by a number of factors; the most relevant of these for the study described below is saliency of the postswitch cues. Highly salient postswitch cues reduce switch costs in both children and adults, presumably because salient cues capture attention automatically and thus reduce demands on executive control (Fisher, 2011; Koch, 2001). For example, postswitch accuracy in the DCCS task can be nearly as high as preswitch accuracy in 3-year-old children when they are switching from sorting by a less salient dimension to sorting by a more salient dimension (Fisher, 2011). Based on the findings that salient postswitch cues reduce switch costs and allow young children to successfully switch to a new sorting task, it is possible to empirically evaluate the possibility that “deep” conceptual features are more salient than “surface” features for young children’s category judgments. If conceptual information is central to categorization early in development (and thus highly salient) whereas perceptual information is peripheral (and thus less salient than), then switch costs should be lower when children switch from categorizing by perceptual information to categorizing by conceptual information. However, if perceptual information is more salient than conceptual information in early categorization, the opposite pattern of performance should be observed. To test these possibilities, Fisher (2009) presented 3-, 4-, and 5-year-old children with a novel task, in which participants were asked to categorize the same set of objects twice and categorization instructions were changed in the middle of the task. Stimuli in this task consisted of iconic images of well-known objects (a familiarity check administered after the experiment, proper confirmed that all children were indeed familiar with the materials). These images were organized into triads that put category information in conflict with appearance similarity. For example, an open red umbrella was paired with a folded blue-and-yellow umbrella to create a category match and with a red mushroom to create an appearance match (see Figure 4). Children were asked to categorize the objects twice: once by grouping together items that “look similar” and once by grouping together items that are “the same kind of thing” (the order of tasks was randomized for each child). Proportion of correct responses in each task was then used to calculate perceptual and conceptual costs of switching. Results of this study indicated that for every age group tested in the study, the costs of switching to categorizing by kind were higher than costs


153

Figure 4 Example of a triad used in Fisher (2009). Target: open red umbrella, appearance match: red mushroom, category match: folded multicolor umbrella. (For interpretation of the references to color in this figure legend, the reader is referred to the Web version of this chapter.)

of switching to categorizing by appearance. Across all three age groups, costs of switching to categorizing by object kind were nearly 25% higher than costs of switching to categorizing by appearance. Furthermore, in 4- and 5-year-old children switching to categorizing by appearance resulted in negligibly low switch costs (8%, not different from zero), whereas switching to categorizing by kind produced high switch costs (39%, significantly above zero). These findings indicate that perceptual information is more salient than conceptual information.

3. The Role of Labels in the Development of Categorization Although young children might have difficulty spontaneously focusing on few category-relevant features, their attention to these features could be guided by words (Gelman, 2003; Jaswal, 2004). In particular, category labels (especially when presented as count nouns) may serve as invitations to form categories and they attract attention to within-category commonalities (Gelman & Markman, 1986; Markman, 1991). As we argue in this section, there is little evidence that this is the case.

3.1. Labels as Features Contributing to Similarity There is much evidence that category labels guide category learning in adults (Yamauchi & Markman, 2000). The overall reasoning behind this work is that if labels are category markers, they should be treated differently

154


from the rest of features (such as shape, color, size, etc.). However, this may not be the case if labels are features. Therefore, inferring a label when features are given (i.e., a classification task) should elicit different performance from a task of inferring a feature when the label is given (i.e., a feature induction task). To test these ideas, Yamauchi and Markman (2000; see also Markman & Ross, 2003 for a review) developed a category learning task that was presented under either classification or feature induction learning condition. There were two categories, C1 and C2, denoted by two labels, L1 and L2. In the classification task, they were presented with a creature that was not labeled, and the task was to predict the category label. The critical condition was the case when an item was a member of C1, but was similar to C2, with the dependent variable being the proportion of C1 responses. The results indicated that there were significantly more category-based responses in the induction condition (where participants could rely on the category label) than in the categorization condition (where participants had to infer the category label). It was concluded therefore that category labels differed from other features in that participants treated labels as category markers. At the same time, there is much evidence that, in contrast to adults, early in development, labels are features of objects, with similarity of compared entities computed over both appearance and labeling attributes (Sloutsky, 2003; Sloutsky & Fisher, 2004; Sloutsky, Lo, & Fisher, 2001). These intuitions were implemented in a similarity-based model of early generalization abbreviated as SINC (similarity-induction-naming-categorization). According to SINC, early in development, induction is a function of the overall perceptual similarity computed over weighted visual and auditory feature (Sloutsky & Fisher, 2004). The theory underlying SINC assumes that auditory information (including linguistic labels) often has greater attentional weights than visual information early in development (support for this assumption is reviewed in the next section). Therefore, when two entities A and B look equally similar to the target entity T, but only one of these entities (e.g., B) is referred to by the same label as the target, perceived similarity of test B to the target will be greater than that of test A to the target. In this situation, children should be highly likely to rely on matching labels to perform tasks such as property induction and categorization. However, when test A looks markedly more similar to the target than test B, but test B shares the name with the target, children may perceive both A and B as equally similar to the target. In this situation, children should be equally likely to rely on matching labels and on matching appearances when performing property induction and categorization tasks. The above predictions were supported in a series of studies in which 4- to 5-year-old children were presented with similarity judgment, property induction, and categorization tasks, with SINC accounting for over 90% of observed variance in performance on these tasks (Sloutsky & Fisher, 2004). Furthermore, SINC accurately predicted a bimodal distribution of

155


responses on the same property induction task and stimuli used in prior research to argue that children are more likely to rely on matching labels than on matching appearances in the course of induction (Gelman & Markman, 1986; see Figure 5A). In particular, as predicted by SINC, preschool-age children were above chance in relying on matching labels

Proportion of label-based responses

A 1

Predicted by SINC Observed in 4- to 5-year-olds in Sloutsky and Fisher (2004)

0.9 0.8

*

*

0.7

*

0.6 0.5 0.4 0.3 0.2 *Indicates above-chance (50%) means, p < 0.05

0.1 0

1

2

3

4

5

6

7

8

9

10 Average

Picture triad B

(Triad 1 on the graph above)

C

(Triad 2 on the graph above)

Bat Starfish Bird

Pinecone

Bird Starfish

Figure 5 Summary of results from Sloutsky and Fisher (2004; Experiment 4). (A) Presents responses observed in preschool-age children by Sloutsky and Fisher (2004) and predicted by SINC on the property induction task and materials originally used by Gelman and Markman (1986). Examples of triads used originally in Gelman and Markman (1986) and then in Sloutsky and Fisher (2004) are presented in (B) and (C). In (B), the test item that was referred to by the label different from the target (i.e., bat) looks overwhelmingly more similar to the target than the test item that was referred to by the same label as the target (i.e., bird). This triad corresponds to triad #1 on the graph in (A). In (C), both test items look equally similar to the target item (for results of the similarity calibration with children, see Sloutsky & Fisher, 2004; Experiment 4). This triad corresponds to triad #2 on the graph in (A).

156


when none of the test items were overwhelmingly perceptually similar to the target (e.g., when one test item had the same shape as the target and the other test item had the same texture and color as the target; see Figure 5C). In contrast, when one of the test items looked overwhelmingly similar to the target and another test item had the same name as the target (see Figure 5B), children’s reliance on matching labels did not exceed chance level. As stated above, SINC assumes that auditory features may have a greater attentional weight than visual features early in development. The reasons as to why this might be the case are reviewed in the next section.

3.2. Labels Overshadow Visual Input There is evidence that auditory input may affect attention allocated to corresponding visual input (Napolitano & Sloutsky, 2004; Robinson & Sloutsky, 2004; Sloutsky & Napolitano, 2003; Sloutsky & Robinson, 2008), and these effects may change in the course of learning and development. In particular, linguistic labels may strongly interfere with visual processing in infants and young children, although these interference effects may somewhat weaken with age (Sloutsky & Robinson, 2008; see also Robinson & Sloutsky, 2007a,b). These issues have been examined in depth in a series of experiments by Sloutsky and colleagues (e.g., Napolitano & Sloutsky, 2004; Robinson & Sloutsky, 2004; Sloutsky & Napolitano, 2003). In these experiments, 4-year-olds and adults were presented with a compound target stimulus, consisting of simultaneously presented auditory and visual components (AUDTargetVISTarget). Participants were presented with a target item, which was followed immediately by a test item (the test item could be either the same as the target or different) and participants had to determine whether the test item was exactly the same as the target. There were four types of test items: (1) AUDTargetVISTarget, which was the old target item; (2) AUDTargetVISNew, which had the target auditory component and a new visual component; (3) AUDNewVISTarget, which had the target visual component and a new auditory component; or (4) AUDNewVISNew, which had a new visual component and a new auditory component. The task was to determine whether each presented test item was exactly the same as the target (i.e., both the same auditory and visual components) or a new item (i.e., differed on one or both components). It was reasoned that if participants process both auditory and visual stimuli, they should correctly respond to all items by accepting old target items and rejecting all other test items. Alternatively, if they fail to process the visual component, they should falsely accept AUDTargetVISNew items,


157

while correctly responding to other items. Finally, if they fail to process the auditory component, they should falsely accept AUDNewVISTarget items, while correctly responding to other items. In one experiment (Napolitano & Sloutsky, 2004), speech sounds were paired with either geometric shapes or pictures of unfamiliar animals. Results indicated that while children ably processed either stimulus in the unimodal condition, they failed to process visual input in the cross-modal condition. Furthermore, a yet unpublished study by Napolitano and Sloutsky indicates that interference effects attenuate gradually in the course of development. There is also evidence that this dominance of auditory input is not under strategic control: even when instructed to focus on visual input young children had difficulties doing so (Napolitano & Sloutsky, 2004; Robinson & Sloutsky, 2004). In one of the experiments described in Napolitano and Sloutsky (2004), 4-year-olds were explicitly instructed to attend to visual stimuli, with instructions repeated before each trial. However, despite the repeated explicit instruction to attend to visual stimuli, 4-year-olds continued to exhibit auditory dominance. These results suggest that auditory dominance is unlikely to stem from deliberate selective attention to a particular modality, but it is more likely to stem from automatic pulls on attention. If linguistic labels attenuate visual processing, such that children ably process a label, but they do so to a lesser extent the corresponding visual input, then these findings can explain the role of labels in categorization tasks. In particular, items that share a label may appear more similar than the same items presented without a label. In other words, early in development, labels may function as features contributing to similarity, and their role may change in the course of development. In fact, there is evidence supporting this possibility (e.g., Sloutsky & Fisher, 2004; Sloutsky & Lo, 1999).

3.3. Effect of Semantic Similarity and Co-Occurrence Probability of Labels If labels are merely features contributing to similarity, then why do semantically similar labels (as opposed to identical labels) affect young children’s generalization? Specifically, Gelman and Markman (1986; Experiment 2) demonstrated that 4- to 5-year-old children generalize unobservable properties from one object to another not only when the objects are referred to by identical labels (which can be readily explained by the label overshadowing hypothesis) but also when objects are referred to by semantically similar labels. For example, in this study, children could be told that a “rabbit eats bugs” whereas a “squirrel eats grass,” and asked whether the target item (called a “rabbit” in the identical labels condition and a “bunny” in the synonyms condition) “eats bugs like the rabbit” or “eats grass like the squirrel.”

158


Gelman and Markman found that children generalized properties to categorically similar items at a level above that expected by chance in both the identical labels and the synonyms conditions. Importantly, children’s performance with synonyms was statistically equivalent to their performance with identical labels (63% and 67% of category-based responses, respectively). These findings suggested that labels not merely promote label-based inferences but truly point to commonalities in kind. However, Fisher (2010a) recently suggested that some label pairs in the Gelman and Markman (1986) study included pairs of labels that were not only semantically similar but also co-occurred in child-directed speech (e.g., bunny-rabbit, puppy-dog) according to the CHILDES database (MacWhinney, 2000; see Fisher, 2010a for details of the co-occurrence analyses). Co-occurrence of words in natural language has been argued to give rise to strong lexical associations (Brown & Berko, 1960; McKoon & Ratcliff, 1992). For instance, when people are presented with the words puppy and bunny, the probability of obtaining the words dog and rabbit in response is 71% and 74%, respectively. In comparison, the probability of producing the word stone in response to rock (a pair that was also used in Gelman and Markman’s synonyms condition) is only 3% (Nelson, McEvoy, & Schreiber, 1998).1 It is possible that effects reported by Gelman and Markman’s (1986) stemmed from labels pointing to commonalities in kind and promoting category-based reasoning. However, it is also possible that responses (at least to some of the items) stemmed from associative priming: children’s responses could be based not on the reasoning that bunnies and rabbits are the same kind of animal, but on the fact that the word bunny primed the word rabbit, but not the word squirrel. Fisher, Matlen, & Godwin (in press) recently obtained evidence in support of the latter possibility. They presented a group of 4-year-old children and adults with an induction task in either an identical labels condition or a semantically similar labels condition. Within each labeling condition, participants were presented either with labels that were likely to co-occur in the CHILDES database (e.g., bunny-rabbit, puppy-dog, kittycat) or with labels that were unlikely to co-occur (e.g., alligator-crocodile, mouse-rat, toad-frog). All labels referred to animal names and all properties consisted of blank predicates. For example, in the semantically similar labels condition, participants could be told the “alligator has matlen inside” and asked whether the “crocodile” (a semantically related test item) or the “hippo” (an unrelated test item) was likely to share this property.

1

Of course, words can become associated by means other than co-occurrence (e.g., thematic relatedness); however, co-occurrence is argued to play an important role in the formation of lexical associations (Brown & Berko, 1960; McKoon & Ratcliff, 1992).


159

Adults were at ceiling in performing category-based induction in all conditions. In the identical labels condition, children’s performance was above chance. However, in the semantically similar labels condition, children performed above chance only with co-occurring labels, but their performance was at chance when nonco-occurring semantically similar labels were used. A picture identification task administered immediately after the experiment proper confirmed that children were well familiar with all the labels used in this study (accuracy on the picture identification task was 99%). Results of a follow-up experiment indicated that less than 20% of 4-year-old children spontaneously utilize semantic similarity when labels are unlikely to co-occur in natural language and 50% of 5-year-old children can do so; however, the majority of children do not spontaneously utilize semantic similarity of nonco-occurring labels in reasoning tasks until 6 years of age (a similar developmental pattern was reported by Ramesh, Matlen, and Fisher (2010) in a study that utilized an analogical reasoning task). These findings pose a serious challenge to the notion that labels guide attention to commonalities in kind from as early as 2 years of age (Gelman & Coley, 1990).

3.4. Flexible Adjustment of Attention Allocated to Labels and Appearances Another challenge to the notion that labels are centrally important for categorization and appearances are peripheral from early in development comes from a series of studies that examined the flexibility with which children can rely on label and appearance attributes in the course of induction. If reliance on labels and appearances in induction and categorization tasks stems from automatic allocation of attention to these attributes, then it should be possible to modify children’s reliance on these attributes by manipulating the amount of attention directed to labels and appearances. If however, children believe that labels are more theoretically central than appearances, then such a change should be difficult (if not impossible), because beliefs are notoriously resistant to change. When children and adults hold strong beliefs, they tend to disregard evidence that conflicts with these beliefs. For example, when children and adults come to a task with preexisting beliefs, they tend to incorrectly encode information in such a way that is consistent with their beliefs—phenomenon known as illusory correlation (Hamilton & Rose, 1980; Johnston & Jacobs, 2003; Lord, Ross, & Lepper, 1979; Meehan & Janik, 1990). The knowledge-based approach predicts that the status of conceptually central attributes (i.e., labels) and conceptually peripheral attributes (i.e., appearances) is relatively fixed and resistant to change, even when new data become available. The coherence-based approach predicts that children’s

160


reliance on various attributes in the course of induction is shaped by data— for instance, data on how well a particular attribute correlates with a particular outcome and thus how useful this attribute might be in predicting the outcome. Therefore, this approach predicts a high degree of flexibility in relying on labeling and appearance attributes in young children who are yet to form beliefs about theoretical importance of labels. There is ample evidence suggesting that attention allocated to different perceptual attributes can be flexibly changed in both animals and humans by manipulating the predictive value of an attribute: when a particular cue is consistently predictive, attention allocated to this cue increases automatically, whereas when a cue is consistently nonpredictive, attention allocated to this cue decreases (see Hall, 1991 for a review). Therefore, manipulating predictive values of labels and appearances in the course of induction should change children’s reliance on these attributes. Evidence in support of this hypothesis comes from the studies in which 4- to 5-year-old children were given evidence that a particular cue—such as a shared label, similar appearance, or shared inheritance (i.e., the items were introduced as mother and baby)—does not correlate with the outcome (Fisher, 2010b; Fisher & Sloutsky, 2006; Sloutsky & Spino, 2004). During the training phase in these studies, children’s task was to predict an unobservable property of a target object (e.g., whether it has thick blood or thin blood) based on the information about the test objects. Children were presented with one target object and up to three test objects, each of which shared only one attribute with the target (e.g., each test object could have the same label as the target, look similar to the target, or have given birth to the target). After each response, children were provided with feedback indicating whether or not the response was correct. The correctness of the response varied based on the condition to which the children were assigned. For example, in the label training condition, children were provided with positive feedback for making label-based predictions, and in the appearance training condition, children were provided with positive feedback for making appearance-based predictions. In some of the studies, the feedback was explicit (i.e., children were told whether their response was correct or not; Sloutsky & Spino, 2004), and in other studies, the feedback was implicit (i.e., children were shown a short cartoon when they made correct predictions; Fisher, 2010b; Fisher & Sloutsky, 2006). The training phase was immediately followed by a transfer phase in which children were tested on an induction task in the absence of feedback. The transfer phase included novel stimuli on which children were not trained (see Figure 6 for examples). A consistent pattern of results emerged from these studies suggesting that the pattern of responses in the transfer phase was (1) different from the baseline pattern of induction and (2) consistent with the training condition to which children were assigned. In other words, children who were assigned to the label training condition relied

161


A

B

a zizi a bala

a gula

a zizi

a fika

a bala

Figure 6 Example of training stimuli (A) and testing stimuli (B) from Fisher and Sloutsky (2006). Children were trained that either similar appearance of matching labels were a predictive cue in a property induction task (Experiment 1) or a similarity judgment task (Experiment 2) using stimuli similar to those in (A). Transfer of this learning was tested in a property induction task using stimuli similar to those in (B).

predominantly on labels during the transfer phase, children who were assigned to the appearance training condition relied predominantly on appearances, and children assigned to the inheritance training relied predominantly on inheritance. In all of these conditions, children’s performance was different from the baseline pattern of induction, which typically reflects children’s integration of information from multiple cues (Sloutsky & Fisher, 2004). Furthermore, Sloutsky and Spino (2004) showed that the trained pattern of performance persisted when preschool-age children were tested in a delayed transfer task, which was administered 3 months after the training phase, by a different experimenter, in a different location, and used a novel set of stimuli. Overall, results described above suggest that reliance on labeling and appearance attributes in the course of induction is flexible: when either attribute becomes nonpredictive in the course of training, reliance on this attribute decreases markedly during testing. These findings challenge the position that effects of labels on induction stem from young children’s belief in their conceptual importance. It is not clear how this hypothesis can account for the fact that a relatively modest amount of training—10 trials in Sloutsky and Spino and 16 trials in Fisher and Sloutsky (2006) resulted in a flexible shift of attention away from predictors that are supposed to be theoretically central (i.e., linguistic labels) to those that are supposed to be theoretically peripheral (i.e., appearances). At the same time, these findings support the idea that reliance on labels and appearances in the course of induction stems from allocation of attention to more predictive (and hence more salient) features and not from theoretical beliefs about centrality of certain attributes over others. In sum, categorization on the basis of few theoretically relevant features may hinge on selective attention that exhibits a great degree of immaturity early in development. In principle, this immaturity could have been

162


compensated for if early categorization were assisted by language with words attracting attention to important within-category commonalities or across-category differences. However, this is not likely due to immaturities in cross-modal processing, with auditory input overshadowing visual input early in development. We presented arguments that instead of being based on few category-relevant features, early categorization is based on automatic attention to bundles of category-relevant features or to few highly salient ones. Although much research is needed to flesh out the details, there is sufficient evidence to consider this mechanism a plausible candidate for how categories are learned early in development.

4. Early Categorization: What Develops? If categorization starts out as implicit learning of perceptual categories, how do people acquire the ability to learn highly abstract categories as well as concepts of mathematics, science, morality, or law? Although we do not have a precise answer to this question, we believe that there are several critical steps in this process. Perhaps, the most important step is the ability to selectively attend to category-relevant information while ignoring category-irrelevant information. This ability is critical for learning categories that are based on few relevant features, and it may be even more critical when these features are unobservable, as is the case of highly abstract concepts. For example, many abstract categories are bound by unobservable relational characteristics, while having very few (if any) perceptual features in common. Consider such concepts as fairness and reciprocity. While these concepts are lexicalized categories, one would be hard pressed to find perceptual commonalities in instances of reciprocity. Therefore, learning of such a category puts high demands on selective attention as one needs to ignore much irrelevant perceptual information. As we discussed in this chapter, the ability to do so has a protracted developmental course and is likely immature long after children make great strides in concept acquisition. As a result of this immaturity of selective attention, some of the abstract concepts are beyond children’s reach. These developmental limitations could be highly adaptive as they prevent young category learners from entertaining one commonality at a time. As a result, young learners have to automatically focus on bundles of overlapping features. Unlike isolated commonalities, that may not generate a clear distinction among categories, bundled commonalities are more likely to do so. Therefore, there is little surprise that the first categories that children acquire are based on multiple overlapping features (e.g., cats, dogs, birds, or people) and that even children of scientists and lawyers do not start their learning with concepts of gravity or litigation.


163

ACKNOWLEDGMENTS Writing of this article was supported by grants from the NSF (BCS-0720135); the Institute of Education Sciences, U.S. Department of Education (R305B070407); and NIH (R01HD056105). The opinions expressed are those of the authors and do not represent views of the awarding organizations.

REFERENCES Baddeley, A. D. (1986). Working memory. Oxford: Clarendon Press. Barsalou, L. W. (1991). Deriving categories to achieve goals. In G. H. Bower (Ed.), The psychology of learning and motivation: Advances in research and theory, 27. New York: Academic Press. Berg, E. A. (1948). A simple objective technique for measuring flexibility in thinking. Journal of General Psychology, 39, 15–22. Bomba, P. C., & Siqueland, E. R. (1983). The nature and structure of infant form categories. Journal of Experimental Child Psychology, 35, 294–328. Booth, A. E., & Waxman, S. R. (2002). Word learning is ‘smart’: Evidence that conceptual information affects preschoolers’ extension of novel words. Cognition, 84, B11–B22. Booth, A., Waxman, S. R., & Huang, Y. T. (2005). Conceptual information permeates word learning in infancy. Developmental Psychology, 41, 491–505. Brown, R., & Berko, J. (1960). Word association and the acquisition of grammar. Child Development, 31, 1–14. Carey, S. (2009). The origin of concepts. New York: Oxford University Press. Chase, A. (2001). Music discriminations by carp (Cyprinus carpio). Animal Learning & Behavior, 29, 336–353. Chater, N., & Christiansen, M. H. (2010). Language acquisition meets language evolution. Cognitive Science, 34, 1131–1157. Chomsky, N. (1980). Rules and representations. New York: Columbia University Press. Demetriou, A., Christou, C., Spanoudis, G., & Platsidou, M. (2002). The development of mental processing: Efficiency, working memory, and thinking. Monographs of the Society of Research in Child Development, 67, SN 268. Elman, J. L. (1993). Learning and development in neural networks: The importance of starting small. Cognition, 48, 71–99. Fisher, A. V. (2009). Does conceptual information take precedence over perceptual information early in development? Evidence from perseveration errors. In N. Taatgen & H. van Rijn (Eds.), Proceedings of the XXXI annual conference of the cognitive science society (pp. 1330–1335). Fisher, A. V. (2010a). What’s in the name? Or how rocks and stones are different from dogs and puppies. Journal of Experimental Child Psychology, 105, 198–212. Fisher, A. V. (2010b). Mechanisms of induction early in development. In M. Banich & D. Caccamise (Eds.), Generalization of knowledge: Multidisciplinary perspectives (pp. 89– 112). New York: Psychology Press. Fisher, A. V. (2011). Automatic shifts of attention in the dimensional change card sort task: Subtle changes in task materials lead to flexible switching. Journal of Experimental Child Psychology, 108, 211–219. Fisher, A. V., & Sloutsky, V. M. (2006). Flexible attention to labels and appearances in early induction. In R. Sun & N. Miyake (Eds.), Proceedings of the XXVIII annual conference of the cognitive science society (pp. 220–225).

164


Fisher, A.V., Matlen, B., & Godwin, K.E. (in press). Semantic similarity of labels and inductive generalization: Taking a second look. Cognition. French, R. M., Mareschal, D., Mermillod, M., & Quinn, P. C. (2004). The role of bottomup processing in perceptual categorization by 3- to 4-month-old infants: Simulations and data. Journal of Experimental Psychology. General, 133, 382–397. Gelman, R. (1990). First principles organize attention to and learning about relevant data: Number and the animate-inanimate distinction as examples. Cognitive Science, 14, 79–106. Gelman, S. A. (2003). The essential child: Origins of essentialism in everyday thought. New York: Oxford University Press. Gelman, S. A., & Coley, J. D. (1990). The importance of knowing a dodo is a bird: Categories and inferences in 2-year-old children. Developmental Psychology, 26, 796–804. Gelman, S. A., & Markman, E. (1986). Categories and induction in young children. Cognition, 23, 183–209. Go´mez, R., & Maye, J. (2005). The developmental trajectory of non-adjacent dependency learning. Infancy, 7, 183–206. Hall, G. (1991). Perceptual and associative learning. New York: Oxford University Press. Hamilton, D. L., & Rose, T. L. (1980). Illusory correlation and the maintenance of stereotypic beliefs. Personality and Social Psychology, 39, 832–845. Jaswal, V. K. (2004). Don’t believe everything you hear: Preschoolers’ sensitivity to speaker intent in category induction. Child Development, 3, 279–300. Johnston, K. E., & Jacobs, J. E. (2003). Children’s illusory correlations: The role of attentional bias in group impression formation. Journal of Cognition and Development, 4, 129–160. Keil, F. C., Smith, C. S., Simons, D., & Levin, D. (1998). Two dogmas of conceptual empiricism. Cognition, 65, 103–135. Kloos, H., & Sloutsky, V. M. (2008). What’s behind different kinds of kinds: Effects of statistical density on learning and representation of categories. Journal of Experimental Psychology. General, 137, 52–72. Koch, I. (2001). Automatic and intentional activation of task sets. Journal of Experimental Psychology: Learning, Memory, and Cognition, 27, 1474–1486. Kruschke, J. K., & Blair, N. J. (2000). Blocking and backward blocking involve learned inattention. Psychonomic Bulletin & Review, 7, 636–645. Lakoff, G. (1987). Women, fire, and dangerous things: What categories reveal about the mind. Chicago: The University of Chicago Press. Lord, C. G., Ross, L., & Lepper, M. R. (1979). Biased assimilation and attitude polarization: The effects of prior theories on subsequently considered evidence. Journal of Personality and Social Psychology, 37, 2098–2109. MacWhinney, B. (2000). The CHILDES project: Tools for analyzing talk (3rd ed.). Mahwah, NJ: Lawrence Erlbaum Associates. Markman, E. M. (1991). Categorization and naming in children: Problems of induction. Cambridge: MA, MIT Press. Markman, A. B., & Ross, B. H. (2003). Category use and category learning. Psychological Bulletin, 129, 592–613. McKoon, G., & Ratcliff, R. (1992). Spreading activation versus compound cue accounts of priming: Mediated priming revisited. Journal of Experimental Psychology: Learning, Memory, and Cognition, 18, 1155–1172. Meehan, A. M., & Janik, L. M. (1990). Illusory correlation and the maintenance of sex role stereotypes in children. Sex Roles, 22, 83–95. Murphy, G. L. (2002). The big book of concepts. Cambridge: MA, MIT Press. Napolitano, A. C., & Sloutsky, V. M. (2004). Is a picture worth a thousand words? Part II: The flexible nature of modality dominance in young children. Child Development, 75, 1850–1870.


165

Nelson, D. L., McEvoy, C. L., & Schreiber, T. A. (1998). The University of South Florida word association, rhyme, and word fragment norms. http://www.usf.edu/FreeAssociation/. Newport, E. L. (1990). Maturational constraints on language learning. Cognitive Science, 14, 11–28. Nosofsky, R. M. (1986). Attention, similarity, and the identification-categorization relationship. Journal of Experimental Psychology: General, 115, 39–57. Pearce, J. M. (1994). Discrimination and categorization. In N. J. Mackintosh (Ed.), Handbook of Perception and Cognition, vol. 9: Animal Learning and Cognition (pp. 110–134). San Diego, CA: Academic Press. Pullum, G. K., & Scholz, B. (2002). Empirical assessment of stimulus poverty arguments. The Linguistic Review, 19, 9–50. Quinn, P. C., Eimas, P. D., & Rosenkrantz, S. L. (1993). Evidence for representations of perceptually similar natural categories by 3- and 4-month-old infants. Perception, 22, 463–475. Rakison, D. H., & Poulin-Dubois, D. (2001). Developmental origin of the animate-inanimate distinction. Psychological Bulletin, 127, 209–228. Ramesh, S., Matlen, B., & Fisher, A. V. (2010). Development of relational reasoning with semantically similar labels. In S. Ohlsson & R. Catrambone (Eds.), Proceedings of the XXXII annual conference of the cognitive science society (pp. 1499–1504). Ribar, R. J., Oakes, L. M., & Spalding, T. L. (2004). Infants can rapidly form new categorical representations. Psychonomic Bulletin & Review, 11, 536–541. Robinson, C. W., & Sloutsky, V. M. (2004). Auditory dominance and its change in the course of development. Child Development, 75, 1387–1401. Robinson, C. W., & Sloutsky, V. M. (2007a). Visual processing speed: Effects of auditory input on visual processing. Developmental Science, 10, 734–740. Robinson, C. W., & Sloutsky, V. M. (2007b). Linguistic labels and categorization in infancy: Do labels facilitate or hinder? Infancy, 11, 223–253. Rogers, T. T., & McClelland, J. L. (2004). Semantic cognition: A parallel distributed processing approach. Cambridge, MA: MIT Press. Rogers, T. T., & McClelland, J. L. (2008). Precis of semantic cognition, a parallel distributed processing approach. Behavioral and Brain Sciences, 31, 689–749. Saffran, J. R., Aslin, R. N., & Newport, E. L. (1996). Statistical learning by 8-month-old infants. Science, 274, 1926–1928. Sloutsky, V. M. (2003). The role of similarity in the development of categorization. Trends in Cognitive Sciences, 7, 246–251. Sloutsky, V. M. (2010). From perceptual categories to concepts: What develops? Cognitive Science, 34, 1244–1286. Sloutsky, V. M., & Fisher, A. V. (2004). Induction and categorization in young children: A similarity-based model. Journal of Experimental Psychology. General, 133, 166–188. Sloutsky, V. M., & Fisher, A. V. (2008). Attentional learning and flexible induction: How mundane mechanisms give rise to smart behaviors. Child Development, 79, 639–651. Sloutsky, V. M., & Lo, Y.-F. (1999). How much does a shared name make things similar? Part 1: Linguistic labels and the development of similarity judgement. Developmental Psychology, 35, 1478–1492. Sloutsky, V. M., & Napolitano, A. C. (2003). Is a picture worth a thousand words? Preference for auditory modality in young children. Child Development, 74(3), 822–833. Sloutsky, V. M., & Robinson, C. W. (2008). The role of words and sounds in visual processing: From overshadowing to attentional tuning. Cognitive Science, 32, 354–377. Sloutsky, V. M., & Spino, M. A. (2004). Naive theory and transfer of learning: When less is more and more is less. Psychonomic Bulletin & Review, 11, 528–535.

166


Sloutsky, V. M., Lo, Y.-F., & Fisher, A. V. (2001). How much does a shared name make things similar: Linguistic labels and the development of inductive inference. Child Development, 72, 1695–1709. Smith, L., Jones, S., & Landau, B. (1996). Naming in young children: A dumb attentional mechanism? Cognition, 60, 143–171. Smith, L. B. (1989). A model of perceptual classification in children and adults. Psychological Review, 96, 125–144. Smith, L. B., Colunga, E., & Yoshida, H. (2010). Contextually-cued attention as a mechanism in development. Cognitive Science, 1287–1314. Smith, L. B., & Yu, C. (2008). Infants rapidly learn word-referent mappings via crosssituational statistics. Cognition, 106, 333–338. Soja, N. N., Carey, S., & Spelke, E. S. (1991). Ontological categories guide young children’s inductions of word meaning: Object terms and substance terms. Cognition, 38, 179–211. Spelke, E. S., & Kinzler, K. D. (2007). Core knowledge. Developmental Science, 10, 89–96. Wyttenbach, R., May, M., & Hoy, R. (1996). Categorical perception of sound frequency by crickets. Science, 273, 1542–1544. Yamauchi, T., & Markman, A. B. (2000). Inference using categories. Journal of Experimental Psychology: Learning, Memory, and Cognition, 26, 776–795. Yao, X., & Sloutsky, V. M. (2010). Selective attention and development of categorization: An eye tracking study. In S. Ohlsson & R. Catrambone (Eds.), Proceedings of the XXXII Annual Conference of the Cognitive Science Society (pp. 1980–1985). Zelazo, P. D., Frye, D., & Rapus, T. (1996). An age-related dissociation between knowing rules and using them. Cognitive Development, 11, 37–63. Zentall, T. R., Wasserman, E. A., Lazareva, O. F., Thompson, R. K. R., & Rattermann, M. J. (2008). Concept learning in animals. Comparative Cognition & Behavior Reviews, 3, 13–45.

C H A P T E R

S I X

Systems of Category Learning: Fact or Fantasy? Ben R. Newell, John C. Dunn, and Michael Kalish Contents 1. Introduction 1.1. Multiple Systems of Category Learning 2. Review and Critique of the Evidence I: Probabilistic Category Learning 2.1. Neuropsychological Dissociations 2.2. Reevaluating the Neuropsychological Evidence 2.3. Behavioral Dissociations 2.4. Reconsidering Behavioral Dissociations 2.5. Neuroimaging 2.6. Reimagining Neuroimaging 2.7. Section Summary 3. Review and Critique of the Evidence II: Deterministic Category Learning 3.1. Neuropsychological Dissociations 3.2. Reevaluating the Neuropsychological Evidence 3.3. Behavioral Dissociations 3.4. Reconsidering Behavioral Dissociations 3.5. Neuroimaging 3.6. Reimagining Neuroimaging 3.7. Section Summary 4. Reexamining Some Fundamental Assumptions 4.1. State-Trace Analysis 4.2. The Inferential Limits of Dissociations 4.3. A State-Trace Reanalysis of Behavioral and Other Dissociations 4.4. Interpretation of Two-Dimensional State-Trace Plots: The Role of Confounds 4.5. Interpretation of Two-Dimensional State-Trace Plots: Systems, Processes, and Latent Variables 5. The Contribution of Mathematical Modeling 5.1. Mathematical/Computational Models of Category Learning 5.2. COVIS Psychology of Learning and Motivation, Volume 54 ISSN 0079-7421, DOI: 10.1016/B978-0-12-385527-5.00006-1

#

168 169 170 171 172 174 175 177 178 178 179 179 181 183 184 186 187 188 188 188 192 194 196 198 199 199 200


167

168

Ben R. Newell et al.

5.3. The Nature and Use of Formal Models 5.4. Model Comparisons 6. Discussion and Conclusions 6.1. Varieties of Explanation 6.2. Explanatory Power 6.3. Final Thoughts Acknowledgment References

201 204 206 207 208 209 210 210

Abstract Psychology abounds with vigorous debates about the need for one or more underlying mental processes or systems to explain empirical observations. The field of category learning provides an excellent exemplar. We present a critical examination of this field focusing on empirical, methodological, and mathematical modeling considerations. We review what is often presented as the “best evidence” for multiple systems of category learning and critique the evidence by considering three questions: (1) Are multiple-systems accounts the only viable explanations for reported effects? (2) Are the inferences sound logically and methodologically? (3) Are the mathematical models that can account for behavior sufficiently constrained, and are alternative (single-system) models applicable? We conclude that the evidence for multiple-systems accounts of category learning does not withstand such scrutiny. We end by discussing the varieties of explanation that can be offered for psychological phenomena and highlight why multiple-systems accounts often provide an illusory sense of scientific progress.

1. Introduction Psychology has seen an upsurge in theories and accounts that use multiple, qualitatively distinct systems to explain empirical phenomena. Researchers in the areas of judgment and decision making, reasoning, social cognition, associative learning, and memory, to name a few, have been swept up in a desire to describe and group behaviors as the products of different systems (see Evans, 2008; Keren & Schul, 2009; Mitchell, De Houwer, & Lovibond, 2009 for relevant discussions). This desire seems to have been exacerbated in recent years by the introduction of increasingly sophisticated brain imaging methods that allow us to “see” which parts of the brain are “recruited” by different tasks. The engagement of neuroanatomically distinct regions of the brain seems to compel the conclusion that tasks are subserved by distinct cognitive systems (Sherry & Schacter, 1987). The underlying rhetoric of such approaches is that the allocation of function to separable systems represents an advance in our understanding of particular phenomena (e.g., Ashby, Paul, & Maddox, in press;

Systems of Category Learning: Fact or Fantasy?

169

Evans, 2008; Poldrack & Foerde, 2008). Thus, describing a cognitive task as being solved by “System 1” which recruits “region X” of the brain is taken to be a better functional explanation than one which does not include such localization or identification with a particular system (but see Coltheart 2006; Page, 2006). Our aim in this chapter is to examine critically the evidence for and the assumptions underlying such multiple-systems views. We are not the first to undertake such a review, but despite the efforts of those who have pointed out the limitations and inconsistencies of the multiple-systems view (e.g., Gigerenzer & Reiger, 1996; Keren & Schul, 2009; Mitchell et al., 2009; Nosofsky & Zaki, 1998; Palmeri & Flanery, 2002; Shanks & St. John, 1994; Speekenbrink, Channon, & Shanks, 2008), it persists as the dominant view and is increasingly presented in popular science as almost accepted “fact” (e.g., Lehrer, 2009). Our vehicle for the exploration of these issues is human category learning. In recent years, increasingly sophisticated multiple-systems interpretations of category learning have appeared (e.g., Minda & Miles, 2010; Poldrack & Foerde, 2008), making category learning an ideal candidate for evaluation. First, we present what we interpret as the proponents’ “best” evidence for the existence of multiple category learning systems. We follow each section with a review of studies that challenge the multiple-systems interpretation of some of the empirical phenomena. In the second half of the chapter, we take a step back from the “for-andagainst” interpretations of particular experiments and examine, thoroughly and critically, the assumptions underlying multiple-system accounts. These include assumptions about what we can infer from behavioral dissociations, assumptions about the “bridges” between neuroscience and behavioral measures, and assumptions about what mathematical models can contribute to our understanding of category learning. We end the chapter by considering the varieties of explanation that can be offered for psychological phenomena and discuss why multiple-systems account can give rise to, what we consider, a false sense of progress and productivity.

1.1. Multiple Systems of Category Learning Many recent reviews of the literature on human category learning conclude that “category learning is mediated by multiple, qualitatively distinct systems” (Ashby & Maddox, 2005, p. 149), or that the multiple-system approach “is superior in its ability to account for a broad range of data from psychology and neuroscience” (Poldrack & Foerde, 2008, p. 197). The overarching theme of these accounts is that there are two independent systems involved in category learning. One system, variously termed explicit, declarative, verbal, or rule-based (RB), relies on working memory, hypothesis testing, and the application of

170


simple rules (Ashby & Maddox, 2005; Minda & Miles, 2010). The other, described as implicit, procedural, nonverbal, or similarity-based, does not involve working memory or attention and learns associations between (motor) responses and category labels (Ashby & Maddox, 2005; Minda & Miles, 2010). The product of learning from the latter system is often assumed to be unavailable to awareness or impossible to verbalize (Ashby, Alfonso-Reese, Turken, & Waldron, 1998; Knowlton, Squire, & Gluck, 1994; Minda & Miles, 2010). Versions of these dual-system accounts abound, and a variety of different tasks have been used in an effort to identify the roles played by “explicit” and “implicit” systems. We focus on two classes of category learning tasks— probabilistic and deterministic—that differ (as the terms suggest) in the relation between the attributes (cues) of a perceptual stimulus and category membership. From the probabilistic domain, we review studies of the popular “weather prediction task” (Knowlton, Mangels, & Squire, 1996), and from the deterministic domain, we examine the “RB” and “informationintegration (II)” tasks popularized by Ashby, Maddox, and colleagues (e.g., Maddox & Ashby, 2004). We have omitted discussion of other tasks such as artificial grammar learning and dot-pattern classification, but many of the arguments of interpretation that we raise apply equally to these tasks (see, e.g., Nosofsky & Zaki, 1998; Palmeri & Flanery, 2002; Shanks & St. John, 1994).

2. Review and Critique of the Evidence I: Probabilistic Category Learning Probabilistic category learning (PCL) involves the presentation of multifeatured stimuli that participants must learn to classify into one of two categories on the basis of trial-by-trial feedback. The feedback is probabilistic, thus discouraging participants from memorizing the outcome associated with a previous encounter with a stimulus. To achieve optimal performance, a participant must integrate information over many trials to establish the appropriate stimulus–response associations. The most commonly used version of the task has a cover story in which participants learn to predict the weather (rainy or fine) on the basis of distinct geometric patterns presented on four individual cards (e.g., a circle card, a square card, etc.). These four cards are presented in all possible combinations (excluding the pattern in which all cards are present on a single trial). Figure 1 shows the basic task along with a brief description of its properties. The PCL task was developed by Knowlton and colleagues (Knowlton et al., 1996, 1994) to provide a human analogue of the gradually acquired, habit learning tasks used in nonhuman animal studies (e.g., maze learning; Poldrack & Packard, 2003). The claim that individuals learn the PCL task

171


Card 1

Card 2

Card 3

Card 4

Cues

0.2

0.4

0.6

Outcome

0.8

Probability of rain given a specific card

Rain or fine

Figure 1 The “weather prediction task”—the most commonly used variant of a probabilistic category learning (PCL) task. The task requires participants to predict the weather on the basis of different cue configurations. Each card has a different predictive validity. Two cards are “strong” predictors of one or other outcome (e.g., the triangle card predicts “rain” with 0.8 probability, and the square card predicts “sun” with 0.8—note in the diagram only validities with respect to rain are indicated). The other two cards are “weak” predictors, with either 0.4 or 0.6 predictive validity. In the standard “feedback” version of the task, participants make predictions and are given trial-by-trial corrective feedback. In the alternative “observation” version, cues and the outcome appear together on each trial and participants are asked to learn (memorize) the cue–outcome pairings.

“without being aware of the information they have acquired” (Knowlton et al., 1996, p. 1400; see also Gluck, Shohamy, & Myers, 2002) fits neatly with the idea that the task is learned by a “primitive” system common to human and nonhuman animals. Studies of the latter indicate that such habit learning tasks rely strongly on the proper function of the dorsal striatum (Poldrack & Packard, 2003); thus, in the first studies that used the PCL task, it was hypothesized that this area would be important for learning the task. Since then, a good deal of evidence has been collected and interpreted as providing support for this hypothesis. The evidence comes from three principal domains: neuropsychological dissociations, behavioral dissociations, and neuroimaging.

2.1. Neuropsychological Dissociations A standard practice in cognitive neuropsychology is to identify tasks that can be learned by some patient groups but not by others and then to use these “dissociations” as evidence for the existence of functionally independent systems. The “holy grail” of such investigations is to find a double dissociation in which patient A can do task 1 but not task 2 and patient B can do task 2 but not 1. Knowlton et al. (1996) claimed to have found a double dissociation in comparing the performance of amnesics and Parkinson’s disease (PD) sufferers on the PCL task. Amnesics showed unimpaired learning of the task, relative to matched controls, but a clear deficit on a

172


questionnaire that assessed declarative memory for features of the task. In direct contrast, PD patients were impaired at learning the task but showed no deficit on the questionnaire. Amnesics have damaged medial temporal lobe (MTL) structures whereas PD sufferers have damage to the dorsal striatum. Thus, the double dissociation reported by Knowlton et al. (1996) supports the notion that these brain structures comprise functionally distinct declarative and procedural memory systems, respectively. This conclusion was supported in another study with PD patients that demonstrated selective impairment contingent on the type of PCL task used. Shohamy, Myers, Grossman, Sage, and Gluck (2004) gave PD patients either the standard version of the PCL task in which learning occurs via trial-by-trial feedback (see Figure 1) or a version in which cues and outcomes are presented simultaneously, and no response is required (an observational or paired-associate version). The PD patients showed the familiar deficit on the feedback version but were relatively unimpaired, relative to controls, in a test following the observational version. This dissociation is consistent with the notion that the dorsal striatum region is necessary for feedback learning (and thus PD patients are impaired) but that observational or paired-associate learning relies on the MTL (which remains intact in PD patients). Shohamy et al. (2004) argued that the selective impairment of PD patients on tasks that rely on learning from feedback is predicted by the underlying neurobiology of the basal ganglia and dopamine system. Specifically, dopamine is released as a result of a prediction-error signal which in turn drives reinforcement learning. When this system is disrupted (as it is in PD), sufferers are unable to learn from feedback. Taken together, proponents argue that these and similar data (e.g., Reber, Knowlton, & Squire, 1996) provide good evidence for deficits in PCL for patients with dorso-striatal damage but relatively intact performance in amnesics (Poldrack & Foerde, 2008).

2.2. Reevaluating the Neuropsychological Evidence One of the key pieces of evidence in the PCL literature is the claimed double dissociation between PD patients and amnesics reported by Knowlton et al. (1996). Given the elevated status of this study (e.g., Knowlton, 1999; Poldrack & Foerde, 2008; Poldrack & Packard, 2003), it is important to scrutinize whether the conclusions are warranted by the data. On first inspection, the fact that amnesic patients learned the task but were impaired in the declarative memory test whereas PD patients showed the opposite pattern appears to provide good evidence for the involvement of distinct independent systems. But, a closer inspection of the data suggests a more complex situation.


173

Although the dissociation pattern held for the early part of training (first 50 trials), Knowlton et al. reported that during an additional training period, PD patients gradually improved on the PCL task to a level of performance that was the same as the amnesics. However, both patient groups displayed lower average accuracy than controls over the last 50 trials (PD, 61.9%; amnesics, 59.2%; controls, 66.1%). This pattern of data is troubling for the multiple-systems view for two reasons. First, as Palmeri and Flanery (2002) point out, if the PCL task is a “processpure” procedural task, then amnesics should be unimpaired relative to controls throughout the task. Knowlton et al.’s explanation for the persistent deficit was that by the end of training, PD patients and controls may have been able to access information from declarative memory (i.e., explicit knowledge of cue-outcome associations) that remained unavailable to amnesics. Such an account becomes very difficult to falsify, however, because any unexpected deficits can simply be attributed to the contributions (or lack thereof) from an alternative memory system (Palmeri & Flanery, 2002). An account in which performance on a task is mediated by multiple systems also undermines a key criterion for the establishment of an independent system— that a system serves a functionally independent role (Sherry & Schacter, 1987). If a PCL task can be learnt equally well by either a procedural or a declarative system, then it suggests that either the task is a nondiagnostic “tool” for identifying the correlates of particular systems, or that the system(s), ultimately, is (are) serving the same role (i.e., learning the cue-outcome contingencies in the task environment; Speekenbrink et al., 2008). Speekenbrink et al. (2008) argued that the difference in levels of performance between amnesics and controls on later trials of the PCL task (e.g., Knowlton et al., 1996) can be captured by differences in the learning rate in a single-system model. Specifically, the learning rate parameter (which in their associative model determined the size of the change in the direction that minimized the prediction error) was lower for amnesics than controls. Moreover, Speekenbrink et al. found in their own study that amnesics and controls had similar high levels of explicit knowledge of cue-outcome contingencies at the end of training. Thus, the learning rate explanation is to be preferred over the suggestion that the later trial advantage for controls over amnesics is due to only the former having access to declarative knowledge (see also Kinder & Shanks, 2003). The second, related, reason these data are troubling is the demonstration that PD patients can learn PCL tasks if given enough time.1 Wilkinson, Lagnado, Quallo, and Jahanshahi (2008) drew the same conclusion in a reexamination of the Shohamy et al. (2004) study. 1

PD patients also appear to be less impaired on PCL from the outset if they are not on medication at the time of testing—a situation which is atypical of the majority of studies ( Jahanshahi, Wilkinson, Gahir, Dharminda, & Lagnado, 2010).

174


Wilkinson et al. highlighted a number of methodological concerns with the Shohamy et al. design and demonstrated in an experiment that dealt with these concerns that the selective impairment disappeared. Wilkinson et al. suggested that their failure to replicate Shohamy et al. (2004) might have been due to the absence of a response deadline. One of the characteristics of PD is slowness, both of movement and information processing (Ransmayr et al., 1990), and thus the selective impairment Shohamy et al. found on the feedback version of the PCL relative to the observation version may have been due to the timed element and not the feedback component. When the response deadline was removed from both versions, PD patients learned both to equivalent levels. This result undermines the claim that an intact dorsal striatum is crucial for learning the feedback version of the task. Rather, differences between PD patients and controls reflect variations in response sensitivity/selection not learning per se (cf. Kinder & Shanks, 2003; Nosofsky & Zaki, 1998). A reevaluation of some pertinent neuropsychological literature reveals that adopting a one-to-one mapping of deficits in PCL with functionally discrete learning and memory systems in the brain is naı¨ve (cf. Palmeri & Flanery, 2002). Purportedly, diagnostic dissociations in which particular patient groups are claimed to be able or unable to acquire different types of knowledge during PCL have been challenged and alternative explanations offered. The most enduring alternative explanation is that all participants, whether brain function is compromised or not, adopt an explicit RB approach to learning the task (Speekenbrink et al., 2008; Speekenbrink, Lagnado, Wilkinson, Jahanshahi, & Shanks, 2010). More evidence for this account can be found from studies examining behavioral dissociations in normal participants.

2.3. Behavioral Dissociations There are surprisingly few studies favoring the multiple-systems view of PCL that rely purely on the presence of behavioral dissociations in humans with unimpaired brain function. An exception to the pattern is a study by Foerde, Poldrack, and Knowlton (2007) which examined how PCL was affected in normal participants by the introduction of a secondary memory load task during learning. A straightforward prediction of the multiplesystem view is that if PCL is primarily mediated by a procedural learning system, then PCL should be relatively unaffected by the introduction of a concurrent memory task. Foerde et al. tested this prediction by giving participants a PCL task to perform on its own or in addition to a secondary tone-counting task. Probe blocks were included at two points in the experiment during which participants who learned under dual-task conditions were released from the secondary task for a given number of trials. Performance was impaired


175

under dual-task conditions, but, when only probe block trials were considered, there was no difference in the accuracy of participants who had learned under single- or dual-task conditions. Foerde et al. (2007) concluded that the secondary task impacted on the expression of learning (i.e., performance) in the PCL task but not on the learning of cue-outcome associations. Moreover, participants who had learned under dual-task conditions had poorer declarative knowledge of cue-outcome associations (in a postexperiment questionnaire) than those who had learned under single-task conditions. This pattern of results was taken as supporting the view that declarative knowledge is acquired by the MTL, which is impacted by the additional secondary load, but that the ability to learn the task is mediated by the dorsal striatum which is unaffected by load.

2.4. Reconsidering Behavioral Dissociations There are two reasons to treat the conclusion drawn by Foerde et al. (2007) with caution. First, the design they used compared performance on intermixed “dual” and “single” task blocks, but on all blocks corrective feedback was given. Thus, it is not clear whether participants learned at a faster rate in the single-task blocks, or that they were “free” to demonstrate the knowledge acquired in the previous dual-task phase. Second, Newell, Lagnado, and Shanks (2007) demonstrated that dual-task conditions led to a clear impairment of the learning of cue-outcome associations in PCL in an experiment in which the number of training trials and no-feedback test trials was matched across load and no-load groups. Their data provided no evidence that a release from the dual task during test facilitates the expression of knowledge acquired during training.2 In a second experiment, Newell et al. found that participants in feedback and observation versions of the PCL did not differ in their explicit knowledge of the task or the strategies they used to solve it, a result which is unexpected if separate systems mediate performance in the two versions (see also Price, 2009). A clear prediction of a purely explicit learning account of PCL is that participants will have accurate insight into what they have learnt in the task. Lagnado, Newell, Kahan, and Shanks (2006) examined this prediction using an innovative approach in which participants were probed throughout training trials for the explicit basis of each prediction. On each trial, participants were asked to rate how much they had relied on each cue in making a prediction. The “explicit” cue ratings were then compared with the “implicit” weights derived from running “rolling” regressions (a series 2

The dual tasks used by Newell et al. (2007) and Foerde et al. (2007) differed considerably—the former used a “numerical Stroop task” whereas the latter used a tone-counting task. Further research is required to establish why these two tasks might have different effects on PCL (see Heffernan, 2009).

176


of regressions from predictions to cues across a moving window of consecutive trials). Figure 2 shows comparison plots of these two dependent measures averaged across participants. In both panels, the values are collapsed across the two “strong cues” and the two “weak cues” (see Figure 1 for an

Strong cues

4 Mean importance ratings

Weak cues

3

2

1 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 Blocks 8

Strong implicit Weak implicit

Absolute regression weight

7 6 5 4 3 2 1 0 6

7

8

9 10 11 12 13 14 15 16 17 18 19 20 Blocks

Figure 2 Top panel: Mean importance ratings averaged across the two strong and the two weak cues in a feedback version of the weather prediction task. Note that values on the y-axis represent how much participants said they relied on each card, with 4 ¼ “Greatly,” 3 ¼ “Moderately,” 2 ¼ “Slightly,” 1 ¼ “Not at all.” Bottom panel: Mean implicit regression weights (absolute values) for strong and weak cards derived from conducting rolling regressions across a moving window of trials. Note in both figures, blocks are 10 trials (adapted from Figure 11 of Lagnado, Newell, Kahan, & Shanks, 2006).


177

explanation of “strong” and “weak”). The top panel shows the explicit cue reliance ratings and the bottom panel, the weights derived from the rolling regressions. The take-home message from these figures is that participants clearly distinguish between strong and weak cues on both the implicit and explicit measures of cue reliance. This ability occurs fairly early in the trials and is maintained, or increases across training. This pattern flies in the face of received wisdom about participants being unaware of what they learn in the PCL task (e.g., Gluck et al., 2002; Knowlton et al., 1996). (An additional experiment demonstrated that overall accuracy in the task was unaffected by the inclusion of the online ratings; Lagnado et al., 2006, Experiment 3.) The advantage of the online measures of cue use is that they interrogate the participant about performance at the time of the prediction rather than at the end of a training episode. Posttraining questionnaires that are typically used in studies of PCL often lead to underestimates of explicit knowledge because their retrospective nature leads to distortions of judgment (Lovibond & Shanks, 2002; Shanks & St. John, 1994). Lagnado et al. (2006) also reported strong positive correlations between individuals’ cue reliance ratings and implicit regression weights. The overall pattern strongly suggests that people have access to the internal states that drive their behavior in PCL and that this access drives both online predictions and verbal reports. This leaves little room for the contribution of an implicit category learning system.

2.5. Neuroimaging Neuroimaging methods such as functional magnetic resonance imaging (fMRI) have been used increasingly in recent years to examine the neural activity of individuals engaged in PCL tasks (e.g., Aron et al., 2004; Foerde, Knowlton, & Poldrack, 2006; Poldrack et al., 2001). These studies reveal that the cortico-striatal circuits and midbrain dopaminergic regions highlighted by the neuropsychological data appear to be actively involved during learning PCL tasks (Poldrack & Foerde, 2008). However, the overall picture is rather more nuanced. Several investigations indicate that dorso-striatal regions and the MTL are activated when participants are learning, but that the relative involvement of the regions is modulated by a variety of factors. For example, Poldrack et al. (2001) demonstrated that early in learning the MTL was active and the dorso-striatal region (caudate nucleus) was inactive, but as learning progressed, this pattern reversed with the caudate becoming active as the MTL deactivated. In a similar vein, Foerde et al. (2006) showed that when participants learned PCL tasks under single- or dual-task conditions, the striatal learning

178


mechanisms were engaged equally. However, in a subsequent test phase, in which no feedback was provided, Foerde et al. found that when participants classified items from the PCL task initially learned under dual-task conditions, accuracy was correlated with activity in the striatum. In contrast, accuracy for items learned under single-task conditions was correlated with activity in the MTL (right hippocampus). These results were interpreted as indicating competition between declarative and procedural memory/learning systems. When factors favor the adoption of explicit, declarative learning (such as early in training, or when attention to the task is undivided), then the MTL region dominates. When declarative processes are compromised by additional tasks, or rendered less important through increased experience with the task, the striatal system takes precedence (Foerde et al., 2006; Poldrack & Foerde, 2008).

2.6. Reimagining Neuroimaging Neuroimaging studies challenging the multiple-systems interpretation of PCL are rare, if not nonexistent. In large part, this reflects the interpretative role played by neuroimaging data in the category learning debate. Sections 5 and 6 discuss this role in more depth; here we note simply that studies showing that the relative involvement of MTL and the striatum can be modulated by task demands (e.g., Foerde et al., 2006) do not by themselves constitute incontrovertible evidence for the operation of dissociable systems (cf. Coltheart, 2006; Page, 2006; Sherry & Schacter, 1987). The preceding sections raise several concerns about interpretations of PCL performance and the neuroimaging data should be viewed with these concerns in mind.

2.7. Section Summary PCL tasks and, in particular, the weather prediction task have been used extensively in recent years to advance a multiple-systems interpretation of category learning. In spite of apparently compelling evidence, this interpretation can be challenged in each of the areas reviewed. In the neuropsychological domain, apparent dissociations between different patient groups dissolve once learning is examined across all trials rather than a subset. Behavioral experiments find that participants have clear insight into their performance and that the effect of increasing cognitive load is consistent with a single rather than multiple-system account. Finally, neuroimaging data while suggestive is by no means conclusive. We now turn to deterministic tasks and argue that similar concerns can be raised there.


179

3. Review and Critique of the Evidence II: Deterministic Category Learning In deterministic category, learning participants learn to assign novel stimuli to discrete categories, but unlike probabilistic tasks, feedback is deterministic (a category “A” stimulus is always in category A). As noted in Section 1.1, our focus is on two classes of deterministic tasks that have been examined extensively in relation to the COVIS (COmpetition between Verbal and Implicit Systems; Ashby et al., 1998) model of category learning: RB and II. In the simplest case, a set of multidimensional stimuli conform to an RB structure if they can be classified on the basis of a single, easily verbalized rule, such as “If the value of X (e.g., height) is greater than c” or “If the value of X is greater than c1 and the value of Y is less than c2.” A categorization problem is II if no easily verbalizable rule allows perfect classification. A familiar example would be family-resemblance categories, such as those used by Reed (1972), or formed by the faces of members of the Hatfield and McCoy clans. A set of canonical RB/II category structures is shown schematically in Figure 3. Here, the filled squares and unfilled circles correspond to stimuli from two different categories. In Figure 3A, the categories are defined by the level of one relevant dimension (spatial frequency of a perceptual stimulus). In Figure 3B, the categories are defined with respect to the levels of two relevant dimensions (spatial frequency and orientation). An example of a typical perceptual stimulus (a Gabor Patch) is shown in Figure 3C. These two classes of tasks are said to engage qualitatively different category learning systems. The RB tasks in which the optimal solution can be described in simple, verbalizable rules recruits the verbal or explicit system which is dependent on working memory and executive attention (for storing and testing rules, respectively). The II tasks for which optimal solutions are difficult or impossible to verbalize (Ashby et al., 1998) do not depend on working memory and attention, but rather learn categories by learning the procedure for generating a response (i.e., the assignment of a category label). As with PCL tasks, the evidence for this differential involvement of distinct systems comes from three main areas: neuropsychology, behavioral dissociations, and neuroimaging.

3.1. Neuropsychological Dissociations Much of the neuropsychological evidence for the involvement of the separable category learning systems identified in COVIS comes from studies of PD patients and so we restrict our review to those studies (see Filoteo &

180


A

B Information integration

0.6

0.6

0.4

0.4 Orientation

Orientation

Rule based

0.2

0

0 2

–0.2

0.2

4

6

8

Frequency

10

2 –0.2

4

6

8

10

Frequency

C

Figure 3 (A): Distribution of stimuli in a rule-based (RB) category structure. (B): Distribution of stimuli in an information-integration (II) category structure. (C) An example of a Gabor patch stimulus. In a typical experiment, a single stimulus is shown on the screen and participants have to categorize it as an “A” or “B” and then receive corrective feedback.

Maddox, 2007; Price, Filoteo, & Maddox, 2009 for more comprehensive reviews). With regard to RB tasks, the degree of impairment shown by PD patients appears to be dependent on the number of irrelevant dimensions present in the to-be-categorized stimuli (Price et al., 2009). For example, Filoteo, Maddox, Ing, Zizak, and Song (2005) gave PD patients and agematched controls a task in which four binary dimensions of a stimulus could vary trial-to-trial and then manipulated how many of the dimensions were irrelevant for classification. When none or only one of the dimensions was irrelevant, PD patients and controls performed at similar levels. However, when two or three dimensions were irrelevant, PD patients were impaired relative to controls. Filoteo, Maddox, Ing, et al. concluded that PD patients have deficits in selective attention leading them to be unable to ignore the irrelevant stimulus attributes (see also Ashby, Noble, Filoteo, Waldron, & Ell, 2003; Channon, Jones, & Stephenson, 1993).


181

In II tasks, in which typically more of the presented stimulus dimensions are relevant to the task, one might expect PD patients to show less of a deficit. The data seem to partially support this prediction. Ashby, Ell, and Waldron (2003) demonstrated that PD patients performed at a similar level to healthy controls in II tasks that involved the integration of three out of four binary valued dimensions. Filoteo, Maddox, Salmon, and Song (2005) contrasted II categories in which the decision bound was linear with those in which it was nonlinear. Figure 3B gives an example of a linearly separable II task; a nonlinearly separable task is one in which the optimal boundary between the categories is described by an alternative function (quadratic in the case of the Filoteo, Maddox, Salmon, et al., 2005 study). Interestingly, PD patients showed no deficit in learning the linearly separable II task (replicating Ashby, Ell, et al., 2003) but were impaired in the more difficult nonlinear task. Filoteo and Maddox (2007) speculate that the deficit is specific to nonlinear tasks because learning nonlinear bounds requires “a greater degree of representation” (p. 16) than learning linear bounds, and this “place[s] more demands on the striatum” (p. 16)—an area known to be damaged in PD patients.

3.2. Reevaluating the Neuropsychological Evidence In line with our reevaluation of the neuropsychological evidence relating to PCL, there are several reasons to be cautious about the one-to-one mapping between deficits and the operation of discrete systems when it comes to deterministic category learning. Price et al. (2009) note that drawing strong conclusions about the ability of PD patients on RB and II tasks is severely complicated because of the role played by medication (see also Footnote 1). Price et al. break down the problems faced by PD patients into four issues: rule generation, rule maintenance, rule shifting, and rule selection. They conclude that the neurochemical changes associated with PD may disrupt rule shifting (ceasing with an unsuccessful rule after negative feedback) and rule selection (finding an appropriate new rule), but that typical treatment to remedy these deficits (e.g., L-Dopa) can cause detriments in rule generation and rule maintenance. Thus, the interpretation of deficits in category learning depends crucially on whether PD patients are tested “on” or “off” medication (cf. Filoteo & Maddox, 2007). The PD patients in the Filoteo, Maddox, Ing, et al. (2005) study of RB task learning were all on dopaminergic medication at the time of testing. Thus, the impairments seen on the tasks in which two or three dimensions were irrelevant could have been due to medication causing increased distractability in the patients thereby reducing their ability to ignore irrelevant information and maintain a correct rule (Price et al., 2009). Such a

182


reinterpretation is important to consider because it suggests that it is not the etiology of PD that leads to impairments, but the nature of the treatment. A similarly complex picture emerges when one reconsiders the evidence for performance of PD patients in II tasks. Price (2005) reported that PD patients on medication were impaired on a linearly separable II task, contrary to the findings of Filoteo, Maddox, Salmon, et al. (2005) who only reported impairments on nonlinearly separable tasks. However, the tasks used in the two studies were very different. The task in Price involved learning the optimal combination of five discrete (presence/absence) cues, one of which was irrelevant to classification. In contrast, Filoteo, Maddox, Salmon, et al. used a task in which classification was based on correctly integrating the orientation and length of presented lines—thus, both dimensions were always relevant. The discrepant results could, therefore, be due to medication impacting on the ability to ignore the irrelevant stimulus dimension in the Price study—something that was not necessary in the study of Filoteo, Maddox, Salmon, et al. Other studies showing inconsistent patterns of results also urge caution. For example, Schmitt-Eliassen, Ferstl, Wiesner, Deuschl, and Witt (2007) failed to replicate Filoteo, Maddox, Salmon, et al.’s (2005) finding of PD patient impairment in learning nonlinear II tasks (though differences in experimental procedures may have led to the discrepancy—see Filoteo & Maddox, 2007). Finally, Swainson et al. (2006) provide some highly diagnostic evidence from a study that contrasted performance of PD patients who were unmedicated or on mild medication. Participants completed two tasks—the “eight-pair task”—a concurrent discrimination task which involved learning from feedback but not selective attention to dimensions of compound stimuli, and the “five-dimensions” task which required compound discrimination and incorporated trial-by-trial feedback. PD patients— regardless of their medication regime—were unimpaired relative to controls on the eight-pair task; a result which Swainson et al. interpret as militating against the claim that the striatum is crucial for feedback learning per se. However, in the five-dimensions task, mild medicated but not unmedicated PD patients were impaired at learning to identify relevant aspects of a compound stimulus. A third group of severely medicated PD patients also performed very poorly on the five-dimensions task. These results converge with those reported above: PD patients on medication are impaired in identifying relevant aspects of multidimensional (or compound) stimuli; however unmedicated patients can learn such tasks (see also Footnote 1). Taken together, these reconsiderations of the neuropsychological literature suggest that drawing simplistic inferences from the etiology of PD (e.g., degeneration in dopamine containing cells) to observed deficits on category learning tasks is naı¨ve for two main reasons. First, differences across tasks that are often grouped together as generic RB and II tasks may show very


183

different patterns of performance (Price, 2005). Second, there is clear evidence that PD patients on and off medication behave very differently and in ways that are often at odds with proposed dissociations ( Jahanshahi, Wilkinson, Gahir, Dharminda, & Lagnado, 2010; Swainson et al., 2006).

3.3. Behavioral Dissociations A huge number of studies document functional dissociations between RB and II categorization tasks in normal, healthy participants (see Ashby & Maddox, 2005; Maddox & Ashby, 2004 for reviews). In a typical experiment, a variable is found to affect learning of either the RB or II structure but to have little or no effect on learning the alternative structure. We focus on four illustrative experiments: two showing the impact of variables on RB (but not II) task learning and two showing the opposite pattern. We acknowledge that this review is selective; however, we have chosen those studies highlighted by proponents as providing particular support for the COVIS model (e.g., Ashby et al., in press). The exclusive reliance of the explicit system on working memory and executive attention led Waldron and Ashby (2001) to predict that increasing cognitive load would have a detrimental effect on RB learning but not II learning. To test this prediction, participants were given a category learning task using geometric patterns which could vary on four binary dimensions (e.g., shape: circle or triangle). In the II task, three out of four of these dimensions were relevant to the classification rule, and in the RB task, only one dimension was relevant. Participants in the load conditions performed a numerical Stroop task concurrently with the category task. Participants performing the RB and the load task concurrently took longer (more trials) to reach criterion than those performing the RB task alone, but load had a negligible effect on II learning. Zeithamova and Maddox (2006) replicated this effect using the Gabor patch stimuli shown in Figure 3. More support for the claim that RB task learning selectively involves working memory comes from a study by Maddox, Ashby, Ing, and Pickering (2004). Maddox et al. manipulated the amount of time participants had to process the corrective feedback delivered after making a categorization decision. Participants were given either RB or II tasks interpolated with a memory scanning task—identifying a probed digit from a set of four briefly presented numbers. In an “immediate” condition, the memory scanning task was presented only 500 ms after category-corrective feedback had been given; in the delay condition, there was a 2500 ms delay. Maddox et al. hypothesized that learning of RB tasks would be detrimentally affected in the immediate condition because there would be no time to explicitly process and reflect on the category feedback. II tasks, in contrast, because they are learned by an implicit procedural system, would be unaffected by the availability of feedback processing time. The results

184


provided support for this hypothesis: the RB immediate group performed less accurately than RB delayed, but the II groups did not differ. The implicit system in COVIS learns categories by learning the procedure for generating a response (Ashby et al., 1998), thus manipulations which interfere with procedural learning should affect performance on II tasks but have little or no impact on RB tasks. Ashby, Ell, and Waldron (2003) conducted experiments in which forms of response-motor interference were introduced. After a period of initial learning of both RB and II tasks, Ashby, Ell, et al. (2003) switched the response buttons used for category assignment. That is, the button that had previously indicated a category A response now indicated a category B response and vice versa. Following the switch, accuracy in II tasks was affected more than accuracy in the RB tasks—a finding consistent with the predictions of COVIS. COVIS also makes clear predictions about performance on II tasks based on the underlying neurobiology of the implicit system (Ashby et al., 1998, in press; provide a formal model of this basis). Specifically, the model assumes that II learning is mediated via reinforcement at cortical-striatal synapses, with dopamine serving as the reinforcement signal. For learning to occur, the appropriate synapses need to be strengthened following a reward. This necessitates that some trace of recently active synapses is maintained in the system, but, because of the morphology of the dendritic spiny cells in the caudate nucleus, this maintenance of activity only lasts for a few seconds. Thus, if feedback is delayed, it will have an adverse effect on II learning but not RB learning because the explicit system can maintain rules in working memory during the delay. This prediction was supported in two studies (Maddox, Ashby, & Bohil, 2003; Maddox & Ing, 2005) which demonstrated that delays of more than 2.5 s had adverse effects on II learning but no effect on learning RB tasks— even when the tasks were matched for difficulty (i.e., number of dimensions relevant to categorization). These demonstrations, and many others, are taken by proponents of the multiple-systems view as conclusive evidence in favor of discrete separable systems underlying category learning. There is even some evidence for “double dissociations” whereby a single manipulation simultaneously enhances performance on RB tasks and impairs performance on II tasks (Maddox, Love, Glass, & Filoteo, 2008).

3.4. Reconsidering Behavioral Dissociations As with the neuropsychological data, there are several reasons to question the empirical evidence for separable systems underlying II and RB learning. Taking the effects on RB learning first, the claim that RB tasks are selectively affected by the addition of a cognitive load is controversial. In an illuminating discussion of the Waldron and Ashby (2001) data, Nosofsky


185

and Kruschke (2002) demonstrated that ALCOVE (a “single”-system model) could naturally predict the behavioral pattern observed by Waldron and Ashby by suggesting that the cognitive load impaired participants’ ability to attend selectively to relevant stimulus dimensions (see Section 5.4 for further discussion of this interpretation). More recently, further doubt has been cast on the interpretation of Waldron and Ashby via a demonstration that the particular learning criterion they adopted (eight trials consecutively correct) is an unreliable measure of II learning for the stimuli that they used (Tharp & Pickering, 2009). Zeithamova and Maddox’s (2006) replication of the selective effect of load on RB tasks has also been challenged. Newell, Dunn, and Kalish (2010) failed to replicate the selective effect in three experiments using a variety of concurrent tasks. They concluded that Zeithamova and Maddox’s original interpretation had been confounded by the inclusion of nonlearners in the analysis (i.e., those participants who had neither learned the category task, nor performed the concurrent task adequately). Once these participants were removed, all evidence for a dissociative effect of load on RB and II tasks disappeared (see Section 4.4 for more discussion of this study). In a similar vein, Stanton and Nosofsky (2007) provide an alternative explanation of Maddox et al.’s (2004) demonstration of the selective effect of a reduction in feedback processing time on RB tasks. Stanton and Nosofsky tested the hypothesis that the dissociation was due to lowered perceptual discriminability of stimuli in the RB structure relative to the II structure. (Perceptual discriminability refers to the distance of items from the category boundary—see Figure 3.) In two experiments, they demonstrated that the interpolated memory scanning task had no effect on the RB task when the RB stimuli were easy-to-discriminate, and that II learning was affected by the memory scanning task when hard-todiscriminate II stimuli were used. This reversal of the “dissociation” is clearly inconsistent with the prediction that the tasks are learned by separate cognitive systems. Nosofsky, Stanton, and Zaki (2005) offer a similar “category complexity” reinterpretation of the Ashby et al. (2003) demonstration that only II tasks are affected by a button-switching manipulation. Nosofsky et al. argued that because the II category structures used by Ashby et al. (2003) were more difficult to learn than the RB structures, the selective interference observed for the II task might simply reflect the fact that more difficult tasks are more susceptible to interference than simpler tasks. By manipulating the difficulty of tasks independently of their RB and II status, Nosofsky et al. demonstrated that complex RB tasks were affected more than simple II tasks by the buttonswitching manipulation. Clearly, such a dissociation is inconsistent with the predictions of COVIS and lends support to the idea that task difficulty rather than the recruitment of different systems is the primary mediator of the observed effects.

186


Such a reinterpretation of the selective effect of delayed feedback on II tasks (e.g., Maddox & Ing, 2005) awaits, but there are grounds to suspect that similar explanations might suffice. For example, in the Maddox and Ing study, although the RB and II tasks were matched for difficulty in the sense that the same number of dimensions was relevant for categorization, the II stimuli were less perceptually discriminable than the RB stimuli. When a delay is imposed between prediction and feedback, the feedback is likely to be combined with an impoverished visual representation of the stimulus (cf. Stanton & Nosofsky, 2007). This process will be severely impaired in cases where the initial difference between categories is low (i.e., II stimuli in the Maddox and Ing experiment). Moreover, Maddox and Ing filled the delay between stimulus presentation and feedback with another Gabor patch stimulus, thus creating obvious potential for confusing the actual stimulus to which the feedback should be attributed with the “mask” presented during the delay period. These speculations await further experimental examination but the increasing number of studies reporting findings at odds with the multiplesystems view implies that a more general reevaluation of the entire hypothesis (of the kind we offer here) is long overdue.

3.5. Neuroimaging Studies examining the neural correlates of the specific RB and II tasks of the kind shown in Figure 3 are relatively scarce. A relevant study was conducted by Nomura et al. (2007; see also Nomura & Reber, 2008), who scanned participants (using fMRI) while they learned to categorize stimuli generated from either II or RB structures. The two groups performed identically in terms of accuracy, but Nomura et al. identified some differences in the patterns of brain activation. Successful categorization in the RB task was associated with activation in the hippocampus, anterior cingulate cortex (ACC) and the medial-frontal gyrus, whereas for II classification, correct answers were associated with activation in the head and tail of the caudate nucleus. Nomura and Reber (2008) explored these patterns of activation further by combining imaging analysis with computational modeling. They defined optimal RB and II models for the tasks and examined participants’ data to find sequences of trials on which clear use of RB and II strategies were evident. On those runs where RB behavior was most clearly shown, the right prefrontal cortex (PFC) showed increased activity compared to the best II runs. In contrast, during clear episodes of II strategy use, the right occipital cortical area appeared more active. Nomura and Reber (2008) interpreted this differential activation as implying a role for working memory (associated with PFC activation) in RB tasks, and the representation of category knowledge acquired by II in the occipital cortex.


187

3.6. Reimagining Neuroimaging Although the patterns of activation found in the Nomura et al. (2007; see also Nomura & Reber 2008) study are broadly consistent with the COVIS model they do present some challenges. For example, the original formulation of COVIS (Ashby et al., 1998) did not include the hippocampus as part of the RB learning system, instead focusing on the ACC, PFC, and the head of the caudate nucleus. (Note that this last area was more active in the II tasks than the RB tasks in the Nomura et al. data.) More recent expositions do include reference to the hippocampus (e.g., Ashby et al., in press), but seem to suggest that it comprises another system separate to the rule and procedural ones described by COVIS (e.g., Maddox et al., 2008). Moreover, one of the clear messages from the Nomura et al. (2007) data was the high degree of overlap in activation in II and RB tasks. Even when targeted, Region of Interest (ROI) analyses were conducted commonality was observed; but rather than interpreting such activity as the operation of a common system, it was taken to reflect “the competition between two simultaneously active categorization systems” (p. 39). As noted in the discussion of the PCL tasks, invoking notions like the simultaneous operation of systems makes the multiple-systems view very difficult to falsify (cf. Palmeri & Flanery, 2002). There also appears to be some confusion about the extent to which imaging results from other tasks can be used to support the multiple-systems hypothesis, in general, and the COVIS model, in particular. Ashby and Ennis (2006) discuss imaging data from PCL tasks (like the weather prediction task) as indicating an MTL-based system (rules) and a caudate-based system (implicit) being recruited at different stages of learning the task. However, earlier in the same chapter, they urge caution about drawing “strong inferences from data collected with [the weather prediction task] because near optimal performance can be achieved by a variety of different strategies (e.g., information-integration, rule-based, explicit memorization)” (p. 11). It is as if the neuroimaging data are somehow privileged in illuminating the involvement of separable systems, even when the behavioral data are entirely equivocal on this issue.3 Finally, the combination of neuroimaging and computational modeling adopted by Nomura and Reber (2008) is commendable but important limitations on the kind of modeling undertaken preclude strong conclusions from being drawn. We return to this issue in more detail in Section 5.

3

Note that the weather prediction task, described in Section 2, is technically a probabilistic II task because the optimal strategy involves integration of information from multiple cues.

188


3.7. Section Summary Our review and critique of deterministic tasks makes it clear that the wealth of evidence interpreted as consistent with multiple systems needs to be treated with caution. Evidence from neuropsychological studies is fraught with difficulties because of the role played by medication, and by variations across tasks and experiments. Key studies that demonstrate “signature” behavioral dissociations (e.g., the effect of increased cognitive load, the effect of procedural interference) have been challenged effectively and alternative explanations proposed. Finally, the to-date limited neuroimaging data highlights a good deal of commonality in activation during tasks that are supposedly supported by anatomically distinct systems.

4. Reexamining Some Fundamental Assumptions As we discussed in the previous sections, much of the evidence that has been used to support multiple-system interpretations of category learning has been based extensively on dissociations in behavioral, neuropsychological, and neuroimaging domains. Our aim in this section is to review the logical status of dissociations and to show that they provide a much weaker form of evidence than has often been supposed. As we have argued elsewhere, while it is possible, in principle, to distinguish single and multiple-system accounts,4 the appropriate logic does not derive from dissociations but rather from state-trace analysis (Bamber, 1979; Newell & Dunn, 2008). The present section is divided into three main parts. In the first part, we present the logic of state-trace analysis and use it to show that dissociations are neither necessary nor sufficient to reject a single-system (or single-process) account of the data. In the second part, we discuss potential difficulties associated with the interpretation of data that are apparently inconsistent with a single-system account. In the final part, we address conceptual issues concerning the nature of the intervening constructs that determine the observed patterns of data. In particular, we examine the question of whether the data alone are sufficient to assert the existence of separate “systems” or “processes” or other theoretical constructs.

4.1. State-Trace Analysis State-trace analysis was originally proposed by Bamber (1979). It is essentially a mathematical analysis of the consequences of manipulating two or more independent variables on two (or more) dependent variables when 4

Although the necessary contrast can be couched in terms of one or two “systems,” we do not commit ourselves to that interpretation and argue later for an alternative conceptualization.


189

these effects are mediated either by one, two, or more intervening latent variables. It can easily be shown that qualitatively different outcomes emerge depending upon the number of such intervening variables. For accessible presentations of this logic, in addition to the original article by Bamber, the reader is directed to papers by Loftus, Dillon, and Oberg (2004), Newell and Dunn (2008), and Heathcote, Brown, and Prince (2010). Newell and Dunn (2008) explicitly applied the logic of state-trace analysis to experiments in category learning. These experiments generally contrast two different kinds of dependent variables. One dependent variable is usually interpreted as primarily reflecting a procedural or implicit learning system and is operationalized in terms of performance on an II category learning task or a PCL task, like the weather prediction task. The other dependent variable is usually interpreted as primarily reflecting a RB or explicit learning system and is operationalized in terms of performance on a RB category learning task or explicit knowledge of the contingencies of the weather prediction task. Typically, two or more independent variables are also manipulated; one of these is frequently the number of learning trials, the others consisting of factors proposed to differentially affect one or other learning system. An illustrative study is that conducted by Maddox et al. (2003), who examined the effect of various delays between response and feedback on performance on RB and II tasks (discussed previously in Sections 3.3 and 3.4). According to COVIS, delay is critical to procedural learning which depends upon time-dependent strengthening of synaptic links in the tail of the caudate nucleus. In contrast, delay should have little or no effect on RB learning which utilizes the storage capacities of working memory. The theoretical structure proposed by the COVIS model is shown in Figure 4A. According to this model, number of trials affects learning by both the explicit and procedural systems while delay differentially affects learning in the procedural system. In addition, learning by the explicit system primarily determines performance on RB tasks while learning by the procedural system primarily determines performance on II tasks. An alternative theoretical structure, proposed by a “single-system” account, is shown in Figure 4B. According to this model, both number of trials and delay affect a common learning system which differentially affects performance on RB and II tasks. That is, performance on these tasks is considered to depend in different ways upon the same underlying degree of learning. That is, performance is assumed to increase with learning but the rate of this increase may change over time in different ways for the two tasks. Although these changes may be complex, one simple expression of this idea is that one task may be more difficult than the other and therefore increase at a slower rate. In terms of state-trace analysis, this corresponds to the assumption that performance on each task is a monotonically increasing function of a single intervening variable.

190


A No. of trials

Delay

Explicit system

RB task

Procedural system

II task

B No. of trials

RB task Learning system

Delay

Il task

Figure 4 Schematic diagrams of the effect of number of learning trials and delay between response and feedback on performance on rule-based (RB) and informationintegration (II) tasks. (A) According to COVIS. (B) According to a single-system model.

An important analytical tool for state-trace analysis is the state-trace plot. This is a scatter plot of the covariation of the two dependent variables across the different experimental conditions. The two models shown in Figure 4 have different consequences for the form of the state-trace plot. These forms may be classified as either one- or two-dimensional (Loftus et al., 2004; Newell & Dunn, 2008). A state-trace plot is one dimensional if the data points fall on a single monotonically increasing or decreasing curve, otherwise it is two dimensional. Importantly, the dimensionality of the state-trace plot cannot exceed the number of intervening latent variables. This means that while the multiple-system model shown in Figure 4A may lead to a two-dimensional state-trace plot, the “single-system” model shown in Figure 4B must always lead to a one-dimensional state-trace plot. Figure 5 shows examples of two state-trace plots. Figure 5A illustrates a two-dimensional state-trace and Figure 5B illustrates a one-dimensional state-trace. In both plots, each point corresponds to the mean level of performance on each of the RB and II tasks for one experimental condition. As these plots are based on the design used by Maddox et al. (2003), there are eight experimental conditions defined by the combination of trial block (1–4, although these are not separately identified in the Figure) and the nature of feedback (immediate vs. delayed). Figure 5A is two dimensional because the data points do not fall along a single monotonic curve. This is shown by the fact that there are pairs of points (i.e., experimental conditions) across which performance on both RB and II tasks increase and pairs of points across which performance on one task increases and the other decreases.

191


A

B 0.8

0.8 Immed.

Immed.

Delay

Delay

0.7

0.7

c

=

= b

0.6

0.6

a 0.5 0.7

a

c

0.8

0.9 RB

1

0.5 0.7

b

0.8

0.9

1

RB

Figure 5 Example state-trace plots. (A) State-trace plot consistent with COVIS and other multiple-system models. (B) State-trace plot consistent with the single-system model shown in Figure 4B. The numbers on each axis refer to the percentage of correct responses on the information-integration (II) and rule-based (RB) tasks with each data point corresponding to performance averaged across one of four blocks of learning trials.

Three such points are labeled a, b, and c in Figure 5A. Between the conditions corresponding to points a and b (as well as to points a and c), both RB performance and II performance increase—the dependent variables are positively associated across these conditions. Between the conditions corresponding to b and c, RB performance increases while II performance decreases—in this case, the dependent variables are negatively associated across these conditions. Similar positive and negative associations are repeated between all the other points in Figure 5A. Dunn and Kirsner (1988) called a combination of positive and negative associations a reversed association, and showed that it is inconsistent with a one-dimensional monotonically increasing state-trace plot which requires differences in each dependent variable always to be in the same direction. In contrast, Figure 5B is one dimensional because, in this case, all the data points fall on a monotonically increasing curve. While there are pairs of points across which performance on both tasks increase, there are no pairs of points across which performance on one increases and performance on the other decreases. This pattern is shown by the three points, again labeled a, b, and c. Between points a and b, RB performance increases but II performance remains the same—it neither increases nor decreases. Similarly, between points b and c, II performance increases while RB performance is constant. Finally, between points a and c, both RB and II performance increases. Thus, although there are pairs of points across which RB performance and II performance are positively associated, there are no pairs of

192


points across which they are negatively associated. These data therefore fail to produce a reversed association and hence are not inconsistent with a onedimensional monotonically increasing state-trace plot.

4.2. The Inferential Limits of Dissociations Newell and Dunn (2008) argued that state-trace analysis supersedes the logic of functional dissociation because dissociations are neither necessary nor sufficient to reject a “single-system” model of the type shown in Figure 4B. For this reason, dissociations cannot be used as evidence to support an alternative “multiple-systems” model of the type shown in Figure 4A. A dissociation is defined as the observation that a factor which affects performance of one task has no effect on a second task. In a state-trace plot, performance on one task is plotted against performance on the other. If there is a dissociation across two conditions, then this means that the data points corresponding to these conditions will be aligned either vertically, if the affected dependent variable is represented by the y-axis, or horizontally, if the affected dependent variable is represented by the x-axis. Dissociations are not necessary to reject a “single-system” model. This is shown by the state-trace plot in Figure 5A. Although this is two dimensional and thus inconsistent with a single-system account, it contains no dissociations as there are no pairs of data points aligned either vertically or horizontally. Dissociations are not sufficient to reject a “single-system” model. This is shown by the state-trace plot in Figure 5B. Although this is one dimensional and thus consistent with a single-system account, it contains several examples of dissociations—pairs of data points aligned either horizontally (e.g., points a and b) or vertically (e.g., points b and c). Very often, the conjunction of these two kinds of dissociation is referred to as a double dissociation and interpreted as providing the strongest evidence in favor of a multiplesystems account (Shallice, 1988). However, this example shows that a dissociation (even a double dissociation) is not sufficient to reject the “single-system” model. State-trace analysis supersedes the logic of dissociations because it specifies a pattern of data that is logically inconsistent with a single-system account. In contrast to single and double dissociations, a two-dimensional state-trace plot is both necessary and sufficient to reject a single-system account. State-trace analysis also implies a different set of statistical questions. A feature of dissociation logic is that it often depends upon showing that conditions do not differ on one or more dependent variable thereby running the risk that any conclusions that may be drawn may be a consequence of

193


one or more Type I errors. In contrast, in order to reject the “single-system” model, state-trace analysis requires that pairs of conditions should differ in systematic ways. Specifically, it is necessary to show that there is at least one pair of points that differ in the same direction on both dependent variables and another pair of points that differ in opposite directions on both dependent variables (Dunn & Kirsner, 1988; Newell & Dunn, 2008). The statistical analysis of state-trace plots is currently under active investigation (Heathcote et al., 2010; Newell et al., 2010). The point we wish to make here is that different conclusions may be drawn from the same data depending upon whether it is analyzed in terms of the logic of dissociation or of state-trace analysis. Figure 6 shows the state-trace plot of data from Experiment 1 of the study by Maddox et al. (2003) averaged over three levels of time interval.5 Two features are apparent in these data. First, as concluded by Maddox et al., and noted above, these data demonstrate a double dissociation—there are pairs of data points aligned, at least

0.8 Immed. Delay

II

0.7

0.6

0.5 0.7

0.8

0.9

1

RB

Figure 6 Observed state-trace plot averaged over delay interval (from Maddox, Ashby, & Bohil, 2003). The numbers on each axis refer to the percentage of correct responses on the information-integration (II) and rule-based (RB) tasks with each data point corresponding to performance averaged across an 80 trial learning block. Error bars correspond to standard errors of the mean.

5

These data were derived from Table 2 of the published paper. This table presents the means and standard errors for each condition which included three levels of time interval (2.5, 5, and 10 s). The data in Figure 6 have been averaged over this factor.

194


approximately, both vertically and horizontally. Second, the data appear to be consistent with a one-dimensional state-trace plot. In fact, the data points shown in Figure 5B correspond to the best-fitting monotonic curve passing through the observed data shown in Figure 6 (for details of how to fit such a curve, see Newell et al., 2010). Although a firm conclusion depends on formal statistical analysis, it is highly unlikely that the small differences between the observed data in Figure 6 and the best-fitting monotonically increasing data in Figure 5B are sufficient to reject a “single-system” account. These data therefore offer little or no support for a multiplesystems view.

4.3. A State-Trace Reanalysis of Behavioral and Other Dissociations We have proposed that arguments for or against multiple-system accounts of category learning should move beyond dissociation logic. As illustrated above, reanalysis of at least one prominent behavioral dissociation using state-trace analysis shows that a quite different conclusion may be drawn from it. This does not imply that all or any other dissociation is open to a similar reinterpretation. While it is beyond the scope of this chapter to examine each dissociation in turn, it is at least possible that some may also be shown not to meet the higher test offered by state-trace analysis. The view that the case for COVIS or other multiple-systems accounts of category learning has been well established is therefore based on an interpretation of the evidence that is currently only provisional. The arguments we have proposed above apply with equal force to dissociations in other domains. For example, in a review of the role of the basal ganglia in category learning, Ashby and Ennis (2006) have drawn attention to dissociations in category learning performance between different patient groups. They reviewed evidence that patients with Huntingdon’s disease, a degenerative disease that affects most of the basal ganglia, are impaired on both RB and II tasks compared to normal controls. In contrast, patients with PD, a degenerative disease that primarily affects the head of the caudate nucleus, are impaired on RB tasks but are relatively less impaired on II tasks (see Sections 3.1 and 3.2 for further discussion). This dissociation is consistent with the COVIS model that proposes that the explicit system that underlies learning in RB tasks is subserved by the head of the caudate nucleus while the implicit system that underlies learning in II tasks is subserved by the tail of the caudate nucleus. Since both patient groups are impaired on RB tasks, evidence for the critical dissociation depends on differences in performance on II tasks. In one study, Filoteo, Maddox, and Davis (2001) compared Huntington’s disease patients with normal controls on two different II tasks that varied in difficulty. In the easier task, the categories were linearly separable; in the

195


harder task, they were separated by a more complex nonlinear boundary. The patient groups were impaired in learning both kinds of task. In a later study, Filoteo, Maddox, Salmon, et al. (2005) compared PD patients with normal controls on two, similarly defined II tasks. In this case, the PD patients were impaired on only the more difficult task (as discussed in Sections 3.1 and 3.2). As noted earlier, dissociations depend upon the failure to reject the null hypothesis of no difference and are thus always subject to a Type I error. For this and other reasons, we have argued that the appropriate statistical approach should be based on state-trace analysis. Figure 7 presents a statetrace plot of the results reported by Filoteo et al. (2001) and Filoteo, Maddox, Salmon, et al. (2005).6 It shows performance on both the linear (easy) and nonlinear (hard) II tasks for each of six blocks of trials for the two sets of patient groups and their corresponding normal controls. Even though the stimulus sets and the nature of the nonlinear bound differed between the

100 HD NC (HD) 90

PD

Nonlinear II

NC (PD) 80

70

60

50 50

60

70

80

90

100

Linear II

Figure 7 Combined state-trace plot from Filoteo, Maddox, and Davis (2001) and Filoteo, Maddox, Salmon, and Song (2005). HD refers to Huntingdon’s disease patients and NC (HD) refers to the corresponding group of normal controls (data from Figures 4 and 7 of Filoteo et al., 2001). PD refers to Parkinson’s disease patients and NC (PD) refers to the corresponding group of normal controls (data from Figure 4 of Filoteo, Maddox, Salmon, et al., 2005).

6

These data were obtained from Figures 4 and 7 of Filoteo et al. (2001) and Figure 4 of Filoteo, Maddox, Salmon, et al. (2005).

196


two experiments, the data nevertheless appear to fall on essentially the same monotonically increasing curve. While this may be a coincidence, the most important point is that this plot reveals that the differences in performance between the two patient groups can be accounted for by differences in the relative difficulty of the II tasks between the two experiments and the shape of the resulting curve. The shape of the curve may be interpreted as resulting from different changes in the performance on the two kinds of task as a function of learning. Because these changes are nonlinear, apparent dissociations may appear. In the case of the PD patients, the difference in their performance relative to the controls appears to be greater on the nonlinear II task than on the linear II task. In the case of the Huntington’s patients, the difference in their performance relative to the controls appears to be approximately the same on both tasks (although the shape of the curve also suggests that they may be relatively more impaired on the linear task). Therefore, while these data may be informative in other ways, state-trace analysis shows that they do not offer strong evidence against a “singlesystem” model.

4.4. Interpretation of Two-Dimensional State-Trace Plots: The Role of Confounds The “single-system” model shown in Figure 4B is a convenient straw man that is, in our opinion, unlikely to be true. This does not mean that we endorse an alternative multiple-systems view. Rather, we are mindful of the fact that any two tasks or dependent variables that are sufficiently dissimilar to warrant investigation are likely to draw upon a (potentially large) number of different systems or processes or be affected by other uncontrolled variables. For this reason, showing that two tasks can, under some circumstances, yield a twodimensional state-trace is an essentially trivial result since it is necessarily true. The challenge in this respect is to show that the two tasks yield a twodimensional state-trace under conditions that are theoretically relevant. This is not always obvious nor an easy thing to achieve. One important advantage of a model such as COVIS is that it clearly specifies a set of minimal contrasts that should differentially affect the different category learning systems it proposes and thereby yield a twodimensional state-trace plot. Nevertheless, the attribution of such a pattern of results to the relevant theoretical constructs is not always straightforward. This concern is nothing new and is simply the extension of the principles of good experimental design to the bivariate domain of state-trace analysis. For this reason, just as in the more familiar univariate domain, it is necessary to mount an additional argument to show that an observed result can be attributed to the theoretical constructs of interest.


197

The potential pitfalls associated with the interpretation of twodimensional state-trace plots can be illustrated by a recent study by Newell et al. (2010).7 The starting point for this study was the observed dissociation, reported by Zeithamova and Maddox (2006) that category learning in an RB task is selectively impaired by concurrent task demands compared to learning in an II task. Reanalysis of these data using statetrace analysis revealed an apparent two-dimensional state-trace plot. Although it was not possible to reject a one-dimensional curve on the basis of formal statistical analysis, the original study was not designed with this analysis in mind and there was enough evidence, in our minds, to suggest the involvement of multiple systems consistent with the predictions of COVIS. However, in a subsequent series of replications of the basic experiment, we consistently observed unambiguous one-dimensional statetrace plots. On examination of this apparent inconsistency, Newell et al. observed that, in their experiment, Zeithamova and Maddox did not partition their participants into those who appeared to learn the task and those who did not. When only the learners were analyzed, the data from Zeithamova and Maddox (Experiment 1) also revealed a clear onedimensional state-trace plot, consistent with the pattern of data found by Newell et al. The inclusion of nonlearners was sufficient to change the dimensionality of the state-trace plot. As pointed out by Newell et al., this was due to the fact that different proportions of nonlearners occurred under different conditions of the experiment. That is, the proportions of nonlearners varied between the load and no-load conditions and between those learning either a RB or II structure. This meant that although the performance of learners was well-described as a function of single learning parameter, consistent with a “single-system” model, the performance of the entire group was a function of both the amount learned and a variable proportion of nonlearners who, by definition, learned nothing. Because this proportion differed between conditions, the resulting state-trace plot was approximately two dimensional leading to the incorrect inference that performance was a function of two latent variables or learning systems.8 This example suggests that caution should be exercised in attributing an observed twodimensional structure to the theoretical mechanism of interest. As in any experiment, confounds are possible (see Section 3.4 for discussion of other “confounds” in experiments comparing II and RB tasks, e.g., Nosofsky et al. 2005).

7 8

We illustrate these problems in relation to state-trace analysis. Exactly the same issues apply to the interpretation of dissociations. As well as to the incorrect inference that the two tasks could be dissociated.

198


4.5. Interpretation of Two-Dimensional State-Trace Plots: Systems, Processes, and Latent Variables Dissociation logic was initially developed in neuropsychology (Lackner, 2009; Teuber, 1955). In that context, dissociations were naturally interpreted in terms of major functional capacities, such as reading, writing, language, and planning, involving relatively large brain regions and functional processing systems. This manner of interpreting dissociations has continued to the present day and forms one of the methodological foundations of the cognitive neuropsychology program where it is used extensively to partition mental function into separate processing modules (Coltheart, 1985; Shallice, 1988). However, there is nothing in the logic of dissociations that compels this interpretation, a point that is made explicit in the logic of state-trace analysis. In this context, the intervening constructs are best described as latent variables or model parameters that have no substantive meaning outside the particular theory to which they are relevant. This point of view was first proposed by Dunn and Kirsner (1988) in their analysis of dissociation logic in which they represented the effects of independent variables on latent variables and the effects of latent variables on dependent variables simply in terms of mathematical transformations between one set of variables and another. Although they referred to the intervening variables as “processes,” these were viewed as abstract constructs or parameters that had no essential meaning. It follows from the above that evidence for multiple intervening constructs, whether based on dissociations or, as we propose, state-trace analysis, is equally consistent with any number of different theoretical interpretations. It is thus equally consistent with an interpretation based on multiple systems, such as COVIS, as with one based on multiple processes or parameters, such as ALCOVE (Kruschke, 1992). ALCOVE proposes that category learning is dependent on a single psychological system governed by four different control parameters. In any one study, if at least two of these parameters are differentially affected by the independent variables and if they, in turn, differentially affect performance on RB and II tasks, then a two-dimensional state-trace plot (as well as potential dissociations) will result. The dimensionality of the state-trace plot reveals the number of underlying latent variables but says nothing about their nature. In particular, state-trace analysis does not offer a principled means of distinguishing between an interpretation of the data in terms of multiple parameters of a single system (as per ALCOVE) or in terms of parameters of multiple systems (as per COVIS). Distinguishing between these interpretations requires additional criteria such as the nature of experimental effects, their internal logic, and their respective abilities to account for the data. This, we suggest, depends upon specification of an explicit mathematical model that precisely defines the theoretical


199

constructs in a form that allows for quantitative evaluation. It is to this issue that we turn in the next section.

5. The Contribution of Mathematical Modeling In the context of a discussion contrasting the multiple- and singlesystem view of categorization, formal models would seem to be an important touchstone. For example, Ashby and Ell (2002) offer the following definition of a system: a system is a mapping of state variables that represent inputs, to state variables that represent outputs, in a manner governed by a vector of parameters. The notion of a system, then, is fundamentally formal. For two systems to be different there must be some parameter values that produce mappings that are unique to each system; if this is the case then the systems are neither identical nor fully nested. For two systems to be systems of the same type (and so potentially “the same”), the input and output variables must be the same. Ashby and Ell further propose that for any candidate systems to be candidate behavioral systems, each must make observable responses. This requirement is taken by Ashby and Ell to be met if the candidate systems produce outputs that can be interpreted as response choices. As suggested in Section 4, any model with multiple parameters can, in principle, produce observable dissociations. The generality of this statement suggests that any initial optimism about mathematical models per se resolving a multiple-systems debate should be tempered with a healthy skepticism about the ability of models to, on the one hand, uniquely describe a pattern of selective influence and, on the other, to unambiguously qualify as models of single or multiple systems. In this section, we briefly try to determine whether the COVIS model (as an exemplary multiple-system model) has advanced the multiple-system discussion.

5.1. Mathematical/Computational Models of Category Learning Excellent, contemporary, reviews of formal models of category learning are readily available (e.g., Kruschke, 2005, 2008). The primary dimensions that differentiate formal models are the sorts of representations, the sorts of learning rules, and the sorts of response-generation processes they employ. The last of these is the least theoretically relevant to category learning, the second is critical for understanding the trajectory of categorization (do people learn more on trials when they are incorrect, or do they learn equally on all trials, or do they update their beliefs rationally, etc.). It is only

200


the first, the representations a model postulates, that is of central relevance to the question of multiple systems. The kinds of representational schemes proposed for categorization break down naturally into two kinds. A pure exemplar model would propose that all items presented for classification are stored in memory, and that classification results in consulting all of these items. A pure abstraction model would propose that only a summary of the presented items is retained; this summary could be a theory, a rule, a prototype, or a boundary of some variety. Impure, or hybrid, models mix these representational formats. Examples of hybrid models include COVIS, RULEX (Nosofsky, Palmeri, & McKinley, 1994), ATRIUM (Erickson & Kruschke, 1998), SUSTAIN (Love, Medin, & Gureckis, 2004), varieties of Rational models (Anderson, 1991; Tenenbaum & Griffiths, 2001), VAM (Vanpaemel & Storms, 2008), Smith and Minda’s (2000) approach, Minda and Miles’ (2010) model, and probably more all the time. Some of these have been very successful in accounting for a variety of data, others far less so. Hybrid models pose a natural group for consideration as multiple-system accounts (leaving aside the question of how one could hope to identify such an account, given the flexibility of any multiple-parameter model). This appearance is deceiving, however, as many hybrid accounts either explicitly deny the possibility, or do not meet the criteria for multiple systems. The first case is presented in Kruschke’s (2008) thoughtful review in this way, “[The] representational assumption for a model does not necessarily imply that the mind makes a formal representation of the stimulus. Only a formal model requires a formal description. . . The representations in the model help us understand the behavior, but those representations need not be reified in the behavior being modelled” (p. 269, fn1). If the representations are only in the model, and not in the brain, then they cannot determine systems of categorization, because it is only the person, and not a formal model, that can produce behaviors. The second case is exemplified by a model like SUSTAIN or VAM, where the component representations are unified into a single mechanism, and do not produce distinct formal predictions. It is only COVIS that claims to be a genuine multiple-systems model.

5.2. COVIS The original formulation of COVIS (Ashby et al. 1998) as a neuropsychological theory was supposed to span all three of Marr’s (1982) levels of description of an information-processing system: task (or computation), algorithm, and implementation. Marr’s notion of a task description is the ideal-observer solution that describes performance; that is, there must be some set of sources of information and some rules for their combination that describe the relationship between the stimulus and the response and these


201

sources and rules constitute the “task.” Ashby et al. provide a different sort of task analysis, by characterizing the global dynamics of COVIS, with a particular focus on the bias the model predicts due to the way it integrates its RB and implicit-boundary-based subsystems.9 The implementation level description of COVIS identifies its formal components (inputs, subsystems, system combination, response generation) one-for-one with brain regions. COVIS (as presented most recently in Ashby et al., in press) is a complicated model. It has at least 19 free parameters, at least eight of which govern the verbal system. It has at least five random factors, which prevents the model from making deterministic predictions, and the competition between the verbal and procedural systems is a complex relationship depending on the performance of each system and the biases of the model as a whole. What COVIS predicts for a given experiment cannot be determined without investigating its parameter space. Thus, predictions from COVIS about the effects of various lesions or diseases on categorization are by virtue, first, of the identification of the formal structure of the capacity to categorize and, second, of the correct identification of the parts of the brain that support (vs. “implement”) the various aspects of that structure. This order is critical—if a model does not get the formal structure correct (if it cannot fit data), then its claims about identity with brain regions are necessarily irrelevant (if it can fit data, then its claims may remain senseless!). Thus, our concern here is with the formal structure (what Ashby et al. take as the algorithmic level10) of the model, to see how it fits data.

5.3. The Nature and Use of Formal Models While the utility of formal models in science is beyond doubt, their nature and use in psychology is problematic. Lewandowsky and Farrell (2010) remind us of the paradigmatic case from astronomy, in which planetary motion models moved from Ptolemaic to Copernican to Keplerian due to two considerations. The first shift was due primarily to the simplicity of the Copernican model, which did not fit the data much better than its geocentric competitor, while the second shift was due to the near-perfect fit of the Keplerian model to the then-available data. The example is attractive for another reason: as Newton showed, the Keplerian model is consistent with 9

The critical nature of these predictions is that what others (e.g., Medin & Schaffer, 1978) have identified as the role of dimensional attention in categorization are actually the result of a dimension-dependent rulebased system competing with an exemplar system. The identification of dimensional attention with explicit, verbalizable rules would suggest that no nonverbal creature could show attention learning, which is obviously false. 10 We will discuss in Section 6 whether this notion of an algorithm is unproblematic. If it is meant to imply that these computations happen “in the mind,” then it is; if it is meant to imply that these computations describe the functional organization of the capacity, then it is not.

202


an underlying causal mechanism (gravity). We now understand this model to be true: The solar system moves the way it does because its bodies have the masses, positions, and velocities that they do. In psychological theory, the parameter vectors of formal models take the role of constructs like “mass,” but we still wish to know whether any particular model is closer to the truth than any other. The promise of a neuropsychological model like COVIS is that it seems to be better positioned to be true than a (mere) psychological model, because it specifies something extra, something akin to adding “gravity” to the explanation of planetary motion. The argument over single- versus multiple-systems has developed (though it need not have) partially around this distinction. Single-systems psychological models, like the GCM and ALCOVE, do not claim that their formalisms are actually entities in some inner mental world (as the quote earlier by Kruschke, 2008, makes clear). Their truth is established only by behavioral experiments; they are in essence measurement models (Stevens, 1946) that tell how a parameter can be measured with an experiment. Neuropsychological models like COVIS, however, do claim that their formalisms are entities, that the components of the model are regions or states or processes in the brain. Their truth is jointly established by behavioral and neuropsychological experiments. The implications of this for evaluating COVIS are difficult to tease out. COVIS, just as it claims to be, is two kinds of model at once. On the one hand, it is a formal measurement model, with equations specifying how its parameters account for data. These equations tell us, for example, how increasing a learning rate parameter, or a decision rule parameter, will change the model’s predicted responses. On the other hand, the parameters are provided meaning by being embedded in an informal theory, a kind of structural model (Busemeyer & Jones, 1983), so that some tasks, or some factors (like working memory capacity, or brain diseases), are taken to cause the parameters to change in certain ways. The presence of a concurrent task is supposed to interfere with the explicit system, reducing some parameters and increasing others, while leaving the implicit system alone. This distinction is clear, in principle, but becomes unclear in practice due to the ambiguous role of the neuropsychological explanations. The point of raising the issue of measurement versus structural models is to provide leverage on the question of the role of computational modeling in establishing the veracity of the multiple-systems theory. This leverage is partially undercut by the dual role of neuropsychological theory in the model—it enters at both the structural level (where, e.g., the locus of the explicit system is identified) and the measurement level (e.g., where the values, or at least names, of parameters in the implicit system are identified). Thus, while the presence of a formal model might suggest a unity between psychological and neural processes akin to the relationship


203

between the geometry of Kepler and the physics of Newton, in the case of COVIS, the relationship is distinctly unclear. To clarify the relationship between the neural and the psychological aspects of COVIS, let us look at a plausible prediction the model might make: Learning to discriminate on the basis of family resemblance (e.g., to tell the Hatfields from the McCoys) depends on the timely delivery of dopamine to the procedural learning system. This prediction involves three things we do not know—and three ways experimentation might enlighten us about categorization. At the highest level, the claim refers to an empirical and testable statement about the world, “learning family resemblances depends on dopamine.” This requires two measurements: one, the degree to which a person has actually learned a family resemblance (vs., e.g., learned a simple rule, or not learned at all), and the other, a measure of the availability of dopamine. Should this relationship be established, we would then be left with two more model-theoretic, (potential) explanations of this fact. The first of these refers to a psychological model of the mental operations that underlie learning family resemblances and states that such learning is “procedural.” The second claim refers to a neuropsychological theory (which is also a model but we will use different terms to avoid confusion) about how the set of operations described by the psychological model depend on or are subserved by physical properties of the brain. This theory states that procedural learning (whatever that might be) depends on dopamine. What is the relationship between the psychological model and the neuropsychological theory? One possibility is that of entailment. For example, our understanding of brain science may be advanced enough to know that if learning is procedural then it has to depend on dopamine, as there is no other way this kind of learning could be implemented in the brain. The neuropsychological theory is thus entailed by the psychological model and is a bona fide prediction of this model. This means that were it to be shown that learning family resemblances does not depend on dopamine, it would follow that it cannot be based on procedural learning. It is highly unlikely, given our present understanding of brain science, that in COVIS or other similar models, the neuropsychological theory is entailed by the psychological model (Coltheart, 2006; Uttal, 2001). If this were so, then the claim that category learning depends upon two separate learning systems would depend upon particular neuropsychological facts. For example, if it were shown tomorrow that parts of the caudate nucleus were not differentially involved in learning RB and non-RB categories, would this challenge the psychological claim that there are two different learning systems? If it did, then what are we to make of the large number of behavioral dissociations used to support this hypothesis? Either they do not distinguish between single- or multiple-system views, in which case they cannot count as evidence one way or the other, or they do, in which case

204


they do so independently of how they may be implemented in the brain. That is, either the neuropsychological data are irrelevant to the claim or the behavioral data are irrelevant. Since it is claimed that neuropsychological and behavioral data are both relevant to the evaluation of COVIS, it follows that the relationship between the two parts of the model cannot be one of entailment (as defined above). What else could it be? One possibility is that of contingency. That is, as far as we know, it happens to be an empirical fact that the RB and procedural systems depend on proper function of the head and tail of the caudate nucleus, for example. In this case, were it shown tomorrow that these parts of the caudate nucleus are not differentially involved in learning RB and non-RB categories, the psychological model would be unassailed with the neuropsychological theory being simply revised accordingly. It follows from this view that while neuropsychological facts may be a useful source of hypotheses concerning the functional characteristics of the psychological model, nothing in the final analysis depends on this— neuropsychological data are ultimately irrelevant to the central claims of the psychological model. It follows from the above argument that unless one were to want to make the strong and, at our present understanding of brain science, unwarranted claim that the psychological model part of COVIS entails the neuropsychological theory part then the only evidence for the former derives from the behavioral data and its ability to account for these data.

5.4. Model Comparisons How well does the formal psychological model part of COVIS, sketched above, do at capturing performance, relative to other models? The short answer is, we do not really know. A full description of the formal structure of the model is beyond the scope of this chapter. Interested readers should consult Ashby et al. (in press) for an in-depth presentation. However, what is clear from the full description of the model is that because it has so many stochastic components (e.g., which rule is chosen for switching, how likely a person is to switch to a new rule, which rule is chosen to be active, what the active rule predicts for a response, what the procedural system predicts for a response) interacting in nonalgebraic fashions, it is probably impossible to fit to data using Maximum Liklihood Estimation methods. Demonstrations of fit are made via MonteCarlo sampling (Ashby et al. 1998), but are not very frequent. Ashby et al. (in press) provide another example of this, in their demonstration of COVIS’s ability to account for behavioral data by approximating the dissociation observed by Waldron and Ashby (2001; discussed in Sections 3.3 and 3.4). Ashby et al. say that, “because the dual task reduces the capacity of working memory and executive attention, we assumed that


205

participants would therefore be less adept at selecting appropriate new rules to try and at switching attention away from ineffective rules.” Thus, the demonstration of the model’s ability to accommodate the data proceeds by modifying two of the explicit system’s parameters (one that governs the chance that the system will perseverate in the face of error, and one that governs the chance that the system will select a random rule instead of a salient rule). The demonstration also involves, due to the nonconfusable nature of the stimuli, both preventing any generalization in the implicit system and making the rule system’s predictions deterministic. This produces the predicted (and observed) near-perfect dissociation, with almost no effect of concurrent task on performance in the II condition. Critically, the model parameters are not discovered by a statistical procedure, but by the application of an informal structural model. It is informative to compare the Ashby et al. (in press) model-based account of the Waldron and Ashby (2001) data with Nosofsky and Kruschke’s (2002) account. Nosofsky and Kruschke explored the model ALCOVE, which is a four-parameter model combining exemplar representations, dimensional attention, and error-driven learning. Nosofsky and Kruschke found that ALCOVE could produce results like Waldron and Ashby’s over a wide range of parameter values, simply by allowing that the concurrent task interfered with participants’ ability to learn to modify their dimensional attention. This would be generally equivalent to COVIS losing the ability to use its verbal rule system, although this is not Ashby et al.’s preferred explanation via COVIS. It is worth noting that ALCOVE could only fit the data when the generalization was rather broad, allowing associations learned for one stimulus to generalize to others even though the items were not confusable. Ashby et al.’s demonstration was obtained in the exact opposite situation, where the generalization was very narrow. This emphasizes Nosofsky and Kruschke’s point that “[i]t seems advisable to test alternative models by using richer sets of parametric data, rather than relying on qualitative patterns of results involving just two data points. Before ruling out an entire class of models, one should conduct an investigation over a reasonably large region of the models’ available parameter space” (p. 171). COVIS (the formal model) has not been tested against data sets in this manner. The argument against formal single-system accounts of category learning is generally that they do not predict the dissociations multiple-system theorists explore. This has nothing to do with the formal models of either single or multiple systems, but entirely to do with the informal theories that surround them. Mathematical models can resolve only so much of the debate about multiple systems. Parameter estimation can tell us which parameter values are required to approximate the data from any given set of manipulations, and informal theory (or common sense) can tell us if these values are meaningful. Model selection can tell us whether one model

206


is generally more accurate given its flexibility, but model selection is incredibly difficult given the complexity and similarity of the formal models and the variability inherent in human category learning. In summary, examination of COVIS as an example of a multiple-systems mathematical model leaves a number of questions unanswered. First, due to the nature of its formalisms, it is very difficult to compare its performance against other models, and thus such rigorous comparisons have yet to be undertaken. Second, the dissociations in behavioral data do not provide unambiguous support for COVIS, and some at least can be explained via other “single-system” models. Third, one of the key selling points of the model—its ability to account for phenomena at both the neural and the psychological level—makes it unnecessarily complex, especially when ultimately the neuropsychological data are irrelevant to the debate about how many systems are involved in our capacity to categorize.

6. Discussion and Conclusions The title of this chapter is deliberately provocative and, of course, a little flippant. However, what we hope to have highlighted in the preceding sections is that determining the “fact” or “fantasy” of discrete, functionally independent systems of category learning requires very careful consideration and integration of a wide range of evidence. This evidence suggests to us that the case for multiple systems underlying either probabilistic or deterministic category learning is still very much open to debate and may be challenged in three main ways. First, in our review of the “best” empirical evidence supporting a multiple-systems interpretation, we observed that the interpretation of much of this research is contentious and open to alternative explanations (Sections 2 and 3). Second, we have argued that since much of this research has relied on the discovery of dissociations, it rests on a flawed logic (Section 4). A feature of this logic is that it depends on the failure to observe an effect and thus is particularly prone to Type I errors. This means that researchers may be easily misled into believing that they have discovered multiple processing systems when they have simply failed to observe a significant effect. Finally, we have argued that dual nature of theoretical accounts such as COVIS that link a psychological model of category learning to proposed neuropsychological structures and processes may serve to complicate rather than to clarify our understanding of category learning at both levels (Section 5). At a fundamental level, the fact that data can be fit by a mathematical model that incorporates multiple systems in no way reifies the existence of these systems in the person learning to categorize. Ultimately, however, the answer to the question we pose in the title requires clarity about what constitutes an explanation of a given phenomenon.


207

6.1. Varieties of Explanation Psychological theory is a very difficult project, and is subject to problems other scientific theories easily avoid. Every field must be clear about what it seeks to explain, and how it seeks to explain it. Substituting an explanandum for an explanation is wrong, as when an effect becomes a cause—for example, when someone identifies “the Stroop effect” as the reason people are slowed by contradictory information. These may pass notice because the effect is a technical term, and technical terms are often explanations. But in psychology, we also must guard against ordinary terms, like “rule,” being mistaken for technical ones. So when we wish to explain how a person learns a category, and we want to say “by testing rules,” we must realize that we have not offered a technical explanation, but an ordinary (perfectly sensible) one. There is only one kind of “system” that can follow a rule, and that is an intelligent creature who can be said to follow it correctly, or not. A volcano cannot erupt in violation of a rule, though its eruption can be inexplicable or unusual. A neuron cannot fire correctly, but only normally. A person can follow a rule, and can forget to do so. Explanations of things that people do are fraught, because we have two ways of offering such explanations. As Bennett and Hacker (2003) so lucidly describe, consider a man who buys a pair of opera tickets. Why does he do so? The explanation, “Because he wishes to give his wife a present” is perfectly valid. His wishes are constitutive of the act of buying the tickets (as opposed to buying them by accident, or being tricked into it). The explanation, “Because his brain was in such-and-such a configuration” might be valid if we understand that the brain is not constitutive (it does not enter into what we mean by “buy the tickets”). Instead, the brain explanation is a description of the causally enabling conditions (Trigg & Kalish, in press) for exercising the capacity to, for example, buy the tickets. In the same way, a formal model can be of one of two types. One can model the constitutive components of a behavior, as we might when modeling how people make choices, reach decisions, test hypotheses, rely on heuristics, or draw conclusions from arguments (e.g., Evans, 2008; Gigerenzer & Todd, 1999; Tversky & Kahneman, 1981). In the case of category learning, such a model could quantify the probability that a person considers some particular simple rule about single features before a more complicated two-feature rule, for example. The other variety of explanation is to model the causally enabling conditions of a behavior, as we might when describing category learning as the association of labels with stored instances. All this kind of model amounts to is a mapping of measured inputs to measured responses; an answer to the question raised in Section 5.3 about what the word “properly” in the phrase “the brain must be functioning properly” means. Models of perception, memory retrieval, and familiar models of category learning

208


(e.g., ALCOVE) can be read as of this variety, as they explicate the mechanisms that support our capacity to perceive, remember, and learn. The possibility that psychology might arrive at this variety of model is thrilling, and to mistake it with the first type, only confusing. Multiple-systems theories of categorization, such as COVIS, are uniformly of this confused sort since they mix constitutive and causal explanations as if they are both the same kind of explanation. On the one hand, we have an explicit “system,” which is really a system that does things that a person does. Only a person can use language (so a “verbal system” is a person), or follow a rule (for a “RB system”), or represent something to him/herself in an explicit way. Explanations based on this “system” can become meaningless: if the explanation of how a person learns a rule is that their “verbal system” learns the rule, then either the explanandum has simply served as an explanation or the explanation is homuncular. On the other hand, multiple-systems models postulate an implicit “system,” which is a description of how our brain has to be organized such that we can learn things like how to shoot a basketball, or go left on a wave, or, maybe, tell a Hatfield from a McCoy. The difference between these kinds of explanations can be seen in what would happen if they were to fail. If the implicit “system” were broken, we could not do the things we do, but if the verbal “system” were broken, we would not be doing anything at all (things might still be happening to us or in us. . .). For a multiple-system theory to make sense (let alone to be true or false), it must respect the difference between who we are, and how we work. An interesting consequence of this distinction is that respecting it does not eliminate theories that propose qualitatively different ways to learn a categorical distinction. The formal model that is COVIS might even be just such a theory if it were interpreted correctly—that is, if it made no reference to “explicit” versus “implicit” systems and if the actions of the RB system were understood as abstractions of function. Many of the hybrid models of Section 5.1 are compatible with just this sort of interpretation, as indeed are even simple models (like ALCOVE) that depend qualitatively on the values of their parameters. But respecting the distinction eliminates the logical possibility of multiple-systems theories that contrast explicit and implicit “systems.”

6.2. Explanatory Power One of the clear impediments to resolving the multiple- versus singlesystems debate in category learning (and many other fields) is that often proposed explanations of phenomena fail to respect the difference between people and their brains. Moreover, as we have been at pains to describe, proponents of multiple-systems explanations often take suggestive, but rather inconclusive evidence at one level (e.g., brain structure) to “support”


209

accounts presented at another level (e.g., functional organization). This combination of explanations gives rise to an illusory veneer of “convergent” evidence. If one steps back and considers the evidence at each level in isolation, the seductive appeal of multiple-system accounts diminishes markedly. At the brain level, there is considerable debate about the reliability and strength of the conclusions that can be drawn from, for example, imaging studies about the operation of distinct systems (e.g., Coltheart, 2006; Page, 2006; Uttal, 2001). At the functional level, the dichotomization of processes and representations into separable systems is muddied considerably by empirical demonstrations that are odds with the putative characteristics of the systems (see Sections 2 and 3). Furthermore, as noted by Keren and Schul (2009), explaining phenomena via the invocation of multiple (most often dual) systems might “feel” productive simply because it serves a basic cognitive function of wanting to structure and order our knowledge (i.e., ironically, categorization). But a good classification scheme combined with a “good story” should not— though it often does—increase our confidence that an explanation is correct. It is this, in our opinion misplaced, sense of confidence that leads Poldrack and Foerde (2008) to offer the following (startling) evaluation: . . .in the end the single- versus multiple-systems approaches must be evaluated on the basis of both their explanatory power and their productivity, in terms of generating new and interesting results. We would argue that the MMS [multiple-memory-systems] view has continued to provoke new and interesting results across human and animal research, whereas the single-system view has largely focused on protecting its central assumptions and attacking results from the MMS approach rather than inspiring novel findings. It is this kind of productivity that we believe argues strongly in favor of the MMS approach in comparison to the single-system approach. (p. 202)

However, good explanations must not only be productive or fruitful in the sense of “inspiring novel findings” (as emphasized by Poldrack and Foerde) but must also satisfy criteria like coherence, empirical accuracy, scope, internal consistency, simplicity, precision of predictions, and formalism (Brewer, Chinn, & Samarapungavan, 1998; Keren & Schul, 2009). As the preceding sections highlight, many multiple-system theories fall short when considering such criteria.

6.3. Final Thoughts Our goal in this chapter has not been to protect assumptions, attack results, and to be generally unproductive. Rather, it has been to highlight fundamental problems we perceive in currently popular multiple-systems

210


accounts of category learning. In so doing, we have aimed to elucidate underlying assumptions, present a balanced interpretation of results, and to suggest an alternative productive program based on the methodological and conceptual analyses we have offered. Because psychological theory is a very difficult project, we need to be as certain as possible about how we interpret evidence and the manner in which we construct our explanatory accounts. For the reasons that we have outlined, we believe that our theoretical understanding of category learning is far from settled. The challenge for researchers in categorization (and many other areas) is to escape the shackles of multiple-system “explanations” and to develop models and measurement procedures that will facilitate a systematic exploration of the body of data before us.

ACKNOWLEDGMENT The support of the Australian Research Council (Grant DP: 0877510 awarded to the three authors) is gratefully acknowledged.

REFERENCES Anderson, J. R. (1991). The adaptive nature of human categorization. Psychological Review, 98, 409–429. Aron, A. R., Shohamy, D., Clark, J., Myers, C., Gluck, M. A., & Poldrack, R. A. (2004). Human midbrain sensitivity to cognitive feedback and uncertainty during classification learning. Journal of Neurophysiology, 2, 1144–1152. doi:10.1152/jn.01209.2003. Ashby, F. G., Alfonse-Reese, L. A., Turken, A. U., & Waldron, E. M. (1998). A neuropsychological theory of multiple systems in category learning. Psychological Review, 105, 442–481. doi:10.1037/0033-295X.105.3.442. Ashby, F. G., & Ell, S. W. (2002). Single versus multiple systems of learning and memory. In H. Pashler & J. Wixted (Eds.), Stevens’ handbook of experimental psychology (pp. 655–691). (3rd ed.). Vol 4: Methodology in experimental psychology. Hoboken, NJ: John Wiley & Sons Inc. Ashby, F. G., Ell, S. W., & Waldron, E. M. (2003). Procedural learning in perceptual categorization. Memory & Cognition, 31, 1114–1125. Ashby, F. H., & Ennis, J. M. (2006). The role of the basal ganglia in category learning. In B. H. Ross (Ed.), The psychology of learning and motivation: Advances in research and theory (Vol. 46) (pp. 1–36). San Diego: Elsevier Academic Press. Ashby, F. G., & Maddox, W. T. (2005). Human category learning. Annual Review of Psychology, 56, 149–178. doi:10.1146/annurev.psych.56.091103.070217. Ashby, F. G., Noble, S., Filoteo, J. V., Waldron, E. M., & Ell, S. W. (2003). Category learning deficits in Parkinson’s disease. Neuropsychology, 17, 115–124. doi:10.1037/08944105.17.1.115. Ashby, F. G., Paul, E. J., & Maddox, W.T. (in press). COVIS. In E. M. Pothos & A. J. Wills (Eds.), Formal approaches in categorization. New York: Cambridge University Press. Bamber, D. (1979). State-trace analysis: A method of testing simple theories of causation. Journal of Mathematical Psychology, 19, 137–181.


211

Bennett, M. R., & Hacker, P. M. S. (2003). Philosophical Foundations of Neuroscience. Oxford: Blackwell Publishing. Brewer, W. F., Chinn, C. A., & Samarapungavan, A. (1998). Explanation in scientists and children. Minds and Machines, 8, 119–136. Busemeyer, J. R., & Jones, L. E. (1983). The analysis of multiplicative combination rules when the causal variables are measured with error. Psychological Bulletin, 93, 549–562. Channon, S., Jones, M., & Stephenson, S. (1993). Cognitive strategies and hypothesis testing during discrimination learning in Parkinson’s disease. Neuropsychologia, 31, 175–182. doi:10.1016/0028-3932(93)90082-B. Coltheart, M. (1985). Cognitive neuropsychology. In M. I. Posner & O. S. M. Marin (Eds.), Attention and performance, Vol. 11 (pp. 3–37). Hillsdale, NJ: Erlbaum. Coltheart, M. (2006). What has functional neuroimaging told us about the mind (so far)? Cortex, 42, 323–331. Dunn, J. C., & Kirsner, K. (1988). Discovering functionally independent mental processes: The principle of reversed association. Psychological Review, 95, 91–101. Erickson, M. A., & Kruschke, J. K. (1998). Rules and exemplars in category learning. Journal of Experimental Psychology: General, 127, 107–140. Evans, J. St. B. T. (2008). Dual-processing accounts of reasoning, judgement, and social cognition. Annual Review of Psychology, 59, 255–278. doi:10.1146/annurev. psych.59.103006.093629. Filoteo, J. V., & Maddox, W. T. (2007). Category learning in Parkinson’s disease. In M. K. Sun (Ed.), Research progress in Alzheimer’s disease and dementia (Vol. 3) (pp. 2–26). New York: Nova Science Publishers, Inc. Filoteo, J. V., Maddox, W. T., & Davis, J. D. (2001). A possible role of the striatum in linear and nonlinear categorization rule learning: Evidence from patients with Huntington’s disease. Behavioral Neuroscience, 115, 786–798. Filoteo, J. V., Maddox, W. T., Ing, A. D., Zizak, V., & Song, D. D. (2005). The impact of irrelevant dimensional variation on rule based category learning in patients with Parkinson’s disease. Journal of the International Neuropsychological Society, 11, 503–513. doi:10.1017/S1355617705050617. Filoteo, J. V., Maddox, W. T., Salmon, D. P., & Song, D. D. (2005). Informationintegration category learning in patients with striatal dysfunction. Neuropsychology, 19, 212–222. doi:10.1037/0894-4105.19.2.212. Foerde, K., Knowlton, B. J., & Poldrack, R. A. (2006). Modulation of competing memory systems by distraction. Proceedings of the National Academy of Sciences of the United States of America, 103, 11778–11783. doi:10.1073/pnas.0602659103. Foerde, K., Poldrack, R. A., & Knowlton, B. J. (2007). Secondary-task effects on classification learning. Memory & Cognition, 35, 864–874. Gigerenzer, G., & Reiger, T. (1996). How do we tell an association from a rule? Comment on Sloman (1996). Psychological Bulletin, 11, 23–26. doi:10.1037/0033-2909.119.1.23. Gigerenzer, G., & Todd, P. M. (1999). Simple heuristics that make us smart. New York: Oxford University Press. Gluck, M. A., Shohamy, D., & Myers, C. (2002). How do people solve the “weather prediction” task?: Individual variability in strategies for probabilistic category learning. Learning & Memory, 9, 408–418. doi:10.1101/lm.45202. Heathcote, A., Brown, S. D., & Prince, M. (2010). The design and analysis of state-trace experiments. Manuscript submitted for publication, University of Newcastle, Australia. Heffernan, M. (2009). An examination of the processes underlying probabilistic category learning. Unpublished PhD thesis, School of Psychology, University of New South Wales. Jahanshahi, M., Wilkinson, L., Gahir, H., Dharminda, A., & Lagnado, D. A. (2010). Medication impairs probabilistic classification learning in Parkinson’s disease. Neuropsychologia, 48, 1096–1103. doi:10.1016/j.neuropsychologia.2009.12.010.

212


Keren, G., & Schul, Y. (2009). Two is not always better than one: A critical evaluation of two-system theories. Perspectives on Psychological Science, 4, 533–550. doi:10.1111/j.17456924.2009.01164.x. Kinder, A., & Shanks, D. R. (2003). Neuropsychological dissociations between priming and recognition: A single-system connectionist account. Psychological Review, 110, 728–744. doi:10.1037/0033-295X.110.4.728. Knowlton, B. J. (1999). What can neuropsychology tell us about category learning? Trends in Cognitive Sciences, 3, 123–124. doi:10.1016/S1364-6613(99)01292-9. Knowlton, B. J., Mangels, J. A., & Squire, L. R. (1996). A neostriatal habit learning system in humans. Science, 273, 1399–1402. doi:10.1126/science.273.5280.1399. Knowlton, B. J., Squire, L. R., & Gluck, M. A. (1994). Probabilistic classification learning in amnesia. Learning & Memory, 1, 106–120. Kruschke, J. K. (1992). ALCOVE: An exemplar-based connectionist model of category learning. Psychological Review, 99, 22–44. Kruschke, J. K. (2005). Category learning. In K. Lamberts & R. L. Goldstone (Eds.), The handbook of cognition (pp. 183–201). London: Sage. Kruschke, J. K. (2008). Models of categorization. In R. Sun (Ed.), The Cambridge handbook of computational psychology (pp. 267–301). New York: Cambridge University Press. Lackner, J. R. (2009). Hans-Lukas Teuber: A remembrance. Neuropsychology Review, 19, 4–7. Lagnado, D. A., Newell, B. R., Kahan, S., & Shanks, D. R. (2006). Insight and strategy in multiple-cue learning. Journal of Experimental Psychology. General, 135, 162–183. doi:10.1037/0096-3445.135.2.162. Lehrer, R. (2009). The decisive moment: How the brain makes up its mind. Melbourne, Australia: Text Publishing. Lewandowsky, S., & Farrell, S. (2010). Computational modeling in cognition: Principles and practice. Thousand Oaks, CA: Sage. Loftus, G. R., Dillon, A. M., & Oberg, M. A. (2004). Linear theory, dimensional theory, and the face-inversion effect. Psychological Review, 111, 835–865. Love, B. C., Medin, D. L., & Gureckis, T. M. (2004). SUSTAIN: A network model of category learning. Psychological Review, 111, 309–322. Lovibond, P. F., & Shanks, D. R. (2002). The role of awareness in Pavlovian condition: Empirical evidence and theoretical implications. Journal of Experimental Psychology: Animal Behavior Processes, 28, 3–26. doi:10.1037/0097-7403.28.1.3. Maddox, W. T., & Ashby, F. G. (2004). Dissociating explicit and procedural-learning based systems of perceptual category learning. Behavioral Processes, 66, 309–332. doi:10.1016/ j.beproc.2004.03.011. Maddox, W. T., Ashby, F. G., & Bohil, C. J. (2003). Delayed feedback effects on rile-based and information-integration category learning. Journal of Experimental Psychology: Learning, Memory, and Cognition, 29, 650–662. doi:10.1037/0278-7393.29.4.650. Maddox, W. T., Ashby, F. G., Ing, A. D., & Pickering, A. D. (2004). Disrupting feedback processing interferes with rule-based but not information-integration category learning. Memory & Cognition, 32, 582–591. Maddox, W. T., & Ing, A. D. (2005). Delayed feedback disrupts the procedural-learning system but not the hypothesis-testing system in perceptual category learning. Journal of Experimental Psychology: Learning, Memory, and Cognition, 31, 100–107. doi:10.1037/ 0278-7393.31.1.100. Maddox, W. T., Love, B. C., Glass, B. D., & Filoteo, J. V. (2008). When more is less: Feedback effects in perceptual category learning. Cognition, 108, 578–589. doi:10.1016/ j.cognition.2008.03.010. Marr, D. (1982). Vision. San Francisco: H. Freeman and Co.


213

Medin, D. L., & Schaffer, M. M. (1978). Context theory of classification learning. Psychological Review, 85, 207–238. Minda, J. P., & Miles, S. J. (2010). The influence of verbal and nonverbal processing on category learning. In B. H. Ross (Ed.), The psychology of learning and motivation: Advances in research and theory (Vol. 52) (pp. 117–162). San Diego: Academic Press. Mitchell, C. J., De Houwer, J., & Lovibond, P. F. (2009). The propositional nature of human associative learning. Behavioral and Brain Sciences, 32, 183–198. doi:10.1017/ S0140525X09000855. Newell, B. R., & Dunn, J. C. (2008). Dimensions in data: Testing psychological models using state-trace analysis. Trends in Cognitive Sciences, 12, 285–290. doi:10.1016/ j.tics.2008.04.009. Newell, B. R., Dunn, J. C., & Kalish, M. (2010). The dimensionality of perceptual category learning: A state-trace analysis. Memory & Cognition, 38, 563–581. doi:10.3758/MC. Newell, B. R., Lagnado, D. A., & Shanks, D. R. (2007). Challenging the role of implicit processes in probabilistic category learning. Psychonomic Bulletin & Review, 14, 505–511. Nomura, E. M., Maddox, W. T., Filoteo, J. V., Ing, A. D., Gitelman, D. R., Parrish, T. B., et al. (2007). Neural correlates of rule-based and information-integration visual category learning. Cerebral Cortex, 17, 37–43. doi:10.1093/cercor/bhj122. Nomura, E. M., & Reber, P. J. (2008). A review of medial temporal lobe and caudate contributions to visual category learning. Neuroscience and Biobehavioral Reviews, 32, 279–291. doi:10.1016/j.neubiorev.2007.07.006. Nosofsky, R. A., & Kruschke, J. K. (2002). Single system models and interference in category learning: Commentary on Waldron & Ashby (2001). Psychonomic Bulletin & Review, 9, 169–174. Nosofsky, R. M., Palmeri, T. J., & McKinley, S. C. (1994). Rule-plus-exception model of classification learning. Psychological Review, 101, 53–79. Nosofsky, R. A., Stanton, R. D., & Zaki, S. R. (2005). Procedural interference in perceptual classification: Implicit learning or cognitive complexity? Memory & Cognition, 33, 1256–1271. Nosofsky, R. M., & Zaki, S. R. (1998). Dissociations between categorization and recognition in amnesic and normal individuals: An exemplar-based interpretation. Psychological Science, 9, 247–255. doi:10.1111/1467-9280.00051. Page, M. P. A. (2006). What can’t functional neuroimaging tell the cognitive psychologist? Cortex, 42, 428–443. doi:10.1016/S0010-9452(08)70375-7. Palmeri, T. J., & Flanery, M. A. (2002). Memory systems and perceptual categorization. In B. H. Ross (Ed.), The psychology of learning and motivation: Advances in research theory (Vol. 41) (pp. 141–189). San Diego: Academic Press. Poldrack, R. A., Clark, J., Pare-Blagoev, E. J., Shohamy, D., Moyano, J. C., Myers, C., et al. (2001). Interactive memory systems in the human brain. Nature, 414, 546–550. doi:10.1038/35107080. Poldrack, R. A., & Foerde, K. (2008). Category learning and the memory systemsdebate. Neuroscience and Biobehavioral Reviews, 32, 197–205. doi:10.1016/j.neurobiorev.2007.07.007. Poldrack, R. A., & Packard, M. G. (2003). Competition among multiple memory systems: Converging evidence from animal and human brain studies. Neuropsychologica, 41, 245–251. doi:10.1016/S0028-3932(02)00157-4. Price, A. L. (2005). Cortico-striatal contributions to category learning: Dissociating the verbal and implicit systems. Behavioral Neuroscience, 119, 1438–1447. doi:10.1037/07357044.119.6.1438. Price, A. L. (2009). Distinguishing the contributions of implicit and explicit processes to performance of the weather prediction task. Memory & Cognition, 37, 210–222. doi:10.3758/MC.37.2.210.

214


Price, A., Filoteo, J. V., & Maddox, W. T. (2009). Rule-based category learning in patients with Parkinson’s disease. Neuropsychologia, 47, 1213–1226. doi:10.1016/ j.neuropsychologia.2009.01.031. Ransmayr, G., Bitschnau, W., Schmidhubereiler, B., Berger, W., Karamat, E., Poewe, W., et al. (1990). Slowing of high-speed memory scanning in Parkinson’s-disease is related to the severity of parkinsonian motor symptoms. Journal of Neural Transmission Parkinsons Disease and Dementia Section, 2(4), 265–275. Reber, P. J., Knowlton, B. J., & Squire, L. R. (1996). Dissociable properties of memory systems: Differences in the flexibility of declarative and nondeclarative knowledge. Behavioral Neuroscience, 110, 861–871. doi:10.1037/0735-7044.110.5.861. Reed, S. K. (1972). Pattern recognition and categorization. Cognitive Psychology, 3, 382–407. Schmitt-Eliassen, J., Ferstl, R., Wiesner, C., Deuschl, G., & Witt, K. (2007). Feedbackbased versus observational classification learning in healthy aging and Parkinson’s disease. Brain Research, 1142, 178–188. doi:10.1016/j.brainres.2007.01.042. Shanks, D. R., & St. John, M. F. (1994). Characteristics of dissociable human learning systems. Behavioral and Brain Sciences, 17, 367–447. Shallice, T. (1988). From Neuropsychology to Mental Structure. New York: Cambridge University Press. Sherry, D. F., & Schacter, D. L. (1987). The evolution of multiple memory systems. Psychological Review, 94, 439–454. doi:10.1037/0033-295X.94.4.439. Shohamy, D., Myers, C. E., Grossman, S., Sage, J., & Gluck, M. A. (2004). Cortico-striatal contributions to feedback-based learning: Converging data from neuroimaging and neuropsychology. Brain: A Journal of Neurology, 127, 851–859. doi:10.1093/brain/ awh100. Smith, J. D., & Minda, J. P. (2000). Thirty categorization results in search of a model. Journal of Experimental Psychology: Learning, Memory, and Cognition, 26, 3–27. Speekenbrink, M., Channon, S., & Shanks, D. R. (2008). Learning strategies in amnesia. Neuroscience and Biobehavioral Reviews, 32, 292–310. doi:10.1016/ j.neubiorev.2007.07.005. Speekenbrink, M., Lagnado, D. A., Wilkinson, L., Jahanshahi, M., & Shanks, D. R. (2010). Models of probabilistic category learning in Parkinson’s disease: Strategy use and the effects of L-dopa. Journal of Mathematical Psychology, 54, 123–136. doi:10.1016/j.jmp. 2009.07.004. Stanton, R. D., & Nosofsky, R. M. (2007). Feedback interference and dissociations of classification: Evidence against the multiple-learning systems hypothesis. Memory & Cognition, 35, 1747–1758. Stevens, S. S. (1946). On the theory of scales of measurement. Science, 103, 677–680. doi:10.1126/science.103.2684.677. Swainson, R., SenGupta, D., Shetty, T., Watkins, L. H. A., Summers, B. A., Sahakian, B. J., et al. (2006). Impaired dimensional selection but intact use of reward feedback during visual discrimination learning in Parkinson’s disease. Neuropsychologia, 44, 1290–1304. doi:10.1016/j.neuropsychologia.2006.01.028. Tenenbaum, J. B., & Griffiths, T. L. (2001). Generalization, similarity and Bayesian inference. Behavioral & Brain Sciences, 24, 629–640. Teuber, H. L. (1955). Physiological psychology. Annual Review of Psychology, 6, 267–296. Tharp, I. J., & Pickering, A. D. (2009). A note on DeCaro, Thomas, and Beilock (2008): Further data demonstrate complexities in the assessment of information-integration category learning. Cognition, 111, 410–414. doi:10.1016/j.cognition.2008.10.003. Trigg, J., & Kalish, M. (in press). Explaining how the mind works: On the relation between cognitive science and philosophy. Topics in Cognitive Science. Tversky, A., & Kahneman, D. (1981). The framing of decisions and the psychology of choice. Science, 211(4481), 453–458. doi:10.1126/science.7455683.


215

Uttal, W. R. (2001). The new phrenology: The limits of localizing cognitive processes in the brain. Cambridge, MA: MIT Press. Vanpaemel, W., & Storms, G. (2008). In search of abstraction: The varying abstraction model of categorization. Psychonomic Bulletin & Review, 15, 732–749. Waldron, E. M., & Ashby, F. G. (2001). The effects of concurrent task interference on category learning: Evidence for multiple category systems. Psychonomic bulletin & Review, 8, 168–176. Wilkinson, L., Lagnado, D. A., Quallo, M., & Jahanshahi, M. (2008). The effect of feedback on non-motor probabilistic classification learning in Parkinson’s disease. Neuropsychologia, 46, 2683–2695. doi:10.1016/j.neuropsychologia.2008.05.008. Zeithamova, D., & Maddox, W. T. (2006). Dual-task interference in perceptual category learning. Memory & Cognition, 34, 387–398.


C H A P T E R

S E V E N

Abstract Concepts: Sensory-Motor Grounding, Metaphors, and Beyond Diane Pecher, Inge Boot, and Saskia Van Dantzig Contents 1. Grounded Cognition 1.1. Introduction 1.2. The Grounded Cognition Framework 1.3. Evidence for the Grounded Cognition View: Concrete Concepts 1.4. Are All Concepts Grounded in Sensory-Motor Processing? The Scope Problem 2. Representing Abstract Concepts: Some Evidence for Grounding 2.1. Emotional Valence 2.2. Abstract Transfer 2.3. Force Dynamics 2.4. Summary 3. Explanations of Abstract Concepts 3.1. Dual Code Theory 3.2. Linguistic Co-occurrence 3.3. Hybrid Models 3.4. Abstract Concepts are Represented by Situations 3.5. Conceptual Metaphor Theory 4. Discussion Acknowledgments References

218 218 219 220 223 224 225 227 227 228 228 228 229 231 231 232 240 242 243

Abstract In the last decade many researchers have obtained evidence for the idea that cognition shares processing mechanisms with perception and action. Most of the evidence supporting the grounded cognition framework focused on representations of concrete concepts, which leaves open the question how abstract concepts are grounded in sensory-motor processing. One promising idea is that people simulate concrete situations and introspective experiences to represent abstract concepts [Barsalou, L. W., & Wiemer-Hastings, K. (2005). Situating abstract concepts. In D. Pecher, & R. A. Zwaan (Eds.), Grounding cognition: The role of perception and action in memory, language, and thinking (pp. 129–163). Cambridge: Cambridge University Press.], although this has not Psychology of Learning and Motivation, Volume 54 ISSN 0079-7421, DOI: 10.1016/B978-0-12-385527-5.00007-3

#


217

218

Diane Pecher et al.

yet been investigated a lot. A second idea, which more researchers have investigated, is that people use metaphorical mappings from concrete to abstract concepts [Lakoff, G., & Johnson, M. (1980). Metaphors we live by. Chicago: Chicago University Press.]. According to this conceptual metaphor theory, image schemas structure and provide sensory-motor grounding for abstract concepts. Although there is evidence that people automatically activate image schemas when they process abstract concepts, we argue that situations are also needed to fully represent meaning.

1. Grounded Cognition 1.1. Introduction How do people think? People may describe their thoughts as mental images, imagined movements through space, simulated sequences of actions, and so on. The idea that the elements of thought are not words or symbols, but visual and motor images is at the core of recent theories of cognition (e.g., Barsalou, 1999b; Glenberg, 1997; Grush, 2004; Zwaan, 1999), that are commonly referred to by the label grounded cognition. According to the grounded cognition view, thinking relies on the activation of the sensorymotor system. When a person thinks about an object such as a banana, the neural patterns in sensory-motor brain areas that have been formed during earlier experiences with bananas are reactivated. This reinstatement of neural activation results in a sensory-motor simulation. The visual system simulates seeing a banana, the motor system simulates the acts of grasping, peeling, and eating the banana, and the olfactory and gustatory system simulate the smell and the taste of the banana. In other words, when thinking about a banana, the brain acts partially as if the person is actually perceiving and interacting with a banana. The grounded cognition approach solves two issues that have not been addressed very thoroughly by previous theories. First, most theories are quiet on the issue of how mental symbols get their meaning, other than by links to other symbols. In order for mental symbols to be meaningful, they need to be grounded in experience. Sensory-motor simulations might provide such grounding. Second, the grounded cognition approach provides an explanation for the flexibility of concepts (i.e., mental representations). Concepts are dynamic constructions that are highly variable, dependent on context and recent experience (Anderson et al., 1976; Barsalou, 1982, 1993; Pecher & Raaijmakers, 2004; Zeelenberg, Pecher, Shiffrin, & Raaijmakers, 2003). Simulations can involve different sensory modalities to different extents (Pecher, Zeelenberg, & Barsalou, 2004). For example, a concert pianist may form different simulations of piano in various contexts. When thinking about an upcoming performance, the simulation

Grounding Abstract Concepts

219

might include the piano’s sound, along with the fine hand movements involved in playing the instrument. When planning to move the piano into a new apartment, however, the simulation might include the shape, size, and weight of the piano, along with the gross movements necessary for lifting and moving it. Furthermore, different instances of a category can be instantiated from one simulation to another. Thus, the simulation mechanism provides a logical explanation of concept variability. These and other reasons have led many researchers to put the grounded cognition framework to the test. Most of the evidence supporting the grounded cognition framework focused on representations of concrete concepts (e.g., objects; Pecher, Zeelenberg, & Barsalou, 2003) or actions (Glenberg & Kaschak, 2002). A frequent criticism of these studies is that they fail to explain how more abstract concepts such as power or democracy are represented. Because people do not have direct sensory-motor experiences with abstract concepts, it is not immediately obvious how they can represent such concepts by sensorymotor simulations. In this chapter, we will discuss various views on how abstract concepts might be represented through sensory-motor simulations. In particular, we will investigate the idea that metaphorically used image schemas may provide a mechanism to ground abstract concepts in embodied experience.

1.2. The Grounded Cognition Framework At first glance it may seem more logical to adopt a symbolic approach to explain abstract concepts. After all, democracy does not have a particular color, shape, smell, sound, or weight. People cannot kick, bite, or squeeze it. It is associated with other abstract concepts such as freedom, majority, or republic. Thus, one may wonder how the sensory-motor systems in the brain can ever represent abstract concepts. On the other hand, the cumulative evidence for the idea that cognition is grounded in sensory-motor systems is too strong to ignore. This theory solves the grounding problem (Harnad, 1990) for concrete concepts, involving objects and actions. If we now require symbolic representations to explain abstract concepts, the grounding problem returns. Therefore, it makes sense to investigate whether the same mechanisms that have proven so successful for concrete concepts can also be used for abstract concepts. Probably the most influential view of grounded cognition is given by Barsalou’s Perceptual Symbol Theory (Barsalou, 1999b). Barsalou proposes that mental representations used in cognitive tasks are grounded in the sensory-motor system. The building blocks of mental representation are perceptual symbols, partial reinstatements of the neural patterns that were stored in perceptual and motor brain areas during actual experience and interaction with the environment. These perceptual symbols are organized

220

Diane Pecher et al.

into simulators. A simulator is a set of related perceptual symbols, representing a particular concept. A simulator should not be viewed as a reenactment of an entire previous experience. Rather, the simulator uses a subset of components that were captured during different experiences. Because perceptual symbols can be combined dynamically to produce a simulation, concepts are specific instances of a category. For example, a person may represent a banana by simulating eating a ripe banana or by simulating seeing a green banana hanging on a tree. This ability to combine components of experience dynamically allows the system to form new concepts. Since simulations are composed of perceptual symbols, they are analogous to actual perception and action. According to Barsalou, a perceptual symbol system is a fully functional conceptual system, which “represents both types and tokens, produces categorical inferences, combines symbols productively to produce limitless conceptual structures, produces propositions by binding types to tokens, and represents abstract concepts” (Barsalou, 1999b, p. 581).

1.3. Evidence for the Grounded Cognition View: Concrete Concepts 1.3.1. Perceptual Simulations On this account, even when people are not actually perceiving or interacting with objects in the world, the sensory-motor systems are still actively involved in cognitive processes such as language comprehension, categorization, and memory retrieval. Because representations are formed in sensory-motor brain areas, they should retain perceptual qualities. This hypothesis has been confirmed by the results of several studies. For example, Solomon and Barsalou (2001) asked participants to verify sentences such as a bee has wings. Earlier in the experiment the same property (e.g., wings) had been presented with a different entity. During this earlier presentation, the property either had a similar perceptual form (e.g., wasp-wings) or a different perceptual form (e.g., butterfly-wings). Performance on the property-verification task was better if the property had a similar form as in the earlier presentation than if it had a different form. Priming for shape similarity has also been observed in tasks that require less semantic processing, such as word naming and lexical decision (Pecher, Zeelenberg, & Raaijmakers, 1998), although the effect tends to be quite fragile in such tasks. Thus, overlap in perceptual features facilitates processing. This effect of shape similarity indicates that visual properties are part of the conceptual structure. This is corroborated by studies showing effects of visual perspective (Borghi, Glenberg, & Kaschak, 2004; Solomon & Barsalou, 2004; Wu & Barsalou, 2009). For example, Borghi et al. presented sentences that evoked an inside perspective (e.g., You are driving a car) or an outside perspective (e.g., You are washing a car). Following the sentence they presented the name of an object part (e.g., steering wheel, antenna). Borghi et al. found


221

that participants were faster to verify properties that were congruent than incongruent with the evoked perspective. This result indicates that participants simulated perceiving the object (e.g., car) from the evoked perspective, and used this simulation to perform the subsequent property verification. Properties that were more likely to be included in the simulation were verified faster than those that were less likely to be included in the simulation. In addition, information from different modalities can be more or less relevant for a task, which affects the composition of a particular representation. A couple of studies showed that there is a processing cost associated with switching the relevance from one sensory modality to another (Marques, 2006; Pecher et al., 2003; Vermeulen, Niedenthal, & Luminet, 2007). For example, participants were faster to verify sentences such as a banana is yellow (visual property) if on the previous trial they had verified that a gemstone is glittering (visual property) than if they had verified that leaves are rustling (auditory property). Studies such as these provide evidence that perceptual features are retained in concepts and strongly suggest that cognition uses the same systems as perception and action. In support of the behavioral findings described above, neuroscientific studies showed that modality-specific sensory brain areas are active during conceptual processing (for a review, see Martin, 2007). Verifying properties from different sensory modalities is associated with the activation of specific sensory brain areas. For example, verifying that a banana is yellow is associated with activation of the visual area, whereas verifying that a banana is sweet corresponds with activity in the gustatory area (e.g., Goldberg, Perfetti, & Schneider, 2006; Kan, Barsalou, Solomon, Minor, & Thompson-Schill, 2003). The retrieval of perceptual knowledge thus appears to rely on the activation of modality-specific brain regions that are associated with experiencing instances of the represented concepts. Because simulations involve processing in sensory-motor systems, the account also predicts that there should be direct interactions between representational processes and perceptual processes. An important line of evidence comes from studies that show effects of mental representation on processing of visual stimuli. In such studies, visual stimuli are typically better processed if their visual features overlap or match with those of the mental representation, For example, Stanfield and Zwaan (2001) presented sentences in which the horizontal or vertical orientation of an object was implied. For instance, the sentence John put the pencil in the cup implies a vertically oriented pencil, whereas the sentence John put the pencil in the drawer implies a horizontally oriented pencil. Immediately following the sentence a picture was presented, and the participant decided whether the object in the picture was mentioned in the preceding sentence. On the trials of interest, the depicted object was presented either in the implied orientation or in the perpendicular orientation. Participants were faster and more accurate to recognize the object when the orientation of the picture

222

Diane Pecher et al.

matched the orientation implied by the sentence than when it did not match. Similar results were obtained for overlap in shape (e.g., sliced vs. whole apple, Zwaan, Stanfield, & Yaxley, 2002) and color (brown vs. red steak, Connell, 2007). These compatibility effects even arise when representation and perception are separated by a delay (Pecher, van Dantzig, Zwaan & Zeelenberg, 2009) or when representation and perception only overlap in sensory modality and not in the specific concept that is simulated (Van Dantzig, Pecher, Zeelenberg, & Barsalou, 2008), which indicates that the effects are not due to strategic imagery. 1.3.2. The Role of Action In addition to perception, action plays an important role in representing concepts (Borghi, 2005). Since the main function of cognition is to support our interactions with the environment (Glenberg, 1997) action should be central to representations. This is supported by multiple demonstrations of interaction between representation and action. For example, in various language comprehension studies, participants are instructed to perform a particular motor action while reading words or sentences. The typical outcome of these studies is that participants read and verify a sentence faster when the performed motor action corresponds to the action described by the sentence. This effect is referred to as the action compatibility effect (ACE). For instance, participants are faster to verify the sentence You handed Jane the ball by making a hand movement away from themselves than by making a hand movement toward themselves. Oppositely, they are faster to verify the sentence Jane handed you the ball by moving the hand toward than away from themselves (Glenberg & Kaschak, 2002; see also Bub & Masson, 2010; Girardi, Lindemann, & Bekkering, 2010; Glenberg et al., 2008; Klatzky, Pellegrino, McCloskey, & Doherty, 1989; Masson, Bub, & Warren, 2008; Pezzulo, Barca, Bocconi, & Borghi, 2010; Van Elk, Van Schie, & Bekkering, 2009; Zwaan & Taylor, 2006). Furthermore, neuroscientific studies showed that motor areas are active while reading action words. Interestingly, a somatotopic organization is found, such that words referring to actions performed by different parts of the body (e.g., hand actions such as “grasp,” foot actions such as “kick,” mouth actions such as “kiss”) are associated with activations of different somatotopic areas of the motor cortex (Hauk, Johnsrude, & Pulvermu¨ller, 2004). To sum up, theories of grounded cognition propose that mental representations are formed by sensory-motor systems. On this account, concepts are simulations of interactive experiences in the real world. This view is supported by empirical studies that show that representations have similar features as perception and action. In addition, many studies show interactions between representation and sensory-motor tasks. Finally, neuroscientific studies indicate that cognition shares neural structures with perception and action.


223

1.4. Are All Concepts Grounded in Sensory-Motor Processing? The Scope Problem The studies presented in the previous sections demonstrate that the sensorymotor system is involved in many cognitive processes, such as conceptual processing and language comprehension. Based on this and related empirical evidence, we can conclude that cognition at least partially involves the systems of perception and action. Clearly, cognition is not completely amodal and symbolic. However, the question remains whether cognition can be completely grounded in perception and action. In other words, can all cognitive tasks be performed with analogue, sensory-motor representations, or do some cognitive tasks require more symbolic, abstract representations? This issue is called the scope problem by Machery (2007). He argues that findings of embodiment in a particular cognitive task cannot automatically be generalized to other tasks. According to Machery, “the interesting question is not, ‘Are concepts amodal or perceptual?’, but rather ‘To what extent do we use reenacted perceptual representations in cognition and to what extent do we use amodal representations?’” (Machery, 2007, p. 42). Some researchers take a wide-scope embodied standpoint on this issue, claiming that cognition does not need any amodal representations. For example, Barsalou used the following argument: “Perceptual symbol systems can implement all of the critical cognitive functions that amodal symbol systems have implemented traditionally. If so, then why did evolution add a redundant layer of amodal symbols?” (Barsalou, 1999a, pp. 75–76). However, others argue that amodal representations may be required for tasks that involve more abstract reasoning. For example, consider thinking about a rather abstract domain such as law or politics. According to Clark, [I]t is not clear how rich sensorimotor simulation could possibly account for all the kinds of moral and abstract reasoning required – reasoning about rights, implications, responsibilities, economics, and so on. (. . .) [I]t is hard to see how sensorimotor simulation could in principle account for all the kinds of thought and reasoning that the problem demands. (. . .). It does seem that the more decoupled and abstract the target contents become, either the less applicable the sensorimotor simulation strategy is, or the less clearly it can then be differentiated from the more traditional approaches it seeks to displace (Clark, 1999, p. 348).

Wilson (2002) has argued that cognitive theories can be placed on a continuum, ranging from a purely embodied account of cognition to a purely disembodied account. On one end of the continuum, pure embodied accounts claim that cognition is completely grounded in the sensorymotor system. Oppositely, pure disembodied accounts claim that cognition is completely symbolic and amodal. These opposite accounts define the ends of the continuum, leaving space for other theories to occupy the middle ground between the two extremes. Several of such intermediate

224

Diane Pecher et al.

theories have been proposed, each suggesting that cognition is partially based on sensory-motor processing and partially based on amodal, symbolic processing. For example, Mahon and Caramazza (2008) proposed the grounding by interaction framework. They hypothesized that concepts partially consist of amodal symbols, and partially of sensory-motor information. While the core of a concept is formed by amodal, symbolic information, sensory-motor information “colors conceptual processing, enriches it, and provides it with a relational context” (Mahon & Caramazza, 2008, p. 10). A related idea was proposed by Barsalou, Santos, Simmons, and Wilson (2008) (see also Simmons, Hamann, Harenski, Hu, & Barsalou, 2008), in their Language and Situated Simulation (LASS) theory. This theory assumes that concepts are represented in two distinct ways; by means of linguistic representations and by means of situated, sensory-motor simulations. During conceptual processing, both the linguistic system and the simulation system become active. The systems interactively contribute to the representation of concepts. The linguistic system is thought to underlie relatively superficial processing, whereas deeper conceptual processing requires engagement of the simulation system (see also Solomon & Barsalou, 2004). The scope problem is related to two questions; the necessity question and the sufficiency question (Fischer & Zwaan, 2008). The necessity question asks whether sensory-motor representations are necessary for cognitive processing. The sufficiency question asks whether sensory-motor representations are sufficient for cognitive processing. Whereas the pure embodied view claims that sensory-motor representations are necessary and sufficient for cognition, the disembodied view claims that sensory-motor representations are neither necessary nor sufficient. The intermediate theories suggest that sensory-motor representations are necessary for deep conceptual processing but not for shallow processing. Additionally, they argue that sensory-motor simulation may not be sufficient for cognition, but that it plays an important role by enriching mental representations and grounding them in experience.

2. Representing Abstract Concepts: Some Evidence for Grounding Are abstract concepts grounded in sensory-motor processes? Answering this question would shed some light on the scope problem. The view that sensory-motor grounding is necessary for cognitive processing predicts that such grounding should be consistently found for abstract concepts. Moreover, such a view predicts that in the absence of grounding full understanding of abstract concepts is impossible. Although the latter claim is very hard to test (it might entail removal of all sensory-motor systems), we


225

argue that consistent findings of sensory-motor effects for abstract concepts provides good indications that grounding is necessary. Because abstract entities by definition have no perceptual or motoric details, we should not expect sensory-motor effects to arise as a mere by-product of processing. Rather, such findings would indicate that sensory-motor processes are fundamental to the representation of abstract (and concrete) concepts.

2.1. Emotional Valence An important line of research that suggests that sensory-motor grounding may extend beyond representations of concrete objects and actions is the one investigating interactions between body and emotional valence. Unlike shape or color, the emotional valence of entities may not be perceived directly. Moreover, interaction effects are also obtained for words that refer to abstract concepts such as peace or hostility. Many studies demonstrated that positive and negative words automatically trigger approach or avoidance actions (e.g., Chen & Bargh, 1999; Lavender & Hommel, 2007; Rotteveel & Phaf, 2004; Solarz, 1960; Wentura, Rothermund, & Bak, 2000). This phenomenon, called the approach/avoidance effect, has been brought forward as an example of grounded cognition (e.g., Niedenthal, Barsalou, Winkielman, Krauth-Gruber, & Ric, 2005). In studies of this phenomenon, participants typically respond to stimuli with positive and negative valence by making an arm movement toward or away from the stimulus. Approach can be defined as the tendency to reduce the distance between a stimulus and the self, whereas avoidance can be defined as the tendency to increase the distance between a stimulus and the self. Pleasant stimuli automatically elicit approach tendencies, while unpleasant stimuli elicit avoidance tendencies. Accordingly, people are faster to respond to positive stimuli by making an approach movement (resulting in reducing the distance toward that stimulus) than by making an avoidance movement (resulting in increasing the distance). Conversely, they are faster to respond to negative stimuli by making an avoidance movement than by making an approach movement. For example, in the study by Chen and Bargh (1999), participants categorized positive and negative words by pulling a joystick toward themselves or pushing it away. In response to positive words, participants were faster to pull the joystick than to push it away. In response to negative words, in contrast, they were faster to push the joystick away than to pull it toward themselves. The effect is not located in the specific arm movement that has to be made, but rather seems to depend on the expected effect of an action. For example, approach/avoidance effects have even been found when participants do not move their arms with respect to the stimulus, but instead make button presses that result in an apparent movement of the stimulus toward or away from the participant (Van Dantzig, Pecher, & Zwaan, 2008).

226

Diane Pecher et al.

In order to investigate whether the approach/avoidance effect truly reflects embodied action, Markman and Brendl (2005) introduced a method in which the participant’s “self” was disembodied. They presented the participant’s name on the computer screen, and argued that this would lead to an abstracted representation of the “self” that was spatially dissociated from the participant’s real body. Markman and Brendl instructed participants to make valence judgments to words by moving a joystick in one of two directions, such that the word would move toward or away from the name. They obtained the usual congruency effect, but with respect to the participant’s name on the computer screen instead of their real body. According to Markman and Brendl, their study demonstrates that approach/avoidance actions are not executed with respect to the body, but with respect to a symbolic, disembodied representation of the self. They conclude that their results constrain grounded theories of cognition, because they show that some cognitive tasks require higher order symbolic representations. However, the results of a study by Van Dantzig, Zeelenberg, and Pecher (2009) (see also Proctor & Zhang, 2010) suggest that the findings of Markman and Brendl are the result of an artifact of their task, rather than a reflection of disembodied, symbolic processing. Van Dantzig et al. showed that similar approach/avoidance effects were obtained with respect to a meaningless rectangle on the computer screen, indicating that the effect was not due to an abstract representation of the self. Thus, the study of Markman and Brendl does not necessarily undermine grounded theories of cognition. These studies show that valence influences action. Effects in the opposite direction -motor actions influencing evaluations- have also been observed. Such effects are found even when participants are unaware of the relation between their action and the evaluated stimulus. In several studies participants received a misleading cover story instructing them to perform certain motor actions, which affected evaluations of unrelated stimuli. For example, valence judgments are more positive when participants are instructed to hold a pen in their mouth using only their teeth, such that their facial expression resembles a smile. Judgments are more negative when participants hold a pen only using their lips, such that their facial expression resembles a frown (Strack, Martin, & Stepper, 1988). Recognition of positive stimuli was better when participants were nodding their head, and better for negative stimuli when they were shaking their head (Fo¨rster & Strack, 1996). Extending the middle finger leads to higher hostility ratings, whereas pointing the thumb upward leads to higher liking ratings (Chandler & Schwarz, 2009). Conversely, when participants are prevented from making spontaneous facial expressions, they are worse at recognizing emotion (Oberman, Winkielman, & Ramachandran, 2007; Havas, Glenberg, Gutowski, Lucarelli, & Davidson, in press). These results provide evidence that representations of emotional valence are grounded in embodied interactions.


227

2.2. Abstract Transfer Evidence for sensory-motor grounding has also been observed for nonemotional concepts. There appears to be an interesting parallel between transfer of objects and transfer of information. The ACE described earlier was also found for sentences describing abstract transfer events (Glenberg & Kaschak, 2002; Glenberg et al., 2008). Similar to responses to sentences such as Jane handed you the ball in which the described event involves actual hand movements, sentences such as Liz told you the story also facilitated responses toward oneself, even though the described situation does not involve any specific hand movements. These results indicate that both concrete and abstract transfer events are represented by simulating concrete transfer, which provides sensory-motor grounding for abstract transfer.

2.3. Force Dynamics Another interesting parallel between the concrete and abstract domain is found in the representation of forces (Talmy, 1988). Talmy’s theory of force dynamics claims that many events, as well as linguistic descriptions of events, can be understood in terms of patterns of forces, both physical and psychosocial. According to this theory, event descriptions are understood in terms of naı¨ve physics and a perceived tendency of described entities to act or rest. In its simplest form, the theory supposes that each event consists of two force-exerting entities. The agonist is a focal entity with a tendency to act or to rest, while the antagonist is a secondary entity that exerts a force to oppose the agonist. The forces exerted by the agonist and antagonist arise through a drive (e.g., to eat, sleep, leave, say something) or a physical circumstance (gravitational force to fall or to stay still). Depending on the relative strengths of these two force-exerting entities, the focused agonist is either able to manifest its force tendency or is overcome by the antagonist. For example, the sentence the gate prevented the boy from leaving the yard can be understood by perceiving the boy as the focused entity (agonist) that has a tendency to leave the yard, and the gate as a stronger entity (antagonist) that instead forces the boy to stay (or rest, in Talmy’s terminology). Talmy shows that this basic idea of opposing agonists and antagonists yields various complex patterns that give rise to such semantic constructs as causation, allowing, and helping (see Wolff, 2007 for a model of causality that relies on force dynamics primitives). According to Talmy’s theory, concrete events (a player kicking the ball) are represented in the same manner as abstract events (a lawyer persuading the jury to convict the defendant). In the case of abstract sentences, the agonist may tend toward rest or action in less physical ways (jury tends toward rest in not convicting the defendant) while the antagonist (the lawyer) opposes this force. Again, whether or not the agonist results in carrying out its own force tendencies or those of the antagonist is

228

Diane Pecher et al.

determined by the relative strength of the agonist and antagonist. If the lawyer is strong enough, he will force the jury to act and convict the defendant. To test Talmy’s theory, Madden and Pecher (2010) presented sentences to participants for sensibility ratings. The sentences described one of four force patterns (force action, force rest, allow action, allow rest) in either a concrete situation (e.g., The bulldozer pushed the pile of dirt across the lot) or an abstract situation (e.g., Her friends persuaded the girl to come to the party). Before each sentence, an animation was shown in which two geometrical shapes interacted in a way that matched or mismatched the force pattern of the following sentence (e.g., a circle moving toward a rectangle, forcing it to topple). The results showed that sensibility ratings were faster and more accurate for sentences that matched with the animation than for sentences that mismatched with the animation, suggesting that prior activation of the correct force pattern made it easier to represent the meaning of the sentences. The advantage of match over mismatch was found for both abstract and concrete situations, indicating that representations of concrete and abstract forces are similar.

2.4. Summary The three cases described above, regarding the representation of emotional valence, abstract transfer events, and force dynamics, show that sensorymotor processing plays a role for some abstract concepts. In these cases, the sensory-motor effects are similar for abstract concepts and for their more concrete counterparts. For example, the approach/avoidance effect extends from concrete things that people want to physically avoid (e.g., a dangerous animal) to more abstract things that are physically harmless (e.g., the word hostility). Thus, these studies suggest that the representations underlying abstract and concrete concepts are at least partially similar. In the next section, we will compare theories that focus on fundamental distinctions between concrete and abstract concepts with theories that focus on similarities between concrete and abstract concepts.

3. Explanations of Abstract Concepts 3.1. Dual Code Theory While the Perceptual Symbols Theory and related frameworks advocate an extreme view where ultimately all representations are grounded in sensorymotor systems, there are others who propose a more moderate view. Mahon and Caramazza (2008) argued that the cognitive system might be only partially grounded in sensory-motor systems. They suggest that


229

sensory-motor information might be coactivated with other, more amodal information during representation, supporting representations through feedback rather than forming the entire basis for representations. Thus, representations themselves could still be largely amodal with only some grounding in sensory-motor systems (see also Dove, 2009). The idea that different representational systems exist is related to Paivio’s dual code theory (e.g., Paivio, 1991). This theory proposes that two independent systems underlie cognitive processing; a verbal and a nonverbal system. The nonverbal system uses sensory-motor analogues (called imagens) for representation, while the verbal system represents the visual, auditory, and perhaps motoric aspects of linguistic units such as words (called logogens). Connections between the systems exist such that nonverbal information can be named and conversely that names can evoke nonverbal experiences such as images. This theory explains why performance in memory tasks is often better for words that refer to concrete entities than for words that refer to abstract entities. According to the dual code theory, this effect occurs because concrete entities can be represented by both systems, whereas abstract concepts can only be represented by the verbal system. As a result, more retrieval cues can be used for concrete than abstract entities. It is important to note that the verbal system does not use abstract or amodal symbols such as propositions to represent verbal information. Rather, it represents the visual and auditory features of words and sentences. Therefore, it is unlikely that this system is suited to represent the meaning of abstract words (Barsalou, 2008). Rather, there may be associations between different verbal representations, without grounding these representations in meaningful experiences. Such associations might support correct performance in tasks that require only superficial activation of meaning, such as the free association task (Barsalou et al., 2008; Solomon & Barsalou, 2004), although Crutch, Connell and Warrington (2009) showed that even in tasks requiring deeper meaning abstract concepts rely more on associative relations whereas concrete concepts rely more on similarity relations.

3.2. Linguistic Co-occurrence Related to this verbal association idea are models based on statistical analyses of language such as LSA (Latent Semantic Analysis; Landauer & Dumais, 1997) and HAL (Hyperspace Analogue to Language; Burgess, 1998). These models represent word meaning as points in a high dimensional space based on their patterns of co-occurrence with other words. The basic principle is the extraction of meaning representations from large corpora of texts. For each word a frequency count is made of its occurrence in many different contexts. Contexts can be the other words within a window (as is done by the HAL model) or paragraphs of texts (as is done by LSA). These frequency counts can be put in vectors such that each word has a single vector

230

Diane Pecher et al.

containing all its frequency counts. This vector of co-occurrence values for that word can be viewed as a reference to a point in multidimensional space that corresponds to semantic memory. Words with highly similar meanings are closer in this semantic space than words with less similar meanings. It is important to note that vector similarity is not a measure of direct cooccurrence. For example, the words kitten and puppy may have very similar co-occurrence vectors, yet they may hardly occur together. Instead, they frequently occur in similar linguistic contexts. LSA can compute vectors for single words and for larger units such as sentences. The cosine between two vectors reflects the similarity of the vectors, and can be used as a measure of semantic similarity or text coherence. Landauer and Dumais (1997) suggest that language learning starts out by learning words in experiential contexts. By the time children have been in school for several years, however, most of their new words are acquired through text reading. According to these researchers, learning the meanings of these new words thus depends on knowing the meanings of other words. These co-occurrence models are quite successful at predicting human performance (Burgess & Lund, 2000) and have been presented as evidence that meaning can be represented by purely symbolic systems (Louwerse & Jeuniaux, 2008). The advantage of these models is that they do not need to distinguish between different types of words such as concrete and abstract, and they have tremendous computational power. One should be cautious in interpreting the co-occurrence vectors as meaningful, however, because it is not clear what they represent. The corpora that provide the input to these models are texts produced by people. Thus, they are measures of how people produce language, which reflects how people think. The abstract nature of the vector representation and the corpus it is based on has no bearing on whether people’s representations are modal or amodal. Thus, the fact that there is systematicity in how words relate to each other, and that two different measures of how people use language (e.g., the co-occurrence vectors and performance in experimental tasks) are related does not by itself explain how meaning is represented. Furthermore, although the co-occurrence vectors are derived from linguistic input, they might reflect the actual distribution of the words’ referents across real-world situations. After all, language typically refers to objects and situations in the world. Words might co-occur in language because their referents co-occur in the real world. For example, co-occurrence models consider the words kitten and puppy as similar because they often occur in similar linguistic contexts (e.g., sentences). However, the reason that they occur in similar sentences might be because these sentences typically describe real-world situations in which kittens and puppies occur. In the real world, kittens and puppies are often found in similar contexts (e.g., in and around the house). As a result, the sentences describing these contexts will be similar. Indeed, a recent study by Louwerse and Zwaan (2009) showed that the spatial distance between pairs


231

of US cities can be derived from their distributions across texts. In the words of Louwerse and Zwaan (2009), “cities that are located together are debated together.”

3.3. Hybrid Models Some researchers suggested that meaning may be partially grounded in sensory-motor processing and partially in relations between words. Andrews, Vigliocco, and Vinson (2009) developed a model in which sensory-motor information and linguistic information are treated as a single combined data set (for similar work, see Steyvers, in press). They used feature production norms as a proxy for sensory-motor experiences, and word co-occurrences as linguistic input. The word co-occurrence data were based on an approach similar to LSA. Andrews et al. argued that a probabilistic model that was based on a combination of linguistic and sensory-motor data was a better predictor of human performance than models based on either source of data alone or based on the two sources independently. Thus, instead of language and sensory-motor representations as being separate, parallel processes, Andrews et al. (2009) propose that they are part of the same network. Language binds different sensory-motor experiences to form coherent categories. During language processing, meaning is represented by a pattern of activation across sensory-motor systems. However, in tasks in which performance can be based on word associations, activation of other words might be sufficient. The advantage of such shortcuts is that the system does not need to wait for full activation of representations in the sensory-motor system. In order to fully represent meaning, however, sensory-motor representations are necessary. Such a mechanism was proposed by Barsalou et al. (2008) and Simmons et al. (2008). They argued that when linguistic materials are processed, linguistic associations are activated first. In turn, these linguistic representations very quickly activate sensorymotor simulations.

3.4. Abstract Concepts are Represented by Situations On such an account, the important issue still remains how sensory-motor systems can represent abstract concepts. Barsalou and Wiemer-Hastings (2005) (see also Barsalou, 1999b) proposed that in the case of abstract concepts, the specific situations in which those concepts occur as well as the introspective experience in response to those concepts might be simulated. Representations might be formed by collections of concrete situations that share the abstract concept. Such a mechanism is central to exemplar models of categorization (Hintzman, 1986; Nosofsky, 1988). In exemplar models, each experience with an exemplar is stored.

232

Diane Pecher et al.

Abstraction across exemplars is achieved when a cue activates several different exemplars. The combined activation as a response to the cue results in a kind of summary representation. In a similar way might a cue for an abstract concept activate several concrete situations in which that abstract concept played an important role. Representation by concrete situations allows these abstract concepts to be grounded in sensory-motor simulations. To investigate the content of abstract concepts, Barsalou and WiemerHastings asked participants to produce properties for concrete and abstract concepts. They found that introspections and situations were important aspects for all types of concepts, but even more for abstract than concrete concepts. The introspective and situational experiences can be simulated by sensory-motor systems because they have the perceptual and motoric details that abstract concepts in isolation lack. The role of situations is suggested by the context availability model proposed by Schwanenflugel and Shoben (1983; Schwanenflugel, Harnishfeger, & Stowe, 1988). They argue that in order to understand a word or sentence, people need to represent a context in which the linguistic material has meaning. They further argue that the main difference between concrete and abstract concepts is that it is more difficult to find an appropriate context for abstract than concrete materials. They presented sentences with and without contexts, and found that the difference in reading time between abstract and concrete materials that existed in the isolated condition disappeared when a context was provided. Barsalou and WiemerHastings (2005) argued that abstract concepts are used in a wider variety of contexts, and that typical events for abstract concepts are much more complex than those for concrete concepts. Barsalou (1999b, 2003) further proposed that abstract concepts might identify complex relations between physical and mental events. Such relations might be spatial, temporal, or causal. These relations are simulated by ‘filling in’ the regions that are connected by the relations with specific entities, events, or mental states. To sum up, abstract concepts might be grounded in sensory-motor simulations in an indirect way. Words for abstract concepts might activate specific, concrete situations that are instances of the concept or that provide a context for the concept. There is at present still very little evidence for this view, so more research will be needed in order to draw stronger conclusions.

3.5. Conceptual Metaphor Theory A very different proposal originated in cognitive linguistics, where researchers have suggested that abstract concepts are understood by way of analogy to representations of concrete, embodied experiences. For example, people may understand the process of solving a problem in terms of traveling from a starting point (the problem situation) to a destination (the solution) along a


233

path (the method that is used to solve the problem), as is illustrated by expressions such as to get sidetracked or to have something get in one’s way (Gibbs, 1994, 2005; Lakoff, 1987; Lakoff & Johnson, 1980, 1999; but see Barsalou & Wiemer-Hastings, 2005). In this view, the concrete concept is used as a metaphor (the ‘vehicle’) to represent the abstract concept (the ‘topic’). Thus, the concrete situation of traveling along a path provides a metaphor for the abstract situation of solving a problem. Lakoff and Johnson used the term conceptual mapping and conceptual metaphors because they argue that metaphorical mappings that are found in linguistic expressions (e.g., love is a journey) are conceptual and occur even beyond language comprehension. The basic claim of the conceptual metaphor theory is that the concrete vehicle partially structures the mental representation of the abstract topic and that the mental representation of the vehicle is necessary to fully understand the topic ( Johnson, 1987; Lakoff & Johnson, 1980, 1999). Because the metaphor’s vehicle refers to a concrete physical experience, conceptual metaphor theory can explain how the sensory-motor system can be used to represent the abstract concept. Lakoff and Johnson (1980) proposed that there is a set of basic concepts that are central to human experience. These basic concepts on which concrete concepts as well as embodied conceptual metaphors (or primary metaphors, Gibbs, 2006; Grady, 1997) are based have been called image schemas ( Johnson, 1987; Lakoff, 1987). These image schemas are concepts such as source-path-goal, containment, or balance. In general, they are analogue representations of mostly spatial relations and movements, although they are not sensory-motor representations themselves (Gibbs, 2006). Rather, image schemas give structure to experiences across modalities, and this multimodal nature makes them more abstract than actual sensory-motor representations. At the same time, they are viewed as grounded because they originate from sensory-motor experiences. Therefore, if such image schemas are grounded in sensory-motor experiences, and abstract concepts are partially represented by image schemas, one could argue that abstract concepts are ultimately grounded in sensory-motor experiences as well. It is important to note, however, that image schemas are abstractions from experience, so this logic relies on a two-step grounding operation. Lakoff and Johnson (see also Gibbs, 1994, 2005) thus argue that image schemas are fundamental to human experience and are used to understand more abstract concepts. They argue that this metaphorical mapping is reflected in language. Therefore, most evidence for this idea has come from analyses of linguistic expressions. Only recently have researchers started to collect experimental evidence for the metaphorical representation of abstract concepts, although some earlier indirect evidence comes from studies on metaphorical language comprehension. These earlier studies investigated whether image schemas were activated after people had read metaphorical statements. For instance, Allbritton, McKoon, and Gerrig

234

Diane Pecher et al.

(1995) obtained priming effects between words in a recognition memory test if these words were related to the same metaphor in a previously read text passage. Gibbs, Bogdanovich, Sykes, and Barr (1997) investigated priming effects for metaphor related words in a lexical decision task. They used metaphorical mappings such as life is a journey. For example, after reading the metaphorical sentence She was over the hill there was more priming for the word journey than after the literal sentence She was getting old. There was no difference in priming between the literal sentence and an unrelated control sentence. Thus, these studies support the idea that metaphorical sentences activate the concrete vehicle (but see Keysar, Shen, Glucksberg, & Horton, 2000; McGlone, 1996 for failures to find evidence for such activation). 3.5.1. Evidence for Conceptual Metaphors and Methodological Issues Most researchers who investigated the role of image schemas were interested in how people process metaphorical language, and did not directly address the representation of abstract concepts beyond metaphorical language. If metaphorical mappings are used only to comprehend metaphorical language, effects of such mapping should be found exclusively during processing of metaphorical sentences (Gibbs, 1992, 1994). When people comprehend metaphorical language, image schemas may be activated simply by words that also literally refer to the concrete vehicle (e.g, prices have risen). On the other hand, the conceptual metaphor theory claims that metaphorical mappings are used more generally to represent abstract concepts. Therefore, the image schema should be part of the representation itself, and activation of the image schema is needed for full understanding of the abstract concept. In order to support the idea that metaphorical mappings are conceptual rather than linguistic we need evidence that image schemas are also activated when no metaphorical language is used. Because the studies described above did not test this directly, they do not allow the conclusion that conceptual metaphors are used outside metaphorical language. Much of the evidence discussed in the previous section has used metaphorical language in some way. The conceptual metaphor theory, however, claims that metaphors are the expression of an underlying conceptual mapping. An interesting line of evidence comes from studies that show image schemas in people’s spontaneous gestures as they talk about abstract concepts (Cienki & Mu¨ller, 2008; Nu´n˜ez & Sweetser, 2006). Future studies should test whether such gestures are nonverbal expressions of the linguistic metaphors (e.g., someone gesturing ‘straight’ to express the linguistic metaphor straight is honest) or reflect embodied image schemas (e.g., the concept honesty is represented by straightness).


235

Spatial image schemas can be mapped on various abstract domains, such as valence (happy is up and sad is down), power (powerful is up and powerless is down), and divinity (god is up and the devil is down). For example, Meier, Hauser, Robinson, Friesen, and Schjeldahl (2007) investigated whether the up–down image schema is activated by the abstract concept divinity. They used god-like, devil-like, and neutral pictures. First, they asked participants to study the pictures that were presented at different, random locations on the screen. Following the study phase, they tested participants’ memory for the location of each picture. Their results showed that participants remembered god-like pictures at a higher location than neutral pictures and devillike pictures at a lower location than neutral pictures. Thus, even when no linguistic reference was made to the up–down image schema, it still affected spatial memory. Comparable results were found by Giessner and Schubert (2007), who investigated the relation between the up–down image schema and power. They demonstrated that the amount of power attributed to a manager was affected by spatial aspects, such as the vertical distance between the manager and his subordinates in an organizational chart. Participants read a brief description of a manager in a fictive organization and were shown a typical organizational tree-diagram in which the members of the organization were depicted as boxes connected by lines. The manager was judged as more powerful if the vertical line between the manager and his subordinates was longer than if it was shorter, all other factors (number of subordinates, leader’s position in the tree-diagram) remaining equal. Another study, by Casasanto and Boroditsky (2008), investigated the spatial representation of time. They presented moving dots and manipulated the spatial displacement and duration of the movement independently. When they asked participants to estimate the duration of the movement, time estimations were affected by the spatial distance that the dot had traveled. This result suggests that people use the mental representation of space in order to fully understand time. An important question regarding these studies is whether the image schema plays a role during representation of the abstract concept or during selection of the response (Boot & Pecher, 2010). There is some evidence that irrelevant information can sometimes affect responses in situations with high uncertainty. For example, participants’ judgments of justice in conditions of uncertainty (e.g., when there is no information about the others’ outcome) was influenced by irrelevant information (e.g., affect, Van den Bos, 2003; Van den Bos, Lind, Vermunt, & Wilke, 1997). In studies investigating activation of image schemas during processing of abstract concepts a similar type of uncertainty might play a role as well. When participants have to choose from many response options (e.g., duration of a stimulus, power of an unfamiliar person) and are uncertain about the accuracy of their choice, irrelevant information may affect responses. For example, in Giessner and Schubert’s study, participants had no idea how

236

Diane Pecher et al.

powerful the person in the chart really was. It is possible that the irrelevant information (position in the chart) might have influenced the response selection in congruence with the metaphor (e.g., people presented at the top are more powerful than people presented at the bottom). In sum, while it is obvious that the conceptual metaphor (e.g., power is up) was active during performance, it is unclear whether the image schema affected the representation of the person’s power or the response selection process. There are some studies in which the image schema was unlikely to have affected response selection. Several researchers have found that identification of a stimulus (e.g., a p or q) at the top or bottom of the screen is affected by a prior event in which a word had to be judged on an abstract quality such as valence (Meier & Robinson, 2004) or power (Van Dantzig, 2009; Zanolie et al., 2010) or even when participants listened to sentences containing abstract verbs such as respect or argue (Richardson, Spivey, Barsalou, & McRae, 2003; Spivey, Richardson, & Gonzalez-Marquez, 2005, but see Bergen, Lindsay, Matlock, & Narayanan, 2007). Van Dantzig asked participants to judge whether a word (e.g., president, slave) presented in the center of the screen represented a powerful or powerless entity. Immediately following the power judgment a letter was presented at the top or bottom of the screen, and participants identified this letter as quickly as possible. Because the target response (the letter identity) was unrelated to the image schema, it is highly unlikely that participants used the image schema in order to select their response. Rather, making the power judgment directed attention to the congruent spatial location (up for powerful and down for powerless). When the subsequent letter appeared in the attended location response was facilitated compared to when it was presented in the unattended location. Using the same paradigm in an EEG study, Zanolie et al. measured ERPs (Event Related Responses, a measure of the brain’s electrical signal in response to an event such as presentation of a stimulus) to the letter targets. They found an early difference in brain activation, the N1, which is a negative deflection occurring at 160–200 ms after stimulus presentation. This component has been observed in many studies of spatial attention, and is assumed to reflect an enhancement of stimuli at the attended location (Luck & Hillyard, 1995; Mangun & Hillyard, 1991). This study thus provides further evidence that the image schema influenced spatial attention. Studies that do not require participants to select a response provide additional evidence that the effects of image schemas are not due to uncertainty. Teuscher, McQuire, Collins, and Coulson (2008) obtained evidence for the automatic activation of a spatial image schema when participants mentally represented time. They presented sentences that described events involving space or time. On half of the trials, the sentence had an ego-moving perspective (She approached the buffalo, We got near the day of the exam) and the other half had an object-moving perspective (The buffalo


237

approached her, The day of the exam was getting closer). Following the sentence they presented a cartoon of a smiley face and a geometric object that represented a congruent or incongruent spatial movement. Rather than asking participants to respond to the cartoons they only asked them to view the cartoons while their EEG was measured. The results showed effects of congruency on the ERPs that were time-locked to the cartoons, with effects of spatial congruency affecting an earlier portion of the ERP waveform for descriptions of events in space than for descriptions of events in time. Finally, several studies have shown effects of image schemas on processing of abstract concepts using tasks that were so easy that it was very unlikely that participants would have used the task-irrelevant image schema in a strategic way. Boot and Pecher (2010) investigated the role of distance for the representation of similarity. They asked participants to make similarity judgments to colored squares that were presented at various distances from each other. The colors were either very similar or very different, making the task quite easy to perform. Still, an interaction effect was found between distance and similarity. Specifically, for similar colors, responses were faster when the spatial distance between the colored squares was small than when it was large, whereas the opposite was found for dissimilar colors (see Casasanto, 2008, for interactions between distance and similarity in more difficult tasks). In another study, Boot and Pecher (in press) investigated the role of the container image schema for the representation of categories. The idea was that categories can be represented as containers (Lakoff, 1987). Things that belong to a category are represented as being inside the container, while things that do not belong to the category are represented as being outside the container. They asked participants to categorize pictures of animals and vehicles. Crucially, pictures could be presented inside or outside a visual frame. Category decisions were faster when pictures were presented congruent with the image schema than when they were presented incongruent with the schema. Again, the task was so easy that it was unlikely that participants used the task-irrelevant frame for response selection. To sum up, many studies have obtained evidence that image schemas play a role in representations of abstract concepts. Effects have been found in a variety of tasks for several abstract concepts. Moreover, several studies have found such effects without using metaphorical language, thus providing evidence for the claim that metaphorical mappings are conceptual rather than linguistic. 3.5.2. Evaluation of Conceptual Metaphor Theory An important issue for theories on grounded cognition is whether abstract concepts automatically activate a concrete vehicle or image schema. The evidence presented in the previous section seems to support this view.

238

Diane Pecher et al.

However, there are also findings that may be more problematic. For example, McGlone (1996) failed to find evidence for activation of the underlying vehicle when subjects were asked to paraphrase metaphors or to come up with another metaphor for the same concept. Keysar et al. (2000) found that processing of sentences containing expressions of metaphorical mappings (e.g., the journey should not last too long for the metaphor argument is a journey) was not facilitated by earlier conventional expressions of the same metaphorical mapping (e.g., point out your position), even when the mapping was explicitly mentioned (e.g., think of an argument as a journey). They did find facilitation when novel expressions of the same metaphor (e.g., parking at a compromise) were used. These results question the assumption of conceptual metaphor theory that metaphorical mappings are activated automatically when people represent the topic of the metaphorical expression. Rather, metaphors may be understood as categorizations (Glucksberg & Keysar, 1990) or may evolve from explicit analogies when they are novel to categorizations or polysemies (i.e., words that have multiple related meanings) when they have become conventional (Bowdle & Gentner, 2005). According to the category view, a metaphorical sentence (e.g., lawyers are sharks) can be understood when a post-hoc category is formed by selecting relevant features and deselecting irrelevant features (e.g., vicious, predatory but not fast swimmer). This selection process, and thus formation of the post-hoc category, can only occur when both the topic and vehicle are activated. According to the analogical view, a metaphorical sentence (e.g., lawyers are sharks) can be understood when the relational structure (e.g., harming others) of the concrete vehicle (e.g., shark) and topic (e.g., lawyer) are aligned. Indeed, being able to see structural similarities might be at the very core of representing and understanding abstract concepts (Gentner, 2003). This alignment process can only occur when both the topic and vehicle are active. Similarly, Coulson and Van Petten (2002; see also Coulson & Matlock, 2001) propose that comprehension of metaphorical mappings requires blending of the topic and the vehicle in a blended space. Such blending is a process by which a subset of features from the vehicle and topic are combined to form a new representation. In order for this to happen, vehicle and topic first need to be represented separately. Coulson and colleagues showed that this process is similar for metaphorical (her voice was sweet syrup) and literal blends (in the movie Psycho, the blood was really cherry syrup). The literal blends were slightly easier to process than the metaphorical blends, but both were harder than simple literal meanings (one of Canada’s major exports is maple syrup). Giora (2002) also noted that there is no fundamental difference between literal and metaphorical meanings. Rather, she posits that saliency (e.g., frequency, contextual fit) determines which meaning is activated first, the literal or metaphorical. On the other hand, according to the conceptual metaphor theory, the concrete vehicle partly structures the abstract topic. Because representation


239

of the topic depends on the vehicle, mapping of the vehicle on the topic domain should occur before the topic is fully represented. Such mapping can only work for conventional metaphors because these are already part of the abstract concept. Therefore, novel metaphor comprehension should rely on a different mechanism such as analogy or blending. Another important difference between novel and conventional metaphors is that novel metaphors may be used deliberately by the speaker in order to force the listener to see a concept in a new perspective, whereas conventional metaphors do not have this communicative goal (Steen, 2008). In contrast, conventional metaphors might reflect conceptual mappings. Some of the experiments described earlier indicate that metaphorical language can lead to the activation of image schemas, although the evidence was mixed. Lakoff and Johnson’s claim is much stronger, however, because it states that people use metaphors to understand concepts. To prove this claim one needs to show not only that abstract concepts activate image schemas, but rather that activation of an image schema precedes full understanding of the abstract concept. For example, processing of the sentence Our relationship was at a crossroads may activate the source-path-goal (or journey) schema. What must be shown, however, is that the source-path-goal schema needs to be activated in order to understand the concept love. In other words, the source-path-goal schema should be necessary for the representation of love. Thus, the image schema should be activated irrespective of whether the abstract concept is presented in a metaphorical or literal sentence. Commenting on this issue, Murphy (1996) argued that the view of conceptual metaphor theory can make strong or weak claims as to how fundamental metaphorical mapping is to the representation of concepts. In the strong view, the abstract concept does not have its own structure but derives all of its structure from the concrete concept. A serious problem with the strong view is that if the abstract concept does not have a structure of its own, it is impossible to interpret the metaphorical relation with the concrete concept. For example, which features of journey overlap with love must be determined by the similarities between the two concepts. How else would someone be able to understand that the statement This marriage has been a long bumpy road means that the relationship has had difficulties but not that it is made of tarmac or has a line down the middle. In order to determine the similarities, however, both concepts need to already have a structured representation, and therefore, Murphy argues, conceptual mapping cannot fully account for the representation of abstract concepts. The weaker version of the conceptual metaphor theory does not hold that representations receive their full structure from conceptual mappings. Rather, the abstract concept has its own structure, but this structure is influenced by the conceptual mapping. It is unclear whether such influence is the result of the metaphorical expression (e.g., love becomes more like a journey because of the way it is talked about) or whether the metaphorical

240

Diane Pecher et al.

expressions are the result of the structural similarity between the abstract topic and the concrete vehicle (e.g., because love and journey share structure people have developed metaphorical expressions that reveal such similarity; Murphy, 1996). Proponents of the conceptual metaphor theory, however, have claimed that the systematicity of image schema mappings for abstract concepts as expressed by metaphorical language provides evidence that the metaphorical mappings are conceptual rather than linguistic. They argue that language reflects the underlying representational structure of concepts, and the systematic use of metaphorical mappings indicates that the mapping must be part of the concept itself. Such a framework predicts that image schemas should be activated whenever the abstract concepts are activated, even when people are not using metaphoric language. In other words, thinking about abstract concepts should activate image schemas directly, and not via language. Language is merely the expression of such activation, not the cause. The studies using nonlinguistic stimuli that we presented in the previous section provide some evidence that image schemas are activated when abstract concepts are processed nonlinguistically.

4. Discussion We started this chapter by asking how abstract concepts might be grounded in sensory-motor processing. Two grounded views have been proposed; representation by concrete situations and the conceptual metaphor view. Whereas only a few studies have investigated the idea that abstract concepts are represented by concrete situations, many more have investigated the role of metaphor or image schemas. As the evidence reviewed above shows, image schemas are relevant for abstract concepts. Image schemas are activated quickly and automatically even when people are performing simple unambiguous tasks that do not involve metaphoric language. Having established that image schemas play a role, we should now turn to the question of what that role is. One of the big challenges for applying conceptual metaphor theory to mental representation is clarifying the process of mapping image schemas onto abstract concepts. This turns out to be nontrivial. One of the main problems is that the concept of image schemas is still shrouded in mystery, and the term has been used to mean quite different things. Hampe (2005; see also Gibbs, 2005), for example, noted that there is no set of clear-cut criteria that distinguishes image schemas from other types of mental representations. Gibbs (2006) described image schemas as embodied but at the same time abstracted from modalityspecific sensory activation, whereas Grady (2005) defined image schemas as mental representations of concrete sensory experience. Dodge and Lakoff


241

(2005) argued that image schemas are categories of experiences for which the same words are used (e.g., high is used for spatial position, quantity, power, divinity, etc.). Mandler (2005), however, viewed image schemas as the earliest concepts that infants acquire. These image schemas do not have perceptual properties, but at the same time they are preverbal. Gentner (2003) argued that image schemas are abstract rather than concrete. Although this short summary of views is not exhaustive, it illustrates that the idea of an image schema is not well specified. The one thing that researchers seem to agree on is that image schemas apply to embodied experiences in different sensory modalities. These image schemas can then be extended to more abstract concepts. Thus, image schemas reflect commonalities between distinct, recurrent experiences in different domains and modalities. The spatial component is an important aspect of image schemas. For example, the up–down, source-path-goal, container, and distance image schemas all refer to spatial experiences, whether they are visual, tactile, or motoric. Image schemas thus may be important for binding perception from different modalities and action together. As an example, in order to have coherent experiences with verticality, the image schema binds visual experiences, proprioceptive experiences, and motor actions so that the representation of visual vertical orientation can be similar to the representation of feeling vertical orientation and the representation of acting in vertical orientations (similar to the idea of event coding, Hommel, 2004). Image schemas thus may be skeletal, providing a basic structure, to which the flesh (specific perceptual detail, valence, and so on) is added by particular meaning and context (e.g., Gibbs, 2006; Johnson, 2005). On this account, image schemas are not identical to concrete concepts, because they lack the perceptual detail of the latter. Rather, they might be viewed as features that provide an element of structure to both concrete and abstract concepts. Experiencing verticality in a particular modality may reactivate previous experiences with verticality in other modalities. We can even extend this mechanism to language. Words that refer to verticality might get bound to physical experiences of verticality because they are used in the same situation. Image schemas may thus reflect associations between words and experiences in different modalities. An important question that remains, however, regards the origin of the connection between image schemas and abstract concepts. This may be a chicken and egg problem. The relation between an image schema and an abstract concept might be formed during early experiences in which the two domains are conflated ( Johnson, 1997). For example, early experiences with power typically involve a perception of verticality, such as when a child looks up to a more powerful adult or taller child. These frequent experiences might underlie the relation between the up–down image schema and power. Alternatively, metaphorical language may have activated representations of physical experiences, which then created a link

242

Diane Pecher et al.

between concrete and abstract domains. For example, expressions such as high power or low status may have activated concrete vertical simulations, thus leading to representations in which power and verticality were combined. Although this view does not explain the historic origin of metaphorical expressions, for individuals it might work because they are exposed to metaphorical language during development. A third possibility is that the image schema and abstract concept have completely separate representations, and people may construct the metaphorical mapping on the fly as an analogy or conceptual blend. In order to fully represent abstract concepts, however, image schemas are not sufficient. The fact that the up–down schema can be mapped on several different concepts such as power, valence, divinity, and quantity shows that more features are needed to distinguish these concepts, just like the concepts blood, apple, traffic light, and carpet all share the feature red, but this feature is not sufficient to represent these different concepts. The up–down image schema, for example, indicates that concepts all have some scale on which things can be compared (e.g., powerful vs. powerless), but for full understanding more is needed. Situations or events might provide such additional representational power. Representations of concrete situations and events may provide a full spectrum of perceptual, motoric, and affective experiences (Barsalou & Wiemer-Hastings, 2005). Except for psychological experiments using single words, abstract concepts are hardly ever processed in isolation. Rather, a concept such as power is represented because there is a specific entity (e.g., mother, boss, country) that has power over another specific entity (e.g., child, employee, other country) making it do specific things (e.g., clean up, write a report, make budget cuts). Image schemas provide similar structure to such diverse situations, but the actual details provide the meaning. On this account, image schemas are not the final solution to the grounding problem for abstract concepts. The pervasiveness of image schemas in processing of abstract concepts shows that similarities between concrete and abstract domains are central to understanding. However, the full grounding of abstract concepts in sensorymotor experience needs an additional step. Most likely, the rich perceptual, motoric, and evaluative details of specific situations provide such grounding.

ACKNOWLEDGMENTS This research was supported by grants from the Netherlands Organization for Scientific Research (NWO) to Diane Pecher. We are grateful to Larry Barsalou and Gerard Steen for stimulating discussions and Seana Coulson for her very helpful comments on an earlier version of this manuscript.


243

REFERENCES Allbritton, D. W., McKoon, G., & Gerrig, R. J. (1995). Metaphor-based schemas and text representations: Making connections through conceptual metaphors. Journal of Experimental Psychology: Learning, Memory, and Cognition, 21, 612–625. Anderson, R. C., Pichert, J. W., Goetz, E. T., Schallert, D. L., Stevens, K. V., & Trollip, S. R. (1976). Instantiation of general terms. Journal of Verbal Learning and Verbal Behavior, 15, 667–679. Andrews, M., Vigliocco, G., & Vinson, D. (2009). Integrating experiential and distributional data to learn semantic representations. Psychological Review, 116, 463–498. Barsalou, L. W. (1982). Context-independent and context-dependent information in concepts. Memory & Cognition, 10, 82–93. Barsalou, L. W. (1993). Flexibility, structure, and linguistic vagary in concepts: Manifestations of a compositional system of perceptual symbols. In A. F. Collins, S. E. Gathercole, M. A. Conway, & P. E. Morris (Eds.), Theories of memory (pp. 29–101). Hove, UK: Erlbaum. Barsalou, L. W. (1999a). Language comprehension: Archival memory or preparation for situated action? Discourse Processes, 28, 61–80. Barsalou, L. W. (1999b). Perceptual symbol systems. Behavioral and Brain Sciences, 22, 577–660. Barsalou, L. W. (2003). Abstraction in perceptual symbol systems. Philosophical Transactions of the Royal Society B: Biological Sciences, 358, 1177–1187. Barsalou, L. W. (2008). Grounding symbolic operations in the brain’s modal systems. In G. R. Semin & E. R. Smith (Eds.), Embodied grounding: Social, cognitive, affective, and neuroscientific approaches (pp. 9–42). New York, NY, US: Cambridge University Press. Barsalou, L. W., Santos, A., Simmons, W. K., & Wilson, C. D. (2008). Language and simulation in conceptual processing. In M. De Vega, A. M. Glenberg, & A. C. Graesser (Eds.), Symbols, embodiment, and meaning (pp. 245–283). Oxford: Oxford University Press. Barsalou, L. W., & Wiemer-Hastings, K. (2005). Situating abstract concepts. In D. Pecher & R. A. Zwaan (Eds.), Grounding cognition: The role of perception and action in memory, language, and thinking (pp. 129–163). Cambridge: Cambridge University Press. Bergen, B. K., Lindsay, S., Matlock, T., & Narayanan, S. (2007). Spatial and linguistic aspects of visual imagery in sentence comprehension. Cognitive Science, 31, 733–764. Boot, I., & Pecher, D. (2010). Similarity is closeness: Metaphorical mapping in a perceptual task. Quarterly Journal of Experimental Psychology, 63, 942–954. Boot, I. & Pecher, D. (in press). Representation of categories: Metaphorical use of the container schema. Experimental Psychology. Borghi, A. M. (2005). Object concepts and action. In D. Pecher & R. A. Zwaan (Eds.), Grounding cognition: The role of perception and action in memory, language, and thinking (pp. 8–34). Cambridge: Cambridge University Press. Borghi, A. M., Glenberg, A. M., & Kaschak, M. P. (2004). Putting words in perspective. Memory & Cognition, 32, 863–873. Bowdle, B. F., & Gentner, D. (2005). The career of metaphor. Psychological Review, 112, 193–216. Bub, D. N., & Masson, M. E. J. (2010). Grasping beer mugs: On the dynamics of alignment effects induced by handled objects. Journal of Experimental Psychology: Human Perception and Performance, 36, 341–358. Burgess, C. (1998). From simple associations to the building blocks of language: Modeling meaning in memory with the HAL model. Behavior Research Methods, Instruments, & Computers, 30, 188–198. Burgess, C., & Lund, K. (2000). The dynamics of meaning in memory. In E. Dietrich & A. B. Markman (Eds.), Cognitive dynamics: Conceptual and representational change in humans and machines (pp. 117–156). Mahwah, NJ: Lawrence Erlbaum Associates.

244

Diane Pecher et al.

Casasanto, D. (2008). Similarity and proximity: When does close in space mean close in mind? Memory and Cognition, 36, 1047–1056. Casasanto, D., & Boroditsky, L. (2008). Time in the mind: Using space to think about time. Cognition, 106, 579–593. Chandler, J., & Schwarz, N. (2009). How extending your middle finger affects your perception of others: Learned movements influence concept accessibility. Journal of Experimental Social Psychology, 45, 123–128. Chen, M., & Bargh, J. A. (1999). Consequences of automatic evaluation: Immediate behavioral predispositions to approach or avoid the stimulus. Personality and Social Psychology Bulletin, 25, 215–224. Cienki, A., & Mu¨ller, C. (2008). Metaphor, gesture, and thought. In R. W. J. Gibbs (Ed.), The Cambridge handbook of metaphor and thought (pp. 483–501). Cambridge: Cambridge University Press. Clark, A. (1999). An embodied cognitive science? Trends in Cognitive Sciences, 3, 345–351. Connell, L. (2007). Representing object colour in language comprehension. Cognition, 102, 476–485. Coulson, S., & Matlock, T. (2001). Metaphor and the space structuring model. Metaphor and Symbol, 16, 295–316. Coulson, S., & Van Petten, C. (2002). Conceptual integration and metaphor: An eventrelated potential study. Memory and Cognition, 30, 958–968. Crutch, S. J., Connell, S., & Warrington, E. K. (2009). The different representational frameworks underpinning abstract and concrete knowledge: Evidence from odd-oneout judgements. Quarterly Journal of Experimental Psychology, 62, 1377–1390. Dodge, E., & Lakoff, G. (2005). Image schemas: From linguistic analysis to neural grounding. In B. Hampe (Ed.), From perception to meaning (pp. 57–91). Berlin/New York: Mouton de Gruyter. Dove, G. (2009). Beyond perceptual symbols: A call for representational pluralism. Cognition, 110, 412–431. Fischer, M. H., & Zwaan, R. A. (2008). Embodied language: A review of the role of the motor system in language comprehension. Quarterly Journal of Experimental Psychology, 61, 825–850. Fo¨rster, J., & Strack, F. (1996). Influence of overt head movements on memory for valenced words: A case of conceptual-motor compatibility. Journal of Personality and Social Psychology, 71, 421–430. Gentner, D. (2003). Why we’re so smart. In D. Gentner & S. Goldin-Meadow (Eds.), Language in mind: Advances in the study of language and thought (pp. 195–235). Cambridge, MA: MIT Press. Gibbs, R. W. J. (1992). Categorization and metaphor understanding. Psychological Review, 99, 572–577. Gibbs, R. W. J. (1994). The poetics of mind: Figurative thought, language, and understanding. New York: Cambridge University Press. Gibbs, R. W. J. (2005). The psychological status of image schemas. In B. Hampe (Ed.), From perception to meaning (pp. 113–135). Berlin/New York: Mouton de Gruyter. Gibbs, R. W. J. (2006). Embodiment and cognitive science. Cambridge: Cambridge University Press. Gibbs, R. W. J., Bogdanovich, J. M., Sykes, J. R., & Barr, D. J. (1997). Metaphor in idiom comprehension. Journal of Memory and Language, 37, 141–154. Giessner, S. R., & Schubert, T. W. (2007). High in the hierarchy: How vertical location and judgments of leaders’ power are interrelated. Organizational Behavior and Human Decision Processes, 104, 30–44. Giora, R. (2002). Literal vs. figurative language: Different or equal? Journal of Pragmatics, 34, 487–506.


245

Girardi, G., Lindemann, O., & Bekkering, H. (2010). Context effects on the processing of action-relevant object features. Journal of Experimental Psychology: Human Perception and Performance, 36, 330–340. Glenberg, A. M. (1997). What memory is for. Behavioral and Brain Sciences, 20, 1–55. Glenberg, A. M., & Kaschak, M. P. (2002). Grounding language in action. Psychonomic Bulletin & Review, 9, 558–565. Glenberg, A. M., Sato, M., Cattaneo, L., Riggio, L., Palumbo, D., & Buccino, G. (2008). Processing abstract language modulates motor system activity. Quarterly Journal of Experimental Psychology, 61, 905–919. Glucksberg, S., & Keysar, B. (1990). Understanding metaphorical comparisons: Beyond similarity. Psychological Review, 97, 3–18. Goldberg, R. F., Perfetti, C. A., & Schneider, W. (2006). Perceptual knowledge retrieval activates sensory brain regions. The Journal of Neuroscience, 26, 4917–4921. Grady, J. E. (1997). A topology of motivation for conceptual metaphor: Correlation vs. resemblance. In R. W. J. Gibbs & G. J. Steen (Eds.), Metaphor in cognitive linguistics. Amsterdam/Philadelphia: John Benjamins publishing company. Grady, J. E. (2005). Image schemas and perception: Refining a definition. In B. Hampe (Ed.), From perception to meaning (pp. 35–55). Berlin/New York: Mouton de Gruyter. Grush, R. (2004). The emulation theory of representation: Motor control, imagery, and perception. Behavioral and Brain Sciences, 27, 377–396. Hampe, B. (2005). Image schemas in cognitive linguistics: Introduction. In B. Hampe (Ed.), From perception to meaning (pp. 1–13). Berlin/New York: Mouton de Gruyter. Harnad, S. (1990). The symbol grounding problem. Physica D, 42, 335–346. Hauk, O., Johnsrude, I., & Pulvermu¨ller, F. (2004). Somatotopic representation of action words in human motor and premotor cortex. Neuron, 41, 301–307. Havas, D. A., Glenberg, A. M., Gutowski, K. A., Lucarelli, M. J., & Davidson, R. J. (in press). Cosmetic use of botulinum toxin affects processing of emotional language. Psychological Science, 21, 895–900. Hintzman, D. L. (1986). Schema abstraction in a multiple-trace memory model. Psychological Review, 93, 411–428. Hommel, B. (2004). Event files: Feature binding in and across perception and action. Trends in Cognitive Sciences, 8, 494–500. Johnson, M. (1987). The body in the mind: The bodily basis of meaning, imagination, and reason. Chicago: The University of Chicago Press. Johnson, C. (1997). Metaphors vs. conflation in the acquisition of polysemy: The case of SEE. In M. K. Hiraga, C. Sinha, & S. Wilcox (Eds.), Cultural, psychological and typological issues in cognitive linguistics. Current issues in linguistic theory, Vol. 152. Amsterdam: John Benjamins. Johnson, M. (2005). The philosophical significance of image schemas. In B. Hampe (Ed.), From perception to meaning (pp. 15–33). Berlin/New York: Mouton de Gruyter. Kan, I. P., Barsalou, L. W., Solomon, K. O., Minor, J. K., & Thompson-Schill, S. L. (2003). Role of mental imagery in a property verification task: FMRI evidence for perceptual representations of conceptual knowledge. Cognitive Neuropsychology, 20, 525–540. Keysar, B., Shen, Y., Glucksberg, S., & Horton, W. S. (2000). Conventional language: How metaphorical is it? Journal of Memory and Language, 43, 576–593. Klatzky, R. L., Pellegrino, J. W., McCloskey, B. P., & Doherty, S. (1989). Can you squeeze a tomato? the role of motor representations in semantic sensibility judgments. Journal of Memory and Language, 28, 56–77. Lakoff, G. (1987). Women, fire, and dangerous things. Chicago: Chicago University Press. Lakoff, G., & Johnson, M. (1980). Metaphors we live by. Chicago: Chicago University Press. Lakoff, G., & Johnson, M. (1999). Philosophy in the flesh: The embodied mind and its challenge to western thought. New York: Basic Books.

246

Diane Pecher et al.

Landauer, T. K., & Dumais, S. T. (1997). A solution to plato’s problem: The latent semantic analysis theory of acquisition, induction, and representation of knowledge. Psychological Review, 104, 211–240. Lavender, T., & Hommel, B. (2007). Affect and action: Towards an event-coding account. Cognition and Emotion, 21, 1270–1296. Louwerse, M. M., & Jeuniaux, P. (2008). Language comprehension is both embodied and symbolic. In M. De Vega, A. M. Glenberg, & A. C. Graesser (Eds.), Embodiment and meaning: A debate (pp. 309–326). Oxford: Oxford University Press. Louwerse, M. M., & Zwaan, R. A. (2009). Language encodes geographical information. Cognitive Science: A Multidisciplinary Journal, 33, 51–73. Luck, S. J., & Hillyard, S. A. (1995). The role of attention in feature detection and conjunction discrimination: An electrophysiological analysis. International Journal of Neuroscience, 80, 281–297. Machery, E. (2007). Concept empiricism: A methodological critique. Cognition, 104, 19–46. Madden, C., & Pecher, D. (2010). The force behind language: Are concrete and abstract sentences understood in terms of underlying force patterns? (Submitted for publication). Mahon, B. Z., & Caramazza, A. (2008). A critical look at the embodied cognition hypothesis and a new proposal for grounding conceptual content. Journal of Physiology Paris, 102, 59–70. Mandler, J. M. (2005). How to build a baby: III. In B. Hampe (Ed.), From perception to meaning (pp. 137–163). Berlin/New York: Mouton de Gruyter. Mangun, G. R., & Hillyard, S. A. (1991). Modulations of sensory-evoked brain potentials indicate changes in perceptual processing during visual-spatial priming. Journal of Experimental Psychology: Human Perception and Performance, 17, 1057–1074. Markman, A. B., & Brendl, C. M. (2005). Constraining theories of embodied cognition. Psychological Science, 16, 6–10. Marques, J. F. (2006). Specialization and semantic organization: Evidence for multiple semantics linked to sensory modalities. Memory & Cognition, 34, 60–67. Martin, A. (2007). The representation of object concepts in the brain. Annual Review of Psychology, 58, 25–45. Masson, M. E. J., Bub, D. N., & Warren, C. M. (2008). Kicking calculators: Contribution of embodied representations to sentence comprehension. Journal of Memory and Language, 59, 256–265. McGlone, M. S. (1996). Conceptual metaphors and figurative language interpretation: Food for thought? Journal of Memory and Language, 35, 544–565. Meier, B. P., Hauser, D. J., Robinson, M. D., Friesen, C. K., & Schjeldahl, K. (2007). What’s “up” with god? Vertical space as a representation of the divine. Journal of Personality and Social Psychology, 93, 699–710. Meier, B. P., & Robinson, M. D. (2004). Why the sunny side is up: Associations between affect and vertical position. Psychological Science, 15, 243–247. Murphy, G. L. (1996). On metaphoric representation. Cognition, 60, 173–204. Niedenthal, P. M., Barsalou, L. W., Winkielman, P., Krauth-Gruber, S., & Ric, F. (2005). Embodiment in attitudes, social perception, and emotion. Personality and Social Psychology Review, 9, 184–211. Nosofsky, R. M. (1988). Similarity, frequency, and category representations. Journal of Experimental Psychology: Learning, Memory, and Cognition, 14, 54–65. Nu´n˜ez, R. E., & Sweetser, E. (2006). With the future behind them: Convergent evidence from aymara language and gesture in the crosslinguistic comparison of spatial construals of time. Cognitive Science, 30, 401–450. Oberman, L. M., Winkielman, P., & Ramachandran, V. S. (2007). Face to face: Blocking facial mimicry can selectively impair recognition of emotional expressions. Social Neuroscience, 2, 167–178.


247

Paivio, A. (1991). Dual coding theory: Retrospect and current status. Canadian Journal of Experimental Psychology, 45, 255–287. Pecher, D., & Raaijmakers, J. G. W. (2004). Priming for new associations in animacy decision: Evidence for context dependency. Quarterly Journal of Experimental Psychology, 57A, 1211–1231. Pecher, D., Van Dantzig, S., Zwaan, R. A., & Zeelenberg, R. (2009). Language comprehenders retain implied shape and orientation of objects. Quarterly Journal of Experimental Psychology, 62, 1108–1114. Pecher, D., Zeelenberg, R., & Barsalou, L. W. (2003). Verifying different-modality properties for concepts produces switching costs. Psychological Science, 14, 119–124. Pecher, D., Zeelenberg, R., & Barsalou, L. W. (2004). Sensorimotor simulations underlie conceptual representations: Modality-specific effects of prior activation. Psychonomic Bulletin & Review, 11, 164–167. Pecher, D., Zeelenberg, R., & Raaijmakers, J. G. W. (1998). Does pizza prime coin? perceptual priming in lexical decision and pronunciation. Journal of Memory and Language, 38, 401–418. Pezzulo, G., Barca, L., Bocconi, A. L., & Borghi, A. M. (2010). When affordances climb into your mind: Advantages of motor simulation in a memory task performed by novice and expert rock climbers. Brain and Cognition, 73, 68–73. Proctor, R. W., & Zhang, Y. (2010). “Mother nature doesn’t have a bullet with your name on it”: Coding with reference to one’s name or object location? Journal of Experimental Social Psychology, 46, 336–343. Richardson, D. C., Spivey, M. J., Barsalou, L. W., & McRae, K. (2003). Spatial representations activated during real-time comprehension of verbs. Cognitive Science, 27, 767–780. Rotteveel, M., & Phaf, R. (2004). Automatic affective evaluation does not automatically predispose for arm flexion and extension. Emotion, 4, 156–172. Schwanenflugel, P. J., Harnishfeger, K. K., & Stowe, R. W. (1988). Context availability and lexical decisions for abstract and concrete words. Journal of Memory and Language, 27, 499–520. Schwanenflugel, P. J., & Shoben, E. J. (1983). Differential context effects in the comprehension of abstract and concrete verbal materials. Journal of Experimental Psychology: Learning, Memory, and Cognition, 9, 82–102. Simmons, W. K., Hamann, S. B., Harenski, C. L., Hu, X. P., & Barsalou, L. W. (2008). fMRI evidence for word association and situated simulation in conceptual processing. Journal of Physiology Paris, 102, 106–119. Solarz, A. K. (1960). Latency of instrumental responses as a function of compatibility with the meaning of eliciting verbal signs. Journal of Experimental Psychology, 59, 239–245. Solomon, K. O., & Barsalou, L. W. (2001). Representing properties locally. Cognitive Psychology, 43, 129–169. Solomon, K. O., & Barsalou, L. W. (2004). Perceptual simulation in property verification. Memory & Cognition, 32, 244–259. Spivey, M. J., Richardson, D. C., & Gonzalez-Marquez, M. (2005). On the perceptualmotor and image-schematic infrastructure of language. In D. Pecher & R. A. Zwaan (Eds.), Grounding cognition: The role of perception and action in memory, language, and thinking (pp. 246–281). Cambridge: Cambridge University Press. Stanfield, R. A., & Zwaan, R. A. (2001). The effect of implied orientation derived from verbal context on picture recognition. Psychological Science, 12, 153–156. Steen, G. (2008). The paradox of metaphor: Why we need a three-dimensional model of metaphor. Metaphor and Symbol, 23, 213–241. Steyvers, M. (2010). Combining feature norms and text data with topic models. Acta Psychologica, 133, 234–243.

248

Diane Pecher et al.

Strack, F., Martin, L. L., & Stepper, S. (1988). Inhibiting and facilitating conditions of the human smile: A nonobtrusive test of the facial feedback hypothesis. Journal of Personality and Social Psychology, 54, 768–777. Talmy, L. (1988). Force dynamics in language and cognition. Cognitive Science, 12, 49–100. Teuscher, U., McQuire, M., Collins, J., & Coulson, S. (2008). Congruity effects in time and space: Behavioral and ERP measures. Cognitive Science, 32, 563–578. Van Dantzig, S. (2009). Mind the body: Grounding conceptual knowledge in perception and action. The Netherlands: Erasmus University, Rotterdam (Doctoral dissertation). Van Dantzig, S., Pecher, D., Zeelenberg, R., & Barsalou, L. W. (2008). Perceptual processing affects conceptual processing. Cognitive Science, 32, 579–590. Van Dantzig, S., Pecher, D., & Zwaan, R. A. (2008). Approach and avoidance as action effects. Quarterly Journal of Experimental Psychology, 61, 1298–1306. Van Dantzig, S., Zeelenberg, R., & Pecher, D. (2009). Unconstraining theories of embodied cognition. Journal of Experimental Social Psychology, 45, 345–351. Van den Bos, K. (2003). On the subjective quality of social justice: The role of affect as information in the psychology of justice judgments. Journal of Personality and Social Psychology, 85, 482–498. Van Den Bos, K., Lind, E. A., Vermunt, R., & Wilke, H. A. M. (1997). How do I judge my outcome when I do not know the outcome of others? The psychology of the fair process effect. Journal of Personality and Social Psychology, 72, 1034–1046. Van Elk, M., Van Schie, H. T., & Bekkering, H. (2009). Action semantic knowledge about objects is supported by functional motor activation. Journal of Experimental Psychology: Human Perception & Performance, 35, 1118–1128. Vermeulen, N., Niedenthal, P. M., & Luminet, O. (2007). Switching between sensory and affective systems incurs processing costs. Cognitive Science, 31, 183–192. Wentura, D., Rothermund, K., & Bak, P. (2000). Automatic vigilance: The attentiongrabbing power of approach-and avoidance-related social information. Journal of Personality and Social Psychology, 78, 1024–1037. Wilson, M. (2002). Six views of embodied cognition. Psychonomic Bulletin & Review, 9, 625–636. Wolff, P. (2007). Representing causation. Journal of Experimental Psychology: General, 136, 82–111. Wu, L., & Barsalou, L. W. (2009). Perceptual simulation in conceptual combination: Evidence from property generation. Acta Psychologica, 132, 173–189. Zanolie, K., Van Dantzig, S., Boot, I., Wijnen, J., Schubert, T. W., Giessner, S., et al. (2010). Mighty metaphors: Behavioral and ERP evidence that power shifts attention on a vertical dimension. (Submitted for publication). Zeelenberg, R., Pecher, D., Shiffrin, R. M., & Raaijmakers, J. G. W. (2003). Semantic context effects and priming in word association. Psychonomic Bulletin & Review, 10, 653–660. Zwaan, R. A. (1999). Embodied cognition, perceptual symbols, and situation models. Discourse Processes, 28, 81–88. Zwaan, R. A., Stanfield, R. A., & Yaxley, R. H. (2002). Language comprehenders mentally represent the shape of objects. Psychological Science, 13, 168–171. Zwaan, R. A., & Taylor, L. J. (2006). Seeing, acting, understanding: Motor resonance in language comprehension. Journal of Experimental Psychology: General, 135, 1–11.

C H A P T E R

E I G H T

Thematic Thinking: The Apprehension and Consequences of Thematic Relations Zachary Estes, Sabrina Golonka,1 and Lara L. Jones Contents 1. Introduction 2. Definition and Differentiation 2.1. Definition of Thematic Relations 2.2. Differentiation from Mere Association 2.3. Differentiation from Scripts 2.4. Differentiation from Ad Hoc Categories 2.5. Differentiation from Taxonomic Relations 3. Dissociating Thematic Relations from Taxonomic (Categorical) Relations 3.1. Neuropsychological Dissociations 3.2. Behavioral Dissociations 4. Apprehension of Thematic Relations 4.1. Uncontrollability 4.2. Speed 4.3. Frequency 4.4. Recency 5. Consequences of Thematic Relations for Cognition 5.1. Similarity 5.2. Memory and Categorization (Conceptual Organization) 5.3. Language 5.4. Inference and Analogy 6. Individual Differences and Cultural Effects 6.1. Individual Differences 6.2. Cultural Effects 7. Conclusion 7.1. Future Directions 7.2. Conclusions References 1

250 251 251 253 254 255 255 256 257 259 261 261 263 265 267 268 268 272 276 279 280 281 283 285 285 286 287

S. Golonka previously published as S. Simmons

Psychology of Learning and Motivation, Volume 54 ISSN 0079-7421, DOI: 10.1016/B978-0-12-385527-5.00008-5

#


249

250

Zachary Estes et al.

Abstract A thematic relation is a temporal, spatial, causal, or functional relation between things that perform complementary roles in the same scenario or event. For example, cows and milk are related by a production theme, and sails and anchors are related via a boating theme. Thematic relations are distinct from mere associations, scripts, and ad hoc categories. They also contrast and complement taxonomic (categorical) relations such as “fruits” and “furniture.” Thematic relations and taxonomic relations arise from distinct processes, as evidenced by numerous neuropsychological and behavioral dissociations. Thematic relations may be apprehended uncontrollably and rapidly according to how frequently and recently they have been encountered. They exert profound effects on many core cognitive processes, including similarity, categorization, memory, language, inference, and analogy, and they exhibit robust processing differences across individuals and cultures. In sum, without such thematic thinking, models of cognition will remain categorically limited.

1. Introduction Thematic relations group objects, concepts, or people together by virtue of their participation in the same scenario or event. This contrasts with taxonomic (or categorical) relations, which group things by common properties. Although taxonomic relations have received considerably more attention within psychology, thematic relations are also essential to cognition. They guide our assessment of similarity, organize our conceptual knowledge, and constrain our comprehension of language, among other cognitive functions. To understand how thematic relations play such an important role in cognition, it is useful to consider the different types of information that taxonomic and thematic relations convey. Taxonomic relations underlie traditional feature-based categories. They allow us to simplify the rich perceptual world by treating nonidentical things as if they are the same, and they support inferential generalizations from one thing to another nonidentical thing. Knowing that avocado and aubergine are both foods, for example, can guide expectations and behaviors: classifying aubergine as a food tells us that it is edible, even if we have never encountered that particular food before. Taxonomic relations thus help us to interact appropriately with classes of objects, concepts, and even people. However, taxonomic relations do not help us generate expectations about events or scenarios. For example, how do we know what to expect when dining in a restaurant? Answering this question requires thematic relations. In the case of a restaurant, thematically related items might include food, menus, waiters, and wine. These items share few features, but they are nonetheless linked by their participation in a common event. Importantly,

Thematic Thinking

251

thematic relations can help guide behavior with respect to events: if someone hands you a menu in a restaurant, you can reasonably expect a waiter to take your order. This inference is based on a thematic, rather than taxonomic, relation. Knowing that a menu is taxonomically related to a book (both contain pages with text) is not a useful basis for generating expectations within this event. So, thematic relations serve an essential organizing function in cognition. They convey knowledge about events and scenarios, which complement one’s knowledge about features and taxonomic relations. The goal of this chapter is to integrate and summarize the literature on thematic relations. We begin by defining thematic relations and by distinguishing them from several other theoretical constructs such as associative relations, scripts, and ad hoc categories (Section 2). Next, we elaborate on the dissociation between thematic and taxonomic relations, arguing that they are distinct constructs that arise from different processes (Section 3). We then consider the processing of thematic relations, with particular emphasis on properties such as controllability, speed, frequency, and recency (Section 4). We subsequently identify the importance of thematic relations for a number of basic cognitive processes, focusing specifically on similarity, memory, categorization, language, and analogy (Section 5). Finally, we discuss individual and cultural differences in the prevalence of thematic thinking. The purpose of this integrative review is to highlight the unique and significant contribution of thematic relations to cognition at large.

2. Definition and Differentiation Before detailing their apprehension and consequences for other cognitive processes, it is necessary to provide a more precise definition of thematic relations and to differentiate them from other theoretical constructs. Throughout the remainder of this chapter, we denote concepts in small caps and thematic relations in underlined text.

2.1. Definition of Thematic Relations Generally speaking, a thematic relation is any temporal, spatial, causal, or functional relation between things. More specifically, things are thematically related if they perform complementary roles in the same scenario or event (Golonka & Estes, 2009; Lin & Murphy, 2001; Wisniewski & Bassok, 1999). For example, COW and MILK are related by a production theme, BOOKS and SPECTACLES are related by a reading theme, and SAILS and ANCHORS are related via a boating theme. In each of these cases, the two things perform complementary thematic roles. COWS are producers and

252


their MILK is the product. A BOOK is the object and SPECTACLES are an instrument of reading. SAILS and ANCHORS are both parts of a boat, but they perform different functions. Note that those thematic roles need not complete the theme; they need only complement one another in the sense of fulfilling distinct roles. For instance, SAIL and ANCHOR complement one another, but they clearly do not complete the boating theme. Among the most typical thematic relations are spatial (e.g., JUNGLE and BIRD), temporal (e.g., SUMMER and HOLIDAY), causal (e.g., WIND and EROSION), functional (e.g., FORK and KNIFE), possessive (e.g., POLICE and BADGE), and productive relations (e.g., COW and MILK). Critically, thematic relations are “external” in that they occur between multiple objects, concepts, people, or events. This contrasts with “internal” features and relations among features, which occur within a single entity. To illustrate, DOGS are furry and have a tail connected to the hindquarters. Both of these are internal properties because they predicate the concept in itself; they entail no other object, concept, person, or event. But the fact that DOGS chase SQUIRRELS is an external property of DOGS because it could not occur without its complementary concept, SQUIRRELS. Thus, the key properties of a thematic relation are (1) Externality—thematic relations occur between two or more things. (2) Complementarity—those things must fulfill different roles in the given theme. As we show in the following sections, these two properties are crucial for differentiating thematic relations from mere association, scripts, ad hoc categories, and taxonomic relations (see Figure 1). Thematic relations can arise from either affordance or convention. Regarding affordance, some things have features that allow them to interact with other things in specific ways (Maguire, Maguire, & Cater, 2010). For instance, because HAMMERS are graspable and have a large, heavy, and flat head, they afford hitting. And because NAILS have a small, flat head, they afford being hit. The thematic relation between HAMMER and NAIL is therefore based on their affordances. Not all thematic relations, however, are affordance based. For instance, a WINE GLASS and a DINNER PLATE are thematically related by convention, in that they frequently co-occur in a meal theme. But their features do not afford specific interactions between GLASSES and PLATES to the same extent that HAMMERS and NAILS interact. GLASSES and PLATES clearly perform complementary roles in the meal theme, but they are less directly interactive than HAMMERS and NAILS. Moreover, thematic relations can arise between objects that have no conventional relationship but do have complementary affordances. One can use a ROCK to hit a NAIL because ROCKS, like HAMMERS, afford hitting (although a ROCK is less well suited than a HAMMER). Thematic relations usually entail some combination of affordance and convention.

253

Thematic Thinking

Themes

Scripts

Associations

Taxons

Figure 1 A typology of semantic relations, illustrating the differentiation of thematic relations from associative, taxonomic, and script relations.

2.2. Differentiation from Mere Association Concepts are associated if one evokes thoughts of the other. Association has been invoked to explain a great many behavioral phenomena (e.g., Grosset, Barrouillet, & Markovits, 2005; Martin & Cheng, 2006; Snyder & Munakata, 2008), but as a theoretical construct it is poorly defined (e.g., Bradley & Glenberg, 1983; Hutchison, 2003; McRae & Boisvert, 1998; Moss, Ostrin, Tyler, & Marslen-Wilson, 1995; Spence & Owens, 1990; Thompson-Schill, Kurtz, & Gabriele, 1998). In practice, most researchers have operationally defined association in terms of free association probabilities, where the likelihood of producing a given target word in response to a specific cue word is their association strength. For example, given the cue word “birthday” in the free association task, the probability of a “cake” response is 0.192 (Nelson, McEvoy, & Schreiber, 1998). There are numerous ways in which concepts may be associated. Associates can be synonyms (e.g., BIG ! TALL), antonyms (e.g., BLACK ! WHITE), category comembers (e.g., HORSE ! COW), or conventional phrases (e.g., FOOT ! BALL), among others. Associated concepts therefore always have some other, more specific relation between them. For many associated concepts, that more specific relation is thematic. For instance, “milk” is strongly associated with “cow” (free association probability ¼ 0.388), and this association is explained by the thematic relation that COWS produce MILK. However, many associated concepts are not thematically related. “Lion” is strongly associated with “tiger” (0.362), yet they are not thematically related. Their associative relationship is based on taxonomic categorization (i.e., both are large cats) and lexical co-occurrence (e.g., “Lions, tigers, and bears. . .”), not upon participation in the same scenario or event. LIONS live on savannahs, TIGERS live in jungles, and they do not interact. So LIONS and

254


TIGERS are neither externally related nor complementary. Moreover, many thematically related concepts are unassociated (Estes & Jones, 2009; Simmons & Estes, 2008). “Milk” and “cat” are not associated (free association probability < 0.01) but are thematically related: like all mammals, CATS also produce MILK and they are renowned for consuming it. Similarly, APPLE and GRAVITY are unassociated, but they are thematically related in the context of Newton’s discovery of gravity. Clearly then, thematic relations are not merely associations between things. This partially overlapping relationship is illustrated in Figure 1. The differentiation from association is important because it indicates that thematic relations may occur not only between concepts that do interact and therefore are associated (e.g., HAMMER and NAIL) but also between concepts that simply could interact and therefore are unassociated (e.g., ROCK and NAIL). Indeed, several studies have shown that thematic relations exert similar effects regardless of whether the related concepts are associated or unassociated (e.g., Estes & Jones, 2009; Hare, Jones, Thomson, Kelly, & McRae, 2009; Jones, 2010; Nation & Snowling, 1999; Scheuner, Bonthoux, Cannard, & Blaye, 2004; Simmons & Estes, 2008). How can unassociated things come to be thematically related? Thematic relations can emerge between unassociated things if their features afford specific interactions (see Section 2.1).

2.3. Differentiation from Scripts A script is a generalized sequence of actions and instruments associated with the execution of some common event (Bower, Black, & Turner, 1979; Schank & Abelson, 1977). For example, a bowling script includes such instruments as a BOWLING ALLEY, BOWLING BALLS, and PINS, and such actions as selecting a BALL and attempting to upend the PINS by bowling the BALL down the ALLEY. The various objects, concepts, people, and actions involved in the execution of a script are externally related by the event itself, and they perform complementary roles in the execution of the script. Thus, scripts are a particular type of thematic relation (see Figure 1). However, not all thematic relations are embedded in scripts. Because scripts involve common events, their actions (e.g., BOWLING) and instruments (e.g., BALL) tend to be associated. But as explained in Section 2.2, many thematically related things are not associated. Even though a DOG and a TATTOO are unassociated, people can readily infer a thematic relation between them. The concepts involved in a script do co-occur, whereas the concepts involved in a thematic relation merely could co-occur. So scripts are a subset of thematic relations, but thematic relations additionally include unassociated things. This generality beyond association lends greater explanatory power to thematic relations.

Thematic Thinking

255

2.4. Differentiation from Ad Hoc Categories An ad hoc category is one that is created spontaneously to achieve some goal (Barsalou, 1983). Examples include THINGS TO REMOVE FROM A BURNING HOUSE and THINGS NOT TO EAT ON A DIET. Because the goal around which an ad hoc category is based may resemble a theme (e.g., a burning house theme or a diet theme), ad hoc categories are easily confused with thematic relations. However, such ad hoc categories differ importantly from themes (Lin & Murphy, 2001). Members of an ad hoc category go together as a result of some internal, goal-based property that they all possess (see Barsalou, 1983, p. 225). All members of THINGS TO REMOVE FROM A BURNING HOUSE have some property (i.e., value) that identifies them for salvaging. It could be monetary value (e.g., JEWELRY), sentimental value (e.g., PHOTOS), or some other value (e.g., PETS). Moreover, the members of such ad hoc categories are noncomplementary. JEWELRY, PHOTOS, and PETS do not functionally complement one another like BOATS, SAILS, and ANCHORS do. Rather, they all serve the same goal of salvaging valuables from a burning house. Without the goal, those things no longer cohere or relate to one another in any obvious way. Themes, in contrast, are networks of external relations in which the constituents fulfill complementary roles. A SAIL and an ANCHOR cohere not because they share some property; in fact, the sail is large and light, whereas the anchor is small and heavy. Rather, they cohere because they perform complementary functions in the sailing theme. So whereas an ad hoc category is based around some shared internal property that serves the same goal among all its members, a theme is based around some external relation in which each constituent performs different roles. That is, ad hoc categories are internal and noncomplementary, and hence they differ fundamentally from thematic relations.

2.5. Differentiation from Taxonomic Relations Taxonomic relations entail membership in a common category on the basis of shared features. For example, WHALES and HORSES share important features (e.g., being warm-blooded and bearing live offspring) and hence belong to the same taxonomic category of “mammals.” PIZZA and CHIPS, due to their shared property of being edible, are both members of the “food” category. Concepts belong in a taxonomic category, and hence are taxonomically related to all other category members, by virtue of shared properties. In order for something to be FOOD, it must be edible. And for something to be a MAMMAL, it must be warm-blooded, produce milk, and bear live young. Moreover, taxonomically related concepts are typically not complementary. WHALES and HORSES do not normally complement one another in any theme. Thus, taxonomic relations are based on the properties of the objects themselves, and taxonomic categories cohere around shared

256


properties (Hampton, 2006; Markman & Wisniewski, 1997; Rosch, 1975). As a consequence, taxonomically related concepts tend to resemble one another. In contrast, thematically related concepts tend not to resemble one another, because thematic categories cohere around complementary roles rather than shared properties. The contrasting thematic roles of OWLS and MICE as predator and prey, respectively, require different features. The OWL must be larger than the MOUSE in order to capture it, and the MOUSE must be quicker than the OWL in order to evade it. To propel a boat, a SAIL must be large and relatively light. But to moor the boat, its ANCHOR must be relatively small and heavy. This is not to say that all thematically related concepts are taxonomically unrelated (see Figure 1). After all, OWLS and MICE are both animals, and MILK and COFFEE are both consumable liquids. HORSES and COWS, while taxonomically related by the “mammal” category, are also thematically related in that HORSES are often used to corral CATTLE. So taxonomic and thematic relations are theoretically orthogonal. Generally speaking though, in order for two things to perform different roles in the same theme, they typically differ in important respects. Consequently, thematically related concepts tend to be featurally dissimilar (Estes, 2003a; Estes & Jones, 2009; Golonka & Estes, 2009; Lin & Murphy, 2001; Wilkenfeld & Ward, 2001; Wisniewski, 1996; Wisniewski & Bassok, 1999; Wisniewski & Love, 1998).

3. Dissociating Thematic Relations from Taxonomic (Categorical) Relations The distinction between thematic relations and taxonomic relations is more than theoretical. Rather, thematic thinking and taxonomic thinking appear to arise from distinct processes. Evidence from neurological impairments and neuroimaging both indicate that thematic processing and taxonomic processing have important differences in neural topography and cortical networks. Purely behavioral studies with neurologically intact participants also suggest that thematic processing and taxonomic processing may be differentially affected by and may have differential effects on other behaviors. Before reviewing these numerous dissociations, a methodological and terminological consideration is necessary. By far, the single most common method used to measure thematic thinking is the matching-to-sample task (see Figure 2). In this task, a base stimulus is presented with two or more option stimuli, and participants are instructed to choose the option that matches the base on some given criterion. For instance, a typical trial might have DOG as the base, CAT as a taxonomically related option, and BONE as a thematically related option, with participants instructed to choose the option that “goes

257

Thematic Thinking

Base

Options Taxonomic

Thematic

Unrelated

Figure 2 Example illustrating the matching-to-sample (a.k.a. matching) task, in which participants are instructed to select the option that matches the base on a given criterion (e.g., “which option goes with the base?”). A typical trial includes only two options, where a taxonomic trial contrasts taxonomic and unrelated options, a thematic trial contrasts thematic and unrelated options, and a conflict trial contrasts taxonomic and thematic options.

with” the base. This paradigm has several parameters, such as the mode of presentation (pictures or words), the number of options (two or more), the relation between the options and the base (e.g., taxonomic, thematic, or unrelated), and the choice criterion (e.g., “goes with,” “is the same kind of thing”). In this paradigm, reliably choosing a thematic option over an unrelated option indicates apprehension of thematic relations, whereas reliably choosing a thematic option over a taxonomic option indicates a preference or processing advantage for thematic thinking (and vice versa for taxonomic choices). Despite some valid criticisms (see Section 5.2), this matching-to-sample task has been ubiquitously employed, and hence we refer to it often throughout the remainder of this chapter. For brevity, we refer to it as the “matching task.”

3.1. Neuropsychological Dissociations Neuropsychological dissociations between taxonomic processing and thematic processing have been observed. Davidoff and Roberson (2004) reported a case study of LEW, who had Wernicke’s aphasia. In a matching task (see Figure 2), LEW was presented pictures of three objects and his task was to indicate which two best “go together.” On different trials, LEW was instructed to respond on the basis of color, size, or function. For example, given a HAMMER, a NAIL, and a SCREW, the correct response on a size trial would be NAIL and SCREW, whereas the correct response on a function trial would be NAIL and HAMMER. LEW performed poorly on color (24%

258


accuracy) and size trials (52%), indicating impaired judgment of basic object features. However, he performed as well as control participants on functionally related, thematic trials (81%). Thus, despite impaired featural categorization, LEW’s thematic categorization was spared. Neuroimaging studies reveal that taxonomic and thematic processing also activate distinct cortical networks in normal participants. Sachs and colleagues have conducted a number of studies contrasting taxonomic from thematic processing. Sachs, Weis, Krings, Huber, and Kircher (2008) used a matching task in which participants selected which of two options best went with a target (e.g., CAR). Choosing a taxonomic match (e.g., BUS) over a thematic match (e.g., GARAGE) was associated with increased activation of the left thalamus, right middle frontal gyrus, and left precuneus. In a lexical decision task, Sachs, Weis, Zellagui, et al. (2008) found greater activation of the right precuneus from taxonomic prime-target pairs (e.g., CAR ! BUS) than from thematic pairs (e.g., CAR ! GARAGE). The increased activation of the precuneus across both studies could be due to the greater reliance of taxonomic processing upon perceptual information (Sachs, Weis, Krings, et al., 2008), or upon less salient meanings of words (Sachs, Weis, Zellagui, et al., 2008). Sass, Sachs, Krach, and Kircher (2009) found that thematic relations activated left superior and middle temporal regions, whereas taxonomic relations activated primarily right-lateralized frontotemporal regions. They concluded that taxonomic relations require more effortful processing than thematic relations. Kalenine et al. (2009) tested whether taxonomic and thematic processing differentially rely upon visual and motor representations, respectively. They hypothesized that taxonomic relations would selectively activate visual networks because they entail featural similarity, whereas thematic relations would selectively activate motor and spatial networks because they support actions. Using a matching task in which participants chose which of two pictures is “semantically related” to the target, they presented either a taxonomic or a thematic option with an unrelated option. In contrast to Sass et al.’s (2009) suggestion that taxonomic processing is more effortful, Kalenine et al. found that taxonomic options were identified more quickly than thematic options. Taxonomic categorization bilaterally activated the visual association networks in the cuneus and lingual gyrus of the occipital cortex, suggesting that taxonomic categorization does indeed rely upon visual processing. Thematic categorization bilaterally activated motor and spatial networks in the posterior middle temporal cortex and inferior parietal lobules. In sum, there is not yet consensus on exactly which cortical structures and networks are required for which mode of processing, but it is clear that taxonomic and thematic processing may be selectively impaired and consistently activate distinct cortical networks. Much remains to be specified neurologically, but the dissociation of taxonomic processing from thematic processing appears incontrovertible.

Thematic Thinking

259

3.2. Behavioral Dissociations Purely behavioral studies of neurologically intact participants have also revealed many differences between taxonomic processing and thematic processing. Several studies have examined the thematic processing of poor readers and normal readers. Children with poor reading abilities are generally less skilled than normally reading children at thematically integrating textual information (Cain, Oakhill, & Elbro, 2003), but providing a thematic organizer facilitates text recall among poor readers (Risko & Alvarez, 1986). And among poor readers, stronger association between prime and target words facilitates lexical decisions for taxonomic pairs (e.g., TABLE ! CHAIR) but not for thematic pairs (e.g., BEACH ! SAND; Nation & Snowling, 1999). Doughty, Lawrence, Al-Mousawi, Ashaye, and Done (2009) presented to schizophrenic (SZ) and control participants 45 objects from five taxonomic categories (i.e., animals, fruits, body parts, clothing, and transport), and they asked participants to sort them into groups that “go together.” Whereas control participants tended to sort all items into their taxonomic categories, SZ participants were more likely to sort thematically. For example, one SZ participant who sorted MONKEY with the group of fruits explained thematically that “monkeys eat fruit.” When subsequently asked to sort the items into taxonomic categories, however, most SZ participants correctly identified all category members. This study indicates that SZ individuals retain relatively normal taxonomic knowledge but exhibit a tendency for thematic processing. However, SZ patients are impaired at thematic sequencing of story events (Matsui et al., 2007). It may be that SZ individuals tend toward thematic processing, but have highly disorganized and idiosyncratic themes that tend not to conform to experimenters’ expectations (see also Titone, Libben, Niman, Ranbom, & Levy, 2007). In a study modeled after Davidoff and Roberson’s (2004) procedure with aphasic patient LEW, Lupyan (2009) presented object triads from which normal undergraduates were asked to choose the one that does not belong. For example, given a triad of BEE, EAGLE, and OWL, the size oddball is BEE. Given a triad of PIG, PENGUIN, and ZEBRA, the color oddball is PIG. And given a triad of POTATO, BALLOON, and CAKE, the thematic oddball is POTATO. On half the trials of this oddball task, participants also rehearsed a string of nine digits (i.e., verbal interference), which they were later prompted to remember. Results are illustrated in Figure 3. Verbal interference significantly slowed detection of size and color oddballs but not thematic oddballs. Thus, like patient LEW, normal undergraduates exhibited impaired featural categorization but preserved thematic categorization. Maki and Buchanan (2008) investigated the latent factors that contribute to the mental representation of word meanings. They submitted five measures of association, three measures of semantic features, and five measures of text-based co-occurrence for each of 629 word pairs to three different

260


Response time (ms)

2800 2600 2400 2200 2000 1800 Color

Size Control

Theme

Interference

Figure 3 Response time to identify a color, size, or thematic oddball among a triad of objects, under normal (control) conditions or with verbal interference. Interference delayed identification of color and size oddballs, whereas identification of thematic oddballs was unimpaired. Results are extrapolated from Lupyan (2009).

statistical analyses (i.e., factor analysis, hierarchical clustering, and multidimensional scaling). Across analyses they found a three-factor structure consisting of separable associative, semantic, and thematic factors, akin to our typology of semantic relations illustrated in Figure 1. This suggests that association strength, semantic (i.e., taxonomic) similarity, and thematic relatedness independently contribute to word meaning. Taxonomic processing and thematic processing can also elicit differential effects on other cognitive processes such as the apprehension of commonalities and differences. Because taxonomic categorization is based on a comparison process (see, e.g., Hampton, 2006; Markman & Wisniewski, 1997), inducing participants to compare objects is commonly assumed to evoke taxonomic processing. Inducing participants to integrate objects, in contrast, is assumed to evoke thematic processing. Estes (2003a) found that comparing concepts decreased participants’ judgments of their similarity, whereas integrating concepts significantly increased judgments of similarity. That is, taxonomic processing and thematic processing, respectively, decreased and increased perceived similarity. Gentner and Gunn (2001) administered a difference listing task, in which participants were given limited time to list a single difference for as many concept pairs as possible. Prior to the difference listing task, participants either compared the concepts or integrated them. Participants listed significantly more differences for pairs that they had compared than for those they had integrated. In other words, relative to taxonomic processing, thematic processing inhibited the detection of

Thematic Thinking

261

differences. Together these experiments reveal that thematic processing decreases perceived difference and increases perceived similarity (see also Golonka & Estes, 2009). Thus, numerous behavioral studies with normal participants have differentiated thematic processing from taxonomic processing.

4. Apprehension of Thematic Relations Now that we have clearly defined what thematic relations are (i.e., spatial, temporal, causal, or functional relations between things that fulfill complementary roles) and what they are not (i.e., mere associations, scripts, ad hoc categories, or taxonomic relations), we will review how they are apprehended. To put it most simply, apprehending a thematic relation entails recognizing that the given concepts could perform different roles in the same scenario. This can be achieved by either retrieving a thematic relation from memory or generating one ad hoc. Conventional relations, such as that between a HAMMER and a NAIL or between a WINE GLASS and a DINNER PLATE (see Section 2.1), can be retrieved directly from memory. These concepts activate their typical roles, and the match between those roles determines whether and how they are thematically related (Estes & Jones, 2009). Unconventional thematic relations, such as that between a ROCK and a NAIL or between a WINE GLASS and COLA, must be generated ad hoc. Such unconventional relations between things arise from their affordances. For instance, a ROCK affords hitting a NAIL, and a WINE GLASS affords containing COLA. These affordances can be perceived directly (e.g., Gibson, 1979), which means that we can tell whether two things could plausibly be related thematically, even if we have no prior knowledge of a thematic link between them. In this section, we consider some basic properties of thematic integration (i.e., its uncontrollability and speed) and key factors of thematic processing (i.e., frequency and recency).

4.1. Uncontrollability Thematic relations are intrusive. They are apprehended involuntarily in tasks for which they are irrelevant and even counterproductive. Bassok and Medin (1997) observed that, when instructed to justify their similarity ratings, participants frequently referred to thematic relations rather than features of the individual stimuli. In a more direct investigation of this phenomenon, Wisniewski and Bassok (1999) showed that not only does thematic processing intrude on a taxonomic task (i.e., similarity ratings) but also that taxonomic processing intrudes on a thematic task (i.e., thematic relatedness ratings). In fact, several other studies have confirmed that

262


thematic relations intrude on similarity judgments (Estes, 2003a; Gentner & Brem, 1999; Golonka & Estes, 2009; Jones & Love, 2007; Simmons & Estes, 2008). These studies are described in more detail in Section 5.1. Golonka (2008) tested whether participants are capable of ignoring thematic relations when judging similarity. Despite instructions not to base their ratings on thematic relations, participants’ similarity ratings nonetheless exhibited a thematic effect of approximately the same magnitude. Ross and Murphy (1999) tested whether thematic information is automatically activated in tasks such as similarity judgments and category decisions. Providing a thematic category label (e.g., “breakfast foods”) increased the similarity of thematically related foods (e.g., BACON and EGGS), and reading a thematic prime (e.g., “The bagel was what he had when he woke up”) facilitated category decisions about those foods (e.g., “is a bagel a breakfast food?”). The finding that thematic labels and primes affected similarity and categorization, respectively, suggests that such thematic information might not be automatically activated under normal (unprimed) circumstances. In contrast, taxonomic primes had no influence on similarity ratings or category decisions, whereas ad hoc category primes induced even larger effects on similarity and categorization. Thus, relative to taxonomic and ad hoc categories, thematic knowledge appears to be moderately activated in similarity and categorization tasks. Gentner and Brem (1999) used a matching task (see Figure 2) in which a taxonomic option was paired with either a thematic option (i.e., conflict trial) or an unrelated option (i.e., taxonomic trial). For instance, the base GARLIC was presented with ONION (taxonomic) and either VAMPIRE (thematic) or CEMENT (unrelated). Participants were instructed to identify the taxonomic option. The rationale was that if thematic relations intrude on taxonomic processing, then participants should exhibit more errors on conflict trials than on taxonomic trials. Indeed, thematic options did intrude on taxonomic processing, as evidenced by more errors on conflict trials. Lin and Murphy (2001) used a matching task with conflict trials, and they asked participants to choose the option that “goes with” the base “to form a category” (Experiment 1) or to choose the two options that “best form a category” (Experiment 2). Participants also were given a definition of “category” that emphasized taxonomic relations. In other studies, participants were instructed to treat the stimuli like representatives of their categories rather than as individuals (Experiment 4), and to justify their choices (Experiment 5). Nevertheless, across studies, participants tended to choose thematic options more often than taxonomic options (see also Murphy, 2001). In another study, Lin and Murphy (2001, Experiment 10) used a speeded categorization task in which participants read a category label (e.g., “animal”) followed by two simultaneously presented options. Participants’ task was to decide whether either option was a member of the target category. On

Thematic Thinking

263

critical trials, one of the options belonged to the category (e.g., DOG) and the other option was either thematically related to that alternative (e.g., LEASH) or unrelated (e.g., NEST). They found that thematic relations facilitated taxonomic categorization. For instance, participants responded that DOG is a member of “animal” more quickly when paired with LEASH than with NEST. Lin and Murphy, like Gentner and Brem (1999), concluded that thematic relations have a fast and automatic influence on taxonomic categorization. Estes and Jones (2009; see also Jones, 2010) showed that a target word is recognized faster after a thematically related prime word (e.g., SOUP ! CAN) than after an unrelated prime (e.g., COW ! CAN). We referred to this effect as integrative priming, because the prime and target concepts were integrated into a single entity (i.e., the word pair denotes a single referent, rather than two independent referents). In subsequent experiments, we also embedded the thematically related word pairs in a list that included either many other thematically related pairs (e.g., BIRTHDAY ! CANDLE) or many thematically unrelated pairs (e.g., LIMB ! CANDLE). The rationale was that if thematic relations were apprehended voluntarily, then integrative priming should only be observed in the list with many thematic pairs. That is, if thematic integration was under participants’ strategic control, then it should not occur in the list with few thematic pairs, because a strategy of thematic integration would rarely succeed in that list. Contrary to this prediction, however, integrative priming was observed across both lists with equal magnitudes (see also Coolen, van Jaarsveld, & Schreuder, 1991). In all of these studies, thematic relations intruded on other cognitive processes (e.g., similarity judgments, categorization, and word recognition) despite being irrelevant to the task.

4.2. Speed Thematic relations appear to be apprehended relatively rapidly. As described in the preceding section, Gentner and Brem (1999) observed interference on a taxonomic categorization task from a thematically related distracter, relative to an unrelated distracter. In fact, to test whether thematic intrusions primarily occur early or late in processing, Gentner and Brem required participants to identify the taxonomic option within either a 1-s or a 2-s response deadline. Results are illustrated in Figure 4. Thematic distracters induced more errors than unrelated distracters at both deadlines, thus indicating that thematic relations are detected early (i.e., 1 s or less; see also Lin & Murphy, 2001, Experiment 10). Interestingly, the magnitude of this thematic intrusion decreased from 16% to 10% from the 1-s deadline to the 2-s deadline. Chwilla and Kolk (2005) created story-like scripts by presenting simultaneously two unassociated words (e.g., DIRECTOR and BRIBE) followed by a third word (e.g., DISMISSAL). Critically, the first two words could either

264


35

Error rate (%)

30 25 20 15 10 5 0 1

2 Response deadline (s) Unrelated

Thematic

Figure 4 Error rates on the matching task at 1-s and 2-s response deadlines. Taxonomic options were paired with either an unrelated option or a thematic option. Thematic options induced more errors than unrelated options at both deadlines. Results are extrapolated from Gentner and Brem (1999).

establish a thematic context for the third word (as in the above example) or they could be unrelated. Across both a lexical decision task and a plausibility judgment task, the thematically related triads elicited faster responses than the unrelated triads. Moreover, the thematically related triads also elicited a smaller N400 effect than unrelated triads. This decreased N400 effect from thematic triads suggests that, given two components of a thematic scenario, participants expected the third concept to also relate thematically. It further indicates that whether the third concept is thematically related to the preceding two concepts can be apprehended in as little as 400 ms (see also Metusalem, Kutas, Hare, McRae, & Elman, 2010). Estes and Jones (2009) directly compared lexical decision times for target words (e.g., CAN) preceded by a prime word that was thematically related (e.g., SOUP), taxonomically related (e.g., JUG), or unrelated (e.g., COW). To compare the time courses of thematic processing and taxonomic processing, we also manipulated the duration between presentation of the prime and target words (i.e., stimulus onset asynchrony or SOA). Results are illustrated in Figure 5. Across SOAs of 100, 500, 1500, and 2500 ms, the thematic and taxonomic primes facilitated recognition of their target words relative to the unrelated primes. However, at no point did the magnitude of the priming effect differ between the thematic and taxonomic conditions. This result has two implications of relevance to our purposes here. First, thematic relations were apprehended rapidly enough to facilitate word recognition when the delay between prime and target onset was only one-tenth of a second. Furthermore, these thematic relations were apprehended just as rapidly as taxonomic relations.

265

Thematic Thinking

60

Priming (ms)

50 40 30 20 10 0 100

500

1500 SOA (ms) Semantic

2500

Integrative

Figure 5 Semantic and integrative priming effects (i.e., baseline RT—experimental RT) across stimulus onset asynchronies (SOA). Both semantic and integrative priming were evident by 100 ms, asymptoted around 500 ms, and persisted to 2500 ms. At no point did the magnitudes of semantic priming and integrative priming diverge. Results are extrapolated from Estes and Jones (2009).

Studies comparing comprehension times for word pairs that are understood by inferring either a thematic relation (e.g., ONION TEARS) or a common feature (e.g., VAMPIRE INSECT) have found that thematic pairs are actually understood more quickly than feature-based pairs (Estes, 2003b; Gagne´, 2000). This finding is consistent with Sass and colleagues’ (2009) suggestion that taxonomic relations require more effortful processing than thematic relations. However, many of the featural pairs used in those studies were more akin to metaphors than to taxonomic pairs (e.g., Estes, 2003a, 2003b; Estes & Glucksberg, 2000), so they are more suggestive than conclusive of a thematic processing advantage. Indeed, in their comparison of thematic choices and taxonomic choices in the matching task, Kalenine and colleagues (2009) found that the taxonomic options were identified more quickly. Thus, much evidence indicates that thematic relations are apprehended relatively rapidly, though it is not yet clear exactly how rapidly.

4.3. Frequency A given concept tends to perform the same role across various contexts, and people are implicitly aware of these thematic roles. For instance, people know from experience that PAPER is often written on. Less often, however, PAPER is also used for making things (e.g., AIRPLANES) and covering things (e.g., GIFTS), among other roles it may serve. Thus, with experience concepts acquire a frequency distribution of thematic relations. We are

266


implicitly aware that PAPER is most frequently written on, is slightly less frequently used to make or cover things, even less frequently causes cuts, and never eats animals. These relation frequencies affect thematic processing in an adaptive way. Whenever we encounter an object, a person, or a concept with which we have sufficient experience, that thing’s frequent relations are activated. Effectively, encountering a familiar concept automatically activates the other concepts with which it is most likely to interact or co-occur, thereby facilitating perception of and responding to those thematically related concepts. Note also that relation frequencies are specific to individual concepts. The subject of a cutting theme is more likely to be a KNIFE than a PAPER, so that thematic relation is more frequent for KNIFE than for PAPER. However, things are very rarely written with knives, so the writing theme is more frequent for PAPER than for KNIFE. Gagne´ and Shoben (1997) demonstrated that people know and use these relation frequencies. In a language comprehension study, they presented word pairs that could be thematically integrated by either a highly frequent relation (e.g., PAPER NOTE) or a less frequent relation (e.g., PAPER CUT), and they asked participants to judge as quickly as possible whether the word pair made sense as a phrase. The phrases were understood more quickly when they instantiated a highly frequent relation than an infrequent relation (see also Gagne´ & Spalding, 2004; but see Maguire, Devereux, Costello, & Cater, 2007). Storms and Wisniewski (2005) replicated this relation frequency effect in the Indonesian language, which differs fundamentally in structure from English, thus revealing that the effect is a general cognitive phenomenon rather than a language-specific idiosyncrasy. Even 4- to 5-year-old children, who possess relatively limited linguistic experience, are able to use relation frequencies in interpreting word pairs (Krott, Gagne´, & Nicolades, 2009). Maguire, Maguire, et al. (2010) and Maguire, Wisniewski, and Storms (2010) demonstrated that relation frequencies are constrained by the semantic categories (or features) of the given concept and that thematic integration depends on the interaction of the two concepts’ categories. For example, LEATHER most frequently serves a compositional role, acting as the substance of which other objects consist. When thematically integrating LEATHER with another concept, then that frequent composition relation is activated. However, some phrases such as LEATHER NEEDLE entail a relation other than composition. This case illustrates the interactive nature of thematic integration: although a NEEDLE cannot be made of LEATHER, LEATHER NEEDLE is nevertheless understood relatively fast because the semantic categories of the two concepts mutually constrain the apprehension of a sensible thematic relation. Thus, the features of a concept constrain the thematic roles that it tends to instantiate, thereby producing a distribution of more and less frequent thematic roles for each concept with which we have sufficient experience. Our implicit statistical knowledge of these relation frequencies guides the apprehension of thematic relations.

Thematic Thinking

267

4.4. Recency Just as the frequency of a thematic relation affects processing, so does its recency. Both children (Smiley & Brown, 1979) and adults (Wisniewski & Love, 1998) are more likely to apprehend thematic relations after a series of other thematic relations than after a series of taxonomic relations. For instance, DOG KENNEL (habitation) is understood faster after OFFICE PLANT (spatial) and WIND EROSION (causal) than after OSTRICH BIRD (taxonomic) and DESK BED (taxonomic). More specifically, even a single thematic relation can facilitate the comprehension of a subsequent relation. For example, DOG KENNEL is understood faster after DOG HOUSE than after DOG FOOD because the first two both instantiate a habitation relation, whereas the third instantiates a different relation (Gagne´, 2001). Gagne´ (2001) initially obtained this relation recency effect only when the same modifier noun was used in both the prime and the target word pairs (e.g., DOG HOUSE ! DOG KENNEL) and not when the modifiers differed (e.g., DOG HOUSE ! CAT KENNEL). However, much subsequent research has revealed that relation priming can in fact be obtained with entirely different and unrelated concepts. For example, despite having no lexical overlap from prime to target, BEAR CAVE facilitates comprehension of BIRD NEST because both use the habitation relation (Estes, 2003b; Estes & Jones, 2006; Spellman, Holyoak, & Morrison, 2001). In an innovative demonstration of this recency effect without lexical repetition, Raffray, Pickering, and Branigan (2007) used a picture matching task in which ambiguous target phrases could be understood in either of two possible ways. For instance, participants were prompted to decide whether a DOG SCARF matched a picture of a dog wearing a scarf (possessor relation) or of a scarf with a dog pattern on it (descriptor relation). These ambiguous targets were preceded by a prime trial that was unambiguously understood by one of those two relations, such as a rabbit wearing a T-shirt (possessor) or a T-shirt with a rabbit pattern on it (descriptor). Participants more frequently matched the ambiguous targets to the picture depicting the same relation as the prime than to the other relation. Hristova (2009) also demonstrated the recency effect in an innovative paradigm. She preceded a thematically related target pair (e.g., BACTERIUM : INFECTION) with a prime pair that used either the same thematic relation (e.g., ACID : CORROSION) or a different relation (e.g., FILTER : WATER). One of the concepts in each pair appeared in either red or green font (the other appeared in black), and critically, the colors of the prime and target were either congruent (i.e., red ! red or green ! green) or incongruent (i.e., red ! green or green ! red). Participants’ task was to identify the color of the font for each word pair. Results are illustrated in Figure 6. When the prime and target appeared in the same color (i.e., congruent trials), relation recency (i.e., same relation) facilitated responding. But when the prime and

268


Response time (ms)

690 680 670 660 650 640 Congruent Same relation

Incongruent Different relation

Figure 6 Response time to identify the color of a target word pair that instantiates either the same or a different relation from a prime, and that appears in the same (congruent) or a different color (incongruent) from the prime. Repetition of the target thematic relation facilitated responding on congruent trials but hindered responding on incongruent trials. Results are extrapolated from Hristova (2009).

target appeared in different colors (i.e., incongruent trials), relation recency actually slowed responding. Evidently, participants apprehended the thematic relations, and if the prime and target instantiated the same thematic relation, participants expected the words to be of the same color. Hristova thus demonstrated the relation recency effect in a paradigm for which thematic relations were irrelevant and, on most trials, counterproductive. Collectively, the findings described above indicate that thematic relations are apprehended uncontrollably and relatively quickly according to their frequency and recency of use.

5. Consequences of Thematic Relations for Cognition Thematic relations are central to many cognitive processes. In this section, we focus on a few basic processes for which thematic relations have particularly profound effects.

5.1. Similarity Traditional models of similarity, such as the contrast model (Tversky, 1977) and the structural alignment model (Gentner & Markman, 1997; Markman & Gentner, 2000), explain similarity only in terms of comparison: To

Thematic Thinking

269

determine the similarity between two things, we compare them, identify their commonalities and differences, and weigh them accordingly. However, to the extent that other cognitive processes are shown to affect perceived similarity, these comparison models may fail to predict similarity judgments and related behaviors such as categorization, preferences, and decisions. Thematic relations reliably affect perceived similarity. To illustrate, consider LEMONADE, COFFEE, and MILK. Pause for a moment to judge which two of these concepts are most similar. All three are drinkable liquids, but of course they differ in taste. They also vary in color and typical serving temperature. LEMONADE and COFFEE differ markedly in both color and typical serving temperature, as do MILK and COFFEE. In contrast, LEMONADE and MILK are both typically served chilled, and they differ only minimally in color. Thus, by feature comparison models such as the contrast model and the structural alignment model, LEMONADE and MILK should be judged most similar. Remarkably, though, MILK and COFFEE are actually judged most similar, despite having the fewest features in common. Why? COFFEE and MILK are perceived to be similar because people often drink them together. More generally, such thematically related concepts are judged more similar than thematically unrelated concepts (Golonka & Estes, 2009; Simmons & Estes, 2008; Wisniewski & Bassok, 1999). Essentially, there are two main sources of similarity. Feature comparison, which is achieved by a process of structural alignment (Markman & Gentner, 2000), reveals the degree of featural commonality between objects or concepts. Thematic relations provide an additional source of similarity. By comparing MILK and COFFEE, we discover their common liquidity, drinkability, and so forth, which endow MILK and COFFEE with some degree of similarity. By thematically integrating MILK and COFFEE, we apprehend their complementary participation in the same scenario, and this boosts their similarity even further. This distinction between feature comparison and thematic integration gives rise to a two-dimensional model of similarity, which for simplicity is conceptualized as a 2 (taxonomic similarity: high, low) 2 (thematic similarity: high, low) similarity space (Wisniewski & Bassok, 1999). Taxonomically similar concepts can be either thematically related (e.g., MILK and COFFEE) or unrelated (e.g., MILK and LEMONADE), and taxonomically dissimilar concepts can also be either thematically related (e.g., MILK and COW) or unrelated (e.g., MILK and HORSE). In a seminal demonstration of this thematic effect, Wisniewski and Bassok (1999) had participants rate the similarity of concepts that varied orthogonally in taxonomic category membership (Cþ, C) and thematic relatedness (Tþ, T). For example, SHIP was compared to TUGBOAT (CþTþ), CANOE (CþT), SAILOR (CTþ), and SOLDIER (CT). They found that participants consistently rated Tþ concepts as more similar than T concepts. This was even true when the concepts were from the same

270


taxonomic category. To illustrate, SHIP was rated more similar to TUGBOAT than to CANOE (see also Golonka & Estes, 2009). A valid criticism of this result is that because different concepts were used across conditions (e.g., TUGBOAT appeared in the CþTþ condition only), the difference in perceived similarity could be attributable to some factor other than thematic relatedness. However, subsequent experiments have established that the exact same concepts are judged more similar when participants thematically integrate them than when participants only compare their features. Estes (2003a) found that thematically integrating concepts (e.g., interpreting DOCTOR LIBRARY as a library for doctors) increases their perceived similarity, relative to a condition where the same items were not integrated prior to the similarity judgment. This suggests that it is the act of thematic integration, rather than a preexisting association, that increased their perceived similarity. Jones and Love (2007) also found a causal effect of thematic integration on similarity. In their experiment, participants judged similarity according to participation in the same thematic context. For instance, participants selected SHEEP as more similar to COLLIE than to GERMAN SHEPHERD when those concepts (i.e., SHEEP and COLLIE) occurred in the same thematic sentence (e.g., “The collie herds the sheep”) rather than in separate, unrelated sentences (e.g., “The German shepherd herds the sheep” and “The collie chases the cat”). Wisniewski and Bassok (1999) argued that whether one compares or thematically integrates a pair of concepts depends upon the compatibility between the stimuli and the processes. A good proxy for featural commonality is taxonomic category membership. Concepts or objects that belong to the same taxonomic category tend to share more commonalities (Hampton, 2006; Mervis & Rosch, 1981) and have more differences related to these commonalities (Markman & Wisniewski, 1997; Pothos & Chater, 2002) than concepts or objects from different taxonomic categories. Wisniewski and Bassok argued that the characteristics of taxonomically related concepts make them highly compatible with the comparison process. For example, MILK and COFFEE are both members of the “beverage” category because they share important commonalities (e.g., liquidity, potability). Comparing MILK and COFFEE draws attention to differences that are related to these commonalities (e.g., caffeine content, taste, and typical serving temperature). These differences are highly informative and can influence how we interact with those objects. For example, one might choose to have MILK instead of COFFEE before going to bed. In such cases, the process of comparison helps to identify salient properties. MILK and COW, however, are difficult to compare because, like all taxonomically unrelated concepts, they have very few properties in common (Markman & Wisniewski, 1997). Furthermore, the differences that result from the comparison of these concepts are relatively uninformative (Wisniewski & Bassok, 1999). Participants who are asked to compare such concepts often simply state the categories to which

Thematic Thinking

271

each belongs (e.g., “milk is a beverage, cow is an animal”). Wisniewski and Bassok (1999) argued that concepts from different categories are not compatible with the comparison process, which requires some basic level of commonality between concepts (see also Bassok & Medin, 1997). Taxonomically unrelated concepts are compatible with thematic integration, however. Because thematic relations occur between concepts that perform different roles, things that share few commonalities are actually easier to thematically integrate than concepts that share many properties. That is, having few commonalities provides an opportunity for different concepts to complement one another thematically. In contrast, taxonomically related concepts tend to have too many commonalities to perform different roles in the same theme. Thus, stimulus compatibility drives process selection (Wisniewski & Bassok, 1999). Two key empirical results support this argument. First, stimulus compatibility explains the tendency to thematically integrate disparate concepts that share no preexisting thematic relation. Both Wisniewski and Bassok (1999) and Bassok and Medin (1997) reported that participants spontaneously generate thematic relations between taxonomically unrelated items. For instance, when asked to describe the similarity between PEDIATRICIAN and CAT, people often respond with statements such as “a pediatrician might own a cat.” These results indicate that, when faced with incomparable stimuli, participants attempt to thematically integrate them. From this perspective, the intrusion of thematic relations on similarity and categorization tasks (Section 4.1) can be interpreted as evidence for a mismatch between task and stimuli. Similarity tasks are intended to tap the comparison process. But if the given stimuli are difficult to compare (e.g., if they are taxonomically unrelated), then participants thematically integrate them instead. Second, as illustrated in Figure 7, thematic relations have a particularly large effect on the similarity of taxonomically unrelated concepts (Golonka & Estes, 2009; Wisniewski & Bassok, 1999). Golonka and Estes (2009) found that thematically related concepts from different taxonomic categories (e.g., SHIP and SAILOR) are judged to be much more similar than thematically unrelated concepts (e.g., SHIP & SOLDIER). For these items, thematic relatedness explained a substantial proportion of the variance in similarity ratings. In contrast, thematically related concepts from the same taxonomic category (e.g., SHIP and TUGBOAT) are judged to be only slightly more similar than thematically unrelated concepts (e.g., SHIP and CANOE). For these items, featural commonality explained the majority of the variance in similarity ratings. To provide another example, MILK is judged much more similar to COW than to HORSE, but MILK is judged only slightly more similar to COFFEE than to LEMONADE. Both COW and COFFEE receive a boost in similarity to MILK due to their respective thematic relations, but that boost is larger for COW because there is little other basis on which to judge its similarity to MILK. In contrast, COFFEE and MILK have many features in

272


7

Similarity rating

6 5 4 3 2 1 Taxonomically related Thematically related

Taxonomically unrelated Thematically unrelated

Figure 7 A typical result illustrating the influence of thematic relations on similarity ratings, with a larger effect among taxonomically unrelated stimuli than among taxonomically related stimuli. Results are extrapolated from Wisniewski and Bassok (1999) and Golonka and Estes (2009).

common, and hence their already-high similarity is boosted only slightly by their thematic relation. The effect of thematic relations on perceived similarity is thus moderated by the concepts’ taxonomic relatedness: taxonomically similar concepts are easily compared and this comparison identifies relevant differences between them. Taxonomically dissimilar concepts are difficult to compare, thus leading participants to thematically integrate them instead (Wisniewski & Bassok, 1999). Collectively, these studies reveal that people spontaneously apprehend thematic relations when judging similarity, and these thematic relations affect similarity judgments. This thematic influence on similarity is particularly pronounced among stimuli that are otherwise difficult to compare (see Figure 7).

5.2. Memory and Categorization (Conceptual Organization) A tremendous amount of research demonstrates that conceptual knowledge is organized, to a large extent, around thematic relations. Evidence of conceptual organization derives primarily from studies showing that thematic relations aid memory and strongly affect categorization. Of particular interest in this area of research has been a claim that conceptual organization changes across the lifespan. Specifically, some have argued that thematic thinking dominates in early childhood, but then becomes secondary to taxonomic thinking in later childhood and into middle adulthood, and finally thematic thinking resuming its dominance in later adulthood (e.g., Smiley & Brown, 1979; see also Nelson, 1977). The cognitive transition in

Thematic Thinking

273

childhood, known variously as the “thematic-to-taxonomic shift” and the “syntagmatic-paradigmatic shift,” has been demonstrated across several memory and categorization paradigms. However, much of that research has been extensively criticized on methodological grounds, and the transition back to thematic thinking in older adulthood is less well researched. Below we summarize this literature. It has long been known that participants spontaneously organize items both taxonomically and thematically during free recall (e.g., Jenkins & Russell, 1952), and that thematic relations facilitate the learning and memory of texts and stories (e.g., Bower et al., 1979; Seifert, McKoon, Abelson, & Ratcliff, 1986). Indeed, much research has demonstrated that when a to-be-remembered list includes words related to a given theme (e.g., THREAD, SEW, SHARP), people often incorrectly remember reading the theme word (NEEDLE; Deese, 1959; Roediger & McDermott, 1995). Ross and Murphy (1999) found that undergraduates spontaneously classify foods via both thematic categories (e.g., breakfast foods) and taxonomic categories (e.g., “vegetables”). Both taxonomic and thematic relations may facilitate memory by evoking greater elaboration during encoding and/or by acting as retrieval cues during recall. For example, Jones, Estes, and Marsh (2008) showed that individual words (e.g., FISH) are more likely to be recalled when the same thematic relation is instantiated at study and at test (e.g., FISH TANK ! FISH POND) than when a different relation occurs at test (e.g., FISH TANK ! FISH FOOD). These studies indicate that thematic relations are a salient way to categorize and remember objects. Lucariello and Nelson (1985) presented 3- and 4-year-old children with two lists, each consisting of nine words for later recall. A taxonomic list consisted of three words from three taxonomic categories (animals, foods, and clothes), and a thematic list consisted of three words each from three thematic categories (zoo animals, lunch food, and clothes put on in the morning). Words on the thematic list were more likely to be recalled, suggesting that thematic relations aid very young children’s memory more than taxonomic categories. Using a matching task (see Figure 2), Waxman and Namy (1997) asked 2-, 3-, and 4-year-old children to choose the option that “goes best with” or that “goes with” a base concept (e.g., DOG). Whereas the 2- and 3-year olds exhibited no clear preference between taxonomic (another dog) and thematic (BONE) options, the 4-year-olds consistently chose the thematic option (see also Blanchet, Dunham, & Dunham, 2001). Several other studies have also found that young children tend to choose a thematic option over a taxonomic option (Lucariello & Nelson, 1985; Nelson & Nelson, 1990), and this thematic preference remains relatively constant from 4 to 7 years of age (Lucariello, Kyratzis, & Nelson, 1992). However, thematic thinking subsequently appears to decline. Siaw (1984) presented to younger (7-year-old) and older (10-year-old) children

274


a set of items that could be grouped either taxonomically or thematically. For example, BUS was included with four other items reflecting a school theme (CLASSROOM, PENCIL, STUDENT, and SWEATER) and four items from the “vehicles” category (CAR, TRUCK, TRACTOR, and WAGON). Siaw examined whether participants were more likely to recall the target item (BUS) along with other members of the school theme or along with other members of the “vehicles” category. The words were less likely to be clustered thematically than taxonomically during recall, and such thematic clustering was more common among the younger children than among the older children. To investigate the presumed conceptual shift across the lifespan, Smiley and Brown (1979) administered a series of conflict trials in the matching task to very young children (4 years), young children (6 years), older children (10 years), young adults (20 years), and older adults (72 years). From the youngest to the oldest age groups, 65%, 70%, 15%, 5%, and 70% of participants exhibited a clear tendency for thematic choices. Examining the stability of this thematic preference in later adulthood, Pennequin, Fontaine, Bonthoux, Scheuner, and Blaye (2006) found that middle-aged (45 years) and older adults (71 years) both tended to choose thematic options over taxonomic options in the matching task. This nonmonotonic pattern demonstrates a strong preference for thematic thinking in early childhood, followed by a strong tendency for taxonomic thinking in later childhood and early adulthood, and finally a reemergence of thematic thinking in middle and late adulthood. However, this developmental shift in conceptual organization has been the subject of much criticism. First, it should be noted that the thematic preference is observed only in particular tasks. Whereas choosing the option that “goes best with” the base tends to be thematic, choosing “another one” of the base tends to be taxonomic (Waxman & Namy, 1997). Likewise, asking participants to choose the picture “that is most like” the base elicits thematic choices, whereas asking them to choose the picture “that is the same kind of thing” elicits taxonomic choices (Dea´k & Bauer, 1995; see also Nguyen & Murphy, 2003). Second, participants’ preferences in such matching tasks are also context dependent. Prior to a matching task, Blaye and Bonthoux (2001) showed 3- and 5-year-old scenes designed to prime either the thematic or taxonomic option. For example, when shown a picture depicting a circus theme, the children chose the thematic pair (TAMER and WHIP) as the best match for the target LION, but when shown a picture depicting a zoo, children chose the taxonomic pair (BIRD and GIRAFFE). Finally, many studies using the matching task have confounded the relation of the options (i.e., taxonomic vs. thematic) with their similarity to the base concept. For instance, the base DOG is more perceptually similar to the taxonomic option CAT than to the thematic option BONE. Indeed, in many of these studies the taxonomic option was not only a category comember with the base (e.g., DOG and CAT) but was actually another

Thematic Thinking

275

version of the same item (e.g., another dog). Because the similarity of the options affects participants’ choices (Markman & Hutchinson, 1984; Osborne & Calhoun, 1998), this prevalent confound renders equivocal many of the conclusions from the matching task. Thus, if anything, the matching task appears to reveal a task-specific and context-dependent processing preference rather than a fundamental aspect of conceptual organization. The matching task is informative only if the options are equated on relevant factors such as their familiarity and attractiveness, and their perceptual similarity to and frequency of co-occurrence with the base. Unfortunately, such experimental controls have rarely been implemented in studies of this type. In contrast, the free association task may provide a simpler and more accurate measure of conceptual organization. Lucariello et al. (1992) used both a matching task and a free association task with 7-year-olds. Thematic responses were favored in the matching task in which they chose which option “goes with” the base, whereas taxonomic responses were more common in the word association task. This discrepancy across tasks provides further support for the conclusion that they reveal processing preferences rather than conceptual organization per se. Borghi and Caramelli (2003) instructed children (5-, 8-, and 10-year-olds) to provide from 5 to 10 associated nouns or sentences for concepts representing nine different kinds, each of which included a superordinate (e.g., FURNITURE), basic-level (e.g., CHAIR), and subordinate (e.g., HIGHCHAIR) concept. Responses were coded as taxonomic if the associate was a superordinate, subordinate, or coordinate of the cue concept, and as thematic if the associate shared a locative (e.g., DOCTOR—“hospital”), temporal (e.g., “BIRD—spring”), action/event (e.g., BIRD—“fly”), or functional (e.g., CHAIR—“to sit on”) relation. Attributive relations such as properties (e.g., CHAIR—“brown”), parts (e.g., BIRD—“beak”), and materials (e.g., CHAIR—“wood”) were scored separately from the aforementioned thematic relations. Results are illustrated in Figure 8. The percentage of taxonomic responses was constant across the three age groups. Thematic responses were the most common, but they decreased across age groups. In contrast, attributive responses increased across ages. Results from the free association task thus indicate a preference for thematic thinking that slightly decreases across childhood. In sum, people are naturally capable of both taxonomic and thematic thinking. Children appear to prefer thematic thinking, but with age, a tendency for taxonomic thinking emerges. However, the extent of this shift may have been overstated in early research, and the consensus is that people are equally capable of taxonomic and thematic thinking (Lin & Murphy, 2001; Nelson, 1977; Ross & Murphy, 1999; Smiley & Brown, 1979; Waxman & Namy, 1997). Priming paradigms such as the naming task and the lexical decision task provide an additional measure of conceptual organization, and as described next (Section 5.3), such studies corroborate the conclusion that thematic relations are highly accessible and influential in cognition.

276


60

Free associations (%)

50 40 30 20 10 0 5 years Taxonomic

8 years Thematic

10 years Attributive

Figure 8 Taxonomic, thematic, and attributive responses in the free association task by 5-, 8-, and 10-year-old children. Whereas taxonomic associations remained constant across age groups, thematic associations decreased and attributive associations increased. Results are extrapolated from Borghi and Caramelli (2003).

5.3. Language Thematic relations are essential to language comprehension on both local (e.g., word) and global (e.g., text passage) levels. As illustrated throughout Section 4 (see also Jones & Estes, in press), thematic relations facilitate the recognition and comprehension of individual words and word pairs. To reiterate, a target word is recognized faster after a thematically related prime (e.g., SOUP ! CAN) than after an unrelated prime (e.g., COW ! CAN; Estes & Jones, 2009; Jones, 2010), and word pairs are more quickly understood if they can be integrated with a thematic relation that has occurred either frequently (Gagne´ & Shoben, 1997; Storms & Wisniewski, 2005) or recently (e.g., Estes, 2003b; Estes & Jones, 2006; Gagne´, 2001). Such thematic priming of language emerges early in childhood. Perraudin and Mounoud (2009) found that 5-year-old children exhibited a robust priming effect in a naming task for thematic word pairs with an instrumental relation (e.g., KNIFE ! BREAD), but only a marginal effect for taxonomic word pairs (e.g., CAKE ! BREAD). In contrast, 7- and 9-year-old children exhibited both instrumental and taxonomic priming. McCauley, Weil, and Sperber (1976) obtained a similar developmental trajectory with associated primes and targets, many of which were thematically related (e.g., BONE ! DOG, NEEDLE ! DOCTOR, MONKEY ! BANANA, FLOWER ! BEE), and such instrumental and script-based priming are also reliably observed among undergraduates (Hare et al., 2009; Moss et al., 1995).

Thematic Thinking

277

In addition to word pairs, thematic integration also occurs among the subject, verb, and object of whole sentences. Nakano and Blumstein (2004) presented sentence frames in which the subject and verb could both be either real words or nonwords (e.g., “The bartender/quajeter is kicking/thazing out the . . .”). They then completed the sentence with either a word or a nonword at the final object position, and participants made lexical decisions to that final word/nonword. Relative to the nonword control sentences (e.g., “The quajeter is thazing out the”), prime sentences with real subject and verb words (e.g., “The bartender is kicking out the . . .”) facilitated responses to target words (e.g., “drunk”), thus indicating facilitation of thematically related sentences. Relative to those control sentences, however, prime sentences with either a nonword subject (e.g., “The quajeter is kicking out the . . .”) or a nonword verb (e.g., “The bartender is thazing out the . . .”) failed to prime the target word (i.e., “drunk”). This latter finding demonstrates that neither the subject nor the verb alone was sufficient for thematic priming. Rather, priming was obtained only with the successive thematic integration of the subject, verb, and object. When a global context conflicts with the local context, thematic integration will occur at the highest possible discourse level. Hess, Foss, and Carroll (1995) found that target nouns (e.g., “poem”) were named faster following a thematically consistent local context (e.g., “The English major wrote the poem”) than after a thematically inconsistent context (e.g., “The computer science major wrote the poem”). But when the thematically related local context followed a scenario in which the global context was thematically unrelated (e.g., a two-sentence description of an English major struggling with a computer science class), target naming times were determined by the global context rather than the local context. In an event-related potential study, Metusalem and colleagues (2010) presented a context story (e.g., playing in the snow) followed by a target word that was globally and locally congruent (e.g., building a SNOWMAN), globally and locally incongruent (e.g., building a TOWEL), or globally congruent but locally incongruent (e.g., building a JACKET). The latter condition is of particular interest here, as its global congruence is based on a thematic relation (i.e., between SNOW and JACKET). As expected, the N400 amplitudes, which are indicative of incongruence between a context and a target word, were large in the globally and locally incongruent condition and were small in the globally and locally congruent condition. Critically though, the globally congruent but locally incongruent words elicited N400 amplitudes of intermediate size. Thus, within 400 ms participants had detected the word’s incongruence with the local context (i.e., building a JACKET), but that incongruence was ameliorated by the word’s thematic congruence with the context (i.e., SNOW and JACKET). As these studies illustrate, plausibility is crucial to thematic integration on both the local and global levels. For example, “The pirate terrorized the sailor” and “The sailor terrorized the pirate” are both syntactically clear and

278


straightforward, but because the former is more thematically consistent (i.e., plausible) than the latter, it is understood more quickly (e.g., Boland, Tanenhaus, Garnsey, & Carlson, 1995). Altmann (1999) demonstrated that we rely on our knowledge of plausible real-world events during the incremental integration of each incoming word of a sentence. Participants read target sentences (e.g., “He delivered some machine guns to the military base next door”) after an antecedent sentence that rendered the target either plausible (e.g., “Hank parked his van outside the local military base”) or implausible (e.g., “Hank parked his van outside the preschool nursery”). As they read the target sentence, participants judged whether each individually presented word continued to make sense. Even though the target sentence itself was entirely sensible, there were more “no” responses for the object of the verb (e.g., “machine” and “guns”) following the implausible antecedent than the plausible antecedent. More locally, Costello and Keane (2000) included plausibility as one of only three major constraints on understanding pairs of individual words. Plausibility and its associated effects depend upon the thematic fit between the various constituents of the phrase, sentence, or passage (for models of plausibility see Connell & Keane, 2006; Pado´, Crocker, & Keller, 2009). Thematic fit is based on one’s knowledge and real-world experience of objects and events (McRae, Spivey-Knowlton, & Tanenhaus, 1998), and consequently, thematic relations constrain the set of candidate words that are likely to follow a preceding word or context (McRae & Matsuki, 2009). For example, Hare and colleagues (2009) showed that object words are understood faster after their typical events (e.g., PICNIC ! BLANKET), locations (e.g., GARAGE ! CAR), and instruments (e.g., OVEN ! COOKIES). Judgments of a target noun are also faster when presented with a verb that sets up a thematic context (e.g., PAYING ! CUSTOMER and SERVING ! CUSTOMER) than when presented with a thematically unrelated verb (e.g., GOVERNING ! CUSTOMER). Such thematically constraining verbs also reliably prime their thematically related instruments (e.g., STIRRED ! SPOON), but surprisingly, no priming occurs between verbs and thematically related locations (e.g., SWAM ! OCEAN). Ferretti, McRae, and Hatherell (2001) argued that this selective priming of instruments but not locations indicates that thematic verbs more strongly constrain the set of prototypical instruments than the set of typical locations. For example, one most frequently stirs things with a SPOON, but one could swim in a number of locations (e.g., POOL, LAKE, RIVER, SEA, OCEAN). Ferretti, Kutas, and McRae (2007), however, found that thematic verbs did reliably prime their thematically related locations when the verb was in the past imperfect form (e.g., WAS SWIMMING ! OCEAN) but not in the past perfect form (e.g., HAD SWAM ! OCEAN). The imperfect aspect of the verb denotes an ongoing action, which renders salient its location. In contrast, the perfect aspect denotes a completed action, which evidently renders its location less relevant.

Thematic Thinking

279

Thus, thematic relations facilitate comprehension of typical agents, patients, instruments, and locations of objects and events. It is no exaggeration to state that thematic integration provides the underlying mechanism by which language is understood. Without thematic integration, individual words could not be integrated into phrases, sentences, and entire texts or conversations.

5.4. Inference and Analogy A rapidly growing body of research has begun to illuminate a deep relationship between thematic thinking and analogical reasoning. In addition, although relatively little research has examined the role of thematic relations in more basic inferential reasoning, preliminary results suggest that thematic thinking also supports some inferences. Inference is typically based on taxonomic knowledge. For instance, the taxonomic knowledge that CRICKET is a sport allows one to validly infer, even if one is unfamiliar with the sport, that it requires physical effort and (often) results in a winner. However, thematic knowledge can also support inference. The thematic knowledge that CRICKET involves a BALL and a BAT allows one to infer that there must an athlete who delivers the ball (i.e., a BOWLER) and another who attempts to hit it (i.e., a BATSMAN). Lin and Murphy (2001) tested for thematic inference by presenting scenarios in which a base animal is related to another animal either thematically (i.e., they interact) or taxonomically. Critically, the base animal was described as having a particular bacterium, and participants judged whether the two other animals were likely to also have the same bacterium. People were more likely to infer that two animals have the same bacterium if those animals were thematically related than if they were taxonomically related (see also Saalbach & Imai, 2007, described in Section 6.2). This makes sense because bacteria are transmitted by proximity, and animals that interact with one another will have opportunities for contact. Chaigneau, Barsalou, and Zamani (2009) found similarly that the accuracy of inferences is improved when participants have knowledge about the events and situations in which objects are used. For example, participants more accurately inferred the function of a novel object when it was presented with other objects used in the same event (e.g., a projectile to be used in a catapult) than when presented in isolation. Thematic relations also support analogical inference. Comprehending an analogy requires one to recognize the relation between two source concepts (e.g., PEN : WRITE) and infer that same relation between two target concepts (e.g., SCISSORS : CUT). Indeed, the relationship between thematic thinking and analogical reasoning appears to be strong and interactive. Doumas, Hummel, and Sandhofer (2008) developed a powerful computational model in which relational concepts (including thematic relations)

280


themselves are abstracted from experience with multiple instances of analogous relations. With sufficient exposure to various causal relations, for instance, one develops CAUSE and EFFECT role concepts, which are then used to more efficiently detect and represent new causal relationships. And conversely, because many analogies involve thematic relations, Leech, Mareschal, and Cooper (2008) developed another powerful computational model in which analogical reasoning develops from the more basic process of relation priming (see Section 4.4). Essentially, they argue that analogical inference is bootstrapped from our natural propensity for apprehending thematic relations. Thus, analogies appear to enable the development of relational themes, and those thematic relations subsequently sustain more advanced analogical inference. Several studies support this presumed link between thematic thinking and analogical reasoning. Understanding an analogy activates the relation between terms (Green, Fugelsang, & Dunbar, 2006), and that relation can be recognized even after the terms have been forgotten (Kostic, Cleary, Severin, & Miller, 2010). Apprehension of the relation between source items not only facilitates relational transfer to a target pair (e.g., Bendig & Holyoak, 2009) but also facilitates retrieval of previously experienced, relationally similar examples (Gentner, Loewenstein, Thompson, & Forbus, 2009; Markman, Taylor, & Gentner, 2007). And just as literal similarity can either help or hinder analogical reasoning (Gentner & Colhoun, 2010), thematic relations can also facilitate a correct response or distract from it (Thibaut, French, & Vezneva, 2010). For example, the highly accessible thematic relation between TRAIN and TRACK induces the correct analogical inference in CAR : ROAD :: TRAIN : ??, but it decreases accuracy in CAR : PETROLEUM :: TRAIN : ??. In addition to the relation itself, the relational roles are also an important factor in analogical reasoning (Estes & Jones, 2008; Hummel & Holyoak, 1997, 2003; Morrison et al., 2004). Faced with the analogy WIND : EROSION :: SMOKE : ??, people tend to incorrectly complete the analogy with the highly accessible and thematically related FIRE. However, given that the direction of the causal relation in the source pair is cause ! effect, a more appropriate response would be SUFFOCATION. Together, these studies suggest that thematic thinking underlies analogical inference and may also influence more basic inferences.

6. Individual Differences and Cultural Effects Although thematic relations are apprehended quickly and automatically (Sections 4.1 and 4.2) from a very young age (Section 5.2), some individuals are more likely than others to think thematically, and some cultures exhibit more thematic thinking than others.

Thematic Thinking

281

6.1. Individual Differences Most people can easily identify thematic relations (see Section 5.2), but there are important and persistent individual differences in thematic thinking. Across multiple studies using the matching task with conflict trials (Figure 2), Simmons and Estes (2008) instructed participants to choose the option that is “most similar to” or most “like” the base concept. Regardless of the precise instruction, participants tended to choose either the taxonomic option across most trials or the thematic option across most trials. In other words, some participants consistently judged similarity on the basis of common features, while other participants consistently judged similarity on the basis of thematic relatedness. Individual variation in how strongly thematic relations affect similarity is quite robust. In a subsequent experiment, Simmons and Estes replicated these individual differences using another task (i.e., similarity ratings) and another set of items that were matched for several lexical variables (i.e., length, written frequency, forward and backward association, and lexical co-occurrence). Furthermore, Golonka and Estes (2009) showed that individual differences in the matching task strongly predict similarity ratings of an additional set of concepts. This indicates that the individual differences persist across tasks. These results are also consistent with the categorization literature, which reports individual differences in the likelihood of forming thematic categories. Lin and Murphy (2001; see also Murphy, 2001) found considerable individual differences in the way participants grouped objects in the triad task. When selecting which option “Goes with [the base concept] to form a category,” many participants choose a thematic option and many others choose a taxonomic option. Lin and Murphy replicated this basic finding with various instructions (“Which two of the three items form a category?”), stimuli (concept labels from Smiley & Brown, 1979), and tasks (speeded category judgments, novel property generalization). Thus, as with similarity judgments, there are pervasive individual differences in how strongly thematic relations affect categorization. Table 1 summarizes the percentages of participants who consistently chose taxonomic options or thematic options, as well as those who exhibited no consistent preference for either option-type, in the various studies by Lin and Murphy and by Simmons and Estes (2008). This table suggests two notable conclusions: (1) about 80% of the participants from these samples of North American undergraduates consistently chose either taxonomic or thematic options and (2) taxonomic thinking and thematic thinking were equally popular, with about 40% of the participants in each group. Simmons and Estes (2008) further examined whether these individual differences were predictable. In addition to administering the matching task, they also measured individuals’ enjoyment of thinking and problem solving (i.e., “Need For Cognition,” NFC; Cacioppo & Petty, 1982), and they had

282


Table 1 Numbers of Participants Who Consistently Chose Taxonomic Options (Taxonomic) or Thematic Options (Thematic), or Who Exhibited No Consistent Preference for Either Option-Type (No Preference), in Conflict Trials of the Matching Tasks Administered by Lin and Murphy (2001) and Simmons and Estes (2008). Source

Study

Taxonomic

No preference

Thematic

Lin and Murphy (2001)

1 2 3 4 5 7 8 1a 1b 2a 2b 3 N %

11 14 2 6 1 7 7 11 11 30 20 12 132 38.37

0 3 10 6 13 0 3 8 4 4 4 15 70 20.35

21 15 6 6 6 9 6 16 20 21 8 8 142 41.28

Simmons and Estes (2008)

Combined

Data for Experiments 7 and 8 from Lin and Murphy include the control condition only.

participants explicitly identify the factors that they considered relevant to similarity judgments. NFC and explicit beliefs both reliably predicted participants’ similarity judgments. Specifically, a preference for thematic relations in similarity judgments was associated with low NFC and an explicit belief that participation in a common scenario is relevant to similarity. There thus appear to be two types of people who prefer thematic thinking to taxonomic thinking: those who do not particularly enjoy engaging in cognitive activities and those who hold a contextual conception of similarity. Some evidence suggests that these individual differences emerge in childhood and reflect differences in language learning and play behavior. Waxman and Namy (1997) found consistent individual differences in thematic and taxonomic choices among 4-year-olds but not among 3-yearolds, indicating that such preferences for thematic thinking (or taxonomic thinking) emerge in early childhood. Dunham and Dunham (1995) found that children who engaged in object identity play (i.e., pointing to objects) and who used language referring to objects tended to categorize taxonomically. In contrast, children who engaged in relational identity play (i.e., focusing on interactions between objects) and who used language referring to the relations between objects tended to categorize thematically. Such individual differences may also be related to formal education. As reported above, enjoyment of thinking and problem solving (i.e., high

Thematic Thinking

283

NFC) is associated with a decreased likelihood of thematic thinking (Simmons & Estes, 2008). Older adults, who are far removed from formal schooling, are also more likely than younger adults and school-age children to classify on the basis of thematic relations (Annett, 1959; Overcast, Murphy, Smiley, & Brown, 1975; Smiley & Brown, 1979). Moreover, adults who have never participated in formal education also tend to categorize on the basis of thematic rather than taxonomic relations. Luria (1976, cited in Lin & Murphy, 2001) found that adults in rural Uzbekistan during the 1930s strongly preferred to sort items thematically. For example, they grouped AXE with TREE rather than SAW, and in fact they denied that AXE and SAW made any sense together. Formal education also predicted the extent of thematic categorization in a rural Mayan population in Mexico (Sharp, Cole, & Lave, 1979). When asked to sort pictures into groups of things that belong together, many participants sorted taxonomically (e.g., grouping all food items together). However, participants with less education tended to sort items into functional, thematic categories (e.g., a food item and a utensil). Thus, there are strong and persistent individual differences in thematic thinking. These individual differences are evident in childhood, with formal education appearing to decrease the tendency for thematic thinking. To the extent that education is culturally mediated, then, cultural differences may also exist (e.g., rural Uzbekistan and rural Mexico). These possible cultural effects are examined further in the following section.

6.2. Cultural Effects Although formal education appears to discourage thematic thinking (Section 6.1), this relationship may vary across cultures. Western cultures emphasize taxonomies by attending to objects and attributes, whereas East Asian cultures emphasize themes by attending to relations and contexts (Nisbett, 2003). To illustrate, Masuda and Nisbett (2001) had Japanese and American participants watch vignettes of a simulated aquarium scene. Afterward, participants recalled what they had seen, and their recognition memory was also tested. Japanese participants were more likely than American participants to recall inert objects (e.g., plants) and background elements of the aquarium, whereas American participants were more likely to recall large moving objects (e.g., fish). Japanese participants were less likely to recognize a previously seen fish if it was presented against a novel background, whereas American participants were unaffected by this change in context. Thus, Japanese participants paid more attention to context, while American participants focused on individual objects. So whereas formal education appears to inhibit thematic thinking in some cultures (Section 6.1), the predominance of thematic thinking among educated Japanese students reveals that thematic thinking is more likely to be mediated by culture than by education

284


per se. Indeed, thematic thinking seems to be more common among welleducated Chinese and Japanese than among well-educated Europeans and Americans. For example, Ji, Zhang, and Nisbett (2004) found that European Americans tended to group triads of objects using taxonomic relations, whereas Chinese preferred thematic categorization (see also Chiu, 1972). Some research suggests that these differences arise from early socialization (Bornstein, Azuma, Tamis-LeMonda, & Ogino, 1990; Fernald & Morikawa, 1993). American parents emphasize object attributes, whereas Eastern parents emphasize relations. In a cross-cultural study of infant-directed language, Fernald and Morikawa (1993) observed differences in the way Japanese and American mothers speak to their children about an object of play (e.g., a toy dog). American mothers used more noun labels (object focused), while Japanese mothers used more onomatopoeic labels and social routines (relation focused). So, an American mother might describe a toy dog with “Look! It has four legs and a tail,” whereas a Japanese mother might identify the dog with “Look! A Woof-woof. Hello, goodbye.” It should be noted that cultural differences in thematic thinking are not yet well understood. Across three different tasks Saalbach and Imai (2007) found inconsistent cultural effects. In a matching task, they found that Chinese participants and German participants did not differ in their preference for thematic and taxonomic options. In a similarity rating task, both groups rated taxonomic pairs (TOWEL and HANDKERCHIEF) to be more similar than thematic pairs (TOWEL and SHOWER), but the magnitude of this difference was greater for German than for Chinese participants, suggesting a greater differentiation of taxonomic and thematic relations among the Germans than among the Chinese. In an induction task (e.g., what is the likelihood that TOWEL and HANDKERCHIEF carry the same bacteria?), however, German participants thought that taxonomic and thematic pairs were equally likely to carry the same bacteria, whereas Chinese participants judged that taxonomic pairs were more likely to share the bacteria. Thus, the categorization task exhibited equivalent preferences for thematic grouping, the similarity task exhibited greater differentiation among German participants, and the induction task exhibited greater differentiation among Chinese participants. So thematic thinking may vary across cultures, but thematic and taxonomic relations are both evident across cultures. Interestingly, language also appears to mediate thematic thinking within a given culture. Ji and colleagues (2004) compared grouping on the matching task by Chinese students who were exposed to English very early in schooling (i.e., from Hong Kong and Singapore) to those who were exposed to English in secondary school or later (i.e., from Mainland China and Taiwan). Regardless of when they learned English, the Chinese students were more likely than American students to categorize thematically. However, depending on the Chinese students’ age of English acquisition, their categorization was affected by the language in which they were

Thematic Thinking

285

tested. Specifically, the Chinese students who acquired English early in life were equally likely to group the objects thematically when tested in Chinese and in English. The Chinese students who acquired English later in life, though, were more likely to group taxonomically when tested in English than in Chinese. These results suggest that culture and language have unique and independent effects on thematic categorization. In summary, the consensus is that individual and cultural differences in thematic thinking reflect subtle biases in people’s tendency to attend taxonomic or thematic relations rather than large differences in conceptual knowledge (Lin & Murphy, 2001; Simmons & Estes, 2008; Smiley & Brown, 1979). However, these individual and cultural differences reflect more than a fleeting preference. The tendency to use thematic relations in similarity and categorization is related to stable phenomena such as NFC (Simmons & Estes, 2008) and cultural norms (Chiu, 1972; Ji et al., 2004), and it is predicted by language learning (Dunham & Dunham, 1995; Ji et al., 2004) and formal education (Luria, 1976; Overcast et al., 1975; Sharp et al., 1979).

7. Conclusion In this section, we highlight some areas that we consider important topics for further research, and finally we conclude by summarizing the current state of knowledge on thematic thinking and by considering its role in cognition more generally.

7.1. Future Directions The current knowledge of thematic thinking has come almost entirely from basic cognitive research. Although this approach may be optimal for understanding the properties and processes of thematic thinking, it yields an unnecessarily limited understanding of the practical implications of thematic thinking. We suggest that thematic thinking is now well enough understood to begin supplementing the basic cognitive research with more applied research. Here we describe some recent investigations of thematic thinking in just two applied domains, but we hope that future studies will address other practical implications. Thematic relations have profound effects on social cognition. For instance, social interactions can be considered thematic relations because they are external (i.e., they arise between two or more people) and complementary (i.e., people typically play reciprocal roles in interactions). The quality of social relations affects attitudes toward members of social or cultural groups, as formalized in intergroup contact theory (Allport, 1954; Crisp & Turner, 2009; Pettigrew, 1998; Stathi & Crisp, 2008; Turner, Hewstone, & Voci, 2007; Turner, Hewstone, Voci, Paolini, & Christ, 2007). Positive

286


interactions between members of different social, religious, and ethnic groups decrease prejudice (Allport, 1954; Pettigrew, 1998; Pettigrew & Tropp, 2006), for instance, by increasing positive intergroup attitudes (Stephan & Rosenfield, 1978) and by decreasing anxiety (Voci & Hewstone, 2003). Negative interactions, in contrast, decrease liking (Allport, 1954). The intergroup contact effect is robust enough that even imagining contact with an outgroup member seems sufficient to decrease prejudice (Stathi & Crisp, 2008; Turner, Crisp, & Lambert, 2007). Social thematic relations thus are powerful moderators of intergroup attitudes. It is likely that thematic relations, in the form of social interactions and relationships, also moderate many other social behaviors involving similarity, categorization, and inference. We view this as a particularly fertile area for more applied research. Thematic relations are also likely to emerge as an important factor in management and marketing research, which is currently dominated by taxonomic thinking. Corporate executives, new product developers, and brand managers are trained not to stray beyond the firm’s taxonomically defined “core competence.” Hence, the current practice is to acquire taxonomically similar companies (e.g., Kraft’s 2010 acquisition of Cadbury), to rejuvenate old products by simply adding taxonomically related features (e.g., camera phones), and to extend one’s brand only to taxonomically related products (e.g., BMW motorcycles). Quite recently, however, a trend has appeared for thematic relations in management. FedEx acquired Kinko’s because their two services complemented one another (i.e., print and ship services), Nikeþ integrates Nike running shoes with the Apple iPod (i.e., many people listen to music while exercising), and several sports brands such as Adidas have recently extended into the deodorant market (i.e., exercise causes sweat, which requires deodorant). In fact, research indicates that consumers actually prefer new products that are taxonomically dissimilar but thematically related (Gibbert & Mazursky, 2009; Gill & Dube, 2007). Thus, we anticipate not only a rapid growth of research on thematic relations in management and marketing but also an increase in thematically related practices in the marketplace. We are ourselves beginning to investigate various applications of thematic thinking in both marketing (e.g., thematic brand extensions and hybrid products) and social cognition (e.g., how social relations affect similarity judgments and group membership). Thematic thinking certainly has strong implications for many other applied domains, which we are optimistic will be advanced theoretically with continued research.

7.2. Conclusions The current state of knowledge on thematic thinking is diverse and well developed. The main conclusions of our review are the following:

Thematic Thinking

287

By virtue of their externality and complementarity, thematic relations are distinct from mere association, scripts, ad hoc categories, and taxonomic categories (Section 2; see also Figure 1). Themes and taxons constitute different modes of thought, as evidenced by numerous neuropsychological and behavioral dissociations of thematic thinking and taxonomic or categorical thinking (Section 3). Thematic thinking typically occurs uncontrollably and quickly, and is guided primarily by the frequency and recency of experience with specific thematic relations (Section 4). Thematic thinking has cascading effects on many basic cognitive processes, including similarity, categorization, memory, language, inference, and analogy (Section 5). People are naturally capable of thematic thinking. This capacity emerges early in childhood and is maintained throughout adulthood. Even when a preference for taxonomic thinking is evident, the capacity for thematic thinking remains undiminished (Sections 4 and 5.2). The propensity to think thematically varies considerably across individuals and cultures. Such variation appears to result from interactions between language, formal education, and cultural norms (Section 6). Thematic thinking complements rather than displaces taxonomic thinking. The two modes of thought develop separately, entail distinct processing mechanisms, and contribute uniquely to cognition. Together they provide a more coherent, cohesive, and complete view of cognition.

In sum, we have considered what thematic relations are, how they are apprehended, and how they affect cognition. Without such thematic thinking, models of cognition will remain categorically limited.

REFERENCES Allport, G. W. (1954). The nature of prejudice. Oxford, England: Addison-Wesley. Altmann, G. T. M. (1999). Thematic role assignment in context. Journal of Memory and Language, 41, 124–145. Annett, M. (1959). The classification of instances of four common class concepts by children and adults. British Journal of Educational Psychology, 29, 223–236. Barsalou, L. W. (1983). Ad hoc categories. Memory & Cognition, 11, 211–227. Bassok, M., & Medin, D. L. (1997). Birds of a feather flock together: Similarity judgments with semantically rich stimuli. Journal of Memory and Language, 36, 311–336. Bendig, B. W., & Holyoak, K. J. (2009). Relational priming of analogical reasoning. In B. Kokinov, K. Holyoak, & D. Gentner (Eds.), New Frontiers in analogy research: Proceedings of the second international conference on analogy (pp. 30–36). Sofia: New Bulgarian University Press. Blanchet, N., Dunham, P. J., & Dunham, F. (2001). Differences in preschool children’s conceptual strategies when thinking about animate entities and artifacts. Developmental Psychology, 37, 791–800.

288


Blaye, A., & Bonthoux, F. (2001). Thematic and taxonomic relations in preschoolers: The development of flexibility in categorization choices. British Journal of Developmental Psychology, 19, 395–412. Boland, J., Tanenhaus, M. K., Garnsey, S., & Carlson, G. (1995). Verb argument structure in parsing and interpretation: Evidence from wh- questions. Journal of Memory and Language, 34, 774–806. Borghi, A. M., & Caramelli, N. (2003). Situation bounded conceptual organization in children: From action to spatial relations. Cognitive Development, 18, 49–60. Bornstein, M. H., Azuma, H., Tamis-LeMonda, C. S., & Ogino, M. (1990). Mother and infant activity and interaction in Japan and in the United States: I. A comparative macroanalysis of naturalistic exchanges. International Journal of Behavioral Development, 13(3), 267–287. Bower, G. H., Black, J. B., & Turner, T. J. (1979). Scripts in memory for text. Cognitive Psychology, 11, 177–220. Bradley, M. M., & Glenberg, A. M. (1983). Strengthening associations: Duration, attention, or relations? Journal of Verbal Learning & Verbal Behavior, 22, 650–666. Cacioppo, J. T., & Petty, R. E. (1982). The need for cognition. Journal of Personality and Social Psychology, 42, 116–131. Cain, K., Oakhill, J. V., & Elbro, C. (2003). The ability to learn new word meanings from context by school-age children with and without language comprehension difficulties. Journal of Child Language, 30, 681–694. Chaigneau, S. E., Barsalou, L. W., & Zamani, M. (2009). Situational information contributes to object categorization and inference. Acta Psychologica, 130, 81–94. Chiu, L. (1972). A cross-cultural comparison of cognitive styles in Chinese and American children. International Journal of Psychology, 7(4), 235–242. Chwilla, D. J., & Kolk, H. H. J. (2005). Accessing world knowledge: Evidence from N400 and reaction time priming. Cognitive Brain Research, 25, 589–606. Connell, L., & Keane, M. T. (2006). A model of plausibility. Cognitive Science, 30, 95–120. Coolen, R., van Jaarsveld, H. J., & Schreuder, R. (1991). The interpretation of isolated novel nominal compounds. Memory & Cognition, 19, 341–352. Costello, F. J., & Keane, M. T. (2000). Efficient creativity: Constraint-guided conceptual combination. Cognitive Science, 24, 299–349. Crisp, R. J., & Turner, R. N. (2009). Can imagined interactions produce positive perceptions? Reducing prejudice through simulated social contact. American Psychologist, 64, 231–240. Davidoff, J., & Roberson, D. (2004). Preserved thematic and impaired taxonomic categorisation: A case study. Language and Cognitive Processes, 19, 137–174. Dea´k, G., & Bauer, P. J. (1995). The effects of task comprehension on preschoolers’ and adults’ categorization choices. Journal of Experimental Child Psychology, 60, 393–427. Deese, J. (1959). On the prediction of occurrence of particular verbal intrusions in immediate recall. Journal of Experimental Psychology, 58, 17–22. Doughty, O. J., Lawrence, V. A., Al-Mousawi, A., Ashaye, K., & Done, D. J. (2009). Overinclusive thought and loosening of associations are not unique to schizophrenia and are produced in Alzheimer’s dementia. Cognitive Neuropsychiatry, 14, 149–164. Doumas, L. A. A., Hummel, J. E., & Sandhofer, C. M. (2008). A theory of the discovery and predication of relational concepts. Psychological Review, 115, 1–43. Dunham, P., & Dunham, F. (1995). Developmental antecedents of taxonomic and thematic strategies at 3 years of age. Developmental Psychology, 31, 483–493. Estes, Z. (2003a). A tale of two similarities: Comparison and integration in conceptual combination. Cognitive Science, 27, 911–921. Estes, Z. (2003b). Attributive and relational processes in nominal combination. Journal of Memory and Language, 48, 304–319.

Thematic Thinking

289

Estes, Z., & Glucksberg, S. (2000). Interactive property attribution in conceptual combination. Memory & Cognition, 28, 28–34. Estes, Z., & Jones, L. L. (2006). Priming via relational similarity: A copper horse is faster when seen through a glass eye. Journal of Memory and Language, 55, 89–101. Estes, Z., & Jones, L. L. (2008). Relational processing in conceptual combination and analogy. Behavioral and Brain Sciences, 31, 385–386. Estes, Z., & Jones, L. L. (2009). Integrative priming occurs rapidly and uncontrollably during lexical processing. Journal of Experimental Psychology. General, 138, 112–130. Fernald, A., & Morikawa, H. (1993). Common themes and cultural variations in Japanese and American mothers’ speech to infants. Child Development, 64, 637–656. Ferretti, T. R., Kutas, M., & McRae, K. (2007). Verb aspect and the activation of event knowledge. Journal of Experimental Psychology: Learning, Memory, and Cognition, 33, 182–196. Ferretti, T. R., McRae, K., & Hatherell, A. (2001). Integrating verbs, situation schemas, and thematic role concepts. Journal of Memory and Language, 44, 516–547. Gagne´, C. L. (2000). Relation-based combinations versus property-based combinations: A test of the CARIN theory and the dual-process theory of conceptual combination. Journal of Memory and Language, 42, 365–389. Gagne´, C. L. (2001). Relation and lexical priming during the interpretation of noun-noun combinations. Journal of Experimental Psychology: Learning, Memory, and Cognition, 27, 236–254. Gagne´, C. L., & Shoben, E. (1997). The influence of thematic relations on the comprehension of modifier-noun combinations. Journal of Experimental Psychology: Learning, Memory, and Cognition, 23, 71–87. Gagne´, C. L., & Spalding, T. L. (2004). Effect of discourse context and modifier relation frequency on conceptual combination. Journal of Memory and Language, 50, 444–455. Gentner, D., & Brem, S. K. (1999). Is snow really similar to a shovel? Distinguishing similarity from thematic relatedness. In S. C. Stoness & M. Hahn (Eds.), Proceedings of the 21 st Annual Conference of the Cognitive Science Society (pp. 179–184). Hillsdale, NJ: Lawrence Erlbaum Associates. Gentner, D., & Colhoun, J. (2010). Analogical processes in human thinking and learning. In A. von Mu¨ller & E. Po¨ppel (Series Eds.) & B. Glatzeder, V. Goel, & A. von Mu¨ller (Vol. Eds.), On thinking: Vol. 2. Towards a theory of thinking (pp. 35–48). Berlin: SpringerVerlag. Gentner, D., & Gunn, V. (2001). Structural alignment facilitates the noticing of differences. Memory and Cognition, 9, 565–577. Gentner, D., Loewenstein, J., Thompson, L., & Forbus, K. D. (2009). Reviving inert knowledge: Analogical abstraction supports relational retrieval of past events. Cognitive Science, 33, 1343–1382. Gentner, D., & Markman, A. B. (1997). Structure mapping in analogy and similarity. American Psychologist, 52, 45–56. Gibson, J. J. (1979). The ecological approach to visual perception. New York, NY: Psychology Press. Gibbert, M., & Mazursky, D. (2009). How successful would a phone-pillow be? Using dual process theory to predict the success of hybrids involving dissimilar products. Journal of Consumer Psychology, 19, 652–660. Gill, T., & Dube, L. (2007). What is a leather iron or a bird phone? Using conceptual combinations to generate and understand new product concepts. Journal of Consumer Psychology, 17, 202–217. Golonka, S. (2008). The influence of thematic relations on similarity and difference. University of Warwick Thesis submitted for PhD.

290


Golonka, S., & Estes, Z. (2009). Thematic relations affect similarity via commonalities. Journal of Experimental Psychology: Learning, Memory, and Cognition, 35, 1454–1464. Green, A. E., Fugelsang, J. A., & Dunbar, K. N. (2006). Automatic activation of categorical and abstract analogical relations in analogical reasoning. Memory & Cognition, 34, 1414–1421. Grosset, N., Barrouillet, P., & Markovits, H. (2005). Chronometric evidence for memory retrieval in causal conditional reasoning: The case of the association strength effect. Memory & Cognition, 33, 734–741. Hampton, J. A. (2006). Concepts as prototypes. In B. H. Ross (Ed.), The psychology of learning and motivation: Advances in research and theory. Vol. 46. (pp. 79–113). San Diego, CA: Elsevier Academic Press. Hare, M., Jones, M., Thomson, C., Kelly, S., & McRae, K. (2009). Activating event knowledge. Cognition, 111, 151–167. Hess, D. J., Foss, D. J., & Carroll, P. (1995). Effects of global and local context on lexical processing during language comprehension. Journal of Experimental Psychology: General, 124, 62–82. Hristova, P. (2009). Unintentional and unconscious analogies between superficially dissimilar but relationally similar simple structures. In B. Kokinov, K. Holyoak, & D. Gentner (Eds.), New Frontiers in analogy research: Proceedings of the second international conference on analogy (pp. 193–203). Sofia: New Bulgarian University Press. Hummel, J. E., & Holyoak, K. J. (1997). Distributed representations of structure: A theory of analogical access and mapping. Psychological Review, 104, 427–466. Hummel, J. E., & Holyoak, K. J. (2003). A symbolic-connectionist theory of relational inference and generalization. Psychological Review, 110, 220–263. Hutchison, K. A. (2003). Is semantic priming due to association strength or feature overlap? A micro-analytic review. Psychonomic Bulletin & Review, 10, 785–813. Jenkins, J. J., & Russell, W. A. (1952). Associative clustering during recall. Journal of Abnormal and Social Psychology, 47, 818–821. Ji, L., Zhang, Z., & Nisbett, R. E. (2004). Is it culture or is it language? Examination of language effects in cross-cultural research on categorization. Journal of Personality and Social Psychology, 87, 57–65. Jones, L. L. (2010). Pure mediated priming: A retrospective semantic matching model. Journal of Experimental Psychology: Learning, Memory, and Cognition, 36, 135–146. Jones, L. L., Estes, Z., & Marsh, R. L. (2008). An asymmetric effect of relational integration on recognition memory. The Quarterly Journal of Experimental Psychology, 61, 1169–1176. Jones, L. L., & Estes, Z. (in press). Lexical priming: Associative, semantic, and thematic influences on word recognition. In J. S. Adelman (Ed.), Visual word recognition, volume 2. Hove, UK: Psychology Press. Jones, M., & Love, B. C. (2007). Beyond common features: The role of roles in determining similarity. Cognitive Psychology, 55, 196–231. Kalenine, S., Peyrin, C., Pichat, C., Segebarth, C., Bonthoux, F., & Baciu, M. (2009). The sensory-motor specificity of taxonomic and thematic conceptual relations: A behavioral and fMRI study. Neuroimage, 44, 1152–1162. Kostic, B., Cleary, A. M., Severin, K., & Miller, S. W. (2010). Detecting analogical resemblance without retrieving the source analogy. Psychonomic Bulletin & Review, 17, 405–411. Krott, A., Gagne´, C. L., & Nicoladis, E. (2009). How the parts relate to the whole: Frequency effects on children’s interpretations of novel compounds. Journal of Child Language, 36, 85–112. Leech, R., Mareschal, D., & Cooper, R. P. (2008). Analogy as relational priming: A developmental and computational perspective on the origins of a complex cognitive skill. Behavioral and Brain Sciences, 31, 357–378.

Thematic Thinking

291

Lin, E. L., & Murphy, G. L. (2001). Thematic relations in adults’ concepts. Journal of Experimental Psychology. General, 130, 3–28. Lucariello, J., Kyratzis, A., & Nelson, K. (1992). Taxonomic knowledge: What kind and when? Child Development, 63, 978–998. Lucariello, J., & Nelson, K. (1985). Slot-filler categories as memory organizers for young children. Developmental Psychology, 21, 272–282. Lupyan, G. (2009). Extracommunicative functions of language: Verbal interference causes selective categorization impairments. Psychonomic Bulletin & Review, 16, 711–718. Luria, A. R. (1976). Cognitive development: Its cultural and social foundations. Cambridge, MA: Harvard University Press. Maguire, P., Devereux, B., Costello, F., & Cater, A. (2007). A reanalysis of the CARIN theory of conceptual combination. Journal of Experimental Psychology: Learning, Memory, and Cognition, 33, 811–821. Maguire, P., Maguire, R., & Cater, A. W. S. (2010). The influence of interactional semantic patterns on the interpretation of noun-noun compounds. Journal of Experimental Psychology: Learning, Memory, and Cognition, 36, 288–297. Maguire, P., Wisniewski, E. J., & Storms, G. (2010). A corpus study of semantic patterns in compounding. Corpus Linguistics and Linguistic Theory, 6, 49–73. Maki, W. S., & Buchanan, E. (2008). Latent structure in measures of associative, semantic, and thematic knowledge. Psychonomic Bulletin & Review, 15, 598–603. Markman, A. B., & Gentner, D. (2000). Structure mapping in the comparison process. American Journal of Psychology, 113, 501–538. Markman, A. B., Taylor, E., & Gentner, D. (2007). Auditory presentation leads to better analogical retrieval than written presentation. Psychonomic Bulletin & Review, 14, 1101–1106. Markman, A. B., & Wisniewski, E. J. (1997). Similar and different: The differentiation of basic level categories. Journal of Experimental Psychology: Learning, Memory, and Cognition, 23, 54–70. Markman, E. M., & Hutchinson, J. E. (1984). Children’s sensitivity to constraints on word meaning: Taxonomic versus thematic relations. Cognitive Psychology, 16, 1–27. Martin, R. C., & Cheng, Y. (2006). Selection demands versus association strength in the verb generation task. Psychonomic Bulletin & Review, 13, 396–401. Masuda, T., & Nisbett, R. E. (2001). Attending holistically versus analytically: Comparing the context sensitivity of Japanese and Americans. Journal of Personality and Social Psychology, 81, 922–934. Matsui, M., Sumiyoshi, T., Abe, R., Kato, K., Yuuki, H., & Kurachi, M. (2007). Impairment of story memory organization in patients with schizophrenia. Psychiatry and Clinical Neurosciences, 61, 437–440. McCauley, C., Weil, C. M., & Sperber, R. D. (1976). The development of memory structure as reflected by semantic-priming effects. Journal of Experimental Child Psychology, 22, 511–518. McRae, K., & Boisvert, S. (1998). Automatic semantic similarity priming. Journal of Experimental Psychology: Learning, Memory, and Cognition, 24, 558–572. McRae, K., & Matsuki, K. (2009). People use their knowledge of common events to understand language, and do so as quickly as possible. Language and Linguistic Compass, 3, 1417–1429. McRae, K., Spivey-Knowlton, M. J., & Tanenhaus, M. K. (1998). Modeling the influence of thematic fit (and other constraints) in on-line sentence comprehension. Journal of Memory and Language, 38, 283–312. Mervis, C. B., & Rosch, E. (1981). Categorization of natural objects. Annual Review of Psychology, 32, 89–115.

292


Metusalem, R., Kutas, M., Hare, M., McRae, K., & Elman, J. (2010). Generalized event knowledge activation during online comprehension. In S. Ohlsson & R. Catrambone (Eds.), Proceedings of the 32nd annual conference of the cognitive science society (pp. 1058–1063). Austin, TX: Cognitive Science Society. Morrison, R. G., Krawczyk, D. C., Holyoak, K. J., Hummel, J. E., Chow, T. W., Miller, B. L., et al. (2004). A neurocomputational model of analogical reasoning and its breakdown in frontotemporal lobar degeneration. Journal of Cognitive Neuroscience, 16, 260–271. Moss, H. E., Ostrin, R. K., Tyler, L. K., & Marslen-Wilson, W. D. (1995). Accessing different types of lexical semantic information: Evidence from priming. Journal of Experimental Psychology: Learning, Memory, and Cognition, 21, 863–883. Murphy, G. L. (2001). Causes of taxonomic sorting by adults: A test of the thematic-totaxonomic shift. Psychonomic Bulletin & Review, 8, 834–839. Nakano, H., & Blumstein, S. E. (2004). Deficits in thematic integration processes in Broca’s and Wernicke’s aphasia. Brain and Language, 88, 96–107. Nation, K., & Snowling, M. J. (1999). Developmental differences in sensitivity to semantic relations among good and poor comprehenders: Evidence from semantic priming. Cognition, 70, B1–B13. Nelson, D. L., McEvoy, C. L., & Schreiber, T. A. (1998). The University of South Florida word association, rhyme, and word fragment norms. http://www.usf.edu/FreeAssociation/ Accessed on 5 January 2011. Nelson, K. (1977). The syntagmatic-paradigmatic shift revisited: A review of research and theory. Psychological Bulletin, 84, 93–116. Nelson, K., & Nelson, A. P. (1990). Category production in response to script and category cues by kindergarten and second-grade children. Journal of Applied Developmental Psychology, 11, 431–446. Nguyen, S. P., & Murphy, G. L. (2003). An apple is more than just a fruit: Cross-classification in children’s concepts. Child Development, 74, 1783–1806. Nisbett, R. E. (2003). The geography of thought: How Asians and Westerners think differently . . . and why. New York, NY: Free Press. Osborne, J. G., & Calhoun, D. O. (1998). Themes, taxons, and trial types in children’s matching to sample: Methodological considerations. Journal of Experimental Child Psychology, 68, 35–50. Overcast, T. D., Murphy, M. D., Smiley, S. S., & Brown, A. L. (1975). The effects of instructions on recall and recognition of categorized lists by the elderly. Bulletin of the Psychonomic Society, 5, 339–341. Pado´, U., Crocker, M. W., & Keller, F. (2009). A probabilistic model of semantic plausibility in sentence processing. Cognitive Science, 33, 794–838. Pennequin, V., Fontaine, R., Bonthoux, F., Scheuner, N., & Blaye, A. (2006). Categorization deficit in old age: Reality or artefact? Journal of Adult Development, 13, 1–9. Perraudin, S., & Mounoud, P. (2009). Contribution of the priming paradigm to the understanding of the conceptual developmental shift from 5 to 9 years of age. Developmental Science, 12, 956–977. Pettigrew, T. F. (1998). Intergroup contact theory. Annual Review of Psychology, 49, 65–85. Pettigrew, T. F., & Tropp, L. R. (2006). A meta-analytic test of intergroup contact theory. Journal of Personality and Social Psychology, 90, 751–783. Pothos, E. M., & Chater, N. (2002). A simplicity principle in unsupervised human categorization. Cognitive Science: A Multidisciplinary Journal, 26, 303–343. Raffray, C. N., Pickering, M. J., & Branigan, H. P. (2007). Priming the interpretation of noun-noun combinations. Journal of Memory and Language, 57, 380–395. Risko, V. J., & Alvarez, M. C. (1986). An investigation of poor readers’ use of a thematic strategy to comprehend text. Reading Research Quarterly, 21, 298–316.

Thematic Thinking

293

Roediger, H. L., III, & McDermott, K. B. (1995). Creating false memories: Remembering words not presented in lists. Journal of Experimental Psychology: Learning, Memory, and Cognition, 21, 803–814. Rosch, E. (1975). Cognitive representations of semantic categories. Journal of Experimental Psychology: General, 104, 192–233. Ross, B. H., & Murphy, G. L. (1999). Food for thought: Cross-classification and category organization in a complex real-world domain. Cognitive Psychology, 38, 495–553. Saalbach, H., & Imai, M. (2007). Scope of linguistic influence: Does a classifier system alter object concepts? Journal of Experimental Psychology. General, 136, 485–501. Sachs, O., Weis, S., Krings, T., Huber, W., & Kircher, T. (2008). Categorical and thematic knowledge representation in the brain: Neural correlates of taxonomic and thematic conceptual relations. Neuropsychologia, 46, 409–418. Sachs, O., Weis, S., Zellagui, N., Huber, W., Zvyagintsev, M., Mathiak, K., et al. (2008). Automatic processing of semantic relations in fMRI: Neural activation during semantic priming of taxonomic and thematic categories. Brain Research, 1218, 194–205. Sass, K., Sachs, O., Krach, S., & Kircher, T. (2009). Taxonomic and thematic categories: Neural correlates of categorization in an auditory-to-visual priming task using fMRI. Brain Research, 1270, 78–87. Schank, R. C., & Abelson, R. P. (1977). Scripts, plans, goals, and understanding: An inquiry into human knowledge structures. Oxford, England: Lawrence Erlbaum. Scheuner, N., Bonthoux, F., Cannard, C., & Blaye, A. (2004). The role of associative strength and conceptual relations in matching tasks in 4- and 6-year-old children. International Journal of Psychology, 39, 290–304. Seifert, C. M., McKoon, G., Abelson, R. P., & Ratcliff, R. (1986). Memory connections between thematically similar episodes. Journal of Experimental Psychology: Learning, Memory, and Cognition, 12, 220–231. Sharp, D., Cole, M., & Lave, C. (1979). Education and cognitive development: The evidence from experimental research. Monographs of the Society for Research in Child Development, 178, 1–109. Siaw, S. N. (1984). Developmental and population comparisons of taxonomic and thematic organization in free recall. Journal of Educational Psychology, 76, 755–765. Simmons, S., & Estes, Z. (2008). Individual differences in the influence of thematic relations on similarity and difference. Cognition, 108, 781–795 [Note: S. Simmons now publishes as S. Golonka]. Smiley, S. S., & Brown, A. L. (1979). Conceptual preference for thematic or taxonomic relations: A nonmonotonic age trend from preschool to old age. Journal of Experimental Child Psychology, 28, 249–257. Snyder, H. R., & Munakata, Y. (2008). So many options, so little time: The roles of association and competition in undetermined responding. Psychonomic Bulletin & Review, 15, 1083–1088. Spellman, B. A., Holyoak, K. J., & Morrison, R. G. (2001). Analogical priming via semantic relations. Memory & Cognition, 29, 383–393. Spence, D. P., & Owens, K. C. (1990). Lexical co-occurrence and associative strength. Journal of Psycholinguistic Research, 19, 317–330. Stathi, S., & Crisp, R. J. (2008). Imagining intergroup contact promotes projection to outgroups. Journal of Experimental Social Psychology, 44, 943–957. Stephan, W. G., & Rosenfield, D. (1978). Effects of desegregation on racial attitudes. Journal of Personality and Social Psychology, 36, 795–804. Storms, G., & Wisniewski, E. J. (2005). Does the order of head noun and modifier explain response times in conceptual combination? Memory & Cognition, 33, 852–861.

294


Thibaut, J.-P., French, R., & Vezneva, M. (2010). The development of analogy making in children: Cognitive load and executive functions. Journal of Experimental Child Psychology, 106, 1–19. Thompson-Schill, S. L., Kurtz, K. J., & Gabriele, J. D. E. (1998). Effects of semantic and associative relatedness on automatic priming. Journal of Memory and Language, 38, 440–458. Titone, D., Libben, M., Niman, M., Ranbom, L., & Levy, D. L. (2007). Conceptual combination in schizophrenia: Contrasting property and relational interpretations. Journal of Neurolinguistics, 20, 92–110. Turner, R. N., Crisp, R. J., & Lambert, E. (2007a). Imagining intergroup contact can improve intergroup attitudes. Group Processes & Intergroup Relations, 10, 427–441. Turner, R. N., Hewstone, M., & Voci, A. (2007b). Reducing explicit and implicit outgroup prejudice via direct and extended contact: The mediating role of self-disclosure and intergroup anxiety. Journal of Personality and Social Psychology, 93, 369–388. Turner, R. N., Hewstone, M., Voci, A., Paolini, S., & Christ, O. (2007c). Reducing prejudice via direct and extended cross-group friendship. European Review of Social Psychology, 18, 212–255. Tversky, A. (1977). Features of similarity. Psychological Review, 84, 327–352. Voci, A., & Hewstone, M. (2003). Intergroup contact and prejudice toward immigrants in Italy: The meditational role of anxiety and the moderational role of group salience. Group Processes & Intergroup Relations, 6, 37–52. Waxman, S. R., & Namy, L. L. (1997). Challenging the notion of a thematic preference in young children. Developmental Psychology, 33, 555–567. Wilkenfeld, M. J., & Ward, T. B. (2001). Similarity and emergence in conceptual combination. Journal of Memory and Language, 45, 21–38. Wisniewski, E. J. (1996). Construal and similarity in conceptual combination. Journal of Memory and Language, 35, 434–453. Wisniewski, E. J., & Bassok, M. (1999). What makes a man similar to a tie? Stimulus compatibility with comparison and integration. Cognitive Psychology, 39, 208–238. Wisniewski, E. J., & Love, B. C. (1998). Relations versus properties in conceptual combination. Journal of Memory and Language, 38, 177–202.

Subject Index

A Abstract concepts abstract transfer, 227 approach/avoidance effects, 225–226 category learning, 162 conceptual metaphor theory (see Conceptual metaphor theory) dual code theory, 228–229 emotional valence, 225–226 force dynamics, 227–228 grounded cognition approach action, role of, 222 democracy, 219 mental symbols, 218 perceptual simulations, 220–222 perceptual symbols, 219–220 thinking, 218 hybrid models, 231 linguistic co-occurrence models, 229–231 sensory-motor processing amodal representations, 223–224 LASS theory, 224 scope problem, 223, 224 situations, role of context availability model, 232 exemplar models, 231–232 Action compatibility effect (ACE), 222 Agenda-based-regulation (ABR) framework habitual responses, 124 information-processing model, 121–123 item selection components, 127 placement of items, 127 presentation format, 126 region of proximal learning, 126 sequential format, 128 STEM effect, 125 learning goals, 123 online monitoring and control, 125 self-paced study accuracy-emphasized instructions, 130 discrepancy-reduction mechanism, 129 jROL, 130 speed-emphasized instructions, 130 C Category learning abstract concepts, 162

cognitive task, 168–169 coherence-based approach bundled features, 145–146 generalizations, 146 processing constraints, 145 selectivity, 146–147 statistical regularity, 145 deterministic category learning behavioral dissociations, 183–186 family-resemblance category, 179 neuroimaging, 186–187 neuropsychology dissociations, 179–183 explanations, variety of, 207–208 fairness and reciprocity, 162 human and nonhuman species, 142 knowledge-based approach conceptual knowledge, 144 poverty of stimulus, 143–144 representational constraints, 144–145 labels, role of auditory and visual input, 156–157 classification task, 154 co-occurrence probability, 158–159 illusory correlation, 159 immaturity, 161–162 induction task, 158–159 picture identification task, 159 predictive values, 160 semantic similarity, 157–159 SINC, 154–156 training and transfer phase, 160–161 mathematical modeling ALCOVE model, 205 candidate systems, 199 COVIS, 200–201, 204–206 formal models, nature and use of, 201–204 hybrid models, 200 MonteCarlo sampling, 204 representational schemes, 199–200 system, definition of, 199 multiple-system approach, 169–170, 206, 208–209 probabilistic category learning behavioral dissociations, 174–177 habit learning tasks, 170–171 neuroimaging, 177–178 neuropsychology dissociations, 171–174 weather prediction task, 170–171

295

296

Subject Index

Category learning (cont.) salience DCCS task, 151–152 switch costs, 152–153 task-switch paradigm, 151 spontaneous learning accuracy, 147–149 contingency learning task, 151 match-to-sample task, 149–150 unsupervised learning, 147–149 state-trace analysis COVIS model, 189–190 dependent variables, 189, 191–192 dissociations, inferential limits of, 192–194 reanalysis, 194–196 single-system model, 189–190 state-trace plot, 190–191 two-dimensional state-trace plots ALCOVE, 198 confounds, role of, 196–197 COVIS and multiple-system models, 190–191 systems, processes, and latent variables, 198–199 Coherence-based approach, category learning bundled features, 145–146 generalizations, 146 labels and appearances, 159–160 processing constraints, 145 selectivity, 146–147 statistical regularity, 145 Conceptual metaphor theory concrete concept, 232–233 evaluation, image schemas literal and metaphorical blends, 238 novel and conventional metaphors, 239 post-hoc category, 238 source-path-goal schema, 239 structure, 239–240 systematicity, 240 evidence, image schemas container image schema, 237 distance and similarity, 237 gesturing, 234 response selection process, 235–236 spatial attention, 236 time, 235–237 up-down image schema, 235, 241–242 priming effects, 234 Contrast model, 268–269 Cumulative distribution functions (CDF), 36 D Deterministic category learning behavioral dissociations, 183–186 family-resemblance category, 179 neuroimaging, 186–187 neuropsychology dissociations

information integration tasks, 180–181 reevaluating, 181–183 rule based tasks, 180 Dimensional change card sort (DCCS), 151–152 Discrepancy-reduction model, 110–111 Driver distraction activity duration, 32 cellular communication brake reaction time, 36 car following paradigm, 35, 36 control and supertasker performance, 52 conversation content analysis, 48 cross-plots driving speed, 38, 39 cumulative distribution functions, 36 driving environment, 44, 45 driving speed profile, 38 dual-task conditions, 45, 51 epidemiological studies, 34 ERP data, 45 fMRI analysis, 54 following distance, 36 half-recovery time, 41 hands-free cell phones, 42 highfidelity driving simulators, 34 inattention blindness hypothesis, 43 methodology, 33 Monte Carlo procedure, 53 observational studies, 34 OSPAN task, 51 patrol-sim driving simulator, 35 recognition memory, 43 single-task condition, 36, 37, 40 time to collision, 41 video analysis, 48 expert cell-phone drivers, 49 exposure rate of activity, 32 hypothetical multitasking situations, 30, 31 impairments, 30, 31 Dual code theory, 228–229 E Electroencephalogram, 68 Emergent coordination, joint action entrainment altruistic behavior, 75 cross-recurrence analysis, 73 eye movements, 73, 76 in-phase/anti-phase coordination, 75 interpersonal, 72 perceptual sensitivity task, 75 verbal interaction, 73 perception–action matching chameleon effect, 76 nonconscious mimicry, 74, 76 partner’s mannerisms, 74 real-time aspects, 91

297

Subject Index F Feedback loops, 16 G Grounded cognition approach action, role of, 222 concepts, flexibility of, 218 mental symbols, 218 perceptual simulations property verification, 220–221 shape similarity, 220 perceptual symbols, 219–220 simulations, 218–219 thinking, 218 H Hierarchical control, typewriting automaticity and procedural memory, 19–20 computational analysis, 5 definition of, 3–4 development of, 20–22 implicit knowledge, 19–20 informational hierarchy, 5 levels of process, 5 nested control loops, 22 preexisting skills, 19 selective influence, 5 two-loop theory (see Two-loop theory) Hybrid models abstract concepts, 231 category learning, 200 Hyperspace analogue to language (HAL), 229–230 I Image schemas, abstract concepts definition, 240–241 evaluation literal and metaphorical blends, 238 novel and conventional metaphors, 239 post-hoc category, 238 source-path-goal schema, 239 structure, 239–240 systematicity, 240 evidence container image schema, 237 distance and similarity, 237 gesturing, 234 response selection process, 235–236 spatial attention, 236 time, 235–237 up-down image schema, 235, 241–242 power, 241–242 priming effects, 234 situations, 242

Information processing model, 121–122 Inner loop theory feedback loops, 17–18 vs. outer loop interkeystroke interval, 7, 8 LEFT/right words, 10 LEght/riFT, 10 Light/rEFT words, 11 priming words, 10 response time, 7 scrambled sentences, 9 selective influence, 8–9 Simon-type effect, 11 J Joint action research action simulation inhibitory attention processes, 72 motor predictions, 71 nonjoint action situations, 72 predictive eye movements, 71 return mechanisms, inhibition of, 71, 72 affordance actor–object relations, 69 common object, 63 joint, 63 psychological experiments, 69 entrainment converging evidence, 67 coupled oscillators, 66 interpersonal, 67 rhythmic behavior, 66 social motor coordination process, 63 synchronization, 69 temporal coordination, 63 unintended coordination, 67 joint perceptions coordination achievement, 84 mental rotation task, 85 planned coordination, 65 perception–action matching automatic activation, 69 intransitive movements, 64 kinematic parameters, 71 lifting/tapping movement, 70 observed actions, 63, 70 similar actions, 69 spreading/grasping movement, 70 representational system, 61 shared intentions, 60 shared task representations cognitive mechanisms, 78–80 ensuing interaction, 77 interpersonal coordination, 65 learning, 81–82 neural mechanisms, 80–81 psychological experiments, 77

298

Subject Index

Joint action research (cont.) social modulations, 82–83 social version, 78 social interaction, 60 K Knowledge-based approach, category learning conceptual knowledge, 144 labels and appearances, 159 poverty of stimulus, 143–144 representational constraints, 144–145 L Language and situated simulation (LASS), 224 Latent semantic analysis (LSA), 229–230 M Mathematical model, category learning ALCOVE model, 205 candidate systems, 199 COVIS, 200–201, 204–206 formal models, nature and use of family resemblance, 203 measurement models, 202 neuropsychological theory, 202–204 planetary motion models, 201–202 procedural learning system, 203 structural model, 202 hybrid models, 200 MonteCarlo sampling, 204 representational schemes, 199–200 system, definition of, 199 Multiple memory systems (MMS), 209 N Nested control loops, 22 Neuroimaging deterministic category learning, 186–187 probabilistic category learning, 177–178 Neuropsychology dissociations deterministic category learning information integration tasks, 180–181 reevaluating, 181–183 rule based tasks, 180 probabilistic category learning double dissocition, 171–172 reevaluating, 172–174 selective impairment, 172 thematic relations, 257–258 O Outer loop theory feedback loops, 15–18 vs. inner loop interkeystroke interval, 7, 8

LEFT/right words, 10 LEght/riFT, 10 Light/rEFT words, 11 priming words, 10 response time, 7 scrambled sentences, 9 selective influence, 8–9 Simon-type effect, 11 P Perceptual symbol theory, 219–220 Planned and emergent coordination, synergy of action simulation individual action performance, 90 neural marker of, 89 temporal coordination, 90 triadic social interaction, 89 affordance, 87 entrainment eye movements, 86 interpersonal coupling, 85 rhythmic movements, 86 mental state attribution, 92 perception–action matching, 88–89 robots, design of, 93 Probabilistic category learning (PCL) behavioral dissociations, 174–177 habit learning tasks, 170–171 neuroimaging, 177–178 neuropsychology dissociations double dissocition, 171–172 reevaluating, 172–174 selective impairment, 172 weather prediction task, 170–171 Probabilistic model, 231 R Region of proximal learning (RPL) components, 132 jROL mechanism, 133 unlearned items, 126, 132 S Self-paced study accuracy-emphasized instructions, 130 discrepancy-reduction mechanism, 129 jROL, 130 speed-emphasized instructions, 130 Self-regulated learning (SRL) agenda-based-regulation framework habitual responses, 124 information-processing model, 121–123 item selection, 125–129 learning goals, 123 online monitoring and control, 125 self-paced study, 129–131

299

Subject Index

definition, 105–107 metacognitive approach, 107–108 study-time allocation common measures, 109 COPES model, 134 discrepancy-reduction model, 110–111 high-learning goal, 110 judgment of learning, 108, 109 monitoring-affects-control hypothesis, 110 review of, 112–118 RPL framework, 132–134 self-paced study, 132 simultaneous vs. sequential format, 110–111, 119 STEM effect, 111, 119 Sensory-motor processing. See Abstract concepts Shift-to-easier-materials (STEM) item selection, 125 RPL framework, 132 study-time allocation, 111, 119 Similarity-induction-naming-categorization (SINC), 154–156 Social thematic relations, 285–286 Spontaneous category learning, 147–151 Structural alignment model, 268–269 Study-time allocation common measures, 109 COPES model, 134 discrepancy-reduction model, 110–111 high-learning goal, 110 judgment of learning, 108 monitoring-affects-control hypothesis, 110 review of, 112–118 RPL framework, 132–134 self-paced study, 132 simultaneous vs. sequential format, 110–111, 119 STEM effect, 111, 119 T Talmy’s theory, 227–228 Taxonomic relations behavioral dissociations, 259–261 contrasting thematic roles, 256 feature-based category, 250–251 matching-to-sample task, 256–257 neuropsychological dissociations, 257–258 shared property, 255–256 Thematic relations analogical reasoning, 279–280 apprehension of conventional relations, 261 frequency, 265–266 recency, 267–268 speed, 263–265 uncontrollability, 261–263 unconventional relations, 261

cultural effects, 283–285 definition affordance/convention, 252 complementarity, 251–252 externality, 252 differentiation of ad hoc category, 255 mere associations, 253–254 scripts, 254 semantic relations, 253 individual differences formal education, 282–283 language learning, 282 matching task, conflict trials of, 281–282 play behavior, 282 similarity, 281, 282 inference, 279–280 language, 276–279 management and marketing research, 286 memory and categorization cognitive transition, 272–273 free association task, 275–276 matching task, 274–275 recall, 273–274 similarity comparison models, 268–269 stimulus compatibility, 270–272 two-dimensional model, 269–270 social cognition, 285–286 taxonomic relations behavioral dissociations, 259–261 feature-based category, 250–251 matching-to-sample task, 256–257 neuropsychological dissociations, 257–258 shared property, 255–256 Two-loop theory outer vs. inner loop interkeystroke interval, 7, 8 LEFT/right words, 10 LEght/riFT, 10 Light/rEFT words, 11 priming words, 10 response time, 7 scrambled sentences, 9 selective influence, 8–9 Simon-type effect, 11 speed and accuracy, 6 Typewriting skills automaticity and procedural memory, 19–20 development of hierarchical control, 20–22 implicit knowledge, 19–20 preexisting skills, 19 skill acquisition, 21 two loop theory, 6–7 U Unsupervised category learning, 147–149


CONTENTS OF RECENT VOLUMES

Volume 40 Different Organization of Concepts and Meaning Systems in the Two Cerebral Hemispheres Dahlia W. Zaidel The Causal Status Effect in Categorization: An Overview Woo-kyoung Ahn and Nancy S. Kim Remembering as a Social Process Mary Susan Weldon Neurocognitive Foundations of Human Memory Ken A. Paller Structural Influences on Implicit and Explicit Sequence Learning Tim Curran, Michael D. Smith, Joseph M. DiFranco, and Aaron T. Daggy Recall Processes in Recognition Memory Caren M. Rotello Reward Learning: Reinforcement, Incentives, and Expectations Kent C. Berridge Spatial Diagrams: Key Instruments in the Toolbox for Thought Laura R. Novick Reinforcement and Punishment in the Prisoner’s Dilemma Game Howard Rachlin, Jay Brown, and Forest Baker Index

Volume 41 Categorization and Reasoning in Relation to Culture and Expertise Douglas L. Medin, Norbert Ross, Scott Atran, Russell C. Burnett, and Sergey V. Blok On the Computational basis of Learning and Cognition: Arguments from LSA Thomas K. Landauer Multimedia Learning Richard E. Mayer Memory Systems and Perceptual Categorization Thomas J. Palmeri and Marci A. Flanery

Conscious Intentions in the Control of Skilled Mental Activity Richard A. Carlson Brain Imaging Autobiographical Memory Martin A. Conway, Christopher W. Pleydell-Pearce, Sharon Whitecross, and Helen Sharpe The Continued Influence of Misinformation in Memory: What Makes Corrections Effective? Colleen M. Seifert Making Sense and Nonsense of Experience: Attributions in Memory and Judgment Colleen M. Kelley and Matthew G. Rhodes Real-World Estimation: Estimation Modes and Seeding Effects Norman R. Brown Index

Volume 42 Memory and Learning in Figure–Ground Perception Mary A. Peterson and Emily Skow-Grant Spatial and Visual Working Memory: A Mental Workspace Robert H. Logie Scene Perception and Memory Marvin M. Chun Spatial Representations and Spatial Updating Ranxiano Frances Wang Selective Visual Attention and Visual Search: Behavioral and Neural Mechanisms Joy J. Geng and Marlene Behrmann Categorizing and Perceiving Objects: Exploring a Continuum of Information Use Philippe G. Schyns From Vision to Action and Action to Vision: A Convergent Route Approach to Vision, Action, and Attention Glyn W. Humphreys and M. Jane Riddoch Eye Movements and Visual Cognitive Suppression David E. Irwin

301

302

What Makes Change Blindness Interesting? Daniel J. Simons and Daniel T. Levin Index

Volume 43 Ecological Validity and the Study of Concepts Gregory L. Murphy Social Embodiment Lawrence W. Barsalou, Paula M. Niedinthal, Aron K. Barbey, and Jennifer A. Ruppert The Body’s Contribution to Language Arthur M. Glenberg and Michael P. Kaschak Using Spatial Language Laura A. Carlson In Opposition to Inhibition Colin M. MacLeod, Michael D. Dodd, Erin D. Sheard, Daryl E. Wilson, and Uri Bibi Evolution of Human Cognitive Architecture John Sweller Cognitive Plasticity and Aging Arthur F. Kramer and Sherry L. Willis Index

Volume 44 Goal-Based Accessibility of Entities within Situation Models Mike Rinck and Gordon H. Bower The Immersed Experiencer: Toward an Embodied Theory of Language Comprehension Rolf A. Zwaan Speech Errors and Language Production: Neuropsychological and Connectionist Perspectives Gary S. Dell and Jason M. Sullivan Psycholinguistically Speaking: Some Matters of Meaning, Marking, and Morphing Kathryn Bock Executive Attention, Working Memory Capacity, and a Two-Factor Theory of Cognitive Control Randall W. Engle and Michael J. Kane Relational Perception and Cognition: Implications for Cognitive Architecture and the Perceptual-Cognitive Interface Collin Green and John E. Hummel An Exemplar Model for Perceptual Categorization of Events Koen Lamberts

Contents of Recent Volumes

On the Perception of Consistency Yaakov Kareev Causal Invariance in Reasoning and Learning Steven Sloman and David A. Lagnado Index

Volume 45 Exemplar Models in the Study of Natural Language Concepts Gert Storms Semantic Memory: Some Insights From Feature-Based Connectionist Attractor Networks Ken McRae On the Continuity of Mind: Toward a Dynamical Account of Cognition Michael J. Spivey and Rick Dale Action and Memory Peter Dixon and Scott Glover Self-Generation and Memory Neil W. Mulligan and Jeffrey P. Lozito Aging, Metacognition, and Cognitive Control Christopher Hertzog and John Dunlosky The Psychopharmacology of Memory and Cognition: Promises, Pitfalls, and a Methodological Framework Elliot Hirshman Index

Volume 46 The Role of the Basal Ganglia in Category Learning F. Gregory Ashby and John M. Ennis Knowledge, Development, and Category Learning Brett K. Hayes Concepts as Prototypes James A. Hampton An Analysis of Prospective Memory Richard L. Marsh, Gabriel I. Cook, and Jason L. Hicks Accessing Recent Events Brian McElree SIMPLE: Further Applications of a Local Distinctiveness Model of Memory Ian Neath and Gordon D. A. Brown What is Musical Prosody? Caroline Palmer and Sean Hutchins Index

303


Volume 47 Relations and Categories Viviana A. Zelizer and Charles Tilly Learning Linguistic Patterns Adele E. Goldberg Understanding the Art of Design: Tools for the Next Edisonian Innovators Kristin L. Wood and Julie S. Linsey Categorizing the Social World: Affect, Motivation, and Self-Regulation Galen V. Bodenhausen, Andrew R. Todd, and Andrew P. Becker Reconsidering the Role of Structure in Vision Elan Barenholtz and Michael J. Tarr Conversation as a Site of Category Learning and Category Use Dale J. Barr and Edmundo Kronmu¨ller Using Classification to Understand the Motivation-Learning Interface W. Todd Maddox, Arthur B. Markman, and Grant C. Baldwin Index

Volume 48 The Strategic Regulation of Memory Accuracy and Informativeness Morris Goldsmith and Asher Koriat Response Bias in Recognition Memory Caren M. Rotello and Neil A. Macmillan What Constitutes a Model of Item-Based Memory Decisions? Ian G. Dobbins and Sanghoon Han Prospective Memory and Metamemory: The Skilled Use of Basic Attentional and Memory Processes Gilles O. Einstein and Mark A. McDaniel Memory is More Than Just Remembering: Strategic Control of Encoding, Accessing Memory, and Making Decisions Aaron S. Benjamin The Adaptive and Strategic Use of Memory by Older Adults: Evaluative Processing and ValueDirected Remembering Alan D. Castel Experience is a Double-Edged Sword: A Computational Model of the Encoding/ Retrieval Trade-Off With Familiarity

Lynne M. Reder, Christopher Paynter, Rachel A. Diana, Jiquan Ngiam, and Daniel Dickison Toward an Understanding of Individual Differences In Episodic Memory: Modeling The Dynamics of Recognition Memory Kenneth J. Malmberg Memory as a Fully Integrated Aspect of Skilled and Expert Performance K. Anders Ericsson and Roy W. Roring Index

Volume 49 Short-term Memory: New Data and a Model Stephan Lewandowsky and Simon Farrell Theory and Measurement of Working Memory Capacity Limits Nelson Cowan, Candice C. Morey, Zhijian Chen, Amanda L. Gilchrist, and J. Scott Saults What Goes with What? Development of Perceptual Grouping in Infancy Paul C. Quinn, Ramesh S. Bhatt, and Angela Hayden Co-Constructing Conceptual Domains Through Family Conversations and Activities Maureen Callanan and Araceli Valle The Concrete Substrates of Abstract Rule Use Bradley C. Love, Marc Tomlinson, and Todd M. Gureckis Ambiguity, Accessibility, and a Division of Labor for Communicative Success Victor S. Ferreira Lexical Expertise and Reading Skill Sally Andrews Index

Volume 50 Causal Models: The Representational Infrastructure for Moral Judgment Steven A. Sloman, Philip M. Fernbach, and Scott Ewing Moral Grammar and Intuitive Jurisprudence: A Formal Model of Unconscious Moral and Legal Knowledge John Mikhail Law, Psychology, and Morality Kenworthey Bilz and Janice Nadler

304

Protected Values and Omission Bias as Deontological Judgments Jonathan Baron and Ilana Ritov Attending to Moral Values Rumen Iliev, Sonya Sachdeva, Daniel M. Bartels, Craig Joseph, Satoru Suzuki, and Douglas L. Medin Noninstrumental Reasoning over Sacred Values: An Indonesian Case Study Jeremy Ginges and Scott Atran Development and Dual Processes in Moral Reasoning: A Fuzzy-trace Theory Approach Valerie F. Reyna and Wanda Casillas Moral Identity, Moral Functioning, and the Development of Moral Character Darcia Narvaez and Daniel K. Lapsley ‘‘Fools Rush In’’: A JDM Perspective on the Role of Emotions in Decisions, Moral and Otherwise Terry Connolly and David Hardman Motivated Moral Reasoning Peter H. Ditto, David A. Pizarro, and David Tannenbaum In the Mind of the Perceiver: Psychological Implications of Moral Conviction Christopher W. Bauman and Linda J. Skitka Index

Volume 51 Time for Meaning: Electrophysiology Provides Insights into the Dynamics of Representation and Processing in Semantic Memory Kara D. Federmeier and Sarah Laszlo Design for a Working Memory Klaus Oberauer When Emotion Intensifies Memory Interference Mara Mather Mathematical Cognition and the Problem Size Effect Mark H. Ashcraft and Michelle M. Guillaume Highlighting: A Canonical Experiment John K. Kruschke The Emergence of Intention Attribution in Infancy Amanda L. Woodward,Jessica A. Sommerville, Sarah Gerson, Annette M. E. Henderson, and Jennifer Buresh


Reader Participation in the Experience of Narrative Richard J. Gerrig and Matthew E. Jacovina Aging, Self-Regulation, and Learning from Text Elizabeth A. L. Stine-Morrow and Lisa M. S. Miller Toward a Comprehensive Model of Comprehension Danielle S. McNamara and Joe Magliano Index

Volume 52 Naming Artifacts: Patterns and Processes Barbara C. Malt Causal-Based Categorization: A Review Bob Rehder The Influence of Verbal and Nonverbal Processing on Category Learning John Paul Minda and Sarah J. Miles The Many Roads to Prominence: Understanding Emphasis in Conversation Duane G. Watson Defining and Investigating Automaticity in Reading Comprehension Katherine A. Rawson Rethinking Scene Perception: A Multisource Model Helene Intraub Components of Spatial Intelligence Mary Hegarty Toward an Integrative Theory of Hypothesis Generation, Probability Judgment, and Hypothesis Testing Michael Dougherty, Rick Thomas, and Nicholas Lange The Self-Organization of Cognitive Structure James A. Dixon, Damian G. Stephen, Rebecca Boncoddo, and Jason Anastas Index

Volume 53 Adaptive Memory: Evolutionary Constraints on Remembering James S. Nairne Digging into De´ja` Vu: Recent Research on Possible Mechanisms Alan S. Brown and Elizabeth J. Marsh Spacing and Testing Effects: A Deeply Critical, Lengthy, and At Times Discursive Review of the Literature


Peter F. Delaney, Peter P. J. L. Verkoeijen, and Arie Spirgel How One’s Hook Is Baited Matters for Catching an Analogy Jeffrey Loewenstein Generating Inductive Inferences: Premise Relations and Property Effects John D. Coley and Nadya Y. Vasilyeva From Uncertainly Exact to Certainly Vague: Epistemic Uncertainty and Approximation in Science and Engineering Problem Solving Christian D. Schunn

305

Event Perception: A Theory and Its Application to Clinical Neuroscience Jeffrey M. Zacks and Jesse Q. Sargent Two Minds, One Dialog: Coordinating Speaking and Understanding Susan E. Brennan, Alexia Galati, and Anna K. Kuhlen Retrieving Personal Names, Referring Expressions, and Terms of Address Zenzi M. Griffin Index

The Psychology of Learning and Motivation, Volume 54: Advances in Research and Theory

Causal Learning (The Psychology of Learning and Motivation: Advances in Research and Theory, Volume 34) (Psychology of Learning and Motivation)

The Psychology of Learning and Motivation, Volume 45: Advances in Research and Theory (Psychology of Learning and Motivation)

The Psychology of Learning and Motivation: Advances in Research and Theory (Psychology of Learning and Motivation, Volume 26)

The Psychology of Learning and Motivation: Advances in Research and Theory (Psychology of Learning and Motivation, Volume 26)

The Psychology of Learning and Motivation, Volume 49: Advances in Research and Theory (Psychology of Learning and Motivation)

Psychology of Learning and Motivation, Volume 40: Advances in Research and Theory (Psychology of Learning and Motivation)

The Psychology of Learning and Motivation, Volume 46: Advances in Research and Theory (Psychology of Learning & Motivation)

Perceptual Learning (The Psychology of Learning and Motivation: Advances in Research and Theory, Volume 36)

The Psychology of Learning and Motivation: Advances in Research and Theory (Psychology of Learning and Motivation, Vol. 18)

The Psychology of Learning and Motivation : Advances in Research and Theory, Volume 31

The Psychology of Learning and Motivation, Volume 53: Advances in Research and Theory

Psychology of Learning and Motivation: Advances in Research and Theory (Psychology of Learning and Motivation) (Vol. 22)

The Psychology of Learning and Motivation, Volume 51: Advances in Research and Theory

Psychology of Learning and Motivation: Advances in Research and Theory (Psychology of Learning and Motivation) (Vol. 21)

Psychology of Learning and Motivation: Advances in Research and Theory (Psychology of Learning and Motivation) (Vol. 20)

Psychology of Learning and Motivation: Advances in Research and Theory (Psychology of Learning and Motivation) (Vol. 24)

Psychology of Learning and Motivation, Volume 39: Advances in Research and Theory

Psychology of Learning and Motivation, Volume 44, Advances in Research and Theory

Psychology of Learning and Motivation: Advances in Research and Theory, Volume 33

Psychology of Learning and Motivation, Volume 35

Psychology of Learning and Motivation, Volume 04

Psychology of Learning and Motivation, Volume 05

Psychology of Learning and Motivation, Volume 30

Psychology of Learning and Motivation, Volume 28

Psychology of Learning and Motivation, Volume 13

Psychology of Learning and Motivation, Volume 08

Psychology of Learning and Motivation, Volume 09

Psychology of Learning and Motivation, Volume 38

Psychology of Learning and Motivation

The Psychology of Learning and Motivation: Advances in Research and Theory

The Psychology of Learning and Motivation, Volume 54: Advances in Research and Theory

Causal Learning (The Psychology of Learning and Motivation: Advances in Research and Theory, Volume 34) (Psychology of Learning and Motivation)

The Psychology of Learning and Motivation, Volume 45: Advances in Research and Theory (Psychology of Learning and Motivation)

The Psychology of Learning and Motivation: Advances in Research and Theory (Psychology of Learning and Motivation, Volume 26)

The Psychology of Learning and Motivation: Advances in Research and Theory (Psychology of Learning and Motivation, Volume 26)

The Psychology of Learning and Motivation, Volume 49: Advances in Research and Theory (Psychology of Learning and Motivation)

Psychology of Learning and Motivation, Volume 40: Advances in Research and Theory (Psychology of Learning and Motivation)

The Psychology of Learning and Motivation, Volume 46: Advances in Research and Theory (Psychology of Learning & Motivation)

Perceptual Learning (The Psychology of Learning and Motivation: Advances in Research and Theory, Volume 36)

The Psychology of Learning and Motivation: Advances in Research and Theory (Psychology of Learning and Motivation, Vol. 18)

The Psychology of Learning and Motivation : Advances in Research and Theory, Volume 31

The Psychology of Learning and Motivation, Volume 53: Advances in Research and Theory

Psychology of Learning and Motivation: Advances in Research and Theory (Psychology of Learning and Motivation) (Vol. 22)

The Psychology of Learning and Motivation, Volume 51: Advances in Research and Theory

Psychology of Learning and Motivation: Advances in Research and Theory (Psychology of Learning and Motivation) (Vol. 21)

Psychology of Learning and Motivation: Advances in Research and Theory (Psychology of Learning and Motivation) (Vol. 20)

Psychology of Learning and Motivation: Advances in Research and Theory (Psychology of Learning and Motivation) (Vol. 24)

Psychology of Learning and Motivation, Volume 39: Advances in Research and Theory

Psychology of Learning and Motivation, Volume 44, Advances in Research and Theory

Psychology of Learning and Motivation: Advances in Research and Theory, Volume 33

Psychology of Learning and Motivation, Volume 35

Psychology of Learning and Motivation, Volume 04

Psychology of Learning and Motivation, Volume 05

Psychology of Learning and Motivation, Volume 30

Psychology of Learning and Motivation, Volume 28

Psychology of Learning and Motivation, Volume 13

Psychology of Learning and Motivation, Volume 08

Psychology of Learning and Motivation, Volume 09

Psychology of Learning and Motivation, Volume 38

Psychology of Learning and Motivation

The Psychology of Learning and Motivation: Advances in Research and Theory

Recommend Documents