THE PSYCHOLOGY OF LEARNING AND MOTIVATION Advances in Research and Theory
VOLUME 37
This Page Intentionally Left Bla...
18 downloads
970 Views
19MB 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
THE PSYCHOLOGY OF LEARNING AND MOTIVATION Advances in Research and Theory
VOLUME 37
This Page Intentionally Left Blank
THE PSYCHOLOGY OF LEARNING AND MOTIVATION Advances in Research and Theory
EDITED BY DOUGLAS L. MEDIN DEPARTMENT OF PSYCHOLOGY NORTHWESTERN UNIVERSITY EVANSTON, ILLINOIS
Volume 37
ACADEMIC PRESS San Diego London Boston New York Sydney Tokyo Toronto
This book is printed on acid-free paper.
@
Copyright 0 1997 by ACADEMIC PRESS All Rights Reserved. No part of this publication may be reproduced or transmitted in any form or by any means, electronic or mechanical, including photocopy, recording, or any information storage and retrieval system, without permission in writing from the Publisher. The appearance of the code at the bottom of the first page of a chapter in this book indicates the Publisher’s consent that copies of the chapter may be made for personal or internal use of specific clients. This consent is given on the condition, however, that the copier pay the stated per copy fee through the Copyright Clearance Center, Inc. (222 Rosewood Drive, Danvers, Massachusetts 01923). for copying beyond that permitted by Sections 107 or 108 of the U.S.Copyright Law. This consent does not extend to other kinds of copying, such as copying for general distribution, for advertising or promotional purposes, for creating new collective works, or for resale. Copy fees for pre-1997 chapters are as shown on the title pages. If no fee code appears on the title page. the copy fee is the same as for current chapters. 0079-742 1/97 $25.00
Academic Press a division of Horcourr Brace & Company 525 B Street, Suite 1900. San Diego, California 92101-4495, USA http://www.apnet.com
Academic Press Limited 24-28 Oval Road. London NWI 7DX. UK http://www.hbuk.co.uk/ap/
International Standard Book Number: 0- 12-543337-9 PRINTED IN THE UNITED STATES OF AMERICA 97 98 99 00 01 02 QW 9 8 7 6
5
4
3 2
I
CONTENTS Contributors
................................................................................
ix
OBJECT-BASED REASONING
Miriam Bassok I. Introduction ........................................................................ 11. Separation and Contrast between Content and Structure ............ 111. Semantic Knowledge Determines How People
Represent Problems ............................................................. IV. Semantic Knowledge Affects Selection of Processing Strategies .... V. Discussion ........................................................................... References ..........................................................................
1 4
7 21 33 35
ENCODING SPATIAL REPRESENTATION THROUGH NONVISUALLY GUIDED LOCOMOTION TESTS OF HUMAN PATH INTEGRATION
Roberta L. Klatzky, Jack M. Loomis, and Reginald G. Golledge I. Navigational Concepts
..........................................................
42
11. General Assumptions and Findings Regarding
Human Navigation
...............................................................
111. Representations and Processes Underlying Shortcuts in Space
.....
IV. The Encoding-Error Model of Path Completion ........................ V. General Methodology of Reported Experiments ........................ VI. Encoding as Inferred from Reproduction and Verbal Report of Simple Paths ....................................................................... VII. Encoding Distances and Turns Inferred from Fitting the Encoding-Error Model .......................................................... V
45
51 53 58
58 62
vi
Contents
VIII. Effects of Experience on Encoding Pathway Parameters ............. IX . When Does Updating of the Homing Vector Occur? Evidence Against Moment-to-Moment Updating .................................... X . Group Individual Differences in Navigation without Vision ......... XI . General Summary and Conclusions ......................................... References ..........................................................................
66 71 74 77 81
AITENUATING INTERFERENCE DURING COMPREHENSION: THE ROLE OF SUPPRESSION
Morton Ann Gernsbacher 1. Introduction ........................................................................ I1. Attenuating Interference during Lexical Access ......................... 111. Attenuating Interference during Anaphoric Reference ................ IV . Attenuating Interference during Cataphoric Reference ............... V . Attenuating Interference during Syntactic Parsing ...................... VI . Attenuating Interference during Metaphor Interpretation ........... VII . Attenuating Interference during Inference Revision ................... VIII . Attenuating Interference and Comprehension Skill .................... IX . Summary ............................................................................ References ..........................................................................
85 88 90 92 94 95 97 98 101 102
COGNITIVE PROCESSES IN COUNTERFACTUAL THINKING ABOUT WHAT MIGHT HAVE BEEN
Ruth M . J . Byrne I . Counterfactual Thinking ........................................................ I1. Mental Models and Counterfactual Thinking ............................ I11. Three Phenomena of Counterfactual Thinking .......................... IV . Conclusions ......................................................................... References ..........................................................................
105 111 123 148 151
EPISODIC ENHANCEMENT OF PROCESSING FLUENCY
Michael E . J . Masson and Colin M . MacLeod I . Introduction
........................................................................
155
I1. Experiment Series 1: Data-Driven and Conceptually Driven
Encoding Tasks ................................................................... I11. Experiment Series 2: Comparing Masked Word Identification and Word Fragment Completion .................................................. IV . Experiment Series 3: The Question of Conscious Recollection .....
160 167 171
vii
Contents
V. Experiment Series 4: Speeded Word Reading as an Indirect Measure of Memory ............................................................. VI. Experiment Series 5: Color Naming versus World Reading and the Specificity of Priming ....................................................... VII. Experiment Series 6: Sources of Priming in Masked Word Indentification ............................................................. VIII. Experiment Series 7: Episodic Effects on Perceptual Judgments IX. Conclusion .......................................................................... References ..........................................................................
184 189
...
193 198 203 206
PRODUCTION, EVALUATION, AND PRESERVATION OF EXPERIENCES CONSTRUCTIVE PROCESSING IN REMEMBERING AND PERFORMANCE TASKS
Bruce W. A. Whittlesea I. Introduction ........................................................................ 11. Separate-Systems Assumptions: A Brief Summary ..................... 111. Selective Construction and Preservation of Experiences: Outline of the Account .................................................................... PART I CONSTRUCTIVE PRODUCTION
...................................
IV. Theme 1: Concepts Are Not Automatically Abstracted across Instances ................................................................... V. Theme 2: Memory Preserves Processing Experiences, Not Stimulus Structures ............................................................... VI. Theme 3: Selective Use of General and Particular Knowledge Is Controlled by the Stimulus Compound .................................... VII. Theme 4: Processing in Large, Familiar Domains Is Also Controlled by Specific Experiences .......................................... PART I1 CONSTRUCTIVE EVALUATION
..................................
VIII. Theme 5: Remembering Is Reconstruction, Not Retrieval ........... IX. Theme 6: The Constructive Nature of Experience ...................... References ..........................................................................
211 213 214 221 221 225 235 238 241 241 248 260
GOALS, REPRESENTATIONS, AND STRATEGIES IN A CONCEPT ATTAINMENT TASK THE EPAM MODEL
Fernand Gobet, Howard Richman, Jim Stastewski, and Herbert A. Simon I. Inter-Subject Differences and Commonalties in Performing Cognitive Tasks ...................................................................
265
Contents
viii
I1. Architecture and Learning in Task Performance ........................
111. Strategy. Goals. Attention. and Task Representation in EPAM
References
..........................................................................
....
268 271 289
AT A LOSS FROM WORDS VERBAL OVERSHADOWING OF PERCEPTUAL MEMORIES
Jonathan W. Schooler. Stephen M . Fiore. and Maria A . Brandimonte I . Three Premises of Verbal Overshadowing ................................ I1. The Modality Mismatch Assumption .......................................
111. The Availability Assumption .................................................. IV. The Recoding Interference Hypothesis .................................... V . How Does Verbalization Disrupt Perceptual Memories? ............. VI . Closing Remarks .................................................................. References ..........................................................................
Index ..........................................................................................
203 297 310 315 318 333 334 341
Numbers in parentheses indicate the pages on which the authors’ contributions begin.
Miriam Bassok (l), Department of Psychology, University of Chicago, Chicago, Illinois 60637 Maria A. Brandimonte (293), University of Pittsburgh, Pittsburgh, Pennsylvania 15260 Ruth M. J. Byrne (105), Psychology Department, University of Dublin, Trinity College, Dublin 2, Ireland Stephen M. Fiore (293), University of Pittsburgh, Pittsburgh, Pennsylvania 15260 Morton Ann Gernsbacher (85),University of Wisconsin-Madison, Madison, Wisconsin 53706 Fernand Gobet (265), Department of Psychology, Carnegie Mellon University, Pittsburgh, Pennsylvania 15213 Reginald G. Golledge (42), Department of Psychology, Carnegie Mellon University, Pittsburgh, Pennsylvania 15213 Roberta L. Klatzky (42), Department of Psychology, Carnegie Mellon University, Pittsburgh, Pennsylvania 15213 Jack M. Loomis (42), Department of Psychology, Carnegie Mellon University, Pittsburgh, Pennsylvania 15213 Colin M. MacLeod (153, Department of Psychology, University of Victoria, Victoria, B.C. V8W 3P5, Canada Michael E. J. Masson (159, Department of Psychology, University of Victoria, Victoria, B.C. V8W 3P5, Canada ix
x
Contributors
Howard Richman (265), Department of Psychology, Carnegie Mellon University, Pittsburgh, Pennsylvania 15213 Jonathan W. Schooler (293), University of Pittsburgh, Pittsburgh, Pennsylvania 15260 Herbert A. Simon (265), Department of Psychology, Carnegie MelIon University, Pittsburgh, Pennsylvania 15213 Jim Staszewski (265), Department of Psychology, Carnegie Mellon University, Pittsburgh, Pennsylvania 15213 Bruce W. A. Whittlesea (211), Department of Psychology, Simon Fraser University, Burnaby, B.C. V5A 1S6, Canada
OBJECT-BASED REASONING Miriam Bassok
I. Introduction Imagine a student who is asked to talk out loud while learning a workedout solution to a mechanics problem (Chi, Bassok, Lewis, Reimann, & Glaser, 1989). The example problem, adapted from a popular physics textbook (Halliday & Resnick, 1981), presents a figure depicting a block hanging from a ceiling (see Fig. 1). The student looks at the block, which is held by three strings tied in a knot, and reads the first sentence of the solution text: “Consider the knot at the junction of the three strings to be the body.” She stops reading and asks: “Why is the knot the body. . . and I thought the block is the body?” This episode is representative of the way in which “good learners” in Chi et al. (1989) learned from examples: They noticed the gaps in their understanding of the worked-out solutions and prompted themselves with questions that they then attempted to answer. Their answers, or “selfexplanations,” were often correct. For instance, the student who asked herself “Why is the knot the body?” realized that, in the strings example, the block exerts a force (i.e., the block’s weight) on the body (i.e., the knot) in the direction of gravity. From this she correctly inferred that a physical body is an arbitrary point of reference for which one has to perform a force analysis. This understanding, in turn, enabled her to solve successfully transfer problems that were structurally isomorphic to the strings example but that involved novel objects and novel spatial configurations (e.g., hockey sticks restraining a puck on a frictionless surface of ice). In general, Chi THE PSYCHOLOGY OF 1,FARNING AND MOTIVATION. VOL 37
1
Copynght 0 1997 by Academic Press All rights of reproduction in any form reserved 0079-7421197 $25 00
2
Miriam Bassok
The figure shows a weight (w) hung by strings. Consider the knot at the junction of the three strings to be the body.” Fig. I. The “strings” figure from a worked-out example of a physics problem used by Chi et al. (1989). From Fundamentals of Physics (p. 70) by D. Halliday and R. Resnick, 1981. New York: John Wiley & Sons. 0 1981 by John Wiley & Sons. Reprinted with permission.
et al. found a high positive correlation between successful transfer and the amount of spontaneous on-line inferences during learning (i.e., selfexplanations). In this chapter I describe results from several studies in which my collaborators and I examined how one type of spontaneous inferences, to which I refer here as object-bused inferences, affect problem solving, analogical transfer, and similarity judgments. Object-based inferences reflect people’s knowledge about the entities that appear in word problems, figures, or statements. Such inferences were responsible for the student’s surprise that the knot rather than the block is the physical body-for raising the very question that initiated the self-explanation episode described earlier. Evidently, this student expected to find correspondence between the abstract physical concepts (i.e., body, forces), to which she was introduced while reading the physics chapter, and attributes of the specific objects (i.e., a block, strings) that instantiated these concepts in the example problem. In particular, because she understood that the block is upheld by the strings rather than vice versa, she expected that the asymmetric functional relation between the block and the strings will be preserved in the example problem: that the block will be the physical body and the strings the physical forces. The studies I describe here are divided into two sets that show, respectively, effects of object-based inferences on the way people represent and process the stimuli they encounter. The first set of studies demonstrates that people include object-based inferences in the mental representations they construct for formally isomorphic word problems. For example, college students in Bassok, Wu, and Olseth (1995) constructed different equations when solving mathematically isomorphic problems involving semantically symmetric versus semantically asymmetric object sets (e.g., two sets of
Object-Based Reasoning
3
children vs children and prizes, respectively). The second set of studies demonstrates that object-based inferences lead to selective application of both basic processing mechanisms (e.g., comparison vs integration) and solution procedures acquired in formal training (e.g., addition vs division). For example, when asked to construct addition problems involving a given pair of object sets, college students readily added apples and oranges but refrained from adding apples and baskets (Bassok, Chase, & Martin, 1997).’ Because people’s experience with the world is not arbitrary, they exploit semantic knowledge much like domain experts exploit nonarbitrary patterns of information in the domain of their expertise (e.g., Chi, Feltovich, & Glaser, 1981;Hinsley, Hayes, & Simon, 1977; Larkin, McDermott, Simon, & Simon, 1980). For example, when selecting an arithmetic operation, people are guided by knowledge that adding apples and oranges is both a more likely and a more meaningful operation than adding apples and baskets. In general, object-based inferences foster selection of stimuli-appropriate processing procedures and thereby ensure useful and reasonable conclusions. Unfortunately, this adaptive aspect of human cognition has been overlooked by most researchers who study learning, problem solving, and reasoning. The studies reported here represent an attempt to incorporate objectbased inferences into research on analogical transfer and similarity judgments. As I describe in the next section, researchers who study analogical transfer attempt to identify regularities in the process by which people retrieve analogous problems from memory (access) and the process by which people align the representations of analogous problems (mapping). In particular, they examine the impact of similarities and differences in the content covers of analogous training (base) and test (target) problems on access and mapping. Although researchers are concerned with effects of content on transfer performance, they treat content as irrelevant information that people fail to ignore. As a result, they fail to notice that content induces semantic knowledge that affects how people represent the base and the target stimuli, and that certain semantic distinctions implied by content lead to significant changes in processing. In order to situate the studies on object-based inferences within the conceptual framework that dominates research on analogical transfer, I begin with a brief description of the paradigmatic contrast between superficial content and solution-relevant structure of analogous problems.
’
The pair apples and oranges alludes to the common saying “It’s like comparing apples and oranges.” As I explain later, the meaning of this saying (i.e., comparison is not independent of the things being compared) captures the gist of the argument I make in this chapter. Hence, I use the pair “apples-oranges” for the purpose of exposition. However, this pair of objects was not used in the experiments reported in this chapter (the closest ohject pair was peaches and plums).
4
Miriam Bassok
11. Separation and Contrast between Content and Structure
Flexible application of knowledge, be it knowledge of Newton’s laws, linear equations, or social rules, requires that people be able to recognize abstract structural similarities between situations that differ in appearance. This ability is measured on most intelligence tests. For example, people are credited with intelligence points on tests of analogical reasoning when in response to the pair “apples : baskets” they choose the pair “students :classrooms” rather than “apples :oranges.” In other words, psychologists reward people for heeding similarities in the way different objects are interrelated (e.g., CONTAIN [container, content]) rather than similarities between attributes of the specific objects that serve as the arguments, or “slot filters,” of different structural relations (e.g., apples). Using a research paradigm similar to that of analogical-reasoning intelligence tests, researchers who study analogical transfer present subjects with a solution to one or more base problems and, after a short delay which may involve an intervening task, ask them to solve one or more target problems. In most studies the base and target problems are constructed such that they share a similar structure but differ in their content covers. Some studies also use base and target problems that share similar content covers but differ in structure. This orthogonal design enables researchers to separate and contrast the impact of similarities and difference in the content and structure of base and target problems on access and mapping. To measure access, or spontaneous retrieval of a base problem from memory, researchers compare the frequencies of transfer solutions with and without informing subjects that the solution to one of the base problems can assist them in solving the target problem (i.e., compare informed and spontaneous transfer). To measure mapping, or people’s ability to successfully align the underlying structures of analogous problems, subjects are explicitly asked to solve the target problem by the method used in solving an analogous base problem. The basic finding in this line of research is that analogical transfer is highly sensitive to similarities and differences in the content covers of base and target problems (see Reeves & Weisberg, 1994, for a recent review). That is, content affects transfer even though it should not affect problem’s solutions. For example, many subjects fail to realize that a convergence solution learned in the context of a military problem can be applied to an analogous medical problem (Gick & Holyoak, 1980);that a statistics principle learned from an example about weather forecasting can be applied to a problem about arrangements of pizza toppings (Ross, 1987); or that a physics equation learned for motion can be applied to an analogous problem dealing with bushels of potatoes (Bassok & Holyoak, 1989).
Object-Based Reasoning
5
Differences between content domains of analogous problems often impair access without affecting mapping (e.g., 30% spontaneous vs 90% informed transfer of the convergence solution between a military base and a medical target in Gick & Holyoak, 1980). However, differences between the specific objects that serve as arguments of analogous problems were found to mediate mapping performance (Gentner & Toupin, 1986; Ross, 1987, 1989). For example, children in Gentner and Toupin’s (1986) study learned a base story in which a squirrel visited a frog. Then they were explicitly asked to enact an analogous target scenario involving a chipmunk and a toad (i.e., mapping). Children’s mapping performance was much better when the chipmunk visited the toad than when the toad visited the chipmunk-when the guest and the host in the base and target stories were similar (e.g., squirrel-chipmunk, frog-toad) rather than different (e.g., squirrel-toad, frog-chipmunk). Theoretical accounts of access and mapping mirror the experimental methodology by contrasting the impact of similarities and differences in content and structure on people’s performance (e.g., Falkenhainer, Forbus, & Gentner, 1989;Forbus, Genter, & Law, 1995;Hofstadter, Mitchell, & French, 1987; Holyoak & Thagard, 1989; Hummel & Holyoak, in press; Thagard, Holyoak, Nelson, & Gochfeld, 1990). Specifically, these accounts hold that access and mapping involve matches and mismatches in two distinct types of aspects that comprise the mental representations of the base and target stimuli: (1)attributes, one-place predicates that take objects as arguments (e.g., FURRY [XI,HOPPING [y]; and (2) relations, predicates that take as arguments two or more objects (e.g., VISIT [x, y]) or propositions (e.g., CAUSE {LIKE [x, y], VISIT [x, y]}). Despite some important differences in the processing assumptions of the proposed computational models, attributional and relational matches and mismatches receive different weights that capture their differential impact on access and mapping. According to Gentner’s (1983) structure-mapping theory, what is important to analogical transfer is correct alignment of the relational structures in the base and target stimuli (e.g., VISIT [x, y]), while attributes of the specific objects that instantiate these structures (e.g., FURRY [XI) can or should be ignored. This is because the distinction between relations and attributes usually coincides with the distinction between what is and what is not relevant to problem solutions (but see Holyoak, 1985). If so, one may wonder why is it that people retain superficial content in problem representations and let similarities and differences in content affect their transfer performance. Medin and Ross (1989; see also Medin & Ortony, 1989) proposed that such behavior is justified by considerations of cognitive efficiency maximization. Specifically, they argued that because content and structure are often correlated in the world (e.g., chipmunks are more likely
6
Miriam Bassok
than toads to behave like squirrels) people use similarities in content as a useful, albeit error-prone heuristic that can help them identify likely similarities in structure. As with every heuristic, people rely on similarities and differences in content to conserve cognitive resources and speed access and mapping (Hinsley et al., 1977; Novick, 1988). However, effects of similarities and differences in content are most pronounced when people do not know which aspects of the base and target problems are relevant to the problems’ solutions. That is, experts, good learners, and older children are more likely than novices, poor learners, and younger children to ignore superficial differences in content, context, and phrasing of analogous problems (e.g., Chi et al., 1981, 1989; Gentner & Rattermann, 1991; Gentner & Toupin, 1986; Novick, 1988, 1992; Schoenfeld & Herrmann, 1982; Silver, 1981). Also, training conditions that encourage abstraction from content (e.g., helping people represent the squirrel as an abstract visitor) decrease the number of mismatches in content and therefore increase the probability of successful transfer (e.g., Bassok & Holyoak, 1989; Brown, 1989; Catrambone & Holyoak, 1989; Gick & Holyoak, 1983; Gick & McGarry, 1992; Needham & Begg, 1991; Reed, 1993; Ross & Kennedy, 1990). To summarize, the conceptual framework that guides research on analogical transfer treats content (object attributes) and structure (relations between objects) as separate aspects of the stimuli and contrasts effects of similarities and differences in these two types of aspects on people’s performance. Within this view, transfer fails because people retain superficial aspects of content in problem representations, and the “undeleted” aspects of content lead to mismatches that impair access and mapping. Such effects are most likely to occur when people lack the ability, the knowledge, or the cognitive resources needed for abstracting the relevant structure from its superficial content. The gist of this view (i.e., the difficulty involved in abstracting, or separating, structure from content) is captured in several metaphors that are implied by the labels with which researchers refer to content and structure (Lakoff & Turner, 1989). Equating content with “surface” implies that structure is like a hidden treasure, and that people must invest effort and ingenuity to discover it. Equating structure with “essence” implies that, like alchemists, people need special knowledge and skills to extract the valuable structural essence of an alloy diluted by content. Describing people’s performance as “content-bound” implies that people are prisoners of content who strive to be “content-free.’’ Whatever their superficial differences, the hidden treasure, alchemist, and jail metaphors share the underlying assumption that content effects arise from failures of an imperfect cognitive system to ignore content.
Object-Based Reasoning
7
111. Semantic Knowledge Determines How People
Represent Problems In the previous section I pointed out that the paradigm by which researchers contrast people’s responses to relational and attributional similarities resembles that paradigm used in psychometric tests of analogical reasoning. Note, however, that by treating content as superficial information, researchers who study analogical transfer exclude from their theoretical accounts the very inferences for which people are credited with intelligence points on Analogies tests (e.g., the inference that apples and baskets imply containment). That is, despite the resemblance in labels and methodology, research on analogical transfer ignores the inferential process that precedes and/or accompanies access and mapping-the interpretive process by which people understand, or abstract, the structures of the base and the target problems. Researchers who study how content affects access and mapping acknowledge that these processing mechanisms operate on mental representations people construct for the base and target problems (e.g., Holyoak, 1985). That is, they acknowledge that the current neglect of interpretation is only temporary. In the meantime, they test their processing assumptions using problem representations that they themselves construct as a working hypothesis about representations that might be constructed by subjects. However, the hypothesized representations constructed by experimenters may not coincide with the representations constructed by subjects. In particular, although researchers assume that the base and the target problems share the same structure and differ only in content (i.e., mismatching object attributes), their subjects may infer from content that the problems differ in structure (i.e., mismatching inferred relations). Hence, in addition t o neglecting potentially important interpretive effects of content, researchers are risking erroneous conclusions about processing. In what follows I describe results from three studies in which the specific objects in analogous or mathematically isomorphic base and target problems induced inferences that led to abstraction of nonisomorphic interpreted structures. To foster comparison with previous work, the studies I described here used base and target problems similar to those used in previous studies on analogical transfer. A. MANNER OF CONVERGENCE
Holyoak and Koh (1987, Experiment 2) compared transfer of a convergence solution learned for two versions of a light bulb problem (base) to Ducker’s (1945) tumor problem (target). The two base problems described a research assistant who fused a broken filament of an expensive light bulb using
8
Miriam Bassok
either low-intensity lasers or low-intensity ultrasound waves. The research assistant applied the lasers (waves) simultaneously from different directions such that they converged on the broken filament. The combined intensity of the lasers (waves) at the point of convergence was sufficiently strong to fuse the broken filament. At the same time, because of their low intensity on the way to the filament, the lasers (waves) did not break the fragile glass surrounding the filament. The analogous tumor problem, which could be solved using the convergence solution, described a doctor who needed high-intensity X rays to destroy a stomach tumor, but was concerned that high-intensity rays may harm the healthy tissue surrounding the tumor. Holyoak and Koh (1987) found that the frequency of convergence solutions to the target tumor problem (i.e., spontaneous transfer) was significantly higher in the lasers than in the waves base condition (69 vs 38%, respectively). They explained this differential magnitude of spontaneous transfer by matches and mismatches between attributes of the converging forces (i.e., lasers are more similar to rays than are ultrasound waves). They also found that informing subjects about the relevance of the base solutions eliminated the effects of matches and mismatches in object attributes (75 vs 81% informed transfer from the lasers and waves versions, respectively). From this they concluded that matches and mismatches in object attributes of structurally isomorphic problems (i.e., problems with matching structures) have a more pronounced impact on access than on mapping. Tammy Smith and I examined the possibility that instead of or in addition to being mediated by direct matches and mismatches in object attributes, the results in Holyoak and Koh (1987) were mediated by matches and mismatches in the object-based interpreted structures of the base and the target problems.2 Specifically, we conjectured that people spontaneously infer the manner in which various forces converge, and treat the inferred manner of convergence as a structural constraint. That is, we conjectured that spontaneous transfer from lasers to X rays was supported by matches in manner of convergence (i.e., matching interpreted structures), while spontaneous transfer from waves to X rays was impaired by mismatches in the manner of convergence (i.e., mismatching interpreted structures). The spatial nature of the convergence solution (Beveridge & Parkins, 1987) suggests that asking subjects to draw a diagram depicting the problem’s solution can capture important characteristics of their mental representations. We employed this methodology to examine how people represent converging forces in the three story problems used by Holyoak and Koh (1987). Specifically, after modifying the text of the tumor problem to include the convergence solution, we asked undergraduate students from the University of Chicago to read one of the three convergence problems ‘This experiment was conducted by Tammy Smith as part of her B.A. honors project.
Object-Based Reasoning
9
and draw a diagram depicting how the lasers, the ultrasound waves, or the X rays converged on the filament (tumor). Analysis of the diagrams revealed that, indeed, subjects included the manner of convergence in their representations of the problem solutions. We classified the diagrams into two categories, Lines or Waves (100% agreement between two independent judges). Diagrams were classified as Lines when the converging forces were represented by straight lines, arrows, or by tunnels filled with dots (representing particles). Diagrams were classified as Waves when the converging forces were represented by wavy lines (smooth or sharp) or by series of expanding arches. Figure 2 presents representative diagrams from the Line and Wave categories. As we expected, subjects constructed different representations of convergence for the two light bulb problems: All 18 subjects in the light bulb-lasers condition (100%) but only 4 of the 22 subjects in the light bulb-waves condition (18%)drew Line diagrams. Moreover, 14 of the 17 subjects (82%) who received the tumor problem represented X rays with straight lines. That is, consistent with the pattern of spontaneous transfer in Holyoak and Koh (1987), the manner of convergence (rays) in the tumor target was similar to the manner of convergence in the lasers-light bulb base but differed from the manner of convergence in the waves-light bulb base. In order to examine whether and how matches and mismatches in manner of convergence affect access and mapping, we replicated Holyoak and Koh’s (1987) transfer experiment with two modifications. The first modification was that subjects were asked to draw the solutions to all the problems they read or solved (base, filler, and target problems). The second modification was that we used two rather than one version of the tumor target problem: a rays version (the forces were labeled “gamma rays”) and a waves version (the forces were labeled “omega waves”). Subjects were 155 undergraduate students from the University of Chicago who were randomly assigned to learn and solve one of the four combinations of two base and two target problems: two matching-manner conditions (lasers-rays: waves-waves) and two mismatching-manner conditions (lasers-waves; waves-rays). Subjects who did not draw diagrams for the base and/or the target problems were excluded from the relevant analyses. The diagrams drawn for the light bulb base problems were similar to those drawn in the first experiment: most subjects drew Line diagrams to represent converging lasers and Wave diagrams to represent converging ultrasound waves (84 vs 31 % Line diagrams, respectively). However, unlike in Holyoak and Koh’s (1987) study, spontaneous transfer was uniformly high in all experimental conditions (above 90%).3That is, matches and It is possible that drawing a diagram of the convergence solution for the base problem induced a more thorough processing and therefore a better understanding of the convergence solution than the task used by Holyoak and Koh (1987).
Miriam Bassok
10 Waves
Lines
x
Fig. 2. Examples of Line and Wave diagrams drawn to represent the convergence solution to the light bulb and tumor analogs.
mismatches in manner of convergence did not affect the frequency of spontaneous transfer. Whatever the reason for this uniformly high level of spontaneous transfer, subjects’ performance cannot be explained by the fact that they ignored the manner of convergence. On the contrary, analysis of the convergence diagrams that accompanied the transfer solutions to the waves and rays versions of the tumor target problems revealed that the manner of convergence had a very interesting and highly significant impact on application of the learned convergence solution.
Object-Based Reasoning
11
Figure 3 depicts the frequency of Line diagrams drawn for the tumor target problems in the four experimental conditions, including diagrams drawn by subjects who applied the convergence solution only after being informed about its relevance (i.e., informed transfer). As can be seen in Fig. 3, subjects in the lasers base condition drew significantly more Line diagrams for the two tumor target problems than did subjects in the waves base condition (75 vs 39%, respectively). That is, subjects actually transferred the manner of convergence from the base to the target problems. Representations of convergence were also affected by the type of force in the tumor target problems: Subjects drew more Line diagrams for the rays (64%) than for the waves tumor problem (49%). However, transfer of manner did not interact with the type of force in the target problems. That is, subjects in the lasers base condition drew more Line diagrams than subjects in the waves base condition both for the rays (80 vs 44%, respectively) and for the waves (65 vs 30%, respectively) versions of the tumor problem. This pattern of transfer indicates that, for many subjects, transfer occurred not because they realized that mismatches in manner of convergence can be ignored as irrelevant to the problem’s solution (i.e., abstraction in the sense of content deletion). Rather, transfer occurred because subjects interpreted the manner of convergence in the mismatching target problems
Fig. 3. Frequency of Line diagrams drawn to represent the convergence solutions for the rays and waves tumor targets in the lasers and waves light bulb base training conditions.
12
Miriam Bassok
as identical to the manner of convergence in the base problems (i.e., interpretation and accommodation). These results are consistent with the findings of Medin, Goldstone, and Gentner (1993) who found that similarity judgments may involve a constructive process of feature interpretation. Specifically, Medin et al. found that people who were asked to judge similarity between geometric figures interpreted ambiguous features of target stimuli such that these features matched the features of the base stimuli (e.g., they interpreted three and a half prongs in a target figure as either three or four prongs, depending on whether the base had three or four prongs). Of course, mismatches between base and target stimuli cannot always be resolved by feature interpretation (or reinterpretation). The next study shows that discrepancies in manner of change, inferred from object attributes (e.g., ice melting is continuous but ice deliveries to a restaurant are discrete), can severely impair both access and mapping.
B. MANNER OF CHANGE The experiments on manner of change follow up on previous work in which Keith Holyoak and I compared transfer of the same solution procedure learned either in the context of algebra or in the context of physics (see Bassok & Holyoak, 1993, for a summary and discussion of this earlier work). We expected that, because algebra training is more conducive to abstraction from content than is physics training, people will be more likely to transfer the learned solution from algebra to physics than vice versa. T o test this hypothesis, Holyoak and I used analogous word problems from algebra and physics. All problems described situations of constant change. The algebra problems asked for the sum of an arithmetic series (e.g., money earned by a person whose yearly salary increased by a constant amount during a 5-year period). The physics problems asked for distance traveled by a moving body during a period of constant acceleration (e.g., distance traveled by a car whose speed constantly increased during a 5-min period). Subjects received the algebra and physics problems either in training (base) or in transfer (target) tests. Consistent with our abstraction-fromcontent hypothesis, most subjects trained in algebra (90%) spontaneously used the arithmetic-series equation to solve the target physics problems. By contrast, only a minority of subjects trained in physics (10%) spontaneously used the distance equation to solve the target algebra problems (Bassok & Holyoak, 1989). Our initial explanation for lack of transfer from physics to algebra was that the physics content (e.g., speed) is included in the representations of the base problems and interferes with people’s ability to recognize that the algebra problems have a similar structure (ie., constant
Object-Based Reasoning
13
change). However, subsequent experiments revealed that this explanation was incomplete because transfer depended on the nature of the changing entities. In Bassok (1990), subjects trained in physics spontaneously transferred the physics solution to algebra problems in which the changing entities were rates (e.g., typing rate in words per minute), but failed to transfer the physics solution to algebra problems in which the changing entities were amounts (e.g., yearly dollar amounts of salary). Because the base physics problems described constant change in a rate entity (e.g., speed in meters per second), the results of Bassok (1990) indicate that people readily ignore mismatches in object attributes (e.g., meters vs words), but they do not ignore mismatches in the internal structure of the changing entities. That is, rate is a two-place predicate (e.g., TYPING RATE (words, minute)), whereas amount is a one-place predicate (e.g., DOLLAR AMOUNT (salary)). In other words, these results suggest that mismatches in local, lower order relations (rate vs amount) impaired recognition of matches in more global, higher order relations (constant change). Karen Olseth and I tested the possibility that the rate and amount entities in constant-change problems impaired transfer performance by creating mismatches in the global representations of constant change (Bassok & Olseth, 1995). Specifically,we conjectured that, as in the convergence experiments described earlier, people inferred the manner of constant change, continous or discrete, and treated differences in the inferred manner of change as differences in structure. Using a categorization task and an analysis of gestures that accompanied verbal descriptions of constant change (Alibali, Bassok, Olseth, Syc, & Goldin-Meadow, 1995), we first examined whether people spontaneously infer the manner in which various entities change with time. We found a pattern of matches and mismatches in manner of change that was consistent with the pattern of transfer in Bassok (1990). For example, both speed and typing rate were understood to be changing continuously (matching interpreted structures), but yearly changes in salary were understood to be changing discretely (mismatching interpreted structures). We then proceeded to examine whether and how matches and mismatches in manner of change affect access and mapping. We used an experimental design that varied orthogonally the manner of change in the base and target problems. Moreover, to control for possible effects of other types of matches and mismatches between the base and the target problems, we compared transfer from continuous and discrete base problems to pairs of continuous and discrete target problems that were equated in content and phrasing. Table I presents one continuous (speed) and one discrete
14
Miriam Bassok
TABLE I A CONTINUOUS AND A DISCRETE BASEPROBLEM, WITH MATCHING PAIRSOF CONTINUOUS AND DISCRETE TARGETS' Speed base (Continuous): The speed of an airplane increases at a constant rate during a period of 12 min from 10 mi/min to 34 mi/min. What distance, in miles, will the plane travel during the 12-min period? Population-rate target (Continuous): The rate of population growth in a certain country increased steadily during the last 12 years from 3000 peoplelyear to 15,000 peoplelyear. How many people total were added to the population during the 12-year period? Attendance-rate target (Discrete): An annual arts and crafts fair is held on November 1 every year. The attendance rate at the annual fair increased steadily during the last 12 years from 3000 people/year to 15,000 people/year. How many people total attended the fair during the 12-year period? Investment base (Discrete): During the last 16 months the monthly deposits into a certain savings account constantly increased from $200/month to $440/month. How much money total was deposited into the account during the 16-month period? Melting-ice target (Continuous): The rate at which ice melts off a glacier steadily increases over an 8-week period from 50 Ib/week to 106 Ib/week. What is the total weight of the ice that will melt off the glacier over the 8-week period? Ice-delivery target (Discrete): Ice is delivered to a restaurant once a week over an 8week period. The weight of the deliveries steadily increases from 50 Ib the first week to 106 Ib the 8th week. What is the total weight of the ice that will be delivered over the 8-week period? 'I From "Object-based reasoning: Transfer between cases of continuous and discrete models of change." by M. Bassok and K. L. Olseth. 1995,Journal of Experimental Psychology: Learning, Memory, and Cognition, 21, p. 1533. Copyright 1995 by the American Psychological Association. Adapted with pcrmission.
(investment) base problem, each with its matching pair of continuous and discrete targets. When the base problems described continuous change (speed), transfer to continuous targets (matching manner) was frequent, but transfer to discrete targets (mismatching manner) was rare. For example, significantly more subjects used the physics distance equation to find the number of people in the continuous-population than in the discrete-attendance target (71 vs 27%, respectively). Because the continuous and discrete target problems were matched in all other aspects of content and phrasing (e.g., both referred to people/year), these results indicate that matches and mismatches in the inferred manner of change were responsible for this overwhelming difference in the frequency of spontaneous transfer. Interestingly, when the base problems described discrete change (investment), mismatches in manner of change did not impair spontaneous transfer. For example, following an economics training session in which subjects learned to solve invest-
Object-Based Reasoning
15
ment problems (discrete), spontaneous transfer was similarly high to the discrete ice-delivery target (100%) and to the continuous melting-ice target (88%). The asymmetric impact of mismatches in manner of change on the frequency of spontaneous transfer (i.e., high transfer in the discrete-tocontinuous direction and low transfer in the continuous-to-discrete direction) was accompanied by an asymmetry in the relative difficulty of mapping (i.e., informed transfer). An analysis of the informed-transfer solutions revealed that subjects had no difficulty aligning the representations of discrete base and continuous target problems, but found it quite difficult to align continuous base and discrete target problems. Table I1 presents a representative excerpt of a verbal protocol from each of the two mismatchingtransfer conditions. The top panel of Table I1 demonstrates a straightforward instantiation of the base sum-of-investments equation (discrete) with values of the target speed problem (continuous). The bottom panel demonstrates repeated attempts to apply the base distance equation (continuous)
TABLE I1
EXAMPLES OF VERBAL PROTOCOLS THAT ACCOMPANIED INFORMED TRANSFER IN THE DISCRETE-TO-CONTINUOUS AND CONTINUOUS-TO-DISCRETE CONDITIONS~ Discrete (investments) to continuous (speed): The sum would be equal to what you start with plus what you end with divided by 2. So, start with 10 + 34 is 44, divided by 2 equals 22, times 12 is 264. . . miles. Continuous (speed) to discrete (investments): Alignment 1, mapping: I guess I’m supposed to apply those equations to this somehow, or they’re analogs. Uhh, all right then: Initial payment would correspond to initial speed and final payment to final speed, and so, average payment would be their sum divided by. . . Alignment 2, repeat 12 times: But wait, I was thinking whether I should use each month as an initial payment, in other words, apply this 12 times. But I don’t want to do that. It’s too much work. Back to Alignment 1, solution: I’m going to see if I can come out with the right answer by just treating it with the first and last months’ payments. So, I add 232 and 100 and get 332, then divide by 2 gives 166. . . Hesitation between Alignment 1 and 2: 1 guess. . . umm, that’s my average payment. How much money total over a year. . . uhh, I 2 average payments? 166 times 12 is 1992. ” From “Object-based reasoning: Transfer between cases of continuous and discrete models of change,” by M. Bassok and K. L. Olseth, 1995,Journal of Experimental Psychology: Learning, Memory, and Cognition, 21, p. 1538. Copyright 19YS by the American Psychological Association. Adapted with permission.
16
Miriam Bassok
to the target investments problem (discrete). These attempts involved shifts in alignment and explicit variable mappings. The asymmetry in transfer between problems with mismatching manner of change cannot be explained by a difference in the level of abstraction from content. Rather, this asymmetry reflects a difference in the possibility of reinterpreting the target such that it fits the representation of the base. As in the convergence experiments, transfer in the discrete-to-continuous condition was possible because continuous change in the target problems could be transformed (parsed) into an arithmetic series of discrete values and therefore fit the discrete structure of the base. For example, if the rate at which ice is melting off a glacier constantly increases over a period of 8 weeks from 50 Ib/week to 106 lb/week, the melting rate increases every week by a constant (8 lb/week), and the consecutive weekly values (e.g., 58 Ib/week, 64 Ib/week . . .) actually exist. This type of transfer was impossible in the continuous-to-discrete condition because, in order to transform discrete targets to fit a continuous base structure, people would have to hypothesize values that are unlikely to exist (e.g., hypothesize continuous deliveries of ice to a restaurant.). It is possible that the asymmetry in transfer between base and target problems with mismatching manner of change was responsible for the overwhelming asymmetry in transfer between algebra and physics in Bassok and Holyoak (1989). This is because the arithmetic-series algebra problems described discrete constant change, whereas the speed physics problems described continuous constant change. Of course, it is also possible that both training type (domain general vs domain specific) and the inferred manner of change (discrete vs continuous) contributed to the ease of transfer from algebra to physics and the difficulty of transfer from physics to algebra.
C. SEMANTIC SYMMETRY A N D ASYMMETRY Object-based inferences often refer to semantic relations between two or more entities. An important aspect of such inferences, to which I refer here as semantic symmetry and asymmetry, is whether the structural roles of arguments in a given semantic relation can or cannot be reversed. For example, PLAY (child A, child B) is a symmetric semantic relation, but PLAY (child A, ball B) is an asymmetric semantic relation. Ling-Ling Wu, Karen Olseth, and I found that inferences about semantic symmetry and asymmetry determined how people represented and solved probability problems (Bassok et al., 1995). Our experiments were motivated by the work of Brian Ross (1987,1989). Ross examined how college students used examples when solving novel
Object-Based Reasoning
17
probability problems. He found that these adult subjects engaged in object mapping much like the children who were aligning a chipmunk with a squirrel rather than with a frog in Gentner and Toupin’s (1986) study. For instance, subjects in Ross (1989) received an example of a permutation problem in which cars were randomly assigned to mechanics (base). Then they were asked to apply the learned solution to an isomorphic permutation problem in which scientists were randomly assigned to computers (target). Subjects received the relevant equation (l/n[n - lJ[n - 21 . . . [n- r + 11) and were asked to instantiate it with the appropriate values (i.e., n = the number of scientists). In the target problem, scientists had to be placed in the structural role of the cars (because both were the randomly assigned sets) and computers in the structural role of the mechanics (because both were the sets of assignees). However, subjects erroneously solved the target problem by placing scientists in the role of the mechanics and computers in the role of the cars. That is, they solved the target problem as if computers were assigned to scientists rather than scientists to computers. Ross explained such erroneous solutions by the fact that subjects tend to place similar objects in similar structural roles (i.e., object mapping). Specifically, he argued that subjects placed scientists in the role of the mechanics because both were animate objects, and placed computers in the role of the cars because both were inanimate objects. The object-mapping explanation accounts quite well for the pattern of transfer results in Ross (1989). However, these results can be also explained by object-based interpretation of the base and the target problems. For example, because subjects know that mechanics receive cars and scientists receive computers, it is possible that they included the asymmetric functional relation cET(receivers, givens) in the mental representations of the base and target permutation problems. From the worked-out solution to the base problem they learned that the givens (cars) were assigned to the receivers (mechanics). Accordingly, in trying to align the asymmetric interpreted GET structures of the base and target problems, they solved the target problem by assigning the givens (computers) to the receivers (scientists). In general, because animate objects were always the receivers and inanimate objects were always the givens in Ross’s problems, subjects’ performance cannot be explained definitively by either object-mapping or alignment of interpreted structures. In Basok et al. (1999, we modifed the problems and the procedure used by Ross (1989) such that we could examine whether transfer is mediated by abstraction of object based interpreted structures (e.g., GET).All base and target problems in our experiments described a person (e.g., a teacher) who randomly assigned three elements from one set (n prizes) to three
18
Miriam Bassok
elements from another set ( m students) and asked for the probabilities of such random assignments. The assignment problems had the same mathematical structure (permutation), but differed in the elements that served as the assigned and assignee sets ( n and m, respectively). Table I11 presents two representative problems from Experiment 1 in Bassok et al. In the first problem, computers are randomly assigned to secretaries; in the second problem, doctors from one hospital are randomly assigned to doctors from another hospital. The paired element sets of objects (0)and people (P) were selected such that, if subjects engage in object-based problem interpretation, they should abstract asymmetric interpreted structures (e.g., GET)for problems involving objects and people (0-P, e.g., computers and secretaries) and symmetric interpreted structures (e.g., PAIR)for problems involving similar sets of people (P-P, e.g., doctors from two hospitals). To test this prediction, we asked undergraduate students to solve either the 0 - P or the P-P permutation problems. Because our subjects never learned how to solve such problems their solutions were incorrect. Yet, consistent with our inter-
TABLE I11 SEMANTICALLY ASYMMETRIC AND SYMMETRIC PERMUTATION PROBLEMS" Computers and Secretaries (0-P, Objects assigned to People) In a big publishing company, some secretaries will get to work on new personal computers. The company received a shipment of 21 computers, with serial numbers in a running order; from 10075 through 10095. There are 25 secretaries in this company that would like to work on a new computer. The names of the secretaries are listed in order of their work experience, from the most experienced secretary to the least experienced one. The manager of the company randomly assigns computers to secreturies according to the work experience of the secretaries. What is the probability that the three most experienced secretaries will get to work on the first three computers (10075, 10076. and 10077). respectively? Doctors and Doctors (P-P, People assigned to People) In a medical meeting, doctors from a Minnesota hospital will get to work in pairs with doctors from a Chicago hospital. There is a list of 20 doctors from Chicago, arranged in alphabetical order. There are 16 doctors from Minnesota that would like to work with the doctors from Chicago. The names of the Minnesota doctors are listed in the order of their social security numbers, from highest to lowest. The chairman of the meering randomly assigns doctors from Minnesota to doctors from Chicago according to the alphabetical order of the Chicago doctors. What is the probability that the first three doctors on the Minnesota hospital's social security number list will get to work with the first three doctors on the Chicago hospital's alphabetical list, respectively? " From "Judging a book by its cover: Interpretative effects of content on problem solving transfer." by M. Bassok. L. L. Wu, and K. L. Olseth, 199.5, Memory. & Cognilion. 23. p. 367. Copyright 1995 by the Psychonomic Society. Inc. Adapted with permission.
Object-Based Reasoning
19
pretation hypothesis, these incorrect solutions captured the inferred semantic symmetry or asymmetry of the paired object sets. Specifically, most subjects (87%) in the 0-P condition constructed equations in which the two sets of elements played asymmetric mathematical roles (e.g., rn3/n!; l/n3). By contrast, most subjects (78%) in the P-P condition constructed equations in which the two sets of elements played symmetric mathematical )~; + n)!). roles (e.g., (rn + n ) / ( r n r ~ 3/(m After they solved (incorrectly) the unfamiliar permutation problems (base), subjects read a chapter that explained the relevant probability concepts. The chapter included the correct equation for solving permutation problems and, for each subject, a worked-out solution to the base problem the subject initially failed to solve correctly. Following this training session, subjects received a target problem together with the relevant equation (1/ n[n - 1][n - 21). They had to instantiate the equation for the target problem, that is, to choose which of the two sets was the randomly assigned set ( n ) .The targets either matched or did not match the interpreted symmetric or asymmetric structure of the base (e.g., 0 - P base and either 0 - P or P-P target). Moreover, as in Ross (1989), the direction of assignment in the base and target 0-P problems was either the same (e.g., objects assigned to people) or reversed (e.g., people assigned to objects). The main finding of interest was that subjects solving the asymmetric 0-P targets had a very strong preference to assign objects to people (e.g., prizes to students) rather than people to objects (e.g., students to prizes). That is, subjects were acting as if the direction of assignment in the permutation problems was consistent with the semantically asymmetric outcome of assignment (e.g., students get prizes rather than vice versa). This interpretive preference mediated the magnitude of object mapping in the 0 -P training conditions in which, as in Ross (1989), animate objects were the receivers and inanimate objects were the givens. In Experiment 2, we unconfounded matches in object attributes (e.g., animate) from matches in interpreted structures (e.g., receivers). For example, subjects received a base with two functionally asymmetric sets of people (i.e., caddies assigned to golfers) and an 0-P target in which either carts were assigned to caddies or caddies were assigned to carts. Although caddies were the assigned set in the base problem, subjects did not assign caddies to carts in the target problems (i.e., object mapping). Rather, consistent with their knowledge about the asymmetric outcome of assignment, they assigned carts to caddies (94 vs 24% correct for the carts assigned to caddies vs caddies assigned to carts targets, respectively). That is, the impact of object-based interpretations was powerful enough to override matches in object attributes.
20
Miriam Bassok
D. SUMMARY Three main findings emerge from the studies described in this section. First, object-based inferences (e.g., lasers proceed in straight lines; speed changes continuously; students get prizes) may lead to abstraction of nonisomorphic interpreted structures for formally isomorphic base and target problems. Second, transfer is sometimes mediated by accommodating an interpreted target to the interpreted structure of the base problem (e.g., treating rays in a target problem as if they were waves; treating continuous change in a target problem as if it were discrete) rather than by abstracting a structure that is common to distinct interpreted structures (e.g., converge; change). Third, object-based interpretive effects (e.g., caddies get carts rather than vice versa) can override effects caused by direct matches and mismatches in object attributes (e.g., caddies are more similar to caddies than are carts). These findings have important implications for instruction in formal domains. They show that, just as the good learner in Chi et al. (1989) expected the block rather than the knot to serve in the role of the physical “body,” students of formal domains expect to find correspondence between the mathematical structure of a word problem and the semantic structure of the situation described in the problems’ cover story. This expectation guides their understanding of worked-out examples and affects transfer to novel problems. Interestingly, such interpretive effects of content are more likely to affect the performance of students with good rather than poor mathematical understanding (Bassok, 1997; Hinsley et al., 1977; Paige & Simon, 1966). The studies described so far addressed a processing stage (i.e., representation) that, at present, remains outside the scope of research on analogical transfer. The interpretive effects found in these studies show that objectbased inferences can affect how people represent base and target stimuli both before and during access and mapping. Note, however, that these studies have no direct bearing on the validity of the processing assumptions in extant models of access and mapping. For instance, it is possible that when people aligned the interpreted representations of the base and target stimuli (e.g., aligned receivers with receivers) they distinguished between relations and attributes in a way that is consistent with the assumptions of the alignment models. The studies I describe in the next section show that, in addition to their effects on representation, objectbased inferences determine how people process the stimuli they encounter. Specifically, these studies show that semantic knowledge affects which processing strategy (e.g., comparison vs integration) is selected for a given pair of base and target stimuli. Unlike representation, the issue of stimuli-appropriate process selection was never on the research agenda of investigators who study analogical transfer or similarity.
Object-Based Reasoning
21
IV. Semantic Knowledge Affects Selection of Processing Strategies The common saying “It’s like comparing apples and oranges” is used to denote an inappropriate comparison. What is interesting about this saying is that apples and oranges, rather than, let’s say, apples and baskets, became idiomatic with a poor comparison. Wouldn’t apples and baskets, which obviously have less commonalities than apples and oranges, serve as a better example of a poor comparison? The answer to this question is that apples and baskets cannot exemplify a poor comparison because they do not exemplify a comparison. Rather, they exemplify a functional (thematic) relation between two entities. This answer emerges from a set of studies in which my colleagues and I found that, irrespective of task instructions, people tend to compare certain stimuli (e.g., apples and oranges) but tend to integrate other stimuli (e.g., apples and baskets). As a result, when people are faced with stimili that do not fit the task requirements (e.g., asked to judge similarity between apples and baskets), they tend to replace a task-appropriate process (comparison) with a stimuli-appropriate process (integration)-performance to which I refer here as processing replucements. Theoretical accounts of similarity seem to hold that the process of comparison is a general-purpose mechanism that can be applied to any arbitrary pair of stimuli (e.g., both apples and baskets and apples and oranges). Most probably, researchers who study similarity or analogical reasoning did not notice that people may integrate rather than compare certain stimuli because, much like their subjects, they typically select comparison-appropriate stimuli (but see Markman & Gentner, 1993). Of course, unintentionally, they may sometimes select stimuli that people tend to integrate rather than compare. However, being unaware of processing replacements, they would try to explain performance that was mediated by integration (e.g., baskets contain apples) in terms of the hypothesized process of feature comparison (e.g., both baskets and apples are objects found in orchards). It was within this conceptual framework that Douglas Medin and I initially found that, for certain pairs of base and target stimuli, subjects systematically arrived at similarity judgments, not by comparing the representations of these stimuli, but rather by integrating them into joint thematic scenarios. Needless to say, these findings caught us by complete surprise. In what follows I first describe these unexpected processing replacements during similarity judgments (Bassok & Medin, 1997). Then, I describe results from a follow-up study that was explicitly designed to examine the impact of semantic knowledge on processing replacements (Wisniewski & Bassok, 1996). Finally, I describe results from a study in which the same
22
Miriam Bmok
semantic distinction that determines which objects are compared and which are integrated led to selective application of arithmetic operations in reasoning about addition and division word problems (Bassok et al., 1997). A. PROCESSING REPLACEMENTS IN SIMILARITY JUDGMENTS Similarity, like analogy, involves comparison and alignment. This observation led several investigators to examine the possibility that the structuralalignment approach to analogical mapping can be adapted to explain similarity judgments (e.g., Goldstone, 1994; Goldstone & Medin, 1994; Goldstone, Medin, & Gentner, 1991; A. B. Markman & Gentner, 1993, 1996; Medin, Goldstone, & Gentner, 1990,1993). It turned out that, indeed, when people judge similarity between various pairs of base and target stimuli they distinguish between relational and attributional matches and mismatches. For example, Medin et al. (1990) presented subjects with stimuli consisting of interrelated geometric shapes. They asked subjects to choose which of two target alternatives, R (a star above a circle) or A (a triangle next to a circle), was more similar to or more different from a standard base (e.g., a triangle above a square). The stimuli were constructed such that R targets had an extra relational match (e.g., above) and A targets had an extra attributional match with the base (e.g., triangle). When the task was to judge similarity, subjects were more likely to choose R rather than A targets as being more similar to the base. However, when the task was to judge difference, this preference was less pronounced or even reversed. In an extreme case, subjects would choose the same R target as both more similar to and more different from the base. From these nonmirroring similarity and difference judgments, Medin et al. concluded that people distinguish between relational and attributional matches: assign higher weights to relational matches in similarity than in difference judgments and/or assign lower weights to attributional matches in similarity than in difference judgments. As in most experiments that document differential effects of relational and attributional matches on similarity judgments, the stimuli used by Medin et al. (1990) were carefully chosen to enable a clear separation and contrast between attributes and relations. For example, the symmetry relation in OX0 and A*A is independent of whether the symmetric shapes in the two figures happen to be circles or triangles, or whether a star is between two triangles rather than vice versa (*A*). However, in most cases it matters which objects serve as arguments in a particular relation. For example, carpenters are known to !ix chairs rather than radios as part of their profession; cutting grass is a different type of cutting than cutting hair; and mockingbirds are believed to sing more nicely than crows. It
Object-Based Reasoning
23
therefore remained unclear whether the processing regularities in similarity and difference judgments observed in the semantically independent case (e.g., “a triangle above a square”) would also hold for semantically interdependent combinations of objects and relations (e.g., “The carpenter fixed a chair”). In Bassok and Medin (1997) we examined how semantic dependencies between attributes and relations affects similarity and difference judgments: whether such dependencies affect only the representations constructed for the base and target stimuli (i.e., interpreted representations) or, in addition, affect the weights assigned to attributional and relational matches (i.e., changes in processing). Our stimuli were simple base and target statements in which the nouns (denoting objects) and verbs (denoting relations) were semantically interdependent. The statements were designed in a way that allowed us to distinguish effects of separate matches in the explicitly stated nouns (A: attributes) and verbs (R: relations) from more abstract matches associated with the interpreted combined meaning of the statements. We assumed that matches in the combined meaning of the statements can be treated as matches in abstract higher order relations (R*) that specify how separate attributes and relations are interrelated. We used 12 quintuplets of statements from which we constructed three types of triplets. Each triplet consisted of a Base, its Attributional target, and one of its three Relational targets. Table IV presents one quintuplet of statements that exemplifies the design of our stimuli. As can be seen in Table IV, the combined meaning of RR* targets always matched the combined meaning of the base (e.g., a professional performing a job-related activity), whereas the combined meaning of R and RA targets did not (e.g, neither plumbers nor carpenters fix radios as part of their profession). As a result, although the three relational targets had the same separate relational
TABLE IV O N E QUINTUPLET OF
BASEA N D TARGET STATEMENTS‘
Base Attributional target (AA) Relational targets Relation inferred Relation* (RR*) Relation (R) Relation + Attribute (RA)
+
The carpenter fixed the chair The carpenter sat on the chair The electrician fixed the radio The plumber fixed the radio The carpenter fixed the radio
Nore. Separate noun (A) and verb (R)matches are typed in boldface. From “Birds of a feather flock together: Similarity judgments with semantically rich stimuli.” by M. Bassok and D. L. Medin, 1997, Journal of Memory and Language, 36, p. 316. Copyright 1997 hy Academic Press.
Miriam Bassok
24
match with the base (e.g., R = fix), RR* formed good analogies to the base, whereas R and R A formed poor analogies. The difference between the poor-analogy R and R A targets was that R A had an additional attributional match with the base (e.g., A = carpenter). The attributional target (AA) had a relational mismatch with the base (e.g., fix # sit) and therefore a combined mismatch. Subjects received one triplet from each of the 12 different quintuplets, and for each triplet they rated the similarity (difference) between the base and its relational and attributional targets. Applying Gentner’s (1983) “systematicity principle” to our stimuli, we predicted that matches and mismatches in the combined meanings of the statements-in the inferred higher order relations (R*)-wil override explicitly stated separate relational (R) and attributional (A) matches and mismatches. Indeed, we found that subjects gave significantly higher similarity ratings to good than to poor analogy targets, both when these targets were equated in separate matches (RR* > R), and even when the pooranalogy targets had an additional separate match (RR* > RA). For example, subjects gave higher similarity ratings to the target “The electrician fixed the radio” (matching R*) than to the target “The carpenter fixed the radio” (mismatching R*) even though a carpenter is clearly morc similar to a carpenter than is an electrician. These findings are similar to the findings described in the previous section (i.e., interpreted structures). They show that semantic dependencies between attributes and relations affect similarity and difference ratings by affecting the representations people construct for the base and target stimuli. To examine whether object-based inferences also affect processing, we compared the ratings in the similarity and difference conditions. However, we did not replicate the nonmirroring pattern of similarity and difference ratings found for arbitrary combinations of separable relations and attributes (Medin et al., 1990). That is, we did not find evidence that subjects gave higher weights to relational matches (R targets) in similarity judgments and/or gave higher weights to attributional matches (A targets) in difference judgments4.These results indicate that, in addition to affecting the representations of the base and the target stimuli, nonarbitrary dependencies between attributes and relations affect processing. In order to better understand the process that mediates comparison of semantically rich stimuli, we asked subjects to explain in writing how they arrived at their similarity and difference ratings. We found that, indeed, subjects compared the inferred combined meanings of the base and the relational targets (RR*, RA, and R). For example: “A lawyer would never design a dance” (unlike engineers who design cars); “Most children do not We found nonrnirroring effects in one cxperirncnl, but only for the good-analogy
(RR*)targets.
Object-Based Reasoning
25
enjoy jobs” (but they enjoy toys); or “The first sentence is a fact, whereas the second is an opinion” (for the base “Equations are more accurate than words” and the target “Equations are more difficult than words”). However, to our great surprise, we found that subjects did not always compare the representations of the base and the attributional targets (AA). Rather, very often they integrated the combined meanings of the base and the attributional-target statements into joint thematic scenarios. For example: “A teacher may have listened to the lecture to prepare” (for the base “The teacher prepared a lecture” and the target “The teacher listened to the lecture”); “Something the child might do if he/she enjoyed the toy, out of selfishness” (for the base “The child enjoyed the toy” and the target “The child hid the toy”); or “A logical step: since equations are more accurate they are more difficult to use” (for the base “Equations are more accurate than words” and the target “Equations are more difficult than words”). Most combined explanations generated for the relational targets involved comparison (94%),whereas a substantial majority of combined explanations generated for the attributional targets involved integration (59%). This systematic difference in the distribution of comparison and integration explanations for the relational and attributional targets was found in the explanations generated by 73% of the subjects ( N = 80 Northwestern University and University of Chicago undergraduates). Moreover, the pattern and magnitude of the ratings were consistent with both types of explanations. Subjects mentioned combined matches (e.g., “Both involve professionals doing their jobs”) when explaining their high similarity ratings for the good-analogy relational targets (RR*) and combined mismatches (e.g., “Not analogous because plumbers don’t fix radios as part of their job”) when explaining their low ratings for the poor-analogy relational targets (R and RA). At the same time, a post hoc analysis of ratings for the attributional targets (AA) revealed that supporting scenarios (e.g., “Similar because the carpenter sat on the chair to see whether he fixed it well”) were associated with significantly higher similarity ratings than impeding scenarios (e.g., “It does not necessarily follow that because he fixed the chair he had to sit in it”). These results strongly suggest that, for both target types, the explanations captured two qualitatively different processes by which subjects arrived at their ratings. When the base and target statements shared a relational match (i.e., common verb), similarity and difference ratings were almost exclusively mediated by a process of comparison. By contrast, when the base and target statements shared only attributional matches (i.e., common nouns), similarity judgments were frequently mediated by a process of thematic integration. That is, the relational and attributional targets induced,
Miriam Bassok
26
or highlighted, two qualitatively different processes. Given that the task (similarity or difference ratings) demanded comparison, subjects’ performance on the attributional targets indicates that they actually replaced a task-appropriate process with a stimuli-appropriate process. Why is it that the attributional targets induced processing replacements? The nature of our stimuli precludes the possibility that subjects were forced to replace comparison with integration because they could not identify salient commonalities and differences between the base and the attributional statements. This is because the attributional targets were explicitly designed such that they had two matching nouns with the base (e.g., carpenter, chair) that could be readily used t o explain similarity, and had a mismatching verb (e.g., fix vs sit) that could be used to explain difference. At the same time, our stimuli were neither designed nor obviously slanted toward familiar scripts o r scenarios. For example, there is no obvious causal or temporal relation between catching a ball and selling it, or between accuracy of equations and their difficulty. In fact, our subjects often used lack of thematic relatedness to explain why the paired statements were not similar (e.g., “It does not necessarily follow that because he fixed the chair he had to sit in it”). Thus, it is not the case that subjects were simply carried away by highly activated thematic scenarios. Rather, it appears that the attributional targets led subjects to actively search for, or construct, possible scenarios. Successful scenarios were taken as supporting similarity (high ratings) and unsuccessful scenarios as impeding similarity (low ratings). The systematic correspondence between processing (comparison vs integration) and target type (relational vs attributional) strongly suggests that processing replacements occurred in response to lack of relational matches between the base and the target stimuli. Interestingly, the pattern of processing replacements strongly supports the structural-alignment approach to analogy and similarity. It suggests that relational commonalities are crucial, if not absolutely necessary, for alignment of semantically structured stimuli-that matching relations (verbs) enable structural alignment, but matching attributes (nouns) do not. At present, it is unclear whether integration comes into play only when comparison fails to achieve some “satisficing” criteria, whether comparison and integration compete with each other as orthogonal alternatives, or whether these two processes operate simultaneously to support and impede similarity. What is quite certain, however, is that comparison is not the only processing mechanism that mediates similarity judgments for semantically rich stimuli. Hence, at the very least, the alignment models would have to be supplemented with a selection mechanism that specifies the conditions that may lead to processing replacements.
B. SEMANTIC COMPATIBILITY OF OBJECT PAIRS WITH COMPARISON AND
INTEGRATION
The alignability account of processing replacements implies that when the base and target stimuli are objects, rather than semantically structured
Object-Based Reasoning
27
combinations of objects, thematic integration should be more prominent when the paired objects are nonalignable (i.e., have few common attributes) than when they are alignable (i.e., have many common attributes). To test this prediction, Edward Wisniewski and I collected similarity ratings for object pairs that differed in alignability (Wisniewski & Bassok, 1996). The stimuli were 12 quintuplets of nouns (denoting objects), each consisting of a base and four targets. Two targets shared many attributes with the base (A+, alignable targets) and the other two shared few attributes with the base (A-, nonalignable targets). The two alignable targets (A+) were selected from the taxonomic category of the base, whereas the nonalignable targets (A-) were selected from a different taxonomic category. We also examined whether familiar thematic relations interact with alignability to inducing processing replacements. Accordingly, one of the A + and A- targets in each quintuplet was related to the base by a familiar thematic relation (T+) and the other was not (T-). The resulting structure of the four targets in each quintuplet was: A + T + , A+T-, A-T+, A-T-. For example, the four targets for the base “milk” were, respectively, “coffee,” “lemonade,” “cow,” and “horse.” The 12 quintuplets used in our study appear in Table V. We presented subjects with 12 base-target pairs, each pair from a different quintuplet. Subjects read instructions telling them that they would see some pairs of common, everyday things, and that they would have to rate TABLE V QUINTUPLETS OF
Base milk ship car chair telephone tie chisel cat CUP
fly peanut butter apple pie
BASEAND TARGET NOUNS~
A+T+ target
A+T- target
A-T+ target
A-T- target
coffee lifeboat tow truck table answering machine suit hammer mouse kettle spider jelly ice cream
lemonade canoe pickup truck bed tape recorder dress screwdriver hamster Pan beetle cream cheese jello
cow sailor mechanic carpenter receptionist man sculpture veterinarian tea screen knife baker
horse soldier plumber electrician waitress woman painting pediatrician wine curtain fork tailor
From “On putting milk in coffee: The effect of thematic relations on similarity judgments,” by E. J. Wisniewski and M. Bassok, 1996. in G. W. Cotrell (Ed.), Proceedings ofthe Eighfeenrh Annual Conference of the Cognitive Science Society (p. 465), Hillsdale. NJ: Erlbaum. Copyright 1997 by Lawrence Erlbaum Associates, Inc. Rcprinlcd with permission.
28
Miriam Bassok
how similar the two things are on a 7-point rating scale. Also, to ensure that thematic replacements do not occur only when subjects explain their ratings, half of the subjects explained in writing why they thought the two things had the degree of similarity that they did and the other half did not. First, consistent with the definition of similarity as a monotonic function of common and distinctive attributes (e.g., Tversky, 1977), A + targets received higher similarity ratings than A- targets (e.g., coffee and lemonade were rated as more similar to milk than cow and horse). Moreover, familiar thematic relations affected similiarity ratings, such that T+ targets received higher ratings than T- targets (e.g., coffee and cow were rated as more similar to milk than lemonade and horse). Because preexisting thematic relations can be treated as commonalities (e.g., milk and coffee appear in the same cup), these results do not prove that similarity ratings were mediated by integration rather than by comparison. However, consistent with the alignability account of processing replacements, the impact of preexisting thematic relations on similarity ratings was much more pronounced for the nonalignable than for the alignable targets. That is, the difference between similarity ratings for A-T+ and A-T- targets (e.g., cow and horse) was larger than that for A + T + and and A + T - targets (e.g., coffee and lemonade). Asking subjects to explain their ratings did not affect the magnitude of ratings. At the same time, as in Bassok and Medin (1997), explanations provided direct evidence for processing replacements. We examined the relative proportion of explanations, both supporting (+) and impeding (-) similarity, that involved either comparison (e.g., “people drink both milk and lemonade,” “a horse can be white like milk”) or fhematic integration (e.g., “people often add milk to coffee,” “you do not milk a horse”). As the alignability account would predict, and consistent with the pattern of ratings, the tendency toward thematic integration was much higher for nonalignable than for alignable targets (52 vs 16% for A- and A + targets, respectively). Importantly, the difference in frequency of thematic replacements between the A- and A + items was larger, but not limited to the T+ targets. That is, preexisting thematic scenarios (e.g., milk and cow vs milk and horse) simply magnified people’s spontaneous tendency to integrate nonalignable stimuli (e.g., “you do not milk a horse”). Obviously, most thematic explanations for the T+ targets supported similarity (e.g., “Carpenters fix chairs”). By contrast, most thematic explanations for the T- items were impeding scenarios, that is, counterfactual arguments that explained why the paired objects were not similar, for example, “Cat and mouse aren’t similar because they are enemies,” or “Women usually do not wear ties.”
Objeet-Based Reasoning
29
To the extent that alignability determines the likelihood of processing replacements in similarity judgments, one would expect a reversed pattern of processing replacements when the task is to judge thematic relatedness. We tested this hypothesis using the same materials (Table V) and asking subjects to rate the degree of functional or causal relatedness between the paired entities. Indeed, we found a pattern of processing replacements that mirrored the pattern observed in similarity judgments. First, attributional replacements (i.e., comparison instead of integration) were much more frequent in A + than in A- targets (50 vs 8%). For example, subjects were more likely to “erroneously” compare milk and lemonade than to “erroneously” compare milk and horse. Importantly, the difference in frequency of attributional replacements between the A + and A - targets was larger, but not limited to the T- targets (a difference of 62 vs 23% for the T- and T+ targets, respectively). That is, lack of preexisting thematic scenarios (e.g., milk and lemonade vs milk and coffee) simply magnified people’s spontaneous tendency to compare alignable stimuli (e.g., “Both are beverages”). To summarize, stimuli differ in their compatibility with processing strategies. When the stimuli are incompatible with a task-appropriate process, or when they are more compatible with an alternative process, people may replace the task-appropriate process with a stimuli-appropriate process (e.g., replace comparison with integration or vice versa). Of course, thematic integration is not a viable processing option for stimuli consisting of arbitrary combinations of interrelated objects (e.g., a triangle above a square; a triangle next to a square). In other words, the type of processing replacements demonstrated in our studies is specific to semantically rich stimuli.’ C. SEMANTIC COMPATIBILITY OF OBJECT PAIRS WITH ADDITION AND DIVISION Processing replacements of the sort documented by Bassok and Medin (1997) and Wisniewski and Bassok (1996) have been previously found only in young children (see E. M. Markman, 1989, for a review) and in adults from illiterate cultures (Luria, 1976).6 In one study, Luria (1976) asked illiterate adults from Uzbekistan to compare pairs of objects-to explain what the objects had in common and in what way they were alike. These adults often responded by noting possible thematic relations between the
’
The difference in processing options available for arbitrary versus semantically meaningful stimuli may have been responsible for the fact that certain processing regularities found for interrelated geometric shapes (e.g.,monitoringsimilarity and difference ratings in Medin et al., 1990) were not replicated with semantically interdependent objects (Bassok & Medin, 1977). In her recent dissertation, Emily Lin (1996) reports similar results for American adults who were performing a categorization task.
‘
30
Miriam Bassok
objects. For example, one illiterate adult who was asked by Luria “What do water and blood have in common?” responded: “What’s alike about them is that water washes off all sorts of dirt, so it can wash off blood too” (p. 82). That is, much like other content effects (e.g., failures of analogical transfer), processing replacements are associated with errors that indicate lack of maturity, poor intelligence, or insufficient knowledge. Because the subjects in our studies were neither young nor illiterate, their performance cannot be explained away by insufficient cognitive or cultural development. Of course, one could maintain that pocessing replacements are errors and reflect confusion. In particular, one could argue that it is relatively easy to confuse integration and comparison because relations between objects (e.g., putting milk in coffee) lead to commonalities (e.g., both milk and coffee are in a cup). However, treating processing replacements as confusion errors does not explain the systematicity in which they occur. Moreover, by treating processing replacements as failures of the cognitive system, the confusion argument marginalizes the importance of this phenomenon and therefore detracts attention from the regularities with which processing is attuned to the processed stimuli. The last study I describe in this review (Bassok et al., 1997) replicated the pattern of processing replacements described earlier in a task that dissociated object-based process selection from confusion errors. In fact, the task was designed such that object-based process selection demanded investment of extra cognitive effort. Specifically, Valerie Chase, Shirley Martin, and I asked undergraduate students to construct simple addition or division word problems for pairs of object sets that we provided (e.g., tulips-daffodils; tulips-vases). Subjects were free to invent the number of elements for each set (e.g., m tulips and n daffodils). The minimal mathematical solution that meets the requirements of the construction task is to relate the given pair of object sets directly by the required arithmetic operation: direct addition (rn + n) when the task is to construct addition problems, and direct division ( d n ) when the task is to construct division problems. However, as I describe below, the object pairs were selected such that half were semantically compatible and the other half were semantically incompatible with direct addition and direct division.
1. Direct Addition (m+n) In the additive mathematical structure, two or more distinct sets of elements (e.g., m and n ) are combined to form a union set (u). The combined sets play symmetric structural roles in the formation of the union set (m+n = n+m) and elements from the constituent sets become equivalent members in the union set ( u ) . The best semantic match to the mathematical con-
Object-Based Reasoning
31
straints of symmetry and membership equivalence is obtained when the added sets consist of identical elements (e.g., m tulips + n tulips = u tulips). When the elements in m and n are not identical, adding elements from the same taxonomic category (e.g., m tulips + n daffodils = u flowers) leads to a more meaningful combined set than adding elements from distinct taxonomic categories (e.g., m tulips + n vases = u things). 2. Direct Division ( d n ) In the multiplicative mathematical structure, elements from two distinct sets (m, n ) are related by a function ( k ) that maps elements from the m set onto elements from the n set. The mapping function, whose value equals the outcome of m divided by n, is directional and usually asymmetric ( d n # n/m). The multiplicative structure is therefore most compatible with situations in which the m and n elements are known to be related by an asymmetric functional relation. For example, the asymmetric functional relation CONTAIN (vases, tulips) naturally corresponds to the asymmetric mapping function and therefore can be directly aligned with the operation of division ( k = m tulipsln vases). Moreover, and unlike in the symmetric additive structure, considerations of semantic compatibility constrain the structural roles of the paired sets in the asymmetric mapping function (e.g., m tulips are contained in and therefore divided by n vases rather than vice versa). Functionally asymmetric relations between elements that belong to the same taxonomic category exist (e.g., CATCH [spiders, flies] and other examples in the A + T + column of Table V). However, most pairs of elements from the same taxonomic category lack a natural functional relation. In this case, the only semantic meaning that corresponds to the mapping function ( k ) is a proportional comparison of set sizes (i.e., the number of tulips is k times the number of daffodils). Unless such proportional asymmetry is supported by semantic knowledge (e.g., there are more robins than bluebirds), the proportional meaning of division will be arbitrary and therefore a less compelling match to the asymmetric mapping function than a functionally asymmetric relation. The difference in compatibility of direct addition and direct division with semantically symmetric and asymmetric relations between the paired object sets afforded us the opportunity to test for object-based process selection that does not lead to mathematical errors. If peole ignore considerations of semantic compatibility in the construction task, then direct addition (m+n) and direct division ( d n ) of members in the given sets should be the least effortful strategy in the addition and division construction tasks, respectively. If, however, people are guided by considerations of semantic
32
Miriam Bassok
compatibility, then they should avoid direct addition of semantically asymmetric sets and direct division of semantically symmetric sets. Instead, they should employ a variety of semantic-escape strategies. For example, they might invent another variable ( p ) that can be added or divided in a semantically compatible way, and fulfill the task requirements by constructing problems with a more complex mathematical structure (e.g., [m+pVn). As predicted, subjects had a very strong preference to construct problems in which the arithmetic operations were compatible with the semantic relations implied by the paired object sets. First, the great majority of problems constructed for semantically compatible object pairs in the addition and the division conditions (70-88%) involved direct addition and direct division of the paired sets, respectively. An example of a directdivision problem constructed for the semantically compatible pair boysteachers in the division condition is ( d n ) :“Three teachers want to evenly divide a class of 60 boys. How many boys should go with each teacher?” At the same time, a significant proportion of problems constructed for semantically incompatible object pairs (27-80%) were semantic escapes. For example, an escape problem constructed for the semantically incompatible pair tulips-daffodils in the division condition also involved addition of the two target sets of flowers ([m+nVp):“Wilma planted 250 tulips and 250 daffodils and it took 20 days to plant them. How many flowers did she plant per day?” Most semantic-escape problems involved the requested arithmetic operation, although this operation did not relate the incompatible object sets. However, in some semantic escapes subjects actually failed to fulfill the task requirements. Instead of constructing complex problems, they replaced the requested but semantically incompatible arithmetic operation with a task-inappropriate but semantically compatible arithmetic operation (i.e., division instead of addition or addition instead of division). An example of such operation replacements is a division escape problem constructed in the addition condition for the semantically incompatible pair peachesbaskets: “Two baskets hold 30 peaches, how many peaches does 1 basket hold‘?’’;or an addition escape problem constructed in the division condition for the semantically incompatible pair peaches-plums: “If there is a basket with peaches and plums in it, and we know that the total number of pieces of fruit is 20, and that there are 5 peaches, how many plums must there be in the basket?” The operation-replacement problems subjects constructed for semantically incompatible object pairs mirror the processing replacements exhibited by subjects who rated similarity and thematic relatedness in Wisniewski and Bassok (1996). That is, subjects both added and compared apples and oranges (alignable stimuli), and they both divided and integrated apples
Object-Based Reasoning
33
and baskets (nonalignable ~ t i m u l i )Importantly, .~ such semantically induced operation replacements occurred even though addition and division are much less confusable than comparison and integration. In other words, the pressure to relate the paired object sets in a semantically meaningful way was powerful enough to override clear task instructions. The correspondence between processing replacements and performance in the construction task highlights the practical implications and the reasonableness of object-based process selection. There is little doubt that the undergraduate students in Bassok et al. (1997) had sufficiently abstract knowledge of arithmetic-that they could readily add and divide abstract Xs and Ys. Nonetheless, they did not treat object sets as abstract variables and did not treat arithmetic operations as general-purpose procedures. Rather, they used addition and division as semantically sensitive modeling tools. By analogy, people do not use comparison and integration as generalpurpose processing mechanisms, but rather, use them as semantically sensitive reasoning tools. While mirroring processing replacements, semantic-escape problems also mirror the pattern of semantic alignments exhibited by subjects who solved the permutation problems in Bassok et al. (1995). Whether solving novel probability problems or constructing familiar arithmetic problems, subjects were trying to align the semantic and the mathematical structures of word problems. By mirroring both types of object-based effects described in this chapter, the Bassok et al. (1997) study ties together the impact of semantic knowledge on representation and process selection. It therefore highlights the inherent dependency between these two aspects of reasoning and implies that they should be studied jointly.
V. Discussion Semantic knowledge is organized such that it affords meaningful and adaptive inferences (e.g., apples and oranges are fruit and therefore can play similar functional roles). The studies reported in this chapter show that such object-based inferences affect how people solve problems, transfer previously learned solutions to novel problems, or judge similarity. Put in this way, the present findings are not very surprising-it is not very surpris-
’
Consistent with this distinction. Wisniewski (1996) found that “similar” and “dissimilar” nouns, which correspond to our alignable and nonalignable object sets, induce two qualitatively different mechanisms by which people form conceptual combinations. When combining alignable nouns (e.g., skunk-squirrel), people tend to replace an attribute of one noun with an attribute of the other noun (e.g., a bad-smelling squirrel). By contrast, when combining nonalignable nouns (e.g., rabbit-box), people tend to form functional conceptual combinations (e.g.. a box for holding rabbits).
34
Miriam Bsssok
ing that people prefer to compare or combine apples and oranges rather than apples and baskets. What is quite surprising, however, is that such effects have been overlooked by researchers who study higher order cognition. It is not that researchers failed to notice that object-based inferences affect reasoning. A classic example of such effects would be Duncker’s (1945) work on “functional fixedness,” whereby the functional role of a box as a container pevented people from using the box as a platform on which they could mount a candle. In a more recent example, Kotovsky, Hayes, and Simon’s (1985) study compared people’s solutions to two versions of the Tower of Hanoi problem. They found that subjects had little difficulty placing a large disk on top of a small disk, but when the disks were labeled “acrobats,” subjects refrained from letting a large acrobat jump on the shoulders of a small acrobat. Unfortunately, such demonstrations were merely added to the list of studies showing that superficial aspects of content and phrasing lead to errors, or make some problem isomorphs more difficult than others. They did not initiate research that looks for regularities in the way people select, or adjust, their reasoning tools (e.g., comparison vs integration) to semantic distinctions they deem important (e.g., functional symmetry vs asymmetry). A line of research that, in its gist, is probably the closest to the studies reported here looks for regularities in the way semantic knowledge affects reasoning about conditional syllogisms (e.g., Cheng & Holyoak, 1985,1989; Cheng, Holyoak, Nisbett, & Oliver, 1986; Cosmides, 1989; Cosmides & Tooby, 1994; Cummins, 1995; Gigerenzer & Hug, 1992). This work has shown that different content instantiations of the material implication (“if p then q”) induce reasoning rules of distinct pragmatic and/or social schemas. According to Cheng and Holyoak (1985), who initiated this line of research, people apply different reasoning rules to formally isomorphic statements such as “if there are clouds, then it rains” or “if you drink beer, then you must be at least 21 years old.” In the first case, people are guided by knowledge that clouds are a necessary albeit insufficient cause for rain (e.g., causation schema), whereas in the second they are guided by knowledge that drinking age is established by a law that might be disobeyed (i.e., permission schema). As in the case of arithmetic operations (Bassok et al., 1997), the rules of formal logic are semantically compatible with the rules of some schemas but not others. For example, Modus Tollens (“if not q then not p ” ) is compatible with the rules of the permission schema (e.g., if you are under 21 [not 41, then you are not permitted to drink alcohol [not p ] ) . However, this rule sometimes conflicts with the rules of the causation schema (e.g.,
Object-Based Reasoning
35
if it doesn’t rain [not 41, it does not necessarily follow that there are no clouds [ n o t p ] ) .Cheng et al. (1986) found that, much as the mathematically sophisticated subjects in Bassok et al. who sometimes “failed” to construct addition or division word problems for semantically incompatible object sets, subjects who were trained in evaluating the validity of conditional syllogisms committed “logical errors” on test problems that induced the rules of incompatible pragmatic schemas. Cheng et al. (1996) argued convincingly that adherence to semantic and pragmatic constraints (i.e., content effects) protects people from arbitrary and anomalous conclusions. In fact, the results of Bassok et al. (1997) strongly suggest that, when application of formal rules conflicts with people’s semantic and pragmatic knowledge, they may prefer arriving at reasonable and logically invalid conclusions to arriving at logically valid but anomalous conclusions. Unfortunately, the notion that abstraction of structure from content is a mark of intellectual achievement appears t o be so appealing that even sensible responses to semantic constraints may be classified by researchers as reasoning errors. The explanatory parsimony that is implied by this notion may also explain why researchers prefer to construct and test content-independent accounts of reasoning. While testing such accounts, they typically average people’s responses to stimuli that differ in content, viewing such responses as measurement errors that obscure basic processing regularities (see Goldstein & Weber, 1995, for an insightful discussion and historical analysis of this view). This practice would most probably average out the very effects that were the focus of the studies described in this chapter. Object-based inferences that reflect adherence to semantic and pragmatic distinctions implied by content are inherent to rather than a deviation from normal processing. Also, such inferences do not indicate poor understanding, lack of maturity, or insufficient cognitive resources. Hence, effects of object-based inferences on reasoning do not fit well the “hidden-treasure” or any other variant of the faiure-of-abstraction-from-content metaphor. Instead, such effects seem to suggest a “toolbox” metaphor, whereby people attempt to find the best fit between their processing tools and the constraints implied by the stimuli they encounter. I strongly believe that, by adopting the toolbox metaphor, researchers are likely to discover interesting regularities in the way people adjust processing to their highly organized semantic and pragmatic knowledge. ACKNOWLEDGMENTS Preparation of this chapter was made possible by a grant from the University of Chicago School Mathematics Project (UCSMP 260039). I want to thank my collaborators for their
36
Miriam Bassok
intellectual and empirical contributions to the work described in this chapter. Special thanks go to Shirley Martin and Edward Wisniewski for their careful reading and excellent suggestions regarding this article.
REFERENCES Alibali, M. W., Bassok, M., Olseth, K. L. Syc, S., & Goldin-Meadow, S. (1995). Gestures reveal mental models of continuous and discrete change. In J. D. Moore & J . F. Lehman (Eds.), Proceedings of the seventeenth annual conference of the Cognitive Science Society (pp. 464-468). Hillsdale, NJ: Erlbaum. Bassok. M. (1990). Transfer of domain-specific problem solving procedures. Journal of Experimental Psychology: Learning, Memory, and Cognition, 16, 522-533. Bassok, M. (1997). Two types of reliance on correlations between content and structure in reasoning about word problems. In L. English (Ed.), Mathematical reasoning: Analogies, metaphors, and images (pp. 221-246). Hillsdale, NJ: Erlbaum. Bassok, M., Chase, V. M., & Martin, S . A. (1997). Adding apples and oranges: Semantic constraints on application of formal rirles. Manuscript under review. Bassok, M., & Holyoak, K. J. (1989). Interdomain transfer between isomorphic topics in algebra and physics. Journal of Experimental Psychology: Learning, Memory, and Cognition, 15, 153-166. Bassok, M., & Holyoak, K. J. (1993). Pragmatic knowledge and conceptual structure: Determinants of transfer between quantitative domains. In D. K. Detterman & R. J. Sternberg (Eds.), Transfer on trial: lnielligence, cognition, and instruction (pp. 68-98). Norwood, NJ: Ablex. Bassok. M., & Medin, D. L. (1997). Birds of a feather flock together: Similarity judgments with semantically-rich stimuli. Journal of Memory and Language, 36, 31 1-336 Bassok, M., & Olseth, K. L. (1995). Object-based representations: Transfer between cases of continuous and discrete models of change. Journal of Experimental Psychology: Learning, Memory, and Cognition, 21, 1522-1538. Bassok, M., Wu, L., & Olseth, K. L. (1995). Judging a book by its cover: Interpretative effects of content on problem solving transfer. Memory & Cognition, 23, 354-367. Beveridge, M., & Parkins, E. (1987). Visual representation in analogical problem solving. Memory & Cognition, 15, 230-237. Brown, A. L. (1989). Analogical learning and transfer. What develops? In S. Vosniadou & A. Ortony (Eds.), Simlarity and analogical reasoning (pp. 369-412). Cambridge, UK: Cambridge University Press. Catrambone, R., & Holyoak, K. J. (1989). Overcoming contextual limitations on problemsolving transfer. Journal of Experimental Psychology: Learning, Memory, iind Cognition, 15, 1147-1 156. Cheng, P. W., & Holyoak, K. J. (1985). Pragmatic reasoning schemas. Cognitive Psychology, 17, 391-416. Cheng, P. W., & Holyoak, K. J. (1989). On the natural selection of reasoning theories. Cognition, 33, 285-313. Cheng, P. W., Holyoak, K. J.. Nisbett, R. E., & Oliver, L. M. (1986). Pragmatic versus syntactic approaches to training deductive reasoning. Cognitive Psychology, 18, 293-328. Chi, M. T. H., Bassok, M., Lewis, M. W., Reimann, P., & Glaser, R. (1989). Self-explanations: How students study and use examples in learning to solve problems. Cognitive Science, I.?, 145-182.
Object-Based Reasoning
37
Chi, M. T. H.. Feltovich. P. J., & Glaser, R. (1981). Categorization and representation of physics problems by experts and novices. Cognitive Sciences, 5, 121-152. Cosmides, L. (1989). The logic of social exchange: Has natural selection shaped how humans reason? Studies with the Wason selection task. Cognition, 31, 187-276. Cosmides, L., & Tooby, J. (1994). Origins of domain specificity: The evolution of functional organization. In L. A. Hirshfeld & S. A. Gelman (Eds.), Mapping the mind: Domain specificity in cognition and culture (pp. 85-1 16). New York: Cambridge University Press. Cummins, D. D. (1995). Naive theories and causal deduction. Memory & Cognition, 23, 646-658. Duncker, K. (1945). On problem solving. Psychological Monographs, 58 (No. 270). Falkenhainer, B., Forbus, K. D., & Gentner, D. (1989). The structure mapping engine: Algorithm and examples. Artificial Intelligence, 41, 1-63. Forbus, K. D., Gentner, D., & Law, K. (1995). MACIFAC: A model of similarity-based retrieval. Cognitive Science, 19, 141-205. Gentner, D. (1983). Structure-mapping: A theoretical framework for analogy. Cognitive Science, 7, 155-170. Gentner, D., & Rattermann, M. J. (1991). Language and the career of similarity. In S. A. Gelman & J. P. Byrnes (Eds.), Perspectives on thought and language: Interrelations in development (pp. 225-277). London: Cambridge University Press. Gentner, D., & Toupin, C. (1986). Systematicity and surface similarity in the development of analogy. Cognitive Science, 10, 277-300. Gick, M. L., & Holyoak, K. J. (1980). Analogical problem solving. Cognitive Psychology, 12, 306-355. Gick, M. L., & Holyoak, K. J. (1983). Schema induction and analogical transfer. Cognitive Psychology, 15, 1-38. Gick, M. L., & McGarry, S. J. (1992). Learning from mistakes: Inducing analogous solution failures to a source problem produces later successes in analogical transfer. Journal of Experimental Psychology: Learning, Memory, and Cognition, 18, 623-639. Gigerenzer, G., & Hug, K. (1992). Domain-specific reasoning: Social contracts, cheating, and perspective change. Cognition, 43, 127-171. Goldstein, W. M., & Weber, E. U. (1995). Content and discontent: Indications and implications of domain specificity in preferential decision making. In J. R. Busemeyer, R. Hastie, & D. L. Medin (Eds.), The psychology of learning and motivation (Vol. 32), pp. 83-136. San Diego, CA: Academic Press. Goldstone, R. L. (1994). Similarity, interactive-activation and mapping. Journal of Experimental Psychology: Learning, Memory, and Cognition, 20, 3-28. Goldstone, R. L., & Medin, D. L. (1994). Time course of comparison. Journal of Experimental Psychology: Learning, Memory, and Cognition, 20, 29-50. Goldstone, R. L., Medin, D. L., & Gentner, D. (1991). Relational similarity and the nonindependence of features in similarity judgments. Cognitive Psychology, 23, 222-262. Halliday, D., & Resnick, R. (1981). Fundamentals of physics. New-York: Wiley. Hinsley, D. A., Hayes, J. R., & Simon, H. A. (1977). From words to equations: Meaning and representation in algebra word problems. In M. A. Just & P. A. Carpenter (Eds.), Cognitive processes in comprehension (pp. 89-106). Hillsdale, NJ: Erlbaum. Hofstadter. D. R., Mitchell, T. M.. & French, R. M. (1987). Fluid concepts and creative analogies: A theory and its computer implementation (Tech. Rep. 87-1). Ann Arbor: University of Michigan, Fluid Analogies Research Group. Holyoak, K. J. (1985). The pragmatics of analogical transfer. In G . H. Bower (Ed.), The psychology of learning and motivation (Vol. 19, pp. 59-87). New York: Academic Press.
38
Miriam Bassok
Holyoak, K. J., & Koh, K. (1987). Surface and structural similarity in analogical transfer. Memory & Cognition, 15, 332-340. Holyoak, K. J., & Thagard, P. (1989). Analogical mapping by constraint satisfaction. Cognitive Science, 13, 295-355. Hummel, J. E., & Holyoak, K. J. (in press). Distributed representations of structure: A theory of analogical access and mapping. Psychological Review. Kotovsky. K., Hayes, J. R., & Simon, H. A. (1985). Why are some problems hard?: Evidence from tower of hanoi. Cognitive Psychology, 17, 248-294. Lakoff, G., & Turner, M. (1989). More than cool reason: The power of poetic metaphor. Chicago: University of Chicago Press. Larkin, J. H., McDermott, J., Simon, D. P., & Simon, H. A. (1980). Models of competence in solving physics problems. Cognitive Science. 4, 317-348. Lin, E. L. (1996). Thematic relations in adult’s concepts and categorization. Unpublished doctoral dissertation, University of Illinois, Urbana. Luria, A. M. (1976). Cognitive development its cultural and social foitndations. Cambridge, MA: Harvard University Press. Markman, A. B.. & Gentner, D. (1993). Structural alignment during similarity comparisons. Cognitive P.yychology, 25, 431 -467. Markman, A. B., & Gentner, D. (1996). Commonalities and differences in similarity comparisons. Memory & Cognition, 24, 235-249. Markman, E. M. (1989). Categorization and naming in children. Cambridge, MA: MIT Press. Medin, D. L., Goldstone, R. L.. & Gentner, D. (1990). Similarity involving attributes and relations: Judgments of similarity and difference are not inverse. Psychological Science, 1, 64-69. Medin, D. L., Goldstone, R. L., & Gentner, D. (1993). Respects for similarity. Psychological Review, 100, 254-278. Medin, D. L., & Ortony, A. (1989). Psychological essentialism. In S. Vosniadou & A. Ortony (Eds.), Similarity and analogical reasoning (pp. 179-195). Cambridge, UK: Cambridge University Press. Medin, D. L., & Ross, B. H. (1989). The specific character of abstract thought: Categorization, problem solving. and induction. In R. J. Sternberg (Ed.), Advances in the psychology o f human intelligence (Vol. 5, pp. 189-223). Hillsdale, NJ: Erlbaurn. Needham, D., & Begg, 1. (1991). Problem-oriented training promotes spontaneous analogical transfer, memory-oriented training promotes memory for training. Memory & Cognition. 19, 543-557. Novick, L. R. (1988). Analogical transfer, problem similarity, and expertise. Journal of Experimental Psychology: Learning, Memory, and Cognition, 14, 5 10-520. Novick, L. R. (1992). The role of expertise in solving arithmetic and algebra word problems by analogy. In J. I. D. Campbell (Ed.), The nature and origins of’mathematicalskills (pp. 155-188). Amsterdam: Elsevier. Paige, J. M., & Simon, H. A. (1966). Cognitive processes in solving algebra word problems. In B. Kleinmuntz (Ed.), Problem solving: Research, method, and theory (pp. 51-119). New York: Wiley. Reed, S. K. (1993). A schema-based theory of transfer. In D. K. Detterman & R. J. Sternberg (Eds.), Transfer on trial: Intelligence, cognition, and instruction (pp. 39-67). Norwood, NJ: Ablex. Rcevcs, L. M., & Wcisbcrg, R. W. (1994). Thc role of content and abstract information in analogical transfer. Psychological Bulletin, 115, 381-400. Ross. B. H. (1987). This is like that: The use of earlier problems and the separation of similarity effects. Journal of Experimental Psychology: Learning, Memory, and Cognition, I S , 629-639.
Object-Based Reasoning
39
Ross, B. H. (1989). Distinguishing types of superficial similarities: Different effects on the access and use of earlier problems. Journal of Experimental Psychology: Learning, Memory, and Cognition, 15, 456-468. Ross, B. H., & Kennedy, P. T. (1990). Generalizing from the use of earlier examples in problem solving. Journal of Experimental Psychology: Learning, Memory, and Cognition, 16, 42-55. Schoenfeld, A. H., & Herrmann, D. J. (1982). Problem perception and knowledge structure in expert and novice mathematical problem solvers. Journal of Experimental Psychology: Learning, Memory, and Cognition, 5, 484-494. Silver, E. A. (1981). Recall of mathematical problem information: Solving related problems. Journal of Research in Mathematics Education, 12, 54-64. Thagard, P. H.. Holyoak, K. J . , Nelson, G., & Gochfeld, D. (1990). Analog retrieval by constraint satisfaction. Artificial Intelligence, 46, 259-310. Tversky. A. (1977). Features of similarity. Psychological Review, 84, 327-352. Wisniewski. E. J. (1996). Construal and similarity in conceptual combination. Journal of Memory and Language, 35, 434-453. Wisniewski, E. J., & Bassok, M. (1996). On putting milk in coffee: The effect of thematic relations on similarity judgments. In G. W. Cotrell (Ed.), Proceedings of the eighteenth annual conference of the Cognitive Science Society (pp. 464-468). Hillsdale, NJ: Erlbaum.
This Page Intentionally Left Blank
ENCODING SPATIAL REPRESENTATIONS THROUGH NONVISUALLY GUIDED LOCOMOTION Tests of Human Path Integration Roberta L. Klatzky Jack M . Loomis Reginald G. Golledge
This chapter describes how people navigate within a space that they do not see, but learn about by walking through it. In doing so, they invoke a broad set of processes that include representing position and orientation with respect to spatial referents, updating the representation over the course of movement, planning routes, and executing those routes. These processes are components of navigation in general. Our particular focus on navigation without sight stems from a long-standing interest in the navigation abilities of the blind. The research described in this chapter is part of a larger project, which has the applied goal of developing a prototype navigation aid for blind people (Loomis, Golledge, & Klatzky, 1996). We emphasize throughout this work that the ability to navigate independently relies on an internal representation of spatial knowledge. Thus, research on navigation involves studying the learning, modification, and use of spatial representations. Our specific interest is in how these processes occur in the absence of visual information. This chapter describes our efforts to (1) acquire basic data about human navigation without vision and (2) theoretically characterize representations and processes associated with it. In keeping with our emphasis on the acquisition of spatial representations, we are particularly interested in the THE PSYCHOLOGY OF LEARNING AND MOTIVATION. VOL. 37
41
Copyright Q tW7 by Academic Press. All rights of reproduction in any form reserved. 0079-7421/97 $25.00
42
Roberta L. Klatzky et al.
process of encoding spatial knowledge. When people are asked to learn about a space by walking through it, in the absence of vision, they experience efferent commands and nonvisual sensory consequences of walking. These cues lead, we propose, to the encoding of an internal representation of the path that has been walked. This initial representation may be expanded by processes of spatial inference. Encoded and inferred knowledge provides the basis for goal-oriented responses. In this chapter we describe and evaluate an encoding-error model (Fujita, Klatzky, Loomis, & Golledge, 1993), which attributes systematic errors in navigation to errors in the process of encoding spatial information through locomotion. We begin the chapter by defining essential concepts and reviewing background findings related to navigation. We then describe a substantial body of work that we have conducted on nonvisual navigation by blind and sighted, blindfolded individuals. These results are of interest in evaluating how well people can track their position in space without sight of landmarks, a process called path integration. We use the data to develop and test the encoding-error model.
I. Navigational Concepts Several concepts are illustrated in Fig. 1, and critical terms are defined in these paragraphs. We begin with the concept of a spatial representation. As a general formalism (Palmer, 1978), a representation can be viewed as a mapping from some “represented world” into a “representing world.” Each of these worlds has (1)elements and (2) interelement relations. In the case of a spatial representation, the source of the mapping-the represented world-includes a collection of points and objects within some region of space, and some set of relations among them. In particular, the relationships between the points and objects in the world are instantiated by their arrangement in physical space. The content of the representing world is some subset of the elements and relations in the represented world. Critical to the nature of a spatial representation is that at least some of its relational components convey information about the spatial relations in the represented world. (Other information, such as semantic associations among objects, might be linked to the spatial knowledge.) In a formal representation of space, it is customary to specify the locations of points by a coordinatesystem. In a plane, two common coordinate systems are Cartesian coordinates and polar coordinates. Both systems specify two parameters, which convey the relation between each point and an origin. From these, additional relations, such as interpoint distances, can be derived.
Encoding Spatial Representations
43 landmark at
(rsp’)
,bearing to landmark (p’ ) origin
G
L
b reference direction heading
relative bearing
to origin
/
..
bearing to origin
0
I I I navigator‘s track
k
course I ’
Fig. 1. Definitions of navigational terms. An origin and reference direction are depicted, which allow the location of a landmark (box) to be specified. A navigator is also shown, following a track from the bottom to the center of the figure. At the final point, the navigator has turned, so that heading is not the same as course.
Within a spatial representation, certain locations may be associated with objects, which have distinctive features and/or labels. Depending on the scale of the representation relative to the size of an object, it might occupy a single point, a sequence of contiguous points, or an area. (We are accustomed to highway maps in which cities, roads, and lakes are so depicted, respectively.) The distinctive features that define objects can be sensed not only with vision, but by other modalities as well. They may be textures under foot, sounds, smells, and characteristic wind currents. The term landmark is sometimes used to refer to any salient location in the space, but it can also more specifically denote an object that is used by a navigator to determine his or her own location. The navigator is an object in the space whose position is variable, and to whom can be ascribed the goal of traveling within the space. The points occupied by objects and navigators in the space are specified by the parameters of the coordinate system. In polar coordinates, a point
44
Roberta L. Klatzky et al.
can be said to be at angle 4 relative to the reference axis and at distance p from the origin. But a navigator within the space does more than occupy a location; he or she also has an orientation in space and may have a path of movement through it. Thus, additional parameters are necessary to describe the situation at any time. One is the navigator’s heading, or direction of body orientation, specified as the angle between a sagittal plane through the navigator’s body and a reference direction. Because human navigators can look around while moving, it is sometimes useful to distinguish between the heading defined by the body and that defined by the head. Finally, if the navigator moves through space, the direction of local movement, expressed relative to the reference axis, is the course. A turn is a change in heading (or course). Navigators can be positioned not only with respect to the origin but with respect to landmarks in the space, and landmarks can be positioned relative to one another. Two parameters used to describe the relative positions of two points in space are the bearing and distance between them. The distance is a metric relation, most commonly Euclidean but sometimes of another type such as city-block. The bearing from location 1 to location 2 is the angle between the reference direction and a line originating at location 1, directed toward location 2. If a navigator’s body is oriented toward an object, the bearing of the object is the same as the navigator’s heading. To describe situations where the heading is different from the bearing, we label the difference as the relative bearing of the object. In order to move to an object in his or her spatial representation, a navigator must know the appropriate turn that would orient him or her to the object’s position, as well as the distance to travel. The desired turn is the difference between the navigator’s current heading and the bearing to the desired location: that is, the navigator should turn the value of the relative bearing. As navigators move through space, they often wish to keep track of their position relative to other objects within it. Two means of doing so have been distinguished: landmark-based navigation and path integration (see Gallistel, 1990). For an example of landmark-based navigation, suppose a navigator knows that landmark A is located at Cartesian coordinates ( x , y ) . Suppose, too, that the navigator can determine the distance ( d ) and bearing (4) of Landmark A relative to his or her current position. Then the navigator’s Cartesian coordinates can be determined. The Cartesian coordinates can also be computed in a number of other ways, including use of the distances to three landmarks or the bearings to two landmarks. Path integration (sometimes called dead reckoning) refers to updating position in space without reference to landmarks, by means of velocity and acceleration signals (for reviews in nonhuman species, see Etienne,
Encoding Spatial Representations
45
Maurer, & SCguinot, 1996; Gallistel, 1990; Maurer & SCguinot, 1995). With both types of signals, the navigator updates position at some interval, such as at each footstep. For example, if the initial position is x meters and y meters from the origin along two orthogonal axes, and the current velocity along the directions specified by the two axes is m meters per footstep and n meters per footstep, respectively, then the position after one footstep will be (x + m, y + n). This calculation integrates velocity over time, in each orthogonal direction, to determine distance traveled. This type of model was suggested by H. Mittelstaedt and Mittelstaedt (1973) for path integration in insects. Path integration provides a navigator with information that will allow a return to the origin of travel. The required information to generate the homeward response is what we call the homing vector (Fujita, Loomis, Klatzky, & Golledge, 1990), which indicates the turn and distance that are required in order to return to the origin. One issue we consider later is how frequently a homing vector is updated in human path integration. It is possible that updating is fine grained, on the order of a footstep, which we will call moment-to-moment updating. Another possibility we consider is that updating occurs at salient points along the route, such as at the ends of straight-line segments where turns or stops occur.
11. General Assumptions and Findings Regarding
Human Navigation We now turn to some general assumptions and findings that have grown out of past psychological research on spatial representation and navigation. These form a starting point for our own work. TO NAVIGATE GROWS OUTOF AN INTERNAL A. THEABILITY OF SPACE REPRESENTATION
This assumption was controversial in the time of Tolman (1948), but most psychologists would now agree that organisms of a reasonable evolutionary status can represent the layout of subjects-including their own bodies-in space. The information conveyed by a spatial representation may vary widely, as we discuss in Section C . The homing vector provides the minimal information needed for a return to the origin, but by itself, it does not allow the navigator to retrace a circuitous pathway that brought him or her to a given point. A more complete representation than the homing vector might provide a prescription of the outbound pathway. Depending on the compu-
46
Roberta L. Klatzky et al.
tational capabilities of the navigator (see later discussion), this representation might permit a direct homeward trajectory to be computed, or it might allow the outbound pathway to be traced in reverse. Still more complete representations would specify information external to experienced pathways and would allow for excursions from known routes toward novel targets.
B. REPRESENTATIONS OF SPACE REFLECT EXPERIENCES WITHIN A N D OUTSIDE IT People can learn about a spatial layout in various ways. They can view a map, bear a verbal description, or view the layout from an external site with a global perspective. They can also learn by traveling through the layout. A considerable body of research has been devoted to differences in the representations of space derived from differing learning histories, such as maps versus travel (e.g.. Evans & Pezdek, 1980;Presson & Hazelrigg, 1984; Sholl, 1987; Thorndyke & Hayes-Roth, 1982). For example, learning from a map appears to produce an advantage in judging straight-line distances between locations, whereas learning from traveling within a space is superior for judging route distances (Thorndyke & Hayes-Roth, 1982). Hart and Moore (1973), Lynch (1960), Piaget and Inhelder (1967), and Siege1and White (1975) were among the first to discuss how representations changed with experience and during maturation. The general view suggested by these theories is that representations progress from an awareness of objects, with minimal spatial structure, to knowledge of the relation of objects to one another and to the observer. Gallistel (1990) presented a model describing how various nonhuman species learn about space from travel within it. The model describes the formation of a cognitive map (sometimes called a survey representation) that represents the distances and directions of objects (or surfaces) in the space relative to one another. According to the model, a critical component of such learning is an underlying path integration process that indicates the animal’s position in the space relative to an external, or geocentric, coordinate system. Whereas sensed landmarks do not contribute to the geocentric localization process, vision and other senses become important in localizing objects relative to the exploring animal (egocentric localization). The egocentric coordinates are computationally converted to geocentric ones, leading to a map of object-to-object relations. According to Sholl (1996), (sighted) humans, in comparison to lower animals, rely more on visual cues and less on path integration to construct a cognitive map. She argues in particular that the invariant structure within optical flow fields is critical to the development of a geocentric cognitive
Encoding Spatial Representations
47
map. This structure remains constant under transformation of the viewer’s perspective and conveys information about interobject relations. She suggests that the level of performance of human subjects in tasks requiring path integration indicates that the process is inadequate to support the formation of a geocentric map. Later, we present data on performance of blind and sighted in such tasks, from our own studies. These data exhibit systematic error, but they indicate that performance is well above chance. C. MULTIPLE REPRESENTATIONS MAYBE AVAILABLE FOR THE SAMESPATIAL LAYOUT
Several distinctions among types of representations that people can hold, after experience with a space, have been suggested. One differentiates between a spatial representation and a motor trace. Whereas a spatial representation conveys the positions of points relative to a designated origin and/or to one another, a motor trace is a memorial representation of previous action within the space. Its sources are efference commands and the sensory consequences of action, including cutaneous, kinesthetic, and vestibular signals. With active movement, the motor trace incorporates both proprioceptive and spatial elements, the latter pertaining to the intended spatial consequences of the action. The importance of spatial intentions is indicated by the superiority of memory for active over passive movements (e.g., Stelmach, Kelso, & McCullagh, 1976), and by the finding that blocking of proprioceptive information has little effect on the reproduction of the end-point location of a movement (Kelso, 1977; Kelso & Holt, 1980). Proprioceptive blocking does, however, impair the ability to reproduce a previously moved distance from a new location. These findings indicate that purely proprioceptive traces support reenactment of previous motions, independent of the spatial location in which they occurred, and that nonproprioceptive information can convey spatial goals of action. Another important distinction is between representations that convey route and survey knowledge of a space. Although these terms have not always been consistently defined, the general idea is that route knowledge is sufficient to direct movement over a fixed sequence of locations, and survey knowledge provides information about points that are not on the route or that are not at adjacent locations within it. The term cognitive map is often treated synonymously with survey knowledge. The distinction between route and survey types of representation is fundamentally one of information content, that is, what the representation conveys. In these terms, the route-survey dichotomy refers to a distinction between representations that (1) convey the distance and relative bearing between only those points in space that are reached successively during
48
Roberta L. Klatzky et al.
travel, and (2) convey relative bearings and distances between additional points. The former is a route representation, the latter a survey one. It is important to note, however, that the completeness of either of these types of content could vary considerably. That is, it is not necessary that a representation that conveys relations between some points, either on or off a route, must convey all possible relations. Thus, the terms route and survey place limits on what spatial relations can be represented, but they should not, in our view, be taken to indicate completeness of representation. The route-survey distinction is sometimes confused with another, based on vantage point. This term refers to the imagined perspective that can be taken by an individual who is retrieving information from a representation (cf. Thorndyke & Hayes-Roth, 1982). The individual may imagine viewing the space from a particular location and orientation within it. Alternatively, he or she may imagine viewing the space from an external location, in particular, a bird’s-eye view. In making a distinction based on vantage point, we do not claim that a navigator must have a phenomenological experience of seeing the space from a specific point of view, when accessing a spatial representation. One can retrieve information content without an imagined visual perspective. A final distinction, which was referred to earlier, is between geocentric and egocentric coordinate systems. The former is centered on some point in space external to the navigator (such as an animal’s nest) and the latter on the navigator. The information content of a representation with an egocentric system is restricted, at most, to the set of bearings and distances from the navigator to each object in the space. These distinctions are somewhat interrelated, as follows: The bird’seye vantage point will provide survey information within a geocentric coordinate system (because the observer is represented as external to the space). The vantage point from within the space provides more limited image content, because it is restricted to those points that can be retrieved from the vantage point (in a visual image, points within the imagined field of view). This content need not be restricted to a route, and hence may constitute survey knowledge by our definition, but it is possible that only egocentric relations (i.e., distances and bearings to objects that are retrievable from the vantage point) are represented. Loomis, Da silva, Fujita, and Fukusima (1992) have found that in viewing physical space, people may have accurate information about self-toobject distances but not about object-to-object (exocentric) distances. This could also be true when distances are retrieved from an adopted vantage point within an internal spatial representation, in which case the representation would be egocentric.
Encoding Spatial Representations
49
CAN BE CHANGED BY PERCEPTUAL D. REPRESENTATIONS UPDATING AND COGNITIVE CONTROL
Phenomenologically, it appears clear that people can move at will between representations of space. One can change vantage point, for example, by imagining one’s kitchen from a bird’s-eye view or from some point within it. One can also imagine oneself moving from one viewpoint to another. In keeping with these intuitions, it has been demonstrated that after learning the locations of landmarks by visual exposure or nonvisually guided travel from a source location, individuals can update their position relative to those landmarks during locomotion without vision (Loomis et al., 1992; 1993; Rieser, 1989; Rieser, Guth, & Hill, 1982, 1986). Updating changes the positions of objects in egocentric coordinates, and if there is an adopted vantage point, that will change concomitantly with movement. Further research has addressed how people update position within a representation over the course of imagined (cf. physical) movement. This introduces some complications, due to the discrepancy between the physical location and orientation of the individual and the imagined location and orientation. Using a task in which individuals pointed without sight to objects in a space, from an imagined position and orientation, Rieser (1989) showed that there is a decrement in accessing information about space if one’s imagined and physical locations differ. The decrement was greater after imagined rotation than after linear translation, and the time to respond after imagined rotation tended to increase with the angular difference between the physical and imagined orientation (Easton & Sholl, 1995; Farrell & Robertson, 1996; May, 1996; Presson & Montello, 1994; Rieser, 1989). Rieser has proposed (1989; Rieser et al., 1986) that the process of spatial updating during real movements differs from the process that accompanies imagined movement. During physical translation or rotation, even without vision, updating of one’s distance and bearing to landmarks is thought to occur through automatic perceptual processes. Updating after imagined rotations, and in some cases imagined translations (Easton & Sholl, 1995), in contrast, apparently requires effortful cognitive processing. A further complexity when movement is only imagined is that the individual maintains two representations of spatial position, actual and imaginal, which may mutually interfere (May, 1996; Presson, 1987). Thus, disorienting a subject before an imagined rotation reduces the difficulty of representing it (May, 1996). It appears important to distinguish between updating of one’s position in space, in terms of a representation of position relative to other objects within it, and updating a homing vector that indicates the turn and distance
50
Roberta L. Klatzky et al.
responses that are needed to return to an origin of travel. Loomis and associates (1992) showed that people who first sighted an object, then traveled without vision along a path that passed by it, could keep their finger pointed at the object as they moved. This indicates that they updated their position relative to the object virtually continuously. Yet we present data suggesting that people who follow routes without vision only intermittently update the responses needed to return to the origin.
E. NOTALLTHATIS IMPLIED BY THE INFORMATION CONTENT OF A REPRESENTATION NEEDBE COMPUTABLE FROM IT Any representation makes available some information content, from which it may be possible to compute further information. For example, a person who knows the bearing and distance of two landmarks from one location in a space can, in theory, compute the bearing of one landmark from the other. A person who is asked to point to targets after assuming an imagined position and orientation (e.g., Rieser et al., 1986) is essentially being asked to do just that. As we have noted, some difficulties arise in doing so, suggesting that theory and reality may differ. In fact, the existence of algorithms for computing new information from a spatial representation does not mean that the cognitive apparatus is capable of doing the computation. That is, it cannot be inferred that we know the information that is implied, as a second-order computation or spatial inference, by information directly available in a representation. For example, if one forms a representation of oneself standing at the sink, and the relative bearing and distances to the refrigerator and stove are represented, it is not necessarily the case that the distance from the sink to the stove is known. As was mentioned above, Loomis and associates (1992) have obtained evidence that the latter, exocentric distances may not be accurately estimated even when the egocentric distances are known and sight of the objects is available.
F. THEREPRESENTATION OF SPATIAL LAYOUT A N D THE AVAILABLE COMPUTATIONS BOTHENABLE AND CONSTRAIN THE ABILITY To NAVIGATE This point may seem obvious, but it is worth emphasizing. Lacking a representation that conveys the distance and relative bearing between each pair of successive points on a route, a navigator will be unable to complete the route. But given such knowledge, the navigator is presumably able to follow the route by walking the distance and turning the amount of the relative bearing, for each segment in succession. In order to trace a route in the reverse direction, more computational processes are necessary. The naviga-
Encoding Spatial Representations
51
tor must know that reversing means signing the relative bearings and distances in the reverse order. (See Fig. 10 for failures to do just that.) Still more is necessary in order to take a shortcut, because a list of relative bearings and distances between landmarks on a route is not sufficient to travel directly between any two of them. It is necessary to have the computational apparatus to determine the direct bearing and distance. And even a representation that conveys all possible shortcuts along a route is not sufficient for the navigator to orient toward landmarks that are off the route.
111. Representations and Processes Underlying Shortcuts
in Space Taking a shortcut between two points in space, after traveling indirectly between them (i.e., via at least one noncollinear point), is often taken to be a sign that a navigator has more than a route representation. In our terms, the navigator has a survey representation, as its content conveys the distance and direction between points that were not encountered sequentially en route. We view short-cutting as a navigational task that incorporates many of the processes described earlier: forming an internal representation based on experience with the space, computing the responses needed to execute the shortcut (if possible), and using the ensuing representation to guide movement. We have used the task in much of our research, because it requires an act of spatial inference, that is, some form of computation from knowledge about the route. Returning directly to the origin of travel, after following a circuitous route, is a specific form of spatial short-cutting, which we call path completion. As we have noted, navigation by path integration should provide the ability to return home by a direct route. The process of path integration provides the information needed to compute a homing vector, which conveys the turn and distance required in order to reach the origin from the navigator’s current location and heading. Path completion tasks have been used by a number of researchers studying animal and human navigation (for reviews of research with mammals and birds, respectively, see Etienne et al., 1996, and Walcott, 1996). Among animals that have been studied are the desert ant (Muller & Wehner, 1988; Wehner & Wehner, 1986), the desert woodlouse (Hoffmann, 1984), the funnel-web spider (H. Mittelstaedt, 1985), the golden hamster (Etienne, Maurer, & Saucy, 1988; SCguinot, Maurer, & Etienne, 1993), the gerbil (H. Mittelstaedt & Mittelstaedt, 1982; M. Mittelstaedt & Glasauer, 1991), and several types of goslings (von Saint Paul, 1982). Even when denied access
Roberta L. Klatzky et al.
52
to landmarks, and hence restricted to navigation by path integration, these species exhibit an impressive ability to return to the origin of travel. There are also several studies in which humans are asked to return or point to nonsuccessive locations along a route that was traveled without vision (Dodds, Howarth, & Carter, 1983; Juurma & Suonio, 1975; Klatzky et al., 1990; Landau, Spelke, & Gleitman, 1984; Loomis et al., 1993; Sauve, 1989; Sholl, 1989; Worchel, 1951, 1952; Yamamoto, 1991). In general, performance declines as the path that was traveled becomes more complex, for example, as it increases from two to four segments (Klatzky et al., 1990; Loomis et al., 1993; Sholl, 1989). A similar effect has been found when subjects point to initially sighted targets that lie off paths of varying complexity (Rieser & Rider, 1991). The computations involved in path integration can be formally specified by trigonometric rules. Above, we gave a formula for path integration on a footstep-by-footstep basis. Alternatively, the computation can be done at the end of a series of linear segments and turns. Specifically, the law of cosines indicates the distance of a shortcut between two points, when they were connected by two linear segments with a turn in between. Where the segments are lengths A and B and the internal angle (the complement of the turn) is 13,the shortcut distance C, is specified by:
C
=
(A’
+ B2
-
2A B cos O).”’
(2)
Once C is known, the same law can be applied to determine the angle between the second and third leg, which is the complement of the required turn toward home. In this case, the shortcut completes a triangle. Although trigonometry captures the relation of the shortcut distance to the initial two legs and turn between them, we do not suggest that people have internalized the law of cosines and can apply it mentally. An alternative mechanism, and one that seems more psychologically plausible, is that people create a spatial representation and compute the size of the third leg and homeward turn by a process analogous to perceiving them. Note, however, that if the representation is a visual image, it presumably rescales the original space, so that only the relative value of the third leg can be determined directly from it (assuming no distortion). In order for the representation to guide production of the shortcut distance, this value must be adjusted for scale by using ad hoc knowledge of the magnitudes of the initial two legs. The homeward turn, in contrast, retains its value under rescaling. Much of our experimental work has focused on the ability of individuals to complete a path that they experienced without sight. In a typical experimental trial, the subject is led through a number of straight-line segments
Encoding Spatial Representations
53
with turns in between, and he or she is then asked to return directly to the origin. Underlying performance in the path-completion task are, at a minimum, four component processes (Fujita et al., 1993; Loomis et al., 1993). These include 1. sensing the pathway, 2. buiding a route representation, 3. computing (and representing) the desired trajectory back to the origin, and 4. executing that trajectory. We have generally treated together the first two of these processes, which culminate in a representation suitable for computing the homeward trajectory. We refer to these processes jointly as encoding. The first three processes, taken together, constitute path integration. This process description assumes that there are two levels of representation involved in path completion. One is encoded from direct experience with the path legs and turns. It is formed as a preliminary to the computation of the homeward trajectory, which expands the representation to indicate the completed path, constituting the second level. Where do errors come from in path completion? Errors are potentially of two types, systematic and stochastic (noise). The former will emerge in measures of average signed error, defined as the observed response (turn or distance) minus the correct value. Stochastic errors emerge in measures of intrasubject variability. (Absolute values of signed error contain both components.) Here, we are concerned with systematic errors. Those could arise at any of the four processes we have proposed. The model discussed in the remainder of this chapter takes the stance that they arise from early processes of encoding. We next describe the model.
IV. The Encoding-Error Model of Path Completion The encoding-error model begins with the multiprocess description of returning directly to an origin of travel, after following a circuitous route. The model attempts to account for systematic errors in observed responses from the encoding processes alone, that is, errors in sensing and deriving the representation from which the homeward trajectory is computed. The model assumes that later stages-trajectory computation and executionare not subject to systematic error. According to the model, all systematic error arises when people are passively guided along an outbound path. The distortions that arise in their encoding of the path lead to computation of an incorrect response
54
Roberts L. Klatzky et al.
trajectory, although the computational process itself is not subject to systematic error. Further, when they actively execute the computed responses, they again do so without introducing systematic error. Note, in particular, that they do not introduce error in the homeward turn and distance by encoding them with the same distortions that were imposed on the outbound path. Figures 2 and 3 indicate how systematic distortions in encoding could cause errors in path completion. In the top panel of Fig. 2, an organism misencodes the traveled distance per unit time (i.e., speed) by 50%. At any point in time, then, the organism thinks it has traveled only half as far as it actually has. Its internal representation is like a small-scale version of
Mental distance = 0.5
X
physical distance
physical
Mental turn = 0.8 x physical turn
Fig. 2. Examples of the effects of mis-encoding on performance. T o p panel: Underencoding of a distance results in a scaled-down representation of the pathway. Botrom panel: Underencoding of turn results in a distortion of pathway shape.
Encoding Spatial Representations
55
the actual path. If it were asked to return home at its stopping point, and if it correctly computed and executed the trajectory from the misencoded position, it would underwalk the required distance because of the reduced scale of the internal representation. In the bottom panel of Fig. 2, the organism misencodes its angular turn per unit time, devaluing it by 20% of the actual value. Now the encoded representation retains the total distance traveled but transforms its shape. If the organism attempted to return to the origin from the stopping point, it would err in both distance and direction, since the changed shape has altered both the bearing and distance of the navigator relative to the origin. Figure 3 shows in detail how errors in encoding could affect completion of a simple triangular shape. The navigator’s traveled path is shown by the solid line and circles; the dotted line and open circles show the encoded path. At the end of the second leg, the navigator has underencoded the lengths of the two legs and the initial turn between them. He or she correctly computes and executes the required distance and turn to return to the origin from the mentally represented stopping point. However, the turn and distance are executed, in physical space, from the actual end of the second leg. As a result, the observed end point is far from the origin of travel. We derive estimates of the accuracy of encoding from performance on completing a multisegment path. We do so by fitting a model to data from observed endpoint
response distance computed turn
distance 1actual path O--- 0 encoded path
Fig. 3. Effects of encoding error on triangle completion. The solid line represents the walked path, the dashed line represents the encoded path. Mis-encoding of the leg lengths and turns results in an erroneous representation of position at the end of the second leg. Although the response turn and distance are correctly computed and executed relative to the representation of the path end point, there is substantial error with respect to its physical location.
56
Roberta L. Klatzky et al.
the task. By this means, parameters are derived that represent the encoded values of the segments and turns experienced by the navigator. By plotting the encoded values against the values that the navigator actually experienced, we compute an encoding function. This function indicates the sources of systematic errors in completing the pathway. More specifically, the encoding-error model is fit t o a set of data as follows: Typically, the data arise from a path-completion task in which the navigator completes each member of a set of paths. The pathways are constructed by combining some values of linear segment lengths and turns. For example, in a study described by Loomis and associates (1993), 27 twosegment pathways were constructed by using 2, 4,or 6 m as the length of each leg and having turns between legs take on the value of 60,90, or 120". The subject is guided along the initial two legs and turn by an experimenter. At the end of the second leg, he or she attempts to return directly to the origin, and the response turn and distance are measured. Errors in these responses are assumed to arise because of misencoding of the path parameters. That is, there is assumed to be an internal, encoded value corresponding to each actual value of leg length and turn. The model predicts the subject's errors by: (1) assuming a set of encoded values; (2) computing the homeward turn and walked distance from these encoded values, without introducing systematic (cf. stochastic) error in the computation: and ( 3 ) assuming that the computed values are executed without systematic error. Thus, the only systematic errors in the process arise through a mismatch between internalized values and correct values of the leg lengths and turns in the experienced paths. The parameters estimated by the model are the set of encoded values in step (1). There is one parameter for each unique value of segment length and one for each unique value of turn that the navigator experienced. The best-fitting parameter values are chosen so as to minimize the difference between performance of the navigator and that predicted by the model. The minimized difference is the model error, used to determine goodness of fit. Our principal measure of model error is based on the subject's stopping point in space. Specifically, the model error is the distance between model-predicted and observed stopping points, summed over paths. Once the model has been fit to the data, and there is an estimate of the encoded value corresponding to each actual value of segment length and turn in the set of pathways, the encoded values can be plotted against the actual values. This gives rise to an encoding function for leg lengths and another for turn. We discuss the form of these functions, as fit to two different data sets, later in this chapter.
Encoding Spatial Representations
51
A. ESTIMATION OF ENCODING FROM SIMPLE AND COMPLEX TASKS Why do we use completion of complex pathways to test the encoding-error model? After all, the model’s principal claim is that systematic error arises during encoding; one might, therefore, think that an appropriate test of that assumption is to examine encoding directly. However, direct examination of encoding is not as straightforward as it might appear. In order to examine encoding directly, subjects can be asked to verbally estimate segment lengths and turns. This process requires a translation into a linguistic representation, which itself could introduce systematic error (Banks & Hill, 1974). As an alternative to verbal report, the accuracy of encoding can be tested by asking subjects to reproduce single pathway segments and turns. However, one should not rely on reproduction tests by themselves as a measure of encoding, because they might obscure systematic error. The problem arises because the error in encoding might be compensated for by a corresponding error in reproduction. Consider, for example, a subject who encodes a 1-m footstep as .5 m. When walking a 10-m path, the subject derives an internal representation of 5 m. But when reproducing the path, the subject also encodes each replication footstep as .5 m; as a result, she must walk 10 m to have a representation of walking 5 m. The subject will thus accurately reproduce the 10-m path, all the while having an inaccurate internal representation of its length. This compensatory error in encoding and reproducing a path is similar to what would happen if a navigator continued misencoding when returning from a circuitous route. In the top panel of Fig. 2, for example, if the navigator continued to encode the path as a 50%-scale version of its actual size while returning to the origin, he or she would still reach the correct origin, albeit with a misrepresentation of his or her history of travel. The reencoding during the return home, which would mask inaccuracies in initial encoding of the outbound path, is specifically excluded in the assumptions of the encoding-error model. When used in tandem, verbal reports and reproduction responses can provide mutual checks on encoding. The reproduction performance indicates biases induced by numerical responses, if any, and the verbal estimates provide a check on whether accurate reproduction is masking grossly inaccurate internalization of stimulus magnitudes. There remains, however, another problem with using either verbal estimation or reproduction as a measure of encoding by means of path integration. The problem is that the estimates or reproduction responses might be based on information other than that derived from path integration, in the form of motor traces. We return to this issue later.
58
Roberta L. Klatzky et al.
V. General Methodology of Reported Experiments In the remainder of this chapter, we present empirical findings related to navigation without sight, and we interpret them in the context of a set of issues regarding encoding. In our experiments, subjects are typically exposed to information about a path in space by walking along it without vision. Auditory cues are minimized by maintaining quiet environments and, in some cases, by using a single omnidirectional microphone connected to headphones, which prevents localization of external sounds. Remote light or sound sources and surface gradients are not present. We therefore require subjects to rely on path integration in the absence of any azimuthal referents (which could be used to define course, without the need to integrate turns). We have, in a few cases, used haptically distinctive landmarks. Path integration is still necessary in such cases, because the landmarks must be proximally sensed. To determine the bearing and distance between points in the space, be they marked by a distinctive feature or undistinguished, navigators must encode some information about the pathway from one to another. In a typical task in our experiments, subjects are guided to some point in a space and then asked to produce a response that brings them to a new stopping point. We measure either the position of the subject at the stopping point or the turn and distance that were executed in order to reach it. For the most part, subjects’ responses have been measured either with a tracking system that records the response trajectory (Loomis, Herbert, & Cicinelli, 1990) or with an ultrasonic distance-measurement system, which allows us to determine the subject’s stopping point and from it infer the response turn and distance. In some tasks, as was noted earlier, we also obtain verbal estimates of turns or distances.
VI. Encoding as Inferred from Reproduction and Verbal Report of Simple Paths In this section, we ask how well people can encode information about a simple path, consisting of a single straight leg or turn. Encoding is assessed by interrogation of knowledge about that path itself; no computational inference step is required. The tasks used are estimating or reproducing single straight-line segments or turns. Because linear and rotational movements are assessed in isolation, the underlying representation is a relatively simple one, and computation of a response route is minimized. If the pathway is to be retraced, one need specify only that the direction of travel
Encoding Spatial Representations
59
in the representation is reversed from the original. If the path is to be reproduced by moving in the same direction as the original, the computation is essentially the same as the encoding process. The verbal report is made by translating the retrieved representation into measurement terms. Given the low demands on computation, errors in reproduction and estimation should predominantly reflect information encoded into memory. In a study of Klatzky et al. (1990), the general procedure for the estimation and reproduction tasks was as follows. Subjects first explored the segment or turn without direct guidance by the experimenter, then made the called-for responses. Verbal estimation preceded the reproduction response. Straight-line segments were initially walked while holding a rope in one hand; the beginning and end of a segment were marked by stops. While walking the segment, subjects were prevented from counting footsteps by imposition of a meaningless vocalization. Reproduction was performed by walking without guidance, after having been led to the center of the workspace. The turns were made in a circular hoop surrounding the subject; the beginning and end of the turn were marked by stops. The stops were removed for the reproduction response, which again occurred within the hoop. (This is one of the few cases in which electronic measurement was not used in our studies.) Average responses as a function of physical stimulus values in the simple reproduction tasks are shown in Fig. 4 (labeled Experiment 1; Experiment 2 refers to Loomis et al., 1993, as described later). In general, reproduction of walked distances and turns could be performed with fairly high accuracy. Turns were reproduced within approximately 5" of the correct response, on average, over a range from 60 to 300". Distances within the range of 4-12 m were reproduced within about 1 m of signed error on average, constituting approximately a footstep. The functions relating responses to actual values of turn and distance were highly linear, as can be seen in Fig. 4. On average, verbal estimations of distance were, like reproduction responses, within 1 m of signed error. Estimations of turns produced somewhat greater signed error than reproduction, with the maximum of about +25" error at the 300" turn. However, the high level of reproduction accuracy may reflect the hoop's helping subjects to control and monitor their motor output. Later, we report higher error in a turn reproduction task performed without the hoop. In any case, the estimation data help to eliminate the ambiguity from the reproduction data. Recall that accuracy in a reproduction task might mask encoding that was actually highly inaccurate, because there was a compensatory misencoding during the response. Although there were differences between verbal report and reproduction responses, they were not so pronounced as to suggest that the reproduction responses should be invalidated.
- -
o- - - Experiment 1
- -
Experiment 2
4
-
-
2
y = 2.14 + 0 . 7 5 ~ y = 0.54 + 0 . 8 1 ~
_ _ _ . _ .
_ _ _ . _ _
6
4
10
8
12
Actual distance (m)
* - - Experiment --* Experiment -
-
- -
0
I 2
60
y = 3.64 + 0 . 9 9 ~ _ _ _ _ _ _ y = 32.86 + 0 . 8 2 ~ ------
120
180
240
Actual turn (deg)
300
360
Encoding Spatial Representations
61
In short, if we estimate the accuracy of encoding from the ability to reproduce either a straight-line translation or a rotation, we find that the process is fairly accurate. The psychophysical functions derived from the reproduction provide a quantitative description of encoding. The linear form of the functions indicates that, within the range of experimental values tested, there is a constant error (the intercept), and each increase in the objective stimulus value produces some proportional increase in the perceived value (that proportion is the slope). The encoding of turns appears to have very modest constant error and a slope near 1.0. The function for straight-line segments indicates more substantial error, with an intercept of 2 m and a slope of .75. In the study of Loomis et al. (1993), we incorporated similar reproduction and estimation tasks, with some methodological differences from the previous study: A sighted guide technique (the subject holds onto the upper arm of the guide) was used to move the subject through the initial distance or turn, turn reproduction was measured from a remote camera system and did not involve placing the subject within a hoop, and turns were estimated with a pointer rather than verbally. The resulting data revealed some differences from the previous study, which might reflect methodological changes or population differences (i.e., the inclusion of blind subjects and a greater age range). The data are shown in Fig. 4 (Experiment 2). It can be seen, from the increased intercept and decreased slope relative to the Klatzky et al. study (1990, Experiment l), that the turns in this experiment were less accurate than those made previously within the hoop. However, the function relating the distance response to actual distance exhibited less distortion overall than in the previous study.
SUMMARY This section provided two estimates of encoding functions for distance and turn, as derived from reproduction tasks. These estimates were supported by verbal (or pointer-setting) responses, which indicated that the relatively high level of reproduction is not due to compensatory errors in encoding and responding. Some differences in the reproduction functions, presumably due to methodology, were evident, but both sets of functions indicated that encoding was fairly accurate. We next consider whether these data Fig. 4. Performance in the task of reproducing walked distances (left panel) and turns (right panel) from two experiments. Experiment 1 is by Klatzky et al. (1990); Experiment 2 is by Loomis et al. (1993). Each figure shows the reproduced response as a function of the actual distance or turn; best-fitting linear functions are also shown. Top panel reprinted with permission of the Helen Dwight Reid Educational Foundation, copyright Heldref Publications 1990.
62
Roberta L. Klatzky et al.
describe the encoding process that occurs during the task of completing a shortcut, as compared to repoducing a single segment or turn.
VII. Encoding Distances and Turn Inferred from Fitting the Encoding-Error Model As we noted earlier, it is not clear that reproduction and estimation are the appropriate tasks to definitively assess encoding by means of path integration, because they might be based on other cues. A person who is reproducing a turn or straight-line segment has access to a motor trace, potentially including kinesthetic, cutaneous, and vestibular components. Traces of lower level efferent commands would be present and could play a role, although, as the initial movement in our task was not performed with a target destination in mind, knowledge of the intended target location presumably does not contribute to performance. Temporal coding is also possible, although footstep counting was not permitted during translation. In this section, we describe an alternative to reproduction performance as a measure of encoding. This measure is derived by fitting the encodingerror model to data from a shortcut task, in this case the task of triangle completion used by Loomis et al. (1993). Because subjects d o not simply reproduce previous responses in this task, contributions from purely motor or temporal sources should be minimized. A more direct estimate of encoding by means of path integration should therefore be provided. The model-based estimate of the underlying encoding functions is the best possible one, under the assumption that the systematic errors in shortcut responses arise from encoding. That is, the model assumes that the response is computed and executed accurately, given the potentially faulty data from the encoding process. All the systematic error in performance arises from the encoding functions that are estimated by fitting the model. Those functions describe the internalization of the initial two legs and turn, which are walked with guidance. It is assumed that the same functions do nor apply when the return path is executed; rather, execution occurs without systematic error. Triangle completion was systematically assessed with congenitally blind, adventitiously blind, and blindfolded, sighted subjects by Loomis et al. (1993). On each trial, the subjects walked a “Leg-A,’’ made a “Turn-1,’’ and then walked a “Leg-B,” after which they attempted to return directly to the origin. In returning, they first made a turn toward the origin, “Turn2,” and then walked a third leg, “Leg-C.” Subjects performed this task with a set of 27 pathways, resulting from the factorial combination of three leg-length values and three turn values.
Encoding Spatial Representations
63
Figure 5 (open circles) shows the observed stopping points, averaged over subjects, for each of the 27 paths in the triangle completion experiment. Figure 6 shows the mean value of the turn and distance responses for each group of subjects. The average turn response is shown as a function of the correct value, and likewise the average distance response is shown as a function of the correct value. Perhaps the most striking aspect of the data is the tendency for low values of the response range to lead to overresponding, and high values to lead to underresponding. As a result, responses are most accurate near the mean of the actual values. We (Fujita et al., 1993) fit the encoding-error model to the data in Fig. 6, averaging over the three subject groups (as they did not differ significantly). As was described earlier, the model first adopts an encoded
Fig. 5. End points predicted by the model (squares) and the average end point (circle) for each of 27 triangular pathways. Within each panel, the length of Leg-A and Turn-1 are held constant, and mean end points for three values of Leg-B are illustrated. The systematic departures of predictions from the data in the lower left panel reflect loss of data due to camera position. From Fujita et al. (1993). copyright Ohio State Univ. Press.
Roberta L. Klatzky et al.
64
CON
CON
y=220.01 + 0.51165xr2 = ,834
E
1200
v=19.532 + 0 . 4 5 5 6 7 ~r2 = ,854
,
/
Y 1000 600 4
400
0
g
200
v)
0 200 400 600 800 1000 1200
0
Distance (cm)
p (degrees)
ADV
y=172.80 + 0.56901xr2 = ,879
E
1200
20 40 60 80 100 120
ADV v=23.856 + 0 . 6 2 1 8 8 ~r2 =.830
400 0
vl
0 200 400 600 800 1000 1200
Distance (cm)
0
20 40 60 80 100 1 0
p (degrees)
SIG
SIG
v=213.38 + 0.63034xr2 = ,918
v=11.135 +0.56770xr2= ,932
0 200 400 600 800 1000 1200 Distance (cm)
0
20 40 60 80 100 120
p (degrees)
Fig. 6. Mean results of the triangle completion task for sighted, congenitally blind, and adventitiously blind subjects (abbreviated SIC, CON, and ADV, respectively). The left column shows the average distance walked as a function of the correct value. The right column shows the average inner angle produced, at the point of completion, as a function of the correct value. (The inner angle, & is 180" minus the response turn.) A best-fitting linear function is shown for each graph. From Loomis et al. (1YY3), copyright by the American Psychological Assn., reprinted with permission.
Encoding Spatial Representations
65
value for each of the actual values the subject experienced-in this case leg lengths of 2,4, or 6 m and turns of 60,90, and 120". Given the assumed encoded values, the model predicts a turn and distance response that the subject would make to complete the triangle (Turn-2 and Leg-C), using appropriate trigonometric formulas. As a simple example, if the actual path had subjects tracing the first two legs of a 3-4-5 right triangle, and the corresponding internalized values were 1.5 for Leg-A, 90" for Turn-1, and 2 for Leg-B, the subject would have rescaled the triangle by 50% and would respond with a Leg-C of 2.5 (rather than 5 ) and a correct value of Turn2 (assuming accurate computation of the response). In general, predicted values of Leg-C and Turn-2, given the assumptions about the encoded values of Leg-A, Leg-B, and Turn-1, can be compared to the empirically determined turn and distance responses. Because the same three values of leg length and three values of turn were used in 27 triangles, the model fit to those data had to estimate the corresponding six encoded values so as to best fit all 27 outcomes. The encoded values corresponding to the three leg lengths and three turns are shown in Fig. 7. A linear function has been fit to them, since it obviously accounts for substantial variance. Unlike the psychophysical functions in Fig. 4, which indicate empirically observed distances and turns, the ordinate on this function is an estimated model parameter, namely, the encoded equivalent of the given objective value. We will call these derived encoding functions. The goodness of fit of the model can be seen in Fig. 5. The stopping points predicted by the model (Fig. 5, filled squares) are
y = .48 x
12
1
" L 5
2
4
Stimulus length (m)
6
+ 50
60
90
120
Stimulus angle (deg)
Fig. 7. The encoding functions for turn and distance, as derived from fitting the encodingerror model to the data of Loomis et al. (1993). Each ordinate value has been derived as a separate parameter; the best-fitting linear function is also shown. From Fujita et al. (1993). copyright Ohio State Univ. Press.
Roberta L. Klatzky et al.
66
very similar to those made by the subjects. In evaluating the model, it is important to compare predicted and actual errors, rather than comparing predicted and actual raw responses, because many models could account for a general increase in response values with correct values. What is important is whether the precise deviations from the correct response are predicted by the assumed errors in encoding. The model was able to account for 93% of the variance in average signed distance error and 92% of the variance in average signed turn error; moreover, the actual error values were fit very well. It is noteworthy that the parameters of the derived encoding functions in Fig. 7, particularly the slopes, indicate more distortion than was evident in the reproduction task (Fig. 4). The slope of .48 for turn encoding is nearly half those obtained from the empirical reproduction data, and the slope of .60 for distance encoding is also substantially less than those obtained empirically. The model indicates, then, that the encoded representation that underlies reproduction is not the same as that used for pathway completion. It is the latter that we take as a more appropriate assessment of human path-integration ability.
SUMMARY Data from the triangle-completion study indicate, first, that responses are systematically related to correct values, but with systematic error present as well. Overall, subjects compressed the range of responses relative to the correct values, by overresponding when low values were called for and underresponding when high ones were called for. They performed correctly near the mean of correct values. These trends were mimicked by the underlying encoding functions extracted with the model. They showed overencoding of low values and underencoding of high ones, with accurate encoding of the mean of the experienced stimulus values. The encoding functions inferred from the path-completion task differed from those obtained from simple reproduction, suggesting that there are contributions to the reproduction task other than path integration.
VIII. Effects of Experience on Encoding Pathway Parameters The pattern in the derived encoding functions in Fig. 5 constitutes regression to the mean of experienced values. That is, subjects are very accurately encoding the mean values of the leg lengths and turns in pathways they have been exposed to. This raises the question as to whether people rapidly assimilate encoding functions to recent context, so that they accurately encode the values at the mean of recently experienced distributions. Alter-
Encoding Spatial Representations
67
natively, the functions could be immutable, not depending on experienced values. In the next study described, we examined the generality of these encoding functions, including the regression that they evidence. In particular, we asked whether the nature of the encoding functions reflects navigators' prior experience with a set of pathways. If so, then manipulating the nature of experience should change the nature of the derived encoding functions. A. TRIANGLE COMPLETION WITH DIFFERENT RANGES OF LEG LENGTH A N D TURN We manipulated the subjects' experience by having them complete a set of pathways with a particular set of leg lengths or turns (Klatzky, Beall, Loomis, Golledge, & Philbeck, 1997). The values in the set of experienced pathways were drawn either from the low region or from the high region of a range of leg lengths and turns. In all cases, subjects completed a triangle after having been led through Leg-A, Turn-1, and Leg-B. To manipulate the region of leg lengths, Leg-A values within a block of trials were all either relatively low (1-3 m) or high (4-6 m). To manipulate the region of turns, Turn-1 values were all at either relatively low (10-70") or high (110-170") levels. When leg-length region was manipulated, the turn angles were held constant at intermediate values, and correspondingly, when turn region was manipulated, the leg lengths were held constant at intermediate values. Region of leg-length values and region of angle were manipulated within subjects. To segregate the subject's experience with a particular region of sampled values, so as to encourage assimilation to local experience, all the trials with that region were presented in a block. There was a distinct break between blocks. Thus, there were four temporally segregated blocks of stimuli, having low turn values and moderate leg lengths, high turn values and moderate leg lengths, low leg-length values and moderate turns, and high leg-length values and moderate turns. How should manipulating the region from which stimulus values were drawn effect the encoding functions? One possibility is that the manipulation would have no effect. This would occur, in particular, if the derived encoding functions from the earlier triangle-completion function were immutable. It is interesting in this regard that both the turn and distance encoding functions in Fig. 7 seem to reflect experience, in that they are centered around the mean of the values that were used in the corresponding experiment. That is, there is highly accurate encoding at the midpoint of stimulus values used in the experiment, with lower values being overestimated and higher ones being underestimated. Thus, there is regression to the mean of the experienced values.
68
Roberta L. Klatzky et al.
Consider what would happen if the same encoding functions were applied to the subsequent experiment in which region of experienced values was manipulated. For distance (i.e., leg length) encoding, the point of accurate encoding in Fig. 7 is 4 m. If the encoding function is invariant, then subjects exposed to pathways that all have leg lengths less than 4 m should still use an encoding function that is most accurate at 4 m and that overestimates lower values. Thus, they should show overencoding at all values in the new experiment. Conversely, those exposed to pathways having leg lengths in the range of 4-6 m, but using the encoding function shown in Fig. 7, should accurately encode the lowest value and underestimate the higher ones in the new experiment. A similar argument would hold for turns. Alternatively, it is possible that adaptation to experienced values, rather than invariance, is characteristic of the encoding process. Adaptation to experience is observed quite generally in perception of stimulus magnitude (e.g., Helson, 1948). In particular, if encoding functions regress to the mean of the experienced values, different encoding functions should be derived from experience with the low and high regions of a path parameter. With respect to distance, for example, exposure to the short legs should lead to accurate encoding of an approximately 2-m length, and exposure to long legs should lead to accurate encoding of about a 5-m length. To determine whether invariance or regression to the mean of experienced values occurs (if either), it is necessary to derive encoding functions by fitting the model. As an alternative to assimilation to local experience, there is another potential reason why encoding functions might differ over low and high levels of turn or distance. The differences might reflect nonlinearities in an underlying, invariant encoding function. It is true that the encoding functions fit to the triangle-completion data were highly linear for both distance and turn. We have noted previously, however (Fujita et al., 1993), that these functions are unlikely to be linear over their entire range. It seems unlikely that, for example, people who made no turn whatsoever at a stopping point would nevertheless estimate having turned 44", the zerointercept of the derived encoding function for turn from the trianglecompletion task in Fig. 7. Presumably, turns of 0 and 180"would be encoded relatively accurately (Klatzky et al., 1990; Sadalla & Montello, 1989). This means that the linear function previously observed would be embedded in a function grounded at (0,O) and (180,180), forming a sigmoid shape. In that case, looking at the encoding of turn values near 0 would reveal a higher slope than looking at the encoding of turn values in a midrange. A similar argument can be made for distance encoding, although the abscissa of that function is unbounded at the positive end. Examination of the data from the experiment gave, at the outset, little reason to think that subjects were adapting their responses separately to
Encoding Spatial Representations
69
each region of distance and turn in the experienced stimuli. In all conditions, a single linear function provided a good fit when the responses were plotted against the actual values. Figure 8 shows a representative function, depicting the average distance response as a function of the correct value. The observations are drawn from two sets of paths, having different regions of Turn1 values. Given the geometric constraints, the lower values of response distance correspond to paths having high values of Turn-1, and the higher values of response distance correspond to paths with low Turn-1 values. What is most important is that a single function can readily be fit across the two regions ( r 2 = .96). The hypothesis that subjects would adopt distinct encoding functions, each regressing to the mean of experienced values within a region, predicts two distinct response functions. Each of those functions would cross the diagonal at approximately the midpoint of correct values within a region. But the data do not indicate two obviously discrete functions, one for each region. B. FIITING THE ENCODING-ERROR MODEL We fit the encoding-error model to the data to further test whether encoding differed according to the region of stimulus values. The process was done independently for the two manipulated parameters, turns and leg lengths. Within each of these conditions, two versions of the model were tested, one fit across both regions of the manipulated stimulus parameters and another within each region. The latter doubled the number of parameters
0
1
2
3
4
5
6
7
Correct distance (m) Fig. 8. Average distance response for subjects completing a triangle, as a function of correct value. Each point represents the observed average for one path. Plotted observations come from two sets of paths, having high or low values of the turn between the initial two legs. The least-squares linear function across observations from both sets is shown.
Roberta L. Klatzky et al.
70
but did not substantially improve the fit of the model, supporting the conclusion that subjects did not adapt responses within each region separately. This same result also does not support the idea that the different regions would be encoded differently due to nonlinearities in the underlying encoding function. We henceforth concentrate on the functions fit to the data combined across regions. The encoding functions (fit with the same measure of model error used for the earlier triangle completion study') are as follows (d = distance and t = turn; primes indicate encoded values and unprimed letters indicate actual values): distance encoding, leg-length manipulation: d ' = .52d + .66m turn encoding, leg-length manipulation: t' = .82t + 0.8" distance encoding, turn manipulation: d ' = .60d + .55 m turn encoding, turn manipulation: t' = .86t + 0.1" The slope of the distance encoding functions is fairly close to the .60 slope obtained in the previous study (see Fig. 7), but the intercept is substantially less than the 1.2-m intercept obtained for distance in that study. The turn functions are quite different from the one obtained previously, which had a slope of .48 and an intercept of 50". It appears, then, that although subjects did not form distinct encoding functions for low and high regions of a manipulated parameter in the stimulus set, the encoding functions obtained in this experiment differed to a considerable extent from those obtained previously. In other words, although encoding does not appear to have been assimilated to the local set of stimuli, it was also not invariant across experiments. The present encoding functions are qualitatively similar to those obtained earlier, in that the slopes tend to be less than 1.0 and the intercepts tend to be positive (although the turn intercepts are near 0). However, they do not regress to the mean of encoded values. The values that would be encoded without error, given the average of the two parameter estimates for distance and turn, are a distance of 1.4 m and a turn of 2.5". These are substantially lower than the mean of the leg lengths and turns as computed across the stimuli presented to the subjects.
'
We have also considered a second measure of model error that uses the turn and distance responses rather than the stopping points. That model error is based on the difference between observed and predicted turns and distances (appropriately normalized to account for variance differences in these measures, and again summed over paths). We have found that predictions derived from the two measures of model error are highly correlated, although the derived encoding functions do differ somewhat.
Encoding Spatial Representations
71
C . SUMMARY The data described in this section suggest that encoding a representation to be used for computing the homing vector is a process that is context sensitive, but not extremely labile. The study did not demonstrate sensitivity to highly local context, as would be evidenced by the formation of a distinct encoding function for each block of trials, corresponding to a distinct region of leg length or turn. Nor did encoding demonstrate immutability, since the functions obtained were different from those previously fit to the data of Loomis et al. (1993).
IX. When Does Updating of the Homing Vector Occur? Evidence against Moment-To-Moment Updating Performance in the path completion task indicates that people can take shortcuts after following pathways without vision, at levels well above chance. This indicates that their representation of the path contains at least the minimal representation needed to return to an origin, the homing vector. We now raise two related issues about the computation of the homing vector. One is whether it is updated frequently, on the order of a footstep (moment to moment), or intermittently. The issue of when the homing vector is updated has implications for a second issue: how much knowledge people have of the pathway they followed. If the homing vector is computed only intermittently, then people must retain more than the homing vector during at least the interval between updates. However, once the homing vector is computed, they might retain only its values. Alternatively, people might retain a fuller history of travel, as well as computing a homing vector. Relevant to these issues are two types of data. One concerns the effects of the configural properties of completed paths on shortcut performance; the other concerns performance on a different task, that of retracing the path. Configural effects are available from the study of Klatzky et al. (1990). In that study, we conducted a path-completion task with multisegment paths, including some that were more complex than a triangle. Subjects walked paths of one, two, or three segments, then were asked to return to the origin. The same configurations were completed at two scales, implemented by having the turning points fall on the circumference of a circle having either a 3-m or 10-m diameter. The one-segment condition of the pathway-completion task is the same as the simple distance reproduction task (except that subjects must make
72
Roberta L. Klatzky et al.
a 180" turn before responding). As would be expected, then, the level of error in the one-segment condition was comparable to that of the simple reproduction task. Error increased substantially with the number of segments walked before returning to the origin, particularly with the largescale pathways. For example, signed turn errors at the large scale increased by approximately 50% as the number of segments in the walked pathway increased from one to three. Signed distance errors changed in direction with an increasing number of segments, from overestimation by approximately 60 cm to underestimation by about the same amount. The increase in error need not indicate, however, that subjects retain a history of travel. As more segments are followed, there are more operations of encoding and computing the homing vector, and error could accumulate. More critical to the issues raised above are effects of the configuration formed by the walked segments. These effects suggest that the particular history of travel is important, not just the number of segments. Figure 9 shows two triangular pathways, together with the trajectories produced by
Fig. 9. Three paths used in the path-completion task of Klatzky et al. (1990) (top row), and the trajectories of the individual subjects' responses for each path (bottom row). In each case, the subject is trying to reach the point 0 from the end point of the path on the circle's circumference. Dark points and bands within the top circles represent the centroid of subjects' responses and 1 standard error in distance and heading. From Klatzky et al. (1990). Reprinted with permission of the Helen Dwight Reid Educational Foundation, copyright Heldref Publications, 1990.
Encoding Spatial Representations
13
the individual subjects in response to the instruction to turn and walk directly to the origin. Pathway 8 produced particularly low turn error, whereas Pathway 7 produced fairly consistent overturning. Another pathway of three segments, in which the third segment crossed over the first, produced highly variable turns. This is Pathway 11 in the figure. Apparently, subjects became confused about the crossover, some audibly complaining as they approached that they no longer knew whether the origin was in front of them or behind. If the homing vector is updated moment to moment, then subjects should not take into account a point of crossover, because the point of intersection was first processed at a moment relatively far back in time. Only if a history of travel is retained and enters into responsecomputation processes should configuration become important. Another important finding pertains to the response latency-the time that the subject paused between the end of the pathway and initiating the pathway-completion response. We examined how response latency changed as a function of the number of segments in the pathway. Considering the data from Klatzky et al. (1990), and averaging over scale (this analysis is reported in Loomis et al., 1993), the latencies were 2.7 s, 2.8 s, and 3.1 s for configurations of one, two, or three segments. The configuration with the crossover was particularly slow, with a latency of 3.3 s. The difference in path-completion latency between paths having two and three legs was replicated by Loomis et al. (1993). In contrast to the effect of number of segments on the latency to initiate a shortcut, Loomis and colleagues further found that path length did not affect the time to initiate a retrace of the outbound path. Why should this difference between the shortcut and retrace tasks arise? A critical difference between tasks lies in the need to compute the homing vector for the shortcut task, but not the retrace. The increase in latency to return home by a shortcut suggests that subjects performed a more complex computation of the homing vector at the end of longer or more complex pathways, that they were more uncertain about the correct response at the end of longer or more complex pathways, or possibly both. However, it appears that the retrieval of the pathway in order to retrace it was no more complex for the longer pathways than for the shorter ones. And because the time to prepare for a retrace did not include the time to update the homing vector, the latency to initiate retracing was unaffected by pathway length. The response latency data clearly argue against moment-to-moment updating of the homing vector. If that were the case, there should be no effect of pathway complexity on latency, because the processing required at the point of initiating the return would be invariant over the complexity and length of the path. At the terminal point, subjects would read out the vector and execute it, regardless of the history of travel. We conclude that
Roberta L. Klatzky et al.
74
completion of complex paths does not rely on a homing vector that is updated from footstep to footstep. Such a vector, we suggest, is computed at the point of response, when the shortcut is called for. If the homing vector is not updated on a moment-to-moment basis, then navigators must have more than this simple representation available. At least, they must retain the history of travel in between updates. In fact, data from the retrace task of Loomis et al. (1993) indicate that people know a substantial amount about their history of travel. In that study, we asked subjects to reproduce the path they had followed on some trials and to return to the origin on others. Subjects were not informed in advance which task would be called for. Representative results are shown in Fig. 10, which gives the initial path and the retrace for a sample of sighted, congenitally blind, and adventitiously blind subjects. The subjects who retraced some segments in the wrong direction (conl) had blindness originating in retrolental fibroplasia (RLF), which has been suggested to carry with it more general spatial deficits. Our RLF subjects were not uniformly so impaired, however, and this claim has been called into question from other sources as well (Dodds, Hellawell, & Lee, 1991).
SUMMARY The latency increases with pathway length and complexity, together with configural effects on error, argue against moment-to-moment updating and indicate some retention of the history of travel. The ability to reproduce a complex path on demand further indicates that the history of a multisegment path can be retained quite well.
X. Group and Individual Differences in Navigation without Vision Much of our work has compared blind and sighted populations in navigation tasks. To date, we have found no obvious differences among these groups. As Fig. 6 indicates, the triangle-completion performances of sighted, adventitiously blind, and congenitally blind groups were indistinguishable in our study (Loomis et al., 1993). Moreover, we found some blind subjects (even those where RLF is the cause of blindness) to be among our best performers, and we found some sighted subjects to be among our worst. There were some indications of group differences in a discriminant analysis performed on the data from Loomis et al. (1993; see Klatzky, Golledge, Loomis, Cicinelli, & Pellegrino, 1995). This analysis incorporated measures from three sets of tasks. In addition to the simple and complex navigation tasks we have described, there were several using geometric stimuli on the
Encoding Spatial Representations
.'&!
Y
75
=I 8
i W
?
P
m
8
x o
X
d
C
8
-8 C
-
?I .-
I > I
Fig. 10. Examples of paths to be retraced (lefr column) and performance of individual subjects on three retracing trials (abbreviated as in Fig. 6). The start of the outbound path (home position) is shown with an X; the origin of the retrace (i.e., the terminus of the outbound path) is shown with a solid circle; and the turning points on the outbound path are shown with open circles. Copyright 1993 by the American Psychological Assn., reprinted with permission from Loomis et al., 1993.
76
Roberta L. Klatzky et al.
scale of a tabletop, including assembly of whole figures from parts, mental rotation, and estimation of the third leg of a triangle after feeling two legs. The discriminant analysis was able to classify subjects according to visual status (sighted, adventitiously blind, congenitally blind) at levels well above chance. We do not suggest, moreover, that there are unlikely to be differences between blind and sighted subjects in navigation, especially over a broader set of tasks. In one relevant study, Rieser, Hill, Talor, Bradfield, and Rosen (1992) reported that subjects with visual field loss of early onset showed deficits in the precision of spatial knowledge about their local communities. Subjects with diminished acuity but normal fields, even those with early visual loss, had spatial precision levels similar to those of fully sighted subjects. The authors suggest that experience with a broad visual field, albeit with reduced acuity, facilitates the development of precise spatial knowledge during locomotion. In another study, Rieser et al. (1986) had subjects learn about an array of objects by traveling to each one from a home location, then point to each object from the location of another object. In some cases, the subject actually moved to the pointing location, and in other cases, he or she imagined being there. The authors found that sighted subjects had more difficulty after imagined movement than after physical relocation, indicating they were better able to update during locomotion than through imagination alone. However, congenitally blind subjects performed almost equally poorly in both conditions, suggesting that they did not automatically update when they moved. We (Loomis et al., 1993) attempted to replicate this experiment and found overall trends consistent with those reported by Rieser et al. (1986). However, there was considerable variability among our subjects, with some congenitally blind subjects performing with high accuracy in both the imagination and locomotion conditions. As a result, group differences were not significant. This may reflect our selection of a generally mobile blind population, which could be more similar to sighted subjects than an entirely random sample. Haber, Haber, Levin, and Hollyfield (1993) found similar equivalence of a mobile blind population to sighted subjects, in a task requiring distance estimation. In any case, our results indicate that some blind subjects can demonstrate considerable navigation ability, including the ability to update self-position in a spatial representation under locomotion (see Thinus-Blanc & Gaunet, 1997, for a broader review of spatial capabilities in the blind). Turning to differences among individuals, an interesting implication of the encoding-error model is its suggestion that individual differences in path completion must reflect differences in people’s encoding functions. The model allows for no systematic error, and hence no basis for systematic
Encoding Spatial Representations
77
variation across individuals, in the processes of computing the homeward trajectory and executing the computed response. To assess individual differences in triangle completion, we fit the encoding-error model to data from 12 individual subjects in the Loomis et al. (1993) triangle-completion experiment (reported in Fujita et al., 1993). We deliberately chose subjects so as to sample a range of performance in the task. Accordingly, there was substantial variation in the parameter estimates across subjects. Slopes of the derived encoding function for turns ranged from 0 to 1.08, and slopes of the derived encoding function for distance ranged from .29 to 1.0. Subjects tended overall, however, to produce slopes less than 1.0 and positive intercept values. Slopes less than 1.0 indicate that low stimulus values were overencoded, and high stimulus values were underencoded, within the constraint of constant error expressed by the intercept. In a factor analysis (Klatzky et al., 1995),we used intersubject correlations to address whether individuals varied in a stable way over tasks. We included the same measures used in the discriminant analysis described above, which were derived from simple navigation, complex navigation, and tabletop tasks. Three underlying factors emerged. The first factor had high loadings for measures of performance in both retracing and completing complex paths. The second factor had loadings from tasks scaled to the tabletop, and the thrid incorporated measures from simple distance and turn reproduction. Thus, complex navigation, simple navigation, and tablescale spatial processing were segregated. These results suggest that individuals vary systematically within classes of tasks, but that performance on one class does not predict performance on another.
XI. General Summary and Conclusions A. How WELLCANPEOPLE PERFORM PATHINTEGRATION WITHOUT VISION? To summarize our principal results, people can do a number of navigational tasks quite well in the absence of visual input. They can reproduce linear segments and turns. Their verbal and symbolic estimations of these same trajectories are also quite accurate, indicating that the reproduction performance is not an artifact of compensatory errors in encoding and responding. People can also reproduce multisegment pathways with substantial accuracy. When asked to complete a path using a shortcut that they have not previously explored, people respond with systematic error, but certainly
78
Roberta L. Klatzky et al.
nonrandomly. There is a general tendency to underrespond when a large value of return distance or turn is required, and to overrespond when a small one is required. In much of our data, people perform most accurately when the required response is near the mean of required responses, suggesting assimilation to stimulus context. Path-completion performance tends to deteriorate as paths become more complex, in terms of segments and configural properties. As we have seen (Fig. 9, Pathway l l ) , it is possible for navigators to appear “completely lost” after following only three linear segments.
B. Do HUMANS PERFORM PATHCOMPLETION BY PATHINTEGRATION? Our experiments set up a context in which navigation by distal landmarks is not possible, and path integration is required. That people can succeed to some extent in completing complex paths provides an argument in favor of path integration as a basis for human navigation without vision. The encoding-error model provides a specific description of how this process occurs. A direct comparison of this model to models of path integration in lower animals suggests potential differences, however. First, we note that the model specifies that the updating of a homing vector does not occur moment to moment, but rather in terms of entire segments and turns. Models of animal navigation, in contrast, have often assumed footstep-by-footstep updating (e.g., Miiller & Wehner, 1988), although Benhamou, SauvC, and Bovet (1990) provide a model where updating occurs at the end of segments and turns. Under footstep-by-footstep updating, the encoding and computational processes are assumed to be combined and performed nearly continuously. That the model we have fit to human data assumes encoding at the end of a segment or turn is apparent, when we consider that the function relating encoded values to actual values can be fit with a substantial positive intercept (see Fig. 7). The intercept is applied once per segment or turn, regardless of the magnitude involved, and hence regardless of the number of footsteps used to produce the translation or rotation. Were the intercept to be added at intervals within a segment, the function would no longer be linear. Were the model to have a zero intercept, it would indicate that the encoded value was simply proportional to the actual value. Subjects would consistently over- or underrespond relative to correct response values, rather than exhibiting the overresponding to low values and underresponding to high ones that characterizes much of our data (see Fig. 6). If the encoding functions were fit with a zero intercept, it would be impossible to discriminate between footstep-by-footstep updating and segment-by-
Encoding Spatial Representations
79
segment (or turn-by-turn) updating, because the proportional multiplier could be applied at any point. But the positive intercept dictates application of the intercept, and hence the encoding function, at the end of a segment or turn. Encoding at the level of whole segments and turns indicates that humans are sensitive to distinctive events in travel, particularly to changes in the travel trajectory. That is, it appears that people mark the terminus of a segment by the fact that they stop translating and begin turning. Similarly, they encode turns at their termini, when the shift to linear translation begins. It is at such distinctive points, according to the encoding-error model, that the encoding function is applied to produce the internalized value. Under the assumption of segment-by-segment updating, a navigator who must complete a triangle would compute a homing vector at the end of the second leg. In the case of a pathway of three or more legs, the homing vector might be computed incrementally (e.g., once two legs are traversed, updating the homing vector after each new turn and leg) or entirely at the point of the final response. Arguing against incremental updating is the increase in response latency with more complex paths. If navigators updated the homing vector following each new turn and walked leg, then at the point of response they would be computing only the effect of the final turn and leg, so there would be no reason for the computation to take longer after three legs than after two. It is also possible, however, that the complexity effect on latency comes from sources other than computation of the final trajectory, such as greater uncertainty about the response trajectory after more complex paths or an increase in memory load. In short, our data suggest that people update the homing vector in our task only intermittently, and possibly only at the point of response. This is somewhat surprising, given data from other tasks indicating that people continually update their positions in space relative to other objects as they move, even without vision. It may be that the featureless environment in which people travel in our path-completion tasks invokes a delay in updating. Or, it may be that regardless of the richness of the environment, computing a homing vector is a process that is quite different from relatively automatic updating of the egocentric coordinates of objects, which appears to occur as one locomotes with or without vision. Finally, we note two other potential differences between human path completion and that of some other organisms. One is knowledge of the history of travel. People’s knowledge of the path is most clearly indicated by the ability to reproduce the path when called upon to do so, rather than to complete it. Secondly, human subjects appear to have sources of error that are not captured in the specific version of path integration that is
Roberta L. Klatzky et al.
80
described by the encoding-error model.* These sources of error come from processing the pathway as a configuration, for example, in detecting a crossover point. We do not know of animal data indicating such configural influences. C. AREENCODING PROCESSES COMMON ACROSSCONTEXTS? We have examined encoding in several contexts, including reproduction of straight-line segments and turns and completing triangles with various ranges of segment lengths and turn values. The encoding functions obtained across these contexts appeared to have certain characteristics in common. They were linear across the range tested. They tended to have nonnegative intercepts and slopes less than or equal to 1.0. As a result, these functions producd inaccuracies in two ways. One is the constant error introduced by the intercept; the other is the systematic underencoding of distance and turn that is introduced by the slope. Beyond the general form of these functions, however, there was considerable variability. Encoding was more accurate, in general, as estimated directly from reproduction tasks than as inferred from fitting the encodingerror model. We attribute the greater accuracy for reproduction tasks to their being based on a readout of motor traces and not purely on path integration processes. It is noteworthy that we did not obtain significant intersubject correlations with respect to performance on reproduction and shortcut tasks (Klatzky et al., 1990,1995). This further supports the idea that the tasks differ in the processes contributing to the encoded representation. The functions inferred from the encoding-error model varied substantially between studies, suggesting that the pathways that were experienced affected the internalized values corresponding to actual leg lengths and turns. However, the assimilation of encoding to experience appears not to be highly local, since a direct manipulation of local experience, by blocking pathways according to whether they had low or high values of distances and turns, failed to produce a regression of the encoding function to the mean of experienced values within a block. We conclude that although much of the systematic error in path completion can be attributed to errors in encoding, there is not a fixed encoding function that transcends tasks and stimulus contexts.
D. CONCLUSION We have found that to varying degrees, humans have the ability to navigate within space without vision, as well as without landmarks or azimuthal cues
’
In Fujita et al. (1990). we reported fitting the encoding-error model to the data of Klatzky et al. (1990), including the complex pathways. The fits were not as good as those to the Loomis et al. (1993) data, although the addition of one parameter, representing an overturn bias, improved the situation.
Encoding Spatial Representations
81
from other sensory modalities. At the heart of nonvisual navigation are the abilities to encode information about space while traveling, to form an enduring representation of that information, and to add additional information to the representation through spatial inference. The resulting representation can incorporate the layout of landmarks in space, traversed and traversable pathways, and the individual’s location and orientation within the space at a given time. Moreover, it appears that multiple representations can be formed and modified, allowing people to execute a variety of navigational skills. They can verbally or manually report about or reproduce prior pathways. They can orient toward objects from their current perspective or an imagined perspective, and they can take shortcuts that they have not directly explored on foot. The accuracy of performance on these tasks varies with the individual, the task, and the stimuli employed, but the human ability to perform them is considerable. At the same time, human path integration exhibits a substantial level of systematic error, much of which can be attributed to errors in encoding a representation of space through nonvisual locomotion.
ACKNOWLEDGMENTS The research reported here was supported by Grants 7022 and 9740 from the National Eye Institute. We thank our many colleagues in this research, including Andrew Beall, Joseph Cicinelli, Sally Doherty, Phyllis Frye, Naofumi Fujita, James Pellegrino, and John Philbeck.
REFERENCES Banks, W. P., & Hill, D. K. (1974). The apparent magnitude of number scaled by random production [Monograph]. Journal of Experimental Psychology, 102, 353-376. Benhamou, S., SauvC, J. P., & Bovet, P. (1990). Spatial memory in large-scale movements: Efficiency and limitation of the egocentric coding process. Journal of Theoretical Eiology, 145, 1-12. Dodds, A. G., Hellawell, D. J., & Lee, M. D. (1991). Congenitally blind children with and without retrolental fibroplasia: Do they perform differently? Journal of Visualhpairment and Blindness, 85, 225-227. Dodds, A. G., Howarth, C. I., &Carter, D. C. (1983). Memory for movement in blind children: The role of previous visual experience. Journal of Motor Behavior, 15, 343-352. Easton, R. D., & Sholl, M. J. (1995). Object-array structure, frames of reference, and retrieval of spatial knowledge. Journal of Experimental Psychology: Learning, Memory, and Cognition, 21, 483-500. Etienne, A. S., Maurer, R., & Saucy, F. (1988). Limitations in the assessment of path dependent information. Behaviour, 106, 81-111. Etienne, A. S., Maurer, R., & SCguinot, V. (1996). Path integration in mammals and its interaction with visual landmarks. Journal of Experimental Biology, 199, 201-209. Evans, G. W., & Pedzek, K. (1980). Cognitive mapping: Knowledge of real-world distance and location information. Journal of Experimental Psychology: Human Learning and Memory, 6, 13-24.
82
Roberta L. Klatzky et al.
Farrell, M. J., & Robertson, 1. H. (1996). Mental rotation and the automatic updating of hodycentered spatial relationships. Unpublished manuscript. Fujita, N., Klatzky, R., Loomis, J. M., & Golledge, R. G. (1993). The encoding-error model of pathway completion without vision. Geographical Analysis, 25, 295-314. Fujita, N., Loomis, J. M., Klatzky, R. L., & Golledge, R. G. (1990). A minimal model for deadreckoning navigation: Updating the homing vector. Geographical Analysis, 22, 326-335. Gallistel, C. R. (1990). The organization of learning. Cambridge, MA: MIT Press. Haber, R. N., Haber, L. R., Levin, C. A., & Hollyfield, R. (1993). Properties of spatial representations: Data from sighted and blind subjects. Percepfion Ce Psychophysics, 54. 1-13. Hart, R. A,, & Moore, G. T. (1973). The development of spatial cognition: A review. In R. Downs & D. Stea (Eds.), Image and environment (pp. 246-288). New York: Aldine. Helson, H. (1948). Adaptation-level a5 a basis for a quantitative theory of frames of reference. Psychological Review. 55, 297-313. Hoffmann, G. (1984). Orientation behaviour of the Desert Woodlouse Hemilepistus Reaumuri: Adaptations to ecological and physiological problems. Symposia of the Zoological Society of London, 53, 405-422. Juurmaa, J., & Suonio, K. (1975). The role of audition and motion in the spatial orientation of the blind and sighted. Scandinavian Journal of Psychology, 16, 209-216. Kelso, J. A. S. (1977). Motor control mechanisms underlying human movement reproduction. Journal of Experimental Psychology: Human Perception and Performance, 3, 529-543. Kelso, J. A. S., & Holt, K. G. (1980). Exploring a vibratory systems analysis of human movement production. Journal of Neurophysiology. 435, 1183-1 196. Klatzky, R. L., Beall, A. C., Loomis, J. M., Golledge, R. G., & Philbeck, J. W. (1997). Ifuniart navigation ability: Tests of the encoding-error model of path integration. Unpublished manuscript. Klatzky, R. L., Golledge, R. G., Loomis, J. M., Cicinelli, J. G., & Pellegrino, J. W. (199.5, January-February). Performance of blind and sighted persons o n spatial tasks. Journal of Visual Impairment & Blindness, pp. 70-82. Klatzky, R. L., Loomis, J. M., Golledge. R. G., Cicinelli, J. G., Doherty. S.. & Pellegrino. J. W. (1990). Acquisition of route and survey knowledge in the absence of vision. Journal of Motor Behavior, 22, 19-43. Landau, B., Spelke. E., & Gleitman, H. (1984). Spatial knowledge in a young blind child. Cognition, 16, 225-260. Loomis, J. M., Da Silva, J. A., Fujita, N., & Fukusima. S. S. (1992). Visual space perception and visually directed action. Journal of Experimental Psychology: Human Perception and Performance, I8, 906-922. Loomis, J. M., Golledge, R. G., & Klatzky, R. L. (1996). GPS for the visually impaired. In Proceedings of the 1996 Trimble Surveying & Mapping Users Conference (pp. 5-9). San Jose, CA: Trimble Navigation. Loomis, J. M., Hebert, C., & Cicinelli, J. G. (1990). Active localization of virtual sounds. Journal of the Acoustical Society of America, 88, 1757-1764. Loomis, J. M., Klatzky, R. L., Golledge, R. G., Cicinelli, J. G., Pellegrino. J. W., & Fry, P. (1993). Nonvisual navigation by blind and sighted: Assessment of path integration ability. Journal of Experimental Psychology: General, 122, 73-91. Lynch, K. (1960). The image ofthe city. Cambridge, MA: MIT Press. Maurer, R., & Seguinot, V. (1995). What is modelling for? A critical review of the models of path integration. Journal of Theoretical Biology, 175, 457-475. May, M. (1996). Cognitive and embodied modes of spatial imagery. Psychologische Beitraege, 38, 418-434.
Encoding Spatial Representations
83
Mittelstaedt, H. (1985). Analytical cybernetics of spider navigation. In F. G . Barth (Ed.), Neiirobiology of arachnids (pp. 298-316). Berlin: Springer-Verlag. Mittelstaedt, H., & Mittelstaedt, M. L. (1973). Mechanismen der Orientierung ohne richtende Aussenreize. Fortschritte der Zoologie, 21, 46-58. Mittelstaedt, H., & Mittelstaedt, M. (1982). Homing by path integration. In F. Papi and H. G. Wallraff (Eds.), Avian navigation (pp. 290-297). Berlin: Springer-Verlag. Mittelstaedt. M.. & Glasauer, S. (1991). Idiothetic navigation in gerbils and humans. Zoologische Jahrbucher, Abteilungfur Algemeine Zoologie und Physiologie der Tiere, 95,427-435. Muller, M., & Wehner, R. (1988). Path integration in desert ants, Cataglyphis fortis. Proceedings of the National Academy of Sciences of the U.S.A., 85, 5287-5290. Palmer, S. E. (1978). Fundamental aspects of cognitive representation. In E. Rosch & B. B. Lloyd (Eds.), Cognition and categorization (pp. 259-303). Mahwah, NJ: Erlbaum. Piaget, J., & Inhelder, B. (1967). The child’s conception of space. New York: W. W. Norton. Presson. C. C. (1987).The development of spatial cognition: Secondary uses of spatial information. In N. Eisenberg (Ed.), Contemporary topics in developmentalpsychology (pp. 7-1 12). New York: Wiley. Presson, C. C., & Hazelrigg, M. D. (1984). Building spatial representations through primary and secondary learning. Journal of Experimental Psychology: Learning, Memory, and Cognition, 10, 723-732. Presson, C. C., & Montello, D. R. (1994). Updating after rotational and translational body movements: Coordinate structure of perspective space. Perception, 23, 147-1455, Rieser, J. J. (1989). Access to knowledge of spatial structure at novel points of observation. Journalof Experimental Psychology: Learning, Memory, and Cognition, I5(6), 1157-1165. Rieser, J. J., Guth, D. A., & Hill, E. W. (1982). Mental processes mediating independent travel: Implications for orientation and mobility. Journal of Visual Impairment and Blindness, 76, 213-218. Rieser, J. J., Guth, D. A,, & Hill, E. W. (1986). Sensitivity to perspective structure while walking without vision. Perception, 15, 173-188. Rieser, J. J., Hill, E. W., Talor, C. R., Bradfield, A., & Rosen, S. (1992). Visual experience, visual field size, and the development of nonvisual sensitivity to the spatial structure of outdoor neighborhoods explored by walking. Journal of Experimental Psychology: General, 121(2). 210-221. Rieser, J. J., & Rider, E. A. (1991). Young children’s spatial orientation with respect to multiple targets when walking without vision. Developmental Psychology, 27, 97-107. Sadalla, E. K., & Montello, D. R. (1989). Remembering changes in direction. Environment and Behavior, 21, 346-363. Sauve, J. P. (1989). L ‘orientationspatiale: Formalisation d’un rnodde de memorisation kgocentree et experimentation chez I’homme. Doctoral thesis, Universitt d’Aix-Marseille 11. SCgunoit, V., Maurer, R., & Etienne, A. (1993). Dead reckoning in a small mammal: The evaluation of distance. Journal of Comparative Physiology A, 173, 103-113. Sholl, M. J. (1987). Cognitive maps as orienting schemata. Journalof Experimental Psychology: Learning, Memory, and Cognition. 13, 615-628. Sholl, M. J. (1989). The relation between horizontally and rod-and-frame and vestibular navigational performance. Journal of Experimental Psychology: Learning, Memory, and Cognition, 15, 110-125. Sholl. M. J. (1996). From visual information to cognitive maps. In J . Portugali (Ed.), The construction of cognitive maps (pp. 157-186). The Hague: Kluwer. Siegel, A. W., & White, S. H. (1975). The development of spatial representations of largescale environments. In H. W. Reese (Ed.), Advances in child development and behavior (Vol. 10, pp. 9-55). New York: Academic Press.
84
Roberta L. Klatzky et at.
Stelmach, G. E., Kelso, J. A. S., & McCullagh, P. D. (1976). Preselection and response biasing in short-term motor memory. Memory & Cognition, 4, 62-66. Thinus-Blanc, C.. & Gaunet, F. (1997). Representation of space in blind persons: Vision as a spatial sense? Psychological Bulletin, 121, 20-42. Thorndyke, P. W., & Hayes-Roth, B. (1982). Differences in spatial knowledge acquired from maps and navigation. Cognifive Psychology, 14, 560-589. Tolman, E. C. (1948). Cognitive maps in rats and men. Psychological Review, 55, 189-208. von Saint Paul, U. (1982). Do geese use path integration for walking home? In F. Papi & H. G. Wallraff (Eds.), Avian navigation (pp. 298-307). Berlin: Springer-Verlag. Walcott, C. (1996). Pigeon homing: Observations, experiments and confusions. Journal of Experimental Biology, 199, 21-27. Wehner, R., & Wehner, S. (1986). Path integration in desert ants. Approaching a long-standing puzzle in insect navigation. Monitore Zoologic0 ffaliano, 20, 309-331. Worchel, P. (1951). Space perception and orientation in the blind. Psychological Monographs, 65, 1-28. Worchel, P. (1952). The role of the vestibular organs in space orientation. Journal of Experimental Psychology, 44, 4010. Yamamoto, T. (1991). A longitudinal study of the development of spatial problem solving ability in the early blind. Japanese Journal of Psychology, 61, 413-417.
ATTENTUATING INTERFERENCE DURING COMPREHENSION The Role of Suppression Morton Ann Gernsbacher
I. Introduction The goal of my research is to identify the cognitive processes and mechanisms that underlie language comprehension and comprehension in general. I have identified a few of those processes and mechanisms in a framework I call the Structure Building Framework (Gernsbacher, l990,1991a, 1995). According to the Structure Building Framework, the goal of comprehension is to build coherent mental representations or structures. These structures represent clauses, sentences, paragraphs, passages, and other meaningful units. To build these structures, first, comprehenders lay foundations for their mental structures (Carreiras, Gernsbacher, & Villa, 1995; Gernsbacher & Hargreaves, 1988,1992; Gernsbacher, Hargreaves, & Beeman, 1989). Then comprehenders develop their mental structures by mapping on information, when that incoming information coheres or relates to the previous information (Deaton & Gernsbacher, in press; Gernsbacher, 1996; Gernsbacher & Givdn, 1995; Gernsbacher & Robertson, 1992; Haenggi, Gernsbacher, & Bolliger, 1993; Haenggi, Kintsch, & Gernsbacher, 1995). However, if the incoming information is less coherent, comprehenders employ a different process: They shift and initiate a new substructure (Foertsch & Gernsbacher, 1994; Gernsbacher, 1985). So, most mental representations comprise several branching substructures. THE PSYCHOLOGY OF LEARNING AND MOTIVATION. VOL. 31
85
Copyright 6 1997 by Academic Press. All rights of reproduction in any form reserved. M)79-7421/97 S2.5.00
86
Morton Ann Gernsbacher
The building blocks of mental structures are what I refer to as memory nodes. According to the Structure Building Framework, memory nodes are activated by incoming stimuli. Once activated, the information they represent can be used by cognitive processes. Furthermore, according to the Structure Building Framework, activated memory nodes transmit processing signals. These processing signals either suppress or enhance the activation of other memory nodes. In other words, once memory nodes are activated, two mechanisms modulate their level of activation. The two mechanisms are suppression and enhancement. Suppression decreases or dampens the activation of memory nodes when the information they represent is no longer as necessary for the structure being built. Enhancement increases or boosts the activation of memory nodes when the information they represent is relevant to the structure being built. By modulating the activation of memory nodes, suppression and enhancement contribute greatly to language comprehension. I want to stress, however, that suppression and enhancement are general cognitive mechanisms. They are not dedicated solely t o language; they play vital roles in nonlinguistic processes, too. But language comprehension draws heavily on these two mechanisms. This chapter focuses on the mechanism of suppression. While I believe that most people can appreciate that we need a mechanism that enhances relevant o r related information, I have suggested that a mechanism that suppresses inappropriate or irrelevant information is perhaps even more crucial to the goal of comprehension, or in the words of the Structure Building Framework, the goal of building coherent mental representations. We need a mechanism of suppression because whenever we comprehend language, we experience various types of interference. Sometimes this interference during comprehension arises from the external environment, as when we conduct a conversation in a noisy restaurant or listen to a lecture with some clod in the audience whispering beside us. At other times, interference during comprehension arises internally, as when we have to deal with the competing meanings of a word or phrase, or the alternate references of a pronoun. Indeed, even in a process as seemingly straightforward as reading a string of letters, such as ROWS, mental information that is related to that string of letters is often activated in our minds. This mental information might be orthographically related (such as the letter string BOWS), or phonologically related (such as the sound/roz/), o r even semantically related (such as the concept “rose”). And indeed, laboratory experiments demonstrate that adults have difficulty quickly rejecting the letter string ROWS as not being a member of the semantic category, flower (van Orden, 1987; van Orden, Johnston, & Hale, 1988).
Attenuating Interference during Comprehension
87
External information often interferes with our comprehension. For example, laboratory experiments demonstrate that it is harder to read a word when it is written within a line drawing of an object, and it is harder to name a line-drawn object if a word is written within it (Rayner & Posnansky, 1978; Rosinski, Golinkoff, & Kukish, 1975; Smith & McGee, 1980). Thus, successful comprehension involves successfully attenuating or inhibiting interfering information. I have argued that a particular cognitive mechanism, what I call the cognitive mechanism of suppression, reduces such interference. In my previous research, I have empirically illustrated the crucial role that suppression plays in many comprehension phenomena. These phenomena include the following: 1. Lexical Access: how comprehenders understand or “access” from their memory the meanings of words (Faust, Balota, Duchek, Gernsbacher, & Smith, in press; Faust & Gernsbacher, 1996; Gernsbacher & Faust. 1991b, 1995; Gernsbacher & St. John, in press); 2. Anaphoric Reference: how comprehenders understand to whom or to what anaphors, such as pronouns, refer (Foertsch & Gernsbacher, 1994; Garnham, Traxler, Oakhill, & Gernsbacher, 1996; Gernsbacher, 1989); 3. Cataphoric Reference: how words that are marked by devices, such as spoken stress, gain a privileged status in comprehenders’ memories (Gernsbacher & Jescheniak, 1995; Gernsbacher & Shroyer, 1989); 4. Syntactic Parsing: how we decode the grammatical forms of sentences into meaning (Gernsbacher & Robertson, 1996); 5. Surface Information Loss: the finding that seemingly superficial information, such as syntactic form, is often forgotten more rapidly than seemingly more important information, such as thematic content (Gernsbacher, 1985); 6. Metaphor Interpretation: how we understand figurative expressions such as “lawyers are sharks” (Gernsbacher, Keysar, & Robertson, 1995); 7. Inferencing: how comprehenders infer information that is only implied by a text or discourse (Carreiras & Gernsbacher, 1992; Gernsbacher, 1991b. 1994; Gernsbacher, Goldsmith, & Robertson, 1992; Gernsbacher & Robertson, 1992; Oakhill, Garnham, Gernsbacher, & Cain, 1992); 8. General Comprehension Skill: skill at comprehending linguistic and nonlinguistic media (Gernsbacher, 1993; Gernsbacher & Faust, 1991a, 1995; Gernsbacher & Robertson, 1995; Gernsbacher, Varner, & Faust, 1990). In the remainder of this chapter, I briefly review some of these experiments. They demonstrate that the mechanism of suppression plays a powerful role in many language comprehension phenomena. Indeed, the role is so crucial that persons who are less skilled at comprehension are marked
88
Morton Ann Gernsbacher
by less efficiency in suppressing or inhibiting interfering information. Let me begin by describing the role of suppression in lexical access.
11. Attenuating Interference during Lexical Access During lexical access, the cognitive mechanism of suppression attenuates the interference caused by other lexical information that is activated when a printed word is read, or a spoken word is heard. This information might be the meanings of a word that are not relevant to the immediate context, for example, the saloon meaning of bar in the pun Two men walk into a bar and a third man ducks. Or the interfering information might be other words or phrases that are related to the sound pattern of a spoken word or phrase, as in the classic new display often erroneously interpreted as nudist play. Most models of lexical access propose that multiple types of information are activated when we read or hear a word; however, my research demonstrates that the mechanism of suppression dampens the activation of the unnecessary information. To examine the role of suppression in lexical access, I have capitalized on a phenomenon that I believe is a quintessential demonstration of the activation of superfluous information during lexical access. The phenomenon involves the comprehension of homonymswords that share the same lexical form but differ in their meaning or origin, for example, chest, tire, bowl, match, organ, head, plot, ring, nail. All languages have homonyms (e.g., cola, sal, calle, hota, trompa in Spanish). Indeed, they are usually the most frequently occurring words in a language. Dahlgren has suggested that the average English word has three different meanings, and Britton has estimated that homonyms comprise about 40 percent of our most common open class words. Indeed, the more common the word, the more likely it is t o have multiple meanings. The homonym phenomenon that I have empirically studied is this: Immediately after we hear or read a homonym such as tnatch or duck, multiple meanings are activated. And more intriguingly, this activation of multiple meanings occurs regardless of the semantic or syntactic context in which the homonym occurs. For example, immediately after we hear or read the homonym match in the sentence He lit the match, both the “fire stick” and the “competition” meanings are activated. Immediately after we hear or read the homonym duck in the sentence He needed to duck, both the “crouching” and the University of Oregon mascot meaning are activated. Swinney (1979) and Tanenhaus, Leiman, and Seidenberg (1979) were among the first researchers to demonstrate this nonintuitive phenomenon, more than 15 years ago; it has been replicated numerous times since. These
Attenuating Interference during Comprehension
89
researchers also demonstrated-in line with our introspections-that the contextually inappropriate meanings of homonyms do not remain activated forever. What happens to these contextually inappropriate meanings? How do they become less activated? I have proposed that the cognitive mechanism of suppression dampens their activation. More specifically, I have hypothesized that memory nodes that represent a higher level structure-in this case the sentence-level structure-transmit processing signals to suppress the activation of the inappropriate lexical-level meanings. My research has provided several sources of converging evidence to support this proposal. For example, in a series of laboratory experiments, postdoctoral fellow Mark Faust and I empirically demonstrated that suppression and not decay reduces the activation of inappropriate meanings (Gernsbacher & Faust, 1991b). That is, inappropriate meanings do not lose activation over time simply because their activation fades with time. We also empirically ruled out a mental “winner takes all” explanation: When inappropriate meanings become less activated it is not because the more appropriate meanings have become more activated. In other words, the inappropriate and appropriate meanings are not “slugging it out mano a mano”; rather, the source of the activation reduction comes from a higher level. Indeed, using a parallel distributed processing network, postdoctoral fellow Mark St. John and I computationally demonstrated how sentencelevel suppression can dampen the activation of contextually inappropriate word meanings (Gernsbacher & St. John, in press). In our connectionist network, suppression driven by a sentence-level representation, what St. John refers to as a gestalt level of representation, was the only type of topdown feedback we allowed, and that alone allowed us to perfectly simulate the behavioral data. Further demonstrating that suppression and not simply decay is the mechanism responsible for decreasing the activation of the inappropriate meanings of homonyms, laboratory coordinator Rachel Robertson and I empirically demonstrated that suppression carries costs (Gernsbacher & Robertson, 1994). After subjects read a sentence such as He lit the match, they were considerably slower and considerably less accurate at simply verifying that the sentence He won the match made sense. If, after reading the sentence He lit the match, the inappropriate meaning of match simply decayed, that is, the competition meaning of match simply returned to base line, that meaning should not have been harder to activate in order to comprehend the subsequent sentence. Indeed, Gernsbacher and Faust (1995) created a laboratory condition in which it behooved subjects to suppress the inappropriate meanings of homonyms, and we discovered that
Morton Ann Gernsbacher
YO
subjects employed suppression more rapidly in this condition than they did in a condition in which the need for suppression occurred only rarely. Furthermore, as 1 describe at the end of this chapter, I have conducted an extensive series of experiments demonstrating that individuals who are less efficient at suppressing many types of information, for example, the color of ink in a Stroop color-naming task, hold onto inappropriate meanings considerably longer than do individuals who are more efficient in suppressing extraneous information. And most recently, postdoctoral fellow Faust and I discovered a right-visual-field, left-cerebral-hemisphere advantage for suppressing the inappropriate meanings of homonyms (Faust & Gernsbacher, 1996). When we presented homonyms to the left visual field (thereby hypothetically stimulating the right hemisphere prior to the left hemisphere), resolution of homonym meanings was slightly delayed. Although we still have miles to go before being able to stake our explorers' flag atop the cerebral location of our putative suppression mechanism, we find it less plausible that a decay mechanism would be similarly lateralized. From all of these findings, I conclude that the mechanism of suppression, which enables the attenuation of interfering mental activation, such a5 the inappropriate meanings of homonyms, plays a crucial role in lexical access. I turn now to a review of the research I have conducted that investigates the role of suppression in anaphoric reference.
111.
Attenuating Interference during Anaphoric Reference
Anaphoric reference is the process by which readers or listeners understand to whom or to what an anaphor, such as a pronoun, refers. In a series of experiments, I discovered that suppressison enables anaphoric reference by attenuating the interference caused by the activation of other referents. By other referents 1 mean the people or things to whom or to which an anaphoric expression does not refer. For example. consider the sentence A m predicted that Pcrrri woicld lose the track race, hut she COFW in first very easily. In this sentence, the pronoun she is an anaphoric device, which most people interpret to refer to the referent Pam. I discovered that correctly interpreting such anaphoric devices is not so much a matter of activating one of the two possible referents: Both are highly activated because they were just mentioned in the first clause. Rather, understanding to whom the pronoun she in the second clause refers depends on how quickly comprehenders can reduce the activation of the referent to whom the pronoun she does not refer (i.e., Arm in the example sentence). In my experiments, subjects read sentences word by word. The first clause of each sentence introduced two participants, for example, Ann and Pam,
Attenuating Interference during Comprehension
91
as in Ann predicted that Pan1 would lose the track race. In the second clause, one of those two participants was referred to anaphorically, using either a very explicit repeated name anaphor, such as Pam, or a less explicit pronominal anaphor, such as she, as in but she came in first very easily. I measured activation of the anaphors’ referents (e.g., P a m ) and what I called the nonreferents (e.g., Ann) using the probe verification task. Subjects were shown a test name, for example, “Pam” or “Ann,” or a name that had not occurred in the sentence, and their task was to verify whether the test name had occurred in the sentence. Presumably, the faster subjects respond to the test name, the more activated the participant represented by that test name is. In half the experimental sentences the referent was the firstmentioned participant, and in half the referent was the second-mentioned participant, as Pam was in the example sentence. In my first experiment I measured activation immediately before versus immediately after the name versus pronoun anaphors occurred. The first test point served as a baseline. I observed that immediately after the very explicit name anaphors were read, the referents were considerably more activated than they were before: that is, reaction times decreased. More intriguingly, immediately after the very explicit name anaphors were read, the nonreferents were considerably less activated than they were before: that is, reaction times increased. By rementioning one participant, the other participant decreased in activation. However, this pattern occurred only for the very explicit name anaphors. For the pronouns, neither the referents nor the nonreferents changed in the activation. This pattern has been replicated in English (MacDonald & MacWhinney, 1990), Spanish (Carreiras, 1997), Korean (Lee, 1992) and American Sign Language (Emmorey. in press). These data suggest that very explicit repeated name anaphors immediately lead to the suppression of nonreferents. In contrast, less explicit-and indeed momentarily ambiguous-pronoun anaphors do not immediately lead to suppression. In a further experiment (Gernsbacher, 1989), I measured activation immediately before repeated-name versus pronoun anaphors, as I did before, and. again, this before-the-anaphor test point served as a baseline. However, in this experiment my comparison test point was at the end of the sentence, after the semantic/pragmatic information (which could disambiguate the syntactically ambiguous pronouns) had occurred. For example, activation was measured at the two test points indicated by asterisks in the following example sentence: Ann predicted that Pam would lose the track race, but “Ptitdshe came in ,first very easily.” By the end of the sentence, even the gender-ambiguous pronoun anaphors had led to a reliable amount of suppression of the nonreferents.
Y2
Morton Ann Gernsbacher
In yet a further experiment (Gernsbacher, 1989),I placed the contextual information before the anaphors, as in Ann lost a track race to Pam. Enjoying the victory, P a d s h e headed toward the shower, or Ann lost a track race to Pam. Accepting the defeat, Andshe headed toward the shower. Despite the context preceding the anaphors, the less explicit pronoun anaphors still did not lead to a reliable amount of suppression until the end of the sentence. Thus, information from outside an anaphor can also trigger suppression, although it does so more slowly and less powerfully. This is good, because with zero anaphors, as in Ann lost a tennis match to Pam and 0 cried all the way home, the anaphor provides no information about its referent. All the information is provided by the semantic, pragmatic, and syntactic context. Therefore, zero anaphors should be the least effective at triggering suppression, a prediction confirmed by Corbett and Chang (1983).Together, the experiments I have described here demonstrate the role of suppression in enabling anaphoric reference: Suppression enables anaphoric reference by attenuating the interference caused by other referents.
IV. Attenuating Interference during Cataphoric Reference Just as anaphoric devices enable reference to previously mentioned concepts, cataphoric devices enable reference to subsequently mentioned concepts. Cataphoric devices include such overt markers as vocally stressing a word in spoken discourse, or boldfacing a word in printed text. Presumably, speakers and writers mark certain concepts with cataphoric devices because those concepts will play a key role in the text or discourse. Thus, it would behoove listeners and readers if those key concepts had a privileged status in their mental structures. Master’s student Suzanne Shroyer and I (Gernsbacher & Shroyer, 1989) demonstrated that in spoken English, the unstressed, indefinite article this, as in So this man walks into a bar, as opposed to So a man walks into this bar, operates as a cataphoric device. The indefinite this is a relative newcomer to English; Wald (1983) dates its use back to the late 1930s. It occurs almost exclusively in informal spoken dialects rather than in formal or written ones, although I have observed personally that many of my e-rnail pen pals use the indefinite this in written e-mail. Because it is an indefinite article, this is used to introduce new concepts into discourse. Indeed, of the 243 occurrences of the indefinite this that Prince (1981) observed in Stud Terkel’s book Working, 242 introduced a distinctly new concept. More interestingly-particularly with regard to my conjecture that the indefinite this operates as a cataphoric device to enable subsequent reference-in 209 of the 243 occurrences of the indefinite this,
Attenuating Interference during Comprehension
93
the concept introduced with the indefinite this was referred to again. Similarly, when Wright and Givon (1987) recorded 8- and 10-year-old children telling one another stories and jokes, they found that when the children introduced concepts with the indefinite this, they referred to those concepts an average of 5.32 times in their next 10 clauses. When the children introduced concepts with the indefinite d u n , they referred to those concepts only .68 times in their next 10 clauses. These descriptive data suggest that speakers use the indefinite this to introduce key concepts. We (Gernsbacher & Shroyer, 1989) tested this proposal experimentally. We presented spoken narratives to college students, telling them that at some point in each narrative the original narrator would stop talking: when that happened, it was their job to continue. For example, subjects heard the following narrative: I swear, m y friend Vicki, every time we go to a garage sale, she just ‘uh, she just goes crazy. I mean like last Saturday we went to one near campus, ‘n she just had to buy thisJan ashtray, n’man . . . . As this example illustrates, the last clause of the part of each narrative that subjects heard introduced a new concept, for example, ashtray. We manipulated whether this concept was introduced with the indefinite this (this ashtray) or the indefinite d u n (an ashtray). When we introduced the concepts with the indefinite this, subjects mentioned those concepts considerably more frequently, virtually always within the first clauses that they produced, and usually with less explicit anaphors such as pronouns. In contrast, when we introduced the concepts with the indefinite d u n , subjects mentioned the concepts less frequently, and typically with more explicit anaphors such as repeated noun phrases. (Through cross-splicing we ensured that the acoustic properties of the matched narratives and their critical concepts were otherwise identical.) These data demonstrate that concepts marked by cataphoric devices, such as the indefinite this, are more salient in listeners’ mental representations. Postdoctoral fellow Joerg Jescheniak and I (Gernsbacher & Jescheniak, 1995) discovered the role that the cognitive mechanism of suppression plays in enabling this privileged status: Suppression enables cataphoric reference by attenuating the interference caused by the introduction of other concepts. In this way, a cataphorically marked concept gains that privileged status in comprehenders’ mental representations, so that it can be referred to more easily. Consider as an analogy a call for volunteers during which the entire line of candidates steps back, save one. The one candidate who did not step back-who was not suppressed-becomes most accessible for selection. Subjects in our (Gernsbacher & Jescheniak, 1995) experiments also heard narratives, like the “Vicki going to a garage sale” narrative. We manipulated the indefinite this in some experiments, and in other experiments we manip-
94
Morton Ann Gernsbacher
ulated a seemingly more powerful cataphoric device, contrastive intonational stress. Using a verification task, we measured activation of the experimental concepts. In addition to cataphorically marked concepts being more activated-in other words, enhanced-we also found that cataphorically marked concepts are very resilient to being suppressed by a subsequently introduced concept. For example, in the “Vicki going to the garage sale” narrative, after we introduced the experimental concept (i.e., this/an ashtray), we introduced a new concept, vase (i.e., ‘n she just had to hiiy this/ an ashtray, ’nman, then shesaw A vase, . .). We observed that the previously mentioned concept, ashtruy, was greatly attenuated in its activation when it was not marked by a cataphoric device. However, when the previously mentioned concept was marked by a cataphoric device, it was just as activated after we introduced a new concept as it was immediately after it was introduced. Thus, cataphoric devices-the indefinite this and contrastive, intonational stress-attenuates the interference caused by introducing other concepts. By attenuating the interference from other concepts, cataphoric devices lead to a privileged status. Furthermore, the two cataphoric devices differ in how powerfully they lead to this privileged status: suppression is more powerfully triggered by contrastive stress than by the indefinite this. This difference makes sense: Contrastive stress is considerably more marked; it is a very iconic way of emphasizing a word in spoken discourse, similar to boldfacing a word in written text. The indefinite this is considerably more subtle; many of us are unaware of our informal use of it. Indeed, our undergraduate research assistants, whom we typically keep blind to our experimental manipulations, were stymied in their attempts to figure out what we were doing in the indefinite this experiments. So, like anaphoric devices, the strength of the suppression signals triggered by cataphoric devices is a function of the cataphoric devices’ markedness. All the experiments that I have described so far demonstrate the role that suppression plays in attenuating lexical- or concept-level interference. I have also examined the role of suppression in attenuating sentencelevel interference.
V. Attenuating Interference during Syntactic Parsing Motivated by the adage Time pies like un arrow; fruit ,flies like a banana, often attributed to Groucho Marx, laboratory coordinator Rachel Robertson and I hypothesized a role that the mechanism of suppression might play in syntactic parsing (Gernsbacher & Robertson, 1996). We proposed that suppression attenuates the interference caused by parsing a previous syntactic form. As the time ffies/fruitflies example demonstrates, once we
Attenuating Interference during Comprehension
95
have parsed the phrase timeflies as a noun and verb, it is difficult not to parse the phrasefruifflies in the same way. We (Gernsbacher & Robertson, 1996) examined a more stringent type of interference by using phrases such as visiting in-laws, which can be interpreted either as a plural noun phrase (i.e., people who are related to one’s spouse and come to visit) or as a gerundive nominal (i.e., the act of visiting people who are related to one’s spouse). In our experiments (Gernsbacher & Robertson, 1996), we preceded sentences containing phrases such as visiting in-laws with sentences that required a similar or conflicting syntactic parse. For example, subjects first read Washing dishes is a drag, and then read Visiting in-laws are, too. Or subjects first read Whining students are a drag, and then read Visiting inlaws is, too. The subjects’ task was to read each sentence and simply decide whether it was grammatical. We found that subjects were considerably slower and frighteningly less accurate to say that a sentence such as Visiting in-laws are, too was grammatical after they read the sentence Washing dishes is a drag. Similarly, subjects were considerably slower and less accurate to say that the sentence Visiring in-laws is, too was grammatical after they read the sentence Whining students are a drag. We interpreted these data as suggesting that correctly responding to the second sentence requires attenuating, or suppressing, the interference caused by the syntactic form in the first sentence. We observed the same effect when we made the second sentences less syntactically dependent on the first sentence, by omitting the ellipses. For example, subjects were again slower and less accurate to say that the sentence Visiting in-laws are a drag, too was grammatical after they read the Washing dishes sentence. And, subjects were considerably slower and less accurate to say that the sentence Visiting in-laws is a drag, too was grammatical after they read the Whining students sentence. Furthermore, we observed the same effect when we made the second sentences syntactically independent of the first sentence, and the verb in the first sentence was not even marked for number. For example, subjects were still slower and still less accurate to say that the sentence Visiting in-laws are a drag was grammatical after they read the sentence Washing dishes can be a bother, and vice versa after they read the sentence Whining students can be a bother. This phenomenon underscores the need for suppression to attenuate the interference caused by a previous syntactic form.
V1. Attenuating Interference during Metaphor Interpretation Rachel Robertson and I, in collaboration with Boaz Keyser (of the University of Chicago) have also explored the role of suppression in metaphor
96
Morton Ann Gernsbacher
interpretation. According to Glucksberg and Keysar (1990), when we interpret a metaphor such as Lawyers are sharks, we should enhance attributes of the metaphor’s vehicle (e.g., sharks) that are common to the metaphor’s topic (e.g., lawyers). So, after comprehending the metaphor Lawyers are sharks, we should enhance the facts that sharks are tenacious, fierce, and aggressive, among other attributes. We augmented Glucksberg and Keysar’s (1990) theory by proposing that when we interpret a metaphor we also suppress the attributes that are not appropriate to (or concordant with) a metaphorical interpretation. So, for example, when we interpret the metaphor Lawyers are sharks, we might suppress attributes such as sharks being good swimmers, having fins, and living in the ocean. We tested both of these hypotheses by asking subjects to read a statement that might be metaphorical such as Lawyers are sharks, and then confirm the verity of a property statement such as Sharks are tenacious. In our first experiment, we used as a control condition statements that contained the same vehicle but a nonsensical topic, such as Notebooks are sharks. We (Gernsbacher et al., 1995) found striking evidence that interpreting a metaphor such as Lawyers Lire sharks leads to both the enhancement of the attributes that are appropriate to the metaphorical interpretation and the suppression of attributes that are inappropriate to the metaphorical interpretation. For instance, subjects were faster to verify the statement Sharks are tenacious after they read the metaphor Lawyers are sharks than after they read the control statement Notebooks are sharks. This finding supports the hypothesis that interpreting a metaphor involves enhancing attributes that are appropriate to the metaphorical interpretation. In contrast, subjects were considerably slower to verify the statement Sharks are good swimmers after they read the metaphor Lawyers are sharks than after they read the control statement Notebooks are sharks. This finding supports the hypothesis that interpreting a metaphor involves suppressing attributes that are inappropriate to the metaphorical interpretation. In a second experiment, we observed identical results when, instead of using a nonsensical statement as a baseline (control), we used a literal statement as a baseline. For example, we presented the literal statement Hammerheads are sharks as a baseline comparison for the metaphorical statement Lawyers are sharks. Again, we found striking evidence to support the hypothesis that interpreting a metaphor leads to both the enhancement of attributes that are appropriate to the metaphorical interpretation and the suppression of attributes that are inappropriate t o the metaphorical interpretation. For example, again, subjects were faster to verify the statement Sharks are tenacious after they read the metaphor Lawyers are sharks than after they read the literal statement Hammerheads are sharks. And conversely, subjects were again considerably slower to verify the statements
Attenuating Interference during Comprehension
97
Sharks are good swimmers after they read the metaphor Lawyers are sharks than after they read the literal statement Hammerheads are sharks. Therefore, both experiments demonstrated that interpreting a metaphor involves both enhancing the attributes that are relevant t o the metaphorical interpretation and more intriguingly, suppressing the attributes that are not relevant to the metaphorical interpretation. VII. Attenuating Interference during Inference Revision When most of us hear or read that George became too bored to finish the history book, we infer that George is reading a very boring book. However, if we later hear or read that George had already spentfive years writing it, we must revise our initially drawn inference because it was inappropriate. Brownell, Potter, Bihrle, and Gardner (1986) found that right-hemispheredamaged patients had a whale of a time revising such inferences. They concluded that right-hemisphere-damaged patients’ difficulty arose because they were unable to “let go of” the initial inferences that they drew. Perhaps revising such an inference is difficult because the revision requires suppressing the initially drawn inference. Thus, another role that suppression might play is to attenuate the interference caused by a previously drawn, but erroneous, inference. We empirically tested this hypothesis, by investigating whether revising such inferences was difficult, not just for right-hemisphere-damaged patients but for “normal” college-aged adults. We constructed 40 two-sentence vignettes, similar to the George became too bored tofinish the history book. He had already spent five years writing it example (other stimuli included Jeff got a ticket after parking his car. As he headed into the movie theater, he handed the ticket to the usher; Sarah drove frantically all the way to the Emergency Room. She was already running 15 minutes late for her shift that evening; Jack painted the boat a bright red. Then he painted the ocean a deep blue and the sun a bright orange). We measured how long subjects needed to read the inference-revising second sentence (e.g., He had already spent five years writing it) after they read the inference-inviting premise sentence (e.g., George became too bored to finish the history book). We compared how long subjects needed to read the inference-revising second sentence (e.g., He had already spentfive years writing it) after they read the inference-inviting premise sentence or after they read a control premise sentence, which was a sentence that did not invite the inference; indeed it explicitly stated a different situation (e.g., George became too bored to finish writing the history book; Jeff bought a ticket after parking his car; Sarah drove frantically all the way to her job at the Emergency Room;
YX
Morton Ann Gernsbacher
In the painting, Jack painted the boat a bright red). We found that subjects required substantially longer to read the second sentence (e.g., He had already spent five years writing it) after they read the experimental (inference-inviting) premise sentence (e.g., George became too bored to,finish the history b o o k ) than after they read the control (inferencenoninviting) premise sentence (e.g., George became too bored to ,finish writing the history hook). We interpreted subjects’ greater latency as reflecting their difficulty in suppressing a previously, but erroneously, drawn inference. Furthermore, we found that members of a particular subject group-a group that we have previously identified to have difficulty quickly employing suppression (as I shall describe later)-were substantially slower to reject a test word that was related to the erroneously drawn inference, even after they read the inference-revising second sentence. For example, members of this group of subjects took longer t o reject the test word READ after they read the inference-revising sentence He hud already spent five years writing it. Members of this subject group were less skilled comprehenders, and this finding leads me to the last section of this chapter, the role of suppression in general comprehension skill.
VIIL Attenuating Interference and Comprehension Skill A few years ago, honors student Kathy Varner, postdoctoral fellow Mark Faust, and I presented evidence in support of a construct we called “General Comprehension Skill” (Gernsbacher et al., 1990). Briefly put, we found that adults’ skill in comprehending written language was highly correlated with their skill in comprehending spoken language, and both skills were highly correlated with comprehending nonverbal picture stories. We also found a critical characteristic of less skilled adult comprehenders: Less skilled adult comprehenders are less able to suppress quickly the inappropriate meanings of homonyms. We (Gernsbacher et al., 1990) discovered this critical characteristic in the following way: We selected 64 more versus less skilled University of Oregon undergraduates on the basis of their performance on our MultiMedia Comprehension Battery (Gernsbacher & Varner, 1988). This battery tests reading, listening, and picture story comprehension. We drew the more skilled comprehenders from the upper third of a distribution of 270 subjects and the less skilled comprehenders from the bottom third. We invited these more and less skilled subjects to return to our lab (which was no easy feat, as the less skilled subjects did not have that much fun the first time they were there). When the subjects returned they read short
Attenuating Interference during Comprehension
99
sentences: after reading each sentence they were shown a test word. Their task was to decide quickly whether the test word fit the meaning of the sentence that they had just read. On experimental trials, the final word of the sentence was a homonym, such as spade, as in He dug with the spade. The test word on these trials was related to a meaning of that homonym, but not the meaning implied by the sentence, for example, ACE. We compared how rapidly the more versus less skilled comprehenders could reject a test word that was related to the inappropriate meaning with how rapidly they could reject the same test word after reading a control sentence, for example, He dug with the shovel. The more time subjects took to reject ACE following the spade versus shovel sentence, the more interference we hypothesized they were experiencing from the inappropriate meaning. We measured this interference immediately (100 ms) after subjects finished reading the sentences and after an 850-ms delay. Immediately after both the more and less skilled comprehenders read the homonyms, both groups experienced a reliable amount of interference, and, indeed, the two groups did not differ in the amount of interference they experienced at the immediate test point. In contrast, after the delay, the more skilled comprehenders were no longer experiencing a reliable interference, suggesting that they had successfully suppressed the inappropriate meanings of the homonyms. But for the less skilled comprehenders, they experienced the same amount of interference after the delay as they experienced immediately, suggesting that they were less able to quickly suppress the inappropriate meanings of the homonyms. This pattern has been replicated by our colleagues around the world with whom we have shared our stimuli. For example, Leslie Twilley and Peter Dixon at the University of Alberta replicated this pattern using our measure of comprehension skill. Harvey Shulman at Ohio State replicated this pattern testing subjects who scored in the top versus bottom half of the ACT verbal test, but not the math test. Francesca Pazzaglia and her colleagues at the University of Padova replicated this pattern using Italian homonyms (and Italian subjects). Natasha Todorov at Macquarie University replicated this pattern with 7th-grade students selected according to their NelsonDenny reading scores. In his dissertation Robert Crane at Washington State replicated this pattern testing university students, with small versus large reading spans. And we replicated this pattern testing United States Air Force recruits. Thus, this pattern replicates with Canadians, Italians, Australians, Buckeyes, and the U.S. military. Furthermore, this pattern occurs when comparing members of other populations who hypothetically suffer from less efficient suppression with members of populations who are hypothesized to have more efficient suppression. For example, using our task and stimulus materials, McDowd and
100
Morton Ann Gernsbacher
Sundry at the University of Southern California found that healthy elderly subjects showed less efficient suppression compared with college-aged subjects. Elizabeth Schaunessy found that children diagnosed with attention deficit disorder showed less efficient suppression compared with children not diagnosed with attention deficit disorder. Mark Faust, David Balota, Janet Duchek, Stan Smith, and I found that patients with severe senile dementia of the Alzheimer’s type showed extraordinarily inefficient suppression compared with patients with only moderate dementia compared with healthy age-matched controls (Faust et al., in press). Indeed, our dementia data show a dosing effect: The more severe the dementia, the more inefficient the suppression. In all the data that I have reviewed, all the subjects showed initial interference from the inappropriate meanings, which I believe is crucial to demonstrate, but the members of the population hypothesized to suffer from less efficient suppression showed continued interference from the inappropriate meanings. Mark Faust and I (Gernsbacher & Faust, 1991b) also observed the same pattern when we examined how quickly less versus more skilled comprehenders could reject test words related to the incorrect forms of homophones, for instance, how quickly they could reject the test word CALM following the sentence He had lots of patients, versus the sentence He had lots of students. (Prior to collecting our data, we conducted a pilot experiment-no pun intended-to ensure that members of this population did know the correct spelling of these homophones.) We (Gernsbacher & Faust, 1991b) also discovered that less versus more skilled comprehenders are not less able to reject the contextually inappropriate meanings of homonyms just because they do not know what is appropriate. We observed that less skilled comprehenders perform equally as well as more skilled comprehenders when the task is to accept the appropriate meaning of a homonym, for example, when their task is to correctly say “yes” that the test word ACE is related to the sentence He dealt the spade. More recently, Rachel Robertson and I (Gernsbacher & Robertson, 1995) replicated our tried and true finding that less skilled comprehenders are worse than more skilled comprehenders when the task is to reject a test word that is related to the inappropriate meaning. For example, less skilled comprehenders are slower to reject the test word ACE after reading the sentence He dug with the spade. Presumably this is because less skilled comprehenders are less able to suppress the activation of the inappropriate meanings. We (Gernsbacher & Robertson, 1995) also replicated the finding I just mentioned, namely that less and more skilled comprehenders do not differ when the task is to accept the appropriate meaning. For example, less skilled comprehenders are just as fast as more skilled comprehenders
Attenuating Interference during Comprehension
101
in accepting the test word A C E after reading the sentence He dealt the spade. Again, this suggests that less skilled comprehenders’ difficulty in rejecting inappropriate meanings is not because they do not know what is appropriate. However, to ensure that we didn’t have a “Nancy Reagan effect” on our hands (i.e., less skilled comprehenders just can’t say “no”), we (Gernsbacher & Robertson, 1995) also created a task in which the goal was to say “yes” to a meaning that was inappropriate, somewhat like what one needs to do to understand a pun. And we again found that less skilled comprehenders were worse than more skilled comprehenders. For example, less skilled comprehenders were slower to accept the test word ACE after reading the sentence He dug with the spade, perhaps because this taskaccepting an inappropriate meaning-requires suppressing the appropriate meaning (recall how difficult it was to understand the pun Two men walk into a bnr and the third man ducks. It is as though to understand the “metal bar” meaning, one needs to suppress the “tavern” meaning).
IX. Summary To summarize, I have suggested that the cognitive mechanism of suppression attenuates interference in many language comprehension phenomena. During lexical access, the mechanism of suppression attenuates the interference caused by the activation of other lexical information, such as the inappropriate meanings of homonyms. During anaphoric reference, the mechanism of suppression attenuates the interference caused by the activation of other potential referents. In this way, the referent to which the anaphor does not refer becomes the most activated concept. Moreover, the strength of the suppression is a function of the markedness of the anaphoric device: More marked anaphors such as repeated proper names immediately lead to suppression; less marked anaphors such as pronouns take longer to enact suppression. During cataphoric reference, the mechanism of suppression attenuates the interference caused by the introduction of other concepts. In this way, a cataphorically marked concept gains a privileged status in comprehenders’ mental representations. The more marked the cataphoric device is, the less interference is caused by the introduction of another concept. More marked cataphoric devices such as spoken stress protect their concepts more than less marked cataphoric devices such as the indefinite this. During syntactic parsing, the mechanism of suppression attenuates the interference caused by a previous syntactic form. During metaphor comprehension, the mechanism of suppression attenuates the interference caused by a literal interpretation. During infer-
102
Morton Ann Gernsbacher
encing, the mechanism of suppression attenuates the interference caused by an initial but inappropriate inference. Thus, my previous research has demonstrated the crucial role that suppression-and by that I mean a general, cognitive mechanism that attenuates interference-plays in many facets of language comprehension. ACKNOWLEDGMENTS Preparation of this chapter was supported by grants from the National Institutes of Health
(R01 NS 29926) and the Army Rcsearch Institute (DASW0194-K-0004).
REFERENCES Brownell, H. H., Potter. H. H., Bihrle, A. M.. & Gardner, H. (1986). Inference deficits in right brain-damaged patients. Brain cirrtf I.angrrrrge. 27, 310-321. Carrciras. M. (1997). Manuscript in preparation. Carreiras, M., & Gcrnsbacher. M. A. ( 1992). Comprehending conceptual anaphors in Spanish. Longiiage arid Cognitive Processes, 7, 281-299, Carreiras, M.. Gernsbacher, M. A., & Villa, V. (1995). The advantage of lirst mention in Sponish. Psychonomic Riilletin & Heview. 2, 124-129. Corbetl, A. T.. & Chang. F. R. (19x3). Pronoun disambiguation: Accessing potential antecedents. Memory & Cognition, I I , 283-204. Deaton, J. A,, & Gernsbacher. M. A. (in press). Causal conjunctions and implicit causality cue mapping in sentence comprehension. Joirrncil of Memory and Lnngiinge. Emmorey, K. A. (in press). Non-antecedent suppression in American Sign Language. Lririgiinge rind Cognitive Processes. Faust. M. E., Balota. D. A.. Duchek, J. A.. Gernsbacher. M. A., & Smith. S. D. (in press). Inhibitory control during sentence processing in individuals with dementia of the Alzheimer type. Brain and Laiigiiage. Faust. M. E., & Gernsbacher. M. A. (1996). Cerebral mechanisms for suppression of inappropriate information during sentence comprehension. Brain and Lmpcrige. 53, 234-259. Foertsch. J., & Gernsbacher, M. A. (1994). In search of complete comprehension: Getting "minimalists" to work. Discourse Processes, pp. 271 -296. Garnham. A.. Traxler. M. J., Oakhill, J.. & Gernshacher, M. A. (1996). The locus o f causality effects in comprehension. Joitrnal of Memory cind Languuge, 35, 517-543. Gernsbacher, M. A. ( 19x5). Surface information loss in comprehension. Cognitive Psychology. 17, 324-363. Gernsbacher. M. A. ( 1989). Mechanisms that improve referential access. Cognilion, 32,99- 156. Gernsbacher, M. A. (1990). Lwzgiicrge comprehension as sfrricture bidding. Hillsdale. NJ: Erlbaum. Gernsbacher, M. A. (199la). Cognitive processes and mechanisms in language comprehension: The structure building framework. In G. H. Bower (Ed.), The psychohgy o,f learning and motivation (pp. 217-263). San Diego, CA: Academic Press. Gernshacher, M. A. (1991b). Comprehending conceptual anaphors. Langriagc? arid Cognitive Processes, 6. 81- 105. Gernshacher. M. A. (1993). Less skilled readers have less efficient suppression mechanisms. Psychological Science, 4, 294-298.
Attenuating Interference during Comprehension
I03
Gernsbacher. M. A. (1994). Activating knowledge of fictional characters' emotional states. In C. A. Weaver. S. Mannes. & C. R. Fletcher (Eds.). Discour.ye conzprehension: Es.says in honor of Wirlter Kintsch (pp. 141-155). Hillsdale, NJ: Erlbaum. Gernsbacher. M. A. (1995). The Structure Building Framework: What it is, what it might also be. and why. In B. K. Britton & A. C. Graesser (Eds.), Models of text rmderstiinding (pp. 289-31 I ) . Hillsdale, NJ: Erlbaum. Gernsbacher, M. A. (1996). Coherence cues mapping during comprehension. In J . Costermans & M. Fayol (Eds.). Processing interclausal relrrtionships in the production and comprehension o f /ex/ (pp. 3-21). Hillsdale, NJ: Erlbaum. Gernsbacher, M. A,, & Faust, M. (IYYla). The mechanism of suppression: A component of general comprehension skill. Journal o f Experimental Psychology: Learning, Memory, and Cognition, 17, 245-262. Gernsbacher. M. A., & Faust, M. (1991b). The role of suppression in sentence comprehension. In G. B. Simpson (Ed.), Unifer.standing word and sentence (pp. 97-128). Amsterdam: North-Holland. Gernsbacher, M. A., & Faust. M. (1995). Skilled suppression. In F. N. Dempster & C. N. Brainerd (Eds.). Interference and inhibilion in cognition (pp. 295-327). San Diego, CA: Academic Press. Gernsbacher, M. A,. & Givon. T. (Eds.). (1995). Coherence in spontatieoirs text. Philadelphia: John Benjamins. Gernsbacher, M. A.. Goldsmith. H. H.. & Robertson, R. R.W. (1992). Do readers mentally represent characters' emotional states'? Cognition and Enlotion. 6, 89-1 11. Gernsbacher. M. A,. & Hargreaves. D. (1988). Accessing sentence participants: The advantage of first mention. Journal of Memory ant1 Liinguage, 27, 699-717. Gernsbacher, M. A,, & Hargreaves. D. (1992). The privilege of primacy: Experimental data and cognitive explanations. In D. L. Payne (Ed.), Prcrgmatics o f word order fkxibility (pp. 83-1 16). Philadelphia: John Benjamins. Gernsbacher, M. A., Hargreaves. D., & Beeman. M. (1989). Building and accessing clausal representations: The advantage of first mention versus the advantage of clause recency. Joirrniil of Meniory uncl Language, 28, 735-755. Gernsbacher. M. A,, & Jescheniak. J. D. (1995). Cataphoric devices in spoken discourse. Cognitive Psychology. 2Y, 24-58. Gernsbacher, M. A,, Keysar, B.. & Robertson, R. R. W. (1995, November). The role of sirppre.s.sion irnd enhancement in nzetiiphor interpretation. Paper presented at the 36th annual meeting of the Psychonomic Society, Los Angeles. Gernsbacher. M. A,, & Robertson, R. R. W. (1992). Knowledge activation versus sentence mapping when representing fictional characters' emotional states. Language and Cognitive Proce.sses, 7, 353-371. Gernsbacher. M. A.. & Robertson. R. R. W. (1994. November). The costs and benefits of merming. Paper presented at the 35th meeting of the Psychonomic Society, St. Louis. MO. Gernsbacher, M. A,. & Robertson, R. R. W. (1995). Reading skill and suppression revisited. P.sychologicir1 Science, 6, 165- 169. Gernsbacher. M. A,. & Robertson, R. R. W. (1997). Parullel form affects sentence comprehrns h 7 . Manuscript submitted for publication. Gernsbacher. M. A,. & Shroyer. S. (1989). The cataphoric use of the indefinite this in spoken narratives. Memory & Cognition, 17, 536-540. Gernsbacher. M. A,. & St. John, M. F. (in press). Modeling the mechanism of suppression in lexical access. In R. Klein & P. McMullen (Eds.), Converging methods for studying rrcidirig mid dvslexio. Cambridge, MA: MIT Press,
104
Morton Ann Gernsbacher
Gernsbacher, M. A.. & Varner, K. R. (1988). The multi-media Comprehension battery (Tech. Rep. No. 88-03). Eugene: University of Orgeon, Institute of Cognitive and Decision Sciences. Gernsbacher, M. A., Varner. K. R., & Faust, M. (1990). Investigating differences in general comprehension skill. Journal of Experimental Psychology: Learning, Memory, and Cognition, 16, 430-445. Glucksberg, S . , & Keysar, B. (1990). Understanding metaphorical comparisons: Beyond similarity. Psychological Review, 07, 3-18. Haenggi, D.. Gernsbacher, M. A,, & Bolliger. C. M. (1993). Individual diffcrcnccs in situationbased inferencing during narrative text comprehension. In H. van Oostendorp & R. A. Zwaan (Eds.), Naturalistic text comprehension: Vul. 53. Advances in tliscoitrse processing (pp. 79-96). Norwood, NJ: Ablex. Haenggi, D., Kintsch, W., & Gernsbacher, M. A. (1995). Spatial situation models and text comprehension. Discurcr.se frucesses, I Y , 173- 199. Lee, J. (1992). On-line processing o,f pronoun resolution in reuding. Unpublished doctoral dissertation, Korea University, Seoul. MacDonald, M. C., & MacWhinney, B. (1990). Measuring inhibition and facilitation effects from pronouns. Journal of'Memory and Language. 29, 460-492. Oakhill. J., Garnham. A., Gernsbacher, M. A., & Cain, K. (1992). How natural are conceptual anaphors? Language and Cognitive Processes, 7. 257-280. Prince, E. F. (1981). On inferencing of the indefinite-this NPs. In A. Joshi, B. Webber. & 1. Sag (Eds.), Elements uf discourse understanding (pp. 231-250). Cambridge, UK: Cambridge University Press. Rayner, K., & Posnansky, C. (1978). Stages of processing in word identification. Journal of Experimental Psychology: General, 107, 64-80, Rosinski, R. R., Golinkoff, R. M., & Kukish. K. S. (1975). Automatic semantic processing in a picture-word interference task. Child Development, 46, 247-253. Smith, M. C., & McGee, L. E. (1980). Tracing the time course of picture-word processing. Journal of Experimental Psychology: General, 109, 373-392. Swinney, D. A. (1979). Lexical access during sentence comprehension: (Re)consideration of context effects. Journal of Verbal Learning and Verbal Behavior, 18, 645-659. Tanenhaus, M. K., Leiman, J. M., & Seidenberg, M. S. (1979). Evidence for multiple stages in the processing of ambiguous words in syntactic contexts. Journal of Verbal Learning (2nd Verbal Behavior, 18, 427-440. van Orden, G. C. (1987). A rows is a rose: Spelling, sound, and reading. Memory & Cognition, 15, 181-198. van Orden, G. C., Johnston, J . C., & Hale, B. L. (1988). Word identification in reading proceeds from spelling to sound to meaning. Journal of Experimenral Psychology: Learning. Memory, & Cognition, 14, 371-386. Wald, B. (1983). Referents and topics within and across discourse units: Observations from current vernacular English. In S. Klen-Andrew (Ed.), Discourse perspectives on synrcix (pp. 91-116). New York: Academic Press. Wright, S., & Givon, T. (1987). The pragmatics of indefinite reference: Quantified text-based studies. Studies in Language. 11, 1-33.
COGNITIVE PROCESSES IN COUNTERFACTUAL THINKING ABOUT WHAT MIGHT HAVE BEEN Ruth M.J. Byrne
I. Counterfactual Thinking Imagine on your way home from work today you walk by a new route and get mugged. While you’re recovering you may find it hard to avoid thinking that if only you’d walked by your usual route, the mugging wouldn’t have happened. Such counterfactual imaginings rest on the spontaneous generation of alternatives to reality that permeates much of our everyday thought (e.g., Kahneman & Miller, 1986). Counterfactual thinking liberates us from thinking just about facts: we can entertain suppositions about possibilities and even impossibilities in our imaginary worlds. Everyday thinking about past or present events is concerned not just with factual reality but also with nonfactual or hypothetical states of affairs. Everyday thinking about factual reality may be based on its present state, for example: (1) Clinton is president of the United States. or its past state, for example: (2) Kennedy was president of the United States. as Table I shows. Everyday thinking about nonfactual or hypothetical states of affairs may be based on present possibilities (that could happen given the actual state of the world), for example: THE PSYCHOLOGY OF L E A R N I N G A N D MOTIVATION. VOL. 37
105
Copyright 0 1YY7 by Academic Press. All rights of reproduction in any form reserved. W79-7421/97 125.W
Ruth M. J. Byrne
I Oh
TABLE 1 EXAMPLES OF FACTUAL A N D NONFACTUAL (HYPOIHEIICAL) EVENTS IN PASTAND THE PRESENT TIME"
THE
Present
Past
Clinton is president of the United States
Kennedy was presidcnt of the United States
Clinton resigns Clititon is presiderir
Clinton resigned in 1995 Clinton W N S presiclmt 0f
~~
Factual Nonfacttial (hypothetical) Possibilities Impossibilities
A irstrrilia "
o,f
A irstrdio
Countcrlactual cvenh are in italics.
(3) Clinton resigns. or on past possibilities (that could have happened given the actual state of the world but did not), for example: (4) Clinton resigned in 1995. (see Johnson-Laird and Byrne, 1991, chap. 4). Everyday thinking about nonfactual states of affairs may also be based on impossibilities (that could never happen in the past or present), such as:
( 5 ) Clinton is president of Australia. This taxonomy of factual and hypothetical thinking also requires the higher order distinction between fact and fiction (Johnson-Laird & Byrne, 1991). Thinking about past possibilities and past or present impossibilities is often called counterfactual thinking.' Counterfactual thinking plays a central role in many aspects of higher order cognition. For example, counterfactual thinking about counterexamples to conclusions helps us to make deductive inferences (e.g., Johnson-Laird & Byrne, 1991); counterfactual thinking about whether an outcome would have happened if another event had not happened helps us to understand the causal relations between events (e.g., Chisholm, 1946; Kahneman & Miller, 1986; Wells & Gavanski, 1989); counterfactual thinking about what would be possible under conditions different from the current conditions helps us to construct subgoals in problem solving (e.g., Ginsberg, 1986; see also Keane, 1996, in press); and to the extent that creativity requires the intention to improve, counterfactual thinking helps in creative discovery (e.g., Hofstadter, 1985). The primary
'
I will not consider thinking about the future in this article: thinking ahout counterfactual future events is similar to thinking about counterfactual events in general but it also displays some intriguing vagaries (see. e g . Lewis. 1979).
Cognitive Processes in Counterfactual Thinking
I07
function of counterfactual thinking may be a preparatory one, to help us formulate intentions to improve our performance (e.g., Markman, Gavanski, Sherman, & McMullen, 1993; Roese, 1994). We think about the way things might have been so that we can learn from our mistakes (e.g., Landman, 1987; see also Roese & Olson, 1995), even if mulling over the past can sometimes prove dysfunctional (e.g., Davis, Lehman, Wortman, Silver, & Thompson, 1995; Wells, Taylor, & Turtle, 1987). As a result of generating counterfactual alternatives to reality, we experience a range of emotions from regret to relief, including blame, guilt, sympathy, and hope (e.g., Gilovich & Medvec, 1994; Johnson, 1986; D. T. Miller & McFarland, 1987). What cognitive mechanisms underlie this fundamental form of human imagination? The processes that underlie thinking about counterfactual situations may be similar to the processes that underlie thinking about factual situations (e.g., Byrne & Tasso, 1994;Johnson-Laird & Byrne, 1991). My aim is to sketch a putative framework within which to understand the nature of the human counterfactual imagination. I wish to argue that thinking about matters of fact and thinking about matters of possibility and impossibility are based on similar sorts of mental representations and cognitive processes. A. CONSTRAINTS ON COUNTERFACTUAL THINKING
What does the mind do when it generates a counterfactual scenario? An answer to this question requires a computational-level theory (e.g., Keane, Legdgeway, & Duff, 1994; Marr, 1982). The informational constraints on what can be thought counterfactually are governed by at least two key principles: counterfactual scenarios are recoverable from the factual scenario, and they are often goddriven to alter the outcomes of the factual scenario. Counterfactual scenarios need to be recoverable from the factual scenario on which they are based. The way in which the factual situation is understood and mentally represented is thus critical. Two people who witness a fatal shooting during a robbery may generate different “if only. . .” scenarios to mentally undo the outcome. One may think: (6) If only the bankrobber hadn’t panicked, the little boy would still be alive.
whereas the other may think: (7)
If only the little boy hadn’t tried to run away, he would still be alive.
Their representation of different aspects of the factual situation may depend on their location and attentional acuity, their knowledge and expectations,
Ruth M. J. Byrne
108
their current goals, and so on: the critical point is that their generation of a counterfactual scenario depends in important ways on their representation of the factual scenario. Their different counterfactual scenarios may lead to differentjudgments of blame, even to the extent of irrational attributions of blame to victims for their own fate (e.g., D. T. Miller & McFarland, 1987). Counterfactual scenarios depend on factual scenarios, and to explain the generation of counterfactual scenarios requires first an explanation of the mental representation of factual scenarios. Counterfactual scenarios must be grounded in the factual reality from which they depart because the essence of counterfactual thinking is the comparison between counterfactual and factual alternatives. Firm links between factual reality and the imaginary alternative must be maintained to ensure that the counterfactual and factual scenarios are mutually recoverable one from the other. As a result, alterations to reality tend to be minimal (Pollock, 1986) along the “joints” of reality (Kahneman & Tversky, 1982), where reality is at its most “slippable” (Hofstadter, 1979). Of course, some counterfactual alternatives may be more similar to the factual reality than others, or more accessible from it (Lewis, 1973; Stalnaker, 1968). The ease with which the route between the counterfactual scenario and the factual one can be traversed may be a function of the number or kind of alterations made to the factual scenario to construct the counterfactual one, and the shared knowledge of speaker and hearer (e.g., Grice, 1975). The generation of counterfactual scenarios is often goal driven to alter the outcomes of the factual scenario. Counterfactual thinking can be undirected, perhaps especially in daydreaming. “What i f . . .” thoughts, where an imaginary antecedent is constructed, for example: (8) If I were to win a million dollars
...
allow various alternative consequences to be explored in a loosely constrained way. Nonetheless, the counterfactual consequents may be directed toward altering factual outcomes, and consequents that leave the factual outcomes untouched seem unsatisfactory, for example: (9)
. . . I would leave the money in the bank and live exactly as I do now.
(see also Wells & Gavanski, 1989). Some “what i f . . .” thoughts have the goal simply of seeing whether the outcome would be similar or different to the current situation. For example, the speculation: (10) What if the allies had delayed the invasion by two weeks in World War II?
might prompt the reply: (11) They would still have won the war.
Cognitive Processes in Counterfactual Thinking
109
and the very word “still” indicates that the counterfactual scenario has not altered the outcome from the factual scenario (see also Kahneman & Varey, 1990). But, many counterfactual thoughts seem to be prompted by a desire to mentally undo an outcome, in particular, unhappy outcomes, for example: (12) I wish I hadn’t been mugged. Although counterfactual scenarios are generated by cognitive processes that are unconscious and automatic, the construction of appropriate antecedents can be driven by the goal set by the undone consequent, for example: (13) If I’d left work while it was still bright, I wouldn’t have been mugged. (e.g., Gleicher, et al., 1990; Roese, 1994). To understand fully what the mind does when it generates a counterfactual scenario, it is necesary to understand why it generates counterfactual scenarios. What is the purpose of counterfactual thinking? Its uses are thrown into sharp relief when one considers what life would be like without hypothetical thinking in general. Even posing the question requires counterfactual thought, such is the pervasive presence and allure throughout our mental life to engage in “subjunctive instant replays” (Hofstadter, 1979). Reports of individuals from other cultures who resist thinking hypothetically and demur from all but the most factual and empirical of inferences are fascinating anthropological conundrums that pose serious questions for our conceptions of rationality (Johnson-Laird & Byrne, 1993). Nonetheless, it may transpire that these data owe more to the social situation of crosscultural testing than to cognitive differences. But suppose there were individuals, say following specific brain lesions, who genuinely did not engage in counterfactual thinking (e.g., Knight & Grabowecky, 1990; Roese & Olson, 1995). Without counterfactual thought, individuals could not think suppositionally about alternatives and so they would be tied to thinking about factual matters only. They would have difficulty in learning from mistakes, they would fail to take alternative perspectives, and they would rarely daydream. They may entertain little ambition or hope, given that these features of mental life are based in part on a comparison of the actual situation with the way it might be improved in the future. For the same reasons, they would share little sense of progress or perfectibility. They may have very little sense of boredom, and for that matter, very little curiosity; they would not tend toward anxiety or worry, as these feelings emerge by comparing the actual situation with a potentially worse one, and they would probably rarely experience surprise, disappointment, or regret. As these conjectures suggest, life without counterfactual thought
110
Ruth M. J. Byrne
would be very different for us cognitively, socially, and emotionally. Counterfactual thinking seems central to the human condition.
B. COUNTERFACTUAL SCENARIO GENERATION
How does the mind generate counterfactual scenarios'? An answer to this question requires an algorithmic-level theory (e.g., Keane et al., 1994; Marr, 1982). The behavioral constraints on what can be thought counterfactually are governed by the detection of mutable aspects of reality, that is, aspects of reality that can be altered or mentally undone most readily. The mutability of aspects of reality may be guided by core categories of mental life, such as space, time, causality, intentionality, and so on (G. Miller & JohnsonLaird, 1976). For example, the causal relations between events affect their perceived mutability: people tend to undo the first cause in a causal sequence more than other causes (e.g., Wells et al., 1987). Likewise, the temporal relations between events also affect their perceived mutability: people tend to undo the most recent event more than earlier events in an independent (noncausal) sequence of events (D. T. Miller & Gunesagaram, 1990). The intentionality of the actions and events affects their perceived mutability. For example, people tend to regret their actions rather than their failures to act (e.g., Kahneman & Tversky, 1982), at least in the short term (Gilovich & Medvec, 1994). Moreover, they tend to undo actions within their voluntary control rather than actions that fall outside of their control (Girotto, Legrenzi, & Rizzo, 1991). They tend to undo the actions of a focal individual more than the actions of an individual in the background (Kahneman & Tversky, 1982), and they tend to undo actions that are unusual for an individual rather than routine actions (Kahneman & Tversky, 3982; see also Bouts, Spears, & van der Pligt, 1992; Gavanski & Wells, 1989), I return to examine these tendencies in more detail shortly. When people think counterfactually. they change some facts to create a new scenario. One possibility is that people find it easier to remove events from the factual scenario rather than to add new events (Kahneman & Tversky, 1982; see also Dunning & Parpal, 1989; Roese, 1994). The deletion of an event may be assisted by an availability heuristic, that is, some information related to the event may be retrieved and can then replace the original event in the counterfactual situation (Kahneman & Tversky, 1982). The most mutable aspect of the factual situation may be the aspect for which alternatives are readily available (Kahneman & Miller, 1986). But, many counterfactuals add a new additional event rather than subtract an event, at least in the case of negative outcomes (e.g., Roese & Olson, 1993; Sanna & Turley, 1996). It is critical to know what facts people start with and how
Cognitive Processes in Counterfadual Thinking
111
they have represented them, if a genuine theory of mutability is to explain and predict a priori the aspects of a factual scenario that people change when they think counterfactually. Toward the development of such a theory of mutability, I take as a starting premise the idea that the nature of the mental representations that reasoners construct of the factual situation may be an important constraint on the alterations that they make to it in order to construct a counterfactual scenario. To understand the generation of counterfactual alternatives to reality it is critical to understand the nature of mental representations of reality, and I turn now to a consideration of the representation of factual situations.
11. Mental Models and Counterfactual Thinking
My argument is that reasoners construct a particular sort of mental representation of factual situations-mental models-and the nature of the mental models they construct constrains the nature of the alterations they make to reality when constructing alternative scenarios. Mental models are mental representations that are close to the structure of the world rather than to the structure of the language that describes the world (JohnsonLaird, 1983). Models may contain as little information as possible explicitly, because of the limitations of working memory and so they may represent as much information as possible in an implicit form (Johnson-Laird & Byrne, 1991). Individuals normally represent explicitly only those situations that are true. To illustrate the sorts of models I take as my starting point, consider the following assertion about the contents of a blackboard, which you cannot see: (14) There is a circle on the blackboard or there is a triangle, or both. Three alternative models represent the three true possibilities: circle circle
triangle triangle
Separate models are represented on separate lines in the diagram: in the first model it is true that there is a circle; in the second model it is true that there is a triangle; in the third model it is true that there is both a circle and a triangle. The set of models represents only those possibilities that are true, that is, the situation where there is neither a circle nor a triangle is not represented in the set of models. The models also represent
Ruth M. J. Byrne
112
only those atomic propositions, affirmative or negative, in the premises that are true in the true possibilities: in the first model it is true that there is a circle, and it is false that there is a triangle, but individuals d o not normally represent this information explicitly. People may make a mental footnote to keep track of how to flesh out implicit information to be fully explicit (see Johnson-Laird & Byrne, 1991 for these more technical details), and they may lose track of these footnotes rapidly (e.g., Johnson-Laird & Savary, 1995). I now describe the models for factual and counterfactual conditionals. As will become clear, the models for both sorts of conditionals are similar but for one very important difference: models of counterfactual situations, unlike models for factual situations, must represent explicitly what is false.
A. MENTALMODELSOF CONDITIONALS Consider the following indicative conditional: (15) If it rained then the children played indoors. The assertion is consistent with a number of different situations, and so people must keep in mind a number of different mental models. The fully explicit set of models can be represented in the following diagram: rain not rain not rain
indoors not indoors indoors
where “rain” stands for “it rained,” “indoors” stands for “the children played indoors,” and “not” is a propositional-like tag to indicate negation (for details, see Johnson-Laird & Byrne, 1991; Johnson-Laird, Byrne, & Schaeken, 1992). The set of models corresponds to the three alternative situations in which the conditional could be true: in one it rained and the children played indoors, in the other two situations it didn’t rain, and in these cases, the children may or may not have played indoors. The models may contain information about who the children are, where they are, what the weather is usually like, what kinds of games they play, and so on, but the content of the models is not my current concern here; instead I wish to focus on their structure. Reasoners may interpret “if” to mean “if and only if,” and in that case they will consider that in the situation in which it didn’t rain the children did not play indoors, that is, their understanding of the assertion will correspond to the first two models only (see JohnsonLaird & Byrne, 1991). The set of models above is fully explicit-each of the three alternatives is represented. But people rarely construct fully explicit models because
Cognitive Processes in Counterfactual Thinking
113
of the constraints of working memory (e.g., Johnson-Laird & Byrne, 1991). Their initial representation is based on a set of models that contains some information implicitly, for example: rain
indoors
... The first model is fully explicit and corresponds to the atomic propositions in the assertion. The second model contains information in an implicit form, represented in the diagram by the three dots, as a “mental footnote” that there may be alternatives to the first model. The initial set of models must contain further footnotes to indicate how to flesh them out to be fully explicit (see Johnson-Laird & Byrne, 1991; Johnson-Laird et al., 1992). The processing assumptions of mental models reflect the constraints of working memory on the construction and revision of mental representations, and as a result, models represent some information explicitly and some information implicitly. The implicit information may be recovered and the models may be fleshed out to be fully explicit. These constraints lead to the prediction that the difficulty of an inference increases as a function of the number of alternatives that must be kept in mind. An inference is more difficult the more models it requires, and inferences that require models to be fleshed out are more difficult than inferences that can be made on the basis of the initial models. The suggestion that reasoners rely on these sorts of models to make deductions has been corroborated experimentally for a range of different sorts of deductions, including conditional inferences (e.g., Byrne & Johnson-Laird, 1992), propositional inferences in general (e.g., Johnson-Laird et al., 1992), spatial inference (e.g., Byrne & Johnson-Laird, 1989), reasoning with quantifiers (e.g., JohnsonLaird & Byrne, 1989; Johnson-Laird, Byrne, & Tabossi, 1989), metadeductive suppositions (Byrne & Handley, 1997; Byrne, Handley, & JohnsonLaird, 1995), and it has been modeled computationally (see JohnsonLaird & Byrne, 1991). People may reason about matters of possibility and impossibility in the same way as they reason about matters of fact: by constructing and revising mental models (e.g., Byrne & Tasso, 1997; Johnson-Laird & Byrne, 1991). As a first step toward illustrating this suggestion, I consider the case of counterfactual conditionals. B. COUNTERFACTUAL CONDITIONALS My suggestion is that reasoning about factual situations and reasoning about counterfactual situations rely on similar sorts of mental representations and
114
Ruth M. J. Byrne
processes. Reasoners make inferences from imaginary situations just as they do from factual situations, by constructing models (e.g., Byrne & Tasso, 1994, 1997). Let me demonstrate my argument by first considering the comparison of factual and counterfactual conditionals. The factual conditional: (16) If Oswald did not kill Kennedy then someone else did. seems to mean something very different from its corresponding counterfactual: (17) If Oswald had not killed Kennedy then someone else would have. (e.g., Lewis, 1973; see also Ayers, 1965; Barwise, 1986). Factual conditionals have received considerable philosophical and psychological scrutiny (for a review, see Evans, Newstead, & Byrne, 1993), but the psychology of counterfactual conditionals is largely undeveloped. The philosophical suggestion that counterfactual conditionals are supported by lawlike generalizations has guided much philosophical thought on the matter (e.g., Adams, 1975; Chisholm, 1946;Goodman, 1973). To appreciate some of the problematical issues surrounding counterfactual conditionals, consider the following example: (18)
If it had rained, the children would have played indoors.
The counterfactual conditional in (18) seems to say something about the situation in which it rained and the children played indoors, but it also seems to presuppose something else: that it did not in fact rain, and the children did not play indoors (see also Carpenter. 1973; Fillenbaum, 1074). For the comparable factual conditional in (15) earlier: (15')
If it rained then the children played indoors.
it is not apparent from the indicative mood whether or not it rained, and whether or not the childrcn played indoors. The presupposition that the antecedent and consequent of a counterfactual conditional are the opposite of what factually occurrcd is partly carried by the subjunctive mood and past tense of the counterfactual conditional: but counterfactuals can be conveyed even without these linguistic cues (e.g., Au, 1983; Comrie. 1986; Dudman, 1988). People may need to suspend their disbelief in the falsc antecedent when they understand a counterfactual, yet maintain consistency between this belief and others in their knowledge base (e.g., Ginsberg, 1986). Can a general theory of conditionals encompass conditionals that are based not only on factual situations but also on the various different sorts of hypothetical situations, such as past and present possibilities and impossibilities? The question has received different answers from different
Cognitive Processes in Counterfactual Thinking
1 I5
researchers interested in counterfactuals from the point of view of philosophy (e.g., Jackson, 1987: Pollock, 1986), psychology (e.g., Braine & O’Brien, 1991; Johnson-Laird & Byrne, 1991), linguistics (e.g., Dudman, 1988), and artificial intelligence (e.g., Ginsberg, 1986: Isard, 1974). Counterfactual conditionals do not seem to have a truth-functional account of their meaning (e.g., Johnson-Laird, 1986). They are not simply true when their antecedents are false or their consequents are true, as is logically the case for factual conditionals (e.g., Jeffrey, 1981). Although reasoners demur from some of the more paradoxical implications for factual conditionals that arise from such a truth-functional account, their reluctance may owe more to the content and context of the inference than their understanding of its essential import (e.g., Grice, 1975; Johnson-Laird & Byrne, 1991). When are counterfactual conditionals true and when are they false? The answer does not seem clear-cut. The question for factual conditions is somewhat more straightforward. Reasoners agree that a factual condition such as (15) earlier: (15’) If it rained then the children played indoors. is clearly true in the situation in which it rained and the children played indoors. It is false in the situation in which it rained and the children did not play indoors. The conditional does not assert that it definitely did rain, and so it may not have. In these cases, the children may have played indoors or they may not. In each of these two situations the conditional is true, on the “material implication” interpretation of the conditional in the propositional calculus (e.g., Jeffrey, 1981). The material implication interpretation may not correspond to the usage of “if” in everyday language, and one possible alternative is the “material equivalence” or biconditional “if and only if ’’ (e.g., Evans et al., 1993). On this account the conditional is also false in the situation in which it did not rain and the children played indoors. On either interpretation there is at least one clear situation where the factual conditional is true, and one situation in which it is false. And, the initial set of mental models of the factual conditional represents the first case explicitly. Because models, unlike truth tables, do not represent what is false, the second case is not represented mentally. The models of the conditional correspond to the situations in which it is true, although content and context can help flesh out or eliminate models. But this truth-based analysis of conditionals does not seem to work readily for counterfactuals (e.g., Lewis, 1973; Stalnaker, 1968). A counterfactual conditional such as (18) earlier: (18’) If it had rained, the children would have played indoors. may not seem to be true in the situation mentioned in the conditional, where the antecedent and the consequent are both true, that is, where it
Ruth M. J. Byrne
116
rained and the children played indoors. The counterfactual seems to suggest that it did not rain, and so it is not clear how such an assertion could be falsified; the clearly falsifying case, where it rained and the children did not play indoors, is ruled out by the presupposition that it did not rain. But people can distinguish between counterfactuals that seem plausible and ones that do not (Miyamoto & Dibble, 1986: Miyamoto, Lundell, & Tu, 1989) and so they cannot simply be relying on the truth of these components. One possibility is that a counterfactual is true if the consequent follows from the antecedent along with any relevant premises in the mental derivation constructed to prove a conclusion (e.g., Braine & O’Brien, 1991; Goodman, 1973; Ramsey, 1931). An alternative view is that a counterfactual is true if the consequent is true in the scenarios or models constructed by adding the false antecedent to the set of knowledge and beliefs it helps retrieve about the actual world-making any adjustments that are necessary to accommodate the antecedent in these models (e.g., Johnson-Laird & Byrne, 1991; Lewis, 1973; Stalnaker, 1968). 1 suggest that the experimental evidence favors the latter view, and I illustrate this point with the case of counterfactual deductions.
Counterfactual Deductions When reasoners understand a counterfactual conditional such as the one in (18) earlier: (18’) If it had rained, the children would have played indoors. they construct an initial representation in which some information is represented explicitly, and other information is represented implicitly, because of the constraints of working memory. They construct an initial set of models, which they annotate with propositional tags to keep track of the epistemic status of the different alternatives: Factual: Counterfactual:
not rain rain
not indoors indoors
They do not represent the alternative that is false given the counterfactual conditional, in which it rained and the children did not play indoors. One of the alternatives, the one in which it did not rain and the children did not play indoors, is presupposed by the counterfactual conditional to be the factual situation. Given this presupposition, the other two situations
Cognitive Processes in Counterfadual Thinking
117
TABLE I1
THEINITIAL A N D FULLY EXPLICIT SETSOF MODELSFOR A FACTUAL AND A COUNTERFACTUAL CONDITIONAL^ 1. If it rained then the children played indoors. Initial sets of models: rain indoors
... Fully explicit set: rain indoors no rain not indoors no rain indoors 2. If it had rained, the children would have played indoors. Initial sets of models: no rain factual: rain counterfactual:
Fully explicit set: factual: counterfactual:
"
n o rain rain no rain
not indoors indoors
not indoors indoors indoors
Adapted from Johnson-Laird & Byrne (1991).
are not, in fact, true. They are true only of the suppositional situation presented by the counterfactual conditional. Counterfactual conditionals require reasoners to keep in mind not only what is presupposed to be true, but also what is suppositionally true but factually false (e.g., Byrne & Tasso, 1994, 1997). as Table I1 shows. When people make inferences from factual conditionals, there is good reason to suppose that they construct mental models of the sorts outlined earlier (e.g., Johnson-Laird et al., 1992).For example, when they understand the premises of a modus tollens inference from the following factual conditional: (19) If Linda was in Dublin then Cathy was in Galway. Cathy was not in Galway.
they construct an initial model of the first premise:
Ruth M. J. Byrne
118
Dublin
Galway
... where “Dublin” represents the proposition “Linda was in Dublin” and “Galway” represents the proposition “Cathy was in Galway.” The model of the second premise is as follows: not Galway The procedures that combine models may fail to yield anything at this point, because the sets of models appear to contain nothing in common. Indeed, the most common error that reasoners make to this inference is to conclude that nothing follows (see Johnson-Laird & Byrne, 1991). A prudent reasoner fleshes out the models to be fully explicit: Dublin not Dublin not Dublin
Galway not Galway Galway
and then the combination of the model of the second premise rules out all but the second model: not Dublin
not Galway
The valid conclusion can be made: Therefore, Linda was not in Dublin. The modiis tollens inference is difficult because reasoners must flesh out their models and keep several alternatives in mind in order to make the deduction. But, on our account modus tollens should be easy to make from a counterfactual conditional: (20)
If Linda had been in Dublin then Cathy would have been in Galway. Cathy was not in Galway.
Reasoners first construct an initial model of the premises; Factual: Counterfact ual:
not Dublin Dublin
...
not Galway Galway
Cognitive Processes in Counterfadual Thinking
119
and they combine it with the model for the second premise: Factual:
not Galway
This time the models can be combined directly, with no need to flesh them out further. Reasoners can eliminate the counterfactual models, and retain the first model only: Factual:
not Dublin
not Galway
which supports the valid conclusion that Linda was not in Dublin. Hence, modus tollens should be easier from a counterfactual conditional than from
a factual conditional. The initial representation of the counterfactual conditional is more explicit than the one of the factual conditional. As a result, modus tollens can be made directly without any need to flesh out the set of models. In an experiment carried out in collaboration with Alessandra Tasso from the University of Padua, we found that only 42% of subjects made the inference from the factual conditional, whereas reliably more, 66%,made it from the counterfactual conditional (see Byrne & Tasso, 1997, for details). Consider now the following modusponens inference from a factual conditional: (21) If Linda was in Dublin then Cathy was in Galway. Linda was in Dublin.
Once again, reasoners construct an initial model of the first premise: Dublin
Galway
... The model of the second premise: Dublin can be combined readily with the first set of models and the combination eliminates the implicit model to leave the first model only: Dublin
Galway
Ruth M. J. Byrne
120
from which the conclusion can be reached: Therefore, Cathy was in Galway. A very similar process is required for the counterfactual conditional: (22) If Linda had been in Dublin then Cathy would have been in Galway. Linda was in Dublin.
Reasoners construct an initial set of models of the premises: Factual: Counterfactual:
not Dublin Dublin
not Galway Galway
and a model of the second premise: Factual:
Dublin
The combination of this model with the first set requires the elimination of some of the models, including the first model, and the updating of the status of the second model to represent an actual situation: Factual:
Dublin
Galway
This model supports the conclusion that Cathy was in Galway. The process of inference is essentially the same for both sorts of conditionals; the nzodus ponens inference can be made from the initial set of models without any need to flesh out the models to be more explicit. In the experiment carried out with Alessandra Tasso, we found that 94% of the subjects made the modus ponens inference from the factual conditional and 81%made it from the counterfactual and there was no reliable difference between them. The same pattern emerged for two further inferences, the denial of die antecedent inference from the minor premise: (23) Linda was not in Dublin. and the ufirmation of the consequent inference, from the minor premise; (24) Cathy was in Galway.
These inferences are fallacies on a “material implication” interpretation of the conditional-which is consistent with a set of three models-because the minor premise picks out two situations and so a single definite conclusion cannot be made. They are valid on a biconditional interpretation of “if”
Cognitive Processes in Counterfadual Thinking
121
as “if and only if,” which is consistent with a set of two models. Reasoners make the inferences when they fail to flesh out their models fully to the conditional interpretation. We predicted that the denial of the antecedent inference would be made more frequently from the counterfactual than from the factual conditional, for the same reasons outlined earlier for the modus tollens inference-because the counterfactual conditional is represented by a more explicit initial set of models. As we expected, we found that the denial of the antecedent inference was made reliably more often from the counterfactual conditional (59%) than from the factual conditional (39%). The affirmation of the consequent inference should be made equally often for both conditionals, again for analogous reasons to the ones outlined for the modus ponens inference, and we found no reliable difference between them (factual: 32%; counterfactual: 50%). The initial models of the counterfactual conditional are richer than the initial models of the factual conditional. As a result, people make more inferences from a counterfactual conditional than from a factual conditional, when the inferences are based on the richer representation, as occurs for modus tollens and the denial of the antecedent:
Factual: Counterfactual:
Modus Tollens 42% 66%
Denial of the Antecedent 39% 59%
People make as many inferences from a counterfactual conditional as from a factual conditional when the inferences are not based on the richer representation, as occurs for modusponens and the affirmation of the consequent:
Factual: Counterfactual:
Modus Ponens 94% 81%
Affirmation of the consequent 32% 50%
We have also found, in this experiment and in an earlier one, similar results for conditionals based not only on past possibilities, as these counterfactual conditionals are, but also for conditionals based on present possibilities, such as: (25) If Linda were in Dublin then Cathy would be in Galway. (see Byrne & Tasso, 1997, for full details). Reasoners think about factual and counterfactual conditionals by relying on the same mechanisms for the construction and revision of mental models.
122
Ruth M. J. Byrne
The representation of counterfactual conditionals requires reasoners to keep track of factual and hypothetical situations and to annotate their models accordingly. The more explicit representation of counterfactual conditionals leads to a greater variety in beliefs about the situations that can verify and falsify them. For example, most people believe that the situation that best verifies a factual conditional is the hypothesized one (e.g., Linda is in Dublin and Cathy is in Galway). In contrast, for the counterfactual conditional, some people believe that the situation that best verifies it is the hypothesized one, whereas other people believe that it is the factual one (e.g., Linda is not in Dublin and Cathy is not in Galway; see Byrne & Tasso, 1997). An explicit representation of the factual situation appears to occur for conditionals with negated components as well as affirmative ones-people make a similar pattern of inferences from counterfactual conditionals that contain negated atomic propositions, such as:
(26) If Mark had not gone to Regina, then Alicia would not have gone to Winnipeg. as they do from counterfactual conditionals that contain affirmative atomic propositions, as indicated by recent experiments carried out in collaboration with Valerie Thompson from the University of Saskatchewan (Thompson & Byrne, 1997). Because the initial set of models for counterfactual conditionals is more explicit than the initial set of models for factual conditionals, the two sorts of conditionals can seem to mean something very different from each other. Nonetheless, a single general theory of conditionals can encompass both factual and counterfactual conditionals (Johnson-Laird & Byrne, 1991). My suggestion is that reasoners reason about matters of possibility in the same way as they reason about matters of fact: by constructing models that correspond to the way the world would be if the assertion were true. The data described here provide some support for this suggestion by showing that inferences from counterfactual conditionals appear to rely on a similar mechanism to inferences from factual conditionals: reasoners make inferences more readily when they are supported by the initial set of models and do not require the models to be fleshed out. An important difference in thinking about factual matters and thinking about counterfactual matters is that the models constructed of factual situations represent explicitly what is true; but the models constructed of counterfactual situations represent explicitly not only what is true, but also what is presupposed to be false. The idea that the nature of the mental representations that reasoners construct of factual situations affects their ability to think and reason about these situations is supported by the results for inferences from factual conditionals (e.g., Johnson-Laird & Byrne, 1991). The idea that the nature
Cognitive Processes in Counterfactual Thinking
123
of the mental representations that reasoners construct of counterfactual situations affects their ability to think counterfactually receives similar support. The nature of the mental representations that reasoners construct provides a general framework for understanding their factual and counterfactual deductions, and I show next that it extends to understanding the generation of counterfactual scenarios. 111. Three Phenomena of Counterfactual Thinking
The mental stimulation that underlies counterfactual thinking may be based on the construction, revision, and fleshing out of mental models of the same sort that I have described here for deductive inference. Heuristics such as the availability heuristic (Kahneman & Tversky, 1982) may help provide the information that becomes explicitly represented in a model. For example, models may be fleshed out by the retrieval of norms during counterfactual thinking about scenarios in which exceptional events have occurred (Kahneman & Miller, 1986). The detection of what is mutable in a factual situation may depend on the nature of the representation of the factual situation (e.g., Byrne, 1996). As a result, what people tend to alter in the factual situation when they think about how it might have been different depends on what they have represented explicitly in their models of the factual situation (see also Legrenzi, Girotto, & Johnson-Laird, 1993). I suggest that “biases” of counterfactual thinking-for example, the tendency to undo actions more than inactions, or the tendency to undo the more recent event in an independent sequence-arise from the nature of the mental models that people construct, just as “biases” in deduction arise from the nature of the mental models that people construct (JohnsonLaird & Byrne, 1991). I sketch this account for three illustrative phenomena of counterfactual thinking: the action effect, the temporality effect, and a new spatial effect predicted by the model theory of counterfactual thinking. A. THEACTION EFFECT People tend to regret their actions more than their failures to act (Kahneman & Tversky, 1982). Consider the following scenario: (27) Mr. Paul owns shares in company A. During the past year he considered switching to stock in company B, but he decided against it. He now finds out that he would have been better off by $1200 if he had switched to the stock of company B. Mr. George owned shares in company B. During the past year he switched to stock in company A. He now finds that he would have been better
124
Ruth M.J. Byrne
off by $1200 if he had kept his stock in company B. Who feels greater regret? The scenario contains two explicit counterfactual conditionals, one about Mr. Paul’s inaction: (28) He now finds out that he would have been better off by $1200 if he had switched to the stock of company B. and one about Mr. George’s action: (29) He now finds that he would have been better off by $1200 if he had kept his stock in company B. Kahneman and Tversky (1982) found that 92% of their subjects believed that the individual who had acted, Mr. George, would feel greater regret. The action effect holds for good outcomes too. People feel better about their actions that lead to good outcomes than their inactions. Consider the following scenario, one of several used by Landman (1987): (30) Paul enrolled in Section 1 of Biology 101; his roommate enrolled in Section 2. At the beginning of the term, Paul considered switching to Section 2, but decided against it. The term is over and Paul just learned that he got an A in the course. His roommate got a B. George and his roommate enrolled in Section 2 of Biology 101. At the beginning of the term, George switched to Section 1. The term is over and George just learned that he got an A in the course. His roommate got a B in Section 2. Who felt better about his section choice, Paul or George?
67% of subjects judged that George, the individual who acted, would feel better about his section choice than Paul, the individual who did not act. However, the action effect for good outcomes is not as strong as the action effect for bad outcomes (Gleicher et al., 1990; Landman, 1987). Gleicher et al. (1990) used Kahneman and Tversky’s (1982) basic scenario about Mr. Paul and Mr. George and their stock to compare actions and inactions that led to bad outcomes and good outcomes for which the counterfactual alternative was made explicit or not (see also Wells & Gavanski, 1989).They also varied the order of the actions and inactions in the scenarios and found no order effect, that is, people are not simply undoing the more recent event, a common tendency in independent sequences of events to which I return in the next section. Most importantly, they examined a scenario with an explicit counterfactual, that is, Kahneman and Tversky’s original Mr. Paul and Mr. George scenario above, for which 91% of their subjects judged that Mr. George would feel worse, and they compared it
Cognitive Processes in Counterfactual Thinking
125
to one with no explicit counterfactual but a focus on the actual outcome instead: (31) Mr. Paul owns shares in company A. During the past year he considered switching to stock in company B, but he decided against it. He now finds out that he is worse off by $1200 because he kept his stock in company A. Mr. George owned shares in company B. During the past year he switched to stock in company A. He now finds that he is worse off by $1200 because he switched to stock in company A. Who feels worse? and they found an equally high percentage of subjects, 90%, judged that Mr. George would feel worse. The provision of an explicit counterfactual alternative has no effect for bad outcomes: Explicit counterfactual: No counterfactual:
91% action effect 90% action effect
They also examined a version of the scenario with a good outcome in which an explicit counterfactual is mentioned: (32) Mr. Paul owns shares in company A. During the past year he considered switching to stock in company B, but he decided against it. He now finds out that he would have been worse off by $1200 if he had switched to the stock of company B. Mr. George owned shares in company B. During the past year he switched to stock in company A. He now finds that he would have been worse off by $1200 if he had kept his stock in company B. Who feels better? and they found that 79% of subjects judged that Mr. George would feel better. However, for a version of the scenario with a good outcome but no explicit counterfactual: (33) Mr. Paul owns shares in company A. During the past year he considered switching to stock in company B, but he decided against it. He now finds out that he is better off by $1200 because he kept his stock in company A. Mr. George owned shares in company B. During the past year he switched to stock in company A. He now finds that he is better off by $1200 because he switched to stock in company A. Who feels better? only 58% of subjects judged that Mr. George would feel better. The provision of an explicit counterfactual alternative has an effect for good outcomes-the action effect is not equally strong in both scenarios:
Ruth M. J. Byrne
126
Explicit counterfactual: No counterfactual:
79% action effect 58% action effect
(Gleicher et al., 1990). The provision of an explicit counterfactual alternative enhances the action effect for good outcomes, in some cases even raising it to a similar level as the action effect for bad outcomes. The final aspect of the action effect to keep in mind is that although people regret their actions more than their failures to act in the short term, they appear to regret their failures to act more than their actions in the long term (Gilovich and Medvec, 1994,1995). Gilovich and Medvec (1994) carried out studies in which they asked subjects to think back over their own lives and to say, for example, what they regretted most. Their subjects tended to think of things they had failed to do, such as not pursuing hobbies, not spending enough time with family and friends, not “seizing the moment.” Their memories for events they regretted from their past lives showed an inaction effect, as did their judgments about other people’s feelings from a long-term perspective. Gilovich and Medvec (1994) demonstrated this reversal of the action effect to an inaction effect from a longterm perspective with the following scenario:
(34) Dave and Jim do not know each other, but both are enrolled at the same elite East Coast university. Both are only moderately satisfied where they are and both are considering transferring to another prestigious school. Each agonizes over the decision, going back and forth between thinking he is going to stay and thinking he will leave. They ultimately make different decisions: Dave opts to stay where he is, and Jim decides to transfer. Suppose their decisions turn out badly for both of them: Dave still doesn’t like it where he is and wishes he had transferred, and Jim doesn’t like his new environment and wishes he had stayed. Seventy-six percent of the subjects who were asked the question “Who do you think would regret his decision more upon learning that it was a mistake?” indicated that Jim, the individual who had acted, would feel more regret. But 63% of the subjects who were asked the question “Who do you think would regret his decision more in the long run?” thought that Dave, the individual who had failed to act, would feel more regret. In summary, the action effect comprises a set of interrelated phenomena: (1) People regret their actions more than their failures to act when they lead to bad outcomes; (2) They feel better about their actions more than their failures to act when they lead to good outcomes too, but the effect is weaker than for bad outcomes; (3) People feel more strongly about their
127
Cognitive Processes in Counterfactual W i g
actions rather than their failures to act for good outcomes just as much as for bad outcomes when the counterfactual alternative is made explicit; (4) Although people regret their actions more than their inactions in the short term, when they take a long-term perspective, they regret their inactions more than their actions. I suggest that an account in terms of the nature of the mental models that individuals construct of the factual situation and hence of the counterfactual situation may help explain each of these four tendencies. 1. Models of the Action Effect
One suggestion has been that the action effect arises because actions are perceived as departures from the status quo (Kahneman & Miller, 1986): Inaction maintains the status quo, whereas actions are a departure from its normality. According to Kahneman and Miller’snorm theory, departures from normality are more easily mutated than normal events because the abnormal event spontaneously recruits its corresponding norm. People may construct models that represent actions explicitly, and inactions implicitly. For example, Dave and Jim are both in college in Gilovich and Medvec’s scenario, and this fact may be represented in the initial models that individuals construct, along with the fact that they are both unhappy. The action of switching to a different college may be represented explicitly,whereas the inaction of staying in the same college may be represented only implicitly: Dave Jim
Factual: Factual:
in college A in college A switch to college B
unhappy unhappy unhappy
Working memory limitations may prevent people from being able to construct all the possible counterfactual models. People could, in principle, construct some possible counterfactual models: Dave
Factual: Counterfactual:
in college A switch to college B switch to college B
unhappy happy unhappy
Jim
Factual:
in college A switch to college B stay in college A
unhappy unhappy happy
Counterfactual:
The models represent the counterfactual possibilities that do not contradict
128
Ruth M. J. Byrne
the factual situation, that is, the models for Dave do not include the counterfactual possibility of him staying in college A and being happy because this counterfactual alters the outcome without altering any of the events that lead to it, and it directly contradicts the factual situation; likewise, the models do not include the counterfactual possibility for Jim of switching to college B and being happy, for the same reason. The critical point to note is that the representation of the factual situation for the individual who acted, Jim, requires multiple models, but there is only a single model required to represent the counterfactual situation. In contrast, the representation of the factual situation for the individual who failed to act, Dave, requires just a single model, but there are multiple models required to represent the counterfactual situations. Although people could in principle construct all possible counterfactual models, keeping track of multiple models in this way far exceeds the constraints of working memory. In practice, reasoners construct minimal models that represent as little information explicitly as possible. People do not think of every counterfactual alternative with equal ease (see JohnsonLaird & Byrne, 1991). The action effect reflects their capacity to construct just a partial fleshing out of the counterfactual models. They may be most able to flesh out the counterfactual situation for Jim because the action is represented explicitly and it is easier to mutate aspects of the factual situation that are represented explicitly in the models of the factual situation (see also Legrenzi et al., 1993). More models must be kept in mind of the factual situation for Jim’s action, which may make it harder t o generate a counterfactual scenario, but this difficulty is compensated for in an important way in this constrained task: multiple models of the factual situation help to rule out the counterfactual possibilities, and so the task of constructing the counterfactual alternatives is an easier one for the action. In contrast, multiple models must be constructed to consider the counterfactual alternatives for Dave’s inaction. Because people flesh out the counterfactual model for Jim, they judge that he will regret his action, not staying in college A, more than Dave will regret his inaction, staying in college A. The action effect results from the properties of the mental representations that people construct. First, they construct an initial set of models that represent as little information as possible explicitly: they mentally represent the action explicitly in their models and the inaction implicitly. Second, they do not construct all the possible counterfactual models. Third, aspects of the factual situation that are represented explicitly in models are easier to mutate than aspects that are not represented explicitly. Finally, the number of models that must be kept in mind to consider the counterfactual alternatives is an important constraint in counterfactual thinking, just as it is in factual thinking. These
Cognitive Processes in Counterfactual Thinking
129
simple principles that underlie the representation of the factual situation and the generation of a counterfactual situation based on it can account for the observation that people tend to regret their actions more than their failures to act. This account also explains the set of related phenomena comprising the action effect. Few people believe the actor will regret his action more than the individual who failed to act when the outcomes turn out to be good, if no explicit counterfactual alternative is given (Gleicher et al., 1990). Counterfactual thinking may be particularly driven to undo negative outcomes (e.g., Sanna & Turley, 1996) and in the absence of bad outcomes, the counterfactual models may not be fleshed out for either protagonist. In contrast, people believe the actor will regret his action more than the individual who failed to act when the outcomes turn out to be good, if an explicit counterfactual alternative is given (Gleicher et al., 1990). When people are given explicit counterfactual alternatives the alternatives are represented explicitly. Once they are prompted to think of the counterfactual negative outcome in this way, the action effect for good outcomes becomes as high as it is for bad outcomes. The action has been represented explicitly in the models of the factual situations, whereas the inaction has not (Gleicher et al., 1990). The provision of the explicit counterfactual alternative has no effect when the factual outcome is bad because these situations spontaneously bring to mind their good counterparts, perhaps because of the goal-driven nature of such tasks of counterfactual thinking. What of the final effect described earlier, that although people regret actions more than inactions in the short term, they regret inactions more than actions in the long term (Gilovich and Medvec, 1994, 1995)? Let us turn now to a closer examination of this effect.
2. When Do People Regret Failures to Act? People regret an action more than a failure to act because they represent an action explicitly in their models of the factual situation: there are more models of the factual situation for the action than for the inaction and accordingly there are fewer models of the counterfactual situation for the action than the inaction. It is well established that people can deal more readily with single models than with multiple models when they make deductions (e.g., Johnson-Laird & Byrne, 1991). It may even be the case that people find single-model counterfactual scenarios more compelling than multiple-model counterfactual scenarios. Actions may be more readily undone mentally because they give rise to a single counterfactual model instead of the multiple counterfactual models for an inaction. If the action effect arises because single-model counterfactual situations are easier than multiple-model counterfactual situations, why does it reverse
130
Ruth M. J. Byrne
to an inaction effect from a long-term perspective? It may be that the shortterm and long-term perspectives require subjects to consider whether the bad outcome from the factual situation was just one possible outcome or a necessary outcome, and whether the outcomes from the counterfactual situations are better, possibly or necessarily. It is easier to infer that a situation is possible (which depends on a single model that is an example) than that it is not possible (which depends on all models), whereas it is easier to infer that a situation is not necessary (which depends on a single model that is a counterexample) than that it is necessary (which depends on all models; Bell & Johnson-Laird, 1997). From a short-term perspective, reasoners may need to construct just a single good-outcome counterfactual alternative to appreciate that the situation could have been different. The short-term perspective may be geared toward learning from mistakes and toward the immediate reparation of bad outcomes, and so it may depend on finding just one good alternative counterfactual situation. In contrast, from a long-term perspective, constructing a counterfactual scenario to a distant bad event may require reasoners to construct multiple counterfactual alternatives to examine which ways the situation would have turned out differently. The long-term perspective may be geared toward a major overhaul of life events, and so it may depend on finding alternative counterfactual situations. From the long-term perspective, when we think about the bad events of our lives, we may feel compelled to leave no stone unturned and no counterfactual model unexamined in our attempt to see whether our lives could have been better. Another possibility is that the temporal proximity of the actions in the short term makes them mutable, just as, for example, it is easier to imagine catching an airplane that one has missed by 5 minutes than an airplane that one has missed by an hour (e.g., D. T. Miller & McFarland, 1987). From the long-term perspective, the temporal distance of our actions may make them seem immutable, as part of the historical presupposed background, much as a perpetrator’s actions presented in the background are assumed to be immutable relative to a victim’s (e.g., Kahneman & Varey, 1990 D. T. Miller & McFarland, 1987).This temporal closeness and distance may affect actions because they are represented explicitly: it may not have the same impact on inactions, which are represented implicitly, and so, over time, inactions loom larger. Gilovich and Medvec (1994) consider several independent, parallel explanations of the reversal of the action effect to an inaction effect in the long term, including the possibility that people feel more responsible for their actions. When actions lead to bad outcomes they engage in more mental work to reduce the “dissonance” of their interpretation of the outcomes as wholly bad: they seek out the silver linings in clouds that arise from their actions more than they do in
Cognitive Processes in Counterfactnd Thinking
131
ones that arise from their failures to act. For example, in the long term, the outcomes from the action carried out by Jim are no longer considered to be wholly bad: people make an attempt to look on the bright side and find some positive outcomes that followed from the action. They do not carry out this mental work for failures to act and so the models for the events surrounding Dave, the individual who failed to act, do not change over time. Because counterfactual thinking is goal driven to undo negative outcomes in this sort of task, the only events to regret from the long-term perspective are those surrounding Dave-only these events result in a bad outcome in the long term. Is the reversal of the action effect in the short term to an inaction effect in the long term a general phenomenon? For example, does it also occur when we think about events that turned out well? In a series of experiments carried out with Alice McEleney and Patrick McAlinney from Dublin University we examined whether there is a reversal to an inaction effect in the long term for good outcomes as well as for bad outcomes. In one experiment, we gave four groups of subjects one of four sorts of scenarios based on Dave and Jim and their college choices: a short-term, bad-outcome version, a long-term, bad-outcome version, a short-term, good-outcome version, and a long-term, good-outcome version. We expected to replicate Gilovich and Medvec’s reversal of the action effect to an inaction effect over time for bad outcomes; we also expected to observe an action effect for good outcomes in the short term because we drew the subjects’ attention to the counterfactual possibilities implicitly with the assertion, “each agonizes over the decision, going back and forth between thinking he is going to stay and thinking he will leave” (e.g. Byrne, & McEleney, 1997). The critical point of the experiment was to examine what happens for good outcomes from a long-term perspective. We replicated Gilovich and Medvec’s reversal of the action effect in the short term to an inaction effect in the long term using a version of their scenario altered in several ways to clarify the finality of the decision, and especially its long-lasting impact: (35) John and Paul do not know each other, but both are enrolled at the same elite university. Both are only moderately satisfied where they are and both are considering transferring to another prestigious university. They are both told they must make their final decision before the end of the year. Each agonizes over the decision, going back and forth between thinking he is going to stay and thinking he will leave. They ultimately make different decisions: John opts to stay where he is, and Paul decides to transfer. Suppose their decisions turn out badly for both of them: at the end of the year, John is even more unhappy where he is and wishes
Ruth M. J. Byrne
132
he had transferred, and Paul is even more unhappy at his new college and wishes he had stayed where he was. However, it is too late for either of them to reverse his decision. As a result they both drop out of college and neither of them ever secures a good job. Who do you think would regret his decision more on looking back on it ten years later? Seventy-six percent of the 46 Dublin University undergraduates who completed this long-term, bad-outcome version indicated that John, the individual who had failed to act, would feel more regret. Only 24% judged that Paul, the individual who acted, would feel more regret. The short-term version continued from the second paragraph in (35) as follows:
(36)
. . . Suppose their decisions turn out badly for both of them: at the end of the year, John is even more unhappy where he is and wishes he had transferred, and Paul is even more unhappy at his new college and wishes he had stayed where he was. However, it is too late for either of them to reverse his decision. Who do you think would regret his decision more on learning it was a mistake?
Sixty-one percent of a second group of 46 subjects who received this shortterm, bad-outcome version indicated that Paul, the individual who acted, would feel more regret. We observed the standard action effect for bad outcomes in the short term, and it reversed to an inaction effect in the long term:
Short term: Long term:
61% action effect 24% action effect
The other two versions of the task had good outcomes instead. For example, the long-term, good-outcome version continued with the second paragraph as follows: (37)
. . . Suppose their decisions turn out well for both of them: at the end of the year, John is happier where he is and is glad he stayed where he was, and Paul is happier at his new college and is glad he transferred. They both do very well at college and secure good jobs after graduating. Who do you think would feel better about his decision looking back on it ten years later?
Seventy-five percent of the 51 subjects who received the long-term, goodoutcome version indicated that Paul, the individual who acted, would feel better about his decision; 70% of the 47 subjects who received the short-
Cognitive Processes in Counterfactual Thinking
133
term, good-outcome version also indicated that Paul, the individual who acted, would feel better about his decision. We observed the standard action effect for good outcomes in the short term, but no reversal to an inaction effect in the long term:
Short term: Long term:
70% action effect 75% action effect
In the short term people regret their actions when they lead to a bad outcome and in the long term they regret their failures to act when they lead to a bad outcome; but in both the short term and the long term they feel better about their actions than their failures to act when they lead to a good outcome. The results suggest that the reversal to an inaction effect in the long term for bad outcomes is not the result of temporal proximity. According to the temporal proximity account, there should be a reversal to an inaction effect for good outcomes, too: temporal distance should cause actions to fade into the background for good outcomes just as much as for bad outcomes, and so inactions would become relatively more mutable for good outcomes, too. But if the effect depends on the number of models people keep in mind, then it is perhaps unsurprising that there is no reversal to an inaction effect in the long term for good outcomes. From the long-term perspective, bad events may lead us to leave no stone unturned to see how things could have been better, but when we think about the good events of our lives, we may be content to find a single counterfactual model in which things turned out worse. As a result, an action effect occurs for good outcomes in both the short term and the long term because a single counterfactual model to the action is easier to construct. In summary, the action effect comprises a set of four core phenomena that I suggest can be explained by the following principles: people construct an initial set of models that represent as little information as possible explicitly: they mentally represent the action explicitly in their models and the inaction implicitly. Reasoners do not construct all the possible counterfactual models in their initial set of models. Aspects of the factual situation that are represented explicitly in models are easier to mutate than aspects that are not represented explicitly. The representation of actions requires multiple models of the factual situation, and accordingly, a single model of the counterfactual situation, whereas the representation of inactions requires a single model of the factual situation, and accordingly, multiple models of the counterfactual situation. I now show how this model theory of counterfactual thinking can be extended to account for a second
134
Ruth M. J. Byrne
phenomenon of counterfactual thinking, the tendency people exhibit to undo recent events.
B. THETEMPORALITY EFFECT People tend to undo the more recent event in an independent sequence of events. D. T. Miller and Gunasegaram (1990, p. 1111) presented subjects with the following coin-toss scenario: (38) Imagine two individuals (Jones and Cooper) who are offered the following very attractive proposition. Each individual is asked to toss a coin. If the two coins come up the same (both heads or both tails), each individual wins E1000. However, if the two coins do not come up the same, neither individual wins anything. Jones goes first and tosses a head; Cooper goes next and tosses a tail. Thus, the outcome is that neither individual wins anything.
Over 80% of their subjects judged that it was easier to undo the outcome by thinking of Cooper tossing a head, rather than Jones tossing a tail. They judged that Cooper would experience more guilt, and would tend to be blamed more by Jones. Kahneman and Miller (1986) found that when subjects were given an ordered sequence of letters (e.g., x f ) on a computer screen and asked to quickly replace one, they tended to mutate the second letter in the sequence. Moreover, D. T. Miller and Gunasegaram (1990) showed that people tended to undo the more recent event more often when they considered how a student who failed an examination might have passed. Their subjects undid the list of examination questions when they believed the questions had been set after the students had prepared for the examination; they undid the nature of the students’ preparation when they believed it had been done after the questions were set (that is, they undid the second event in each case). Miller and Gunasegaram suggest that early events tend to be presupposed or taken for granted more than later events and the temporality tendency may play a role in many everyday judgments, such as the tendency for blackjack players to be averse to playing on the last box, for teams to sport their faster runner last in a relay race, and for people to wager more on their predictions than their postdictions. 1. Mental Models and the Temporality Effect
In the coin-toss scenario, earlier events in a sequence provide the context against which subsequent events are interpreted. The second event, Cooper tossing a tail, is interpreted in the context of the first event, Jones tossing a head, and the interpretation results in knowing that Cooper did not toss heads also. I suggest that the reason the earlier event is presupposed is
Cognitive Processes in Counterfactual Thinking
135
that the earlier event initializes the model, that is, it provides the cornerstone of its foundation. Reasoners may construct a representation of the cointoss scenario in which they explicitly represent the situation that is described where each individual and their coin-toss outcome is represented in the set of models, and the counterfactual possibilities are implicit in the models: Factual: Counterfactual:
Jones heads
Cooper tails
...
Cooper’s toss of tails may even be interpreted in the context of Jones’s toss of heads to the extent that it is represented as its negation. Factual: Counterfactual:
Jones heads
Cooper not heads
...
Models may be anchored just as they are in numerical domains, by the earlier information. Consider, for example, the task of estimating quickly the answer to 8 X 7 X 6 X 5 X 4 X 3 X 2 X 1. Subjects produce larger estimates than those produced by subjects asked to estimate the answer to 1 X 2 X 3 X 4 X 5 X 6 X 7 X 8 (Tversky & Kahneman, 1982). The earlier information in the numerical sequences anchors their judgments. Similarly, people initialize their models of the coin-toss scenario by the first event, and subsequent information is integrated into the model against this backdrop. Subsequent events may modify the context, or require some adjustment from the anchor (Byrne, Segura et al., 1997; Tversky & Kahneman, 1982). Models may be continually changing to deal with new situations, and so the cornerstone of a new situation may help to initialize a new model. The counterfactual possibilities in the coin-toss scenario may be fleshed out to be fully explicit, and they may be annotated to indicate the outcomes: Factual: Counterfactual:
Jones heads Jones heads Jones not heads Jones not heads
Cooper not heads Cooper heads Cooper not heads Cooper heads
lose win win lose
But once again, the temporality effect indicates that people do not consider all counterfactual alternatives explicitly, and instead they flesh out their models to be consistent with just one of the alternatives:
Ruth M. J. Byrne
136
Factual: Counterfactual:
Jones heads Jones heads
Cooper not heads Cooper heads
...
Because of the goal-driven nature of counterfactual thinking the counterfactual alternative in which the players both lose (the last in the fully fleshed out set) is not readily constructed as it does not undo the outcome of losing. The temporality effect arises because models of an independent sequence of events are initialized or contextualized by the first event against which subsequent events are interpreted. When reasoners undo the outcome, they flesh out the counterfactual model in which the context or initializing event is left untouched. Some events are conferred the status of initializing the model, to provide an integrating function that constrains the interpretation of subsequent events inserted into the model. In the absence of knowledge about, for example, the normality of the situation, the first event in an independent sequence of events is designated the context against which subsequent events are interpreted. The effect arises because of the need to integrate information into a coherent representation. The role of representational integration has been demonstrated clearly in the primary domains of deduction. In particular, combining incoming information adequately with information already represented has been shown to be critical to each of the main sorts of deduction: relational inference (e.g., Maybery, Bain, & Halford, 1986), syllogistic inference, (e.g., Johnson-Laird & Bara, 1984; Johnson-Laird & Steedman, 1978), and propositional inference (e.g., Byrne, 1989a, 1989b).The necessity for representational integration ensures that some elements within a model are immutable. The initializing events in a model are less mutable than other events represented in the model, which may help to ensure that the recoverability constraint alluded to earlier is satisfied: mutations to the factual model are minimal in that they do not tamper with critical aspects of the factual model. 2.
Context Effects in the Temporality Effect
Our suggestion is that the more recent event is more available for mutation because the earlier event provides the context against which the subsequent event is interpreted (e.g., Byrne, Culhane, & Tasso, 1995; Byrne et al., 1997). In an experiment with Ronan Culhane from Dublin University and Alessandra Tasso from the University of Padua, we have shown that it is possible to eliminate the temporality effect by decoupling or separating the contextualizing role of the first event from its position in the sequence and its contribution to the outcome. In one experiment we provided a separate contextualizing event, prior to the two target events, and found
Cognitive Processes in Counterfadual Thinking
137
that both the first and second target event were equally mutable. In the experiment we provided a separate contextualizing event by devising a scenario in which the players take part in a game-show in which there is a technical hitch: (39) Imagine two individuals (Jones and Brady) who take part in a television game-show, on which they are offered the following very attractive proposition. Each individual is given a shuffled deck of cards, and each one picks a card from their own deck. If the two cards they pick are of the same color (i-e., both from black suits or both from red suits) each individual wins 21000. However, if the two cards are not the same color, neither individual wins anything. Jones goes first and picks a black card from his deck. At this point, the game-show host has to stop the game because of a technical difficulty. After a few minutes, the technical problem is solved and the game can be restarted. Jones goes first again, and this time the card that he draws is a red card. Brady goes next and the card that he draws is a black card. Thus, the outcome is that neither individual wins anything. The first player’s prehitch choice initializes the model, and subsequent choices are interpreted in the light of this context. For example, Jones’s subsequent posthitch choice of red may be interpreted as not-black: Factual: Counterfactual:
Jones not black
...
Brudy black
The conflict between, on the one hand, the immutability of the initializing event, Jones’s selection, which renders it less likely to be fleshed out in the counterfactual models and, on the other hand, the negation, not-black, which may call to mind its affirmative counterpart (e.g., Johnson-Laird & Byrne, 1991), which renders it more likely to be fleshed out, cancels out the temporality effect or any reversal of it. Subjects who were given this “different-context’’ scenario showed no temporality effect in their completions of the sentence: “Jones and Brady could each have won 21000 if only one of them had picked a different card, for instance i f . . . .” Forty-four percent of the 36 Dublin University undergraduates given this version of the scenario focused on the second event only, or mentioned it first, and almost as many, 42% of the subjects, focused on the first event only, or mentioned it first. Their judgments of emotions followed suit, for example, 44% of the subjects judged that Brady, the second player, would feel more guilt, and almost as many, 31%, judged that Jones would feel more guilt.
Ruth M. J. Byrne
138
These results contrast with the results from the second group of 39 subjects given a “same-context’’ scenario in which the context was the same as the first target play: (40) Imagine two individuals (Jones and Brady) who take part in a television game-show, on which they are offered the following very attractive proposition. Each individual is given a shuffled deck of cards, and each one picks a card from their own deck. If the two cards they pick are of the same color (i.e., both from black suits or both from red suits) each individual wins f1000. However, if the two cards are not the same color, neither individual wins anything. Jones goes first and picks a black card from his deck. At this point, the game-show host has to stop the game because of a technical difficulty. After a few minutes, the technical problem is solved and the game can be restarted. Jones goes first again, and this time the card that he draws is a black card. Brady goes next and the card that he draws is a red card. Thus, the outcome is that neither individual wins anything. In this situation 59% of subjects undid the second event overall, reliably more than the 23% of subjects who undid the first event overall. Their judgments of emotion followed suit: for example, 77% of the subjects judged that Brady, the second player, would feel more guilt, reliably more than the 10% who judged that Jones would feel more guilt. The results show that the temporality effect arises because of the contextualizing role of the first event, which gives rise to a perception of the immutability of the first event and a tendency to represent subsequent events with reference to the first event. The effect is eliminated when this initializing role is decoupled from the serial position of the event contributing to the outcome. The elimination occurs for the completion of “if only” sentence stems: Different context
Same context
42% 44%
23% 59%
Different context
Same context
31% 44%
10% 17%
First event overall Second event overall
and it also occurs for judgments of guilt:
First player Second player
Cognitive Processes in Counterfactual Thinking
139
The elimination of the temporality effect in the experiment described here rules out the explanation of the temporality effect that the more recent event is “fresh” in working memory, or more available to a backward search through the entries. D. T. Miller and Gunasegaram briefly consider the possibility that “later events in a temporal sequence may be assigned more causal responsibility because they are more available in memory” (1990, p. 1117). It is a plausible suggestion that people mutate events that lead to the outcome by using a backward search strategy that encounters the most recent entry first (e.g., Wells et al., 1987). However, if the temporality effect arose because the more recent event was more available in working memory, then the effect should continue to be observed even in the different-context scenario: the second player’s play is most recent in either version of the story. Moreover, we have also found that there is a temporality effect in thinking about good outcomes and it can be eliminated in the same manner as for bad outcomes (see Byrne et al., 1997, for details). In another experiment carried out with Susana Segura and Pablo Berrocal from the University of Malaga, and Patrick McAlinney from Dublin University, we have shown that the temporality effect depends on the structure of the events in the world, not the structure of the language used to describe the world (Byrne et al., 1997). We gave subjects scenarios in which the first event to occur in the world was also the first to be described in the scenario and the second event to occur in the world was also the second to be described in the scenario. An example of one of the three scenarios given to the subjects is as follows: (41) Imagine two individuals who are offered the following proposition. Each individual is given a sack of marbles, and each one draws a marble from her own sack. If the two marbles they draw are of the same color (i.e., both are blue or both are white) each individual wins 21000. However, if the two marbles are not the same color, neither individual wins anything. Anne had her turn and drew a blue marble from her sack; after her, Joan had her turn and drew a white marble from her sack. Thus, the outcome is that neither individual wins anything.
The results reveal the standard temporality effect: 50% of the completions for the sentence “They could each have won flOOO if only one of them had picked a different card, for instance if. . .” for the 20 Dublin University undergraduates given this version of the scenarios focused on the second event, whereas only 37% of their completions focused on the first event. Their judgments of emotions also support this trend: 73% of the subjects judged that Joan, the second player, would feel more guilt, more than the 5% who judged that Anne would feel more guilt. We gave another group
Ruth M. J. Byrne
140
of 20 subjects versions of the three scenarios in which the first event to occur in the world was the second to be described in the scenario and the second event to occur in the world was the first to be described in the scenario, as follows: (42) Imagine two individuals who are offered the following proposition. Each individual is given a sack of marbles, and each one draws a marble from her own sack. If the two marbles they draw are of the same color (i.e., both are blue or both are white) each individual wins f1000. However, if the two marbles are not the same color, neither individual wins anything. Joan had her turn and drew a white marble from her sack; before her, Anne had her turn and drew a blue marble from her sack. Thus, the outcome is that neither individual wins anything.
In this case 48% of their completions focused on the second event, even though it was mentioned first in the scenario, whereas only 28% focused on the first event, even though it was mentioned second. Intriguingly, their judgments of emotions show no preference: 27% of subjects judged that Joan, the second player, mentioned first, would feel more guilt, and as many, 33%,judged that Anne, the first player, mentioned second, would feel more guilt. In summary, people undo the second event to occur in the world even if it is mentioned first:
First-mentioned event Second-mentioned event
After
Before
37% 50%
48% 28%
Their judgments of guilt follow suit for the standard “after” scenario but the effect is eliminated for the “before” scenario:
After
Before ~
First player Second player
5% 73%
27% 33%
The results of the experiment show that reasoners undo the second event regardless of whether it is mentioned first or second. The finding is consistent with our proposal that models correspond to the structure of the world
Cognitive Processes in Counterfactual Thinking
141
rather than to the structure of the language used to describe the world (Johnson-Laird & Byrne, 1991). In summary, people have a tendency to undo the more recent event in an independent sequence of events (D. T. Miller & Gunasegaram, 1990; Sherman & McConnell, 1996). This temporality effect arises because reasoners construct models that correspond to the world rather than to the language used to describe the world. They initialize or contextualize their models of an independent sequence of events by the first event, against which subsequent events are interpreted. Reasoners undo the outcome by fleshing out the counterfactual model in which the initializing event is left unaltered. Some events provide an integrating function in the set of models and this status results in some events that are immutable. The initializing event is the event that occurred first in the world, regardless of whether it occurred first in the description of the world. I turn now to the final phenomenon of counterfactual thinking that I address in this article-a new phenomenon predicted by the model theory.
C. THESPATIAL EFFECT One of the primary issues in counterfactual thinking is to establish the perceived “joints” of reality (Kahneman & Tversky, 1982), where reality is at its most “slippable” (Hofstadter, 1985). The joints of reality may correspond to core categories of mental life, such as time, space, causality, and intentionality (G. Miller and Johnson-Laird, 1976). People may tend to mutate their representations of reality along these fault lines, as demonstrated by the existence of a temporality effect (D. T. Miller & Gunasegaram, 1990), the tendency to mutate the more recent event in an independent sequence; a causality effect (Wells et al., 1987),the tendency to mutate the first event in a causal sequence; various intentionality effects, including an action effect (Kahneman & Tversky, 1982), the tendency to mutate actions rather than inactions; and a controllability effect (Girotto et al., 1991), the tendency to mutate actions within an individual’s control rather than outside the individual’s control. If people do mutate their representations of factual reality along these fault lines, then clearly there should be spatial mutability effects. Are there any effects of spatial mutability? The question remains largely unanswered. I turn now to an examination of the mutability of space and to some novel predictions made by the model theory of counterfactual thinking that I have sketched in this article. Spatial aspects of reality seem to be mutable, at least when they are exceptional. Kahneman and Tversky (1982) found that people tend to undo what is exceptional rather than what is routine. They demonstrated this phenomenon in a scenario in which they varied the exceptionality of either
142
Ruth M. J. Byrne
the route taken by an individual or the time at which the individual took it, as follows: (43) Mr. Jones was47 yearsold, the father of three, and a successful banking executive. His wife had been ill at home for several months. On the day of the accident, Mr. Jones left his office at the regular time. H e sometimes left early to take care of home chores at his wife’s request, but this was not necessary on that day. Mr. Jones did not drive home by his regular route. The day was exceptionally clear and Mr. Jones told his friends at the office that he would drive along the shore to enjoy the view. The accident occurred at a major intersection. The light turned amber as Mr. Jones approached. Witnesses noted that he braked hard to stop at the crossing, although he could easily have gone through. His family recognized this as a common occurrence in Mr. Jones’ driving. As he began to cross after the light changed, a light truck charged into the intersection at top speed, and rammed Mr. Jones’ car from the left. Mr. Jones was killed instantly. It was later ascertained that the truck was driven by a teenage boy, who was under the influence of drugs. As commonly happens in such situations, the Jones family and their friends often thought and often said, “if only. . .” during the days that followed the accident. How did they continue this thought? In the time version of the scenario, the following insertion was made: (44) On the day of the accident, Mr. Jones left the office earlier than usual, to attend to some household chores at his wife’s request. He drove home along his regular route. Mr. Jones occasionally chose to drive along the shore, to enjoy the view on exceptionally clear days, but that day was just average. Fifty-three percent of the 62 subjects who were given the exceptional-route version of the scenario undid this spatial aspect of the scenario, compared to only 13% of the 61 subjects who were given the normal-route version of the scenario. Twenty-six percent of the subjects given the exceptionaltime version undid the time at which Mr. Jones left the office, compared to only 3% of subjects who were given the normal-time version (Kahneman & Tversky, 1982):
Exceptional route, normal time Exceptional time, normal route
Undo route
Undo time
53% 13%
3% 26%
Cognitive Processes in Counterfactual T h i i n g
143
The conceptual proximity to a goal also has an effect on the mutability of the events. Proximity may be established temporally, as in the example outlined earlier, in which it is easier to imagine catching an airplane that one has missed by 5 minutes, than an airplane that one has missed by an hour. Proximity can also be established spatially, for example, it is easier to imagine someone being saved when they died just 5 miles from help rather than when they died 75 miles from help (e.g., Kahneman & Varey, 1990; D. T. Miller & McFarland, 1987). I report the results of an experiment on the mutability of space, which provides a test of the basic tenets of the model theory of counterfactual thinking that I have tried to advance in this article. A core prediction of the model theory is that inferences that require multiple models to be kept in mind are harder than inferences that require single models to be kept in mind. Problems that require multiple models exceed the limitations of working memory, and reasoners make more errors and take longer to solve them, and this prediction has been corroborated in the main domains of deduction (e.g., Johnson-Laird & Byrne, 1991). It follows that a model theory of counterfactual thinking will predict that, all other things being equal, counterfactual scenarios that require multiple models to be kept in mind will be harder to construct than counterfactual scenarios that are based on a single model. One example where multiple models lead to more errors than single models is the domain of spatial inference. I will outline these effects in spatial deduction first, and then show how they apply in counterfactual thinking about spatial arrangements. Consider the following description of the layout of three objects: (45) The tree is behind the postbox. The tree is in front of the lamppost. The spatial description is determinate and it requires a reasoner to keep a single model in mind, which can be captured in the following diagram, from a bird’s-eye view: lamppost tree postbox Compare the description in (45) above to the following description: (46) The tree is behind the postbox. The postbox is in front of the lamppost.
The indeterminacy in the description in (46) renders it compatible with at least two alternative models:
Ruth M. 1. Byrne
144
lamppost tree postbox
tree lamppost postbox
An indeterminate spatial description requires multiple models to be kept in mind, and working memory limitations make it difficult to keep multiple models in mind (e.g., Johnson-Laird & Byrne, 1991). The determinacy and indeterminacy of a spatial description has a profound impact on comprehension and reasoning (see Chapter 6 in Evans et al., 1993, for a review). People remember descriptions of spatial layouts better in terms of their gist when they are consistent with a single model than with multiple models (Mani & Johnson-Laird, 1982; see also Ehrlich & Johnson-Laird, 1982). The number of models also affects the accuracy of their deductions, when they are asked what relation, if any, they can infer between the tree and the lamppost in problems such as the examples in (45) and (46). They make more correct deductions when the description is consistent with a single model that has a valid conclusion (69%), than when the description is consistent with multiple models and has no valid conclusion (19%, see Byrne & Johnson-Laird, 1989, for details). One can construct a counterfactual alternative to the spatial layouts described in the examples earlier. For both the determinate and the indeterminate spatial descriptions, there are many, many counterfactual alternatives that could be constructed by altering one, two, or all three of the objects (e.g., deleting the tree), and one or both of the relations (e.g., transposing the tree and the lamppost). For the determinate spatial description, there is a single model of the factual situation, and multiple models of the counterfactual possibilities. For the indeterminate spatial description, there are multiple models of the factual situation, and multiple models of the counterfactual possibilities. (Unlike the domain considered earlier, based on an action and an inaction such as Dave and Jim’s college choices, multiple models of the factual situation do not result in an appreciable reduction in the number of counterfactual models.) Thinking about a factual situation is easier when the situation is a determinate one for which just a single model needs to be kept in mind. I have argued that thinking about matters of fact and thinking about matters of possibility depend on the same sorts of representations and processes, and so the prediction is clear: thinking counterfactually should be easier when it can be based on situations that are determinate, and that require a single model to be kept in mind. It should be more difficult to generate one of the multiple counterfactual possibilities when the factual situation is consistent with multiple models, than when it is consistent with a single
Cognitive Processes in Counterfactual Thinking
145
model (see Byrne, 1997, for details). Given a single-model description, individuals must keep the single model of the factual situation in mind, make alterations to it, and generate counterfactual alternatives on the basis of these alterations. Given a multiple-model description, individuals must keep at least two alternative models of the factual situation in mind. The limits of working memory may make it difficult to make alterations to the models, and to generate other models as the counterfactual alternatives. As a result, the prediction is that counterfactual mutations of multiplemodel factual descriptions should be of a more simple sort than counterfactual mutations of single-model descriptions. The aim of the experiment was to test this prediction. The experiment relied on two sorts of vignettes, one with a determinate spatial layout and one with an indeterminate spatial layout. One group of 21 undergraduates from Dublin University received the following determinate vignette: (47) Joanne drove home from work one cold wintry night. She was listening to the radio as she drove along a familiar stretch of road. As she scanned the bend ahead she saw a postbox on the left-hand side of the road. There was a large tree behind the postbox. The tree was in front of a silver lamppost. She slowed approaching the bend and it was then that the accident happened. She hit a large patch of black ice and lost control of her car. She hit the postbox first. By an unlucky coincidence more ice caused the car to spin and she hit the tree. Once again by a freak coincidence she spun on more ice and hit the lamppost. She sustained multiple injuries including a broken leg and a fractured collarbone. In the weeks she spent recovering in hospital she often became upset thinking about the accident and thinking about how things could have turned out differently. She often thought “the accident wouldn’t have been so bad if only . . . .” How do you think Joanne completed this thought? Please list six possible ways. The layout of obstacles into which Joanne crashed can be captured in a single model: lamppost tree postbox Joanne’s car where the diagram also represents the location of Joanne and her car when she scans the bend ahead and sees the three obstacles. The experiment
Ruth M. J. Byrne
146
used three versions of the vignette, which differed only in which of the three objects was assigned to which of the three positions, to control for any tendency subjects may have to more readily undo, say, a lamppost than a tree. The causality of the situation is separated in these vignettes from its spatial characteristics, in the sense that crashing into one object does not directly cause the individual to crash into another object. The description of the layout of the objects in the scenario in (47) is a determinate description, which gives rise to a single mental model. A second group of 20 subjects received the same sort of scenario but in which the description of the spatial layout was indeterminate: (48) Joanne drove home from work one cold wintry night. She was listening to the radio as she drove along a familiar stretch of road. As she scanned the bend ahead she saw a postbox on the left-hand side of the road. There was a large tree behind the postbox. The postbox was in front of a silver lamppost. She slowed approaching the bend and it was then that the accident happened. She hit a large patch of black ice and lost control of the car. She hit the postbox first. By an unlucky coincidence more ice caused the car to spin and she hit the tree. Once again by a freak coincidence she spun on more ice and hit the lamppost. She sustained multiple injuries including a broken leg and a fractured collarbone. In the weeks she spent recovering in hospital she often became upset thinking about the accident and thinking about how things could have turned out differently. She often thought “the accident wouldn’t have been so bad if only. . . .” How do you think Joanne completed this thought? Please list six possible ways.
The indeterminacy in the description renders it compatible with at least two alternative models:
lamppost tree postbox
tree lamppost postbox
The subjects’ task was to generate six counterfactual scenarios. One point of interest of the results is that the rate of mutation of spatial aspects of the scenario in the counterfactual alternatives generated by each subject was relatively high. Over the two versions of the scenario a major category of mutations, 33%, concerned the spatial aspects of the scenario. The remaining mutations included 28% of mutations that focused on the ice and
Cognitive Processes in Counterfactual Thinking
147
factors associated with it, such as the weather, wintertime and night time, 16% focused on factors associated with Joanne’s control of the car, including skid management, experience with driving, and so on, and 13%focused on matters of attention, especially listening to the radio. The primary focus of interest is the nature of the mutations in the 33% of counterfactual alternatives that undid the spatial aspects of the scenario. The expectation is that counterfactual mutations to indeterminate multiplemodel descriptions will be minimal because of the constraints of working memory, and these minimal mutations may concern, for example, simple deletions of an object: “if only the tree hadn’t been there.” Mutations to determinate single-model descriptions will be more complex, for example, changes not only to objects but also to relations between objects: “if only the tree hadn’t been behind the postbox.” As expected, counterfactual mutations to multiple-model factual situations were minimal: 73% of the counterfactual alternatives were simple mutations that mentioned just a single object (e.g., deleting it), whereas 27% of the alternatives were more complex mutations that mentioned two or more of the objects (e.g., altering their relations), and the difference is reliable. But for the determinate single-model descriptions a different pattern emerged. Counterfactual mutations to single-model factual situations were more complex: 53% of the counterfactual alternatives were simple mutations that mentioned just a single object, but just as many, 47% of the counterfactual alternatives, were more complex mutations that mentioned two o r more of the objects:
Simple mutation Complex mutation
Single model
Multiple model
53% 47%
73% 27%
The experiment shows that people can mutate spatial aspects of scenarios when they think counterfactually about what might have been. Space can join the lineup of time, cause, and intentionality as fundamental joints of reality along which we compute counterfactual alternatives. But more importantly, the experiment provides a test of the model theory of counterfactual thinking. The model theory of factual thinking predicts that problems that require multiple models to be kept in mind will be more difficult than problems that require a single model. The extension of the model theory from factual thinking to counterfactual thinking is based on the premise that thinking about matters of fact and thinking about matters of possibility are based on the same sorts of mental representations and processes. The model theory of counterfactual thinking therefore predicts
Ruth M. J. Byrne
148
that when the generation of multiple counterfactual scenarios requires multiple models of the factual situation to be kept in mind it will be harder than when it requires a single model of the factual situation. The experiment shows that people make both simple and complex mutations to a single factual model of a determinate description to generate a counterfactual alternative. In contrast, they make many more simple mutations than complex mutations to the multiple factual models of an indeterminate description to generate a counterfactual alternative to it. In summary, the determinacy of the description of the factual situation has a profound impact on people’s ability to think of counterfactual alternatives to it. We can be more imaginative when we think about what might have been for situations that are consistent with a single factual model than for those that are consistent with multiple factual models. This novel effect of the number of models that must be kept in mind is a prediction from the basic tenets of the model theory (e.g., Johnson-Laird & Byrne, 1991). It may be the case that the number of models to be kept in mind affects other aspects of counterfactual thinking, too, not just spatial scenarios. The finding that the number of models that must be kept in mind of the factual situation affects the ease with which individuals can construct counterfactual scenarios supports the general suggestion that reasoning about matters of fact and matters of possibility and impossibility rely on the same sorts of representations and processes. It also supports the more specific suggestion that these representations and processes are procedures that construct and revise mental models. IV. Conclusions My aim in this article has been to show that a theory developed to account for deduction, the model theory (Johnson-Laird, 1983; Johnson-Laird & Byrne, 1991), can be extended to account for everyday sorts of thinking such as counterfactual thinking about what might have been. I began this article by outlining the model theory of reasoning about matters of fact, taking as an example, a factual conditional, such as: (15’) If it rained then the children played indoors.
which is represented in an initial set of models in which some information is represented explicitly and other information is represented implicitly because of the constraints of working memory: rain
indoors
...
Cognitive Processes in Counterfactual Thinking
149
I indicated that the model theory of factual conditionals could also account for reasoning about matters of possibility, such as deductions from a counterfactual conditional, for example: (18’) If it had rained, the children would have played indoors.
which is represented in a richer initial set of models: Factual: Counterfactual:
no rain rain
...
not indoors indoors
The deductions people make from counterfactual conditionals are different from the deductions they make from factual conditionals only because they construct an initial representation of counterfactuals that corresponds to a more fleshed out version of the initial set of models for factual conditionals. The representations and processes required for reasoning about matters of fact and matters of possibility are essentially the same. What differs is what is represented explicitly in the initial set of models and this difference can have a profound impact on the inferences that reasoners make. Reasoners usually represent what is true in their models. A unique aspect of counterfactual reasoning is that it requires individuals to keep track of what is true and also what is false, but temporarily supposed to be true. My emphasis has been on highlighting the similarities between thinking about facts and thinking about possibilities and impossibilities. I have taken as a starting premise the following idea: a theory of the mental representations and processes that underlie the generation of counterfactual scenarios requires a theory of the mental representations and processes for factual scenarios. Reasoners generate counterfactual scenarios of many different sorts, and many factors guide the manner in which they generate the scenarios. Among the factors that guide the generation of counterfactual scenarios are factors associated with the nature of the mental representation of the factual situation. I have examined three primary phenomena of counterfactual thinking: the action effect, the temporality effect, and a new spatial effect, to try to sketch how the representation of the factual situation may guide the counterfactual scenarios individuals construct in each of these domains. People tend to regret their actions more than their inactions. The action effect can be explained by the following principles: people construct an initial set of models that represent as little information as possible explicitly; they mentally represent the action explicitly in their models and the inaction implicitly. They do not flesh out all the possible counterfactual models
150
Ruth M. J. Byrne
because of the constraints of working memory. Aspects of the factual situation that are represented explicitly in models are easier to mutate than aspects that are not represented explicitly. The representation of an action requires more models than the representation of an inaction, and accordingly, the multiple models of an action reduce the number of counterfactual possibilities, whereas the single model of an inaction leaves multiple counterfactual possibilities. People tend to mentally undo the more recent event in an independent sequence of events. This temporality effect arises because reasoners construct models that correspond to the world rather than to the language used to describe the world. They initialize or contextualize their models of an independent sequence of events by the first event, against which subsequent events are interpreted. They do not flesh out all the possible counterfactual models. The models they flesh out are the models in which the initializing event is unaltered. The initializing event is the event that occurred first in the world, regardless of whether it occurred first in the description of the world. People can undo spatial aspects of scenarios just as they undo temporal, causal, or intentional aspects. Factual scenarios may contain spatial layouts, for which there are many counterfactual alternatives. Individuals generate more complex counterfactual scenarios when the factual situation requires them to keep just a single model in mind compared to when the factual situation requires them to keep multiple models in mind. The number of models that must be kept in mind may have as powerful an effect on counterfactual thinking as it does on factual thinking. The biases that people exhibit when they generate counterfactual scenarios, for example, their tendency to undo actions more than inactions, their tendency to undo the more recent event in an independent sequence of events, or their tendency to construct more complex counterfactual alternatives for determinate spatial scenarios rather than indeterminate spatial scenarios, can be understood with reference to the nature of the models that they construct of the factual situation. Reasoning about matters of fact and matters of possibility and impossibility relies on the same sorts of mental representations and processes, and the underlying mechanisms may be procedures that construct and revise mental models. The sketch of a model theory of counterfactual thinking that I have tried to provide here suggests that counterfactual thinking, which lies at the heart of the human imagination, may share fundamental properties with logical thought. ACKNOWLEDGMENTS I am very grateful to Phil Johnson-Laird, Vittorio Girotto and Mark Keane for their many helpful comments on this paper. I also thank Ronan Culhane, Tim Dalgleish, Orlando
Cognitive Processes in Counterfactual T h i n g
151
Espino, Simon Handley, Patrick McAlinney. Rachel McCloy, Liz McLoughlin, Sergio Moreno Rios, Susana Segura, Alessandra Tasso, and Valerie Thompson, for discussions of counterfactual thinking. Some of the research described here was supported by grants from the Spanish and Italian governments, and the Dublin University Arts and Social Sciences Benefactions Fund, Ireland. The ideas here have benefited from discussions of some of the experiments at the International Workshop on Reasoning and Decision-Making in Siena in 1995; the 16th and 17th Annual Conferences of the Cognitive Science Society in Atlanta and Pittsburgh in 1994 and IYY5: the International Conference on Thinking, and the Mental Models in Cognitive Science Symposium in Honour of Phil Johnson-Laird, both in London, in 1996; and the Spatial Mental Models Workshop in Freiburg in 1996.
REFERENCES A d a m , E. W. (1975). The logic of conditionals. Dordrecht, The Netherlands: Reidel. Au. T. K. ( 1083). Chinese and English counterfactuals: The Sapir-Whorf hypothesis revisited. Cognition, 15, 155-187. Aycrs. M. R. (1965). Counterfactuals and subjunctive conditionals. Mind. pp 347-364. Barwise. J. (1986). Conditionals and conditional information. In E. C. Traugott, A. ter Meulen, J. S. Reilly, and C. Ferguson (Eds.). O n conditionals. Cambridge, UK: Cambridge University Press, pp. 21-53. Bell, V.. & Johnson-Laird, P. N. (1997). A model theory of modal reasoning. Unpublished manuscript, Princeton University, Princeton, NJ. Bouts, P., Spears, R., & van der Pligt, J. (1992). Counterfactual processing and the correspondence between events and outcomes: Normality versus value. European Journal uf Social Psychology, 22,387-396. Braine, M. D. S.. & O’Brien, D. P. (1991). A theory of I F A lexical entry, reasoning program, and pragmatic principles. Psychological Review, Y8, 182-203. Byrne. R. M.J . ( 198Ya). Everyday reasoning with conditional sequences. Quarterly Journal o,f Exprimenttrl Psychology, 4 / A , 141- 166. Byrne, R. M. J. (1989b). Suppressing valid inferences with conditionals. Cognition, 31, 61-83. Byrne. R. M.J. (1996). Towards a model theory of imaginary thinking. In J. Oakhill & A. Garnham (Eds.), Mental models in cognitive science: Essays in honour of Phil Johnson/.uird (pp. ISS-174). Hove, UK, Hillsdale, NJ: Taylor & Francis, Erlbaum. Byme. R. M. J . (1997). Spatial effects in counterfactual thinking. Kognitionswissenschaft. In press. Byrne. R. M. J., Culhane, R., & Tasso, A. (1995). The temporality effect in thinking about what might have been. In J. D. Moore & J. F. Lehman (Eds.), Proceedings of the seventeenrh annual cvnference of the Cognitive Science Society (pp. 385-390). Hillsdale, NJ: Erlbaum. Byrne, R. M. J., & Handley. S. J. (1997). Reasoning strategies for suppositional deductions. Cognition, 62, 1-49. Byrne, R.M. J.. Handley, S. J.. & Johnson-Laird, P. N. (1995). Reasoning with suppositions. Quarterly .loirmal uf Exprrimenrirl P s y c h ~ l o g y@A, ~ 915-944. Byrne. R. M. J.. & Johnson-Laird, P. N. (1989). Spatial reasoning. Journal uf Memory and Langiruge, 28, 564-575. Byrne. R. M. J., and Johnson-Laird, P. N. (1992). The spontaneous use of propositional connectives. Quarterly Joitrnul of Experimental Psychology, 44A, 89-110. Byrne. R. M. J.. & McEleney, A. (1997). Cognitive processes in regret for actions and inactions. In Langley, P. & Shafto, M. (Eds). Proceedings of the 19th Annual Conference of the Cognitive Science Society. Hillsdale, NJ: Erlbaum.
152
Ruth M. J. Byrne
Byrne, R. M. J., Segura, S., Culhane, R., Tasso, A., McAlinney, P., & Berrocal, P., (1997). Context alters temporal mutability in imaginary counterfactual thinking. Manuscript submitted for publication. Byrne, R. M. J., & Tasso, A. (1994). Counterfactual reasoning: Inferences from hypothetical conditionals. In A. Ram & K. Eiselt (Eds.), Proceedings ofthe sixteenth annual conference of the Cognitive Science Society. Hillsdale. NJ: Erlhaum. pp. 124-129. Byrne, R. M. J., & Tasso, A. (1997). Counterfactual conditionals: Deductive reasoning in imaginary scenarios. Manuscript submitted for publication. Carpenter, P. A. (1973). Extracting information from counterfactual clauses. Journal of Verbal Learning and Verbal Behavior, 12, 512-520. Chisholm, R. M. (1946). The contrary-to-fact conditional. Mind, 55, 289-307. Comrie, B. (1986). Conditionals: A typology. In E. C. Traugott, A., ter Meulen. J. S. Reilly, & C. Ferguson (Eds.), On conditionals. Cambridge, UK: Cambridge University Press. pp. 77-99. Davis. C. G., Lehman. D. R., Wortman, C. B., Silver, R. C., & Thompson, S. C. (1995). The undoing of traumatic life events. Personality and Social Psychology Bulletin, 21, 109-124. Dudman, V . H. (1988). Indicative and subjunctive. Analysis, 48, 113-122. Dunning, D.. & Parpal, M. (1989). Mental addition versus subtraction in counterfactual reasoning: On assessing the impact of personal actions and life events. Journal ofPer.sona1ity and Social Psychology, 57, 5-15. Ehrlich, K., & Johnson-Laird, P. N. (1982). Spatial descriptions and referential continuity. Journal of Verbal Learning and Verbal Behavioitr, 21, 296-306. Evans, J. St. T., Newstead, S., & Byrne, R. M. J. (1993). Human reasoning: The psychology of deduction. Hillsdale. NJ: Erlbaum. Fillenbaum, S. (1974). Information amplified: Memory for counterfactual conditionals. Journal of Experimental Psychology, 102, 44-49. Gavanski, I., & Wells. G. L. (1989). Counterfactual processing of normal and exceptional events. Journal of Experimental Social Psychology, 25, 314-325. Gilovich, T.. & Medvec, V. H. (1994). The temporal pattern to the experience of regret. Journal of Personality and Social Psychology, 67, 357-365. Gilovich, T., & Medvec, V. H. (1995). The experience of regret: What, when and why. Psychological Review, 102, 379-395. Ginsberg, M. L. (1986). Counterfactuals. Artificial Intelligence, 30, 35-79. Girotto, V., Legrenzi, P., & Ruzo, A. (1991). Event controllability in counterfactual thinking. Acta Psychologica, 78, 111-133. Gleicher, F., Kost, K. A., Baker. S. M., Strathman, A. J., Richman, S. A.. & Sherman, S. J. (1990). The role of counterfactual thinking in judgments of affect. Personality and Social Psychology Bulletin, 16, 284-295. Goodman, N. (1973). Fact, fiction and forecast, (3rd ed.). New York: Bobbs-Merrill. Grice, H. P. (1975). Logic and conversation. In P. Cole & J. L. Morgan (Eds.), Syntax and Speech acts. New York: Academic Press. pp. 41-58. Hofstadter, D. R. (1979). Godel, Escher, Bach:An eternalgolden braid. New York: Basic Books. Hofstadter, D. R. (1985). Metamagical themas: Questing for the essence of mind and pattern. London: Penguin. Isard, S. D. (1974). What would you have done if . . . ? Journal of Theoretical Linguistics, 1. 233-255. Jackson, F. (1987). Conditionals. Oxford: Basil Blackwell. Jeffrey, R. (1981). Formal logic: Its scope and limits (2nd ed.). New York: McGraw-Hill. Johnson, J. (1986). The knowledge of what might have been: Affective and attributional consequences of near outcomes. Personality and Social Psychology Bulletin, 12, 51-62.
Cognitive Processes in Counterfactual Thiiking
153
Johnson-Laird. P. N. (1983). Mental models. Cambridge, UK: Cambridge University Press. Johnson-Laird, P. N. (1986). Conditionals and mental models. In E. C. Traugott, A. ter Meulen, J. S. Reilly, & C. Ferguson, (Eds.), On conditionals. Cambridge, UK: Cambridge University Press. pp. 55-75. Johnson-Laird, P. N., & Bara, B. G. (1984). Syllogistic inference. Cognition, 16, 1-62. Johnson-Laird, P. N., & Byrne, R. M. J. (1989). Only reasoning. Journal of Memory and l.unguage, 28, 313-330. Johnson-Laird, P. N., & Byme, R. M. J. (1991). Deduction. Hove, UK, Hillsdale, NJ: Erlbaum. Johnson-Laird, P. N., & Byrne, R. M. J. (1993). Models and deductive rationality. In K. I. Manktelow & D. E. Over (Eds.), Rationality: Psychological andphilosophicalperspectives (pp. 177-210). London: Routledge. Johnson-Laird, P. N., Byrne, R. M. J., & Schaeken, W. (1992). Propositional reasoning by model. Psychological Review, 99, 418-439. Johnson-Laird, P. N., Byrne, R. M. J., & Tabossi, P. (1989). Reasoning by model: The case of multiple quantification. Psychological Review, 96, 658-673. Johnson-Laird, P. N., & Savary, F. (1995). How to make the impossible seem probable. In J. D. Moore & J. F. Lehman (Eds.), Proceedings ofthe seventeenth annual conference of the Cognitive Science Society (pp. 381-384), Hillsdale, NJ: Erlbaum. Johnson-Laird, P. N., & Steedman, M. J. (1978). The psychology of syllogisms. Cognitive Psychology, 10, 64-99. Kahneman, D., & Miller, D. (1986). Norm theory: Comparing reality to its alternatives. Psychological Review, 93, 136-153. Kahneman, D.. & Tversky, A. (1982). The simulation heuristic. In D. Kahneman, P. Slovic, & A. Tversky (Eds.), Judgment under uncertainty: Heuristics and biases (pp. 201-208). New York: Cambridge University Press. Kahneman, D., & Varey, C. A. (1990). Propensities and counterfactuals: The loser who almost won. Journal of Personality and Social Psychology, 59, 1101-1110. Keane, M. T. (1996). On adaptation in analogy: Tests of pragmatic-importance and adaptability in analogical problem solving. Quarterly Journal of Experimental Psychology, 49A, 1062108.5. Keane, M. T. (in press). What makes an analogy difficult?: The effects of order and causal structure in analogical mapping. Journal of Experimental Psychology: Language, Memory, und Cognition. Keane. M. T., Ledgeway, T., & Duff, S. (1994). Constraints on analogical mapping: A comparison of three models. Cognitive Science, 18, 387-438. Knight, R. T., & Grabowecky, M. (1990). Escape from linear time: Prefrontal cortex and conscious experience. In M. Gazzaniga (Ed.), The cognitive neurosciences. Cambridge, UK: Cambridge University Press. pp 1357-1371. Landman, J. (1987). Regret and elation following action and inaction: Affective responses to positive versus negative outcomes. Personality and Social Psychology Bulletin, 13, 524-536. Legrenzi, P., Girotto, V., & Johnson-Laird, P. N. (1993). Focusing in reasoning and decisionmaking. Cognition, 49, 37-66. Lewis, D. (1973). Counterfactuals. Oxford, England: Blackwell. Lewis, D. (1979). Counterfactual dependence and time’s arrow. Nous, X f f l , 455-76. Mani, K., & Johnson-Laird, P. N. (1982). The mental representation of spatial descriptions. Memory & Cognition, 10, 181-187. Markman, K. D., Gavanski, I., Sherman, S. J., & McMullen, M. N. (1993). The mental simulation of better and worse possible worlds. Journalof ExperimenfalSocialPsychology, 29. 87-109.
154
Ruth M. J. Byrne
Marr. D. (1982). Vision. New Ynrk: Freeman. Maybery, M. T., Bain, J. D.. & Halford. G . S. (1986). Information processing demands of transitive inference. Journal of Experimental Psychology: Learning, Memory, and Cognition, 12, 600-613. Miller. D. T.. & Gunasegaram, S. (1990). Temporal order and the perceived mutability of events: Implications for blame assignment. Journal o,f Personality and Social Psychology. 59, 1111-1118. Miller. D. T., & McFarland. C. (1987). Counterfactual thinking and victim compensation: A test of norm theory. Personality and Social Psychology Bulletin. 12, 513-519. Miller, G., & Johnson-Laird, P. N. (1976). Lunguuge and perception. Cambridge, UK: Cambridge University Press. Miyamoto, J. M., & Dibble, E. (1986). Counterfactual conditionals and the conjunction fallacy. In Proceedings of the eighth annual conference of the Cognitive Science Society. Hillsdale, NJ: Erlbaum. Miyamoto. J. M., Lundell. J . W., & Tu, S. (1989). Anomalous conditional judgments and Ramsey’s thought experiment. In Proceedings of the eleventh annual conference of the Cognitive Science Society. Hillsdale, NJ:Erlbaum. Pollock, J. L. (1986). Subjunctive reasoning. Dordrecht, The Netherlands: Reidel. Ramsey, F. P. (1931). The foundations of mathematics and other logical essays. London: Kegan Paul. Roese. N. J. (1994). The functional basis of counterfactual thinking. Journal of Personality arid Social Psychology, 66, 805-818. Roese, N. J., & Olson, J. M. (1993). The structure of counterfactual thought. Personality and Social Psychology Bulletin, 19, 312-319. Roese, N. J., & Olson, J. M. (1995). What might have been: The .socialpsychology of coiinterfactun1 thinking. Mahwah, NJ: Erlbaum. Sanna, L. J., & Turley, K. J. (1996). Antecedents of spontaneous counterfactual thinking: Effects of expectancy violation and outcome valence. Personality and Social Psychology Bulletin, 22, 906-919. Sherman, S. J., & McConnell, A. R. (1996). The role of counterfactual thinking in reasoning. Applied Cognitive Psychology. 10, 113-124. Stalnaker, R. C. (1968). A theory of conditionals. In N. Rescher (Ed.), Studies in logical theory. Oxford, England: Basil Blackwell. pp. 98-1 12. Thompson. V.. & Byrne, R. M. J. (1997). Counterfactrial conditionals. Manuscript in preparation. Tversky. A., & Kahneman, D. (1982). Judgment under uncertainty: Heuristics and biases. In D. Kahneman, P. Slovic. & A. Tversky (Eds.), Judgment under uncertainty: Heuristics and biases (pp. 201-208). New York: Cambridge University Press. Wells, G . L., & Gavanski. I. (1989). Mental simulation of causality. Journal of Personaliry and Social Psychology, 55, 161-169. Wells, G . L., Taylor, B. R., & Turtle, J. W. (1987). The undoing of scenarios. Journal of Personality and Social Psychology. 53, 421-430.
EPISODIC ENHANCEMENT OF PROCESSING FLUENCY Michael E. J. Masson Colin M. MacLeod
I. Introduction About 25 years ago, the prevalent view of memory began to change. Prior to that time, the predominant theories were framed in terms of engrams or traces that were stored in memory at the time of encoding and recovered from memory at the time of retrieval, much as money is deposited in and withdrawn from a bank. This can be seen early on in the writings of Semon (1904/1921),but this style of theory prevailed into the 1960s (see textbooks of that time, e.g., Kintsch, 1970). Encoding processes laid down traces in memory and retrieval processes searched for them. Traces were products of encoding and objects of retrieval. Clearly, processes were involved in both encoding and retrieval, but what was stored in memory was the consequences of initial processing, not records of the processing itself. This trace view is well captured by the best known models of that period, the twostate or “buffer” models of memory (Atkinson & Shiffrin, 1968; Waugh & Norman, 1965). Then, in the early 1970s, a new perspective emerged. In large part, this stemmed from research carried out at the University of Toronto. It began with two key elements: the renewed emphasis on processing that formed the basis of the levels of processing framework (Craik & Lockhart, 1972); and the increased stress on the interplay between encoding and retrieval THE PSYCHOLOGY OF LEARNING AND MOTIVATION, VOL. 37
155
Copyright 0 1997 by Academic Press. All rights of reproduction in any form reserved. W79-7421/97 $25.00
156
Michael E. J. Masson and Colin M. MacLeod
that culminated in the encoding specificity principle (Tulving & Thomson. 1973). These fundamental ideas came together in several places, notably in the work on transfer of skills (Kolers, 1976; Kolers & Roediger, 1984) and in the concept of transfer-appropriate processing (Morris, Bransford, & Franks, 1977). Under these views, memory did not record the products of processing; rather, memory was a chronicle of the processes themselves. In this new theoretical framework, encoding does not extract and store away a trace based on analysis of the stimulus situation; instead, the analysis itself is recorded in memory. Kolers made this point especially forcefully: “Memory is not a place in the mind where things are stored, but a way of responding to events with skills acquired in previous encounters with like circumstances; recognition is not a matter of matching but a process of transferring skills across occasions” (1976, p. 564). Remembering, then, consists of processes occurring at the time of retrieval, and is successful t o the extent that the same processes are applied at both points in timeencoding and retrieval-resulting in successful transfer of processing. Morris et al. put this quite simply: “the value of particular types of acquisition activities must be defined relative to the type of activities to be performed at the time of test” (1977, p. 531). This analysis of remembering has come even more to the fore as cognitive psychologists have begun to explore not just conscious remembering, but also unconscious remembering. Almost certainly, we use memory without being aware of it a great deal of the time, probably more than we do for deliberate recollection of our past. As just one example, when we are working at our computers, there are a great many remembered skills and facts that we use without being aware that memory is being engaged. This conscioushnconscious distinction is now firmly entrenched in the memory literature as the explicithmplicit distinction (Graf & Schacter, 1985) and is increasingly influencing our theorizing about the operation of memory. As just two illustrations, both Jacoby (1983) and Roediger (1990) have argued that the nature of the processing at the time of encoding strongly influences performance both on direct tests of conscious recollection and on indirect tests of unconscious remembering. The essence of the argument is that different types of memory tests selectively map onto different processes executed at the time of encoding.
A. FLUENT REMEMBERING These ideas about the connection between encoding and retrieval are intimately tied to the concept of fluency in remembering, a concept that Jacoby and his colleagues have championed for some time. In the words of Jacoby and Brooks (1984, p. 35), “When using the fluency heuristic, the person
Episodic Enhancement of Processing Fluency
157
infers that an item must have occurred before if it can be processed relatively more easily.” When an earlier processing episode relates to a later processing episode, that subsequent processing will benefit because the processes can be rerun more easily the second time. This benefit may be quite unconscious. Alternatively, the subject may experience a sense of fluency, and this impression may lead to an attribution of “remembering.” So, from this perspective, conscious remembering is an attribution contingent upon fluent reprocessing. There is already considerable evidence in accord with this position. Jacoby and Dallas (1981, p. 333) invoked this account early on in contrasting performance on recognition (explicit) and masked word identification (implicit) tests of memory: “Subjects may base their recognition memory decision on judgments of the relative fluency of their own performance of a task.” More recently, in their study of false recognition, Jacoby and Whitehouse (1989) referred to an “illusion of memory” wherein “fluency of processing can give rise to a feeling of familiarity” (p. 133) that in turn affects conscious recognition judgments. In their studies of reading words aloud and making judgments about those words, Whittlesea, Jacoby, and Girard (1990, p. 716) reinforced this idea: “fluent performance is unconsciously attributed to whatever source is apparent and . . . feelings of familiarity result when fluency is attributed . . . to past experience.” As will become clear, the ideas of Jacoby and his colleagues have heavily influenced the account of memory we have been developing. VERSUS CONCEPTUALLY DRIVEN PROCESSING B. DATA-DRIVEN
The dominant interpretation of implicit versus explicit remembering has grown out of the transfer-appropriate processing framework (Morris et al., 1977), initially through the work of Jacoby (1983) and since then through the work of Roediger (1990). On this view, the majority of indirect tests that measure memory implicitly, where an individual need not be aware that he or she is remembering, are “tuned” principally to data level or perceptual encoding operations. Thus, changes between study and test in stimulus form (e.g., between fonts, or from word to picture) or modality (e.g., from auditory to visual) reduce performance on tasks such as word fragment completion, where a stimulus such as . r a t _ r must be completed with a word such as “sweater” (see Roediger & McDermott, 1993, for a review). In contrast, standard direct tests that measure memory explicitly, tests such as recall and recognition, are tuned primarily to conceptual level or meaningful encoding operations. Elaborations that direct attention toward meaning during encoding, such as producing semantic associates of the study items, especially benefit conscious recollection. In this way,
158
Michael E. J. Masson and Colin M. MacLeod
transfer-appropriate processing has been modified to capture both ways of remembering, implicit and explicit. One procedure that has seen extensive use in exploring this processing distinction is the read versus generate manipulation. During encoding, subjects either read isolated words (e.g., umbrella) or produce those words from semantic clues (e.g., “This keeps the rain off your head-u?”). It is well established that generated words are much better remembered than are read words on direct tests like recall or recognition (Slamecka & Graf, 1978). But the reverse is sometimes true for certain indirect tests: Reading can actually lead to better remembering than generating (e.g., Jacoby, 1983: Roediger, 1990). Of course, both types of study activity are superior t o baseline performance for unstudied words. This improvement due to repeated processing of a stimulus has come to be called priming. In developing their data-driven versus conceptually driven account, Roediger and his colleagues at one point suggested that the read/generate pattern be used as a benchmark for defining whether a particular test was data driven or conceptually driven (Roediger, Weldon, & Challis, 1989). Expressing this proposal as a conditional rule, we have: if read > generate, then data-driven: if generate > read, then conceptually driven. The majority of their results, particularly those emphasized in theory development, do fit this rule. Roediger and McDermott (1993: see also Roediger, 1990) review many of the relevant studies, especially studies using the indirect word fragment completion test. The repeated finding is that having read the critical words earlier leads to more priming than does having generated them, whereas the opposite is true on direct tests of recognition or recall.
c.
INITIAL INTERPKEIIVE ENCODING A N D SUBSEQUENT ELABORATIVE ENCODING
In our initial examination of these ideas regarding data-driven and conceptually driven processing, we used a different indirect test and obtained a different pattern of results (Masson & MacLeod, 1992). The test we used was masked word identification (also called perceptual identification), in which the subject must identify a word visually presented for about 30 ms and followed by a pattern mask. We were able to replicate the findings of Jacoby (1983) when we used the same sort of antonym generation scheme (ha-c; generate the opposite beginning with the specified letter). Recall that Jacoby’s results were consistent with the rule: Read words were more often identified than generated words, despite the opposite pattern in recognition. But it turned out that antonyms were virtually unique in producing this pattern. For almost every other type of generation rule that we examined in a series of 11 experiments, from definitional phrases t o famous
Episodic Enhancement of Processing Fluency
159
names to synonyms, we found a different pattern: Read and generated words produced similar amounts of priming in masked word identification. The overwhelming majority of results on our indirect test, therefore, did not fit the rule. Our results led us to propose an alternative account of how encoding and retrieval operated. This account owes a considerable intellectual debt to an explanation first set out by Graf and Mandler (1984), who discriminated between integrative and elaborative encoding. We also suggested that subjects produced two encodings at the time of study. The initial interpretive encoding was made immediately upon perceiving the stimulus, and incorporated both conceptual and perceptual aspects of the stimulus as processed (and, indeed, any other available aspects). Reenacting this initial interpretive encoding was seen as underlying any priming on subsequent indirect tests, such as masked word identification, that involved identifying or in some other way producing a specific target stimulus. Thus, a studied word should show priming on such a test, regardless of whether it had been generated or read. An elaborate encoding could also be created, but this was subsequent to interpretive encoding and more subject to deliberate control. It was this “deeper” encoding, to use levels of processing language, that dominated performance on direct tests, explaining why generated words showed better explicit retention than did read words. The general principle that we intended to emphasize by proposing the concept of an initial interpretive encoding is that the primary goal of encoding a stimulus is to interpret or classify the stimulus, so that appropriate knowledge about it can be called to mind. Adoption of that general principle had its roots in early work that demonstrated the significant influence that context has on encoding and retrieval (e.g., Masson & Sala, 1978; Meyer & Schvaneveldt, 1971; Neely, 1991; Rumelhart, 1977). As an aside, we should note that our view actually is somewhat more radical than we have set out this far. In fact, we do not see the distinction between encoding and retrieval as particularly valuable any longer. Each encounter with a stimulus involves interpretation and potentially elaboration, whether that encounter is nominally the first one (encoding) or a subsequent one (retrieval). Furthermore, each encounter necessitates operations corresponding to what traditionally have been called encoding and retrieval. Identifying the operations is an important goal of research; maintaining an artificial dichotomy between encoding and retrieval operations would not appear to provide additional benefits. Processing overlap across episodes supports remembering whether that remembering is measured implicitly or explicitly. The “value added” of the encodinghetrieval distinction is unclear.
160
Michael E. J, Masson and Colin M. MacLeod
We return now to the concept of initial interpretation as involving both perceptual and conceptual aspects. More recent theoretical developments provide additional encouragement for our view that both aspects of a stimulus are collaboratively involved in the encoding operations that underlie repetition priming effects. Two articles are especially relevant in this regard. First, Strain, Patterson, and Seidenberg (1995) have shown that a conceptual variable, imageability of a word, influences speeded word reading even when words are presented in isolation. The important implication of this result is that even though the subject’s task is to read aloud, with no inducement to engage in semantic elaboration of the word, conceptual knowledge associated with the word appears to be recruited. Second, Becker, Moscovitch, and Behrmann (in press) have developed a connectionist model of semantic priming based on long-term changes to connection weights. One important prediction derived from this model is that semantic priming can have long-lasting effects. More important for our purposes, though, is the fact that the model constitutes an implementation of the idea that a specific processing episode can have a long-term effect on the fluency of word identification. Moreover, with a model such as theirs, that long-term influence involves not only perceptual aspects of a word’s representation, but conceptual aspects of that representation as well. Although we have yet to explore this very promising development, we expect that a model of this class should be useful in deriving testable predictions regarding repetition priming effects. In the remainder of this chapter, we describe several series of experiments that we have conducted in our laboratories and that reveal the principles determining the influence of specific processing episodes on subsequent word identification. We begin with a summary of the initial series of experiments that led to our adoption of the distinction between interpretive and elaborative encoding. Subsequent series of experiments to be described were aimed at (1) providing converging evidence that conceptual and perceptual processes jointly contribute to the priming observed in tasks such as masked word identification, and (2) testing hypotheses regarding how memory for prior episodes modifies word identification processes. All of these issues come together in the concepts of fluency and interpretive encoding.
11. Experiment Series 1:Data-Driven and Conceptually Driven
Encoding Tasks The experiments reported by Jacoby (1983) provided evidence for two fundamentally important claims regarding episodically enhanced word
Episodic Enhancement of Processing Fluency
161
identification. First, improved identification of briefly exposed target words was dependent on prior perceptual processing of those words. Second, improved target identification was not a result of conscious recollection of the previously presented words. Jacoby used three encoding tasks, two of which are important for the current discussion: reading a target word in isolation and generating a target word from its antonym (e.g., hot-???). In the latter task, the target itself was not actually presented to the subject, but was experienced only in the act of generating it in response to its antonym. Subjects were then tested on the encoded words and a set of nonstudied words in one of two ways. One test was a yesho recognition test in which subjects decided whether each target word had been presented earlier (either read or generated). The other test was the masked word identification task. Jacoby found that recognition memory was more accurate for generated than for read targets, but the reverse was true for performance on the masked word identification task. Moreover, little or no priming was found on masked word identification for targets that had been generated in the encoding phase. The double dissociation between the recognition memory and masked word identification tests provided powerful evidence that conscious recollection was not responsible for the pattern of results observed in the masked word identification task. The finding that generation at encoding led to little or no enhancement on the masked word identification task indicated that performance on that task was affected almost exclusively by prior perceptual (data-driven) experiences. This outcome fits nicely with the transfer-appropriate processing framework (Morris et al., 1977; Roediger et al., 1989), by which the influence of prior processing episodes on subsequent tests is assumed to depend on the involvement of similar processing operations at study and test. The read encoding task was deemed to involve data-driven processing and the generation task was confined to conceptually driven processing. The fact that the masked word identification task showed substantial benefit only from the read encoding task implies that the masked word identification task should be classified as data driven. Although the results obtained by Jacoby (1983) provided convincing support for the view that enhanced performance on the masked word identification task could be brought about only by perceptual processing experience, we were concerned about another factor that was at work in those experiments. Items encoded in the read task might have been placed at an advantage because the context in which they appeared during study and test was the same, namely, the word appeared by itself with no surrounding context. In the case of words encoded in the generation task, however, targets were processed in conjunction with a strongly associated context word (the antonym cue) but tested in isolation. The change in
162
Michael E. J. Masson and Colin M. MacLeod
contextual environment experienced with items in the generation encoding condition might have contributed to the lack of priming for those items. To test this possibility, we designed an initial experiment in which the contextual information used to generate targets during study was present or absent at the time of test (Masson & MacLeod, 1992). A related study by Toth and Hunt (1990) provided preliminary support for this approach. They presented subjects with a semantic associate for each target at the time of study. The study tasks involved reading some targets and generating others from fragments in which one or two letters of the target were missing. The context word might be expected to play a larger role in the generate condition because some additional work had to be done to determine the target’s identity. Subjects were then given a masked word identification test in which the target word was tested either in the presence or absence of its semantic associate. Toth and Hunt found that identification of items in the generate condition was improved by the presence of the relevant associate to a greater extent than was the case for items in the read condition. Although this result suggests that generated targets have the potential to benefit more from the reinstatement of contextual information provided during study, the generation task used by Toth and Hunt was not optimal for addressing the issue of concern here. Generating targets from word fragments with only one or two letters missing seems to involve a strong data-driven component because it was found that even when context was not reinstated at test, identification performance on generated targets was at least as high as performance on read items. In our initial experiment (Masson & MacLeod, 1992, Experiment l), subjects generated targets from a brief definition and the first letter of the target word (e.g., “the orange vegetable a rabbit eats-c?”). By providing only the first letter of the target word, we avoided the strong data-driven processing that apparently occurred in Toth and Hunt (1990) study. For generation cues, we used definitions instead of antonyms, as Jacoby had done, because pilot testing revealed that when an antonym was presented as a context word during the masked word identification test, there was a strong propensity for subjects to respond with the context word’s antonym even though that was not the target word. By using definitions as cues, we were able to select a representative word from the definition cue to use as a context word during the masked word identification test. The selected context words would not have sufficiently powerful associations with the targets to induce subjects to adopt the strategy we had observed with antonyms as context words. The design of the experiment was a 3 X 2 factorial within subjects design. Three encoding conditions were used: generate, read, and new (items not presented during the encoding phase). The second factor was test condition:
Episodic Enhancement of Processing Fluency
163
context and no context. In the study phase of the experiment, subjects were presented with two kinds of trials randomly intermixed. For targets assigned to the read condition, the target word appeared by itself at the center of a computer monitor until the subject read it aloud. On generation trials, a cue consisting of a brief definition and the first letter of the target appeared on the monitor and the subject attempted to name the intended target word. The encoding phase was followed by a test phase involving the masked word identification task. Three types of target were tested: generate, read, and new. Presentation of a target was preceded by a context stimulus consisting either of a row of Xs (the no context condition) or a relevant context word (a word taken from the definition cue for that target). In the case of generate targets, the context word had appeared during the study phase as part of the target’s encoding episode. For the read and new targets, the context word was related to the target only by preexperimental experience. The context stimulus was in view for 250 ms. The computer monitor was then blank for 250 ms, before the target word was presented in lowercase letters. The target appeared for 33 ms at the center of the monitor and was immediately followed by a pattern mask. The subject then attempted to identify the target. In the study phase, subjects typically had no difficulty in determining the correct target on generation trials. Across the experiments described in this chapter, subjects usually averaged over 90% correct when generating targets in the study phase. The results we report from the generate condition of these experiments are based only on generate items that were correctly generated during the study phase. To ensure that this conditionalized scoring of generate targets did not lead to item selection artifacts, we conducted parallel analyses that included all generate targets, regardless of whether they were correctly generated at study. These analyses almost invariably produced the same pattern of results found with the analyses of conditionalized data, indicating that item selection effects played little if any role in our comparison of generate and read encoding conditions. The mean proportions of targets correctly identified as a function of study and test condition are shown in Fig. 1A. There was a significant effect of study condition, indicating that prior exposure led to higher probability of identification on the masked word identification task. There was also an effect of context, with better identification performance observed when targets were preceded by related context words. These two factors, however, did not interact. The lack of interaction means that the prior study episodes involving generation of target words from semantic cues provided no additional potency to the context words drawn from those cues. Rather, the benefit of the context words was derived entirely from preexperimental experience. This finding seriously compromised the attempt to test the
164
Michael E. J. Masson and Colin M. MacLeod
Fig. 1. Mean proportion correct on masked word identification and recognition memory tasks as a function of encoding task. (A) The generation cues were definitions and targets were tested in the context of a related word taken from the relevant definition or with no relevant context word. (B) The generation cues were antonyms. (C) The generation cues were definitions that specified either the dominant or the secondary meaning of the homograph target words. (D) The generation cues were idioms in which the target word served as the last word in the idiom. Error bars indicate the 95% within-subject confidence intervals for the means. Data are adapted from Masson and MacLeod (1992).
context reinstatement hypothesis. One other important but unexpected result emerged in this experiment. The amount of priming observed for generate targets was just as large as the priming found with read targets. Thus, unlike Jacoby (1983), we found no advantage on the masked word identification test for targets that had been encoded perceptually, relative to targets that had been encoded without being perceptually processed. Given the unexpected anomalous finding of equal priming for generate and read targets, we examined the implications of that result rather than the question of context reinstatement. Our concern was that the relatively weak amount of priming found with the generate encoding task in the Jacoby (1983) study might not be a general result. It was important, therefore, to replicate the finding of greater priming in the read condition. To
Episodic Enhancement of Processing Flnency
165
do so, we turned to the materials used by Jacoby, which consisted of antonyms as generation cues. In our next experiment (Masson & MacLeod, 1992, Experiment 2), then, we adapted these materials by presenting a target word’s antonym and first letter as generation cues (e.g., true-7). In the study phase, subjects read some targets and generated others from their antonym cues. The masked word identification task was then presented with three sets of target words: generate, read, and new. The mean proportion of identified words in each study condition is shown in Fig. 1B. The results clearly replicated Jacoby’s (1983) finding of little or no priming when targets were previously generated from antonym cues. This result is important because it rules out uninteresting explanations of the difference between the results of our first experiment and the results obtained by Jacoby. The primary difference between these two studies is the nature of the materials. In subsequent experiments in the Masson and MacLeod (1992) article, we tested various kinds of generation cues in an effort to discover the factor that determined whether generating targets would yield as much priming as reading them. Most of the generation cues we examined produced as much priming for generate words as for read words (e.g., synonym and associate cues, hammer-n?; famous last names as cues for first names, M. Twain). For instance, in Experiment 5, subjects generated homographs from phrases that specified either the primary or a secondary meaning of the homograph. To help constrain responses to the intended target word, additional letters besides the first letter of the target were provided (e.g., “the sprinter got ready for her r-c-”; “we are all part of the human r-c-”). Subjects read or generated from one of its possible cues each of a set of homograph target words. Another set of homographs was not presented during the study phase but appeared as new words in the test phase. Subjects then were given a masked word identification test in which studied and nonstudied targets were presented in isolation, as in the earlier experiments. Finally, subjects were given a yeslno recognition test in which each of the targets from the identification test was presented. The task was to indicate whether a target had been presented during the study phase. This recognition test was partly a source discrimination test because all of the new items had appeared on the immediately preceding masked word identification test (although not all had been identified). The proportions of correctly identified words in each encoding condition of this experiment are shown in Fig. 1C. Once again, generated targets produced as much priming as read targets. Note also that it did not matter whether a homograph was generated in the context of a cue that specified its primary or secondary meaning. In either case, the same amount of priming was observed on the identification task. This lack of a dominance
166
Michael E. J. Masson and Colin M. MacLeod
effectcontrasts with the results of the recognition test, which are also shown in Fig. 1C. On that test, subjects were more likely to correctly recognize a generated target when it had been generated from a cue that specified its primary meaning. The dissociation between the masked word identification test and the recognition test revealed by the meaning dominance factor suggests that subjects were not relying to any substantial degree on conscious recollection when performing the identification test (see Experiment Series 3 for further discussion of this issue). Had conscious recollection played an important role on the identification test, we should have seen an effect of meaning dominance among the generated targets just as was found on the recognition test. Instead, the absence of the dominance effect on the identification test indicates that the brief presentation of a target is capable of recruiting a recent episode involving that word, regardless of which of the target’s meanings was invoked by that episode. This possibility is consistent with the proposal that all meanings of a homograph are initially retrieved before one of its meanings is selected (Seidenberg, Tanenhaus, Leiman, & Bienkowski, 1982; Swinney, 1979). We suggest that it is the early stages of word identification that are captured by the masked word identification task and that prior processing of either meaning of a homograph can be contacted by the brief presentation of the target word. In contrast to the masked word identification task, we assume that decisions on the recognition test are based for the most part on a selected meaning of the homograph and most typically the selected meaning is the dominant one. Thus, targets generated from a cue that specifies the dominant meaning are more likely to be recognized (for a related finding, see Light & Carter-Sobell, 1970). The demonstration that various generation tasks produce as much priming as reading a target word forces the question as to why the antonym task fails to do so. In Masson and MacLeod (1992), we suggested that one of the factors governing the amount of priming created by generation encoding tasks is the extent to which the generation cue and the target word are integrated during encoding. Our hypothesis was that if cue and target are integrated at encoding, then the brief presentation of the target word on the masked word identification test would retrieve an integrated record of that encoding event, making it difficult to discriminate between the cue and the target word. In support of this proposal, we demonstrated that other generation cues that were assumed to lead to strong integration of cue and target behaved like antonym cues. That is, generation of target words from these cues yielded less priming on the masked word identification test than did the read encoding task. An example of this result occurred in Experiment 6 of Masson and MacLeod (1992), in which generation cues were idiomatic expressions and
Episodic Enhancement of Processing Fluency
167
targets were the final words of those expressions (e.g., “kick the b-”). This experiment used the same procedure as Experiment 5 , including a yesho recognition test following the masked word identification test. The mean proportion of identified targets in each encoding condition is shown in Fig. 1D. Although the generate task produced reliable priming compared to new targets, it was significantly less than the priming found in the read condition. The opposite pattern was obtained in the recognition test, in which a higher proportion of correct responses was given to generate than to read items. This crossed dissociation replicates the pattern found with the antonym generation task (Jacoby, 1983; Masson & MacLeod, 1992, Experiment 2). Taken together, the series of experiments reported in Masson and MacLeod (1992) provided evidence to challenge the view that the masked word identification task reflects exclusively the effects of prior perceptual processing episodes. More generally, the experiments illustrated the plausibility of the claim that fluent word identification performance is sensitive to components of processing episodes that are responsible for constructing an interpretation of a presented word. These components include not only the perceptual aspects of the episode, but also the conceptual processes that are required for the construction of a meaningful interpretation of the stimulus. 111. Experiment Series 2: Comparing Masked Word
Identification and Word Fragment Completion Given the masked word identification studies described in Experiment Series 1 and their apparent contrast to other studies using such implicit tests as word fragment completion (see, e.g., Roediger & McDermott, 1993), we suspected that different implicit tests could behave quite differently. Witherspoon and Moscovitch (1989) had already shown that word fragment completion and masked word identification tests were stochastically independent under a variety of conditions. On this basis, they concluded that it is the overlap in the processes used, as well as differences in the information that they summon from memory, that differentiates memory tests. But there was also a significant conflict in the literature in the form of a study by Weldon (1991, Experiment 1). Weldon directly compared the masked word identification and fragment completion tests in her first experiment and found a gradient of priming for both tests, with the read condition showing more priming than the generate condition. Had her materials comprising the generation clues been antonyms, we would have expected this pattern (Jacoby, 1983; Masson &
168
Michael E. J. Masson and Colin M. MacLeod
MacLeod, 1992, Experiment 2). But she had used short definitional phrases, materials that we had shown (Masson & MacLeod, 1992, Experiments 1 and 7) produced equal priming for read and generate words in masked word identification. Therefore, we set out to determine the cause of these incongruent results. In a set of experiments (MacLeod & Masson, 1997b) using the procedures described in Experiment Series 1,we set out to compare directly the masked word identification and fragment completion tests under similar conditions. In our first experiment, we used Weldon’s (1991) materials to replicate Experiment 1 of Masson and MacLeod (1992). We noted that Weldon had required her subjects to respond to every test item in masked word identification (forced report), whereas we had permitted passing (free report), so we tested two groups of subjects, one with each type of report instruction. Our concern was that criterion differences might underlie the different patterns of results. But that was not the answer to the puzzle. As can be seen in Fig. 2A, both versions of the experiment produced results just like those of Masson and MacLeod’s Experiment 1: Performance for the two encoding conditions was equivalent in the free-report and forcedreport conditions. Thus, we replicated Masson and MacLeod’s finding, but failed to replicate Weldon’s. A discomforting possibility is that somehow subtle differences in the procedures being used in our respective laboratories were the source of the discrepancy in results. Because we also generally used different tasks, fragment completion in the case of Weldon and Roediger and masked word identification in our case, what looked like a task difference might actually be caused by some entirely different factor. In a sense, there was “their pattern,” with read > generate in amount of priming, and “our pattern,” with read = generate in amount of priming. The best way to discount this possibility was to carry out both tests in a single laboratory under identical conditions. When we did this (MacLeod & Masson, 1997b, Experiments 2 and 3), we obtained different patterns for the two tests. As indicated in Fig. 2B, masked word identification showed equal priming for read and generate encoding tasks, regardless of whether encoding task was manipulated within or between subjects. For fragment completion, shown in Fig. 2C, there was reliably more priming in the read condition than in the generate condition, both for within and between subject manipulations of encoding task. Thus, we consistently produced the read advantage for the fragment completion test and equivalent read and generate priming for the masked word identification test. Unfortunately, these results do not explain Weldon’s (1991, Experiment 1) finding of a read advantage for priming in the masked word identification test.
Episodic Enhancement of Processing Fluency
169
Fig. 2. Mean proportion correct on masked word identification and word fragment completion tasks as a function of encoding task. (A) Masked word identification performance under free and forced report instructions. (B) Masked word identification with encoding task manipulated within or between subjects. (C) Word fragment completion with encoding task manipulated within or between subjects. (D) Masked word identification with encoding tasks presented in blocked or mixed format. Error bars indicate the 95% within-subject confidence intervals for the means. Data are adapted from MacLeod and Masson (1997b).
We next carefully examined Weldon’s (1991) procedure to see whether there was some element that discriminated her experimental approach from ours. We found two differences. First, Weldon used blocked presentation of read and generate trials during encoding, whereas we had used mixed presentation. Second, she used four encoding conditions (the additional ones being auditory presentation and pictorial presentation), whereas we had always used only two encoding conditions. It seemed most plausible to us that blocking was the potent factor, with mixing perhaps encouraging a kind of “leakage” in encoding between read trials and generate trials (e.g., subjects might have tried to form a mental image of a generated target’s orthography). So, we conducted an experiment in which we blocked
170
Michael E. J. Masson and Colin M. MacLeod
the presentation of read and generate trials during encoding (MacLeod & Masson, 1997b, Experiment 4). In the subsequent masked word identification test, we once again found a pattern of equivalent priming for read (.28) and generate (.28) encoding conditions, almost perfectly reproducing the free-report condition in Experiment 1 of this series (priming in read = .26; generate = .27). These results showed that blocking alone was not the answer. Although it was hard to accept, we decided that the critical manipulation might not be a main effect, but rather could be an interaction. Perhaps it was the combination of multiple encoding conditions with blocking that caused Weldon’s (1991, Experiment 1) pattern for masked word identification to diverge from ours for the same task (Masson & MacLeod, 1992). In Experiments 5 and 6, we test this possibility directly. To the standard read and generate conditions, we added a third encoding condition: Read the word and produce an associate to it (e.g., for watermelon, the subject would say “watermelon” and then an associate such as “seeds”). In Experiment 5, the three encoding conditions were blocked; in Experiment 6, they were mixed. Figure 2D displays the results. When the three encoding conditions were blocked, there was more priming for the read condition (.14) than for the generate condition (.05);when they were mixed, priming was equal (read = .09; generate = 0.8). The answer, then, appears to be that the use of more than two encoding conditions in blocked presentation during study produces the pattern of priming that was observed in Weldon’s (1991, Experiment 1) masked word identification study. To summarize, the word fragment completion test reveals a highly consistent pattern, with more priming for words that were read than for words that were generated during encoding. It is important to realize, however, that generated words do produce priming quite reliably in word fragment completion; they simply produce less priming than read words. The masked word identification task also produces a very consistent pattern: Ordinarily, there is equivalent priming for read and for generated words, perhaps helping to explain why Witherspoon and Moscovitch (1989) found fragment completion and mask word identification to be independent. But there do appear to be exceptions in the case of masked word identification: When the generation rule involves antonyms (Jacoby, 1983; Masson & MacLeod, 1992, Experiment 2), there is little or no priming for generated words, and when there are more than two encoding conditions and these are blocked, there is a gradient of priming, with more for read words than for generated words. Taken together, these results convince us that these two implicit memory tests, certainly two of the most widely used tests in the implicit memory literature, ordinarily behave differently. We hypothesize that this has to
Episodic Enhancement of Processing Fluency
171
do with the different form of the test stimulus confronting subjects in the two tests. The whole word is presented in masked word identification; only part of the word is presented in word fragment completion. A whole word can summon both the perceptual and the conceptual aspects of the initial interpretive encoding, whereas a partial word primarily calls forth the perceptual elements of that encoding. Thus, the read condition is favored in fragment completion, which invokes a strategy of letter insertion and lexical search (cf. Nelson, Keelean, & Negrao, 1989, Experiment 4), but not in masked word identification, where access to both form and meaning is involved. Finally, we must admit that we do not have a compelling explanation for the two exceptions to the general finding of equivalent priming in masked word identification. We suggested in Masson and MacLeod (1992) that generation from antonyms may lead to an especially integrated encoding of the clue and the target words that then suffers at the time of test when the clue is no longer presented with the target in the generate condition. This decoupling of context for the generate condition impairs identification of generated words relative to read words. Such a contextual-integrative explanation probably could be developed for the multiple blocked encoding conditions exception as well, but it would be decidedly post hoc at this point. Instead of trying to explain both exceptions in the same way, it is important to note that it is the generate condition that is impaired in the antonym case, whereas the read condition shows somewhat less priming in the mixed encoding task situation, relative to the blocked encoding case (Fig. 2D). Perhaps, as we suggested in the MacLeod and Masson (1997b) article, subjects treat the words in the read condition in a more cursory, less conceptual fashion when they occur in a mixture of trials that includes other encoding conditions that emphasize conceptual processing. The solution to explaining these discrepancies from the general pattern remains to be seen, but at least we do know the standard patterns now. Moreover, we can see that the two implicit tests show different patterns of priming. Like Witherspoon and Moscovitch (1989), we take this as evidence that the tests invoke different collections of processes, consistent with a transferappropriate processing account. We have begun to suggest what those differences are, emphasizing through the idea of initial interpretive encoding the role played by conceptual processing in such indirect tests.
IV. Experiment Series 3: The Question of Conscious Recollection One of the most contentious issues associated with studies of automatic or unconscious influences of memory is the question of whether task perfor-
112
Michael E. J. Masson and Colin M. MacLeod
mance is mediated at least in part by conscious recollection. In the quest for pure measures of unconscious influences, this problem has been referred to by such terms as “conscious contamination.” The concern is that even in indirect tests of memory that do not instruct subjects to recollect specific prior events, the ability to consciously remember those events may in some way influence current task performance. If conscious recollection contributes to performance, any putative unconscious influence of memory is assumed to be contaminated. A widely used technique for addressing the problem of conscious influences on indirect tests of memory is to establish dissociations between indirect and direct tests of memory. A dissociation consists of demonstrating that an independent variable affects one of these two types of test, but not the other. Preferably, this dissociation is established under conditions in which the retrieval cues are the same for both the direct and indirect test. This definition of task dissociation is known as the retrieval intentionality criterion (Schacter, Bowers, & Booker 1989). The only difference between the two tasks is the instruction to the subjects: On the direct test, subjects are instructed to intentionally retrieve items from a prior study context, whereas on the indirect test no such instruction is given. For example, in an indirect test of memory involving word stem completion, subjects would be instructed to complete each stem with the first word that comes to mind; no instruction to remember previously presented words is given. On the direct version of this test, the same cues would be provided, but subjects would be instructed to use the cues consciously to remember the words that were studied earlier. If a factor such as modality of the presentation of the words during study affects the probability of producing target completions in one of these tests but not the other, the preferred conclusion is that different processes determined performance on the two tasks. In particular, such dissociations are used to argue that conscious recollection did not affect the indirect test. Although the establishment of task dissociations has played a significant role in the study of conscious and unconscious influences of memory, two important challenges to the task dissociation logic have been put forward. First, Dunn and Kirsner (1988) have shown that dissociations in which a variable affects performance on one test but has no effect on another test, or even a crossed double dissociation, in which a variable affects two tests in opposite ways, do not logically rule out the possibility that performance on the two tests is determined by a single process. If a single process can account for task dissociations, it is difficult to argue that one task involves a conscious influence of memory while the other involves an unconscious influence.
Episodic Enhancement of Processing Fluency
173
Second, Jacoby and his colleagues (Jacoby, 1991; Jacoby, Lindsay, & Toth, 1992; Jacoby, Toth, & Yonelinas, 1993) have argued that direct and indirect tests of memory are not process pure with respect to conscious and unconscious influences on test performance, even when task dissociations are established. To demonstrate this point, and to provide a means of estimating the contribution of conscious and unconscious influences of memory to task performance, Jacoby (1991) proposed the method of process dissociation. In this method, it is assumed that controlled (conscious) and automatic (unconscious) memory processes make independent contributions to task performance. A crucial aspect of the procedure involves placing controlled and automatic processes in opposition to one another, as explained below. By using this procedure to obtain estimates of controlled and automatic influences, it can be shown that these estimates are differentially affected by certain independent variables (e.g., divided vs full attention), thereby establishing process dissociations and supporting the assumption of independence ( Jacoby, Yonelinas, & Jennings, 1996). The process dissociation procedure is of particular interest here because it has been used to estimate quantitatively the automatic and controlled influences of memory for generate and read encoding tasks. More specifically, Toth, Reingold, and Jacoby (1994) found that the generate encoding task had no automatic memory influence on a word stem completion task. This result is a striking contrast to the findings we have described so far in this chapter and deserves careful consideration. Toth et al. (1994, Experiment 2) used short definitions as generation cues, just as Masson and MacLeod (1992, Experiments 1and 5 ) and MacLeod and Masson (1997b) have done. During a study phase, subjects generated some target words from their definition cues and read aloud other target words that were presented in isolation. In a subsequent word stem completion test, stems were tested under two different instructional conditions. In the inclusion condition, subjects were instructed to use the word stem as a recall cue in an attempt to remember the word from the study list that fit the stem. If a subject was unable to recall a previously studied word, he or she was to complete the stem with the first word that came to mind. On the assumption that controlled and automatic influences of memory operate independently to determine which word will be produced to complete a stem, the probability of completing a stem with the studied target under inclusion instructions can be expressed as
Z= C
+ (1 - C)A,
where C is the probability of consciously recollecting the target completion
174
Michael E. 1. Masson and Colin M. MacLeod
(controlled responding) and A is the probability of the target completion coming to mind automatically. In the second instructional condition, subjects were given exclusion instructions, whereby they were told to attempt to recall the studied word that would fit the stem, but not to give the word as a completion. Instead, subjects were to give a different word as the completion for that stem. Again, assuming independence between controlled and automatic influences of memory for the studied words, the probability of completing a stem with a studied word in the exclusion condition can be expressed as
E = (1 - C)A. According to this equation, a target completion will be given only if it comes to mind through an automatic influence of memory and is not consciously remembered. By obtaining the probability of completing word stems with target words in the inclusion and exclusion conditions, Toth et al. (1994) were able to compute estimates of controlled and automatic influences of memory by solving the two process dissociation equations. The mean proportion of word stems completed under inclusion and exclusion instuctions and the mean estimate of controlled and automatic influences of memory reported by Toth et al. are shown in Fig. 3A. The inclusion and exclusion task performance is of interest primarily for the sake of producing estimates of controlled and automatic influences of memory. But one aspect of the completion probability data is important for another reason. The completion probabilities for new items were not reliably different in the inclusion and exclusion conditions. This result is important because it was put forward by Toth et al. as evidence that subjects were not engaging in a generate-recognize strategy during stem completion, particularly in the exclusion condition. The validity of the process dissociation equations depends on the assumption that controlled and automatic influences of memory operate in the same way under inclusion and exclusion instructions. That is why, for example, subjects are instructed in the exclusion condition to attempt to recall studied words (then not to give those words as completions), just as they are instructed in the inclusion condition. There is concern, however, that subjects might operate according to a different agenda in the exclusion condition. In particular, they may rely on a generate-recognize strategy in which potential completions come to mind without the subject necessarily trying to remember words from the study phase. If the subject fails to recognize that a potential completion is a studied word, then he or she uses it to complete the stem. This characterization of performance in the exclusion condition is contrary to what is assumed to
Episodic Enhancement of Processing Fluency
175
Fig. 3. Mean word stem completion and estimates of controlled and automatic influences of memory based on the process dissociation procedure as a function of encoding task. (A) Data are adapted from Toth et al. (1994, Experiment 2). (B) Replication of Toth et al. adding the associate encoding task. (C) Second replication of Toth et al. using exclusion task instructions intended to induce a high criterion for excluding previously studied targets. (D) Application of the independence rememberlknow procedure to estimate controlled and automatic influences of memory. Error bars indicate the 95% within-subject confidence intervals for the means. Data in (B), (C), and (D) are adapted from Bodner et al. (1997). Inc. = inclusion; Exc. = exclusion; Cont. = controlled; Auto. = automatic; Rem. = remember.
happen by the equation for exclusion task performance given earlier. Thus, it is important to show that subjects did not engage in this approach to the exclusion task. Evidence that can be used in this cause consists of demonstrating that the probability of completing stems whose target completions were not studied with those nonstudied (new) targets is the same in the inclusion and exclusion conditions. The rationale is that if subjects had applied a generate-recognize strategy in the exclusion condition, they should be less likely to complete stems that correspond to new targets because of instances of false recognition. That is, some target completions in the new condition will falsely be classified as old and consequently withheld by the subject. This recognition check would not be invoked in the inclusion condition.
176
Michael E. J. Masson and Colin M. MacLeod
Finding lower completion performance in the exclusion condition than in the inclusion condition, then, would constitute evidence for the operation of a generate-recognize strategy in the exclusion condition. On the other hand, finding no difference between the completion rates for new items in these two conditions implies that subjects did not use the generaterecognize strategy. The completion rate for new items was not reliably different in the inclusion and exclusion conditions in the Toth et al. (1994) study, suggesting that subjects did not rely on a generate-recognize strategy. The results shown in Fig. 3A indicate that the generate task led to greater control in responding than did the read task, as subjects were much more sensitive to the effect of inclusion/exclusion instructions in the case of generate items. This differential sensitivity translates into a higher estimate of conscious influences of memory in the generate condition, which is computed as
C=I-E. The results in Fig. 3A clearly show that the conscious contribution of memory was estimated to be much larger following the generate encoding task than following the read task. In contrast, the estimates of automatic influences of memory showed the opposite pattern as a function of encoding task. The read encoding task produced a substantial estimate of automatic influences of memory, over and above the preexperimental influences as assessed by completion probabilities in the new condition. The generate condition, however, led to an automatic estimate that was no different from the completion probability for new items, indicating that there was no automatic influence of memory for the generate encoding episode. Toth et al. (1994) concluded that the automatic influence of memory on the stem completion task operated through data-driven processes. Because the generate task did not engage data-driven processing of targets, no automatic influence of those encoding episodes was observed. Toth et al. also suggested that demonstrations of enhanced word identification performance on tasks such as masked word identification following generate encoding tasks (e.g., Masson & MacLeod, 1992) were due to conscious recollection of target words and not to automatic influences of memory for the encoding episodes. We have two concerns regarding the conclusions reached by Toth et al. (1994). First, the generalization of their conclusion regarding the lack of automatic influences of memory for generate encoding episodes to word identification tasks other than stem completion is questionable. In the version of the stem completion task used by Toth et al., only the first three letters of a target word were provided. This kind of retrieval cue may
Episodic Enhancement of Processing Fluency
177
not provide enough constraints to permit the recruitment of conceptual knowledge about a target word, including episodic memory for conceptual processing operations performed during the generation task. In the masked word identification task, however, the target word is in full view, although for only a brief time. Evidence suggests that even under these conditions, the brief exposure of the entire word is capable of making contact with conceptual knowledge associated with that word. For example, semantic or associative priming effects have been found with brief, masked presentations of primes (Carr & Dagenbach, 1990;Dagenbach, Carr, & Wilhelmsen, 1989; de Groot, 1983; Lukatela & Turvey, 1994). Even if it is the case that stem completion does not present an opportunity for memory for conceptually driven processing episodes to influence task performance automatically, the masked word identification task may do so. Our second concern is that the estimate of automatic influences of memory following the generate encoding task, as determined by the process dissociation procedure, may not be accurate. Various criticisms of the procedure and the estimates it yields have been raised, particularly with respect to the assumption that automatic and controlled contributions of memory operate independently (e.g., Curran & Hintzman, 1995; Joordens & Merikle, 1993). Our concern is not with the question of independent processes and we are willing not to question that assumption for the time being. Rather, we suspect that the estimate of automatic influences of memory is underestimated by the process dissociation procedure, particularly when assessing the effect of conceptually driven encoding tasks, because of what Richardson-Klavehn and colleagues have referred to as involuntary conscious memory (e.g., Richardson-Klavehn & Gardiner, 1995; Richardson-Klavehn, Gardiner, & Java, 1996). Involuntary conscious memory involves two components. The first is an automatic influence of memory, perhaps causing a potential word stem completion to come to mind. The second component is a subsequent awareness that that word had appeared in the study phase of the experiment. Should this process occur during word stem completion under exclusion instructions, subjects very likely will withhold that word and seek an alternative completion. Notice that by the usual meaning of an automatic influence of memory, generation of a completion under these circumstances is a valid instance of an automatic influence of memory. The subsequent involuntary awareness of that word’s prior occurrence, however, is alleged to mask that influence. The result is an underestimate of the true automatic influence of memory. Moreover, the effect of involuntary conscious memory is likely to be stronger for words encoded with conceptually driven study tasks because these tasks most strongly promote conscious recollection.
178
Michael E. J. Masson and Colin M. MacLeod
Although Toth, Reingold, and Jacoby (1995; Reingold & Toth, 1996) have argued that there is no clear evidence that involuntary conscious memory has a substantial influence on estimates of automatic influences of memory obtained with the process dissociation procedure, we have recently conducted a series of experiments that justifies this concern (Bodner, Caldwell, & Masson, 1997). Our first experiment was a replication and extension of the Toth et al. (1994, Experiment 2) study in which generate and read encoding tasks were compared using a word stem completion task. We extended that original study by including a third encoding task that was expected to be revealing with respect to the purported influence of involuntary conscious memory. In that condition, called the associate condition, subjects read aloud a target word just as in the read task, but then also reported the first word that came to mind upon reading the target word. It was intended for the associate task to provide subjects with a datadriven processing experience identical to that in the read condition. Furthermore, the production of an additional, probably related, word was expected to add a conceptually driven processing component to the encoding task. Our expectation was that if process dissociation procedure estimates are unaffected by involuntary conscious memory, then the associate and read encoding tasks should lead to equivalent estimates of automatic influences of memory. On the other hand, if the estimate of automatic influences can be artifactually lowered by involuntary conscious memory, because of its conceptual component the associate condition would be more strongly subject to that influence than would the read condition. As a result, the associate condition should produce a lower estimate of automatic influences, perhaps as low as that found in the generate condition whose estimated automatic influence also is assumed to be powerfully affected by involuntary conscious memory. Although in our first experiment we used inclusion and exclusion instructions similar to those used by Toth et al. (1994), we made one modification to the exclusion instructions. Whereas the exclusion task instructions used by Toth et al. appear to be vague with respect to whether the subject is permitted knowingly to use a studied word to complete a stem, we directly informed subjects that they were not to use previously studied words as completions. The mean proportion of stems completed with target words under inclusion and exclusion instructions and the corresponding estimates of controlled and automatic influences of memory are shown in Fig. 3B. Performance under inclusion instructions for items in the generate and read encoding conditions was similar to that reported by Toth et al., and items in the associate condition led to the highest proportion of target completions. This result is consistent with the idea that the associate condi-
Episodic Enhancement of Processing Fluency
179
tion bestowed the advantages of both conceptually driven and data-driven components upon the encoding episode. Somewhat to our surprise, however, we were unable to replicate the results reported by Toth et al. (1994, Experiment 2) under exclusion instructions. Instead, we found that subjects rarely used target words as stem completions, even in the read and associate conditions, which should have produced a strong automatic influence of memory. The success of subjects in our experiment regarding exclusion of target completions might be taken as evidence that they used a generate-recognize strategy when under exclusion instructions. Were that the case, application of the process dissociation equations would be rendered invalid. The evidence on this question is suggestive. First, the completion rate for new items was somewhat different in the inclusion (.36) and exclusion (.32) conditions, a difference that approached statistical significance. Second, the estimates of automatic influence of memory in the generate and associate conditions were both reliably lower than performance in the new condition. That latter result is anomalous in the sense that valid estimates of automatic influences should not drop below baseline. These anomalous estimates suggest that subjects were relying on a generate-recognize strategy. A similar result was obtained by Richardson-Klavehn, Gardiner, and Ramponi (1996) in a replication of a different experiment from Toth et al. (1994), in which level of processing in the study phase was manipulated. Richardson-Klavehn et al. used exclusion instructions in which subjects were instructed to use the stems to recollect studied words but not to give those words as completions for the stems. Subjects were also clearly told that they should give no response in the exclusion condition if only a studied word completion came to mind. Richardson-Klavehn et al. found that stem completion performance under exclusion instructions was very low, just as did Bodner et al. (1997). They concluded that the higher completion rate observed by Toth et al. (1994, Experiment 1) under exclusion instructions was a result of subjects not consistently following instructions. In a second experiment, we persevered in an attempt to replicate the Toth et al. (1994, Experiment 2) results, while including the associate condition. We hypothesized that subjects in the Toth et al. study may not have been as thorough in withholding studied words as completions under exclusion instructions because they had set a higher criterion for deciding that a potential completion was an old word. Further, we supposed that under exclusion instructions, subjects would be unlikely to seriously attempt direct retrieval of old words using word stems as cues. Rather, we considered it likely that subjects would generally follow a generate-recognize strategy in which potential completions would be rejected if they seemed to be old words. Therefore, in Experiment 2 of Bodner et al. (1997), we tested the
180
Michael E. J. Masson and Colin M. MacLeod
possibility that the results reported by Toth et al. were the product of a generate-recognize strategy in the exclusion condition, combined with a high criterion for deciding that a completion was an old word. We used exclusion instructions that told subjects to provide the first completion that came to mind, unless they were sure that word was an old one. If a potential completion was considered to be an old word, a different completion was to be provided. We expected that if the Toth et al. (1994, Experiment 2) results were due to undetected use of a generate-recognize strategy, our Experiment 2 should replicate the pattern of data obtained by Toth et al. A successful replication, however, could also arise if, despite providing instructions that encourage a generate-recognize strategy, subjects fail to adopt that strategy. The associate encoding condition plays an important role in resolving this potential ambiguity. If subjects operate as assumed by the process dissociation equations, rejecting under exclusion instructions only those completions obtained through direct retrieval, the estimate of automatic influences of memory should be the same for the associate and read conditions. In contrast, if subjects actually use a generate-recognize strategy with a high recognition criterion, the estimate of automatic influences in the associate condition should be reduced relative to the read condition. At the same time, the other markers of the generate-recognize strategy would be absent (i.e., differential completion of new items across inclusion and exclusion instructions, automatic estimates falling below baseline) because the high recognition criterion would prevent exclusion scores from approaching zero and would make rejection of completions for new items unlikely. This outcome would suggest that the Toth et al. results were also due to an undetected influence of involuntary conscious memory operating under exclusion instructions. The results of our second experiment are shown in Fig. 3C. Performance under inclusion and exclusion instructions in the generate and read encoding conditions was very similar to that reported by Toth et al. (1994, Experiment 2). The estimates of controlled and automatic influences of memory also replicated the pattern found by Toth et al. There was a higher estimate of controlled processing for the generate than for the read condition, but the reverse was true for the automatic estimate. Despite the use of generaterecognize instructions, the data from the generate and read conditions showed no evidence that the boundary conditions for applying the process dissociation equations had been violated. The stem completion rates for new items under inclusion and exclusion instructions were not significantly different from each other and the automatic estimate for the generate condition was not significantly different from the completion rate for new items.
Episodic Enhancement of Processing Fluency
181
Although the generate and read encoding conditions replicated the Toth et al. (1994) results, the associate condition revealed a set of estimates that clearly indicates subjects were influenced by involuntary conscious memory. The controlled estimate for the associate condition was higher than for either generate or read, indicating that subjects consciously remembered a substantial number of words encoded in that condition. More important, the automatic estimate for the associate condition was significantly lower than the estimate for the read condition, and not different from the generate estimate or from the completion rate for new items. Thus, despite applying the same data-driven encoding operations as in the read condition, the associate condition failed to produce evidence for an automatic influence of memory. This result supports the hypothesis that encoding tasks that involve substantial conceptually driven processing, such as the generate and associate tasks used here, are subject to much larger effects of involuntary conscious recollection on later indirect tests of memory than are encoding tasks that do not involve such processing. Even though a subject may experience a word completion as coming to mind automatically, without deliberate recollection, he or she may subsequently recognize that word as having appeared in the study phase and therefore withhold it when completing its stem under exclusion instructions. This situation creates a serious problem when applying the process dissociation procedure and equations to obtain estimates of automatic influences of memory. In particular, automatic influences can be greatly underestimated for encoding tasks that include a strong conceptually driven processing component. The results of the second experiment of Bodner et al. (1997) may be criticized on the grounds that subjects were induced to use a generaterecognize strategy in the exclusion condition and therefore it was inappropriate to apply the process dissociation equations to obain estimates of automatic and controlled influences of memory. Our counterargument is that although the instructions we used encouraged such a strategy, none of the behavioral markers of that strategy was apparent in the data, just as was the case in Toth et al. (1994). We suggest that the Toth et al. results may have arisen from undetected application of a generate-recognize strategy. Consequently, the parameter estimates they obtained are no more valid than those obtained in our second experiment. We sought to reinforce our argument in a third experiment by using a procedure that does not involve testing subjects under exclusion instructions. Jacoby et al. (1996) developed an alternative technique for estimating the independent contributions of automatic and controlled processes in remembering prior events based on the rememberknow distinction developed by Tulving (1985) and Gardiner (1988). In its original formulation,
182
Michael E. J. Masson and Colin M. MacLeod
remembedknow judgments were applied during a recognition memory test in which subjects were asked to classify into two categories items they recognized as having been studied previously: those for which some aspect of the original encoding episode could be remembered and those for which no such memory was available but the subject simply “knew” that the item had been studied. In the procedure developed by Jacoby et al. (1996), called the independence remembedknow (IRK) procedure, subjects are given direct retrieval instructions, like those used in the inclusion component of the standard process dissociation procedure. Subjects are told that if they cannot remember a word from the study list, they are to complete the stem with the first word that comes to mind. Once a completion is provided, the subject then classifies the completion into one of three categories: remember (R), know (K), and new (N). These categories are defined as in the remembedknow procedure. To obtain estimates of controlled and automatic influences of memory, the proportion of remember responses is treated as a pure estimate of controlled influences of memory (i.e., C = R). The estimate of automatic influences is derived by combining K and N responses, and dividing by 1 - R (i.e., A = [K + N]/[1 - R]). The relation between these equations and the usual process dissociation equations is readily seen if one takes the combination of all three categories as an approximation to inclusion performance (R + K + N) and the combination of know and new categories as an approximation to exclusion performance (K + N). In the third experiment of Bodner et al. (1997), subjects were tested using the IRK procedure following a study phase in which words were encoded using the generate, read, and associate tasks as before. We expected that the estimates of controlled and automatic influences of memory would follow the pattern seen in the second experiment, in which generate and associate encoding conditions would yield high controlled estimates but very low automatic estimates that are not different from the baseline stem completion rate. The proportions of stems that were completed and subsequently assigned to each of the three response categories are shown in Fig. 3D. The generate and associate conditions produced a similar profile of responding across the three categories, with a much higher proportion of stems completed and assigned to the remember category than was the case for the read condition. In contrast, items in the read condition were more likely to be completed and assigned to either the know or new categories than was the case for items in the generate and associate conditions. The corresponding estimates of controlled and automatic influences of memory are also shown in Fig. 3D. The pattern of these estimates replicated the results from our second experiment. The controlled estimates for the generate and associate conditions were substantially higher than for the
Episodic Enhancement of Processing Fluency
183
read condition, but the reverse was true for the automatic estimate. The automatic estimate for the associate condition was even significantly lower than the completion rate for new items. This outcome was primarily due to a few subjects assigning all completions in the associate condition to the remember category, thereby producing an estimate of zero for the automatic parameter. Exclusion of these subjects from the analysis produced an automatic estimate of .24, which was not significantly different from the completion rate for new items. The experiments using the process dissociation procedure were successful in replicating the original Toth et al. (1994, Experiment 2) result, which has been taken as evidence that the generate encoding task does not lead to any automatic influence of memory on stem completion. But these experiments also call into question the validity of that conclusion because they demonstrate that another encoding task, one that has both conceptually driven and data-driven processing components, shows the same pattern of estimates of automatic influences as the generation task. This result is paradoxical on the view that automatic influences of memory on the stem completion task are based on data-driven processes. The paradox is readily resolved, however, by the observation that estimates of automatic influences produced by the process dissociation procedure can be seriously underestimated because of involuntary conscious recollection. We argue that when operating under exclusion instructions or when assigning stem completions to remember, know, or new categories, subjects often have the experience of a completion automatically coming to mind, followed by awareness that the completion had appeared in the study phase. This involuntary conscious remembering of prior episodes is systematically excluded from estimates of automatic influences of memory by the process dissociation procedure. The exclusion of such instances from consideration as examples of automatic influences of memory leads to the observed underestimate of those influences. The estimates of automatic influences of memory derived from the process dissociation procedure are more properly considered estimates of involuntary memory influences that are not accompanied by awareness. Encoding tasks that vary in the amount of conceptually driven processing they require will quite naturally lead to differential estimates of this subclass of automatic influence; greater conceptually driven processing at encoding is going to reduce the likelihood that subsequent involuntary influences of memory will go forward without themselves prompting awareness of the relevant prior experience. Awareness of prior occurrence that arises after the automatic or involuntary influence of memory has occurred is most often not regarded as grounds for ascribing that influence to conscious recollection. Rather, by most accounts, these instances of involuntary conscious memory are considered
184
Michael E. J. Masson and Colin M. MacLeod
valid examples of an automatic influence of memory on task performance. The difficulty encountered in applying the process dissociation procedure is that these instances are inadvertently attributed to conscious recollection. Application of the procedure therefore creates the likely possibility that the automatic influence of memory for conceptually driven encoding episodes, such as those created by the generation task, are substantially underestimated. Although the pattern of estimates obtained by Toth et al. (1994) fits well with the transfer-appropriate processing view, at least under the assumption that the stem completion task relies exclusively on data-driven processing, the Bodner et al. (1997) results cast doubt on those estimates. Those results do not, however, cause a problem for the general transfer-appropriate processing framework. Rather, we see the results as fitting with that framework and as constituting a warning about the dangers of classifying memory tests as exclusively or even predominantly determined by either data-driven or conceptually driven processes. We suggest that in many indirect tests of memory that involve identification of a target stimulus, both types of process are invoked, and that prior memory episodes involving either process in the course of identifying a target can enhance subsequent identification of that same item.
V. Experiment Series 4: Speeded Word Reading As an Indirect Measure of Memory Aside from our work with the process dissociation procedure, we have pursued another approach to address the question of whether repetition priming effects we have found on word identification tasks following a generation episode could be the result of conscious recollection. Consider first how conscious recollection could affect performance in a task such as masked word identification. When subjects have difficulty identifying a masked word, they may resort to trying to remember words from the study phase. Given the much better conscious recollection of generated words as compared to read words (Slamecka & Graf, 1978), subjects are more likely to succeed in recollecting generated words, which then augment performance in the generate condition. It is this augmentation that leads to equal amounts of priming in the read and generate conditions. On this account, were it not for that extraneous influence, substantially more priming would be observed in the read condition than in the generate condition; indeed, the generate condition might produce no priming at all (e.g., Toth et al., 1994).
Episodic Enhancement of Processing Fluency
185
In Experiment Series 4 (MacLeod & Masson, 1997a), we tested this proposal by moving to a word identification task that would be very unlikely to engage conscious recollection processes. Our goal was to select a task for which interpretive encoding would be the dominant process, just as we propose is the case for the masked word identification task, then to use this task to confirm the equivalent priming we saw in masked word identification for read versus generated words. In so doing, we would both generalize the findings we described in Experiment Series 1 and 2 and help to dispel concerns that the pattern we observed hinged on conscious recollection selectively assisting the generate condition. Ideally, this new task would be one in which the test word would be presented without any form of degradation, and where its processing would be rapid and seemingly effortless. A near perfect candidate to meet these criteria is speeded oral word reading, a task often called “naming” in the literature (e.g., Balota & Chumbley, 1984; Forster & Chambers, 1973). The subject’s task is simply to read aloud as quickly as possible a clearly presented word, and the dependent measure is latency to read. There is no problem solving, and the task is certainly not difficult. In fact, based on the interference that occurs in color naming in the Stroop (1935) task when the subject must say “red” to the stimulus word green printed in red, it is frequently held that word reading is automatic, in the sense that it requires little attention (see MacLeod, 1991, for a review of the Stroop literature). Word reading thus seems to be an ideal implicit measure of memory, a claim supported by existing research (e.g., Carr, Brown, & Charalambous, 1989; Logan, 1990). Word reading should not promote conscious recollection both because the task is extremely easy and because it is performed very quickly. With this logic in mind, we set out to extend our studies of masked word identification to speeded oral word reading (MacLeod & Masson, 1997a). These new experiments adopted the same study procedure as in the masked word identification experiments, but substituted a word reading test as the indirect test of memory. Note that because the measure is now latency to read words, priming will now appear in the form of shorter bars in the studied conditions. The first experiment in this series used the definition materials that Weldon (1991) had used, and that we had borrowed for our Experiment Series 2 (MacLeod & Masson, 1997b). As can be seen in Fig. 4A, there was reliable priming for both read and generate words, and that priming did not reliably differ across those conditions, precisely the pattern we have seen previously for masked word identification. When we switched to antonyms as our materials in Experiment 2, we also found equivalent priming for read words and generated words that was similar in magnitude to the priming observed with definitions (see
186
Michael E. J. Masson and Colin M. MscLeod
Fig. 4. Mean speeded word reading time as a function of encoding task. The generation cues were (A) definitions or (B) antonyms. Error bars indicate the 95% within-subject confidence intervals for the means. Data are adapted from MacLeod and Masson (1997a).
Fig. 4B). Intriguingly, then, even antonyms, which showed no priming for generated words in masked word identification, now showed equivalent priming whether they had been read or generated during encoding. In Experiment Series 1, where masked word identification was the test (Masson & MacLeod, 1992), we explained the absence of priming for words generated from antonym clues as possibly being due to the integrated encoding of clue and target, which hampered the target when it stood alone at the time of test. Perhaps the automaticity of speeded word reading minimizes the relevance of this integration at encoding because now retrieval is trivially easy because the target word is clearly in view. In the third experiment in the speeded word reading series, we removed the generate condition and replaced it with a condition we had used in Experiment Series 2 and 3: the associate condition, in which subjects read the target then generated an associate. The rationale for making this substitution was that now the subject would have the same perceptual experience in both encoding conditions, but more apparent emphasis on conceptual encoding in the associate condition. If we are correct, however, that both
Episodic Enhancement of Processing Fluency
187
perceptual and conceptual processing routinely contribute to the initial interpretive encoding of each studied word anyway, then this change should not alter the equivalence of priming. Indeed, as shown in Fig. 4C, we obtained similar amounts of priming for the associate condition (50 ms) and the read only condition (46 ms), further generalizing the finding. Given the consistency with which these experiments produced equal amounts of priming across various study conditions, one might doubt that speeded word reading is a good measure of priming, in that it perhaps resonates consistently to any prior experience with an item. To address this possibility, we turned to one of the manipulations that has been most reliable in producing differential repetition priming effects, that of modality (e.g., Bassili, Smith, & MacLeod, 1989; Roediger & Blaxton, 1987; see Richardson-Klavehn & Bjork, 1988; Roediger & McDermott, 1993, for reviews). It has been shown repeatedly that changing the modality (e.g., from auditory to visual) between study and test reduces priming' sharply and sometimes eliminates it. This effect may have to do with the decreased contribution of perceptual aspects of the encoding when modality shifts. In any event, modality change virtually always decreases priming. We chose to manipulate modality in our paradigm to determine whether the speeded word reading measure would be sensitive to a manipulation that routinely affects other implicit tests. Subjects either saw a word or heard it during study, with the auditory and visual encoding conditions blocked for ease of administration. As Fig. 4D shows, the modality manipulation was successful in producing differential priming. With the word reading test involving visual presentation of the targets, words studied in the visual modality produced significantly more (about twice as much) priming as those studied in the auditory modality, though auditorily presented words still produced reliable priming (as in many previous studies, such as Bassili et al., 1989). Word reading is, therefore, a sensitive measure of implicit remembering. More recently, one further result has been obtained to complement the findings of Experiment Series 4. MacLeod and Daniels (1997) manipulated two variables at encoding prior to a speeded word reading implicit test. One of these was the now familiar read versus generate manipulation. Factorially combined with this was a manipulation of whether each word was followed by an instruction to remember the word for a later memory test or to forget it. This latter manipulation, called directed forgetting, is known to have a powerful effect on explicit remembering (for a review, see MacLeod, in press). The materials were the Weldon (1991) definition items of Series 2. The question was how these two variables would affect explicit remembering (free recall) as opposed to implicit remembering
Michael E. J. Masson and Colin M. MacLeod
188
(speeded word reading). Figures 5A and 5B provide the answer. Figure 5A shows that directed forgetting had a strong effect on recall of words read during study, but not on words generated during study. Apparently directed forgetting cannot affect an item that is already richly encoded. But for present purposes, the more crucial result concerns the speeded word reading implicit measure. Once again, Fig. 5B shows that all studied conditions led to priming, but that there was no differential priming as a function of how the word was encoded. Neither manipulation, despite its renowned strong impact on explicit measures, had any impact on priming in speeded word reading. Experiment Series 4 (MacLeod & Masson, 1997a) plus the MacLeod and Daniels (1997) study show us that the masked word identification pattern is a general one. At least one other implicit measure, speeded word reading, shows the same equivalence in priming for read and generated words as does masked word identification. Furthermore, speeded word reading would appear to be a very good implicit measure in that the task does not foster a problem-solving set that in turn could lead to conscious recollection in the face of failure to identify a particular target word. Instead, the task is easily and quickly performed, perhaps automatically. It is even conceivable that conscious recollection would interfere with the smooth performance of word reading, slowing it down. Yet, latencies in our experiments are very similar to those seen in typical word reading studies, and certainly do not seem unusually elevated. These results make arguing for intrusion of conscious recollection into performance under the generate condition much less feasible.
;0.5 8 0.4
590 2.
2
2
0.3 0 .-
5 Q
2 n
580
0
c
2
(II 0)
2
0.1
0.0
570
c
0.2
560
550
Generate Read Encoding task
Generate Read Encoding task
New
Fig. 5. Mean proportion of words correctly recalled and mean speeded word reading latency following remember and forget instructions in a directed forgetting paradigm. Error bars indicate the 95% within-subject confidence intervals for the means. Data are adapted from MacLeod and Daniels (1977).
Episodic Enhancement of Processing Fluency
189
VI. Experiment Series 5: Color Naming versus Word Reading and the Specificity of Priming
Our view of the encoding-retrieval interaction that constitutes remembering is rooted in the idea of transfer-appropriate processing (Morris et al., 1977; Roediger, 1990). For transfer to be appropriate, there must be at least some correspondence between the processes engaged by the initial episode and those engaged by the subsequent episode. We are reminded of Osgood’s (1949) “transfer surface” and the notion that transfer varies as a function of a number of factors relating to the stimuli and responses. In our case, the “surface” of transfer would be dictated not by the stimuli and responses, however, but by the processes involved. An obvious question concerns just how related the processes on the two occasions must be for transfer to occur. In the series of experiments to be described in this section (Mackod, 1996), two different implicit tests were compared. The first test was speeded word reading, familiar from Experiment Series 4. The second was color naming, modeled after the classic Stroop (1935) task. Color naming has previously been used to measure semantic priming in studies by Warren (1972) and others (for a review, see MacLeod, 1991, pp. 173-174). Warren showed subjects three words from a category (e.g., aunt, uncle, cousin) and then had them name the print color of a word that could be from either the presented set (e.g., uncle), their superordinate (e.g., relative), or an unrelated control (e.g., horse). He showed that subjects were slowed in naming the color of print when the meaning of the target word printed in color had been primed by immediately prior study. Initially, the goal of the experiments in this section (reported by MacLeod, 1996) was to try to develop an implicit measure in which the priming from prior study of a word was expressed as interference rather than facilitation. Although the results turned out rather differently than Warren’s (1972) work suggested they would, they are nevertheless informative about the specificity of priming. In the first experiment, subjects studied 48 high-frequency words under instructions to remember them for a later test, characterized as an explicit test. Following study, subjects were tested on 48 words in a speeded word reading test and 48 words in a color naming test, with test order counterbalanced. On each 48-word test, there were 24 studied words and 24 new words, with no overlap in materials between the two tests. For both tests, the words were presented in color, with subjects instructed to ignore the color in the word reading test and to ignore the word in the color naming test. To discourage the use of conscious recollection during these tests, subjects were told that the tests were filler activities before the real memory test.
Michael E. J. Masson and Colin M. MacLeod
190
As shown on the left side of Fig. 6A, the word reading test revealed reliable facilitation, with the studied words responded to on average about 9 ms faster than the new words. There was, however, no corresponding enhancement of interference: The colors of studied words were, if anything, named slightly faster than those of new words. The same pattern was evident in a second experiment using low-frequency words, which generally show more priming (e.g., MacLeod & Kampe, 1996). Word reading showed facilitation of about 15 ms (right side of Fig. 6A), but color naming again showed no interference. An experiment not included in the MacLeod (1996) article extended this result by repeating Experiment 2 but with the addition of some standard incongruent Stroop trials (e.g., red printed in the color green, say “green”). Because such trials always cause substantial interference, the idea of includ-
660c
A High Frequency
I
LOW
Frequency
Read
4
1
Color word
2
8
760 740
c 720 g 700 680 660
Word
Word reading
Color Word Color Response attribute
C
Color naming
D ,
I
I
Encoding
Encoding
Fig. 6. Mean response latency on speeded word reading and color naming tasks with word stimuli printed in color as a function of prior exposure. (A) Word reading and color naming tasks with high- and low-frequency words. (B) Word reading and color naming tasks, including a set of color words as targets. (C) Color naming task, including control items consisting of rows of Xs.(D) Color naming task. Error bars indicate the 95% within-subject confidence intervals for the means. Data in (A) and (C) are adapted from MacLeod (1996), data in (D) are adapted from MacLeod and Masson (1997a).
Episodic Enhancement of Processing Fluency
191
ing them was to highlight the presence of words, perhaps drawing more attention to the noncolor words as well. As shown in Fig. 6B, however, although color naming responses on the incongruent Stroop trials were about 100 ms slower than on the other trials, indicating the usual Stroop effect, the studied words still did not interfere more than the new words. This was true despite the strong facilitation of word reading for studied words as compared to new words. The third experiment in the published series examined only color naming. This time, the number of intervening items as well as the time between the study of a word and its test were reduced by using multiple “minilists” of nine studied words and nine color naming test items. Of the nine test items in each minilist, three were studied words, three were new words, and three were control items (i.e., XXXX).The purpose of these modifications was to give priming more of a chance to express itself by moving the study and test of individual items closer together, and to determine whether noncolor words did in fact interfere with color naming relative to a neutral baseline. The color naming results shown in Fig. 6C indicate that although both studied words and new words did interfere substantially relative to the control, once again there was no enhancement of interference for studied words. To bridge the experiments in the previous series with those in this series, we reintroduced our standard read versus generate manipulation and followed our standard procedure. This experiment is Experiment 2 in the MacLeod and Masson (1997a) set, for which the word reading data have already been described in the previous section. Recall that subjects showed 12 ms of priming for generated words and a virtually identical 15 ms of priming for read words, relative to new words. But these subjects also did a color naming task, with the order of the two tests counterbalanced. As shown in Fig. 6D, there was no interference for either generated words or read words relative to new words. Once again, word reading showed facilitation and color naming showed no effect due to priming. Taken together, then, the MacLeod (1996) experiments and the related experiments just described show that primed words facilitate speeded word reading but do not interfere with color naming. The failure to find increased interference in color naming for studied words is not an isolated one, as Burt (1994) has reported similar findings. These findings have a number of implications, but only one is especially relevant here. Apparently, studying a word in all of these experiments primed that word sufficiently to produce facilitation in word reading. Of course, word reading at test is essentially the same task as word reading at study, so tranfer would be expected. But color naming at test has little in common with word reading at study in the way of shared processes. Thus, it would appear that the priming attached
192
Michael E. J. Masson and Colin M. MacLeod
to the words during study is specific to reading those words, and does not spill over onto a color naming task where word reading is irrelevant and indeed discouraged by the instructions to subjects. The explanation offered by MacLeod (1996) was that priming is process specific. Priming will ordinarily be expressed as facilitation on only a very similar processing task, and should not appear as interference with another task, even one that involves the same studied materials. It is not the materials that are primed, but rather the processes that were applied to those materials on initial contact. This argument is supported by another related set of data. MacLeod, Grant, and Ridout (1996) report three experiments in which they had subjects study a list of words and then take part in a flanker task (see Eriksen, 1995, for a review of this paradigm). In this task, there were always two stimuli on each test trial, one above the other. The target, varying over trials from top to bottom, was enclosed in arrowheads; the subject’s task was to decide whether the target was a word. This target could be either a studied word, a new word, or a nonword. The adjacent flanker could be either a studied word, a new word, a nonword, or a neutral control row of Xs. Although targets benefited from prior study (by about 50 ms), flankers were totally unaffected. Priming a flanker did not enhance the interference it caused with the decision about the target. These two studies both suggest that priming is specific not to the studied item (or its representation), but to the processes in play when that item was analyzed upon its initial occurrence. To the extent that these processes are recruited at test, as when a word that was read during study must be read again at test or must have its lexical status judged (which necessitates reading it), primed words will cause facilitation. But when attention at test is directed to a dimension of the stimulus display other than the primed one, for example, to the color of print of a studied word or to an adjacent target, priming is irrelevant. Indeed, in such cases, priming would have a negative impact, so it is probably good that priming does not routinely affect elements of the stimulus display other than the target. Our goal in introducing this set of experiments that contrasts color naming and word reading was to show the specificity of the processes involved in priming. These experiments nicely illustrate that it is not words or their representations that are primed, but rather the operations that were applied to those words during the previous encounter. If the words were read at study and must be read again at test, then facilitation will appear because of the repetition of the reading operations. If, however, something else must be done to the words upon second encounter, then the prior processing may not be relevant and so priming will not be evident. We certainly intend to explore this (hyper)specificity more, but for the present only wish to note that it is entirely consistent with the explanation we have been elucidat-
Episodic Enhancement of Processing Fluency
193
ing here. The initial interpretation records how the stimulus was processed. When this encoding is summoned later because of similar processing of the same stimulus, priming will appear in the form of facilitation. That priming will not leak through to disrupt processing on another stimulus dimension, thereby protecting us from undesirable interference when a stimulus recurs in a different setting that requires different types of processing.
VII. Experiment Series 6: Sources of Priming in Masked Word Identification In a recent special issue of the Canadian Journal of Experimental Psychology devoted to implicit memory, we (Masson & MacLeod, 1996) tackled
a question that had become a growing concern to us in our studies of priming in masked word identification. Put simply, the question is this: Is the priming we have observed repeatedly for studied words in masked word identification the result of increased efficiency in extracting perceptual information from studied words (i.e., sensitivity in signal detection theory terms), or does that priming result from some kind of bias favoring studied words? Our interest in this question was spurred by the recent work of Ratcliff and McKoon (1993; Ratcliff, McKoon, & Verwoerd, 1989). They had subjects decide which of two probe items was identical to a just-presented masked target. Using this two-alternative forced choice (2-AFC) test, they found that prior exposure to the masked target item did not increase accuracy of recognition, arguing against the view that priming is the result of increased perceptual clarity for studied targets. Instead, Ratcliff et al. (1989) suggested that a perceptual bias was operating. They based this argument on the fact that, when tested with orthographically similar test pairs (e.g., lied, died), one of which was old (i.e., had been previously seen), subjects tended to choose the old alternative regardless of whether the masked target word was old or new. Ratcliff et al. (1989) went on to propose that this perceptual bias happened early in perception, and cited two main reasons for this claim. First, response times were very fast, leaving little room for strategic processing. Second, subjects said that the old words seemed to “pop out.” Faced with a degraded stimulus as in masked word identification, subjects prefer to select an old word when the possible responses are similar. That this is a bias was underscored by the finding in their Experiment 5 that there was no bias in favor of old words when the two alternatives were not orthographically alike (e.g., died, love).
194
Michael E. J. Masson and Colin M. MacLeod
We wondered whether the free report procedure that had been used in our previous experiments might be harboring such a bias toward old, studied words or whether the old words were actually seen better, as argued by Reinitz and Alexander (1996). To address this question, we conducted a series of seven experiments (Masson & MacLeod, 1996). In all of these, subjects first studied a set of words by reading them; the generation encoding task was not included here because our question was about why priming occurs in masked word identification, and did not require comparing two encoding conditions. Subjects then were given one or two forms of masked word identification test. There were three different forms of test in all. One form of the test was the “free report” version of identification that we have described in prior sections of this chapter: Subjects saw the masked word and then tried to say what it was. Regardless of whether priming results from sensitivity or bias, free report should lead to better performance for studied words. To discriminate these two explanations, we used two other testing procedures following the masked word. One of these was a 2-AFC test and the other was a single probe test. Incorporation of the single probe test allowed us to examine whether any bias observed on the 2-AFC test stemmed from strategic use of discriminative information in the two choices, a strategy that could not operate in the case of a single probe. We first describe the pattern of results over the entire series of experiments, then return to what those results tell us about sensitivity versus bias. Experiment 1 was essentially a replication of the Ratcliff et al. (1989) study with one key change. Except for a baseline condition, Ratcliff et al. had one old and one new alternative on every trial, whether the masked target was old or new. This was what allowed them to detect the bias favoring old items. We chose instead to prevent this particular bias from operating by always having both alternatives on the recognition test have the same study status, either both old or both new. Like Ratcliff et al., our test pairs were orthographically similar, differing by only one internal letter (e.g., cage, cave).Subjects studied 80 words (consisting of 40 pairs of orthographically similar words) one at a time in random order. They then took part either in a mixture of free report and 2-AFC masked word identification trials, each using 20 studied targets and 20 nonstudied targets, or a set of test trials using the single probe procedure (with 40 studied and 40 nonstudied targets). For the 2-AFC and single probe tests, studied words that were not tested as targets were used as foils. Figure 7A shows that for Experiment 1, free report produced a strong advantage for studied words, consistent with our masked word identification studies (i.e., MacLeod & Masson, 1997b; Masson & MacLeod, 1992). But this advantage disappeared on the 2-AFC test. When, in Experiment 2, we switched from related probe pairs to using pairs of words on the 2-AFC
Episodic Enhancement of Processing Fluency
195
B
A I
Free report
0.6
2-AFC
y q 7
1
Free report
D
o.61
0.7
1_1I 2-AFC
,,,,,
,'\'\'\'\' 1 1 1 ,
, 1 1 1 ,
,,,,,
'\'\'\'\' ,,,,, ,',',',', ,',',',', ,,,,,
'.\\.\ '.'.'.* '.'.'.'\' ,,,,, ,,,,, \ \ \ \
,,,,, 2-AFC \ \ \ \
Free report
2-AFC
Free report
Fig. 7. Mean proportion correct on free report and two-alternative forced choice (2-AFC) versions of masked word identification as a function of prior exposure. (A) Probe pairs on the 2-AFC test were orthographically similar. (B) Probe pairs on the 2-AFC test were not similar in any way. (C) A I-s pause was inserted between the masked target and presentation of the probes on the 2-AFC test. (D) A 2-s pause was inserted between the masked target and presentation of the probes on the 2-AFC test, during which subjects attempted to report the target word; these report attempts constitute the free report data. Error bars indicate the 95% within-subject confidence intervals for the means. Data are adapted from Masson and MacLeod (1996).
test that were not orthographically related, we obtained the same result (see Fig. 7B). In Experiment 4,we inserted a 1-s delay between the target and the test probes on the 2-AFC trials. We expected that subjects would attempt to identify the target during that interval before the probes appeared. If subjects could more successfully identify studied as compared to nonstudied targets under these conditions, the advantage for studied targets ought to be revealed on the 2-AFC test. That effect did not appear, as can be seen by examining Fig. 7C. In Experiment 6 , a further attempt to produce a priming effect on the 2-AFC test was made by inserting a 2-s delay between the masked target and the probes, and by instructing subjects to attempt to identify the target
Michael E. J. Masson and Colin M. MacLeod
196
before the two alternatives appeared. To increase the power of this experiment, all targets were tested using the 2-AFC method and there were no free report trials. Nevertheless, this experiment also failed to produce a priming effect on the 2-AFC test, despite the fact that the free report attempts were more often correct for studied than for nonstudied targets. That advantage of prior study, however, was not robust enough to produce an effect on the 2-AFC test, probably because correct free report occurred on only about one third of the trials. The relatively low free report performance was a result of using display parameters that were sufficiently demanding that 2-AFC performance did not reach ceiling. Thus, subjects were often unsuccessful in their preliminary free report attempt and therefore frequently waited for the probes to appear before making a response to the masked target word. The remaining experiments involved the single-probe test, where a single word followed the masked target and the subject had to indicate whether that word was identical to the target. The study status of the probe was always the same as that of the target, to prevent the Ratcliff et al. (1989) type of bias from operating. The left side of Fig. 8 shows the data from Experiment 3, where the probe word immediately followed the mask. Accuracy on the probe task was measured using A', a measure of sensitivity that is not subject to the assumptions of normally shaped and equal-variance signal and noise distributions (e.g., Grier, 1971). Clearly, there was no evidence of priming. The same thing happened in Experiment 5 with a 2-s delay inserted between the masked target and the probe (see the middle section of Fig. 8). Again, we had expected that subjects would attempt to
a
0.90
Y
6 2
0.85
a
0 0
0.80
8
'
0.75
Immediate
2-s Delay
Attempt ID
Fig. 8. Mean accuracy on the single-probe version of the masked word identification task as a function of prior exposure. Dependent measure is A', a signal detection measure of sensitivity. Left section: the probe immediately followed the masked target. Middle section: there was a 2-s pause between the masked target and the probe. Right section: there was a 2-s pause between the masked target and the probe, during which subjects attempted to report the target word. Error bars indicate the 95% within-subject confidence intervals for the means. Data are adapted from Masson and MacLeod (1996).
Episodic Enhancement of Processing Fluency
197
identify the target during that delay interval and that greater success at identifying a studied target compared to a nonstudied target would carry over to the single-probe decision. In Experiment 7, however, when we combined a 2-s delay with instructions to attempt identification of the target prior to the appearance of the probe, we now observed priming in the probe task, as shown in the right section of Fig. 8. The single-probe procedure permitted use of less demanding display conditions than was the case with the 2-AFC test. Therefore, subjects were more frequently able in Experiment 7 to identify targets during the 2-s delay (about 45% of the time). We suspect that this greater level of preliminary report success led subjects’ decisions about the subsequent probe to be significantly affected by what was experienced during the preliminary report attempt. This series of experiments presents a quite simple picture. The better performance for studied words than for nonstudied words seen in the free report data for masked word identification both in this series and in others (MacLeod & Masson, in press; Masson & MacLeod, 1996) does indeed reflect a kind of bias. The advantage for studied words disappears in 2AFC and single-probe versions of the testing procedure unless, as in Experiment 7, sufficient time and adequate display conditions are provided to allow frequent identification of the masked target word before the probe word appears. The important remaining task is to make sense of this pattern of results. On the basis of their results, Ratcliff et al. (1989) suggested that performance differences favoring studied items were not being driven by heightened perceptual sensitivity to those items. We agree with this conclusion. Ratcliff et al. went on to propose a type of early perceptual bias in favor of studied words. It is not clear that this form of bias can account for our results, however, given that we prevented that bias from operating by always having test pairs and probes match the target with respect to study status. Instead, we argued in the Masson and MacLeod (1996) article for a fluency heuristic, as outlined in the introduction to this chapter. Our suggestion is that a masked target word that was studied earlier comes to mind more easily than one that is new, and that this is especially likely to be observed under impoverished conditions such as masking. A studied word comes to mind more easily because of the automatic retrieval of the previous encoding episode involving that word, cued by the masked target presentation. Because the studied word comes to mind more fluently, subjects are more likely to conclude that it was the presented target and to output it in free report identification. On a 2-AFC or a single-probe test, however, subjects are instead led to use discriminative information in comparing test items to the masked target word. As a result, the processing fluency generated
198
Michael E. J. Masson and Colin M. MacLeod
by studied words is ignored. Only when fluency is forced back into the probe task by adding a delay and urging subjects to attempt to identify the masked target prior to the probe’s appearance do subjects again revert to relying on fluency and again show priming. There is a tendency to think of a bias explanation as a disappointing one, the interpretation being that the situation under study does not represent a very interesting psychological phenomenon. We would argue quite vigorously against this view. The fluency heuristic may indeed be a very fundamental aspect of remembering, helping us to repeat actions more smoothly each time and even permitting us to make attributions about whether a situation or stimulus has been encountered previously. Surely these are the primary functions of memory. The experiments in this series tell us that priming in masked word identification is a memory phenomenon, not a perceptual one. Consequently, we must disagree with the conclusion reached by Reinitz and Alexander (1996) that priming occurs via the increased rate of gain of visual information from studied items. The evidence presented in this section in large part underlies why we prefer not to call the masked word identification task “perceptual identification.” The lesson is that we should avoid yoking our theories to our task designations because the tasks ordinarily outlive the initial theories proposed to account for performance in them.
VIII. Experiment Series 7: Episodic Effects on Perceptual Judgments We have argued that a fluency heuristic is responsible for the repetition priming observed in the masked word identification task. Moreover, this heuristic is assumed to operate similarly following data-driven and conceptually driven encoding tasks. The consequence of either type of prior encoding episode is that a target word comes to mind more fluently when that word is later presented under perceptually challenging conditions. In this final series of experiments, we report evidence to support the claim that this fluency can be experienced by subjects as a perceptual phenomenon regardless of whether its source is an earlier perceptual encoding, as in the read task, or a conceptual encoding with no direct perceptual processing, as in the generate task. The experiments in this series build on results reported by Witherspoon and Allan (1985) and others (Jacoby, Allan, Collins, & Larwill, 1988; Whittlesea et al., 1990). These researchers have shown that subjects can be led to attribute the processing fluency arising from prior exposure of target items to the perceptual conditions under which target items are
Episodic Enhancement of Processing Fluency
199
tested rather than to prior exposure. In the Witherspoon and Allan study, for example, subjects’ estimates of the exposure duration of briefly presented, masked target words were increased if the targets had appeared in an earlier study list. Similarly,Jacoby et al. found that when sentences were presented auditorily in a background of white noise varying in loudness, ratings of the loudness of the noise were lower if sentences had been heard in an earlier study phase. In general, it appears that attributions of processing fluency to perceptual characteristics of stimuli can be induced if subjects are asked to report on those characteristics. We attempted to capitalize on this phenomenon in a comparison of the influence of read and generate encoding episodes. The question we addressed was whether both of these types of encoding are capable of affecting subjects’ perceptual judgments. If there is a qualitative difference in what subjects experience as a result of a data-driven encoding task as opposed to a conceptually driven task, we might expect that perceptual judgments would be influenced only (or at least primarily) by the former type of task. On the other hand, if both classes of encoding task create the potential for processing fluency that is qualitatively similar, both kinds of task should affect perceptual judgments about subsequent target displays in the same way. These ideas were tested in a series of experiments by Caldwell and Masson (1997) that adopted the general procedure developed by Witherspoon and Allan (1985). In the first experiment, subjects read or generated from definitions a series of target words. Next, subjects were shown a series of filler words, each displayed for 30,60, or 90 ms and followed by a pattern mask. Subjects were informed that three different durations were being used and the task was to rate the duration of each display on a three-point scale. After the filler trials, critical targets began to appear, some of which had been read or generated in the study phase, whereas others were new words. At this point, however, unknown to the subjects, only 30-ms and 60-ms exposure durations were used. It was expected that prior exposure to targets would lead subjects to experience greater processing fluency and to attribute that experience to longer exposure duration, as had been observed by Witherspoon and Allan. That effect would be expressed as longer duration ratings for previously studied targets. By withholding 90-ms exposure trials, we hoped to increase the likelihood that subjects would be induced by experienced fluency to assign the highest duration rating to trials on which old targets were shown for only 60 ms (or even 30 ms). Of particular interest was the possibility that both read and generate encoding tasks would produce a similar effect on the duration ratings. On a three-point scale, the mean duration ratings for generate, read, and new critical targets presented in the 60-ms condition showed little variation
200
Michael E. J. Masson and Colin M. MacLeod
(2.25-2.27) and did not reliably differ as a function of encoding condition. In the 30-ms condition, however, there were significant differences between the mean ratings for the three conditions. In particular, the generate and read targets produced mean ratings that were not reliably different (1.55 and 1.60, respectively), although each of these conditions was significantly higher than the average for new items (1.46). The results of the first experiment were encouraging if not fully convincing. The failure to find an effect of prior presentation among the 60-ms targets may have been a result of the fact that none of the trials in the critical part of the test phase used the longest exposure duration of 90 ms. Therefore, subjects may have been reluctant to use the highest rating. An examination of the rating values shows this was not the case. For critical targets shown in the 60-ms condition, nearly 40% were assigned the maximum duration rating of three but these assignments were evenly spread across old and new targets alike. Another possibility is that in the duration judgment task subjects may not have processed the target words in a manner that would consistently lead them to experience the fluency produced by prior encoding episodes. Given that the only requirement was to rate the exposure duration of the words, subjects often may not have engaged in attempts to identify the words. On our account, it is the attempt to identify a target that will induce processing fluency and lead to the subjective experience of longer exposure duration. To test this idea, a second experiment was conducted in which subjects again judged exposure duration, but they did so only after first attempting to identify the target word. Another change from the first experiment involved the exposure durations used in the test phase. For the filler words presented to acquaint subjects with the various exposure durations, the durations were 15,45, and 75 ms. When the critical target words began to appear, mixed in with some filler targets, the critical items all were shown at a single duration midway between two of the practice durations. Again, this change was not announced to the subjects. The set of filler items was shown at the third duration. For half of the subjects, the critical targets were shown for 30 ms and the fillers were shown for 75 ms. For the other half of the subjects, the critical targets were shown for 60 ms and the fillers were shown for 15 ms. Thus, we created a between-subjects replication of the critical exposure durations used in the first experiment. In this experiment, however, the critical target duration was situated between two previously experienced durations. By making the actual exposure duration (30 or 60 ms) “ambiguous” with respect to the durations with which the subjects initially had been trained, the expectation was that subjects would tend to assign new
Episodic Enhancement of Processing Fluency
201
targets to the lower duration category and old targets to the higher duration category that flanked the actual exposure duration. The mean proportion of correctly identified targets and the mean duration ratings are shown in Figures 9A and 9B, respectively. Subjects in the 60-ms group correctly identified nearly all of the critical target words, leading to little difference in accuracy as a function of prior exposure. In the 30-ms group, subjects identified about half of the critical targets; the generate and read conditions led to very similar identification performance, which was substantially better than identification accuracy in the new condition (our typical pattern). These different patterns of identification accuracy led to a highly reliable interaction between encoding task and target duration. Despite the difference in the pattern of identification performance in the two duration groups, the profile of duration ratings was remarkably consistent across duration. Duration ratings were, of course, much higher for the 60-ms group than for the 30-ms group. The effect of prior exposure was robust, with higher duration ratings for previously studied items, regardless of whether the encoding episode involved generation or reading. There was no interaction between encoding task and duration, indicating that the pattern of effects on duration ratings was the same for both groups. Planned comparisons indicated that the duration ratings did not differ between the generate and read encoding tasks, but the average rating across these two prior exposure conditions was reliably higher than for the new condition. This pattern held even when the 60-ms group was considered alone. Prior exposure to target words led to a small but highly reliable increase in rated exposure duration, regardless of whether subjects engaged in per-
Duration
Duration
Fig. 9. (A) Mean proportion of correctly identified targets and (B) mean target duration ratings as a function of encoding task. Error bars indicate the 95% within-subject confidence intervals for the means. Data are adapted from Caldwell and Masson (1997).
202
Michael E. 1. Masson and Colin M. MacLeod
ceptual processing of the targets during the study phase. The importance of inducing subjects to identify the targets is apparent from the fact that the effect of prior exposure was more robust here than in the first experiment of the series, when subjects were not instructed to identify the targets. In that first experiment, we failed to find an effect of prior exposure on duration ratings when the actual exposure duration was 60 ms. As expected, having subjects attempt to identify the targets made the effect of prior exposure on duration ratings more robust. We suggest that this occurred because the influence of prior exposure on duration ratings operates through fluency experienced when viewing and attempting to identify the target words. One might argue, however, that the effect of prior exposure on duration ratings was a result of either (1) conscious recollection of the words’ prior occurrence or (2) a strategy whereby subjects systematically gave higher duration ratings to words they successfully identified. Various aspects of the results of the second experiment in this series are inconsistent with these possibilities. First, if subjects were relying on conscious recollection to determine duration ratings, we should have observed a stronger effect of prior exposure in the generate than in the read condition. Many previous studies have shown that awareness of prior occurrence as measured by direct tests of memory is much more likely to follow the generate encoding task than the read encoding task (e.g., Jacoby, 1983; Masson & MacLeod, 1992). Instead, the generate and read conditions consistently produced similar duration ratings as well as similar identification success. Even in the 60-ms group, where nearly all targets were correctly identified, the mean duration ratings were nearly identical for targets in the generate and read conditions. On the question of strategically based ratings, tied to successful identification, we note that a reliable effect of prior exposure on duration ratings was found in the 60-ms group, even though identification accuracy in that group was uniformly high so that accuracy was unaffected by prior exposure. This dissociation between identification accuracy and duration ratings suggests that differential ratings for studied and nonstudied targets was not simply a result of successful identification. Rather, the pattern of ratings indicates that subjects were sensitive to differences in identification fluency, not merely to the success or failure of their identification attempts. Finally, it is particularly important that the generate and read encoding conditions produced similar results in both identification accuracy and duration ratings. We take these results as evidence that the processing fluency experienced by subjects during the masked presentation of target words is qualitatively similar for items that were generated or read at the time of encoding. Thus, prior perceptual experience of a target word was not required to produce enhancement of either identification or exposure dura-
Episodic Enhancement of Processing Fluency
203
tion ratings. Instead, an episode involving conceptually driven processing of a target word (i.e., generation from a semantic cue) was sufficient to produce both enhanced identification and inflated duration judgments. These results lend further support to our argument that skilled word identification recruits relevant prior experiences that are either perceptually or conceptually relevant to the current identification attempt. To the extent that especially relevant prior episodes are recruited, a target word will be experienced as more readily perceived, and that experience may be attributed to whatever dimension the subject is asked to judge.
IX. Conclusion We have attempted to build a case for the view that memory for prior encoding episodes increases the fluency with which the current encoding of a stimulus goes forward. In tasks involving the identification of a target word, either through a brief, masked presentation, or through completion of a word fragment or stem, prior experiences with the target word that share aspects of the current processing task will be recruited and used in the construction of an interpretation of the target stimulus. The availability of such prior episodes will enable a more fluent construction of an interpretive encoding. On this account, prior episodes do not contribute to more efficient perceptual processing of the target’s physical features (cf. Reinitz & Alexander, 1996). Rather, the recruitment of a relevant prior encoding episode contributes to the fluent interpretation of those features. For this reason, prior episodes that emphasize a word’s meaning also have the potential to improve identification of that word when it is presented under degraded conditions on a subsequent test. To the extent that the form in which a word is presented supports the recruitment of conceptual aspects of prior processing, even prior episodes that involved no direct perceptual experience (such as the generation task) can enhance identification performance. The general effect of this initial interpretive encoding of a target word is subjectively experienced as the word’s identity coming to mind fluently. That experienced fluency is in turn subject to attribution on the part of the subject. For example, if the task is to identify the target word, experienced fluency is likely to increase the subject’s confidence that the correct candidate has been selected. If the task is to make a judgment about some perceptual aspect of the target, such as the duration of its presentation, fluent reprocessing may be attributed to that aspect of the target (e.g., Experiment Series 7). Finally, if the task is to determine whether the target word had occurred earlier in the context of the experiment, fluent identifi-
204
Michael E. J. Masson and Colin M. MacLeod
cation provides evidence of such an occurrence. Thus, artificially created processing fluency can create illusions of prior occurrence (e.g., Whittlesea et al., 1990). On this account, conscious awareness of prior occurrence results from an evaluation of the current encoding episode (see Whittlesea’s chapter, this volume), rather than contributing to its success. More generally, we see the influence of prior encoding episodes as the basis for the development of skilled word reading (e.g., Masson, 1986). Although models of word reading typically ascribe skilled performance to activation of stable, abstract word representations such as logogens (e.g., Morton, 1969), there is mounting evidence to suggest that specific episodes powerfully affect word identification even among highly accomplished readers. Some of the earliest research demonstrating the effects of repetition on word identification is remarkable for this very reason (e.g., Scarborough, Cortese, & Scarborough, 1977). In our view, the surprisingly large benefit of a single additional exposure to a word derives from the substantial similarity in contextually bound episodes that constitute the initial and repeated presentations of a word in the repetition priming paradigm. That strong similarity gives the first presentation of the target in the context of the experiment a privileged role relative to the many hundreds or thousands of previous experiences with that target word. Thus, an episode with strong contextual relevance produces a much more substantial influence on performance than would be expected from a simple increment in the number of generic encounters with a word (Kirsner & Speelman, 1996). O n this view, although skill can be generalized, it is at the same time subject to contextual specificity in the same way and for the same reason as other memory phenomena. A. CONSCIOUS RECOLLECTION On the account advocated here, conscious recollection of a target’s prior occurrence is an attribution that arises in part from the fluency experienced during the encoding of the target. Further grounds for deciding that a target had been studied earlier are likely to be present if that study experience involved conceptually driven processing. Elements of such encoding episodes are typically well remembered, particularly when the target word that was the core of the episode is available as a cue. An alternative view of the role played by conscious recollection is that this type of remembering can inflate or contaminate performance on indirect tests of memory. That is, conscious recollection is assumed to have the potential to contribute to successful generation of a target word on tasks such as masked word identification or stem completion. The influence of conceptually driven encoding tasks on subsequent word identification
Episodic Enhancement of Processing Fluency
205
tests that are assumed to be dominated by data-driven processes is therefore ascribed to conscious recollection, rather than to an automatic influence of memory for conceptually driven processing (Toth et al., 1994). On this view, our demonstrations of similar amounts of repetition priming following generate and read encoding tasks are the result of conscious recollection artifactually enhancing priming effects among targets encoded with the generate task. The most clear statement of how this contamination could come about has been developed by Jacoby and his colleagues (Jacoby et al., 1992) in the context of the process dissociation procedure. In that approach, it is assumed that conscious recollection and automatic influences of memory independently contribute to responding on word identification tests. This position contrasts with the account we prefer, in which conscious recollection or awareness of prior occurrence is a constructed attribution that follows successful word identification. In Experiment Series 3, we showed how the process dissociation procedure can yield underestimates of the automatic influence of memory for conceptually driven encoding episodes because of this attribution process. Moreover, by assessing the effects of generate and read encoding tasks on a speeded word reading test in Experiment Series 4,we were able to create a situation that should have minimized if not prevented any useful contribution of conscious recollection to test performance. Those experiments replicated our findings with the masked word identification test, in which the generate encoding task led to reliable repetition priming that was not significantly less than that observed for targets that had been encoded in the read task. We take these results as strong evidence that conscious recollection is not responsible for producing the repetition priming on word identification tasks that follows generate encoding episodes. In fact, it is just the opposite: Successful identification of such word targets on these indirect tests of memory leads to conscious recollection of prior occurrence.
B. CHALLENGING THE TASKPURITY ASSUMPTION In our research on the influence of processing episodes on word identification, we have emphasized the notion of interpretive encoding. We have done so in an effort to highlight the important role played by conceptual processes in constructing an interpretation of a stimulus. Our view is that this construction is a crucial part of encoding and identification. This approach challenges the assumption that certain word identification tasks are pure with respect to data-driven and conceptually driven processing operations. Earlier demonstrations of modality specificity and weak or nonexistent benefits of prior conceptual processing for tasks such as word
206
Michael E. J. Masson and Colin M. MacLeod
stem completion and masked word identification have encouraged the view that such tasks are perceptual in nature (e.g., Blaxton, 1989; Jacoby, 1983; Weldon, 1991). Without denying the validity of modality-specific effects on these tasks, we maintain that contributions from conceptually driven processes are also genuine. We have argued that the transfer-appropriate processing framework should be extended to include the important role played by conceptual processes in the construction of interpretive encodings of stimuli (Masson & MacLeod, 1992). Demonstrations of greater enhancement produced by data-driven encoding experiences indicate that certain word identification tasks (e.g., word fragment completion) provide less opportunity for conceptual processes to take hold than do other tasks, such as masked word identification or speeded word reading. In these latter tasks, the availability of the entire word is presumed to enable greater conceptually based recruitment of prior episodes. Accordingly, our research has shown that there are word identification tasks in which conceptually driven encoding episodes make a substantial contribution to later processing fluency. More generally, then, we expect that the influence of conceptually driven encoding episodes on subsequent word identification performance will vary, depending on the extent of conceptual processing afforded by the cues available in the identification task. ACKNOWLEDGMENT Preparation of this chapter and the experiments reported here were supported by research grants from the Natural Science and Engineering Research Council of Canada to Michael Masson and to Colin MacLeod. We are grateful to the many graduate students and research assistants who have contributed to this project, particularly Natasa Arbutina, Glen Bodner, Graham Brown, Judy Caldwell, Carrie Hicks, Shelley Hodder, Anne Kelly, Alyssa Megale, and Tonya Stokes.
REFERENCES Atkinson, R. C., & Shiffrin, R. M. (1968). Human memory: A proposed system and its control processes. In K. W. Spence & J. T. Spence (Eds.), The psychology of learning and motivation: Advances in research and theory (Vol. 2, pp. 89-195). New York: Academic Press. Balota, D. A,, & Chumbley, J. I. (1984). Are lexical decisions a good measure of lexical access? The role of word frequency in the neglected decision stage. Journal of Experimental Psychology: Human Perception and Performance, IO, 340-351. Bassili, J. N., Smith, M. C., & MacLeod. C. M. (1989). Auditory and visual word-stem completion: Separating data-driven and conceptually driven processes. Quarterly Journal of Experimental Psychology, 41 A, 439-453.
Episodic Enhancement of Processing Fluency
207
Becker, S., Moscovitch, M., & Behrmann, M. (in press). Long-term semantic priming: A computational account and empirical evidence. Journal of Experimental Psychology: Learning, Memory, and Cognition. Blaxton, T . A. (1989). Investigating dissociations among memory measures: Support for a transfer-appropriate processing framework. Journal of Experimental Psychology: Learning, Memory, and Cognition, 15, 657-668. Bodner, G. E., Caldwell, J. I., & Masson, M. E. J. (1997). Involuntary conscious memory affects parameter estimates in the process dissociation procedure. Manuscript submitted for publication. Burt, J . S. (1994). Identity primes produce facilitation in a colour naming task. Quarterly Journal of Experimental Psychology, 47A, 957-1000. Caldwell, J. I., & Masson, M. E. J. (1997). Conceptually driven encoding episodes creale perceptual rnisartributions. Manuscript in preparation. Carr, T. H., Brown, J. S., & Charalambous, A. (1989). Repetition and reading: Perceptual encoding mechanisms are very abstract but not very interactive. Journal of Experimental Psychology: Learning, Memory, and Cognition, 15, 763-778. Carr, T. H., & Dagenbach, D. (1990). Semantic priming and repetition priming from masked words: Evidence for a center-surround attentional mechanism in perceptual recognition. Journal of Experimental Psychology: Learning, Memory, and Cognition, 16, 341-350. Craik, F. I. M., & Lockhart, R. S. (1972). Levels of processing: A framework for memory research. Journal of Verbal Learning and Verbal Behavior, 11, 671-684. Curran, T., & Hintzman, D. L. (1995). Violations of the independence assumption in process dissociation. Journal of Experimental Psychology: Learning, Memory, and Cognition, 21, 531-547. Dagenbach, D., Carr, T. H., & Wilhelmsen, A. (1989). Task-induced strategies and nearthreshold priming: Conscious effects on unconscious perception. Journal of Memory and Language, 28, 412-443. de Groot, A. M. B. (1983). The range of automatic spreading activation in word priming. Journal of Verbal Learning and Verbal Behavior, 22, 417-436. Dunn, J. C., & Kirsner, K. (1988). Discovering functionally independent mental processes: The principle of reversed association. Psychological Review, 95, 91-101. Eriksen, C. W. (1995). The flankers task and response competition: A useful tool for intestigating a variety of cognitive problems. Visual Cognition, 2, 101-118. Forster, K. I., & Chambers, S. M. (1973). Lexical access and naming time. Journal of Verbal Learning and Verbal Behavior, 12, 627-635. Gardiner, J. M. (1988). Functional aspects of recollective experience. Memory & Cognition, 16, 309-313. Graf, P., & Mandler, G. (1984). Activation makes words more accessible, but not necessarily more retrievable. Journal of Verbal Learning and Verbal Behavior, 23, 553-568. Graf, P., & Schacter, D. L. (1985). Implicit and explicit memory for new associations in normal and amnesic subjects. Journal of Experimental Psychology: Learning, Memory, and Cognition. 11, 501-518. Grier, J. B. (1971). Nonparametric indexes for sensitivity and bias: Computing formulas. Psychological Bulletin, 75, 424-429. Jacoby, L. L. (1983). Remembering the data: Analyzing interactive processes in reading. Journal of Verbal Learning and Verbal Behavior, 22, 485-508. Jacoby, L. L. (1991). A process dissociation framework: Separating automatic from intentional uses of memory. Journal of Memory and Language, 30, 513-541. Jacoby, L. L.. Allan, L. G., Collins, J. C., & Larwill, L. K. (1988). Memory influences subjective experience: Noise judgments. Journal of Experimental Psychology: Learning, Memory, and Cognition, 14, 240-247.
208
Michael E. J. Masson and Colin M. MacLeod
Jacoby, L. L., & Brooks, L. R. (1984). Nonanalytic cognition: Memory, perception, and concept learning. In G. H. Bower (Ed.), The psychology of learning and motivation (Vol. 18, pp. 1-47). New York: Academic Press. Jacoby, L. L., & Dallas, M. (1981). On the relationship between autobiographical memory and perceptual learning. Journal of Experimental Psychology: General, 110, 306-340. Jacoby, L. L., Lindsay, D. S., & Toth, J. P. (1992). Unconscious influences revealed. American Psychologist, 47, 802-809. Jacoby, L. L., Toth, J. P., & Yonelinas, A. P. (1993). Separating conscious and unconscious influences of memory: Measuring recollection. Journal of Experimental Psychology: General, 122, 139- 154. Jacoby, L. L., & Whitehouse, K. (1989). An illusion of memory: False recognition influenced by unconscious perception. Journal of Experimental Psychology: General, 118, 126-135. Jacoby, L. L., Yonelinas, A. P., & Jennings, J. (1996). The relation between conscious and unconscious (automatic) influences: A declaration of independence. In J. Cohen & J. W. Schooler (Eds.), Scientific approaches to the question of consciousness (pp. 13-47). Mahwah, NJ: Erlbaum. Joordens, S., & Merikle, P. M. (1993). Independence or redundancy? Two models of conscious and unconscious influences. Journal of Experimental Psychology: General, 122, 462-467. Kintsch, W.(1970). Learning, memory, and conceptual processes. New York: Wiley. Kirsner, K., & Speelman, C. (1996). Skill acquisition and repetition priming: One principle, many processes? Journal of Experimental Psychology: Learning, Memory, and Cognition, 22, 563-575. Kolers, P. A. (1976). Reading a year later. Journal of Experimental Psychology: Human Learning and Memory, 5, 554-565. Kolers, P. A., & Roediger, H. L., 111. (1984). Procedures of mind. Journal of Verbal Leurning and Verbal Behavior, 23, 425-449. Light, L. L.,& Carter-Sobell, L. (1970). Effects of changed semantic context on recognition memory. Journal of Verbal Learning and Verbal Behavior, 9, 1-11. Logan, G. D. (1990). Repetition priming and automaticity: Common underlying mechanisms? Cognitive Psychology, 22, 1-35. Lukatela, G., & Turvey, M. T. (1994). Visual lexical access is initially phonological: 1. Evidence from associative priming by words, homophones, and pseudohomophones. Journal of Experimental Psychology: General, 123, 107-128. MacLeod, C. M. (1991). Half a century of research on the Stroop effect: An integrative review. Psychological Bulletin, 109, 163-203. MacLeod, C. M. (1996). How priming affects two speeded implicit tests of remembering: Naming colors versus reading words. Consciousness & Cognition, 5, 73-90. MacLeod, C. M. (in press). Directed forgetting. In J. M. Golding & C. M. MacLeod (Eds.), Intentional forgetting: Interdisciplinary approaches. Mahwah, NJ: Erlbaum. MacLeod, C. M., & Daniels, K. A. (1997). Implicit vs. explicit remembering: Directed forgetting meets the generation effect. Manuscript in preparation. MacLeod, C. M., Grant, S. C., & Ridout, S. M. (1996). Primingflankers and targets in a lexical decision flanker task. Manuscript submitted for publication. MacLeod, C. M., & Kampe, K. E. (1996). Word frequency effects on recall, recognition, and word fragment completion tests. Journal of Experimental Psychology: Learning, Memory, and Cognition, 22, 132-142. MacLeod, C. M., & Masson, M. E. J. (1997a). Fluent word reading as a measure of implicit remembering. Manuscript in preparation. MacLeod, C. M., & Masson, M. E. J. (1997b). Priming patterns are different in masked word identification and word fragment completion. Journal of Memory and Language 36, 461-483.
Episodic Enhancement of Processing Fluency
209
Masson, M. E. J. (1986). Identification of typographically transformed words: Instance-based on.Journal of Experimental Psychology: Learning, Memory, and Cognition, 12, 479-488. Masson, M. E. J., & MacLeod, C. M. (1992). Re-enacting the route to interpretation: Enhanced perceptual identification without prior perception. Journal of Experimental Psychology: General, 121, 145-176. Masson, M. E. J., & MacLeod, C. M. (1996). Contributions of processing fluency to repetition effects in masked word identification. Canadian Journal of Experimental Psychology. 50, 9-21. Masson, M. E. J., & Sala, L. S. (1978). Interactive processes in sentence comprehension and recognition. Cognitive Psychology, 10, 244-270. Meyer. D. E., & Schvaneveldt, R. W. (1971). Facilitation in recognizing pairs of words: Evidence of a dependence between retrieval operations. Journalof Experimental Psychology, 90, 227-234. Morris, C. D., Bransford, J. P., & Franks, J. J. (1977). Levels of processing versus transfer appropriate processing. Journal of Verbal Learning and Verbal Behavior, 16, 519-533. Morton, J. (1969). Interaction of information in word recognition. Psychological Review, 76, 165178. Neely, J. H. (1991). Semantic priming effects in visual word recognition: A selective review of current findings and theories. In D. Besner & G. W. Humphreys (Eds.), Basicprocesses in reading: Visual word recognition (pp. 264-336). Hillsdale, NJ: Erlbaum. Nelson, D. L., Keelean, P. D., & Negrao, M. (1989). Word-fragment cuing: The lexical search hypothesis. Journal of Experimental Psychology: Learning, Memory, and Cognition, 15, 388-397. Osgood, C . E. (1949). The similarity paradox in human learning: A resolution. Psychological Review, 56, 132-143. Ratcliff, R.. & McKoon, G. (1993, November). Bias in implicit memory tests. Paper presented at the annual meeting of the Psychonomic Society, Washington, DC. Ratcliff, R., McKoon, G., & Verwoerd, M. (1989). A bias interpretation of facilitation in perceptual identification. Journal of Experimental Psychology: Learning, Memory, and Cognition, 15, 378-387. Reingold, E. M., & Toth, J. P. (1996). Process dissociations versus task dissociations: A controversy in progress. In G. Underwood (Ed.), Implicit cognition (pp. 159-202). New York: Oxford University Press. Reinitz, M. T., & Alexander, R. (1996). Mechanisms of facilitation in primed perceptual identification. Memory & Cognition, 24, 129-135. Richardson-Klavehn, A,, & Bjork, R. A. (1988). Measures of memory. Annual Review of Psychology, 39, 475543. Richardson-Klavehn, A., & Gardiner, J. M. (1995). Retrieval volition and memorial awareness in stem completion: An empirical analysis. Psychological Research, 57, 166-178. Richardson-Klavehn, A., Gardiner, J. M., & Java, R. I. (1996). Memory: Task dissociations, process dissociations and dissociations of consciousness. In G. Underwood (Ed.), lmplicit cognition (pp. 85-158). New York Oxford University Press. Richardson-Klavehn, A,, Gardiner, J. M., & Ramponi, C. (1996). Does depth of processing inpuence involuntary retrieval in stem completion? Manuscript submitted for publication. Roediger, H. L., 111. (1990). Implicit memory: Retention without remembering. American Psychologist, 45, 1043-1056. Roediger, H. L., 111, & Blaxton, T. A. (1987). Effects of varying modality, surface features, and retention interval on priming in word fragment completion. Memory & Cognition, 15, 379-388.
210
Michael E. J. Masson and Colin M. MacLeod
Roediger, H. L., 111, & McDermott, K. B. (1993). Implicit memory in normal human subjects. In F. Boller & J. Grafman (Eds.). Handbook of neuropsychology (Vol. 8, pp. 63-131). Amsterdam: Elsevier. Roediger, H. L., 111, Weldon, M. S., & Challis, B. H. (1989). Explaining dissociations between implicit and explicit measures of retention: A process account. In H. L. Roediger & F. I. M. Craik (Eds.), Varieties of memory and consciousness: Essays in honour of Endel Tulving (pp. 3-14). Hillsdale, NJ: Erlbaum. Rumelhart, D. E. (1977). Toward an interactive model of reading. In S . Dornic (Ed.), Attention and performance Vl. Hillsdale, NJ: Erlbaum. Scarborough, D. L., Cortese, C., & Scarborough, H. S. (1977). Frequency and repetition effects in lexical memory. Journal of Experimental Psychology: Human Perception and Performance, 3, 1-17. Schacter, D. L., Bowers, J., & Booker, J. (1989). Intention, awareness, and implicit memory: The retrieval intentionality criterion. In S. Lewandowsky, J. C. Dunn, & K. Kirsner (Eds.), Implicit memory: Theoretical perspectives (pp. 47-65). Hillsdale, NJ: Erlbaum. Seidenberg, M. S., Tanenhaus, M. K., Leiman, J. M., & Bienkowski, M. (1982). Automatic access of the meanings of ambiguous words in context: Some limitations of knowledgebased processing. Cognitive Psychology, 14, 489-537. Semon, R. (1921). The mneme. London: Allen & Unwin. (Originally published in 1904 as Die Mneme. Leipzig: Engelmann) Slamecka, N. J., & Graf. P. (1978). The generation effect: Delineation of a phenomenon. Journal of Experimental Psychology: Human Learning and Memory, 4, 592-604. Strain, E., Patterson, K., & Seidenberg, M. (1995). Semantic effects in single-word naming. Journal of Experimental Psychology: Learning, Memory, and Cognition, 21, 1140-1 154. Stroop, J. R. (1935). Studies of interference in serial verbal reactions. Journal of Experimental Psychology, IS, 643-662. Swinney, D. A. (1979). Lexical access during sentence comprehension: (Re)consideration of context effects. Journal of Verbal Learning and Verbal Behavior, IS, 645-660. Toth, J. P., & Hunt, R. R. (1990). Effect of generation on a word-identification task. Journal of Experimental Psychology: Learning, Memory, and Cognition, 16, 993-1003. Toth. J. P., Reingold, E. M., & Jacoby, L. L. (1994). Toward a redefinition of implicit memory: Process dissociation following elaborative processing and self-generation. Journal of Experimental Psychology: Learning, Memory, and Cognition, 20, 290-303. Toth, J. P., Reingold, E. M., & Jacoby, L. L. (1995). A response to Graf and Komatsu’s critique of the process dissociation procedure: When is caution necessary? European Journal of Cognitive Psychology, 7, 113-130. Tulving, E. (1985). Memory and consciousness. Canadian Psychologist, 26, 1-12. Tulving, E., & Thomson, D. M. (1973). Encoding specificity and retrieval processes in episodic memory. Psychological Review, SO, 352-373. Warren, R. E. (1972). Stimulus encoding and memory. Journal of Experimental Psychology, 94, 90-100. Waugh. N. C.. & Norman, D. A. (1965). Primary memory. Psychological Review, 72, 89-104. Weldon, M. S. (1991). Mechanisms underlying priming on perceptual tests. Journal of Experimental Psychology: Learning, Memory, and Cognition, 17, 526-541. Whittlesea, B. W. A,, Jacoby, L. L, & Girard, K. (1990). Illusions of immediate memory: Evidence of an attributional basis for feelings of familiarity and perceptual quality. Journal of Memory and Language, 29, 716-732. Witherspoon, D., & Allan, L. G. (1985). The effects of a prior presentation on temporal judgments in a perceptual identification task. Memory & Cognition, 13, 101-111. Witherspoon, D., & Moscovitch, M. (1989). Stochastic independence between two implicit memory tests. Journal of Experimental Psychology: Learning, Memory, and Cognition, 1.5, 22-30.
PRODUCTION, EVALUATION, AND PRESERVATION OF EXPERIENCES: CONSTRUCTIVE PROCESSING IN REMEMBERING AND PERFORMANCE TASKS Bruce W. A. Whittlesea
I. Introduction Memory is the faculty of learning, the capacity of mind to be changed by an experience, such that its potential to respond to stimuli is altered. Memory is the sum of all one’s experiences: it is the locus of all skills, all knowledge. Memory controls all of behavior except reflex activity. In understanding human behavior, it is therefore extremely important to have a proper understanding of the structure and function of memory. Unfortunately, memory is involved in a bewildering variety of activities, differing on many dimensions: perceptual, conceptual, and motor activities; assimilation of information and control of performance; social, intellectual, and emotional interactions; simple responses and complex chains of behavior; abstract reasoning and automatic reactions. Each of these activities and dimensions has its own quirks and special conditions. It is thus a considerable challenge to educe simple, fundamental principles of memory as a whole. In attempting to deal with the complexity of memory’s specific interactions with various tasks and stimuli, investigators of mind have long treated memory as serving two basic and qualitatively different functions, namely remembering and the control of performance. Remembering is the use of memory as a record of events to reconstruct past experience, in the act THE PSYCHOLOGY OF LEARNING AND MOTIVATION. VOL. 37
211
Copyright 0 1997 by Academic Press. All rights of reproduction in any form reserved. W79-7421/97SZS.00
212
Bruce W. A. Whittlesea
of recall or recognition: it is the attempt to think about some particular experience of an object, in some specific context. Isolated or unusual events are remembered better than regular, normal events: it is hard to remember any specific experience of riding a bus, distinct from all other such experiences, unless something unique happened during a particular journey. This ability to think selectively about particular past events appears to imply that memory retains knowledge of specific events, which can be retrieved to consciousness. In contrast, other tasks, such as classification and identification, require the person to perform conceptual, categorical, or perceptual judgments, without reflecting on the specific past events in which they acquired the knowledge that enables them to perform the judgments. These tasks call on memory to provide information about the meaningful properties of objects, such as their identity or utility to satisfy some current purpose. Performance in these tasks is optimal for objects that are regular, average, or typical of their class. For example, people classify a typical fruit, such as an apple, faster than an atypical fruit, such as a fig (Rosch & Mervis, 1975); orthographically regular words are identified more efficiently than orthographically irregular words (Wheeler, 1970). Such observations seem to imply that memory also possesses abstract, general knowledge about concepts and categories of objects. Thus, the fundamental fact about memory seems to be that it possesses two qualitatively different types of knowledge, specific and general, selectively used to perform two types of task, one with respect to particular events and the other with respect to the generic identity of objects. This dichotomy of knowledge types entails a host of other assumptions about the mechanisms by which each is acquired, stored, and applied. In fact, remembering and control of performance appear to depend on completely separate memory systems, operating under qualitatively different sets of principles. This reasoning is the cornerstone of theories of memory such as Anderson’s (e.g., 1980) associative network account and Tulving’s (e.g., 1985, 1995) multiple-systems account. I contend that dichotomizing memory in this way seriously miscasts its essential nature, and results in a series of mistaken ideas about each of memory’s subsidiary functions. Instead, in this chapter I propose that the prime functions of memory are the construction of psychological experiences and the preservation of those experiences. The construction of an experience of a stimulus is performed through the same set of principles both in remembering specific events from the past and in identifying or classifyingpresent stimuli. The detail of the constructive process on different occasions is extremely various, responding to differences in task, stimulus structure, context, and the availability of similar prior experiences, but the
Constructive Processing in Remembering and Performance
213
fundamental principles are the same in all cases. Further, I contend that memory preserves only one type of information, namely whatever a person does in constructing an experience of a stimulus. Fundamentally, I argue, memory is unitary in structure and operation. I describe a formulation of memory, which I call SCAPE (for Selective Construction And Preservation of Experiences). This theory is a marriage of ideas from many sources, including instance theory (e.g., Brooks, 1978; Jacoby & Brooks, 1984; Medin & Schaffer, 1978), the episodic-processing account of concept acquisition (e.g., Whittlesea & Dorken, 1993), the attribution theory of remembering (e.g., Jacoby & Dallas, 1981; Jacoby, Kelley & Dywan, 1989; Whittlesea, 1993), skill transfer (e.g., Kolers & Smythe, 1984) and transferappropriate processing (e.g., Masson & MacLeod, 1992; Morris, Bransford, & Franks, 1977; Roediger & Challis, 1992). 11. Separate-Systems Assumptions: A Brief Summary
Separate-systems accounts assume that memory contains two qualitatively different kinds of information, respectively, of specific events and general knowledge. They thus have to explain how these two types of knowledge are acquired separately, preserved separately, and applied selectively in different tasks or situations. Event knowledge is thought to be acquired in the act of processing stimuli in specific contexts for particular purposes. Memory registers those aspects of the stimulus and context that happen to be processed during that encounter, given the purpose of the encounter, the attentional resources available, and the availability of prior experiences to guide current processing. Event knowledge is thus encoded directly from experience: it is a record of what the person actually experienced in the encounter (the encoding variability principle; e.g., Craik & Lockhart, 1972). Remembering consists of activating a specific representation, independent of all other representations that include similar elements: for example, remembering that TABLE was in a training list requires the person to access that event distinct from all other experiences of that word. Accurate retrieval thus depends on the distinctiveness of the original experience, and on the availability of cues reinstating distinctive aspects of that experience (the encoding specificity principle; e.g., Tulving & Thompson, 1973). The acquisition, contents, and retrieval of event knowledge are thus sensitive to the idiosyncratic properties of particular experiences of stimuli. In contrast, performance in classification and identification tasks is sensitive to general, summary properties of stimuli, such as frequency, typicality, and regularity. These attributes are abstract properties of events taken as classes, not properties that could be directly apprehended in any actual
214
Bruce W. A. Whittlesea
encounter with an object: They emerge across a succession of events, but are not present in any one of them. Sensitivity to these properties must therefore mean that memory has some mechanism that computes general properties across particular experiences, abstracting the common meaning or structure of successive experiences, and leaving behind idiosyncratic properties, such as context and purpose. This abstraction is unconscious and automatic: it is driven by the structural similarities between members of a class rather than by intention. This general knowledge is preserved separately from knowledge about the events from which it was abstracted, in the form of logogens, prototypes, rules, or linkages in a semantic network. Initially labile, these abstract summaries become strong and stable as they accrete the common properties of hundreds or thousands of instances of a category. Once acquired, they take control of the perceptual and cognitive processing of new stimuli, providing the person with knowledge of the meaning, class, or identity of the stimuli. The efficiency and success of this processing depends on the extent to which the properties of a new stimulus match those represented in the abstract summary, giving rise to regularity and typicality effects. Accounts of memory based on such automatic abstraction of general knowledge have been proposed by many investigators, studying quite different aspects of performance, including dissociations between remembering and identification (Tulving, 1995; Tulving, Schacter & Stark, 1982), implicit learning (Reber, 1989, 1993), concept acquisition (Rosch, 1978), semantic priming (Meyer & Schvaneveldt, 1971), word identification (Paap & Noel, 1991), repetition blindness (Park & Kanwisher, 1994), and many other phenomena. It is still the dominant theory of memory representation, as demonstrated by the fact that it is taught, in all of its ramifications, in most of the current textbooks, with only minor acknowledgment of an alternative.
111. Selective Construction and Preservation of Experiences: Outline of the Account Despite the wide success of separate-systems accounts in explaining cognitive phenomena, I argue that those accounts are based on three fundamental errors about the structure and function of memory. The first error is that there is no mechanism in memory to perform chronic, unconscious abstraction.’ In consequence, there is no semantic memory, no type representa-
’
People can obviously perform abstractions on demand: if asked what writing paper, swans, and chalk all have in common, people can produce the answer “white.” However, that abstraction is performed in response to a particular demand, not computed automatically and unconsciously through the operation of a chronic abstraction mechanism. It is a specific experience of three stimuli, which will be recorded by memory, like any other experience of the same stimuli for any other purpose. Also, like any other event, it will continue to exist in memory as a potential to perform a similar activity on similar stimuli in the future.
Constructive Processing in Remembering and Performance
215
tions, no lexicon, logogens, prototypes, or other forms of automatically computed summary information. Instead, memory simply preserves particular experiences. Each encounter with a stimulus will produce some specific experience, dictated by the physical structure of the stimulus, the processing performed on it to satisfy the current purpose, the physical and meaningful context in which it is encountered, and the availability of prior similar experiences to guide current processing. Memory preserves that experience, and that is all that it records. In consequence, there is no distinction to be made between a semantic and an episodic memory system. The second error is the idea that memory is a store of knowledge, that is, a library where facts and concepts are stored, which can be accessed as needed to recall prior events or identify current objects. Separate-systems accounts not only make a distinction between general and particular knowledge, but also between declarative and procedural knowledge (e.g., Squire, 1992; Tulving, 1995). “Declarative knowledge” is meant to be the body of conceptual and factual information that one has to draw upon to identify objects or recall events, whereas “procedural knowledge” consists of the skills necessary to perform operations on stimuli, such as riding a bicycle. However, I argue, in any interaction with a stimulus, representations of prior experiences guide current processing of the stimulus, producing a mental and/or motoric event. Naming a bicycle or classifying it as a means of transportation is as much a skilled operation as riding it (cf. Kolers, 1973; Kolers & Smythe, 1984). The fact that one operation consists of producing a verbal label for a stimulus and the other consists of producing a motor response to the stimulus is not a fundamental difference. In each case, the person encounters the stimulus for some purpose and in some context: that stimulus complex cues representations of prior similar experiences with similar stimuli, which in turn guide processing of the current stimulus. I argue that memory does not directly contain any knowledge about what things are. Instead, memory preserves records of operating on stimuli, constructing cognitive events. In turn, those records drive performance in later encounters with other stimuli. In that sense, all knowledge is procedural. The library metaphor had misled us. In fact, memory is dynamic and interactive with the world. It controls performance in interacting with stimuli, rather than existing as a set of passive records to be consulted. The third error is that remembering consists of the retrieval of a representation of some prior event. In separate-systems accounts, it is this idea that separates memory’s function of remembering from its function of controlling perceptual and conceptual performance: remembering consists of finding and elevating a memory representation to consciousness, whereas in performance of the latter tasks, a summary, abstract representation that is not itself retrieved to consciousness directs the processing of some
216
Bruce W. A. Whittlesea
stimulus. I argue that nothing is ever retrieved from memory, in the sense of finding a representation and bringing it intact to consciousness. Instead, memory produces behavior, both overt and mental, and it does so not only in the control of performance tasks like identification, but also in remembering. The mental event that occurs while remembering may be very similar to the actual past event one is trying to remember, if it is selectively guided by the representation of that experience. However, it is not a copy of the representation of that event, stimulated into consciousness by a cue. Instead, it is a construction imposed on the stimulus, a construction guided by the representation of the prior event, but also influenced by expectations and attitudes aroused by the detail of the current stimulus encounter. In general, I argue that all mental events, whether consisting of the identification of a common object, the remembrance of an event, or the experience of pleasure in smelling a rose, are constructions created by the interplay between stimuli, tasks, contexts, and prior experiences: No mental event reflects direct knowledge of the stimulus. I contend that there is one memory system, that preserves all experiences and controls all behavior. Every processing event involves a stimulus compound, consisting of a stimulus (either an external object or an internal mental event), a purpose (either given by some external agency or arising from prior processing), and a context (both the physical, external world and the set of expectations that mind carries to the event). This stimulus compound is the thing on which memory operates. The concatenation of properties of the stimulus compound selectively cues memory representations of events that involved similar tasks, stimulus structures, and contexts. The cued traces control the processing of the stimulus compound, producing a new experience. And memory preserves the new processing experience. At the most fundamental level of analysis, this is all that happens in any encounter with the world, whether the purpose of that encounter is to identify or use a stimulus or to remember some prior event. The processing of a stimulus compound is an essentially constructive event. The stimulus itself possesses a set of physical properties, but not meaning or organization. For example, ORCHID, as a stimulus, is a physical entity. Its physical properties support the computation of many nonphysical properties, such as location, extent, unitariness, sound, pleasantness, familiarity, identity, and meaning. However, those are not properties of the stimulus object itself instead, they are constructed from cued memory representations of prior experiences and attributed to the stimulus, in the act of processing the physical stimulus for some purpose. The construction of an experience of a stimulus occurs through two semidistinct activities, namely production and evaluation. In terms of the processing performed, the only difference between these activities is
Constructive Processing in Remembering and Performance
217
whether the target stimulus is an external or internal one: the mechanism of processing is the same in the two cases. However, they are very different in effect. The former permits the person to deal with properties of the external world, and the latter permits the person to experience subjective states about his or her performance. The production activity consists of attributing properties to a stimulus, properties that are not physically present in the stimulus. Production of these properties begins when the stimulus compound (the concatenation of stimulus structure, purpose, and context) selectively cues representations of similar experiences. These cued representations preserve a record of performing specific perceptual and cognitive operations on specific stimulus structures. They control subsequent processing of the current stimulus, directing the construction of an organized percept of the stimulus, the imposition of conceptual properties such as identity on the percept, and the production of a response. The production activity ends with the occurrence of a mental or behavioral event. These end events can be exceedingly various, such as speaking a name in greeting a friend, the coming-to-mind of the meaning of a word or phrase in reading a novel, the act of patting in encountering a dog, or the perception of relationship in admiring a picture. They can be as simple as imposing a global perceptual organization on a circle of dots or as elaborate as picturing oneself using a stimulus object in an imaginary environment. The mental or behavioral event produced in a stimulus encounter can be an end in itself, as for example in classifying a stimulus as a WUG in a psychological experiment. However, the event will often become part of the context of another stimulus complex, and help to determine the processing of another stimulus. For example, in reading a novel, the coming-to-mind of the meaning of one word in a sentence influences the interpretation of the next word, and so on. The production of mental events is always controlled by representations of specific prior experiences, selectively cued by the details of the stimulus complex. This is often not obvious in performance in tasks such as naming and classification, which appear to be fairly impervious to changes in the context or perceptual manifestation of test stimuli. Instead, as discussed earlier, performance in such tasks is correlated with the structural similarity of a stimulus to the mass of past experiences of the same class. However, I argue in Part I of this chapter that this apparent stability is illusory: it occurs not because production of identity is insensitive to contextual detail or driven by an alternate knowledge base, but because similarity to specific experiences and to abstract summaries of those experiences is usually confounded in the experiments used to study those tasks. When specific experiences diverging from the average are supplied in an experiment, performance in identification and classification is observed to be controlled
218
Bruce W. A. Whittlesea
through an interaction of the specific cues offered at test with the specific characteristics of prior experiences. The act of remembering is also a constructive activity. In this case, the person is not asked to produce the identity or class of a stimulus, but instead some property that was associated with the stimulus on a prior occasion. In recognition, a person is given an item as a stimulus and asked to produce the identity of a context in which it has been encountered previously: in recall, the person is given a context as a stimulus and asked to produce the identity of an item that occurred earlier in that context. These properties are constructed in the same way that the meaning or class of an item is produced in identification and classification. The specific characteristics of the present stimulus compound selectively cue representations of prior experiences, which direct the construction of a mental event. The only difference is that, owing to the demand of the task, the person constructs a mental experience of some context, rather than a mental experience of the identity of the stimulus. The constructive nature of remembering is discussed in Part I1 of this chapter. The production function permits people to interact efficiently with the world, generating identities and remembrances. When such productions are performed fluently, one can usually trust them to be accurate. If the word dog comes to mind on seeing a small, hairy quadruped, it is probably the correct label; if I am trying to remember where my glasses are, and an image of them lying on the hall table comes to mind, I will probably find them there. The accuracy of these productions stems from the selectivity of the stimulus compound in cuing similar experiences, which will usually involve relevant information. Such fluent productions do not seem to be accompanied by any strong feeling-state: in fact, they are often barely noticed. In reading the word grass in a novel, one accepts the coming-to-mind of the meaning of “grass” without question, and without experiencing any particular emotion. One does not even experience a feeling of confidence that one is correct: one simply understands the word, and passes on. In remembering to buy milk on the way home, I may experience annoyance at having to go out of my way, but the remembrance itself is not accompanied by any strong feeling, such as a feeling that the idea of buying milk is awfully familiar. So long as the production occurs fluently, one “just knows” the information produced, and proceeds to the next activity. However, there are many occasions when the production operations fail in a surprising way. For example, one can think of a familiar meaning, but fail to produce the word that has that meaning, or encounter a familiar face but fail to produce the context from which one knows that person. On these occasions, people experience a tip-of-the-tongue feeling, or a
Constructive Processing in Remembering and Performance
219
feeling of familiarity, that can be very powerful (and annoying). They also interrupt ongoing activities to concentrate on the offending stimulus: stop speaking to concentrate on that word, or stop looking around the bus to stare at that individual. Such failures of production can occur in a variety of ways. For example, the production function may succeed in generating a complete mental event, but may do so with unexpected fluency or unexpected awkwardness. Alternatively, I may discover that the mental event was in error: at closer range, the quadruped turns out to be a cat, or the glasses are not on the table. An external agency, such as an experimenter, may also question the result of the production operation, asking, “Did you produce the right answer? Are you sure?” Under any of the circumstances, memory performs a second function, namely evaluation? The evaluation function is performed through exactly the same mechanism as the production function: a stimulus compound is processed within a context and for a purpose, guided by cued representations. However, the object of attention in this case is the quality of the mental event that was just produced in encountering a stimulus (the fluency, completeness, or elaborateness of processing), rather than the external stimulus itself. The purpose is also different: instead of producing a response to an external stimulus, the object now is to evaluate the goodness or source of the mental event that just occurred. For example, in a recognition experiment, a subject shown a test word may experience a mental image of seeing that word in the training context. On that basis, they can infer that the item is old. However, they could produce such a mental event whether or not they had actually experienced the test word in that context, just as one can imagine the word horse written in red gothic letters, even if one has never seen it that way. To decide between these alternatives, the subject may evaluate the fluency with which that image comes to mind, or the clarity, elaborateness, or completeness of that image. He or she could expect that an actual prior experience of the word in that context would sponsor more vivid, efficient, and detailed production of the mental image than would ad hoc construction. Similarly, in classifying objects, a person can use the ease with which a category label comes to mind as an indicator of whether that label is correct. Thus, people can use the efficiency of the production of a mental event to judge the source or accuracy of that event. This evaluation function
* I am unsure whether to suggest that evaluation inevitably accompanies every production, or only productions that are surprising or questioned by an outside agency. I do not yet fully understand the mechanism through which a person is spontaneously surprised by processing they have just performed. It may turn out that there is a chronic evaluation of every act of production, which becomes obvious only on those occasions on which processing produces unexpected results; alternatively, the mechanism may act only on the latter occasions.
220
Bruce W. A. Whittlesea
enables memory to make attributions either about the nature of the stimulus or about general or specific aspects of their previous experience of the stimulus. The evaluation process can be performed deliberately and consciously, but also occurs automatically and unconsciously when processing is surprisingly fluent or nonfluent. This evaluation is a heuristic, inferential process. Just as many different meanings or organizations can be imposed on a stimulus in the production function, depending on the task and context, so many different meanings can be imposed on unexpectedly fluent or nonfluent processing. In the context of a recognition task, fluent processing may be attributed to the influence of past episodes, and be experienced as a feeling of familiarity: in the context of a classification judgment, the same fluent processing may instead be attributed to structural goodness of the stimulus, and be experienced as a feeling of rightness. Unexpectedly nonfluent processing may be interpreted either as novelty of a stimulus or as an error in producing a name or behavior. The effect of evaluative processing is thus to produce feelings about stimuli, feelings of rightness and wrongness, pleasantness and unpleasantness, novelty and familiarity, and so on. These feelings provide an important second basis for performing both remembering and nonremembering tasks, beyond the simple production of mental events. They also appear to serve an alerting function, warning that there is something wrong or special about the processing just conducted. They are not just information about the nature of the stimulus (that it is unpleasant or familiar): they are motivating, driving the system to solve a problem. For example, when a face at a bus stop feels familiar, one can guess that one has seen that person before (the information function): but one will also be driven to search for the reason that face feels familiar, attempting to produce the context from which that face is known, or to identify the friend whose face that face resembles. If one succeeds in producing the source of the feeling (e.g., produces a mental image of seeing that face on the clerk at the corner store), the feeling of familiarity disappears, replaced by “just knowing” the person. If one fails to produce the source, then the feeling of familiarity persists, warning about an unsolved problem. The evaluative function thus serves as an important backstop to production, alerting the system to stimuli that may be of special significance. To summarize, performance in remembering tasks is sensitive to the episodic detail of particular experiences, whereas performance in nonremembering tasks is often observed to be sensitive to general properties of the mass of prior experience. According to separate-systems accounts, this evidence means that memory serves two functions, retrieving specific events from the past and producing knowledge about present objects. These func-
Constructive Processing in Remembering and Performance
221
tions are served by separate memory systems that preserve qualitatively different kinds of information, and acquire and apply their information through qualitatively different principles. In contrast, the SCAPE account argues that the differences in performance in remembering and nonremembering tasks are due to the differences in the cue compound and the interpretive contexts made available in the tasks. It also argues that memory performs two functions, construction and preservation of experiences, but that those functions occur in the same way in remembering and performance tasks. Fundamentally, there is only one kind of representation in memory, and one set of processing principles: a current stimulus compound selectively cues similar prior experiences; those experiences guide current processing; memory preserves a representation of that experience; and the new representation serves as a resource for perception and performance on further occasions. The principles of the SCAPE account are simple, but have wide and diverse implications. I illustrate the theory through six themes that have controlled my research program. In Part I, I attempt to demonstrate the constructive nature of performance in tasks such as classification and identification. The experiments show that automatic, unconscious abstraction is an unnecessary and insufficient assumption, and that instead, performance in such tasks depends on the preservation of particular experiences of particular stimuli. They demonstrate the utility of the “episodic-processing’’ account of acquisition, representation, and production, on which SCAPE theory is partly based. In Part 11, I illustrate the constructive nature of remembering. The experiments also demonstrate the interplay between the production and evaluation functions of memory, and the “attribution” account of decision making, which makes up the remainder of SCAPE theory.
PART I CONSTRUCTIVE PRODUCTION IV. Theme 1:Concepts Are Not Automatically Abstracted across Instances Wittgenstein (1953) pointed out that many natural categories have an internal family resemblance structure, consisting of the fact that each member
222
Bruce W. A. Whittlesea
shares some features with some other members, although any two members may share no features in common. For example, the overlap of features in the items {ABC, CDE, EFG} binds them together as a set, distinct from the set {KLM, MNO, OPQ}. Further, the overlap of features that defines the category means that some features, like C and E, are more common than others, so that items bearing them, like CDE, are more typical of the set as a whole than are other items. Rosch (e.g., 1977, 1978) demonstrated that this internal structure of categories influences people’s behavior. Rosch and Mervis (1975) asked people to rate the typicality of members of the “furniture” category, and also to list the features of each item. They observed that people’s typicality judgments could be predicted from the number of features that any member shared with other members of the category. For example, the item “chair,” which shares features with many other members, including “table,” “desk,” and “sofa,” was rated as highly typical of the category, whereas “carpet,” which shares few features with any other member, was judged to be atypical. Rosch concluded that people’s judgments of typicality must be controlled by the distribution of features across members of the category. However, people never directly encounter that distribution: instead, they encounter only the individual members from which the distribution can be calculated. The question was thus how people come to be sensitive to the abstract properties of the category as a whole. In any such category, there is usually one item, called the prototype, that shares more features with all other items than any other does, and so is most typical of the set as a whole. Posner and Keele (1968) observed that the prototype has a favored status among category members: even if it is not shown in the training phase of an experiment, it is classified as well as any training item, and better than novel items. Moreover, the probability of correctly classifying other members of the category can be predicted from their similarity to the prototype (e.g., Franks & Bransford, 1971; Neumann, 1974; Rosch, Simpson, & Miller, 1976). This correlation of typicality of test stimuli with accuracy of classification convinced many researchers that the prototype must be directly represented in memory, even if it was never presented (e.g., Homa, Sterling, & Trepel, 1981; Rosch, 1977, 1978). It was thought to be computed through an unconscious, automatic abstraction mechanism, which tallies the frequencies with which various features occur in successive training stimuli. The prototype would embody the set of most typical features, and become the standard that people use to classify other members of the category. Despite wide acceptance of the idea that category learning proceeds through the abstraction of prototypes, the evidence for it remained correlational. There was thus the possibility of an alternate explanation of the
Constructive Processing in Remembering and Performance
223
relationship between typicality and accuracy. Brooks (1978) and Medin and Schaffer (1978) suggested that the relationship between typicality and accuracy could be explained equally well if subjects simply encoded each presented instance of the category, and compared test items to the set of encoded instances. Because the prototype is the average of those instances, a test item that is similar to the prototype is also necessarily similar to many instances of the category, whereas an atypical item would be similar to fewer instances, and also probably less similar to any specific instance. The prototype itself is more similar to other instances than is any other instance, and so would benefit most from this comparison in a test. The question thus became whether the relationship between typicality and accuracy is direct and actual, resulting from abstraction of the prototype during the learning phase, or indirect and potential, resulting from simply encoding the training instances. No direct evidence of prototype abstraction has yet been presented. However, there is now much support for the alternative hypothesis, that typicality effects in category learning result from encoding and preserving representations of individual training instances. For example, Whittlesea (1987) unconfounded the typicality of test stimuli (their similarity to the prototype) from their similarity to individual training items. I created two categories of items, based on the prototype strings FURIG and NOBAL. Training stimuli all differed by two letters from their prototype; for example, the FURIG-based training set was {FEKIG, FUTEG, PURYG, FYRIP, KURIT}. Subjects were shown each item three times, and pronounced and copied them on each occasion. Test items were presented tachistoscopically: the dependent variable was the subjects’ accuracy in identifying the component letters. In this test, there were two types of novel stimuli. One type (e.g., FUKIP) shared three letters with its prototype, the other (e.g., PEKIG) shared only two. However, the first type shared only three letters with the most similar training item (FEKIG), whereas the second shared four. This provided a critical test of the alternatives: if people abstract prototypes across instances, and use them to perform judgments about category members, then performance should be superior for the more typical items, whereas if they simply preserve representations of the individual stimuli, then performance should be superior for items more similar to actual training instances. Contrary to prototype theory, but in support of instance theory, the latter pattern of performance was observed. Further experiments in this series demonstrated that, although the most similar training stimulus exercises most control over performance on a test item, other instances can as well. For example, in one study, subjects were shown two types of novel test items (e.g., FUKIG and PEKIG) that were
224
Bruce W. A. Whittlesea
equally similar to the most similar training instance (FEKIG). However, FUKIG was more similar to the remaining instances of that category than PEKIG: it was also more accurately identified. In five further experiments, I observed that the subjects’ relative accuracy of identification could be accurately predicted from the simultaneous similarity of those items to the set of training instances. The observation that most impressed early investigators of the typicality effect was that the prototype item could be classified with equal, or sometimes greater, accuracy than items that had actually been seen in training. In contrast, in the two studies just reported, training items (e.g., FEKIG) were also presented in test, and were always better identified than novel test items, even items closer to the prototype (e.g., FUKIG). However, in those studies, training items presented again as test items had the distinct advantage of being identical to 1 out of the 10 items presented in training: In consequence, their average similarity to the set of training items was in fact greater than that of more typical items. I reduced this advantage by increasing the stock of training items to 15 per category. Now, any training item that occurred again as a test item was identical to only 1 out of 30 training items, and instead, items closer to the prototype were more similar on average to the training set. In this case, the pattern reversed: the novel, more typical items were at last classified more accurately than items the subject had actually seen in training. The results of these three studies lead to several important conclusions. First, the idea that categories are represented by an abstracted prototype is unnecessary and insufficient to explain people’s performance in tasks involving members of well-structured categories. Instead, memory preserves representations of particular instances of categories, and those representations control performance on novel instances in a perception or classification test. Second, neither similarity to a prototype nor prior presentation as a training instance is sufficient to predict relative success in identifying test instances. Instead, judgments about category members are driven simultaneously by all of the instances that memory has previously encoded. In small domains, or domains containing relatively distinctive members, this can cause instances that were actually experienced previously to be processed more accurately or efficiently than more typical instances. In contrast, in larger domains, containing less distinctive members, items that are more typical of the category as a whole will be more similar to the set of encoded instances, and will thus be processed optimally. Finally, these results stand as a warning against interpreting correlations of performance with general properties of a domain as evidence that people have actually abstracted those properties. Typicality is ordinarily a good predictor of performance in family-resemblance categories. However, no
Constructive Processing in Remembering and Performance
225
matter how strong the correlation, it is not direct evidence about the actual basis of performance. Unfortunately, as described later, investigators of numerous phenomena of category and concept learning have accepted such correlational evidence as evidence of automatic, unconscious abstraction, without performing critical tests. To understand the structure and operation of memory, it is necessary to perform direct experimental manipulations of the conditions of learning and test.
V. Theme 2: Memory Preserves Processing Experiences, Not Stimulus Structures
The original motivation of instance theory, as proposed by Brooks (1978) and Medin and Schaffer (1978), was to account for people’s ability to generalize from experience of a set of instances to novel instances of the same category; that is, it was a theory of concept formation and classification. Originally, the theory consisted of the idea that people learn about categories by learning their instances, as illustrated in the last section. That idea turned out to be very powerful, and has proved to be applicable in all manner of tasks and domains. Over the years, that original conception has been expanded and modified as its application was extended from learning about family-resemblance categories to repetition priming, word identification, implicit learning, and also to remembering. In particular, instance-minded investigators began to realize the importance of context, rather than of instances treated in isolation, and of the specific processes applied to stimuli in encountering them for different purposes. The focus shifted from thinking primarily about the structure of stimuli and the structural similarity of category members to thinking about the experience of a stimulus in some task and context, and the similarity of that experience to other experiences in the same or some other task and context. I chose to refer to this idea as episodic-processingtheory (e.g., Whittlesea & Dorken, 1993), rather than instance theory, to reflect the new emphasis. In fact, a similar idea emerged in parallel in a different literature. Morris et al. (1977) proposed transfer-appropriate processing theory, which states that performance in memory tests succeeds to the extent that the mental operations required in the test match those the subject performed on that item earlier. As originally conceived, transfer-appropriate processing was a theory of remembering, of recognizing the same item on a later occasion. More recently, it has also been applied to cases in which the specific nature of a prior experience influences later nonremembering performance on that item, such as identification (e.g., Roediger & Challis, 1992). From their
226
Bruce W. A. Whittlesea
different roots, episodic-processing and transfer-appropriate processing theory are converging. At the time that Brooks (1978), Medin and Schaffer (1978), and Morris et al. (1977) proposed their respective theories, concept formation and remembering were treated as utterly different topics: they involved very different tasks, and respectively focused on learning for the future versus retrieving the past, and the ability to handle generalization between stimuli versus knowing the contexts of specific events. Concept formation studies tended to focus on the structure of categories, and the structural relations among instances; remembering studies focused on processing variations at training (e.g., levels of processing; Craik & Lockhart, 1972) and test (encoding specificity;Tulving & Thompson, 1973).Investigators of one topic rarely thought about the other. However, during the 1980s,numerous investigators began to break down the barriers between these areas, and began using tasks that had been previously considered more appropriate for the other area. For example, Jacoby and Dallas (1981) trained subjects under a levelsof-processing procedure, but then tested them not only in a recognition task but also in a flash identification task. Unlike recognition, the identification task does not require the subject to think about his or her previous experience of the test items in any way. Tachistoscopic identification had long been used in another tradition, namely word identification, in which it was considered to be a direct measure of a subject’s knowledge of the general properties of words (e.g., Wheeler, 1970). Jacoby and Dallas instead used it as an indirect measure of the subjects’ specific learning about a word in the training phase of a remembering study, contrasted with the direct measure of asking the subjects whether they recognized the item. The use of indirect measures in remembering studies became increasingly common during that decade. Investigators of remembering began to use fragment completion, listening through noise, pleasantness judgments, and a host of other nonremembering tasks to find out about the learning that permits people to recognize specific events. At the same time, investigators of concept formation began to use a variety of indirect measures of category learning, including pleasantness judgments and recognition, in addition to the traditional classification task: these measures are indirect in the sense that they ask the subject to evaluate category members as individual entities or events, rather than to classify them as examples of a category. Very broadly (and subject to important exceptions), performance on these indirect measures was found to be different in some important ways from that on direct measures, but similar in other important ways. For example, Jacoby and Dallas (1981) observed that prior experience of an item in a training phase influenced performance in both identification and recognition tests, but that the levels-of-processing training affected only
Constructive Processing in Remembering and Performance
227
recognition. This kind of dissociation sparked an explosion of investigation and theory. Some investigators, such as Tulving et al. (1982), took it as strong support for multiple, separate systems of memory, arguing that the different tasks tapped into qualitatively different knowledge bases? Others, such as Jacoby (1983; Jacoby & Witherspoon, 1982), took it as support for a unitary memory preserving highly specific experiences of stimuli, experiences that could satisfy the requirements of different tasks to differing degrees. Despite this cross-borrowing of tasks, research on how information about isolated stimuli is acquired, stored, and accessed has continued largely in isolation from research about categories of stimuli, and principles important in one area are largely ignored in the other. For example, there exist two literatures, one on “implicit memory” (e.g., Roediger & Challis, 1992) and one on “implicit learning” (e.g., Reber, 1993), which have almost no cross talk despite obvious common interest. The former is characterized by a concern with the encoding and representation of individual stimuli (usually words), and with dissociations between direct (recognition) and indirect (e.g., identification) tests of a specific prior experience of those items. The chief tool used to understand these issues is variations in the processing required to be performed on an item at training and at test. Examining transfer from an experience with one stimulus to performance on a structurally similar stimulus is rare, although transfer along semantic dimensions is commonly tested. In contrast, investigators of “implicit learning” are concerned with how people encode and access information about classes of structurally related stimuli. The stimuli used are letter strings, numeral strings, or sequences of tones or colors: words are almost never used, so that semantic relationships cannot be examined. The most common manipulations in that area are of the deep structural similarity between test and training items and the identity of surface features used in training and test. Manipulations of the processing that subjects do in training and test through variations in task demands are rare. Even manipulation of the structure of the training set is rare: for example, most investigations of Tulving et al. (1982) were forced to go beyond the two stores, of semantic versus episodic information, that had previously been accepted. Jacoby and Dallas’ (1981) observation that a single, extra presentation of a well-known word in the training phase of an experiment could influence identification of that word in test could not be explained by either semantic or episodic memory. Tulving’s (e.g., 1985) solution was a third memory system, the Perceptual Representation System. Further observations led Tulving (e.g., 1995) to increase the number of proposed memory systems to five. As argued by Witherspoon and Moscovitch (1989) and Roediger (1990). there is no reason to suppose that the eventual number of hypothetical memory systems will remain manageably small, because dissociations between tasks is the major indicant of a new division of memory, and because dissociations are being increasingly observed.
228
Bruce W.A. Whittlesea
artificial grammar learning have used one artificial grammar, the one given by Reber and Allen (1978). In my own studies of “implicit learning,” described later, I have used designs that are rare in that paradigm, but common in the “implicit memory” literature. These include manipulating the processing at training and test so that they match or mismatch (as in transfer-appropriate processing studies), looking for dissociations between dependent variables, and comparing direct versus indirect measures of learning. These studies have convinced me that the emphasis on structural relations, so endemic to studies of “implicit learning,” has led the field astray. Implicit learning has been extensively studied using the “artificial grammar paradigm.” In this paradigm, stimuli are generated from a grammar (a set of structural rules dictating the sequence and repetition of stimulus elements), generating items such as MlTV, VXTVT, and VXVRXRM. The basic “implicit learning” phenomenon consists of the observation that subjects exposed to members of a rule-governed domain can later discriminate above chance between novel legal items and items violating the rules, even though they were not aware of the existence of rules during the training, and cannot state the rules in test (e.g., Reber & Allen, 1978). Two rival factions have emerged to explain this finding. One, the abstraction camp, argues that memory is directly sensitive to the structure of a stimulus domain: it automatically abstracts general information about the deep structure of the domain across the particular instances of the category that the subject experiences in the training phase. That is, in addition to people consciously learning something about the surface structures of individual items, they also compute some aspects of the general deep structure of the domain, either the rules themselves (Reber & Allen, 1978), or common letter sequences (Mathews et al., 1989; Servan-Schreiber & Anderson, 1990), or the covariance among features (Cleeremans, 1993). The opposing faction denies the necessity of abstracting general structure of the domain, either consciously or unconsciously, to produce the observed sensitivity to the rules. For example, Brooks (1978,1987; Vokey & Brooks, 1992) argued that if subjects simply memorized the training stimuli, as they were instructed to do, they would have some success in discriminating legal from illegal test stimuli, because of necessity novel stimuli that satisfy the rules will be more similar to training stimuli than illegal items. Similarly, Dulany, Carlson, and Dewey (1984) argued that the subjects’ knowledge consisted of salient subunits of items, segments that captured the effect of the rules, and that the subjects were quite aware of having this knowledge. Perruchet and Pacteau (1991) argued that even encoding no more than some random bigrams from the stimuli would enable subjects to discriminate legal from illegal items, because illegal items often contain pairwise violations,
Constructive Processing in Remembering and Performance
229
and legal items do not. All of these forms of information are available to be encoded directly in separate stimulus encounters, without any need for computation of abstract properties across items. Most of the debate over the problem of “implicit learning” has thus concentrated on the structure of the stimuli as individuals and as a group, and on what units of that structure people learn when they do not know the domain has a deep structure. The “instance theory” that I described in the last section would argue that people simply encode the structures of particular exemplars of the domain, and classify structurally similar test instances as legal. In contrast, its descendent, the “episodic-processing account” (Whittlesea & Dorken, 1993), argues that people encode processing experiences, not knowledge structures. According to that account, every learning event consists of processing the structure of some stimulus in some context and for some purpose: different contexts and purposes cause the person to impose different organizations on that stimulus structure. The subject’slater ability to perform a classification test depends on the demands of that test, and on how they processed training items, given the task at that time, not just on the structure of the individual instances (cf. Vokey & Brooks, 1992). To demonstrate this, Whittlesea and Dorken (1993) created two grammars, and required subjects to spell instances of one grammar and to pronounce instances of the other. After this training, we explained that the items were taken from two categories, created by two sets of rules. We then showed subjects novel instances of each grammar, asking them to classify these items as belonging to the category they had spelled earlier, or to the category they had pronounced earlier. Before making that decision, they were asked to pronounce or to spell the test item. This activity was crossed with what they had done earlier with similar items of the same category: We required them to spell half of the items taken from the spelling category and to pronounce the rest, and to pronounce half of the items from the pronounced category and to spell the rest. That is, the processing that subjects performed on test items just before classifying them either matched or mismatched the processing that they had done on members of the same category; but the structural similarity of a test instance to other members of its category was the same however they were processed at training or test. If learning about the structure of training items, either at an abstract, deep structural level or at the level of individual instances or bigrams, is the source of “implicit learning,” then our manipulation should have no effect. But if the implicit development of sensitivity to categories is mediated by representations of particular experiences, then we could expect to observe processing-specific transfer.
230
Bruce W. A. Whittlesea
The subjects had some ability to discriminate the test items of the two categories: Overall, their accuracy was 64%, against a chance rate of 50%. However, their success in classifying the test items depended on whether the task they performed on the test items (spelling or pronouncing) was the same as the task they had earlier performed on other members of that grammar. When the processing at test matched that at training, their success in classification was 6796, but when the tasks were different, their success was only 61%. In a second experiment, the test task was changed, from discriminating between the categories of grammatical items to discriminating grammatical items (of either category) from nongrammatical items. Again, subjects spelled or pronounced novel test items before classifying them, and the test task was congruent with the training task for half of the grammatical test items. The subjects showed excellent ability to disciminate legal from illegal test items: overall, they claimed grammatical items to be legal on 57% of trials, and nongrammatical items on only 27%. However, once again we observed that the match between training task and test processing task was important. The subjects correctly classified novel items as grammatical on 66% of trials when the training and test tasks matched, and on only 48% of trials when they did not. These results demonstrate that subjects exposed to well-structured stimuli in an implicit learning experiment do not simply acquire knowledge units at some level of abstraction, such as bigrams, instances, or abstract rules. Instead, in satisfyingthe purpose of the encounter, they impose organization on the structure of the stimuli, an organization specific to the task. Performance on subsequent stimuli does not depend simply on the structural similarity of those stimuli to training stimuli, either individually or as a set, but instead on the similarity of those structures as processed in the training and test tasks. Successful transfer to novel stimuli depends on having performed appropriate processing in the training. Proponents of the abstraction hypothesis agree that such results do demonstrate that some part of the transfer observed in implicit learning experiments is due to encoding specific experiences, but argue that there is always residual variance in such experiments that cannot be explained in that way (e.g., Reber, 1993). For example, in the experiments just described, the match-mismatch manipulation does not account for all of the subjects’ ability to discriminate legal from illegal items. The residual variance is argued to be due to abstraction of general structure. That argument is weak: the residual variance could as easily be explained through variability in experiencing and encoding stimuli that was controlled by factors other than the experimental manipulation, such as attention, salience, pronounce-
Constructive Processing in Remembering and Performance
231
ability, and the orthographic similarity of various items to natural words. In any case, it is negative evidence at best. Whittlesea and Dorken (1993) examined the “mixture of knowledge” idea in a different way. If abstraction of general structure inevitably occurs when people process members of a category attentively, then changes in the purpose for encountering the training items should make no difference to that activity. We asked three groups of subjects to study instances of a grammar, such as RXXCTXT and MTTMRCT, for one of three purposes (none of the groups was informed about the existence of categorical rules). One group was asked to memorize the stimuli for a later test (the usual induction task in implicit learning experiments) by rehearsing them aloud. Another group was also asked to rehearse the items aloud, but that rehearsal was incidental to their overt task: Subjects were given a 3-digit number to remember, then shown a grammatical instance and asked to rehearse it for 10 s as a distractor, and then asked to recall the number. A third group was asked to say whether each letter of each stimulus was repeated elsewhere in the item. Subjects in all three groups thus processed the entire structure of each stimulus, but understood their task differently. At test, all subjects were told about the grammar and asked to classify novel items as legal or illegal. All groups were tested on novel legal and illegal items presented in the same letter set as training items, and on the same items presented in a novel set of letters (e.g., the item RXXCRCT being replaced with the item FPPLFLS, which has the same deep structure but a novel surface structure). These tests present different demands. The former can be accomplished simply by encoding some information about the surface structures of some stimuli, coded in terms of the identities of the letters; but the latter requires the subject to have already encoded, or to be able to generate at test, some information about the abstract pattern of repetition of letters within items. All three groups performed above chance in the same-letter test, and to about the same degree. However, in the novel-letter test, the rehearsal-asdistraction group showed no reliable ability to discriminate the items. The memorization group showed a reliable ability to do so, but less than in the same-letter test. The analytic group performed at identical levels in the two tests, and that group performed reliably better in the changed-set test than did the memorization group. That is, the first group showed no ability to discriminate items on the basis of their deep structure: the memorization group could do so to some degree, but less than subjects who had analyzed the patterns of training items. We concluded that the different purposes that subjects had been given for encountering the stimuli had caused them to process training items in different ways, that differentially prepared them for the different demands of the tests. We further concluded that the implicit
232
Bruce W. A. Whittlesea
development of the ability to discriminate legal from illegal items does not demonstrate the operation of automatic, chronic abstraction of deep structure. Instead, that ability, when it occurs, is an incidental by-product of processing training stimuli for a specific purpose. Some proponents of the abstraction hypothesis argue that memory does not abstract information directly from the stimulus, but instead from the subject’s experience of the stimulus (e.g., Mathews & Roussel, 1993). This might go some way to explaining why the subjects who rehearsed instances incidentally in a number remembering procedure were later unable to discriminate items on their deep structure, even though they had repeatedly rehearsed those items. However, this argument makes the abstraction function seem a bit odd: instead of granting implicit sensitivity to the real structure of the world, as envisioned by Reber and Allen (1978), it instead is a running summary of the average properties of what one has done with stimuli. By this argument, memory codes its experiences twice: once in full episodic detail, and, again, computing the average of those experiences. Such redundancy is not impossible, of course: but given that that summary is implicitly represented in the set of episodic representations, so that memory can be sensitive to those average properties without directly computing them (as demonstrated in the last section, on learning of family resemblance categories), it seems peculiar that memory would have developed a special mechanism to perform that activity. However, such speculations are less important than direct evidence. Whittlesea and Wright (1997) tested the abstraction hypothesis in a different way. We used only one training procedure. All subjects were asked to memorize a set of grammatical instances for a later test (without, of course, being informed about rules). In consequence, all subjects had the same purpose for processing, and would not experience the stimuli in qualitatively different ways. The stimuli were letter strings, such as RMRMCTX and RMCCTXT. We examined what subjects had learned using direct and indirect tests of grammatical sensitivity. The direct test was classification. The indirect test was a pleasantness judgment, in which subjects were asked to describe each stimulus as pleasant or dull. Pleasantness judgments have been observed to be affected by prior experience of grammatical instances in the same way that classification is. For example, Gordon and Holyoak (1983) found that subjects claimed to find novel instances of a grammar more pleasant than novel illegal items. However, in that case, test items were presented in the same surface features as training items: that is, using the same set of letters in both cases. Our test items were either identical in deep structure to training items (legal), or simply different (illegal). But in either case, test items were presented in a novel surface structure, being shown either in a novel set
Constructive Processing in Remembering and Performance
233
of letters or as patterns of color patches instead of letter strings. For example, an item shown as RMRMCTX in training was now shown either as QFQFSPL or as a row of squares colored: red, yellow, red, yellow, green, blue, brown. When test items were presented in novel letters, subjects had a small but reliable tendency to find grammatical items pleasant (about 53%), but a greater ability to classify them as grammatical (about 60%). When test items were presented as strings of color patches, they were still able to classify stimuli as legal or illegal above chance (about 59%), but their pleasantness ratings fell to chance levels. That is, the actual status of the item as conforming to the rules of the grammar influenced the subjects’ behavior in the direct, classification test, but not in the indirect, pleasantness test. These results are difficult to explain through the abstraction account. That account argues that subjects abstract the common structure of their experiences of a set of stimuli, and apply that knowledge in dealing with novel members of the domain. Although the acquisition of that knowledge may be conditional, based on the subject’s specific processing and experience of the training stimuli, the application must be automatic, driven by the structure of the test stimuli. After all, in the standard implicit learning experiment, subjects acquire their knowledge in a memorization task, but are supposed to apply it unconsciously to novel stimuli in a completely different task, namely classification. If the application of that knowledge truly is implicit,then it must also be automatic: in fact, the idea that implicitly acquired knowledge of general structure could control people’s behavior without their awareness was the promised magic of “implicit learning.” Moreover, Gordon and Holyoak (1983) showed that memorization of grammatical instances can control performance in indirect tests such as pleasantness judgments as well as in direct tests, if test stimuli bear the same surface structure as training stimuli. If the basis of performance in the classification tests were the automatic application of abstracted knowledge in the presence of familiar deep structure, then we ought to observe an effect on the indirect test whenever we see an effect in the direct test. Instead, I argue that test performance is based on the cues presented by various test tasks and stimuli, and the ability of those cues to access specific representations of prior experiences: that is, the stimulus compound, not just the stimulus. The classification task does not simply offer subjects legal and illegal stimulus structures. It also comes with the instruction that the discrimination is to be made relative to a specific set of prior stimuli. This instruction cues the subjects to recall any possible information about those stimuli. If the subjects can recall even a few of those items, they can succeed in the classification task above chance, even though the test stimuli are presented in a novel surface structure, by performing what Brooks and
234
Bruce W. A. Whittlesea
Vokey (1991) called an “abstract analogy”; that is, having recalled a training stimulus, they can analyze the repetition pattern of that item, and use that information to evaluate the legality of the test stimuli. They thus achieve some abstract knowledge, but about individual items rather than the general properties of the set. Moreover, they do so deliberately and consciously, and at the time of test, rather than unconsciously and automatically during exposure to training stimuli. In contrast, the indirect, pleasantness judgment does not provide that direct cue back to the training set. Instead, experiences of training items will be cued to control processing in this task only if the stimuli are themselves specifically similar to those experienced in training, not only in formal structure but also in perceptual manifestation. In Gordon and Holyoak’s (1983) study, presenting test stimuli in the same letters as training items made them specifically similar in both ways. In Whittlesea and Wright’s (1997) changed-letter test, the perceptual manifestation was different, but at least the surface structure still consisted of letters, cuing the previous experiences to a slight extent. However, when we changed to color-patch presentations, the only cue to access the prior representations was the fact that legal test stimuli were of identical deep structure to training items: and that was not sufficient to cue those representations. I conclude that the phenomenon of “implicit learning” does not require that memory have the function of automatic, unconscious abstraction of structure. Instead, that phenomenon is a by-product of encoding stimuli for particular purposes, experiences that can facilitate other performance in unanticipated tasks on a later occasion. On every occasion of processing a stimulus, one becomes aware of some aspects of processing (those related to the purpose of the encounter), but also performs and records other activities without awareness. The preserved record of those activities can support processing on a later occasion, without the person becoming aware of the source of influence. For example, in reading the word tiger, I may become aware of the meaning of the word, because that is the purpose of the encounter, but I’m also acquiring a representation of its orthographic code, without being aware that I’m doing so. That representation can prime later identification of that word in a similar context: that is, I have implicitly acquired the potential to perform better on a later, unanticipated test of identification. When that priming occurs, I may become aware that I have performed surprisingly well, but not be able to determine the source of that effect (either because I cannot recall the earlier experience, or because I do not understand the effect that that experience could have on my current performance). However, the lack of awareness of the creation and realization of the potential to perform does not call for a separate form of memory. Instead, as I argue later, it simply demonstrates the general prob-
Constructive Processing in Remembering and Performance
235
lem that people have in becoming aware of the sources of their performance, a problem that occurs in classification and identification as well as in the act of remembering.
VI. Theme 3: Selective Use of General and Particular Knowledge Is Controlled by the Stimulus Compound The previous two sections have concentrated on experiences of individual stimuli as the basis of performance in nonremembering tasks, to counter the abstractionist claim that performance in those tasks depends on abstracted general knowledge. However, the SCAPE account does not argue that people possess knowledge only about individual stimuli: quite obviously, people can abstract general properties of their experience, and do know information about classes per se. However, the account claims that they do so only in the service of some particular purpose for processing, and not as an automatic, chronic function. Acquisition of such knowledge may be deliberate, as in analyzing instances of a category to discover what they have in common, or computed incidental to performing some other judgment that requires direct knowledge about abstract properties of the domain, such as deciding whether a person is relatively short for an athlete. The means of acquiring representations of such knowledge is different than that of learning about individual stimuli, as it necessarily involves comparison of stimuli and analysis of their common or different properties.The contents of the representations are also quite different, consisting respectively of experiences of individual instances and of class properties of a domain. However, representations of general knowledge are acquired through the same general principles as knowledge about individual stimuli, through the construction of an experience of a stimulus in some task and context, guided by similar experiences in the past, and are accessed in the same general way, through the occurrence of a further stimulus complex that cues similar prior experiences. This account conflicts with the separate-systems explanation of the acquisition, representation, and application of general and particular knowledge. By that account, as discussed earlier, those kinds of knowledge are acquired by different mechanisms and preserved in qualitatively different stores. Moreover, in that account, the chief factor that dictates which type of knowledge is accessed on any occasion is the type of task. Tasks that specify particular contexts, such as remembering, will selectively invoke representations of specific experiences; in contrast, tasks that present stimuli as instances of a general class will cue the application of general knowledge.
236
Bruce W. A. Whittlesea
Whittlesea, Brooks, and Westcott (1994) examined the selective use of general and particular knowledge using the family resemblance domain described earlier, based on the FURIG and NOBAL prototypes. In this study, the FURIG stimuli were presented as artificial nouns, bearing an -ISM suffix (e.g., FEKIGISM, KURITISM), and members of the NOBAL category were presented as artificial verbs, bearing an -1NG suffix (e.g., NEKALING, KOBATING). In all experiments of this series, the subjects were required to learn about the stimuli as individual entities, but were also required to compute abstract knowledge about the set. In one training phase, they pronounced and copied each stimulus (complete with suffix), thereby experiencing it as a unique entity. In another, they were shown the same stimuli, told whether each was a noun or verb, but shown only the stem (e.g., FEKIG). On each trial, they were asked to judge how frequently each letter of the stem had occurred in stimuli of that class on preceding trials, thus computing a running summary of the typicality of each letter across the set. The question was which form of knowledge, about specific whole items or the typicality of their features, the subjects would apply when they had encoded both, and when they could use either to perform a test task. This was evaluated through comparison of two types of items, one (e.g., PEKIG) that was more similar to a particular training instance (FEKIG) but less typical of the whole set, and another (e.g., FUKIP) less similar to any training instance but more typical of the category (with prototype FURIG). At test, one group of subjects was shown whole items (e.g., PEKIGING) and asked whether they were completed with the correct suffix. Another group was shown only the stems of items (e.g., PEKIG), and asked whether they were nouns or verbs. Although these tasks are logically identical, and can be performed using either knowledge about individual features or knowledge of whole items, the former presents itself as item verification, the latter as verification of membership in a class. In the former task, subjects relied on their knowledge about particular exemplars; but when asked to classify the stem as a noun or verb, they instead relied on the typicality of the features. That is, the subjects used particular and general knowledge selectively, depending on whether the task involved treating a stimulus as an entity in its own right or as an instance of a class. This result is consistent with the assumptions of the separate-systems account. However, in four other experiments, we observed that the use of those forms of knowledge was also controlled by a variety of other characteristics of the task. In one study, subjects were offered a forced choice between two novel stimuli, complete with suffixes (e.g., FEKIGING and
Constructive Processing in Remembering and Performance
237
NEKALING): they had to decide which had the correct suffix. The stimuli were shown side by side for half of the subjects; for the other half, the stimuli were presented one above the other. Although both tasks treated the items as individual entities, the subjects relied on similarity to particular training items in the former case, but on the typicality of separate features in the latter. In another study, we presented a category name (e.g., “verb”) and a whole item (e.g., NEKALING), asking subjects to verify that the stem belonged to that category. Half the subjects received the category name in advance of the item; half received the item before the category name. Although the demand was to classify in both cases, subjects relied on typicality when the category name was supplied first, but on item knowledge when the item was presented first. Contrary to the separate-systems account, these studies demonstrate that the selective application of general and particular knowledge is not unilaterally controlled either by the structures of stimuli or by the type of task. Instead, selective use of the two levels of knowledge was cued by the specific perceptual organization and cognitive operations that the subject imposed on stimuli in the test. When stimuli were presented side by side, it was easier to compare them by coding each as a whole item, whereas vertical presentation made it easier to compare them a letter at a time. Those organizations of the stimuli were, respectively, more similar to representations of whole training items or of separate features, and selectively cued those representations, regardless of the demand to verify items as individuals. Similarly, when the name of the class was presented first, it cued class-level information, and caused classification to be based on that knowledge; whereas presenting the item first cued representations of similar items, causing classification to rely on that basis, regardless of the overt type of task. In general, the selective use of prior experience is controlled by the similaritybetween a stimulus compound (the structure of a present stimulus, as processed under the perceptual and cognitive operations performed to satisfy the current task) and specific representations of prior experiences (the structures of previous stimuli, as encoded given the operations that were performed at that time). That same principle applies whether the current stimulus is treated as a unique entity, occurring in some specific time and place, as in a remembering task, or as an instance of a generic class, in a classification or identification task. The fundamental organization of memory is not a dichotomy of knowledge types, acquired, stored, and applied by different principles. Instead, memory is organized through the specific similarity of the current experience of a stimulus to prior experiences of other stimuli.
238
Bruce W. A. Whittlesea
VII. Theme 4: Processing in Large, Familiar Domains Is Also Controlled by Specific Experiences Previous sections of this chapter have dealt with the acquisition and use of information about novel domains. A subject’s total experience of such domains during an experiment may amount to no more than thirty or forty instances. However, people deal every day with familiar domains, such as words, that have thousands of separate identities, and uncountable numbers of specific instances. It is worth asking whether processing in artificial domains is representative of processing in familiar, natural domains, such as the identification of natural words and objects. People’s success in identifying a word is to a large extent predictable from its typicality or regularity. For example, orthographically regular words (Wheeler, 1970) and phonologically regular words (Glushko, 1979) are easier to identify than other words: so are words that occur with high frequency (McClelland & Rumelhart, 1981) or belong to high-density orthographic neighborhoods (Andrews, 1992). Such observations have led many researchers to conclude that these abstract properties are directly represented in memory, and directly control the act of identification. This assumption is the basis of the “dual-route” hypothesis of word identification (e.g., Paap & Noel, 1991). Familiar words have a representation in a “lexicon,” a mental table connecting the perceptual manifestations of particular words with their pronunciation and meaning codes. These lexical representations correspond to prototypes in the category literature: They are representations of the essential, abstract structure of the word, summarized across thousands of particular experiences of that word. In consequence, a word like table is represented only in terms of the compound of its features {T, A, B, L, E}; no information about the variations in letter size, font, handwriting style, or text context in which the word has been encountered is included in the representation. This form of knowledge is thought to be responsible for the word-frequency effect. Unfamiliar words are thought to be identified through the other route, consisting of a set of spelling-to-sound rules that can be used to construct a pronunciation for a letter compound not found in the lexicon. The evidence that people have such rules for constructing pronunciations, even though they cannot describe them, is the same as the evidence for the abstraction of rules from instances of artificial grammars: the subjects’ performance is predictable from the rules. The dual-route hypothesis suffers from the same problem as the prototype- and rule-abstraction hypotheses of category formation: It is based on correlational evidence. The fact that people behave the way they would if they had abstracted logogens or rules does not prove they have
Constructive Processing in Remembering and Performance
239
done so: The abstract properties of words are confounded with the specific experiences that they summarize. High-frequency words have actually been experienced in more events and contexts than less frequent words; regular words are structurally similar to more words, encountered in more specific contexts, than are irregular words. In consequence, effects associated with frequency, typicality, regularity, and neighborhood size would occur whether people actually abstract general, deep structural properties across their experiences, or encode and preserve only their specific experiences with words. These alternatives can be disambiguated by providing subjects with new, distinctive experiences of familiar words. The SCAPE account assumes that each occurrence of a word will be encoded as it was experienced. That means that aspects of the experience that are definitionally irrelevant to naming it, such as the context or font, will be incorporated into the representation, and will influence later interactions. In contrast, the dual-route account argues that the effective knowledge controlling identification consists of abstract information about the structural relationships among features, coded in a lexical entry or spelling-to-sound rule. In consequence, if the dual-route account is correct, then performance in identification tasks should be broadly stable. A single specific experience of a word that has been seen many times in the past should not greatly affect identification of that word on a later occasion. In contrast, if performance depends on the similarity of the current stimulus to prior processing episodes, then identification performance should be highly modifiable, simply by providing one new but distinctive experience. Whittlesea and Brooks (1988, Experiment 7) investigated the influence of context on word identification. In a study phase, we presented common words either in isolation (e.g., rough, steel) or in a simple phrase (e.g., dirty black wall, big wooden box). The question was whether this presentation of a common word, in a specific context, would have an effect on the subjects’ ability to identify that word later. In the subsequent test phase, we presented training words in a variety of contexts, some presented alone and some in phrases. These test presentations endured for only 30 ms and were followed by a pattern mask. We observed that a word that had originally been presented alone (rough) was identified very accurately (70% of trials) if it was again presented in isolation at test, but less often if it was now presented in a novel context (dry rough planks: 34%) or in a context that had been associated with a different item in training (dirty rough wall: 28%). Similarly, words that had been shown originally in a phrase (big wooden box) were well identified when presented in that same context (61%), but less often when presented alone (52%). Subjects had even more difficulty when such words were presented in a new context (nice wooden
240
Bruce W. A. Whittlesea
chair: 43%) and had most difficulty when they were shown in an old but re-paired context (dirty wooden wall: 31%). These results demonstrate the phenomenon of repetition priming (cf. Jacoby & Dallas, 1981). However, they go beyond the standard demonstration of that effect, showing that the effect of a prior experience is not simply to facilitate current performance on a word. Instead, the amount of facilitation of a repeated word depends on the similarity of the test context to the original training experience. Prior presentation of a word in a phrase can even interfere with later identification of that word: target words inserted into contexts originally associated with a different word were harder to identify than words presented in novel contexts or in isolation. We concluded that memory preserves specific experiences of words in particular contexts, and that word identification is critically influenced by these particular prior processing experiences, not just by the average, abstract structural properties of the mass of prior experience. Forster and Davis (1984) argued that repetition priming effects occur because specific experiences of words are stored in an episodic memory system. In contrast, they argued, identification of words not re-presented in specific contexts is performed through access to the lexicon, which is part of a semantic memory system. Similar arguments, implicating multiple memory systems, have been made by other investigators, including Tulving (e.g., 1985,1995). By this argument, priming is a special case, that does not inform us about the ordinary activity of identifying a word. However, that argument is based on the idea that priming occurs only under exact representation of a word, which would be a rare and exceptional event. The evidence mentioned earlier demonstrates that prior specific experiences can influence current identification in various ways, even when the test presentation is not identical to the training. That suggests that the phenomenon of priming is much more common than initially thought. In fact, it suggests that every occasion of processing a word is influenced by prior experiences of that word in similar but different contexts, and probably by many prior specific experiences in concert. This effect can be documented only when a person’s prior experience of a word is known in detail, as in a priming study: but the fact that an investigator does not know in detail the mass of a person’s experiences that influence the identification of a regular word does not mean that that identification is based on knowledge abstracted across those experiences. I therefore argue that the basis of word identification cannot be deduced from comparing performance on words differing in frequency or regularity, because the person’s specific history that makes those words frequent or regular is not known. Finding that performance is correlated with aspects of general structure does not mean that the original learning of that domain
Constructive Processing in Remembering and Performance
241
proceeded through abstraction of that structure, or that performance in that domain is now controlled by knowledge of abstract structure. Instead, the principles of word identification must be inferred from observations about the acquisition of new words and new experiences of old words, in studies controlling not only the structural properties of items but also the specific processing that people apply to those structures and the contexts in which they are presented in training and test. Under those circumstances, as demonstrated by the experiments described in the previous three sections, it is clear that memory preserves specific processing experiences, and that later processing is controlled by the similarity of that event to earlier ones, not only in structure, but also in context and the purpose of the encounter. This can be observed whenever the prior experiences are unique and distinctive, so that their effects can be individually assessed. When instead the person has multiple similar experiences, those experiences will exert simultaneous influence on test performance, so that their individual influences are harder to detect. Performance will be correlated with their average properties, although those properties are not directly computed or directly responsible for performance.
PART I1 CONSTRUCTIVE EVALUATION VIII. Theme 5: Remembering Is Reconstruction, Not Retrieval One of the major barriers to producing a unitary account of memory is the idea that remembering involves retrieval of information from memory to consciousness,whereas in nonreflective tasks such as perception and classification, memory’s role is to guide processing, without the guiding representations emerging in consciousness. In this section, I argue that “retrieval” is a faulty metaphor for remembering: that in the act of remembering, just as in perception, memory representations cued by a current stimulus guide further processing of that stimulus without themselves being transported to consciousness.Further, I argue that the feeling of “pastness” that separates remembering from other memory-supported activities is not a direct product of interacting with memory, but the result of evaluating that interaction. In the next section, I extend this argument, suggesting that memory-guided
242
Bruce W. A. Whittlesea
processing followed by interpretation of that processing is a very general phenomenon, and is responsible for all manner of conscious states. The essential quality of remembering is the feeling of familiarity. A person who tries to remember a past event can tell themselves a story about what might have happened. If that story is experienced without a feeling of familiarity, then the person will not feel that they are remembering, even if the story exactly corresponds to the event. In contrast, the experience of a feeling of familiarity in encountering a stimulus is usually sufficient to make people feel that they have prior experience of that stimulus or one like it, even if they cannot recall any details of the event. The fundamental problem in understanding remembering is thus to understand the source of the feeling of familiarity. Intuitively, a feeling of familiarity occurs when one encounters a stimulus that one has encountered before. By this intuitive understanding, the earlier occurrence establishes a trace in memory: the second encounter activates that trace, and the activation is experienced consciously as a feeling of familiarity. By this account, the possession of a trace of an earlier experience is a necessary and sufficient cause of the feeling of familiarity. Contrary to that idea, Jacoby and his colleagues argued that the feeling of familiarity is the product of an unconscious attribution process (cf. Jacoby, Kelly, & Dywan, 1989). Jacoby and Dallas (1981) showed subjects a list of words, and then showed old and new words in both a recognition and a tachistoscopic identification test. They observed that subjects were more likely to claim to recognize test items that they could identify more readily. They suggested that the subjects were employing a “fluency heuristic” to perform the recognition task. The idea was that the earlier experience of an item facilitated later processing of that same item presented in the same context (the phenomenon of repetition priming). That additional fluency assisted subjects in identifying the words: They also used it as a basis for deciding that they had seen that item previously. Jacoby and Dallas concluded that the latter inference was performed unconsciously: In the context of a recognition judgment, the extra fluency simply felt like familiarity. Although that original observation was based on correlational evidence, direct experimental evidence for the fluency heuristic was later provided by several investigators. For example, Whittlesea (1993) showed subjects short lists of words, each word in a list presented for 67 ms. Following each list, subjects were shown a sentence stem, such as She saved her money and bought a . . ., and were asked to read it aloud. The last word of the sentence was then presented, lamp in the example. The subjects named this word, and were then asked whether that word had been shown in the preceding list.
Constructive Processing in Remembering and Performance
243
The subjects could discriminate new and old words to some degree, claiming to recognize repeated words 16% more often than new words. They also pronounced repeated words about 100 ms faster than new words. The question was whether that difference in processing fluency produced a feeling of familiarity, a feeling that the subjects could use to perform the recognition task. Unknown to the subjects, half of the sentence stems were designed to facilitate pronouncing the target word. For example, one stem was The stormy seas tossed the. . ., completed with the terminal word boat. Compared to the first example, this stem is to some degree predictive of the target word. That covert manipulation reduced the latency of pronouncing target items: subjects were about 130 ms faster in pronouncing target words presented in a predictive context. That is, the test context affected the fluency of naming in much the same way as did a prior presentation of the stimulus. If extrafluent performance creates a feeling of familiarity, and if people cannot discriminate fluency resulting from a prior exposure from fluency caused by the predictive test context, then manipulation of the test context should produce false feelings of familiarity. On trials when the target word was presented in the preceding list, the subject could claim to recognize the target either through a feeling of familiarity alone, or by using familiarity coupled with production of some contextual detail of the earlier exposure. However, on trials when the target stimulus was not presented in the preceding list, claims of recognition could be based only on a feeling of familiarity: there was no earlier event to recall. On those trials, the subjects claimed to recognize target words in predictive contexts 18%more often than words in nonpredictive contexts. Apparently, the subjects were sensitive to the fluency of their test performance, and used differences in that fluency to perform the recognition judgment. However, they were not sensitive to the source of the fluency, mistaking enhancement of fluency that was due to the covert manipulation of test context for fluency due to a prior presentation. I am not suggesting that people usually make errors in feeling objects to be familiar: ordinarily, test contexts are not engineered to provide a spurious source of fluency. However, the illusion of familiarity caused by enhancing the processing of novel test items demonstrates that the feeling of familiarity is not a direct result of cuing a prior experience of the stimulus. Instead, the ordinary development of a feeling of familiarity is a twostep process, a production followed by an evaluation. In the first step, on presentation of the stimulus, cued memory representations construct a percept of the stimulus, and perhaps a full identification of its name, depending on the task. This production does not of itself produce a feeling of familiarity, but it occurs with some fluency. This fluency is likely to be
244
Bruce W. A. Whittlesea
greater if the person has previously experienced that same stimulus in a similar context. In the second step, the person evaluates the fluency of production. From many prior experiences, the person has an expectation about the fluency with which he or she could expect to name that word on the basis of his or her general experience of that word in other contexts. If the fluency of performance in the present context exceeds that expectation, the additional fluency will be attributed to a prior encounter with the word in that same context. This evaluation and attribution are not conducted deliberately or consciously; instead, extrafluent processing is simply experienced as a feeling of familiarity. Whittlesea and Williams (1996) discovered that the evaluative process is driven by the surprise value of the fluency of processing, not by the fluency per se. In one experiment, we asked subjects to memorize a list consisting of natural words (e.g., RAINBOW, TABLE, CANDLE), nonwords that are easy to pronounce (e.g., HENSION, FRAMBLE, BARDEN), and hard nonwords (e.g., JUFICT., STOFWUS, LICTPUB). At test, subjects were shown those items plus an equal number of novel items, of all three types. In this test, they were asked to pronounce each item, decide if it was a word or nonword, and then decide if they had seen it earlier. Novel words were pronounced about 150 ms faster than novel easy nonwords, which in turn were pronounced about 400 ms faster than hard nonwords. Thus, if fluency per se is responsible for feelings of familiarity, then we could expect most false alarms on the words, followed by easy nonwords, followed by hard nonwords. That prediction was disconfirmed: false alarms for easy nonwords (37%) were much higher than for words (13%) or hard nonwords (9%). We concluded that the false feeling of familiarity was produced by surprise. Pronouncing the words was easy, but the subjects expected it to be: they knew that they had much prior experience of those words. Pronouncing the hard nonwords was difficult, but again, having determined in the lexical decision task that they were nonwords, the subjects were not surprised at their slow pronunciation. However, the easy nonwords were pronounced fairly fluently, although the subjects knew from their lexical decision that they were nonwords that they had never seen prior to the experiment. Their pronunciation of those items was thus not just fluent, but surprisingly fluent. Normatively, this extra fluency should have been attributed to the orthographic regularity of the items, not to a prior exposure. However, in the context of a recognition judgment, the subjects experienced this unexpected fluency as a feeling of familiarity. Looking back, we realized that surprise was also the basis of the illusion of familiarity in the Whittlesea (1993) study, just described. In that case, the subjects were unaware of the context manipulation, and so were sur-
Constructive Processing in Remembering and Performance
245
prised when a test item in the predictive context was processed with extra fluency. The idea that surprise is the basis of familiarity helps to explain spontaneous feelings of familiarity, experienced when a person is not being interrogated about their prior experience of stimuli. The feeling of familiarity does not occur when encountering factually familiar (and fluently processed) stimuli in expected contexts, for example, meeting one’s spouse at home or the clerk in the corner store. One knows who those people are, but does not experience any pressing feeling of “I’ve seen you before.” Instead, that feeling occurs when those stimuli are encountered in an unexpected context, for example meeting the clerk on a bus. In such cases, processing proceeds much more efficiently, completely, and elaborately than one would expect for stimuli encountered in that context. Given that there is nothing else obvious to attribute that greater fluency to, it will be attributed to prior experience, and experienced as a feeling of familiarity. Moreover, this helps to explain why the feeling of familiarity stops when one is finally able to produce the context from which one knows the person (the corner store): in that case, the fluent processing is no longer surprising, just as it is not in actually encountering that clerk in the store. There is, of course, more to remembering than familiarity, although that is essential. As Mandler (1980) argued, recognition can be performed on either of two bases: by a feeling of familiarity alone, or through additional recall of the contextual detail of the earlier experience. The recall of the detail of an event certainly sounds like retrieval. And indeed, if it could be established that the content of mind on some later occasion was coextensive with the content of some representation of a prior experience, then it would make sense to speak of that as retrieval of the representation. However, evidence going back to Bartlett (1932) demonstrates that the detail of an experience that is recalled is not the same as the original experience. Instead, it is subject to various inferences, omissions, and additions (cf. the variety of distortions of remembering documented by Franks and Bransford, e.g., 1971). Even in the immediate case, experiments by Lindsay and Read (1994) and Roediger and McDermott (1995) suggest that the act of recall is as inferential and remote from the representations as is the feeling of familiarity, and for the same reason. Given a list of words, such as bed, night, pillow, dream, and blanket, and later asked to recall the list, people often falsely claim to recall seeing the word sleep in the list. Because that word is thematically related to the others, it is easy for the person to generate sleep in thinking about the possible words that might have occurred in the list. Once generated as a candidate for recall, the word must be evaluated: did it come to mind because it was in the list, or because it was thematically related to the other words? The only way that the person can decide this issue is through experiencing a feeling of
246
Bruce W. A. Whittlesea
familiarity for the word or for the contextual detail that comes to mind with the word. But as just demonstrated, that feeling is itself the product of an attribution about the ease with which the word can be processed, not an independent source of evidence about the existence of a representation of that word in that context. Thus, familiarity is fundamental to remembering. In an extended act of remembering, a current stimulus and context cause further ideas to come to mind. Those ideas cue production of further ideas yet, until one has a full-blown story in mind. However, each of those component acts of recall consists in principle of the same thing: a current stimulus complex engages memory, which produces a new mental event. If that new mental event is evaluated as feeling familiar, on the ease with which it comes to mind, it will be used as part of a new stimulus complex to stimulate memory into further production. Never in this chain is there any retrieval. At each stage, memory produces a mental event, and that event is evaluated and attributed to some ~ o u r c e . ~ The attribution of surprising fluency to a particular source, either in the present or in the past, is heuristic and inferential: it is also guided by the context in which the decision takes place. For example, in the Whittlesea (1993) experiment described earlier, subjects were asked whether the target word was pleasant or dull prior to judging whether they had seen it in the list. On trials when the target word had been shown in the preceding list, subjects were 12%more likely to call it pleasant than when it had not been shown (cf. the mere exposure effect: Zajonc, 1980). Similarly, when the ease of identifying the word was enhanced through a predictive context, subjects were 16%more likely to call it pleasant. Both of these effects are illusions: the words’ actual pleasantness was controlled through random assignment to conditions. However, they demonstrate that the same functions that produce feelings of familiarity, namely production followed by evaluation and attribution, also control feelings about the present quality of a stimulus: which the person experiences, a feeling of familiarity or of present quality, depends only on the purpose for which the processing is evaluated. A large number of judgments about the present quality of stimuli have been shown to be affected by a single exposure of an item earlier in the experiment, including brightness or darkness (Mandler, Nakamura, & van The production and evaluation of mental events is reminiscent of Anderson and Bower’s (1972) generate-recognize theory. However, the underlying assumptions about representation and processing are quite different in the two accounts. Anderson and Bower assumed that knowledge of concepts and events is preserved in an associative network, and that generation and recognition occurred through the spread of activation through the branches of that network. The current chapter disavows such a global organization of memory.
Constructive Processing in Remembering and Performance
241
Zandt, 1987), clarity (Whittlesea, Jacoby, & Girard, 1990), duration (Witherspoon & Allen, 1985),loudness (Jacoby, Allan, Collins, & Larwill, 1988), truth (Begg & Armour, 1991), pleasantness (Zajonc, 1980), understanding (Carroll & Masson, 1992), and knowing (Jacoby, Woloshyn, & Kelley, 1989). Whittlesea et al. (1990) demonstrated symmetrical illusions, such that judgments about past experience were influenced by the clarity of test presentations as well as prior presentation, and judgments of clarity were affected by prior presentation of test items as well as actual clarity. Which feeling the evaluation function will produce appears to depend only on what possible source of influence the subject is aware of Surprising fluency of production will be attributed to any likely source, causing a feeling of whatever that source represents. Although I lack direct evidence as yet, I suspect that the evaluation function is also an important basis of performance in tasks such as classification. In another implicit learning study, Whittlesea and Dorken (1993) showed subjects novel legal and illegal stimuli, but told subjects that half were items they had seen in training and half were new, without mentioning the existence of rules. Subjects were asked to pronounce each item, and then judge it as old or new. Because all items were novel, they could be judged as new or old only on the basis of a feeling of familiarity (rather than recall of context). Items possessing well-formed structure were (falsely) judged old on 63% of trials, whereas illegal items were judged old on only 8%of trials. We concluded that subjects were able to pronounce legal items more fluently than illegal items, because of their greater similarity to training items, and that this caused legal stimuli to feel more familiar. In a paired experiment, subjects were instead correctly informed that some test items conformed to the same rules as training items, whereas others were illegal. In this classification test, subjects judged legal items to be legal on 64% of trials, whereas illegal items were judged legal on only 27% of trials. We argued that the basis of this effect was similar to that of the recognition effect: subjects produced the names of legal test items with greater fluency than those of illegal items, attributed this greater fluency to legality, and experienced a conscious feeling of goodness. Under the demand to pronounce and classify the items, the mass of prior experience directly controls the integration of stimulus elements into a percept and the production of a corresponding sound: the fluency of this production can be attributed either to familiarity or to structural goodness, depending on the task. I therefore argue that remembering and performance tasks do not differ in fundamental process. It is not the case that remembering occurs through retrieval of a specific memory trace, whereas performance tasks are supported by application of general knowledge. Instead, both remembering and nonremembering performances are supported by the production of
248
Bruce W. A. Whittlesea
responses toward stimuli, including perceptual, cognitive, and motoric responses, driven by representations in memory, and by evaluation and attribution of the fluency of that production. One can use the fluency of producing a name for a stimulus object to decide whether the name is correct (“I think that spice is called coriander . . . yeah, that sounds right”), or the fluency of creating a perceptual organization of a stimulus to decide which category a stimulus belongs to (“That one’s certainly a wug: it’s so wuggish”), or the fluency of integrating ideas to decide whether some statement is true (“I’d guess that kangaroos are poor swimmers”), or the fluency of processing surface structure to decide that some deep concept is understood (“I had no trouble reading the textbook: I must have understood the material”). The only difference between performance in remembering and nonremembering tasks is the source to which fluent processing is attributed, under the control of the current context and task.
IX. Theme 6: The Constructive Nature of Experience A fundamental idea behind the SCAPE account is that consciousness is not in direct touch with either events of the past or even current objects. Instead, the contents of experience are constructed, built up through the interaction of representations of prior experiences with the current stimulus compound, but also interpreted within the current context. In the previous section, I demonstrated how the evaluative function of memory can produce feelings:feelings of familiarity, goodness, pleasantness, and so on. However, its role is greater than simply to produce feelings. It can actually interact with the production function, determining the contents of people’s consciousness. This is nicely illustrated by the “repetition blindness” phenomenon. Repetition blindness was first documented by Kanwisher (1987). She showed subjects sentences, presented one word at a time on a monitor, at a rate of 117 ms/word. Some sentences contained a repeated word, for example, When she spilled the ink there was ink all over. In other sentences, one occurrence of the repeated word was replaced by a synonym, for example, When she spilled the ink there was liquid all over. Kanwisher observed that subjects reported both target words at a high rate when they were different words, but that when a word was presented twice in a sentence, subjects often reported only the first occurrence, even though that rendered the second half of the sentence ungrammatical. To explain this finding, Kanwisher (1987) appealed to the typekoken distinction (e.g., Anderson & Bower, 1973). A “type” is the hypothetical representation of a general concept, a node in a semantic network. The
Constructive Promsing in Remembering and Performance
249
meaning of that type is given by the semantic linkages by which it is connected to other nodes in the network. The strength of these connections is a reflection of the frequency with which those ideas have been associated over the life-span of the individual. The type thus represents general knowledge, abstracted across many experiences of instances of the concept. In contrast, a token is a representation of a particular event in which the concept participated. The token links contextual information, such as time and place, to the relevant type. When a person encounters a stimulus, the corresponding type representation is automatically activated, allowing the person to know the identity of the stimulus. To remember that they had some particular experience of that stimulus, the person must additionally have created a token on the original occasion, and that token must be activated by the contextual properties at test. Kanwisher (1987; Park & Kanwisher, 1994) applied this dual-knowledge account to the repetition blindness effect. She reasoned that, in sentences not containing repetition, each target word activated its respective type, and also caused the creation of a token representing the Occurrence of that word, so that the subject could later report both events. The first occurrence of a repeated word also activated its type and created a token, allowing the subject to report the occurrence of that word later. In contrast, she argued, formation of a second token for a type is inhibited if it occurs too soon after the first. In consequence, in rapid list presentation, the second instance of a repeated word does not produce a token, so that the subject has no record of that event having taken place, thus explaining the selective “blindness” that she observed. The typekoken account of repetition blindness suggests a tight linkage between consciousness and memory. By that account, the formation of a memory trace is necessary and sufficient for reporting that one has just perceived a word. Similarly, the formation of two separate traces is necessary to report that the word occurred twice. Having two traces enables the person to remember the word twice, and thereby to know it was repeated in the list. Thus, if a word presented twice in a rapid list is reported only once, one can conclude that the person had formed only one representation of the occurrence of those words. Fundamentally, the typeltoken explanation treats reporting repetitions in rapid lists as a problem of detection, of registering the event at the time it occurs. Mental events occurring after the list are simply readouts of traces that were formed earlier. In contrast, Whittlesea, Dorken, and Podrouzek (1995) and Whittlesea and Podrouzek (1995) supplied a “constructive” explanation of the effect. We argued that failure to report both occurrences of a repeated word is not due to failure to encode the second presentation at the time it occurred. Both occurrences are encoded, and representations of both experiences
250
Bruce W. A. Whittlesea
drive that word to come to mind after the list. Instead, failure to report the word twice results from failure to interpret the memory-controlled production of that word as evidence that two occurrences had been encoded. By this account, the relationship between memory representations and mental events is less direct than that suggested by the typekoken account. When cued, memory can cause thoughts to come to mind, but these mental events are not necessarily isomorphic with the representations in memory, either in number or content. The occurrence or failure of a conscious mental event is thus not a direct indicator of the past, even of the immediate past. Instead, the person is in the position of trying to make decisions about the nature of their past experience, based on the evidence of what comes to mind now. That is, according to this account, the subject does not have direct access to memory contents, but instead must judge what they just experienced on the basis of their current performance.’ We argued that, during the presentation of a rapid list, each occurrence of a repeated word is encoded in much the same way as a nonrepeated word presented in the same list location. However, items in rapidly presented lists are not processed extensively; neither is the association between any item and its context (its adjacent words). In consequence, the representations of the two occurrences of a repeated word do not contain much distinctive information. Thus, even if both occurrences are encoded, and both later produce a mental event when the person attempts to recall the list, the mental events they produce will often not be distinctive enough for the person to interpret them as being produced by different sources. In consequence, the person is likely to decide that they saw that word only once, even though their recall of that word is supported by two separate representations. Moreover, the first occurrence of the repeated word, presented earlier in the list, is likely to be more extensively processed with its context. In consequence, people are more likely to be able to retrieve that context than the context of the second occurrence, and will therefore more likely report that word as having occurred in the first context than in the second. I also argue that the person is in that same position even when a stimulus is physically present. In looking at a word like business, I have the idea come to mind that it is the word that means “business.” I can then say that I know the meaning of that word. However, this claim to knowing is inferential, based on the ease with which that meaning comes to mind. This “knowledge” may be spurious: on seeing the word entomology, the meaning “study of the origins of words” may spring to my mind, driven by the phonological similarity of that word with eiymology, creating the false feeling that I know that word. The fact that the meaning of words usually comes to mind easily and accurately does not mean that one is in direct contact with the memory base: It simply means that masses of prior experience with that word cause its meaning to come to mind fluently when it is presented. Memory drives performance efficiently and generally accurately. However, any reflcction on the source of that performance is inferential.
Constructive Processing in Remembering and Performance
251
Whittlesea and Podrouzek (1995) demonstrated that reporting repetitions is based on the same attributional mechanism as familiarity. The procedure was almost identical to that of the experiment by Whittlesea (1993) reported in the previous section, in which the fluency of pronouncing a target word was sometimes enhanced through a predictive context: The only difference was that target words were presented in the list either once or twice, rather than presented or not presented. Repeated presentation of a word within the list enhanced the latency of pronouncing that word in the later test sentence; so did presentation of the target word in a predictive context. Subjects were 15% more likely to claim that a target word was repeated in the previous list if it was, showing some ability to perform the actual discrimination. However, the subjects were also highly influenced by the enhanced fluency of processing targets in predictive test contexts: On trials when the target word was shown only once, the subjects claimed to have seen it twice 19% more often when it was presented at test in a predictive context than in a nonpredictive context. This illusion of experiencing repetition demonstrates that the feeling of repetition is not directly produced by the possession of multiple traces, but is instead mediated by an attribution of the fluency of performance to a source that makes sense. This experiment, of course, does not directly explain the repetition blindness effect: It actually shows the opposite effect, of reporting nonexistent repetition. However, it demonstrates that the logic on which the typekoken explanation is based, that reports of occurrences of stimuli are directly related to encoded representations, is false. Whittlesea et al. (1995) investigated the formation of representations of repeated items. We presented sentences containing a repeated word, such as When she spilled the ink there was ink all over. We also presented two control conditions, each missing one word required by syntax and meaning: They differed in that the word was omitted from either an early location, called C1 (The red passed our car on the left) or a later location, called C2 (The boy hit the little in the playground). If the contents of consciousness truly reflect the contents of memory immediately after an event, then we should observe that subjects report nonrepeated words (the control conditions) accurately. Further, if memory does not encode a second occurrence of a word occurring soon after the first, we should not be able to predict reports of repeated words from performance in the control conditions, in which repetition-induced failure of encoding would not occur. First, we found that when subjects reported repeated words only once, they tended to report them in C1 rather than in C2,just as observed by Kanwisher (1987): words were reported only in C1 24% more often than only in C2. Such asymmetric reporting is the basis of Kanwisher’s claim that subjects selectively report the first occurrence of a repeated word.
252
Bruce W. A. Whittlesea
However, when we examined the control conditions, we observed the same phenomenon. When a word was omitted early in the sentence, subjects often erroneously reported the target that had been presented in C2 as being in C1 (producing The red car passed our on the left). However, they rarely made the reverse error of reporting a word presented early as being in the late location. Averaged across the two control conditions, subjects were 23% more likely to report a target word in C1 than in C2, similar to their early-report rate in the repetition condition. Clearly, in the control conditions, when there were two locations in which they could report a word, but only one word was actually presented that could fit those locations, the subjects had a bias to report that word in C1. That bias is easily understood. In reporting the sentence, they came to that location first: the meaning and syntax of the sentence demand a word at that point, they had available a word to fill it, and so they reported the word at that time. Coming to the second location, needing a word again, but having no other word come to mind and having no evidence that that word had been presented twice, they left that location blank. That is, the subjects were not able to simply read out the sentences from active memory representations: Instead, they constructed a remembrance of the sentence, based partly on what words came to mind, but also driven by their experience of meaning and standard syntax. We argued that that same constructive bias could explain the similar effect in reporting one occurrence of a repeated word. If the subject believed that that word had been presented only once, they would tend to report it in C1, for the same reasons they tend to report once-presented words there. The asymmetry in reporting repeated words only in C1 versus only in C2 is thus not indicative of which occurrence the subjects are reporting: They could be reporting the second occurrence in the first target location. We also suspected that at such fast presentation rates, processing of the second occurrence of a repeated word might be independent of processing the same word presented in C1, rather than being either facilitated or inhibited by the earlier experience. To test that hypothesis, we attempted to predict performance on trials containing a target word in both C1 and C2 from the trials containing a target only in C1 or only in C2. We knew that processing of words presented in C1 on nonrepeated trials was necessarily independent of processing words presented in C2, because they were different words on different trials. In the simulation, we combined the data of these conditions as though they were independent contributors to reporting a word presented twice on a single trial. For example, the word would be reported only in C1 either if only the first target was encoded, and was reported only in that location, or if only the second target was encoded, but was reported only in C1, or if both occurrences were encoded, but both
Constructive Processing in Remembering and Performance
253
were interpreted as a presentation only in the first location. We computed similar probabilities for report only in C2 and for report in both locations. The simulated probabilities matched the actual data for repeated trials very closely: average error was 4% in one study, and less than 1%in a second. We concluded that we could predict performance on repeated trials quite well from knowing performance on trials not containing a repetition: apparently, nothing happens on repeated trials that does not also happen on nonrepeated trials. We further concluded that there was no evidence of inhibition of encoding a second occurrence of a repeated word. Instead, the second occurrence appeared to be processed independent of the first. If there is no inhibition of encoding repeated presentations, why do people report both occurrences of repeated words less often than two nonrepeateed words presented in one sentence, as observed by Kanwisher (1987)? If the mental events that occur during the act of recall had unique correspondence to representations, as assumed by the typeltoken argument, then each representation of the repeated word would sponsor a different mental event, and the person could recall that they had experienced two occurrences. However, as I have tried to argue in this chapter, recall is not retrieval of a trace, but instead the production of a mental event, controlled by representations in memory, followed by evaluation of the source of that event. Both representations of a repeated word would cause that word to come to mind after the list. However, to interpret that coming-to-mind of the word as due to two prior occurrences, the word would have to come to mind as two different experiences, But words presented in high-speed lists are often not well integrated with their context. (It is this poor encoding of the contexts of words that permitted the subjects to report C2 presentations of nonrepeated words in Cl.) In consequence, the mental event produced by either of these representations is not much different to that which would be produced by the other: Acting simultaneously, they cause the person to remember the word, but to interpret and experience the comingto-mind of that word as being produced by a single prior event. In contrast, presentation of two different words in C1 and C2 does not present this problem. After the list, the two representations cause two different words to come to mind, so that the person can conclude they observed two occurrences. Whittlesea and Wai (in press) provided further evidence that the report of words in rapid sentences is based on construction of a remembrance rather than simply reporting encoded occurrences. In one study, we presented target words twice in a sentence, once in a syntactically and meaningfully congruent context, and once in an incongruent context. The incongruent occurrence always occurred first, for example, The diamond ring gift was a gift from his heart. In control sentences, the first occurrence of the
254
Bruce W. A. Whittlesea
repeated word was replaced by a completely different word, for example, The diamond ring sock was a gift from his heart. We observed the repetition blindness effect: the subjects reported both target nonrepeated words on 22% of trials, but both occurrences of repeated words on only 7% of trials. However, target words of both kinds were reported much more often in the second than in the first presentation location. Moreover, the deficit in reporting repeated words twice occurred selectively through subjects failing to report the word in the first location rather than in the second. This “reverse repetition blindness’’effect demonstrates that when people report a repeated word only once, the location of that report is controlled by their understanding of the meaning and syntax of the sentence: They reconstruct the sentence, using available evidence about what words occurred and how often, and where they make most sense. In another study, we provided congruent contexts for all occurrences of target words, whether repeated or not. The manipulation was of the distinctiveness of the contexts of first and second target words, for example, The teacher gave the test (quiz) today and the test was easy (less distinctive) versus He failed the history test (quiz) but the math test was easy (more distinctive). Using the less distinctive contexts, we observed the standard repetition blindness effect: Subjects reported both nonrepeated targets 25% more often than both occurrences of repeated targets, and selectively reported repeated targets in the first presentation location when they reported them only once. When contexts were made more distinctive, reports of the nonrepeated target in the second location increased by 2%. In contrast, reports of the repeated word in that location increased by 15%, reducing the repetition blindness effect by half, from 25% to 12%. We concluded that, in reconstructing the sentences, the subjects were more likely to interpret the fluent coming-to-mind of a word as indicating repeated experience when the syntax and meaning of the sentence more clearly supported a second occurrence of the word. These studies demonstrate that repetitiqn detection, like longer term remembering, occurs through the preservation of the repeated presentations, as they were processed at the time, followed by reconstruction of the experience, based on what comes to mind later, with what fluency, and in what context. In general, memory drives performance very efficiently, allowing people to deal with the outside world with robust accuracy. However, in making judgments about their own experience, people can only interpret their performance toward stimuli, under some theory about how they would behave in this context if their performance were controlled by this source or that; they have no direct or private access to the source of their behavior. Their beliefs about the nature of their experience are inevitably interpretations and attributions, not direct knowledge. In consequence, those beliefs
Constructive Processing in Remembering and Performance
255
are highly malleable, subject to modification through changes in the meaning or form of the test context. GENERAL SPECULATION The SCAPE theory, based on the selective production, evaluation, and preservation of specific experiences, can handle phenomena in a wide variety of areas, through a small set of assumptions. I suspect that it is capable of doing a good deal more work yet. For example, I suspect that it can explain the phenomena of attention, emotion, and motivation, without any increase in the fundamental assumptions. The constructs of memory and attention have been dealt with quite separately in the history of psychology, just as remembering and category learning have been. The topic of attention is the selective control of current processing, whereas memory has been treated as the residue of that experience, the storage of whatever was computed when attending a stimulus. Attention is the spotlight, selector, or gateway to mind: memory is the set of knowledge resources that can be drawn on to process selected aspects of the stimulus environment. I recently witnessed a compelling demonstration, given by Cavanaugh (1995). Cavanaugh pointed out that one of the basic phenomena of attention is to dilate time. To illustrate this, he showed a blinking blue dot, projected on a wall, turned on and off at a stimulus onset asynchrony of about half a second. After some cycles, the dot unexpectedly expanded until it covered the entire wall. Subjectively, this presentation appeared to be much longer than the other presentations, although it was actually of the same duration. Cavanaugh explained that movement captures attention, and attention dilates subjective time. Such demonstrations seem to call for an explanation of attention as a function of mind separate from memory. On returning from the conference, I replicated Cavanaugh’s demonstration on my laptop. However, in addition, I programmed it in reverse: a dot expanded for a number of cycles, and was then presented as a static display. The reverse phenomenon occurred: the static dot now appeared to endure longer. Moreover, on repeating either of the demonstrations, the subjective dilation of time no longer occurred. This suggested to me that motion is not the key component in the phenomenon of dilated subjective time, and that it is unnecessary to appeal to an extra set of attentional principles to explain it. Instead, I interpret the phenomenon as one of learning and evaluation. Earlier, I described a study by Whittlesea and Williams (1996), in which subjects were shown words (e.g., RAINBOW), as well as easy (HENSION) and hard (STOFFWUS) nonwords. In that study, we observed large false-
256
Bruce W. A. Whittlesea
familiarity effects for easy nonwords, which we interpreted to mean that subjects were surprised by their fluency of processing those items, and, in the context of a remembering task, attributed that unexpected fluency to familiarity. In another study, we switched to examining duration judgments. There was no training phase: but the three kinds of stimuli were presented for either 100 or 200 ms, and postmasked. Subjects were asked to pronounce each item, and then say whether it had been presented for a short or a long duration. Long presentations were judged to be long about 20% more often than short presentations, regardless of stimulus type. However, in addition, short presentations of words and easy nonwords were judged long 15% more often than short presentations of hard nonwords; that is, the subjects suffered an illusion of increased duration with the more fluently processed stimuli. I do not think that one would want to try to explain this finding in terms of differential attention to the stimuli. Instead, we concluded that the subjective experience of duration is the product of an attribution about the fluency of processing, an inferential process that attributes surprising fluency to a source suggested by the test context. I suggest the same basis for Cavanaugh’s (1995) demonstration. The first time a subject watches it, they learn from the induction series of static dots to expect further static dots. Moreover, because the dot is static, it can be completely processed through the production of a single percept. The expanding dot is unexpected. Moreover, because it keeps changing, the subject keeps processing it; and because it expands smoothly, that processing is fluent. This surprising fluency is attributed within the interpretive context to increased duration, rather than to other possible sources, such as familiarity. On a second pass through the demonstration, the subject knows to expect an expanding dot. No longer unexpected, the processing of that stimulus is no longer experienced as taking more time. I am not trying to downplay the interest of attentional phenomena. Instead, I am trying to integrate the issues of that area with our growing understanding of memory as an active agent in stimulus processing. I suspect that most, if not all, of the phenomena of selective and divided processing can be understood through the constructs of learning and evaluation, without appealing to a separate set of attentional principles. The phenomenon of negative priming (Tipper, 1985) provides a good example. In that paradigm, a person ignores one stimulus while attending another: for a short while afterward, the ignored stimulus is harder to identify than is a novel stimulus. This phenomenon initially seemed to require an explanation in terms of inhibitory processing of the ignored stimulus (e.g., Neuman & DeSchepper, 1992), an attentional control process quite unlike learning processes. Like abstraction of general structure, inhibitory processing is a
Constructive Processing in Remembering and Performance
257
conceivable mental activity: but, to avoid the unnecessary proliferation of entities, before accepting inhibitory processing as real, we should investigate whether negative priming might not be the result of some other kind of process. The inhibitory-processing explanation rests on the assumption that knowledge about familiar words and objects resides in generic representations in a semantic memory system, representations that have been formed over many prior experiences, and that are little changed by one further experience. The presentation of a word in a priming trial is thus thought of as simply activating the lexical representation of that word, making its identity available. The same logic dictates that if the identity of the word is not available after the word is presented but actively ignored, then the lexical representation must have been inhibited. However, according to the SCAPE account, there is no generic representation of a word to be activated or inhibited: knowledge of a word is distributed across hundreds or thousands of specific prior experiences. Further, those separate experiences are not uniformly cued in a further encounter with a word. Instead, as demonstrated under Theme 4, specifically similar experiences of a word can outweigh the mass of prior experiences. Moreover, every experience of a stimulus is a learning event, in which the specific activities performed on a stimulus in a specific context are recorded. This provides an alternate interpretation of the negative priming phenomenon, that it is a learning effect, as suggested by Williams (1996).6 According to this alternate approach, the priming presentation, in which one item is attended and another ignored, is a learning trial, on which the subject learns to produce the identity of one stimulus and not to produce the identity of the other. The probe presentation is a transfer trial. As on any other occasion of encountering a stimulus, memory traces are cued by context, and impose operations on the current stimulus. In this case, one of the prior experiences that is very likely to be cued is the preceding prime trial, on which the person learned not to name that stimulus. That processing is recapitulated in the probe trial, resulting in difficulty in producing the name of the stimulus. Negative priming can thus be understood as a learning phenomenon, a side effect of constructing and preserving particular experiences, rather than as the product of separate attentional processes. In general, I think that the study of cognitive psychology is suffering from misapplied notions of modularity. Clearly, mind can be thought of a serving a number of separate functions, including perception, attention, learning (about both events and concepts), storage, remembering (short Neil1 and Valdes (1992) also describe a learning account of negative priming. However, their account includes the construct of inhibitory processing. Williams’ (19%) account explains the effect simply as an example of transfer-inappropriate processing.
258
Bruce W. A. Whittlesea
and long term), reasoning, and awareness. To a great extent, psychologists have elected to study each of these functions in isolation from the others. In doing so, we have also drifted into thinking about them as existing in self-contained units, trading products with each other, but performing qualitatively different activities through qualitatively different mechanisms and by different principles. This drive toward modularity represents good engineering science, as practiced by Western engineers: each part of the machine performs one and only one function, its unique contribution to the overall function of the system. Modularity also makes systems more understandable. If the perceptual system is independent of the remembering module, then they can be studied independently, without worrying that one’s conclusions about remembering are true only in some specific perceptual modality. The drive toward modularity has been fueled by the observation of dissociations between tasks. The idea is that if two tasks are dependent on the same memory representation, then performance in one should be predictable from performance in the other. In consequence, dissociations between tasks can be taken as evidence that they rely on different kinds of representation. For example, Tulving et al. (1982) observed that subjects’ performance in an indirect test of memory, such as completion of word fragments, cannot be predicted from their behavior in a direct test, such as recognition, that is performed on the same items. This approach has produced the strongest reason to think of memory as consisting of separate systems, operating by different principles on different aspects of learning and performance (e.g., Schacter, 1987; Tulving, 1983,1985). Our neuropsychological colleagues have eagerly assisted in this enterprise, supplying physiological correlates to back up the performance dissociations observed by cognitive psychologists (e.g., Knowlton & Squire, 1994; Squire, 1992; Weiskrantz, 1987). The problem is that there are too many dissociations. For example, Jacoby and Dallas (1981) also observed a dissociation between direct and indirect tests of memory, but their indirect test consisted of tachistoscopic identification rather than fragment completion. Tulving et al. (1982) explained that dissociation in the same way, concluding that recognition is served by one mechanism (episodic memory), and fragment completion and tachistoscopic identification by another (later, two others: the semantic and perceptual representation systems). Then Witherspoon and Moscovitch (1989) showed a dissociation between tachistoscopic identification and fragment completion of the same stimuli. They pointed out that Tulving’s dissociation logic requires that flashed versus incomplete presentations of stimuli are processed by different memory subsystems. Hayman and Tulving (1989) even found a dissociation between two successive fragment comple-
Constructive Processing in Remembering and Performance
259
tion tests on the same items if the subjects were shown different fragments of each item in the two tests. Tulving and his colleagues currently argue for 5 systems and 12 subsystems within memory, each performing specific, limited functions (e.g., Schacter & Tulving, 1994; Tulving, 1995). There is no obvious end to the number of dissociations and consequent splitting of systems into subsystems, each new pair smaller and more limited in function than the one they replace. Conceptualizing memory as composed of separate functional systems thus seems to be of diminishing utility. Instead, I am impressed by Jacoby’s (e.g., 1983; Jacoby & Witherspoon, 1982) and Roediger’s (e.g., Roediger & Challis, 1992) interpretation of the array of observed dissociations. Each task that a person can be set requires some specific resources. Perceptual tasks require prior experience of the perceptual aspects of a stimulus;conceptual tasks require conceptual experience; selective remembering tasks require the encoding of distinctive information about particular events. Any specific experience of a stimulus will involve some combination of these types of information. Any later experience of that same stimulus will benefit to differing degrees from the earlier experience, depending on the type of later task and the resources it requires, in interaction with the resources made available by the specific nature of the earlier experience. This account suggests that the number of observable dissociations is limited only by the imagination of experimenters to contrive combinations of tasks. I agree that the modular approach is a practical tool for splitting off manageable hunks of the psychological system for study. However, I argue that the various mental functions that psychologists have identified are not real, separate, and endemic properties of mind, but rather convenient categories that we have imposed on mind in trying to understand it. As demonstrated by the discussions earlier, I am attempting to take the opposite approach, treating mind as a fundamentally unitary whole. Perhaps the most important difference between my approach and that driving the development of separate-systems accounts is our assumptions about the directness of the relationship between a cognitive function and the mechanism that serves that function. The idea of separate systems is founded on the assumption that each function is directly served by a mechanism dedicated to that function. Thus, if people are sensitive to abstract prototypes, they must have a prototype-abstraction mechanism; sensitivity to the frequency of words must mean they have a frequency counter (the strength of a logogen); failure to report second occurrences of rapidly presented stimuli must mean there is an inhibitory mechanism suppressing encoding of repetitions. In contrast, the SCAPE theory suggests that mind performs most of its functions indirectly, almost by accident: All that it ever really does is construct an experience of a stimulus, under the interact-
260
Bruce W. A. Whittlesea
ive control of the current stimulus complex and the mass of prior specific experiences, and preserve the new experience. The various sensitivities of the system to abstract properties of experience, although extremely valuable to the person, are side effects, produced by the distributions of properties across prior experiences, and the cue properties of the current situation. The system can perform the function of identifying a stimulus, producing its name: but it does not do that through an identification mechanism. Instead, the stimulus complex, including the demand to name, cues prior experiences of producing names for objects that look similar to this object, directing the system to impose the same processing on the current stimulus. It can also perform the function of remembering, although it has no retrieval mechanism. Instead, given the task of remembering as part of the stimulus compound, it constructs images of contexts. It can also experience familiarity, by interpreting fluent processing as due to a source in the past; but the same mechanism serves the functions of classification, judgments of temporal duration, and feelings of pleasantness, depending only on the task context in which the fluency is experienced. Memory is fundamentally very simple. Human performance derives its complexity not from the architecture or processing of memory but from the variety tasks, stimulus structures, and contexts to which memory is exposed. REFERENCES Anderson, J. R. (1980). Cognitive psychology and its implications. San Francisco: Freeman. Anderson, J. R., & Bower, G. H. (1972). Recognition and retrieval processes in free recall. Psychological Review, 79, 97-123. Anderson, J. R., & Bower, G. H. (1973). Human associative memory. Washington, DC: Winston. Andrews, S. (1992). Frequency and neighborhood effects on lexical access: Lexical similarity or orthographic redundancy. Journal of Experimental Psychology: Learning, Memory, and Cognition, 18, 234-254. Bartlett, F. C. (1932). Remembering: A study in experimentalandsocialpsychology.Cambridge, UK: Cambridge University Press. Begg, I., & Armour, V. (1991). Repetition and the ring of truth: Biasing comments. Canadian Journal of Behavioral Science, 23, 195-213. Brooks, L. R. (1978). Non-analytic concept formation and memory for instances. In E. H. Rosch & B. B. Lloyd (Eds.), Cognition and categorization (pp. 169-211). Hillsdale, NJ: Erlbaum. Brooks, L. R. (1987). Decentralized control of categorization: The role of prior processing episodes. In U. Neisser (Ed.), Concepts and conceptual development: Ecological and intellectual factors in categorization (pp. 141-174). Cambridge, UK: Cambridge University Press. Brooks, L. R., & Vokey, J. R. (1991). Abstract analogies and abstracted grammars: Comments on Reber (1989) and Mathews et al. (1989). Journal of Experimental Psychology: General, 120, 316-323.
Constructive Processing in Remembering and Performance
261
Carroll, M., & Masson, M. E. J. (1992). Reading fluency as a basis for judgments of text comprehension: Misattributions of the effects of past experience. Unpublished manuscript. Cavanaugh, P. (1995, May). Attention-based visual processes. Paper presented at the Banff annual seminar in Cognitive Science, Banff, Alberta Canada. Cleeremans, A. (1993). Mechanisms of implicit learning: Connectionist models of sequence processing. Cambridge, MA: MIT Press. Craik, F. I. M., & Lockhart, R. S. (1972). Levels of processing: A framework for memory research. Journal of Verbal Learning and Verbal Behavior, 11, 671-684. Dulany, D. E., Carlson, R. A,, & Dewey, G. I. (1984). A case of syntactical learning and judgment: How conscious and how abstract? Journal of Experimental Psychology: General, 113, 541-555. Forster, K. I., & Davis, C. (1984). Repetition priming and frequency attenuation in lexical access. Journal of Experimental Psychology: Learning, Memory, and Cognition, 10, 680-698. Franks, J. J.. & Bransford, J. D. (1971). Abstraction of visual patterns. Journal of Experimental P.yychology, 90, 65-14. Glushko, R. J. (1979). The organization and activation of orthographic knowledge in reading aloud. Journal of Experimental Psychology: Human Perception and Performance, 5, 674-691. Gordon, P. C., & Holyoak, K. J. (1983). Implicit learning and generalization of the “mere exposure” effect. Journal of Personality and Social Psychology, 45, 492-500. Hayman, C. A. G., & Tulving, E. (1989). Is priming in fragment completion based on a “traceless” memory system? Journal of Experimental Psychology: Learning, Memory, and Cognition, 15, 941-956. Homa, D., Sterling, S., & Trepel, L. (1981). Limitations of exemplar-based generalization and the abstraction of categorical information. Journal of Experimental Psychology: Human Learning and Memory, 2, 322-330. Jacoby, L. L. (1983). Remembering the data: Analyzing interactive processes in reading. Journal of Verbal Learning and Verbal Behavior, 22, 485-508. Jacoby. L. L.. Allan. L. G., Collins, J. C., & Larwill, L. K. (1988). Memory influences subjective experience: Noise judgments. Journal of Experimental Psychology: Learning, Memory, and Cognition, 14, 240-247. Jacoby, L. L., & Brooks, L. R. (1984). Nonanalytic cognition: Memory, perception and concept formation. Psychology of Learning and Motivation, 18, 1-47. Jacoby, L. L.. & Dallas, M. (1981). On the relationship between autobiographical memory and perceptual learning. Journal of Experimental Psychology: General, 110, 306-340. Jacoby, L. L., Kelley, C . M., & Dywan, J. (1989). Memory attributions. In H. L. Roediger & F. I. M. Craik (Eds.), Varieties of memory and consciousness: Essays in honor of Endel Tulving (pp. 391-422). Hillsdale, NJ: Erlbaum. Jacoby, L. L., Witherspoon, D. E. (1982). Remembering without awareness. Canadian Journal of Psychology, 36, 300-324. Jacoby, L. L.. Woloshyn, V., & Kelley, C. M. (1989). Becoming famous without being recognized: Unconscious influences of memory produced by dividing attention. Journal of Experimental Psychology: General, 118, 115-125. Kanwisher, N. G. (1987). Repetition blindness: Type recognition without token individuation. Cognition, 27, 117-143. Knowlton, B. J., & Squire, L. R. (1994). The information acquired during artificial grammar learning. Journal of Experimental Psychology: Learning, Memory, and Cognition, 20, 79-91. Kolers, P. A. (1973). Remembering operations. Memory & Cognition, 12, 347-355.
262
Bruce W.A. Whittlesea
Kolers, P. A., & Smythe, W. E. (1984). Symbol manipulation: Alternatives to the computational view. Journal of Verbal Learning and Verbal Behavior, 21, 289-314. Lindsay, D. S., & Read, J. D. (1994). Psychotherapy and memory for childhood sexual abuse: A cognitive perspective. Applied Cognitive Psychology, 8, 28 1-337. Mandler, G. (1980). Recognizing: The judgement of previous occurrence. Psychological Review, 87, 252-271. Mandler, G., Nakamura, Y.,& van Zandt, B. J. S. (1987). Nonspecific effects of exposure on stimuli that cannot be recognized. Journal of Experimental Psychology: Learning, Memory, and Cognition, 13, 646-648. Masson, 1. E. J., & MacLeod, C. M. (1992). Re-enacting the route to interpretation: Enhanced perceptual identification without prior perception. Journal of Experimental Psychology: General, 121, 145-176. Mathews, R. C., Buss, R. R., Stanley, W. B., Blanchard-Fields, F., Cho, J. R., & Druhan, B. (1989). Role of implicit and explicit processes in learning from examples: A synergistic effect. Journal of Experimental Psychology: Learning, Memory, and Cognition, 15, 10831100. Mathews, R. C., & Roussel, L. G. (1993). Automatic abstraction of stimulus structure from episodes: Comment on Whittlesea and Dorken (1993, pp. 397-400). McClelland, J. L., & Rumelhart, D. E. (1981). An interactive-activation model of context effects in letter perception: Part 1. An account of basic findings. Psychological Review, 88, 375-407. Medin, D. L., & Schaffer, M. M. (1978). Context theory of classification learning. Psychological Review, 85, 207-238. Meyer, D. E., & Schvaneveldt, R. W. (1971). Facilitation in recognizing pairs of words: Evidence of a dependence between retrieval operations. Journalof Experimental Psychology, 90, 227-234. Morris, C. D., Bransford, J. D., & Franks, J. J. (1977). Levels of processing versus transferappropriate processing. Journal of Verbal Learning and Verbal Behavior, 16, 5 19-533. Neill, T., & Valdes, L. A. (1992). Persistence of negative priming: Steady state or decay? Journal of Experimental Psychology: Learning, Memory, and Cognition, 18,565-576. Neuman, E., & DeSchepper, B. G. (1992). An inhibition based fan effect: Evidence for an active suppression mechanism in selective attention. Canadian Journal of Psychology, 46, 1-40. Neumann, P. G. (1974). An attribute frequency model for the abstraction of prototypes. Memory & Cognition, 2, 241-248. Paap. K. R., & Noel, R. W. (1991). Dual-route models of print to sound: Still a good horserace. Psychological Research, 53, 13-24. Park, J., & Kanwisher, N. (1994). Determinants of repetition blindness. Journalof Experimental Psychology: Human Perception and Performance, 20, 500-519. Perruchet, P., & Pacteau, C. (1991). Synthetic grammar learning: Implicit rule abstraction or explicit fragmentary knowledge? Journal of Experimental Psychology: General, I 19, 264-275. Posner, M. I., & Keele, S. W. (1968). On the genesis of abstract ideas. Journal of Experimental PSycholOgy, 77, 353-363. Reber, A. S. (1969). Transfer of syntactic structure in synthetic languages. Journal of Experimental Psychology, 77, 115-119. Reber, A. S. (1989). Implicit learning and tacit knowledge. Journalof Experimental Psychology: General, 118, 219-235. Reber, A. S. (1993). Implicit learning and tacit knowledge: A n essay on the cognitive rtnconscioirs. New York: Oxford University Press.
Constructive Processing in Remembering and Performance
263
Reber, A. S., & Allen, R. (1978). Analogic and abstraction strategies in synthetic grammar learning: A functionalist interpretation. Cognition, 6, 193-221. Roediger, H. L. (1990). Implicit memory: A commentary. Bulletin of the Psychonomic Society, 28, 373-380. Roediger, H. L., & Challis, B. H. (1992). Effects of exact repetition and conceptual repetition on free recall and primed word fragment completion. Journal of Experimental Psychology: Learning, Memory, and Cognition, 18, 3-14. Roediger, H. L., & 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. H. (1977). Human categorization. In N. Warren (Ed.), Advances in crosscultural psychology (Vol. 1, pp. 1-49). London: Academic Press. Rosch, E. H. (1978). Principles of categorization. In E. H. Rosch & B. B. Lloyd (Eds.), Cognition and categorization (pp. 27-48). Hillsdale, NJ: Erlbaum. Rosch, E. H., & Mervis, C. B. (1975). Family resemblances: Studies in the internal structure of categories. Cognitive Psychology, 7, 573-605. Rosch, E. H., Simpson, C., & Miller, R. S. (1976). Structural bases of typicality. Journal of Experimental Psychology: Human Perception and Performance, 2, 491-502. Schacter, D. L. (1987). Implicit memory: History and current status. Journal of Experimental Psychology: Learning, Memory, and Cognition, 13, 501-518. Schacter, D. L., & Tulving, E. (1994). What are the memory systems of 1994? In D. L. Schacter and E. Tulving (Eds.), Memory systems 1994. Cambridge, MA: MIT Press, pp. 1-24. Servan-Schreiber, E., & Anderson, J. R. (1990). Learning artificial grammars with competitive chunking. Journal of Experimental Psychology: Learning, Memory, and Cognition, 16, 592-608. Squire, L. R. (1992). Declarative and nondeclarative memory: Multiple brain systems supporting learning and memory. Journal of Cognitive Neuroscience, 4, 232-243. Tipper, S. (1985). The negative priming effect: Inhibitory priming by ignored objects. Quarterly Journal of Experimental Psychology, 37A, 571-590. Tulving, E. (1983). Elements of episodic memory. Oxford, England: Clarendon Press. Tulving E. (1985). How many memory systems are there? American Psychologist, 40,385-398. Tulving, E. (1995). Organization of memory: Quo vadis? In M. Gazzaniga (Ed.), The cognitive neurosciences. Cambridge, MA: MIT Press, pp. 839-847. Tulving, E., Schacter, D. L., & Stark, H. A. (1982). Priming effects in word-fragment completion are independent of recognition memory. Journal of Experimental Psychology: Learning, Memory, and Cognition, 8, 336-342. Tulving, E., & Thompson, D. M. (1973). Encoding specificity and retrieval processes in episodic memory. Psychological Review, 80, 352-373. Vokey, R. R., & Brooks, L. R. (1992). The salience of item knowledge in learning artificial grammars. Journal of Experimental Psychology: Learning, Memory, and Cognition, 18, 328-344. Weiskrantz, L. (1987). Neuroanatomy of memory and amnesia: A case for multiple memory systems. Human Neurobiology, 6, 93-105. Wheeler, D. D. (1970). Processes in word recognition. Cognitive Psychology, I, 59-85. Whittlesea, B. W. A. (1987). Preservation of specific experiences in the representation of general knowledge. Journal of Experimental Psychology: Learning, Memory, and Cognition, 13, 3-17. Whittlesea. B. W. A. (1993). Illusions of familiarity. Journal of Experimental Psychology: Learning, Memory, and Cognition, 19, 1235-1253.
264
Bruce W. A. Whittlesea
Whittlesea, B. W. A,, & Brooks, L. R. (1988). Critical influence of particular experiences in the perception of letters, words, and phrases. Memory & Cognition, 16, 387-399. Whittlesea, B. W. A., Brooks, L. R., & Westcott, C. (1994). After the learning is over: Factors controlling the selective application of general and particular knowledge. Journal of Experimental Psychology: Learning, Memory, and Cognition, 20, 259-274. Whittlesea, B. W. A., & Dorken, M. D. (1993). Incidentally, things in general are particularly determined: An episodic-processing account of implicit learning. Journal of Experimental Psychology: General, 122, 227-248. Whittlesea, B. W. A,, Dorken, M. D., & Podrouzek, K. W. (1995). Repeated events in rapid lists, Part 1: Encoding and representation. Journal of Experimental Psychology: Learning, Memory, and Cognition, 21, 1670-1688. Whittlesea, B. W. A., Jacoby, L. L., & Girard, K. (1990). Illusions of immediate memory: Evidence of an attributional basis for feelings of familiarity and perceptual quality. Journal of Memory and Language, 29, 716-732. Whittlesea, B. W. A,, & Podrouzek, K. W. (1995). Repeated events in rapid lists, Part 2: Remembering repetitions. Journal of Experimental Psychology: Learning, Memory, and Cognition, 21, 1689-1687. Whittlesea, B. W. A., & Wai, K. H. (1996). Reverse repetition blindness and release from repetition blindness: Constructive variations on the repetition blindness effect. In press; Psychological Research. Whittlesea, B. W. A., & Williams, L. D. (1996). Surprising illusions of familiarity. Manuscript submitted for publication. Whittlesea, B. W. A., & Wright, R. (1997). Implicit (and explicit) learning: Acting adaptively without knowing the consequences. Journal of Experimental Psychology: Learning, Mernory, and Cognition, 23, 1-20. Williams, L. D. (1996). Process transfer: A reinterpretation of the negative priming effect. Unpublished manuscript. Witherspoon, D., & Allan, L. G. (1985). The effects of a prior presentation on temporal judgments in a perceptual identification task. Memory & Cognition, 13, 101-111. Witherspoon, D., & Moscovitch, M. (1989). Stochastic independence between two implicit memory tasks. Journal of Experimental Psychology: Learning, Memory, and Cognition, 15, 22-30. Wittgenstein, L. (1953). Philosophical investigations. New York: Macmillan. Zajonc, R. B. (1980). Feeling and thinking: Preferences need no inferences. American Psychologist, 35, 151-175.
GOALS, REPRESENTATIONS, AND STRATEGIES IN A CONCEPT ATTAINMENT TASK The EPAM Model Fernand Gobet Howard Richman Jim Staszewski Herbert A. Simon
Any explanation of human performance in cognitive tasks must take account of a broad range of factors. In memory or concept learning tasks these begin, on the stimulus side, with the structure of the task itself and the knowledge that is embedded in it, then move to the subject side-to the ways in which subjects interpret the task (their representation of the goal and of the task structure)-thence to the strategies they adopt for attacking it, and finally to the capabilities, some of them “built in,” some of them acquired through previous learning, that they apply to it (Newel1& Simon, 1972, ch. 14).
I. Inter-Subject Differences and Commonalities in Performing Cognitive Tasks Subjects who represent the same task in an experiment differently, employ different strategies in approaching it, or have different capabilities in the form of memory and processing capacities and relevant knowledge can be expected to exhibit different behavior while performing it. If the differences THE PSYCHOLOGY OF LEARNING AND MOTIVATION, VOL. 37
265
Copyright 0 1997 by Academic Press. All rights of reproduction in any form reserved. M)79-7421/’)7 $25.00
266
Fernand Gobet et al.
among subjects are small, we may be satisfied in averaging over our sample in order to provide a general, and approximate, explanation of the behavior of “typical subjects.’’ If the differences are large, averaging over subjects will provide only a rough and fuzzy picture of the behavior and a correspondingly incomplete account of it. Even if we are primarily interested in those mechanisms that are largely shared among subjects, we are more likely to get a clear view of them if we strip off the differences before trying to discern the similarities. One approach to sorting out the individual differences and commonalities we can expect to encounter among subjects in a particular task is to undertake deliberately to magnify the differences by inducing different groups of subjects to adopt different problem representations or different strategies, or by selecting groups of subjects who can be expected to possess quite different bodies of knowledge about the task. In this way, the remaining human commonalities are revealed as the residual, so to speak, that is shared by the different experimental conditions. At least three significant lines of experimental work in recent decades have pursued this strategy: one line focusing on the knowledge subjects have (comparing the performance of experts with that of novices) (e.g., Chase & Simon, 1973; de Groot, 1946; Ericsson & Staszewski, 1987); a second focusing on differences in problem representation, induced by presenting the task through distinct, but isomorphic, instructions (e.g., Hayes & Simon, 1974); a third focusing on differences in strategy, induced by including strategy recommendations as part of the problem instructions (e.g., Medin & Smith, 1981). In this chapter, we discuss this third line of work, taking as an example for analysis the paper by Medin and Smith (1981) on concept attainment, and using the EPAM (Elementary Perceiver and Memorizer) theory, a familiar computer simulation model that has been used to simulate a wide and growing range of memory tasks. We describe how EPAM can model the different strategies subjects employed-differences induced by the three sets of task instructions. In doing this, we show how a formal simulation model can embody not only mechanisms that explain human commonalities (“invariant psychological laws”), but also the mechanisms that interpret different representations and different strategies for the same problem. The model thereby provides a way of unifying the theory of the behaviors of individuals who approach a task with wide differences in knowledge, previous experience, and task interpretation. Changes in behavior, including those we describe as “learning” may result from changes in any one or more of these components. Differences among subjects in the representations and strategies they employ may
The EPAM Model
261
reflect differences in the knowledge they bring to the task, or differences in what they learn from feedback provided by or during the task performance. NEEDS,GOALS,TASKREPRESENTATIONS, STRATEGIES
When attention has been turned to a particular need or task, one or more goals may be evoked from memory that, if attained, would meet the need or complete the task. With the goal in place, a subgoal may then be set of representing the task in such a way that the cognitive processes can go to work on it. The task representation is usually called the problem space, and task activity can be thought of as a search through the problem space that, if successful, attains the goal. As it is seldom either possible or efficient to search the problem space in a routine, systematic way, the system must also generate a strategy to guide the search (Newel1 & Simon, 1972,pp. 788789). We can view the process as involving the sequence: need + attention + goal + problem space .--) strategy + search
We should not be misled, however, into thinking that problem-solving efforts typically follow these steps in an inexorable linear order. “Linear thinking” (often contrasted with “creative thinking”) is a much maligned phrase that is irrelevant to the picture we have just drawn. The process does not usually advance without many detours and retreats. To the diagram above, we must add numerous feedback loops. Ideas evoked while designing the problem space may lead back to reformulation of the goal; strategies often show the problem space to be incomplete or inappropriate and cause a return to restructuring it. The system that behaves in this way therefore requires a metastrategy to monitor its progress and decide when it should reconsider earlier steps. Efforts toward structuring the problem (creating and modifying the problem space) usually predominate during the early stages of activity, and efforts toward searching a particular problem space predominate during the later stages, but with much intermingling of all the processes, particularly when a problem presents difficulty or novelty. In a difficult problem like the Mutilated Checkerboard Problem, which requires discovery of a nonobvious problem space for solution, subjects, who initially adopt an “obvious” problem space, generally return to searching for a new problem representation after an hour or more of unsuccessful and ultimately frustrating search in the “obvious” but inappropriate space; and if they then discover a more appropriate representation, solve the problem after one or two minutes of search in the new space (Kaplan & Simon, 1990). Failure-produced frustra-
Fernand Gobet et al.
268
tion followed by sudden success is characteristic of so-called insight problems. Of course, failure is not always followed by success, nor does frustration always evoke a new problem space. Prolonged failure may simply lead to a brief, lengthy, or permanent abandonment of the task. Switches from one task to another are mediated by the whole structure of emotions, motives, and external demands, which must be included in any comprehensive model of the system and its behavior. For example, Johannes Kepler, in one of his early works, announced a law stating that the periods of revolution of the planets increased with the square of their distances from the Sun; but, after about a decade, decided that the fit of the law to the data was unsatisfactory. Resuming his search, in about a month’s time he found the law that we now regard as correct (Kepler’s Third Law: the period varies as the 3/2 power of the distance). We know also that during the intervening decade, Kepler’s attention was distracted by other pressing matters-not the least that his mother was being tried for witchcraft! He resumed his search shortly after she was found innocent. We have laboratory notebooks for extended periods of a few scientists (e.g., Darwin, Faraday, Krebs, but not Kepler) that cast light on these attention-control processes, at least on a coarse time scale.
11. Architecture and Learning in Task Performance
In accounting for behavior in cognitive tasks, we need to deal with several channels of causation: the influences of the task domain, the problem space, and the subjects’ strategies.
A. THETASKDOMAIN First, the behavior will depend on the characteristics of the task domain (Newel1 & Simon, 1972). In the case of very simple tasks, the goal and task domain essentially determine the behavior: the actor takes the “obvious” action that achieves the goal. If we know that someone’s goal is to reserve time on a parking meter, we readily predict that he or she will insert one or more quarters in the slot. Notice that, even in this case, we must make a number of implicit assumptions: that the actor is familiar with parking meters, knows what denomination of coin is needed, knows where the slot is and how to insert the coin, or can read the appropriate instructions on the meter to obtain this information. Notice also that our prediction of the behavior depends on
The EPAM Model
269
our own access to the knowledge we assume the actor is using and our ability to apply it appropriately: in particular, that we use the same “obvious” representation for the problem that the actor uses. The use of these assumptions to make predictions of behavior is what Newel1 (1990) has called “prediction at the knowledge level.” A great many, perhaps most, of our predictions of behavior in everyday life are made in this way, that is, by emulation; and we are often quite unaware of the assumptions we are making about the actor’s knowledge and ability (or inability) to reason. B. REPRESENTING THE TASK:THE PROBLEM SPACE Second, in all cases beyond those in which the behavior can be predicted at the knowledge level, the behavior will depend on the way in which the actor represents the task: the problem space. Generating an appropriate problem space for a task may vary in difficulty from the trivial (as in the previous example) to the essentially impossible (a problem space for inferring Kepler’s laws of planetary motion from gravitational attraction before the invention of the calculus). In the earlier example of the mutilated checkerboard problem, we have already shown how critical the selection of problem space can be in determining the subsequent behavior-and its failure or success. C. STRATEGIES
Third, because the problem space prescribes only the representation of the task and not the precise way in which it will be attacked, the actor’s behavior will depend on the search strategies that are adopted for exploring the space for a solution. It is almost always inefficient, and usually infeasible, for people to search a problem space either exhaustively or randomly. Various procedures are adopted to guide the search in productive directions, so that the task can be accomplished after a very small part of the total space has been examined. In favorable circumstances (e.g., solving a linear equation in algebra), the actor may know a systematic procedure (algorithm) that is guaranteed to lead to the solution after a few steps. In most less formal domains, no such algorithms exist, and problem solvers must be satisfied with selective rules of thumb (heuristics) that often lead to solutions without excessive search, but are not guaranteed to do so. Interwoven with the goals, the problem space, and the strategies is the knowledge the actor possesses that is relevant to understanding and formulating these elements. Some of this knowledge will be obtained from the task instructions (thereby involving learning processes); much of it will already be stored in memory (as the result of previous learning).
270
Fernand Gobet et el.
D. ALLBEHAVIOR Is SOCIAL It follows from these characteristics of the situation that in the laboratory for cognitive psychology we are studying the behavior not of otherwise undefined specimens of Homo supiens, but of particular sets of human beings who, as the result of both initial endowment and a mountain of experiences since birth, carry around in their heads a large body of knowledge and skills that is mostly social in origin and is enormously variable among subject populations as we move from one culture to another or from one time to another. In this important sense, all cognitive psychology is social psychology, and we have to look hard to discover invariants of behavior that remain stable over cultures and eras. In cognitive research, we have typically used two means to finesse this problem of the social relativity of behavior. First, we summarize the knowledge and skills of subjects drawn from a relatively homogeneous population under brief, but informative, labels like “college sophomores at XYZ University.” Second, we devote a large part of our research energy (or did so traditionally) to tasks that call largely for knowledge that all members of the subject population can be presumed to possess (e.g., puzzles like The Tower of Hanoi or Missionaries and Cannibals). In recent years, in contrast, we have also moved into research on knowledge-rich task domains by studying “expert” and “novice” populations-again assuming homogeneity within each population, but admitting the possibility of large differences between the populations. More and more, especially as our research moves in the direction of knowledge-rich tasks, we will be obliged to learn, by pretesting and in other ways, what attitudes toward goals, what information about problem representations, and what information about strategies subjects bring to the tasks we set for them. We will also use the experimental instructions and various kinds of pretraining to alter the goals, problem spaces, and strategies that subjects have available. E. THEMODELMUSTINCORPORATE A ~ E N T I OCONTROL N Human bounded rationality requires people to perform their tasks using only the knowledge they have, however much that may depart from the reality it purports to describe. Moreover, the inferences they will draw from their knowledge will be severely limited by their computational capabilities. Finally, of the body of knowledge they have stored in memory, only a fraction-often a very small fraction-of the knowledge potentially relevant to a particular task will be evoked initially, or even in the course of time, by the presentation of the task and the instructions.
The EPAM Model
27 1
Under these circumstances, to understand and predict behavior requires us to understand and predict what part of the information in memory will actually be evoked and applied in the course of task performance. What will subjects attend to and when? Under what circumstances will shifts of attention occur and evoke new information or lead to the loss of information previously evoked? A theory of cognition must incorporate a theory of attention, and, as we saw earlier, attention is closely linked to motivation and even to emotion. In the remainder of this chapter, we undertake to make these ideas more concrete by showing how they enter into the modeling of behavior observed in a well-known piece of experimental work on concept attainment and categorization. As the basis for our modeling, we use the EPAM system, which, first developed about 1959to account for a number of the phenomena of rote verbal learning, has been extended in the succeeding 35 or more years to account for a progressively wider range of phenomena of perception and memory.
111. Strategy, Goals, Attention, and Task Representation in EPAM
Each EPAM simulation has three components: (1) an experimenter module representing the activities of the experimenter; (2) a subject module representing the activities of the subject; and (3) a manager model, which coordinates the two whenever time is added to the simulated clock. The subject module requires problem spaces and strategies much as a mathematical model requires parameters. EPAM has not yet matured to the point where it determines its own goals and builds its own problem spaces. Currently, its problem spaces and strategies are programmed for each experiment and form an adjustable component of the model. Just as mathematical model builders seek to keep the number of numerical parameters low and their values consistent across many tasks, the EPAM programmer seeks to keep the strategies as simple and constant as is possible. Strategies are changed from one simulation to another only when necessary to correspond with differences in the experimental task or differences in the directions given to the subjects. A. STRATEGY Often the results produced by EPAM are direct outcomes of the strategies chosen, and are interpreted as such. For example, the very first simulation using EPAM predicted the invariant serial position curve as a consequence
212
Fernand Gobet et al.
of an attention strategy, called the “anchor point” strategy, used by people when familiarizing themselves with an ordered list (Feigenbaum & Simon, 1962). EPAM was programmed to learn the list by working from the anchor points inward. For most lists this led EPAM to choose the first and last elements as the obvious anchor points. The program thus demonstrated that an anchor point strategy could match quantitatively the ubiquitous serial position curve. Introduction of other, attention-attracting anchors (e.g., a word printed in red) automatically produced a von Restorff effect. Similarly, a 1967 simulation (Gregg & Simon, 1967) cast light upon a contradictory pattern of results where otherwise-similar experiments sometimes produced “one-trial’’ learning and sometimes incremental learning. When EPAM employed a strategy of rehearsing one syllable pair at a time until they were completely learned, it produced one-trial learning. Where it used an “all-at-once’’ strategy, stopping rehearsal of one pair whenever a new pair was presented, its learning was incremental. Postexperimental reports by subjects confirmed that this choice in rehearsal strategy corresponded to individual success with one-trial learning. EPAM explained how the strategy chosen by human subjects would determine whether they learned a syllable in a single trial or by increments over several trials. B. GOALS The goal of the subject module in many EPAM simulations is to add information to long-term memory and then later retrieve that information. For example, in a rote learning task the system gradually creates a net for paired associates indexed by the stimuli, so that when a stimulus is presented, a chunk, comprising the stimulus together with the response, can be accessed and the system can give the correct response. Similarly, in a concept-formation task the system creates a net or nets such that the stimulus will elicit a chunk or chunks indicating the correct category, and the system can respond with that category. In both tasks the subject has to associate stimuli with correct responses. The main difference between the tasks is that, whereas each stimulus elicits a different response in the pairedassociate task, several stimuli all elicit the same response in the categorization task. Subjects form generalizations that put many stimuli in the same class in the latter task. Studies of subjects’ strategies in concept attainment (e.g. Bruner, Goodnow, & Austin, 1956; Hunt & Hovland, 1961; Hunt, Marin, & Stone, 1966) have shown that people try to find a simple economical criterion involving only a few features that enables the system to categorize a great many stimuli. Such economy is impossible in the paired-associate task. Within a system like EPAM, such economy can be achieved if the system finds efficient tests while building the net. For example, if all members of
The EPAM Model
273
a “Category A” are red, and all members of a “Category B” are blue, an EPAM-like system would seek to put a test for color at the top of its discrimination net. Then all red items would sort to a single node and all blue items to a different node. In earlier versions of the EPAM model (versions I through IV), the system could not replace an inefficient test after it had been added to the discrimination net. The most recent version, EPAM V, permits the insertion of effective tests above inefficient tests in the discrimination net. This can occur whenever a generalization is studied. For example, a universe of simple geometric figures defined by three attributes, size, color, and shape, could include large red circles, small red circles, large blue squares, and so on. If all red squares, regardless of size, are in category A, then a generalization can be formed that “red squares regardless of size” elicit Category A. Assume that EPAM V had created the inefficient discrimination net shown in Fig. 1. Although EPAM can recognize small red squares as members of Category A, it currently misrecognizeslarge red squares as members of Category C. EPAM V, at this point, needs to replace an inefficient test for size in its discrimination net with a test for color, which it does by discovering the appropriate generalization and then studying it. There are many processes that concept formation systems could use to discover that red squares of any size are members of Category A. For example, the CLS (concept learning) system of Hunt et al. (1966) examined members and nonmembers of a category in order to discover a combination of features that appear together in all members but not together in any
Categoty C
Category B
B R E ; t w w c RED
Category A
Category B
Fig. 1. An inefficient discrimination net.
274
Fernand Gobet et al.
Category A
Category B
Fig. 2. EPAM’s modified discrimination net.
nonmember. The RULEX (Rule-Plus-Exception) system of Nosofsky, Palmer & McKinley (1994) systematically tried single features or combinations of features in order to find a feature or a set of features that was sufficiently predictive. A third process, that used by EPAM V in Medin and Smith’s experiment, compares succeeding pairs of members of a category and forms a rule that classifies together two members (e.g., a large red square and a small red square) that are found to differ by only a single feature. We do not hold that this method is used most often by people, nor is it the only system that could be used by EPAM in order to discover rules; in fact, versions of the other systems programmed for EPAM work at least as well. The advantage of the minimal-pairs system is that it enables EPAM to simulate both the concept formation task and the rote learning task with the same default strategy. Once a rule has been formed, EPAM studies it (see Fig. 2). When it studies “red square of any size” it sorts to the test for SIZE and inserts a test for the relevant attribute, COLOR, and a branch for red, producing the net shown in Fig. 2. Now all red squares will be correctly sorted to Category A.’
C. THEMEDINAND SMITH TASK In this section, we discuss the model’s simulation of the experiment conducted by Medin and Smith (1981). Our main purpose is not to compare
’
Implementational note: EPAM V permits multiple tests at a node. The new test for COLOR is simply added to the top of a list of tests at the node. The old test for SIZE remains at the same node in the net under the test for COLOR and is used when none of the branches for the test for COLOR applies.
The EPAM Model
275
EPAM with other models of categorization in their goodness of fit to the data, but to illustrate how directions given to subjects may be incorporated into the EPAM subject module as different strategies. These diverse strategies produce different patterns of response. A sequence of stylized (Brunswick & Reiter, 1938) faces (see Fig. 3) are presented to subjects, the faces varying with respect to eye height (EH), eye separation (ES), nose length (NL), and mouth height (MH), and subjects are instructed to assign them to Category A or Category B. In the five stimuli that the experimenter assigned to Category A (see Table I), four have high eyes and four have long noses, while three have wide eye separation and three have high mouths. Thus, positive values (1s) on these traits tend to indicate membership in A; negative values (Os), membership in B, with high eyes and long noses being the more reliable indicators. Four of the five stimuli have three of the positive traits, and one (Face 7) has only two. Faces 4, 7, and 15 all have both of the “reliable” traits; Faces 5 and 13 have only one each. In the four stimuli that the experimenter assigned to Category B, three have low eyes, three have short noses, three have low mouths, and two have narrow eye separation. Hence, the first three of these characteristics are criterial for B. Two of the Category B faces have two traits each that are criterial for A, and one for B (Faces 12,2); one face (14) has one trait that is criterial for A, and two for B; one (10) has no traits criterial for A, and three for B. We might expect that almost any scheme for learning to assign the faces to these categories will place especial weight on the two reliable traits for A and on the relative number of A traits and B traits a stimulus possesses. By whatever mechanisms these criteria are implemented, one could therefore predict that 4, 7, and 15 among the A faces, and 10 and 14 among the B
Category A
Face 4
Face 5
Face 7
Face 13
Face 2
Face 10
Face 12
Face 14
Face 15
Fig. 3. Sequence of stylized (Brunswick) faces assigned to categories based on eye, nose, and mouth characteristics.
Fernand Gobet et d.
276
TABLE I ATTRIBUTE STRUCTURE OF CATEGORIES USEDIN THE EXPERIMENT Attribute value Face number
EH
ES
NL
MH
A exemplars
4 7 15 13 5 B exemplars
1 1
1 1 0
12
1
2 14
0
0 0 New transfer items 1 1 3 1 6 1 8 0 9 0 11 0 16 0 10
Note. EH = eye height; ES = eye separation; NL = nose length: MH = mouth height. See text for explanation of binary notation. From Medin & Smith (1981).
faces would be easy to learn; while 5 and 13 among the A faces and 2 and 12 among the B faces would be hard to learn. We would not expect experimental instructions or other similar interventions to change very much this division of stimuli between “easy” and “hard.” We will see that, in general, these predictions hold up well, but that we can refine them a bit further if we infer from specific experimental instructions the corresponding learning strategies the subjects will use.
D. EPAM SIMULATION OF MEDINAND SMITH (1981) Medin and Smith presented three groups of subjects with separate instructions for the categorization task: (1) “standard instructions,” (2) “rulesplus-exception instructions,” and (3) “prototype instructions.” We have translated each set of instructions into a strategy programmed into the EPAM module. Each strategy has two interrelated components: (1) a recog-
The EPAM Model
277
nition strategy and (2) a learning strategy. First, we look at each of these instructions and our interpretation of it by a program for the EPAM subject module. Then we compare the results of employing each of the three strategies to simulate the experimental behavior with the behavior of Medin and Smith’s subjects in the corresponding experimental condition. In proceeding in this way, we are adding additional degrees of freedom to the EPAM structure (the strategies we assume), thereby reducing its parsimony and predictive force. We have to assume that the strategies we construct are veridical interpretations of the task instructions, and a plausible case can usually be made for more than one strategy as compatible with the instructions. Selecting a particular strategy from this set is like selecting particular parameter values to fit a theory that contains parameters. We will be interested in the sensitivity or insensitivity of the predictions to postulated strategy differences. A more powerful theory, a future EPAM, would generate the strategies automatically from the task instructions, in a manner similar to the way that the UNDERSTAND program generates problem representations from verbal task instructions (Hayes & Simon, 1974). Such a theory would have substantially fewer degrees of freedom, leaving no room for the modeler’s judgment in associating strategies with instructions. But this more complete version of the EPAM theory has yet to be constructed. The Medin and Smith experiment had three phases: a learning phase with feedback, which continued for 32 trials or until the subject had a single perfect trial; a transfer phase without feedback, which continued for 2 trials; and a speeded-classification phase, with feedback, for 16 trials during which time the subject’s response latencies were measured. The subject routine responds with the same algorithm for every stimulus or stimulus-response pair presented: 1. Subject waits until the next stimulus appears in its visual sensory store. 2. If the stimulus that appears states that the experiment is done, subject exits this loop. 3. Subject responds to the stimulus, using its find-category routine. This routine can vary with different instructions. 4. Subject waits for the next stimulus or response to be presented in the visual sensory store. 5. If the stimulus that appears states that the experiment is done, subject exits this loop. 6. If the next stimulus or response is a stimulus rather than a response, subject returns to Step 3. (This occurs during the transfer phase, in which feedback is not given.) 7. Subject studies the contents of the visual sensory store and the visual imagery store using the study-category routine. This routine can vary with the instructions.
278
Fernand Gobet et al.
8. Subject replaces the current contents of the visual-imagery store with a copy of the current contents of the visual-sensory store. Step 9. Subject returns to Step 1.
Different find-category routines and study-category routines were used by the subject in the different experimental conditions.
E. STANDARD INSTRUCTIONS We have identified the standard instructions in the Medin and Smith experiment with the default concept attainment strategy of the EPAM model. The standard instructions simply tell the subjects to guess at first, but then to pay attention to the feedback so that they can assign each face to its appropriate category. EPAM’s default find-category routine is to sort the stimulus in the net and report the category at the node it reaches. If it cannot find a category associated with the stimulus, then it guesses. EPAM’s default study-category routine is the following algorithm: 1. If the find-category routine responds correctly, do nothing. 2. Pick a random number from 0 to 99, if it is over 17, then do nothing. 3. If a previous study-category routine is busy transferring information to long-term memory (it takes EPAM 5 s of background learning time to add a new chunk such as a new node or response to its discrimination net), do nothing. 4. If the last stimulus (currently in the visual-imagery store) and the present stimulus (currently in the visual-sensory store) share the same response and have at least three features in common, then form and memorize a generalization consisting of the features that are on both stimuli and associate that generalization with the correct response. 5. Otherwise associate the present stimulus with the correct response.
As an example, the discrimination net produced by a particular run of EPAM in the standard instructions condition is illustrated in Fig. 4. The net is shown as it was at the conclusion of the learning phase of the experiment. This net has a test EH for eye height at its top node and additional tests for nose length and mouth height within the net. Examination of the net reveals the following: 1. Faces 4,7, and 15, which have E H = 1 but do not have NL = 0, are sorted to a node that states that they belong to Category A. EPAM must apply two tests in order to reach this terminal. Assuming that it takes EPAM about 10 ms to notice that a new face has been presented, 100 ms to enter a discrimination net, and 250 ms to find each of the attribute
The EPAM Model
FACE B 2
279
A FACE 13
\ ? FACE 5
B FACE 12
Fig. 4. Discrimination net produced by EPAM in the standard instructions condition in the Medin and Smith experiment.
values (EH and NL) using tests that require separate eye fixations, with the discrimination net of Fig. 4, it takes EPAM 610 ms to categorize Face 4, 7, or 15 as Category A. 2. Face 13 has EH = 1 and NL = 0 but not MH = 0. It is sorted to a terminal that is labeled as belonging to Category A. It takes EPAM about 860 ms to categorize this face: 10 ms to notice that a new face has been presented, 100 ms to enter the net, and 750 ms to perform the three tests. 3. Face 12 has EH = 1, NL = 0, and MH = 0. It takes EPAM 860 ms to categorize it as a member of Category B. 4. Faces 10 and 14 (EH = 0 and NL = 0) are each sorted in 610 ms to a terminal labeled for Category B. 5. Face 2 (EH = 0, NL = 1, and MH = 0) is sorted in 860 ms to a terminal labeled for Category B. 6. Face 5 (EH = 0, NL = 1, and MH = 1) sorts to an ambiguous node that is not labeled with a category. During the 26th trial of the learning phase, EPAM guessed correctly that Face 5 was a member of Category A and as a result, EPAM was able to categorize all of the faces correctly during that trial, even though it had not yet learned the correct categorization of Face 5. During both of the transfer phase trials, EPAM guessed incorrectly that Face 5 was a member of Category B. At the beginning of the 16-trial speeded-classification phase of the experiment, it took EPAM 860 ms to sort Face 5 to the node for its face, and an additional 1000 ms to guess the value of a category, for a total of 1860 ms. During this phase of the experi-
280
Fernand Gohet et al.
ment, however, EPAM learned the correct category for Face 5 , and then only 922.5 ms was taken by EPAM to categorize Face 5.
F. RULES-PLUS-EXCEPTION INSTRUCTIONS The rules-plus-exceptions instructions that Medin and Smith (1981) gave their second group of subjects were much more complex than the standard instructions. The former described a two-stage learning process and a complex recognition process. During the first stage of the learning process the subjects were told to create a rule, based on nose length, find which category long noses usually belong to, and then associate short noses with the other category. In this first stage, EPAM’s subject module creates a rule net and as the first stimuli come in, determines which category is most often associated with short noses and which category with long noses. At the completion of this stage, the subject module creates a net to cache the results of this rule: If NL = 1, the net sorts to a node labeled for Category A. If NL = 0. the net sorts to a node labeled for B. During the second learning stage subjects were instructed to memorize exceptions to the rule. EPAM stores exceptions in a second net for exceptions. The terminals in this net reports “Yes” if the item that sorts to the node is an exception and “No” if it is not. Medin and Smith describe the rules-plus-exceptions strategy to subjects in the following general fashion: When you have mastered the task, you will be doing something like looking to see if the face is one of the exceptions, if so, make the correct response, if not, apply the rule for short and long noses. (p. 247)
EPAM’s find-category routine uses a recognition strategy that corresponds closely to these instructions. First, it looks in the exceptions net to see whether the stimulus is an exception. Then it sorts in the rule net to find out whether the rule classifies the face as A or B. If the item is not an exception, EPAM categorizes the item according to the decision of the rule net. On the other hand, if the item is an exception, EPAM reverses this decision. If the rule has not yet been learned, EPAM accesses the hypothesis currently being held in short-term memory and categorizes the item according to the prediction that would be made by that hypothesis. EPAM’s study-category routine for the rules-plus-exceptions condition is also a two-stage algorithm. In the first stage, if the subject has not yet learned the rule, it forms a hypothesis, and studies the rule following a strategy very much like that outlined in the instructions. Specifically, it accepts a nose-length rule, such as “short nose predicts Category A,” when
The EPAM Model
281
its excess of correct over incorrect predictions exceeds 3, and accepts the opposite nose-length rule (i.e., "short nose predicts Category B") when the excess of correct over incorrect predictions dips below zero. The second stage is very much like the find-category routine except that the system is determining whether or not the face is a member of the exceptions net. Specifically: 1. If the find-category routine responds correctly, do nothing. 2. Pick a random number from 0 to 99, if it is over 17, do nothing. 3. If a previous study-category routine is busy transferring information to long-term memory, do nothing. 4. If the last stimulus (which is currently in the visual-imagery store) and the present stimulus (which is currently in the visual-sensorystore) share the same response and at least three features in common, fdrm and completely memorize a generalization consisting of the features that are on both stimuli and label that generalization as exception or not, consistently with the present stimulus. 5. Otherwise label the present stimulus as an exception or not, as the case may be.
The two discrimination nets after completion of EPAM's learning stage in a run of the rules-plus-exceptions condition are illustrated in Fig. 5. The net on the left is a rules net with a single test for nose length. The net on the right is an exceptions net with a top test for mouth height. In this case,
R"Lp\ B
A
Jy? FACE 15
No FACE 10 FACE 7
No FACE 2 FACE 4
No FACE 5 FACE 12 FACE 13
Fig. 5. Two discrimination nets after completion of EPAM's learning stage in a run of the rules-plus-exceptions condition.
282
Fernand Gobet et al.
EPAM went through all 32 learning trials without a perfect trial, and, indeed, using this net, EPAM currently misclassifies both of the exceptions to the rule, Faces 13 and 2. It currently takes EPAM about 960 ms to categorize each face: 10 ms to react to the stimulus, 100 ms to enter each of the two nets, and 250 ms to sort through each of the three tests (the test for NL in the Rule net and the tests for MH and ES in the Exception net. During the speeded-classification phase of the experiment, the system adds additional tests to the net, and these permit it to discriminate the exceptions. As a result, the average categorization latencies for Faces 4,5, 13, 2, and 12 are higher than 960 ms and with the guessing that occurs before the categories for the new nodes are learned, the average latencies for the two exceptions, Faces 13 and 2, are over 1350 ms each. As Faces 13 and 2 are the “exceptions” in this condition, we would expect them, with almost any strategy consistent with the instructions, to be more difficult to learn than the others. In all three experimental conditions these faces are among the three most difficult for EPAM, but they are especially difficult in the Rules and Exceptions condition. On the other hand, Face 12, which is not an exception to the long-nose rule, but is first or second in difficultyin the other two conditions, is fourth (and very much easier than Faces 13 and 2) in this condition.
G . PROTOTYPE INSTRUCTIONS Medin and Smith’s instructions for their “prototype” subjects were to memorize what A faces look like and what B faces look like. They were told that they later would have to answer questions about the characteristics of each type of face. EPAM memorizes types of faces by making a separate net for each type. Medin and Smith’s instructions were: “we want you to use these general impressions to help you classify these faces.” EPAM’s find-category routine for the prototype condition does this by sorting each stimulus in both nets. If a face is found to be a member of one category but not the other, EPAM responds with the former category. If it is found to be a member of both or is not found to be a member of either, the subject module guesses the category. If it guesses wrong, it elaborates the net for the correct category in which to include this stimulus. EPAM’s study-category routine for the prototype condition follows almost the identical strategy as the study-category routine for the standard condition, except that there is the additional step: the study-category routine must determine which net to use for studying. Specifically: 1. If the find-category routine responds correctly, do nothing. 2. Pick a random number from 0 to 99, if it is over 17, do nothing.
The EPAM Model
283
3. If a previous study-category routine is busy transferring information to long-term memory, do nothing. 4. If the last stimulus (which is currently in the visual-imagery store) and the present stimulus (which is currently in the visual-sensory store) share the same response and at least three features in common, form a generalization consisting of the features common to both stimuli. If the stimulus was not identified as a member of the correct net, then memorize the generalized stimulus completely in the correct net and associate it with “Yes” in that net. On the other hand, if the stimulus was identified as a member of the correct net, then it must have been misidentified as a member in the other net, so memorize the generalized stimulus completely in the other net and associate it with “No” in that net. 5. If the stimulus was not identified as a member of the correct net then associate it with “Yes” in that net. If the stimulus was identified as a member of the correct net, then memorize the stimulus completely in the other net and associate it with “No” in that net.
The two discrimination nets that resulted from one particular run of EPAM in the prototype-instructions condition are illustrated in Fig. 6. 1. Face 10 sorts to a node that identifies members of the B Net and to another node that does not identify it as a member of the A Net. The subject responds with category B in 1210 ms: 10 ms to react to the stimulus, 200 ms to enter the two nets, and 1000 ms to sort through the four tests. 2. Faces 4,7, 13, and 15 are members of the A Net but are ambiguous in the B Net. They are correctly identified as members of Category A. ANET
BN%
Case 14 Case 13 Case 15
t
1
$
Case14 ? Case 12 case 10
case 10
Case 7 case 4
case2
? Case 7 case 4 Case 12
?
Case 15 Case 13
Fig. 6. Two discrimination nets resulting from one run of EPAM in the prototypeinstructions condition.
284
Fernand Gobet et al.
3. Face 12 is ambiguous in both nets, so it is categorized by guessing. It was correctly categorized by guessing on the 32nd learning trial. 4. Faces 2,5, and 14 are identified as members of both nets, so they are categorized by guessing. All three were correctly categorized by guessing on the 32nd learning trial.
H. COMPARISON WITH HUMAN DATA A single free parameter called the “study parameter” was adjusted so that approximately the same proportion of EPAM-simulated subjects as real subjects would attain the same overall result in the learning condition. Medin and Smith reported that 14 of 32 people met the criterion of a perfect trial in both the standard and rules-plus-exceptions conditions, while only 8 of 32 people met the criterion in the prototype condition. With the study parameter set at 18, so that EPAM studied only 18% of the time when studying was possible, 151 of 320 simulated subjects met the criterion in the standard condition, 129 of 320 met the criterion in the rules-plusexceptions condition, and 63 of 320 met the criterion in the prototype condition, closely matching the ratios for the human subjects. Both EPAM and people found it easier to meet the criterion in the standard and rulesplus-exceptions condition than in the prototype condition. Table I1 compares Medin and Smith’s human subject error data with the results from 100 runs of the EPAM model for each condition. There was no special adjustment of parameters for this simulation, and simple and straightforward interpretations of the instructions were used for EPAM. For all three experimental conditions, the Pearsonian correlation between human subjects and EPAM of the numbers of errors for the various faces is very high: .93, .93, and .77 for the three conditions. In EPAM as in the human experiments, Faces 13, 2, and 12 (except 12 in the Rules and Exceptions condition) produced by far the largest number of errors. There was little difference in rank order among the several conditions for either the human subjects or EPAM. In this important respect, the instructions had little effect on the outcomes, affecting only the overall level of difficulty of the task as a whole. The relative reduction in difficulty of Face 12 in the Rules and Exceptions condition was reflected in the performance of both subjects and EPAM. Face 7 produced fewer errors than Face 4 in all conditions, which, as Medin and Smith point out, is consistent with context models but not with independent-cue models, where the net effect of cues is additive. As EPAM is, in many respects, highly nonlinear and nonadditive in its operation, hence a “context” model in the sense of Medin and Smith, we would predict this result. On average, both the subjects and EPAM found the prototype
The EPAM Model
285
TABLE I1 MEANNUMBER OF ERRORS FOR EACH FACEDURING INITIAL LEARNING As A FUNCTION OF INSTRUCTIONS Instruction Standard Face number 4 5 7 13 15 2 10 12 14 M S
Pearson's r a
Rules and Exceptions
Prototype
People
EPAM
People
EPAM
People
EPAM
4.5 8.2 4.2 11.9 2.8 12.9 4.4 15.2 6.6 7.9 4.5
9.0 10.8 6.6 11.3 5.6 14.0 6.6 12.8 8.4 9.5 2.9
3.9 5.9 3.3 10.7" 2.8 13.8" 3.8 6.3 6.8 6.3 3.1
6.5 8.5 4.0 18.8" 4.5 18.5" 4.9 1.7 5.3 8.1 5.8
1.7 9.2 6.7 13.7 4.9 10.3 4.2 17.4 8.7 9.2 4.2
11.2 11.9 9.9 12.8 8.6 16.0 10.4 15.1 12.4 12.0 2.4
.93
.93
.I7
Face was an exception in rules-plus-exceptions condition.
condition hardest, the standard condition next hardest, and the rules-andexceptions condition easiest. The range of errors from the easiest to the most difficult faces was smaller for EPAM than for the human subjects in the standard and prototype conditions, but not in the rules-plus-exception conditions. The simulation of the prototype condition is arguably the least satisfactory of the three. The errors for subjects in the standard and prototype conditions were closely similar ( r = .92), suggesting that some subjects in the prototype condition may have ignored the instructions and followed a strategy much like that used by subjects in the standard condition. The correlation between the standard EPAM condition and the prototype subject condition is .81, higher than the correlation between the EPAM and subject prototype conditions (.78). To obtain a closer fit of EPAM to the human data in the standard and prototype conditions would require EPAM to respond less promptly to the need to add new branches to the net. A change in strategy in this direction would probably also increase the relative difficulty of the harder over the easier faces. However, we have preferred to show the quite good results obtained with a strategy that was not specially "tuned" to the data.
Fernand Gobet et al.
286
I. TRANSFER TASK After completing their initial learning of the classification of the nine faces, subjects were given a transfer task, in which they were asked to categorize the same faces again, intermingled with examples of seven new faces. The results of the transfer experiment are shown in Table 111. Again, there is a close relation between the subjects' data and the EPAM simulations on the transfer test, the relation being somewhat closer for the old than for the new faces. EPAM tends to move closer to chance (50%) on the new faces, which is consistent with the fact that it sometimes guessed TABLE 111 OBSERVED A N D PREDICTED PROPORTIONS OF CORRECT CATEGORIZATIONS FOR EACHFACEDURING TRANSFER Instruction Rules and Exceptions
Standard Face number Old faces 4A 7A 15A 13A 5A 12B 2B 14B 1OB
M M abs. diff. New faces 1A 6A 9A 11A 3B 8B 168 M abs. diff.
Prototype
People
EPAM
People
EPAM
People
EPAM
.97 .97 .92 31 .72 .67 .72 .97 .95 .86
.83" .96 .96 .78 .76 .71 .68 .89 .96 .84
.89 .94 .94 .72 .78 .73 .70 .91 .95 .84
.87 .98 .94 .64 .I7 .80 .67 .94 .96 .84
.77 .97 .88 .70 .60 .45 .72 .83 .87 .77
.75 .X5" .90 .68 .68 .57* .57" .75 .89 .74
.05 .72 .98 .27 .39 .44 .77 .91
.03 .54" .93 .52" .58" .63a .55a .83
.I7
.45
.08
.26" .74a .33" .88" .91" .28" .66"
.88
.08 .75 .80
.42 .88 .17
.73
.56"
.87
.nn so"
.28 .52 .35 .78
.61
55" .53" .74*
.88
.15
Note. In EPAM, the current hypothesis is lost from short-term memory before the transfer test. Thus, both the items that were correctly identified via the hypothesis and those correctly identified via guesses often produce errors on the transfer test. a Difference between subjects and model exceeds .lo.
The EPAM Model
287
the categories correctly in the learning experiment without extending the differentiation net to classify them unambiguously. Nevertheless, the correlation over all conditions combined between errors made by EPAM and errors made by subjects for all the different faces was 32, but .88 for the faces seen previously and .73 for the seven new faces. Medin and Smith used regression models (a “context” or multiplicative model, and an “independent cue” or additive model) to fit the data from the transfer experiment, obtaining average absolute deviations about a third as large as EPAM’s in the first case, and about half as large in the second. However, each of these models had 4 free parameters that were used in fitting the data, and were estimated separately for the three experimental conditions-a total of 12 parameters. Hence, it is hard to conclude that the regression models did a better job than EPAM of fitting the facts. Medin and Smith remark that the results were very sensitive to the exact values of the parameters, which suggests that the parameters were doing much of the work. This kind of flexibility was not available to EPAM, for the same interpretations of the instructions were used to model both the learning and transfer experiments. J. SPEEDED CLASSIFICATION Finally, after the subjects had completed the transfer task, they were asked to perform the classification task again with the original nine faces, but respond as rapidly as they could. In Table IV, we show the average reaction times of the subjects in responding to each face for each set of instructions, and compare these with EPAM’s reaction times, without modifying any of EPAM’s time parameters from their usual values. In the three conditions, the subjects took, on average, 31, 28, and 24% longer than the EPAM simulation. Hence, the times predicted with parameters obtained from earlier studies of rote verbal learning provided a reasonable fit to the data. There is a high rank-order correlation between the times, averaged over subjects, taken to respond to the different faces and the numbers of errors they had made while learning the faces. In the three experimental conditions, there are rank-order correlations between EPAM’s times on individual faces and the subjects’ times of .71, S O , and .13, respectively. Thus, the speeded-classification task shows much the same pattern of findings as the two previous tasks.
K. DISCUSSION In this chapter we have described how EPAM, a program originally constructed to predict the behavior of human subjects in verbal learning experiments, can be used to predict behavior in categorization experiments, with-
Fernand Gobet et al.
288
TABLE IV
MEANREACTION TIMES FOR CORRECT RESPONSES FOR EACH OLDFACE DURING SPEEDED CLASSIFICATION As A FUNCTION OF INSTRUCTIONSO Instruction Rules and Exceptions
Standard Face number 4 5 7 13 15
2 10 12 14 M
Prototype
People
EPAM
People
EPAM
People
EPAM
1.11 1.34 1.08 1.27 1.07 1.30 1.08 1.13 1.19 1.17
.96 1.02 .70 1.03 .64 1.14 .65 1.10 .I7 .89
1.27 1.61 1.21 1.87' 1.31 1.97' 1.42 1.58 1.34 1.51
1.23 1.33 .93 I .44h .95 1.53h .96 1.30 .95 1.18
1.92 2.13 1.69 2.1 2 1.54 1.91 1.64 2.29 1.85 1.90
1.65 1.68 1.39 1.64 I .22 1.78 1.25 1.75 I .44 1.53
Note. Reaction times are calculated using 250 msltest node traversed plus 250 mslnet utilized. plus 250 ms if the system guesses the category. " Mean reaction times in seconds. Face was an exception in the rules-plus-exceptions condition.
out the need to modify substantially the basic learning and performance mechanisms of the system or the time parameters that predict rate of learning and speed of response. To illustrate how EPAM accomplished this, we took as an example a task that had been studied by Medin and Smith under three different conditions that corresponded to three different sets of task instructions. This task is of special interest because it requires subjects to use a different strategy for each of the experimental conditions. Hence, the data of the subjects' performance reflect not only their own learning and response capabilities, but also differences of difficulty in categorizing the individual stimuli (characteristics of the task domain) and differences in the strategies they adopt. Although the importance for task difficulty of the task domain, the subjects' representation of the task domain (the problem space), and the strategies employed by subjects has been known for a long time (see, e.g., Newel1 & Simon, 1972, especially Chapter 14), there are still relatively few published experiments in which these variables have been manipulated, or in which the subjects' behavior on these dimensions have been recorded and reported.
The EPAM Model
289
The complex, and sometimes apparently conflicting, results that appear in the literature on concept attainment and categorization underscore the importance of understanding in as much detail as possible the processes that subjects use to formulate and attack the problems presented to them, and the differences in performance that can be produced by different choices of problem space and strategy. But in addition, as this particular set of experiments shows, the detailed structure of the task domain can itself show through into the behavior of subjects: here, largely determining the relative difficulties of the different stimulus items in a way that is predictable from the structure of the stimuli. In our reexamination of the Medin and Smith data, we have also addressed the issues that must be faced in applying models like EPAM that, while they simulate processes in some detail, have considerable generality enabling them to model behavior over a wide range of laboratory tasks. Before any model possessing substantial generality can be employed in a particular task, a component must be added to the model to represent the task definition-its goals and constraints-and another component to represent the subjects’ strategies. In traditional mathematical modeling, these adaptations are achieved by manipulating parameters that are built into the model structure. In modeling symbolic processes, they are achieved by constructing and inserting subroutines corresponding to these components of the task. Of course, the degrees of freedom available for shaping the components cause a loss of parsimony in the theory. In the long run, the added components should not be built on an ad hoc basis for each task, but should emerge from the workings of learning processes that constitute a permanent part of the model itself. In the absence of such learning processes, effort must be taken to obtain direct evidence from the behavior of the human subjects of the problem representations and strategies they are actually using to perform the task.
REFERENCES Bruner, J. S., Goodnow, J. J., & Austin, G. A. (1956). A study of thinking. New York: Wiley. Brunswik, E., & Reiter, L. (1938). Eindruckscharaktere schematisierter Gesichter. Zeitschrifi fur Psychologie, 142, 67-134. Chase, W. G . ,& Simon, H. A. (1973). The mind’s eye in chess. In W. G . Chase (Ed.), Visual information processing (chap. 5). New York: Academic Press. de Groot, A. D. (1946). Het denken van den schaker. Amsterdam: North-Holland. Ericsson, K. A,, & Staszewski, J. (1989). Skilled memory and expertise: Mechanisms of exceptional performance. In D. Klahr & K. Kotovsky (Eds.), Complex infornration processing: The impact of Herbert A. Simon (pp. 235-267). Hillsdale, NJ: Erlbaum. Feigenbaum, E. A., & Simon, H. A. (1962). A theory of the serial position effect. British Journal of Psychology, 53, 307-320.
290
Fernand Gobet et el.
Gregg, L. W., & Simon, H. A. (1967). An information-processing explanation of one-trial and incremental learning. Journal of Verbal Learning and Verbal Behavior, 6, 780-787. Hayes, J. R., & Simon, H. A. (1974). Understanding written problem instructions. In L. W. Gregg (Ed.), Knowledge and cognition (chap. 8). Potomac, MD: Erlbaum. Hunt, E. B., & Hovland, C. I. (1961). Programming a model of human concept formulation. Proceedings of the Western Joint Computer Conference, 1961, pp. 145-155. Hunt, E. B., Marin, J., & Stone, P. J. (1966). Experiments in induction. New York: Academic Press. Kaplan, C. A,, & Simon, H. A. (1990). In search of insight. Cognitive Psychology, 22(3). 374-419. Medin, D. L., & Smith, E. E. (1981). Strategies and classification learning. Journal of Experimental Psychology: Human Learning and Memory, 7(4), 241 -253. Newell, A. (1990). Unified theories of cognition. Cambridge, MA: Harvard University Press. Newell, A., & Simon, H. A. (1972). Human problem solving. Englewood Cliffs, NJ: Pren tice-Hall. Nosofsky, R. M., Palmeri, T. J., & McKinley, S. C. (1994). Rule-plus-exception model of classification learning. Psychological Review, 101( l), 53-79.
AT A LOSS FROM WORDS: Verbal Overshadowing of Perceptual Memories Jonathan W. Schooler Stephen M. Fiore Maria A. Brandimonte
There are two broad, and seemingly contrasting, themes that characterize many discussions of the relationship between language and thought. O n one hand, many theorists have proposed that language represents the central scaffolding for cognition. In this vein, Wittgenstein (1922/1961) observed, “The limits of my language mean the limits of my world” (p. 115) and Sapir (1921, cited in Hardin & Banaji, 1993) proclaimed, “We see and hear and otherwise experience very largely as we do because the language habits of our community predispose certain choices of interpretation” (p. 277). An equally acclaimed tradition, however, has argued that there are many thoughts that transcend words. So, for example, James (1891) noted, “Great thinkers have vast premonitory glimpses of schemes of relations between terms, which hardly even as verbal images enter the mind, so rapid is the whole process” (p. 255). Einstein (cited in Schlipp, 1949), in a striking fulfillment of James’ characterization, reported, “These thoughts did not come in any verbal formulation. I very rarely think in words at all. A thought comes, and I may try to express it in words afterwards” (p. 228). Although these two depictions of the relationship between language and thought might seem at odds, research in a number of domains of perceptual memory suggests that they may both be accurate. For example, individuals’ ability to successfully recognize difficult-to-verbalize colors (Heider, 1972), THE PSYCHOLOGY OF LEARNING AND MOIIVATION. VOL. 37
291
Copyright 0 1997 by Academic Press. All rights of reproduction in any form reserved. 0079-7421197 $25.00
292
Jonathan W. Schooler et al.
faces (Polanyi, 1966), and nonverbal forms (Attneave, 1957) reveals the substantial degree to which knowledge can often transcend linguistic skill. Yet these domains are not immune to the influence of language, as revealed by the recognition advantage of easily named colors (Lucy & Schweder, 1979), the impact of postevent verbal information on memory for faces (Greene, Flynn, & Loftus, 1982), and the influence of verbal labels on memory for form (Carmichael, Hogan, & Walter, 1932). In short, cognitive representations can exceed, and yet still be influenced by, language. The confluence of these two premises raises an intriguing question: What happens when one attempts to articulate cognitions that cannot be fully captured in words? More specifically, what happens when one attempts to describe their memory for an indescribable perceptual experience? If perceptual memories exceed words, and yet can be constrained by language, then describing one’s recollections of perceptual experiences might actually impede later access to the nonreportable aspect of those experiences. There has been a growing accumulation of evidence that verbalization of perceptual memories can interfere with subsequent memory performance. This verbal disruption of nonverbal cognition was initially examined by J. W. Schooler and Engstler-Schooler (1990) in the domain of face recognition. Face recognition is an ideal area in which to examine the impact of verbalization on nonverbal cognition because of the marked disparity between nonverbal memory, as revealed by recognition ability, and verbal memory, as revealed by the (in)ability to describe faces. Indeed, this disparity has served as the jumping-off point for prior philosophical discussions of the relationship between verbal and nonverbal thought. For example, Polanyi (1966) began his seminal discussion of the nature of tacit knowledge with the following observation: I shall reconsider human knowledge by starting from the fact that we can know more than we can tell. . . . Take an example. We know a person’s face, and can recognize it among a thousand, indeed a million. Yet we usually cannot tell how we recognize a face. So most of this knowledge cannot be put into words. (p. 4)
This self-evident disparity between the verbalizable and nonverbalizable aspects of face memory has also been demonstrated empirically. Although face recognition is typically quite good (Shapiro & Penrod, 1986), verbal descriptions of faces are often not precise enough to enable judges to distinguish target faces from similar distractors (e.g., Ellis, Shepard, & Davies, 1980). If face recognition performance markedly exceeds the ability to articulate the basis for that performance, and if the use of language can influence the application of nonverbal knowledge, then describing one’s memory for
At a Loss from Words
293
a face might actually hamper subsequent recognition. It is important to note that such a prediction is contrary to the standard view that verbal rehearsal generally improves memory performance (e.g., Darley & Glass, 1975; Glenberg & Adams, 1978; Maki & Schuler, 1980). Nevertheless, J. W. Schooler and Engstler-Schooler (1990) found that verbalizing the appearance of a previously seen face significantly impaired subjects’ subsequent ability to discriminate the target face from verbally similar distractors. Schooler and Engstler-Schooler termed this interference “verbal overshadowing” on the basis of two hypotheses regarding the nature of this interference; first, that the disruptive effects of verbalization are specifically the consequence of verbalizing nonverbal cognition, and, second, that verbalization overshadows but does not eradicate the original perceptual memory. In addition, they introduced a general account of verbal overshadowing effects, termed the “recoding interference hypothesis” based on the premise that the act of verbal recall produces a verbally recoded memory representation that interferes with access to the original visual representation. We review Schooler and Engstler-Schooler’s original studies in the context of these three characterizations of verbal overshadowing effects in order to provide a conceptual foundation on which to base our review of the more recent findings on this topic.
I. Three Premises of Verbal Overshadowing A. THEMODALITY MISMATCH ASSUMITION A central component of Schooler and Engstler-Schooler’s original account of the negative effects of verbalizing a face is the assumption that the disruptive effects of verbalization are specifically the result of a mismatch between the nonverbal perceptual knowledge associated with the original memory and the verbal knowledge associated with the act of verbalization. This premise, which we here term the modality mismatch assumption, is based on the standard premise of many cognitive models that memory can involve two distinct types of knowledge: nonverbal/perceptual information and verbal/conceptual information (e.g., Bartlett, Till, & Levy, 1980 G. R. Loftus & Kallman, 1979; Mandler, 1980; Paivio, 1986). One central prediction of the modality mismatch assumption is that the effects of verbalization should critically depend on the degree to which memory performance relies on nonverbal knowledge. Schooler and Engstler-Schooler provided several strands of evidence for the role of memory verbalizability in mediating the effects of verbalization. For example, they investigated the impact of verbalizing a previously seen color. As with faces, color represents another
294
Jonathan W. Schooler et al.
domain of perceptual memory that, although influenced by language, cannot be fully captured in words (cf. Heider, 1972; Lucy & Schweder, 1979). And, as with faces, Schooler and Engstler-Schooler observed that verbalization of a previously seen color impaired subsequent memory performance. When subjects saw a shade of a color (e.g., army green) and then attempted to describe it, their subsequent ability to distinguish that particular shade from other similar shades was impaired relative to control subjects who did not describe the color. In contrast to the disruptive effects of verbalizing memories for nonverbal stimuli such as faces and colors, Schooler and Engstler-Schooler observed a rather different pattern of results when they examined the impact of verbalization on the recognition of a more verbal stimulus. In this study, after subjects viewed a video tape of a bank robbery, they were asked to engage in one of three tasks: verbally recalling the appearance of the robber’s face, verbally recalling what he said, or engaging in an unrelated verbal activity (e.g., naming states). Subjects were then given recognition tests for both the target face and the target statement. As can be seen in Fig. 1, although verbalization markedly impaired recognition of the target face, it numerically improved recognition of the verbal statement, thereby providing further support for the premise that the effects of verbalization critically depend on the degree to which the memory task requires nonreportable knowledge. When nonreportable knowledge is required, as in the case of face and color recognition, verbalization disrupts performance. However, when nonreportable knowledge is not required, as in the case of statement recognition, verbalization is benign. Additional evidence for the modality mismatch assumption was provided by an examination of the impact of engaging in nonverbal recall, that is, visualization. If the consequences of verbalization are specifically the result of attempting to commit nonverbal knowledge to words, then visual recall should not disrupt performance. Consistent with this prediction, Schooler and Engstler-Schooler observed that, in contrast to verbalization, visualization of a previously seen face did not impair performance. Visualization was also found to have no effect on color recognition.
B. THEAVAILABILITY ASSUMPTION Another premise of verbal overshadowing is that verbal knowledge overshadows but does not eradicate the original nonverbal memory. The central claim of this assumption is that the original memory remains available (Tulving & Pearlstone, 1966), and thus effects of verbalization should be reversible if conditions can be introduced that favor retrieval of nonverbal knowledge. Schooler and Engstler-Schooler addressed the availability of
At a Loss from Words
295
L
0
0
Condition Fig. 1. Percentages of correct face and statement recognition. From J. W. Schooler & Engstler-Schooler (1990, Experiment 4). Copyright 1990 by Academic Press, Inc.
the original memory by drawing on the premise, suggested by a variety of investigators, that visual components of perceptual memories are accessed prior to the verbal components (e.g., Bartlett et al., 1980; Paivio, 1986; Rabinowitz, Mandler, & Barsalou, 1977). This differential access rate of the two types of information suggests that, if subjects are limited in the amount of time given to make a recognition response, they might be compelled to rely primarily on their visual representation, thereby avoiding the disruptive consequences of verbalization. In a final experiment, Schooler and Engstler-Schooler tested this premise by introducing a new condition in which subjects were given only five seconds in which to recognize the target face. Under the standard recognition conditions, the negative impact of verbalization was observed. However, when subjects were forced to make very quick decisions, thereby presumably constraining them to the quickly accessed visual representation, no effect of verbalization was shown. This finding supported the availability assumption and thus the claim that
296
Jonathan W. Schooler et al.
verbalization overshadows but does not eradicate the original visual memory. INTERFERENCE HYPOTHESIS C. THERECODING Schooler and Engstler-Schooler accounted for the findings just described by suggesting that the act of verbally recalling a nonverbal memory results in the formation of a new verbally biased representation, which interferes with access to the original visual memory. This recoding interference hypothesis is generally consistent with various memory theories that assume that perceptual memories are often recoded into verbal representations (e.g., Bransford & Franks, 1971; Glanzer & Clark, 1963). The verbal recoding hypothesis also enables verbal overshadowing effects to be readily related to more standard memory interference results. For example, the verbally recoded representation could be viewed as a form of self-generated misinformation (e.g., E. F. Loftus, Miller, & Burns, 1978; J. W. Schooler, Gerhard, & Loftus, 1986), which interferes with the visual memory. Most importantly, the recoding interference hypothesis readily accounted for most of Schooler and Engstler-Schooler’s findings. First, it explained why verbalization interferes with visual but not verbal memories. A verbally biased representation is unlikely to provide a veridical description of a nonverbalizable stimulus but should maintain the critical information necessary for the recognition of a verbal stimulus. Second, it explained why visualization does not impair recognition of visual stimuli. Recollection within the same modality is more likely to be veridical and thus less likely to lead to an ill-matched representation. And, third, recoding interference helped to account for why limiting recognition time facilitates performance. With limited recognition time, subjects have less opportunity to access their verbally biased representation. Although the recoding interference hypothesis was effective in accounting for most of Schooler and Engstler-Schooler’s findings, there was one finding for which it was less successful. Specifically, if the negative effects of verbalization are a consequence of relying on a memory representation corresponding to the verbal recollection, then one might reasonably expect a relationship between the contents of subjects’ verbalizations and their recognition accuracy. In other words, the recoding interference hypothesis seems to suggest that more accurate verbalizations should lead to greater recognition accuracy than less accurate verbalizations. However, in several experiments, Schooler and Engstler-Schooler failed to find a relationship between subjects’ face recognition performance and the accuracy of their descriptions, as assessed by judges using an independently devised coding scheme. Schooler and Engstler-Schooler noted that this lack of a relationship was somewhat problematic for the recoding interference hypothesis,
At a Loss from Words
297
but suggested that it could be explained by assuming that the recoded representation includes both verbal and visual elements combined in an idiosyncratic manner. As a consequence, they suggested that “the retrieved recoded memory may neither resemble the original visual memory nor the subsequent verbalization” (1990, p. 65), thereby accounting for the absence of a relationship between the contents of subjects’ verbalization and their recognition performance. D. SUMMARY OF THE ORIGINAL PREMISES OF VERBAL OVERSHADOWING
In sum, Schooler and Engstler-Schooler’s examination of the disruptive effects of verbalization on perceptual memories introduced three general premises regarding the manner in which language may overshadow nonverbal cognition. The first premise, which we term the modality mismatch assumption, presumes that language specifically interferes with the application of the nonreportable aspects of perceptual memories. This assumption was supported both by stimulus differences (i.e.’ verbal rehearsal disrupts the recognition of several types of nonverbal stimuli, but not verbal stimuli) and by processing differences (i.e., verbal rehearsal impairs performance but visual rehearsal does not). A second premise, which we term the availability assumption, asserts that verbalization overshadows but does not eradicate the original visual memory. This assumption was supported by the finding that limiting recognition time attenuates the negative effects of verbalization, presumably by constraining the retrieval of verbal information. A third premise, termed the recoding interference hypothesis, assumes that the disruptive effects of verbalization are a consequence of retrieving a nonveridical, verbally biased representation generated during the process of verbalization. This interpretation accounted for all of the primary findings, although it did make one prediction that was not supported, that is, that the contents of subjects’ verbal descriptions would be predictive of their recognition accuracy. Since the publication of Schooler and Engstler-Schooler’s original series of studies, additional research, across a variety of domains of perceptual memory, has further investigated all three of the above claims. As will be seen in the following review, there is now substantially more evidence for both the modality mismatch and availability assumptions. The recoding interference hypothesis, however, has proven to be inadequate for accounting for some findings, and has thus required some supplementation. 11. THE MODALITY MISMATCH ASSUMPTION
The claim that the effects of verbalization are a result of a disparity between verbal and nonverbal knowledge has proven to be a rather powerful princi-
298
Jonathan W. Schooler et al.
ple that has led to a number of successful predictions. In this section we review the evidence for three distinct predictions of the modality mismatch assumption. (1) the generality of verbal overshadowing-if verbalization disrupts the application of nonverbal knowledge, then the effects of verbalization should generalize across domains that rely on nonverbal knowledge; (2) processing differences-if it is specifically the language component of verbal rehearsal that produces the interference, then the effects of rehearsal on nonverbal stimuli should be shown to specifically depend on whether or not verbal processes are engaged; (3) expertise differences-if verbalization specifically disrupts the application of nonverbal knowledge, then its effect should depend on individuals’ relative verbal and nonverbal expertise. Accordingly, individuals whose perceptual expertise markedly exceeds their verbal expertise should be vulnerable to verbalization. In contrast, individuals whose verbal and nonverbal expertise is more commensurate should be relatively unaffected by verbalization. We now consider the evidence in support of these various predictions of the modality mismatch assumption. OF VERBAL A. THEGENERALITY
OVERSHADOWING
The claim that language interferes with the application of nonverbal knowledge makes a rather strong prediction about the generality of verbal overshadowing effects, namely that they should apply across the domains of cognition known to rely on nonverbal knowledge. Although this claim naturally follows from the modality mismatch assumption, we have nevertheless been rather surprised to discover the extent to which verbal overshadowing effects apply across domains of perceptual memory, as well as to other areas of cognition known to rely on nonverbal knowledge. We briefly review these various domains.
I. Memory for Forms Memory for forms has long been known to be influenced by language. For example, Carmichael et al. (1932) demonstrated that verbal labels associated with nonverbal forms during encoding could bias the manner in which such forms were subsequently reproduced (see also Daniel, 1972; Riley, 1962). Memory for forms has also been shown to involve knowledge, which, at least for some items, is inherently nonverbalizable. For example, Attneave (1957) found that individuals were able to recognize hard-toname forms despite their inability to label them. Thus, according to the modality mismatch assumption, memory performance that relies on the nonreportable aspects of memory for visual forms should be vulnerable to verbalization. Recently, several studies by Brandimonte and colleagues (e.g., Brandimonte & Gerbino, 1993,1996; Brandimonte, Hitch, & Bishop,
At a Loss from Words
299
1992a, 1992b; Brandimonte, Schooler, & Gabbino, 1997) have provided support for this claim. The basic paradigm used by Brandimonte and colleagues involves having subjects study a set of visual forms like those illustrated in Fig. 2. During this study phase several manipulations have been used to moderate the degree to which individuals verbalize the forms. In some studies, verbalization is manipulated covertly by using either easyor hard-to-name forms, with the assumption being that subjects will be more likely to spontaneously verbalize easy-to-name as compared to hardto-name forms. In other studies, verbalization is overtly manipulated by either presenting or not presenting verbal labels with the forms during encoding. After subjects learn the forms under either verbal or nonverbal conditions, they are then given an imagery task that requires them to manipulate their veridical visual memories. For example, in a study by Brandimonte et al. (1992b), subjects were asked to mentally rotate the forms and determine the constituent letters contained in the rotated forms (see Fig. 2). The standard result from these studies is that individuals’ ability to successfully complete the memory-based imagery task critically depends on whether the prior encoding conditions encouraged or discouraged verbalization. Under conditions in which verbal encoding occurs (e.g., easy-to-name forms or labeled hard-to-name forms) visual imagery performance is impaired relative to conditions in which verbal encoding is not encouraged (e.g., unlabeled hard-to-name forms).
2. Memory for Macrospatial Relationships Macrospatial memory (i.e., memory for the relationship between spatial locations in the environment) has also been shown to involve distinct types of representations that vary in their reliance on verbal and nonverbal knowledge (e.g., route vs configural representations, Hirtle & Hudson, 1991; procedural vs survey knowledge, Siege1 & White, 1975; Thorndyke & Hayes-Roth, 1982). According to the modality mismatch assumption, the existence of such dichotomies suggests that certain aspects of a macrospatial memory should be susceptible to verbalization, while others may be invulnerable. Consistent with this prediction, Fiore and Schooler (1997) found that verbalizing one’s memory for the route on a map impaired later performance on a measure of the configural aspects of that map (Euclidean distance estimations), while having no effect on a measure of the more verbalizable featural aspects (route distance estimations).
3. Memory for Taste Taste is another domain of perceptual memory that is known to transcend linguistic depiction, as revealed by common allusions to the “indescribabil-
300
Jonathan W. Schooler et al. Easy nameability
Difficult nameability
mask
door
lamp
window
ball
i
Q I chair
bridge
music stand
shell
drop
roof
Fig. 2. Easy and difficult nameability stimuli. Originally used in Brandimonte et al. (1992b). Copyright 1992 by the American Psychological Association. Reprinted with permission.
At a Loss from Words
301
ity” of tastes. Although surprisingly little research has been conducted on taste memory per se, olfaction is a fundamental component of the taste experience (Lawless, 1985) and considerable research has examined the relationship between language and olfaction memory. This literature indicates that, as with other perceptual experiences, memory for smell is frequently influenced by language (e.g., Engen & Ross, 1973). Nevertheless, there are many cases in which individuals’ memory for smells exceeds their ability to describe them, that is, they can recognize previously encountered scents that they were unable to label (Lawless & Eagen, 1977). This nonverbalizability of olfactory memory, and by extension taste memory, suggests that it too should be vulnerable to verbalization. Consistent with this prediction, Melcher and Schooler (1996) found evidence that describing the taste of a wine can interfere with with its subsequent recognition. 4. Memory for Audition
As with other perceptual domains, auditory memory is known to be influenced by, yet often exceed, linguistic expression. For example, although memory for musical segments can involve both verbal and nonverbal codes (e.g., Krumhansl, 1991; Samson & Zatorre, 1991), novice listeners have been shown to be unable to make use of the verbal code (e.g., Makumo, 1992; Zatorre & Beckett, 1989). Thus, according to the modality mismatch assumption, memory for music also should be vulnerable to verbalization. Consistent with this prediction, Houser, Fiore, and Schooler (1997) found that verbalizing a previously heard musical segment significantly impaired subjects’ ability to distinguish it from similar distractors. 5.
Other Domains of Nonverbal Cognition
If verbalization interferes with application of nonverbalizable knowledge, then, in principle, the negative effects of verbalization should extend beyond perceptual memory. In further support of the modality mismatch assumption, we, and others, have now found evidence of negative effects of verbalization across several other domains that rely on nonverbal processes. a. Affective Decision Making Knowledge of the basis of one’s judgments is a type of cognition that is notoriously difficult to verbalize (e.g., Nisbett & Wilson, 1977). At the same time, affective judgments are known to be influenced by language (e.g., Cooper & Fazio, 1984). Thus, affective judgments represent another domain of cognition that, according to the modality mismatch assumption, might be vulnerable to verbalization. Consistent with this prediction, Wilson and Schooler (1991) found that, relative to nonverbalizing controls, subjects who verbally analyzed the reasons for
302
Jonathan W. Schooler et al.
affective judgments (e.g., taste quality of strawberry jams) made decisions that were less in line with the opinions of experts. Wilson et al. (1993) further found that verbally analyzing reasons causes individuals to make decisions that result in less postchoice satisfaction (see Wilson & Dunn, 1986; Wilson, Dunn, Kraft, & Lisle, 1989; Wilson & Lafleur, 1995, for additional research on the impact of verbally analyzing reasons for affective decisions). b. Insight Problem Solving Although problem solving is generally characterized as readily lending itself to verbal exposition (Ericsson & Simon, 1980, 1984), insight problem solving (“Aha!”-type problems that require finding alternative ways of conceptualizing the problem) have frequently been claimed to involve nonreportable cognitive processes (Ohlsson, 1992). However, like the other nonverbalizable domains reviewed here, insight problem solving is also known to be influenced by language (e.g., Glucksberg & Danks, 1968). Thus, insight problem solving represents yet another domain that, according to the modality mismatch assumption, may be disrupted by verbalization. Consistent with this prediction, J. W. Schooler, Ohlsson, and Brooks (1993) found that both retrospective and concurrent verbalization impaired individuals’ ability to find solutions to insight-type problems. In contrast, logical problem solving, which relies on more verbalizable knowledge (cf. J. W. Schooler & Melcher, 1995), was found to be unaffected by verbalization.
6. Domains Not Affected by Verbalization If verbalization specifically disrupts the application of nonverbalizable knowledge, then tasks that rely on more verbalizable knowledge should be relatively invulnerable to verbalization. We have already noted a number of situations in which tasks that rely on verbalizable knowledge were found to be relatively invulnerable to verbalization including: word recognition (e.g., Darley & Glass, 1975; Glenberg & Adams, 1978; Maki & Schuler, 1980); statement recognition (J. W. Schooler & Engstler-Schooler, 1990); route distance estimation (Fiore & Schooler, 1997); and logical problem solving (J. W. Schooler et al., 1993; see also Gagne & Smith, 1962). In addition, there are also a number of other domains that rely on verbalizable knowledge for which verbalization has been found to be, at a minimum, benign and often helpful, including the learning of declarative knowledge (Chi, Le Leeuw, Chiu, & LaVancher, 1994) and medical decision making (Henry, LeBreck, & Hozemer, 1989). B. PROCESSING DIFFERENCES In addition to predicting the domains in which verbalization is likely to be disruptive, a second implication of the modality mismatch assumption is
At a Loss from Words
303
that the effects of verbal rehearsal should critically depend on whether or not language processes per se are employed. As noted, Schooler and Engstler-Schooler provided some initial evidence for this prediction, showing that, in contrast to verbal rehearsal, visual rehearsal did not impair memory for either faces or colors. Since this initial demonstration additional investigations of the impact of manipulating the use of language during rehearsal have further implicated the critical role of verbalization in disrupting perceptual memory performance, thereby further supporting the modality mismatch assumption.
1. The Effects of Visualization on Map Memory One possible concern with the visualization manipulations used by Schooler and Engstler-Schooler is that there was no manipulation check to insure that subjects were in fact engaging in visual recall. It is possible that subjects, knowing that their performance could not be monitored, were not adequately engaged in the visual recollection. In a recent map memory study, Fiore (1994) addressed this possible concern by giving subjects a visual scanning task (cf. Kosslyn, Ball, & Reiser, 1978) that directly monitored visualization performance. Subjects studied a map of a small town and then either verbalized their memory for the route shown on the map, engaged in an unrelated verbal activity (describing memory lapses they have experienced), or visualized moving a black dot from one landmark on the map to another. In the mental scanning condition, subjects were presented with pairs of landmarks and were to imagine scanning from one landmark to the other, pressing a button when they “reached” each new destination. As shown in Fig. 3, in this study verbalization significantly hindered subjects’ performance on a later map-drawing task compared to both the visualization and control conditions. This study illustrates that even when subjects are given a demanding visualization task that requires compliance with the visual recall instructions, subjects still fail to show a detriment of visual recall. Thus, this study provides further evidence that it is specifically the verbal aspect of verbal recall that produces interference.
2. The Effects of Verbal Suppression on Form Memory Additional evidence that verbalizing nonverbal knowledge is the critical source of interference in verbal overshadowing effects comes from an examination of the effects of a manipulation known to minimize verbal processing. Verbal suppression (e.g., repeating the phrase “lala” while encoding a stimulus) is a well-known technique for reducing the extent of spontaneous verbal rehearsal (Murray, 1967). Thus, according to the modality mismatch assumption, engaging in verbal suppression during the learning of nonverbal
304
Jonathan W. Schooler et al.
Condition Fig. 3. Difference scores between actual and reproduced maps. From Fiore (1994).
stimuli should attenuate verbal processing and thereby attenuate the effects of verbalization. To investigate this issue, Brandimonte et al. (1992b) had subjects study either easy or hard-to-name forms (see Fig. 2), under either standard encoding conditions or articulatory suppression (repeating the phrase “lala”). After learning the various forms, subjects were given the mental rotation task described earlier. As can be seen in Fig. 4, for subjects who did not engage in articulatory suppression, a verbal overshadowing effect was observed, such that imagery performance was worse for easy-to-name forms relative to hard-to-name forms. In contrast, under verbal suppression conditions, performance on the easy- and hard-to-name forms was comparably high. These findings suggest that preventing verbal processing through verbal suppression reduces the spontaneous verbal labeling of visual stimili, and thereby prevents verbal overshadowing of nonverbal stimuli.
C. EXPERTISE DIFFERENCES A third general prediction of the modality mismatch assumption is that relative differences in verbal versus nonverbal expertise should mediate
At a Loss from Words
305
1.0 7
0
Articulatory suppression No Articulatory suppression
-
0.8
. I -
0
0.6L
0 0 w
c
a,
2 a, a
0.4-
0.2-
0.0I Easy nameability
Difficult nameability
Stimulus type Fig. 4. Effect of articulatory suppression on easy and difficult nameability stimuli. From Brandimonte et al. (1992b). Copyright 1992 by the American Psychological Association. Reprinted with permission.
verbal overshadowing effects. Accordingly, if verbalization specifically interferes with the application of nonverbal knowledge, then the magnitude of such effects should critically depend on individuals’relative verbal versus nonverbal expertise. When nonverbal expertise markedly exceeds verbal expertise, verbal overshadowing should be observed; however, when the two types of knowledge are more commensurate, verbal overshadowing should be avoided. This interaction between verbal overshadowing and expertise has been observed in several domains. 1. Own-versus Other-Race Face Recognition Generally speaking, face recognition represents a classic example of a situation in which perceptual expertise exceeds verbal expertise. Individuals are experts at the nonverbal task of recognizing faces, although most are quite unskilled at the verbal task of describing faces (e.g., Ellis et al., 1980). The magnitude of this discrepancy, however, depends on individuals’degree
306
Jonathan W. Schooler et al.
of familiarity with different types of faces. Individuals are typically better at recognizing members of their own race, as compared to members of other races (e.g., Brigham & Malpass, 1985; Rhodes, Tan, Brake, & Taylor, 1989). This expertise specifically involves an increased sensitivity to the configural properties of a face (i.e,, the interrelationship between features, Rhodes et al., 1989), which are particularly difficult to verbalize (e.g., Wells & Turtle, 1987). If the negative effects of verbalization specifically pertain to nonverbal expertise, and if own-race face recognition particularly relies on such expertise, then own-race face recognition should be especially vulnerable to verbalization. To examine the relationship between verbalization and perceptual expertise, Fallshore and Schooler (1995) replicated the standard verbal overshadowing paradigm using both same- and other-race faces. Caucasian subjects' were presented with Caucasian and African-American stimulus faces and later were asked either to verbally describe the faces or to perform an unrelated, interpolated task. Finally, all subjects were presented with a forced-choice recognition test including the target and similar distractors. As can be seen in Fig. 5 , Fallshore and Schooler replicated the standard verbal overshadowing effect for own-race faces. However, for other-race faces, subjects' performance was completely unaffected by verbalization. This interaction between verbalization and race of face can be readily accounted for by the view that the particular expertise associated with recognizing own-race faces is uniquely nonverbal in nature. Along these lines, it is worth noting that in the verbalization conditions, the generally observed own-race face advantage (seen in the control conditions) was entirely eliminated. Indeed, if anything, in the verbalization condition there was a trend for superior performance for other-race faces. The notion that the interaction between verbalization and race of face was specifically due to the greater role of nonverbal knowledge in ownrace face recognition suggests several predictions that were tested in a second study. First, if the expertise that differentiates own- from otherrace face recognition is primarily perceptual in nature, then the descriptions of these two types of faces should be relatively comparable in quality. Second, if recognition of other-race faces relies to a greater degree on more I Ideally it would have been quite informative to use non-Caucasian subjects as well. However, African-American populations in the United States, and particularly those attending predominantly Caucasian college campuses, have experience with own- and other-race faces that is quite different from that of Caucasian subjects. The fact that African-American subjects' experience with other-race faces does not mirror that of Caucasian subjects indicates that African-American subjects would not have served as the appropriate comparison group in this study. Rather, what would be needed is individuals of African descent who live in countries where they are the distinct majority. Unfortunately, inclusion of such a population was beyond the means of this study.
At a Loss from Words
307
0
1 .o
Verbalization Control
0.8 +-
0
2
z
0.6
0 c C
a,
2
2
0.4
0.2
0.0
Other race
Own race Race of face
Fig. 5. Percentage of correctly recognized own- and other-race faces. From Fallshore & Schooler (1995, Experiment 1). Copyright 1995 by the American Psychological Association. Reprinted with permission.
readily verbalized information, then the descriptions of other-race faces may be more predictive of subjects’ actual recognition performance than those associated with own-race faces. Fallshore and Schooler tested these two predictions using a communication accuracy paradigm (e.g., Lantz & Volney, 1964; Lucy & Shweder, 1979) wherein subject-judges were yoked with each of the verbalization subjects from Experiment 1. Each subjectjudge read the verbal description generated by their yoked verbalization subject counterpart and attempted to use the description to identify the target face from the recognition array. Comparison of subject-judges’ overall identification performance indicated that there was little difference between the original verbalization subjects’ ability in describing own- and other-race faces. Subject-judges’ ability to use verbalization subjects’ descriptions to identify the target face was actually numerically, though not significantly, greater for other- versus own-race faces (with mean identifica-
308
Jonathan W. Schooler et al.
tion rates of 32.1 and 26.8%, respectively, with chance equaling 16.7%). This finding suggests that, although individuals’ recognition performance tends to be better for own- versus other-race faces, their ability to describe the two types of faces is quite comparable. In other words, as predicted, the increase in expertise associated with own-race face recognition appears to be primarily nonverbal in nature. If, as argued, face recognition expertise is associated with an increased reliance on nonreportable information, then subjects’ verbal descriptions of own-race faces should be less predictive of their recognition performance than their descriptions of other-race faces. To explore this issue, Fallshore and Schooler examined the relationship between the identification performance of the subject-judges and the recognition accuracy of the subjects who generated those descriptions. Interestingly, the correlation between the identification accuracy of the subject-judges and the recognition accuracy of their yoked verbalization subject counterparts was significant for otherrace faces (r = .36), but no such relationship for own-race faces was found (r = .12). The fact that verbal descriptions were predictive only of otherrace recognition thus provides further evidence that other-race face recognition is distinguished from own-race recognition with respect to its greater reliance on verbal knowledge. Thus, the absence of a verbal overshadowing effect for other-race faces can be seen as yet another source of evidence for the hypothesis that verbal overshadowing specifically involves the verbal disruption of nonverbal knowledge.
2. Wine Expertise The premise that verbal overshadowing occurs when nonverbal expertise exceeds verbal expertise suggests that we should be able to document such a relationship between expertise and verbalization in other domains besides face recognition. Moreover, if verbal expertise is truly a critical component of this phenomenon, then we should find that the effects of verbalization depend not only on individuals’ nonverbal expertise in a given domain, but also on how fluent they are in articulating that knowledge. The challenge, therefore, is to find a domain in which individuals can excel in perceptual abilities both with and without commensurate verbal ability; wine tasting, it turns out, is just such a domain. Many wine drinkers develop a palate for wine such that they can distinguish between many fine wines. Despite possessing a developed perceptual palate, novice wine drinkers, while perhaps being familiar with a few technical wine terms, do not really know how to describe wines with much precision. In contrast, wine professionals and those who have participated in extensive wine tasting training do, with time, develop a vocabulary that
At a Loss from Words
309
enables them to significantly exceed their novice counterparts in describing wines (e.g., Solomon, 1990). Thus, wine tasting skill provides us with three theoretical levels of expertise in which to examine the hypothesized relationship between expertise and verbalization: non-wine drinkers who possess minimal perceptual or verbal expertise; untrained wine drinkers who have some perceptual expertise but minimal verbal expertise; and trained experts who have both perceptual and verbal expertise. According to the present approach, these three populations should be differentially affected by verbalization. Non-wine drinkers, like individuals verbalizing other-race faces, should show minimal effects of verbalization because their perceptual expertise may fail to markedly exceed their verbal expertise. In contrast, untrained wine drinkers, like individuals recognizing own-race faces, may show a substantial effect of verbalization because they have developed a degree of perceptual expertise (a palate) but lack the vocabulary to express their knowledge. Finally, wine experts possess both perceptual and verbal expertise and consequently should exhibit strong performance regardless of verbalization. In a recent study, Melcher and Schooler (1996) examined the impact of describing a previously tasted wine on three such populations: non-wine drinkers (individuals who drank less than once a month); untrained wine drinkers (those who drank more than once a month, but had little or no formal training); and experts (individuals who were either in the wine profession or had taken multiple wine courses). Subjects tasted a wine, either described it or not, and then rated four wines (the target and three distractors) for how closely they matched the taste of the target wine. Performance was gauged by taking the difference between the rating given to the target wine and the mean given to the three distractors. As can be seen in Fig. 6, subjects’ discrimination performance largely supported the predictions outlined earlier. Neither the nondrinkers nor the trained experts were impaired by verbalization. In fact, if anything, both groups tended to improve following verbalization. In contrast, the untrained wine drinkers showed a marked decline in performance following verbalization. Our interpretation of these findings is that verbalization reduced subjects’ ability to draw on their perceptual expertise, and thus primarily impacted that population of subjects for whom their verbal and perceptual expertise was least commensurate. In further support of this interpretation, Melcher and Schooler examined the correlation between subjects’ performance in the two conditions and their scores on independent measures of verbal and perceptual expertise. Verbal expertise was gauged by subjects’ responses to a wine knowledge questionnaire. Perceptual expertise was determined by how often subjects reported drinking red wines. In the nonverbalization condition, perceptual expertise was the best predictor of discrimination
Jonathan W. Schooler et al.
310
0 Verbalization Control
4
n
3
C
.-0 .-2 .-E
2
t v)
b
1
C
m
r" 0 U
-1
Nondrinkers
Nonexpert drinkers
Expert drinkers
Fig. 6. Mean discrimination score for wine drinkers of differing levels of expertise. From Melcher & Schooler (1996). Copyright 1996 by Academic Press, Inc.
performance, suggesting that when subjects do not verbalize they tend to rely on their perceptual experience. In contrast, in the verbalization condition, verbal expertise was the best predictor of performance, suggesting that engaging in verbalization forces subjects to rely on their verbal knowledge. OF EVIDENCE FOR THE MODALITY D. SUMMARY ASSUMPTION MISMATCH
In sum, three distinct strands of evidence converge on the general claim that verbalization uniquely disrupts nonreportable knowledge: (1)the disruptive effects of verbalization have been found to generalize across a surprisingly wide array of domains that rely on nonverbal knowledge; (2) manipulations that minimize verbal processing prevent the disruptive effects of verbalization; and (3) the effects of verbalization are limited to situations in which nonverbal expertise exceeds verbal expertise. 111. The Availability Assumption
A long-standing issue in investigations of memory interference is what happens to the original memory. This issue dates back to classic "A :B
At a Loss from Words
31 1
A : C”list-learning studies. For example, Melton and Irwin (1940) suggested that encountering new associations extinguishes the earlier associations making them irrecoverable. In contrast, McGeoch (1942) argued that both associations existed in memory and that impairment of the original association was a consequence of response competition. More recently, this debate has played itself out in the context of accounting for the negative effect of misinformation on memory. For example, like Melton and Irwin (1940), E. F. Loftus and Loftus (1980) argued that the original memory is lost following exposure to misleading postevent information. In contrast to this “destructive updating” account, others offered a view comparable to that of McGeoch, arguing that the original memory and the postevent memory coexist but compete or are confused at the time of retrieval (e.g., Bekerian & Bowers, 1983; Christinaasen & Ochalek, 1983; Lindsay & Johnson, 1989; Zaragoza & Lane, 1994). As mentioned at the outset, a central assumption of the verbal overshadowing framework is that verbalization overshadows but does not eradicate the original visual memory. In support of this claim Schooler and EngstlerSchooler observed that limiting subjects’ recognition time attenuated the negative effects of verbalization, presumably by reducing access to the more slowly retrieved verbal knowledge.* Several studies have provided further support for the availability assumption and the general claim that verbalization overshadows but does not eradicate the original memory. We briefly review these more recent studies. A. FACERE-PRESENTATION
If verbalization overshadows but does not eradicate the original visual memory, then manipulations that reinstate the original visual memory should reverse the negative effects of verbalization. Recently J. W. Schooler, Ryan, and Reder (1996) described a study that provided evidence that reinstating the original visual memory can eradicate the negative effects of verbalization. Specifically, Schooler et al. examined the effect of representing the target face following verbalization. Subjects viewed a tar-
’
The limited response time manipulation appears to be a somewhat delicate intervention. On one hand. limiting response time has also been found to attenuate the effects of verbalization in the related domain of attitude judgments (T. Y. Schooler, 1990; T. D. Wilson, personal communication, October, 1996).On the other hand, using stimulus/test face recognition materials different from those used by J. W. Schooler and Engstler-Schooler (1990). several studies have observed negative effects of verbalization even in the limited response time condition (Dodson, Johnson, & Schooler, 1997; Read & Schooler, 1994). It seems likely that the impact of limiting response time may depend on the specific stimulus/test materials that are used. Such materials undoubtedly influence the test inspection time necessary for successful recognition, and may thereby determine the critical duration at which limiting response time is most apt to be beneficial. Future research might profitably investigate this issue.
Jonathan W. Schooler et al.
312
get face, and then either described it or engaged in an unrelated filler activity. After the control or verbalization activities, subjects assigned to the re-presentation condition were shown the target photo again. Finally, all subjects were given the recognition array that included a different photo of the target face and five similar distractors. If verbalization eradicates the original visual memory, then one would expect that, even after re-presentation of the target face, subjects in the verbalization condition should show poorer performance than subjects in the control condition. Alternatively, if verbalization simply makes the visual memory less accessible, then reinstating it through re-presentation should make the visual memory more accessible, thereby attenuating the negative effects of verbalization. As can be seen in Fig. 7, our prediction was generally observed, but with a twist. To our surprise, re-presentation not only eliminated the verbalization effect, it reversed it! This reversal in the effects of verbalization following re-presentation was somewhat unexpected, although subsequent studies proved it to be reliable. We defer our present account for this reversal until the final section of this chapter. Suffice it to say, for the present purposes, verbalization subjects’ superior performance in the representation condition strongly suggests that their original memories remained intact, thereby enabling them to show such substantial gains following re-presentation. 1.0-
0 Verbalization H Control
r
0.8c
0
E 0 c
C
No re-presentation
Re-presentation
Fig. 7. Effects of verbalization and re-presentation on recognition performance. From Schooler, Ryan, & Reder, 1996. Copyright 1995 by Lawrence Erlbaum Associates. Reprinted with permission.
At a Loss born Words
313
B. CONTEXT REINSTATEMENT OF MEMORY FOR FORMS The issue of the availability of the original visual memory is of particular pertinence to the verbal overshadowing paradigm as it has been applied to memory for forms. As described earlier, Brandimonte and colleagues have reported numerous demonstrations in which either implicit or explicit verbal labeling of forms during encoding hampers subsequent imagery performance. Their interpretation of this finding is similar to that of other verbal overshadowing experiments in which the verbalization is completed after the stimulus is no longer in view; namely that verbalized information interferes with the successful retrieval of potentially available visual information. However, because the verbalization in the Brandimonte et al. studies (1992a, 1992b) occurs at the time of encoding, rather than postencoding (as it does in the other perceptual memory paradigms described in this chapter), there is an alternative account of their findings. It is possible that rather than interfering with the retrieval of visual information, verbalization may interfere with the encoding of such information. Indeed, an encoding disruption account was suggested by several prior researchers who found evidence that verbal processing during encoding interferes with visual memories (e.g., Bahrick & Boucher, 1969; Nelson & Brooks, 1973; Pezdek et al., 1986). For example, Nelson and Brooks suggested “forced involvement of the verbal system may have reduced the time available for coding the superior pictorial representation” (p. 48). One important prediction of an encoding disruption account is that any reduction in visual memory performance following verbalization should be irreversible. If the information never made it in to memory, then it cannot be expected to ever be retrieved from memory. In short, an encoding disruption account of verbalization effects makes a very unambiguous prediction regarding the fate of the original visual information; it should be unavailable under any retrieval conditions. In contrast, if, as we have hypothesized, the locus of verbalization effects is at retrieval, then in principle, given the appropriate retrieval conditions, verbal overshadowing effects should be attenuated. Recently, Brandimonte, Schooler, and Gabbino (1997) conducted several experiments to assess the availability of the original visual memory within the form memory paradigm. Brandimonte et al. investigated the effects of introducing, at the time of retrieval, visual cues that were present during encoding. The logic of this manipulation was that re-presentation of the visual cue during retrieval should increase the likelihood of accessing any intact visual memory. Accordingly, if verbalization interferes with the retrieval of intact visual information, then emphasizing visual components of the memory during retrieval may attenuate the effects of verbalization.
Jonathan W. Schooler et al.
314
If, however, verbalization reduces the quality of the initial encoding of visual information, then visual retrieval cues should be ineffective. In one study, subjects learned either easy- or hard-to-name line drawings (see Fig. 2), which were drawn on colored cards (the assumption being, as mentioned earlier, that easy-to-name drawings are more likely to be spontaneously verbalized than hard-to-name drawings). Subjects were then asked to complete the mental imagery task, which required them to recall detailed visual characteristics of the stimuli in order to identify the letters hidden in the figure. In each condition, just before performing the imagery task, half the subjects were re-presented with the color of the card on which each picture was originally viewed and half performed the task without exposure to the color cues. As can be seen in Fig. 8, the results provided evidence that the original visual information does in fact remain potentially available following verbal0.8
0
Cue Nocue
0.6
0.4
0.2
0.0 Easy nameability
Difficult nameability
Stimulus type Fig. 8. Percentage of letters correctly identified in mental imagery task in the Cue and No-cue conditions as a function of nameability. From Brandimonte, Schooler, & Gabbino (1997, Experiment 1). Copyright 1997 by the American Psychological Association. Reprinted with permission.
At a Loss from Words
315
ization. As in prior studies, (e.g., Brandimonte et al., 1992a, 1992b) under standard retrieval conditions, subjects who encoded easy-to-name figures had greater difficulty subsequently deciphering the letters in their images relative to subjects who encoded hard-to-name figures. On the assumption that easy-to-name figures are more likely to be spontaneously verbalized than hard-to-name figures, this result suggests that verbalization during encoding hampered subsequent memory performance in the no-cue condition. However, when a color cue was presented prior to the imagery task, the difference between imagery performance with easy- versus hard-toname pictures was attenuated. This latter finding suggests that introducing retrieval conditions that favor access to the visual information enables subjects to retrieve their intact visual memories. A second study further demonstrated the availability of visual form memory following verbalization. This study was similar to the first, except the method of inducing verbalization was changed. Rather than covertly inducing verbalization through the use of easy-to-name forms, verbalization was overtly induced by supplying verbal labels to hard-to-name forms. As can be seen in Fig. 9, the results of this second experiment paralleled those of the first. As predicted, in the no-cue condition, overt verbalization impaired imagery performance. However, when subjects were supplied with visual cues, the impaired imagery performance associated with labeled forms was attenuated. These findings thus provide further evidence that verbalization does not interfere with either the formation or the storage of visual memory. Rather, the effect of verbalization, even when introduced at encoding, is to overshadow an intact visual representation.
IV. The Recoding Interference Hypothesis As mentioned at the outset, J. W. Schooler and Engstler-Schooler (1990) suggested a recoding interference account for why verbalization interfered with access to nonverbal memory. It was hypothesized that verbalization of a nonverbal memory results in the formation of a nonveridical, verbally biased representation that is accessed instead of the original visual memory. This notion of recoding interference is consistent with many theories of memory interference based on competition between distinct memory representations (e.g.,Johnson, Hashtroudi, & Lindsay, 1993; Morton, Hammersley, & Bekerian, 1985), and also accounted for most of Schooler and Engstler-Schooler’s original findings. However, as it has turned out, the results of a number of studies suggest that, at least in the case of memory for faces, the recoding interference hypothesis cannot account for the detrimental effects of verbalization. We review the various sources of difficulty
Jonathan W. Schooler et al.
316 0.8
[7 Label No label
0.6
c
0
2
L
0
0.4
0 4-
c a, $2 a,
a
0.2
0.0 No cue
Cue
Test condition Fig. 9. Percentage of letters correctly identified in mental imagery task in the Cue and No-cue conditions as a function of presence of labels. From Brandimonte, Schooler, & Gabbino (1997, Experiment 2). Copyright 1997 by the American Psychological Association. Reprinted with permission.
for the recoding interference hypothesis and then consider several potential alternatives. A. THERELATION BETWEEN VERBALIZATION RECOGNITION PERFORMANCE
CONTENT AND
As previously noted, there was one thorn in the side of Schooler and Engstler-Schooler’s original recoding interference interpretation of verbal overshadowing-the absence of a relationship between the contents of subjects’ verbal descriptions and their subsequent performance. If the negative effects of verbalization are the consequence of subjects’ reliance on a nonveridical memory representation corresponding to their memory for the face description, then one might expect that the quality of the descriptions would be predictive of recognition performance. One possible explanation for why Schooler and Engstler-Schooler failed to find such a relation-
At a Loss from Words
317
ship is that they used a potentially weak measure of description accuracy: coding the number of correct and incorrect features in each description. A limitation of coder-ratings of verbal descriptions is that such ratings fail to take into account the possibility that some features may be more discriminating than others. A potentially superior technique is the communication accuracy approach (e.g., Lantz & Volney, 1964; Lucy & Schweder, 1979), mentioned earlier, in which the verbal descriptions generated by subjects are given to subject-judges who must identify the target on the basis of the description alone. This technique has the advantage of clearly determining the degree to which reliance on the verbal description alone is sufficient for making a correct identification. However, as noted, Fallshore and Schooler (1995) observed that even with this more sensitive measure, there was no relationship between the contents of subjects’ verbal descriptions and recognition performance. Moreover, this lack of a relationship cannot be attributed to a lack of sensitivity of the communication accuracy paradigm, because Fallshore and Schooler did find a communication accuracy correlation for other-race faces. However, these faces were not associated with negative effects of verbalization. In this context it is worth noting that Schooler and Engstler-Schooler also found a relationship between verbal description quality and recognition performance only for the stimuli for which they did not find a negative effect of verbalization: verbal statements. Thus, it appears that a relationship between description quality and recognition performance is observed only for relatively verbalizable stimuli, which are precisely the stimuli that are invulnerable to verbalization. In contrast, memory stimuli that are difficult to translate into words are both vulnerable to verbalization and associated with an absence of a correspondence between verbalization performance and recognition performance. Although this pattern of findings is quite consistent with the modality mismatch assumption, it is somewhat harder to reconcile with the claim that verbal overshadowing effects result from an inappropriate reliance on memory representations corresponding to the verbal description.
B. THEEFFECTS OF WARNING SUBJECTS Another prediction of the recoding interference hypothesis is that the negative effects of verbalization may be attenuated if subjects are warned to ignore their memory for how they described the face. Warnings of this sort have been found to be quite effective in other paradigms involving source confusions between distinct memory representations. For example, a number of researchers have found that the negative effects of verbal postevent misinformation on visual memory can be attenuated if subjects are explicitly instructed to distinguish between what they originally saw
318
Jonathan W. Schooler et al.
and what they later read (e.g., Lindsay & Johnson, 1989; Zaragoza & Koshmider, 1989;Zaragoza & Lane, 1994;see also Christaansen & Ochalek, 1983). Given the effectiveness of such source monitoring instructions in the context of misinformation studies, it follows that if the negative effects of verbalization involve the equivalent of a self-generated misinformation effect (as the recoding interference hypothesis suggests), then source monitoring instructions might be comparably effective in reversing verbal overshadowing effects as well. To test this hypothesis, Dodson et al. (1997) conducted a verbal overshadowing study in which they explicitly warned subjects that “In completing the following task you should ignore your memory for how you described the face and only rely on your memory for seeing the face” (p. 5 ) . Strikingly, Dodson et al. found that these instructions did not at all reduce the magnitude of the verbal overshadowing effect; in fact, if anything, warning subjects to ignore their verbalizations increased the negative effects of verbalization.
C . THEEFFECTS OF VERBALIZATION ON NONVERBALIZED STIMULI Still further evidence against the original formulation of the recoding interference hypothesis comes from studies investigating the impact of verbalization on nonverbalized stimuli. If verbal overshadowing effects result from retrieving a memory representation corresponding to the process of describing the stimulus, then one would expect that the effects of verbalization would be limited to the stimulus that was described. However, in several experiments Dodson et al. demonstrated that describing one face can actually interfere with the recognition of a different face. For example, in one experiment subjects viewed two faces, one male and one female. Subjects were then instructed to verbalize either the male face, the female face, or to engage in an unrelated verbal activity. Remarkably, Dodson et al. found that describing a nontarget face produced impairment comparable to that associated with describing the target face. In another experiment, Dodson et al. found that verbally recalling the appearance of a parent’s face impaired recognition of a previously seen face. These findings strongly argue against the premise of the recoding interference hypothesis that the disruptive effects of verbalization are a consequence of subjects’ specific reliance on their memory of the verbal description.
V. How Does Verbalization Disrupt Perceptual Memories? The findings reviewed thus far provide strong support for the claim that verbalization specifically disrupts the application of nonverbal knowledge
At a Loss from Words
319
(the modality mismatch assumption) and the premise that this interference is not permanent, but can be attenuated if conditions encourage retrieval of nonverbal information (the availability assumption). However, contrary to our initial conceptualizations, the disruption associated with verbalization does not, at least in the case of face recognition, appear to involve a reliance on a recoded memory representation corresponding to the verbalization process. The question thus arises: How does verbalization disrupt the use of nonverbal knowledge? A. DOESVERBALIZATION CAUSEMEMORY INTERFERENCE OR A STRATEGY SHIFT? The finding that verbalization of one face can interfere with the recognition of a different face raises the possibility that the disruptive effects of verbalization may not involve memory interference at all, but may rather reflect a shift in subjects’ recognition strategies. Accordingly, verbalization may produce a general predisposition to favor verbal knowledge over nonverbalizable knowledge. To the degree that such verbalizable knowledge is nondiscriminating, such a shift in recognition strategies could lead to impaired performance. This verbal strategy explanation could readily account for why verbalization exclusively disrupts nonverbal processses in that excessive consideration of verbal knowledge would be detrimental only for nonverbalizable stimuli. The verbal strategy explanation could also account for the reversibility of verbal overshadowing effects with the assumption that presenting retrieval conditions that favor perceptual information attenuates the bias toward verbal information. A verbal strategy account could also explain why verbalization of one face might interfere with the recognition of a different face by simply assuming that verbalization of any nonverbal stimulus may be sufficient to induce the adoption of verbal recognition strategies. The resolution of this issue requires the indentification of a measure that can specifically assess the type of knowledge that individuals are employing in making their recognition decisions. There have been several major advances in the use of self-report measures to assess the degree to which subjects’ recognition strategies rely on verbalizable versus nonverbalizable knowledge. Tulving (1985) introduced the knowhemember distinction that uses a self-report measure to differentiate between “remember” recognition judgments based on specific episodic cues and “know” recognition judgments that are not based on awareness of any specific cues. Since Tulving’s original introduction of the know/ remember distinction, a number of studies have demonstrated that this distinction interacts with a variety of variables in a manner quite consistent with other memory measures known to distinguish between the reliance
320
Jonathan W. Schooler et al.
on explicit (reportable) versus implicit (nonreportable) knowledge (for reviews, see Gardiner & Java, 1993; Rajaram & Roediger, 1997). A similar self-report measure was introduced by Dunning and Stern (1994) in the context of a multiple-choice face recognition paradigm. In this paradigm, reliance on nonreportable versus reportable knowledge was determined by subjects’ agreement with statements such as “His face just popped out” versus “I compared the photos to each other in order to narrow the choices.” Consistent with the claim that face recognition critically relies on nonreportable knowledge, Dunning and Stem found that subjects’ reported reliance on nonverbalizable strategies was predictive of accuracy. Although the self-report measures used in the Know-Remember and Dunning and Stern (1994) paradigms have been applied in different settings (i.e., yes/no vs multiple choice recognition), invoke different verbalizable strategies (e.g., reliance on episodic context vs use of process of elimination), and vary with respect to their emphasis (i.e., the nature of the memory experience vs the recognition decision), they both suggest the value of using self-reports to distinguish between recognition decisions made on a verbalizable versus nonverbalizable basis. Indeed, the two approaches are strikingly similar with respect to their characterization of the use of nonreportable knowledge, with Know classification involving judgments of subjects who “are fully aware that the memory belongs to their personal past [but] they are unable to determine the basis of this conscious experience” (Rajaram & Roediger, 1997, p. 235) and automatic classifications corresponding to judgments in which subjects “just recognized him, I cannot explain why” (Dunning & Stern, 1994, p. 818). This similarity suggests that these two approaches may correspond to a more general distinction, which we term the Just Know/Reason distinction, between judgments that rely on a nonreportable versus reportable knowledge. 1. Verbalization, Face Recognition, and the Just
Know/Reason Distinction Recently we (J. W. Schooler, Fiore, Melcher, & Ambadar, 1996) developed and employed a self-report measure based on the Just Know/Reason distinction, in order to assess the impact of verbalization on the use of strategies involving verbalizable and nonverbalizable knowledge. In this paradigm subjects were first shown a video tape of a bank robbery including a target individual. Subjects were then introduced to the Just Know/Reason distinction in the context of a word learninghecognition task modeled after Gardiner (1988). Prior to engaging in recognition, subjects were instructed: Sometimes when you make a judgment, you are aware of specific reasons for that judgment, this is what we call a “reason” decision. Other times, your decision may be
At a Loss from Words
321
just based on a “gut” reaction without any specific reasons. This is what we call a “just know” decision.
After completing the word recognition task? subjects participated in either the verbalization or the control activities. Finally, subjects were given a recognition test and were asked to “assess how you just made this face recognition decision and decide, as you previously did with the words, whether it was a ‘just know’ or ‘reason’ decision.” Before reviewing the results, consider again the respective predictions of the verbal strategy and memory interference hypotheses. If verbalization causes subjects to adopt a verbally oriented recognition strategy, then it should alter their relative inclination to rely on verbal versus nonverbal knowledge, as would be reflected in the frequency with which they report making reason versus just know-based recognition decisions. In contrast, if verbalization produces some form of memory interference that influences the relative accessibility of verbal and nonverbal knowledge, then the effects of verbalization should be associated with changes in the relative accuracy of reason versus just know-based judgments. As can be seen in Fig. 10A, contrary to the predictions of the verbal strategy hypothesis, verbalization had no effect whatsoever on the frequency with which individuals reported relying on verbal and nonverbal knowledge. However, as can be seen in Fig. 10B, verbalization had a marked effect on the accuracy of just know versus reason judgments. Verbalization substantially impaired the accuracy of recognition decision judgments classified as just know, while having no effect whatsoever on decisions classified as reason. This latter finding suggests that the reduction in recognition performance resulting from verbalization is specifically the consequence of a reduced accessibility of nonverbalizable knowledge. Following verbalization, subjects continue to attempt to use nonverbalizable knowledge to make recognition judgments; they are simply less successful at doing so. Although we used instructions that were somewhat different (and much simpler) than those used by Tulving (1985) and Gardiner (1988) (that is, we made no mention of the construct of remembering nor did we give subjects any indication of the type of reasons on which they might have relied), we nevertheless found word recognition findings that were virtually identical to these earlier studies. The advantage of semantic encoding was exclusively limited to reason-based judgments. This interaction parallels that of other measures of explicit (reportable) and implicit (nonreportable) knowledge (i.e., semantic elaboration influences recall but not implicit priming, e.g., Graf, Mandler, & Haden, 1982; Jacoby & Dallas, 1981). and thus helps to validate that the Just Know/Reason instructions used here were in fact distinguishing between the use of reportable and nonreportable knowledge.
Jonathan W. Schooler et al.
Verbalization
-
0.6 0.5
-
0.4
-
0.3
-
Control
-
0.2 0.1
-
0.0 Verbalization
Control Condition
.8,
c)
0
2 b 0
.6,
I
C
a,
Ea,
.4-
Q
.2.
0 Just know
Reason Judgment
At a Loss from Words
2.
323
Verbalization, Voice Recognition, and the Just Know/Reason Distinction
The finding that verbalization does not alter subjects’ self-reported recognition strategies but rather influences the ability to rely on nonreportable knowledge, is of marked value in helping us to conceptualize verbal overshadowing of face recognition. However, to demonstrate its general applicability to verbal overshadowing effects, it is important to show that this characterization applies to the various other domains in which verbal overshadowing has been observed. Schooler, Fiore, Melcher, and Ambadar (1996) completed a second study indicating that the unique effects of verbalization on “just know” recognition judgments extend to at least one other domain: voice recognition. In this study, subjects listened to a brief audio recording of a spoken statement. After either describing the speakers voice or engaging in an unrelated activity, subjects were given a voice recognition test consisting of the same statement heard before, spoken by the person heard at encoding and by three similarly sounding foil voices. As Fig. 11A and 11B illustrate, this study revealed the same relationship between verbalization and just knowheason judgments as was observed with faces. Verbalization impaired just know judgments, without influencing either the accuracy or the frequency of reason judgments, thus once again demonstrating that verbalization does not alter subjects’ self-reported recognition strategies, but rather reduces their access to nonverbalizable knowledge. B. TRANSFER-INAPPROPRIATE RETRIEVAL
The unique effects of verbalization on just know recognition decisions suggest that the primary effect of verbalization is to disrupt individuals’ ability to apply nonverbal knowledge. However, contrary to Schooler and Engstler-Schooler’s original recoding interference hypothesis, at least in the case of face recognition, this disruption does not appear to involve the inappropriate emphasis on a recoded memory representation corresponding to the verbalization activity;rather, it appears to involve a more general form of interference. In reconceptualizing the nature of the interference that results from verbalization it may be useful to briefly revisit Schooler and Engstler-Schooler’s original account. According to the original verbal overshadowing theory: “visual memory interference following verbal processing may occur as a result of the interac-
Fig. 10. (A) Frequency of “Just Know” judgments by condition for face recognition task. (B) Accuracy of “Just Know/Reason” judgments by condition for face recognition task.
Jonathan W. Schooler et al.
324
0 Verbalization IControl
Verbalization
Control Condition
b
.9
-
.8-
.6 -
1
.7
-
0
-
.5
!L 0
0
.4
-
c.
c
al
eal
n
.2.3
.l0-
Just know
Reason Judgment
At a Loss from Words
325
tion of two processes: a) the influence of retrieval cues, and b) the consequences of recollection” (J. W. Schooler & Engstler-Schooler, 1990, p. 40). Verbal retrieval cues were hypothesized to elicit a nonveridical, verbally biased recollection. The resulting inaccurate recollection was then hypothesized to interfere with the original visual memory (cf. J. W. Schooler, Foster, & Loftus, 1988) by causing subjects “. . . to generate a recoded memory, disproportionately emphasizing the verbal code. This verbally biased recoding may then interfere with the application of the original memory” (p. 41). As noted, current evidence reveals that this recoding interference account is inadequate for accounting for a variety of verbal overshadowing findings. Nevertheless, we suggest that Schooler and Engstler-Schooler may still have been correct in attributing the negative effects of verbalization to an interaction between verbal retrieval cues and the act of recollection. They simply did not get the impact of these two factors quite right. 1. The Role of Retrieval Cues: Transfer-Appropriate Processing
A central assumption of the recoding interference account of verbal overshadowing is that verbalization produces a conflict between two memory representations, one corresponding to the original memory, and another corresponding to the verbalization activity. Although it remains possible that interference between distinct memory representations may characterize some verbal overshadowing effects, such an account simply cannot accommodate a variety of findings reported in this chapter. There is, however, another general way of conceptualizing conflicts in memory that does not entail the assumption of distinct memory representations. Specifically, rather than involving a competition between memory representations, it is possible that verbal overshadowing effects result from a conflict between memory processes (Kolers, 1973; Kolers & Roediger, 1984). A central premise of transfer-appropriate processing theories (e.g., Morris, Bransford, & Franks, 1977; Roediger, Weldon, & Challis, 1989) is that memory performance depends on the “extent to which operations required at test recapitulate or overlap the encoding operations performed during learning” (Roediger et al., 1989, p. 16). If retrieval conditions fail to elicit the processing operations involved during encoding, then retrieval failures can ensue. Such processing mismatches have, in the past, been revealed following
Fig. 11. (A) Frequency of “Just Know” judgments by condition for voice recognition task. (B) Accuracy of “Just Know/Reason” judgments by condition for voice recognition task.
326
Jonathan W. Schooler et al.
disparities between the operations encouraged by encoding and retrieval conditions. For example, Morris et al. (1977) found that the recognition of acoustically encoded information was maximized if the test emphasized acoustical operations, whereas the recognition of semantically encoded information was maximized if the test emphasized semantic operations. In principle, however, it is possible that other factors besides the similarity between encoding and retrieval conditions could also influence the match between encoding and retrieval processes. A central assumption of transfer-appropriate processing theory is that memory retrieval does not necessarily invoke all of the processes entailed during encoding, rather the processes elicited during retrieval depend on the nature of the retrieval cues. Thus, transfer-appropriate processing theory suggests that verbalization instructions are likely to maximize verbal memory processes but not perceptual memory processes. If a primary emphasis on verbal operations can influence the memory processes that are invoked on a subsequent memory test, then this could provide a foundation for verbal overshadowing effects. Indeed, this possibility seems especially plausible in light of recent studies examining the impact of retrieval on subsequent memory performance. 2.
The Role of the Act of Recollection: Retrieval-Induced Forgetting
There have been some impressive demonstrations of the surprising impact that retrieval can have on the accessibility of nonretrieved information. Anderson and Spellman (1995) report that the cued retrieval of some members of previously studied categorical word lists can impair the subsequent recall of the nonretrieved items. So, for example, studying fruit: ,impairs the subsequent orange, banana, and then retrieving fruit or recall of fruit: banana. More strikingly, Anderson and Spellman further find that this retrieval-induced forgetting generalizes beyond the nonretrieved members of the retrieved category. So, for example, if subjects study two category lists (e.g., green: emerald, lettuce and soups: mushroom, chicken) and are then encouraged to retrieve some members of one of the lists (e.g., green: emerald), this can impair the subsequent recall of both the nonretrieved member of the retrieved category (e.g., lettuce) and its categorical associates in the nonretrieved list (e.g., mushroom). Anderson and Spellman’s results illustrate that the retrieval of information from memory can reduce the accessibility of both nonretrieved items and items that are related to nonretrieved items. As we show, when combined with the construct of transfer-appropriate processing, this generalized but category-
At a Loss from Words
327
bound retrieval-induced forgetting may provide the necessary ingredients for an account of many verbal overshadowing findings? 3. Transfer-Inappropriate Retrieval
Consideration of the principles of transfer-appropriate processing together with the findings of retrieval-induced forgetting suggest a relatively straightforward approach for conceptualizing verbal overshadowing, which we term “transfer-inappropriate retrieval.” This approach is based on the following four assumptions: (1) verbal recall encourages the application of verbal processes and consequently the retrieval of the verbalizable aspects of the memory. This premise follows naturally from the assumption of transferappropriate processing, that retrieval processes are determined by retrieval cues; (2) selective recall of the verbal aspects of a memory can reduce the accessibility of the nonrecalled information (i.e., of nonreportable knowledge). This assumption is supported by the various demonstrations that partial retrieval of prior experiences can hamper access to nonretrieved One possible concern in drawing parallels between retrieval-induced interference effects and verbal overshadowing is that such effects are often observed with recall measures but not with recognition measures (Anderson Kc Spellman, 1995; Bjork, 1989; Slamecka, 1975). In contrast, verbal overshadowing effects typically involve recognition paradigms. The standard explanation of the greater sensitivity to interference of recall measures relative to recognition measures is that the latter introduces more retrieval cues, which thereby attenuate interference effects. In this respect, it is worth noting that in all of the verbal overshadowing studies using face recognition, the target individual is depicted in a different manner at encoding and test (either two different photos, or a video during encoding and a photo at test). In so doing, we may have minimized the degree to which the recognition test provides unique retrieval cues (cf. Read, Hammersley, Cross-Calvert, & McFadzen, 1989). It is also worth noting that verbal overshadowing effects, though replicated in many studies and in many different labs (R. Chaffin, personal communication, November, 1990; Dodson et al., 1997; C. M. Kelley, personal communication, November, 1992; Halberstadt, 1996; Lovett, Small, & Engstrom, 1992 [Experiment 21; K. Pezdek, personal communication, November, 1996; Read & Schooler, 1994 Westerman, 1991), have been found to be somewhat fragile. Occasionally verbal overshadowing effects have not been observed under situations in which they would have been expected (e.g., D. S. Lindsay, personal communication, January, 1989; Lovett et al., 1992 [Experiment 11; Yu & Geiselman, 1993). Indeed, it would be quite consistent with the present theorizing to suggest that the fragility of verbal overshadowing effects is due to the fact that most of the verbal overshadowing paradigms typically rely on recognition measures. Consistent with this prediction, using the form memory paradigm, Brandimonte, Schooler, and Gabbino (1997) observed verbal overshadowing with a recall measure (deciphering embedded forms) but not with a recognition measure. It seems quite likely that verbalization effects, like other memory effects that depend on the absence of adequate retrieval cues (e.g., context reinstatement effects, cf. Murnane & Phelps, 1994; Smith, 1988), can be observed with both recall and recognition, but are apt to be inherently more robust when tested with recall measures. Unfortunately by the very nature of nonverbal stimuli, it is often difficult to find recall measures that enable subjects to do justice to their memories.
328
Jonathan W. Schooler et al.
information (e.g., Anderson & Spellman, 1995; Roediger, 1974); (3) the interference resulting from verbal retrieval is relatively broad in scope involving a disruption in the application of the type of nonreportable processes omitted in the initial verbal retrieval. This assumption is perhaps the most controversial of the set; however, it is generally consistent with Anderson and Spellman’s (1995) finding that retrieval-induced forgetting can generalize to information that is related to nonretrieved information; (4) the reduced accessibility of nonverbal knowledge/processes can be reversed if retrieval conditions are introduced that favor the application of perceptual/nonverbal processing. This assumption follows quite naturally from the principles of transfer-appropriate processing. As the following brief review illustrates, this transfer-inappropriate retrieval account of verbal overshadowing may help to explain many of the findings that we have reported in this chapter.
a. The Generality of Verbal Overshadowing According to the transferinappropriate retrieval approach, verbal overshadowing should be limited to situations in which memory performance relies on knowledge/processes not invoked by the initial process of verbal recall. Thus, this approach can account, in principle, for why verbal overshadowing effects are observed across a variety of domains of nonverbal memory, but do not apply to domains that rely on more verbalizable knowledge. b. Processing Differences A central premise of transfer-inappropriate retrieval is that verbal overshadowing effects result from a mismatch between the processes used during encoding and verbal retrieval. Thus, this approach readily explains why nonverbal retrieval does not produce comparable interference. Accordingly, visual recall should encourage retrieval processes that are consistent with the original visual encoding operations, and thus should not impair the subsequent retrieval of information associated with such information. The transfer-inappropriate retrieval approach further suggests that visual retrieval of verbal stimuli might interfere with the subsequent access to verbal knowledge. Indeed such “visual overshadowing” might be a fruitful topic for further investigation.
c. Expertise Differences Transfer-inappropriate retrieval can also account for the relationship between verbalization and expertise. If verbal recall reduces access to nonverbal knowledge, then the costs of verbal recall should be greatest to the degree that nonverbal expertise exceeds verbal expertise. In addition, transfer-inappropriate retrieval also suggests an additional reason for why verbal recall does not interfere with the perceptual memory performance of wine experts. With training, individuals may be more likely to engage in a combination of perceptual and verbal processes
At a Loss from Words
329
during encoding. Thus, for experts, verbal retrieval may engage processes that are more appropriately matched with those initiated during encoding. As a consequence, for experts, verbal recall may be more apt to compliment rather than to clash with the processes invoked during the initial encoding of the wine.
d. The Availability of Verbalized Memories The transfer-inappropriate retrieval approach readily accounts for the effectiveness of manipulations that attenuate verbal overshadowing effect^.^ Accordingly, introducing retrieval conditions that encourage the engagement of nonverbal perceptual processes reinstates access to the perceptual knowledge associated with such processes. In addition to providing a general explanation for the attenuation of verbal overshadowing effects, the transfer-inappropriate retrieval approach may also help to explain two previously anomalous findings regarding the reversal of verbal overshadowing: the beneficial effects of re-presentation and the elimination of verbal overshadowing effects over repeated trials. As mentioned previously, J. W. Schooler et al. (1996) observed that representation of the target face increases the performance of verbalization subjects to such a degree that they exceed that of nonverbalization subjects (see Fig. 7). In the context of the transfer-inappropriate retrieval framework, this finding can be explained as follows. When subjects engage in verbal recall of a visual stimulus they rehearse and potentially strengthen the verbal knowledge, but at the expense of impairing access to visual knowledge. However, when the target is re-presented the perceptual operations are reinstated, giving verbalization subjects the best of both worlds: rehearsed verbal knowledge and refreshed perceptual operations. Another previously perplexing situation in which verbal overshadowing effects have been observed to reverse is following repeated participation in the verbal overshadowing paradigm. Specifically, in a number of studies we have observed marked verbal overshadowing effects for the first stimuludtest set, and with little or no verbal overshadowing effects on
’
At first blush, the availability assumption could be viewed as providing a possible challenge to the applicability of retrieval-induced interference. Specifically, a central premise of the availability assumption is that verbal overshadowing can be attenuated if conditions are introduced that cue the perceptual processes used during encoding. However, Anderson and Spellman posit that retrieval-induced forgetting may be cue independent (Tulving, 1974). that is, it should occur regardless of the nature of the retrieval cues. Nevertheless, they note that “more systematic exploration of which cues do and do not reinstate the ability to recall the impaired items is clearly desirable” (p. 92). In addition, they speculate that re-presenting the target stimulus may be one condition that reverses retrieval-induced forgetting, suggesting that such “dissipation is consistent with the notion that the representations of items are inhibited by retrieval but not damaged in any permanent sense” (p. 93). Thus, in fact the availability assumption is actually quite consistent with retrieval-induced forgetting literature.
330
Jonathan W. Schooler et al.
subsequent trials. The attenuating effects of verbal overshadowing over trials has now been observed in a number of domains of perceptual memory including memory for faces (Fallshore & Schooler, 1995; J. W. Schooler, Ryan, & Reder, 1991), tastes (Melcher & Schooler, 1996), and audition (Houser et al., 1997). Although the precise reason for this trial effect remains unknown, it stands to reason that, over trials, the processes elicited during encoding, verbalization, and test may become more synchronized, making transfer-inappropriate retrieval less apt to occur. e. The Lack of a Relationship between Verbal Descriptions and Performance The transfer-inappropriate retrieval approach provides a straightforward account for the lack of a relationship between verbal descriptions and performance. A central premise of this approach is that the interference associated with verbalization is not a consequence of an excessive reliance on a memory representation corresponding to the verbal activity. Rather, verbal recall is hypothesized to interfere with the successful application of nonreportable processes. As a consequence, there is no reason to expect a relationship between the specific contents of verbalization and performance.
$ The Effect of Warnings The transfer-inappropriate retrieval framework also accounts for the ineffectiveness of warnings in preventing the negative effects of verbalization. Accordingly, if verbal retrieval reduces the accessibility of the perceptual aspects of a memory, then admonishing subjects to rely exclusively on their perceptual memories should not be helpful. Indeed, the transfer-inappropriate retrieval framework may help to explain the tendency for warnings to exacerbate negative effects of verbalization. Accordingly, if verbalization impairs access to the perceptual aspects of a memory while maintaining some potentially useful verbal information, then the verbalization warning may compel subjects to ignore a potentially viable source of information. g. The Effects of Verbalizing a Nonverbalized Face One of the greatest strengths of the transfer-inappropriate retrieval approach is its ability to account for the effects of verbally recalling one face on the recognition of a different face. As discussed earlier, retrieval-induced forgetting has been shown to involve a generalized form of interference that hampers the retrieval of information categorically related to the nonretrieved information (Anderson & Spellman, 1995). The assumption of transferinappropriate retrieval is that this generalized interference can apply to not only semantic categories, but also types of processes. Thus, this approach specifically predicts that engaging in retrieval that exclusively emphasizes verbal processes should produce a generalized interference that hampers
At a Loss from Words
331
the subsequent retrieval of perceptual information associated both with the verbalized stimulus and with related nonverbalized stimuli. The negative effects of verbalizing one face on the subsequent recognition of a different face clearly support this aspect of the approach.
h. The Effects of Verbalization on Just Know/Reason Judgments Finally, the transfer-inappropriate retrieval approach readily accounts for the manner in which verbalization interacts with just know and reason judgments. This approach predicts that verbal retrieval should specifically interfere with the subsequent application of nonreportable knowledge/ processes. And that is precisely the pattern of findings suggested by the Just Know/Reason paradigm. Verbalization impairs just know judgments, without influencing either the accuracy or the frequency of reason judgments. Thus, in accord with the assumptions of transfer-inappropriate retrieval, it appears that verbalization specifically disrupts the type of knowledge that is least likely to be accessed during the initial verbal retrieval process, that is, nonreportable knowledge that subjects “just know.” 4.
Caveats and Future Directions
The transfer-inappropriate retrieval framework provides a reasonably compelling account of many of the complex findings surrounding the impact of verbalization on perceptual memories. Nevertheless, in its present form, it remains a relatively rough-hewn framework that will need further testing and refinement. One critical issue that awaits further research is the applicability of the transfer-inappropriate retrieval account to the various domains in which verbal overshadowing has been observed. As noted, several of the key sources of evidence for abandoning a recoding interference account in favor of the transfer-inappropriate retrieval approach come from investigations in the domain of face recognition. On one hand, as the present review has shown, there are good reasons for suspecting that very similar mechanisms may underlie verbal overshadowing across the various perceptual memory recognition paradigms reviewed in this chapter. The paradigms are quite similar, and even idiosyncratic findings such as the interactions between verbalization and expertise, the trial effect, and the unique relationship between verbalization and just knowheason judgments have been observed across domains. On the other hand, as noted, other verbal overshadowing paradigms, such as the Brandimonte et al. (1992a, 1992b, 1997) visual form procedure, differ in more notable respects. For example, in the Brandimonte et al. paradigm, verbalization is introduced at encoding rather than postencoding, and verbalization involves labeling the stimuli rather than describing them. Although, as argued, it seems likely that all verbal overshadowing effects involve a conflict between verbal and nonverbal
332
Jonathan W. Schooler et al.
sources, the precise nature of this conflict may depend on the particular paradigm. In some cases (e.g., the face recognition paradigm), the conflict appears to involve an interference between verbal and nonverbal processes. In other paradigms, however (e.g., the form imagery paradigm), the conflict may still be best characterized by the original notion of recoding interference, that is, as competition between distinct verbal and nonverbal memory representations. Future research will be needed to determine the conditions under which verbalization elicits processing conflicts, representational conflicts, or some combination of the two (cf. Dodson et al., 1997). If, as suggested, verbal overshadowing effects can involve a general disruption in the application of nonverbal processes, then future research will also be needed to determine the scope of this disruption. One possibility is that the processing disruption is stimulus specific. For example, the generalized effects of verbalizing a face might involve processes that are unique to face recognition (e.g., configural face processing, cf. Fallshore & Schooler, 1995; Rhodes et al., 1989). Alternatively, verbal recall might produce a more global disruption in the application of nonverbalizable perceptual processes. Consistent with this latter alternative is an unpublished finding by Westerman (1991) suggesting that even verbalizing the appearance of a previously seen car is sufficient to interfere with subsequent face recognition. This finding suggests that the processing disruption associated with verbalization may, like the verbal overshadowing effect itself, generalize broadly across domains of nonverbal processing. However, confirmation of this claim also awaits further research. In addition to investigating the scope of the processing interference introduced by verbal retrieval, additional research is also needed to clarify the premises postulated by the transfer-inappropriate retrieval approach. For example, Anderson and Spellman (1995) argued that retrieval-induced forgetting may specifically involve inhibitory proceses. Although we have been cautious in adopting this inhibitory assumption due to the fact that many seemingly inhibitory processes can be accounted for otherwise (cf. Cohen & Servain-Schreiber, 1992),it nevertheless remains a real possibility that inhibitory processes may be involved in verbal overshadowing effects. In this regard, it may be useful to find converging evidence for such inhibitory processes using a combination of neurological (e.g., Shimamura, Jurica, Mangels, & Gershberg, 1995) and individual difference (e.g., Hasher, Stoltzfus, Zacks, & Rypma, 1991) measures of inhibition. Finally, if transfer-inappropriate retrieval does in fact turn out to be a critical mechanism underlying at least some verbal overshadowing effects, then it seems likely that it may be operative in other situations as well. For example, transfer-inappropriate retrieval may provide a way of accounting for Graf and Mandler’s (1984) finding that instructing subjects to complete
At a Loss from Words
333
word fragments with previously seen words actually impairs fragment completion performance. Accordingly, explicit memory retrieval may emphasize conceptualhemantic memory processes that may thereby hamper subjects’ ability to draw on the perceptuaYnonreportable processes necessary for implicit priming. More generally, the transfer-inappropriate retrieval account suggests that perceptuallnonverbal and conceptualherbal processes are not, as is often assumed, entirely independent (e.g., Jacoby, Yonelinas, & Jennings, 1997; Paivio, 1986). Rather, at least in some situations, the application of verbal processes may be at the expense of the subsequent application of nonverbal processes.
VI. Closing Remarks In closing it may be useful to revisit the issue with which we opened this chapter, namely, the relationship between language and thought. In the past, the primary approach for investigating this relationship has involved examination of the linguistic relativity hypothesis, that is, whether concepts differ as a function of the language that one speaks (see Hunt & Agnoli, 1991, for a recent review). However, the present analysis suggests that evidence pertinent to the relationship between language and thought may be found much closer to home, by investigating the impact of committing nonverbal thoughts to words. Moreover, verbal overshadowing findings illustrate that the effects of language on thought are not, as is often assumed in Whorfian accounts (Whorf, 1956),a necessary consequence of the particular words available to a language. It seems that the impact of language on thought may not simply depend on whether words exist for a particular experience but additionally on whether or not words are applied to that experience. The central conclusion of verbal overshadowing research is that various forms of nonverbalizable knowledge may be best served by avoiding the application of language. In his seminal treatise on the relationship between language and thought, Wittgenstein (1922/1961) observes that it is possible to apprehend experiences that transcend language, noting, “there are indeed things that cannot be put into words” (p. 156). However, he argues that such experiences are beyond the purview of philosophy and, by extension, science, observing, “The correct method in philosophy would really be the following: to say nothing except what can be said” (p. 151). Although Wittgenstein may not have had in mind the types of indescribable cognitions that we have been exploring here, his conclusions nevertheless have relevance. On one hand, contrary to Wittgenstein claims, we have found that inexpressible experiences are quite amenable to empirical analysis and scientific discussion (see
334
Jonathan W. Schooler et al.
also J. W. Schooler & Fiore, 1997; J. W. Schooler & Melcher, 1995). At the same time, our research suggests new merit to Wittgenstein’s closing admonition, “Where of one cannot speak, there of must one be silent” (cited in Black, 1964, p. 377). However, given that describing the indescribable is not merely futile, but actually disruptive, a more prescriptive variant may be in order: Where of one cannot speak, there of should one be silent. ACKNOWLEDGMENTS Much of the research reported in this chapter was supported by an NIMH grant to J. W. Schooler. The authors thank Tonya Schooler for her helpful comments on an earlier draft. The authors are particularly grateful to Zara Ambadar, Marte Fallshore, Joseph Melcher. and Robert Ryan for their many contributions to the research reported in this chapter.
REFERENCES Anderson, M. C., & Spellman, B. A. (1995). On the status of inhibitory mechanisms in cognition: Memory retrieval as a model case. Psychological Review, 102, 68-100. Attneave, F. (1957). Transfer of experience with a class-schema to identification-learning of patterns and shapes. Journal of Experimental Psychology, 54, 81-88. Bahrick, H. P., & Boucher, B. (1969). Retention of visual and verbal codes of the same stimuli. Journal of Experimental Psychology, 78, 417-422. Bartlett, J. C., Till, R. E., & Levy, J. C. (1980). Retrieval characteristics of complex pictures: Effects of verbal encoding. Journal of Verbal Learning and Verbal Behavior, 19,430-449. Bekerian, D. A,, & Bowers, J. M. (1983). Eyewitness testimony: Were we misled? Journal of Experimental Psychology: Learning, Memory, and Cognition, 9, 139-145. Bjork, R. A. (1989). Retrieval inhibition as an adaptive mechanism in human memory. In H. L. Roediger & F. I. M. Craik (Eds.), Varieties of memory and consciousness: Essays in honour of Endel Tulving (pp. 309-330). Hillsdale, NJ: Erlbaum. Black, M. (1964). A companion to Wittgenstein’s Tractatus. Ithaca, NY: Cornell University Press. Brandimonte, M. A., & Gerbino, W. (1993). Mental image reversal and verbal recoding: When ducks become rabbits. Memory & Cognition, 21, 23-33. Brandimonte, M. A., & Gerbino, W. (1996). When imagery fails: Effects of verbal recoding on visual memory accessibility. In C. Cornoldi, R. H. Logie, M. A. Brandimonte, G. Kaufman, & D. Reisberg (Eds.), Stretching the imagination: Representations and transformations in mental imagery (pp. 31-76). Oxford, England: Oxford University Press. Brandimonte, M. A., Hitch, G. J., & Bishop, D. V. M. (1992a). Influence of short-term memory codes on visual image processing: Evidence from image transformation tasks. Journal of Experimental Psychology: Learning, Memory, and Cognition, 18, 157-165. Brandimonte, M. A., Hitch, G. J., & Bishop, D. V. M. (1992b). Verbal recoding of visual stimuli impairs mental image transformations. Memory & Cognition, 20, 449-455. Brandimonte, M. A., Schooler, J. W., & Gabbino, P. (1997). Attenuating verbal overshadowing through visual retrieval cues. Journal of Experimental Psychology: Learning, Memory, and Cognition, 23, 915-931. Bransford, J. D., & Franks, J. J. (1971). The abstraction of linguistic ideas. Cognitive Psychology, 2, 331-350.
At a Loss from Words
335
Brigham, J. C., & Malpass, R. S. (1985). The role of experience and contact in the recognition of faces of own- and other-race persons. Journal of Social Issues, 41, 139-155. Carmichael, L., Hogan, H. P., & Walter, A. A. (1932). An experimental study of the effect of language on the reproduction of visually perceived forms. Journal of Experimental PSyChOlOgy, 15, 73-86. Chi, M. T. H., de Leeuw, N., Chiu, M., & LaVancher, C. (1994). Eliciting self-explanation improves understanding. Cognitive Science, 18, 439-477. Christiaansen, R. E., & Ochalek, K. (1983). Editing misleading information from memory: Evidence for the coexistence of original and postevent information. Memory & Cognition, 11, 467-475. Cohen, J. D., & Servain-Schreiber, D. (1992). Context, cortex, and dopamine: A connectionist approach to behavior and biology in schizophrenia. Psychological Review, 99, 45-77. Cooper, J., & Fazio, R. H. (1984). A new look at dissonance theory. In L. Berkowitz (Ed.), Advances in experimental social psychology (Vol. 17, 229-265). New York: Academic Press. Daniel, T. C. (1972). Nature of the effect of verbal labels on recognition memory of form. Journal of Experimental Psychology, 96, 152-157. Darley, C. F., & Glass, A. L. (1975). Effects of rehearsal and serial list position on recall. Journal of Experimental Psychology: Human Learning and Memory, 104, 453-458. Dodson, C. S., Johnson, M. K., & Schooler, J. W. (1997). The verbal overshadowing effect: Source confusion or strategy shift? Memory & Cognition, 25, 129-139. Dunning, D., & Stern, L. B. (1994). Distinguishing accurate from inaccurate eyewitness identifications via inquiries about decision processes. Journal of Personality and Social Psychology, 65, 818-835. Ellis, H. D., Shepard, J. W., & Davies, G. M. (1980). The deterioration of verbal descriptions of faces over different delay intervals. Journal of Police Science and Administration, 8, 101-106. Engen, T., & Ross, B. M. (1973). Long-term memory of odors with and without verbal descriptions. Journal of Experimental Psychology, 100, 221-227. Ericsson, K. A,, & Simon, H. A. (1980). Verbal reports as data. Psychologicul Review, 87, 215-251. Ericsson, K. A,, & Simon, H. A. (1984). Protocol analysis: Verbal reports as data. Cambridge, MA: MIT Press. Fallshore, M., & Schooler, J. W. (1995). The verbal vulnerability of perceptual expertise. Journal of Experimental Psychology: Learning, Memory, and Cognition, 21, 1608-1623. Fiore, S. M. (1994). Verbal overshadowing of macro-spatial memory. Unpublished master's thesis, University of Pittsburgh, Pittsburgh, PA. Fiore, S. M., & Schooler, J. W. (1997). How did you get herefrom there: Verbal overshadowing of macro-spatial memory. Manuscript in preparation. Gagne, R. H., & Smith, E. C. (1962). A study of the effects of verbalization on problem solving. Journal of Experimental Psychology, 63, 12-18. Gardiner, J. M. (1988). Functional aspects of recollective experience. Memory & Cognition, 16, 309-313. Gardiner, J. M., & Java, R. I. (1993). Recognizing and remembering. In A. Collins, M. A. Conway, S. E. Gathercole, & P. E. Morris (Eds.), Theories ofmemory (pp. 163-188). Hillsdale, NJ: Erlbaum. Glanzer, M., & Clark, W. H. (1963). The verbal loop hypothesis: Binary numbers. Journat of Verbal Learning and Verbal Behavior, 2, 301-309. Glenberg, A,, & Adams, F. (1978). Type I rehearsal and recognition. Journal of Verbal Learning und Verbal Behavior, 17, 455-463.
336
Jonathan W. Schooler et al.
Glucksberg, S., & Danks, J. (1968). Effects of discriminative labels and of nonsense labels upon availability of novel function. Journal of Verbal Learning and Verbal Behavior, 7, 72-76. Graf, P., & Mandler, G. (1984). Activation makes words more accessible, but not necessarily more retrievable. Journal of Verbal Learning and Verbal Behavior, 23, 553-568. Graf, P., Mandler, G., & Haden, P. (1982). Simulating amnesic symptoms in normal subjects. Science, 218, 1243-1244. Greene, E., Flynn, M. S., & Loftus, E. F. (1982). Inducing resistance to misleading information. Journal of Verbal Learning and Verbal Behavior, 21, 207-219. Halberstadt, J. B. (1996). Effects of verbalization on configural and featural face recognition. Unpublished doctoral dissertation. Indiana University, Bloomington, IN. Hardin, D., & Banaji, M. R. (1993). The influence of language on thought. Social Cognition, 11, 277-308. Hasher, L., Stoltzfus, E. R., Zacks, R. T., & Rypma, B. (1991). Age and inhibition. Journal of Experimental Psychology: Learning, Memory, and Cognition, 17, 163-1 69. Heider, E. R. (1972). Universals in color naming and memory. Journal of Experimental Psychology, 93, 10-20. Henry, S . B., LeBreck, D. B., & Holzemer, W. L. (1989). The effect of verbalization of cognitive processes on clinical decision making. Research in Nursing and Health, 12, 187-193. Hirtle, S. C., & Hudson, J. (1991). Acquisition of spatial knowledge for routes. Journal of Environmental Psychology, 1I , 335-345. Houser, T., Fiore, S. M., & Schooler, J. W. (1997). Verbal overshadowing of music memory: What happens when you describe that tune? Manuscript in preparation. Hunt, E., & Agnoli, F. (1991). The Whorfian hypothesis: A cognitive psychology perspective. Psychological Review, 98, 377-389. Jacoby, L. L., & Dallas, M. (1981). On the relationship between autobiographical memory and perceptual learning. Journal of Experimental Psychology: General, 110, 306-340. Jacoby, L. L., Yonelinas, A. P., & Jennings, J. M. (1997). The relation between conscious and unconscious (automatic) influences: A declaration of independence. In J. C. Cohen & J. W. Schooler (Eds.), Scientific approaches to consciousness: The 25th Carnegie Syrnposium on Cognition (pp. 13-47). Hillsdale, NJ: Erlbaum. James, W. J. (1891). The principles of psychology. New York: Holt. Johnson, M. K., Hashtroudi, S., & Lindsay, D. S. (1993). Source monitoring. Psychological Bulletin, 114, 3-28. Kolers, P. A. (1973). Remembering operations. Memory & Cognition, I , 347-355. Kolers, P. A., & Roediger, H. I., 111. (1984). Procedures of mind. Journal of Verbal Learning and Verbal Behavior, 23, 425-449. Kosslyn, S . M., Ball, T. M., & Reiser, B. J. (1978). Visual images preserve metric spatial information: Evidence from studies of image scanning. Journal of Experimental Psychology: Human Perception and Performance, 4, 47-60. Krumhansl, C. L. (1991). Music psychology: Tonal structures in perception and memory. Annual Review of Psychology, 42, 277-303. Lantz, D., & Volney, S. (1964). Language and cognition revisited. Journal of Abnormal Psychology, 69. 472-481. Lawless, H . T. (1985). Flavor description of white wine by “expert” and nonexpert wine consumers. Journal of Food Science, 49, 120-123. Lawless, H. T., & Eagen, T. (1977). Associations to odors: Interference, mnemonics, and verbal labeling. Journal of Experimental Psychology: Human Learning and Memory, 3, 52-29.
At a Loss from Words
337
Lindsay, D. S., & Johnson, M. K. (1989). The eyewitness suggestibility effect and memory for source. Memory & Cognition, 17, 349-358. Loftus, E. F., & Loftus, G. R. (1980). On the permanence of stored information in the human brain. American Psychologist, 35, 409-420. Loftus, E. F., Miller, D. G., & Burns, H. J. (1978). Semantic integration of verbal information into a visual memory. Journal of Experimental Psychology: Human Learning and Memory, 4, 19-31. Loftus, G . R., & Kallman, H. J. (1979). Encoding and use of detail information in picture recognition. Journal of Experimental Psychology: Human Learning and Memory, 5, 1 97-21 1. Lovett, S. B., Small, M. Y.,& Engstrom, S. A. (1992). The verbal overshadowing effect: Now you see ir, now you don't. Paper presented at the annual meeting of the Psychonomic Society, St. Louis, MO. Lucy, J. A., & Shweder, R. A. (1979). Whorf and his critics: Linguistic and non-linguistic influences on color memory. American Anthropologist, 8, 581-613. Maki, R. H., & Schuler, J. (1980). Effects of rehearsal duration and levels of processing on memory for words. Journal of Verbal Learning and Verbal Behavior, 19, 36-45. Makumo, M. (1992). Encoding strategies for tonal and atonal melodies. Music Perception, 10, 73-82. Mandler, G. (1980). Recognizing: The judgment of previous occurrence. Psychological Review, 87, 252-271. McGeoch, J. A. (1942). The psychology of human learning. New York: Longmans, Green. Melcher, J. M., & Schooler, J. W. (1996). The misremembrance of wines past: Verbal and perceptual expertise differentially mediate verbal overshadowing of taste memory. Journal of Memory and Language, 35, 231-245. Melton, A. W., & Irwin, J. M. (1940). The influence of degree of interpolated learning on retroactive inhibition and overt transfer of specific responses. American Journal of Psychology, 53, 173-203. Morris, C. D., Bransford, J. D., & Franks, J. J. (1977). Levels of processing versus transfer appropriate processing. Journal of Verbal Learning and Verbal Behavior, 16, 519-533. Morton, J., Hammersley, R. H., & Bekerian, D. A. (1985). Headed records: A model for memory and its failures. Cognition, 20, 1-23. Murnane, K., & Phelps, M. P. (1994). When does a different environmental context make a difference in recognition? A global activation model. Memory & Cognition, 22, 584-590. Murray, D. J. (1967). The role of speech responses in short-term memory. Canadian Journal of Psychology, 21, 263-276. Nelson, D. L., & Brooks, D. H. (1973). Functional independence of pictures and their verbal memory codes. Journal of Experimental Psychology, 98, 44-48. Nisbett, R. E., & Wilson, T. D. (1977). Telling more than we can know: Verbal reports on mental processes. Psychological Review, 84, 231-259. Ohlsson, S. (1992). Information-processing explanations of insight and related phenomena. In M. Keane & K. Gilhooley (Eds.), Advances in the psychology ofthinking (pp. 1-44). London: Harvester-Wheatsheaf. Paivio, A. (1986). Mental representations: A dual coding approach. New York Oxford University Press. Pezdek, K., Maki, R., Valencia-Laver, D., Whetstone, T., Stoeckert, J., & Dougherty, T. (1986). How well d o we remember the visual details in pictures? Paper presented at the annual meeting of the Psychonomic Society, New Orleans. Polanyi, M. (1966). The tacit dimension. Garden City, NY: Anchor Books.
:: ,
'.
'
'
338
Jonathan W. Schooler et al.
Rabinowitz, J. C., Mandler, G., & Barsalou, L. W. (1977). Recognition failure: Another case of retrieval failure. Journal of Verbal Learning and Verbal Behavior, 16, 639-663. Rajaram, S., & Roediger, H. L. (1997). Remembering and knowing as states of consciousness during retrieval. In J. C. Cohen & J. W. Schooler (Eds.), Scientific approaches to consciousness: The 25th Carnegie Symposium on Cognition (pp. 213-240). Hillsdale, NJ: Erlbaum. Read, J. D., Hammersley, R., Cross-Calvert, S., & McFadzen, E. (1989). Rehearsal of faces and details in actions events. Applied Cognitive Psychology, 3, 295-31 1. Read, J. D. & Schooler, J. W. (1994). Verbalization decrements in long-termperson identification. Paper presented at the Third Practical Aspects of Memory Conference, College Park, MD. Rhodes, G., Tan, S., Brake, S., & Taylor, K. (1989). Expertise and configural coding in face recognition. British Journal of Psychology, 80, 313-331. Riley, D. A. (1962). Memory for form. In L. J. Postman (Ed.), Psychology in the making: Histories of selected research problems (pp. 402-465). New York: Knopf. Roediger, H. L. (1974). Inhibiting effects of recall. Memory & Cognition, 2, 261-269. Roediger, H. L., Weldon, M. S., & Challis, 8 . H. (1989). Explaining dissociations between implicit and explicit measures of retention: A processing account. In H. L. Roediger & F. 1. M. Craik (Eds.), Varieties of memory and consciousness: Essays in honour of Endel Tulving (pp. 3-41). Hillsdale, NJ: Erlbaum. Samson, S., & Zatorre, R. J. (1991). Recognition memory for text and melody of songs after unilateral temporal lobe lesion: Evidence for dual encoding. Journal of Experimental Psychology: Learning, Memory, and Cognition, 17, 793-804. Sapir, E. (1921). Language: A n introductionto thestudy ofspeech. New York: Harcourt, Brace. Schlipp, P. A. (1949). Albert Einstein: Philosopher-scientist.Cambridge, UK: Cambridge University Press. Schooler, J. W., & Engstler-Schooler, T. Y. (1990). Verbal overshadowing of visual memories: Some things are better left unsaid. Cognitive Psychology, 22, 36-71. Schooler, J. W., Fiore, S. M., Melcher, J., & Ambadar, Z. (1996). Verbal overshadowing and the “Just Know/Reason” distinction. Paper presented at the annual meeting of the Psychonomic Society, Chicago. Schooler, J. W., Foster, R. A., & Loftus, E. (1988). Some deleterious consequences of the act of recollection. Memory & Cognition, 16, 243-251. Schooler, J. W., Gabbino, P., & Brandimonte, M. A. (1997). [Verbal overshadowing of outof-focus pictures]. Unpublished raw data. Schooler, J. W., Gerhard, D., & Loftus, E. (1986). Qualitiesof the unreal. Journal of Experimental Psychology: Lemming, Memory, and Cognition, 12, 171- 181. Schooler, J. W., & Melcher, J. M. (1995). The ineffability of insight. In S. M. Smith, T. B. Ward, & R. A. Finke (Eds.), The creative cognition approach (pp. 97-133). Cambridge, MA: MIT Press. Schooler, J. W., Ohlsson, S., & Brooks, K. (1993). Thoughts beyond words: When language overshadows insight. Journal of Experimental Psychology: General, 122, 166-183. Schooler. J. W., Ryan, R. S.. Fallshore, M. F., & Melcher, J. M. (in press). Knowing more than you can tell: The relationship between language and expertise. In R. E. Nisbett & J. Caverni (Eds.). The psychology of expertise. Amsterdam: Elsevier. Schooler, J. W., Ryan, R., & Reder, L. M. (1991, July). Better the second time around: Representation reverses verbalization’s impairment of face recognition. Paper presented at the International Conference on Memory. Schooler, J. W., Ryan, R. S., & Reder, L. (1996). The costs and benefits of verbally rehearsing memory for faces. In D. Herrmann, M., Johnson, C. McEvoy, C. Hertzog, & P. Hertel (Eds.), Basic and applied memory: New findings (pp. 51-65). Hillsdale, NJ: Erlbaum.
At a Loss from Words
339
Schooler, T. Y. (1990). Do discrepant attitudinal components become consistent over time. Unpublished master’s thesis, University of Pittsburgh, Pittsburgh, PA. Shapiro. P. N., & Penrod, S. (1986). Meta-analysis of facial identification studies. Psychological Bulletin, 100, 139-156. Shimamura, A. P., Jurica, P. J., Mangels, J. A., & Gershberg, F. B. (1995). Susceptibility to memory interference effects following frontal lobe damage: Findings from tests of pairedassociate learning. Journal of Cognitive Neuroscience, 7, 144-152. Siegel. A. W., & White, S. H. (1975). The development of spatial representation of largescale environments. In H. W. Reese (Ed.), Advances in child development and behavior (pp. 9-55). New York: Academic Press. Slamecka, N. J. (1975). Intralist cuing of recognition. Journal of Verbal Learning and Verbal Behavior, 14, 630-637. Smith, S. M. (1988). Environmental context-dependent memory. In G. Davies & D. Thomson (Eds.), Memory in context: Context in memory (pp. 13-34). New York: Wiley. Solomon, G. E. A. (1990). Psychology of novice and expert wine talk. American Journal of Psychology, 103, 495-517. Thorndyke, P. W., & Hayes-Roth, B. (1982). Differences in spatial knowledge acquired from maps and navigation. Cognitive Psychology, 14, 560-589. Tulving, E. (1974). Cue-dependent forgetting. American Scientist, 62, 74-82. Tulving. E. (1985). Memory and consciousness. Canadian Psychologisr, 26, 1-12. Tulving, E., & Pearlstone, Z. (1966). Availability versus accessibility of information in memory for words. Journal of Verbal Learning and Verbal Behavior, 5, 381-391. Wells, G. L., & Turtle, J. W. (1987). What is the best way to encode faces? In M. M. Gruneberg, P. E. Morris, & R. N. Sykes (Eds.), Practical aspects of memory: Current research and issues (pp. 163-168). New York: Wiley. Westerman, D. L. (1991). The effects of verbal description on the recognition of faces and other visual material. Paper presented at the North East Ohio Undergraduate Psychology Conference, Kent. Whorf, B. L. (1956). Language, thought, and reality: Selected writings. Cambridge, MA: MIT Press. Wilson, T. D., & Dunn, D. S. (1986). Effects of introspection on attitude-behavior consistency: Analyzing reasons versus focusing on feelings. Journal of Experimental Social Psychology, 22, 249-263. Wilson, T. D., Dunn, D. S., Kraft, D., & Lisle, D. J. (1989). Introspection, attitude change, and attitude-behavior consistency: The disruptive effects of explaining why we feel the way we do. In L. Berkowitz (Ed.), Advances in experimental social psychology (Vol. 22, pp. 287-343). Orlando, FL: Academic Press. Wilson, T. D. & Lafleur, S. J. (1995). Knowing what you’ll do: Effects of analyzing reasons on self-prediction. Journal of Personality and Social Psychology, 68, 21-35. Wilson, T. D., Lisle, D. J., Schooler, J. W., Hodges, S. D., Klaaren, K. J., & Lafleur, S. J. (1993). Introspecting about reasons can reduce post-choice satisfaction. Personality and Social Psychology Bulletin, 19, 331-339. Wilson, T. D., & Schooler, J. W. (1991). Thinking too much: Introspection can reduce the quality of preferences and decisions. Journal of Personality and Social Psychology, 60, 181-192. Wittgenstein, L. (1961). Tractatus Logica-Philosophicus.New York: Humanities Press. (Original work published 1922) Yu,C. J., & Geiselman, R. E. (1993). Effects of constructing identi-kit composites on photospread identification performance. Criminal Justice and Behavior, 20, 280-292.
340
Jonathan W. Schooler et al.
Zaragoza, M. S., & Koshmider, J. W. (1989). Misled subjects may know more than their performance implies. Journal of Experimental Psychology: Learning, Memory, and Cognition, 15, 246-255. Zaragoza, M. S., & Lane, S. M. (1994). Source misattribution and the suggestibility of eyewitness memory. Journal of Experimental Psychology: Learning, Memory, and Cognition, 20, 934-945. Zatorre, R. J., & Beckett, C. (1989). Multiple coding strategies in the retention of musical tones by possessors of absolute pitch. Memory & Cognition, 17, 582-589.
A Access, 3, 7, 9 Action effect, counterfactual thinking, 123-1 33 Addition, object-based reasoning, 30-31 Affect access, manner of change, 13 Affective decision making, verbalization and, 301-302 Affirmation of the consequent inference, 120 Anaphoric reference attenuating interference during, 90-92, 101 defined, 87.90 Arithmetic operations, object-based reasoning, 30-33, 34 Attenuating inference, 98-101 anaphoric reference, 87, 90-92, 101 cataphoric reference, 87, 92-94, 101 inference revision, 87, 97-98, 102 lexical access, 87, 88-90, 101 metaphor interpretation, 87, 95-97, 101-102 syntactic parsing, 87, 94-95, 101 Attributes, 5, 6 Audition, memory for, 301 Availability assumption, verbal overshadowing, 294-296,297,298, 310-315
B Bearing (navigation), 44
C Cataphoric reference attenuating interference during, 92-94, 101 defined, 87,92 Classification, 212 speeded classification, 287 Cognitive map, 47 Color naming, 189-193 Color recognition, 293-294 Computer simulation models, EPAM, 266-289 Conceptually driven encoding, 157-158, 160- 167 Conditionals counterfactual, 113-122 mental models, 111-112 Conscious recollection, 171-184.204-205 Content, 3, 6 Context experiment design, 162-163 temporality effect and, 136-141 word identification and, 238-241 Convergence, object-based reasoning, 7-12 Coordinate system, 42,46 Counterfactual conditionals, 113-122 Counterfactual deductions, 116-122, 149 Counterfactual thinking, 105-107, 123, 148-150 action effect, 123-133 constraints, 107-110 deductive inferences and, 106 functions, 106-107 341
342
Index
generation of scenario, 107-109, 110-11 1, 123 mental models of conditionals, 111-122 spatial effect, 141-148 temporality effect, 134-141 Course (navigation), 44
D Data-driven encoding, 157-158, 160-167 Dead reckoning, 44 Deductions, counterfactual, 116-122, 149 Deductive inferences, counterfactual thinking and, 106 Denial of the antecedent inference, 120 Direct addition, object-based reasoning, 30-31 Direct division, object-based reasoning, 31-33
E Egocentric coordinate system, 46, 48 Elaborative encoding, 158-160 Encoding, 42, 53 accuracy, 80 conceptually driven, 157-158, 160-1 67 data-driven, 157-158, 160-167 elaborative, 158-160 errors in, 55 estimation of, 57 experience and, 66-71 interpretive, 159-160, 203, 205 memory, 155, 156, 158-160 path completion, 62-66 path encoding, 42.53, 58-71, 78-79.80 recoding interference hypothesis. 293, 296-297,315-318, 325 reproduction tasks, 58-62 verbalization and, 313 Encoding-error model, 42,53-58.62-66 Encoding specificity principle, 156 Enhancement, memory nodes, 86 EPAM model, 266,271,287-289 comparison with human data, 284-285 goals, 272-274 Medin and Smith task, 274-284 prototype instructions, 282-284 rules-plus-exception instructions, 280-282 speeded classification, 287
strategy, 271-272 transfer task. 286-287 EPAM V. 273 Episodic-processing theory, 225 Explicitlimplicit distinction, 156
F Face recognition Just Know/Reason distinction, 320-322 Medin and Smith task, 274-280 own versus other-race recognition, 305-308 re-presentation, 311-312 verbalization, 292-293, 305-308, 31 1-312, 317-319,330-331 Failure to act, See also Counterfactual thinking regretting, 129-133 Fluent remembering, 156-157 Forgetting, retrieval-induced, 326-327 Form, memory for, 298-299, 300,303-304, 313-315 Functional fixedness, 33
G Geometric coordinate system, 46-47
H Heading (navigation), 44 Homing vector, 45.49-50, 71-74,79 Human path integration. See Path integration
I Identification, 212 Implicit learning, 228-230, 234 Independence remernbedknow (IRK) procedure, 182 Inference revision, attenuating interference during, 97-98, 102 Inferences, counterfactual thinking and, 117-122 Inferencing, defined, 87 Insight problem solving, verbalization and, 302 Interpretive encoding, 158-160, 203, 205 Involuntary conscious memory, 177-178
Index
J Just KnowlReason distinction face recognition, 320-322 voice recognition. 323
L Landmark, 43 Landmark-based navigation, 44 Language, thought and, 291,333 Language comprehension, 85-88 anaphoric reference, 87, 90-92, 101 attenuating inference and, 87, 98-101 cataphoric reference, 87, 92-94 inference revision, 87, 97-98, 102 lexical access, 87, 88-90, 101 metaphor interpretation, 87, 95-97. 101-102 syntactic parsing, 87, 94-95, 191 Lexical access attenuating interference during, 88-90, 101 defined, 87 Lexicon, 238
M Macrospatial relationships, memory for, 299 Map memory, visualization and, 303 Mapping, 3, 5, 7 manner of change, 13 manner of convergence, 9 Masked word identification, 158, 161, 164, 166 experiments, 156-171 priming, 193-198 Medin and Smith task, 274-284 Memory for audition, 301 availability assumption, 294-296, 297, 310-315 color naming, 189-193 conscious recollection, 171-184, 204-205 defined, 156,211-212 encoding, 155, 156, 158-160 experiments, 160-1 67 episodic effects on perceptual judgments, 198-203 episodic enhancement of processing fluency, 155-206
343
explicitlimplicit distinction, 156 fluent remembering, 156-157 for forms, 298-299,300,303-304, 313-315 involuntary conscious memory, 177-178 for macrospatial relationships, 299 map memory, 303 masked word identification, 158, 161, 164, 166 experiments, 156-171 priming, 193-198 modality mismatch assumption, 293-294, 297-310 for music, 301 priming, 164, 187, 189-198,240 processing, 156, 225-235 data-driven vs. conceptually driven, 157-158 recoding interference hypothesis, 293, 296-297,315-318 repetition blindness, 248-250 selective construction and preservation of experience (SCAPE), 213,214-260 separate-systems assumptions, 213-214, 235, 237 speeded word reading, 184-193 for taste, 299, 301 Memory nodes, 86 Metaphor interpretation attenuating interference during, 95-97, 101-102 defined, 87 Modality mismatch assumption, 293-294, 297-310 Modus ponens inference, 119-120, 121 Modus rollens inference, counterfactual thinking and, 117-119, 120, 121 Moment-to-moment updating, 45, 71-74 Motor trace, 47 Music, memory for, 301 Mutilated checkerboard problem, 267
N Naming, 185 Navigation, 41-45 encoding, 42,53,58-71, 78-79, 80 encoding-error model, 42, 53-58,62-66 homing vector, 45,49-50,71-74, 79 shortcuts, 51-53, 62, 71, 77-78
344 spatial representation, 45-51 without sight, 58-81
0 Object, 43 Object attributes, 6 Object-based inferences, 2, 16, 19-20, 24 Object-based reasoning, 1-3 content and structure, separation and contrast between, 4-6 problem representation, 2, 7-20 processing strategies, 2-3, 21-33 semantic knowledge, 33-35 manner of change, 12-16 manner of convergence, 7-12 symmetry and asymmetry, 2. 16-19
P Path completion, 51-53 encoding, 62-66 encoding-error model, 42, 53-58, 62-66 Path integration, 41-42.44-45, 77-81 encoding, 42, 53, 58-71,78-79, 80 homing vector, 45,49-50,71-74,79 navigation, 42-51 without vision, group and individual differences, 74-77,79-80 Shortcuts, 42,51-53,62-66 Perceptual identification, 158 Perceptual memories verbalization, 292-293 verbalization and, 318-334 verbal overshadowing of, 290-334 Priming, 160, 164, 187 masked word identification, 193-198 repetition priming, 240 specificity of, 189-193 Problem space, 267, 269 Processing replacements. 21 in similarity judgments, 22-26 Prototype abstraction, 222-223
R Recoding interference hypothesis, verbal overshadowing, 293,296-297,315-318. 325
Index Relations, 5 Relative bearing (navigation), 44 Remembering. See Memory Repetition blindness, 248-250 Repetition priming, 240 Retrieval cues, transfer-appropriate processing, 325-326 Retrieval-induced forgetting, 326-327 Retrolental fibroplasia (RLF), 74 Route (navigation), 47-48 RULEX system, 274
S Selective construction and preservation of experience (SCALE), 213, 214-260 constructive evaluation, 241-260 constructive production, 221 -241 Semantic-escape strategies, 31-32 Semantic knowledge, 2-3 problem representation, 2, 7-20 processing strategies, 2-3, 21-33 Semantic priming, 160 Semantic symmetry and asymmetry, 2, 16-19 Separate-systems explanation, 235, 237 Sound, memory for, 301 Spatial effect, counterfactual thinking, 141- 148 Spatial representation, 42, 45-47 computations, 50 multiple representations, 47-48 navigation, 50-51 updating, 49 Speeded classification, 287 Speeded word reading, 184-193 Structure, 6 Structure building framework, 86 Structure-mapping theory, 5 Suppression, See also Attenuating interference language comprehension, 85-101 memory nodes, 86 Surface information loss, defined, 87 Survey (navigation), 47-48 Syntactic parsing attenuating interference during, 94-95, 101 defined, 87 Systematicity principle, 24
Index
T Task domain, 268-269 Task performance attention control, 270-271 EPAM model, 271-289 problem space, 267, 269 social behavior and, 270 strategies, 267, 269 task domain, 268-269 task representation, 267-268 Task representation, 267 Taste, memory for, 299, 301 Temporality effect, counterfactual thinking, 134-141 Thought, language and, 291,333 Traces, 155 Transfer-inappropriate retrieval, 323, 327-333 Transfer task, EPAM, 286-287
V Verbalization, See abo Verbal overshadowing affective decision making and, 301-302 domains not affected by, 302 encoding and, 313 face recognition, 292-293, 305-308, 311-312,317-319.330-331 form recognition, 298-299,300,303-304, 313-315
345
insight problem solving and, 302 macrospatial relationships, 299 music, 301 nonverbalized stimuli and, 318 perceptual memories and, 318-334 taste, 299, 301,308-310 wine tasting, 308-310 Verbal overshadowing availability assumption, 294-296, 297, 298,310-315 defined, 293 modality mismatch assumption, 293-294, 297-310 recoding interference hypothesis, 293, 296-297,315-318,325 transfer-inappropriate retrieval, 328 Vision, path integration without, 77-78 Visualization, map memory and, 303 Voice recognition, Just Know/Reason distinction, 323
W Wine tasting, verbalization, 308-310 Word identification context and, 238-241 masked word identification, 158, 161, 164, 166 experiments, 156-171 priming, 193-198
This Page Intentionally Left Blank
CONTENTS OF RECENT VOLUMES Volume 26 Spatial Memory in Seed-Caching Corvids Alan C. Kamil and Russell P. Balda Detecting Response-Outcome Relations: Toward an Understanding of the Causal Texture of the Environment E. A. Wasserman Priming of Nonverbal Information and the Nature of Implicit Memory Daniel L. Schacter, Suzanne M. Delaney, and Elizabeth P. Merikle Metamemory: A Theoretical Framework and New Findings Thomas 0. Nelson and Louis Narens The New Multimodal Approach to Memory Improvement Douglas J. Herrmann and Alan Searlemann A Triphasic Approach to the Acquisition of Response-Selection Skill Robert W. Proctor, T. Gilmour Reeve, and Daniel J. Weeks The Structure and Formation of Natural Categories Douglas Fisher and Pat Langley Index
Volume 27 Deriving Categories to Achieve Goals Lawrence W. Barsalou Learning and Applying Category Knowledge in Unsupervised Domains John P. Clapper and Gordon H. Bower
Spatial Mental Models Barbara Tversky Memory’s View of Space Timothy P. McNamara Made in Memory: Distortions in Recollection after Misleading Information Elizabeth F. Loftus Cognitive Processes and Mechanisms in Language Comprehension: The Structure Building Framework Morton Ann Gernsbacher Temporal Learning John E. R. Staddon and Jennifer J. Higa Behavior’s Time Peter R. Killeen Index
Volume 28 Conditioned Food Preferences Elizabeth D. Capaldi Classical Conditioning as an Adaptive Specialization: A Computational Model C. R. Gallistel Occasion Setting in Pavlovian Conditioning Peter C. Holland Pairings in Learning and Perception: Pavlovian Conditioning and Contingent Aftereffects Shepard Siege1 and Lorraine G . Allan Evolutionary Memories, Emotional Processing, and the Emotional Disorders Susan Mineka Investigations of an Exemplar-Based Connectionist Model of Category Learning Robert M. Nosofsky and John K. Kruschke 347
348
Contents of Recent Volumes
Reconstructing the Past: Category Effects in Estimation Janellen Huttenlocher and Larry V. Hedges Index
Volume 29 Introduction: A Coupling of Disciplines in Categorization Research Roman Taraban Models of Categorization and Category Learning W. K. Estes Three Principles for Models of Category Learning John K. Kruschke Exemplar Models and Weighted Cue Models in Category Learning Roman Taraban and Joaquin Marcos Palacios The Acquisition of Categories Marked by Multiple Probabilistic Cues Janet L. McDonald The Evolution of a Case-Based Computational Approach to Knowledge Representation, Classification, and Learning Ray Bareiss and Brian M. Slator Integrating Theory and Data in Category Learning Raymond J. Mooney Categorization, Concept Learning, and Problem-Solving: A Unifying View Douglas Fisher and Jungsoon Park Yo0 Processing Biases, Knowledge, and Context in Category Formation Thomas B. Ward Categorization and Rule Induction in Clinical Diagnosis and Assessment Gregory H. Mumma A Rational Theory of Concepts Gregory L. Murphy Concept Structure and Category Boundaries Barbara C. Malt Non-Predicating Conceptual Combinations Edward J. Shoben Exploring Information about Concepts by Asking Questions Arthur C. Graesser, Mark C. Langston, and William B. Baggett
Hidden Kind Classifications Edward Wilson Averill Is Cognition Categorization? Timothy J. van Gelder What Are Concepts? Issues of Representation and Ontology William F. Brewer Index
Volume 30 Perceptual Learning Felice Bedford A Rational-Constructivist Account of Early Learning about Numbers and Objects Rochel Gelman Remembering, Knowing, and Reconstructing the Past Henry L. Roediger 111, Mark A. Wheeler, and Suparna Rajaram The Long-Term Retention of Knowledge and Skills Alice F. Healy, Deborah M. Clawson, Danielle S. McNamara, William R. Mamie, Vivian I. Schneider, Timothy C. Rickard, Robert J. Crutcher, Cheri L. King, K. Anders Ericsson, and Lyle E. Bourne, Jr. A Comprehension-Based Approach to Learning and Understanding Walter Kintsch, Bruce K. Britton, Charles R. Fletcher, Eileen Kintsch, Suzanne M. Mannes, and Mitchell J. Nathan Separating Causal Laws from Causal Facts: Pressing the Limits of Statistical Relevance Patricia W. Cheng Categories, Hierarchies, and Induction Elizabeth F. Shipley Index
Volume 31 Associative Representations of Instrumental Contingencies Ruth M. Colwill A Behavioral Analysis of Concepts: Its Application to Pigeons and Children Edward A. Wasserman and Suzette L. Astley
Contents of Recent Volumes The Child’s Representation of Human Groups Lawrence A. Hirschfeld Diagnostic Reasoning and Medical Expertise Vimla L. Patel, Jose F. Arocha, and David R. Kaufman Object Shape, Object Name, and Object Kind: Representation and Development Barbara Landau The Ontogeny of Part Representation in Object Concepts Philippe G. Schyns and Gregory L. Murphy Index
Volume 32 Cognitive Approaches to Judgment and Decision Making Reid Hastie and Nancy Pennington And Let Us Not Forget Memory: The Role of Memory Processes and Techniques in the Study of Judgment and Choice Elke U. Weber, Wiliam M. Goldstein, and Sema Barlas Content and Discontent: Indications and Implications of Domain Specificity in Preferential Decision Making William M. Goldstein and Elke U. Weber An Information Processing Perspective on Choice John W. Payne, James R. Bettman, Eric J. Johnson, and Mary Frances Luce Algebra and Process in the Modeling of Risky Choice Lola L. Lopes Utility Invariance Despite Labile Preferences Barbara A. Mellers, Elke U. Weber, Lisa D. Ordofiez, and Alan D. J. Cooke Compatibility in Cognition and Decision Eldar Shafir Processing Linguistic Probabilities: General Principles and Empirical Evidence David V. Budescu and Thomas S. Wallsten
349
Compositional Anomalies in the Semantics of Evidence John M. Miyamoto, Richard Gonzalez, and Shihfen Tu Varieties of Confirmation Bias Joshua Klayman Index
Volume 33 Landmark-Based Spatial Memory in the Pigeon Ken Cheng The Acquisition and Structure of Emotional Response Categories Paula M. Niedenthal and Jamin B. Halberstadt Early Symbol Understanding and Use Judy S. DeLoache Mechanisms of Transition: Learning with a Helping Hand Susan Goldin-Meadow and Martha Wagner Alibali The Universal Word Identification Reflex Charles A. Perfetti and Sulan Zhang Prospective Memory: Progress and Processes Mark A. McDaniel Looking for Transfer and Interference Nancy Pennington and Bob Rehder Index
Volume 34 Associative and Normative Models of Causal Induction: Reacting to versus Understanding Cause A. G. Baker, Robin A. Murphy, and FrBdBric VallBe-Tourangeau Knowledge-Based Causal Induction Michael R. Waldmann A Comparative Analysis of Negative Contingency Learning in Humans and Nonhumans Douglas A. Williams Animal Analogues of Causal Judgment Ralph R. Miller and Helena Matute Conditionalizing Causality Barbara A. Spellman Causation and Association Edward A. Wasserman, Shu-Fang Kao, Linda J. Van Hamme, Masayoshi Katagiri, and Michael E. Young
350
Contents of Recent Volumes
Distinguishing Associative and Probabilistic Contrast Theories of Human Contingency Judgment David R. Shanks, Francisco J. Lopez, Richard J. Darby, and Anthony Dickinson A Causal-Power Theory of Focal Sets Patricia W. Cheng, Jooyong Park, Aaron S. Yarlas, and Keith J. Holyoak The Use of Intervening Variables in Causal Learning Jerome R. Busemeyer, Mark A. McDaniel, and Eunhee Byun Structural and Probabilistic Causality Judea Pearl Index
Volume 35 Distance and Location Processes in Memory for the Times of Past Events William J. Friedman Verbal and Spatial Working Memory in Humans John Jonides, Patricia A. Reuter-Lorenz, Edward E. Smith, Edward Awh, Lisa L. Barnes, Maxwell Drain, Jennifer Glass, Erick J. Lauber, Andrea L. Patalano, and Eric H. Schumacher Memory for Asymmetric Events John T. Wixted and Deirdra H. Dougherty The Maintenance of a Complex Knowledge Base After Seventeen Years Marigold Linton Category Learning As Problem Solving Brian H. Ross Building A Coherent Conception of HIV Transmission: A New Approach to Aids Educations Terry Kit-fong Au and Laura F. Romo Spatial Effects in the Partial Report Paradigm: A Challenge for Theories of Visual Spatial Attention Gordon D. Logan and Claus Bundesen
Structural Biases in Concept Learning: Influences from Multiple Functions Dorrit Billman Index
Volume 36 Learning to Bridge Between Perception and Cognition Robert L. Goldstone, Philippe G. Schyns, and Douglas L. Medin The Affordances of Perceptual Inquiry: Pictures Are Learned From the World, and What That Fact Might Mean About Perception Quite Generally Julian Hochberg Perceptual Learning of Alphanumeric-Like Characters Richard M. Shiffrin and Nancy Lightfoot Expertise in Object and Face Recognition James Tanaka and Isabel Gauthier lnfant Speech Perception: Processing Characteristics, Representational Units, and the Learning of Words Peter D. Eimas Constraints on the Learning of Spatial Terms: A Computational Investigation Terry Regier Learning to Talk About the Properties of Objects: A Network Model of the Development of Dimensions Linda B. Smith, Michael Gasser, and Catherine M. Sandhofer Self-organization, Plasticity, and Low-Level Visual Phenomena in a Laterally Connected Map Model of the Primary Visual Cortex Risto Mikkulainen, James A. Bcdnar. Yoonsuck Choe, and Joseph Sirosh Perceptual Learning From Cross-Modal Feedback Virginia R. de Sa and Dana H. Ballard Learning As Extraction of Low-Dimensional Representations Shimon Edelman and Nathan lntrator Index
This Page Intentionally Left Blank
I S B N 0-12-5q3337-7