Psychology of Learning and Motivation, Vol 17

THE PSYCHOLOGY OF LEARNING AND MOTIVATION Advances in Research and Theory VOLUME I7 CONTRIBUTORS TO THIS VOLUME Har...

Author: Gordon H. Bower

117 downloads 1164 Views 14MB Size Report

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

DOWNLOAD PDF

THE PSYCHOLOGY OF LEARNING AND MOTIVATION Advances in Research and Theory

VOLUME I7

CONTRIBUTORS TO THIS VOLUME

Harry P . Bahrick William F . Brewer Michael Domjan Terry R . Greene Marcia K . Johnson David Kieras

John R . Pani Barbara C . Penner Timothy A . Post James F . Voss

THE PSYCHOLOGY OF LEARNING AND MOTIVATION Advances in Research and Theory

EDITEDBY GORDON H. BOWER STANFORD UNlVERSlTY, STANFORD, CALIFORNIA

Volume 17

I983

ACADEMIC PRESS A Subsidiary of Harcourt Brace Jovanovich, Publishers

New York London Paris San Diego San Francisco SBo Paulo Sydney Tokyo Toronto

COPYRIGHT @ 1983, BY ACADEMIC PRESS,INC. ALL RIGHTS RESERVED. NO PART O F THIS PUBLICATION MAY BE REPRODUCED OR TRANSMITTED IN ANY FORM OR BY ANY MEANS, ELECTRONIC O R MECHANICAL, INCLUDING PHOTOCOPY, RECORDING, OR ANY INFORMATION STORAGE AND RETRIEVAL SYSTEM, WITHOUT PERMISSION IN WRITING FROM THE PUBLISHER.

ACADEMIC PRESS,INC.

11I Fifth Avenue, New York, New York 10003

United Kingdom Edition published by ACADEMIC PRESS, INC. ( L O N D O N ) LTD. 24/28 Oval Road, London N W l

7DX

LIBRARY OF CONGRESS CATALOG CARDNUMBER:66-30104 ISBN 0-1 2-54331 7-4 PRINTED IN THE UNITED STATES OF AMERICA 83 84 85 86

9 8 7 6 5 4 3 2 1

CONTENTS

Contributors.. ...............................................................

ix

Contents of Previous Volumes.. ...............................................

xi

TY William F . Brewer and John R . Pani

Introduction .......................... An Example ......................... ........................... Ill. A Botany of Memory ................................................. IV. A Structural Account of Human Memory ............................... V. Structure of Memory: Implications ..................................... V1. Conclusions. ............. ............................... References ........................................................ I.

11.

2 5 14

29 33 34

A SIMULATION MODEL FOR THE COMPREHENSION OF TECHNICAL PROSE David Kieras

I. 11. 111. IV.

Introduction ......... Description of the Sim .......................................... Comparisons with Dat Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . References . . . . . . . . . . V

46

78

vi

Contents

A MULTIPLE-ENTRY, MODULAR MEMORY SYSTEM Marcia K . Johnson I. 11. 111.

IV.

Introduction .......................................................... The Model ........................................................... The Multiple-Entry Model and Other Theoretical Issues .................. Summary ............................................................ References ...........................................................

81 82 103 115 116

THE COGNITIVE MAP OF A CITY: FIFTY YEARS OF LEARNING AND MEMORY Harry P . Bahrick I. 11. 111.

IV. V.

Introduction .......................................................... Subject Recruitment and Test Administration. .............. ........ Composition of the Tests.. ...................... Scoring Procedures ... .... .... ....... Results and Discussion.. ..............................................

125

127 134

138

......................................................... 161 ........................ . . . . . 163

PROBLEM-SOLVING SKILL IN THE SOCIAL SCIENCES James F . Voss, Terry R . Greene, Timothy A . Post, and Barbara C . Penner I. 11. 111. IV.

V. VI. VII. VIII.

Introduction ............................. . . . . . . . . . . . . . . 165 The Information-Processing Model ........ ...................... 166 Social Science Problems and Their Solutions ............................ I68 The Problem-Solving-Reasoning Model .................... ................................ 173 m-Solving Skill . . . . . . General Considerations .................... Concluding Remarks ....................... References. . . . ............................................ 212

Contents

vii

BIOLOGICAL CONSTRAINTS ON INSTRUMENTAL AND CLASSICAL CONDITIONING: IMPLICATIONS FOR GENERAL PROCESS THEORY Michael Domjan I. I1 . I11 . IV . V. V1 . VII . VIII .

Introduction.......................................................... Constraints on Positive Reinforcement .................................. Constraints on Punishment ............................................ Constraints on Avoidance Learning ..................................... Selective Associations in Classical Conditioning.......................... LongDelay Learning ................................................. Potentiation in Classical Conditioning ................................... Implications for the Study of General Process Learning Theory ............ References ...........................................................

I n ~ e x.......................................................................

216 217 227 228 235 248 259 263 268 219

This Page Intentionally Left Blank

Numbers in parentheses indicate the pages on which the authors' contributions begin.

Harry P. Bahrick, Department of Psychology, Ohio Wesleyan University, Delaware, Ohio 43015 (125) William F. Brewer, Department of Psychology, University of Illinois at Urbana-Champaign, Champaign, Illinois 61820 (1) Michael Domjan, Department of Psychology, The University of Texas at Austin, Austin, Texas 78712 (215) Terry R. Greene, Learning Research and Development Center, University of Pittsburgh, Pittsburgh, Pennsylvania 15260 (165) Marcia K. Johnson, Department of Psychology, State University of New York at Stony Brook, Stony Brook, New York 11794 (81) David Kieras, Department of Psychology, Faculty of Social and Behavioral Sciences, College of Arts and Sciences, University of Arizona, Tucson, Arizona 85721 (39) John R. Pani,' Department of Psychology, University of Illinois at UrbanaChampaign, Champaign, Illinois 61820 (1) Barbara C. Penner, Learning Research and Development Center, University of Pittsburgh, Pittsburgh, Pennsylvania 15260 (165) Timothy A. Post, Learning Research and Development Center, University of Pittsburgh, Pittsburgh, Pennsylvania 15260 (165) James F. Voss, Learning Research and Development Center, University of Pittsburgh, Pittsburgh, Pennsylvania 15260 (165)

'Present address: Department of Psychology, University of Massachusetts, Amherst, Massachusetts 01003. ix

This Page Intentionally Left Blank

Volume 1

Volume 3

Partial Reinforcement Effects on Vigor and Persistence Abram Amsel A Sequential Hypothesis of Instrumental Learning E. J . Capaldi Satiation and Curiosity Harry Fowler A Multicomponent Theory o f t h e Memory Trace Cordon Bower Organization and Memory George Mandler Author Index-Subject Index

Stimulus Selection and a "Modified Continuity Theory Allan R. Wagner Abstraction a n d the Process of Recognition Michael I. Posner Neo- Noncont i n ui t y Theory Marvin Levine Computer Stimulation of Short-Term Memory: A Component-Decay Model Kenneth R. Laughery Replication Process in Human Memory and Learning Harley A. Bernbach Experimental Analysis of Learning t o Learn Leo Postman Short-Term Memory in Binary Prediction by Children: Some Stochastic Information Processing Models Richard S. Bogartz Author Index-Subjuct Index "

Volume 2 Incentive Theory and Changes in Reward Frank A. Logan Shift in Activity and the Concept of Persisting Tendency David Birch Human Memory: A Proposed System and Its Control Processes R. C, Atkinson and R. M. Shiffrin Mediation and Conceptual Behavior Howard K. Kendler and Tracy S. Kendler Author Index-Sub~ect Index

Volume 4 Learned Associations over Long Delays Sam Revusky and John Garcia On the Theory of Interresponse-Time Reinforcement xi

xii

Contents of Previous Volumes

G.S. Reynolds and Aiastdir McLeod Sequential (:hoice Behavior Jerome L. Meyers The Role of Chunking and Organization in the Process (if Recall Neal F. Johnson Organization o f Serial Pattern Learning Frank Restle and Eric Brown Atilltor Index-Subject Index

Volume 5 Conditioning and a Decision Theory of Response Evocation G. Robert (;rice Short-Term Memory Bennet B. Murdock.fr. Storage Mechanisms in Recall Murray Glanzer By-Products of Discrimination Learning H. S. Terrace Serial Learning and Diniensional Organization Sheldon M. Ebenholtz FRAN: A Simulation Model of Free Recall John Robert Anderson Author Index-SubjPct Index

Volume 6 Informational Variables in Pavlovian Conditioning Robert A. Rescorla T h e Operant Conditioning of Central Nervous System Electrical Activity A. H. Black T h e Avoidance Learning Problem Robert C. Bolles Mechanismsof Directed Forgetting William Epstein Toward a Theory of Redintegrative Memory: Adjective-Noun Phrases Leonard M. Horowitz and Leon Manelis

Elaborative Strategies in Verbal Learning and Memory William E. Montague Author IndPx-Subjert Index

Volume 7 Grammatical Word Classes: A Learning Process and Its Simulation George R. Kiss Reaction Time Measurements in thc: Study of Memory Processes: Theory and Data John Theios lndividiial Differences in Cognition: A New Approach to Intelligence Earl Hunt, Nancy Frost. and Clifford Lunneborg Stimulus Encoding Processes in Human Learning and Memory Henry C. Ellis Subproblem Analysis of Discriminalion Learning Thomas Tighe Delayed Matching and Short-Term Memory in Monkeys M. R. D’Amato Percentile Reinforcement: Paradigms for Experimental Analysis of Response Shaping John R. Platt Prolonged Rewarding Brain Stimulation J. A. Deutsch Patterned Reinforcement Stewart H. Hulse Author Index-Subjert Index

Volume 8 Semantic Memory and Psychological Semantics Edward E. Smith, Lance]. Rips, and Edward J. Shoben Working Memory Alan D. Baddeley and Graham Hitch T h e Role of Adaptation Level in Stimulus Generalization David R. Thomas


Recent Developments in Choice Edmund Fantino and Douglas Navarick Reinforcing Properties of Escape from Frustration Aroused in Various Learning Situations Helen B. Daly Conceptual and Neurobiological Issues in Studies of Treatments Affecting Memory Storage James L. McCaugh and Paul E. (;old T h e Logic of Memory Representations Endel Tulving and Gordon H. Bower Subject Itidrx

...

Xlll

Toward a Framework for Understanding Learning John D. Bransforcl and Jeffrey ,I. Franks Econoniic Demand Theory and Psychological Studies o f Choice. Howard Raclilin, Leonard Green, John H. Kagel, and Raymond C. Battalio Self-punitive Behavior K. Edward Renner and,leanne B. Tinsley Reward Variables in Instrumental Conditioning: A Theory Roger W. Black Subject Indm

Volume 9 Prose Processing Lawrence T. Frdse Analysis and Synthesis of Tiitorid\ Dialogues AllanCollins, Eleanor H. Warnock. andJosephJ. Passafiunie On Asking People Questions about What They Are Reading Richard C. Anderson and W. Barry Biddle The Analysis of Sentence Production M. F. Garrett Coding Distinctions and Repetition Effects in Memory Allan Paivio Pavlovian Conditioning and Directed Movement Eliot Hearst A Theory of Context in Discrimination Learning Douglas L. Medin SUhJerl I n d a

Volume 10 Sonie Functions of Memory in Probability Learning and Choice Behavior W. K. Estes Repetition and Memory Douglas L. tlintzman

Volume 11 Levelsof Encoding and Retentionof Prose D.Janies Doolingand Robert E. Christiaansen Mind Your p's and q's: T h e Role o f Content and Context in Sonie Uses o f And, Or, and If Saniiie\ Fillenbdum Encoding and Processing o f Symbolic Information in Comparative Judgments William P. Banks Memory for Problem Solutions Stephen K. Reed and Jeffrey A. Johnson Hybrid Theory of Classical Conditioning Frank A. Logan Internal Constructions o f Spatial Patterns Lloyd R. Peterson, Leslie Rawlings. and Carolyn Cohen Attention and Preattention Howard Egeth Subject Indm

Volume 12 Experimental Analysis of Imprinting and Its Behavioral Effects Howard S. Hoffman Memory, Tempoid\ Distxiiiiiilation, and

xiv


Learned Structure in Behavior Processing Charles P. Shimp Robert J. Jarvella T h e Relation between Stimulus Subject Index Analyzability and Perceived Dimensional Structure Barbara Burns, Bryan E. Shepp, Dorothy McDonough, and Willa K. Volume 14 Wiener-Ehrlich Mental Comparison Robert S. Moyer and Susan T. A Molar Equilibrium Theory of Learned Dumais Performance T h e Simultaneous Acquisition o f Multiple William Timberlake Memories Fish as a Natural Category for People and Benton J . Underwood and Robert A. Pigeons Malmi R. J. Herrnstein and Peter A. The Updating o f H tiinan Memory de Villiers Robert A. l3jork Freedom of Choice: A Behavioral Analysis Subject Index A. Charles Catania A Sketch of an Ecological Metatheory for Theories of Learning Timothy D. Johnston and M. T. Volume 13 Turvey SAM: A Theory of Probabilistic Search of Associative Memory Pavlovian Conditioning and the Mediation Jeroen G. W. Raaijmakers and of Behavior Richard M. Shiffrin J. Bruce Overmier and Janice A. Memory-Based Rehearsal Lawry Ronald E. Johnson A Conditioned Opponent Theory of Pavlovian Conditioning and Habituation Individual Differences in Free Recall: When Some People Remember Better Jonathan Schull Than Others Memory Storage Factors Leading to Marcia Ozier Infantile Amnesia Index Norman E. Spear Learned Helplessness: All of U s Were Right (and Wrong): Inescapable Shock Has Multiple Effects Steven F. Maier and Raymond L. Volume 15 Jackson On the Cognitive Component of Learned Helplessness and Depression Conditioned Attention Theory Lauren B. Alloy and Martin E. P. R. E. Lubow, I. Weiner. and Paul Seligman Schnur A General Learning Theory and Its A Classification and Analysis of ShortApplication to Schema Abstraction Term Retention Codes in Pigeons John R. Anderson, Paul J. Kline, and Donald A. Riley, Robert G. Cook, and Charles M. Beasley. Jr. Marvin R. Lamb Similarity and Order in Memory Inferences in Information Processing Robert G. Crowder Richard J. Harris Stimulus Classification: Partitioning Many Are Called but Few Are Chosen: Strategies and Use of Evidence The Influence of Context on the Effects Patrick Rabbitt of Category Size Immediate Memory and Discourse Douglas L. Nelson


Frequency, Orthographic Regularity, and Lexical Status in Letter and Word Perception Dominic W. Massaro, James E. Jastrzembski, and Peter A. Lucas Self and Memory Anthony G. Greenwald Children's Knowledge of Events: A Causal Analysis of Story Structure Tom Trabasso, Nancy L. Stein. and Lucie R. Johnson Index

Volume 16 Skill and Working Memory William G. Chase and K. Anders Ericsson

xv

The Impact of a Schema on Comprehension and Memory Arthur C. Graesser and Glenn V. Nakamura Construction and Representation of Orderings in Memory Kirk H. Smith and Barbee T. Mynatt A Perspective on Rehearsal Michael J. Watkins and Zehra F. Peynircioglu Short-Term Memory for Order Information Alice F. Healy Retrospective and Prospective Processing in Animal Working Memory Werner K. Honig and Roger K. R. Thompson Index

This page intentionally left blank This Page Intentionally Left Blank

William F. Brewer and John R . Pani UNIVERSITY OF ILLINOIS AT URBANA-CHAMPAIGN CHAMPAIGN. ILLINOIS I . Introduction ........................................................... I1 An Example ........................................................... A . The Episode ...................................................... B. Three Types of Memory ............................................ C . Strategy .......................................................... D . Data from Phenomenal Reports ..................................... I11. A Botany of Memory ................................................... A Purpose .......................................................... B. Six Types of Memory .............................................. C Reflections of the Classification in Ordinary Language ................. D Memory Classifications by Psychologists .............................. E Memory Classifications by Philosophers .............................. IV . A Structural Account of Human Memory ................................. A . Purpose .......................................................... B Overview of Structural Account ..................................... C. A Structural Account of Human Memory ............................. V. Structure of Memory: Implications ....................................... A Relation of the Botany of Memory to the Structure of Memory . . . . . B Mental Imagery in the Transfer of Procedural Memory to Semantic Memory ................................................. C Multiple Forms of Representation .................................... D . Copy Images versus Reconstructed Images ........................ V1 Conclusions ........................................................... References ............................................................

.

. . . . .

. . .

.

I

.

i 2 2 3 4 4 5

5 6

9 9 13 14 14 14 18

29 29 30 31 32 33 34

Introduction

The overall purpose of this article is to provide an analysis of the structure of human memory . We shall focus primarily on the process of recall of information from long-term memory . In Section I1 we examine a hypothetical episode in the life of an undergraduate . The episode is intended to provide a clear example of personal memory. a type of memory rarely studied in experimental psychology . It also shows how one episode can give rise to three different forms of memory: personal memory. semantic memory. and rote linguistic skill. I THE PSYCHOLOGY OF LEARNING AND MOTIVATION.VOL . 17

Copyright 0 1983 by Academic Press. Inc. All rights of reproduction in any form reserved. ISBN 0-12-543317-4

2

William F. Brewer and John R. Pani

In Section 111 we develop a “botany” of important naturally occurring forms of memory. We make an explicit methodological commitment to use the phenomenal reports of subjects as one major class of evidence in our analysis. Our description reveals six important types of memory: personal memory, semantic memory, generic perceptual memory, motor skill, cognitive skill, and rote linguistic skill. We contrast this description with the traditional distinction between episodic and semantic memory. We conclude that the term episodic memory, as it is currently used, includes two very different forms of memory-personal memory and skill. In Section IV we provide a more analytic approach to the structure of human memory. We decompose the important naturally occurring types of memory and attempt to construct a table of the logically possible types of human memory. This analysis organizes human memory in terms of the types of inputs and types of acquisition conditions, and proposes an account of the possible forms of memory representation in terms of the intersections of these two factors. A systematic attempt is made to examine both imaginal and nonimaginal forms of representation for each form of memory. The analysis captures a wide variety of types of memory and the forms of representation postulated to underlie each type. In Section V we relate the initial botany of memory to our more analytic classification scheme. We discuss the mental processes involved in transferring information from procedural memory to semantic memory. We point out the complexity that can arise from our assumptions about multiple forms of representations and finally we discuss the problem of the veridicality of mental images. 11.

An Example

The analysis of memory outlined in this article is quite different from most current approaches in psychology, and so to display some of the differences, we shall work through a concrete example illustrating three of the basic forms of memory that will occur in our treatment. A. THEEPISODE

Consider the following event. A University of Illinois undergraduate comes in the side door of the Psychology Building. He takes the elevator to the fourth floor. He pulls a slip of paper out of his pocket, checks the room number, and walks down the corridor to the experimental room. He hesitates a minute, knocks on the door, and goes inside. He sees the experimenter standing behind a table that contains a memory drum. He sits

The Structure of Human Memory

3

down and is given 20 trials of practice on a long paired-associate list. One of the items on the list is the pair DAX-FRIGID. After the experiment is over, he gets up, gives a sigh of relief, and leaves the experimental room.

B. THREETYPESOF MEMORY This episode can be used to illustrate how the same event can lead to the development of three forms of memory: personal memory, semantic memory, and rote linguistic skill.

I . Personal Memory If, the next day, we were to ask this undergraduate, “DO you remember the psychology experiment you were in yesterday?” he might say something like: “Sure, I remember coming in the side door on Sixth Street. I turned to the right and took the elevator up. It was my first experiment. I couldn’t remember the room number so I had to check my experiment notice. I remember feeling nervous as I stood there in front of the door. I remember opening the door and seeing the experimenter standing behind the table. I remember being surprised that the experimenter was a woman. She had a white laboratory coat on, etc.” If we asked the undergraduate, “Was anything going through your mind while you were telling us about the experiment?” he would probably say something like: “Yes, as I was recalling the information I could see in my mind’s eye much of what I told you. I could see the door on Sixth Street. I could see the expression on the experimenter’s face when I opened the door.” It is this type of memory that will be called personal memory in this chapter. 2.

Semantic Memory

If, some months later, we were to ask this undergraduate, “Do you remember what psychology experiments you were in last semester?” he might say, “Sure, there was a verbal learning experiment, a perception experiment, and two social psychology experiments.” If we asked him, “Was anything going through your mind when you told me you were in the verbal learning experiment?” he would probably say something like: “NO, I just know that there were four experiments and one of them was a verbal learning experiment. Now that we are talking about it, I can see the experimenter in her white coat standing behind the table, but nothing like that was happening when I answered your question.” His initial recall is an example of the type of memory that we will call semantic memory.


4

3. Rote Linguistic Skill

We can illustrate the third form of memory by asking our undergraduate to perform the following task: “I am going to give you a series of nonsense syllables from that experiment you were in several days ago, and when I give you a nonsense syllable, I want you to tell me the word that was paired with it.” We then give the undergraduate a series of items from the experimental list including the item DAX. When presented with the nonsense syllable DAX, our undergraduate says “FRIGID.” If we ask him, “Was anything going through your mind when you gave the response ‘FRIGID’?” he might say something like: “NO, I had been over that blasted list so many times that I was able to say it as soon as you showed me the stimulus.” This type of memory is rote linguistic skill.

C. STRATEGY The purpose of this description of the visit to the psychology laboratory has been to provide a detailed example of some of the types of memory that will be discussed later in this article. However, it also illustrates two general strategies adopted throughout this article. We are looking for naturally occurring categories of memory phenomena, and we take the phenomenal reports of subjects as one important class of data to be used in the study of human memory. D.

DATAFROM PHENOMENAL REPORTS

In the last decade there has been a growing acceptance of the position that reports of phenomenal experience can be used in scientific psychology (Ericsson & Simon, 1980; Hilgard, 1980; Natsoulas, 1970, 1974; Radford, 1974). The general line of argument is that phenomenal reports are as acceptable as any other type of data. As long as the data from phenomenal reports enter into lawful relations with other data, and as long as theoretical constructs derived from phenomenal experience interact in a meaningful fashion with other theoretical constructs, there is no reason to exclude them from scientific psychology. We agree with these arguments derived from philosophy of science and from methodological considerations, but we wish to push the issue one step farther. We take the position that a complete scientific psychology must be able to account for the data from phenomenal experience and that an information-processing account of the mind that excluded the data from phenomenal experience would be an incomplete science (for a similar line of argument in philosophy, see Block, 1980; Shoemaker, 1980).


5

It also seems to us that there has been some divergence between the acceptance of phenomenal reports in theory and the actual use of them in practice. Even though cognitive psychology is considered to be a mentalistic psychology, the focus on unconscious mental processes within the information-processing tradition has led to remarkably little serious use of data from phenomenal experience for an early counterexample, (see Dulany, 1968). This avoidance of data from phenomenal experience is very pervasive. Our analysis in this article of some recent experiments by the senior author relies on unsystematic phenomenal reports made by the subjects after the formal experiment was over. Clearly, in the course of gathering the data in those experiments the phenomenal report data were not considered to have the same scientific status as the data on the number of correct responses in recall. Recently we initiated a series of experiments explicitly designed to gather phenomenal reports during a variety of recall tasks (Brewer & Pani, 1982, 1983b). The basic methodology is to ask subjects a memory question (e.g., “What is the opposite of false?” or “Which is farther south, the tip of Texas or the tip of Florida?”) and then request descriptions of their mental experience during recall.

111.

A.

A Botany of Memory

PURPOSE

In this section of the article we take an explicitly morphological approach to human memory. We want to find the common forms of human memory and provide careful descriptions of them, much as a biologist might describe the obvious species occurring on a newly discovered island. In this section we shall not discuss how the types of memory might have developed or what mechanisms might underline their operation. This approach is rarely taken by experimental psychologists, and so to help inform our observations we have explored a number of literatures outside of current cognitive psychology. The particular description of memory that we outline below has been most strongly influenced by: (1) our own introspections; (2) the work of philosophers on memory (e.g., Bergson, 191 1; Furlong, 1951; von Leyden, 1961; Locke, 1971; Malcolm, 1963; Russell, 1921); (3) the early research of introspective psychology (e.g., Crosland, 1921; Humphrey, 1951; Kuhlmann, 1906, 1907, 1909; Titchener, 1910); and (4) current cognitive psychology (Neisser, 1976; Norman, 1976; Tulving, 1972).

We consider our proposed classification to be tentative. One reason for

William F. Brewer and John R. Pan1

6

this is that we do not have systematically obtained phenomenal reports concerning memory, and therefore we have had to rely on our own and others’ unsystematic observations.

B. SIXTYPESOF MEMORY We shall now turn to the botany of memory and describe six types of human memory. Table I gives examples of questions intended to elicit these six types of memory. TABLE I

BOTANYOF MEMORY:EXAMPLES Personal memory “When was the last time you spent cash for something?” “Who was the last person you saw before reading this article?” “Did you see anyone on the ground floor of your office building when you came to work today?” Generic memory

Semantic memory “What part of speech is used to modify a noun?” “What is the opposite of falsehood?” “Which is faster, the speed of sound or the speed of light?” Perceptual memory “In which hand does the Statue of Liberty hold the torch?” “How many windows are there in your house?” “What shape are a German shepherd’s ears?” Skill

Motor skill Typing a sequence of random letters from copy Riding a bicycle Signing your name Cognitive skill Speaking a sentence with a verb in the past tense Adding a column of two-digit numbers Using a text editor Rote linguistic skill Giving your phone number Multiplying 2 x 2 Recalling a list of nonsense syllables


7

1, Personal Memory

A personal memory is a recollection of a particular episode from an individual’s past. Personal memory seems always to be experienced in terms of some type of mental imagery-predominantly visual, since vision is the dominant sense (e.g., “I could see the expression on the experimenter’s face” in the above example). Personal memory includes some nonimaginal information also (in the above example, “It was my first psychology experiment”). The memory is experienced as the representation of a particular time and location. Indeed, it often seems to be a kind of “reliving.” In the case of time, this does not mean that the individual can assign an actual date to the memory, just that it is experienced as having been a unique time. For location, the ability to actually recall a particular place seems much stronger, but data are needed here. The personal memory episode is accompanied by a propositional attitude (cf. Fodor, 1978) that “this episode occurred in the past.’’ A personal memory is accompanied by a belief that the remembered episode was personally experienced by the individual (thus the term “personal memory”). A personal memory is also frequently accompanied by a belief that it is a veridical record of the original episode. This is not to say that personal memories are veridical, just that they are frequently believed to be. We shall discuss the veridicality issues later in the article.

2. Generic Memory A generic memory is the recall of some item of the individual’s general knowledge. Generic memory is not experienced as having occurred at a particular time and location. Two important forms of generic memory are semantic memory and perceptual memory. a. Semantic Memory. Semantic memory is the subclass of generic memory that involves memory of abstract knowledge. Examples are the knowledge underlying the statement, “the speed of light is a constant,” and “I have always avoided abstractions.” Philosophers, logicians, and psychologists have frequently represented this type of abstract knowledge with some form of propositional notation. Recalling information from semantic knowledge is not typically accompanied by an experience of mental imagery. However, if the knowledge required is strongly associated with highly imageable information, one may experience imagery during recall (i.e., in answering the question “What is the capital of France?” one might have an image of the Eiffel Tower). b. Perceptual Memory. Perceptual memory is the subclass of generic memory that involves the memory of generic perceptual information. Ex-

8


amples are the information contained in a generic perceptual memory of a map of the United States or of the capital letter “E.” Recalling information from generic perceptual memory is typically experienced in terms of mental imagery. For example, if asked “What state is directly to the south of Oklahoma?” or “How many corners in a capital letter ‘E’?” most people report experiencing a “generic visual image.” The generic images are not typically experienced as involving a particular time and location. Both personal memory and generic perceptual memory have consistent mental image properties, but they involve somewhat different phenomenal experiences. For example, a generic image will tend to be a figure without an imaginal ground; irrelevant attributes may not be present, and it more often occurs in a single modality. 3. Skill

A skill is the ability t o carry out a practiced motor performance or cognitive operation. When skilled actions are carried out, there is typically no experience of mental imagery. Three important types of skills are motor skills, cognitive skills, and rote linguistic skills. a. Motor Skill. Motor skills can involve the execution of a single motor action or a complex sequence of motor actions. An example of a simple motor skill would be pressing the “K” key on a computer to make the Pac Man figure go right, or pushing the gear shift lever to put a car in reverse. An example of a more complex motor skill would be the skill involved in swimming or playing tennis. Note that the complex motor skills are generative in the sense that if a tennis ball arrives in some unique position, a skilled tennis player can hit it with a motor action never previously produced. b. Cognitive Skill. Cognitive skills involve the execution of practiced cognitive operations. These skills are generative in the sense that the cognitive operations can be applied to a class of new instances, and that class may be indefinitely large. Examples of cognitive skills are taking the square root of a number, and making the subject and verb of English sentences agree in number. c. Rote Linguistic Skill. Rote linguistic skill involves the ability to produce surface structure linguistic objects. This skill differs from motor skills and cognitive skills in several important respects. The skill deals with the meaningless surface structure aspects of particular linguistic objects and it is not generative. Having learned a rote skill is simply to have mastered a given set of surface linguistic objects, and it does not allow generative trans-


9

fer to a new set of surface linguistic objects. Examples of rote skills are the ability to say the alphabet and to give one’s social security number. c .

REFLECTIONSOF THE CLASSIFICATION IN ORDINARY LANGUAGE

A number of philosophers (Locke, 1971; Malcolm, 1963) have suggested that there are linguistic “tests” for the three fundamental categories of memory outlined above (i.e., personal memory, generic memory, and skill). Apparently these memory categories are fundamental enough so that the ordinary language reflects the differences among them. The three linguistic frames are: “I remember X”; “I remember that X”; and “I remember how to X.” Personal memory statements tend to be acceptable in the first frame, but not the second two. Thus, “I remember the expression on the experimenter’s face,” but *“I remember that the expression on the experimenter’s face” and *“I remember how to the expression on the experimenter’s face.” Generic memory statements tend to be acceptable in the first and second frames, but not in the third. Thus, for semantic memory, “I remember the speed of light is a constant” and “I remember that the speed of light is a constant,” but *“I remember how to the speed of light is a constant.” Similarly for generic perceptual memory, “I remember Texas is directly to the south of Oklahoma” and “I remember that Texas is directly to the south of Oklahoma,” but *“I remember how to Texas is directly to the south of Oklahoma.” Motor and cognitive skill statements tend to be acceptable in the third frame, but not the first two. Thus, for motor skills, “I remember how to swim,” but *“I remember swim” and *“I remember that swim.” For cognitive skills, “I remember how to take the square root of a number,” but *“I remember take the square root of a number” and *“I remember that take the square root of a number.” Rote linguistic statements tend to be acceptable in the first and third frames. Thus, “I remember the alphabet” and “I remember how to say the alphabet,” but *“I remember that the alphabet.” Although these tests do not work all the time, the fact that they work as well as they do is impressive. The fact that the ordinary language reflects the memory distinctions provides independent evidence that these are important categories of our mental life.

D. MEMORYCLASSIFICATIONS BY PSYCHOLOGISTS In this section we want to examine two major landmarks in the analysis of memory phenomena: the position of Ebbinghaus (1885/1964) in the first

William F. Brewer and John

10

R. Pani

experimental investigation of memory, and Tulving’s (1972) more recent distinction between episodic and semantic memory. 1. Ebbinghaus

In the first chapter of the Ebbinghaus monograph on human memory, he discusses three forms of memory. He identifies one form that seems closest to personal memory as outlined above. He says: Mental states of every kind,-sensations, feelings, ideas,-which were at one time present in consciousness and then have disappeared from it, have not with their disappearance absolutely ceased to exist. [We] can call back into consciousness by an exertion of the will directed to this purpose the seemingly lost states (or, indeed, in case these consisted in immediate sense-perceptions, we can recall their true memory images). (Ebbinghaus, 1885/1964, p. I ) .

...

Ebbinghaus’ second form of memory was an involuntary type of personal memory which is not relevant to this discussion. The final form of memory outlined by Ebbinghaus was close to skill as it was discussed above. He states, “There is a third and large group to be reckoned with here. The vanished mental states give indubitable proof of their continuing existence even if they themselves do not return to consciousness at all. . . . The boundless domain of the effect of accumulated experiences belongs here” (Ebbinghaus, 1885/1964, p. 2). In the section of the monograph related to the methods of the natural sciences Ebbinghaus (1 885/ 1964) argues that psychologists should study skills because the study of this type of memory requires “less dependence upon introspection” (p. 8). In fact, Ebbinghaus went on to suggest that in studying skill the method of recall was too likely to be influenced by conscious mental processes, and so he chose to use the method of savings (i.e., the improvement in the speed of learning of a task resulting from previous trials with the task). While later researchers decided that Ebbinghaus had been a little too limited in not allowing recall techniques, they essentially accepted the methodologically motivated focus on skill. For 80 years the experimental study of memory was the study of rote linguistic skill, with an occasional study of motor skills (e.g., McGeoch & Irion, 1952; Melton, 1964).

2. Tubing

In the late 1960s a few psychologists (e.g., Collins & Quillian, 1969) were able to break out of the Ebbinghaus emphasis on skill and began to carry out experiments that tested semantic memory. The relationship of these


I1

studies to the traditional verbal learning experiments remained a puzzle for a few years until Tulving’s insightful paper in 1972. Tulving pointed out the differences between the traditional verbal learning experiments and the new semantic memory experiments, and proposed that the differences be formulated in terms of a distinction between semantic memory and episodic memory. a. Semantic Memory. Tulving (1972) states that semantic memory is “the memory necessary for the use of language . . . [the] organized knowledge a person possesses about words and other verbal symbols, their meaning and referents, about relations among them” (p. 386). The definition of semantic memory given in our botany clearly follows Tulving’s usage, although we tend to de-emphasize the focus on linguistic information and instead treat semantic memory as memory of all abstract things. The other major way in which our classification differs from Tulving’s is that we consider that an individual’s overall general knowledge covers more than just semantic memory. Thus, in our classification we have adopted the term generic memory for the broader class of general knowledge (for a similar view see Hintzman, 1978; and Schonfield & Stones, 1979) and retained the term semantic memory for the subclass of memory for abstract things. One important advantage for our approach is that it allows us to treat the important class of generic perceptual information and thereby to incorporate ordinary memory phenomena such as the occurrence of “mental maps.” While we have disagreed with some aspects of Tulving’s construct of semantic memory, one should not lose sight of the importance of this construct in the development of psychological theories of memory. By distinguishing semantic memory from other types of memory Tulving recognized that recall of general knowledge is one important type of memory that must be included in a successful description of the forms of human memory. 6. Episodic Memory. While Tulving’s description of semantic memory clarified thinking about memory for generic knowledge, his account of episodic memory has definite problems. Tulving (1972) states that episodic memory “stores information about temporally dated episodes or events and temporal-spatial relations among these events” (p. 385) and proposes that instances of episodic memory refer “to a personal experience that is remembered in its temporal-spatial relation to other such experiences” (p. 387). It seems fairly clear that when Tulving gives an abstract definition of episodic memory, he is describing personal memory as outlined in our classification. The problem arises when one examines the examples of episodic memory

12


given in his paper. One of four examples was the statement “Last year, while on my summer vacation, I met a retired sea captain who knew more jokes than any other person I have ever met” (Tulving, 1972, p. 386). Taken at face value this appears to be an example of generic memory as we have used the term. The statement seems to refer to Tulving’s knowledge that he met a sea captain during his last summer vacation. A clear example of personal memory would have been a statement such as “I remember sitting on the bar stool, drinking a hot toddy, while he told me the travelling sailor joke, etc.” One of the other of the four examples suggests a deeper problem. This example is “I know the word that was paired with DAX in this list was FRIGID” (Tulving, 1972, p. 387). In terms of our classification this is either an example of generic memory (“I remember that DAX was the syllable paired with FRIGID”) or an example of a rote linguistic skill (given the item DAX the subject produces “FRIGID”). Since Tulving was using this example as an instance of episodic memory, he must not have intended the generic memory interpretation. This leaves the rote skill interpretation. This classification of an instance of rote skill under the heading of episodic memory apparently reflects a general decision on Tulving’s part to classify rote skill as a type of episodic memory, since Tulving (1972, p. 402) explicitly states that traditional verbal learning experiments are to be considered to be experiments investigating episodic memory. Thus, in terms of the memory classification we have outlined above, Tulving’s treatment of episodic memory is inconsistent. Tulving’s formal definition of episodic memory seems very close to our definition of personal memory, yet the examples given and the classification of the traditional laboratory experiments as instances of episodic memory are inconsistent with his definition. Nevertheless, virtually every psychology text written since Tulving’s classic paper quotes his definition of episodic memory, and then states that the memory experiments in the Ebbinghaus tradition are all examples of episodic memory. Examination of our initial example of the undergraduate going to the psychology experiment shows the problems produced by this inconsistency with respect to personal memory and skill. Our hypothetical undergraduate had 20 trials on a long paired-associate list that resulted in the development of the rote verbal skill of producing the responses when given the stimuli. We argued that the undergraduate would probably have a strong personal memory of coming to the building and starting the experiment, but it seems to us highly unlikely that the undergraduate could have a personal memory for a particular trial, say Trial 13, in the series. It would appear that the conditions for the development of skill are, in fact, antithetical to the development of personal memory (this issue will be discussed again later in the article). Thus, it seems to us that the treatment of episodic memory in


13

current discussions of human memory contains a conceptual inconsistency and that an analysis more like the one we have proposed is needed to resolve this inconsistency.

E. MEMORYCLASSIFICATIONS BY PHILOSOPHERS In carrying out investigations of memory, philosophers have tended to use a more differentiated classification scheme than that of psychologists. The first modern philosophical discussion of the issues is that of Henri Bergson (191 1). Bergson distinguished two forms of memory, “memory par excellence” and “habit memory”; these correspond fairly closely to our personal memory and skill memory. Bertrand Russell (1921) retained the division of memory into two forms. His “true memory” and “habit memory” correspond closely to our personal memory and skill memory. In somewhat more recent times a number of philosophers added memory for knowledge into their classification schemes and have adopted a distinction that corresponds to our personal memory, semantic memory, and skill memory. Furlong (1951) uses the terms “retrospective memory,” “nonretrospective remembering that,” and “nonretrospective remembering how.” Ayer (1956) uses the terms event memory, factual memory, and habit memory, while Locke (1971) adopts the terms personal memory, factual memory, and practical memory. For the most part these theoretical discussions of memory by philosophers have had little impact on psychological research. However, it is interesting to note that the one recent revision of memory classification, that of Tulving (1972), may derive indirectly from the philosophers. In Tulving’s paper he gives credit to an earlier distinction between “remembrances” and “memoria” by Reiff and Scheerer (1959). This distinction corresponds roughly to our personal versus nonpersonal memory. Examination of the section of the Reiff and Scheerer (1959) monograph on memory distinctions shows that they based their treatment on the early work of Bergson, thus showing a fairly direct link between the episodic/semantic distinction and the philosophical tradition. There is a striking contrast between psychology and philosophy in what types of memory have been the focus of interest. Most of the first 80 years of research on memory in psychology were directed at the problems of rote verbal skill. (There were exceptions, such as the work on memory of the Wurzburg psychologists, of the Functionalists in the United States, and of the Gestalt psychologists.) The emphasis on skill by psychologists was driven by methodological and metatheoretical considerations. The study of semantic memory seems to require the introduction of abstract entities; and the study of personal memory seems to require the introduction of mental


14

images; neither of these was acceptable to most memory researchers during this period. The research of Collins and Quillian (1969) and Rumelhart, Lindsay, and Norman (1972) opened up the study of semantic memory in psychology, and the present article argues for empirical work on personal memory. The philosophers have taken a very different approach. In general, they have tended to find skill the least interesting form of memory. Initially, with the work of Bergson and Russell the focus was on personal memory. Thus, for example, Russell (1921) called personal memory “the essence of memory” (p. 167). In the more recent work philosophers have continued to discuss the problem of personal memory and its degree of veridicality, but they also have focused on the problems of memory for knowledge. Our conclusion from this brief historical sketch is that current experimental and theoretical work on memory by psychologists should be more pluralistic. In particular, more attention should be given to the study of personal memory.

IV. A.

A Structural Account of Human Memory

PURPOSE

The purpose of this section is to develop a more analytic account of the structure of human memory. In this section we attempt to work out the logical possibilities of the forms of human memory instead of simply describing a number of types that occur in our normal interchange with the world. We also intend that our structural model reflect some aspects of the processes that lead to various types of memory representations. Finally, we try to follow our own suggestion and take the data of phenomenal experience as a fundamental aspect of a description of the structure of human memory. B.

OVERVIEW OF STRUCTURAL ACCOUNT

The essence of our organization of memory is given in Table 11. This table is structured with types of input to the memory system along the top and types of acquisition conditions along the side. Within are the hypothesized mental events resulting from the conjunction of the particular type of input and type of acquisition condition. 1. Acquisition Conditions (Rows)

We consider three important types of acquisition conditions: exposure t o a single instance of the input, exposure to multiple instances of the input


1s

without variation (e.g., ten exposures to the same picture or ten trials on the same serial list of nonsense syllables), and exposure to input that is repeated with variation. The category of repetition with variation is intended to cover a range of levels of abstraction of the input. Thus, multiple exposures to the same dog on different occasions in various circumstances would be repetition with variation and so would the more abstract level provided by exposure to a number of instances of different types of dogs. When analyzing the types of memory in the single instance condition we will assume that the memory tasks will be directed at information specific to the single instances. In analyzing the types of memory in both of the repeated items conditions, we shall assume that the memory tasks will be directed at recall that utilizes the experiences with the set of repeated items and not at one of the instances. In each input condition we have divided the resulting mental events into imaginal events and nonimaginal events. This is motivated by our desire to treat the phenomenal data as a serious part of theory construction. Note that the division is between imaginal and nonimaginal and not between phenomenal and nonphenomenal. We adopted this approach primarily for methodological reasons. We believe that there are phenomenal states that are nonimaginal (e.g., the imageless thoughts of the Wurzburg psychologists, Woodworth, and Binet; see Calkins, 1909; Humphrey, 1951; Ogden, 1911). However, this is a difficult area and the data have been hard to interpret (see Pani, 1983). Thus, until clarifying data are obtained on this issue, we shall restrict our analysis primarily to phenomenal reports of mental images, where the data are clearer and easier to obtain. In most cases the types of mental representations we postulate for the nonimaginal cells are schemata (Brewer & Nakamura, 1983; Minsky, 1975; Rumelhart, 1980; Schank & Abelson, 1977). Schemata are nonphenomenal mental representations of organized knowledge. When an input occurs and activates a schema, then the organized knowledge can be related to the input. This process makes possible: expectations, inferences, and active anticipations. The term “schema” will be used to cover a wide range of knowledge structures-from object schemata that allow one to infer what the nonvisible side of an object might look like, to motor production schemata, which allow the smooth output of a particular motor action. This discussion of schemata raises an interesting problem. How do these abstract schemata differ from the “abstract knowledge” that we referred to in our discussion of semantic memory? It may be that these two types of mental representation should actually be considered to be of one type. However, we would like to distinguish between them. We propose that semantic memory is knowledge of abstract things, whereas schemata are abstract knowledge of things.

TABLE I1 A STRUCTURAL ACCOUNT OF HUMAN MEMORY Types of input Visual-spatial Acquisition conditions

Single instance Imaginal

Nonimaginal

Meaningful (objects, places) Particularized visual images

Particularized visual images

Instantiated schemata

Partially instantiated schemata

Repeated without variation Imaginal Reduced particularized visual images Nonimaginal

Repealed with variation Imaginal Nonimaginal

Meaningless

Instantiated schemata and development of rigid schemata

Particularized visual images

Partially instantiated schemata and development of rigid schemata

Generic visual Generic visual images images Instantiated Schema schemata and development schema development

Visual-temporal Meaningful (events, actions)

Auditory-nonlinguistic

Meaningless

Meaningful (common sounds)

Meaningless

Particularized sequence of visual images Instantiated schemata

Particularized sequence of visual images Partially instantiated schemata

Particularized auditory images


Instantiated schemata

Partially instantiated schemata

Reduced particularized sequence of visual images Instantiated schemata and development of rigid schemata (scripts)

Particularized sequence of visual images

Reduced particularized auditory images Instantiated schemata and development of rigid schemata


-?-

Instantiated schemata and schema development (plans)


-7-

-7-

Generic auditory images Instantiated schemata and schema development


Generic auditory images Schema development

Types of input Linguistic Acquisition conditions Single instance Imaginal

Nonimaginal

Repeated without variation Imaginal

Emotional situations

Meaningful (expository discourse)

Emotion images Emotion schemata

None

-7-

None

-3-

Facts Propositions

-7-

None

-7-

Thoughts

Facts Propositions

Meaningless (surface structure) Auditory or visual images Incomplete surface structure production schemata

Cognitive operations -?-

Motor performance -7-

Plan production -?-

Incomplete cognitive production schemata

Incomplete motor production schemata

Little or no imagery



Surface structure production schemata

Rigid cognitive production schemata

Rigid motor production schemata

Rigid plan productions (awareness of intentions and goals) (scripts)

Little or no imagery Generative cognitive production schemata

Little or no imagery Generative motor production schemata

-?-

Plan productions (awareness of intentions, goals and subgoals)

Nonimaginal

Repeated with variation Imaginal

Nonimaginal

Plan production (awareness of intentions and goals)


18

2.

Types of Input (Columns)

The columns in the analysis are organized according to our view of the fundamental types of input that lead to the various forms of memory representation. In those cases where the input involves forming a representation of a content from the external world, we have subdivided the input into meaningful stimuli and meaningless stimuli. We do this because there are important differences between memory for meaningful stimuli, which are easily encoded into preexisting schemata, and memory for meaningless stimuli, which are more difficult to encode into such schemata. In addition, the distinction is of practical value, since how one interprets a particular memory experiment is frequently determined by the nature of the input with respect to this distinction.

c.

A STRUCTURAL ACCOUNTOF HUMANMEMORY

In this section we examine the hypothesized mental events resulting from the conjunction of the particular types of input and types of acquisition condition. We shall proceed through the table by types of input (i.e., column by column).

I.

Visual-Spatial

a. Meaningful. We postulate that a single exposure to a meaningful visual-spatial input will lead to a particularized visual image. This is, of course, one of the strong components of personal memory as described in the botany of memory. In his classic review of introspective methods for the study of mental imagery, Angel1 (1910) suggests that brief exposure to an arbitrary array of objects or pictures, followed by a memory task requiring information about the concrete properties of the display, is one of the best ways to elicit visual imagery (also see Kuhlmann, 1909). There have been a number of recent experiments examining memory for single exposures to meaningful visual-spatial input (e.g., Brewer & Treyens, 1981; Hock & Schmelzkopf, 1980; Mandler & Parker, 1976), but these experiments rarely include data concerning the phenomenal experiences of the subjects during the recall task. However, from the informal comments of the subjects in the Brewer and Treyens (1981) experiment on memory for rooms and from the fact that they sometimes pointed to an imaginary position in space when answering a question, we think that appropriately designed experiments will support the assertion that this type of input leads to particularized visual image representations.


19

The mental images associated with personal memory appear to be very vivid and to include much “irrelevant” detail. It is not clear that one can show increased recall of information based on these mental experiences. However, if one could show such evidence in recall, then one might want to hypothesize that the representations for personal memories are less reworked by schema processes and somehow closer to the initial perceptual input than other forms of recall. This difficult and controversial issue is clearly in need of additional study. Exposure to a single instance of a meaningful visual-spatial input leads to schema instantiation. The individual uses generic schema information to interpret the particular visual-spatial input. The resulting instantiated schema representation consists of an integration of the information contained in the new instance and the information from the generic schema. Thus, in the Brewer and Treyens (1981) experiment subjects attempted to recall an office in which they had been for a brief period. The information given in recall was a mixture of information that was clearly from the particular room (e.g., it contained a Skinner box) and information that was not actually in the particular room, but was derived from their general office schema (e.g., it contained books). Note that the case of schema instantiation is part of the larger issue of the interaction of particular input with generic knowledge. It is likely that both generic visual images and generic schemata can interact with the information contained in input from single instances to produce partially reconstructed visual images and instantiated schemata (e.g., Neisser, 1982, pp. 43-48). Multiple exposures without variation of a meaningful visual-spatial input should lead to a more articulate image. However, if there is a consistent focus of attention on particular items or properties, then the meaningful nature of the material may lead to a reduction from the image of less relevant properties. Thus, while we would expect context to occur in the visual-spatial component of a personal memory, it may not always remain in cases of multiple exposures. If an object or class of objects is repeated with variation, we postulate that a generic visual image results. This is a topic that needs research. For highly variable classes such as “furniture” it seems unlikely that one forms a generic image (e.g., Rosch, Mervis, Gray, Johnson, & Boyes-Braem, 1976); for other less variable classes (e.g., “dog,” “triangle”) more data are needed. Essentially we are assuming that the process of abstraction (e.g., Gibson, 1969) leads to the production of generic images from experience with multiple differing particulars. This suggests an interesting problem. Does the process of abstraction lead to a generic image and multiple particular images, or is there loss of the particularized images? Medin and Schaffer (1978) have proposed a theory

20


of concept classification that emphasizes recall of specific instances. Brewer and Dupree (1981, 1983b) have carried out an experiment on the development of generic place representations and have found that there is apparently some loss of the particularized instances that go into making up the generic representation. Clearly this is another area that needs additional research. When meaningful visual-spatial inputs are repeated with variation, we assume the development of schemata in addition to generic images. The reason that we postulate nonimaginal schema representations for objects and places is that we do not think generic images are abstract enough to account for much of our visual-spatial knowledge. We feel that there is abstract knowledge about objects and places that is somehow specific to them and not a part of semantic memory. Thus, for objects we would consider the classic Piagetian object schema to contain a nonimaginal schema component. For places, consider the following question, “Which is closer, your bathroom or the post office?’’ It seems to us that one may answer this question with nonimaginal place schema information and without generic image information or information from semantic memory. These are difficult problems, and clearly more theory development is needed to make progress on these issues. b. Meaningless. We turn now to the second column in Table 11. A meaningless visual-spatial input is one that has little schema information already existing in long-term memory. We assume that exposure to stimuli of this type results in an attempt to build a new schema or impose an old schema (see Piaget, 1952, 1954; Rumelhart & Norman, 1977). Our position here is similar to that taken by Bartlett (1932) in his discussion of the use of nonsense syllables in memory experiments. Bartlett argues that when €aced with “meaningless” material subjects would attempt to impose meaning on the stimuli. He referred to this process as “effort after meaning” (Bartlett, 1932, p. 20). In more recent times investigators in the Ebbinghaus tradition have shown, in some detail, the powerful effects that effort after meaning has on learning meaningless linguistic material (Montague, Adams, & Kiess, 1966; Prytulak, 1971). The property of meaningless items that makes them meaningless is that the imposition of prior schemata is only partially successful. Thus, to the degree that schema instantiation is inadequate, material must be newly learned from immediate perception. Several investigators have pointed out that the memory for such cases should be relatively imagistic and depictive, since no other form of memory representation is available (Kosslyn, 1980, 1981; Kosslyn & Jolicoeur, 1980; Pani, 1982, 1983). The occurrence of repetition with variation of instances of an initially meaningless pattern should lead to the development of a new generic image


21

and a new schema. The classic experiments of Posner (Posner, 1969; Posner & Keele, 1968) on recognition memory for dot patterns generated from an underlying pattern represent an attempt to study this process. Posner (1969) suggests that the form of representation in these tasks may be abstract nonimaginal schemata. From our perspective it would be interesting to carry out these experiments in a recall paradigm, with a variety of types of schemata, and obtain phenomenal reports from the subjects during recall. Do the subjects report generic mental images, images of particular instances, or no images at all? 2.

Visual-Temporal

a. Meaningful. We now begin the third column. We assume that a single exposure to a meaningful visual-temporal event leads to the development of nonimaginal schemata. Observed causal events lead to event schemata and observed goal-directed actions lead to plan schemata (cf. Schank & Abelson, 1977). An empirical study of memory for goal-directed actions by Lichtenstein and Brewer (1980) supports our assumptions on this issue. These investigators had subjects view a videotape of an actor carrying out a series of goal-directed actions and showed that a wide range of recall data could be accounted for by assuming that the subjects had developed plan schemata for the observed actions. There are few data on the imaginal properties resulting from observed visual-temporal events. Lichtenstein and Brewer (1980) did not, unfortunately, gather any systematic data on this issue. However, some informal observations during those experiments suggest that few subjects report the phenomenal experience of being able to “replay in their mind’s eye” a smooth version of what they saw. Instead, they tended to report sequences of static images with some limited movement. An early discussion of mental imagery by Ladd (1894) supports this view. Ladd believed that the progressive “condensation” of imaginal representation extended in time is a fundamental principle of the development of cognition. Pani (1983) has suggested that the deletion of redundant material from the imaging of visual-temporal input would result in savings of time and effort. These claims are reminiscent of the views of Attneave (1954) and Hochberg (1968) on the perception of visual-spatial structure. They point out that there are particular parts of items that convey relatively large amounts of information about the nature of an object, and other parts that are relatively uninformative. This suggests that the remaining information in the imaginal representation of visual-temporal inputs consists of images of the more informative stages of events. On the basis of these various considerations we have tentatively assumed

22


that single exposures to meaningful visual-temporal events lead to sequences of static visual images, rather than to single temporally continuous images. This view contrasts with the inclinations of much current imagery theory (e.g., Kosslyn, 1980; Shepard, 1978), although Kosslyn does discuss the “blink” transformation. Clearly more data are needed here. If a meaningful sequence is repeated without variation, we assume that it may be converted into a more rigid script. Examples of events repeated with little variation might be religious rites and mechanical processes. Again, we have no data relevant to the issue of the imaginal consequences of repetition without variation. However, it is possible that condensation may continue for irrelevant properties. We would also suspect that repetition would lead to stronger visual imagery for the information that is retained. For meaningful events repeated with variation we assume the development of more abstract plan and event schemata (Schank & Abelson, 1977; Schmidt, 1976). It is not clear what the imaginal properties would be for this condition. It is likely that even after a great deal of condensation, highly informative generic reference points remain imageable. However, it also is possible that events can differ enough among themselves so that they are encoded at a level of abstraction that cannot be captured in an image. There is a subset of meaningful visual-temporal events that we wish to distinguish-memory for personal actions. After an individual has carried out some goal-directed action, the individual can recall what he or she did. This type of recall seems similar to the recall of the actions of another person. However, the actor has direct access to the actor’s plans, to knowledge about intentions not acted upon, and to other aspects of conscious mental life that occurred during the action. We shall provisionally assume that memory for personal actions can be treated as essentially similar to memory for the observed actions of others. b. Meaningless. We assume that subjects exposed to meaningless visual-temporal events will attempt to impose causal and plan schemata on the events. However, there will be numerous cases in which unique perceptual properties of a particular event are remembered. As we have argued before, imagery may be the primary way in which such properties as these will be represented at first. A recent experiment by Brewer and Dupree (1983a) supports these assumptions. Brewer and Dupree obtained data showing that subjects attempted to provide plan schemata for relatively meaningless actions and that when they did it improved recall. No phenomenal report data were obtained in this experiment; however, the recognition data and the overall pattern of results can be used to draw inferences about the imagery for meaningful and meaningless actions. It is possible that after viewing an action there is visual information that is retained over a period of hours, but that after several days the information


23

is greatly reduced and underlying plan information is predominantly what is retained.

3. Auditory (Nonlinguistic) The auditory input conditions were filled in by analogy with the visual columns. We have little specific to contribute to the analysis of this type of input, and have included it primarily for consistency. However, there are a number of studies that suggest that the auditory columns will be analogous to the visual columns (e.g., Garner, 1974; Williams & Aiken, 1975). A complete description of memory would also include an account of memory for music. This is a complex issue. For example, we suspect that an analysis of music should share some of the characteristics of our analysis of linguistic input. However, we know so little about these issues at the present that we are not willing to speculate. 4. Emotional Situations

In an earlier version of this article we omitted memory for emotions because so little is known about the topic. However, we have decided to include it because just making the attempt seemed to force us to ask interesting questions. In one of the few recent discussions of memory for emotions, Bower (1981) proposes that memory for emotions should be analyzed in terms of “emotion nodes.” Given the framework adopted in this article, there are additional issues that must be resolved. Does the memory for an emotion have an emotion reliving component (an “emotional image”), and does the memory for an emotion have a nonimaginal “emotion schema” component? There are formidable problems here. For example, if one attempts to carry out introspective studies of memory for emotion, it is necessary to distinguish the current emotions from the recalled emotions. The problem arises owing to the fact that recalling a situation that made you angry can cause you once again to become angry about the situation. Try recalling “your most embarrassing moment” for an intuitive example of the difficulty. A second issue is a theoretical one. What does it mean to talk about “emotion schemata”? Clearly one can come to have semantic knowledge about any type of input in our table. Thus, one can explicitly enter into semantic memory the fact that “the state to the south of Oklahoma is Texas.’’ Similarly, one can have semantic knowledge that “I was angry when I received the letter last week.’’ The theoretical puzzle is whether it makes any sense to postulate something called an emotion schema inde-

24


pendently of the knowledge that you felt a particular emotion. The most sophisticated treatment of this issue that we know of is by St. Augustine. In the Confessions, Augustine discussed the representation issue that we just outlined, and he rejected (on the basis of his own introspections) the view that one relives an emotion when remembering it. However, he also rejected the view that one simply has semantic knowledge of the emotions. He postulated a third form of representation, “notions,” to deal with the problem. In our terms, he was apparently suggesting that memory for emotions consists of emotion schemata without emotion images or semanitc memory of emotions. Obviously, the issue of memory for emotion is in need of empirical and conceptual clarification. 5. Linguistic

a. Special Properties of Language. Memory for linguistic input is the most thoroughly studied area in the experimental study of memory. It is, however, one of the most subtle in terms of the structure of memory. First, one has to take into account that language input can be used to convey many different types of information. This means that, in remembering what was conveyed by a linguistic input, the memory representations themselves may not be linguistic in form. Brewer (1980) has argued that descriptive discourse is represented in terms of visual-spatial schemata, while narrative discourse is represented in terms of plan schemata. There is considerable experimental evidence that can be interpreted to support this assertion. In a study by Bransford, Barclay, and Franks (1972) subjects given sentences describing objects in simple spatial arrangements produced recall data that was similar to what one might have expected if the subjects had actually seen pictures of the scenes described by the sentences. For narrative discourse, Lichtenstein and Brewer (1980) have carried out an explicit test. In Experiments 4 and 5 of that paper subjects were given narratives describing a series of goal-directed actions. The pattern of recall data was essentially the same as that produced by subjects who saw actual videotapes of the goal-directed actions. Thus, for our purposes, any time linguistic input conveys information characteristic of some other type of input, we shall assume that the form of representation in memory is the form postulated for that type of input (e.g., visual-spatial for linguistic descriptions). For example, in terms of Table 11, this approach means that meaningful linguistic input of narrative form should be analyzed as if it occurred in the visual-temporal (meaningful) input column. Note, however, that we d o not assume such a shift for expository text. Brewer (1980) has argued that the underlying representations for expository text are abstract propositions or thoughts. To put it another way, expository text is linguistic input that encodes semantic memory information.


25

A second way in which meaningful linguistic input differs from the other forms of input is that we assume that, in addition to perceptual images (e.g., the sound of a word), there are two abstract levels of representation arranged in a hierarchical fashion: surface structure production schemata, and thoughts. In particular, we are making the assumption that there is a separate abstract level of representation that is a nonimaginal surface structure production schema. This allows the overt recall of a nonsense syllable such as DAX without imaging that syllable first. This contrasts with the view that surface structure must be encoded in terms of auditory or visual images. Thus, from our perspective the auditory occurrence of a meaningful word such as “truth” leads to three levels of representation: an auditory image, a surface structure production schema, and the thought (meaning) expressed by the word. b. Meaningful. Our mode of analysis leaves memory for expository linguistic input with no image properties. We have assumed that the basic meaning for this material is in terms of abstract nonimaginal thoughts. Imaginal representation of the visual or auditory properties of the presented word would be treated under memory for surface structure information. If the word were concrete and gave rise to visual imagery, then that would be treated under memory for visual-spatial information. As mentioned earlier, we believe that one may want to consider the existence of phenomenal but imageless thoughts for this type of representation, but for now we shall ignore that possibility. Our analysis of the representation of meaningful linguistic input in terms of an abstract nonimaginal representation is consistent with the standard approach in current cognitive psychology (e.g., Anderson & Bower, 1973; Bransford & Franks, 1971; Kintsch, 1974), and with the earlier work on imageless thoughts (Calkins, 1909; Humphrey, 1951; Ogden, 1911). There is a large experimental literature on memory for thoughts supporting the position outlined above (Anderson, 1974; Bock & Brewer, 1974; Brewer, 1975; Graesser & Mandler, 1975; Sachs, 1967). The typical approach in these experiments is to give subjects a sentence memory task (recall or recognition) and to show that the subjects retain the ideas expressed in the sentences even when they d o not retain the particular input surface form. We have included abstract nonimaginal representations under the heading of input from meaningful expository linguistic discourse because that is probably the most frequent form of input for these processes. However, we believe that information can enter the system of abstract nonimaginal representations (semantic memory) through a variety of nonlinguistic interchanges with the world and through internal reworkings of the information already in the system. When one has a single exposure to a trivial piece of knowledge, it is easy

26


to become confused about the appropriate form of representation (thought versus surface structure). Take the example of someone who has learned the names of the state capitals so that when given “Illinois,” this individual says “Springfield.” We would argue that this performance typically requires two levels of representation (surface structure and thought) and so should be distinguished from the case of learning to say “DAX” when given “ZEQ.” Evidence that some abstract factual knowledge had been acquired in the first case would be shown by the individual’s ability to paraphrase the information and draw inferences. Thus, having learned Illinois-Springfield, our subject could paraphrase the information, “Springfield is the capital of Illinois” or “the capital of Illinois is Springfield.” Similarly the subject should be able to make the inference that “Chicago is not the capital of Illinois.” Note that for someone who did not know English the ability to give “Springfield” when presented with “Illinois” would be merely an example of memory represented in the form of a surface structure production schema. The effect of repetition with variation for meaningful linguistic input can have a different effect than it has for the other inputs. There is not necessarily a shift toward more abstract representations. With even a single instance of expository language the initial representation is already abstract, and its content may be extremely abstract (e.g., “religion stems from the need to know”). When there is repetition with variation, there is an increase in the richness and complexity of the representation. c. Meaningless (Surface Structure), We postulate that after exposure to a single instance of a meaningless linguistic input individuals have an auditory or visual image representation. Some of the early introspective studies of memory support this position (Fernald, 1912). In addition to image representations, a single exposure to a meaningless linguistic input leads to the beginning of a surface structure production schema. The development of surface structure production schemata for more than a few items is a skill that takes a number of repetitions to develop. The contrast between the ability to produce thoughts and to produce surface structure schemata was noted by Ebbinghaus (1885/1964, p. 50), and was studied by a number of investigators as the difference between “logical” and “rote” memory (Cofer, 1941; Welborn & English, 1937). For purposes of clarity we have been using examples of meaningless linguistic input to discuss the development of surface structure production schemata. However, because of the hierarchically organized nature of the two forms of linguistic representations, one can also investigate the development of surface structure production schemas for meaningful sentences. There is a wide range of studies showing that for meaningful sentences the memory for the underlying thoughts is better than the memory for the surface structure (e.g., Brewer, 1975; Sachs, 1967).


21

Consideration of what it means to repeat a surface structure with acceptable variation (change the type face, shift speakers) shows that this type of input does not lead to the same level of abstraction as the other inputs. There is one area where rote linguistic skill is very important. Each native speaker of a language has to master the tens of thousands of lexical forms that make up the vocabulary of the language. Clearly the ability to develop surface structure production schemata plays a crucial role in learning a spoken language.

6. Cognitive Operations Cognitive skills involve the execution of practiced cognitive operations. Cognitive skills differ from rote skills in that cognitive skills are generative and rote skills are not. Once an individual has learned a cognitive skill, that individual can typically apply it to a large class of new objects (e.g., taking square roots); but once an individual has learned a rote skill, that individual has the ability to produce only one set of surface structure objects. The rote skill of saying the multiplication table in English does not allow one to say the alphabet. The distinction between cognitive skill and recall of information from semantic memory can sometimes be unclear. We tend to classify a task as an instance of cognitive skill if the task is procedural, if it is knowledge how rather than knowledge that. The difference is clear in the case of the rules of syntax of one’s native language. A child has the cognitive skill of performing many syntactic operations before entering school and in the course of formal education the child comes to develop knowledge that about some of the rules. This is presumably the distinction that Chomsky (1965) was intending when he stated, “a generative grammar attempts to specify what the speaker actually knows, not what he may report about his knowledge” (p. 8). There recently has been a renewed interest in the study of cognitive skill in psychology (Anderson, 1981; Card, Moran, & Newell, 1980). Cognitive skills, like the other skills, require a number of repetitions before smooth, successful operation. Thus, it is difficult to discuss the representation that results after a single operation of a cognitive skill. However, in a recent study of the early stages of learning to use a text editor, Ross (1982) has obtained verbal protocols suggesting that the subjects attempt to supplement the missing cognitive skill with other types of knowledge. They use personal memories, “Oh yes, I remember when I pressed that button over there, the whole screen went blank” and semantic knowledge “Let’s see, the rule is that to change a word in the text, select the word, press capital ‘R,’ type the new word and press the ESC key.” One has the feeling that

28


what is going on here is similar to Bartlett’s “effort after meaning,” perhaps “effort after production.” When cognitive operations have been repeated many times, there is little or no imaginal accompaniment (Book, 1908). It is presumably this observation that led Lashley (1960) t o state: “No activity of mind is ever conscious” (p. 532). It may be the case that cognitive operations are a type of mental occurrence that is intrinsically nonphenomenal. On the other hand, it may be that they are phenomenally experienced only during the early stages of the acquisition of a skill and not later (see Pani, 1983). 7. Motor Performance The issues relating to the acquisition of motor skill are similar to those discussed for cognitive skill. Many investigators in this area have suggested that, when motor actions have been practiced, the conscious correlates of performing the action are reduced or eliminated (Adams, 1971; Book, 1908; Fitts & Posner, 1967). The classic discussion is in James’ chapter on habit. He states, “habit diminishes the conscious attention with which our acts are performed” (James, 1890, Vol. 1, p. 114). While there is agreement that conscious processes occur during the early stages of the acquisition of a motor skill, it is not clear exactly what types of processes these are. For example, they may be motor imagery, imageless thoughts, or other types of memory representation used in “effort after production.” 8. Plan Production

The carrying out of intended activities involves the production of complex sequences of actions (e.g., driving to a new restaurant). We assume that these intentional actions are structured in terms of plan production schemata. Plan production schemata organize actions in terms of hierarchically structured goal-subgoal relations. Plan production is intended to allow us to include the memory component that is involved in walking home from the office, baking a cake, dancing in a square dance, for example. One might want to argue that plan production is a complex mixture of cognitive operations and motor performance, but we prefer t o treat plan production as a separate category. Carrying out a single instance of a new plan seems to be a memory task in only a limited fashion. In performing a new plan, say finding your way for the first time from Heathrow airport to the British Museum, much of the performance seems to be problem solving with very general generic memory input. It is not clear how much imaginal activity occurs when one carries out a new plan of this type, but there does appear to be a large


29

amount of nonimaginal phenomenal experience. It seems to us that one is aware of intentions, the goal, and many subgoals (e.g., “I need to get from here to the museum. . . I wonder how I can get my money changed. . . . How do I get to the underground, etc.”). In plans carried out with little variation (taking the same route home from the office every day), it would appear that the awareness of the subgoals and subplans tends to decline (Shallice, 1972). It seems likely that it is these fixed plans that are most likely t o lead to “actions slips” (Norman, 1981) where the individual carries out an action that was not intended. Carrying out a variety of intentional actions of a given type leads to the development of generic plan schemata (e.g., going to restaurants, traveling to new cities).

.

V.

Structure of Memory: Implications

In this section of the chapter we shall relate the earlier, more descriptive botany of memory to our analysis of the structure of memory. We shall also work out some of the implications of our structural account for a number of particular issues in the study of human memory. A.

RELATIONOF T H E BOTANYOF MEMORY T O T H E STRUCTURE OF MEMORY

Our intent in outlining the botany of memory was to describe common types of human memory. Our intent in the structural account was to give an analytic account of possible types of human memory along with some indication as to how different forms of memory are acquired. We think that the types of memory discussed in the botany are the ecologically important subset of the possible types of memory given in Table 11. They comprise the subset that tends to occur in the normal ecological interactions with the environment. In moving around in the world one tends to be exposed t o many unique co-occurrences of meaningful visual-spatial input, meaningful visual-temporal input, meaningful auditory input, and linguistic input. It is roughly this set of representations (the single-instance rows in Table 11) that go into making up personal memory. In our dealings with the world, and in particular in our dealings with the products of culture, we are exposed to much abstract knowledge (facts, propositions, thoughts). It is this type of knowledge that constitutes semantic memory (the meaningful linguistic input column). In moving through the visual world, we tend to view constant objects and constant places in the environment from a variety of perspectives. It


30

is this type of interaction with the world that leads to the development of generic visual memory (the repeated with variation rows in the visual-spatial input columns). In learning to speak a language and in memorizing nonsense syllable lists for experimental psychologists, we develop the surface structure production schemata that make up rote linguistic skill (the repeated without variation rows in the meaningless linguistic input column). In carrying out some of the complex repetitive processes that are part of modern civilization (arithmetic, text editing) we come to develop cognitive skills, and finally when we repeatedly manipulate objects in the world we come to develop motor skills. Thus, by taking the analysis of the structure of memory, and looking at naturally occurring human actions, we find the botany of memory to be a natural consequence of the operation of the human memory system and the normal organism-environment interactions in our culture. MENTALIMAGERY I N THE TRANSFER OF PROCEDURAL MEMORYTO SEMANTIC MEMORY

B.

In our analysis of memory we noted that the knowledge involved in practiced skills is represented in production schemata and little imaginal experience is reported during a skilled performance. However, in the course of pilot work for an experiment on phenomenal experience during memory (Brewer & Pani, 1983b), we have uncovered an interesting class of mental processes. If one asks a subject for a propositional account of information that “resides in” procedural (skill) memory, then there is a striking occurrence of appropriate mental imagery. Thus, for example: a.

Rote linguistic skill (1) “What is the seventh letter of the alphabet?” (2) “What is the next t o last digit of your phone number?” b. Cognitive skill (1) “What is the sum of 78 and 43?” (2) “What are the last three letters of the plural of irony?” c. Motor skill (1) “Which finger do you use to type an ‘r’?” (2) “When backing a car which direction do you turn the steering wheel in order to make the back of the car go to the left?” It appears that in these cases one is able to divide one’s conscious mental processing into two parts. One part of the mind carries out the procedural task in imagistic form, and the other part of the mind notes the contents

The.Structure of Human Memory

31

of the images and gives the required propositional answer. It seems to us that this class of phenomena shows the qualitative difference between knowledge how and knowledge that, and suggests that the mental imagery might play a functional role in performing the memory task. In the course of everyday life one rarely needs to perform procedural tasks in imagistic form. However, mental arithmetic is an exception. Most mental skills are carried out in interaction with cultural objects (e.g., a computer terminal for text editing, pencil and paper for square roots), and we have argued that during skilled performance little imagery occurs. However, in everyday life one occasionally needs to carry out the task of simple arithmetic without paper and pencil and so resorts to “mental arithmetic” or “doing the problem in your head.” In keeping with our account of mentally performed skills, there appears to be strong imagery in this task. The phenomenon is so powerful in this case that when B. F. Skinner (1957) was attempting to work out a radical behaviorist approach to psychology, he was forced to note that “In intraverbal chaining, for example, necessary links are sometimes missing from the observable data. When someone solves a problem in ‘mental arithmetic,’ the initial statement of the problem and the final overt answer can often be related only by inferring covert events” (p. 434).

c.

MULTIPLEFORMS OF REPRESENTATION

One of the obvious consequences of our analysis of memory is that there are many different forms of memory representation. The same event can result in different memory representations (as in the initial example of the undergraduate going to the psychology experiment), and a given recall performance can be based on a variety of forms of mental representation. For example, consider a typical semantic memory task where the subject is asked “What color is a canary?” and responds correctly. In terms of our analysis the subject’s response could have been based on: (1) a particularized image, (2) a schema, (3) a generic image, (4) semantic memory, or ( 5 ) rote linguistic skill. Clearly, if one is going to construct adequate models of the memory process, one must be sensitive to this issue and attempt to establish what form of representation the subject is using in a given performance (for a similar position see Kosslyn, 1980). In general, the proposal we have outlined is going to be hard on the “nothing but” theorist (e.g., the theorist who says that the form of representation is nothing but X).For example, when Begg and Paivio (1969) postulated that abstract sentences are represented in memory as nothing but surface structures, Brewer (1975) was able to show the problems with this position by providing memory data (synonym substitutions) that seem to

32

William F. Brewer and John R. Pnni

require an abstract nonimage form of representation in the recall of meaningful abstract sentences. To take another example, it seems to us that many types of reasoning problems can be solved with both imaginal and nonimaginal representational processes. Thus, in studying reasoning problems one must find out what forms of representation are being used in a particular performance and why (see Banks, 1977; Clark, 1969; Huttenlocher, 1968; Moyer & Dumais, 1978). The approach that we have adopted here can account for many individual differences in the performance of a given task. When we ask people to tell us the seventh letter of the alphabet, we usually get long reaction times and strong reports of auditory and/or visual imagery. However, one individual we tested gave the response immediately and with little report of imagery. When we asked the subject some questions to find out why he differed from our other subjects, we found that he was an amateur cryptographer and had the letter-number correspondences stored in rote linguistic form. To take another example, we have recently tried to elicit personal memory by asking a question such as, “What did you have for breakfast?’’ The subjects tested gave personal memory reports, but suppose that a subject had given a response such as “eggs” very rapidly and with little report of imagery, or feeling of reliving. We suspect that further questioning would show that this subject had eggs for breakfast every day and was using information from semantic memory to answer the question.

D.

COPY IMAGES VERSUS RECONSTRUCTED IMAGES

We find the logical and empirical arguments of Pylyshyn (1973, 1981) and others against pure copy theories of imagery to be compelling. It must be the case that at least part of the phenomenally experienced image is reconstructed from information of a nonimaginal kind. We recently carried out some experimental work on this topic. In an earlier study, Brewer and Treyens (1981) showed that schema-driven inferences occurred in the recall of information about a room that subjects had been in briefly. The subjects frequently recalled having seen books in the room, even though there were no books present. Brewer and Pani (1983a) have replicated the Brewer and Treyens (1981) study, but included detailed questions about mental imagery experiences after each recall trial. We found that for present and inferred items of equivalent memory strength the subjects reported roughly equivalent amounts and quality of imagery. In other words, the schema-driven inferences were apparently incorporated into the phenomenally experienced image of the room. An important area for this issue is the study of autobiographical personal


33

memories. As discussed earlier, these memories are accompanied by a strong belief that they are veridical. Neisser (1982, pp. 43-48) recently argued that even the intense “flashbulb” form of personal memories resulting from a highly emotional event (e.g., “Where were you when you heard that Kennedy had been shot?”) are not veridical. He has also shown that John Dean’s recall of specific events at the Watergate hearings was a complex reworking of information from a number of different occasions (Neisser, 1981). Except for Neisser’s study of John Dean, most of the data here remain anecdotal, and the standard techniques for studying autobiographical memory (e.g., Robinson, 1976) do not allow one to resolve this issue. Brewer (1983) has developed a technique which should allow a more careful examination of the veridicality of personal memory. He has subjects carry a random alarm device and has them record what is occurring when the alarm goes off. By comparing personal memories occuring at the time of test with the original record of the event this technique makes possible the gathering of systematic data on the issue of veridicality of personal memory.

VI. Conclusions In this article we have tried to take a fresh approach to the problem of human memory. We first attempted to provide a description of the common forms of memory. We adopted this strategy because we think that research in memory has frequently cut short the process of description and moved too soon to the job of detailed analysis and model building. We have argued, on theoretical grounds, that the data from phenomenal experience should be given equal status with the other forms of data typically gathered in experiments on human memory. In carrying out our analysis we have attempted to provide an example of how this data can be used in theory construction. In working out our analysis of the structure of memory we felt a constant tension between a view of memory as the reliving of earlier perceptions and a view of memory as a schema-based reconstructive process. We hope the analysis succeeds in providing a synthesis of these two positions. Compared to other recent theories of memory our position looks somewhat complex. It seems to us that the complexity in our analysis is simply a reflection of the complexity of the problem. We think that many of the classic theories of human memory have achieved simplicity by ignoring the actual complexity of the phenomena and by attempting to give a simple image account, or a simple interference account, or a simple propositional account.

34


At the end of many sections of this article we found ourselves saying that more empirical and theoretical work was needed. We hope that this was not merely ritualistic language on our part and that, in fact, the framework provided in this contribution does lead one to see new problems and new issues in the study of memory.

ACKNOWLEDGMENTS We would like to thank Brian Ross, Glenn Nakamura, Carolyn Mervis. Demetrios Karis,

Don Dulany. and Ellen Brewer for comments on an earlier version of this paper. Some of the ideas in this article first appeared in a paper presented by the senior author at the Fourth Annual Conference of the Cognitive Science Society, 1982.

REFERENCES Adams, J.A. A closed-loop theory of motor learning. Journal of Motor Behavior, 1971, 3, I 11-149. Anderson, J .R. Verbatim and propositional representation of sentences in immediate and longterm memory. Journal of Verbal Learning and Verbal Behavior, 1974, 13, 149-162. Anderson, J.R. (Ed.). Cognitive skills and their acquisition. Hillsdale. New Jersey: Erlbaum, 1981. Anderson, J.R., & Bower, G.H. Human associative memory. Washington, D.C.: Winston, 1973. Angell, J.R. Methods for the determination of mental imagery. Psychological Monographs, 1910, 13 (1, Whole No. 53), 61-108. Attneave, F. Some informational aspects of visual perception. Psychological Review, 1954, 61, 183-193. St. Augustine. The confessions (Great Books of the Western World, Vol. 18). Chicago, Illinois: Encyclopedia Britannica, 1952. Ayer, A.J. The problem of knowledge. Baltimore, Maryland: Penguin, 1956. Banks, W.P. Encoding and processing of symbolic information in comparative judgments. In G.H. Bower (Ed.), Psychology of learning and motivation (Vol. 1 I). New York: Academic Press, 1977. Bartlett, F.C. Remembering. London and New York: Cambridge Univ. Press, 1932. Begg, I., & Paivio, A. Concreteness and imagery in sentence meaning. Journal OJ Verbal Learning and Verbal Behavior, 1969, 8, 821-827. Bergson, H. Matter and memory. London: Allen & Unwin. 1911. Block, N. Troubles with functionalism. In N. Block (Ed.), Readings in philosophy of psychology (Vol. 1). Cambridge, Massachusetts: Harvard Univ. Press, 1980. Bock, J.K., & Brewer, W.F. Reconstructive recall in sentences with alternative surface structures. Journal of Experimental Psychology, 1974, 103, 837-843. Book, W.F. The psychology of skill. University of Montana Publications in Psychology. 1908, 1. Bower, G.H. Mood and memory. American Psychologbt, 1981, 36, 129-148. Bransford, J.D., Barclay, J.R., & Franks, J.J. Sentence memory: A constructive versus interpretive approach. Cognitive Psychology, 1972, 3, 193-209.


35

Bransford, J.D., & Franks, J.J. The abstraction of linguistic ideas. Cognitive Psychology. 1971, 2, 331-350.

Brewer, W.F. Memory for ideas: Synonym substitution. Memory di Cognition, 1975, 3 , 458464.

Brewer, W.F. Literary theory, rhetoric, and stylistics: Implications for psychology. In R.J. Spiro, B.C. Bruce, & W.F. Brewer (Eds.), Theoretical issues in reading comprehension. Hillsdale, New Jersey: Erlbaum, 1980. Brewer, W.F. Autobiographical memory. In preparation, 1983. Brewer, W.F., & Dupree, D.A. Episodic and generic memory for places. Paper presented at the twenty-second annual meeting of the Psychonomic Society, Philadelphia, November, 1981.

Brewer, W.F., & Dupree, D.A. Use of plan schemata in the recall and recognition of goaldirected actions. Journal of Experimental Psychology: Learning, Memory and Cognition, 1983, 9, 117-129. (a) Brewer, W.F., & Dupree, D.A. Schema acquisition: Episodic and generic memory for places and events. In preparation, 1983. (b) Brewer, W.F., & Nakamura, G. V. The nature and function of schemas. In R.S. Wyer & T.K. Srull (Eds.), Handbook of social cognition (Vol. 1). Hillsdale, New Jersey: Erlbaum, 1983, in press. Brewer, W.F., & Pani, J.R. Personal memory, generic memory. andskill: An empiricalstudy. Paper presented at the twenty-third annual meeting of the Psychonomic Society, Minneapolis, November, 1982. Brewer, W.F., & Pani, J.R. Evidence for reconstructive imagery in memory for rooms. In preparation, 1983. (a) Brewer, W.F., & Pani, J.R. Phenomenalreportsduringmemoryrecall. In preparation, 1983. (b) Brewer, W.F., & Treyens, J.C. Role of schemata in memory for places. Cognitive Psychology. 1981, 13, 207-230.

Calkins, M. W.The abandonment of sensationalism in psychology. American Journal of Psychology, 1909, 20, 269-277. Card, S.K.,Moran, T.P., & Newell, A. Computer text editing: An information-processing analysis of a routine cognitive skill. Cognitive Psychology, 1980, 12, 32-74. Chomsky, N. Aspects of the theory of syntax. Cambridge, Massachusetts: MIT Press, 1965. Clark, H.H. Linguistic processes in deductive reasoning. Psychological Review, 1969,76,387-

404. Cofer, C.N. A comparison of logical and verbatim learning of prose passages of different lengths. American Journal of Psychology. 1941, 54, 1-20. Collins, A.M., & Quillian, M.R. Retrieval time from semantic memory. Journal of Verbal Learning and Verbal Behavior, 1969, 8, 240-247. Crosland, H.R. A qualitative analysis of the process of forgetting. Psychological Monographs, 1921, 29 (1, Whole No. 130). Dulany, D.E. Awareness, rules, and propositional control: A confrontation with S-R behavior theory. In T.D. Dixon & D.L. Horton (Eds.), Verbal behavior and general behavior theory. Englewood Cliffs, New Jersey: Prentice-Hall, 1%8. Ebbinghaus, H.Memory. New York: Dover, 1964. (originally published 1885.) Ericsson, K.A., & Simon, H.A. Verbal reports as data. Psychological Review, 1980.87, 215251.

Fernald, M.R. Thediagnosis of mentalimagery. PsychoIogicalMonographs, 1912,14(1, Whole No. 58). Fitts, P.M., & Posner, M.I. Human performance. Belmont, California: BrooksICole, 1%7. Fodor, J.A. Propositional attitudes. Monist, 1978, 61, 501-523. Furlong, E.J. A study in memory. London: Thomas Nelson, 1951.

36


Garner, W.R. The processing of information and structure. Potomac, Maryland: Erlbaum 1974. Gibson, E. J. Principles of perceptual learning and development. New York: Appleton, 1969. Graesser, A., & Mandler, G. Recognition memory for the meaning and surface structure of sentences. Journal of Experimental Psychology: Human Learning and Memory, 1975, 104, 238-248. Hilgard, E.R. Consciousness in contemporary psychology. Annual Review of Psychology, 1980, 31, 1-26. Hintzman. D.L. The psychology of learning and memory. San Francisco, California: Freeman, 1978. Hoch, H.S., & Schmelzkopf, K.F. The abstraction of schematic representations from photographs of real-world scenes. Memory & Cognition, 1980, 8, 543-554. Hochberg, J. In the mind’s eye. In R.N. Haber (Ed.), Contemporary theory and research in visual perception. New York: Holt. 1968. Humphrey, G. Thinking. New York: Wiley, 1951. Huttenlocher, J. Constructing spatial images: A strategy in reasoning. Psychological Review, 1968, 75, 550-560. James, W. Principles of psychology (Vol. 1). New York: Macmillan, 1890. Kintsch, W. The representation of meaning in memory. Hillsdale, New Jersey: Erlbaum, 1974. Kosslyn, S.M. Image and mind. Cambridge, Massachusetts: Harvard Univ. Press, 1980. Kosslyn, S.M. The medium and the message in mental imagery: A theory. Psychological Review, 1981. 88, 46-66. Kosslyn, S.M., & Jolicoeur, P. A theory-based approach to the study of individual differences in mental imagery. In R.E. Snow, P.A. Federico, & W.E. Montague (Eds.), Aplitude, learning. and instruction: Vol. 2. Cognitiveprocess analyses of learning and problem solving. Hillsdale, New Jersey: Erlbaum, 1980. Kuhlmann, F. On the analysis of the memory consciousness: A study in the mental imagery and memory of meaningless visual forms. Psychological Review, 1906. 13, 316-348. Kuhlmann, F. On the analysis of the memory consciousness for pictures of familiar objects. American Journal of Psychology, 1907, 18, 389-420. Kuhlmann, F. On the analysis of auditory memory consciousness. American Journal of Psychology, 1909, 20, 194-218. Ladd, G.T. Psychology, descriptive and explanatory. New York: Scribner, 1894. Lashley, K.S. Cerebral organization and behavior. In F.A. Beach, D.O. Hebb, C.T. Morgan, & H.W. Nissen (Eds.), The neuropsychology of Lashley. New York: McGraw-Hill, 1960. Leyden, W. von. Remembering. New York: Philosophical Library, 1961. Lichtenstein, E.H., & Brewer, W.F. Memory for goal-directed events. Cognitive Psychology, 1980, 12, 412-445. Locke, D. Memory. Garden City, New York: Anchor, 1971. Malcolm, N. Knowledge and certainty. Englewood Cliffs, New Jersey: Prentice-Hall, 1963. Mandler, J.M., & Parker, R.E. Memory for descriptive and spatial information in complex pictures. Journal of Experimental Psychology: Human Learning and Memory, 1976, 2, 38-48. McGeoch, J.A., & Irion, A.L. Thepsychology of human learning. New York: David McKay, 1952. Medin, D.L., & Schaffer, M.M. Context theory of classification learning. Psychological Review. 1978, 85, 207-238. Melton. A.W. Categories of human learning. New York: Academic Press, 1964. Minsky, M. A framework for representing knowledge. In P.H. Winston (Ed.), Thepsychology of computer vision. New York: McGraw-Hill, 1975. Montague, W.E., Adams, J.A., & Kiess, H.O. Forgetting and natural language mediation. Journal of Experimental Psychology. 1966, 72, 829-833.


37

Moyer, R.S., & Dumais, S.T. Mental comparison. In G.H. Bower (Ed.), Psychologyof learning and mofivafion(Vol. 12). New York: Academic Press, 1978. Natsoulas, T. Concerning introspective “knowledge.” Psychological Bulletin, 1970, 73, 89111. Natsoulas, T. The subjective, experiential element in perception. Psychological Bulletin, 1974. 81, 611-631. Neisser, U. Cognifion and reality. San Francisco, California: Freeman, 1976. Neisser, U. John Dean’s memory: A case study. Cognifion, 1981, 9, 1-22. Neisser, U. hfemory observed. San Francisco, California: Freeman, 1982. Norman, D.A. Memory and atfention (2nd ed.). New York: Wiley, 1976. Norman, D.A. Categorization of action slips. Psychological Review, 1981. 88, 1-15. Ogden, R.M. Imageless thought: Rtsumk and critique. Psychological Bulletin, 191 I , 8, 183197. Pani, J.R. A functionalist approach fo mental imagery. Paper presented at the twenty-third annual meeting of the Psychonomic Society, Minneapolis, November, 1982. Pani, J.R. The fheory of menfal imagery in psychology: A consisfentparadigm. Unpublished manuscript, University of Illinois, Urbana-Champaign, 1983. Piaget, J. The origins of infelligencein children. New York: International Universities Press, 1952. Piaget, J. The construction of reality in fhe child. New York: Basic Books, 1954. Posner, M.I. Abstraction and the process of recognition. In G.H. Bower & J.T. Spence (Eds.), The psychology of learning and motivation (Vol. 3). New York: Academic Press, 1969. Posner, M.I., & Keele, S.W. On the genesis of abstract ideas. Journal of Experimenfal Psychology, 1968, 77, 353-363. Prytulak, L.S. Natural language mediation. Cognitive Psychology, 1971, 2 , 1-56. Pylyshyn, Z.W. What the mind’s eye tells the mind’s brain: A critique of mental imagery. Psychological Bulletin, 1973, 80, 1-24. Pylyshyn, Z. W. The imagery debate: Analogue media versus tacit knowledge. Psychological Review. 1981, 88, 16-45. Radford, J. Reflections on introspection. American Psychologist, 1974, 29, 245-250. Reiff, R., & Scheerer, M. Memory and hypnofic age regression. New York: International Universities Press, 1959. Robinson, J.A. Sampling autobiographical memory. Cognitive Psychology, 1976.8, 578-595. Rosch, E., Mervis, C.B., Gray, W.D., Johnson, D.M., & Boyes-Braem, P. Basic objects in natural categories. Cognifive Psychology, 1976, 8, 382-439. Ross, B.H. The role of reminding in learning a cognifiveskill. (Doctoral dissertation, Stanford University, 1982.) Dissertation AbsfracfsInfernafional, 1982, 43, 1286B. Rumelhart, D.E. Schemata: The building blocks of cognition. In R.J. Spiro, B.C. Bruce, & W.F. Brewer (Eds.), Theoretical issues in reading comprehension. Hillsdale, New Jersey: Erlbaum, 1980. Rumelhart, D.E., Lindsay, P.H., & Norman, D.A. A process model for long-term memory. In E. Tulving & W. Donaldson (Eds.), Organization of memory. New York: Academic Press, 1972. Rumelhart, D.E., & Norman, D.A. Accretion, tuning and restructuring: Three modes of learning. In J.W. Cotton & R.L. Klatzky (Eds.), Semanficfactors in cognifion. Hillsdale, New Jersey: Erlbaum, 1977. Russell, B. The analysis of mind. London: Allen & Unwin, 1921. Sachs, J.S. Recognition memory for syntactic and semantic aspects of connected discourse. Perception & Psychophysics, 1967, 2, 437-442. Schank, R.C., & Abelson, R.P. Scripts, plans, goals, and undersfanding.Hillsdale. New Jersey: Erlbaum, 1977. Schmidt, C.F. Understanding human action: Recognizing the plans and motives of other per-

38


sons. In J.S. Carroll & J.W. Payne (Eds.), Cognifion and social behavior. Hillsdale, New Jersey: Erlbaum, 1976. Schonfield, D.. & Stones, M.J. Remembering and aging. In J.F. Kihlstrom & F.J. Evans (Eds.), Funcfional disorders of memory. New York: Wiley, 1979. Shallice, T. Dual functions of consciousness. Psychological Review, 1972, 79, 383-393. Shepard, R.N. The mental image. American Psychologist, 1978, 33, 125-137. Shoemaker, S . Functionalism and qualia. In N. Block (Ed.), Readings in philosophv of psychology (Vol. 1). Cambridge, Massachusetts: Harvard Univ. Press, 1980. Skinner, B.F. Verbal behavior. New York: Appleton, 1957. Titchener, E.B. A tesf-book of psychology. New York: Macmillan, 1910. Tulving, E. Episodic and semantic memory. In E. Tulving & W. Donaldson (Eds.), Organization of memory. New York: Academic Press, 1972. Welborn, E.L., &English, H. Logical learning and retention: A general review of experiments with meaningful verbal materials. Psychological Bullefin, 1937, 34, 1-20. Williams, T.M.. & Aiken, L.S. Auditory pattern classification: Continuity of prototype use with development. Developmental Psychology, 1975, 11, 715-723.

A SIMULATION MODEL FOR THE COMPREHENSION OF TECHNICAL PROSE David Kieras UNIVERSITY O F ARIZONA TUCSON, ARIZONA Introduction ........................................................... A. Criteria for Simulation Models ...................................... B. Issues in Evaluating Simulations ..................................... C. Overview of the Simulation ......................................... 11. Description of the Simulation.. .................. ..................... A. Memory Structure ......................... ..................... B. Parsing Processes.. ................................................ C. Memory Search.. .................................................. 111. Comparisons with Data. .... ......................................... A. Task and Coherence Effects.. ....................... B. Prior Knowledge Effects.. .......................... C. Problems with the Model ........................... IV. Conclusion. ........................................... References ................................ ......................

I.

39 39 42

44 46 46 48

66 72

79

1. Introduction

The purpose of this article is to describe a simulation model for the comprehension of simple prose that has been in use for several years, and was described briefly in Kieras (1977a, 1981). Because this model has been compared to data in some depth, a fairly complete description of it would be useful. To give an initial flavor of the capabilities of the model, Table I gives an example of a passage that the current model can process reasonably successfully. It should be kept in mind however, that the model is still under development, and so its current form undoubtedly contains errors. A.

CRITERIA FOR SIMULATION MODELS

Despite their now long history of use in cognitive psychology, simulation models of cognitive processes, and the techniques used to develop them, are by no means standardized and uncontroversial. For example, one still hears the comment that simulation models are essentially untestable because 39 THE PSYCHOLOGY OF LEARNING AND MOTIVATION, VOL. 17

Copyright 0 1983 by Academic Press, Inc. All rights of reproduction in any form reserved. ISBN 0-12-5433174

David Kieras

40

TABLE I EXAMPLE PASSAGE 1. 2.

3. 4. 5. 6.

7.

a.

USED IN

JOHNSONAND KIERAS (1983)

Modern timekeeping devices are extremely accurate. An inexpensive quartz-crystal watch has one-second accuracy for several months. Proper adjustment of the watch can improve the accuracy. Atomic resonance clocks measure atomic vibrations and they are incredibly accurate. The theory of relativity predicts that tiny distortions of time would be produced on a long trip in a commercial airliner. Atomic resonance clocks could measure the distortions of time and confirm the theory. A hydrogen-maser clock has one-second accuracy for 10 million years. A hydrogen-maser clock is used today by the National Bureau of Standards.

they have so many “degrees of freedom” that they can be made to fit any set of data. The work described here was guided by some general criteria for what it means to have a meaningful simulation model of a cognitive process. These criteria can be briefly summarized in order to set the stage for the presentation of the model itself.

1.

Theoretical Basis

A simulation model should be based on relatively few theoretical principles. If this is so, then the simulation will be easy to describe, because it will be possible to describe the psychological principles that underlie the model, and then as much, or as little, of the technical details of how the simulation program implements these principles can be supplied as desired. This strategy perhaps works best if the simulation model is based on some existing theoretical ideas that have already appeared in the traditional verbal or formal form. For example, the Kintsch and van Dijk theory of prose comprehension was first developed in the standard modes of psychological theory, and then was used in simulation models (Kintsch & van Dijk, 1978; Kieras, 1982). When such a theory is used as the basis for a simulation model, the division between the psychological principles and the technical details of the realization of them in the simulation is especially clear. In contrast, when a simulation model is developed from scratch, without the guidance or clear domination of a few psychological principles, the model is much more prone to be an ad hoc concoction of special case solutions. Or the mechanisms of the model may be so closely tied to the implementation that it is difficult to see the psychological content of the model as distinct from its programming.

A Simulation Model for the Comprehension of Technical Prose

41

2. Realistic Behavior and Simplicity A simulation model should display at least apparently realistic behavior when based on a few principles. By apparently realistic is meant a loose, but obvious, similarity at the gross level. For example, since people clearly interpret active and passive forms of sentences similarly, the model should do the same. Or, since people integrate the content of the sentences of a passage together, so should the model. While it is obvious that this is a goal for a simulation model, it is not a goal for the other forms of models, which do not actually include analogs of the mental processes. For example, this criterion could not apply to Kintsch’s (1974) system, which did not specify processes for converting surface sentences into a propositional representation. Hence, although people map active and passive sentences into similar meaning representations, this early model does not, although it is valuable in its own right. If this criterion is combined with the first one, then the process of developing a simulation model actually becomes quite difficult, meaning that the “degrees of freedom” one supposedly has with a simulation model are mostly illusory. It is in fact quite difficult to start with a few general principles and then write a program that will produce apparently realistic behavior. The important psychological content of these two criteria is that if the general principles have been well chosen, then it is possible to produce apparently realistic behavior with a program that is relatively simple and easy to describe. If the principles were not correct, or not good choices, then to produce apparently realistic behavior requires a baroque, overly complex, and difficult to describe simulation. Thus, an important feature of the simulation approach is that an initial check on the correctness of the theoretical principles can be done by simply attempting to build the model using the principles. If this is too difficult, or yields an overly complex model, the validity of the principles is cast in doubt.

3.

Use of Detail

A simulation model is most valuable when compared to very detailed data. One of the major virtues of simulation modeling, as compared to other forms of theorizing, is that the model contains a rigorous and extremely detailed specification of the hypothesized mental processes. If, however, the model is not compared to data at a corresponding level of detail, then much of the value of the model as an empirical explanation has been lost. Only recently have there been efforts to use simulation models in such precise ways (see Kieras, 1983); the standard use of simulations has

42

David Kieras

been to simply demonstrate the sufficiency of a set of principles. However, simulation models have certainly developed to the point where they can be taken as seriously in an empirical context as the classical mathematical models. One of the problems, of course, with trying to use simulation models as detailed explanations for data is the issue of just how the comparison of the model with the data should be carried out. Some recent work on the subject is reviewed in the volume by Kieras and Just (1983). One very useful technique will be summarized later in this chapter. But once a model has been compared to data, some important issues arise that have been controversial for some time. B. ISSUES IN EVALUATING SIMULATIONS 1.

Testability and Modifications

What is the status of a model if it has been modified in order to get a good fit to the data? This situation seems to be very serious with simulation models since they apparently have so many “degrees of freedom” that once the model has been modified repeatedly, there is a temptation to say that the model is no longer the same model as it originally was, and that in some way the modifications to the model must be counted against the quality of the fit. In the context of the earlier mathematical models, the amount of information taken from the data to estimate the parameters of the model could be quantitatively assessed in the statistical techniques used to compare the model to the data. This does not seem to be possible when the model involved is a simulation model, despite the emergence of some statistical methods for comparing the model to data. The reason, of course, is that the entire simulation program can be considered as a vast parameter set which can be modified in arbitrary ways, thereby guaranteeing a good fit. But, as applied to current simulation efforts, this is not really a serious criticism. First, if the above criteria are recognized, arbitrary modifications to the model are not permissable. Second, to be immune to this criticism, the model would have to be one whose very first statement turned out to be exactly right. This is clearly not a sensible approach to the development of scientific theory, and simulation models are no exception. We naturally expect our theories to be modified to take data into account. The fact that simulation models can change so quickly in the course of their development is a reflection of the fact that simulation modeling is an extremely powerful and efficient theoretical tool, which makes it possible to evolve theoretical ideas much more rapidly than simple verbal reasoning would allow. The real criterion for whether a modified model is an acceptable explanation for a set of data is whether the general principles underlying the


43

model can still be clearly seen as being responsible for the model. In other words, if in order to get the model to fit the data, only minor adjustments to the model processes, as opposed to major reformulations of its basic principles, were necessary, then in fact we simply have a corrected version of a model based on the same theoretical concepts. If, in contrast, the basic principles or fundamental assumptions had to be modified, then we can say that we have arrived at a different theory. For example, in the work described in Kieras (1977b), a change was made from a spreading activation memory search process that had unlimited capacity to one that had limited capacity, and in the process an apparently better fit to data was obtained. This would have to be classed as a different theory, at least with regard to the memory search process, because a fundamental assumption was involved. On the other hand, modifying the parser, so that it properly handles sentences of a particular structure, counts as merely correcting a detail, or extending the same theory.

2. Explaining Individual Differences The second issue in the empirical status of simulation models is the issue

of individual versus group data, and whether the model is intended to describe individuals or some average or typical individual. Because simulation models possess such precise detail, it is possible to compare the detailed behavior of the model to the detailed behavior of individual subjects, making this problem more serious than it is for traditional theories. But if this comparison is done for more than one subject, the model is almost certain to be embarrassed. Each individual subject is almost certain to do something different from the others. Many times, especially in the context of prose comprehension, these differences apparently are profound differences in strategy or background knowledge and are therefore not realistically considered as simply being due to random “error.” A good example appears in Kieras (1982) and Kieras and Bovair (1981). In that work, 10 subjects produced think-aloud protocols while reading simple technical prose one sentence at a time. Each individual subject often appeared to be following a definite strategy, but this strategy would be very different from that used by other subjects, and apparently would even vary from one passage to the next. As long as simple verbal models for grouped or aggregated data are used, this problem of individual variability is easily avoided. The precision of the theoretical approach matches up well with data obtained by averaging or aggregating over subjects. Most such work is based on simple directional hypotheses. However, the commitment to the level of detail available in a simulation model implies a commitment to a detailed understanding of the

David Kiem

44

data. Asking a detailed model to make merely a directional prediction is thus underusing its precision considerably. But how does one deal with the fact that the model is very likely at best to fit only some of the subjects some of the time, and be quite incorrect for the other subjects? A possible general approach to this problem is what could be called the “erector set” approach to modeling. The traditional goal of constructing a model is to be able to present a single model as an explanation for all examples of a given class of phenomena. That is, most models, either traditional or simulation, are presented as reasonable descriptions for what all, or most, subjects will do in a particular task. The erector set approach would be to define a set of processes, which can be combined in certain ways according to a set of rules, that will suffice to explain the behavior of most subjects doing a particular task. Thus, one subject may call for a certain set of processes that are combined in a way determined by that subject’s strategy, while another subject may require only a subset of those processes, combined in a different way according to that subject’s strategy. Thus, an adequate theory of the behavior consists not of a single set of processes combined in a single way, but rather a set of processes which can be combined according to certain rules. The theoretical task then consists of determining which processes are needed and what the rules of combination are. However, the erector set approach to modeling has not yet been used ‘systematically (but see Kieras, 1982), and clearly would be of most value only when individual subject data is being considered. But a similar approach seems to be very useful in describing the effects of reading task manipulations on comprehension data. For example, in Kieras (1981), different reading tasks are assumed to “switch in” different processes. This is an erector set approach applied to account for task, rather than subject, differences.

c.

OVERVIEW OF THE SIMULATION

1.

Organization and Goals

The simulation model consists of a parsing process and a memory storage and retrieval system. The parsing process is based on the augmented transition network (ATN) method of parsing. The memory system consists of a semantic network, based on the Anderson (1976) ACT system. The memory retrieval process consists of a spreading activation process, similar to that originally suggested by Collins and Loftus (1975) and Anderson (1976), along with a specialized process which examines the paths of intersection of activation in the memory system to determine whether the desired information is present and retrieve it for further use.


45

The original intent in developing the simulation was to work on some problems in the comprehension of prose (see Kieras, 1977a). Thus, the major concern with the model was some of the simpler integrative processes involved in prose comprehension, and not so much with the details of the semantic representation for sentences. More specifically, where this emphasis shows up is that the simulation has relatively elaborate processes for dealing with noun phrases, but very limited processes for dealing with verb phrases, with many issues simply ignored. 2.

Theoretical Basis

In the introduction to this chapter it was claimed that one of the most important criteria for a good simulation model is that it be based on a relatively few theoretical principles, and that the status of the model depends heavily on whether these theoretical principles are still readily apparent even after the model has been developed and modified. Here will be summarized the principles underlying the present simulation model. The principles can be roughly subdivided into three sets, concerned with parsing, memory, and integration processes. a. Parsing. The simulation is based on a bottom-up, syntactically based approach to parsing. The principle is that the surface form of sentences serves an important role in specifying the semantic content of the sentence. Thus, in formal prose, such as technical prose, the syntactic information in the individual sentences is a quite reliable description of the immediate semantic content of the sentence. Furthermore, it is assumed that much of the important comprehension processes involved in prose comprehension can be handled with the same approach. Namely, the ATN parsing mechanism is used not only to process individual sentences, but to process a series of sentences making up a passage. Finally, the noun phrase is assumed to be the basic parsing unit, and all integration and memory search processing for a noun phrase is completed before the parse continues further into the sentence. b. Memory. The model uses a knowledge representation system consisting of a semantic network. The network is searched by a spreading activation process, and it is an important assumption of the model that the major context effects in comprehension can be handled by the use of activation which spreads not only from the concepts being processed, but also from the passage topic. Since technical prose refers not only to instances of concepts but also very often to the concepts themselves, the memory system must make a clear distinction between concepts and instances, and the memory search involved in comprehension will have to take this into account.

David Kieras

46

c. Integration. The model assumes that the integrative processes in comprehension consist basically of attaching new memory structure to existing memory structure, usually the structure resulting from the comprehension of the previous sentences in the passage. The model is based on the principle that noun phrases perform the major integrative work in comprehension. More specifically, each noun phrase identifies either a new referent, or a given referent (Clark & Haviland, 1977), whose representation must then be located in memory. Most passage sentences contain at least one given referent, and thus the new information in the sentence can be added onto the preexisting representation in memory. A major point of this model is that resolving most such references in noun phrases can be done by a relatively small set of uniform rules for searching memory. An assumption of the model is that the specifications for the search for noun phrase referents can be derived from the syntactic properties of the sentence in the passage. 3.

The Importance of Bottom-Up Processes

In many ways, these principles are somewhat old-fashioned or reactionary. The basic underlying theme is that a substantial portion of comprehension processes in fact rely upon the syntactic information in the input. This stems from a belief that the rich and relatively well-defined syntactical structure of written language has a definite function; namely, it permits comprehension to be accomplished with relatively simple and very efficient bottom-up processes. The very complex top-down processes that have been of great interest to many comprehension researchers are the “heavy artillery” that should be reserved for those occasions when the surface form of the input has failed to successfully convey the intended meaning. This model undoubtedly overestimates the sufficiency of syntactic information. However, as will be illustrated below, the model provides a valuable service of showing how much of comprehension can be accomplished on a syntactic basis, and in the process makes extremely clear when phenomena appear that cannot be based simply on syntax.

11.

A.

Description of the Simulation

MEMORYSTRUCTURE

The memory structures used in the model are based on those presented in Anderson (1976), being a slight modification of his ACT system. The memory structures are semantic networks consisting of nodes and links. A


41

memory node itself is simply an arbitrary symbol. The term node name will be used to refer to the actual symbol for the node. Using the property list facility of the LISP language, such a symbol may have other information attached to it, which in this case is a list of pointers, which are the links, to other memory nodes. Each link actually consists of two one-way pointers, one on each node. The different link types are listed in Table 11. All the memory nodes are formally the same, but have different roles depending on their link connections. There are proposition nodes, which are the sources for subject (S) and predicate (P) links. There are relational predicate nodes, which appear at the end of a predicate link and emit relation (R), or argument (A) links. Finally, there are referent and concept nodes. These nodes represent “things”, or more exactly, sets. Each such node has a specified cardinality, specified by a number link (N) to a number. These numbers either indicate the actual number of items in the set of things that the node represents, or else have certain conventional values. A cardinality of 1 means that the node represents a unique individual or instance of a concept. A cardinality of 99 indicates that the node represents a general concept. A cardinality of 7 is used to represent an unspecified plural number of members. For example, if the sentence mentioned some plural noun, but without a definite number specified, a cardinality of 7 would be attached to the new referent node. According to J. R. Anderson (1980, personal communication), revisions of the ACT system have the set cardinality specified as a separate proposition. This would avoid the somewhat awkward special case nature of the cardinality specification as done in this model. TABLE I1 SUMMARY OF LINKTYPESAND

STRUCTURE-BUILDING FUNCTIONS

Function

Link type

Node A

($SUBJ A B)

S

Proposition node

(OPRED A B)

P

Proposition node

($NUM A B) ($REL A B) ($ARC A B)

N R A

Referent or concept Relational predicate Relational predicate

(SARG1 A B)

A1

Relational predicate

($ARG2 A B)

A2

Relational predicate

Node B Referent, concept, predicate, proposition Concept or rational predicate Cardinality number Relation concept Single argument of relation First argument of relation Second argument of relation

48

David Kieras

A key feature of how this system represents objects is that the main class membership of an object, the IS-A relation, is represented simply with a proposition whose subject is the node for the object, and whose predicate is the concept node for this main class. This main class proposition often corresponds directly to the head noun of a noun phrase, and so plays an important role in resolving references (for more discussion see Bach, 1968; Kieras, 1977a). The links between nodes are either permanent or temporary. The permanent links make up structures that are part of long-term memory (LTM). The temporary links make structures in working memory (WM). Such structures are built in the course of processing not only individual sentences, but also parts of sentences, and the whole passage, and can be converted into permanent structures by changing the links to permanent ones. In describing memory structures, it is important and convenient to be able to describe a structure without having to specify the actual memory nodes involved. In the original HAM model (Anderson & Bower, 1973), this was done by means of probe structures, which were simply pieces of temporary semantic network structure. In retrieving information from memory, each node and link in the probe structure would be matched to corresponding nodes and links in memory in order to locate the actual memory nodes corresponding to those in the probe structure. In this simulation, the function of the probe structures is replaced by the structure-building function list. This is a specification of each individual link of memory structures, with variables taking the place of the actual memory nodes. If the name of a memory node is already known, then this node name would appear as a constant rather than a variable. Table 111 provides an example. Retrieving information from memory is then a matter of matching this structure-building function list to the actual memory structure, and in the process replacing the variables with actual memory node names. During parsing, the structure-building function list is constructed, but any actual structure-building is deferred until the system has made a final decision on the parsing solution. B. PARSING PROCESSES The parsing process is the part of the simulation that is of the most general interest, because it represents a set of hypotheses about how simple technical prose can be analyzed and integrated. The presentation here is not intended to allow the reader to reprogram the model, but rather to allow a psycholinguist to understand what the abilities and limitations of the parser are, and what assumptions are being made about the semantic content of various sentence forms, and to be able to “hand-simulate” the parsing


49

TABLE 111 STRUCTURE-BUILDING FUNCTION LIST FOR “A green frog did not eat quickly a bug.” (SSUBJ *PI *Rl)(SPRED *PI Frog-concept)(SNUM *RI 1) (SSUBJ *P2 *RI)(SPRED *P2 Green-concept) (SSUBJ *P3 *Rl)(SPRED *P3 *Nl)(%REL*NI Eat-concept)($ARG *N1 *R2) (SSUBJ *P6 *Nl)($PRED *P6 Quick-concept) (SSUBJ *P4 *P3)(SPRED *P4 False-concept) (SSUBJ *P5 *R2)(SPRED *P5 Bug-concept)(SNUM *R2 I )

of a sentence using the ATNs as described here. Consequently, as much detail as possible has been eliminated, and many difficult programming issues have been glossed over. More thorough detail, including LISP sources, can be obtained from the author. I.

The A TN Concept

The augmented transition network (ATN) approach to parsing consists of specifying a set of parsing rules in the form of a transition network. Each node, or state in the network, represents a state of the parse, and the links, or arcs, between nodes represent ways in which the parsing process can go from one state to the next. Roughly speaking, a transition from one node to the next is made for every word in the sentence. Thus, the progress of the system from node to node in the network corresponds to a process of processing each individual word in the sentence from left to right. The ATN approach was first used by Woods (1970) as a natural language interface to a question-answering system. Since then, ATN parsing has been a relatively popular approach to syntactic analysis (see Rumelhart, 1977). The use of ATNs in this simulation is both a simplified, and generalized, form of the original Woods ATN approach, taking advantage of the very general computing power of an ATN. That is, there is a common tendency to confuse the ATN principle with the specific set of parsing rules stated in ATN form that have appeared in the literature. However, an ATN is actually a very general way to express sequential processes, and so can be used to implement many different specific sets of parsing rules. In fact, the augmented transition network can actually be viewed as a specialized type of production system representation. The set of arcs leaving a node can be considered as a set of production rules, with the nodes corresponding to control variable values. In fact, the notation used in Anderson (1976) to diagram the control structure of a set of production rules closely resembles an ATN. So, in this simulation, an ATN is used not only to specify the

David Kieras

50

parsing rules for sentences, but also the rules for parsing a series of sentences or text. The ATN is also used to specify the other processes that go on in the course of comprehending and integrating prose. For example, an ATN is used to decide what type of memory search must be done in order to determine the referent of a noun phrase. Although ATNs have usually been described with diagrams, a tabular form is used here to save space. In each table such as Table IV in Section II,B,4, a series of nodes are listed, each node name being followed by one or more arcs. Each arc consists of a condition (shown in parentheses), an action (described in the table), and a next node (prefixed with >). The arcs are tested in a fixed order, which is represented by the order in which they are listed under each node. Each node has a name for reference purposes; certain nodes are named POP and FAIL, and do not actually appear, except as a next node. These nodes are special states which mean, respectively, for the system to POP up to the calling network with a SUCCESS status or to POP up with a FAIL status. Calls to embedded networks are represented by a condition that begins with the key word CALL followed by the name of the first node in the subnetwork that is being called. If that subnetwork returns successfully, that is, enters a POP state, then the condition on the calling arc is considered to be satisfied and the transition will be made. CALLX is a special form of network call; it has the same effect as a CALL condition as described before, with the important difference that the embedded network called is a path examiner network. This will be described in more detail below. In the course of parsing a sentence, the system attempts to find a complete pathway through the network to a final POP state. If the condition on the first arc is satisfied, the system will make the transition to the next state and perform the action on the arc along the way. Then it tries to leave the next state by the same process. If an arc has a nil condition, then that arc is crossed unconditionally when it is tested; such an arc must be the last one listed for a node. If at any point the system is not able to leave a state because none of the arcs have a condition that is met, the system will then back up to the previous state and try the next arc following the last arc previously tested that leaves that state. If the system fails to leave that previous state, it will then back up to the one before that, and so forth. If the system is not able to find a complete transition through the network that terminates in a POP state, then it fails to parse the sentence. 2.

The Lexicon

The basic information used by the parsing process is the grammatical word class of each word known to the system. The organization of the system’s lexicon can be briefly summarized. Each individual word appears


51

as a node in the long-term memory semantic network. A set of special syntactic-class links connect each word node to the corresponding semantic memory concept node. These syntactic-class links correspond to basic parts of speech such as noun, adjective, adverb, preposition, and so forth. A word node may have connections to more than one concept node, and likewise a concept node may have more than one word node associated with it. In addition to these links to concepts, words may have some additional syntactic properties. For example, a verb could be a past participle form, or an auxiliary verb. A noun could be a plural form. The model has no mechanisms for forming or recognizing plurals of words, or of different forms of the same underlying verb. Each individual word that the model can recognize must be explicitly represented along with its relevant syntactic properties.

3.

Variables and Functions

In order for the parsing process to perform its work, it is necessary for it to have available variables and functions. In the networks shown in the tables, the variables are the symbols that begin with an asterisk. The rules for a calling process are that the variables that are listed in a CALL condition are passed by name into the subnetwork, which also passes those variables back. Thus, variables listed in the CALL can be used to communicate information in both directions. Generally, variables are redefined every time a CALL is executed, unless they are listed in the CALL condition. There is a family of functions which are used to assist in the writing of conditions and actions. The functions used in the conditions mostly consist of the special function $PO% $POS evaluates to TRUE if the current word has the grammatical property specified in the argument to $POS. Thus, $POS DEFDET will be true if the current word is a definite determiner. For brevity, $POS is not written into the conditions in the ATN tables, since its presence is implied by the condition being some grammatical word class. The current word is always the first element of the current string being parsed. There is a special action function, called CHOP, that when executed causes the string to be changed so that the current word will be the second element in the string when processing begins at the next node. That is, the first element of the string is removed when the transition is made. Then the parse leaving the next node will be using the next word in the string as the current word. The current word is always contained in the variable *WORD. Since CHOP is almost always part of the actions on an arc, it is not shown in the ATN tables. There is an important function used in the ATN actions which constructs

52

David Kiens

a list of structure-buildingfunctions. This list is a variable called *SBFL. Once the parsing process decides that the sentence content has been properly extracted, the structure-building functions are actually executed by the function $BUILD, which executes the individual functions in *SBFL, resulting in the nodes and links of semantic network structure being added to memory. Tables I1 and 111, discussed above, describe the different structure-building functions, and the type of semantic memory link they construct. 4. Description of the Parsing Network

a. Text and Statements. Table IV shows the top level of the parsing ATN. This network defines a simple text grammar. For simplicity, some of the top level of the network that is concerned with implementing several convenience features has been eliminated. The topmost network starts with the node labelled PROSE-READ. The first arc leaving PROSE-READ performs a special READ-SENTENCE function; then the user is allowed to terminate the simulation run by entering the key words END or QUIT. If the user has entered a sentence, the system goes to PROSE-READ1, where it calls PROSE-CON. PROSE-CON first checks for propositional connectives. These are of two types: a PCON-A, such as therefore or however, which take only one argument, and PCON-A-B, such as although, which takes two arguments. After noting the presence and type of any propositional connective, the parser then calls the embedded network STATEMENT, which is the basic network for parsing sentences. When this subnetwork finishes by entering a POP node, it will return the system back to the PROSE-CON network that called it, which will then POP back to the arc on PROSE-READ1. Then the parse will go to the node CHECKCOHERENCE. CHECK-COHERENCE has a set of arcs which are responsible for maintaining or updating the topic list (*TOPICS) and determining whether the sentence topic (*T) is coherent with the current topic list. This is mainly done by determining whether the sentence topic is a member of the topic list, or whether the sentence topic received any activation from the topic list during the memory activations performed for the sentence. The sentence is decreed either to be coherent with the previous topics, or else not coherent, and the topic list is updated accordingly. Then the system goes on to read a new sentence by going back to the first node PROSE-READ. The STATEMENT network is shown in Table V. A statement consists of a noun phrase, indicated by the CALL NP, followed by a clause predicate, indicated by the CALL CLSPRED. After the clause predicate is parsed, the system does a CALLX to TRIM. This is part of the prior knowl-

A Simulation Model for the Comprehension

of Technical Prose

53

TABLE IV TOPLEVELOF ATN _ _ _ _ ~

Node with arcs

Action and comments

PROSE-READ (read sentence & no command) > PROSE-READ1 (END or QUIT command) > POP PROSE-READ1 (CALL PROSE-CON *E *T) > CHECK-COHERENCE CHECK-COHERENCE (*TOPICS empty) > PROSE-READ

Execute $BUILD on *SBFL if END, discard *SBFL if QUIT

Main proposition *E and sentence topic *T are returned

*T is main topic, add to *TOPICS

(*T related to *TOPICS) > PROSE-READ

Sentence is coherent

(NIL) > PROSE-READ

Sentence is not coherent, add new topic to *TOPICS

PROSE-CON (PCON-A)

> PROSE-CON2

Current *E will be subject of relation Given by propositional connective with one argument

(PCON-A-B) > PROSE-CON3

Next statement will be subject of relation

(CALL STATEMENT *E *T >POP

Parse the statement, with no leading connective

PROSE-CON2 (COMMA) > PROSE-CON2 (CALL STATEMENT *E *T

> POP PROSE-CON3 (CALL STATEMENT *E *T) >PROSE-CON2

Skip past a COMMA Build structure showing relation between previous main proposition and this *E

Parse first of two statements, this *E will be subject of relation

David Kieras

54

TABLE V

STATEMENT ATN Node with arcs STATEMENT (CALL NP *XO) > STATEMENT2 STATEMENT2 (CALL CLSPRED *REF *E)

>STATEMENTZA STATEMENT2A (CALLX TRIM *SBFL >STATEMENT3

Action and comments Sentence topic *T = *XO, referent of NP *REF = *XO

*REF is referent for clause predicate *E is main proposition node Activate from all sources

Build new structure identified by TRIM

STATEMENT3 (PERIOD) >POP (COMMA) >POP EMBED-STATEMENT (CALL NP *XO) > EMBED-STATEMENT2

*REF = *XO

EMBED-STATEMENT2 (CALL CLSPRED *REF *E) > POP

*XO = *E

edge mechanism, and will be discussed later. The last thing done in the statement network is to examine for the presence of a comma or period. If found, the network pops and it may of course be called again by the PROSECON network. The EMBED-STATEMENT network is called by the verb phrase network, to be described below. It is similar to the ordinary statement network, except that it is not concerned with the TRIM process, and returns in *XOthe name of the memory node that corresponds to the main proposition of the statement. b. Noun Phrases. There are two noun phrase networks shown in Tables VI and VII. The NP network, shown in Table VII, is the most important network in the simulation. Its basic structure is to call the ANP network shown in Table VI, which does the actual parsing. After ANP finishes its work, NP examines the combination of determiner, plurality of the head noun, the context of the noun phrase, and other information. to

55


TABLE VJ NOUNPHRASE PARSING NETWORK Node with arcs ANP (DEFDET) >ANPI

Action and comments *DET-PRN = DEFDET

(INDEFDET) >ANPI

*DET-PRN = INDEFDET

(NAME) >POP

'DET-PRN = NAME *CONCEPT = concept for name

(PROPPRN) >POP

'DET-PRN = PROPPRN *CONCEPT = proposition antecedent

(NIL) >ANP I

*DET-PRN = NO DETERMINER

ANPl (ADJ) > ANP 1 (NOUN) > ANP2

Add proposition to *SBFL showing referent modified by concept of adjective Set *NUMBER to SING or PLU Add proposition to *SBFL showing referent as instance of *CONCEPT with singular or plural cardinality

ANP2 (OF) > ANPOF (POSSessive ending) > ANPPOSS (NIL) > ANP3 ANPOF (CALL PREDNP *XO) > ANP3

ANPPOSS (CALL PREDNP *XO) > ANP3

Add structure to *SBFL showing that *XO,referent of the PREDNP possesses the referent of the main NP Principal referent *REF is then main NP referent Add structure to *SBFL showing that referent of PREDNP is possessed by the referent of the main NP Principal referent *REF is then the PREDNP referent

(continued)

David Kieras

56

TABLE VI (continued) Node with arcs ANP3 (CALL RESRELCLS *REF) >POP (NIL) > POP PREDNP (CALL NP 0x0) >POP

Action and comments Additional structure will be added to +REF

PREDNP flag is carried into NP referent returned in 0x0

decide what kind of memory search must be done in order to resolve the reference contained in the noun phrase. The ANP network will be the first one described. As shown in Table VI, ANP first checks for, and makes a note of, the presence of a determiner, either definite or indefinite, or some other pronoun, or whether the current word is a name, or is a propositional pronoun, such as the use of this to refer to the previous statement. Other pronouns are allowable, but the pronoun mechanism used in this simulation is very crude, and so not described. A variable +DET-PRN is set to signal the calling network concerning the nature of the determiner or pronoun. The variable *CONCEPT can be set by ANP to the memory node name corresponding to the concept denoted by the word. The system then goes to the node ANPl where an adjective is tested for, and if so a loop is made to the same node. The second arc at ANPl is to test for the presence of a noun, which would be the head noun for the phrase. The actions performed on these two arcs are shown in very condensed form. More detail will be provided to illustrate how the structure-building function list is constructed. The first step in the action for an adjective is to obtain a name for a new proposition node. Then, two structure-building functions are added to *SBFL, $SUBJ and $PRED, which specify the construction of a subject link from the new proposition node to the current referent of the noun phrase and a predicate link to the concept referred to by the adjective. A list, *SOURCES, is updated to include the adjective concept node; this list contains the nodes used to activate memory. The actions performed for the noun arc are very similar, only somewhat more complicated. The basic difference is that the cardinality of the head noun is used to construct a number link with the function $NUM. Continuing at ANP2, the system tests for the presence of the word oft or a special signal for possession that crudely represents the appearance of a possessive word ending. If so, another network PREDNP, shown at the

TABLE VII NOUNPHRASE MEMORYSEARCH ATN Node with arcs NP (CALL ANP) > NP2

NP2 (PROPPRN OR NAME) >POP (NIL) > NP3 NP3 (PREDNP & INDEFDET & SING) > NP6

Action and comments ANP returns information in *CONCEPT, *DET-PRN, and *NUMBER concerning main noun concept, determiner or pronoun type, and singular-plural nature of head noun

*XO = *CONCEPT Activate from *SOURCES collected by ANP and *TOPICS

Change cardinality to indicate concept

(PREDNP & NO DET & PLU) > NP6

Change cardinality to indicate concept

(CALLX FINDV 'SBFL) > POP

If individual matching *SBFL is found, *XO = memory node for individual

(CALLX FCON) >POP

If concept representation matching *SBFL is found, and match is complete, *XO = memory node for concept Change cardinality to indicate concept Fail to resolve reference

NPS (CALLX FCON *SBFL) > POP

I(IP6 (CALLX FCON *SBFL) > POP

(NIL) > POP

If search for concept matches completely referent *XO is the concept If there are unmatched predicates, create a new concept that is subset of referenced concept. *XO is new concept

Referent *XO is a new instance of located concept, with cardinality of head noun, and possibly with additional predicates *XO does not correspond to any existing memory node; new structure will be built for entire NP

58

David Kieras

end of Table VI, is called. PREDNP simply calls the NP network with a signal that the noun phrase appears in a predicate. Both the word of and explicit possession are represented as a possession relationship between the referent of the main noun phrase and the referent of the predicate noun phrase. At this point it can be decided what the principal referent of the noun phrase is; this node name is placed in *REF. At ANP3 the system looks for a restricted relative clause (described later), and simply POPS if one is not present. Referring to Table VII, the main noun phrase network calls ANP at the node NP. If ANP is successful, the system then decides which form of memory search should be done to identify the referent of the noun phrase. At NP2, if the noun phrase consisted only of a name, or pronoun reference to the previous statement, the system POPS, with the name of the corresponding node in *XO. If not, the system activates memory, and then goes to NP3. The activation is spread from the nodes in *TOPICS and *SOURCES. At NP3, if the embedded noun phrase was an indefinite singular noun phrase appearing in a predicate, then the assumption is that this noun phrase refers to a new instance of a concept. Thus, the transition is made to the node NP6 where an examination process, FCON, is called. Likewise, if the noun phrase appeared in a predicate and there was no determiner and the number of the head noun was plural, a similar assumption is made that a new instance of a concept is being referred to, and the system goes to NP6. At NP6, the CALLX to the examination network FCON will identify an existing node in semantic memory that corresponds to the referred-to concept. If only some of the predicates were matched to the definition, the propositions corresponding to the remainder are kept in *SBFL. If the CALLX is successful, the system specifies that a node should be built to refer to the noun phrase referent that is simply shown as a subset of the referred-to concept, and any additional predicates are attached to the new instance. The new instance node is given a cardinality of either 1 or 7, depending on whether the head noun was singular or plural. The next possibility at the node NP3 is to determine if the noun phrase is referring to a known individual. This is done by a CALLX to the examination network FINDV. This network will be described later. If it is successful, the individual found is considered to be the referent of the noun phrase and this node name is put into the variable named *XO. The next possibility checked for at NP3 is a CALLX to the same FCON examination network used before. Thus, it is assumed that if none of the previous tests have been successful, then the noun phrase is referring to a concept directly. In this case, if all of the predicates of the concept are found, the concept itself is taken to be the referent of the noun phrase, and the corresponding node name is put into the variable *XO.


59

The next arc at NP3, which goes to NPS, takes into account the fact that a noun phrase may refer to a concept that is not yet known to the system. The actions construct the memory structure for the concept and then designate this concept as the referent of the noun phrase. The arc that leaves NP3 changes the cardinality of the to-be-looked-for referent to 99, to indicate that we are looking for a concept. Then at the node NPS, an attempt is made to find the concept with the FCON examination network. The system will then construct a new concept, which is a subset of any existing concept that was found in memory, but with the new predicates, if any, added to it. This new concept is then the referent of the noun phrase. c. Clause Predicates. The second major network is the clause predicate network CLSPRED, shown in Table VIII, along with RESRELCLS. CLSPRED has two major branches, reflecting a basic division between sentences of a copular nature, in which the main verb is a form of to be, and

TABLE VIII NETWORKFOR CLAUSE PREDICATES AND RESTRICTEDRELATIVECLAUSES Node with arcs CLSPRED (CALL COPPHR *REF) >CLSPREDZ (CALL VERBPHR *REF) > CLSPREDZ

Action and comments Parse clause based on “to be” Parse verb-based clause

CLSPREDZ (AND) > CLSPRED (NIL) >POP RESRELCLS (RESRELPRN) > RESRELCLSZ (Not in PREDNP) > RESRELCLSI RESRELCLSZ (CALL CLSPRED *REF) > POP RESRELCLS1 (CALL PREPPHR *REF) > POP (CALL VERBCLS *REF) > POP

Check for “that” “That” not needed if in predicate noun phrase Clause predicate structure will be added to *REF in *SBFL Add structure to *SBFL Add new structure to *SBFL

David Kietas

60

sentences based on ordinary verbs. CLSPRED simply tries first the copular form, then the verb form, and goes to CLSPRED2. If the word and appears, it will parse another predicate. Then it POPS, returning to the calling network the structure-building function list corresponding to the propositions it has found. The restricted relative clause network, RESRELCLS, examines first for the presence of a restricted relative pronoun, such as the word that, and then calls the clause predicate network. If the system is not in a predicate noun phrase, either a prepositional phrase or a verb clause may follow without a restricted relative pronoun. The network COPPHR, in Table IX, parses the copular clause predicate. This network strips off any auxiliary verbs, and then tests for the presence of a copula, and then goes on to CP1. The system at this point defines a node for the appearance of the main proposition for the clause. At CP1, appearance of a negation causes a proposition to be constructed onto the main proposition of the clause indicating that the main proposition is false. At CPlA, either an adverb, adjective, prepositional phrase, or a definite or indefinite noun phrase may be found. The adjectives are simply attached TABLE I X

NETWORKFOR COPULATIVE PHRASES Node with arcs

Action and comments

This proposition will be main proposition *E

Add proposition structure negating original proposition, and make it proposition +E

Add proposition to +SBL showing modification of original proposition by concept of adverb (ADJ) > CP3

Add proposition to +SBFL showing modification of +REF by concept of adjective


61

TABLE IX (continued) Node with arcs

Action and comments

(CALL PREPPHR *REF) > CP3

Add new structure modifying *REF to *SBFL

(DEFDET) >CPREDNOM

Do not advance word, note definite determiner

(INDEFDET) >CPREDNOM

Do not advance word, note indefinite

CPREDNOM (CALL NP *XO) >CP3

determiner

If DEFDET noted, add proposition to *SBFL that *REF is SAME-AS NP referent

*XO If INDEFDET noted, add proposition to *SBFL that *REF is instance of NP referent *XO

Additional predicates can be supplied following AND followed by a copulative

as a proposition to the clause subject, which is in the variable *REF. The same is true of the propositions specified by prepositional phrase. The nodes CP3 and CP4 permit multiple items to be specified in a copular sentence by connecting them with the word and. If the predicate of the sentence is a noun phrase, a rather different type of action is performed at CPREDNOM. A definite noun phrase is taken to mean that the subject of the clause is the same thing as the referent of the predicate noun phrase. Rather than merging the two nodes in memory, a proposition is constructed which shows that there is a SAME-AS relationship between the two referents. If the predicate noun phrase is an indefinite noun phrase, then a subset relationship is being indicated between the subject referent and the predicate referent. In this case, the system merely constructs a proposition showing that the subject referent, *REF, is an instance (or subset) of the set indicated by the predicate noun phrase.

62

David Kieras

The verb phrase network shown in Table X is the most complicated in the model. Note that tenses and auxiliary verbs, with their complexities, are igndred. The network has two entries, VERBCLS and VERBPHR. VERBCLS is the entry for the network that is used where an -ing form of a verb may appear, as in RESRELCLS. VERBPHR is the main entry for a verb phrase. This network mainly determines first of all, whether a passive construction, or an active construction, appears and then parses the rest of the phrase accordingly. A direct object or logical subject noun phrase is optional. Two important nodes are defined at VERBPHR: *E for the main proposition for the “event” represented by the verb phrase, and *X2, which is the relational predicate for this main proposition. Adverbial information is represented as propositions attached to the *X2relational predicate node. At VERBPHRI, the signal for a passive construction is the presence of a form of to be. Notice that the COPPHR network would have parsed a truly copular sentence previously. At the node VERBPASS-ING, a passive verb phrase is distinguished from a phrase with an -ing verb. I f there TABLE X

NETWORK FOR VERB PHRASES AND PASSIVE VERB PHRASES Node with arcs VERBCLS (VERBING) > VERBPHR

VERBPHR (NIL) > VERBPHRl VERBPHRl (NEG) > VERBPHRI

Action with comments -1ng form of verb can introduce a verb phrase Define main proposition *E and node *X2 for main relation concept of verb Make main proposition *E a negation proposition attached to original main proposition

Appearance of copulative here (after COPPHR already tried) means a passive construction or a modification using -ing form of verb (NIL) > VERBACTIVE

Otherwise. have an active construction


63

TABLE X (continued) Node with arcs VERBPASS-ING (NEG) > VERBPASS-ING

Action and comments Negate as above

(CALL ADV *X2) > VERBPASS-ING

Add new structure to *X2 in *SBFL

(VERBING) >VERBACTIVE

If -ing form of verb, this is active construction

(VERBPP) > VERBPASSZ

Note passive construction Add structure showing verb concept as relation, and *REF as argument

VERBPASS2 (BY) > VERBPASS3 (PERIOD, COMMA, AND) > VERBEND

Sentence may be continued

(CALL VPREPPHR *X2) >VERBPASSZ

Add new structure to *SBFL modifying *X2 node

(CALL ADV *X2) > VERBPASS2

Add new structure modifying *X2 node

(NIL) > VERBEND VERBPASS3 (CALL PREDNP *XO) > VERBPASSZ VERBEND (NIL) >POP

Add new structure showing that *XOis subject of verb relation

If no noun phrase found following verb, add structure for a dummy subject or argument node as appropriate for passive or active construction

is the past participle of a verb (VERBPP), a structure is prepared anticipating that the logical subject of this verb will appear later, and a note made that a passive construction is being parsed, and a transition is made to VERBPASS2. At VERBPASS2, several possibilities are checked for. Adverbs will result in the corresponding structure. If the word by appears, a predicate noun phrase will then be parsed, and will be made the subject of the main proposition. A prepositional phrase in a verb phrase context might

David Kieras

64

appear, and will be parsed by VPREPPHR (see Table XII), and taken to be a modification of the relational predicate in the main proposition. Finally, at the node VERBEND the decision is made whether a noun phrase was encountered that was either the direct object or the logical subject in a passive phrase. If not, the corresponding place in the structure will be filled by a dummy memory node, with no attached propositions. The active verb phrase construction is processed with the network shown in Table XI, starting from the node VERBACTIVE. As before, when the verb appears, a relational structure is prepared. Then at VERBACT2, various forms of the remainder of the verb phrase are recognized. For example, if the word that appears, an embedded statement is present. Or if the word to appears, an embedded verb phrase may be present. These two networks for these possibilities are shown in Table XII. The simple direct object of the verb phrase is parsed by a call to PREDNP. Again, a prepositional phrase in the context of the verb phrase can be processed by a call to VPREPPHR, shown in Table XI1 also. Notice that throughout the network

TABLE XI NETWORKFOR ACTIVEVERB PHRASES Node with arcs VERBACTIVE (CALL ADV +X2) >VERBACTIVE (VERB) > VERBACT2 VERBACT2 (THAT) > STATEMENT-OBJECT

Action and comments As above Add structure showing +REF as subject of relation given by verb concept Object of verb may be main proposition of another statement

(TO) > VERB-INF-OBJ

Object of verb may be main proposition of another verb phrase

(CALL PREDNP *XO) > VERBACT2

Add structure showing +XO as argument of verb relation

(PERIOD, COMMA, or AND) > VERBEND (CALL ADV +X2) > VERBACT2

As above

(CALL VPREPPHR +X2) > VERBACT2

Prepositional phrase here is assumed to modify relational predicate node 0x2

(NIL) > VERBEND

TABLE XI1 MISCELLANEOUS PARSING NETWORKS Node with arcs STATEMENT-OBJECT (CALL EMBED-STATEMENT *XO) > VERBACTZ VERB-INF-OBJ (CALL EMBED-VERBPHR *REF *XO) > VERBACTZ

Action and comments Main proposition 0x0is argument of verb Main proposition is argument of verb, while *REF for the VERBPHR is same as current *REF Add structure showing that relational predicate node is modified by adverb concept

EMBED-VERBPHR (CALL VERBPHR *REF *E) >POP VPREPPHR (NIL) > VPREPPHRI

Add new structure to *SBFL, *XO = *E *REF = 0x2

VPREPPHRI (CALL PREPPHR *REF) > POP

New structure is added to *X2 relational predicate node

PREPPHR (PREP-A-B) > PREPPHRl

Note that preposition takes two arguments

(PREP) > PREPPHRI

Ordinary prepositions take one argument

PREPPHRI (NIL) > PREPPHRZ (CALL PREDNP *XO) > PREPPHR3

Add structure that *REF is subject of relation given by preposition concept If one-argument preposition noted, make *XOthe simple argument If two arguments noted, make *XO the ARGl

PREPPHR3 (one-argument) >POP (AND) >PREPPHR4 PREPPHR4 (CALL PREDNP 0x0) >POP

Make 0x0 the ARC2 of the relation

66

David Kieras

an adverb may appear at various places. The network ADV in Table XI1 interprets adverbs as a modification of whatever proposition or relational structure is being constructed, and PREPPHR in Table XI1 parses prepositional phrases in which the preposition specifies a relation with either one or two arguments.

C. MEMORYSEARCH The major process involved in prose comprehension is integration of the content of the individual sentences. The model assumes that this is mostly a matter of identifying the referents of noun phrases with nodes in memory, usually nodes created in processing the previous sentences. This requires searching memory to locate these prior nodes based on the information in the noun phrase. Memory search in this model consists of two steps. The first is a spread of activation through memory; the second is an examination of the results of the spreading activation. 1. Activation

The activation process corresponds to the conventional idea of a limitedcapacity spread of activation in (simulated) parallel from a set of activation sources. When the activation from the individual sources intersect, the resulting pathway to the point of intersection from each source is noted. In the model, the variable * SOURCES a list of the memory nodes from which activation will be spread during the spreading activation process. In practice, this list normally contains all of the concepts that were referred to in the noun phrase, or larger unit, being processed. Also, the passage topic is routinely included in *SOURCES in order to guide the memory search according to the overall passage context (see Kieras, 1977a). In somewhat more detail, the spreading activation process consists of a series of cycles in which activation is spread from each one of the activation sources. In a cycle, each activation source is allowed to spread its activation to any connected nodes, over a single link step, from the previously activated nodes for that source. There is a limit on how many nodes may be activated in each cycle. The total allocation is evenly divided between the sources, each of which is allowed to activate its share of the allocation in each cycle. Activation consists of placing a tag on the node. There is a unique tag for each activation source, making it possible to determine whether a previously activated node was activated from a certain source or not. If an attempt is made to activate a node that has already been activated from some other source, then an intersection has occurred. The two tags are merged, and thereafter both activation sources continue to activate connected nodes using the joint tag.


67

When all sources have intersected, the process stops. A new tag, called the PATH-TAG, is placed on the node of intersection and then the links are followed backwards to the original activation sources, with this tag being put on each node. Thus, the path of intersection between the sources is marked with a single tag. If a certain number of cycles passes without any intersections occurring, the process stops. The PATH-TAG is then placed on any paths of intersection that were discovered. Thus, the search process uncovers any partial intersections between the activation sources, as well as complete ones. The tagged paths are then passed to the path examination process, which determines whether the paths of intersection meet the specifications required for the particular memory search being done. 2. Path Examination In Kieras (1977a), it was argued that the search for referents in text comprehension could be very complex. As a simple example, notice that it is permissible to refer to an instance of a set by giving the name of its superset. For example, in a passage about Lassie, one might refer to the dog, or even the animal. However, it is impossible to tell from the syntactic features of the text itself how many levels or layers of subset-superset relations there will be in long-term memory between the immediate reference in the noun phrase (the animal), and the representation of the referent itself (Lassie). For this reason, resolving even these simple forms of reference requires a method of searching memory that is not committed to a specific structure of set relations being present. This is in sharp distinction to the MATCH process described by Anderson and Bower (1973), which requires that the configuration of structure being sought be specified exactly. A very flexible way of searching memory structures was developed for this model. The basic idea is that patterns of memory structure can be specified by parsing rules for the intersection paths uncovered by the spreading activation process. The rules can allow arbitrary repetitions of patterns, such as superset-subset relations. Transition networks are used to specify these parsing rules for memory structure patterns. These examination networks are processed by an interpreter in a manner very similar to that used for the ATNs, but with some important differences. In the examination network, each node represents a set of memory nodes, possibly consisting of only one memory node, and the arcs between nodes specify conditions by which the interpreter can go from one set of memory nodes to another. Usually the conditions on these arcs consist of tests for a certain type of memory link. For example, the interpreter can start from one set of memory nodes and go to the set of memory nodes that were connected to the first set by a subject link. Such tests usually specify that the memory links be tagged, meaning that the link is connected to a memory

68

David Kiem

node that is tagged with the special PATH-TAG from the activation process. The examination network interpreter works as follows: To leave an examination network node, the interpreter starts with the first memory node in the set associated with the examination network node. Using this first memory node, the interpreter applies the conditions specified on the first arc that leaves the examination network node. If this test is for a memory link of a certain type, the interpreter determines if there is at least one memory node connected to the present memory node by that type of link. If so, the interpreter makes a transition to the next examination network node, where the new set of memory nodes will be the set that was connected to the previous memory node by the specified type of link. If the condition on the arc is not satisfied, the interpreter tries the next arc with the same memory node. If that memory node does not meet the conditions specified on any of the arcs, the interpreter will take the next memory node in the list, and start again with the set of arcs. If the interpreter cannot leave an examination network node at all, it backs up to the previous examination network node, and tries the next arc or memory node for that previous examination network node. As with an ATN, the goal of the interpreter is to arrive at a final POP state. If it cannot, then the implication is that the sought-for pattern of memory structure cannot be found. The typical use of the path examiner is in the memory searches implied by a noun phrase. After activation has been spread from all the content words in the noun phrase, and the paths of intersection marked, the path examiner is started from the main class concept, which is often indicated by the head noun of the phrase. By following the activated paths in the manner specified by the examination network, it tries to match portions of the structure-building function list for the noun phrase with portions of the memory structure. Since the examination network can include recursive calls and loops, it is possible for the examiner to perform this matching in ways that are much more flexible than could be done by simply matching the structure-building function list directly against the memory structure, in a manner analogous to Anderson and Bower’s (1973) MATCH function. In more detail, in processing a noun phrase, the examination process attempts to match the propositions attached to the head noun against portions of the memory structure. Depending on the search requirements, some or all of these propositions must be matched for the search process to succeed. Certain variables are set then to the names of memory nodes identified by the matching process. For example, in searching for the referent of a definite noun phrase, all of the propositions associated with the referent of the noun phrase must be matched against structures in memory in order to conclude that the referent of the noun phrase has been located;


69

once the match is accomplished, the name of the node representing the noun phrase referent is known. But in defining a new concept, the propositions that cannot be matched against the memory structure define a new concept to be added to memory. Achieving the goal of allowing memory structures to be located without a commitment to the exact configuration of the desired memory structure is mainly a matter of how deeply nested the subset or superset relations are. The basic mechanism for this in the model is that if the examination network cannot locate a referent at the first level below the concept referred to by the head noun, it follows a subset relationship downward to the next concept and tries to match there. Thus, a referent that was at the bottom of a deep hierarchy of subsets below the main noun concept will require a lengthier path examination process. 3. Description of the Examination Network

a. General Description. Unlike the parsing process, the path examination process is not of much general interest. Thus, the description of it will be very superficial and will leave out many details. Table XI11 lists the major networks used in the path examination process, along with a concise summary of their function, and which other networks they rely upon. There are two major networks, FCON and FINDV. The input to these networks consists of the structure-building function list, *SBFL, for the sentence unit being processed, and the memory node that it should start from, which is usually the concept node corresponding to a main noun. FCON is used to find a concept; it returns the memory node name for the referred-to concept, after matching as many of the supplied propositions as it can. It also returns a list of the structure-building functions that specify the propositions that it could not match. The network FINDV is used to locate a specified individual or set of individuals. This network only succeeds if all of the propositions specified in the structure-building function list are matched against memory structure. The TRIM network in Table XI11 is supplied with a structure-building function list for the entire sentence after parsing and previous calls to FINDV and FCON have been completed, and activation has been spread from all of the concepts or referents mentioned in the sentence. Starting from the memory node for the sentence topic, TRIM determines which portions of the structure-building function list can be matched against material already in long-term memory. It then returns the subset of the structurebuilding function list that was not already present in memory, which specifies the unique new structure that must be added. This network was developed as part of the study of prior knowledge effects summarized below.

David Kieras

70

TABLE XI11 SUMMARY OF THE EXAMINATION NETWORK FCON

FINDV

FIND

FPREDS

FARGFORM

TRIM

Finds a concept Returns name of concept node list of unmatched predicates Calls FIND Finds individual referent node Succeeds only if all predicates matched Returns name of referent node Calls FIND Starting from noun concept, locates referent or concept node matching structure-building function list Calls FPREDS, FARGFORM, self. Starting from a referent or concept node, follows S-inverse-P links to match predicates in structure-building function list Calls FARGFORM, self. Like PFREDS, only follows A-inverse link to match relational structure attached to a referent or concept node Calls FPREDS, self Starting from a referent or concept node, trims redundant information from structure-building function list Calls FPREDS, FARGFORM.

6. Detailed Example. Both FCON and FINDV call a more basic network, FIND, which will be described in some detail to give the flavor of the examination network approach. FIND is shown in simplified form in Table XIV. Starting from the concept node for the head noun, the first arc leaving the node FIND follows an inverse predicate link to a proposition node, and then follows the subject link down to what should be the referent node corresponding to the head noun. The first attempt is to verify that this is the actual referent node. This is done by the subnetwork beginning at FI4, which tests this node for a match with the structure-building function list in terms of cardinality and the main noun concept proposition attached to it. Then, other propositions are matched at node FIS. This node either goes to POP if all of the items in the structure-building function list have been accounted for, or calls FPREDS. FPREDS has the function of finding and removing all propositions connected to the current node from the structure-building function list, by repeatedly attempting to find a pair of inverse subject and predicate links, and then matching the predicate that it finds with an item in the structure-building function list. The predicate

TABLE XIV

EXAMINATION NETWORKFOR FIND FUNCTION Node with arcs

Action and comments

FIND (tagged P-inverse) > FI2

Start from concept node

F12 (tagged S) > F13

Now at proposition node

F13 (CALL FI4) > POP (CALL FIE) > POP

If successful, current memory node is referent node Try a deeper level

F14 (node has correct cardinality and main noun concept) > FIS

Main noun concept proposition and cardinality are now matched

FIS (all structure matched) > POP

Finished if match complete

(CALL FPREDS) > FI5

Match predicate propositions

(CALL FARGFORM) > FIS (NIL) >POP FIE (current node is a concept) >F19

Have matched main noun concept proposition, now try to match referent to a subset

F19 (tagged P-inverse) >FI10 FIlO (tagged S) >FII1 FIll (CALL FI4) >POP (NIL) > F19

Attempt match at this node Descend another level

12

David Kieras

items might either be simple references to other concepts, or relational predicates. FPREDS might ultimately call itself in order to match up arguments found in the relational predicates. Once FPREDS is finished, F1S then looks for a structure in which the current node lies at the end of an argument link, using the network FARGFORM. FARGFORM operates in a manner similar to FPREDS, except that instead of matching propositions about the current node, it matches structures in which the current node appears as an argument of a relational predicate. After all of these matching operations have been applied as fully as possible, the network POPS at this level using the final arc. The network at the node F14 attempts to determine whether the current node is the desired referent node. If this attempt fails, F18 is called from FI3, and investigates the possibility that the head noun is actually a superordinate concept of the desired referent. This initiates a search down through a series of inverse predicate and subject links at FI9 and FIlO, with the F14 network called at FIll to attempt a match with the new candidate referent node. If necessary, the network may descend yet another .level in the hierarchy. Once FIND completes its work, it will POP back to its caller. If FIND succeeds in matching all of the items in the structure-building function list, FINDV will succeed. FCON will succeed even if FIND was not completely successful, but the ATN caller of FCON may or may not accept the result. 111. Comparisons with Data

The simulation model has been compared to reading time data using the multiple regression methodology described in Kieras (1983). The basis for the comparison is that the amount of processing that the model has to do on any given sentence in a passage should be related to the amount of processing that a person has to do reading the same sentence, and therefore the amount of processing in the model should predict the reading time taken by a subject to read each sentence in a self-paced reading procedure. These comparisons have only been concerned with the reading time variable; but notice that the model conforms to the above criterion of producing apparently correct behavior at the gross level. In the work reported in Kieras (1981), the amount of processing taken by the model is measured by counting how many ATN transitions must be made to parse the sentence, how many one-way links of memory structure must be built in order to represent the content of the sentence, and how many memory nodes must be activated while searching for referents, and how many items must be maintained in the topic list during processing the


73

sentence. These results can be used as independent variables in a multiple regression analysis to predict the reading times on each sentence. The quality of prediction that can be obtained depends on whether the data is collapsed across subjects or not (see Kieras, 1983), and ranges from 15% of the variance in the work reported in Johnson and Kieras (1983), to 70% of the variance when applied to the individual tasks described in Kieras (1981). When auxiliary variables (see Kieras 1981, 1983) are added to account for task differences that are not reflected in the model, the proportion of variance accounted for can range up to 95%. The results of this work will be briefly summarized, and then more general issues will be considered. A.

TASKAND COHERENCE EFFECTS

In Kieras (1981), the passages consisted of 11 sentences of very simple form, such as in Table XV, being based roughly on the Bransford and Franks (1971) passages. The passages were prepared in such a way that the sentences could be presented in any order, without disrupting the integrated content. That is, there were no causal chains or sequential events in the passage, so their interpretation could not be substantially affected by the order in which the sentences were viewed. However, some presentation orders had the property that each sentence has at least one given referent in it, and so the new sentence information can be simply attached to what is already present in working memory. Such orders were highly coherent. However, in other sentence orders, the initial several sentences of the passage had the property that only new referents are mentioned. This means that the reader must construct several unconnected pieces of memory struc-

TABLE XV EXAMPLE OF PASSAGES USED IN KIERAS (1981) The ants were eating the jelly. The ants were hungry. Thg jelly was grape. The ants were in the kitchen. E. The jelly was on the table. F. The kitchen was spotless. G. The table was wooden. H. The kitchen was equipped with the blender. 1. The table was against the stove. J. The blender was white. K. The stove was hot. A. B. C. D.

14

David Kieras

ture, and keep track of them until later sentences allow the structures to be tied together. Such incoherent orders take longer to read and are more difficult to recall (Kieras, 1978). The passages, in several different presentation orders, were presented to subjects one sentence at a time in three different self-paced reading tasks. In the free reading task, the subjects merely read each sentence for as long as they felt necessary to understand the meaning of the sentence. They were instructed that no later recall or test of any sort was to be given. In the topic choice task, after reading all the sentences in a passage, subjects had to produce a short title for the passage. The third task was a standard gist recall task, in which subjects had to recall the content of the passage immediately after reading it. The dependent variable of interest is the reading time, the time each sentence was viewed by the subject. The basic rationale for applying the model to the data obtained in the Kieras (1981) study was that the simulation model could explain the longer reading time in incoherent passages as being due to a combination of the reader having to construct more memory structure for incoherent sentences, since they contained only new referents, and also because of the time-consuming rehearsal processes required to maintain extra topics in the topic list. The different sentence orders had the effect of strongly varying the amount of new structure in the sentences, while leaving the surface form of the sentences intact. This made it possible to distinguish the processes involved in parsing, memory structure building, and topic list maintenance. Furthermore, the orders used had the property of placing the major thematic sentence either at the beginning or the end of the passages. This made it possible to determine whether there were any unique processes associated with highly thematic sentences. The results were generally accounted for very well by the model, in terms of proportion of variance accounted for. Table XVI lists the regression coefficients obtained in Kieras (1981). Apparently, highly thematic sentences do involve unique processes. The first sentence in a passage, or a highly thematic sentence appearing at the end of the passage, was read longer in the topic choice task than it should be, based on the amount of parsing and memory structure building involved. Parsing appeared to be a very fast process, whereas memory structure building was relatively slow. If the subject was in a recall task, additional time was required in order to encode each memory link for later recall. The memory search process, measured by the number of nodes activated, contributed a very small amount to the portion of variance accounted for. In fact, its effects were detectable only in the topic choice tasks and the free-reading tasks, which had much less variable and shorter reading times than the recall task. Other aspects of these results present problems for the model, and are discussed below.


75

TABLE XVI SUMMARY OF COMPONENT PROCESSING TIMESFROM KIERAS (1981) Process Parsing Memory activation WM structure building LTM Encoding Topic list processing Perceptual, response, and other processes Topic identification on each sentence in a topic choice task Processing a highly thematic sentence in a topic choice task Rehearsal processing during each sentence in a recall task

Time per step, item or sentence .006 sec/ATN transition

.003 sec/node .029seclone-way link .046 sec/one-way link .051 sec/topic entry 1.010 sedsentence

.232 sedsentence .441 sedsentence 1.887 sedsentence

B. PRIORKNOWLEDGE EFFECTS The work reported in Johnson and Kieras (1983) consisted of using the simulation model to explain the effects of prior knowledge of individual facts in the study of simple technical prose, such as the passage shown in Table I. Again, using a self-paced, sentence-at-a-time paradigm, subjects studied several passages for later gist recall. Unlike the work in Kieras (1981), there was no manipulation of passage sentence order. Again the variable of interest is the reading (study) time on each sentence. In addition, each subject provided measures of fiow much of the passage information they knew before the experiment. In the form of concern here, this consisted of a multiple-choice test for prior knowledge of the propositions in the passages. The amount of prior knowledge was a significant predictor of the reading time for the sentences. The simulation model was used to process the passages, and quantify the amount of parsing involved and the amount of memory structure that had to be constructed for each individual sentence. The value for the amount of structure built for each sentence was then modified for each individual subject, depending on which specific propositions that the individual subject already knew, based on the knowledge test. Thus, the simulation provided a specification of the full structure underlying each sentence, and then the subject’s test score specified how much of this structure could be reused, because it was already present in LTM, in representing the passage content. When the measures provided by the simulation, the knowledge test, and some additional empirical variables were used in a multiple regression anal-

David Kieras

76

ysis to predict the reading times, about 15% of the variance of the uncoffapsed data was accounted for. If subject variability had been removed, analogous to the collapsing of the data done in Kieras (1981), this would be expected to rise to approximately 30-4090 (see Johnson & Kieras, 1983). As it happens, this is as good a fit as could be obtained by a set of purely empirical predictor variables. The coefficients resulting from this fit are shown in Table XVII. Thus, this work showed that a certain interpretation of the effects of prior knowledge, namely that prior memory structure could be reused to represent the content of a passage, could be represented in a stimulation model to quantitatively account for the amount of processing that must be done on individual sentences, and could so at the level of individual subjects.

c.

PROBLEMS WITH THE

MODEL

1. Consistency and Adequacy

The results reveal several problems with the model. Since this simulation model handles only the microlevel processes in prose comprehension, the issue arises as to how important these processes are in the overall reading process. According to the above results, parsing is a very fast process that by itself accounts for very little of the variance in reading times. However, the coefficients for parsing shown in Tables XVI and XVII are substantially different. This could be due to processes not represented in the model, such as lexical access, which would differ substantially for the two different kinds of prose. Nonetheless, these two times should agree with each other more than they do. In contrast, the structure-building times are comparable. In the simple passages in Kieras (1981), the microlevel processes are apparently very important, because the simulation could account for at least half of the variance in each of the three tasks used, and in combination TABLE XVII COEFFICIENTS FROM JOHNSONAND KIERAS (1983) Process

Time per step or scale point

Parsing Structure building, including LTM encoding Deduct for known structure that can be reused Content word frequency Sentence importance

0.028 sec/ATN transition .068 sec/one-way link - .049 sec/one-way link

- . I 4 0 sedindex point .290 sechcale point

A Simulation Model for the Comprehension of Technicid Prose

71

with auxiliary variables, could account for over 90% of the variance in the three tasks combined. However, in the simple technical prose used in Johnson and Kieras (1983), only about 30-40% of the variance could be accounted for after removing subject variability. Apparently, the processes represented in the model account for only a small part of the reading times. However, it must be pointed out that the Johnson and Kieras work used a recall task, which in Kieras (1981) produced data with variability far exceeding that of the other tasks, and rather more complex materials were used as well. The data from Johnson and Kieras may simply be far noisier. But this suggests that the standard gist recall task is probably a poor choice for studying basic or microlevel processes in comprehension (cf. Johnson & Kieras, 1983). The memory search processes appeared to be very subtle and accounted for very little of the variance. This does not agree with the work of Miller and Kintsch (1980), which showed that the number of reinstatement searches, which would require considerable memory search, was an important predictor of the readability of somewhat complicated passages. Clearly, further work is needed on the role of memory search and inference processes. 2.

Recall Phenomena

In the Kieras (1981) work, the recall task seemed to involve additional processes which, while not represented directly in the model, could be tied to the model. That is, as shown in Table XVI,each link of memory structure required an additional increment of time in a recall task. Thus, the model could quantify this extra processing, even though it was not directly represented. However, the model-fitting work showed that the recall task involved a substantial amount of extra time across the board for each sentence (see Table XVI). Nowhere in the model is a corresponding process represented, or even hinted at. One interpretation is that in a recall task, subjects spend a certain amount of time on each sentence that is not related to the amount of information in the sentence. One possibility is that subjects rehearse the passage information for a certain constant amount of time on each sentence before going on to the next one. This process is revealed by attempting to fit the model to the data, but the model does not include it. An additional aspect of the recall task has to do with the phenomena of incoherence in Kieras (1981). The simulation quantified how many items had to be kept in the topic list and rehearsed for each of the incoherent passage orders that were used. However, the study times produced by subjects did not conform to the predicted pattern. The model predicted a

78

David Kieras

steady increase in memory load up to a certain sentence, where the disconnected portions of information would begin to be merged together, leading to descending memory loads through the rest of the passage. Subjects showed a reading time pattern consisting of an increase followed by a decrease, but the peak was much less pronounced than would be expected, given the usual estimates of short-term memory rehearsal times, and appeared much too early in the passage. The clear interpretation is that subjects did indeed slow down due to the increased need for rehearsal for the first few incoherent sentences, but then adopted some other strategy as the load increased. Perhaps they just ignored the additional items, and so leveled off their processing load much earlier than the model would expect. Again, applying the model revealed this process, but the model cannot account for it.

3. Macrolevel Processes Perhaps the most important processes that the model does not include are the macrolevel processes. A good example is the extra processing required on thematic sentences mentioned above. While these macrolevel processes can be approached with simulation models (see Kieras, 1982), the model described in this article does not perform any such processing. Macrolevel processes are extremely complex, and appear to be much more governed by the particular inference rules and general knowledge available to the reader than are the microlevel processes captured in the present simulation.

IV. Conclusion The simulation model described here certainly has its problems, but is fairly successful at accounting for microlevel processes in comprehension, not only at the level of the apparently correct gross behavior, but also in the quantitative detail of reading times of individual sentences. This shows that the simulation modeling approach can be used successfully at a detailed level to study comprehension processes. The work reported in Kieras (1982), tells a similar story for macrolevel processing. The problems in the model, revealed by where it fails to fit the data, point out, and can even quantify, phenomena that probably would have remained obscure without the model. For example, the effects of incoherence in passages are not as simple as it first appeared. Also, the recall task commonly used in comprehension research has substantial differences from other tasks; some of these differ-


79

ences appear to be straightforward additions of additional encoding processes, while others are as yet unexplained. Further, the thematic material in a passage, and the initial position in which it usually appears, are associated with special processes beyond the microlevel captured in the model. Thus, the simulation model can be used as a research tool, to reveal empirical phenomena, as well as a realization of a cognitive theory. Given that such work is possible, it seems likely that the future of research in prose comprehension lies with the more widespread and more thorough application of simulation models to account for performance data in comprehension tasks.

ACKNOWLEDGMENTS This work was supported by the Office of Naval Research, Personnel and Training Research Program, under Contract Number N00014-78-C-0509. Contract Authority Identification Number NR 157-423. Requests for further information should be directed to David Kieras, Department of Psychology, University of Arizona, Tucson, Arizona.

REFERENCES Anderson, J. R. Language, memory, and thought. Hillsdale, New Jersey: Erlbaum, 1976. Anderson, J. R.. & Bower, G . H. Human associativememory. Washington, D.C.: Hoit, 1973. Bach, E. Nouns and noun phrases. In E. Bach & R. T. Harms (Eds.), Universals in linguistic theory. New York: Holt, 1968. Bransford. J. D., & Franks, J. J. The abstraction of linguistic ideas. Cognitive Psychology, 1971, 2, 331-350. Clark, H. H., & Haviland, S. E. Comprehension and the given-new contract. In R. 0. Freedle (Ed.), Discourse processes: Advances in research and theory (Vol. 1). Norwood, New Jersey: Ablex, 1977. Collins, A. M., & Loftus, E. F. A spreading activation theory of semantic processing. Psychological Review, 1975, 82, 407-428. Johnson, W., & Kieras, D. E. Representation-saving effects of prior knowledge in memory for simple technical prose. Memory & Cognition, 1983, in press. Kieras, D. E. Problems of reference in text comprehension. In M. Just & P. Carpenter (Eds.), Cognitive processes in comprehension. Hillsdale, New Jersey: Erlbaum, 1977. (a) Kieras, D. E. Comparing a simulation to data using multiple regression. Presented at the 10th annual Mathematical Psychology Meetings, University of California. Los Angeles, August, 1977. (b) Kieras, D. E. Good and bad structure in simple paragraphs: Effects on apparent theme, reading time, and recall. Journal of Verbal Learning and Verbal Behavior, 1978, 17, 13-28. Kieras, D. E. Component processes in the comprehension of simple prose. Journal of Verbal Learning and Verbal Behavior, 1981, 20, 1-23.

80

David Kieras

Kieras, D. E. A model of reader strategy for abstracting main ideas from simple technical prose. Text, 1982, 2 , 47-82. Kieras, D. E. A method for comparing a simulation model to reading time data. In D. Kieras & M. Just (Eds.), New methods in comprehension research. Hillsdale, New Jersey: Erlbaum, 1983, in press. Kieras, D. E., & Bovair, S . Strategiesfor abstracting main ideasfrom simple technical prose. Technical Report No. 9 (UARZ/DP/TR-81/9), University of Arizona, November, 1981. Kieras, D. E., & Just, M. A. (Eds.), New methods in comprehension research. Hillsdale, New Jersey: Erlbaum, 1983. in press. Kintsch, W. The representation of meaning in memory. Hillsdale, New Jersey: Erlbaum, 1974. Kintsch, W., & van Dijk, T. A. Toward a model of discourse comprehension and production. Psychological Review, 1978, 85, 363-394. Miller, J. R., & Kintsch, W. Readability and recall of short prose passages: A theoretical analysis. Journal of Experimental Psychology: Human Learning and Memory, 1980, 6 , 335-354. Rumelhart, D. E. Introduction to human information processing. New York: Wiley, 1977. Woods, W. A. Transition network grammars for natural language analysis. Communications oftheACM, 1970, 13, 591-606.

Marcia K. Johnson STATE UNIVERSITY OF NEW YORK AT STONY BROOK STONY BROOK, NEW YORK I. Introduction ...........................................................

............................................................ A. Overview ......................................................... B. The Reflection System.. . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . . C. The Perceptual System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . D. The Sensory System.. . . . . . . . . . . . . . . . .. . . .. . . . . . . . . . . . . . . . . , . . . . . . . . E. Differentiating among Subsystems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . F. Activated Memory and Attention . . . . . , . . . .. . . . . . . . . . . . . . . . . . . . . . . . . . G. Stability of Entries.. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . H. The Concept of an “Entry” versus the Concept of “The Trace”. . . . . . . . . The Multiple-Entry Model and Other Theoretical Issues. . . . . . . . . . . . . . . . . . . . . A. Memories as Records of Complete Experiences . . . . . . . . . . . . . . . . . . . . . . . . B. Emotion.. ..... ... ....... ... . .. .... .. ... .... .... ........ . ....... .. C. Other Typologies of Memory.. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . , . . . . D. Amnesias .. . . . . . . . . .. . .. . . . . . .. . . . .. . . . . .. .. . . . . . . . . . . . . . .. .. . .. .. Summary .............................................................. References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

81 82 82 86 87 88

11. The Model

111.

IV.

90 95

99 101 103 103 104 106

110 115 116

I. Introduction Memory is what allows our past t o modify how we deal with events in the present. The most impressive thing about memory is the range of functions that it supports. The same memory system that recalls your vacation learns to play racketball. The same system that memorizes a part in a play is startled by faces that resemble the mugger who got your wallet. The same system that can instantly classify a strange animal as a bird struggles to identify the pharmacist when you run into him in the grocery store. Once it was thought that a few universal laws would apply to all of these situations: A continuous series of cases extends from the revival of one’s own experiences at one extreme to the automatic performance of a learned movement at the other, and the whole series belongs together. The difference between the extremes is a matter deserving of attention, but the likeness is more fundamental. (Woodworth, 1938, p. 5)

81 THE PSYCHOLOGY OF LEARNING AND MOTIVATION,VOL. 17

Copyright 0 1983 by Academic Press, Inc. All rights of reproduction in any form reserved.

ISBN 0-12-543317-4

82

Marcia K. Johnson

The search for general laws of learning was sometimes expressed as an attempt to explain all learning in terms of a single paradigm, for example, in terms of trial and error learning or in terms of classical conditioning (see McGeoch, 1952). This strategy was not successful, and movement toward specialization intensified (e.g., classical conditioning, operant learning, verbal learning, semantic memory). This specialization is good for many purposes, but there is a cost. A memory system designed for one function (e.g., language comprehension) would not necessarily be the same as one designed for multiple functions (e.g., language comprehension, pattern recognition, concept learning, dancing, developing preferences). Thus, it should not surprise us that theories created to account for a relatively limited range of facts ultimately prove to be limited. On the other hand, acknowledging the range of activities memory must perform does not necessarily commit us to looking for a single mechanism for all of them (Tolman, 1949). In this article I describe a general approach to memory in the form of a model of memory called a multiple-entry, modular memory system (MEM). The present model proposes that the multiple functions that memory serves are accomplished by several interacting, but distinguishable subsystems. To a large extent, these subsystems respond to different aspects of experience; hence, any particular event is likely to create multiple entries, that is, entries in more than one subsystem. This model tries to resolve or clarify a number of issues, for example, the controversy between trace theories and constructive theories of memory, the role of attention in establishing long-term memories, and the relations among various measures of memory (e.g., recall and recognition). In addition, MEM provides a framework for generating new hypotheses about a number of other areas, for example, the relation between memory and emotion, the relation between specific, autobiographical memories and general knowledge, and the problem of characterizing memory disorders.

11. The Model

A. OVERVIEW

Any model of memory should address a fundamental question: What is the relation between what we remember and what “really” happened? Memory theorists essentially ignored this question for years by explicitly or tacitly adopting a sort of naive realism. According to naive realism, our memory stores copies or traces of “stimuli,” and the “strength” of a mem-

A Multiple-Entry, Modular Memory System

83

ory is directly related to such things as how frequently and how recently we have perceived the corresponding stimulus. This view is the one that dominated work on memory from the time Ebbinghaus (1885/1964) first initiated experimental work on memory in the late 1800s until quite recently. While there have always been critics who questioned this view (e.g., Bartlett, 1932), it was not until the late 1960s and early 1970s that naive realism was finally overtaken by what now might be called “naive constructivism.” Although not identical to information processing, constructivism was fueled by the information-processing approach to cognition. In the view of many psychologists, the central feature of memory is rapid decay of information in the physical stimulus unless it is recoded. Recoding is viewed as a series of processes in which earlier products are discarded as successively higher recoding operations take place. For example, visual features detected by feature detectors become letters, letters become words, and words become meanings. The idea of recoding can be extended so that the highest levels of code are “holistic ideas” or “propositions” arranged according to familiar “schemata.” From this viewpoint, the representation of a complex event is a highly abstract, “constructed” representation and one that is not necessarily veridical or true to reality. In the constructivist view, we do not remember what we saw, but what we thought. Neither of these approaches can account for all of the data. People do embellish information and then sometimes mistake their embellishment for fact (e.g., Johnson, Bransford, & Solomon, 1973); at the same time, very specific sensory detail appears to be preserved in memory (e.g., Hintzman & Summers, 1973). Naive realists did not generate theories that could easily account for the sometimes dramatic errors and distortions in memory. Naive constructivists have not generated theories that can easily accommodate the sensitivity and sometimes remarkable accuracy of the memory system. In MEM, both accurate and inaccurate memory are consequences of a system that evolved for multiple functions. A basic idea represented in the MEM model is that multiple functions are very likely accomplished by a memory composed of independent, interacting subsystems. In MEM, there are three major subsystems, the sensory system, the perceptual system, and the reflection system. The first two abstract, store, and revive external, perceptually derived experiences, and the third creates, stores, and revives internally generated events (Johnson & Raye, 1981).’ ‘I should acknowledge at the outset that the boundaries between the sensory, perceptual. and reflection subsystems are not clear. However, I think these subsystems comprise useful ‘‘fuzzy sets” for organizing findings, hypotheses, and speculations.

Marcia K. Johnson

a4

Each of the subsystems is specialized for a number of functions, for example, the sensory system for detection of stimuli and the development of certain sensory-motor skills, the perceptual system for identifying relationships among objects and recognizing the familiar, the reflection system for planning and for voluntary recall of events. A particular event may be processed by all subsystems, creating memory traces, or “entries” in all subsystems (see Fig. 1). If you stare at Fig. 1, it reverses in Necker-cube fashion: each of the subsystems overlaps with the other and each can be seen as in front, in back, or in the middle. As this characteristic of the figure suggests, the various systems interact continuously. The subsystems should be imagined as working more or less simultaneously rather than serially; they are more like light filters responding to different aspects of experience than stages in a transformation. To illustrate the idea of multiple entries, consider the activity of learning to play tennis. This is a complex skill in which various components are probably largely mediated by different subsystems: learning t o anticipate the trajectories of tennis balls is a sensory function; learning to see relations among the opponent’s position on the court, posture, and racket orientation that signal his or her probable shot (lob, down the line, etc.) is a perceptual function; learning to recognize the opponent’s strategy or to plan one (vary pace, baseline game, etc.) is a reflection function. Consider the activity of reading a story. Sensory processes create sensory entries that reflect the configuration of light and dark on the page, and the

rm Stimulus-driven motor skills

Percealual thresnold Identification

of degraded stimuli

Sensory

Generative motor skills

x x x Recognition

Perceptual

Reality monitoring

Reflection Recall

Fig. 1. A particular event creates entries in three subsystems of memory, sensory, perceptual, and reflection. The shading indicates activated entries; the darker the shading the greater the likelihood that the activation will recruit attention. Various memory tasks are listed near the subsystem(s) that they are most likely to draw upon.


85

visual scanning operations induced by this configuration as various parts of the stimulus are localized and figure-ground relations emerge. Perceptual processes create perceptual entries that reflect specific instances of identifiable patterns, for example, words and phrases in a particular typeface, in a particular spatial arrangement on the page. In addition, reflective entries preserve a record of internally generated thought processes such as imaging, drawing inferences, and other embellishing. Later, subjects might show reduced perceptual thresholds for words in the stories, be able to discriminate exact words from paraphrases in a forced-choice recognition test, and yet produce pragmatic inferences in recall (i.e., claim that they read information that they only inferred). These results are not contradictory; they are all possible because memory is a multiple-entry system. We have evolved memory structures that tie us to reality in fairly direct ways through our sensory and perceptual systems. These systems allow us to detect highly probable recurrences and invariants (Brunswik, 1956; Gibson, 1966) in an external reality to which we must adapt. However, we also have evolved mental structures that allow us to produce and retain a “selfgenerated” reality as well. These allow us the independence from ongoing perceptual stimulation that is necessary for anticipating, drawing inferences, reminiscing, planning, and otherwise manipulating ideas. The benefits of the reflection system for creative invention are clear, but we pay for them. The cost is occasional confusion between the real and the imagined in memory (Johnson & Raye, 1981). In general, failure on a memory task does not necessarily imply a loss of information from memory. Different memory tasks draw differentially on different subsystems in MEM; hence, an entry not revealed by one memory task may be revealed by another. This means that while subsystems interact continually, different subsystems may predominate at different times. The memory subsystems also interact with attention mechanisms. At any given moment, only part of memory (potentially consisting of entries and functions from all subsystems) is activated-this is “activated memory.” Only a subset of activated memory receives attention, that is, reaches awareness. Information within and between the different subsystems differs in ability to recruit attention; thus, some entries inhibit others from gaining attention. These ideas are represented graphically in Fig. 1. They are developed in more detail in the body of the article, along with a consideration of the way in which MEM relates to certain other issues, for example, emotion and memory, semantic versus episodic memory, and amnesia. I shall begin with a brief characterization of each subsystem and then consider evidence bearing on differentiating among them.

86

Marcia K. Johnson

B. THEREFLECTION SYSTEM The term “reflection” is taken from the British philosopher, John Locke, who suggested that our knowledge originates from two sources, experience and reflection. In MEM, reflection is the active thinking, comparing, and judging function of the mind. It includes planning, creating images, organizing, elaborating, rehearsing-the processes emphasized by cognitive theorists. What these processes all have in common is that they are generated by the subject. Compared to sensory and perceptual processes, they depend less on immediate perceptual data. The reflection record preserves our interpretation of and “commentary” on perceptual events, our fantasies, efforts to understand, and our attempts to control what happens to us. Many activities that we must perform without the support of external stimuli depend heavily on the reflection record. For example, the reflection system allows us not only to generate hypotheses, but to keep track of the ones we have already tested and evaluate new ones in light of prior evidence and goals (e.g., Levine, 1966; Newel1 & Simon, 1972). It allows us to construct mental maps and other representations of related facts and use them to guide actions or responses in memory tests (e.g., Bransford, Barclay, & Franks, 1972; Levine, Jankovic, & Palij, 1982). The reflection system helps us to find relationships between new information and old knowledge so that we can comprehend and draw inferences (e.g., Bransford & Johnson, 1973; Kintsch, 1974). Free recall of events especially depends on entries in the reflection system. Reflection processes integrate and organize (e.g., Mandler, 1967; Tulving, 1968) by creating or reactivating relationships between one event and another (e.g., between target items and other information, such as other targets, elements of the experimental context, or episodic events from the subject’s life). The associations produced by reflection activities have been given various names, for example, interitem (Mandler, 1980), contextual (Jacoby & Dallas, 1981), elaborative (Craik & Tulving, 1975), and vertical (Wickelgren, 1979). There is considerable evidence that high levels of free recall depend on such reflective activities (e.g., Bellezza, Cheesman, & Reddy, 1977; Tversky, 1973). Not all reflective activities are organizational. Some are less likely to create interrelations among events. Covert rehearsal that is not elaborative (“maintenance rehearsal,” Craik & Watkins, 1973) does little to improve recall, but increases recognition (Glenberg, Smith, & Green, 1977; Rundus, 1977). Thus, recognition can draw on the reflection record, although, as emphasized in the next section, the perceptual record appears to be particularly influential in recognition under many circumstances (cf. Mandler, 1980).


87

Both elaborative and nonelaborative reflection typically require “effori,” use “cognitive capacity,” occupy “attention,” or depend on “controlled processes” (Hasher & Zacks, 1979; Kahneman, 1973; Shiffrin & Schneider, 1977; Tyler, Hertel, McCallum, & Ellis, 1979). However, reflection should not be equated with cognitive capacity or attention; many perceptual processes use capacity or recruit attention as well.

C. THEPERCEPTUAL SYSTEM The information comprising particular events varies in degree of “organization” much as a random dot pattern (e.g., Julesz, 1971) differs from a meaningful scene (Biederman, Rabinowitz, Glass, & Stacy, 1974; Friedman, 1979; Mandler & Stein, 1974) or a fly buzzing erratically around the room differs from a ping pong game. The distinction between sensory and perceptual systems in MEM attempts in part to capture this sort of difference. As the discussion below of the sensory system suggests, random dot patterns may create permanent memories (Stromeyer & Psotka, 1970), and you could get better at tracking the fly; however, phenomenal experience, including the sort of remembering that is associated with a sense of pastness, is ordinarily dominated by the more organized products of the perceptual system. That is, perceptual functions give our experiences the characteristically organized and relational quality that they have, qualities emphasized by the rationalist philosophers (e.g., Descartes and Kant) and Gestalt psychologists (cf. Goldmeier, 1982). Perceptual functions involve both innate and learned ways of organizing stimuli. For example, certain innate, automatic coding categories or processes may register experiences as “causal,” or “similar,” or “symmetrical,” or, in the case of humans, as “face-like,” or “language-like.” Other temporal and spatial aspects of experience might be given by innate qualities of perceptual functions as well. The fact that elements are near each other or have common fate would automatically make them cohere into organized percepts. Learned perceptual categories build up with experience as well. Hence, at some time we begin to perceive whole words, rather than individual letters, etc. The perceptual record is particularly important in accounting for the storage of complex patterns and in producing a sense of familiarity, that is, the sense of having seen something before. Perceptual entries create and are in turn guided by schemas or mental structures (Hochberg, 1981). However, these are perceptual schemata and should not be too casually equated with other sorts of mental contexts (Bransford & Johnson, 1973), schemata (Bartlett, 1932) and scripts (Schank & Abelson, 1977) that have been shown to influence memory, and are much more likely to involve reflective pro-

88

Marcia K. Johnson

cesses. The point is important because current textbooks tend to stress findings suggesting that memory is based only on meaning and not on perceptual characteristics of stimuli, but this is not the case. People are better at recognizing that they have seen words before if the typeface in which the words are displayed on the test is exactly the same as the typeface in which the words were shown originally (Hintzman & Summers, 1973). The same is true for auditory features-subjects are better able to recognize a word they heard before if it is spoken in the same voice on the test (Craik & Kirsner, 1974; Geiselman & Bjork, 1980). For prose material, originally presented sentences can be discriminated from paraphrases over intervals at least as long as a week (Bates, Masling, & Kintsch, 1978; Dorfman, 1979; Keenan, MacWhinney, & Mayhew, 1977; Kintsch & Bates, 1977). Memory for exact wording and specific details has been reported many times (e.g., Hasher & Griffin, 1978; Tyler & Ellis, 1978; Tzeng, 1975), even for situations very like those originally interpreted as refutations of verbatim memory (e.g., Bransford & Franks, 1971; Sachs, 1967). This sensitivity of memory to repetition of surface features and specific detail challenges the view that all we store are abstract “meanings.” Specific perceptual aspects of events are stored as well. D. THESENSORYSYSTEM It has often been proposed that sensory information that is not encoded beyond the sensory level is rapidly lost (e.g., Atkinson & Shiffrin, 1968; Broadbent, 1958; Craik & Lockhart, 1972; Neisser, 1967; Sperling, 1960). An experiment by Turvey (1967) using the Sperling task illustrates this view: on each trial, subjects received a 50-msec exposure of a 3 x 5 matrix of randomly selected digits. The subjects were cued with a tone to recall one of the three rows of digits. One matrix was repeatedly interspersed among the others, but recall was not significantly better on tests involving the repeated matrix than on tests involving nonrepeated matrices. Turvey (1967) suggested that the Sperling task produces preperceptual traces that “should be excluded from the domain of memory (p. 292).” In contrast, MEM assumes that elementary sensory information is stored in a permanent form. However, “stored” does not necessarily imply that an entry is accessible to voluntary recall processes. Changes in memory as a function of experience may reveal themselves in some tests, but not others, for example, in reduced perceptual threshold or faster processing (e.g., in lexical decision or naming tasks) but not recognition, or in recognition but not recall (Johnson, 1977). According to MEM, performance on recall tests is particularly dependent on prior reflection. The Sperling procedure, involving brief presentations and massive interference from recombinations


89

of the same items, obviously is designed to prevent reflection activities. Yet, with enough repetitions, even in this task recall does improve (Merluzzi & Johnson, 1974). Tests other than recall, however, should be more sensitive to effects of prior sensory processing. This prediction receives support from an experiment by Kunst-Wilson and Zajonc (1980). They found that subjects liked previously presented random polygons better than new ones, even though the prior exposure duration of the “old” polygons was only 1 msec! This result contradicts the idea that brief exposures that presumably produce only sensory processing have only transient consequences, and supports the idea that a sensory record is created by such sensory processing. There is other evidence that relatively “low-level” processing leaves an entry. For example, Haber and Hershenson (1965) showed that subjects’ ability to identify a word increased over trials even though the duration of presentation on each trial was briefer than what would be necessary to identify the word on the first trial. People are faster in deciding a letter string is a word the second time it is presented, and the facilitation is greater when the word is presented in the same modality (e.g., visual-visual rather than auditory-visual) on both occurrences (Forbach, Stanners, & Hochhaus, 1974; Kirsner & Smith, 1974; Scarborough, Cortese, & Scarborough, 1977). Evidence continues to accumulate showing that relatively “shallow” levels of processing produce quite long-lasting consequences in memory (Jacoby & Dallas, 1981). Further demonstrations of memory for specific aspects of events await only the application or development of more sensitive memory tests (Johnson, 1977). The sensory system in MEM is presumed to be sensitive to quite elementary properties or changes in the stimulus array. Just what these sensory properties are remains to be clarified. Some suggestions are made in the next section. However, it is clear that we shall not look for them unless we propose that they are there. In addition, those properties that create memory entries will probably have to be specified for each sensory system individually. Whatever these properties, the fact that they accumulate in the sensory system allows us to get better with experience at dealing with stimulus information of which we are rarely, if ever, aware. Sensory properties allow us to know that an event has occurred without necessarily knowing what the event was. Furthermore, the sensory system is probably involved in establishing associations or schemata relating stimulus properties and some responses, for example, between the sounds that we make and articulatory movements that produce them. In general, the sensory record probably plays a large role in improvements in various motor skills, such as developing hand-eye coordination, learning to make appropriate postural adjustments to changes in external cues, adjusting to weightlessness, improvements in

90

Marcia

K. Johnson

tracking tasks, and other largely stimulus-driven tasks (e.g., Kowler & Martins, 1982). In short, information that is not accessible to voluntary recall and not necessarily recognizable as comprising familiar patterns can be extremely useful information in that it can support the performance of rather complex skills. The sensory record preserves and accumulates such information.

E. DIFFERENTIATING AMONG SUBSYSTEMS The difficult problem is to analyze the memory system into component processes. How many different processes do we need? Surely we need more than one undifferentiated process, but not a separate process for every unique event. We need a way of classifying or grouping some processes together because they have something in common. One classification scheme is suggested by MEM; processes are grouped into the major categories of sensory, perceptual and reflective; within each of these are subprocesses, for example, seeing, hearing, planning, and comparing, and each of these could be further subdivided. Unfortunately, we cannot simply classify these systems in terms of tasks because performance in almost any task will be supported by more than one subsystem. For example, there are sensory and perceptual components in tasks that we think of as largely reflective (e.g., reading and speaking), and there are reflective components in tasks we think of as largely sensory and perceptual (e.g., playing tennis and driving). To fully characterize learning and memory in such complex situations would require an understanding of the way that separate subsystems work, the way that they interact, and the way that their interaction changes with practice. For example, reflection may help define perceptual patterns to look for, but after extended practice, perception in terms of these patterns may proceed without reflection.’ However, although the job of disentangling sensory, perceptual, and reflection subsystems is difficult, there is some reason to believe that thinking in terms of these subsystems might result in a useful conceptual framework. 1. Evidence from Studies of Cognition

Several lines of evidence suggest that it is reasonable to categorize memory functions and records in terms of the subsystems proposed in MEM. lAn arrangement in which different subsystems had substantial amounts of independence would be especially valuable for doing more than one thing at a time (e.g., chewing gum, driving a car, and rehearsing the day’s lecture).


91

First, while reflectively generated (e.g., imagined) and externally derived (perceived) events are clearly similar, they are not exactly the same. If you take fairly clear cases of perceived and imagined events, such as a visually presented picture and an imagined picture, the features do not seem to be equally distributed across perceived and imagined memories. For example, memories of perceptions typically have more specific sensory features than memories of imaginations, and imagination creates a more embellished record of the operations involved in generating the image. These differences can be used by subjects to discriminate between perceived and self-generated events in memory (“reality monitoring,” Johnson & Raye, 1981). If you make imagined events more like perceptual events by increasing their sensory-perceptual characteristics (e.g., Johnson, Raye, Wang, & Taylor, 1979; Foley & Johnson, 1982) or by decreasing the reflective operations that went into producing them (Johnson, Kahan, & Raye, 1981; Johnson, Raye, Foley, & Foley, 1981), you will decrease the accuracy of reality monitoring. Thus, as shown in Fig. 1, reality monitoring draws on both perceptual and reflection records. The present model, along with the reality-monitoring model, provides a resolution of the controversy between realists and constructivists: both perceptual and reflective entries are preserved. A test drawing on the perceptual system is likely to yield evidence for physical features; a test drawing on the reflection system is likely to yield the sorts of omissions, elaborations, and distortions introduced during reflection. Failures in reality monitoring will sometimes lead people to treat reflective entries as if they had been perceptions, and vice versa (for a description of decision processes involved in reality monitoring, see Johnson & Raye, 1981; for a review of evidence related to the realist/constructivist controversy, see Alba & Hasher, 1983). As a task, reality monitoring can potentially help to illuminate the similarities and differences in entries derived from perception or generated via reflection because it explicitly requires discrimination between the two. Second, the subsystems can at least partially be dissociated on the basis of the patterns of relationships among various memory tasks. If we were dealing with a unitary memory system in which information existed only at various levels of strength, some memory tests would appear easier than others because they would have lower thresholds for successful performance. However, overall, particular variables should have the same effect on performance on all measures. Furthermore, any memory that exceeded the threshold for a more difficult test should exceed the threshold for an easier test. Both of these criteria for a unitary system can be shown to be false. For example, reality monitoring is not necessarily correlated with measures of recall or recognition (Johnson & Raye, 1981). Furthermore, recognition

92

Marcia

K. Johnson

almost always increases with frequency of occurrence, and recall often does not. In fact, measures of cued recall and recognition for the same items indicate that success on one task is largely independent of success on the other, suggesting that different features are sampled by the two memory tasks (Flexser & Tulving, 1978). This independence is consistent with the suggestion made above that recall draws heavily on the reflection system while recognition draws heavily on the perceptual system. Not only are recognition and recall at least partially independent, but measures of recognition and measures related to perceptual threshold also appear to show some independence (Jacoby & Dallas, 1981). This would be expected if threshold tasks draw heavily from the sensory system and recognition tasks from the perceptual system. Third, it is possible that selective attention tasks can be developed to help illuminate the difference between the sensory and perceptual systems. For example, Rock and Gutman (1981) showed subjects a series of overlapping green and red nonsense figures and asked the subjects to attend selectively either to the green figures or the red figures. On a later recognition test, subjects could not distinguish the unattended shapes from other, similar shapes. Rock and Gutman (1981) use these data to make the point that perception of shape requires attention: “Phenomenal shape entails an apprehension by the observer of the exact spatial interrelationships of the parts of the figure to one another and of the relationships of these parts to the up-down, left-right spatial coordinates (p. 282).” It would also be consistent with the present model if objects seemed familiar only if they had been a phenomenal object in the past in the sense described by Rock and Gutman (see also Hochberg, 1971); the sort of processing that produces a unique and organized whole would be characteristic of the MEM perceptual system. As has already been suggested, successful recognition that an object has occurred before very likely depends on perceptual entries. Equally important, however, is that in a subsequent experiment by Rock and Gutman specifically designed to discover what, other than shape, might have been processed, subjects did remember some things about an unattended figure-the size, whether its shape was open or closed, and whether the contour of the shape was a continuous line, a dotted line, a dashed line or composed of small xs. Rock and Gutman (1981) proposed that there was “a failure of form perception simultaneous with successful perception of other properties of the object’’ (p. 283). It is possible that these other properties were picked up by what in MEM is called the sensory system. Similarly, Treisman and Gelade (1980) have proposed that there are a number of elementary dimensions including orientation, brightness, direction of movement, and texture segregation or figure-ground grouping that are combined to produce conscious percepts. These elementary dimensions


93

are the sorts of properties proposed here to be included in the MEM sensory system. The present framework proposes, in addition, that sensory entries accumulate over successive occurrences and that they can enter into functional associations with other stimuli and responses, including movement and emotions. In summary, while there is still much to be done in the way of establishing criteria for deciding when we are dealing exclusively (or even primarily) with one subsystem in memory and not another, the present framework provides a vehicle for organizing some of the available evidence and for suggesting further lines of research. Without further work, only a tentative grouping of tasks that might help us explore characteristics of the various subsystems can be offered. Tasks that appear to reflect the influence of the sensory record include measures of the effects of prior exposure on perceptual threshold or identification of very briefly presented stimuli (Jacoby & Dallas, 1981), on identification of degraded stimuli (Warrington & Weiskrantz, 1970), lexical decision (Scarborough et al., 1977), naming (Durso & Johnson, 1979), and the development of preferences (Kunst-Wilson & Zajonc, 1980). Selective attention tasks (e.g., Rock & Gutman, 1981; Treisman & Schmidt, 1982) are promising as well. Properties of the perceptual record are likely to influence performance in recognition tests involving complex stimuli, for example, sentences (Keenan et al., 1977), pictures (Biederman et al., 1974; Tulving, 1981), faces (Light, Kayra-Stuart, & Hollander, 1979), and nonsense forms (Rock & Gutman, 1981). The characteristics of the perceptual record in memory might also be clarified by studying “implicit learning” (Reber & Lewis, 1977) and the development of perceptual categories (e.g., Cerella, 1979; Herrnstein & devilliers, 1981; Posner & Keel, 1968) and their role in the memory of experts (e.g., Chase & Simon, 1973). Memory for perceptual information also can be assessed fairly directly, for example, by testing what subjects remember about the color (Nilsson & Nelson, 1981) or location of items (Johnson, Raye, Foley, & Kim, 1982; Rothkopf, 1971). One of the major messages of cognitive research has been the importance of what are here called reflective activities in remembering, and the free recall task has been particularly revealing in this regard (e.g., Bartlett, 1932; Bower, 1972; Mandler, 1967; Tulving, 1968). The properties of reflective entries have been explored in the context of studies of cognitive maps (e.g., Hanley & Levine, 1983) and other representations constructed from sequentially presented information (e.g., Bransford et al., 1972), and in studies of integration (e.g., Loftus, 1975), imagery (e.g., Paivio, 1971), and of the role of inferential processes, schemata and scripts (e.g., Bower, Black, & Turner, 1979; Bransford & Johnson, 1973). There have been relatively few attempts to explicitly compare the properties of memory entries created by

94

Marcia K. dohnson

reflection and those created by perceptual processes (e.g., Jacoby, 1978; Johnson & Raye, 1981; Peterson, 1975; Slamecka & Graf, 1978), but this provides a potentially powerful way of exploring characteristics of reflective entries. 2. Neuropsychological Evidence

The general scheme proposed in MEM also finds some support from research in neuropsychology. Some cortical zones are composed of neurons that are responsive to a specific sensory modality. Lesions in these areas produce fundamental sensory losses, such as the absence of sensation in an area of the visual field (scotomas). Other zones are also modality specific, but synthesize sensory input into more complex relations. Lesions in these areas produce agnosias, “the inability to combine individual impressions into complete patterns” (Kolb & Whishaw, 1980, p. 194). For example, Luria (1973) showed a patient a picture of a watch with several lines superimposed over it. The patient said it was a “chick hatching from an egg and some funny circles.” It is as if the patient was basing an interpretation on some physical features of the stimulus, but suffered a severe disruption of the ability to resolve relational aspects of the picture. In still other zones “sensory modalities overlap, enabling the sensory systems to integrate their input and to work in concert with one another and with information already stored in the nervous system” (Kolb & Whishaw, 1980, p. 244). Lesions in these zones disrupt cross-modality matching, for example, identifying the unfamiliar object in a visual array that is the same as an object you feel with your hand but that you are not allowed to see. These lesions do not, on the other hand, disrupt basic vision, hearing, or somatic sensation. There are other lines of neuropsychological evidence suggesting that different aspects of perception are mediated by different anatomical systems: in vision, the detection of intensity and location seem to be a function of one system, whereas pattern discrimination seems to be a function of another. For example, experiments with monkeys, rats, and hamsters indicate that lesions of the visual cortex disrupt pattern discrimination but not discrimination among lights differing in location. On the other hand, lesions of the superior colliculus disrupt localization, but not pattern discrimination (Schneider, 1969). The “fibers leaving the colliculus appear to connect with the motor control system for eye movements, head orientation, and postural adjustments” (Lindsay & Norman, 1977, p. 77). Patients have been reported who could grasp moving objects and report the direction of motion while at the same time claiming that they did not “see” the object (Weiskrantz, Warrington, & Sanders, 1974). The point here is not to suggest that the sensory system and perceptual


95

system can be equated with the specific brain structures mentioned above, but rather to point out that some aspects of perception function in the absence of others. This indicates that perception, even within a single modality such as vision, is composed of component processes with some degree of modularity. Thus, the conceptual division of perceptual memory into the subsystems used in MEM is at least physiologically plausible. Furthermore, both the physiological data and results of cognitive studies provide some reason to group together the detection of changes in brightness and elementary figure-ground relationships, localization of stimuli, and some basic motor functions in one system (sensory), and the detection of relational attributes, complex pattern perception, object identification, and familiarity in another system (perceptual). In a later section, evidence about amnesia patients will be discussed that is consistent with the idea that a third system (reflection) is largely responsible for establishing the conditions of voluntary recall and relating events to personal identity. F.

ACTIVATED MEMORYAND ATTENTION

Cognitive activities such as perceiving, thinking, or remembering create patterns of neural activation in the brain. This ongoing activation is sometimes called short-term memory, working memory, or activated memory. In MEM, activated memory consists of currently activated information from all subsystems, sensory, perceptual, and reflection. Activated memory is created by ongoing entry processes. But we are not equally aware of all activated entries; only a subset receive conscious attention (e.g., Posner, 1978).

1. Attention and Acquisition

Some investigators have proposed that only what is attended to or subjected to controlled processing is stored permanently in long-term memory (LTM) (e.g., Broadbent, 1971; Shiffrin & Schneider, 1977). In contrast, MEM proposes that whatever is processed by a subsystem is entered into the corresponding record. Several different types of findings have been taken as evidence for the idea that storage depends on attention; however, an argument can be made that this evidence has either been contradicted by other findings or does not constitute the most stringent test of the attention-dependent storage hypothesis. In studies of memory for prose, people typically remember general ideas better than exact wording (Bransford & Franks, 1971; Sachs, 1967). Because it is reasonable to assume that people pay attention to meaning and

96

Marcia K. Johnson

not to physical characteristics of the words that convey the meaning, such findings have been used to support attention-dependent storage (Shiffrin & Schneider, 1977). However, as mentioned previously, more recent evidence (e.g., Hasher & Griffin, 1977; Keenan et al., 1977; Kintsch & Bates, 1977; Tyler & Ellis, 1978; Tzeng, 1975) shows considerable retention of specific detail, including exact wording, in memory for prose. The results of studies in which orienting tasks are used to specifically direct attention to one attribute or another (e.g., sensory versus semantic or color versus form) of events suggest a similar conclusion. While what is attended to is usually remembered best, attention directed at a particular stimulus feature does not eliminate encodings based on other features. For example, Nelson and Walling (reported in Nelson, 1979) demonstrated that biasing a sensory encoding of a target word by presenting it with a rhyming context word (TOWER-FLO WER)did not eliminate the effectiveness later of semantic cues (ROSE) on a recall test. Some of the most compelling evidence for the attention-dependent storage position comes from studies using the dichotic listening technique. When subjects shadow information presented to one ear, their memory for information presented to the other ear may not be above chance, even for items repeated several times in the unattended ear (Moray, 1959). One explanation is that such stimuli contact a representation in LTM, but are filtered out by preattentive processes and hence do not leave a trace of their occurrence. In the multiple-entry model, these are stimuli that become part of activated memory, but that do not recruit attention; they are “nonattended” stimuli, but they are entered in memory. According to MEM, both subthreshold perception (in the sense of activated but not attended to) and subthreshold reactivation (memory revival) should have consequences for thought and behavior (see also Erdelyi, 1974). Furthermore, these consequences are viewed as permanent, not transient; for example, they should cumulate over successive subthreshold occurrences (e.g., Haber & Hershenson, 1965). Admittedly, there is not much evidence for this proposition. However, investigators have used relatively insensitive tests (recall and recognition) to look for memory for unattended information. Unattended information is not reflected upon and therefore would not be expected to be recalled. To the extent that reflective processes play some role in recognition, or to the extent that recognition depends on prior phenomenal perception, recognition would be poor as well. However, automatic activation of pathways (Posner, 1978), or sensory and perceptual processes that are stimulus driven (Norman & Bobrow, 1975), whether or not the subject attends, should create entries in memory. Thus, tests that are more likely than recall or recognition tests to draw on these entries, such as measures of perceptual threshold or processing time in lexical de-


91

cision or naming tasks, should detect prior exposure to unattended information.' Along these lines, evidence is accumulating that information that is not perceived in the sense of consciously identified can affect subsequent processing. For example, subjects in a lexical decision task are asked to press one button if a word occurs and another if a nonword letter string occurs. Responses to a word such as doctor are faster if a related word precedes it, for example, nurse (Mayer & Schvaneveldt, 1976). Facilitation from related words occurs to about the same degree even if the first, prime word is masked so that it cannot be identified (Fowler, Wolford, Slade, & Tassinary, 1981). Similarly, naming a picture is faster if it is preceded by a related picture, even if the prime picture was presented at an exposure duration brief enough to preclude identification of the prime (McCauley, Parmelee, Sperber, & Carr, 1980). These effects are consistent with MEM. Furthermore, MEM would expect that these effects are not necessarily restricted to a few milliseconds but should persist over substantial intervals and build up with repetitions. Consistent with this prediction, Wilson (1979) reported that subjects had a greater preference for melodies previously repeated on the unattended ear during a dichotic listening task, compared to new melodies. In addition, Kellogg (1980) makes the important point that the nature of the new items on a test will influence whether subjects show memory for unattended information. New items that are very similar to targets will yield low scores (this is true even for attended information); new items that are less similar may reveal that some characteristics of the unattended information were stored. Kellogg did find significant memory for the class of faces that had been presented as unattended distracting stimuli during a mental arithmetic task. A similar conclusion can be drawn from the previously mentioned study by Rock and Gutman (1981); an appropriate selection of new items (e.g., differing from the target in size) revealed that subjects stored certain characteristics of unattended visual figures. In short, it seems likely that if investigators looked for evidence of memory for nonattended stimuli with more sensitive tests, they would find it. Finally, as another argument against attention-dependent storage, consider complex tasks such as playing a piece on the piano, typing, reciting a 'Eich (1982) has recently reported evidence consistent with the present prediction. Subjects shadowed a passage presented in one ear while two-word phrases (e.g., taxi fare) were presented in the other ear. Later, subjects were not able to recognize words presented in the unattended ear. However, the results of a spelling test indicated that the unattended information had been stored; compared to unpresented control words, subjects were more likely to spell homophones (e.g., fair/fare) in the way consistent with the interpretation that had been biased by context during presentation in the unattended ear (e.g., taxi fare).

98

Marcia K. Johnson

poem, baking bread, or reading. With extended practice, such activities become relatively “automatic” in that they can be done while other “secondary” tasks are performed also (e.g., Spelke, Hirst, & Neisser, 1976). Extended practice beyond the point of “mastery” when performance is already largely “automatic,” that is, overlearning, very likely continues to have effects (e.g., Bahrick, 1982). Theoretically, these effects in memory should show up as continued reduction in response times even after the trial of last error, reduced forgetting over long intervals as measured by savings methods (Ebbinghaus, 1885/1964; Nelson et al., 1979), or as increases in the difficulty of the secondary task that can simultaneously be performed along with the primary task. Any of these findings would constitute evidence that events that do not require attention are nevertheless stored. The question of attention-dependent versus attention-independent storage is related to the more general issue of whether there is a separate “memory mechanism” that is added to perceptual or self-generated experience in order to “store” it. That is, do we perceive and think by some mechanisms and have memory established by others (e.g., a “Now Print” mechanism, Brown & Kulik, 1977)? In MEM,the processes that produce perception and thought result in changes in potential future perception and thought (i.e., produce entries); there is no separate storage mechanism. 2. Priority of Access An important issue for memory theories is specifying conditions affecting the relative availability of entries, and predicting the situations in which we are influenced by some but not other entries. For example, one task for theories of attention is to specify the mechanisms by which some entries recruit attention while others d o not (the problem of “selection”). Models of attention have tended to concentrate on selectivity of response to external stimuli during ongoing perception or comprehension. However, it is equally important to account for selectivity of response to internally cued activation during less perceptual tasks such as reminiscing. In MEM, information from all memory subsystems is presumed to produce activated memory, but these entries are not equal in their ability to recruit attention; this varies with characteristics of the entries, characteristics of the test context, and the state of the subject. Access to attention is greatly affected by the general circumstances during remembering. For example, if you close your eyes and try to remember a specific event, for example, a party that you attended recently, the recollection will very likely be dominated by what you did and thought rather than by what you saw and heard (assuming, of course, that you were an active participant during the party) (Johnson, Raye, Foley, & Foley, 1981;


99

Keenan et al., 1977; Raye, Johnson, & Taylor, 1980; Slamecka & Graf, 1978). Activated information from the reflection system has priority, indicated by the darker shading in Fig. 1. Information from the perceptual system has somewhat less priority, and information from the sensory system is still less able to gain attention, or can only do so under special circumstances. Of course, attention can be recruited by a sufficiently “richyyperceptual record. And, of course, perceptual input at the time of the memory test would also increase the availability of the perceptual record; hence, recognition tests are typically more sensitive than recall tests to physical features of events. Also, increasing the salience of perceptual features by making targets and distractors more similar may improve recognition under some circumstances (Tulving, 198l), as may activating related perceptual information in memory (Malpass & Devine, 1981). However, in the absence of specific perceptual stimuli, as is often the case during recall, our consciousness is dominated by the reflection record. It is not that sensory and perceptual information fades, but rather, according to MEM, that it is inhibited by information from the reflection system. The sensory and perceptual systems are “low-access’ ’ systems to voluntary processes. Under some conditions, the inhibition from the reflection system might be lifted, allowing entries from sensory and perceptual systems to have more influence, and perhaps accounting for sudden vivid recollections (Salaman, 1970), eidetic imagery (Stromeyer & Psotka, 1970), and some aspects of hypnagogic images, dreaming, and hallucinations. The idea that one aspect of cognition might inhibit another has, of course, been suggested by many investigators working on various problems (e.g. , Freud, Hughlings-Jackson, and Pavlov). The general concepts of excitation and inhibition have been basic building blocks in a number of theories of cognitive function. The mechanisms by which attention is recruited are probably related to quantitative changes in patterns of excitation and inhibition (e.g., changes in intensity, or marked dispersion or specificity of activation).

G . STABILITY OF ENTRIES 1.

Time-DependentProcesses

There are two time-dependent processes that have been important in theorizing about memory, decay and consolidation. The multiple-entry model does not assume that memories decay. Aside from brain damage, storage is essentially permanent. Similarly, MEM assumes that beyond the relatively short time it takes

100

Marcia K. Johnson

for chemical processes in the brain to respond to present external and internal events, memories do not further consolidate. Electroconvulsive shock or certain drug treatments have been administered to animals in order to disrupt a hypothetical consolidation process. However, there is some evidence of spontaneous recovery after such treatments and some evidence that “unconsolidated” memories are made more available by reminders (e.g., Spear, 1973). These results are consistent with the view presented here that whatever has been processed has created entries. Furthermore, improvements in performance without further stimulus input that suggest a consolidation process, for example, reminiscence and hypermnesia, would be the byproduct of further processing (e.g., rehearsal or revival) (cf. Roediger & Payne, 1982) that would not necessarily involve attention.

2. Integration Some investigators assume that successive external events that are related create an integrated representation that replaces the individual representations (e.g., Bransford & Franks, 1971; Loftus, 1975; Loftus, Miller, & Burns, 1978). In contrast, the multiple-entry model proposes that both the earlier, individual representations and the integrated, constructed information produced by reflection are entered in memory (Johnson & Raye, 1981). Earlier memory entries are not lost by virtue of having been included in a construction (e.g., Beckerian & Bowers, 1983). This point can be illustrated by considering our everyday use of cognitive maps. Suppose that you learn your way from your hotel to the conference center by a “route map” in a strange town. Later, you find your way to the zoo and imagine a “survey map” that puts the hotel, conference center, and zoo all in relation to each other. If subsequently, someone asks you the way to the drugstore, which is located between your hotel and the conference center, you would not necessarily access your new comprehensive survey map of the area, but might well access your earlier, more restricted, but still sufficient-for-the-task route map. From the present view, research on conditions affecting the relative dominance of perceptually derived and reflectively generated information would be interesting.

3. Mechanisms of Forgetting Assuming that decay and integration are not the major mechanisms of forgetting, what are? There are many potential mechanisms for forgetting in a system with essentially stable entries. For example, as previously discussed, confusion between the perceived and the imagined (failures in real-


101

ity monitoring) can introduce distortion and inaccuracy in remembering (Johnson & Raye, 1981). Interference processes, such as competition and blocking from more dominant memories within subsystems (McGeoch, 1932) or inhibition from activation from other subsystems, could produce omissions and intrusions. Clearly, without loss or degrading of entries, forgetting occurs when the appropriate stimulus conditions are not recreated (McGeoch, 1932; Tulving & Thomson, 1973). In fact, it could be argued that other mechanisms of forgetting are special cases of failure to reinstate appropriate stimuli. For example, consider the case of response competition. Most investigators would agree that responses are associated with functional rather than nominal stimuli. Consequently, apparent competition between two different responses to the “same” nominal stimulus may reflect cue-dependent forgetting (e.g., failure to activate the appropriate interpretation of a stimulus) rather than direct competition between nominal responses (e.g., Hasher & Johnson, 1975; Hashtroudi & Johnson, 1976). The critical point is that entries can be relatively stable and yet the system can yield imperfect memory. Interference may occur at many levels. It may be quite general (as when one system recruits attention at the expense of another), more specific (as when some entries have an advantage over others within the same subsystem), or even more specific (as when one set of entries associated with a stimulus are overshadowed by another set associated with a similar stimulus configuration). Furthermore, in MEM, activation of entries within any particular subsystem depends largely on the degree to which present stimulus conditions match those coded in the memory entries. Finally, not all tests are equally sensitive to information stored in memory (e.g., Bahrick, 1967; Johnson, 1977; Nelson, 1978), nor (as emphasized here) are they equally revealing about characteristics of different functional subsystems.

H. THECONCEPT OF AN “ENTRY”VERSUS THE

CONCEPT OF “THE TRACE”

Memory theorists often talk about “the trace.” For most, this is a shorthand way of referring to whatever unknown physiological changes underlie changes in behavior produced by experience. At this extremely abstract level, there is not much disagreement that traces exist. However, the concept of a memory trace has been often challenged, primarily because traces tend to be thought of as “copies” of external events. Furthermore, the concept of trace tends to be associated with “episodic tasks” rather than “semantic tasks.” The recent surge of interest in knowledge, skill learning, and strategies-the sorts of products of memory that do not appear to be mediated

102

Marcia K. Johnson

by a single episodic experience-has lead some investigators to look for some metaphor to replace the idea of a trace (e.g., Bransford, McCarrell, Franks, & Nitsch, 1977). While there may be general consensus in the field that it is not fruitful to speak of “the” trace, this does not mean that there is not “any” trace. The point here is similar to Pavlov’s (1932) criticism of Lashley. Lashley (1930) reported a series of studies showing that rats found mazes more difficult to learn in proportion to the amount of their cerebral hemispheres that had been destroyed. Lashley used these data to deemphasize functions of specific cortical areas in favor of a principle of mass action. Pavlov found Lashley’s conclusion “original” but “quite inconceivable.” He pointed out that different receptor systems (olfactory, auditory, visual, cutaneous, kinesthetic) are located in different parts of the cerebral hemispheres (along with possible dispersion of elements from each system throughout the entire mass of the hemispheres). If the rat draws on all of these receptors in learning the maze, then destroying any or several of them will hurt performance, and in proportion to the number destroyed. But it does not follow that there are no specific entries anywhere. The multiple-entry model assumes a similar distinction between localization topologically and localization within functional memory subsystem^.^ According to the multiple-entry model, all events are complex and therefore entries are potentially formed in a number of subsystems (sensory, perceptual, reflection). Our subjective experience when we are remembering depends on the particular combination of information from all of these subsystems that reaches activation and, most importantly, attention, at any given moment. The fact that our subjective experience is a blend of these factors (i.e., “cross-modal”) does not mean that different single entries do not exist. Many investigators have emphasized the integrated, holistic qualities of memories. However, we know that the memory trace for an event is not typically unitized in any technical sense; revival of one type of information ‘While sensory, perceptual, and reflection subsystems may not be localized in different places, they might be mediated by different types of structures or processes in the brain. Also, you might expect different species to have characteristically different relative distributions of these structures. Thus, a species that was good at recognition would not necessarily be good at recall and vice versa. It might even be some advantage in accessing sensory and perceptual entries to have relatively little competition from reflection entries (Shettleworth, 1983). However, there is no reason to believe that the selective pressures that produced the evidently “higher” reflection functions would eliminate the structures upon which the sensory and perceptual records and functions depend. In fact, a “multiple-entry” memory would be extremely valuable. Redundant information is, of course, less vulnerable. Furthermore, a malfunction in one system could perhaps be partially compensated for by the continued functioning of other processes and records.


I03

(e.g., semantic) does not necessarily result in revival of other information (e.g., graphemic). That is, all features are not equally good redintegrative cues for the whole event, nor are events remembered “all or none.” Even at the same apparent “level,” not all aspects of a memory are activated at once (Loftus, 1982). Thus, events must create functionally distinct codes (Posner, 1978) or levels (Craik & Lockhart, 1972) or entries (the present model). We should keep in mind the distinction between the phenomenology of memory, where information from various sources is integrated and gestalt-like, and the mechanisms of memory, which may include functionally separate systems. In this regard memory is like perception: if we fell in a pigsty, we would have a unified experience, but we would not seriously mean that there were no differences in the mechanisms of seeing, smelling, touching, and hearing that go into that “holistic” experience. One advantage of thinking in terms of subsystems is that it helps us imagine the independent operation of different aspects of memory. This potentially may help us organize such diverse facts as the independence of recall and recognition (Flexser & Tulving, 1978), improvement of amnesics on motor tasks that they do not remember engaging in previously (e.g., Milner, 1965), and the apparent dissociation between changes in reported fear and changes in responses to feared objects (Lang, 1969). In sum, there is not one trace from a complex event, but multiple “entries.” This is how memory derives its redundancy and flexibility. I have introduced the term “entry” rather than “trace” to avoid the connotation of an exclusive concern with externally derived information. In fact, a major emphasis of the present model is that self-generated information is a pervasive and powerful source of entries.

111. The Multiple-Entry Model and Other Theoretical Issues

A. MEMORIES AS RECORDS OF COMPLETE EXPERIENCES There is a tendency for researchers to characterize memory representations in terms of abstractions such as “information” or “propositions.” The general idea is that memory representations consist of a “deep structure” for events, rather than the events themselves. While useful hypotheses follow from this approach, such abstractions tend to divorce memories from the mechanisms by which they were established. The multiple-entry model adopts the opposite approach emphasized by Johnson and Raye (1981): memory reflects the origin of information. Memory entries are the records of the specific processes that created the entries (Kohlers, 1975). Insofar as different events engage different processes (e.g., seeing a word versus hear-

104

Marcia K. Johnson

ing it versus thinking of it), the memory entries will be different. There is nothing “beyond” the record of these processes that “is” the memory representation. In addition, cognitive models tend to paint a picture of a complex but essentially actionless and emotionless subject. However, memory is not composed simply of “pure” perceptions and thoughts, but includes actions and feelings as well. The simplest events involve movements and feelings (if only eye movements and boredom). Furthermore, actions and feelings are not transient. We are changed (perhaps permanently) as a consequence of their having happened. For example, such things as posture and mood are part of an event and may influence our ability to recollect the rest of the event just as “semantic” cues do (Bower, 1981; Rand & Wapner, 1967). Therefore, an argument could be made that action and feeling cannot simply be grafted onto a finished cognitive model, but should be part of its ongoing development. In MEM it is assumed that all subsystems are involved with the initiation and maintenance of movement and with the experience and “reexperience” of emotion. However, it is further proposed that the various subsystems play different roles in action or in emotion. For example, as mentioned earlier, certain types of stimulus-driven movements may largely be controlled by entries in the sensory system. Movement in response to complex stimulus patterns, for example, a skilled musician reading a piece of music for the first time or a skilled artist drawing “from life,” may primarily involve the perceptual system. Movement that is more generative, that is, it requires new organization and planning (e.g., choreographing a dance), is more likely to involve the reflection system as well (see Fig. 1). Of course, “control” can be exercised by two or three systems simultaneously, or pass from one to another in different phases of learning; however, entries are created and preserved whenever a subsystem has been involved. Similarly, emotion very likely is influenced by or interacts with all subsystems. Some further thoughts on emotion and memory are presented in the next section. B. EMOTION

Zajonc (1980) has suggested that affect does not depend on “semantic” processing. Evidence supporting this idea is that preference judgments can be made more quickly than recognition judgments and that preferences build up with repeated exposure even when the subject does not recognize that the stimulus has occurred before. However, as Lazarus (1982) has pointed out, if a serial stage view of information processing (in which meaning is the end product of a complex sequence of transformations) is rejected, it is easier to see how cognition could be implicated in emotion with the immediacy and lack of awareness that is seemingly required. In terms of the

A Mukiple-Entry. Modular Memory System

LO5

present model, Zajonc’s results suggest that emotion may be differentially associated with the memory subsystems. In fact, many emotional responses do seem strongly associated with information in the sensory and perceptual systems. In these cases, related information from the reflection system typically produces less emotion when it gains attention. Therefore, when we voluntarily recall an event, it often has a muted emotional quality. On the other hand, when we see something that reminds us of the event, such as a person involved, or the room where the event took place, we sometimes feel overcome by what appears to be a recreation of the initial emotional reaction. What would be the functional value of such an arrangement? As Zajonc (1980) points out, “Before we evolved language and our cognitive capacities, which are so deeply dependent on language, it was the affective system alone upon which the organism relied for its adaptation” (p. 170). Clearly, it would be valuable to have an associative system that would initiate fear and running the moment a tiger came into view. At the same time, it would not be very functional if voluntarily thinking of a tiger created the same emotional responses. If you thought of a tiger while you were planting crops you might run off and hide; the crops would fail with obvious dire consequences. One solution would be for evolution to select for an animal that simply did not think of tigers when they were not present; thus, inappropriate actions would not interfere with more functional, ongoing, activities. Another solution would be to select for an animal that could think of tigers without the full emotional responses that are associated with real tigers. The value of this second solution is that it would allow you to think of tigers in their absence in order to plan what to d o when they are there (to build a safe hiding place, to invent a weapon). Such planning is one primary responsibility of the reflection system. Thus, the ability to “dispassionately” contemplate events has its advantages (although people vary greatly in their ability to do so!). However, sometimes it is instructive, or useful therapeutically, to remember our emotional responses to specific events, objects, or people. There are probably more and less effective ways to help someone do this. For example, “Tell me how you felt when you realized you hurt his feelings,” would probably be less successful than getting the person to first recreate the details of the event-“What was the look on his face when you said . . . .” Through revival of sensory-perceptual records, the original emotion often follows relatively automatically. Some emotions, although firmly entrenched in memory, are not easily accessed voluntarily.s ’“Delay of gratification” studies may make a similar point about the importance of sensory/ perceptual stimuli for emotional-motivational responses (Mischel. 1981). The above discussion

106

Marcia K. Johnson

On the other hand, certain emotions (e.g., some types of anxiety or depression) do not seem to fit the above description. They seem less affected by the presence or absence of specific, external stimuli, and more dependent on the nature of reflective activity. It is not that some emotions (e.g., fear) are inherently “in” the sensory system and others (e.g., depression) “in” the reflective system. For example, depression and anxiety can be elicited by actual dreadful circumstances as well as from our misconstrual of (and “compulsing over”) objectively favorable circumstances (Sampson, 1981). However, thinking in terms of the MEM model suggests that methods of accessing and/or influencing feelings may depend on the primary entry system to which they are attached (cf. Lang, 1969).

c.

OTHER TYPOLOGIES OF MEMORY

The memory system has been divided up in a number of ways in the last 20 years: into sensory buffers, short-term memory (STM), and LTM (Atkinson & Shiffrin, 1968); by levels of processing (Craik & Lockhart, 1972), or attributes (Bower, 1967; Underwood, 1969; Wickens, 1970); in terms of visual and verbal modes of representation (Paivio, 1971); into semantic and episodic memories (Tulving, 1972); and into the results of controlled and automatic processes (Shiffrin & Schneider, 1977).

The multiple-entry model explicitly differs from some of these views, for example, by denying the assumption of differential permanence of different types of information in Atkinson and Shiffrin (1968), Craik and Lockhart (1972), and Shiffrin and Schneider (1977). With respect to other ways of characterizing the memory system, MEM is more orthogonal than contradictory. For example, the dual code hypothesis proposes that there are functionally separate verbal and visual memory systems (Paivio, 1971). The multiple-entry model emphasizes a different organization of events. That is, words can be generated internally or perceived from external sources. In both cases, they will have linguistic properties. Similarly, pictures may be perceived or they may be generated (imaged), and in both cases they will have pictorial properties, for example, spatial characteristics (Kosslyn, 1980). However, according to MEM, perceived words and perceived pictures have something in common as well, their external source and dependence on perceptual processing for establishing representational

Footnote 5 (conlinued) is related to the general issue of what makes a cognition “hot” (Abelson, 1963). Apparently it is concreteness or specificity, at least in part. That is, a person is not so much angered by abstractions such as “communism,” as at their bohemian son who lives in a dusty house with a lot of other people.


107

entries. Likewise, generated words and generated pictures have something in common, their occurrence in the absence of perceptual stimuli and their control via reflection mechanisms. Thus, depending on the theoretical problem being addressed, at one time the visual qualities of an entry might be important, and at another time its perceptual origin might be important. To fully specify a “complete memory,” many aspects would have to be taken into account (e.g., Underwood, 1969). However, it does seem possible to make progress in understanding memory by pursuing the implications of each of several different ways of characterizing the memory system. The multiple-entry model is also orthogonal to the distinction between semantic and episodic memory (Tulving, 1972). However, how MEM and the semantic-episodic distinction might fit together will be considered in the next two sections to help clarify additional ideas embodied in the present model. Finally, MEM is particularly compatible with recent work emphasizing that memory should not be equated with “awareness” (e.g., Jacoby & Witherspoon, 1982), nor with “effort” (Hasher & Zacks, 1979), and with approaches that characterize memory in terms of functional systems rather than in terms of global capacities (Spear, 1983).

1.

Generic Memory and the Problem of Abstraction

There are at least two possible meanings of “semantic.” First, the term semantic may denote having to do with word meanings. Second, the term semantic is sometimes used to refer to our knowledge of other types of meaning relations as well. In the multiple-entry model, meaning of this second sort is distributed throughout all subsystems, and is not exclusively associated with one (e.g., a conditioned emotional response is meaningful). As Hintzman (1978) points out, “generic” memory is probably a better name for what most people (including Tulving, 1972) mean by “semantic” memory, because most do not limit it to word meanings but include other sorts of general knowledge (e.g., days are shorter in winter, strategies for solving problems, etc.). In the present framework, generic memories are created by sensory, perceptual, and reflection functions. The knowledge represented by generic memories would not all be available to introspection. Generic memories consist of well-learned, or readily inferred information; by definition, they are summations across two or more episodes. They include the rules, schemata, prototypes, and modal or averaged information that create “classes” or “categories” of events. Perhaps, as Wickelgren (1981) has suggested, all memories are on a continuum of generic memory; others have made a similar suggestion that generic and episodic memories represent different types

108

Marcia K. Johnson

of knowledge rather than separate memory systems (McCloskey & Santee, 1981). A major problem in understanding generic memory is the process of abstraction (Brooks, 1978; Herrnstein & devilliers, 1981; Mervis & Rosch, 1981). How are concepts and schemata created? In the present framework,

it seems reasonable to assume that abstraction results from both sensoryperceptual and reflective processes. Some abstraction will be the by-product of perceptual experience, based, for example, on frequency of occurrence, stimulus generalization, and the overlap of activation of features or relations upon repeated experience with similar events. Such schemata are created by processes that are more automatic than consciously constructive. We might call this “perceptual abstraction.” In contrast, other schemata are created by processes that are more consciously constructive. These abstractions will be the product of reflection, the self-generative processes that involve active search, comparison, analysis, and criticism (“reflective abstraction”). From the present point of view, neither type of abstraction replaces the specific, episodic memories on which they are based, but either might typically be more accessible than any specific episodic memory. What is particularly interesting from the present perspective is the possibility that abstractions created perceptually and those created reflectively might have characteristically different properties, and might be activated by different types of cues. Such a difference might help explain why consciously constructed and maintained views or schemata (e.g., theories) often seem resistant to counterexamples, and, conversely, why unconsciously abstracted generalized reactions (e.g., fear of furry animals) often seem resistant to argumentation. The optimum strategy for changing a schema might depend on whether it is largely perceptually, or largely reflectively, based. 2. Episodic versus Personal Memory

In MEM, episodic memories are also created by all subsystems. Episodic memories are the consequences of unique experiences; each is distinctive by virtue of the complex activation pattern created when it was established (Hintzman, 1976). As Tulving pointed out, part of what makes a memory seem episodic rather than semantic is that we can remember the time and place the event occurred (e.g., the word “table” occurred on List A in the cognition laboratory a few minutes ago). These sorts of contextual features of episodic memories have sometimes been treated as the defining features of episodic memories and are at least emphasized by most investigators. However, a memory may seem episodic even if not well localized in time and place, for example, when a face in a crowd seems suddenly familiar,


109

the familiarity may arise from the distinctive features of the face, not from the identification of the time and place where it was previously experienced. Similarly, you might suddenly be reminded of a dream that you had sometime in the past, and have a vivid sense of some highly particular image, and yet have no idea when and where the dream occurred. (In fact, you may wonder if the image did not come from a painting rather than a dream, after all.) These are instances of what it seems reasonable to call episodic memories; the sense that they are derived from particular episodes gives us the feeling that we should be able to recall when and where they were previously experienced. However, their familiarity and unique, “episodic” character is not dependent on such contextual information. (Dkja vu-the sense of having experienced something before-is an instance of memory that seems specifically “episodic” while time and place information escape us.) What makes a memory “personal” rather than simply episodic? Here the identification of time and place clearly play an important role. However, temporal and spatial information comprise only a subset of a more general class of associated information that may personalize a memory. According to MEM, this more general class of information is created by reflection activities. These are activities that go beyond immediate perceptual processes. They serve to identify relations among events from the past and the anticipated future of an individual. A particularly important function of the reflection system is that it helps create and maintain personal identity or a sense of “self.” For example, creating plans, initiating action, and evaluating events in relation to plans are the sorts of activities that differentiate the self from the outside world. They define the self as a locus of power or energy, as a constant in time and space, as an object in relation to other objects. Consider two different situations. In Situation A, you anticipate getting hungry, look for food, prepare it, become hungry, and eat the food. In Situation B, food magically appears on a random schedule often enough so that you are physically satisfied most of the time. In Situation A, you become an agent acting in and on the world, with some degree of control over your destiny. In Situation B, the food is happening to you-it is externally derived and you play no role in bringing it into existence. Anticipation is not necessary; feedback loops between acts and consequences are not established (Miller, Galanter, & Pribram, 1960). The ability to obtain food, and whatever specific skills and general competence that implies, does not become part of your “selfschema” (Markus, 1977). In addition to defining the self in relation to possible acts and consequences, plans serve another important function-they order events. Suppose that you had to determine the relative order of getting hungry and

Marcia K. Johnson

110

obtaining food in the two situations above. It would be much easier to do this in the case where the two events occurred in the context of a planned sequence than in the case where they did not. A similar case could be made for other reflective activities such as searching, comparing, criticizing, and plotting. They relate present events to the past and future. These reflective processes create a sense of continuity of experience by explicitly bridging gaps between distinct episodes. What would happen if these reflective functions did not take place? Episodic memories would still be established in sensory and perceptual subsystems. Similarly, generic memories derived from these episodes via perceptual abstraction would continue to be established. However, the ability to voluntarily recall would be severely disrupted because recall depends greatly on reflection records. Those episodic memories that were remembered, for example, via recognition, while perhaps seeming distinctive or unique, would be difficult to localize in time and, especially, would not seem to have specific implications for the “self.” They would not have been related to or embedded in a particular past or future. A pattern very much like this appears in some cases of clinical amnesia, the topic of the next section.

D. AMNESIAS I.

Clinical Amnesias

Differences in terminology, style of reporting, and the like make it difficult to fully integrate work on clinical amnesia and experimental work on memory (Schachter & Tulving, 1982a). However, a reasonable minimal requirement of a general memory model is that it not flagrantly contradict available clinical evidence. More desirable still, it would suggest particular ways of looking at memory dysfunction and help sharpen the discussion of theoretical issues. The present model can satisfy, I think, both of these criteria. Amnesia may be produced various ways-a blow to the head, surgically induced lesions, brain damage associated with prolonged excessive drinking (Korsakoff’s syndrome), and functional amnesias precipitated by traumatic personal events. Amnesias are commonly divided into two major types, retrograde and anterograde. Retrograde amnesia refers to forgetting events prior to the onset of amnesia; anterograde refers to disruption of memory for events that occur after the onset of amnesia. While it seems reasonable to suppose that both types of amnesia might be explained with a common mechanism (Wickelgren, 1979), there does seem to be some evidence that the severity of retrograde and anterograde amnesia are not necessarily cor-


111

related, and these two types of amnesia may sometimes reflect different mechanisms (Hirst, 1982; Moscovitch, 1982). People with anterograde amnesia often have normal immediate memory spans: they can keep a reasonable number of unrelated items in mind as long as they can keep rehearsing them. However, after a short distraction, they may not be able to recall any of the information, or in fact, that they met the experimenter previously or engaged in a memory task. This pattern of performance seemed at one point to implicate a failure to transfer information from STM to LTM (Atkinson & Shiffrin, 1968). The idea that transfer from STM to LTM was disrupted was what might be called a “strong encoding” hypothesis. It tacitly assumed that information is a homogeneous commodity flowing through the system. A block should therefore eliminate any memory in such a system. However, evidence began to accumulate indicating that something of the experience was stored after all. For example, cues, either in the form of partially degraded stimuli or the first three letters of words, improve performance (Warrington & Weiskrantz, 1970). Furthermore, amnesics have shown improvement in tasks such as rotary pursuit, reading mirror images of words, identifying the objects in degraded pictures, and solving number sequences according to an addition rule (e.g., see Baddeley, 1982). These results are hard to interpret if one imagines a disruption in a flow of homogeneous information from short-term memory to long-term memory. Savings in these various tasks indicate that memories are created, even though subjects do not appear specifically to recall the sessions during which the memories were created. A major alternative to the disruption of transfer theory is the retrieval deficit theory. Warrington and Weiskrantz (1970), for example, suggested that amnesics suffer during retrieval from increased proactive interference. The problems with the encoding versus retrieval distinction have been discussed by Tulving (1979; see also Kinsbourne & Wood, 1982). Basically, the point is that neither encoding nor retrieval can independently be assessed without the other. The problem with trying to separate encoding from retrieval effects can be seen clearly in the present framework. Suppose a patient with Korsakoff’s syndrome cannot recall but can recognize an event. This is not sufficient evidence to conclude that “the” memory is there and the problem is “merely” one of retrieval. The event may not have been encoded in such a fashion as to permit recall even in a normal person. That is, it may not have been entered by the reflection system. As long as it is reasonable to suppose that different memory tests draw differentially on different subsystems, the information that a particular test gives us about memory is specific to the subsystem that it draws upon.

112

Marcia K. Johnson

A Reflection System Deficit. A kind of “encoding deficit” position makes sense if we consider the operations needed to engage a subsystem in the first place. That is, we can meaningfully speculate about deficits of encoding within a system if systems are defined in terms of these operations. In this context, it might, for example, be reasonable to propose that Korsakoff patients have relatively intact sensory and perceptual subsystems, but have an encoding deficit that is primarily associated with the reflection subsystem. Consistent with this idea is the fact that people with Korsakoff’s syndrome do not seem to be able to think of and describe strategies for remembering (Hirst & Volpe, cited in Hirst, 1982). When amnesics are required by an orienting task to respond semantically to target items, their performance is improved; but they do not benefit more than control subjects d o (Cermak & Butters, 1973). From the present point of view, it is not surprising that inducing item-by-item semantic processing with an orienting task does not eliminate the difference between people with Korsakoff’s syndrome and normal people. Normal people would be expected to engage in additional reflective activities that build relations among items or between items and other potential recall or recognition cues. That is, it would be difficult to fully equate the reflective activities engaged in by normal people and amnesics with an orienting task. Proactive inhibition can also be viewed as an encoding deficit, that is, as a consequence of poor elaborators (Hasher & Johnson, 1975; Keppel, 1968; Postman & Underwood, 1973). There is some evidence that the quality of elaborators generated by subjects does decline under conditions of interference (Hasher & Johnson, 1975). Presumably, the search for effective elaborators that will withstand increasing interference from other items is an active, reflective process. Thus increased proactive inhibition in Korsakoff’s patients would be expected. In sum, the proposition here is that Korsakoff amnesics and other amnesics do not spontaneously engage in a range of reflection functions that create a reflection record like the one to which normals have voluntary access. A number of other recent suggestions for how to characterize the deficit that produces amnesia similarly converge on initial processing. For example, it has been characterized as a deficit in semantic processing (Cermak, 1977), a deficit in “strategic processing” (Crowder, 1982), a deficit in episodic processing (Kinsbourne & Wood, 1982), a deficit in distinctive encodings (Jacoby, 1982), a deficit in “vertical processes” (Wickelgren, 1979), and a deficit in initial learning (Huppert & Piercy, 1982). The localization of the deficit in the reflection subsystem would be consistent with many of these suggestions. The same deficit that did not engage reflection functions at acquisition


113

would not engage them on a memory test. Hence, even assuming that reflection could be induced at acquisition, a deficit might still be seen unless similar operations were induced at recall as well. That is, it is reasonable to suppose that within subsystems, some kind of encoding specificity holds (Tulving & Thomson, 1973). To reactivate an entry, you need a sufficient match between processes at acquisition and processes at test. While a consensus seems to be developing that amnesics have particular problems with the kind of active processing that would be characteristic of the reflection system, in a recent review of the amnesic literature, Hirst (1982) seems to come to a quite different conclusion. Hirst places great weight on data suggesting that amnesics are particularly poor at judging the relative temporal order of events. Following Hasher and Zack’s (1979) suggestion that temporal information is encoded automatically, Hirst proposes that automatic coding is disrupted in amnesics. However, suppose that a substantial amount of temporal information is not “automatically” encoded in the sense that it is a fixed trace property. Suppose instead that temporal judgments depend also on memories created by earlier reflections. For example, temporal order between A and B would be specified if you remembered thinking of A when you saw B (when you saw the movie “Raiders of the Lost Ark,” you thought that you liked it better than “Close Encounters”) (cf. Hintzman, Summers, & Block, 1975; Tzeng, Lee, & Wetzel, 1979; Wickelgren, 1977). If you do not engage in that kind of comparative thinking, you will not later have available some potentially powerful temporal dating cues. Thus, I would draw just the opposite conclusion from that of Hirst: relatively automatic encoding (as evidenced by skill learning and perceptual learning) is more or less intact, but reflection functions (particularly strategic activities) are disrupted. The present argument depends, of course, on the idea that temporal judgments are often “derivative” from past or current reflection functions, and are not automatically encoded, simple properties of memory entries. Why might someone cease making the sorts of connections illustrated by the movie example? There are several possibilities: (1) the associations that cause one episode (or idea) to activate another, or two episodes to be simultaneously activated by the same stimulus, may no longer exist; (2) the activation of related information may occur but not recruit attention; (3) attention may be recruited by related information, but additional reflective processes may not take place (e.g., elaboration or explicit comparison). Thus, the person may not try to figure out why one thing brought another to mind. The second and third possibilities seem more plausible as explanations of amnesia than the first. However, the exact nature of the disruption still remains to be specified. For example, in (2), disruption in attention re-

114

Marcia K. Johnson

cruitment could be produced by unusually high background levels of activation, making it harder to detect any changes in activation (i.e., a WeberFechner function). Or if overall activity in the entire system were low, changes in activation patterns would generally fall below the threshold values necessary to engage attention, except perhaps for very well-learned information. Either of these possibilities would functionally disconnect attention from the sort of meaningful associations that go beyond the present stimulus, especially during remembering. On the other hand, disruption of reflective functions (3) without a deficit in attention recruitment may signal a malfunction of reflective functions specifically responsible for the covert manipulation of ideas. The type of amnesia would depend on the specific nature of the disruption: (2) would produce both retrograde and anterograde amnesia, while (3) alone would produce only anterograde amnesia. If a breakdown in reflective activities developed gradually (as is very likely the case in patients with Korsakoff’s syndrome), the result would be an “apparent” temporal gradient of retrograde amnesia. That is, increasing failure over a period of years to engage in reflective processing would produce a temporal gradient in recall that would appear to be a consequence of greater forgetting of more recent compared to more remote events (see also Butters & Albert, 1982). In any case, (l), (2), or (3), comparisons between present and past events would be unlikely. These sorts of comparisons normally embed an event in time and interweave it within the personal past. Without reflective activity, the event would not be recalled, and perhaps not recognized. If it did seem familiar later, it would not seem to have any connection with the self, or in Schacter and Tulving’s (1982b) phrase, the memory would be “free floating.” 2. Childhood Amnesia

A good deal of recent work in developmental psychology indicates that children are less likely than adults to engage in reflection functions in memory tasks (e.g., Brown, 1975; Kail, 1979). Or perhaps they engage in different reflection functions. In any event, in tasks that do not appear to depend so heavily on reflection, such as recognition or frequency judgments, children do quite well (Brown, 1975; Hasher & Zacks, 1979). In the present framework, because children do not set up the conditions for later voluntary retrieval, memories from childhood are relatively inaccessible (“childhood amnesia”). However, when these memories are activated by appropriate stimulus conditions (often specific physical stimuli such as Proust’s madeleine), the memory may be intensely sensory (Salaman, 1970).


115

According to MEM, this intensity is because sensory aspects do not have as much competition in attention from associated reflection entries. The greater salience of sensory features in children’s memories gives the impression that children’s memories are “unschematized,” while adults are “schematized” (e.g., Allport & Postman, 1947; Schachtel, 1947). From the present perspective, it would be more likely that in the adult the relatively “unschematized” entries from sensory and perceptual functions are suppressed, rather than nonexistent.

IV.

Summary

This article presents a general framework for integrating memory research from a number of areas. The model is called “a multiple-entry, modular memory system” (MEM). It characterizes long-term memory as three distinguishable, though clearly interacting, subsystems: the sensory record, the perceptual record, and the reflection record. The phenomenal experience of remembering is created by the particular blend of information from these subsystems that reaches awareness. The multiple-entry model differs from certain other cognitive models in several ways: memory is not thought of as the last stage of a serial sequence of transformations but rather as the product of each of several processes; storage in memory does not depend on attention; and entries are permanent, including those that would be characterized in some models as “early stage” or “shallow,” and therefore transient. Not all tasks equally reveal what has been entered into the subsystems comprising memory. This helps to clarify certain controversies (e.g., between naive realists and naive constructivists) and helps to explain the independence among certain memory tasks, for example, recall, recognition, and perceptual threshold. I also suggested that the idea of partially modular subsystems is consistent with certain clinical findings, for example, agnosias and amnesias. I further proposed that a distinction should be made between episodic and personal memory: episodic memories (like generic memories) are created by all subsystems, whereas personal memory is largely created by the reflection system. A malfunction in the reflection system would tend to result in a disconnection between memories and personal identity such as is found in amnesics. In addition a number of additional proposals were made: I suggested that generic memories, for example, abstractions and schemata, differ in the relative roles that perceptual versus reflective processes played in their development. Similarly, emotions differ in the extent to which they seem

1 I6

March K. Johnson

tied to perceptual events versus reflective events. Following this line of thought, I speculated that the most effective way to access or change schemas or emotions would depend on which subsystem was most involved. ACKNOWLEDGMENTS Preparation of this article was supported by grants from the National Science Foundation (BNS 79-26145) and the National Institutes of Mental Health (MH 37222-01). I would like to thank the colleagues and students who read an earlier draft of the manuscript and provided comments and/or encouragement. I would especially like to acknowledge the helpful suggestions and criticisms of Evan Chua-Yap, Marvin Goldfried. Lynn Hasher, Shahin Hashtroudi. Julian Hochberg, Robert Liebert, Carol Raye, and Rose Zacks.

REFERENCES Abelson, R. P. Computer simulation of “hot cognitions.” In S. Tomkins & S. Mesick (Eds.), Computer simulation of personality. New York: Wiley, 1%3. Alba, J. W. & Hasher, L. Is memory schematic? Psychological Bulletin, 1983,93,203-231. Allport, 0 . W., & Postman, L. The psychology of rumor. New York: Holt, 1947. Atkinson, R. C., & Shiffrin, R. M. Human memory: A proposed system and its control processes. In K. W. Spence & J . T. Spence (Eds.), Thepsychology of learning and motivation: Advances in research and theory (Vol. 2). New York: Academic Press, 1968. Baddeley, A. Amnesia: A minimal model and an interpretation. In L.S. Cermak (Ed.), Human memory ond amnesia. Hillsdale, New Jersey: Erlbaum, 1982. Bahrick. H. P. Relearning and the measurement of retention. Journal of Verbal Learning and Verbal Behavior, 1967,6 , 89-94. Bahrick. H. P. Maintenance of long-term ecological memory-methods and findings. Paper presented at the annual meeting of the American Psychological Association, August 1982. Bartlett, F. C. Remembering: A study in experimental and social psychology. London and New York: Cambridge Univ. Press, 1932. Bates, E.. Masling, M. & Kintsch, W. Recognition memory for aspects of dialog. Journal of Experimental Psychology: Human Learning and Memory, 1978,4, 187-197. Bekerian. D. A., & Bowers, J. M. Eyewitness testimony: Were we mislead? Journal of Experimental Psychology: Learning, Memory, and Cognition, 1983,9, 139-145. Bellezza, F. S., Cheesman, F. L., 11, & Reddy, B. G. Organization and Semantic Elaboration in Free Recall. Journal of Experimental Psychologv: Human Learning and Memory, 1977, 3, 539-550. Biederman, I., Rabinowitz. J., Glass, A.. & Stacy, E. On the information extracted from a glance at a scene. Journal of Experimental Psychology, 1974, 103, 597-600. Bower, G. H.A multi-component theory of the memory trace. In K. W. Spence & J. T. Spence (Eds.), The psychology of learning and motivation (Vol. I). New York: Academic Press, 1967. Bower, G. H. A selective review of organizational factors in memory. In E. Tulving & W. Donaldson (Eds.), Organization of memory. New York: Academic Press, 1972. Bower, G. H. Mood and memory. American Psychologist, 1981,36, 129-148. Bower, G.H..Black, J. B., & Turner, T. J. Scripts in memory for text. CognitivePsychology, 1979, 11, 177-220.


I I7

Bransford, J. D., Barclay, R., & Franks, J. Sentence memory: A constructive versus interpretative approach. Cognitive Psychology, 1972,3, 193-209. Bransford, J. D., & Franks, J. J. The abstraction of linguistic ideas. Cognitive Psychology. 1971,2, 331-350. Bransford, J. D., & Johnson, M.K. Considerations of some problems of comprehension. In W. Chase (Ed.), Visual information processing. New York: Academic Press, 1973. Bransford, J. D., McCarrell, N. S., Franks, J . J., & Nitsch, K. E. Toward unexplaining memory. In R. Shaw & J. Bransford (Eds.), Perceiving. acting. and knowing. Hillsdale, New Jersey: Erlbaum. 1977. Broadbent, D. E. Perception and communication. Oxford: Pergamon, 1958. Broadbent, D. E. Decision and stress. New York: Academic Press, 1971. Brooks, L. Nonanalytic concept formation and memory for instances. In E. Rosch & B. B. Lloyd (Eds.), Cognition and categorization. Hillsdale, New Jersey: Erlbaum, 1978. Brown, A. L. The development of memory: Knowing, knowing about knowing, and knowing how to know. In H.W. Reese (Ed.), Advances in child development and behavior (Vol. 10). New York: Academic Press, 1975. Brown, R., & Kulik, J. Flashbulb memories. Cognition, 1977,5 , 73-99. Brunswik, E. Perception and the representative design of psychological experiments (2nd ed.). Berkeley, California: Univ. of California Press, 1956. Butters, N., & Albert, M. S. Processes underlying failures to recall remote events. In L. S . Cermak (Ed.), Human memory and amnesia. Hillsdale, New Jersey: Erlbaum, 1982. Cerella, J. Visual classes and natural categories in the pigeon. Journal of Experimental Psychology: Human Perception and Performance, 1979,5, 68-77. Cermak, L. S. The contribution of a “processing” deficit to alcoholic Korsakoff patients’ memory disorder. In 1. M. Birnbaum & E. S. Parker (Eds.), Alcoholand human memory. Hillsdale, New Jersey: Erlbaum, 1977. Cermak, L. S., & Butters, N. Information processing deficits of alcoholic Korsakoff patients. Quarterly Journal of Studies on Alcohol. 1973,34, 1110-1 132. Chase, W. G., & Simon, H. A. The mind’s eye in chess. In W. G. Chase (Ed.), Visual information processing. New York: Academic Press, 1973. Craik, F. I. M., & Kirsner, K. The effect of speaker’s voice on word recognition. Quarterly Journal of Experimental Psychology, 1974, 26, 274-284. Craik. F. I. M., & Lockhart, R. S. Levels of processing: A framework for memory research. Journal of Verbal Learning and Verbal Behavior, 1972, 11, 671-684. Craik, F. I. M., & Tulving, E. Depth of processing and the retention of words in episodic memory. Journal of Experimental Psychology: General, 1975, 104,268-294. Craik, F. 1. M., & Watkins, M. J. The role of rehearsal in short-term memory. Journal of Verbal Learning and Verbal Behavior. 1973, 12, 599-607. Crowder, R. G. General forgetting theory and the locus of amnesia. In L. S. Cermak (Ed.), Human memory and amnesia. Hillsdale. New Jersey: Erlbaum, 1982. Dorfman, D. Recognition memory for the surface form of sentences in connected discourse. Paper presented at the 50th annual meeting of the Eastern Psychological Association, Philadelphia, April, 1979. Durso, F. T., & Johnson, M. K. Facilitation in naming and categorizing repeated pictures and words. Journal of Experimental Psychology: Human Learning and Memory, 1979, 5 , 449-459. Ebbinghaus, H. Memory: A contribution to experimentalpsychology. New York: Dover, 1964. (Originally published 1885.) Eich, E. Memory for vs. awareness of unconscious events. Paper presented at the annual meeting of the Psychonomic Society, Minneapolis, Minnesota, November, 1982. Erdelyi, M. A new look at the new look. Psychological Review, 1974, 81, 1-25.

118

Marcia K. Johnson

Flexser, A. J., & Tulving, E. Retrieval independence in recognition and recall. Psychological Review, 1978, 85, 153-171. Foley, M. A., & Johnson, M. K. The consequencesfor memory of imagining in another person’s voice. Paper presented at the annual meeting of the Eastern Psychological Association, April, 1982. Forbach, G. B.. Stanners, R. F., & Hochhaus, L. Repetition and practice effects in a lexical decision task. Memory & Cognition, 1974, 2, 337-339. Fowler, C. A., Wolford, G., Slade, R., & Tassinary, L. Lexical access with and without awareness. Journal of Experimental Psychology: General, 1981, 110, 341-362. Friedman, A. Framing pictures: The role of knowledge in automatized encoding and memory for gist. Journal of Experimental Psychology: General, 1979, 108, 316-355. Geiselman, R. E., & Bjork, R. A. Primary versus secondary rehearsal in imagined voices: Differential effects on recognition. Cognitive Psychology, 1980, 12, 188-205. Gibson, J. J. The senses considered as perceptual systems. Boston, Massachusetts: Houghton, 1966. Glenberg, A., Smith, S. M., & Green, C. Type I rehearsal: Maintenance and more. Journal of Verbal Learning and Verbal Behavior, 1977, 16, 339-352. Goldmeier, E. The memory trace: Its formation and its fate. Hillsdale, New Jersey: Erlbaum, 1982. Haber. R. N.. & Hershenson. M. Effects of repeated brief exposures on the growth of a percept. Journal of Experimental Psychology, 1965, 69, 40-46. Hanley, G. L., & Levine, M. Spatial problem solving: The integration of independently learned cognitive maps. Memory & Cognition, in press. Hasher. L.,& Griffin, M. Reconstructive and reproductive processes in memory. Journal of Experimental Psychology, 1978, 4, 318-330. Hasher, L., & Johnson, M. K. Interpretive factors in forgetting. Journal of Experimental Psychology: Human Learning and Memory. 1975, 1, 567-575. Hasher, L.. & Zacks, R. T. Automatic and effortful processes ih memory. Journal of Experimental Psychology: General, 1979, 108, 356-388. Hashtroudi, S., & Johnson, M. K. Transfer and forgetting: Interpretive shifts and stimulus reinstatement. Journal of Experimental Psychology: Human Learning and Memory, 1976, 2, 262-272. Herrnstein, R. J., & de Villiers, P. A. Fish as a natural category for people and pigeons. In G. H. Bower (Ed.), The psychology of learning and motivation: Advances in research theory (Vol. 14). New York: Academic Press, 1980. Hintzman, D. L. Repetition and memory. In G. H. Bower (Ed.), Thepsychology of learning and motivation (Vol. 10). New York: Academic Press, 1976. Hintzman. D. L. The psychology of learning and memory. San Francisco, California: Freeman, 1978. Hintzman. D. L.. & Summers, J. J. Long-term visual traces of visually-presented words. Bulletin of the Psychonomic Society, 1973, l, 325-327. Hintzman, D. L., Summers, J. J., & Block, R. A. Spacing judgments as an index of studyphase retrieval. Journal of Experimental Psychology: Human Learning and Memory, 1975, 1, 31-40. Hirst, W. The amnesic syndrome: Descriptions and explanations. Psychological Bulletin, 1982, 91, 435-460. Hochberg, J. Perception 1. Color and shape. In J. W. Kling & L. A. Riggs (Eds.), Woodworth & Schlosberg’s Experimental psychology (3rd ed.). New York: Holt, 1971. Hochberg, J. On cognition in perception: Perceptual coupling and unconscious inference. Cognition. 1981, 10, 127-134.


I I9

Huppert, F. A., & Piercy, M. In search of the functional locus of amnesic syndromes. In L. S . Cermak (Ed.), Human memory and amnesia. Hillsdale, New Jersey: Erlbaum, 1982. Jacoby, L. L. On interpreting the effects of repetition: .Solving a problem versus remembering a solution. Journal of Verbal Learning and Verbal Behavior, 1978, 17,649-667. Jacoby, L. L. Knowing and remembering: Some parallels in the behavior of Korsakoff patients and normals. In L. S . Cermak (Ed.), Human memory and amnesia. Hillsdale. New Jersey: Erlbaum, 1982. Jacoby, L. L., & Dallas, M. On the relationship between autobiographical memory and perceptual learning. Journal of Experimental Psychology: General, 1981, 110, 306-340. Jacoby, L. L., & Witherspoon, D. Remembering without awareness. Canadian Journal of Psychology, 1982,36, 300-324. Johnson, M. K. What is being counted none the less? In I. M. Birnbaum & E. S. Parker (Eds.), Alcohol and human memory. Hillsdale, New Jersey: Erlbaum, 1977. Johnson, M. K., Bransford, J. D., & Solomon, S. K. Memory for tacit implications in sentences. Journal of Experimental Psychology, 1973,98, 203-205. Johnson, M. K., Kahan, T. L.,& Raye, C. L. Dreamsand reality monitoring. Paper presented at the annual meeting of the Psychonomic Society, November, 1981. Johnson, M. K., & Raye, C. L. Reality monitoring. Psychological Review, 1981, 88, 67-85. Johnson, M. K., Raye, C. L., Foley, H. J.. & Foley, M.A. Cognitive operations and decision bias in reality monitoring. American Journal of Psychology, 1981,94, 37-64. Johnson, M. K., Raye, C. L., Foley, M. A., & Kim, J. Pictures and images: Spatial and temporal information compared. Bulletin of the Psychonornic Society, 1982, 19, 23-26. Johnson, M. K. , Raye, C. L., Wang, A. Y., & Taylor, T. H. Fact and fantasy: The roles of accuracy and variability in confusing imaginations with perceptual experiences. Journal of Experimental Psychology: Human Learning and Memory, 1979,5 , 229-240. Julesz, B. Foundations of cyclopean perception. Chicago, Illinois: Univ. of Chicago Press, 1971. Kahneman, D. Attention and effort. New York: Prentice-Hall. 1973. Kail, R. The development of memory in children. San Francisco: Freeman, 1979. Keenan, J. M., MacWhinney, B., & Mayhew, D. Pragmatics in memory: A study of natural conversation. Journal of Verbal Learning and Verbal Behavior, 1977, 16, 549-560. Kellogg, R. T. Is conscious attention necessary for long-term storage? Journal of Experimental Psychology: Human Learning and Memory, 1980,6 , 379-390. Keppel, G.Retroactive and proactive inhibition. In T. R. Dixon & D. L. Horton (Eds.), Verbal behavior and general behavior rheory. New York: Prentice-Hall, 1968. Kinsbourne, M., & Wood, F. Theoretical considerations regarding the episodic-semantic memory distinction. In L. S . Cermak (Ed.), Human memory and amnesia. Hillsdale, New Jersey: Erlbaum, 1982. Kintsch, W. The representution of meaning in memory. Hillsdale, New Jersey: Erlbaum, 1974. Kintsch, W., & Bates, E. Recognition memory for statements from a classroom lecture. Journal of Experimental Psychology: Human Learning and Memory, 1977,3. 150-168. Kirsner, K., & Smith, M. C. Modality effects in word identification. Memory & Cognition, 1974,2, 637-640. Kolb, B., & Whishaw, I. Q.Fundamentals of human neuropsychology. San Francisco, California: Freeman, 1980. Kohlers, P. A. Specificity of operations in sentence recognition. Cognitive Psychology, 1975, 7, 289-306. Kosslyn, S. M. Image and mind. Cambridge, Massachusetts: Harvard Univ. Press, 1980. Kowler, E.,& Martins, A. J. Eye movements of preschool children. Science, 1982,215, 997999.

I20

Marcia K. Johnson

Kunst-Wilson, W. R., & Zajonc, R. B. Affective discrimination of stimuli that cannot be recognized. Science, 1980, 207, 557-558. Lang, P. J. The mechanics of desensitization and the laboratory study of fear. In C. M. Franks (Ed.), Behavior therapy: Appraisal and status. New York: McGraw-Hill, 1969. Lashley, K. S. Basic neural mechanisms in behaviour. Psychological Review., 1930,37,1-24. Lazarus, R. S. Thoughts on the relations between emotion and cognition. American Psychologist, 1982,37, 1019-1024. Levine, M. Hypothesis behavior by humans during discrimination learning. Journal of Experimental Psychology, 1966. 71, 331-336. Levine, M., Jankovic, 1. N.. & Palij, M. Principles of spatial problem solving. Journal of Experimental Psychology: General, 1982, 111, 157-175. Light, L. L., Kayra-Stuart, F., & Hollander, S. Recognition memory for typical and unusual faces. Journal of Experimental Psychology: Human Learning and Memory, 1979, 5, 212-228. Lindsay, P. H., & Norman, D. A. An introduction to psychology (2nd ed.). New York: Academic Press, 1977. Loftus, E. F. Leading questions and the eyewitness report. Cognitive Psychology, 1975, 7, 560-572. Loftus, E. F. Remembering recent experiences. In L. S. Cermak (Ed.). Human memory and amnesia. Hillsdale, New Jersey: Erlbaum, 1982. Loftus, E. F.,Miller, D. G., & Burns, H. J. Semantic integration of verbal information into a visual memory. Journal of Experimental Psychology: Human Learning and Memory, 1978,4, 19-31, Luria, A. R. The working brain: An introduction to neuropsychology. New York: Basic Books, 1973. Malpass, R. S., & Devine, P. G. Guided memory in eyewitness identification. Journal of Applied Psychology, 1981,66, 343-350. Mandler, G . Organization and memory. In K. W. Spence & J. T. Spence (Eds.), The psychology of learning and motivation (Vol. I). New York: Academic Press, 1967. Mandler. G . Recognizing: The judgment of previous occurrence. Psychological Review, 1980, a7,2~2-271. Mandler, J., & Stein, N. Recall and recognition of pictures by children as a function of organization and distractor similarity. Journal of Experimental Psychology, 1974,102,657669. Markus, H. Self-schemata and processing of information about the self. Journal of Personality and Social Psychology, 1977, 35, 63-78. McCauley, C., Parmelee, C. M., Sperber, R. O., & Carr, T. H. Early extraction of meaning from pictures and its relation to conscious identification. Journal of Experimental Psychology: Human Perception and Performance, 1980, 6 , 265-276. McCloskey, M., & Santee, J. Are semantic memory and episodic memory distinct systems? Journal of Experimental Psychology: Human Learning and Memory, 1981, 7,66-71. McGeoch, J. A. Forgetting and the law of disuse. Psychological Review, 1932. 39, 352-370. Merluui, T. V., & Johnson, N. F. The effect of repetition on iconic memory. Quarterly Journal of Experimental Psychology, 1974, 26, 266-273. Mervis, C. B., & Rosch, E. Categorization of natural objects. AnnualReview of Psychology, 1981,32,89-115. Meyer, D. E., & Schvaneveldt, R. W. Meaning, memory structure, and mental processes. Science, 1976, 192, 27-33. Miller, G. A., Galanter, E.. & Pribram, K. A. Plansand thestructure ofbehavior. New York: Holt, 1960.


121

Milner, B. Memory disturbance after bilateral hippocampal lesions. In P. M. Milner & S. E. Glickman (Eds.), Cognitive processes and the brain. Princeton, New Jersey: Van Nostrand. 1965. Mischel, W. Objective and subjective rules for delay of gratification. In G. D’Ydewalle & W. Lens (Eds.), Cognition in human motivation and learning. Hillsdale, New Jersey: Erlbaum, 1981. Moray, N. Attention in dichotic listening. Affective cues and influence of instructions. Quarterly Journal of Experimental Psychology, 1959, 11, 56-60. Moscovitch, M. Multiple dissociations of function in amnesia. In L. S . Cermak (Ed.), Human Memory and Amnesia. Hillsdale, New Jersey: Erlbaum, 1982. Neisser, U. Cognitive psychology. New York: Appleton, 1967. Nelson, D. L. Remembering pictures and words: Appearance, significance, and name. In L. S. Cermak & F. I. M. Craik (Eds.), Levels of processing in human memory. Hillsdale, New Jersey: Erlbaum, 1979. Nelson, T. 0. Detecting small amounts of information in memory: Savings for nonrecognized items. Journal of Experimental Psychology: Human Learning and Memory. 1978,4453468.

Nelson, T. 0..Fehling, M. R., & Moore-Glascock, J. The nature of semantic savings for items forgotten from long-term memory. Journal of Experimental Psychology: General, 1979, 108, 225-250.

Newell, A., & Simon, H. A. Human problem solving. New York: Prentice-Hall, 1972. Nilsson, T. H.. & Nelson, T. M. Delayed monochromatic hue matches indicate characteristics of visual memory. Journal of Experimental Psychology: Human Perception and Performance, 1981, 7 , 141-150. Norman, D. A., & Bobrow. D. G. On data limited and resource limited processes. Cognitive PSyChOlOgV, 1975, 7 , 44-64. Paivio, A. Imagery and verbalprocesses. New York: Holt, 1971. Pavlov, I. P. The reply of a physiologist to psychologists. Psychological Review, 1932, 39, 91-127.

Peterson, M. J. The retention of imagined and seen spatial matrices. Cognitive Psychology, 1975, 7, 181-193.

Posner, M. I. Chronometric explorations of mind. Hillsdale, New Jersey: Erlbaum, 1978. Posner, M. I., & Keele, S. W. On the genesis of abstract ideas. Journal of Experimental psycho lo^, 1968, 77, 353-363. Postman, L., & Underwood, B. J. Critical issues in interference theory. Memory and Cognition, 1973, 1, 19-40. Rand, G., & Wapner, S. Postural status as a factor in memory. Journal of Verbal Learning and Verbal Behavior, 1967, 6 , 268-27 1. Raye, C. L., Johnson, M. K., & Taylor, T. H. Is there something special about memory for internally-generated information? Memory and Cognition, 1980, 8, 141-148. Reber, A. S., & Lewis, S. Implicit learning: An analysis of the form and structure of a body of tacit knowledge. Cognition, 1977, 5, 333-361. Rock, I., & Gutman, D. The effect of inattention on form perception. Journalof Experimental Psychology: Human Perception and Performance, 1981, 7 , 275-285. Roediger, H. L., & Payne, D. G. Hypermnesia: The role of repeated testing. Journal of Experimental Psychology: Learning, Memory, and Cognition, 1982, 8, 66-72. Rothkopf, E. Z . Incidential memory for location of information in text. Journal of Verbal Learning and Verbal Behavior, 197I , 10, 608-61 3. Rundus, D. Maintenance rehearsal and single-level processing. Journal of Verbal Learning and Verbal Behavior, 1977, 16, 665-681.

Marcia K. Johnson

122

Sachs, J. S. Recognition memory for syntactic and semantic aspects of connected discourse. Perception & Psychophysics. 1967, 2, 437-442. Salaman, E. A collection of moments. London: Longman, 1970. Sampson. E. E. Cognitive psychology as ideology. American Psychologist, 1981.36.730-743. Scarborough, D., Cortese. C., & Scarborough, H.Frequency and repetition effects in lexical memory. Journal of Experimental Psychology: Human Perception and Performance. 1977. 3, 1-17.

Schachtel, E. G. On memory and childhood amnesia. Psychiatry, 1947, 10, 1-26. Schacter, D. L., & Tulving. E. Amnesia and memory research. In L. S. Cermak (Ed.), Human memory and amnesia. Hillsdale, New Jersey: Erlbaum, 1982. (a) Schacter, D. L., & Tulving, E. Memory, amnesia, and the episodic/semantic distinction. In R. L. lsaacson & N. E. Spear (Eds.), The expression of knowledge. New York: Plenum, 1982. (b) Schank, R. C., & Abelson, R. Scripts, plans, goals, and understanding. Hillsdale, New Jersey: Erlbaum, 1977. Schneider, G. E. Two visual systems. Science, 1969, 163, 895-902. Shettleworth, S. J. Memory in food-hoarding birds. Scientific American, 1983, 248, 102-1 10. Shiffrin, R. M., & Schneider, W. Controlled and automatic human information processing: 11. Perceptual learning, automatic attending, and a general theory. Psychological Review, 1977. 84, 127-190.

Slamecka, N. J., & Graf, P. The generation effect: Delineation of a phenomenon. Journal of Experimental Psychology: Human Learning and Memory, 1978, 4, 592-604. Spear, N. E. Retrieval of memory in animals. Psychological Review, 1973, 80, 163-194. Spear, N. E. The future study of learning and memory from a psychobiological perspective. In V. Sarris & A. Parducci (Eds.), Perspectives in Psychological Experimentation. Hillsdale, New Jersey: Erlbaum, 1983, in press. Spelke, E., Hirst, W., & Neisser, U. Skills of divided attention. Cognition, 1976, 4, 215-230. Sperling, G. The information available in brief visual presentations. Psychological Monographs, 1960, 74, (Whole No. 498). Stromeyer, C. F., Ill, & Psotka, J. The detailed texture of eidetic images. Nature. (London), 1970, 225, 346-349.

Tolman, E. C. There is more than one kind of learning. Psychological Review, 1949, 56, 144-155.

Treisman, A. M., & Gelade, G. A feature-integration theory of attention. Cognitive Psyc h o l o ~ 1980, , 12. 97-136. Treisman. A. M., & Schmidt, H.Illusory conjunctions in the perception of objects. Cognitive Psychology, 1982, 14, 107-141. Tulving, E. Theoretical issues in free recall. In T. R. Dixon & D. L. Horton (Eds.), Verbal behavior and general behavior theory. New York: Prentice-Hall, 1968. Tulving, E. Episodic and semantic memory. In E. Tulving & W. Donaldson (Eds.), Organization of memory. New York: Academic Press, 1972. Tulving, E. Relation between encoding specificity and levels of processing. In L. S. Cermak & F. I. M. Craik (Eds.). Levels of processing in human memory. Hillsdale, New Jersey: Erlbaum, 1979. Tulving, E. Similarity relations in recognition. Journal of Verbal Learning and Verbal Behavior, 1981, 20, 479-496. Tulving, E., & Thomson, D. M. Encoding specificity and retrieval processes in episodic memory. Psychological Review, 1973, 80, 352-373. Turvey, M. T. Repetition and the preperceptual information store. Journal of Experimental Psychology, 1967, 74, 289-293. Tversky, B. Encoding processes in recognition and recall. Cognitive Psychology, 1973, 5, 275-281.


123

Tyler, S. W., & Ellis, H. C. A test of alternative models of sentence recognition. Bulletin of the Psychonomic Society, 1978, 12, 375-377. Tyler, S. W., Hertel, P. T., McCallum, M. C., & Ellis, H. C. Cognitive effort and memory. Journal of Experimental Psychology: Human Learning and Memory, 1979, 5 , 607-617. Tzeng, 0. J. L. Sentence memory: Recognition and inferences. Journal of Experimental Psychology: Human Learning and Memory, 1975, 1. 720-726. Tzeng, 0. J. L., Lee, A. T., & Wetzel, C. D. Temporal coding in verbal information processing. Journal of Experimental Psychology: Human Learning and Memory, 1979, 5 , 52-64.

Underwood, B. J. Attributes of memory. Psychological Review, 1969, 76, 559-573. Warrington. E. K., & Weiskrantz, L. Amnesic syndrome: Conso!idation or retrieval? Nature (London), 1970, 228, 628-630. Weiskrantz, L., Warrington, E. K., &Sanders, M. D. Visual capacity in the hemianopic field following a restricted occipital ablation. Brain, 1974, 97, 709-728. Wickelgren, W. A. Learning and memory. New York: Prentice-Hall, 1977. Wickelgren, W. A. Chunking and consolidation: A theoretical synthesis of semantic networks, configuring in conditioning, S-R versus cognitive learning, normal forgetting, the amnesic syndrome, and the hippocampal arousal system. Psychological Review, 1979, 86, 44-60. Wickelgren, W. A. Human learning and memory. Annual Review of Psychology, 1981, 32, 21-52.

Wickens, D. D. Encoding categories of words: An empirical approach to meaning. Psychological Review, 1970, 77, 1-15. Wilson, W. R. Feeling more than we can know: Exposure effects without learning. Journal of Personality and Social Psychology, 1979, 37, 81 1-821. Woodworth, R. S . Experimental psychology. New York: Holt, 1938. Zajonc, R. B. Feeling and thinking: Preferences need no inferences. American Psychologist, 1980, 35, 151-175.


THE COGNITIVE MAP OF A CITY: FIFTY YEARS OF LEARNING AND MEMORY Harry P. Bahrick OHIO WESLEYAN UNIVERSITY DELAWARE, OHIO

I . Introduction........................................................... 11. Subject Recruitment and Test Administration. . . . . . . . . . . . 111. Composition of the Tests . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . A. Free Recall of Streets.. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . B. Free Recall of Buildings and Other Landmarks . . . . . . . . . . . . . . . . . . . . . . . . C. Visually Cued Recall ............................................... D. Verbally Cued Test ................................................ E. Matching Test ..................................................... F. Questionnaires .................................................... IV. Scoring Procedures ..................................................... A. Spatial Sequence Scores . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . B. Questionnaire Scoring . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . V. Results and Discussion.................................................. A. Acquisition ....................................................... B. Retention ......................................................... VI. Summary ............................................................. References ............................................................

.

..

..

. ..

125 i27 130 130 131 131 131 134 134 134 135 137 138 138 150 161 163

I. Introduction The study of space perception has a long history in philosophy and psychology, but empirical work has only recently been extended from the microspace of the laboratory to the macrospace in which people live. Tolman (1948) first used the term “cognitive map” in an important paper in which he discussed how rats find their way in a maze. Since then the cognitive maps of children and adults have become a topic of increasing interest to geographers, architects, urban planners, sociologists, and others, but these scholars looked in vain to the psychological literature for information about the process by which such maps get into peoples’ heads. The pioneer work of Lynch (1960) emphasized the need for more ecologically oriented research, and during the last 20 years investigators have responded to this 125 THE PSYCHOLOGY OF LEARNING AND MOTIVATION, VOL. 17

Copyright 0 1983 by Academic Press, Inc. All rights of reproduction in any form reserved. ISBN 0-12-54331 7-4

126

H i n y P. Bibrick

need with a spate of studies of environmental space perception. This literature has been ably reviewed by Evans (1980) and Siege1 and White (1975). Psychologists were slow to respond to this need because learning about ecological space is an extended process that does not lend itself readily to the type of controls deemed essential for scientific study. The recent impetus to ecological research reflects new and broader criteria for evaluating scientific contributions and a growing acceptance of some sacrifices of controls in a tradeoff which encourages the investigation of important problems that can only be studied in an ecological setting (Bahrick & Karis, 1982). The research reported here deals with the acquisition of spatial information about a city during a 4-year period of residence, and the subsequent loss of this information over a period of 46 years. The investigation differs in several ways from earlier work on environmental cognition. Earlier investigations have emphasized the acquisition process, but have provided little information about long-term retention. This is a study of long-term memory. Other major differences are methodological. The present study involves the administration of 5 subtests, which yield 23 indicants of knowledge. The interrelations among these indicants provide a broad view of the available spatial information. Finally the method of cross-sectional adjustment (Bahrick, 1979) is used in an attempt to compensate for the lack of control over variables of acquisition and of rehearsal. The cognitive map of a city has special appeal to us as a vehicle for the study of long-term ecological memory because this type of knowledge requires the integration of linguistic with nonlanguage encoded information. The content areas that we had previously investigated were either based completely on language (the retention of the Spanish language) or they largely excluded linguistic organization, for example, memory for people’s faces (Bahrick, Bahrick, & Wittlinger, 1975; Bahrick, 1983). To conduct a cross-sectional investigation of the acquisition and maintenance of ecological knowledge, one must find a situation in which many individuals have acquired certain knowledge, but at different times. Further requirements are that one can estimate with acceptable reliability the knowledge originally acquired by each individual, and the exposure to relevant information during the retention interval. Knowledge of a college town on the part of those who were students at that college meets most of these requirements. Students come to the town at about the same age, they typically leave 4 years later, and the time of residence, as well as visits to the city after graduation, can usually be established with reasonable accuracy. Delaware, Ohio, turned out to be particularly well suited to the purpose, because the central portion of the city, containing the campus, had changed very little during a period of almost 50 years preceding the study.

The Cognitive Map of a City

11.

127

Subject Recruitment and Test Administration

Eight hundred fifty-one individuals (359 males, 492 females) participated in the investigation. Of these, 275 were undergraduate students at Ohio Wesleyan University, who were tested to obtain data concerning the acquisition of the cognitive map. Fifty-one of these were freshman tested during the fourth week of residence; 63 were freshman tested during the eighth month of residence; 64 were sophomores tested during the 17th month of residence; 53 were juniors tested during the 26th month of residence; and 44 were seniors tested during the 35th month of residence. Five hundred seventy-six were alumni of Ohio Wesleyan University who attended the college between 1929 and 1979 and were tested to obtain data concerning retention of knowledge. The alumni were assigned to one of 8 groups based upon their date of graduation. Table I shows the number of years intervening between graduation and the time of testing, and specifies various subject characteristics. Freshmen and sophomores were recruited from introductory psychology classes on the basis of a course requirement. Juniors and seniors were recruited from advanced undergraduate psychology courses and from sororities and fraternities. Most junior and senior students were paid for participating in the study. Later comparisons revealed no significant effects on performance due to the type of incentive offered for participation. Recruiting of alumni subjects turned out to be much more difficult. All alumni subjects were unpaid volunteers. Some were tested on campus on the occasion of a visit, others were tested at alumni meetings conducted in other cities, and still others self-administered the tests in their home and returned the completed test by mail to the investigator. The place of test administration and the presence or absence of a test administrator constitute potentially significant procedural variables. It turns out that being tested in Delaware significantly improved performance. The effect was dealt with by assigning those subjects tested on campus the extreme score (0) on the variable specifying the interval since the most recent visit to Delaware prior to taking the test. The variable of recency of visit was included in the multiple regression analysis described in more detail later. The decision to have subjects administer their own tests involves other considerations. These alumni were recruited by mail, and problems of sampling bias will be considered first. Sampling error could result from the selection of alumni who were asked to be subjects, from the fact that not all of those who were asked agreed to serve, and from the fact that some of those who agreed to serve failed to return their completed tests. It is therefore possible that the respondents were not a representative sample of alumni and that their performance may not be typical. Although some sampling bias is almost cer-

TABLE I

MEANSAND STANDARD DEVIATIONS OF KEY INDEPENDENT VARIABLES FOR THE EIGHTTIME GROUPSO Retention interval (months)

Frequency of visits (number per year)

Recency of visits (months)

Duration of visits (days per year)

Visits with car

Car on campus (months)

(W)

Group

Mean

SD

Mean

SD

Mean

SD

Mean

SD

Mean

SD

Mean

SD

1

13.66 44.35 76.46 126.25 179.20 249.20 340.58 556.42

2.43 7.39 9.38 15.95 16.52 20.66 49.47 48.98

2.32 .90 .41 .I8 .I8 .21 .62 1.04

3.24 1.38 .52 .2l .36 .35 .93 2.11

6.44 25.79 30.03 71.67 106.85 112.69 79.45 20.17

4.98 17.19 28.04 52.48 67.60 97.68 116.09 49.46

5.22

7.05 3.23 .77 .38 .36 .55 I .28 2.15

63.89 47.41 65.03 54.87 52.73 53.85 61.22 83.47

46.35 49.04 45.35 48.87 49.31 45.98 43.41 29.01

9.11 9.61 11.53 7.40 7.16 6.22 3.60 2.84

10.43 10.13 10.15 9.33 11.15 11.65 8.15 6.98

2 3 4 5 6 7

8

1 .%

.57 .27 .I9 .32 .8f

1.32

“From Bahrick (1979). Copyright 1979 by the American Psychological Association. Reprinted by permission of the publisher.


I29

tainly present, several considerations suggest that the effects are minor. Solicitations were extended to all alumni who had majored in psychology and whose addresses were available. In addition, invitations were extended to a large number of alumni who had taken a course in psychology from the investigator. Most often this was an introductory course elected by approximately 90% of students at the college. Of those who were asked 81070 agreed to serve, and 84% of those who agreed ultimately returned the completed self-administered tests. Among those 19% who failed to answer the original request were many who did so for reasons that are probably uncorrelated with test performance, for example, they were abroad or they had moved without leaving a forwarding address. Failure to return tests that were mailed (16%) constitutes a more serious selection problem, because of the likelihood that at least some of these alumni attempted the test, but found it too difficult or tiresome to complete. The overall effect on the data cannot be estimated with precision, but it is probably small. This is so even if the missing data represent a biased subsample, since the overall mean would be weighted more heavily by the 84% for whom data are complete, than by the 16% for whom data are missing. Additional problems associated with self-administration of tests may have arisen because of failure to follow some of the printed instructions sent with the test. Consulting others about answers or using maps or other sources of information are among the obvious examples. No time limits were set for the tests so that there are no problems of adherence to such limits, but it is possible that the average time spent with the test is significantly longer for those who administered their own tests, and that this affected the data. Alumni subjects were unpaid volunteers whose motivation was primarily a desire to support the research efforts of the investigator. This condition favors compliance with the directions of the test. The best evidence regarding the aggregate effect of the several sources of bias which have been discussed comes from comparisons of test performance of alumni who were tested by an administrator with performance of those alumni who administered their own tests. Although these differences generally favor those who administered their own test, the differences failed to reach statistical significance, and are generally not large enough to cast doubt on or alter the major conclusions of this study. The previous discussion provides an illustration of problems likely to be encountered in ecologically oriented research involving large numbers of participants. The solutions adopted here represent compromises with optimum procedures with regard to subject selection and control over test conditions. The acceptability of such compromises is a matter which does not allow facile generalizations. The decisions made here reflect subjective judgments concerning the value of data not obtainable by standard meth-

130

Harry P. Bahrick

ods, and expectations regarding the magnitude of effects associated with various sources of bias. Such decisions must usually be made without adequate information, and penalties for errors in judgment are likely to be severe. The consequences of errors may not be evident until a time when much effort and expense have been invested. The option to start over is less attractive than in the case of most laboratory investigations in which mistakes have been made. These considerations must be included in evaluating large scale ecologically oriented research, and they have been discussed elsewhere (Bahrick & Karis, 1982).

111. Composition of the Tests

Based on our own exploratory work and the reports of others, we decided against the use of hand-drawn sketch maps in this investigation. Problems reported by others include the confounding of drawing ability with spatial knowledge (Golledge, 1976; Blaut & Stea, 1974) and the lack of independence among drawn elements (Evans, 1980). Our own experience showed that the hand-drawn sketch maps interact with other tests that we wanted to administer, and that they would excessively prolong the time required for total testing. We administered the following 5 subtests: (1) free recall of streets, (2) free recall of buildings and other landmarks, (3) visually cued recall, (4) verbally cued recall, and ( 5 ) matching. All subjects took 4 out of these 5 tests in the order listed, with one-half of the subjects arbitrarily assigned to the visually cued recall test and the other half to the verbally cued test. An effort to minimize interaction effects among subtests guided the sequence in which the subtests were administered, and the decision to administer either a verbally or a visually cued recall test, but not both to the same subjects. A. FREERECALLOF STREETS Subjects were required to list all street names of Delaware that they could recall. They then were required to classify each listed street as running in either a north-south direction or an east-west direction, and to arrange the streets assigned to each category in correct order. The streets running northsouth were to be ordered beginning with the street farthest west (assigned number 1) to the street farthest east; the streets categorized as running eastwest were to be ordered from the street farthest north to the street farthest south.


B.

131

FREERECALLOF BUILDINGS AND OTHERLANDMARKS

This subtest consisted of two sections. The first section required free recall of names of buildings and landmarks in the city of Delaware; the second section required free recall of the names of buildings and landmarks on the campus of Ohio Wesleyan University. Each section of the test further required that the recalled items be ordered separately with regard to the north-south dimension and with regard to the east-west dimension. Subjects who listed 10 or fewer responses were required to order all of their responses in both dimensions. Subjects who listed 11-20 responses were required to order only their odd-numbered responses, ignoring the evennumbered responses. Subjects who listed between 21 and 30 responses ordered every third listed response, ignoring the others. The same procedure was applied to both sections of the subtest. We chose the technique of two-dimensional spatial ordering of locations and streets and sacrificed direct distance estimation because preliminary data indicated that the ordinal sequences obtained were more reliable than distance estimates, and yet correlated highly with distance estimates. This finding probably applies only when distances are relatively small, and can be inferred from the number of city blocks separating locations. Our findings also agree with those of Baird, Merrill, and Tannenbaum (1979), who report that either pairwise judgment or direct mapping yield accurate representations of spatial relations in a familiar environment. C. VISUALLYCUEDRECALL

Subjects were provided the outline map of the city of Delaware shown in Fig. 1. The' map identifies streets by letters, and buildings and landmarks by numbers. Subjects were instructed to write down the names of any streets or buildings that they were able to recall with the help of the visual cues provided by the map. Subjects were encouraged to guess if necessary, and to recall in any order. They were also told that only buildings and places which had existed since 1930 were included on the map. D.

VERBALLY CUEDTEST

This test consisted of two sections. The first section provided subjects with an alphabetized sequence of names of 20 Delaware streets. The subjects were instructed to cross out street names that they did not recognize, and to assign the remaining streets either to a category of streets running north-south, or to a category of streets running east-west. Within each category subjects were then required to order the streets as in the test of

132

Harry P. Bahrick

Fig. 1 . Outline map of Delaware.


133

TABLE I1 ALUMNI QUESTIONNAIRE Male 1.

Female

Years of attendance at Ohio Wesleyan University.

to 19

19

2. College major

-Years

3.

How many academic years (Sept.-June) did you live in Delaware?

4.

How many summer months did you live in Delaware in addition to the above?

5.

If you lived in Delaware before or after college, please state number of years Before Years After

Months

Years

6. How much of the time did you drive a car while living in Delaware? Years Months 7.

How much of the time did you use a bicycle for transportation while living in Delaware? Years Months

8.

How much of the time did you have a job which involved visits to parts of town other than campus or dormitory? Years Months

9. Please check the place where you lived while in college. Check more than one if applicable. Dormitory Private home Fraternity house 10.

How many times have you visited Delaware since you left Ohio Wesleyan?

11.

What are the approximate dates and duration of your visits to Delaware?

I Date

I

No. of hours

I

No. of days

I

Date

No. of hours

12. On how many of the above visits did you drive a car when you visited Delaware? Times 13. 14. 15.

How often have you used a map to find your way around in Delaware? Frequently Occasionally How often have you used a map to find your way around other cities? Frequently Occasionally Did you work as a tour guide while at Ohio Wesleyan? Yes

Never Never

-No

Times

No. of days

Harry P. Bahrick

I34

free recall of street names. Section 2 provided a list of names of 26 buildings and other landmarks in the city and on the Ohio Wesleyan University campus. Subjects were instructed to cross out names that they failed to recognize, and to order all of the remaining landmarks independently along the two compass dimensions, as in the subtest for free recall of buildings and landmarks. E. MATCHING TEST Subjects were provided with the outline map shown in Fig. 1 and with an answer sheet that listed the names of all streets identified by letter on the map, and of all landmarks identified by number on the map. The instructions required subjects to place the appropriate letters and numbers from the map next to the name of each street and landmark listed on the answer sheet. Guessing was encouraged. No time limits were imposed for any subtests. Subjects returned the answer sheets and instructions for each subtest before receiving the next subtest. Total testing time varied considerably, but averaged about 55 minutes. F. QUESTIONNAIRES In addition to completing 4 subtests, all subjects completed a questionnaire. A student questionnaire contained questions concerning duration and location of residence in Delaware, type of transportation used, use of city or campus maps, and other questions concerning conditions which we though might affect the exposure to information about the campus and city. An alumni questionnaire listed additional questions concerning dates and durations of visits to the city during the retention interval. The alumni questionnaire is shown in Table 11.

IV. Scoring Procedures Objective scoring was relatively simple for the visually and verbally cued subtests as well as the matching test, but somewhat complicated for the subtests of free recall. The matching and cued recall subtests were limited to those aspects of the city which had remained unchanged over a 50-year period, but no such restrictions could easily be applied to the free recall of landmarks and buildings. Alumni often listed the names of landmarks which had been renamed, had disappeared, or could not be identified. Extensive research was conducted to develop semiobjective scoring of these responses. Among the sources used to determine the accuracy of names and locations


135

given were old city directories, telephone books, college catalogues, yearbooks, and interviews with archivists and local history buffs. Based on these sources, a composite map was constructed. This map showed the names and locations of all verified landmarks. The map was updated as previously unlisted responses were verified. An arbitrary system of scoring was used which awarded one-half point for names of landmarks given incompletely or partly correct, for example, “high school” instead of “Hayes High School,” or “park” instead of “Blue Limestone Park.’’ Responses were awarded one point credit if they were complete and correct. Subscores were calculated for recall of names of streets with north-south orientation, with east-west orientation, and for the total number of buildings and landmarks recalled. A.

SPATIAL SEQUENCE SCORES

The rationale for scoring spatial sequences is based on the assumption that spatial knowledge is revealed by any correct listing of a street or landmark in relation to any other street or landmark with regard to the compass direction being ordered. One point was therefore awarded for each correct relational listing. Thus, if 5 streets are to be arranged in their correct northsouth sequence, the highest score obtainable is 10, that is, one point for each of the N(N- 1)/2 possible paired comparisons, where N i s the number of streets to be ordered. If the correct sequence is 1, 2, 3, 4, 5 , a listing of 1, 3, 2, 4, 5 would yield a score of 9. Streets that were assigned incorrectly with regard to their direction were omitted from the sequence used to calculate the score. Separate sequence scores were calculated for the northsouth and for the east-west sequence. Some subjects mislabeled a sequence, that is, they listed all streets with a north-south orientation in the east-west category, and vice versa. It was generally not possible to determine whether such errors reflected problems of disorientation with regard to compass directions or careless use of the label. The compass directions are part of most Delaware street names and signs, for example, North Sandusky Street, but several subjects reported that they are habitually unaware of compass directions. In all instances of systematic 90” errors, the errors were recorded but subjects were given credit for the full sequence score, as if the street sequence had been placed in the correct category. This scoring strategy was also followed when subjects committed errors of 180”, that is, when they listed streets appropriately in regard to the north-south and east-west directional categories, but reversed the numbering sequence so that the number 1 was assigned to the street farthest west, rather than the street farthest east. This scoring decision is based on the rationale that the sequence score should reflect knowledge of locations within the city relative to each other,

I36

Harry P. Bahrick

rather than information concerning locations in the city relative to a larger geographic space. Thus, a subject instructed to list streets in order from north to south who produces a sequence of 5 , 4, 3, 2, 1 instead of 1, 2, 3, 4, 5 is judged to have knowledge about the location of streets in relation to each other and receives full credit for the listing. Data concerning the frequency and significance of 90" and 180" errors of orientation were analyzed separately. The decision to score spatial knowledge for only a portion of the free recall responses also requires comment. Subjects varied greatly as to the number of free recall responses listed. If all responses are scored with regard to the correctness of the spatial sequence, the attained sequence score is primarily a function of the number of items listed, rather than the accuracy of knowledge of their relative location. Thus, a subject who lists the names of only 5 landmarks and arranges them in the correct sequence receives a score of 10, but a subject who lists the names of 30 landmarks receives a score of 218 solely on the basis of chance success in arranging a random sequence. A guessing correction would not eliminate this dependence of spatial scores on the number of items listed. Prorating spatial scores in relation to the number of items listed was considered, but discarded. This system has the disadvantage that individuals who list only two or three items and arrange their spatial sequence correctly would receive the same spatial knowledge score as those who correctly arrange the sequence of a large number of items. Another problem of comparability of spatial sequence scores arises from the variation in average distance between the listed locations. This distance tends to diminish if more items are listed, and thus more detailed spatial knowledge is required for each correct sequence element in a longer list of items. The strategy of basing the sequence score on a limited number of items, arbitrarily selected from among those listed, considerably reduces the problems discussed above. Furthermore, this procedure reduces the excessive amount of time required to complete the tests for those individuals who listed a large number of items. Preliminary data indicated that older subjects were particularly unlikely to complete a test which required more than 1 hour time, and it became necessary to limit the required tasks to accommodate this constraint. The above scoring method was applied to the subtests for free recall of streets and to the free recall of buildings and landmarks in the city and on the campus. The visually and verbally cued tests limited the number of possible responses and therefore did not present the scoring problems discussed above. The score obtained from the visually cued test was the number of buildings and streets correctly identified with the help of the map. The verbally cued test yielded separate spatial sequence scores for streets and for


137

landmarks. Further subscores were calculated for the east-west and northsouth dimension, and these calculations were performed as before, that is, by awarding one point for each correct paired comparison. Errors of compass orientation were treated as before. The matching test yielded two scores, one for streets and the other for landmarks, with one point awarded for each correct response. In several instances subjects failed to take one of the subtests or failed to follow directions and a subtest had to be discarded. To accommodate various computer-aided analyses, the missing data were supplied by substituting the mean score on that subtest obtained by the group to which the subject belonged. Less than 1% of the data were supplied in this manner.

B. QUESTIONNAIRE SCORING Data pertaining to residence in DeIaware were converted to months, as were data on the use of a car or bicycle and the length of the retention interval. The total residence score was the aggregate number of months of residence as a student, in addition to any summer residence, or residence prior or subsequent to the enrollment at the university. Individuals with more than 2 years of residence in the city prior to or subsequent to their enrollment at the college, or with residence in the city during their childhood years, were excluded from the study. For most subjects, the residence score is based upon four periods of 9-month residence with interruptions during the summer vacation period. Fewer than 3% of subjects reported interruptions of a year or longer because they transferred or dropped out of school. Four scores were calculated from the data specifying visits to Delaware during the retention interval. A frequency of visit score was obtained by counting the number of separate trips made to the city, with no regard to the duration of each visit. This score was divided by the retention interval for each subject and is expressed in terms of the mean number of visits per year of the interval. A recency of visit score was defined as the number of months elapsed since the last visit prior to taking the test. As previously mentioned, a score of zero on this variable indicates that the test was taken on campus. A duration of visit score was calculated by counting the aggregate number of days spent in Delaware on all visits subsequent to termination of residence (usually the time of graduation). This score was also divided by the retention interval and expressed in terms of the mean number of visit days per year. Calculation of a distribution of visits score involved the preliminary step of dividing the retention interval of each subject into ten equal time segments. Next, the number of visit days which fell into each of these time segments was established, and the deviation of these numbers

Harry P. Bahrick

138

from an equal distribution of visit days determined. The equal distribution score was determined by assigning each interval one-tenth of the total number of visit days. The deviations between the actual visit days in each time interval and the equal distribution days were averaged for the ten intervals to obtain an average deviation. This method of calculating a distribution of visits score turned out to be preferable to several alternatives considered, but it did not resolve some problems, for example, uncertainty on the part of the subject with regard to the exact dates of certain visits. Results and Discussion

V.

A. ACQUISITION Figures 2 and 3 present acquisition functions based upon 14 indicants of knowledge. Figure 2 shows acquisition of knowledge of street names, and Fig. 3 knowledge of the names of buildings and other landmarks on campus and in the city. Street names are learned at a fairly steady rate of about 2-3 names per year, following the first 3 weeks of residence. This rate shows no apparent diminution during the 36 months of residence, and is largely independent of which measure is used; that is, the rate of gain following the first month of residence is about the same for free recall, visually cued recall, and the matching indicant. Absolute differences in

0 FREE RECALL 14{

AMATCHING STREET N M E S

0 VISUALLY CUED STRZfT NAMES 12-

03

W

z

10-

Z

l0 W

8-

[L

a: 0 0

6-

!L

0

L L " w

m

z

3

z

2-

0 , 75

8

17

26

MONTHS OF RESIDENCE

Fig. 2. Learning of street names.

35


139

0 FREE RECALL OF ClTV LPNDMMlKS A FflEE RECALL OF CAUPUS LANDMARKS 0 MATCHING CAMRIS BvlLDlffiS 0 MATCHING CITY LANDMARKS A VISUALLY CUED CAMPUS BUILDINGS

LL

0

E 3

5

2

05 75

0

17

I

I

26

35

MONTHS OF RESIDENCE

Fig. 3. Learning the names of buildings and landmarks.

number of correct responses yielded by the various indicants are also smaller than those typically reported in laboratory investigations, and the directions of differences are the reverse of the ones usually found. The number of free recall responses exceeds the number of correct responses on the matching and on the visually cued recall test. This reversal of typical findings reflects the nature of the tests, as well as the circumstances of naturalistic learning. The set of responses available for the free recall test is the entire set of names of city streets, while the visually cued responses and the responses on the matching test represent a smaller subset, that is, the 18 principal streets in the center of the city. Performance on the matching and cued tests would presumably be somewhat higher if the names of all city streets were cued. Thus, the tests are not completely comparable. More important, free recall responses require only knowledge of the street name, but correct responses on the matching and visually cued tests require in addition an association between street name and location. It follows that some learning of street names occurs without spatial association, or in the context of spatial associations insufficient to mediate retrieval on the basis of the cues provided on the map. A related condition limiting performance on the cued tests involves the nature of the cues. The outline map of the city is used to provide the cues, and some subjects report that they have had limited previous experience with the use of maps. Typically, interindividual variance of performance is larger on these tests than on the free recall test. Furthermore, many of the cues provided on the outline map can be generated

140

Harry P. Bahrick

by subjects without the benefit of the map. Most student subjects (but not long-term alumni) report that their free recall of street names is based upon their own cognitive map of the city, and the validity of this claim is supported by the fact that the order in which street names are given in free recall often corresponds to the correct sequence. The superiority of cued recall and matching scores over free recall typically reported in laboratory investigations depends upon providing cues that are not easily self-generated. In our previous investigation of memory for high school classmates (Bahrick et al., 1975), performance on picture cued recall and performance on a name-face matching task greatly exceeded the free recall of names. In that situation the pictorial cues yield performance superior to free recall because subjects cannot easily generate these cues on their own. Figure 3 shows the acquisition function for the names of landmarks on campus and in the city. The rate of learning new locations diminishes throughout the residence period, and comparatively little new information is added during the last two academic years. This is in contrast to the slow, steady learning of street names. Knowledge of campus landmarks exceeds knowledge of city landmarks for all indicants during the early residence period, but this superiority is lost by the end of the first year of residence. Campus locations are apparently so important in the daily lives of students that a large amount of information is acquired early on, while learning of city landmarks proceeds at a slower and more even pace. Comparisons of free recall, cued recall, and matching performance show the results previously noted, but differences are even more pronounced. The number of free recall responses exceeds the number of correctly matched responses by a factor of more than 2: 1. We explain this as before. The cued and matched responses can only be effective if there is some spatial knowledge. Although free recall responses are frequently mediated by self-generated spatial cues, some free recall responses apparently occur with very little spatial knowledge. Furthermore, the size of the free recall set (all names of buildings and other landmarks in the city) exceeds by a very large ratio the size of the subset of 26 landmarks cued on the outline map. Thus, the previously noted lack of comparability of the tests has become a more important influence on relative performance. Figure 4, which, like Figs. 2 and 3, also is based upon 14 indicants, shows the acquisition of spatial knowledge as reflected by various sequence scores. It was noted earlier that knowledge of spatial sequence also plays a role in the free recall of names of streets and landmarks, even though no requirement of spatial order is imposed in that task. The high correlation between free recall of street names and street sequence scores is of course also a consequence of requiring subjects to specify the sequence of those street names that they generate in free recall. The larger the number of


.75

a

17

26

141

35

MONTHS OF RESIDENCE

Fig. 4. Learning spatial order.

names generated, the larger is the possible sequence score. However, the correlation is almost equally high in the verbally cued test where the dependence is not built in. This suggests that street names are typically learned in the context of spatial learning. Following the initial spurt, sequence learning continues at an even rate throughout the 4-year residence period, with no indication of diminution. The rate of gain of information is approximately the same for sequences based on the free recall of landmarks and the free recall of streets and verbally cued streets. Scores for the spatial order of verbally cued landmarks were much higher than scores for any other indicant. They were divided by a constant of 4 in order to accommodate the function within Fig. 4. Direct comparison of this acquisition function with the function for spatial sequence of freely recalled landmarks is questionable because the free recall sequence is based upon a sample of no more than 10 landmarks. In contrast, 26 landmarks were verbally cued, and subjects arranged a sequence for all of the landmarks they recognized. No such differential ceilings affect the comparison of functions for the spatial order of streets. Subjects were instructed to arrange a sequence of all street names that they recalled rather than a sample of 10, because the number of streets recalled is only a fraction of the number of recalled landmarks. As a result, the sequence scores based on streets are relatively free of a ceiling effect, or a sampling effect. Apparently, knowledge of streets is acquired at a much slower rate than knowledge of landmarks. These results also show that names are more effective cues in accessing spatial knowledge than the visual cues provided in our map. Performance based

Harry P. Bahrick

142

upon verbal cues equals or exceeds performance based upon free recall, while performance based upon the outline map is substantially below free recall performance. Of course, the visual cues provided on the outline map are only useful as indicants of location; they do not give a visual representation of the landmark or building, and thus they permit no general conclusion in regard to the relative effectiveness of verbal versus visual cues.

I . Directional Differences Subjects have more information about streets running in the east-west direction than in the north-south direction. This is consistent for all stages of learning and affects knowledge of street names as well as knowledge of their spatial sequence. The difference is most pronounced for the spatial sequence of verbally cued streets because more verbal cues are given for the east-west than for the north-south streets, but it is also observed for the sequence of street names given in free recall. No consistent directional differences are observed for knowledge of the spatial sequence of landmarks on campus and in the city. Scores based on the east-west sequence are somewhat higher for campus landmarks as are scores based on the north-south dimension for city landmarks. The best explanation for these directional findings is based upon the geographic separation of the residential and academic buildings of the college, and the travel routes that students usually follow on their way to and from classes. Most residential buildings are located approximately 1 mile west of the academic buildings, so that students take one of several streets running from east to west to reach the academic buildings. Much more is learned about the names and sequence of these streets than about the names and sequence of streets running north-south, which are crossed, but rarely traveled. Knowledge of city locations also reflects this difference, favoring the east-west dimension, but to a lesser degree. The difference is absent or reversed in relation to campus landmarks. Travel on campus apparently follows east-west routes and northsouth routes equally frequently. 2.

Comparisons with Other Findings

Although the literature contains no other accounts of large-scale investigations of the acquisition and retention of the cognitive map of a city, data from several other investigations are relevant for comparison of the acquisition process. Our data are in close agreement with those reported by Herman, Kail, and Siege1 (1979). They tested college freshmen for their knowledge of the campus of the University of Pittsburgh after 3 weeks, 3 months, and 6 months of residence. They avoided the use of sketch maps


143

and used a number of tests similar to ours. Their subjects recalled the names of 15 campus landmarks after 3 weeks of residence (as compared to 20 landmarks for our subjects) and they report that knowledge about the campus begins to approach an asymptote after 3 months. Our acquisition function for campus landmarks supports this conclusion. Their subjects also performed better on a free recall task than on a matching task, but the difference is less pronounced than the one that we obtained because they used more realistic visual representations of buildings for their matching tests. Devlin (1976) reports data based upon sketch maps of the city of Idaho Falls, Idaho. The maps were hand drawn by 26 wives of servicemen after 2% weeks of residence, and again 3 months later. She thus used a longitudinal design in which the same subjects are tested repeatedly. This has the great advantage of avoiding differences in subject characteristics across time groups as a source of error, but the possible disadvantage of interaction effects among successive tests. Longitudinal designs are practical for investigations which cover relatively short time spans, but extremely difficult when an investigation spans several decades so that subject attrition becomes a major factor, and the time needed to conclude a longitudinal study is a major deterrent. Devlin's data are also based upon a somewhat larger city (40,000 versus 15,000) and upon the use of hand drawn maps. In spite of these differences, comparisons of her results with ours are worthwhile. The most striking difference between the present findings and those reported by Devlin pertains to knowledge of streets. Devlin reports that her subjects were able to draw and identify an average of 20.3 streets after 2% weeks of residence, and this number nearly doubled 3 months later. In contrast, our subjects recalled the names of only 4 streets after 3 weeks of residence, and only 6 streets after 8 months of residence. Our matching and visually cued tests produced even lower results. This difference is almost certainly due to the alphabetical and numerical designation of many street names in Idaho Falls. The great majority of streets identified in the sketch maps shown by Devlin run parallel and are consecutively numbered or lettered. This simple organization greatly facilitates acquisition and retrieval of street names and of street sequences. In addition, our method of testing does not permit subjects to demonstrate spatial knowledge of streets for which they could not recall, or recognize the name. The numerical organization of streets in Idaho Falls, and the use of the sketch map technique allows subjects to demonstrate such knowledge. Devlin does not report the average number of landmarks shown in the sketch maps of her subjects, but it would appear that the number is substantially smaller then the average number of names of landmarks and buildings recalled'by our subjects. Her 26 subjects drew a total of 40 dif-

144

Harry P. Bihrick

ferent landmarks after 2% weeks of residence. Since all subjects do not report the same landmarks, the average number per subject is presumably very much smaller. During the fourth week of residence, our subjects recalled on the average the names of 15 landmarks in the city and 20 landmarks on campus. It is apparent that what was learned by each group of subjects reflects their needs, and their opportunities to learn. The superior early learning of landmarks of our subjects reflects the needs of college students to identify many campus buildings by name during the first weeks of college life. The close proximity of many buildings encountered by walking over a relatively small campus area facilitates learning the names of buildings, rather than the names of many streets. In contrast, Devlin’s subjects presumably learned about the city by driving from their home to various areas of the city for purposes of shopping, etc., and in the process had more need for and greater opportunity to become acquainted with a wide network of streets, but comparatively less need and opportunity for learning the names of many buildings. The importance of organizational aids (e.g., having parallel streets numbered or lettered consecutively) should be stressed again in accounting for these differences. In a somewhat different context, Maki’s (1981) research on distance estimation shows that an organization based upon categorizing locations (with regard to areas) can greatly facilitate acquisition and retrieval of spatial information. The spatial map of Delaware does not offer obvious opportunities for organizing names of streets, and our results reflect this fact. Very rapid learning of major streets and landmarks is also shown in data reported by Pearce (1977), whose subjects were tourists in Oxford. They greatly improved their sketch maps between the second and sixth day of their visit and demonstrate that spatial learning reflects the needs of the individual and not simply the frequency of exposure to information. 3. Intercorrelations among Indicants

Tables 111 and IV show intercorrelations among 13 indicants of learning during the 4th week of residence and during the 35th month, respectively. The intercorrelations show the extent to which individual differences of performance on each indicant are predictable from a knowledge of individual differences on the other indicants. It is not surprising that the acquisition of information about the city is relatively independent of acquisition of information about the campus, and that learning the names of streets is independent of learning the names of landmarks. Comparison of comparable cells in the two tables permits an analysis of changes of intercorrelations during acquisition. This reveals that a majority of intercorrelations

TABLE 111 INTERCORRELATIONS AMONG INDICANTS OF PERFORMANCE DURING THE

FOURTH WEEK

OF RESIDENCE ~

Free recall of all street names Free recall of east-west streets Spatial order for all streets Spatial order for east-west streets Free recall of city landmarks Spatial order of city landmarks East-west order of city landmarks Free recall of campus landmarks Spatial order of campus landmarks East-west order of campus landmarks Matching streets Matching university buildings Matching city landmarks

1

2

3

1.oo .17 .71 .28 .17 - .01 - .01 .16 .05 .02 .30

.I7 1.oo .36 .79

.71 .36 1.oo .58 .I5

.40 .32

.40 .16 .07 .34 - .03 -.16 .56 .28 .29

4 .28 .79 .59

.08

1.oo .26 .10

.oo

.oo

.02 .12 .07 .44 .20 .27

.17 .05

- .03 .63 .23 .30

5

6

7

8

.I7 .40 .15 .26 1.oo .49 .49 .43 .12 - .01 .07 .42 .39

- .01 .16

- .01 .07

.08

.oo

.10 .49 1.oo .70 .13 - .06 -.14

.oo

.16 .34 .02 .17 .43 .13 .16 1.oo .21 .09 .16 .42 .I4

.oo

.49 .70 1.oo .16 .17 .02 - .03

.18 .29

.35 .35

9 .05

- .04 .12 .05 .12 - .06 .17 .21 1.oo .85

.07 .4 I .12

10 .02 - .16 .07 - .03 - .01 -.14 .02 .09 .85

11

12

13

.30 .56 .44 .63 .07

.40

.oo

.18

- .03 .I6 .07

.35 .42 .41 .29 .30

.32 .29 .27 .30 .39 .29 .35 .14 .12 .01 .33

1.oo

.oo

.29

1.oo .30 .33

.oo

.01

.28 .20 .23 .42

1.oo

.55

.55 1.oo

TABLE IV INTERCORRELATIONS AMONG INDICANTS OF PERFORMANCE DURING THE 35TH


1

2

1 .oo .71 .94 .I7 .26 .29 .28 .44 .33 .35 .59 .09 .35

.71 1 .oo .72 .93 .24 .13 .13 .39 .15 .23 .59 .03 .35

4

5

6

7

8

9

10

11

12

.77 .93 .84

.26

.29 .13 .I8

.28 .13 .25 .24 .42 .59

.44 .39 .35 .38 .52 .43 .53

.33 .I5 .22 .22 .49 .64

.59 .59 .53 .64 .I6 .39 .41

.09

1 .oo .69 .74

.69 1.00 .92 .48 .26 .23

.35 .23 .22 .25 .53 .66 .60 .74 .92 1.00 .48 .30 .24

3

.94 .72 1

.oo .84 .21 .I8

.25 .35 .22 .22 .53 - .02 .40

MONTHOF RESIDENCE

.24

.08

.21 .20 1 .00 .41 .42 .52 .49 53 .16 .07

.32

.06

.oo

1

.20 .I8 24 .38 .22 .25 .64

.I8

.41 I .00 39 .43 .64 .66 .40 .10 .28

1

.oo .53 -58

.60 .41 .25 .18

.55

.31 .33

.58

.55

.48

.a 1 .00 .45 .41

.03 - .02 .08

.07 .I0

.25 .31 .26 .30 .45 1 .OO - .20

13 .35 .35 .40 .32 .06 .28 .18 .33 .23 .24 .41 - .20 1 .oo


147

increase during the residence period, but some decrease, and others remain fairly stable. Figure 5 illustrates each of these cases, using data from intercorrelation matrixes obtained at all five stages of acquisition. The correlation between the north-south and the east-west sequencing scores for university buildings remains stable and high, the correlation between the free recall scores of the names of university buildings and the east-west sequencing score for university buildings increases steadily, and the correlation between the matching scores for university buildings and the matching scores for city buildings declines. It is useful to consider correlations among indicants as a product of several types of causes. The most obvious of these is the structure of the cognitive system and the interdependencies that this structure imposes among indicants. Once performance has stabilized and the individual indicants are reliable, intercorrelations reflect this structure, and factor analysis can reflect the structure parsimoniously. When scores on individual indicants are unstable, either as the result of learning or forgetting, or because of unreliability of performance, the intercorrelations among indicants are strongly affected. The influence of these changes on intercorrelations is easily illustrated with the present data. Typically, reliability of indicants increases as information is added during acquisition, and, other things equal, this causes the intercorrelations among indicants to increase. This is so because correlations reflect stable, overlap-

0 NORTH-SCUTH SWTIAL ORDER OF CAMWS

- 20

A

u

1001,

75

L A N C M A W LvlD EAST-WEST ORDER MATWING CITY LWDMARKS AND MATCHING CAMPUS LANDMARKS FREE RECALLOF CAMPUS L A N W K S AND SWTIAL ORDER OF CAMPUS LWDM@RC

I

I

8

17

26

35

MONTHS OF RESIDENCE

Fig. 5. Changes of correlations among indicants of performance during four academic years of residence.

148

Harry P. Bahrick

ping elements, and are attenuated by random elements. Thus, during the fourth week of residence, free recall of street names correlates only .17 with free recall of the names of streets which have an east-west orientation. This seems surprising at first, since the east-west score is a component of the total free recall score. The low correlation is explained by the fact that the mean total free recall score is only four streets at this stage of learning. Thus, knowing whether a particular individual recalls three, four, or five street names does not permit a reliable prediction of how many of these streets will have an east-west orientation. The comparable correlation has increased from .17 to .71 by the 35th month of residence because the mean free recall score has risen to 13 and the east-west and north-south subscores have become more reliable. At this point the correlation reflects the interdependencies imposed by the knowledge and scoring system, rather than the unreliability of the scores. In the example given, both indicants increased at a comparable rate during acquisition. The effect on the correlation is quite different when this is not the case. If the acquisition functions for the two indicants differs markedly, correlations between the indicants decrease during acquisition. Thus, the matching score for city landmarks increases throughout the residence period, but the matching score for campus buildings shows little gain after the eighth month of residence. Early during the residence period the two indicants correlate moderately (39, reflecting the fact that fast learners tend to progress more quickly than slow learners with regard to both types of knowledge. This is no longer true at a time when learning continues on one, but not on the other indicant. Fast learners continue to improve their scores on one, but not the other of the indicants and the correlation between the indicants has disappeared (.20). In fact, because performance on matching university buildings shows no improvement after 8 months of residence, but most other indicants do show continued improvement, the intercorrelations of this matching indicant with other indicants generally decline during acquisition. The above examples illustrate that intercorrelations reflect the nature of the knowledge and scoring system as well as individual differences in rates of learning and forgetting. When performance is stable, the intercorrelations reflect primarily the nature of the knowledge and scoring system, but when performance changes, intercorrelations are strongly affected by individual differences with regard to these changes. Correlations are augmented if the rates of gain or loss of information are comparable for two indicants, and correlations are diminished if the rates of gain or loss differ. Because knowledge systems are dynamic, and various aspects of knowledge stabilize at different times, the intercorrelations are almost certain to reflect individual differences in rates of learning and forgetting.


I 49

4. Sex Differences

Investigators of environmental cognition have found little evidence of siknificant differences based upon gender. Although some trends suggesting male superiority were reported by Appleyard (1976) and by Herman et al. (1979) for adults, and by Siege1 and Schadler (1977) for children, most cognitive mapping research has found no significant sex differences (Francescat0 & Mebane, 1973; Maurer & Baxter, 1972; Orleans & Schmidt, 1972). Our own findings strongly support this negative conclusion. We compared acquisition functions for male and female subjects on the basis of all indicants and found no consistent superiority of either sex on any indicant. Of 115 comparisons between the sexes (23 indicants at five stages of acquisition) only eight yielded a significant difference (p < .05), with four comparisons in favor of each gender, and no consistent trend with regard to any individual indicant.

5. Other Correlations between Independent and Dependent Variables Intercorrelations were calculated among all independent variables listed on the questionnaire (Table 11) administered to student subjects, as well as all performance variables. During the first month of residence, students who reported riding a bicycle or working on a job, acquired more knowledge of the city, its streets and landmarks, but not of the campus. After the first month of residence, the use of a bicycle no longer correlated significantly with indicants of knowledge. Surprisingly, the reported frequency of use of maps correlated significantly with performance late, but not early during the acquisition period. Sophomores, juniors, and seniors, but not freshmen, knew more about city streets if they reported frequently having used maps of any kind. The use of Delaware city maps did correlate with knowledge of the city, but few students reported using such maps and, therefore, the overall correlations were low. The use of cars began to have a significant influence on knowledge of the city during the junior and senior years, as expected, since freshmen and sophomores reported very low use of cars. Interpretation of intercorrelations among independent and dependent variables was made more difficult by the fact that the correlations confounded the degree of association between the variables with the extent to which the independent variable applied to the individuals in question. Thus, even though being a tour guide on campus may have a great deal of influence on acquiring information about the campus, the overall correlation was low because only a very small number of students were engaged

Harry P. Bahrick

150

as tour guides. The use of cars accounted for more variance of knowledge late during the residence ptriod, not because they were intrinsically greater sources of knowledge at that time, but because more individuals were involved in driving cars. B.

RETENTION

Average scores were calculated for each indicant of performance for each of the 8 groups of alumni described in Table I. The unadjusted averages for the various indicants are shown in Table V. To facilitate comparison of retention among various indicants, the values are expressed as a percentage of the score on the same indicant attained by senior students in their last month of residence in Delaware. Inspection of Table I shows that the eight alumni groups differed widely with regard to their visits to Delaware during the retention interval, that is, in their exposure to information about the city after the original acquisition period. The frequency and duration of visits per year, for example, is much greater for the extreme time groups (1, 2, 7, 8) than for the middle time groups (3, 4, 5 , 6). Alumni who happened to be available as subjects for this study differed systematically between groups, and these differences are reflected in varying degrees by different indicants. Thus, on the basis of most indicants the unadjusted values in Figs. 6 and 7 show terminal rises, reflecting the fact that subjects in Groups 7 and 8 had more information about the city than subjects in Groups 4 and 5 . Uncontrolled group differences on key variables can, of course, be present in any study of learning and memory, but they are unlikely to be large if the assignment of subjects to groups is random, as is TABLE V UNADJUSTED RETENTION OF EIGHTALUMNI GROUPS

1 1


11.1 6.3 31.2 19.8 25.6 72.4 36.0 27.0 74.8 38.7 9.8 7.5 14.8

2

5.5 3.0 6.5 4.4 20.9 58.6 29.2 22.4 61.4 31.0 6.4 7.1 11.4

3

5.0 2.6 5.3 2.6 19.5 62.7 31.7 21.6 66.3 33.2

5.1 6.7 10.3

4

4.0 2.2 2.5 1.6 13.5 45.4 22.6 17.6 54.2 28.2 3.3 5.9 8.0

5

4.2 2.4 3.1 2.1 12.6 42.7 21.4 18.2 54.5 28.2 3.5 5.3 8.1

6

4.8 2.7 4.0 2.5 11.4 41.7 20.9 16.8 55.6 28.0 3.3 5.5 7.3

7

8

5.5 3.1 6.9 4.3 15.4 43.7 22.3 17.4 51.2 26.1 4.6 6.0 9.7

7.3 4.3 13.3 8.9 12.9 37.9 18.8 16.0 47.2 23.5 5.7 5.9 10.4


WADJUSTED Y W S ADJUSTED YEPNS FOR ZERO VISITS O O E W ADJUSTED YEINS FOR ZERO VISITS GROUP SPECIFIC)

-__

'?\

LOG (TIMEtI)

(YEARS)

Fi 6. Unadjusted and adjusted retention curves for the recall of stre names (from Bahrick, 1979. Copyright 1979 by the American Psychological Association. Reprinted by permission of the publisher.

the case in most laboratory investigations. Even if subjects are assigned to groups on the basis of their a priori status on the independent variable, as is common in naturalistic research, the results need not show systematic

-8 v

-UNADJUSTED MEANS ADJUSTED MEANS FOR ZERO VISITS PO OLEO) -.---ADJUSTED MEANS FOR ZERO VI I T S (GROUP S PEClF IC?

80. \ \

i

LOG (TIMEtl)

(YEARS)

Fig. 7. Unadjusted and adjusted retention curve for the recall of the spatial order of streets (from Bahrick, 1979). Copyright 1979 by the American Psychological Association. Reprinted by permission of the publisher.

152

Harry P. Bahrick

bias. Comparisons of the characteristics of freshmen, sophomores, juniors, and seniors in this investigation, for example, showed some significant differences with regard to variables defining exposure to information, for example, driving a car on campus, but no adjustments to the data were required since group differences attributable to such variation are minor. To obtain estimates of retention relatively free of the known inequalities among the groups, the retention values were adjusted by the method of cross-sectional adjustment (Bahrick, 1979;Bahrick & Karis, 1982). The adjustment estimates are obtained by multiple-regression equations that predict performance for each group on the basis of intercorrelations among independent and dependent variables. Adjustments can be calculated in several ways for the purpose of establishing various types of estimates. Groupspecific adjustments are based upon intercorrelations among independent and dependent variables separately calculated from the data for each time group. This means that the multiple-regressionequations for calculating the adjustment are based on partial-regression coefficients peculiar to that group. Pooled adjustments are based upon intercorrelations among independent and dependent variables calculated for the pooled data from all groups, with indicator variables designating time groups (Neter & Wasserman, 1974). This approach yields common partial regression coefficients for the regression equations applicable to all groups, and adjustments are a function of those coefficients and of the separate intercept values characteristic of the data for each group. The pooled adjustment method entails the risk of systematic error if regression changes systematically across groups, but it has the advantage of minimizing sampling errors, since the partial correlations are based upon a much larger number of observations. Both methods were applied to the present data and for the most part yielded similar outcomes. Figures 6 and 7 illustrate the effect of adjustments for two indicants. The values shown in Figs. 8-10 are based upon the pooled adjustment method. This method was used because there was no evidence of systematic changes of regression across groups and because the estimates of regression based upon all 576 alumni subjects are far more reliable. Table VI shows partial regression coefficients and multiple correlations for two indicants of performance (Bahrick, 1979). The adjusted values in Figs. 6 and 7 were calculated so as to achieve estimates of retention, discounting the effects of all visits to Delaware during the retention interval, and at the same time equating the number of months a car was driven during the years of college attendance. Variables were included in the regression equation predicting a dependent variable if the inclusion significantly (p c .lo)increased the percentage of variance accounted for. Since the adjustments discount the estimated knowledge acquired on visits to Delaware during the retention interval, the adjusted val-


153

0 FREE RECALL

a

MATCH!NG STREET NAMES

0 VIYJALLV CUED STREET NAMES

1

a

z

40-

a

0 20

0

-

,

z 0 + 5

n

I

r 0

0

0

I

I

N

rn

d

2 L

a

U

U

a

m >

(D

I

a

o ,

p

D

b

o

0

a

t

?

LI

=

t Na

m N

LD

”>

I

s

-f

0

LOG ( T I M E + I )

(YEARS)

Fig. 8. Adjusted retention curves of street names.

LOG ( T I M E + I)

(YEARS)

Fig. 9. Adjusted retention curves of the names of landmarks.

Harry P. Bahrick

154

-

'00

-$!

0 A

ORDER OF STREETS IN FREE RECALL ORDEROF VERBALLY CUED STREETS ORDER OF CITY W D M A R K S IN FREE RECALL 0 ORDEROF VERBALLY CUED W D M b R K S ORDER CF CAMRJS UNDMARKS IN F E E

u

80-

Y

w

a 60-

0 -I

a

f

40-

c3 a 0

LOG ( T I M E + I )

(YEARS)

Fig. 10. Adjusted retention curves for spatial order.

ues are lower than the unadjusted ones. The magnitude of the adjustments, however, depends upon the magnitude of the multiple correlation with the dependent variable, and upon the mean scores of each group for the variables included in the regression equation. Large adjustments result when a group made many visits to Delaware during the retention interval, and when scores on the indicant are highly correlated with such visits. Small adjustments are typical if the multiple correlation is low, or if the group made only a few visits. TABLE VI REGRESSIONCOEFFICIENTS AND MULTIPLE CORRELATIONS FOR CALCULATING ADJUSTMENTS AND MAINTENANCE PREDICTIONS'

PARTIAL

(%)

Driving car on campus (months)

Distribution of visits

R

.2646

.0026

.0325

-.7832

.6323

.9169

.0026

.0735

-1.2619

.6035

Frequency of visits

Recency of visits

Duration of visits

Recall of street names

.2163

-.9934

Recall of street sequence

.3892

-1.2373

Visits with car

"From Bahrick (1979). Copyright 1979 by the American Psychological Association. Reprinted by permission of the publisher.


1.

155

Recall of Street Names

Of the 13 street names that graduating seniors can recall, only 4 can still be recalled 3 years and 8 months after graduation. The loss of information is approximately the same for streets running north-south and east-west. The effectiveness of map cues in retrieving street names diminishes even more rapidly. At graduation subjects can retrieve 7-8 street names with the help of the outline map, and only 1 or 2 of those are still retrievable 3 years and 8 months later. After 46 years the visual cues typically yield recall of a single street name, or none at all. Based on the matching-recognition task forgetting of street names is at first somewhat more gradual. About 9 names are correctly matched at graduation, and 3 years later performance has declined about 50%. Ten years after graduation performance stabilizes at 20070, and this is somewhat below the free recall level. Thus, most information about street names is lost within 10 years, and it matters little whether free recall, visually cued recall, or recognition-matching tasks are used.

2. Memory for Buildings and Landmarks The names and locations of campus buildings are forgotten more slowly than the names and locations of streets. Free recall of the names of campus buildings is an approximately linear function of the log of time, and has fallen to about 40% of the graduation level after 46 years. The names of city landmarks are forgotten at a somewhat faster rate; the recall level at the end of 46 years is about 20% of the level at graduation, and performance also declines at an approximately linear function of the logarithm of time. Visual map cues lose their effectiveness at about the same rate as the cues used to generate free recall responses for the names of university buildings. The loss rate is higher for the recall of city landmarks with the aid of visual map cues. The latter function has reached zero retention at the end of 46 years. Retention based on the matching task shows the least decline. Forty-six years after graduation the names of university buildings are matched with locations on the map at 50% of the original level of accuracy, and city buildings are matched at a somewhat lower level. Thus, the matching task yields substantially higher retention than the recall task, even when corrections are made for chance success. Variation of the rate of decline for various indicants is closely related to variation of the corresponding acquisition functions. Steep acquisition functions of constant slope produce rapid rates of forgetting. Negatively accelerated acquisition functions produce comparatively slower rates of forgetting. Street names are learned at an even rate (Fig. 2), with little evidence of negative acceleration, and they are forgotten most quickly (Fig. 8). The acquisition function for the names of buildings is more negatively accel-

I56

Harry P. Bahrick

erated (Fig. 3), so that few items are added during the last two years of residence. This means that a large portion of the material known at graduation has been overlearned, and this higher degree of overlearning diminishes the rate of loss during the retention interval (Fig. 9). Thus, retention losses are greatest for the recall of street names, less for the recall of city buildings, and least for the recall of campus landmarks. The acquisition function for matching college buildings is virtually flat after the first year of residence (Fig. 3), and the corresponding retention function (Fig. 9) shows the slowest rate of loss of information. Street name matching declines to 20% of the graduation level within 10 years (Fig. 8), and the acquisition function (Fig. 2) shows that much of the knowledge is acquired during the last two years of residence. 3. Retention of the Spatial Sequence

The spatial order of street names generated in free recall declines to the level of chance success within 3 years of graduation (Fig. 10). When street names are verbally cued, the accuracy of sequencing declines to the chance level within 10 years. Only the matching task with both visual and verbal cues provided yields evidence of spatial knowledge retained beyond a 10year interval. Spatial knowledge of buildings is preserved much better. Forty-six years after graduation, sequence scores based on city landmarks generated in free recall are still at approximately one-third of the original level and about the same results are obtained for verbally cued city landmarks. The decline of performance is approximately linear with the log of time. Retention is even better for the sequence of university buildings. At the end of 46 years, scores are still at 40% of the graduation level. The previous conclusions relating retention functions to the corresponding acquisition functions apply equally to the spatial data. Retention losses are inversely related to the degree of negative acceleration of the respective acquisition functions. The acquisition functions for street sequences show little evidence of negative acceleration (Fig. 4). This is true for the sequence based on free recall and for the sequence based on cued street names, and this information is lost most rapidly. The acquisition functions for locating landmarks are more negatively accelerated, particularly for verbally cued information and for campus buildings, and this information is retained best. To calculate a rough index of the degree of negative acceleration of each acquisition function, performance after 3 weeks and after 8 months of residence can be expressed as a fraction of the performance attained after 35 months of residence. These indexes predict with fair accuracy the order of retention among indicants 46 years later.


4.

157

Gender Differences in Retention

Mean scores for male and female subjects were calculated for each group, for each indicant of retention, and for each of the independent variables. The results support the conclusion reached on the basis of the acquisition data. There are few, if any, systematic differences between the sexes with regard to the adjusted scores, and sex differences on independent variables account for the minor sex differences in the unadjusted data. Males were more likely to have driven a car on campus; they were more likely to have more visits during the retention interval; their visits were on the average of longer duration and more likely to involve the use of a car. These differences have a cumulative effect, so that unadjusted retention of the street sequence and the spatial order of landmarks is superior for male subjects for the long retention intervals. These differences do not survive adjustments based upon differential regression equations appropriate to the data for subjects of each gender. 5.

Various Projections of Retention Based upon Regression Analysis

The adjusted retention functions discussed thus far estimate retention on the basis of constant conditions of original learning and no opportunities for rehearsal during the retention interval. Thus, the adjustments estimate what each group of alumni subjects would have remembered if they had all resided in Delaware for 36 months (the overall mean), had driven a car for 7.1 months during their residence (the overall mean), and had never visited the city during the retention interval. Before discussing some limitations of this method of estimating long term retention, other types of projection based upon the same method will be illustrated. a. The Projection of Contour Retention Curves. There is, of course, no reason to calculate adjustments only for the average overall value of a variable determining acquisition, or for the case of retention without any opportunity for rehearsal, as was done so far. In order to estimate the effects on retention based on other assumptions concerning a particular variable, it is only necessary to substitute the assumed values of that variable in the multiple regression equation used to calculate the adjustments. For example, the equation can be solved for incremental values of a particular variable, and the resulting adjustments, when added to the unadjusted means, yield contour retention curves such as those illustrated in Figs. 11 and 12. The curves in Fig. 11 estimate recall of the street sequence for individuals who made a single yearly visit to Delaware during the retention interval, with the duration of the yearly visit increased from 1 to 4 days. The lowest function is the one shown in Fig. 8 and is based upon no visits.

158

Harry P. Bahrick

0 0 A 0

2 D A Y VISIT DURATION 3 DAY VISIT DURATION

0

4 DAY VISIT DURATION

NO VISITS I DAY VISIT DUPATION

F o

4

w a

LOG (TIME+I)

F l

f

F

E

w

w

e0

(YEARS)

Fig. 11. Contour retention curves for the recall of the street sequence based on incremental duration of a yearly visit during the retention interval (from Bahrick & Karis, 1982).

100

80

2 W K

0

NO

4

1

VISIl

2

V181lS

A

3

VISITS

0

1 VISITS

VISITS

PER PER

YEAR YEAR

P E R YEAR

PER

YEAR

60

0 0

In J

40

q

I

l3

;2 0 0

0

LOG (TIME+Il

(YEARS)

Fig. 12. Contour retention curves for the recall of the street sequence based on incremental frequency of yearly visits during the retention interval (from Bahrick & Karis, 1982).


I59

Figure 12 shows the estimated effect of increasing the number of yearly visits from 1 to 4, with the duration of each visit held constant at one day. 6. Predictions Regarding Maintenance of Knowledge and Trade-offs among Independent Variables. Instead of estimating the effect of various assumed values of an independent variable on long term retention, it is possible to change the dependent variable and estimate the value of an independent variable necessary to achieve a given level of retention. In this case the question asked is: How much rehearsal or reexposure to knowledge is necessary in order to yield a certain level of retained knowledge? To answer this type of question the desired value of the indicant of retention is substituted on the left side of the regression equation for the prediction of the indicant; arbitrary values are substituted for all other independent variables, and the equation is solved for the independent variable in question. Trade-off functions among independent variables can be determined by changing the arbitrary value assigned to one variable (e.g., assuming a visit duration of 3 days instead of 2 days), leaving all other values of the equation the same, and again solving the equation for the value of the independent variable in question. Table VII reprinted from Bahrick (1979) gives the resulting trade-off estimates among recency and duration of visits for the recall of street names, with frequency of visits per year as the dependent variable. The details of this procedure are discussed in Bahrick (1979) and Bahrick and Karis (1982). c. Variations of the Cross-Sectional Adjustment Method. The estimates for contour retention curves and for maintenance of knowledge were based on the assumption of rectilinear regression and of homogeneity of regression across groups of subjects. If regression varies systematically across groups, then the data for each group must be treated independently, TABLE VII NUMBEROF

VISITS PER YEAR NEEDEDTO MAINTAIN FREERECALLOF STREET NAMESAT THE GRADUATION LEVEL

~

Most recent visit (months)

0

.so 1 .OO

1 .so 2.00

Duration of each visit (days) 1

2

3

4

5

12.57 12.93 13.19 13.39 13.55

8.11 8.34 8.51 8.64 8.74

5.98 6.16 6.28 6.37 6.45

4.74 4.88 4.98 5.05 5.1 I

3.93 4.04 4.12 4.18 4.23

OFrom Bahrick (1979). Copyright 1979 by the American Psychological Association. Reprinted by permission of the publisher.

160

Harry P. Bahrick

and the partial regression coefficients will differ in the regression equation for each group. This option was previously discussed; it is attractive only if the individual groups are large enough to provide reliable estimates of correlations. If the partial regression between variables is not a straight line over the range of values covered, then the regression equation should include quadratic, or higher order terms for that variable, and the inclusion of these higher order terms will significantly increase the proportion of variance that is accounted for on the dependent variable. An alternative is also available to the treatment of the retention interval. The technique that we have described so far assigns subjects to a number of groups in accord with their date of graduation, and the unadjusted means for each group are used as the initial estimate of retention for the interval corresponding to that group. Adjustments are limited to corrections based upon the regression equation, which compensates for inequalities among the groups on the independent variables included in the equation. Alternatively, time can be treated as a separate variable for each subject, and included in the regression analysis. If this is done, subjects are not assigned to groups. Rather, a single regression equation is obtained for all subjects. The equation provides retention estimates for all points in time. The estimates are obtained by substituting the desired values of retention in the term of the equation specifying time and solving the equation for those intervals. This equation can also be used to generate contour retention curves, and estimates of the amount of rehearsal needed to obtain desired levels of knowledge. Advantages and shortcomings of these alternatives are illustrated and discussed by Bahrick and Karis (1982). d. Limitations of Cross-Sectional Adjustments. It is very likely that adjustments calculated by any of the methods described here are conservative; that is, they do not fully correct systematic inequalitiesamong groups. A primary reason for this is unreliability of measurement of the adjustor variables. Such unreliabilities arise because data are frequently unverifiable estimates obtained from each subject, for example, the duration or frequency of visits to Delaware. Such data are subject to the inaccuracies of long-term memory. Errors of measurement, particularly random errors, will lower the intercorrelations among variables, and this in turn will diminish adjustments based upon the magnitudes of correlations. As a result, adjustments are likely to correct only in part for the inequalities among groups. The data in Fig. 7 illustrate this problem. It can be seen that the adjusted curves correct for most but not all of the terminal rise of the curve based on unadjusted data. Predictions for the maintenance of knowledge and for the amount of rehearsal needed to achieve a given level of knowledge are also likely to be affected by this problem and therefore are likely to provide

The Cognitive Map of a Clty

161

estimates of rehearsal that are too large if the correlations between rehearsal variables and retention performance are underestimates. In spite of these limitations, the validity of the method of cross-sectional adjustment is apparent from the high multiple correlations achieved by the regression equations used to calculate the adjustments and from the relatively smooth adjusted retention functions. These adjusted functions are based upon independent data from the various groups, and their regularity compares favorably with the regularity of memory functions obtained from laboratory investigations.

VI.

Summary

Eight hundred fifty-one students and alumni of Ohio Wesleyan University were tested for their knowledge of streets and landmarks of Delaware, Ohio. Students were tested at one of five stages during their period of attendance at the college, and alumni were tested for retention 1-46 years after their graduation. Alumni also answered questions about their visits to Delaware during the retention interval and about other opportunities to relearn their knowledge. The method of cross-sectional adjustment was used to correct retention data statistically for inequalities among subjects in regard to such variables. The tests yielded 23 indicants of knowledge, including free recall, visually cued recall, and matching, and indicants of spatial location of streets and landmarks on the campus and in the city. No hand-drawn sketch maps were used. Instead, subjects indicated their spatial knowledge by arranging streets and landmarks in two separate sequences, corresponding to their relative positions along the north-south and east-west compass dimensions. The acquisition data for most indicants show that more knowledge is gained during the first 3 weeks of residence than during later periods of comparable length. After the first 3 weeks, street names and locations are learned at an even rate of about 2-3 streets per academic year, without diminution at the end of 4 years. This rate is approximately the same for indicants of free recall, verbally or visually cued recall, or matching tests, and includes indicants of verbal and spatial knowledge. Acquisition functions for names of buildings and other landmarks are more negatively accelerated, with little knowledge added during the last two academic years. Learning of the spatial order of streets continues without indication of diminution throughout the residence period. The acquisition function for learning the spatial order of landmarks is more negatively accelerated. Intercorrelations among all indicants of performance were calculated at five stages of learning. When performance has stabilized, these intercorrelations

I62

Harry P. Bahrick

reflect primarily the nature of the knowledge and scoring system, but during the acquisition process the correlations reflect individual differences in the speed of learning. Correlations among indicants increase during the 4 years of acquisition if the indicants show similar acquisition functions, but they decrease if learning continues for one indicant, and not for the other. Performance of male and female subjects was compared on all indicants at all stages of learning, and no consistent sex differences were found. Subjects in this investigation learn much more about landmarks and buildings and less about the network of streets than subjects in other investigations of acquisition of information about cities. This reflects the needs and experiences of college students, who spend much time walking over a relatively small area of city and campus, and relatively less time driving cars over a wider area. It also reflects the fact that the streets of Delaware are not arranged in a numerical or alphabetical sequence, which facilitates rapid learning and retrieval. The rate of forgetting based upon various indicants is closely related to variation of the corresponding acquisition functions. Steep acquisition functions of constant slope produce rapid rates of forgetting. Negatively accelerated acquisition functions produce comparatively slower rates of forgetting. Street names and the spatial order of streets are forgotten quickly; the names of landmarks and buildings are forgotten more slowly. This is consistent with the established principle that overlearning retards forgetting, since the degree of negative acceleration of the acquisition function is associated with the degree of overlearning. The method of cross-sectional adjustment is used to project estimates of the effects of various degrees of rehearsal (e.g., visits to the city during the retention interval) on long-term retention. The estimates are obtained by substituting incremental values for a given rehearsal variable in the multiple regression equation predicting memory. The result is a family of contour retention curves reflecting the estimated effect of changes in one rehearsal variable on the particular indicant of retention. By changing the values of more than one rehearsal variable in the multiple regression equation, tables estimating trade-off effects among the variables are produced.

ACKNOWLEDGMENTS This research was supported by U. S. Public Health Service Research Grant HD00926-15 from the National Institute o f Child Health and Human Development. The author wishes to express appreciation to Phyllis Bahrick and Melva Hunter, who supervised the collection of data, and to the following individuals who assisted in the collection and analysis of data: Scott Henkle, Mitzi Leedy, and Tom Rollins.


I63

REFERENCES Appleyard, D. Planning a pluralistic city. Cambridge, Massachusetts: MIT Press, 1976. Bahrick, H. P. Maintenance of knowledge: Questions about memory we forgot to ask. Journal of Experimental Psychology: General, 1979, 108(3), 296-308. Bahrick, H. P. Memory for people. In J. Harris (Ed.), Everyday memory, actions and absentmindedness. New York: Academic Press, 1983. Bahrick, H. P., Bahrick, P. O., & Wittlinger, R. P. Fifty years of memory for names and faces: A cross-sectional approach. Journal of Experimental Psychology: General, 1975, 104(1), 54-75. Bahrick, H. P., & Karis, D. Long-term ecological memory. In R. Puff (Ed.), Handbook of research methocis in human memory and cognition. New York: Academic Press, 1982. Baird, J. C., Merrill, A. A,, & Tannenbaum, J. Studies of the cognitive representation of spatial relations: 11. A familiar environment. Journal of Experimental Psychology: General, 1979, 108, 92-98. Blaut, J . M., & Stea, D. Mapping at the age of three. Journal of Geography, 1974,73, 5-9. Devlin, A. S. The “small town” cognitive map: Adjusting to a new environment. In G. Moore & R. Golledge (Eds.), Environmental knowing. Stroudsburg, Pennsylvania: Dowden, Hutchinson & Ross, 1976. Evans, G. W. Environmental cognition. Psychological Bulletin, 1980,88, 259-287. Francescato, D., & Mebane, W. How citizens view two great cities: Milan and Rome. In R. Downs & D. Stea (Eds.), Image and environment. Chicago: Aldine 1973. Golledge, R. G. Methods and methodological issues in environmental cognition research. In G. Moore & R. Colledge (Eds.), Environmental knowing. Stroudsburg, Pennsylvania: Dowden, Hutchinson & Ross, 1976. Herman, J. F., Kail, R. V., & Siegel, A. W. Cognitive maps of a college campus: A new look at freshman orientation. Bulletin of the Psychonomic Society, 1979, 13, 183-186. Lynch, K. The image of the city. Cambridge, Massachusetts: MIT Press, 1960. Maki, R. H. Categorization and distance effects with spatial linear orders. Journal of Experimental Psychology: Human Learning and Memory, 1981,7 , 15-32. Maurer, R., & Baxter, J. C. Images of the neighborhood and city among black, anglo, and MexicawAmerican children. Environment and Behavior, 1972,4, 351-388. Neter, N., & Wasserman. W. Applied linear statistical models. Homewood, Illinois: Irwin, 1974. Orleans,P., & Schmidt, S. Mapping the city: Environmental cognition of urban residents. In W. Mitchell (Ed.), Environmental design: Research andpractice. Los Angeles: University of California Press, 1972. Pearce, P. L. Mental Souvenirs: A study of tourists and their city maps. Australian Journal Of PSyChOlOgy, 1977, 29, 203-210. Siegel, A. W., & Schadler, M. Young children’s cognitive maps of their classroom. Child Development, 1977, 48, 388-394. Siegel, A. W. & White, S. H. The development of spatial representations of large-scale environments. In H. W. Reese (Eds.), Advances in child development and behavior (Vol. 10). New York: Academic Press, 1975. Tolman, E. C. Cognitive maps in rats and men. Psychological Review, 1948,55, 189-208.


James F. Voss. Terry R . Greene. Timothy A . Post. and Barbara C. Penner UNIVERSITY OF PITTSBURGH PITTSBURGH. PENNSYLVANIA

I . Introduction .......................... ........................... I1 . The Information-Processing Model ..................................... A . Task Environment ................................................ B Problem Space ................................................... C . Problem Representation. .......................................... D Problem Solution Activity ............................. E . Evaluation ...................................................... Ill . Social Science Problems and Their Solutions . A . General Characteristics of Social Science Problems . . .......... .......... B . Specific Considerations of Social Science Problems . . IV . The Problem-Solving-Reasoning Model ................................. ..... ............. V . Protocol Collection and Analyses ......... A . Participants ..................................................... B The Collection of Protocols .................... .............. C . Protocol Analysis ................................................ D . Expert Protocols .... .......................................... E Novice and Postnovice Protocols ................................... F Graduate Student Protocols ........................................ G Nonexpert Expert Protocols ...... ............................. H Chemist Protocols ........................................... 1 Miscellaneous Protocols ................................. ..... v1. The Acquisition of Social Science Problem-Solving Skill . . . . . . . . . . . . . . . . . . . VII . General Considerations ................. A Task Environment ................. B Problem Space ............................... C Problem Representation ........................................... D Solution Activity ..................................... E Evaluation ............................................ VIII . Concluding Remarks . ..................................... References ........................................

. .

.

. . . . . . . . . .

I

.

165 166 166 167 167 167 168 168 I68 169 171 173 173 174 174 175 193 195 197 201 202 203 208 208 209 209 210 211 211 212

Introduction

In recent years problem solving has been studied in the context of particular subject matter domains such as physics (e.g., Larkin. McDermott. THE PSYCHOLOGY OF LEARNING AND MOTIVATION. VOL . 17

165

Copyright 0 by Academic Press. Inc . All rights of reproduction in any form reserved. ISBN 0-12-543317-4

166

1. F. Voss, T. R. Greene, T. A. Post, and B. C. Penner

Simon, & Simon, 1980), geometry (e.g., Greeno, 1978), and software design (e.g., Jeffries, Turner, Polson, & Atwood, 1981). The present research is concerned with problem solving in yet another domain, social sciences. There are a number of reasons for studying social science problem solving. One is to examine how recent theoretical developments in problem solving may be extended to the social sciences. A second and related reason is that problems of the social sciences tend to be ill structured, and the study of solving such problems has received relatively little attention (cf. Reitman, 1965; Simon, 1973). Finally, many problems of everyday life, for example, economic and political, are in the social sciences, and a better understanding of problem solving in such domains may help to improve instruction in these areas. Within the social sciences, we have concentrated on problems in political science, and especially on problems related to the Soviet Union. While we have employed a number of different problems, our most extensive analyses have been conducted upon the Soviet agriculture problem. In this problem we indicate that Soviet crop productivity is not able to meet the country’s needs, and we ask the solver what he or she would do to increase agricultural productivity in the Soviet Union, given that the solver is the Head of the USSR Ministry of Agriculture. The methodology that we have employed is protocol analysis, collecting the protocols from individuals varying in training and experience. The organization of this article is as follows. The present research was conducted within the framework of the general information processing model of problem solving, and the model is briefly described (cf. Newell & Simon, 1972; Simon, 1978). Next, there is a discussion of some characteristics of social science problems. Subsequently we present a description of the more specific model of the solving process we have developed as a result of this research, and this is followed by a summary of the protocols that we have collected. The final sections present some ideas regarding how social science problem-solving skill is acquired and how the present work is related more generally to problem-solving research.

11. The Information-Processing Model A.

TASKENVIRONMENT

The task environment is the problem statement and the context in which the statement is presented, that is, the “objective” statement of the problem as found under particular conditions.

Problem-Solving Skill in Socinl Sciences

B.

167

PROBLEM SPACE

When presented with a problem, the solver establishes a problem space that consists of the information known or potentially available to the solver that may be useful in solving the problem. This information typically includes the problem goal and subgoals, as well as the possible states of the problem that may occur as the problem solver moves toward the solution. In addition, the problem space contains operators which enable the individual to move from state t o state. Finally, the problem space contains the solver’s knowledge of the constraints under which a problem is to be solved. Thus, the solver is assumed to begin at an initial problem state and move along a solution path from state to state until the goal state is reached (Simon, 1978). Furthermore, this movement is assumed to be constrained by the short-term memory limitations of the individual. Such limitations are assumed to limit the number of problem components that the solver is able to consider at any point in the solving process. C.

PROBLEM REPRESENTATION

When presented with a problem, the individual develops a problem representation which, in a sense, is the solver’s interpretation of the problem statement. For a number of problem classes, for example, puzzle problems (Greeno, 1974; Simon, 1975), the solver’s representation consists of little more than an understanding of the problem statement. However, as shown by Larkin et al. (1980), problem representation is generally more complex for physics problems. In this case, the solver must interpret the problem statement by studying the features of the problem and their interrelations. Thus, a novice tends to use the problem’s surface information to form a representation, while the expert’s representation is usually based upon the conceptual relations underlying the surface statement (cf. Chi, Feltovich, & Glaser, 1981). Moreover, the expert’s representation may include classifying the problem, for example, as one of “mechanics.” Thus, the expert is able to represent the problem appropriately because of the knowledge and experience in dealing with that particular type of problem.

D.

PROBLEM SOLUTION ACTIVITY

While general problem solving may be viewed as stepwise movement toward a goal, a more complete understanding of the solving process must include an understanding of the more specific strategies that may be employed. Sacerdoti (1977), for example, developed a model that consists of decomposition of the problem into a number of subproblems, and solutions are sought for the subproblems. The decomposition occurs via a top-down.

168

J. F. Voss, T. R. Creene, T. A. Post, and B. C. Penner

breadth-first process, in the sense that the individual specifies the major subproblems and subsequently solves them. The solver may work on the subproblems iteratively, refining the subproblem solutions with each iteration (cf. Jeffries et al., 1981; Sacerdoti, 1977). Other descriptions of problem solving strategies have been developed by Hayes-Roth and Hayes-Roth (1979) and Stefik (1981). The primary issue of concern in these models is subproblem interaction. The concept of opportunistic solving has been offered by Hayes-Roth and Hayes-Roth, and a somewhat similar notion, constraint posting, has been suggested by Stefik. Opportunistic solving refers to taking the opportunity of solving one subproblem while the solver is in the process of solving another subproblem. Similarly, constraint posting refers to not solving a subproblem until it becomes necessary to deal with a constraint during the course of solving the subproblem.

E. EVALUATION In the information-processing framework, evaluation of the steps taken during the solving process is often assumed to take place via means-ends analysis. In this method, a step is evaluated in terms of whether it moves the solver closer to the goal. Indeed, the solver may look at the difference between the current state and the goal and then try to select which step reduces this difference. Hence, this process is sometimes referred to as working backwards. Another form of evaluation is the generate-test process in which the solver may generate a number of solutions and evaluate (test) the solution via the application of some criterion. On the other hand, Larkin et al. (1980) have shown that once physics experts are able to classify a problem, they go through the solution steps in a routine manner. Because the solution is reasonably stereotypical and does not usually require evaluation, the process is referred to as working forward. In summary, for relatively complex problems, the solver develops a problem representation and then offers a solution to the problem and a number of strategies have been proposed to describe how this takes place. Moreover, as the solution activity is taking place, the steps taken are often evaluated, except in cases where working out the solution is a routine process. 111. Social Science Problems and Their Solutions

A. GENERAL CHARACTERISTICS OF SOCIAL SCIENCE PROBLEMS Many social science problems, including those employed in the present research, involve the existence of an undesirable state of affairs (the prob-

Problem-Solving Skill in Social Sciences

169

lem) that requires improvement (the solution). Thus, a high crime rate requires reduction, an excessive inflation rate should be decreased, agricultural productivity is below the nation’s needs and must be increased, etc. This type of problem may also be found, of course, in other fields such as medicine. In general, solving such problems consists of isolating the cause(s) of the problem and eliminating or at least reducing the effects. In social sciences, however, solving such problems is often made quite difficult by two factors, namely, the lack of agreed-upon solutions and delay in implementing solutions. 1.

The Lack of Agreed-upon Solutions

While virtually all problems used thus far in problem-solving research have one o r more known solutions, social science problems seldom have solutions about which experts are in complete agreement. An effect of such lack of agreement is that the solver must provide arguments supporting why the particular proposed solution should be adopted. Indeed, as shown later in this article, such argument development is a major part of social science problem solving, especially for experts. 2.

The Delay of Implementation

In most social science problem solving there is a relatively long delay from the time a solution is proposed and accepted to when it is fully implemented; for example, an economic policy designed to reduce inflation may take years to implement. Moreover, the effects of such implementation delay are indeed profound. First, relatively few solutions may be tried for any given problem because of time, cost, and other factors. Second, the opportunity to try so few solutions contributes t o the relatively slow development of the data base of social sciences. Third, the conditions existing when the solution is proposed may change during the course of implementation. Naturally, a good solution anticipates changes in conditions, but anticipation can be quite difficult. Finally, implementation delay affects solution evaluation, a point discussed in more detail in the next few pages.

B. SPECIFIC CONSIDERATIONS OF SOCIAL SCIENCE PROBLEMS 1.

The Goal

The goal of many social science problems is often vaguely stated, for example, “reduce the crime rate,” “decrease inflation,” and “improve crop productivity.” Moreover, what constitutes an “acceptable” goal, for example, how much inflation needs t o be reduced in order to consider the

I70

J. F. Voss, T. R. Greene, T. A. Post, and B. C. Penner

problem “solved,” can be a point of considerable controversy (and a political football).

2. Causes, Constraints, and Subproblems As previously noted, solving the type of problem employed in the present work typically involves isolating cause(s) of the problem, and this strategy raises questions about causation that are beyond the scope of this article. For our purposes, however, causes are viewed as factors that play a role in producing the problem, and these factors are taken to be subproblems and constraints. Subproblems are typically regarded as problems that are subordinate to a more general problem. Constraints, on the other hand, are usually regarded as factors assumed to be invariant over the course of solving a particular problem that in some way restrict the range of solutions. In the present research, for example, the lack of peasant knowledge regarding how to use technical farm equipment is typically viewed as a subproblem that could be solved by education. Soviet ideology, on the other hand, is viewed as a constraint because solving the Soviet Union’s agricultural problem must be done within the framework of the Soviet system, and it is generally assumed that there is no chance of changing that system. However, a factor identified as constraint early in a protocol may be converted to a subproblem if the subject attempts to provide a solution. In principle then, while one factor may be assumed to be invariant over the course of solving the problem, it is possible that any constraint could be subsequently interpreted as a subproblem. (Soviet ideology could be attacked via shifts toward capitalism or even revolution.) The distinction of subproblem and constraint in social sciences thus appears to be less definite than in other areas. For example, the operation of physical laws can only serve as constraints in physics problem solving. Finally, we note that in social sciences constraints and subproblems are usually not given in the problem statement, and the solver must rely upon his or her knowledge of the field to identify these factors.

3. Evaluation

The previously mentioned implementation delay found in social science problem solving has an important influence upon the evaluation process. Specifically, after proposing a solution, the solver (especially the expert): (1) may provide support for the proposed solution, (2) may isolate subproblems that need to be considered in order for the proposed solution to be implemented and may offer solutions to these subproblems, and (3) may evaluate any of the solutions, with one or more constraints serving as the


171

evaluation criterion. Thus, social science problem solving includes argument development which involves “building a case” for a proposed solution. The account presented thus far suggests that social science problem solving is an integration of two processes, problem solving and verbal reasoning. In an early attempt to describe the solving process, Voss, Tyler, and Yengo (1983) used an extension of a model of argument, that is, a jurisprudence model (Toulmin, 1958; Toulmin, Rieke, & Janik, 1979). The model described by Voss and his co-workers had deficiencies, however, the most serious of which was the lack of a problem-solving control structure. The model presented in this article takes this deficiency into account by assuming two structures, a problem-solving control structure and a reasoning structure, each with its own set of operators.

IV.

The Problem-Solving-Reasoning Model

The model consists of a problem-solving control structure (G) and a reasoning structure (R). Here, G is viewed as a goal structure which controls the problem-solving process. It consists of operators that act upon the individual knowledge base and generate the problem solution. The G operators are presented in Table I. The first operator is State Constraint (GCON). This operator is applied when a solver explicitly or implicitly indicates a problem constraint. The second operator is State Subproblem (GSUB), applied when the solver indicates that a particular factor is being used as a subproblem. The third operator, State Solution (GSOL), is applied when the solver states a solution, either to the given problem or to a subproblem, explicitly or implicitly expressed. The three operators stated thus far constitute the “hard-core” operators, which in some form or another may be found in most descriptions of problem solving. The remaining four G structure operators are TABLE I

G STRUCTURE OPERATORS GCON GSUB GSOL GIPS GSUP GEVA

GSUM

State constraint State subproblem State solution Interpret problem statement Provide support Evaluate Summarize

172

J. F. Vow., T. R. Creene, T. A. Post, and B. C. Penner

supportive to these three operators and are used in conjunction with Reasoning Structure operators. The operator Interpret Problem Statement (GIPS) is applied when the solver considers how the problem is to be interpreted. The Provide Support operator (GSUP), for reasons described below, is applied only when the solver is using some type of argument to support the existence of a subproblem or constraint. The Evaluate (GEVA) operator is applied when (1) the solver develops an argument that supports or rejects a solution, and (2) when the solver evaluates a solution in relation to a particular constraint, Thus, GEVA is used in the first case rather than GSUP because evaluation is provided which is not necessarily supportive. Furthermore, to distinguish support from positive evaluation under such circumstances is virtually impossible. Therefore, we used GSUP only when supportive evidence is stated for the existence of a subproblem or constraint and GEVA when an evaluation of a solution occurs. The Summarize (GSUM) operator is applied when the solver presents a summary of a relatively large portion of the protocol. It seems that GSUM has an integrative function in that the solver is stating, “Here is where I am so far.” The R Structure operators are listed in Table 11. These operators are applied in conjuction with the GIPS, GSUP, GEVA, and GSUM operators. Typically, the application of the R structure begins with an argument made by the solver, that is, the application of the RARG operator of Table 11. Subsequently, a combination of the remaining operators is applied in the argument development. The State Assertion operator (RSAS) is applied when the solver refers to a constraint, subproblem or solution, but the factors referred to are not in the goal structure of the solution process. This occurs primarily when the solver refers to one of these factors as part of argument development, not

TABLE I1 R STRUCTURE OPERATORS RARG RSAS RFAC RPSC RREA ROUT RCOM RELA RCON RQUA

-

State argument State assertion State fact Present specific case State reason State outcome Compare and/or contrast Elaborate and/or clarify State conclusion State qualifier

Problem-Solving SLul in Social Sciences

173

actually dealing with the factor as part of his or her proposed solution. The State Fact (RFAC) operator is used when the solver supports another statement via the statement of a fact. The Present Specific Case (RPSC) operator is used for stating a specific case or example which demonstrates the contents of a previous statement. The State Reason (RREA) operator is applied when the individual states a reason for a previous statement. This operator often involves use of terms such as “because.” The State Outcome (ROUT) operator is applied when the solver states an outcome of a previous statement. The Compare and/or Contrast (RCOM) operator is applied when the solver compares a previous statement with some other entity related to the statement. The Elaborate and/or Clarify (RELA) operator is used when the solver attempts to elaborate or clarify a previous statement while essentially not adding anything new. The State Conclusion (RCON) operator is employed when a concluding statement is provided after a series of previous statements. It typically terminates a line of argument and is more specific than the GSUM operator. The State Qualification (RQUA) operator involves the use of a statement to restrict the range of application of the previous statement. Finally, we would note that if one were to develop a simulation model involving the use of the G and R structure operators, it would be necessary to derive a much more precise statement for the input and output operator conditions. Our analysis, however, had as its goal a reasonable exposition of the solving processes found in the protocols. V. A.

Protocol Collection and Analyses

PARTICIPANTS

Six experts and ten novices participated in the study. The experts included five faculty members of the University of Pittsburgh and one person finishing his dissertation who is now on the Ohio State University faculty. Six of the novices were individuals in a course on Soviet domestic policy. These individuals were given the Soviet agriculture problem both at the beginning and at the conclusion of the course, the latter responses being referred to as postnovice protocols. The purpose of repeating the protocol collection was to get an idea of what effect the course may have had upon the solving process. Four additional novice protocols were obtained from individuals attending another section of the same course. In research using the constrastive method, experts and novices typically differ in a number of ways in addition to their acknowledged difference in expertise. In an effort to gather data that would help to “unconfound” some of the expert-novice differences, we obtained protocols from three first or second year graduate students in political science whose field of

174

J. F. Vow, T. R. Greene, T. A. Post, and B. C. Penner

interest was the Soviet Union. We also obtained protocols from four advanced political science graduate students whose field of expertise was Latin and/or South America. We obtained three protocols from faculty whose field of expertise was Latin America or American Government and Policy and we also obtained four protocols from faculty of the Chemistry Department. Finally, we obtained one protocol from a foreign service officer of the State Department whose service is in South America and one protocol from a visiting scholar of an Eastern European country. With these protocols we thus were able to compare not only expert and novice performance, but were able to compare protocols of: (1) experts on the Soviet Union with individuals who were not experts on the Soviet Union but whose field of expertise was in the same general discipline, (2) experts on the Soviet Union t o that of experts of a different discipline (chemistry), and (3) undergraduate, graduate, and faculty within a given discipline, and thereby use a cross-sectional approach t o develop some ideas regarding how expertise develops.

B. THECOLLECTION OF PROTOCOLS Except where noted, the problem given to the participants was the previously stated Soviet agricultural problem. There were two variations of the problem. One version was as follows: “Suppose you were the Minister of Agriculture in the Soviet Union and assume that crop productivity has been low over the past several years. You have the responsibility to increase crop production. How would you go about solving this problem?” The second added that because of political reasons, “productivity had to be increased as soon as possible.” In addition one expert, the foreign service officer, and the European scholar received probes during the course of generating the protocol. This procedure was used to determine whether they knew more problem-related information than they utilized during the generation of the protocol. Individuals were instructed to “think out loud” while generating their solution, being encouraged to state “whatever comes into their head.” A tape recorder was started and the participant was given the problem on a sheet of paper. In some cases individuals were allowed to use pencil and paper to make notes, but in only one case was there significant use of these materials. C.

PROTOCOL ANALYSIS’

The first group of protocols we gathered, which included four experts and the six novice-postnovices, were initially divided into idea units. Typ‘Copies of the original protocols and the analyses of the protocols are published in an LRDC Series publication by Penner and Voss. This publication is available upon request.

Problem-Solving SkUl in Social Sciences

175

ically, a unit was a sentence, but on occasion it was of clause length or a number of sentences. The interrater reliability of segmenting units was in excess of .90.For the remaining protocols, we used a more “top-down” method, analyzing the protocols in relation to the previously mentioned G and R structures, but again segmenting the protocols in relation to assertions germane to the problem solving process. When the first group of protocols was analyzed in the same manner, the two methods led t o equivalent segmentation. Finally, in analyzing protocols, every effort was made to stay close to the protocol contents, and contents were discarded only rarely and only when they were not related to the solution process.

D. EXPERTPROTOCOLS Expert A began with an explicit statement of constraints: 1 think that as Minister of Agriculture, one has to start out with the realization that there are certain kinds of special agriculture constraints within which you are going to work. The first one, the most obvious one, is that by almost every count only ten percent of the land in the Soviet Union is arable. This is normally what is called the Blackland in the Ukraine and surrounding areas. And secondly, even in that arable ten percent of the total land surface. you still have climate for instance, problems over which you have no direct control. Okay, so that is sort of the overall parameter in which we are working.

Subsequently, Expert A analyzed the problem from a historical perspective, concluding that inadequate technological modernization is a major factor in low productivity, thus representing the problem as technological. Table 111 presents the historical analysis in relation to the R structure. Immediately thereafter Expert A stated his basic solution, greater capital investment in agriculture: I think as a starting point as Minister of Agriculture, my first aim would be to get monies to invest in this further mechanization and further application of scientific techniques of agriculture, to agriculture situations.

This statement was followed by argument supporting the solution. The argument is focused upon specific problems that by implication would be solved by greater investment: Even though we have mechanized to some extent, it has been a rather crude form of mechanization, it has been rather low-level, it is not coherent, it is not consistent. We have the same old problem that we’ve had. If we develop tractors or we produce tractors, we produce a thousand tractors, we have no parts to service them when they break down. We don’t have adequate transportation supplies or transportation networks to carry the produce we do have to the urban markets. We have been woefully lacking in a methodical application of fertilizers to our agricultural sectors in society. We are much more like a third world than an industrial world in terms of the way, the lack of use of

176

J. F. Voss, T. R. Greene, T. A. Post, and 9. C. Penner

TABLE 111 EXPERTA’S R

STRUCTURE FOR

HISTORICAL ANALYSIS SEGMENT

(GIPS) Interpret problem statement (RARG) Historically, agriculture has been a problem in the Soviet Union (RFAC) Problem has been inherited from the time the czars freed the serfs (RFAC) Agricultural production was low even before then (RREA) Historically, the aristocracy had no need to fend for itself (RCOM) Was not like English aristocracy (RPSC) Never introduced modern methods of fertilization (RPSC) Never went to enclosures or consolidation of land (RPSC) Never experimented with crop rotation (RFAC) Agriculture problem was passed onto peasants so they could do what they willed with the land (ROUT) They responded with old, inefficient ways (RFAC) USSR had three different policies to increase agricultural production (RPSC) Exhortation (RELA) Campaign for more effort on the part of the peasants (RCON) Was waste of time and energy (RREA) Only gave the party a sense of false importance (RREA) It is encumbant upon the party to develop these campaigns (but they haven’t paid o f 0 (RREA) Party believes that ideological policies can overcome objective limitations (RCON) I would not use exhortation (RPSC) Reorganization (ROUT) Leads t o confusion, mismanagement. makes peasants laid back (RPSC) Reorganized collective state farms, machine tractor stations (RPSC) Now have agroindustrial complexes, reducing number of collectives and making farmer a wage earner (RQUA) Have always allowed private crop to exist (RQUA) Except in stringent ideological periods Is taken to be a more primitive form of production (RELA) Private crops account for 40% of food staples (RFAC) (RPSC) Mechanization (RCON) This is where I would start my solution (RELA) Have tried to mechanize agricultural production more (RELA) Tried t o introduce scientific advances in agricultural production

fertilization. We still don’t have very scientific management in terms of crop rotation and because of all this, we still have a rather labor intensive agricultural production system and therefore, production per unit is very marginal.

An important point is that the subproblems, for example, inadequate transportation system and poor management of crop rotation, are regarded as subordinate to the more general lack of technological development, and it is further implied that these problems can be resolved by the solution of capital investment. As such, the subproblems are not basic in the solution


I77

structure. Instead, the solution to these subproblems is being used by Expert A to support his more abstract solution. This is a type of reasoning commonly employed by experts, that is, showing how a more general solution will act to solve a number of subordinate subproblems. In the next part of the protocol, Expert A stated a subproblem that he would encounter, namely, there would be a need t o reorder financial priorities and this would mean interacting with other agencies. The primary line of argument used by Expert A is to convince those in power to reorder priorities so that agricultural development will lead to less dependency on the West and a more stable regime: I would have to fight very strongly in the party and the government to redirect the investment ratio going to agriculture over heavy industry, and even light industry from the way it has been in the past. Though we have in the last ten years shifted investment policies so that more and more is coming into agriculture and away from industrial production per se, it still is not anywhere near the break even point even though for both international, political and domestic reasons, agriculture is our “Achilles Heel.” If we simply don’t produce enough we will become more and more vulnerable to dependency upon the West for agricultural production, and we will, if we don’t have enough food stuffs internally, we will develop a more unstable regime. There will be lack of support. We have developed a program where the support from the people is going to come at least in some basic way from demand satisfaction and yet we haven’t been able to satisfy a lot of the basic food demands of large parts of our population. So, I would first have to fight in the decision-making circles to once and for all, recognize that the construction of socialism is the construction of an entire society that is self-sufficient and does not have to depend upon, especially potential adversaries, for crucial parts of its basic resources. That means that we have a high enough industrial base, we have the war technology equivalent to the West, we must now redirect in a major way, our investment policy toward agriculture, toward mechanization, toward the infrastructure around mechanization, toward the transportation program, toward the plastic bag program so that whenever the damn fertilizer is packaged it doesn’t sit out in the lot as it does. We lose one-half of our fertilizers because it rains on paper packages. What I mean by infrastructure, there is a whole series of secondary enterprises that we have to make.

Finally, Expert A defined a number of subproblems that should be considered in relation to the proposed solution. These include the need for education, the need to improve wages and incentive, the need to make private plots a more integral part of the economy, and the need for a modified market mechanism. A solution is provided for each of these issues and each solution also involves agrument to support it. Figure 1 presents a diagram of Expert A’s G structure. Brackets indicate that the protocol contained R structure information which was used to support the contents of the node having the brackets. As shown, Expert A’s protocol contained extensive reasoning to support various phases of the problem solving activity. Also, a dashed line depicts an implied subprob-

piGGi-1 STATEMENT

e 7 10% arable

]+FO&

Achilles Heel

I

Invest in mechanization & technolosv

?I[

I

mechanization

GSUB ---- 4--GSUB --7 I--------wags too low I Existenceof 1 I private plots I ~~

r----

,

I

+

GSOL

Develop educational programs

I

temporary

I

Raise wager and other benefits more a part of

I

A

GSUB

Marginal increasing arable land

sibly arable land and

GEVY Will work as

Fig. 1. G structure of Expert A.

A

GSUB

Market mechanism insufficient

Make collectives more rewarding

[ e l By increasing


179

lem, for example, Expert A’s statement, “We will need to educate the peasants,” implies that lack of peasant knowledge is a subproblem. We note the following additional points regarding Expert A’s protocol: 1. In discussing increased wages he made the statement “ideology aside” and discussed the need for agricultural workers to receive the same benefits as industrial workers. However, he did not fully explore how he may be violating the ideology constraint in his solutions involving an increase in wages, making the private plot a greater part of the system, and modifying the market mechanism. 2. Expert A provided an example of converting a constraint into a subproblem when initially he stated limited arable land as a constraint but later considered increasing the amount of arable land via irrigation. 3. Expert A’s protocol depicts a problem-solving strategy in which the given problem was converted into a problem for which a solution may be offered; that is, the stated problem was converted to inadequate technological development and the solution proposed was greater agricultural investment.

Expert B began with a strong statement of the political constraint: 1 think the most difficult part of this sort of problem for a Soviet leader is that the objective is not as clear as your question hints, because solving the productivity of agriculture in technical terms, even in economic terms, is no big problem, given that you can get the land fertile and can rotate it. But in the Soviet Union all these economic problems are political issues and so the objectives are not at all clear because you can not raise agriculture productivity in the most simple fashion without damaging certain political priorities that you will not be able to ignore.

Subsequently, Expert B, in a quite interesting manner, suggested a number of solutions to the problem that could be offered, if it were not for the political constraint. (Among these is Expert A’s solution, which Expert B rejected.) Table IV presents a description of the R structure of this analysis. From this analysis, Expert B developed his problem representation, which was stated immediately after the conclusion of the Table IV contents: So our problem is really not technological, but it is social-how to make labor work harder on the collective, as opposed to the private plot.

A general solution is subsequently proposed, followed by a statement of support: Maybe one way we could do it is to make the incentives in the collective really competitive with the incentives on the private plot. As it is now, people on the private plot

TABLE IV. SAMPLE OF EXPERT B’s R STRUCTURE (GIPS) Interpret problem statement (RARG) Agriculture problem would be easy to solve if it were not for politics involved (RSAS) Could demand finances for agriculture (RSAS) Could use money for technology (RPSC) Fertilizer (RPSC) Machinery (RPSC) Heaters (RSAS) Could use money for training (RCON) These solutions are technically okay, but politically bad (RCON) Therefore, cannot use them (RSAS) Could use money to industrialize agriculture (ROUT) Would kill off jobs, creating unemployment (RREA) Force people to move to the city (RREA) Cities can’t absorb the people (ROUT) Would create an even more serious social problem (RCOM) This is problem in advanced industrialized countries ’ (ROUT) Need to keep labor in countryside (RREA) Cities can’t absorb them (RREA) Not enough industry jobs (ROUT) Create large unemployed mass (RCOM) This is alright for country like Turkey (RCON) We can’t do that (RREA) Nationalism too strong (RREA) Can’t ignore peasants (RREA) Already have high unemployment Invisible though (RREA) People have senseless jobs Don’t want to drive it up higher (RREA) Cities overcrowded already (RREA) Can’t build cities overnight (RCON) This is why we have tried to build up industrial base to siphon off agricultural labor (RSAS) Could allow for incentive system (RSAS) Would have to prevent capital accumulation (RREA) Increased incentives allow some to get rich (RSAS) If rich can buy land it results in problems (RREA) Few people needed in countryside if aggregates exist (RREA) Creates serfdom and poor (RSAS) Could allow incentives on collectives only (ROUT) Would increase incentives on collectives only (RSAS) Have problem of where to get resources for incentives (RSAS) Can’t raise price of food (RREA) Tenet of socialist system to provide food for everyone so there is no starvation (RFAC) Real achievement of our system (RCON) Need t o increase incentives. but not raise food prices (RSAS) Could let (RSAS) (RSAS) (RFAC)

private plot become more important source of agricultural output Have same problem of allowing some to get very rich Undermines effort to make collectives the area of real productivity Private plot doesn’t take advantage of economies of scale

(RSAS) Take advantage of labor and keep it on the farm

Problem-Solving SkUl in Socinl Sciences

181

can take their goods and sell them at the bazaars around the countryside and sell them at much, much higher prices than the collective earns and therefore make much more money. Therefore, it is in their interest to not work too hard on the collective to keep the supply of food down so that the food that comes off their private plots can be sold at a higher price. We have got to provide them with an incentive on the collective which equals the incentive on the private plot and actually attracts them away from the private plot and to the collective, which simply means much higher incentives.

As illustrated by Expert B’s G structure found in Fig. 2, the remainder of the protocol is concerned with how to implement the proposed solution. Expert B first considered raising food prices, rejected this, and then proposed education and subsidization: To do that the obvious way would be to raise prices of food, which would allow the collectives to sell at a higher price and therefore reap a greater profit and pour it back into the collective. But we can’t do that. We can’t allow prices of food to rise very much. Maybe minimally, but otherwise it would be socially unacceptable. People already are at 8 minimal level of wages. So the way it seems we will have to do it is to take the money that we would like to put into technological efficiency and instead pour it into the labor force. Pour into the labor force in two ways: (1) education, but (2) even more importantly, into simple subsidies for the collective to allow the prices to stay low.

A brief argument of support followed the solution, which in turn was followed by a statement of a problem that may arise, namely, the accumulation of capital or wealth. The remainder of the protocol is then directed toward how to prevent such accumulation, which of course is in opposition to the ideology. Expert B indicated that preventing accumulation may be accomplished by forced reinvestment, limiting options with respect to how money may be spent, for example, not permitting the purchase of land. He also argued that accumulation may also be prevented by the development of consumer products and indicated that in the short run such products would need to be imported, but in the long run they may be domestically manufactured and thus help to develop the industrial base of the economy. In summary, Expert B developed a problem representation via analysis of solutions that would not work because of ideology, the representation being “social,” and dealing with the lack of productivity on collectives, compared to private plots. Expert B thus converted the problem in a manner similar to Expert A. Expert B’s solution was general, and the bulk of his protocol was concerned with how it could be implemented, that is, solving a major subproblem generated by the solution. Expert C provided the longest protocol, although the G structure presented in Fig. 3, is relatively brief. After beginning with a constraint clarification, Expert C stated that research needed to be done to study productivity in various parts of the country, in order to determine whether

PROBLEM STATEMENT

Solution is easy except for political problem

labor education

Subsidize agriculture

Keeps prices down

I work done on private plots

Raise food prices

Reject

Avoids large land concentrations

,

i Import

T

“SOL,

Divert $ to create them

Want $spent in state industry

Fig. 2. G structure of Expert B.

Re-allocate $

I PROBLEM STATEMENTI

Multiple opinions will exist People stand up for their own interests Primary job of Minister e

m

exist on funda: mental ideologi-

Solution involves

W

Involves future ideas of Communist expansion

limited in what Provides examples

I 11

Have to convince peasants to implement new policies Give and take process to be reevaluated according to various needs

L

Fig. 3. G structure of Expert C.

I 11

I84


productivity is near its potential. If it is not, then it must be determined why the productivity is low. Expert C then presented a number of examples pertaining to various parts of the country and discussed how different solutions could be developed to the problems in various geographical regions. Expert C also pointed out that this work should be done by the staff of the Ministry of Agriculture: Once we identify those levels that are less productive than other levels, or less productive than their potential at least, then we have to figure out why. Now, hopefully all this has been done by the myriad of staff which surround the Minister of Agriculture. (This is followed by specific hypothetical examples.)

Expert C subsequently shifted to a second issue, namely, that the issues go beyond the Ministry of Agriculture: But those are issues which the Ministry of Agriculture itself wouldn’t have the final word on. That would involve other people as well. It would involve people who would be involved in what we would call marketing and so on, but is something which would be pursued for sure.

This statement was followed by a lengthy set of arguments showing how agriculture issues go beyond the Ministry of Agriculture. But subsequently Expert C reduced the problem of working with other agencies to the first person level: There is the whole political angle. That is, how does the Minister of Agriculture actualize whatever he sees as problem areas involved with other agencies. To a certain extent that is going to depend on how powerful he is, or I am, it is going to depend on how much pull he has. What kind of committees does he have that can kind of help him along within the Central Committee? What kind of access does he have to the Politburo? What kind of access does he have to not only the heads of ministries, which he would not just have on a formal basis, but personal access to them? What kind of favors do they owe each other? So there is the whole political thing.

Expert C, to this point, thus decomposed the problem into two subproblems. The first was the need to conduct extensive productivity analyses which then could be used to improve productivity in areas that were producing below their potential. This would be done by the ministry staff. The second was the problem the Ministry Head faces in dealing with other agencies. Furthermore, Expert C refined this subproblem by defining it as a first person political problem. Expert C thus represented the problem for the Ministry Head as political. Expert C subsequently proposed a solution that was especially oriented for the short-run aspects of the problem, namely, since the problem is po-


I85

litical, the solution consists of the Minister of Agriculture making sure his voice is heard. This statement was followed by an extensive discussion of how the Ministry Head may do this. The subsequent portions of the protocol were dominated by the political theme. Expert C returned to his initial concern of having the agency staff analyze productivity geographically, but this time Expert C emphasized how improving productivity in these areas would involve political interaction. He further provided examples regarding this matter as well as how one might deal with peasants in implementing policies. Finally, he considered another issue that could arise geographically, namely, the shifting population of the Soviet Union may require decentralization of agriculture control and this would have political ramifications. Expert C’s solution strategy differed from the strategies of Experts A and B. While the latter converted the given problem into a problem for which a solution could be proposed, Expert C decomposed the problem into two subproblems, focusing especially on the second, which yielded a political representation of the problem. Expert C’s solution was concerned primarily with the Ministry Head needing to act politically to implement the agricultural policies which presumably were to be developed without saying much at all about the policies themselves. Then Expert C supported his solution by showing how political skill would operate in relation to the analysis of agricultural problems by the ministry staff and this way brought his solution to bear on the solution to the first subproblem; that is, he discussed what one must do politically in various geographical regions in order to implement the policies for increasing productivity in the respective regions. Finally, we note that Expert C gave lengthy examples of specific cases which illustrated his respective points. Experts D and E employed the strategy of problem decomposition. They differ, however, in the extent to which they provided solutions. The G structure of Expert D is presented in Fig. 4.2 This expert was the only person to use notes in any serious way, using them as a reminder to comment on something later. These comments were not included in our analyses. Expert D decomposed the problem into five components: two constraints, ideology and elements, and three subproblems, bureaucracy, peasant mentality, and lack of infrastructure. A relatively long initial section of Expert D’s protocol is presented because it shows first, a historical analysis which helps to decompose the problem, and second, a quite explicit ’The reader is asked to note that numbers refer to the solver’s returning to issues. Basically, the G structure runs left to right in a temporal order, but when the solver returns to deal specifically with a particular issue, a number is placed at that point in the order and a jagged line followed by the number denotes the issue. Statements not analyzed are included in parentheses.

6 Have capaci-

t y to ba self-

Fig. 4. G structure of Expert D.


187

statement of problem decomposition, including a summary statement at the end of the section: The problem of Soviet agriculture is a nagging and historical one which has plagued the Soviet Union virtually since its founding. By way of thinking through possible solutions, let me review what I remember of various Soviet attempts to deal with this problem (as I sit in my office surveying the tundra, and deciding). Lenin came to power with the promise of bread, peace, and land and promptly tried to requisition grain from the peasants, that is to say, to force them to give their grain to the state. That didn’t work, and it was forced on them mostly by the contingencies of World War I and the Civil War. That failed, and Lenin had the foresight to replace it with what he called the tax in kind, that is, peasants could give a certain amount of their grain to the state, and then after they were able to grow anything they wanted, sell it and dispose of it as they wished. In fact, at one point in time a member of the government even said “enrich yourselves.” He said to the peasants, “make a lot of money.” Grain production improved, more grain was delivered to the cities, but it wasn’t, it wasn’t, of course socialism. It was by no stretch of the imagination Communism, it was socialism, but it wasn’t even what you’d call pure socialism. Essentially what he was creating was, or what he was accused of creating, was a class of little rich private farmers, something that a socialist revolutionary government would presumably not want. Stalin changed all that at the end of the twenties and the beginning of the thirties by forced collectivization, actually forcing people to yield their land up to collective farms, or state farms, in which people worked for the state or for the collective farms, much as a factory worker, and received a salary. This was pure socialism, but it also created a tremendous havoc in the countryside, produced revolution, peasants slaughtered their livestock, burned their crops, and resisted as long as they could. It is estimated that in the course of this process, perhaps 10 million people were killed or deported, or in some way deprived of either their livelihood or their life, for the greater good of Socialism and moving the country forward. That has remained the practice in the Soviet Union through the War until the middle of the fifties. And, what has also remained has been a consistent lack of efficient production and delivery. (So I’m making a note to myself that 1 should also comment on the fact that it is not only the system, but also natural aspects, whether infrastructure, questions, traditions, so I’m making the note to myself not to blame it all on government policy.) Well, in the fifties as the Soviet Union attempted to bring a better life to its people, especially under the leadership of Khrushchev, and pursue what was known as the New Course in the Soviet Union, a shifting away from consistent investment in heavy hand of the state on all aspects of the economy, including agriculture, Khrushchev came up with a number of plans to try to solve this problem, and one of which is known as the Virgin Lands Plan (which I usually joke to my students is a plan by which the fields are cultivated by 13 year old girls). In fact it was a program to bring into production land that had not been used before. It failed. Agricultural production improved, but not dramatically over the course of Khrushchev’s time and Khrushchev was, among other things when he fell, attacked for producing what were known as harebrained schemes, among them the Virgin Lands Plan. There is probably no greater problem that has persisted in the Soviet Union, that has nagged the system and its governing body, than what to do about agriculture. Now, in addition to the problems of trying to reconcile a Socialist policy with the need for production and, I hesitate t o say man’s instinct because, let’s confine it to peasants, that the peasants’ instinct seems to be, across countries in history, forlorned, for his

188

J. F. Voss, T. R. Creene, T. A. Post, and B. C. Penner

own land and for some share in his own production. That conflicts with a collective notion of some sort of government control in production, at least, making sure that people get fed. If grain is produced in the countryside, and people live in the cities, some method has to be found to get that grain into the cities, and at a reasonable cost, in reasonable proportion, or otherwise, there will be chaos, or even revolution and civil war, the cities of Russia emptied out as people desperate for grain left the cities in a desperate attempt to get grain. The peasant can be encouraged to grow, but will he sell, will he sell what he makes to the government? Or will he simply hoard the grain, knowing that the grain has a value itself beyond whatever he might be paid for it? Let me leave the government policy for a moment and talk about other things that affect agriculture. First, as I’ve indicated, or sort of hinted at, there is the peasant’s own frame of mind. It’s a traditional frame of mind. The peasant is typically a very religious person, very close to his land, his family, his village, and his church. He’s very suspicious about outsiders. This is also clear throughout Russian history, when various movements would arise. Go to the People was one of them, by the sort of Russian populists, known as the Urudniks, in the nineteenth century. Go to the People, and bring them the New World, or the new revolutionary mode. I forget exactly what they were supposed to be brought, most of them never had the chance, because the peasants simply killed many of the people who came to them, certainly the tax collector, any government agent is always suspicious. So that’s problem number one. The peasantry is not only suspicious of the government, but any new methods, any new introduction of new ways of producing agriculture, he would always grow enough for himself and for his family, and, if motivated properly, might grow enough to feed other people. The question is how to motivate him properly and retain some of the Socialist principles upon which our government is based. But there are problems as well, which are less malleable, I think, and less the fault of mismanagement of the government. One is the fact that agriculture is a particularly difficult business because of questions of weather, of soil development, of soil erosion. The Soviet Union is, for example, two and one half times the size of the United States, roughly one-seventh of the earth’s surface and yet, in terms of arable land, it roughly has about the same as the United States and less than a country like Australia, for example. So there is a problem simply of the elements, how to deal with that, in an effective way. Third, there is the problem of once the grain is grown, how to get it to the markets. The Soviet Union is a vast country with roughly, three hundred thousand miles of paved road, on a country two and one half times the size of the United States. The United States, just by contrast, has roughly some three million miles of paved road. So there’s a significant difference. When you have one railroad line going to Siberia and back, you recognize the size of the problem. There’s a similar disparity in terms of rail passage. The country covers thirteen time zones. Delivering, getting grain from the farm to the land, to the people who need it is a significant problem, even if there were not other problems of government mismanagement. Finally, this sort of gets back to the question of Socialist principles. This is a command economy in which the government makes a plan and it’s carried out by way of decisions, that are determined by the government, how much is to be grown, how much is to be invested which is crucial in terms of agriculture, how much is to go into fertilizer, how much is to go into tractors, how much is to go into every aspect of the agricultural question. That produces an enormous bureaucracy, and enormous sluggishness, an inflexibility to the market and to the needs of the country. So, if I were just to review the four sets of problems (and I’m making a note to help me to remember what I just said) there’s a question of the elements, there’s a question of the peasants’, let’s say, culture (and by that I don’t mean arts and sciences, I mean

Problem-Solving Sldll in Social Sciences

I89

his orientation to the world, his frame of mind) and finally, there’s the question of . . . well, two finallys, one, the question of infrastructure, that is, getting to market, railroad cars, the fuel, provisions, infrastructure and let me go back to point one, that is, the principles of Socialism, let’s call it the principles of Socialism management, meaning that there is a question of ideology, and question of management, that is, of plans, solving them, the plan, a problem of bureaucratic management.

The decomposition is followed by an interesting solution process. Although the next excerpt is also lengthy, the reader is asked to note how Expert D began his solution by stating that he would first deal with the peasants’ frame of mind, and then stopped and said, “I’ve changed my mind.” He then switched to the topic of what the Minister of Agriculture would “push for” politically. However, as the protocol continued, the expert brought in bureaucracy and infrastructure, and finally returned to the peasants’ frame of mind. Indeed, this section of the solution is a fascinating example of interaction among the solutions being provided to three subproblems: Let me begin with the peasants’ frame of mind. The peasant is a suspicious person. This is, of course, a stereotype, but we’ll operate on the basis of it to begin with, and, like all stereotypes it contains a kernel of truth. If we can assume, let’s say, that the peasant is suspicious of the government, suspicious of government actions, what needs to be done is to provide him with some sense of security. In our system, we allow peasants to have a certain amount of private land, in addition to the collective land on which he works. In fact, if it were not for that amount of private land, and it’s a small amount, that the peasant is allowed to have usually behind his house, or something like that, that he can cultivate. If it were not for that, our system would be in much worse shape, we would not be able to provide even the level that we do, which is, incidently, better than it was ten years ago, and much better than it was twenty years ago, but still, by modern standards, poor. And, more importantly than that, volatile. 200 million tons of wheat harvest one year, and then 180 million tons the next year, a vulnerable, volatile system. I changed my mind. I won’t start with the peasants’ frame of mind. Let’s start at the top with the question of policy. Since I’m the Minister of Agriculture, what kinds of things can 1 begin to push for? By saying that, I mean to indicate that I can’t make any of these decisions by fiat, by simply deciding on them myself. I can only lobby, and try to get the Government and the Politburo to go along. I think that’s likely to happen because I think that everybody recognizes the seriousness and the vulnerability of the problem. Things that I would push for are greater investment in agriculture. There must be more of the state budget, which is. after all, decided by the Party, put into agriculture, put into investment, into machines, tractors, fertilizer. If the peasantry will misappropriate that and use it for their private plots, well, we have to wink at that. It seems to me that we have to look the other way. But there has to be more of that so that agriculture becomes more modern. We have missiles, we have satellites, we have extraordinary advances in computers and other technologies. There’s no reason that we can’t have the same, just in terms of the nuts and bolts, of farming in our country. So, point one would be greater investment in agriculture. Point two, point one-b, probably would be greater investment in the things that would allow the supplies, once grown, to make it to the cities. Greater investment in infrastructure, paving the roads to the countryside, certainly electrification where it doesn’t already exist, railroad cars, trucks, specifically

J. F. Voss, T. R. Greene, T. A. Post, and B. C. Penner for the purpose of agriculture. Getting things that are grown in the land to the cities. Okay. (I’m making a note to myself that this means taking it from somewhere else and so I’m writing down the word budget to remind myself to talk about the politics of the budget.) The question is how to simply encourage people. Suppose we have all this investment. and railroad cars. and tractors and fertilizer, we still can’t get people to use it, to pour it into the farms, especially in the state and collective farms. How can we get them to use it in an efficient, effective way and encourage people to make efficient use and productive use of resources rather than wrathful use of resources or no use of resources at all? We have to, well, I think, allow for incentives. This is under what I’ve written down as private enterprise. but that’s not really what I mean. Well, what I mean is, private incentives which allow greater proportions, let’s say, of the crops grown on the collective farm to be retained, or to be sold. Maybe make collective farms self- governing enterprises or, that’s not what I mean enterprises whose profit and loss is a function of their own success, and to suggest to them that if they can grow more and they can sell more at a profit, then they can keep a certain amount and distribute it among themselves. To provide what might be called by someone else, certainly not me in the Ministry of Agriculture, capitalist incentives, incentives to increase production. To go along with offering people medals of Hero Socialist Labor if they’ll produce more, we need to offer people real incentives that encourage that. At the same time, we must recognize that the peasants’ production on his own private plot is crucial to our element, if it doesn’t get out of hand and people don’t start becoming rich landowners, and speculators in currency, and completely destroying social strata. Then 1 think that we should tolerate a greater degree of profit and initiative on the part of the private plots. We should allow use of state equipment on the private plots, and lower what taxes there are on the sale of things, to allow, what in America would be called truck farmers, to bring in goods to the cities more efficiently, more easily, even perhaps to encourage that. Along these lines, a somewhat more radical proposal, which might be a couple of more years down the road, would be to allow for the formation of genuine cooperatives. A cooperative is different than a collective farm in that it presumes the voluntary association of the farmers, their voluntary input and some degree of self-governing of their entities, genuine determination by themselves of how much they’re going to make and how much they’re going to invest. I think that we should begin along that to happen, perhaps in an experimental area of the country, to see if it works. It’s being done in China, it’s being done in Yugoslavia, with some success. I think we ought to allow that to happen too. While being careful, of course, that this doesn’t lead to a run on the collective farms and people leaving those in droves, but to see if it works, and if it increases production, and if it does, then reconsider to see i f we might want to do it on a broader scale. This is related to the third point, which is the one about the peasants’ orientation, the peasants’ feeling of security, the farmers’ feeling of security. One of the reasons, in Poland, for example, that the farms, although private, although owned by families, though also small and inefficient, have not produced as much as they might, is the constant feeling that the government might sooner or later take away what the peasant has invested. So that the peasant, for example, who makes a lot of money by selling his grain, is loathe to reinvest it into the farm, and into equipment, because he’s afraid-why should he do that-because the government will take it away. Why should he raise livestock beyond what he needs, or what he can sell for a few dollars. or a few zloty because there’s no point in that because the government might come along and take it away. That same sense of insecurity, I think, needs to be assuaged in our country probably by the measures that I indicated before, encouraging the use of private plots, making it easier. By making a series of reassuring policy deci-

..

...

Problem-Solving Skill in Social Scieoces

191

sions, which may involve making changes in the Ministry of Agriculture where people who are more inclined to reassure the private farmer, any of the peasants, that their production goals will be achieved, that they can invest some of their money, that they will have a feeling that the government will not be there to take away the fruits of their labor (I’m beginning to sound like Ronald Reagan). But there is no place, perhaps, that the government is more involved in the people’s lives than here in the Soviet Union, and if we have to ease up on that, while being assured of course that norms of socialistic legality are adhered to, it might reassure the farmers, and it will take years to do so, but I think the steps have to be taken so that the peasant’s feeling of insecurity will be lessened.

The G structure of Expert E, found in Figure 5 , indicates that two subproblems were delineated, one involving the need t o deal with issues beyond the Ministry of Agriculture per se and the other with issues of domestic and international interpretation. Following this, Expert E delineated a number of subproblems which have existed historically, and he refined some of these into more specific problems. Finally, when the issue of budget was considered as a subproblem, Expert E stated a solution which essentially consisted of presenting a plan to the Politburo. However, he pointed out in the evaluation of this solution that acceptance of the plan is quite related to the overall Politburo policies regarding investment and international relations. Expert E then proposed t o try to do something to help solve the various subproblems, but he recognized that he had not really given a solution and instead had shown how extensive and involved the problem is. Expert E thus decomposed the problem into a number of subproblems, but was not able to specify a general solution. The protocol of Expert F is not discussed in detail. The mode of solution was to decompose the problem into two subproblems, one dealing with geography and the other with worker incentives. Particular aspects of the expert protocols are now presented, including some that were not discussed in relation t o the specific protocols. 1. Experts did not articulate their highest level plans. While an expert may have mentioned the need to consider constraints, no expert spoke of a general strategy or plan to be used to solve the problem. 2. During the initial phases of protocol generation, experts showed no evidence of having a well-developed solution plan. Instead, they tended to do some type of review of the problem (three used a historical approach) from which they developed a general representation of the problem. Under what conditions experts would have a well-articulated plan is discussed later in the article. 3. As previously noted, the solution strategy to problems such as the Soviet Union agricultural problem is to isolate one or more of the factors producing the problem and to solve the problem by eliminating the effects

Fig. 5 . G structure of Expert E.


193

of the factors. While this strategy was used by everyone, two more specific problem solving strategies were employed. Problem conversion was used by Experts A and B and, in a sense, by Expert C in that Expert C converted the problem into a first person political issue. The other strategy employed was problem decomposition, used by Experts C, D, E, and F. (How consistently a particular expert would use the same strategy for a number of problems is a question of interest, but it is beyond the scope of this contribution.) 4. Experts generated subproblems in two ways; they either stated a subproblem as part of a decomposition process or they “encountered” subproblems when they were exploring the implications of a proposed solution. This is an important distinction, for it is primarily the experts that generated subproblems via the second mechanism. 5 . A large proportion of the contents of expert protocols consisted of R structure information. Possible reasons for this were discussed earlier in the article, but the protocols make it abundantly clear that the expert frequently explores the ramifications of the proposed solutions and provides extensive supportive argument. 6. The nature ofthe solutions offered by experts, regardless of strategy, is to find one general solution which is able to solve the problem as it had been represented. Indeed, one expert, E, indicated that he was unable to provide a solution when he could not come up with this type of general statement even though he suggested that subproblems could be handled on an individual basis. E.

NOVICE A N D POSTNOVICE

PROTOCOLS

Novice protocols are characterized by problem decomposition in which solutions are proposed for a number of relatively low-level subproblems, for example, need more fertilizer, need more tractors. This decomposition, on occasion, even takes the form of a listing. Figure 6 presents the G structure of the least developed novice protocol, while Fig. 7 presents the G structure for one of the most developed novice protocols. With respect to the R structure, novice protocols are characterized by a lack of argument. While some novices provided support for their solutions, the support was generally weak. Furthermore, novices quite typically did not evaluate solutions in terms of ideology or some other constraint and they did not specify subproblems that could be encountered when proposed solutions were implemented. The protocol for the novice whose G structure is presented in Fig. 6 had no R structure, while the R structure for the novice of Fig. 7 is presented in Table V. As previously mentioned, six novices received the Soviet agriculture prob-

PROBLEM STATEMENT

other areas to make up for things they

from other parts

Fig. 6. G structure of least developed novice.

LLd


195

lem both at the beginning and the end of a course on Soviet domestic policy. An interesting finding is that there was relatively little difference in the form and contents of the protocols.

F. GRADUATE STUDENT PROTOCOLS The protocols of the three first- and second-year graduate students form an interesting transition between novice and expert performance. While the graduate student protocols involved problem decomposition in a manner similar to the novices, the subproblems were more abstract than those of the novices but not as abstract as those of the experts. For example, while novices suggested the need for tractors, and the experts suggested the need for infrastructure development, graduate students stated the need for better machinery, better transportation, and more money for agricultural development. Figure 8 presents the G structure for one of the graduate student protocols.

PROBLEM STATEMENT

I

I

Use Nationalistic incentives

I

GSOL Use technology and research to find new land and control climate Land arid and cold

art of land

7 -i1

and incentives

Fig. 7. G structure of well-developed novice.

196

J.

F. Voss, T. R. Creene, T. A. Post, and B. C. Penner

Fig. 8. Sample of graduate

An aspect of the graduate student protocols is that, on occasion, when a solution was proposed, implications of the solutions were considered, including subproblem development. However, while graduate students examined such implications, they provided nowhere near the argument development found in expert protocols. Table VI presents one of the more extensively developed graduate student arguments. Briefly stated, the picture that emerges regarding the protocols discussed thus far is that, with the training and experience involved, the path to expertise includes: (1) isolating more abstract or general problems as being the primary factors producing the given problem, and in so doing, making lower level problems subordinate to the more abstract problem, (2) in a related way, stating a general or abstract solution which also solves subordinate


197

1 Too many I

flatstarms 1

creased

down on num. bsr of state farms and collectivm

+,+, break down

of farm labor

,-

Lators-

1

$GWL,

Increa~

military

US is in. creasing dc-

f 1.nu .nnnrl,na

Take money out 01 defenn a d put it into riculture

Military power not the only way 10 solve world pro blemr

student G structure.

problems rather than stating, as solutions, the solutions t o more specific problems, and (3) developing much greater skill in examining the implications of stated solutions, especially in providing support for particular solutions, stating what subproblems may emerge in relation t o the solutions, and evaluating solutions in relation to constraints.

G. NONEXPERTEXPERTPROTOCOLS This category of participants refers to four advanced graduate students and two faculty members who are in political science but whose field of expertise was not the Soviet Union. Figure 9 presents the most developed

198

J.

F. Voss, T. R. Creene, T. A.

Post, and

B. C.

Penner

TABLE V

SAMPLE OF NOVICER

STRUCTURE

(GEVA) Evaluation of solution (RARG) Would have to use nationalistic incentives (RREA) Personal incentives are capitalistic, not communistic (RELA) Nationalistic incentives-for the better of the people and the USSR (RELA) Convince people to work harder for more growth of the USSR

G structure of an advanced graduate student. His area of specialization is Latin America. The protocol indicates the following: 1. Problem decomposition occurred, with solutions offered to the subproblems. There is, however, no general solution as found in expert protocols. 2. The subproblems are more general than those defined by novices and resemble more those defined by experts. 3. There is a larger argument development than found in novice or beginning graduate student protocols, including some argument by analogy. Furthermore, the argument development sometimes refers to general political issues.

An example of this person’s R structure is found in Table VII. Figure 10 presents the G structure of one of the faculty. This person’s area of expertise involved domestic policies such as taxation. As shown, this solver has a relatively undeveloped protocol, emphasizing the ideology constraint and what solutions may be proposed while working within that constraint. The solutions are of a general political nature. Argument development is relatively weak, as shown by an R structure presented in Table VIII. TABLE VI SAMPLE OF GRADUATE STUDENT R STRUCTURE (GSUP) Support that subproblem of opposition exists (RARG) There is going to be a lot of opposition (RSAS) Can increase incentives, but against the socialist system (RSAS) Measures I might propose might be rejected by Politburo (RPSC) Kieryenka might reject measures

STATEMENT

*?, problem

Have to $ resources

ivization

Examples

systematic investigation to study

Know what problem is so can campaign for resources

when they

world staple Problems

CaUSeS

farmers are demoralized

pziiiiq STATEMENT

Fig. 9. G structure of nonexpert expert.

I am a Communist party loyalist

Ideology

Options limited for solutions

GEVA option

GEVA May be way of getting around ideological problems

Fig. LO. G structure of domestic policy expert (agriculture problem).

I


20 1

TABLE VII SAMPLE OF NONEXPERT EXPERTR STRUCTURE (RARG) Have to examine problem within the Latin American context with which I am familiar (RSAS) Not enough government investment in infrastructure (RPSC) Granting credit for fertilizers (RPSC) For roads and transportation (ROUT) People can’t get their things to market

We also gave this solver a problem in her own area of expertise, a problem involving income tax. Figure 11 presents the G structure for the obtained protocol. While the form of the G structure is similar to that shown in Figure 10, the contents are more abstract in the individual’s own area of expertise. There is a top-down, breadth-first decomposition defining two problem areas termed “horizontal equity” and “vertical equity.” The solver then indicated what could be done to solve problems related to both of these areas. Furthermore, there is a much more developed R structure than found in this person’s Soviet agriculture protocol, as shown by the account presented in Table IX. Comparing the two protocols generated by this individual provided some interesting information. First, a general problem-solving strategy was applied in both problems. However, the subproblems and solutions are quite generally political for the Soviet problem, and little more than a superficial knowledge is demonstrated with respect to subproblems and solutions and R structure contents. In contrast, the G structure contents in the person’s area of expertise are more specifically related to the issues, even though they are stated in an abstract manner similar to that of Soviet experts. In addition, the R structure development, including issues of support, evaluation, and subproblem development are also highly similar to the Soviet experts’ protocols. H.

CHEMISTPROTOCOLS

Of the four protocols of chemists, three strongly resembled those of novices, especially in relation to the level of problem and solution considered as well as to the lack of exploring solution implications. The fourth chemist provided a much more developed protocol, which, while not demonstrating an extensive knowledge of the Soviet Union, nevertheless demonstrated a


202

sensitivity to bureaucracy. Figure 12 presents the G structure of this protocol. With respect to the R structure, this solver at times demonstrated evaluation based upon ideology and government bureaucracy. I.

MISCELLANEOUS PROTOCOLS

The protocol of the foreign service officer was quite pragmatic in that he indicated what would be done if the Ministry Head was a technocrat. This person also stressed the need for more information. As to structure, the foreign service officer isolated a few subproblems, and these are presented in Fig. 13. In addition, there was relatively extensive argument development concerning why particular steps could not be taken. The Eastern European scholar made a strong statement regarding the

doing thiws right

Do not handle so that mutually

I

years do correla. tional study to ree whm a f f v t r

Crnni

Have to decide what to tell superiors

I with scientific approach

Ifcorrelations significant. than can do somethinp, if

information

Give reasons for directives

Have discussions w/regional directors to get information

Get opinions of why they think yields are low

Again, be honest

Given information you have, 2 possibilit i n for real answers

Fig. 12. G structure of chemist’s protocol.


203

ideology constraint, isolated the problems of infrastructure and incentive, and then proposed the interesting solution of funding two "think tanks," one to deal with technological issues and the other to deal with sociopsychological problems. This person's G structure is presented in Fig. 14. The R structure of this person showed supportive argument and evaluation, but it was not developed as extensively as experts.

VI.

The Acquisition of Social Science Problem-Solving Skill

The question considered in this section is what the present research suggests regarding how social science problem-solving skill develops. Before

PROBLEM

i-l

Is real goal to increase production or other goals above that which must be met?

Probe Given your experience

GSUB

Have to respond to politics and ideology

Would have to set up my own scenario GSUP

Solution depends on Minister beina wlitician or technocrat

Would have to manupulate several factors. Examples

Have to look at what drives the policy maker

Would be sure to understand real dvnamics of problem

I am technocrat. I would say "Damn the ideology" and would put in system to stimulate farm production I

r.

1

11

GSUP Therefore, can't always get understand problem I

4 Would have to provide i ncent ives

Would need more information

GSOL

Find out why production

push and say everything will be produced to maximum

Examples

Fig. 13. G structure of career foreign service officer.

-I

J. F. Vow, T. R. Greene, T. A. Post, and B. C. Penner

204

TABLE VIII SAMPLE OF DOMESTIC POLICY EXPERT R STRUCTURE (AGRICULTURE PROBLEM) (GIPS) Interpret problem statement (RARG) Would have to assume that I am a Communist Party loyalist (RSAS) 1 have been brainwashed for years and years with Marxism-Leninism IRSAS) Certain options would be available in other countries, but not USSR (RPSC) Private incentives (RPSC) Payments to farmers (RELA) Subsidies to encourage individual effort (RSAS) I know our balance in payments is screwed up (RREA) Because we have to import so much I am not allowed to do anything about it though (RSAS) (RREA) Because I would be shot down

getting into this issue, however, we consider which components of such skill appear to be domain specific and which are more general. The present reseach suggests that two aspects of the skill are general. One is the individual's knowledge of the physical and social world, which may be assumed to constitute a data base common to all individuals. The other is knowledge of and ability to use quite general problem-solving strategies. Thus, all participants tried to isolate one or more factors producing the agriculture problem, and they did this in a reasonably straightforward man-

1

PR0BL EM STATEMENT

I

To what extent can agriculture be improved?

P w r infraitructure

Workers' motivation

Labor f a l s that what belongs to states belongs to rI

- -State - -run- _ - - _ ,

Area of deciiion making i i divided into 2 parti. techno. logical and w i o piychologiul

Form 2 think tanks

I

Have to create feeling that peasants can participate

Create feeling that somathing dependi

Fig. 14. G structure of East European scholar.

One for each type of problem and then follow their supaeitionr

Problem-Solving Skill in Social Sriences

205

TABLE IX

EXPERTR STRUCTURE (INCOMETAXPROBLEM)

SAMPLE OF DOMESTIC POLICY

(GlPS) Interpret problem statement (RARG) Given my political interests, I will have to think very differently if I am going to head this committee (RSAS) 1 am going to have to think like a conservative, business-oriented supply-sider (RQUA) These are values I question seriously (RSAS) Some of the serious problems with income tax are not entirely ideological (ROUT) Therefore, experts will be able to agree on some solutions (RPSC) Recent tax increase passed had reforms that liberals have been supporting for years (RELA) Improving compliance, withholding of savings and loan interest (RCON) These were goals of tax policy professionals and liberals (RPSC) Trying to improve fairness and rationality of the system

ner. The exception is the experts whose method of isolating the primary factor(s) usually included a rather extensive problem analysis. On the other hand, one component taken to be domain specific is knowledge of the subject matter, that is, declarative knowledge (Anderson, 1982; Ryle, 1949) related to the Soviet Union. A second domain-specific component consists of problem-solving strategies that tend to be used within a particular subject matter domain. Furthermore, two types of such strategies may be discerned. The first involves strategies that are not content specific, but nevertheless tend to be used in particular subject matter areas. Problem conversion is one such strategy, and is demonstrated by some of the Soviet experts. Thus, while problem conversion is no doubt used in many domains, the present research suggests that the more sophisticated individuals working within the social science domain may use this strategy while the less sophisticated do not. In this sense the use of the strategy is taken to be a function of experience working in the domain. The second class of strategies that is taken to be domain specific consists of those modes of problem analysis found in particular domains, an example being the historical analysis of the experts on the Soviet Union. Other domain-specific strategies would include how to set up proofs in geometry (Greeno, 1978) and how to decompose a software design problem in computer programming (Jeffries et al., 1981). Thus, while the “world” information and general prob-

206

J . F. Voss, T. R. Creene, T. A. Post, and B. C. Penner

lem-solving strategies may be acquired from a number of sources, domainrelated information and strategies are assumed to be acquired within the domain via experiences which provide opportunities to utilize and organize the information. Such experiences no doubt include reading, writing, listening, and solving problems within the particular domain. We now turn to a cross-sectional type of analysis of the current findings in order to present some ideas regarding the acquisition of social science problem-solving skill. The novices of the present research had some knowledge of the Soviet Union but were not effective in using it in the problem-solving context. Moreover, their knowledge seemed to consist of “bits and pieces” of information that were not well integrated. The postnovice protocols were not appreciably different from those of the novices, a finding which suggests that, although the individuals acquired information related to the agriculture problem during the academic course, they were not successful in utilizing this information when given the problem at the end of the course. We would suggest that this result occurred because the course provided little opportunity to organize information in relation t o issues such as agricultural productivity. Compared to the novices, first- and second-year graduate students showed three changes, although these did not apparently occur as well-defined stages. First, some knowledge of subproblem interaction was shown, even though the various subproblems were described in a manner similar to novices. Second, the subproblems were stated at a little more abstract level than those stated by the novices and postnovices. Third, more reasoning was shown, in the sense of support and evaluation. The three changes suggest that learning within the domain includes the use of declarative knowledge t o construct three types of structures. First, conceptual networks are constructed and continually expanded. This construction provides the individual with increasing knowledge of relations among concepts, facts, principles, etc. Second, in a related way, the individual constructs causal relations among these factors which provides the knowledge of the interdependencies that exist within a domain. It is this knowledge, moreover, that enables the individual to develop argumentation. Third, the individual develops hierarchical structures, a finding of particular interest because such development is quite clearly an important aspect of expert performance. (A possible reason for developing hierarchical structures is to handle the magnitude of information one is arguing; cf. Simon, 1969.) Thus, the expert sees the lack of fertilizer, lack of repair parts, lack of infrastructure, etc. as elements related to a more abstract concept, lack of capital investment. Furthermore, the hierarchical organization in all likelihood enhances the ability to retrieve information.


207

The performance of graduate students and faculty with expertise in social science areas other than the Soviet Union and the performance of chemists also suggests some points regarding the acquisition process. First, when expertise is acquired in a field such as political science, it may be assumed that one acquires domain-related strategies and a data base within one’s field of expertise. However, when confronted with a political science problem outside of one’s specialization, and for which one does not have a substantial data base, the individual tends to fall back upon more general knowledge of the field and of the topic in question when utilizing the domain-related strategies. Furthermore, while the social scientists demonstrated the use of solution strategies but lacked a Soviet Union data base, the chemists apparently lacked both the data base and the domain-related strategies. However, we d o not know of course whether the chemists “really” lacked the strategies or whether the lack of the data base did not provide them with the opportunity to utilize the strategies. Turning now to the experts, in addition to the components of skill development discussed thus far, the most obvious factor contributing to expertise is the varied exposure to issues and problems of the particular field. Such experiences, it may be assumed, provide multiple patterns of organization that are constructed and stored. Furthermore, these patterns are probably highly context dependent. For example, let us assume that the experts of the present study had never really considered specifically why agricultural productivity was low in the Soviet Union and that the protocols generated represent what happens. The expert typically performs an analysis of the problem, for example, an historical analysis, and the expert presents a solution, combined with an examination of the implications of the solution. But now assume that the expert is asked repeatedly how agricultural productivity in the Soviet Union could be increased. In all likelihood, an expert’s answer would become less variable with repetition; whereas the first time that the expert was confronted with the problem it was necessary to conduct an “organized search,” this becomes less necessary with repetition, and whenever the issue of Soviet agriculture comes up, the expert has readily accessible organized information. Thus, as the expert considers more issues such as the agricultural problem, the expert organizes information according to the problem context. Finally, the patterns of organization also give the expert an added advantage, for if the expert acquires new information about a particular issue, it is quite likely that the information is stored in relation to the existing structure and thus is also readily accessible when that issue arises again. The experiences that enhance the expert’s knowledge organization no doubt cover a wide range of issues and problems, and the expert thus develops a highly flexible information-processing system. It is flexible in the

208


sense that domain-related information, whatever it is, may be readily interpreted in terms of existing structures, and new information may be assimilated into the appropriate structures. In a general sense then, the expert social science problem solver thus begins to resemble the chess expert (Chase & Simon, 1973) in that experience provides for the development of a large number of patterns which the expert may identify. (If the reader wishes to assume the expert has many Soviet Union-related schemata, we have no objection. We do, however, feel that the schema concept and especially the script concept have come to connote more rigidity than actually found with expertise.) Our discussion of skill development has been at a global level. The more detailed mechanisms by which the structures are developed and knowledge i s organized are not considered here. However, we would note that mechanisms such as those proposed by Anderson (1982) are compatible with the notions expressed in this paper. For example, in Anderson’s model, the proceduralization mechanism involves converting declarative knowledge into procedural knowledge. What we referred to as patterns of organization developed by the experts’ application of strategies would in Anderson’s model be regarded as procedural knowledge in that the knowledge is employed as a set of procedures to solve particular problems.

VII. General Considerations A. TASKENVIRONMENT Two interesting statements were made by one of our participants not mentioned previously. He is a Latin American expert who has frequently appeared on television and radio interviews discussing such issues as the policies of the United States in El Salvador. The problem that he was given was as follows: “Assume that you were a high-ranking State Department official and were asked by the President to suggest how the United States should change its policies toward El Salvador. What would you suggest?” Contrary to the protocols generated by all other experts, this individual immediately stated his solution, which was relatively abstract, and which emphasized the need to assemble the leaders of various political factions and enter into negotiations. For our purposes, the important point is that this individual later indicated that the wording of the question made it unnecessary for him to go through a preliminary development of what the situation in El Salvador was, how it had developed historically, and what the current United States policy is and why it is not working. The solver indicated that the fact that the problem stated a new policy was needed presumed this. Furthermore, this individual indicated that he had been


209

thinking about this issue for a long time and that he knew what he felt the solution should be and he therefore was able to state it without hesitation. This supports our notion that when an issue has been considered often, a solution is readily accessible. Another interesting observation is that this solver indicated that in giving his protocol he was, in the context in which the protocol was given, not hesitant to make statements which would suggest his own political inclinations. However, he indicated that if he were in a public interview, he would only make statements which he would be able to support by argument and he would try not to show his own political preferences. In a similar vein, at a recent conference held at the University of Pittsburgh, a State Department official, in replying to a question pointed out how problem solving is influenced by what the next in command would accept or reject. (This factor is no doubt highly significant in government and industry problem solving, but of course not in academia.) These observations are highly instructive. They suggest that experts need not develop extensive problem representations when ( I ) the wording of the problem suggests that this is not necessary, and (2) when the solver already has a solution in mind. Also, the observations indicate how the surroundings in which the problem is presented may provide constraints and influence what the individual states. Indeed, there apparently are “audience effects” in social science problem solving just as in writing and speaking. We might add that the bulk of our research has been conducted in a relatively “open” atmosphere and the solver is aware that anonymity will be preserved. In this respect constraints were hopefully reduced.

B.

PROBLEM SPACE

The problem spaces established by novices and postnovices appear to be quite limited, while those of the more experienced solvers are of course much more filled with information. Furthermore, there is some evidence that the problem space may expand as the expert solver advances his or her solution. This occurs when the solver comes to issues that were not part of the original representation.

C. PROBLEM REPRESENTATION The issue of problem representation is central to the solving process, especially for experts. The following additional finding is of interest with respect to this matter. We presented experts and novices with an international problem in which we asked that one assume herself or himself t o be a high-ranking U.S. State

210

J . F. Voss, T. R. Greene, T. A. Post, and B. C. Penner

Department official who is asked to formulate a workable foreign policy with respect to the Soviet Union. However, in so doing, the individual was asked to take into account a large number of factors, including Eastern European countries, Western European countries, one’s own political party, popular opinion, position on SALT, cultural exchange, etc. Without exception, novices went down the list of things to consider and indicated what should be done with respect to each item. Experts, on the other hand, spent considerable time on problem representation. One expert (Expert B in the present research) gave a highly instructive protocol. He initially indicated that the policy will depend upon one’s perception of the actions and motives of the Soviet Union and whether one thinks that the Soviet Union is basically aggressive or reactive (political science distinction). This person said that he felt that the Soviet Union is basically reactive and provided extensive argument for why he thought this to be the case, using primarily examples of Soviet actions. Only after much argument development did this solver get to specific issues which in his protocol were subordinate to one’s attitude toward the Soviet Union. This protocol and those of other experts illustrate a point which could not be demonstrated in the Soviet agriculture problem solutions, namely, that the problem representation and what follows can be highly influenced by one’s attitudes and motivations. Indeed, one has the idea that government policy development is apparently quite influenced in this way.

D. SOLUTION ACTIVITY While decomposition was clearly the predilection of the less experienced solvers, this strategy was also found in expert protocols. Thus, while decomposition may be assumed to be a quite general problem-solving strategy, which is used for the “something is wrong, do something about it” problem, the experience of social science experts, as previously noted, provides the experts with strategies that may be used effectively in their domain. Furthermore, the faculty member having United States policy expertise demonstrated a top-down, breadth-first strategy for a problem with which she was highly familiar. Finally, we note that it was primarily the experts who showed evidence of opportunistic solving and constraint posting. This finding is reasonable because only experts examined the implications of their solutions, and therefore had an opportunity to encounter a subproblem or constraint. In these instances, the experts handled the new subproblem or constraint before proceeding further with the argument. One of the major findings of the present research is that experts spend a large amount of time on argumentation. At first glance, one may conclude that natural sciences and social sciences may be fundamentally different in


21 I

this regard, with little argumentation found in natural sciences. However, this is not really the case. Our view is that the major factor producing the social science and natural science differences in solving activity is that, as noted earlier in this article, problems in fields such as physics have been well worked out. This is particularly true for those problems used in problem-solving research, (e.g., Larkin et al., 1980). On the other hand, few social science problems have agreed upon solutions. While we are not aware of any research on the problem-solving activity of a scientist when trying to solve a problem that has not been worked out, some interesting work by Tweney (1981) is germane. Tweney has studied the notebooks of Michael Faraday, especially in reference to Faraday’s process of confirming or rejecting hypotheses. The interesting observation in the present context is that in studying the (previously unsolved) problem of electromagnetic induction, Faraday apparently proposed one or possibly two basic solutions and studied the implications of these solutions. Faraday’s notebooks thus apparently portray a solving process similar to the present work. A difference, however, is that Faraday developed arguments which led t o experimental tests that could be conducted and refined within a short period of time in his laboratory. Testing hypotheses in social sciences, however, tends to be a much more protracted process. Thus, the relatively large amount of argumentation found in social sciences may be attributed to the fact that the problems d o not have agreed upon solutions, whereas the problems studied thus far in natural sciences have well-formulated solutions. E.

EVALUATION

Evaluation was employed in the present research primarily by the solvers who had more domain-related experience. There was little evidence of a straightforward means-ends analysis. The evaluation process more closely resembled a generate-test method, although the solutions were generated in a highly developed way. The evaluation process occurred via two mechanisms: direct, that is, evaluating a solution in relation to a constraint, and indirect, that is, providing support for a solution by exploring its implications with respect to what it would accomplish. In a sense then, evaluation is a jurisprudence process in that a person builds a case for a particular solution.

VIII.

Concluding Remarks

We close this article with three brief comments. First, while this work is exploratory, it nevertheless does demonstrate that the solving of complex


212

and ill-defined as well as “real-world” problems is amenable to study. Second, the research provided information concerning protocol analysis as a method, and suggested some ways in which protocol contents may be influenced. Third, the research provided a sketch of the processes involved in learning the skill of social science problem solving, emphasizing the importance of experience in providing ways of organizing information so that it subsequently may be utilized in further processing.

ACKNOWLEDGMENTS The research reported in this paper was supported by the Learning Research and Development Center (LRDC) at the University of Pittsburgh, Pittsburgh, Pennsylvania. The LRDC is supported, in part, as a research and development center by funds from the Advanced Research Projects Agency (ARPA), an office under the Secretary of Defense, and the National Institute of Education (NIE), United States Department of Education. The opinions expressed do not necessarily reflect the position or policy of ARPA or NIE, and no official endorsement should be inferred. The authors wish to thank Dr. Sherman Tyler and Laurie Yengo for their contributions to the present research. The authors also wish to thank the University of Pittsburgh faculty, graduate students, and undergraduates, and visiting scholar, as well as the foreign service officer, who participated in this work.

REFERENCES Anderson, J. R. Acquisition of cognitive skill. Psychological Review, 1982, 89, 369-406. Chase, W. G., & Simon, H. A. The mind’s eye in chess. In W. G. Chase (Ed.), Visual information processing. New York: Academic Press, 1973. Chi, M. T. H., Feltovich, P., & Glaser, R. Categorization and representation of physics problems by experts and novices. Cognitive Science, 1981, 5 , 121-152. Greeno, J. G. Hobbits and orcs: Acquisition of a sequential concept. Cognitive Psychology, 1974, 6, 270-292.

Greeno, J. G. A study of problem solving. In R. Glaser (Ed.), Advances in instructional psychology (Vol. I). Hillsdale. New Jersey: Erlbaum, 1978. Hayes-Roth, B., & Hayes-Roth, F. A cognitive model of planning. Cognitive Science, 1979, 3, 275-310.

Jeffries, R., Turner, A. A., Polson, P. G., & Atwood, M. E. The processes involved in designing software. In J. Anderson (Ed.), Cognitive skills and their acquisition. Hillsdale. New Jersey: Erlbaum, 1981. Larkin, J., McDermott, J., Simon, D. P., & Simon, H. Expert and novice performance in solving physics problems. Science, 1980, 208, 1335-1342. Newell, A., & Simon, H. Human problem solving. New York: Prentice-Hall, 1972. Reitman, W. Cognition and thought. New York: Wiley, 1965. Ryle, G. The concept of mind. London: Hutchinson, 1949. Sacerdoti, E. D. A structure for plans and behavior. Amsterdam: Elsevier. 1977. Simon, H. A. The science of the artificial. Cambridge, Massachusetts: MIT Press, 1969.


213

Simon, H. A. The structure of ill-structured problems. Artifcial Intelligence, 1973. 4, 181201.

Simon, H. A. The functional equivalence of problem solving skills. Cognitive Psychologv, 1975, 7, 268-288.

Simon, H. A. Information-processing theory of human problem solving. In W. K. Estes (Ed.), Handbook of learning and cognitive processes: Human information processing (Vol. 5 ) . Hillsdale. New Jersey: Erlbaum, 1978. Stefik, N. Planning and meta-planning (MOLGEN: Part 2 ) . Artificial Intelligence. 1981, 16, 14 1- 170. Toulmin, S. E. The uses of argument. London and New York: Cambridge Univ. Press, 1958. Toulmin, S. E., Rieke, R.,& Janik, A. An introduction to reasoning. New York: Macmillan, 1979.

Tweney, R. D. Confirmatory and disconfirmatory heuristics in Michael Faraday 's scientific research. Paper presented at meeting of the Psychonomic Society, Philadelphia, 1981. Voss, J. F., Tyler, S . , & Yengo, L. Individual differences in social science problem solving. In R. F. Dillon & R. R. Schmeck (Eds.), Individual differences in cognitive processes (Vol. I). New York: Academic Press, 1983.


BIOLOGICAL CONSTRAINTS ON INSTRUMENTAL AND CLASSICAL CONDITIONING: IMPLICATIONS FOR GENERAL PROCESS THEORY Michael Domjan THE UNIVERSITY O F TEXAS AUSTIN. TEXAS Introduction ......................................................... Constraints on Positive Reinforcement . . . . . . A. “Misbehavior” in Depositing Tokens for B. Differential Conditionability of Various Responses with Positive Reinforcement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111. Constraints on Punishment . . . . . . . . . . . . . . . . . . . . . . . . , . . . . . . . . . . . . . . . . . . , 1v. Constraints on Avoidance Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . A. Techniques that Facilitate Avoidance Conditioning . . . . . . . . . . . . . . . . . . B. Species-Specific Defense Responses and Avoidance Learning . . . . . . . . . . C. The Consummatory Stimulus Reward Hypothesis. . . . . . . . . . . . . . . . . . . . . D. Further Investigations of Stimulus Factors .................. in Avoidance Learning . . . . . . . . . . . . . . . . . . . . . E. Concluding Comments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . V. Selective Associations in Classical Conditioning . , . .. . . . . . . . , . . . . . . . . . . . . . A. Selective Aversion Conditioning with Shock and Illness . . . B. Other Studies of Selective Conditioning of Ingestive Behavior.. . . . . . . . . C. Selective Appetitive and Aversive Conditioning in Pigeons.. D. Selective Conditioning of Stimuli as Danger and Safety Sign

I.

11.

. . .

v1. A.

VII.

Mechanisms Contributing to the Delay of Reinforcement Gradient in Taste-Aversion Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

A. Facilitation of Direct Association with Toxicosis . . . . . . . . B. Mediated Conditioning of Nongustatory Aspects of Food C. Instances of Potentiation outside the Feeding System . . . . . . . . . . . . . . .._.

VIII.

A.

Levels of Analysis of Conditioning Procedures

References . . . . . . . . .

...........

. . . . . . . . . . . . .. . ................

216 217 217 219 227 228 228 229 23 1 233 234 235 235 245 246 247 248 248 249 253 259 260 26 I 262 263 263 264 265 268

215 THE PSYCHOLOGY OF LEARNING AND MOTIVATION, VOL. 17

Copyright 0 1983 by Academic Press, Inc. All rights of reproduction in any form reserved. ISBN 0- 12-543317-4

Michael Domjan

216

I.

Introduction

Investigators of learning traditionally have sought to discover general principles of behavior that would describe and explain the course of conditioning and learning in a wide variety of species and situations. They were mindful of the existence of species-specific adaptations in sensory, motor, and motivational systems. However, such specializations were assumed to influence only the manifestations of learning and not its mechanisms. The laws of learning were assumed to be independent of the particular cues, responses, and reinforcers that were employed in an experiment (see Schwartz, 1974, 1978). The assumption that the laws of learning are independent of the cues, responses, and reinforcers involved came under vigorous attack about ten years ago with the publication of a number of important theoretical papers and books (Hinde & Stevenson-Hinde, 1973; Rozin & Kalat, 1971; Seligman, 1970;Seligman & Hager, 1972; Shettleworth, 1972). These writings were stimulated by certain observations about learning that were contrary to the principles of learning widely espoused at the time and that appeared to illustrate adaptive specializations in learning. On the basis of these observations, writers argued that the traditional pursuit of general process learning theory may be an ill-fated venture and that future studies of learning will have to consider constraints on instrumental and classical conditioning that result from evolutionary adaptation. Biological constraints on learning quickly became treated as a fundamental problem in the study of learning and were discussed in numerous introductory psychology books and more advanced books on learning (e.g., Adams, 1976; Bolles, 1975; Flaherty, Hamilton, Gandelman, & Spear, 1977; Gleitman, 1981; Hall, 1976;Houston, 1976;Hulse, Deese, & Egeth, 1975;Morgan & King, 1975; Nevin & Reynolds, 1973;Rachlin, 1976;Schwartz, 1978;Smith, Saranson, & Saranson, 1982;Tarpy & Mayer, 1978;Wickelgren, 1977). Some of the most famous examples of biological constraints on instrumental and classical conditioning include “misbehavior” in depositing tokens for positive reinforcement (Breland & Breland, 1961), constraints on conditioning of grooming with positive reinforcement (Thorndike, 191 l), constraints on punishment (e.g., Walters & Glazer, 1971), constraints on avoidance learning (Bolles, 1970), selective associations in aversion learning (Garcia & Koelling, 1966), long-delay poison-avoidance learning (Revusky & Garcia, 1970), and potentiation of food-aversion learning (e.g., Galef & Osborne, 1978). Given the great impact that discussions of biological constraints on learning have had during the past decade, it seems appropriate to reconsider some of the prominent examples of constraints on instrumental and classical conditioning. The purpose of this article is to review the empirical and theoretical analyses of these phenomena that have taken

Biological constraints on Learning

217

place since the issue of biological constraints was brought into focus about 10 years ago. We also consider what impact these investigations have had on the search for a general process learning theory. The review indicates that recent studies of biological constraints on instrumental and classical conditioning have stimulated modifications of previously accepted ideas about learning. They have greatly increased the range of factors that are considered to be relevant to predicting learning in various situations. However, such studies have not required abandonment of the pursuit of general theories of learning. In fact, recent experiments illustrate that detailed empirical investigation of a biological constraint on learning often leads to formulation of new general principles of behavior.

11.

Constraints on Positive Reinforcement

Although positive reinforcement is remarkably effective in increasing the probability of a variety of responses, notable exceptions to this general rule have been evident from the inception of research on instrumental conditioning. Thorndike (1911) observed, for example, that cats will readily manipulate a latch or a string to be released from a puzzle box, but licking and scratching responses are much more difficult to maintain with reinforcement. Other famous examples of constraints on positive reinforcement have been described by Breland and Breland (1961). In one experiment, they tried to reinforce a raccoon for putting coins into a box and found that the animal would engage in excessive manipulation of the coins but would not release them into the box. The Brelands called this activity “misbehavior” because it violated conventional predictions about reinforcement: The behavior occurred even though it was not required for reinforcement and in fact postponed and sometimes prevented delivery of the reinforcer. Such constraints on learning may occur because the responses in question cannot be conditioned by the reinforcer (an associative deficit) or because the procedures used lead to activities that interfere with increases in the reinforced behavior (a performance deficit). Most of the recent research on constraints on positive reinforcement has been devoted to exploring various performance deficits that might contribute to the observed findings. However, associative deficits have been also discussed. A.

“MISBEHAVIOR” IN DEPOSITING TOKENS REINFORCEMENT

FOR POSITIVE

Recent research suggests that “misbehavior” in depositing tokens for positive reinforcement reflects performance factors rather than a limitation on association learning. Boakes, Poli, Lockwood, and Goodall (1978) ob-

218

Michael Domjan

served “misbehavior” in rats required to carry and release a ball bearing into a hole for food or water reward and suggested that the phenomenon arises from a conflict between behavior maintained by instrumental and stimulus-reinforcer contingencies. The instrumental contingency required subjects to release the ball. The stimulus-reinforcer contingency arose from pairings of the ball with delivery of food or water. Such pairings presumably conditioned the ball to become a food or water incentive stimulus, with the result that subjects tended to approach and maintain contact with the stimulus (cf. Hearst & Jenkins, 1974). Timberlake, Wahl, and King (1982) investigated more explicitly the role of stimulus-reinforcer relations in the acquisition of “misbehavior” directed toward ball bearings in rats. The floor of their experimental chamber had a groove in which ball bearings could roll from one end of the chamber to the other. They found that instrumental contingencies are not necessary to produce activities very similar to what Breland and Breland (1961) originally described as “misbehavior.” Rather, it appears that such activities are the result of Pavlovian pairings of the ball bearings with food. When contact with the ball was required for food reinforcement, the stereotypy and reliability of the ball-directed behavior increased. However, instrumental and Pavlovian procedures produced fundamentally similar interactions with the ball bearings. Furthermore, many aspects of the results were inconsistent with the idea that instances of misbehavior are the product of adventitious instrumental reinforcement. Although the results obtained by Timberlake et al. (1982) strongly implicate Pavlovian conditioning as the mechanism responsible for the development of misbehavior, their findings are not consistent with the traditional stimulus-substitution view of classical conditioning. Conditioned responses directed toward the ball bearings tended to be more time consuming and elaborate than responses directed toward the food pellets. A strict stimulus substitution view of classical conditioning has been also found to be inadequate in explaining various other instances of classical conditioning (e.g., Timberlake, 1982; Timberlake & Grant, 1975; Wasserman, 1973). The results of Timberlake et al. (1982) are consistent with a broader conception that assumes that through conditioning with food, the ball bearings come to elicit various responses that are part of the subject’s innate appetitive behaviors related to obtaining and handling food (e.g., carrying and chewing). As is true for other recently documented instances of Pavlovian conditioning (e.g., Holland, 1977), the nature of the responses elicited by the ball bearings will be in part determined by the nature of the ball stimuli. Consistent with this view, Boakes et al. (1978) found more extensive chewing of nylon than of stainless steel balls in their study of misbehavior.

Biological Constraints on Learning

219

Procedures in which subjects are required to deposit tokens in a slot t o obtain reinforcement (Boakes et al., 1978; Breland & Breland, 1961) are in some ways ideally suited for observation of classically conditioned responses directed to the tokens. In such procedures, the food or water reinforcer is not delivered until the subject has let go of the token. Therefore, responses directed toward the reinforcer do not occur until after the tokenrelated activities take place and cannot interfere with these behaviors. The view that misbehavior results from stimulus-reinforcer contingencies in the absence of competition from responses directed toward the reinforcer (Timberlake et al., 1982) attributes instances of misbehavior to generally applicable principles of behavior. This analysis suggests that misbehavior observed in token reinforcement experiments can be accommodated with modifications of certain generally applicable mechanisms and does not require abandonment of the pursuit of general theories of learning. B.

DIFFERENTIAL CONDITIONABILITY OF VARIOUS RESPONSES WITH POSITIVE REINFORCEMENT

The most systematic recent demonstrations and analyses of the differential conditionability of various responses with positive reinforcement have been conducted by Shettleworth and her collaborators using the golden hamster. In one experiment (Shettleworth, 1975, Experiment 4) fooddeprived hamsters were reinforced for performing one of six action patterns. Reinforcement of open rearing, scrabbling, and digging produced large increases in the amount of time subjects spent performing these responses. A much smaller increase occurred in face washing, and scratching and scent marking were hardly influenced at all by reinforcement. The action patterns that showed the strongest effects of reinforcement also increased in bout length during the reinforcement phase, whereas the bout lengths of action patterns that were not greatly increased by reinforcement (face washing, scratching, and scent marking) decreased. Generally, if an action pattern was not reinforced, it remained unchanged or decreased. The only exceptions to this general pattern were that reinforcement of scent marking increased digging, and reinforcement of face washing increased open rearing. Results comparable to those obtained by Shettleworth have been also observed with rats. For example, Pearce, Colwill, and Hall (1978) reported that food reinforcement was much more effective in increasing lever pressing than scratching in rats. Annable and Wearden (1979) compared the effects of reinforcing various types of grooming in rats with food and found that rates of paw washing and body washing increased when these responses were required for reinforcement in independent groups. However, reinforcement did not increase the rate of face washing. Efforts to explain these

220

Michael Domjnn

and other constraints on instrumental conditioning have involved consideration of a variety of performance and association factors. I shall first discuss performance variables that may limit certain types of instrumental behavior and shall then consider factors that may limit association learning with positive reinforcement.

1. Behavioral Repertoire Limited by Unconditioned Motivational States One prominent performance variable that may constrain instrumental behavior is the subject’s motivational state. Instrumental conditioning often involves motivating the subject in some way, and the motivational state of the organism may limit the type of activities it is likely to perform. Consistent with this idea, Shettleworth (1975, Experiment 1) found that responses that are difficult to reinforce with food (face washing, scratching, and scent marking) were decreased by food deprivation, whereas responses that are readily increased by reinforcement (digging, open rearing, and scrabbling) were increased by food deprivation. However, other evidence suggests that the responses elicited by unconditioned motivational states cannot account for all aspects of constraints on instrumental conditioning. One implication of the motivational state hypothesis is that constraints on learning should be less evident with reinforcers, such as electrical brain stimulation, that do not require experimentally induced deprivation states. Shettleworth and Juergensen (1980) investigated the reinforcement of various action patterns in nondeprived hamsters with electrical stimulation of the medial forebrain bundle. Even though food deprivation was not used, the same pattern of differences in the instrumental conditionability of various action patterns was observed as had been found with food reinforcement. Electrical stimulation of the brain was more effective in reinforcing digging, open rearing, and scrabbling than in reinforcing face washing, scratching, or scent marking. In another study, sunflower seeds and nest paper were used to reinforce open rearing, scrabbling, and face washing in nondeprived hamsters (Shettleworth, 1978a). Open rearing and scrabbling were again substantially increased by the instrumental conditioning procedures. However, little, if any, reinforcement effect was obtained with face washing. Although it is difficult to reach definitive conclusions on the basis of comparisons of the effectiveness of different reinforcers (Hogan & Roper, 1978), these studies with nondeprived subjects suggest that limitations in the subject’s response repertoire produced by experimentally induced deprivation states are not entirely responsible for constraints on instrumental conditioning.


22 I

2. Behavioral Repertoire Limited by Conditioned Motivational States Competing responses that interfere with increases in reinforced action patterns such as face washing, scratching, and scent marking may also arise from conditioned motivational states. In all instrumental conditioning procedures, the reinforcer is delivered only contingent on performance of a specified activity. Therefore, the various exteroceptive and proprioceptive stimuli that accompany the target response may become classically conditioned by the reinforcer (e.g., Bindra, 1974, 1978). These conditioned incentive stimuli may in turn elicit responses that are incompatible with performance of the instrumental behavior, thus constituting a constraint on learning. a. Effects of a Noise Conditioned with Food. In one investigation of the types of action patterns that are elicited by classically conditioned food incentive stimuli, a pulsating white noise conditioned stimulus (CS) was repeatedly paired with food in hamsters (Shettleworth, 197813). Following conditioning, the noise produced different effects on the various activities of the hamsters. However, the results did not correspond to the outcome of instrumental conditioning studies. Of the responses that are readily increased by food reinforcement (digging, scrabbling, and open rearing), presentation of the food-conditioned noise did not influence digging, descreased scrabbling, and increased open rearing. Particularly distressing for the hypothesis under consideration, grooming and scrabbling were both decreased by the classically conditioned noise, even though of these two action patterns only grooming is difficult to increase with food reinforcement. 6. Effects of the Anticipation of Daily Feeding. It might be suggested that responses elicited by a food-conditioned noise do not reflect the role of classical conditioning in instrumental procedures because in the typical instrumental procedure the classically conditioned stimuli are not as explicit as a discrete exteroceptive noise. More relevant evidence might be obtained from studies that d o not employ explicit exteroceptive conditioned stimuli. In one such study (Shetleworth, 1975, Experiment 2), hamsters were observed just before their daily feeding, immediately after feeding, and several hours after feeding. The anticipation of food increased environment-directed activities such as gnawing, open rearing, wall rearing, and climbing, and decreased self-maintenance activities such as grooming and lying or standing in the nest. These observations are compatible with the proposition that environment-directed activities (digging, scrabbling, open rearing) are readily increased by food reinforcement because they are compatible with responses that occur during the anticipation of food, whereas self-directed activities such as grooming are difficult to reinforce because they are de-

222

Michael Domjan

creased by the anticipation of food. However, these results do not help in trying to understand why scent marking is difficult to increase with reinforcement. Scent marking occurred only rarely in the study and was not altered by the anticipation of food. c. Effects of Fixed-Time Presentations of the Reinforcer. Another approach to the study of the role of classical conditioning in the differential conditionability of instrumental responses involves observing the effects of periodic presentations of the reinforcer. If food, for example, is delivered independently of behavior every 30 sec, temporal conditioning may occur, with temporal cues near the end of the 30-sec period coming to signal the impending presentation of food, and temporal cues earlier in the interval coming to signal the unavailability of food. The effects of such fixed-time food delivery schedules were first extensively investigated by Staddon and Simmelhag (197 1) in pigeons. They found that certain activities (labeled “terminal responses”) came to predominate near the end of the interval between food presentations, and other activities (labeled “interim activities”) had their peak probability earlier in the interval. Based in part on the distinction between terminal and interim responses, Staddon and Simmelhag (197 1) suggested that instrumental conditioning procedures simply select the response increased by reinforcement from the available activities of the subject. Since the subject’s behavior at the time of an impending delivery of the reinforcer is characterized by terminal responses, these responses should be much more amenable to instrumental reinforcement than interim responses. Anderson and Shettleworth (1977) tested the above prediction with hungry hamsters given food every 30 sec independent of behavior. In general the results supported the proposition that instrumental reinforcement selects from available terminal responses. Digging and scrabbling (which were previously shown to rapidly increase with food reinforcement) were terminal responses, and grooming action patterns (which were previously shown to be difficult to reinforce) were interim responses. However, the correspondence between terminal and interim responses and instrumental conditionability was not perfect. The most notable discrepancy occurred with open rearing. Open rearing was predicted to appear as a reliable terminal response but was observed only in some of the subjects, and then as an interim response. Another experiment evaluated the effects of response-independent fixedtime deliveries of brain stimulation reinforcement in hamsters (Shettleworth & Juergensen, 1980). The terminal response category again included action patterns that were easily increased by instrumental reinforcement (open rearing and scrabbling), and the interim response category included action patterns that were resistant to the effects of instrumental reinforcement (face


223

washing). However, not all of the data were consistent with the view that the distinction between terminal and interim responses allows predicting the outcome of instrumental conditioning. Digging, scratching, and scent marking all occurred infrequently and without temporal patterning, suggesting these three responses should be equally susceptible to reinforcement. However, digging is much more readily increased by instrumental conditioning than is scratching or scent marking. d . Summary of Classical Conditioning Effects. The above experiments provide some evidence of a correspondence between responses that occur because of classically conditioned anticipation of the reinforcer and the instrumental conditionability of various action patterns. The evidence is clearest for grooming responses (face washing, scratching). Grooming is probably difficult to increase with instrumental procedures because stimuli classically conditioned by the reinforcer usually suppress grooming. Grooming was suppressed by a noise that had been classically conditioned with food, it was suppressed immediately preceding daily feedings, and it occurred as an interim rather than a terminal response with periodic deliveries of food or brain stimulation reinforcement. These experiments did not help isolate the source of constraints on scent marking because scent marking occurred rarely in the experiments and was not altered by the various classical conditioning manipulations. The experiments also did not help clarify why digging, open rearing, and scrabbling are readily increased by reinforcement; these responses did not always occur as classically conditioned anticipations of the reinforcer. A likely possibility is that classical conditioning mechanisms are only partly responsible for the differential conditionability of various instrumental responses.

3. Supporting Stimulation Another variable that may limit performance of instrumental behavior is the absence of supporting stimulation. Supporting stimulation may be particularly important for a response such as scratching that typically occurs in reaction to some type of cutaneous irritation. This suggests that the instrumental conditioning of scratching may be facilitated by skin irritants that promote scratching. Consistent with this prediction, Konorski (1967) noted that scratching is increased by food reinforcement in cats if the subjects have a piece of cotton in their ears to create irritation. More recently, Pearce et al. (1978) observed that cutaneous stimulation provided by a plastic collar greatly facilitated instrumental conditioning of scratching in rats. Removal of the collar after acquisition resulted in a rapid decrease in the reinforced behavior despite continued reinforcement. This suggests that the

224

Michael Domjan

collar was primarily involved in supporting performance of food-reinforced scratching. The importance of supporting stimulation for performance of instrumental behavior has been also documented in studies of the instrumental conditioning of licking in rats (Young & Black, 1977). In one experiment, licking a drinking tube was reinforced using a discriminated shock-avoidance procedure. The rate of discriminative instrumental licking was influenced by the nature of the solution that was provided in the drinking tube and by the subjects’ state of deprivation. A 10% sucrose solution in the drinking tube supported higher response rates than deionized water, and higher rates of discriminative licking occurred when subjects were water deprived than when they were water satiated. In a related experiment, water deprivation and sucrose in the drinking tube similarly increased the rate of licking for food reinforcement. However, because this experiment did not involve a discrimination procedure, it is not clear whether the results reflected licking for food reinforcement or just licking for the fluid that was provided in the drinking tube. The studies by Pearce et al. (1978) and Young and Black (1977) demonstrate that supporting stimulation can be important in the performance of instrumental responses, but such positive results have not been always obtained. Shettleworth (1978a) observed that spraying hamsters with a light mist of water elevated baseline rates of face washing, as expected, but did not facilitate instrumental conditioning of face washing. In another study (Shettleworth & Juergensen, 1978), placing subjects in the home cage of an unfamiliar conspecific increased baseline rates of scent marking without substantially improving the instrumental conditionability of the response. These negative findings indicate that only some instances of constraints on instrumental behavior can be understood in terms of a lack of required supporting stimulation.

4. Role of Innate Behavioral Sequences

A particular response may be difficult to increase with instrumental reinforcement if it is part of an innate behavioral sequence and cannot occur independent of other component actions. Strong innate sequential organization of behavior may constrain independent variation of component responses when one or another component is reinforced. Not much evidence relevant to this hypothesis is available. However, a recent study by Annable and Wearden (1979) can be interpreted in these terms. Independent groups of rats were reinforced with food for performing various components of grooming: paw washing (licking of the paws), face washing (bringing paws over the ears and down over the snout), or body


225

washing (bending the head forward into the belly). Reinforcement of paw washing and body washing increased these responses without causing increases in nonreinforced grooming responses. In contrast, reinforcement of face washing did not increase the rate of face washing but resulted in more frequent paw washing. This effect may have occurred because of the sequential organization of grooming. Rats may always begin grooming with paw washing, followed by face washing. If paw washing is strictly preliminary to face washing, reinforcement of face washing would not be expected to increase face washing independent of paw washing. In fact, reinforcement of face washing may be expected to increase paw washing, as was observed. Sequential organization of behavior may have been also responsible for observed relations between scent marking and digging in hamsters. Reinforcement of scent marking with either food or electrical brain stimulation produced very little, if any, increase in scent marking but was accompanied by increases in digging (Shettleworth, 1975; Shettleworth & Juergensen, 1980). This may have occurred because digging is preliminary to scent marking in the activities of hamsters.

5. Differences in Operant Level Conditioned and unconditioned motivational states, absence of supporting stimulation, and innate behavioral sequences may constrain performance of instrumental behavior but do not necessarily limit the learning of response-reinforcer contingencies. One variable that may constrain instrumental learning is the operant level of a response. None of the studies of constraints on instrumental conditioning involved reinforcement of successive approximations to the target responses. Therefore, the frequency with which a response was reinforced depended on the frequency of its occurrence when the reinforcement procedure was introduced. However, differences in the operant level of various responses cannot account for the observed constraints on instrumental conditioning. In many of the studies, responses that were not increased by reinforcement were not less likely to occur prior to introduction of the conditioning procedures than easily modified action patterns (e.g., Annable & Wearden, 1979; Charlton & Ferraro, 1982; Pearce et al., 1978: Shettleworth, 1975, 1978a). In other studies a systematic relationship was not found between the operant level of a behavior and its susceptibility to reinforcement (Shettleworth & Juergensen, 1980). In still other work, instrumental conditioning of an action pattern such as face washing was not facilitated by procedures (e.g., spraying the subjects with a mist of water) that substantially elevated the operant level of the behavior (Shettleworth, 1978a; see also Shettleworth & Juergenson, 1980, Experiment 5).

226

Michael Domjan

6. The Discriminability of Instrumental Behavior Some responses may be difficult to increase with instrumental reinforcement because their performance is not sufficiently discriminable. If a subject cannot readily discriminate when it performs a response, the behavior may not be available for association with reinforcement. Morgan and Nicholas (1979) evaluated this possibility with rats. One group of subjects was required to discriminate between face washing and open rearing. Following performance of one or the other of these responses, two response levers were inserted into the experimental chamber. Pressing one of the levers was reinforced with food following face washing, and pressing the other lever was reinforced following open rearing. A second group of subjects was required to discriminate between face washing and scratching. Choice behavior did not exceed chance levels following a scratch response. In contrast, subjects were very accurate in their selection of the correct response lever following instances of open rearing and also gradually learned to make the correct choice more than 50% of the time following instances of face washing. These results clearly indicate that scratching is much less discriminable to rats than face washing and open rearing. However, the study did not provide independent evidence of corresponding differences in the instrumental conditionability of these responses. Face washing and rearing increased more than scratching during the course of the experiment, but this effect may have been due to the fact that subjects were more likely to select the correct lever after face washing and rearing and were therefore more likely to get food reinforcement following these responses than following scratching. 7. Historical Response-Reinforcer Independence

Constraints on instrumental conditioning may also result from the previous history of a subject. Mackintosh (1974) suggested that self-care responses such as face washing may be difficult to increase with reinforcement because such responses are unrelated to procurement of food and water reinforcers during a subject’s lifetime. Research has shown that responsereinforcer independence can interfere with subsequent instrumental conditioning (e.g., Maier & Jackson, 1979; Maier & Seligman, 1976). Whether this type of mechanism is responsible for any of the constraints on instrumental conditioning described above remains to be investigated. 8. Concluding Comments

Research on constraints on positive reinforcement has highlighted the importance of a variety of performance and associative factors that may


227

influence instrumental behavior in addition to the response-reinforcer contingency. Conditioned and unconditioned motivational states, presence or absence of supporting stimulation, and innate behavioral sequences may all provide constraints on the performance of instrumental behavior. Constraints on association learning in instrumental conditioning may be provided by a low operant level, poor response discriminability, and historical response-reinforcer independence. Research has provided considerable support for explanation of constraints on positive reinforcement in terms of performance factors. Additional work is required to obtain clear evidence of associational constraints. Consideration of performance and associational variables indicates how constraints on positive reinforcement may be explained in terms of general mechanisms of learning and behavior.

111. Constraints on Punishment As is the case with positive reinforcement, not all responses are equally susceptible to the effects of punishment. Shettleworth ( 1 9 7 8 ~ for ) ~ example, observed that in golden hamsters punishment consistently suppressed scrabbling. Less consistent results occurred with open rearing. The total time spent open rearing and bout lengths of open rearing were suppressed by punishment in all experiments, but in some studies the number of bouts of open rearing was unaffected or only temporarily suppressed by punishment. The effects on face washing were even more unexpected. Although the total time and bout lengths of face washing were always suppressed by punishment, the number of bouts of face washing was increased by punishment in three experiments and remained unchanged in a fourth. The susceptibility of various responses to punishment evidently does not reflect their general modifiability or “voluntariness.” Open rearing in hamsters was less influenced by punishment than scrabbling even though the two responses were both easily increased by various types of positive reinforcement in other studies (see Section 11). More convincing evidence exists indicating that some responses are difficult to suppress with punishment because they are elicited by aversive stimulation. Open rearing is probably less susceptible to punishment than scrabbling in part because open rearing is increased by aversive stimulation (response-independent brief shocks and a shock-conditioned noise) whereas scrabbling is not (Shettleworth, 1978~). Instances in which punishment increases the punished responses are probably also due to the fact that the response in question is elicited by aversive stimulation (e.g., Melvin & Ervey, 1973; Walters & Glazer, 1971). Although some of the constraints on punishment can be understood by considering what responses are elicited by aversive stimulation or the an-

228

Michael Domjan

ticipation of aversive events, other variables are no doubt also important. For example, increases in the number of bouts of face washing produced by punishment cannot be explained in this way because face washing is not increased by response-independent shocks or presentations of a shock-conditioned stimulus (Shettleworth, 1978b,c). Recent research indicates that some constraints on punishment also reflect an associative rather than a performance deficit. Shettleworth (1981) measured the strength of response-shock associations in hamsters in terms of the ability of a punished response to interfere with conditioning of a tone that was presented between occurrences of the response and consequent shock (cf. Kamin, 1969). Scrabbling was suppressed by punishment, and occurrences of scrabbling blocked conditioning of the tone stimulus. In contrast, punishment did not suppress open rearing, and open rearing did not block conditioning of the tone. Further research is required to determine the bases of such associative deficits.

IV. Constraints on Avoidance Learning Some of the most dramatic examples of constraints on instrumental conditioning are found in avoidance learning. Depending on what response is required to prevent aversive stimulation, avoidance learning may occur very rapidly, at an intermediate rate, or very slowly (see reviews by Bolles, 1970, 1971). Rats, for example, can learn to jump out of a shock box to avoid shock in one trial (Maatsch, 1959) but have much more difficulty in learning to press a response lever (e.g., D’Amato & Schiff, 1964; Meyer, Cho, & Wesemann, 1960). Running in a running wheel is acquired more rapidly as avoidance behavior than rearing (Bolles, 1969), and remaining still is learned more rapidly than movement (Brener & Goesling, 1970). In pigeons, oneway locomotion, general activity, and treadle pressing can all be conditioned as avoidance behavior (Foree & LoLordo, 1970; Graf & Bitterman, 1963; Macphail, 1968; Smith & Keller, 1970). In contrast, pecking a key is much more difficult to condition (e.g., Hoffman & Fleshler, 1959; Schwartz, 1973; Smith & Keller, 1970). A. TECHNIQUES THATFACILITATE AVOIDANCE CONDITIONING

The difficulties encountered in avoidance conditioning of certain responses have stimulated both technological and theoretical developments. Numerous efforts have been made to find procedures that facilitate lever pressing in rats and pecking in pigeons as avoidance behavior. In one successful procedure pigeons were exposed to a train of shock pulses of grad-


229

ually increasing intensity which was reduced to zero contingent on key pecking (Rachlin & Hineline, 1967). In an elaboration of this procedure (Ferrari, Todorov, & Graeff, 1973), successive approximations to pecking were reinforced by reductions in shock intensity, and competing responses (such as attempts to escape from the chamber) were met with increased shock. Another technique that facilitated key pecking as avoidance behavior involved pretraining subjects to peck for food reinforcement (Foree & LoLordo, 1974; Lewis, Lewin, Stoyak, & Muehleisen, 1974). Food reinforcement pretraining has been also found to facilitate shockavoidance lever pressing in rats (Giulian & Schmaltz, 1973; Kulkarni & Job, 1970; Riess, 1970). Lever-press shock avoidance is also facilitated by punishment of inactivity and bar holding (Feldman & Bremner, 1963), delivery of a brief response-independent shock shortly (2.5 sec) before the warning stimulus (Delprato & Holmes, 1977), permitting subjects t o escape temporarily from the shock environment contingent on lever pressing (Crawford & Masterson, 1978), and using a strong shock (2 mA) and a relatively long interval (60 sec) between presentation of the warning signal and subsequent shock (Berger & Brush, 1975). B.

SPECIES-SPECIFIC DEFENSERESPONSES AVOIDANCE LEARNING

AND

Although the above studies demonstrate that under certain circumstances pigeons can learn t o peck and rats can learn to bar-press to avoid shock, the experiments do not contradict the original observations of differential effectiveness of avoidance contingencies in conditioning various responses. The dominant theoretical interpretation of constraints on avoidance learning is provided by a theory proposed by Bolles (1970, 1971). The starting point of the theory is that delivery of aversive stimulation restricts the subject’s range of activities to species-specific defense responses (SSDRs), such as flight, freezing, and aggression. The relative probabilities of these responses depend on the circumstances. In the absence of an obvious escape route, freezing may be the most likely response; in the presence of a prominent target for aggression (such as a conspecific), some form of aggression may occur. The theory assumes that only SSDRs can be rapidly acquired as avoidance responses. Furthermore, avoidance procedures are assumed to select from available SSDRs by punishment rather than negative reinforcement. If the most likely SSDR (or some close approximation of it) is not effective in satisfying the avoidance contingency, this behavior will be punished and suppressed by the unavoided shocks, allowing the next most likely SSDR to predominate. Responses that are not a part of the subject’s species specific defense repertoire are difficult to increase with avoidance

230

Michael Domjan

procedures because SSDRs first have to be suppressed before a nondefensive response can occur frequently. Although species-specific defense responses have not been extensively investigated within the context of laboratory avoidance conditioning research, some evidence is available. Freezing has been observed in rats as a defensive response to conditioned aversive situational cues (Blanchard & Blanchard, 1969a), and withdrawal has been observed as a defensive response to a localized conditioned aversive stimulus (a shock prod) (Blanchard & Blanchard, 1969b). An unconditioned aversive stimulus (exposure to a cat) elicits freezing in rats when escape is not possible and fleeing when an escape route is provided (Blanchard & Blanchard, 1971). Observations of differential effectiveness of various responses in avoidance situations are compatible with the SSDR theory. For example, rapid acquisition in rats of jumping out of a shock box to avoid subsequent shocks can be explained by assuming that the jump-out response is a component of the fleeing SSDR in this situation. Poor acquisition of lever pressing in many experiments can be explained by pointing out that this response is not an SSDR. In addition, freezing is assumed to occur as an SSDR in leverpress avoidance situations, and lever pressing is incompatible with this freezing (Delprato & Holmes, 1977; Feldman & Bremner, 1963). Running in a wheel is assumed to be more easily learned as an avoidance response than rearing (Bolles, 1969) because running is closer to the SSDR of fleeing. Poor acquisition of pecking in pigeons as avoidance behavior is explained by assuming that pecking is not an SSDR but is part of the behavioral system involved in obtaining food (Schwartz, 1973). Certain aspects of techniques used to facilitate lever pressing in rats and key pecking in pigeons as avoidance behavior may be also understood in terms of the SSDR theory. Pretraining subjects to perform these responses for food reinforcement may facilitate acquisition and maintenance of leverpress and key-peck avoidance by reducing the aversiveness of the situation when the avoidance procedure is introduced so that fewer incompatible SSDRs are elicited. Appetitive pretraining may disrupt aversive conditioning of situational cues (cf. Dickerson & Pearce, 1977) and may reduce the number of shocks that subjects receive by increasing rates of lever pressing and key pecking at the start of avoidance training. Punishment of freezing and bar holding probably facilitates lever press avoidance learning by suppressing the freezing and bar holding that occur as SSDRs in the situation. The delivery of a response-independent shock shortly before the warning stimulus may also facilitate bar-press avoidance by disrupting competing freezing behavior. Although the above accounts of various aspects of avoidance learning are plausible in terms of the SSDR theory, the explanations are post hoc.


23 1

Better evidence is provided by studies that test predictions of SSDR theory. The theory predicts that remaining still will be easily acquired and maintained as avoidance behavior when escape is not possible. Consistent with this prediction, Brener and Goesling (1970) found that rats learn to remain still to avoid shock more quickly than they learn to engage in general activity. Bolles and Riley (1973) also reported that freezing is readily increased and maintained by an avoidance procedure, and Davis and Burton (1976) found effective avoidance conditioning of a form of freezing-continuously holding down a response lever. The SSDR theory not only predicts rapid acquisition of a defensive response such as freezing but makes the added claim that increased freezing will develop not because it is effective in preventing shocks but because it is elicited by shock. Various aspects of the experiments by Bolles and Riley (1973) confirm this claim. With the exception of the experiment by Bolles and Riley (1973), the above studies of freezing did not document freezing as an SSDR in the absence of the avoidance procedures used. The strongest test of the SSDR theory is provided by studies in which the nature of SSDRs are first established with response-independent aversive stimulation, and predictions are then made about what responses will be rapidly or slowly acquired during avoidance conditioning. In one such study (Grossen & Kelley, 1972), responseindependent shock caused rats to freeze near the side walls of a large test arena. Subsequently, rats that were required to jump on a platform placed near the side walls to avoid shock learned the avoidance response faster than rats that were required to jump on a platform placed in the center of the arena. Thus, avoidance performance was accurately predicted by the subjects’ reactions to response-independent aversive stimulation. Furthermore, the results were specific to aversive motivation. The position of the platform did not make a difference when subjects were reinforced with food for making the jump response.

C. THECONSUMMATORY STIMULUS REWARD HYPOTHESIS The SSDR theory focuses on responses involved in avoidance conditioning experiments. However, different responses also result in different feedback stimulation. Freezing, for example, results in maintained contact with a stable configuration of proprioceptive, vestibular, and environmental stimuli. In contrast, fleeing results in a host of changing visual, acoustical, vestibular, and proprioceptive cues. Bolles was sensitive to these differences and noted that the critical feature of SSDRs was not their topography but their functional effectiveness (Bolles, 1970). However, he did not elaborate on the range of topographical variations possible for SSDRs or pre-

232

Michael Domjan

cisely how feedback stimuli operated. Feedback from fleeing, for example, presumably shapes the topography of the flight behavior. Given the general tenor of SSDR theory, one might assume that such shaping is restricted to a small range of possible fleeing responses. Could the presentation of feedback stimuli that normally accompany flight be used to reinforce a nonbiological defensive response such as lever pressing in rats? This question was addressed by Crawford and Masterson (1978). Crawford and Masterson (1978) trained rats to press a lever using a discriminated avoidance procedure. Upon making the avoidance response, some animals remained in the shock chamber, others were allowed to run to a safe compartment for the intertrial interval, and a third group of rats was carried there. Getting to the safe compartment promoted lever-press avoidance learning whether or not subjects performed the flight response. A second study (Masterson, Crawford, & Bartter, 1978) showed that, although remaining in the safe compartment during the intertrial interval facilitated avoidance performance, getting to the safe area was sufficient to produce rapid avoidance learning even if subjects were returned to the shock box for most of the intertrial interval. The above studies indicate that stimuli that normally accompany an SSDR such as fleeing are very important for the acquisition and maintenance of avoidance behavior. In fact, under certain circumstances, the motor aspects of the flight SSDR did not even contribute to avoidance learning. Reinforcement of nonbiological avoidance responses by flight-related stimuli cannot be explained by the SSDR theory. Therefore, Masterson and his colleagues proposed an alternative view, the consummatory stimulus reward (CSR) hypothesis (Crawford & Masterson, 1978; Masterson et al., 1978; Masterson & Crawford, 1982). The consummatory stimulus reward hypothesis assumes that aversive stimulation induces a defense motivational system which not only increases the probability of certain responses (species-specific defensive responses) but also activates representations of relevant consummatory stimuli. A match between stimuli encountered by the subject and the central representation of relevant consummatory stimuli reduces the motivational state and provides positive reinforcement. The model applies drive-reduction views of food reinforcement to aversive situations. Just as food deprivation is assumed to induce food-seeking responses and make food-related stimuli positively reinforcing, aversive stimuli are assumed to induce defensive responses and make flight-related stimuli positively reinforcing. This added assumption of positive reinforcement by flight-related stimuli permits the instrumental conditioning of nondefensive responses (such as lever pressing) in avoidance situations. The consummatory stimulus reward hypothesis is an attractive addition


233

to the SSDR theory. However, numerous questions remain. We do not know, for example, what aspects of complex flight-related stimuli are critical for reinforcement of avoidance behavior. We also do not know whether stimuli related to other SSDRs such as freezing are similarly rewarding, what factors determine whether flight- or freezing-related stimuli will act as reinforcers in a particular case, whether stimuli can become conditioned to serve as consummatory reward stimuli in aversive situatjons, and what constraints exist on the conditioning of such stimuli.

D. FURTHERINVESTIGATIONS OF STIMULUS FACTORS IN

AVOIDANCE LEARNING

The consummatory stimulus reward hypothesis focuses on the importance of stimuli that follow the avoidance response in avoidance conditioning. Avoidance learning is also often governed by stimuli that signal the impending delivery of aversive stimulation. Just as various stimuli may be differentially effective as reward stimuli for avoidance behavior, various events may be differentially effective as warning stimuli. Furthermore, certain types of warning stimuli may be most effective in combination with particular types of response feedback cues. Frontali and Bignami (1973, 1974) compared the differential effectiveness of visual and auditory cues as warning stimuli in procedures involving a discrimination between active and passive avoidance in rats and found that the two types of cues were not interchangeable. However, the complicated design of their experiments made it difficult to identify the source of group differences. In a more definitive study, Jacobs and LoLordo (1977) compared visual and auditory cues as signals for danger and safety in leverpress avoidance conditioning of rats (see also Jacobs & LoLordo, 1980). The onset of white noise was found to be more effective as a warning stimulus than as a safety signal. In contrast, the termination of white noise, onset of houselights, and offset of houselights were more likely to become conditioned as safety signals than as warning stimuli for shock. These results indicate that various auditory and visual stimuli are not equally effective as warning and safety signals, and stimuli that are especially effective in one context can be much less effective in the other. We do not know precisely why auditory and visual cues are differentially effective as warning and safety stimuli. The studies by Jacobs and LoLordo were encouraged by preliminary observations that sharp sounds and sudden movements in the visual field elicited freezing and running in rats. If running occurred, the rat consistently ran along a wall or vertical surface and toward a dark hole. If it was startled in an open area, the rat ran toward the nearest shaded area. These observations do not permit precise predic-

234

Michael Domjnn

tion of the experimental results that were obtained. However, they loosely implicate the use of auditory cues in the stimulation of defensive behavior and visual cues in the guidance of such behavior. Regardless of the causes of differential utilization of auditory and visual cues, as Jacobs and LoLordo (1977, 1980) point out, these findings help understand previous successes and failures in avoidance conditioning of lever pressing in rats. Investigators who have reported difficulty in conditioning lever-press avoidance (D’Amato & Schiff, 1964; Feldman & Bremner, 1963; Meyer et al., 1960) employed visual warning stimuli. Better acquisition was obtained in studies that used an auditory warning stimulus (Myers, 1959, 1964), and acquisition was even faster when the full complement of supporting stimuli was used, the onset of an auditory cue as a warning stimulus and a visual safety signal (Berger & Brush, 1975). Research has also identified differential effectiveness of various types of stimuli in aversive conditioning of pigeons (Foree & LoLordo, 1973, 1975; LoLordo & Furrow, 1976; LoLordo, 1979; see also Shapiro, Jacobs, & LoLordo, 1980). These studies compared the effectiveness of a 440-Hz tone and red houselights as discriminative stimuli for treadle pressing reinforced with food presentation or shock avoidance. The light gained greater discriminative control over responding than the tone during appetitive conditioning. In contrast, the tone was more effective than the light as a discriminative stimulus for shock avoidance. These findings are similar to the results obtained with rats in showing that in comparison to visual stimuli, auditory cues are more effective as warning stimuli in aversive conditioning (see also Delius & Emmerton, 1978; Schindler & Weiss, 1982). E. CONCLUDING COMMENTS

The research reviewed above indicates that the SSDR theory is incomplete as an explanation of constraints on avoidance conditioning because it does not consider stimulus aspects of avoidance learning. Avoidance learning very much depends on the nature of warning and safety signals. The consummatory stimulus reward hypothesis addresses this issue partially by noting the importance of stimuli that follow avoidance responses in reinforcing the avoidance behavior. However, the hypothesis does not consider the importance of warning stimuli and the fact that various cues are differentially effective as warning stimuli. One is tempted to attribute the differential effectiveness of various cues as warning and safety signals to the evolutionary history of animals and their unconditioned defensive behavior systems. However, much more detailed information about natural defensive behavior is required to substantiate such claims. Cross-species comparisons would be particularly interesting in this context because species differences in natural defensive behavior should be reflected in corresponding species

Biological Constraints on Lenrning

235

differences in laboratory avoidance conditioning performance. Other factors that have to be considered in an analysis of the differential effectiveness of stimuli will be discussed in the following section.

V.

Selective Associations in Classical Conditioning

Research on instrumental avoidance learning has revealed instances in which certain cues are more easily conditioned as warning stimuli than as safety signals, whereas other cues are more easily conditioned as safety signals than as warning stimuli. Selective association effects have been more extensively investigated using classical conditioning procedures in which stimuli are presented independent of behavior. These studies have involved comparisons of the relative effectiveness of two different types of conditioned stimuli (CSl and CS2) in conditioning with two types of unconditioned stimuli (US1 and US2). The outcome of such studies is interpreted as illustrating a constraint on classical conditioning if CS1 is more readily conditioned with US1 than with US2, and CS2 is more readily conditioned with US2 than with US1. Given such results, instances of poor learning (in our example the conditioning of CSl with US2 and CS2 with USl) cannot be attributed to the general ineffectiveness of either the conditioned or unconditioned stimuli involved because the same stimuli are found to be processed well when they participate in other associations (CSI with US1 and CS2 with US2). Thus, the critical evidence for a constraint on learning is the finding that conditioning depends on the particular combination of conditioned and unconditioned stimuli employed rather than on their individual properties (for a further discussion of this issue see LoLordo, 1979; Schwartz, 1974). A.

SELECTIVE AVERSION CONDITIONING SHOCK AND ILLNESS

WITH

One of the first and perhaps most famous examples of a constraint on classical conditioning involves selective aversion conditioning with shock and illness in rats. In a landmark experiment, Garcia and Koelling (1966) found that following conditioning with immediate or delayed shock, rats evidenced more of an aversion to an audiovisual CS than to a taste CS. In contrast, following conditioning with illness induced by radiation exposure or lithium chloride, the rats drank less when licks produced the taste stimulus than when licks produced the audiovisual cue. In a subsequent study (Garcia, McGowan, Ervin, & Koelling, 1968), shock was observed to condition a differential aversion to the size of food pellets but not to their flavor, whereas exposure to X irradiation conditioned a differential aver-

236

Michael Domjan

sion to the flavor of the pellets but not to their size. These results were considered to illustrate the differential processing of interoceptive and exteroceptive stimuli (e.g., Garcia, Hankins, & Rusiniak, 1974). According to this interpretation, conditioned and unconditioned stimuli are more easily associated with one another if they are both exteroceptive (auditory/ visual cues and shock) or both interoceptive (taste and illness). Since the initial demonstrations of selective aversion conditioning with shock and illness, numerous efforts have been made to discover the bases of these results. Some experiments have been concerned with whether the differential aversions were produced by associative or nonassociative processes. Others have questioned whether the effects represent differences in the modality of the stimuli involved or other features such as how the stimuli are presented, the similarity of the conditioned and unconditioned stimuli, and the presumed durations of their stimulus traces. Still other experiments have been concerned with whether selective conditioning is the result of the learning history of the subjects or reflects inborn characteristics of neural functioning. I . Nonassociative Effects of the Unconditioned Stimuli

Results of selective aversion conditioning experiments are difficult to interpret entirely in terms of nonassociative factors. In one study (Domjan & Wilson, 1972a, Experiment 2), for example, all subjects received equal exposure to taste and buzzer conditioned stimuli but had only one of the stimuli paired with either a shock or lithium unconditioned stimulus. Only subjects that had the taste of saccharin paired with lithium learned an aversion to saccharin, and only subjects that had the buzzer paired with shock learned an aversion to the buzzer. Studies of this sort clearly show that taste aversions observed in poisoned animals and audiovisual aversions observed in shocked rats reflect an association between the respective conditioned and unconditioned stimuli. However, they do not prove that the selectivity of the aversion learning is produced by associative processes. Some have suggested that shock and illness unconditioned stimuli may induce different types of orientation and selective attention, and such differential orientations may be responsible for selective conditioning effects (Gillette, Martin, & Bellingham, 1980; Rescorla & Holland, 1976). For example, shock may cause rats to attend or orient selectively to auditory and visual stimuli, and perhaps that is why shocked rats learn selective aversions to audiovisual cues. In contrast, lithium-induced malaise may cause subjects to attend or orient selectively to taste stimuli, and perhaps that is why lithium-injected rats learn selective aversions to tastes. Several approaches have been used to investigate the possible involvement of differential orientations in selective aversion learning. Miller and


231

Domjan (1981a) sought to obtain independent evidence of differential orientations or attention by measuring increased neophobia for taste and auditory/visual cues following foot-shock and lithium treatments in rats (cf. Aitken & Sheldon, 1970; Domjan, 1977). Exposure to shock decreased the subjects’ preference for novel auditory/visual cues but did not influence their reactivity to a novel saccharin solution. In contrast, treatment with lithium decreased preference for a novel saccharin solution but did not influence choice between novel and familiar light and noise stimuli. The above findings are consistent with suggestions that shock produces selective orientation or attention to exteroceptive cues, whereas lithium-induced malaise results in selective attention to taste stimuli. However, the selective sensitization effects were short lasting. Shock increased neophobia for auditory/visual cues for less than 5 min, and lithium increased neophobia for taste for less than 6 hr. Studies of selective aversion conditioning typically employ much longer intervals between conditioning trials and between conditioning and testing. Therefore, it is unlikely that the selective sensitization effects identified by Miller & Domjan (1981a) were responsible for previous reports of selective aversion conditioning (Domjan & Wilson, 1972a; Garcia & Koelling, 1966; Garcia et al., 1968). The differential orientation hypothesis does not specify how differential orientations are manifest in behavior. Therefore, one might argue that the performance measures employed by Miller and Domjan (198 1a) did not adequately reflect possible selective orientations involved in previously reported aversion conditioning experiments. Another approach to this control problem is to design aversion conditioning experiments in a way that precludes interpretation of the results in terms of US-induced differential orientations. One possibility is to equate all subjects in terms of whatever differential orientations might be produced by shock and illness (Rescorla & Holland, 1976). If all subjects have equal exposure to shock and to lithium, they should be equivalent in terms of increased attention to exteroceptive and taste cues. In an application of this strategy, Miller and Domjan (1981b) administered shock and lithium to the same animals on successive days, but had only one of the unconditioned stimuli paired with a CS. Aversions to a light CS were evident only in animals that had the light paired with shock, and aversions to a taste CS occurred only in animals that had the taste paired with lithium. Thus, the selective aversion conditioning effect was obtained even though all animals received the identical pretest history of exposure to the shock and lithium unconditioned stimuli. 2. Methods of Presentation of Conditioned Stimuli The research reviewed above indicates that selective aversions do not occur because of differential orientations to the conditioned stimuli caused

238

Michael Domjan

by the unconditioned stimuli. Differential orientation to conditioned stimuli may be also caused by differences in how the stimuli are presented. In the study by Garcia et al. (1968), for example, subjects experienced taste stimuli contingent on eating food pellets, but exposure to the exteroceptive cues (size of the food pellets) did not require ingestion. Therefore, differences in aversion learning to taste and exteroceptive cues might have reflected differential involvement of ingestion in exposure to the two types of stimuli. In other demonstrations of selective aversion conditioning in rats, the method of presentation of taste and auditory/visual cues was carefully equated by presenting both types of stimuli contingent on licking a drinking tube (Garcia & Koelling, 1966; Miller & Domjan, 1981b) or independent of behavior (Domjan & Wilson, 1972a). The studies by Garcia and Koelling (1966), Miller and Domjan (1981b), and Domjan and Wilson (1972a) make it unlikely that selective aversion conditioning in rats results from differences in the method of presentation of the conditioned stimuli. However, recent research indicates that the method of stimulus presentation may produce selective conditioning effects in infant chicks (Gillette et al., 1980). The conditioned stimulus consisted of either a distinctive color or flavor added to particles of solid food or to water. The chicks readily learned an aversion to a distinctive color paired with lithium toxicosis when the color was added to solid food, but did not learn an aversion when a distinctive flavor was added to the food. In contrast, they learned a significant flavor aversion when the flavor was added to water, but did not learn an aversion when the water had a distinctive color. Differences in the conditioning of color and taste aversions were related to differences in stimulus orientation during eating and drinking. When the chicks ate, they used an aim-and-peck response sequence in which they appeared to maintain continuous visual contact with the food particles. In contrast ,during drinking, visual orientation toward the water occurred only with initiation of the drinking response. The continuous visual contact that occurred when subjects were eating may have contributed to the selective conditioning of food color but not food taste with toxicosis. In contrast, the minimal visual orientation when the birds were drinking may have contributed to selective conditioning of the taste cues in the drinking experiments. Consistent with this interpretation, the chicks acquired stronger aversions to the color of water when the water was presented in a way that required closer visual orientation. 3. Stimulus Similarity

Another factor that may lead to selective associations independent of the specific modality of the stimuli involved is the similarity between the con-


239

ditioned and unconditioned stimuli. Some have suggested that tastes are readily conditioned with toxicosis and audiovisual cues are readily conditioned with shock because these pairs of stimuli have similar temporalintensity patterns (e.g., Garcia & Koelling, 1966; Testa, 1974; Testa& Ternes, 1977). The sensations of taste and toxicosis may be characterized as starting slowly, persisting for a relatively long period, and ending slowly. In contrast, audiovisual and shock stimuli used in aversion conditioning experiments are usually brief and abruptly introduced and terminated. Although evidence for the importance of stimulus similarity in conditioning has been obtained in other conditioning situations (see below), strong evidence for stimulus similarity in the selective conditioning of taste and audiovisual aversions is unavailable. a. Duration and Pattern of Unconditioned Stimuli. The stimulus-similarity hypothesis suggests that shock-conditioned taste aversion learning should be facilitated by increasing the duration of shock to more closely resemble the duration of illness. Attempts to do this have not been successful (e.g., Green, Bouzas, & Rachlin, 1972), probably because it is difficult to specify what parameters foot shock should have to mimic the temporal-intensity pattern of illness. Furthermore, it seems doubtful that such information will be readily forthcoming. Drug and radiation have multiple physiological effects with varying temporal-intensity patterns, and which aspect of illness is critical for producing a taste aversion has not been isolated. Therefore, we do not know which physiological index of illness should be followed in designing a shock schedule that will be similar to the temporal-intensity pattern of illness. Although the effective temporal-intensity pattern of illness is difficult to identify, indirect support for the stimulus similarity hypothesis is provided by some evidence that taste-aversion learning is influenced by the duration and pattern of drug-induced malaise. Goudie and Dickins (1978) found that the strength of taste aversions conditioned with exposure to nitrous oxide was directly related to the duration of inhalation of the drug. In another study, the pattern of illness caused by lithium chloride was manipulated by varying the schedule of drug administrations (Domjan, Foster, & Gillan, 1979). Rats received 3.0 meq/kg lithium in either one injection immediately after exposure to a saccharin solution or in two half-dose injections separated by 35, 70, or 140 min. Spacing of the drug administrations facilitated aversion learning when a strong taste solution was used as the CS (1.0% saccharin) or when subjects were allowed to drink substantial amounts (10 ml) of a weak flavor (.15% saccharin). Another approach that has been used to study the role of the duration of illness in taste-aversion learning involves comparing the effectiveness of certain drugs and their long-acting analogs as aversion-inducing agents. D’Mello, Goldberg, Goldberg, and Stolerman (1979, 1981), for example,

240

Michael Domjnn

have used this strategy in comparing cocaine with the longer acting analog Win 35,428 and apomorphine with the longer acting diisobutyrylapomorphine (DBA). Such comparisons have failed to provide strong evidence that drug duration facilitates taste-aversion learning independent of drug dose, and Stolerman and D’Mello (1981) concluded that “even marked increases in duration of action d o not necessarily have a specific effect on potency in [conditioned taste aversions]” (p. 202). b. Stimulus Similarity in First-OrderFear Conditioning. Although the development of the stimulus similarity hypothesis was encouraged by selective aversion learning to taste and audiovisual cues paired with malaise and foot shock, the best evidence for the hypothesis comes from other conditioning preparations. One of the first studies designed to evaluate the hypothesis employed two types of air blast as unconditioned stimuli and two types of visual cues as conditioned stimuli (Testa, 1975). The air blast was either pulsed and originated from the ceiling or was presented in a waveform and originated from the floor. Similarly, the visual CS was either pulsed and emanated from the ceiling or had a waveform pattern and emanated from the floor. Rats learned stronger aversions to a visual CS if the CS and US were similar in location and temporal pattern than if the temporal pattern and spatial location of the CS and US were different. However, the experiment did not include procedures to rule out possible sensitization effects. c. Stimulus Similarity in Second-Order Conditioning. The most systematic evidence that stimulus similarity can lead to selective associations has been obtained in second-order conditioning experiments with pigeons and rats. In some studies (Rescorla & Furrow, 1977, Experiments 1 and 2), stimulus similarity was varied by using first- and second-order conditioned stimuli that were of the same modality (both visual or both auditory) as compared to stimuli that were of different modalities (one visual and the other auditory). In other studies, the dimension of localized visual stimuli (color or line orientation) (Rescorla & Furrow, 1977) and the location of visual cues (Rescorla & Cunningham, 1979) were varied as manipulations of stimulus similarity. Similarity between first- and second-order conditioned stimuli in terms of stimulus modality, stimulus dimension, and stimulus location all facilitated second-order conditioning. Furthermore, the design of the experiments precluded interpreting these results in terms of stimulus generalization between first- and second-order CSs. d. Mechanisms of Stimulus-Similarity Effects. Stimulus similarity may be a primary determinant of associations that cannot be analyzed in terms of other more elementary behavioral processes. Alternatively, the effects of stimulus similarity may be a product of other associative mechanisms that are unequally manifest in subjects conditioned with similar as com-


24 1

pared to dissimilar stimuli. Rescorla and Gillan (1980) recently explored the latter possibility in the following manner. If two stimuli are similar, they presumably have some stimulus features in common. If a, b, c, and d represent stimulus elements, two similar stimuli may be represented as ab, and ac, with a serving as the common element. Two dissimilar stimuli may be represented as ab, and dc. Second-order conditioning is typically conducted by presenting the second-order CS followed in succession by the first-order CS. If the two stimuli are similar, the proac, where ab is the second-order CS cedure may be represented as ab and ac is the first-order CS. Rescorla and Gillan (1980) suggest that in this case the second-order CS will become associated primarily with the unique element c of the first-order stimulus. This assumption is made because in the transition ab -,ac, the only change in stimulation is produced by introduction of the new element c, ensuring that the subject will pay closer attention to the new element c than to the constant element a. In addition, c is in a more favorable temporal relation for conditioning of the secondorder CS. Only element c is paired sequentially with the second-order CS. These factors favoring association of the second-order cue with the unique stimulus element c of the first-order CS do not exist for subjects that receive a dissimilar first-order CS. In a second-order conditioning trial involving ab followed by dc, for example, both d and c of the first-order CS involve a stimulus change, ensuring attention to both stimulus elements. In addition, in this case both d and c are experienced sequentially relative to the second-order cue. The above analysis predicts that similarity should facilitate second-order conditioning only if the unique elements of the first-order CS are more strongly conditioned with the unconditioned stimulus than are the common elements. This was probably true in the first experiment conducted by Rescorla and Gillan (1980), as well as in previous studies demonstrating facilitory effect of stimulus similarity on second-order conditioning. In a subsequent experiment, Rescorla and Gillan (1980) conducted first-order conditioning in a way that ensured that the unique features of the firstorder CS would not be as well conditioned by the US as the elements that the first- and second-order CSs had in common. Under these circumstances the usual facilitory effect of stimulus similarity on learning was lost. Another approach to the analysis of stimulus similarity effects on conditioning concerns the possible role of common stimulus elements as retrieval cues for the CS when the unconditioned stimulus is presented. Holland (1 98 1) recently demonstrated that exteroceptive cues conditioned with a specific food can mediate a learned aversion to the food if the exteroceptive stimuli are subsequently paired with lithium malaise. Comparable mediation could occur with stimulus elements that conditioned and +

242

Michael Domjan

unconditioned stimuli have in common. If these common elements become associated with the unique features of the CS, their presence when the US is administered may mediate the association of the CS with the US. Such mediational processes may be especially relevant when a long delay separates the conditioned and unconditioned stimuli, making direct associations between CS and US less likely. If the conditioned and unconditioned stimuli are presented with a long CS-US interval, the common and unique elements of the US are both likely to command attention when the US is introduced. Therefore, the differential attention mechanism discussed by Rescorla and Gillan is not as likely to be important. 4.

The Temporal Relationship of CS-and US-Induced Sensations

The stimulus similarity hypothesis considers the similarity between the temporal-intensity patterns of the conditioned and unconditioned stimuli as critical for association of the two events. Another analysis of selective associations focuses on when the CS and its trace are experienced relative to when the US and its traces are experienced (Krane & Wagner, 1975). According to this account, conditioning is best if the conditioned stimulus and its trace are experienced before introduction of the US, and none of the CS trace extends past the end of the US. To apply this idea to selective associations with taste and audiovisual cues, one has to make the oftenstated assumption that tastes have longer stimulus traces than audiovisual cues. Strong audiovisual aversions are conditioned by shock presumably because the trace of audiovisual cues is very brief and does not extend past the shock, which is typically administered at the end of the audiovisual CS. In contrast, shock does not condition a good aversion to taste presumably because the trace of gustatory cues lasts longer than the effects of shock administered at the end of a taste exposure. Strong taste aversions are conditioned by toxicosis because even though gustatory cues have long traces, the taste experience presumably does not outlast the effects of an aversive drug treatment. Poor audiovisual aversion conditioning with toxicosis is explained by the long CS-US interval that results when an audiovisual cue (with a short trace) is followed by a drug treatment that has delayed effects. As is evident from the above discussion, a full account of selective associations in terms of the temporal relationship of CS- and US-induced sensation requires many assumptions about the latency, duration, and traces of various sensations. However, this type of account makes clear differential predictions about the effects of the CS-US interval on aversion conditioning of taste and audiovisual cues with shock. Because of the brief trace of audiovisual cues, the strongest audiovisual-shock conditioning is


243

expected to occur with a short CS-US interval. In contrast, with a short CS-US interval the stimulus trace of gustatory cues is expected to persist after the shock and disrupt conditioning. This source of interference is expected to be reduced as the CS-US interval is increased, making longer CSUS intervals more effective for taste-shock conditioning. These predictions were confirmed in an experiment by Krane and Wagner (1975). However, additional research is necessary to decide how much of the selective conditioning effect with shock and toxicosis can be attributed to the temporal relation of CS- and US-induced sensations.

5. Role of Previous Experience Another approach to explaining selective aversion conditioning in terms of more elementary associative processes assumes that the effect reflects the acquisition of learning sets (e.g., Dickinson, 1980; Mackintosh, 1974). Nearly all selective conditioning experiments have employed adult subjects. The learning set account assumes that during an animal’s normal experiences, taste sensations are highly correlated with subsequent visceral and postingestive events, and these experiences lead to the acquisition of a learning set that facilitates the later association of tastes with interoceptive unconditioned stimuli such as toxicosis. Correspondingly, audiovisual and other exteroceptive stimuli are assumed to be highly correlated with consequent cutaneous and other exteroceptive pain during the subject’s lifetime, and these experiences presumably lead to the formation of a learning set that facilitates the acquisition of aversions to audiovisual cues paired with shock. Although the learning set hypothesis was proposed to explain selective taste and audiovisual learning, it can be also applied to examples of selective associations in second-order conditioning based on stimulus similarity. In this case, one has to assume that similar stimuli are more likely to be experienced together than dissimilar ones during an animal’s normal life. Experimental tests of the learning set interpretation may be conducted in two ways. One strategy is t o alter normal ontogenetic experiences so as to break up the stimulus correlations that are assumed to lead to selective associations. This approach is difficult to carry out because it requires the construction of an abnormal environment, and one must make sure that any change in learning is due to the unusual stimulus correlations in the environment rather than other abnormal experiences the subjects may receive during the treatment. Another approach is to investigate selective conditioning at a very early age, before subjects have had a chance to acquire the learning sets that are assumed to form the basis of selective associations. This strategy was followed in two recent studies with rats. In one experi-

244

Michael Dom]an

ment (Gemberling, Domjan, & Amsel, 1980), aversion conditioning was conducted at 5 days after birth, and in the other experiment (Gemberling & Domjan, 1982) conditioning was conducted 1 day postpartum. The flavor of a .5% saccharin solution and the texture of smooth cardboard served as conditioned stimuli. At both ages, lithium toxicosis conditioned an aversion to the saccharin flavor but not to the cardboard texture. In contrast, shock conditioned texture aversions but not flavor aversions. The above findings strongly suggest that the acquisition of learning sets is not necessary for selective aversion conditioning. However, they allow the possibility that prior learning experiences can contribute to the effect. Consistent with this possibility, Dalrymple and Galef (1981) recently found that rats pretrained to select palatable as opposed to unpalatable food on the basis of visual cues were subsequently more likely to associate visual stimuli with toxicosis. Since only one pretraining task was used, this effect may not have reflected the acquistion of a learning set. Nevertheless, the finding illustrates that previous experience can facilitate later aversion learning with particular combinations of stimuli.

6. Measurement Problems in Selective Conditioning Experiments Two types of measurement problems have been also discussed in efforts to explain selective conditioning in terms of more elementary processes (LoLordo, 1979). One of these assumes that selective conditioning effects do not reflect different rates of learning but different starting points for learning (cf. Konorski, 1967). Presumably an association will be evidenced in behavior only if its strength is sufficient to exceed the threshold for a behavioral effect. If an association is closer to this threshold at the start of conditioning, learning of that association will produce a behavioral effect sooner than the learning of associations that are not as close initially to the behavioral threshold. Using this idea, one may explain observations of selective aversion conditioning by assuming that taste-toxicosis and audiovisual-shock associations are closer to the threshold for a behavioral effect at the start of conditioning than associations of taste with shock and audiovisual cues with toxicosis. Although this is a plausible explanation, it is difficult, if not impossible, to verify experimentally. Behavioral techniques are not well suited to test the hypothesis because the events assumed to be responsible for selective conditioning involve differences below a behavioral threshold. Neurophysiological techniques are also difficult to use because our knowledge of the neurophysiology of learning is inadequate to tell us where or how to look for initial subthreshold associations. A second measurement problem arises from the responses chosen to pro-


245

vide evidence of conditioning. Behavioral changes that develop during classical conditioning depend on the nature of both the conditioned and the unconditioned stimulus (e.g., Holland, 1977; Jenkins & Moore, 1973). Learning in selective conditioning experiments is often measured in terms of a single behavioral index, such as suppression of drinking. This index may be sensitive to taste-toxicosis and audiovisual-shock associations but may not reflect taste-shock and audiovisual-toxicosis associations. Several strategies can be used to search for behaviorally silent associations. One approach involves using a behaviorally silent CS to influence other more easily measured instances of learning. For example, a behaviorally silent CS may be used to block the conditioning of a readily acquired conditioned response (LoLordo, Jacobs, & Foree, 1982), or it may be used to produce second-order conditioning. A second approach involves observing the total behavioral repertoire of the subjects in an effort to detect any change in behavior that may result from conditioning (Shapiro eta/., 1980). Neither of these procedures have been used yet in studies of the selective aversion learning phenomenon.

B. OTHERSTUDIES OF SELECTIVE CONDITIONING OF INGESTIVE BEHAVIOR Differential aversions conditioned to taste and audiovisual cues and studies of the importance of stimulus similarity in higher order conditioning are the most extensively investigated instances of selective learning in classical conditioning. However, numerous other examples of selective conditioning have been also reported, and some of these have been concerned with selective conditioning of ingestive behavior. Green, Holmstrom, and Wollman (1974) investigated selective learning of preferences conditioned by recuperation from illness and concluded that tastes are selectively associated with recovery. However, their experiments allow other interpretations because nonassociative effects of recuperation from illness were not evaluated. Weisinger, Parker, and Skorupski (1974) compared aversions to salt and sugar solutions conditioned by injections of formalin and insulin. Because formalin causes a sodium appetite, they predicted a constraint on conditioning an aversion to salt with formalin treatment, and because insulin reduces blood sugar, they predicted a constraint on conditioning an aversion to sugar with insulin injections. The experiments confirmed these predictions. However, subsequent work has failed to replicate the results. No limitations on conditioning of salt aversions with formalin treatment were found in a variety of circumstances, and experiments suggested that the failure of Weisinger and his co-workers to condition an aversion to the taste

Michael Domlan

246

of sugar with insulin injections was due to the weak sugar flavor that they employed (Domjan & Levy, 1977).

c.

SELECTIVE APPETITIVE AND AVERSIVE CONDITIONING IN PIGEONS

I noted in the earlier discussion of constraints on avoidance learning that discriminative shock-avoidance behavior is more strongly controlled by auditory than by visual cues, whereas discriminative food-reinforced behavior is more strongly controlled by visual than by auditory cues (Foree & LoLordo, 1973, 1975; LoLordo, 1979; LoLordo & Furrow, 1976; Schindler & Weiss, 1982). Selectivity in appetitive and aversive conditioning has been also investigated using classical conditioning procedures. Shapiro et ai. (1980) compared the relative effectiveness of a tone and a red light in classical conditioning of pigeons with food and shock. For food-conditioned animals, pecking in and around the food hopper developed as a conditioned response. In contrast, the most prominent conditioned response in pigeons conditioned with shock was raising of the head and prancing (rapid sideto-side movements, and lifting of the feet higher than in normal walking). When food served as the unconditioned stimulus, conditioned responses developed for subjects that had the red light paired with food. When shock served as the unconditioned stimulus, conditioned responses were acquired by subjects that had the tone paired with shock. Much less conditioned responding occurred when the tone was paired with food or the light was paired with shock. The selective conditioning effect observed by Jacobs et al. (1980) may have been produced by differential attention caused by the food and shock unconditioned stimuli. Presentations of food may have caused the subjects to attend selectively to visual cues, and presentations of shock may have caused them to attend selectively to auditory stimuli. Another possibility is that the visual and auditory cues became conditioned equally rapidly, but the pigeons did not show changes in behavior in response to auditory stimuli conditioned with food and visual cues conditioned with shock. Evidence for such behaviorally silent associations may be obtained with the use of indirect tests of association. In one such test, Shapiro and LoLordo (1982) evaluated the conditioned reinforcing properties of food-conditioned visual and auditory cues and found that visual stimuli became conditioned as reinforcers more rapidly than did auditory cues. In other experiments, the conditioning of a visual stimulus with food was not blocked by the presence of a tone previously conditioned with food, and the conditioning of a tone with shock was not blocked by the presence of a visual stimulus previously conditioned with shock (LoLordo et a!., 1982). These results confirm the


247

dominance of visual cues in conditioning with food and auditory cues in conditioning with shock.

D.

SELECTIVE CONDITIONING OF STIMULI AS AND

DANGER

SAFETY SIGNALS I N RATS

Jacobs and LoLordo (1977) observed that in instrumental avoidance conditioning of rats the onset of noise is more likely to become conditioned as a warning stimulus than as a safety signal. In contrast, the termination of white noise, and the onset or termination of a visual stimulus were more likely to become conditioned as safety signals than as warning stimuli. Similar selective stimulus effects have been observed using classical conditioning procedures with rats (Jacobs & LoLordo, 1980). The results of conditioning were measured in terms of changes in the rate of a free-operant wheel-turn avoidance response. The onset of the tone readily became conditioned as a warning stimulus and increased wheel-turn avoidance behavior but did not become conditioned as a safety signal. In contrast, the termination of the tone, illumination of the house lights, and extinction of the house lights all became readily conditioned as safety signals and decreased wheel-turn avoidance but did not become conditioned as warning stimuli. The above findings represented associative processes because comparable results were not produced by random exposures to the CSs and shock. The selective stimulus effects also could not be attributed to different initial effects of the stimuli on wheel-turn avoidance because three of the CSs (onset of a soft tone, termination of a loud tone, and extinction of the house lights) had no effect on avoidance behavior initially but still became differentially conditioned as warning and safety signals. The onset of a soft tone was more readily conditioned as a warning than as a safety signal, and the other two CSs were more readily conditioned as safety signals than as warning stimuli. In addition, the classical conditioning phase was rather extensive (30trials), which should have overcome any differences in initial response to the stimuli. The selective conditioning effects also could not be attributed to differential orientations caused by the unconditioned stimulus because warning and safety signal groups received the identical exposure to shock. The investigations of Jacobs and LoLordo indicate that changes in illumination are more likely to become conditioned as safety signals than as warning stimuli for aversive events. However, this outcome does not necessarily imply that visual stimuli cannot be conditioned as warning signals for shock in rats. Using a different type of behavioral assessment technique (suppression of food-reinforced lever pressing), Welker and Wheatly (1977) found that visual cues can become conditioned to signal shock. Further-

248

Michael Domlan

more, an increase in illumination was more effective than a decrease in lighting, even though the two types of cues were equally effective in signaling the absence of food reinforcement. These results may have occurred because increased illumination is slightly aversive in comparison to decreased illumination. The aversiveness of the increased illumination was insufficient to produce unconditioned suppression of lever pressing. However, this subthreshold initial aversiveness may have summed with fear conditioning to yield the differences in conditioned suppression that were observed.

E. CONCLUDING COMMENTS Research on selective conditioning has provided us with numerous examples of such phenomena. Rather than being an adaptive specialization for poison-avoidance learning, cue-consequence specificity appears to be a more general characteristic of conditioning. Investigations have identified a variety of mechanisms that may contribute to selective conditioning. The critical issue is whether selective conditioning effects are a direct result of the stimuli involved or are produced by more elementary general processes activated by the stimuli. Research on selective associations in higher order conditioning has identified some of the elementary processes responsible for those effects. Similarly, selective conditioning of visual and taste cues in feeding and drinking situations in chicks has been shown to result from more elementary mechanisms. In contrast, the analysis of other examples of selective conditioning has not progressed as far. Therefore, it is premature to conclude that those effects represent unique ways in which the stimuli are processed or special ways in which the stimuli activate more elementary general mechanisms. VI. Long-Delay Learning

One of the recurrent themes in the history of the study of associations is that two events have to occur close together in time to become associated with one another. This requirement of a close temporal relationship between conditioned and unconditioned stimuli is evident in most types of classical conditioning (see Kimble, 1961). However, animals can learn aversions to the taste of poisonous food even if the illness does not occur for many hours after the taste exposure (Garcia, Ervin, & Koelling, 1966; Smith & Roll, 1967). Because long-delay learning of taste aversions is highly unexpected on the basis of studies of trace conditioning in other situations, the phenomenon has been extensively investigated. The research has ad-


249

dressed two types of questions. Some investigations have focused on the mechanisms responsible for long-delay learning in the feeding system. Other studies have been devoted to finding out whether long-delay learning is unique to taste or feeding-related stimuli or represents more general processes. A.

MECHANISMS CONTRIBUTING TO THE DELAY OF REINFORCEMENT GRADIENT IN TASTE-AVERSION LEARNING

Instances in which a conditioned stimulus becomes associated with a delayed unconditioned stimulus have been explained traditionally by postulating that the conditioned stimulus has a decaying trace which persists long enough to overlap and become associated with the US. The associative strength of the CS trace was then assumed to generalize to the CS. The CS trace may be considered to be a temporally decaying neuronal process or a physical stimulus aftereffect. With taste stimuli, physical stimulus aftereffects (aftertastes) are a distinct possibility, and various strategies have been employed to show that long-delay taste-aversion learning is not mediated by the aftertaste of the flavor exposure (see Revusky & Garcia, 1970; Rozin & Kalat, 1971). However, disproof of aftertaste mediation does not leave the neuronal trace-decay hypothesis as the only alternative explanation. Several associative interference interpretations of the delay gradient have been also entertained.

I.

The Learned Safety Hypothesis

One alternative to the trace decay explanation of long-delay learning attributes the delay gradient to learning that the CS flavor is safe (Kalat & Rozin, 1973). Kalat and Rozin suggested that animals regard new foods as potentially poisonous and monitor their state of well-being after the first sampling to see if the food makes them sick. As time passes without ill effects, subjects gradually learn that the food is safe, and this interferes with learning an aversion to the food. Thus, the delay gradient in tasteaversion learning is attributed to the gradual learning of safety rather than to trace decay, and the association of taste with long-delayed aversive consequences is explained by assuming that the learning of safety occurs very slowly. Initial support for the learned safety hypothesis was provided by an experimental design that pitted the learned safety hypothesis against the tracedecay hypothesis (Kalat & Rozin, 1973; Bolles, Riley, & Laskowski, 1973). However, subsequent investigations have questioned the utility of this ex-

250

Michael Domjan

perimental design for studies of the role of learned safety in the delay of reinforcement gradient (Best & Barker, 1977; Best & Gemberling, 1977; Domjan & Bowman, 1974). A more direct test of learned safety was conducted by Nachman and Jones (1974). They reasoned that if the absence of illness following the first exposure to a novel or slightly aversive taste progressively produces learned safety, then the longer one waits before a second exposure to the flavor, the more subjects will drink during the second exposure. Consistent with this prediction, intakes during a second exposure to a slightly aversive sucrose solution and a novel saccharin solution were directly related to the interval between the first and second stimulus presentations, and this outcome could not be explained in terms of changes in thirst (see also Green & Parker, 1975). Although the above findings indicate that some type of learning occurs as subjects fail to experience ill effects following exposure to a novel flavor, other evidence suggests that learning of “safety” is insufficient to explain the delay of reinforcement gradient in taste-aversion learning. A delay gradient is obtained even if a highly familiar flavored solution is employed in aversion conditioning, one that is entirely “safe” to the subjects (Nachman & Jones, 1974, Experiment 3). A learned safety type of process also cannot explain instances in which a brief reintroduction of the CS flavor during a long CS-US interval facilitates aversion conditioning (Best & Gemberling, 1977; Domjan & Bowman, 1974). Both of these findings are compatible with the idea that trace decay is also involved in the delay of reinforcement gradient. Other evidence indicates that if a process like learned safety is involved in delay learning, the process is misnamed. If exposure to a flavor without ill effects leads to learning that the flavor is safe, such an experience should facilitate later association of that flavor with the absence of expected illness. Contrary to this prediction, Best (1975) has shown that nonpoisoned exposure to a flavor in fact interferes with the subsequent conditioning of that flavor as a signal for the absence of illness. Based on such evidence, Kalat (1977) has suggested that “learned safety’’ should be renamed “learned noncorrelation” or “learned irrelevance.” According to this reformulation, exposing subjects to a flavor in the absence of consequent aversive (or positive) unconditioned stimulation teaches the subjects that the flavor is unrelated to potential unconditioned stimuli, and this interferes with subsequent attempts to condition the flavor (cf. Mackintosh, 1973). 2. Concurrent Interference Theory

Another alternative to the trace decay model of long-delay learning also attributes the delay gradient to interfering associations (Revusky, 1971, 1977). According to this formulation, a target stimulus A will become as-


25 1

sociated with a subsequent target event B to the extent that A does not become associated with other events that follow it and B does not become associated with other events that precede it. Associations between the target stimuli and nontarget events are assumed to provide “concurrent” interference for the formation of an association between A and B. The concurrent interference principle explains long-delay learning of taste-toxicosis associations in terms of the phenomenon of selective aversion conditioning. In typical demonstrations of long-delay taste-aversion learning, subjects are exposed to a novel taste, followed several hours later by some type of toxicosis. During the interval between these target stimuli, the subjects are likely to experience various visual, auditory, and other exteroceptive stimuli but are not exposed to other novel tastes or other sources of malaise. Selective aversion learning presumably insures that the various nontarget exteroceptive stimuli subjects experience during the taste-toxicosis delay interval do not become associated with either the target CS or target US. Therefore, these extraneous stimuli do not provide interference for the tastetoxicosis association, and learning can occur with long intervals between the two target stimuli. Such long-delay learning is not possible in conditioning an audiovisual cue with shock because many of the various nontarget stimuli subjects are likely to experience during the CS-US interval (other auditory and visual cues) presumably can become associated with shock, and this interferes with association of the target CS with shock. The best evidence for concurrent interference mechanisms in taste-aversion learning is provided by experiments in which interfering stimuli and interfering associations are directly manipulated. Revusky (1971) has reviewed a number of instances of this type. For example, he has shown in several ways that the introduction of a novel flavor during the CS-US interval in long-delay taste-aversion learning disrupts conditioning of the target CS flavor to the extent that the interposed taste can be conditioned with toxicosis. Kalat and Rozin (1971) also found that the introduction of several novel flavors during a long-delay interval disrupts aversion conditioning of the target CS flavor. In contrast to the above evidence, the concurrent interference theory has not been as successful in predicting long-delay learning in situations where interfering stimuli are not explicitly introduced by the experimenter. For example, because the theory considers long-delay learning to be a by-product of selective associations, it predicts that the ontogenetic development of these two phenomena will be similar. Contrary to this prediction, longdelay flavor-aversion learning has been found to appear much later in development than selective aversion learning. Strong (all or none) selectiveaversion learning effects have been observed in rats conditioned 1 day after birth (Gemberling & Domjan, 1982). However, l-day-old rats did not learn

252

Michael Domjnn

a flavor aversion if lithium was administered 30 min after the flavor exposure instead of without a delay. Aversion learning with 30-min delayed toxicosis was observed in 5-day-old rats, but at this age no learning occurred with delays of 60 or 90 min. Another study compared 12- and 15-day-old rats and observed flavor-aversion learning with a 120-min interval only in the 15-day-old subjects (Gregg, Kittrell, Domjan, & Amsel, 1978). The absence of long-delay learning in very young rats cannot be attributed to their general inability to remember taste information. Conditioned aversions acquired at 1 day of age are retained when subjects are tested 5 days later (Gemberling& Domjan, 1982), and 5-day-old rats exposed to a novel flavor show an attenuated neophobia to the flavor when tested 12 hr later (Gemberling et al., 1980). To explain the gradual ontogenetic development of long-delay learning in terms of concurrent interference, one must assume that various tactile, visual, auditory, and olfactory stimuli that subjects may encounter during a long flavor-toxicosis interval provide more concurrent interference early in life than they do later. For this to happen, interfering stimuli would have to be more easily conditioned with toxicosis at 1 day of age than at older ages. The available evidence suggests that this is not true. Tactile stimuli have been found not to be conditionable with toxicosis at either 1 or 5 days of age (Gemberling & Domjan, 1982; Gemberling et al., 1980). Other studies have shown that in the absence of a delay between odor exposure and lithium treatment, rats acquire comparable odor aversions between 2 and 14 days of age; with longer CS-US intervals, learning improves with age (Rudy & Cheatle, 1979). The concurrent interference theory is also challenged by demonstrations of long-delay aversion learning to nongustatory aspects of food. Rats, guinea pigs, fish, and birds have all been shown to be able to associate the visual aspects of food with subsequent poisoning (e.g., Bravemen, 1974; Galef & Osborne, 1978; MacKay, 1977; Martin & Bellingham, 1979). Longdelay learning of aversions to olfactory cues in rats (e.g., Palmerino, Rusiniak, & Garcia, 1980) and tactile aspects of food in rats and monkeys (Domjan &Hanlon, 1982; Domjan, Miller & Gemberling, 1982) have been also observed. In addition, rats have been shown to be able to learn an aversion to their drinking cup paired with delayed toxicosis (Revusky & Parker, 1976; see also Nachman, Rauschenberger, & Ashe, 1977). In all of these cases, no special effort was made to limit the visual, olfactory, and tactile stimuli subjects experienced during the delay interval. Exposure to numerous visual, olfactory, and tactile cues during the delay interval should have provided strong concurrent interference and made it difficult, if not impossible, to obtain these instances of long-delay learning. The above discussion indicates that neither learned safety nor concurrent


253

interference processes provide an entirely satisfactory account of long-delay learning. Long-delay visual- and odor-aversion learning may be mediated by certain novel mechanisms to be discussed in the following sections. In contrast, it appears that we cannot dispense with the traditional trace-decay idea in interpreting long-delay taste-aversion learning. To explain all of the data, we must assume that the memory of a taste experience remains available for association with toxicosis for a relatively long period. In rats, the neural mechanisms involved in this memory process seem to develop slowly during the first 2 weeks of life.

B.

LONG-DELAYLEARNING OUTSIDE FEEDING SYSTEM

STUDIES OF THE

Efforts to gain insight into the mechanisms of long-delay learning have also involved studies of long-delay learning outside the feeding system. These investigations have been motivated by the assumption that food-aversion experiments employ procedures that optimize the operation of general mechanisms of long-delay learning that can occur in a variety of response systems. If the procedural parameters critical for long-delay learning can be identified, these parameters presumably can be applied to produce longdelay learning outside the feeding system.

I . Minimizing Concurrent Interference The first systematic effort to investigate long-delay learning outside the feeding system was stimulated by the concurrent interference theory. The theory predicts that long-delay learning will occur whenever interfering associations are minimized by selective conditioning mechanisms. Revusky (197 1) proposed a new selective conditioning mechanism, situational relevance, according to which exteroceptive stimuli experienced in one situation are not likely to become associated with events that are experienced in a distinctively different situation. The concurrent interference theory, together with situational relevance, predicts that subjects will learn to associate one event with another over a long delay provided that the two events are experienced in the same environment and the delay interval is spent in a distinctively different place. Lett has used this strategy in demonstrations of long-delay learning in the T maze. In her initial experiment, Lett (1973) reinforced rats for either turning right or left in a T maze, with reinforcement delayed up to 8 min. After making the correct choice, the subjects were returned to the home cage for the delay interval, at the end of which they were placed in the start area of the maze to receive the reinforcer, wet mash. In a subsequent study (Lett,

254

Michael Domjan

1974), animals received 1-min delayed reinforcement for selecting the white or black arm of a T maze. The reward was again presented in the start area of the maze, and subjects spent the delay interval in the home cage. A later experiment (Lett, 1975, Experiment 3) evaluated the importance of placing subjects in the home cage for the delay interval. Some subjects were confined to the chosen arm of the T maze for the first 15 or 60 sec of a 2-min delay interval before being placed in the home cage for the rest of the interval. Consistent with predictions of the concurrent interference theory, the longer subjects spent in the T maze after a correct choice, the less likely they were to learn the spatial discrimination task with 2-min delayed reinforcement. Although the findings of Lett (1973, 1974, 1975) are consistent with the concurrent interference theory, the interpretation of these results has been called into question by subsequent investigators. With certain refinements of the procedures that had been used by Lett (1974), Roberts (1976) failed to find convincing evidence that rats can learn a black-white visual discrimination with 1-min delayed reinforcement. Lett (1977) has suggested that the data obtained by Roberts (1976) in fact provide evidence of learning. However, Roberts (1977) suggests that what might appear to be learning in fact represents nonassociative stimulus satiation effects. D’Amato and his co-workers also encountered difficulties in demonstrating longdelay instrumental learning in the T maze (D’Amato, Buckiewicz, & Puopolo, 1981; D’Amato, Salmon, & Puopolo, 1981). Using Cebus monkeys, they found that strong response biases often obscured evidence of conditioning. Other investigators have focused on the mechanisms of long-delay spatial discrimination learning in rats and the importance of removing the subject from the T maze during the delay interval (Lieberman, McIntosh, & Thomas, 1979). These studies have found that delayed spatial discrimination learning is a reliable phenomenon but call into question concurrent interference as a mechanism for the learning. In one experiment, subjects were picked up when they made a correct choice in the T maze and were either placed in the choice area of the maze or in the home cage for a 1min delay interval before receiving reinforcement in the start box. Both groups learned the discrimination equally well. Subsequent experiments showed that the critical event for long-delay spatial discrimination learning was picking up the subjects immediately after they made a choice. Where they spent the delay interval was much less, if at all, important. These results suggest that the poor learning Lett (1975) obtained when she allowed subjects to remain in the chosen arm of the maze during part of the delay interval was not due to remaining in the maze but was caused by the fact that the subjects were not handled immediately after making a choice response.

Biological Constraints on haming

2.

255

The Marking Hypothesis

Lieberman et al. (1979) proposed the marking hypothesis to explain their findings of long-delay learning in the T maze. According to the marking hypothesis, an event is remembered longer and is available for association longer if it is “marked” by a salient, unexpected stimulus. The common feature of procedures that produced successful long-delay spatial discrimination learning in the T maze was that subjects were handled immediately after making a choice response. Given this handling, subjects learned the discrimination regardless of where they spent the delay interval. This outcome is explained by assuming that handling of the subjects immediately after a choice response served to “mark” the behavior and thereby made it more memorable. Subjects that were not handled after a choice (animals that remained in the chosen arm for a while or were allowed to walk to a delay chamber) did not learn the discrimination presumably because their choice response was not “marked” and was therefore more easily forgotten. According to the marking hypothesis, handling facilitates long-delay learning only because it is a salient and unexpected event following choice responses. Therefore, other stimuli may be just as effective in “marking” and enhancing the memory of instrumental behavior. Consistent with this possibility, Lieberman et al. (1979, Experiment 4) found that marking a choice response with a noise burst or flash of light facilitated long-delay learning even more than handling of the animals after each response. These results provide considerable support for the marking hypothesis. We noted earlier that studies of long-delay learning outside the feeding system represent an effort t o discover general mechanisms of long-delay learning that may occur in a variety of response systems. Is the marking hypothesis applicable to long-delay ingestional aversion learning? Flavoraversion conditioning experiments typically do not involve handling the subjects or presenting a noise burst or flash of light immediately after exposure to the conditioned stimulus. What serves to mark the conditioned stimulus to enhance its memorability? A strong possibility is that ingestive behavior and its accompanying sensations provide this function (Domjan, 1980). Contact with the conditioned stimulus typically occurs in the context of the complex series of responses involved in approaching an ingestible substance, taking it in the mouth, licking or masticating, and swallowing. These responses, in turn, are accompanied by a host of stimuli provided by the texture, temperature, viscosity, and other features of the ingested substance. This complex series of responses and stimuli may serve to mark the distinctive features of a food stimulus and make these features sufficiently memorable t o allow association with delayed toxicosis. Consistent with this

256

Michael Domjan

view, rats for which a flavored solution is rinsed over the tongue in the absence of approach, drinking, and swallowing responses acquire weaker aversions to the taste than subjects that are allowed to consume the flavored solution in the normal fashion prior to delayed toxicosis (Domjan & Wilson, 1972b). The idea that ingestive behavior serves to mark food-related stimuli and make them available for association with delayed toxicosis also helps in understanding instances of long-delay visual, olfactory, and tactile aversion learning. As we noted earlier, long-delay aversion learning to nongustatory aspects of food presents a special theoretical challenge because animals are bound to encounter a variety of visual, olfactory, and tactile stimuli during the delay interval, and these would be expected to interfere with learning an aversion to the target CS. However, the intervening stimuli are not likely to be food related and are not likely to be experienced in conjunction with ingestion. If ingestive behavior serves to mark nongustatory aspects of food for association with delayed toxicosis, intervening stimuli that are unrelated to eating are not likely to constitute a source of interference.

3.

The Memory Retrieval Hypothesis

Long-delay learning in the T maze has been also discussed in terms of a memory retrieval hypothesis (Lett, 1979; Lieberman et al., 1979). Procedures that resulted in successful long-delay learning involved presenting the marking stimulus (handling, for example) both after the subject made a choice response and just before it received the reinforcer. The choice response may have become associated with the marking stimulus, and the presence of the marking stimulation when the reinforcer was delivered may have reactivated the memory of the choice response and thereby permitted association of the response with the reinforcer. Subjects in these studies were also handled (or received the noise or light stimulus) on trials when they made an incorrect choice. Presumably the memory of incorrect choices is not retrieved by the handling cues during reinforcement because the correct response is the most recent one when reinforcement is delivered. Further research is required to experimentally distinguish between the memory retrieval and marking hypotheses. One clear differential prediction is that for memory retrieval to take place, the stimulus that follows a choice response also has to be presented when the reinforcer is delivered. However, for the marking function to occur, the stimulus only has to be presented with the instrumental response. Studies evaluating the importance of presenting the “marking” stimulus during reinforcement are not yet available. It is important to point out that mediation of long-delay learning by


251

memory retrieval is different from mediation by secondary reinforcement. The results obtained by Lett (1973, 1974, 1975) and Lieberman et al. (1979) cannot be explained in terms of secondary reinforcement because subjects received the same stimulation immediately after correct and incorrect choices. Therefore, any secondary reinforcement would have reinforced correct and incorrect choices equally, and discrimination learning would not have been evident. 4. Affective versus Instrumental Response Learning

The experiments reviewed demonstrate that rats can learn instrumental discriminations with delayed reinforcement. However, most of the studies employed delays of only several minutes (but see Lett, 197% and acquisition was typically not evident until subjects received more than 10 reinforced trials. These delays of reinforcement are longer than what was previously known to permit instrumental learning, but neither the delays used nor the speed of acquisition are comparable to corresponding features of taste-aversion learning, which can occur in one trial with reinforcement delays of several hours (e.g., Revusky, 1968; Smith & Roll, 1967). D’Amato and his associates hypothesized that the critical aspects of food-aversion experiments are the conditioning of an affective response (an aversion to the food), and the use of a simple response measure (suppression of intake) that directly reflects what is learned. The affective response that they endeavored to condition was preference for a visually distinctive arm of a T maze, and conditioning was measured by giving subjects unrestricted access to the two maze arms for several minutes. In one study (D’Amato & Buckiewicz, 1980), Cebus monkeys were confined to the nonpreferred arm of a T maze for 1 min before being placed in a holding cage for a 30-min delay interval. The monkeys were then returned to the start area of the maze to receive reinforcement. Control groups were treated comparably except they either did not receive the CS exposure or were not reinforced upon return to the start box. Subjects in the experimental group showed a dramatic increase in preference for the CS arm of the maze after just one conditioning trial, and this preference persisted when subjects were tested 4 months after having received three conditioning trials. Control subjects exposed to the CS without reinforcement did not show any increase in preference for the CS during the experiment. Control subjects that received reinforcement unpaired with the CS arm of the maze showed a moderate increase in preference for the CS, but this change was not as great as what occurred in the experimental group. These results provide the first demonstration of one-trial spatial preference learning with a 30-min delay.

258

Michael Domjan

Safarjan and D’Amato (1981) investigated one-trial long-delay spatial preference learning more extensively using rat subjects. Preferences were not influenced by the delay intervals tested (30 and 120 min). However, experimental subjects showed greater preference for the CS arm of the T maze than controls only if they had been exposed to the CS for 20 or 40 min during the conditioning trial. A 10-min CS exposure during conditioning was insufficient to permit spatial preference learning in rats. In another study, D’Amato and Safarjan (1981) provided direct evidence that instrumental response learning is more easily disrupted by a delay of reinforcement than affective conditioning and suggested that this occurs because instrumental behavior requires learning of an anticipatory affective response. To make the correct instrumental choice in a T maze, for example, the subject has to learn to anticipate the stimuli that it will experience after the choice response. In contrast, spatial preference (or aversion) learning does not require such anticipations because subjects can manifest a conditioned spatial preference simply by remaining longer in the preferred arm than they would otherwise. Long-delay learning is easier to detect in the conditioning of affective responses presumably because conditioned anticipations are not involved, and the learning can be measured more directly. 5. Relation of Response, Cue, and Reinforcer

Other investigators have been also concerned with the role of the response in long-delay learning and have focused on the relationship of the response to the cue and the reinforcer in the situation. Testa and Ternes (1977) suggested that long-delay poison-avoidance learning may be considered an instance of object learning in which the object is the poisonous food. In nature, both the cue subjects learn about and the poison are attributes of the poisonous food object. The response modified by conditioning (ingestion) is also strongly determined by the food object. If having the response, the cue, and reinforcer all centered on the same object facilitates long-delay learning, long-delay learning should be evident in learning about objects outside the feeding system. Consistent with this prediction, Sullivan (1979) reported that rats can learn an aversion to touching small objects if touching these objects results in periodic grid shock presented 35 min later. However, Sullivan did not manipulate the relationship between cue and indicant response. Direct evidence of the importance of having the cue and response related to the same object is provided by Galef and Dalrymple (1981), who found that rats are more likely to learn an ingestional aversion to a visual cue paired with toxicosis if the visual cue is a property of the ingested food than if the visual stimulus is a property


259

of the food dish or feeding area. In a related experiment, Logue (1980) found that pigeons are more likely to learn an aversion to a red color paired with toxicosis if the red color is contained in ingested water or is provided by a light that illuminates the water spout than if the red color is located on a response key the pigeons have to peck to obtain water. The suggestion that long-delay learning is facilitated by a special relationship among response, cue, and reinforcer is essentially a description of procedural parameters that are expected to promote learning and not a statement about theoretical mechanisms. One possibility is that the object acts as a mediator or retrieval cue for the learning. Since the cue and the reinforcer are attributes of the same object, they are presumably part of an associative network involving various attributes of the object. Since they belong to the same associative network, presentation of one of the stimuli should result in retrieval of the other. Perhaps long-delay learning is facilitated under these circumstances because presentation of the reinforcer results in retrieval of the memory of the cue. Having the conditioned response involve contacts with the same object may facilitate obtaining evidence of learning because contacts with the object help to reactivate the memory of the cue-reinforcer association. Thus, the common object may mediate both the establishment of the cue-reinforcer association and the behavioral change used to obtain evidence of this association.

6. Concluding Comments Studies of long-delay learning outside the feeding system have shown that such learning is not unique to the conditioning of ingestive responses and has suggested general variables and processes that may lead to long-delay learning in a variety of response systems. These factors include the use of simple and direct measures of learning, the conditioning of affective responses, enhancing memory through marking or the presentation of retrieval cues, and having cue, response, and reinforcer centered on the same object. Additional research is required to delineate the procedural parameters that optimize these factors. We also do not know what the relative contributions of these variables are to long-delay learning and whether these variables are sufficient to explain fully the occurrence of one-trial longdelay learning of taste aversions.

VII.

Potentiation in Classical Conditioning

The potentiation effect was discovered in poison-avoidance learning and consists of the facilitation of learning about nongustatory aspects of poisonous food by the presence of a novel taste. The presence of a novel taste

260

Michael Domjan

was found to facilitate the conditioning of an aversion to odor cues in rats (Palmerino et af., 1980; Rusiniak, Hankins, Garcia, & Brett, 1979; but see Bouton & Whiting, 1982), visual stimuli in pigeons and quail (Clarke, Westbrook, & Irwin, 1979; Lett, 1980a), and visual stimuli in rats (Galef & Osborne, 1978). These instances of potentiation attracted a great deal of attention because they involved a facilitory interaction between two cues that are presented during a conditioning trial instead of the more familiar interference effects that are expected and often found under such circumstances (e.g., Kamin, 1%9; Mackintosh, 1975; Pavlov, 1927; Rescorla & Wagner, 1972). The facilitation of olfactory and visual aversion learning by the presence of a novel taste is clearly useful in avoiding poisonous foods. By responding to the olfactory and visual aspects of a poisonous food, animals can reject the food without contacting it. However, the learning of olfactory and visual aversions has to be modulated by some process so that subjects do not learn aversions to odors and visual cues that are not related to the poisonous food. Having nongustatory aversions facilitated by the presence of a novel taste helps restrict aversion learning to the olfactory and visual cues that are relevant to poison avoidance. A.

FACILITATION OF DIRECTASSOCIATIONS WITH TOXICOSIS

Some have suggested that a novel taste facilitates the direct association of olfactory and visual aspects of food with consequent poisoning. Facilitation of direct associations could occur in several ways. Galef and Osborne (1978) and Lett (1980a) suggested a directed attention hypothesis according to which the taste of food directs attention to nongustatory (e.g., visual) food cues and thereby facilitates the association of the nongustatory cues directly with subsequent toxicosis. Taste cues may also facilitate the direct association of visual and olfactory aspects of food with poisoning by facilitating retrieval of the relevant nongustatory stimulus when delayed toxicosis occurs. After ingesting a poisonous food with distinctive visual and olfactory features, animals are likely to experience numerous other nongustatory cues before the onset of toxicosis. Taste may mark nongustatory aspects of food for retrieval when the subject subsequently gets sick (Clarke, et al., 1979; Palmerino et al., 1980; Rusiniak et al., 1979). If taste increases attention to nongustatory food cues, it should facilitate conditioning of these cues no matter what unconditioned stimulus is used. In contrast, if taste serves as a retrieval cue activated by poisoning it should facilitate conditioning only with poisoning. Consistent with the second of these alternatives, taste has been found to potentiate odor aversions con-


26 1

ditioned with poisoning but not odor aversions conditioned with shock (Rusiniak, Palmerino, Rice, Forthman, & Garcia, 1982). B.

MEDIATED CONDITIONING O F NONGUSTATORY ASPECTSOF FOOD

A second category of explanations of potentiation assumes that the enhanced aversion learning to nongustatory aspects of food is mediated by an association between these cues and the taste of the food. The toxicosis delivered during the conditioning trial is assumed to condition an aversion to the taste cue. Because nongustatory cues are associated with the taste, the aversion response also comes to be evident in response to the nongustatory stimuli. Some investigators have considered and rejected two models of how this mediation might take place, sensory preconditioning and second-order conditioning (Galef & Osborne, 1978; Rusiniak et al., 1979). Regardless of what other details a mediational mechanism may have, the mediational hypothesis attributes potentiated nongustatory aversions to the aversiveness of the taste stimulus associated with the nongustatory cues. Therefore, extinction of the taste following conditioning should also attenuate aversions to the nongustatory stimuli. Durlach and Rescorla (1980) obtained evidence confirming this prediction in studies of odor aversion learning in rats. Additional evidence in support of the mediational hypothesis is provided by studies of the effects on potentiation of preconditioning exposure to odor and taste cues separately as compared to simultaneously. The mediational hypothesis predicts that compound preexposure to odor and taste will facilitate subsequent odor-aversion learning when odor plus taste is paired with toxicosis because such preexposure allows formation of the mediating odor-taste association. In contrast, preexposure to the odor and taste cues separately is expected to interfere with learning the mediating odor-taste association. Consistent with these predictions, Palmerino et af. (1980) found that compound preexposure resulted in stronger odor-aversion learning than preexposure to odor and taste separately. In fact, following separate preexposure to the odor and taste stimuli, the presence of taste during the conditioning trial did not significantly potentiate odor-aversion learning. The above evidence clearly indicates that within- compound associations can contribute to the potentiation effect in poison-avoidance learning but does not prove what such associations are necessary. The results cited were obtained in experiments in which animals were preexposed to odor and taste prior to conditioning. Lett (1980b) has argued that potentiation is mediated by within-compound associations only with such preexposure. In support of this claim, she found that in the absence of preexposure extinction of

262

Michael Domjan

the taste stimulus does not attenuate potentiated aversions to an odor or visual cue in rats and pigeons, respectively. The procedures (and in one case the species) Lett used differed from those employed by Durlach and Rescorla in numerous respects. Therefore, one cannot easily identify the source of the discrepant findings. However, the fact that the potentiation effect can remain unchanged by extinction of the taste stimulus indicates that potentiation can occur in the absence of mediation by an association between the olfactory and visual aspects of food and taste.

c.

INSTANCES OF POTENTIATION OUTSIDE THE FEEDING SYSTEM

The procedures used to produce potentiation in food-aversion learning may be considered to involve serial compound conditioning. When subjects approach and ingest a poisonous food, they experience first its visual/olfactory properties, then its taste, and then its poisonous effects. Recent reports indicate that potentiation effects can be also obtained with serialcompound procedures in nictitating membrane conditioning of rabbits (Kehoe, Gibbs, Garcia, & Gormezano, 1979), fear conditioning of rats (Pearce, Nicholas, & Dickinson, 1981), and autoshaping in pigeons (Rescorla, 1982). In one of the nictitating membrane conditioning experiments (Kehoe ef al., 1979; Experiment 3), for example, a 400-msec tone (Sl) and a 400-msec flashing light (S2)served as conditioned stimuli. For serial compound conditioning, S1 was presented 1800 msec before the unconditioned stimulus, and S2 was presented 400 msec before the US. Responding to S1 was far greater in subjects given serial compound conditioning than in subjects that received conditioning of S1 and S2 on separate trials with the same CSUCS intervals. The potentiated conditioning of S1 in the serial compound group could not be attributed to stimulus generalization from S2. Results of the serial compound conditioning experiments are closely analogous to findings of potentiation in poison-avoidance learning. In the poison-avoidance experiments, the presence of a stimulus that readily becomes associated with the US (taste) facilitates the conditioning of another cue that is ordinarily more difficult to condition (e.g., odor). In the serial conditioning experiments, S2 was more readily conditioned than S1 because it was presented in a more favorable temporal relationship to the unconditioned stimulus. The presence of this more easily conditioned stimulus facilitated the conditioning of the less easily conditioned S1. Another similarity is that more potentiation occurs in both nictitating membrane and poisonavoidance conditioning when the interval between S1 and the unconditioned stimulus is longer (see Kehoe et al., 1979, Experiment 2; Palmerino ef al., 1980, Experiment 2). Investigators have also tested whether potentiation is


263

mediated by an association between S2 and S1 by extinguishing S2 following serial compound conditioning. As Lett (1980b) observed in poisonavoidance learning, extinction of S2 was observed not to reduce potentiated responding to S1 in both conditioned suppression and autoshaping (Pearce et al., 1981; Rescorla 1982). Results analogous to potentiation in food-aversion learning have been also observed in appetitive conditioning of honeybees (Couvillon & Bitterman, 1982). Honeybees were reinforced with sucrose for landing on a jasmine-scented target. The presence of an orange-colored disk on the target potentiated conditioning of a preference for the jasmine odor.

D. CONCLUDING COMMENTS The potentiation effect has not been investigated sufficiently extensively in either the feeding system or in other conditioning preparations to permit firm conclusions about its mechanisms. We do not know, for example, whether the potentiation effects observed in nictitating membrane conditioning of rabbits, fear conditioning of rats, autoshaping of pigeons, and appetitive conditioning of honeybees are produced by the same factors that are responsible for potentiation in poison-avoidance learning. However, given the large range of situations in which potentiation effects occur, it seems premature to regard potentiation as an adaptive specialization of the feeding system. The similarities discovered to date in the parametric features of potentiation in the feeding system and nonfeeding situations suggest that more general mechanisms may be invovled.

VIII. Implications for the Study of General Process Learning Theory The issue of biological constraints on learning attracted a great deal of attention when it was brought into focus about 10 years ago because it highlighted numerous phenomena that were unexpected on the basis of commonly held theories and beliefs about learning. In particular, these phenomena challenged the assumption that learning occurs the same way regardless of the cues, responses, and reinforcers involved in a learning situation. This equipotentiality principle was viewed as a fundamental assumption of general process learning theory, and the inadequacy of the principle was considered good reason to dramatically alter traditional approaches to the study of learning (e.g., Seligman & Hager, 1972; Shettleworth, 1972; Rozin & Kalat, 1971). Given what we have learned about the phenomena that originally stimulated thinking about biological constraints,

Michael Domjan

264

do these phenomena still require rejection of the equipotentiality principle, and if so does this justify abandoning the search for a general theory of learning? Our review of the literature suggests that the answer to the first question is “Yes,” and the answer to the second question is “No.” OF ANALYSIS OF CONDITIONING A. LEVELS

PROCEDURES

Research during the past 10 years has added substantially to the mounting body of evidence showing that learning does not occur the same way regardless of the cues, responses, and reinforcers employed. We have numerous new and well-documented instances of the differential effectiveness of positive reinforcement and punishment in changing the probability of various responses. We also have numerous new instances of the differential conditioning of cues as a function of the reinforcer. This body of evidence strongly suggests that a general theory of learning built on the assumption that cues, responses, and reinforcers are entirely interchangeable cannot succeed. Given that we cannot make generalizations about learning at the level of cues, responses, and reinforcers, are generalizations possible at some other level? Recent research on biological constraints on learning illustrates that to explain the outcome of instrumental conditioning procedures one has to consider more than just the response-reinforcer contingency, and to understand the outcome of classical conditioning procedures one has to consider more than just the cue-reinforcer contingency. For example, studies of the excessive handling of tokens that occurs when animals are required to deposit tokens into a hole for food or water have shown that it is insufficient to consider only the experimentally defined aspects of this task (token release for food or water). One also has to consider the stimuli the subjects experience in connection with executing the instrumental response (features of the tokens) and the relationship of these cues to presentations of the reinforcer. Such a more detailed analysis of the situation led to explanation of the “misbehavior” in terms of stimulus-reinforcer contingencies (Timberlake el al., 1982). Understanding of the differential effectiveness of positive reinforcement and punishment in modifying various responses also required more detailed analyses than what is provided by merely considering the response and reinforcer involved. In this case we saw that whether learning occurred was influenced by the unconditioned motivational state in which the instrumental conditioning was conducted and the motivational state that was conditioned by presentations of the reinforcer. Learning was also related to the discriminability of the responses involved, the supporting stimulation pro-


265

vided for the responses, and possibly the innate behavioral sequence to which the responses belonged. Studies of the differential effectiveness of avoidance procedures in conditioning various responses showed that the particular stimuli animals experienced prior to and following an avoidance response are critical for predicting learning. Studies of constraints on classical conditioning showed that whether one stimulus (S 1) becomes conditioned by another (S2) depends on orientation to the cues, the stimulus features S1 and S2 have in common, the features that command attention, and the power of these features to produce learning. Analyses of long-delay learning indicated that we must consider the stimuli that occur immediately after the cue or response to be conditioned, the stimuli that occur during the delay interval, and the cues that accompany the reinforcer. We must also consider whether we are conditioning affective or instrumental responses and the possible interrelation of cue, response, and reinforcer in the situation.

B. GENERALITY OF MOLECULAR MECHANISMS OF CONDITIONING The brief summary provided above indicates that understanding of various instances of constraints on learning has required more detailed analyses than what are provided by simply focusing on the response-reinforcer and cue-reinforcer contingencies in instrumental and classical conditioning procedures, respectively. Conceivably such detailed analyses might reveal that various instances of constraints on learning are governed by specialized mechanisms that cannot be applied to other situations. However, this has not been the thrust of the research during the past decade. The factors that have been identified as relevant to the “misbehavior” observed when animals are reinforced for depositing tokens into a hole are part of a new general conception of classical conditioning (Timberlake, 1982; Timberlake ei al., 1982). The motivational and stimulus factors that have been discussed as relevant to the understanding of constraints on positive reinforcement and punishment are likewise potentially relevant to all instances of positive reinforcement and punishment. The new explanation of constraints on avoidance learning, the consummatory stimulus reward hypothesis, represents the extension of a general formulation about reinforcement to avoidance learning. The stimulus similarity hypothesis and various mechanisms of it that have been proposed to explain selective aversion conditioning are also general mechanisms. Finally, many of the various factors that have been discussed as responsible for long-delay taste-aversion learning sought to explain this effect in terms of mechanisms that presumably can produce long-delay learning in other situations as well. Thus, detailed

266

Michael Domjan

experimental analyses of various constraints on learning have led to explanations of these phenomena in terms of generally applicable principles of behavior. The general-process explanations offered for various instances of biological constraints are consistent with many of the facts known about these phenomena. The explanations are not stated in terms of interchangeable cues, responses, and reinforcers, but in terms of more detailed features of a learning situation. If these elementary features are in fact responsible for the constraint phenomena, similar results should occur when the critical features appear in a new learning situation, with different cues, responses, and reinforcers. In fact, the strongest evidence for a general-process explanation of a biological constraint phenomenon is the accurate prediction of similar effects in new situations. Such evidence has been obtained in several aspects of research in constraints on learning. Studies of the mechanisms of avoidance learning, for example, suggested that transient stimuli encountered during flight serve to reinforce avoidance behavior (Masterson et al., 1978). This led to formulation of the consummatory stimulus reward hypothesis, which predicts that any response can be acquired in avoidance conditioning provided the response is reinforced with flight-related stimuli. Consistent with this prediction, Crawford and Masterson (1978) found that rats can rapidly learn to press a lever to avoid shock when the lever-press response is reinforced by flight-related cues. Many of the examples of selective associations in classical conditioning were discovered during tests of hypotheses proposed to explain selective conditioning of taste with illness and audiovisual cues with shock (Garcia & Koelling, 1966). Studies of selective associations in higher order conditioning (Rescorla & Cunningham, 1979; Rescorla & Furrow, 1977) and fear conditioning (Testa, 1975) were designed to test the stimulus similarity explanation of the Garcia-Koelling effect. Selective conditioning of audiovisual cues with immediate shock and flavor cues with delayed shock was discovered as confirmation of the idea that the Garcia-Koelling effect is in part due to the temporal relationship of CS- and US-induced sensations (Krane & Wagner, 1975). Investigations of the mechanisms of long-delay poison-avoidance learning similarly yielded instances of the phenomenon in new situations. The concurrent interference explanation of long-delay taste-aversion learning stimulated a search for instances of long-delay learning in the T maze. These investigations demonstrated that rats can learn spatial discriminations in the T maze with delays of reinforcement of several minutes (e.g., Lett, 1975; Lieberman et al., 1979). Considerations of the role of affective learning and the interrelation of cue, response, and reinforcer in long-delay food-aversion learning also stimulated demonstrations of long-delay learning outside


267

the feeding system. D’Amato and Buckiewicz (1980) observed spatial preference learning with a 30-min delay of reinforcement in monkeys, and Safarjan and D’Amato (1981) reported spatial preference learning in rats with delays of reinforcement as long as 120 min. Finally, Sullivan (1979) observed conditioned aversions to a small object in rats when touching the object resulted in shock 35 min later. New discoveries of selective conditioning and long-delay learning predicted by general mechanisms proposed to explain these effects in poisonavoidance learning provide strong support for the explanations. In other cases, phenomena similar to a constraint on learning have been discovered in new situations without being predicted by explanations of the original findings. For example, selective conditioning of the color of food and the taste of water with poisoning in young chickens (Gillette et al., 1980) was not discovered during tests of hypotheses to explain other instances of selective associations. Studies of selective conditioning of visual cues with food and auditory cues with shock in pigeons (Shapiro et al., 1980) and the differential conditionability of visual and auditory stimuli as cues for danger and safety in rats (Jacobs & LoLordo, 1980) were also not conducted as tests of explanations of selective associations in other situations. Finally. studies of the potentiation effect in rabbit nictitating membrane conditioning, rat fear conditioning, pigeon autoshaping, and honeybee appetitive conditioning (Couvillon & Bitterman, 1982; Kehoe et al., 1979; Pearce et al., 1981; Rescorla, 1982) were conducted independently of studies of potentiation in poison-avoidance learning. Nevertheless, these demonstrations of selective conditioning and potentiation in diverse situations challenge the proposition that the phenomena represent unique adaptations rather than elementary general processes. The fact that selective conditioning, long-delay learning, and potentiation can be observed in a variety of species and with diverse stimuli, responses, and reinforcers might be seen as a return to the equipotentiality principle. However, these effects probably do not occur in diverse situations because the particular stimuli, responses, and reinforcers involved are irrelevant to the underlying mechanisms of the effects. Rather, the phenomena appear to be the products of elementary behavioral processes that may be activated by more than one but not all combinations of cues, responses, and reinforcers. Cues, responses, and reinforcers are interchangeable only to the extent that they continue to activate the elementary general processes responsible for the phenomena. The above discussion suggests that investigations of biological constraints on learning have not stimulated a retreat from the search for general principles of behavior. However, the pursuit of a general-process learning theory has not been unaffected by these investigations. Studies of constraints

268

Michael DomJan

on learning have provided numerous new insights into the mechanisms of classical and instrumental conditioning. Thus, investigations of biological constraints have significantly contributed to our knowledge of elementary general mechanisms of learning.

ACKNOWLEDGMENTS This chapter was prepared while the author was visiting McMaster University, supported by a Faculty Research Assignment from the University of Texas. The author wishes to thank S. Siegel, M. Daley, B. G. Galef, Jr., and others for their hospitality during his stay at McMaster University.

REFERENCES Adams, J. A. Learning and memory: An introduction. Homewood, Illinois: Dorsey, 1976. Aitken, P. P., & Sheldon, M. H. Electric shock and rats’ preference for the familiar arm of a maze. British Journal ojPsychology, 1970, 61, 95-97. Anderson, M. C., & Shettleworth, S. J. Behavioral adaptation to fixed-interval and fixedtime food delivery in golden hamsters. Journal of the Experimental Analysis of Behavior, 1977, 21, 33-49.

Annable, A., & Wearden, J. H. Grooming movements as operants in the rat. Journal of the Experimental Analysis of Behavior, 1979, 32, 297-304. Barnett, S. The rat: A study in behavior. Chicago: Aldine, 1963. Berger, D. F., & Brush, F. R. Rapid acquisition of discrete-trial lever-press avoidance: Effects of signal-shock interval. Journal of the Experimental Analysis of Behavior, 1975, 24, 227-239.

Best, M. R. Conditioned and latent inhibition in taste-aversion learning: Clarifying the role of learned safety. Journal of Experimental Psychology: Animal Behavior Processes, 1975, 1, 97-113. Best, M. R., & Barker, L. M. The nature of “learned safety” and its role in the delay of reinforcement gradient. In L. M. Barker, M. R. Best, & M. Domjan (Eds.), Learning mechanisms in food selection. Waco, Texas: Baylor Univ. Press, 1977. Best, M. R., & Gemberling. 0 . A. Role of short-term processes in the conditioned stimulus preexposure effect and the delay of reinforcement gradient in long-delay taste-aversion learning. Journal of Experimental Psychology: Animal Behavior Processes, 1977, 3, 253-263.

Bindra, D. A motivational view of learning, performance, and behavior modification. Psychological Review, 1974, 81, 199-213. Bindra, D. How adaptive behavior is produced: A perceptual-motivational alternative to response-reinforcement. The Behavioral and Brain Sciences, 1978, 1, 41-91. Blanchard, R. J., & Blanchard. D. C. Crouching as an index of fear. Journal of Comparative and Physiological Psychology, 1969, 61, 370-375. (a) Blanchard, R. J.. & Blanchard, D. C. Passive and active reactions to fear-eliciting stimuli. Journal of Comparative and Physiological Psychology, 1969, 68, 129-135. (b) Blanchard, R. J.. & Blanchard, D. C. Defensive reactions in the albino rat. Learning and Motivation, 1971, 2, 351-362.


269

Boakes, R. A., Poli, M., Lockwood, M. J., & Goodall, G. A study of misbehavior: Token reinforcement in the rat. Journal of the Experimental Analysis of Behavior, 1978, 29, 115-134.

Bolles, R. C. Avoidance and escape learning: Simultaneous acquisition of different responses. Journal of Comparative and Physiological Psychology, 1969, 68, 355-358. Bolles, R. C. Species-specific defense reactions and avoidance learning. Psychological Review, 1970. 17, 32-48.

Bolles, R. C. Species-specific defense reaction. In F. R . Brush (Ed.), A versive conditioning and learning. New York: Academic Press, 1971. Bolles, R. C. Learning theory. New York: Holt, 1975. Bolles, R. C., & Riley, A. L. Freezing as an avoidance response: Another look at the operantrespondent distinction. Learning and Motivation, 1973, 4, 268-275. Bolles, R. C., Riley, A. L., & Laskowski, B. A further demonstration of the learned safety effect in food aversion learning. Bulletin of the Psychonomic Society, 1973, 1, 190-192. Bouton, M. E., & Whiting, M. R. Simultaneous odor-taste and taste-taste compounds in poison-avoidance learning. Learning and Motivation, 1982, 13, 472-494. Braveman, N. S. Poison-based avoidance learning with flavored or colored water in guinea pigs. Learning and Motivation, 1974, 5 , 182-194. Breland. K., & Breland, M. The misbehavior of organisms. American Psychologist, 1961, 16, 681-684.

Brener, J., & Goesling, W. J . Avoidance conditioning of activity and immobility in rats. Journal of Comparative and Physiological Psychology, 1970, 70, 276-280. Charlton, S. G., & Ferraro, D. P. Effects of deprivation on the differential conditionability of behaviour in golden hamsters. Experimental Animal Behavior, 1982, 1, 18-29. Clarke, J. C.. Westbrook, R. F., & Irwin, J. Potentiation instead of overshadowing in the pigeon. Behavioral and Neural Biology, 1979, 25. 18-29. Couvillon, P. A., & Bitterman, M. E. Compound conditioning in honeybees. Journal of Comparative and Physiological Psychology, 1982, 96, 192-199. Crawford, M., & Masterson, F. Components of the flight response can reinforce bar-press avoidance learning. Journal of Experimental Psychology: Animal Behavior Processes, 1978, 4, 144-151.

Dalrymple, A. J., & Galef, B. G.. Jr. Visual discrimination pretraining facilitates subsequent visual cue/toxicosis conditioning in rats. Bulletin of the Psychonomic Society, 1981, 18. 267-270.

D’Amato. M. R., & Buckiewicz, J . Long-delay, one-trial conditioned preference and retention in monkeys (Cebus apella). Animal Learning & Behavior, 1980, 8, 359-362. D’Amato, M. R., Buckiewicz, J., & Puopolo, M. Long-delay spatial discrimination learning in monkeys (Cebus apella). Bulletin of the Psychonomic Society, 1981, 18, 85-88. D’Amato, M. R., & Safarjan, W. R. Differential effects of delay of reinforcement on acquisition of affective and instrumental responses. Animal Learning & Behavior, 1981, 9. 209-21 5.

D’Amato, M. R., Salmon, D. P., & Puopolo, M. Long-delay visual discrimination learning in monkeys (Cebus apella). Bulletin of the Psychonomic Society, 1981, 18, 89-91. D’Amato, M. R., & Schiff, D. Long-term discriminated avoidance performance in the rat. Journal of Comparative and Physiological Psychology, 1964, 51, 123-126. Davis, H., & Burton, J. Reinforcement of lever holding by avoidance of shock. Bulletin of the Psychonomic Society, 1976, 8, 61-64. Delius, J. D.. & Emmerton, J. Stimulus-dependent assymetry in classical and instrumental discrimination learning by pigeons. Psychological Record, 1978, 28, 425-434. Delprato, D. J., & Holmes, P. A. Facilitation of discriminated lever-press avoidance by noncontingent shock. Learning and Motivation. 1977, 8, 238-246.

270

Michael Domjan

Dickinson, A. Contemporary animal learning. London and New York: Cambridge Univ. Press, 1980. Dickinson, A., & Pearce, J. M. Inhibitory interactions between appetitive and aversive stimuli. Psychological Bulletin, 1977, 84, 690-71 I . D’Mello, G. D., Goldberg, D. M., Goldberg, S. R., & Stolerman, I. P. Conditioned taste aversion and operant behavior in rats: Effects of cocaine and a cocaine analogue (Win 35.428). Neuropharmacology, 1979, 18, 1009-1010. D’Mello, G. D., Goldberg, D. M., Goldberg, S. R., & Stolerman, 1. P. Conditioned taste aversion and operant behavior in rats: Effects of cocaine, apomorphine and some longacting derivatives. Journal of Pharmacology and Experimental Therapeutics, 1981, 219, 60-68. Domjan, M. Selective suppression of drinking during a limited period following aversive drug treatment in rats. Journal of Experimental Psychology: Animal Behavior Processes, 1917, 3, 66-76. Domjan, M. Ingestional aversion learning: Unique and general processes. In J. S. Rosenblatt. R. A. Hinde, C. Beer, & M.-C. Busnel (Eds.), Advances in the study of behavior (Vol. 11). New York: Academic Press, 1980. Domjan, M., & Bowman, T. G. Learned safety and the CS-US delay gradient in taste-aversion learning. Learning and Motivation. 1974, 5, 409-423. Domjan, M.,Foster, K., & Gillan, D. 0.Effects of distribution of the drug unconditioned stimulus on taste-aversion learning. Physiology & Behuvior, 1979, 23, 931-938. Domjan, M., & Hanlon, M. J . Poison-avoidance learning to food-related tactile stimuli: Avoidance of texture cues by rats. Animal Learning & Behavior, 1982, 10, 293-300. Domjan, M., & Levy, C. J. Taste aversions conditioned by the aversiveness of insulin and formalin: Role of CS specificity. Journal of Experimental Psychology: Animal Behavior Processes, 1977. 3, 119-131. Domjan, M., Miller. V., & Gemberling, G. A. Note on aversion learning to the shape of food by monkeys. Journal of the Experimental Analysis of Behavior, 1982, 38, 87-91. Domjan, M., & Wilson, N. E. Specificity of cue to consequence in aversion learning in the rat. Psychonomic Science, 1972, 26, 143-145. (a) Domjan, M., & Wilson, N. E. Contribution of ingestive behaviors to taste-aversion learning in the rat. Journalof Comparative and Physiological Psychology, 1972,80,403-412. (b) Durlach, P. J., & Rescorla, R. A. Potentiation rather than overshadowing in flavor-aversion learning: An analysis in terms of within-compound associations. Journal of Experimental Psychology: Animal Behavior Processes, 1980, 6 , 175-187. Feldman, R. S., & Bremner, F. J . A method for rapid conditioning of stable avoidance bar pressing behavior. Journal of the Experimental Analysis of Behavior, 1963, 6 , 393-394. Ferrari, E. A., Todorov, J. C.,& Graeff, F. G. Nondiscriminated avoidance of shock by pigeons pecking a key. Journal of the Experimental Analysis of Behavior, 1973, 19, 211-218. Flaherty, C. F., Hamilton, L. W., Gandelman, R. J., & Spear, N. E. Learning and memory. Chicago, Illinois: Rand McNally. 1977. Foree, D. D., & LoLordo, V. M. Signalled and unsignalled free-operant avoidance in the pigeon. Journal of the Experimental Analysis of Behavior, 1970, 13. 283-290. Foree, D. D., & LoLordo, V. M. Attention in the pigeon: Differential effects of food-getting versus shock-avoidance procedures. Journal of Comparative and Physiological Psychology, 1973, 85, 551-558. Foree, D. D., & LoLordo, V. M. Transfer of control of the pigeon’s key peck from food reinforcement to avoidance of shock. Journal of the Experimental Analysis of Behavior, 1974, 22, 251-259.


27 1

Foree, D. D., & LoLordo, V. M. Stimulus-reinforcer interactions in the pigeon: The role of electric shock and the avoidance contingency. Journal of Experimental Psychology: Animal Behavior Processtq 1975, 1, 39-46. Frontali, M., & Bignami, G. Go-no go avoidance discriminations in rats with simple “go” and compound “no go” signals: Stimulus modality and stimulus intensity. Animal Learning & Behavior, 1973, 1, 21-24. Frontali, M., & Bignami, G. Stimulus nonequivalences in golno-go avoidance discriminations: Sensory, drive, and response factors. Animal Learning & Behavior, 1974, 2, 153-160. Galef, B. G., Jr., & Dalrymple, A. J. Toxicosis-based aversions to visual cues in rats: A test of the Testa and Ternes hypothesis. Animal Learning & Behavior, 1981, 9, 332334.

Galef, B. G., Jr., & Osborne, B. Novel taste facilitation of the association of visual cues with toxicosis in rats. Journal of Comparative and Physiological Psychology, 1978, 92, 907-916.

Garcia, J., Ervin, F. R., & Koelling, R. A. Learning with prolonged delay of reinforcement. Psychonomic Science, 1966, 5 , 121-122. Garcia, J., Hankins, W. G., & Rusiniak, K. W. Behavioral regulation of the milieu interne in man and rat. Science, 1974, 185, 824-831. Garcia, J., & Koelling, R. A. Relation of cue to consequence in avoidance learning. Psychonomic Science, 1966, 4, 123-124. Garcia, J., McGowan, B. K., Ervin, F. R., & Koelling. R. A. Cues: Their relative effectiveness as a function of the reinforcer. Science, 1968, 160, 794-795. Garcia, J., & Rusiniak, K. W. What the nose learns from the mouth. In D. Muller-Schwarze & R. M. Silverstein (Eds.), Chemical signals. New York: Plenum, 1980. Gemberling, 0 . A., & Domjan, M. Selective associations in one-day-old rats: Taste-toxicosis and texture-shock aversion learning. Journal of Comparative and Physiological PsyChOlOgV, 1982, 96, 105-1 13. Gemberling, G . A., Domjan, M., & Amsel, A. Aversion learning in 5-day-old rats: Tastetoxicosis and texture-shock associations. Journal of Comparative and Physiological Psychology, 1980, 94, 734-745. Gillette, K., Martin, G. M., & Bellingham, W. P. Differential use of food and water cues in the formation of conditioned aversions by domestic chicks (Callus gallus). Journal of Experimental Psychology: Animal Behavior Processes, 1980, 6 , 99-1 11. Giulian, D., & Schmaltz, L. W. Enhanced discriminated bar-press avoidance in the rat through appetitive preconditioning. Journal of Comparative and Physiological Psychology, 1973, 83, 106-112.

Gleitman, H. Psychology. New York: Norton, 1981. Goudie, A. J., & Dickins, D. W. Nitrous oxide-induced conditioned taste aversions in rats: The role of duration of drug exposure and its relation to the taste aversion-self-administration “paradox.” Pharmacology Biochemistry & Behavior, 1978, 9, 587-592. Graf, V., & Bitterman, M. E. General activity as instrumental: Application to avoidance training. Journal of the Experimental Analysis of Behavior, 1963, 6 , 301-305. Green, K. F., Holmstrom, L. S., & Wollman, M. A. Relation of cue to consequence in rats: Effect of recuperation from illness. Behavioral Biology, 1974, 10, 491-503. Green, K . F., & Parker, L. Gustatory memory: Incubation and interference. Behavioral Biology, 1975, 13, 359-367. Green, L., Bouzas, A., & Rachlin, H. Test of an electric-shock analog to illness-induced aversion. Behavioral Biology, 1972, 7 , 513-518. Gregg, B., Kittrell, E. M. W., Domjan, M., & Amsel, A. Ingestional aversion learning in preweanling rats. Journal of Comparative and Physiological Psychology, 1978, 92, 785-795.

Michael DomJan

272

Grossen, N. E.,& Kelley, M. J. Species-specific behavior and acquisition of avoidance behavior in rats. Journal of Comparative and Physiological Psychology, 1972,81,307-3 10. Hall, J. F. Classical conditioning and instrumental learning: A contemporary approach. Philadelphia, Pennsylvania: Lippincott. 1976. Hearst, E.. & Jenkins, H. M. Sign-tracking: The stimulus-reinforcer relation and directed action. Austin, Texas: Psychonomic Society, 1974. Hinde, R. A., & Stevenson-Hinde, J. (Eds.). Constraints on learning. New York: Academic Press, 1973. Hoffman, H. S., & Fleshler, M.Aversive control with the pigeon. Journal ofthe Experimental Analysis of Behavior. 1959, 2,213-218.

Hogan, J. A., & Roper, T. J. A comparison of the properties of different reinforcers. In J. S. Rosenblatt, R. A. Hinde, C. Beer, & M.-C. Busnel (Eds.), Advances in thestudy of behavior (Vol. 8). New York: Academic Press, 1978. Holland, P. C. Conditioned stimulus as a determinant of the form of the Pavlovian conditioned response. Journal of Experimental Psychology: Animal Behavior Processes, 1977, 3, 77-104.

Holland, P. C. Acquisition of representation-mediated conditioned food aversions. Learning and Motivation. 1981, 12, 1-18.

Houston, J. P. Fundamentals of learning. New York: Academic Press, 1976. Hulse, S. H.. Deese. J., & Egeth, H. The psychology of learning. New York: McGraw-Hill, 1975.

.Jacobs, W. J., & LoLordo, V. M. The sensory basis of avoidance responding in the rat. Learning and Motivation, 1977, 8, 448-466.

Jacobs, W. J., & LoLordo, V. M. Constraints on Pavlovian aversive conditioning: Implications for avoidance learning in the rat. Learning and Motivation. 1980, 11, 427-455. Jenkins, H.M., & Moore, B. R. The form of the autoshaped response with food or water reinforcers. Journal of the Experimental Analysis of Behavior, 1973, 20, 163-181. Kalat, J. W.Status of “learned safety” or “learned non-correlation” as a mechanism in taste aversion learning. In L. M. Barker, M. R. Best, & M. Domjan (Eds.). Learning mechanisms in food selection. Waco, Texas: Baylor Univ. Press, 1977. Kalat. J. W., & Rozin, P. Role of interference in taste-aversion learning. Journal of Comparative and Physiological Psychology, 1971, 77, 53-58.

Kalat, J. W., & Rozin, P. “Learned safety” as a mechanism in long-delay taste-aversion learning in rats. Journal of Comparative and Physiological Psychology, 1973, 83, 198-207.

Kamin, L. J. Predictability, surprise, attention. and conditioning. In B. A. Campbell & R. M. Church (Eds.), Punishment and aversive behavior. New York: Appleton, 1%9. Kehoe, E. J., Gibbs. C. M., Garcia, A.. & Gormezano, I. Associative transfer and stimulus selection in classical conditioning of the rabbit’s nictitating membrane response to serial compound CS. Journal of Experimental Psychology: Animal Behavior Processes, 1979, 5, 1-19.

Kimble. G. A. Hilgard and Marquis’ conditioning and learning. New York: Appleton. 1961. Konorski, J. Integrative activity of the brain. Chicago, Illinois: Univ. of Chicago Press, 1967.

Krane, R. V., & Wagner, A. R. Taste aversion learning with a delayed shock US: Implications for the “generality of the laws of learning”. Journal of Comparative and Physiological Psychology, 1975, 88, 882-889. Kulkarni, A. S., & Job, W. M. Instrumental response pretraining and avoidance acquisition in rats. Journal of Comparative and Physiological Psychology, 1970, 70, 254-257.


273

Lavin, M. J. The establishment of flavor-flavor associations using a sensory preconditioning training procedure. Learning and Motivation, 1976, 7 , 173-183. Lett, B. T. Delayed reward learning: Disproof of the traditional theory. Learning and Motivation. 1973, 4, 237-246. Lett, B. T. Visual discrimination learning with a I-min delay of reward. Learning and Motivation, 1974, 5, 174-181. Lett, B. T. Long delay learning in the T maze. Learning and Motivation, 1975, 6 , 80-90. Lett, B. T. Regarding Robert’s reported failure to obtain visual discrimination learning with delayed reward. Learning and Motivation, 1977, 8, 136-139. Lett, B. T. Long-delay learning: lmplications for learning and memory theory. In N. S. Sutherland (Ed.), Tutorial essays in psychology (Vol. 2). Hillsdale, New Jersey: Erlbaum, 1979. Lett, B. T. Taste potentiates color-sickness associations in pigeons and quail. Animal Learning & Behavior, 1980, 8, 193-198. (a) Lett, B. T. Tustepotentiation in poison avoidance learning. Paper presented at the third annual Harvard Symposium on Quantitative Analysis of Behavior, Harvard University, June, 1980. (b) Lewis, P., Lewin, L., Stoyak, M., & Muehleisen, P. Negatively reinforced key pecking. Journal of the Experimental Analysis of Behavior, 1974, 22, 83-90. Lieberman, D. A., Mclntosh, D. C., & Thomas, G. V. Learning when reward is delayed: A marking hypothesis. Journal of Experimental Psychology: Animal Behavior Processes, 1979, 5, 224-242. Logue, A. W. Visual cues for illness-induced aversions in the pigeon. Behavioral and Neural Biology, 1980, 28, 372-377. LoLordo, V. M. Selective associations. In A. Dickinson & R. A. Boakes (Eds.), Mechanisms of learning and motivation. Hillsdale, New Jersey: Erlbaum, 1979. LoLordo, V. M., & Furrow, D. R. Control by the auditory or the visual element of a compound discriminative stimulus: Effects of feedback. Journal of the Experimental Analysis of Behavior, 1976, 25, 251-256. LoLordo, V. M., Jacobs, W. J., & Foree, D. D. Failure to block control by a relevant stimulus. Animal Learning & Behavior. 1982, 10, 183-193. Lubow, R. E. Latent inhibition. Psychological Bulletin, 1973, 79, 398-407. Maatsch, J. L. Learning and fixation after a single shock trial. Journal of Comparative and Physiological Psychology, 1959, 52, 408-410. MacKay, B. Visual and flavor cues in toxicosis conditioning of codfish. Behavioral Biology, 1977, 19, 87-97. Mackintosh, N. J. Stimulus selection: Learning to ignore stimuli that predict no change in reinforcement. In R. A. Hinde & J. Stevenson-Hinde (Eds.), Constraints on learning. New York: Academic Press, 1973. Mackintosh, N. J. The psychology of animal learning. New York: Academic Press, 1974. Mackintosh, N. J. A theory of attention: Variation in the associability of stimuli with reinforcement. Psychological Review, 1975, 82, 276-298. Macphail, E. M. Avoidance responding in pigeons. Journal of the Experimental Analysis of Behavior, 1968, 11. 629-632. Maier, S. F., & Jackson, R. L. Learned helplessness: All of us were right (and wrong): lnescapable shock has multiple effects. In G . H. Bower (Ed.), The psychology of learning and motivation (Vol. 13). New York: Academic Press, 1979. Maier, S. F., & Seligman, M. E. P. Learned helplessness: Theory and evidence, Journal of Experimental Psychology: General. 1976, 105. 3-46.

Michael Domjnn

274

Martin, G. M., & Bellingham, W. P. Learning of visual food aversions by chickens (Callus gallus) over long delays. Behavioral and Neural Biologv. 1979, 25, 58-68. Masterson, F. A., & Crawford, M. The defense motivation system: A theory of avoidance behavior. The Behavioral and Brain Sciences, 1982, 5, 661-696. Masterson, F. A., Crawford, M., & Bartter, W.D. Brief escape from a dangerous place: The role of reinforcement in the rat's one-way avoidance acquisition. Learning and Motivation, 1978, 9, 141-163. Melvin, K. B., & Ervey, D. H. Facilitative and suppressive effects of punishment of speciestypical aggressive display in Betta splendens. Journal of Comparative and Physiological Psychology, 1973, 83, 451-457. Meyer, D. R.. Cho, C., & Wesemann, A. S. On problems of conditioning discriminated leverpress avoidance response. Psychological Review, 1960, 67, 224-228. Miller, V.. & Domjan, M. Selective sensitization induced by lithium malaise and footshock in rats. Behavioral and Neural Biology, 1981. 31, 42-55. (a) Miller, V.. & Domjan, M. Specificity of cue to consequence in aversion learning in the rat: Control for US-induced differential orientations. Animal Learning & Behavior, 1981, 9, 339-345. (b) Morgan, C. T., & King, R. A . Introduction to psychology. New York: McGraw-Hill, 1975. Morgan, M. J., & Nicholas, D. J. Discrimination between reinforced action patterns in the rat. Learning and Motivation, 1979, 10, 1-22. Myers, A. K. Avoidance learning as a function of several training conditions and strain differences in rats. Journal of Comparative and Physiological Psychology, 1959, 52, 381-386.

Myers, A. K. Discriminated operant avoidance learning in Wistar and G-4 rats as a function of type of warning stimulus. Journal of Comparative and Physiological Psychology, 1964, 58, 453-455.

Nachman, M., & Jones, D. R. Learned taste aversions over long delays in rats: The role of learned safety. Journal of Comparative and Physiological Psychology, 1974, 86, 949-956.

Nachman, M., Rauschenberger, J., & Ashe, J. H. Studies of learned aversions using nongustatory stimuli. In L. M. Barker. M. R. Best, & M. Domjan (Eds.), Learning mechanisms in food selection. Waco, Texas: Baylor Univ. Press, 1977. Nevin, J. A., & Reynolds, G . S . (Eds.), The study of behavior. Glenview, Illinois: Scott Foresman, 1973. Palmerino, C. C., Rusiniak, K. W.,& Garcia, J. Flavor-illness aversions: The peculiar roles of odor and taste in memory for poison. Science. 1980, 208, 753-755. Pavlov, I. P. Conditioned reflexes. London and New York: Oxford Univ. Press, 1927. Pearce, J. M.. Colwill, R. M., & Hall, G. Instrumental conditioning of scratching in the laboratory rat. Learning and Motivation, 1978, 9, 255-271. Pearce, J. M., Nicholas, D. J., & Dickinson, A. The potentiation effect during serial conditioning. Quarterly Journal of Experimental Psychology, 1981, 33B, 159-179. Rachlin, H.Behavior and Learning. San Francisco, California: Freeman, 1976. Rachlin, H., & Hineline, P. N. Training and maintenance of key pecking in the pigeon by negative reinforcement. 'Science, 1967, 157, 954-955. Rescorla, R. A. Pavlovian second-order conditioning. Hillsdale, New Jersey: Erlbaum, 1980. Rescorla, R. A. Effect of a stimulus intervening between CS and US in autoshaping. Journal of Experimental Psychology: Animal Behavior Processes, 1982, 8, 131-141. Rescorla, R. A., & Cunningham, C. L. Spatial contiguity fatilitates Pavlovian second-order conditioning. Journal of Experimental Psychology: Animal Behavior Processes, 1979. 5, 152-161.

Rescorla, R. A., & Furrow, D. R. Stimulus similarity as a determinant of Pavlovian condi-


275

tioning. Journal of Experimental Psychology: Animal Behavior Processes. 1977, 3, 203-215. Rescorla, R. A., & Gillan, D. G. An analysis of the facilitative effect of similarity on secondorder conditioning. Journal of Experimental Psychology: Animal Behavior Processes. 1980,6, 339-351. Rescorla, R. A., & Holland, P. C. Some behavioral approaches to the study of learning. In M. R. Rosenzweig & E. L. Bennett (Eds.), Neural mechanbms of learning and memory. Cambridge, Massachusetts: MIT Press, 1976. Rescorla, R. A., & Wagner, A. R. A theory of Pavlovian conditioning: Variations in the effectiveness of reinforcement and nonreinforcement. In A. H. Black & W. F. Prokasy (Eds.), Classical conditioning 11: Current research and theory. New York: Appleton, 1972. Revusky, S. H. Aversion to sucrose produced by contingent x-irradiation: Temporal and dosage parameters. Journal of Comparative and Physiological Psychology, 1968, 65, 17-22. Revusky, S. H. The role of interference in association over a delay. In W. K. Honig & P. H. R. James (Eds.), Animal memory. New York: Academic Press, 1971. Revusky, S. H. The concurrent interference approach to delay learning. In L. M. Barker, M. R. Best, & M. Domjan (Eds.), Learning mechanisms in food selection. Waco, Texas: Baylor Univ. Press, 1977. Revusky, S. H., & Garcia, J. Learned associations over long delays. In G. H. Bower (Ed.), Thepsychology of learning and motivation (Vol. 4). New York: Academic Press, 1970. Revusky, S. H.,& Parker, L. Aversions to unflavored water and cup drinking produced by delayed sickness. Journal of Experimental Psychology: Animal Behavior Processes. 1976, 2, 342-355. Riess, D. A shaping technique for producing rapid and reliable Sidman barpress avoidance. Journal of the Experimental Analysis of Behavior, 1970, 13, 279-280. Roberts, W. A. Failure to replicate visual discrimination learning with a I-min delay of reward. Learning and Motivation. 1976, 7, 313-325. Roberts, W. A. Still no evidence of visual discrimination learning: A reply to Lett. Learning and Motivation, 1977, 8, 140-144. Rozin, P., & Kalat, J. W. Specific hungers and poison avoidance as adaptive specializations of learning. Psychological Review, 1971. 78, 459-486. Rudy, J. W., & Cheatle, M. D. Ontogeny of associative learning: Acquisition of odor aversions by neonatal rats. In N. E. Spear & B. A. Campbell (Eds.), Ontogeny of learning and memory. Hillsdale, New Jersey: Erlbaum, 1979. Rusiniak, K. W., Hankins, W. G . , Garcia, J., & Brett, L. P. Flavor-illness aversions: Potentiation of odor by taste in rats. Behavioral and Neural Biology, 1979,25, 1-17. Rusiniak, K.W., Palmerino, C. C., Rice, A. G . , Forthman, D. L., & Garcia, J. Flavor-illness aversions: Potentiation of odor by taste with toxin but not shock in rats. Journal of Comparative and Physiological Psychology, 1982,96, 527-539. Safarjan, W. R., & D’Amato, M. R. One-trial, long-delay, conditioned preference in rats. The Psychological Record, 1981,31, 341-426. Schindler, C. W., & Weiss, S. J. The influence of positive and negative reinforcement on selective attention in the rat. Learning and Motivation, 1982, 13, 304-323. Schwartz, B. Maintenance of key pecking in pigeons by a food avoidance but not a shock avoidance contingency. Animal Learning & Behavior, 1973, 1, 164-166. Schwartz, B. On going back to nature: A review of Seligman and Hager’s Biological boundaries of learning. Journal of the Experimental Analysis of Behavior. 1974,21, 183-198. Schwartz, B. Psychology of learning and behavior. New York: Norton, 1978. Seligman, M. E. P. On the generality of the laws of learning. Psychological Review, 1970, 77, 406-418.

Michael Domjan

276

Seligman, M. E. P., & Hager, J. L. (Eds.), Biological boundaries of learning. New York: Appleton, 1972. Shapiro, K. L., Jacobs, W. J., & Lolordo, V. M. Stimulus-reinforcer interactions in Pavlovian conditioning of pigeons: Implications for selective associations. Animal Learning & Eehavior, 1980, 8. 586-594. Shapiro, K. L., & LoLordo, V. M. Constraints on Pavlovian conditioning of the pigeon: Relative conditioned reinforcing effects of red-light and tone CSs paired with food. Learning and Motivation. 1982, 13, 68-80. Shettleworth, S. J. Constraints on learning. In D. S. Lehrman. R. A. Hinde, & E. Shaw (Eds.), Advances in the study of behavior (Vol. 4). New York: Academic Press, 1972. Shettleworth, S. J. Reinforcement and the organization of behavior in golden hamsters: Hunger, environment, and food reinforcement. Journal of Experimental Psychology: Animal Behavior Processes, 1975, 1, 56-87. Shettleworth, S. J. Reinforcement and the organization of behavior in golden hamsters: Sunflower seed and nest paper reinforcers. Animal Learning & Behavior, 1978, 6 , 352-362. (a) Shettleworth, S. J. Reinforcement and the organization of behavior in golden hamsters: Pavlovian conditioning with food and shock unconditioned stimuli. Journal of Experimental Psychology: Animal Behavior Processes, 1978,4, 152-169. (b) Shettleworth, S. J. Reinforcement and the organization of behavior in golden hamsters: Punishment of three action patterns. Learning and Motivafion, 1978. 9, 99-123. (c) Shettleworth, S. J. Reinforcement and the organization of behavior in golden hamsters: Differential overshadowing of a CS by different responses. Quarterly Journal of Experimental Psychology, 1981, 33B, 241-255. Shettleworth, S. J., & Juergensen, M. R. Reinforcement and the organization of behavior in golden hamsters: Brain stimulation reinforcement for seven action patterns. Journal of Experimental Psychology: Animal Behavior Processes, 1980, 6 , 352-375. Smith, J. C..& Roll. D. L. Trace conditioning with x-rays as an aversive stimulus. Psychonomic Science. 1967, 9, 11-12. Smith, R. E., Sarason, I. G.. & Sarason, B. R. Psychology: The frontiers of behavior. New York: Harper, 1982. Smith, R. F., & Keller, F. R. Free-operant avoidance in the pigeon using a treadle response. Journal of the Experimental Analysis of Behavior. 1970, 13, 211-214. Staddon, J. E. R., & Simmelhag, V. L. The “superstition” experiment: A reexamination of its implications for the principles of adaptive behavior. Psychological Review, 1971. 78, 3-43.

Stolerman, I. P., & D’Mello. G. D. Oral self-administrawn and the relevance of conditioned taste aversions. In D. Blackman & S . Sanger (Eds.), Advances in behavioralpharmacology (Vol. 3). New York: Academic Press, 1981. Sullivan, L. Long-delay learning with exteroceptive cue and exteroceptive reinforcement in rats. Australian Journal of Psychology, 1919, 31, 21-32. Tarpy, R. M., & Mayer, R. E. Foundations of learning and memory. Glenview, Illinois: Scott Foreman, 1978. Testa, T. J . Causal relationships and the acquisition of avoidance responses. Psychological Review, 1974, 81, 491-505. Testa, T. J. Effects of similarity of location and temporal intensity pattern of conditioned and unconditioned stimuli on the acquisition of conditioned suppression in rats. Journal of Experimental Psychology: Animal Behavior Processes, 1975, 1, 114-121. Testa, T. J., & Ternes, J. W. Specificity of conditioning mechanisms in the modification of


277

food preferences. In L. M. Barker, M. R. Best, & M. Domjan (Eds.), Learning mechanisms in food selection. Waco, Texas: Baylor Univ. Press, 1977. Thorndike, E . L. Animal intelligence: Experimental studies. New York: Macmillan, 191 1. Timberlake, W. The functional organization of appetitive behavior: Behavior systems and learning. In M. D. Zeiler & P. Harzem (Eds.), Advances in analysis of behavior (Vol. 3): Biologicalfactors in learning. New York: Wiley, 1981. Timberlake, W., & Grant, D. L. Auto-shaping in rats to the presentation of another rat predicting food. Science, 1975, 190, 690-692. Timberlake, W., Wahl, G., & King, D. Stimulus and response contingencies in the misbehavior of rats. Journal of Experimental Psychology: Animal Behavior Processes, 1982, 8. 62-85.

Walters, G. C., & Glazer, R. D. Punishment of instinctive behavior in the Mongolian gerbil. Journal of Comparative and Physiological Psychology, 1971, 75, 33 1-340. Wasserman, E. A. Pavlovian conditioning with heat reinforcement produces stimulus-directed pecking in chicks. Science, 1973, 181, 875-877. Weisinger, R. S., Parker, L. F., & Skorupski, J. D. Conditioned taste aversions and specific need states in the rat. Journal of Comparative and Physiological Psychology, 1974, 87, 655-660.

Welker, R. L., & Wheatley, K. L. Differential acquisition of conditioned suppression in rats with increased and decreased luminance levels as CS+ s. Learning and Motivation, 1977, 8, 247-262. Wickelgren, W. A. Learning and Memory. New York: Prentice-Hall, 1977. Young, G. A., & Black, A. H. A comparison of operant licking and lever pressing in the rat. Learning and Motivation, 1977, 8, 387-403.


INDEX A

“misbehavior” in depositing tokens and, 217-219 potentiation in classical conditioning and,

Abstraction, generic memory and, 107-108 Acquisition attention and, 95-98 cognitive map of city and, 138-150 structure of memory and, 14-17 Amnesia, modular memory and, 110-1 I5 Appetitive conditioning in pigeons, 246-

259-260, 263

association with toxicosis and, 260-261 nongustatory aspects of food and, 261262

outside the feeding system, 262-263 on punishment, 227-228 on selective conditioning, 235, 248 appetitive and aversive, in pigeons,

247

Attention, modular memory and, 95-99 Auditory input, structure of memory and linguistic input, 24-27 nonlinguistic input, 23 Aversion conditioning in pigeons, 246-247 with shock and illness, 235-245

246-247

aversive, with shock and illness, 235245

of danger and safety signals in rats, 247-248

of ingestive behavior, 245-246

B

C

Biological constraints, 215-277 on avoidance learning, 228, 234-235 consummatory stimulus reward hypothesis and, 231-233 facilitation of learning and, 228-229 species-specific defense responses and,

Cognitive maps, of city, 125-163 acquisition and, 138-150 free recall of landmarks and, 131 free recall of streets and, 130 matching test and, 134 questionnaires and, 134 retention and, 150-161 scoring of questionnaires and, 137-138 spatial sequence scores and, 135-137 subject recruitment and test administration and, 127-130 verbally cued test and, 131-134 visually cued recall and, 131 Cognitive operations, structure of memory and, 27-28 Cognitive skill, 8 Coherence effects with comprehension simulation model, 73-75 Comprehension simulation model, 39-80 comparisons with data, 72-73

229-23 1

stimulus factors and, 233-234 general process learning theory and, 263264

generality of molecular mechanisms of conditioning and, 265-268 levels of analysis of conditioning procedures and, 264-265 on longdelay learning, 249, 259 in taste-aversion learning, 249-253 outside the feeding system, 253-259 on positive reinforcement, 217 differential conditionability of responses and, 219-227

279

280

Index

Comprehension simulation model (cont. ) prior knowledge effects, 75-76 problems with model, 76-78 task and coherence effects, 73-75 criteria for, 39-42 description of memory search, 66-72 memory structure, 46-48 parsing processes, 48-66 issues in evaluating, 42-44 overview of, 44-46 Concurrent interference theory, 250-254 Conditioning, constraints on, see Biological contraints Constraints, see Biological constraints Consummatory stimulus reward hypothesis, 231-233 D

Danger signals, selective conditioning in rats, 247-248

E Emotion(s) modular memory and, 104-106 structure of memory and, 23-24 Entry, modular memory and concept of, 101-103 stability of, 99-101

F Forgetting, mechanisms of, modular memory and, 100-101 I

Illness, aversion conditioning with, 235-245 Imagery, structure of memory and copy images versus reconstructed images, 32-33 mental imagery in transfer of procedural memory to semantic memory, 30-31 multiple forms of representation and, 3132

Ingestive behavior, selective conditioning Of, 245-246 Input type, structure of memory and, 18 auditory nonlinguistic, 23 cognitive operations, 27-28 emotional situations, 23-24 linguistic, 24-27 motor performance, 28 plan production, 28-29 visual-spatial input, 18-21 visual-temporal input, 21-23 Integration, modular memory and, 100

L Learned safety hypothesis, 249-250 Learning, longdelay taste aversion, 249-253 outside the feeding system, 253-259 Linguistic input, structure of memory and, 24-27 Linguistic skill, rote, 4, 8-9 Long-delay learning in taste aversion, 249-253 outside the feeding system, 253-259

M Maps, see Cognitive maps Marking hypothesis, 255-256 Memory, 1-38, see ulso Recall activated, modular memory and, 95-99 episodic, 2-3, 11-13 personal memory versus, 108-1 10 phenomenal data and, 4-5 strategy and, 4 types of, 3-4 generic abstraction and, 107-108 perceptual, 7-8 semantic, 3, 7, 11, 30-31 modular. see Modular memory morphological approach to classification in ordinary language, 9 classifications by philosophers, 13-14 classifications by psychologists, 9-13 purpose of, 5-6 relation to structure of memory, 29-30 types of memory and, 6-9

Index

28 1

personal, 3, 7 P episodic memory versus, 108-1 10 procedural, transfer to semantic memory, Parsing processes, comprehension simula30-3 1 tion model and, 48-49 as record of complete experiences, 103ATN concept and, 49-50 104 description of network, 52-66 skill lexicon and, 50-5 1 cognitive, 8 variables and functions and, 51-52 motor, 8 Perceptual system, modular memory and, rote linguistic, 8-9 87-88 structure of, 14-29 Plan production, structure of memory and, comprehension simulation model and, 28-29 46-48 Potentiation, in classical conditioning, 259copy images versus reconstructed 260, 263 images and, 32-33 association with toxicosis and, 260-261 mental imagery in transfer of procenongustatory aspects of food and, 261dural memory to semantic memory 262 and, 30-31 outside the feeding system, 262-263 multiple forms of representation and, Prior knowledge effects, with comprehen31-32 sion simulation model, 75-76 relation of morphology of memory to, Problem solving, see Social science prob29-30 lem solving Memory retrieval hypothesis, 256-257 Punishment, biological constraints on, 227Memory search, comprehension simulation 228 model and, 66 activation and, 66-67 description of examination network, 69R 72 path examination and, 67-69 Recall, see also Memory Modular memory, 81-123 with comprehension simulation model, activated memory and attention and, 9577-78 99 free, cognitive map of city and, 130-131 amnesias and, 110-1 15 visually cued, 131 concept of “entry” versus concept of Reflection system, modular memory and, “the” trace and, 101-103 86-87 differentiating among subsystems and, amnesia and, 112-1 14 90-95 Reinforcement, positive emotion and, 104-106 differential conditionability of responses memories as records of complete experiand, 219-227 ences and, 103-104 “misbehavior” in depositing tokens for, other typologies of memory and, 106217-219 110 Retention, see also Memory overview of, 82-85 cognitive map of city and, 150-161 perceptual system and, 87-88 reflection system and, 86-87 sensory system and, 88-90 stability of entries and, 99-101 5 Motor performance, structure of memory and, 28 Safety signals, selective conditioning in rats, Motor skill, 8 247-248

Index

282

Sensory system, inodular memory and, 8890 Shock, aversion conditioning with, 235-245 Social science problem solving, 165-213 acquisition of skill in, 203-208 characteristics of problems general, 168-169 specific, 169-171 general considerations in evaluation, 21 1 problem representation, 209-210 problem space, 209 solution activity, 210-21 1 task environment, 208-209 information-processingmodel of evaluation of, 168 problem representation and, 167 problem solution activity and, 167-168 problem space and, 167 task environment and, 166 problem-solving reasoning model and, 171-173

protocol collection and analyses chemists’ protocols, 201-202 collection of protocols, 174

expert protocols, 175-193 graduate student protocols, 195-197 miscellaneous protocols, 202-203 nonexpert expert protocols, 197-201 novice and postnovice protocols, 193195

participants, 173-174 protocol analysis, 174-175 Species-specific responses, avoidance learning and, 229-231

T Task effects. with comprehension simulation model, 73-75

V Visual-spatial input, structure of memory and, 18-21 Visual-temporal input, structure of memory and, 21-23

Psychology of Learning and Motivation, Vol 17

Psychology of Learning and Motivation (Vol. 6)

Psychology of Learning and Motivation Vol 16