Table Of ContentLANGUAGE, MEMORY,
AND THOUGHT
THE EXPERIMENTAL PSYCHOLOGY SERIES
ArthurW.Melton . Consulting Editor
MELTON AND MARTIN • Coding Processes in Human Memory, 1972
McGUIGAN AND LUMSDEN . Contemporary Approaches to Conditioning and
Learning, 1973
ANDERSON AND BOWER' Human Associative Memory. 1973
GARNER . The Processing ofInformation and Structure, 1974
MURDOCK' Human Memory: Theory and Data, 1974
KINTSCH . TheRepresentation ofMeaning in Memory, 1974
KANTOWlTZ • Human Information Processing: Tutorials in Performance and
Cognition, 1974
LEVINE . A Cognitive Theory ofLearning: Research onHypothesis Testing,
1975
CROWDER • Principles ofLearning and Memory, 1976
ANDERSON • Language, Memory, and Thought, 1976
LANGUAGE,
MEMORY,
AND THOUGHT
John R. Anderson
YALE UNIVERSITY
LAWRENCE ERLBAUM ASSOCIATES, PUBLISHERS
1976 Hillsdale, New Jersey
Copyright© 1976by Lawrence Erlbaum Associates, Inc.
All rights reserved. No part of this book may be reproduced in
any form, by photostat, microform, retrieval system, or any other
means, without the prior written permission of the publisher.
Lawrence Erlbaum Associates, Inc., Publishers
62 Maria Drive
Hillsdale, New Jersey 07642
Libraryof Congress CataloginginPublication Data
Anderson,John Robert, 1947-
Language,memory,andthought.
(TheExperimentalpsychologyseries)
Includesbibliographicalreferencesandindexes.
1. Humaninformation processing. 2.Languages
Psychology. 3. Memory. 4. Thoughtandthinking.
BF4SS.AS2 001.53 76-21791
ISBN0-89859-107-4
Printed in the United States of America
Contents
Preface xi
1 Preliminary Considerations 1
1.1 Introductionto This Book 1
1.2 Limitationsof a CognitiveTheory 4
1.3 Goals for a CognitiveTheory 15
1.4 Strategiesfor Theoretical Advance 20
Summaryof Chapter 1 23
References 24
2 Propositional Theories of Knowledge 26
2.1 Clark's LinguisticTheory 26
2.2 Summaryof HAM 39
2.3 Kintsch's PropositionalTheory of Memory 48
2.4 The LNR Project 55
2.5 SummaryConclusionsabout Propositional
Theories 68
References 75
3 Models of Procedural Knowledge 78
3.1 The Role of ProceduralKnowledge 78
3.2 Stimulus-SamplingTheory 81
3.3 Newell's ProductionSystem 89
3.4 Conclusions about ProductionSystems 106
Summary of Chapter3 110
References 111
v
vi Contents
4 Overview of ACT....................................... 114
4.1 Predisposing Biases 114
4.2 Assumptions of the ACT Model 122
4.3 Examples of ACT Production Systems 125
4.4 The Computational Power of ACT 140
Summary of Chapter 4 144
References 145
5 ACT's Propositional Network........................... 146
5.1 Propositional Networks as Psychological
Models 146
5.2 Representational Assumptions 154
5.3 Formal Syntax of ACT's Network 173
Summary of Chapter 5 180
References 181
6 ACT's Production System 182
6.1 Principles of Operation 182
6.2 Subroutine Structuring for Productions 196
6.3 Experimental Tests of the Properties of
Productions 205
Summary of Chapter 6 218
References 218
7 Formal semantics of ACTRepresentations .............. 220
7.1 A Model-Theoretic Semantics for ACT 221
7.2 Expression of Predicate Calculus in ACT 231
7.3 Act Representations for Natural Language
Concepts 240
Summary of Chapter 7 250
References 251
8 The Activation of Memory 252
8.1 An ACT Model for Fact Retrieval 252
8.2 HAM's Serial Graph-Searching Algorithm 270
8.3 A Review of Some Fact-Retrieval Data 273
8.4 Manipulations of Probe Size 293
8.5 Generality of the ACT Activation Model 310
Summary of Chapter 8 317
References 318
9 Inferential Processes ................................... 320
9.1 Mechanical Inference Making 321
9.2 Natural Inferential Processes in ACT 335
Contents vii
9.3 Researchon SyllogisticReasoning 345
9.4 StrategicInferenceMakinginACT 355
Summaryof Chapter9 375
References 376
10 learning and Retention 378
10.1 EncodingSpecificityand Encoding
Variability 379
10.2 Depthof Processing 390
10.3 Associative versusGestaltTheoriesof
Memory 406
10.4 Effectsof LinguisticStructures 423
Summaryof Chapter10 432
References 433
11 LanguageComprehensionand Generation 438
11.1 ComputerModelsof Language
Processing 439
11.2 AugmentedTransitionNetworks 449
11.3 An ACT Production SystemModelfor Language
Processing 460
Summaryof Chapter11 485
References 486
12 Inductionof Procedures. ............. ............ ...... 490
12.1 The GeneralProblemof Induction 491
12.2 Applicationof Resultson Inductionto Language
Acquisition 506
12.3 Suggestionsfor Inductionin ACT 512
Summaryof Chapter12 527
References 528
13 ProvisionalEvaluations 530
TheDifficulties 531
PracticalApplications? 535
References 535
AuthorIndex 537
SubjectIndex 542
Preface
This book presents a theory about human cognitive functioning, a setof experi
ments testing that theory, and a review of some of the literature relevant to the
theory.The theoryisembodied inacomputersimulation modelcalled ACT. The
ACT model consists of two major components. There is anassociative network
model of long-term memory that encodes the model's propositional knowledge
about the world, and a production system thatoperates on thepropositional net
work to perform various cognitive tasks. The model is a theory of "higher
mentalprocesses." It isveryconsciously notconcerned withperception orother
nonsymbolic aspects of cognition. However, within the domain of "higher
mental processes" this is a very general theory and is proveably capable of all
specifiable behaviors. The principal empirical areas of concern are indicated by
the title of this book-there are chapters dealing with retrieval from long-term
memory, inference making, learning and retention of prose, language under
standingandgeneration, induction, andlanguageacquisition. Anattempt ismade
inthese chapters to provide ACT models for these tasks.
The theory presented in this book represents a continuation of my efforts to
understand the nature of human intelligence by building models of cognitive
processes and testing them. In 1973 Gordon Bower and I published a book
describing a model of human memory called HAM, which represented the out
come of our four years of collaboration at Stanford. Gordon Bower and I were
agreedthattheHAMmodelhadanumberofseriousdeficiencies. ACTevolvedas
an attempt to produce a model that would deal with HAM's problems. Unfor
tunately, the end of my graduate career meant anend to the close collaboration
I had with Gordon Bower. So, while he agrees as to what these problems
are, it is not so clear that he would want to endorse all the solutions adopted
in ACT.
The central problem with HAM was that, while there was a well worked out
theoryofhowmemoryoperates, therewasnowell-definedtheoryoftheprocesses
ix
x Preface
thatusedthiscomponent. Theconsequence wasthatitwasdifficult toknowhow
to apply HAM to many empirical domains and as a result difficult to perform
definitive empirical tests of the model. Our primary goal in the HAM project
wastodevelop amodelforprosememory. However, anadequate modelrequired
thatwespecifyprocesses of languagecomprehensionandinference making. The
HAMframework seemed rather bankrupt in providing us withideas about these
processes. There werealsoseriousproblems withHAMbecauseitspropositional
networklackedaformal semantics. In addition tothegeneral problems, over the
periodofthe threeyearssince theformulation ofHAM weandother researchers
have found a great many specific empirical inadequacies with that model.
There areanumberofwaystodeal withtheinadequacies found inHAM. One
is to give up the attempt to develop a large scale general theory and focus on
tractiblesubproblems. Anotheristotakeamoreartificialintelligence bent, andto
begin to develop large and complex models to deal with the richness and com
plexityof humancognition. Forreasons setforth inChapter 1,Irejected both of
these approaches and instead attempted to find a small set of basic mechanisms
that would provide adequate building blocks for human cognition. To provide
some focus to my efforts I decided to concentrate on inference making and
language acquisition. My efforts to understand inference making led me into
modem logic, mechanical theorem proving, and into research on human in
ferential processes. My attempts todeal with language acquisition led me much
moredirectlytoanewcomputermodel. ThismodelwascalledLAS(seeChapters
II and 12ofthisbook). Itusedanaugmented transition network(ATN)grammar
likethatadvocated byRonKaplan, Eric Wanner, andWilliam Woods. The LAS
model had moderate success in simulating the learning of simple languages. At
first it seemed that the ATN model had some promise as a general model for
cognitive processing, notjust language processing.
Meanwhile empirical work by myself and others on memory processes had
indicated some serious defects in the HAM conception of memory, particularly
with respect to how memory was searched. Rather than HAM's logical serial
search, the data seemed to indicate that memory was searched by means of a
parallel activation process, somewhat along the line suggested earlier by Ross
Quillian. This was the state of affairs when I first considered writing this book
in the summer of 1974. The HAM model was beginning to crumble under
empirical onslaught, andheadway wasbeing madeindeveloping anATN model
for cognitive procedures. It seemed that there was a potential for integrating an
ATNmodelwitharevisedmemorymodel. Suchamodelhadthepromiseofbeing
very powerful. The book writing enterprise seemed like the ideal means for
disciplining this theoretical endeavor.
However, as soon as I had begun to set the theory down, I encountered two
stumblingblocks. The firstoccurred whenItriedto applyanATNmodeltodeal
withthe setof ideas and data Ihad beencollectingabout inference making. The
model did not seem to have any natural application to this domain. Moreover,
theproblems uncovered intheapplication ofATNstoinferencemakingmade me