5‘:

£53.

fiAY ‘

 

 

 

WC . a

 

 

 

3.

 

 

 

 

 

 

 

 

 

ABSTRACT

A THEORETICAL MODEL OF HUMAN LANGUAGE PROCESSING

by
Jeffrey H. Katzer ‘

The purpose of this study was to develop and test a theoretical
model of continuous-free-association behavior. 'Ihe model is in the
form of an information processing model; which may be thought of as
a oonputer program. The model consists of six related hierarchical
routines. The time executive routine controls the parallel process-
ing of the other routines . Macroprocessing routine oversees the timed
routines. The stimulus sorting routine takes a coded input stimulus
word and attenpts to recogfize it in the verbal nerrory. The net sort-
ing routine cmtrols the sorting of stimults and re5ponse codes through
the binary discrimination net nemory . Finding terminal routine is call-
ed whenever an msatisfactory terminal in the menory is readied. It
atten'pts to find a satisfactory terminal. 'Ihe major routine in the nodel
is the response giving routine. Over tine it initiates associated poten-
tial respcnses to the stimults words. One at a time they are examined
to see if their item-availability is sufficient for evocation. If suf-
ficient for evocation, the potential responses may serve as internal me-
diating stilmlus words .

'Ihe mt model uses a hypothetical nenory. When presented with

a stimulus word it evokes non-trivial responses. In producing these

 

 

 

 

Jeffrey H. Katzer

responses the nodel Operates in a conplex manner. It learns over tine:
short-term-nenory and reinforcement of internal processing have a pro-
found effect cm the responses evoked. Part of the discussion is concern—
ed with the problems of net building and with obtaining measures of word

meaning from the model by a deterministic process—oriented method.

 

 

A_THBORETICAL MODEL OF HUMAN LANGUAGE PROCESSING

By
Jeffrey H. Katzer

A.THBSIS

Smeitted to
Michigan State University
in partial fUlfillment of the requirements
fer the degree of

Doctor of Philosophy

Department of communication

1970

 

 

 

Accepted by the faculty of the Department

of Communication , College of Communication Arts ,
Michigan State University, in partial fulfillment
of the requirements for the Doctor of Philoscphy

aw QM

Director of Thesis J

degree.

Guidance Committee: fluid! flélukb Chairman
IQ / aqua/I
2w. 4’ /). W

 

 

 

 

 

 

ACIQQWIL'DGH'IENTS

Supposedly a doctoral thesis is the work of one individual. How-
no one who has worked on a thesis believes that literally. The

se of this section is to identify and thank those people who were
cularly helpful to me throughout my graduate education , and who

helpful in shaping this thesis from its nebulous beginning to its

, form.

First of all I would like to acknmledge my debt to two peOple who

ed the environment necessary for this endeavor to fluorish. Dr.

. K. Berlc, Chairman of the DePartment of Communication is largely

nsible for the academic environment in which an inquiry of this

is encouraged and supported. My wife, Linda, provided the moral

rt and understanding necessary . Without her help many germinative
could not have reached fruition. Because of this and because she

erything that she is, I would like to dedicate this work to her.

I want to thank Dr. Erwin P. Bettinghaus, chairman of my guidance

ttee, and Dr. Randall Harrison, Dr. R. Vincent Farace and Dr. Richard
members of my guidance committee . Singly and together they performed
f the sundry and sometimes cnercts tasks required to get this thesis
its current fern.

In addition to the guidance committee, there is for every student an

mal group of friends and advisors who in many different ways support

ffcrts. In my case I want to thank Dr. Miles Martin, Dr. Clyde Morris

ii

i
i.
J
I

!
L.’

 

 

 

 

and David Beatty for handball , for laughter and for serious advice .

It is hard to say which was more important.
Finally, I want to thank Mrs. Jessica Gard, my typist, who single-
handedly made sense out of nonsense and order out of chaos.

 

 

 

 

 

TABLE OF CONTENTS

LIST OF TABLES ...................................................... vi
LIST OF FIGURES ..................................................... vii
INTRODUCTION ........................................................ ix
CHAPTER I ....... . ...... . ............................................ l
Orientation . . . . . ................................................ l
Mediation . . . ............... . .................................... ‘4
Association . . . . . ............................................... 7
Relationships Between Association and Mediation Approaches ...... 12
Summary . . . ...... . ............... . ........... . ................... 16
CHAPTER II ................. . . . ................. . .................... 18
Models and Simulation . . ................................. . ....... 18
Information Processing Models ...... . ......................... 20
Evaluation of IPMs ....................... . ................... 22
Construction of IPMs ...... . . . . . .............................. 29
Computer Simulation .......................................... 31
Operaticnalizing the Simulation .............................. 32
Testing the Simulation ....................................... 35
S ................................................... 38
IPMs of Verbal Behavior ......................................... 38
Hm I I I ..... I I I I OOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOO no
EPAM II ......... . ................................. . . ......... 1m
EPAM III . . . . . ................................................ 59
WEPAM ....................... . ............................... 73
SAL I-III ooooooooo c o I o ooooooooooooooo o nnnnnnnnnn o nnnnnnnnnnnn 77
Smy I I I I I I I I I IIIIIIIII I I I I I I I I I I IIIIIIIIIIII I IIIIIIIIIIIII 81
CHAPTER III. .......................... . ..... . ..................... 83
Scope of the Problem ..................... . ................... 83
Ar‘ 1m Of C-F-A BehaViOI' a a o c o u a o ooooooooooooooooooooooooooooo 87
8mm I I I I I I O I I I I C I I I I I I IIIIIIII I ....... I I O I I l I IIIIIIIIIIIII 107

iv

 

 

 

10.

12.
13.

l”.

 

 

 

LIST OF TABLES

Processes in an IPM of Simple Concept Learning ............. 23
An Outline of EPAM II Processing in Serial Anticipation . . . . 52-5u
An Example of EPAM III Paired-Associate Learning . . . . . . ..... 6H—65
Summary of Abbreviations Used to Describe the C—F—A Model . . 88
A Terminal in the Units' Part of the Discrimination Net . . . . 90
Identification of Attribute Testing Nodes in Letters '

Part OfNe‘t..... ..... .................... ............... 111+
Organization of Hypothetical Memory . . . . . . .................. 115
An Example of C-F-A Item Recognition for LEG ............... 117
Time Units Required to Recognize All Items in

Hypothetical Memory ........ ...................... 118
Value of Parameters for Simulation ......................... 118
I—AV and UT of First Items in Memory ....................... 120
An Outline of C-F—A Response Giving .......... . ............. 121—129
Responses Evokedduring Simulation.................... ..... 130
Time Units Different Numbers of Markers were Active ........ 139

vi

 

.A_._.- u» .-

 

 

10.

12.

13.
11+.
15.
16.
17.

18.

19.

 

LIST OF FIGURES

Item Selection and Learning in EPAM I ......................... |+3
EPAM II Performance Processes ................................. 1+5
Discrimination Net after the Learning of the

First TWO Items . ................ . ................ . ........ l+7
Discrimination Net of Figure 3 after the Learning of

Stimulus Item, PIB ........ . ............................... u?
Discrimination Net of Figure 4 after the Learning of

the Response Item, JUK ........... . ........................ ’49
Discrimination Net after Trial 1 ..... . ............... . ........ 55
Discrimination Net after Trial 2 .............................. 55
Discrimination Net after Trial 3 ..... . ..... . .................. 56
Changes in Discrimination Net after Trial 5 . ......... . ..... 57
Discriminatim Net after Entire List is Learned ............... 58
A Sample Discrimination Net Grown by EPAM III ................. 61
Flowchart of F20, the Paired—Associate Learning Routine
ofEPAMIII.................. ....... . ...... . .................. 66-69
Discrimination Net of Figure 11 after Trial 1 . . ............. . . 70
Discrimination Net of Figure 13 after Trial 2 . .. . ............. 71
Discrimination Net of Figure 11+ after Trial 3 ................. 72
Discrimination Net of Figure 15 after Trial l4 . .. . ............. 72
Portim of a Typical WEPAM Net ................................ 75
Syllables' Portion of Discrimination Net of Figure 17

afterlaterLearning.......... ..... ........... . ..... 76

InterrelationsAmngoutinesinC—F-A............ ........ 93

vii

 

20.
21.
22.
23.
24.
25.

26.

Flowchart of Time Executive Routine of C—F—A ................... 95

Flowchart cf Macroprocessing Routine of C— F-A .................. 96
Flowchart of Stimulus Sorting Routine of C—F—A ................. 98
Flowchart of Response Giving Routine of C—F—A . ....... 1 ......... . 101
Flowchart of Net Sorting Routine cf C-F-A ..... .. ...... . ........ lea—101+
Flowchart of Finding Terminal Routine of C—F—A ................. 105
Discrimination Net Used in Example . . . . . . . ...................... 112-113

viii

 

INTRODUCTION

approach to meaning. The theory is in the form of an information pro—
cessing model.

TWO concerns motivated the construction of this theory. In the
first place, an adequate theory of language behavior is essential to
the general understanding of an individual ' 5 communication behavior.

The model examines the relationships between a measure of the meaning
of a lexical item (e. g. word, syllable, etc.) and the generation of
similar items in an association task. It seems reasonable to assume
that an understanding of language implies an understanding of sentences;
”hi-Ch in turn implies an understanding of simpler lexical forms (q.v.
Osgood, 1963).

Meaning is typically considered to be a major variable in the study
of human communication (q.v. Mowrer, 195% Berlc, 1960), and if an ex—
tmm Stinnllus-response position is not taken, it has a similar role
in a more general study of language behavior. Osgood comments forcefully
on the inpor'tance of meaning:

22222.22 .2 mxmtm. .mmm2

édjustment is mainly a matter of acquiring and HDdlny-ng the Slgnlf‘

leance of signs and learning how to behave in ways appropriate to
these significances. (1961, p. 91)

ix

 

 

 

A second goal of this study is to evaluate information process-

ing models as models. Some researchers in the social—behavioral sci—
ences (e.g. J.G. Miller, 1963; Mackay, 1968; Miller, Galanter 8 Pribram,
1960) have argued that man can be profitably viewed as a general inform-
ation processor. This view could be adopted by more communication re—
searchers. Oertainly, none of the social—behavioral sciences deals with
phenomena more complex than those studied by commmication scholars.

Nowhere are the concepts of process and information more central
than in commmication. Information processing models are a viable alter—
native to the linear additive models so commonly used. This is especial—
ly true when the phenomenon modeled is a complex, interactive process.
For example,

It has been argued that the problem of meaning is of major

importance in the study of the nature of intelligence, and

that a useful definition of meaning must include not only

denotaticn but connotation and implication as well. To

handle these important questions it is necessary to study

cognitive organizations which are more complex than those

upon which most psychological theories are based. (Lindsay,

1963a, p. 233)

This study is organized into five chapters. In the first, an out-
line of a mediation theory of meaning is presented. Certain empirical
relationships found between measures of meaning and association behaviors
are discussed. It is these relationships that a fully developed and
fully validated model will have to duplicate, and thereby offer a suf—
ficient explanation of their causes. Chapter 2 evaluates information
processing models in terms of their potential contribution to science.
The relative advantages of these models compared with other models is
discussed. The last half of Chapter 2 presents several related inform-

ation processing models of verbal behavior. These models form the

X

 

   

framework of the theory developed in Chapter 3. Chapter 3 presents a
family of information processing models which, hOpefully, will become
a part of a general theory of individual language behavior. These
models seek to explain some of the empirical and theoretical relation-
ships found between free association behavior and several measures of
meaning. In Chapter 1+, one of the models will be examined by means of
hand simulation. That is, the model will be followed step-by—step to
see what outputs are related to what inputs . Chapter 5 evaluates the
models, explores their consequences , and points the way for further

research in the area.

 

 

 

 

CHAPTERI

This chapter presents some psychological contributions to the
definition and measurement of meaning. By focusing on psychological
investigations I do not want to imply that other studies of meaning
(notably the philosophic, linguistic and anthropologic) are of no
import . Currently , the study of free association behavior and the
operational definitions of the meaning of individual linguistic units
(e.g. words) are mainly within the domain of experimental psychology.
These are the major topics of this thesis.

The material in this chapter is organized into four major sections:
(1) an orientation to the psychological study of meaning, (2) the me-
diation approach to meaning, (3) the association approach, and (14) re-

lationships between the mediation and association approaches .

Orientation

 

Psychologists who study language are behavioral theorists -— in
Alston's sense of the word. The behavioral theory of meaning identifies
the meaning of a linquistic item "with the stimuli that evoke its utter—
ance and/or the responses that it in turn evokes" (Alston, 1960; p. 12).
Operaticnally, psychological studies of meaning seem to stem from
Bloomfield's definition of the meaning of a linguistic form: "the situ-
ation in which the speaker utters it and the re5ponse which it calls forth

in the listener" (1933, p. 139).

 

 

Psychological discussions of meaning may center on underlying

processes of meaning acquisition and comprehension and on indices
(dimensions) of meaning. A major assumption underlying these types of
psychological studies is that words are the basic units of language and
are, therefore, central to any investigation of verbal behavior. This
assumption is also trne for those studies in which, for experimental-
control reasons, non-words (also called nonsense syllables) such as
consonant-vowel—ccnsonants (e.g. XOJ ), consonant—consonant—consonants
(e.g. XRV), and disyllables (e.g. GOJEY) have been used. Those inves—
tigators who use words and those who use non—words are equally and
ultimately concerned with human processing of real languages. While
concerned with both meaning acquisition and measurement, this chapter
does not deal with original language learning (e. g. Brom's 1958 "Orig—
inal Word Game") nor the studies of developmental differences in lan-
guage behavior (e.g. Piaget, 1955, Vygotsky, 1962). The focus of this
chapter will be the theoretical and empirical relationships between two
approaches to meaning: the associative approach and the mediation ap-
proach. First, however, antecedents of these methods and these theories
must be discussed.

In terms of methods, Creelman (1966) traced the American investiga—
tions of the experimental study of meaning from the earlier work based
on classical condition to the later studies concerned with scaling, associ-
ation, and operant conditioning. The work in semantic generalization
(q.v. Razran, 1939) typifies the conditioning approach. In such studies
a word (or object) is the conditional stimulus (C81). A test is then

given to see if the conditioned response will generalize to a new stimulus

 

 

(C82) whose primary relationship with the old stirmilus is semantic

(e.g. CS1 is the word "ball" and CS2 is a ball, or vice versa). In con—
trast with these procedures, contemporary approaches to meaning are based
upon scaling and/or association techniques. This chapter is focused upon
these two methods and their relationship with each other.

In terms of theory, forerunners of current psychological positions
are the substitution theories of the early behaviorists and the diSposi-
tional view of Morris. The Watsonian behaviorists considered a linguistic
item to refer to an object (i.e. name the object) if the item elicited in
the receiver the same behaviors as the object itself elicited. For example,
the word "food" would be considered to refer to food if upon hearing the
word, the receiver salivated, chewed, digested, etc. This view is not
generally held today because "it is well known that the conditioned re—
sponse [to the lexical item] is seldom precisely the same as the uncondi—
tioned response [to the object]" (Carroll, 196”; p. 36). The trouble with
the behaviorist view is that total equivalence of reactions (to the word
and to the object) is required. It is certainly true that the receiver
may have some of the reactions to the word as he would have to the object
(e. g. a hungry person upon hearing the word "food" might start to salivate,
but probably would not start chewing). Morris (19%) tried to avoid this
problem by equating reference with an internal "disposition" on the part
of the language user to react to the lexical item as if it were the object
itself. This position has been criticized in depth by Alston (196%, pp. 28—
30) who considers it oversimplified. In their review of psycholinguistics,

Erwin—Tripp and Slobin have traced the problem of behavioral correlates

 

  

I
1
\

  

 

of meaning,

from 'conditioned response' through 'respcnse disposition' ,
'fractional anticipatory goal response' , 'representat ional
mediating response' , to the most recent candidate. Stats
and Stats ' 'conditioned sensory, motor, and autmmmic
response' (1966, p. I+50).

The two most frequently used definitions of meaning are the topics of

the next sections of this chapter.

Mediation

The mediation approach to meaning has been presented by Osgood
(1952, 1957) and Osgood Suci and Tannenbaum (1957). The meaning of
a stimulus word in this approacln is the representational mediated re-
sponses which are elicited in the person upon presentation of the stim—
ulus word. ‘ A representational mediated response is the internal stim-
ulus-response (hence mediaticnal) which is part of (hence representational)
the total response the person has toward the word's referent.

Mediated meaning acquisition , according to this view depends upon
the develOpment of mediated responses. Part of the total reaction to
a stimulus word or object is classically conditioned to the new word
developing meaning. This self—stimulating conditioned response is the
mediating respmse. Through numerous yet varied pairings of the new
word with other words or objects , a complex pattern of mediating respon-
ses will be conditioned to the new word and will, in fact, be the meaning
of the new word.

Some examples are in order. Osgood distinguished between two types
of language learning. Sign learning is a process in which the meaning of
a word is learned through repeated pairings of the word with the object it

names. Assign learning occurs whenever one learns the meaning of a word

 

 

by means of other words -- a verbal definition. Suppose one has ex—

perienced a lemon (e.g. drank lemonade squeezed lemons, etc.) but has
no name for it. Through repeated pairings of the word "lemon" with

fine object lemon (or lemonade, etc.) the mediation principle posits
that certain portions of one's reactions to the object lemon will be-
come conditioned to the word "lemnon" and will mediate between the word
"lemm" as a stimulus and the reaction to the word. This is the pro—
cess of sign learning. In assign learning, the meaning of the word "lem-
on" may be obtained by placing it in temporal, spatial, or semantic con-
tiguity with other words such as, citrus, tart, yellow, sour, etc. The
mediation approach claims that portions of the intermediate reactions
which constitute the meaning of these other words, become part of the
intermediate reactions to the neW'word, "lemon."

A criticism of mediaticns approaches to meaning comes from Fodor
(1965) who claims that the two-stage models (q.v. Osgood, 1952; Mowrer,
1951+) differ from fine Watsonian one-stage model only in terms of obser-
vability of response. In general, two-stage models posit at least one
stimulus-response sequence intervening between the overt stimulus and
fine overt response. The one-stage models of Watson and Pavlov do not
posit such intermediaries. Since the difference of observability of re-
sponse is considered insignificant by Fodor, he argues that the newer
mediation models are susceptible to fine same criticisms as the older
Pavlovian ones. Sudn a position, however, was not readily agreed upon by
the mediationists (q.v. Osgood, 1956; Berlyne, 1966) who consider Fodor's
interpretations inaccurate: a one-stage model cannot functionally sep-

arate decoding and encoding behaviors.

 

 

Osgood and his associates posit that fine meaning of a word can be

operationalized by its location in n—dimensional semantic space. Each
dimension of this space is defined by a bipolar adjectival scale pass—
ing through the origin. Consider a 2—dimensional semantic space defin-
ed by fine adjective scales sweet-sour and strong-weak. The meaning of
the word "lemon" could be quantified as a Cartesian point in the plane
defined by finese scales. Presumably, such a point would be more toward
the sometrong quarter of the plane finan fine sweet-weak quarter. The
method described by Osgood to locate a word in semantic space is by means
of a semantic differential. A semantic differential is a paper and pen—
cil instrnment consisting of a set of bipolar adjective scales on which a
person rates a word or a concept. The distance between fine ends of each
scale is broken into (usually seven) supposedly equal intervals. The
rater indicates which inter-val reflects his reaction to fine word or con-
cept. A typical analysis of this data entails the computation of a cor—
relation matrix between scales. This matrix is factor analyzed. The re—
sulting factors form the dimensions of semantic space. In finis manner,
ratings on a semantic differential are convertible to locations in semantic
space and, finerefore, constitute fine meaning of the word or concept rated.
There is an assumed relationship between the mediation theory of meaning
acquisition and the semantic differential.

Corresponding to each major dimension of the semantic space,

defined by a pair of polar terms, is a pair of reciprocally

antagonistic mediating reactions, which we may_symbolize as

i‘é'loaiédalfleé‘i’ifhifiiitfwEarmai‘ii‘v‘eififthe

judgnent by fine subject using the semantic differential, in

whidn a sign is allocated to one or fine other direction of

a scale, corresponds to_fine acquired capacity of that sign

to elicit either rm or rm, and the extremeness of fine sub-

ject's judgment corresponds to fine intensity of reaction

associating the sign with either rIn or rm. (Osgood, Suci
and Tannenbaum, 1957, p. 27)

 

 

One frequent criticism of fine semantic differential concerns the

approPriateness of calling fine measurement "meaning". This criticism
is supported by two types of arguments. The intuitive argument claims
that what a person means by "lemon" is more than a coordinate position
in a hypothetical space -- there is more to the meaning of lemon than
can be shown with adjectives. The second argument stems from the mea—
surement of fine relationship between the meanings of words in semantic
space: words lying far apart in semantic Space are less related than
those close together. If two words lie in the same position of seman-
tic space (within the limits of measurement error) then one would have
to conclude that the two words have the same meaning. However, few
peOple would be willing to say that "nurse" and "success" mean the same
thing even though they occupy the same position in semantic space.
Criticisms similar to these have led to a re—interpretation of what is
being measured with a semantic differential. The current position is
finat connotative meaning or affective reactions is being measured. That
is, no claim is made that "nurse" and "success" refer to the same object
(same denotative meaning). Rather, bofin words name concepts which peo-

ple react to similarly (same connotative meaning) .

Association

 

A second way to look at meaning from a psychological point of view
is fine association approach. This is based upon the reaction of an indi-
vidual to a word. TWO words, for example, may be said to have the same
meaning if finey evoke the sane total reaction pattern within the indivi-
dual. Since fine associationists of interest here study verbal behavior,
they limit themselves to intraverbal meaning —- the verbal reactions to

a word.

 

 

Noble (1952) defines meaning in a Hullian framework as the several

habit (tendency for a stimulus to evoke a particular reSponse) strengths
between the stimulus word and the class of corresponding conditioned
verbal responses. Deese (1962, 1965) and Garskof and Houston (1963) com—
pare meanings of stimulus words by comparing the patterns of free associ-
ates elicited by each. The totality of free associates elicited is , ac—
cording to Deese, a sample of the intra-verbal meaning of a word.

One difficulty encountered in defining meaning as re8ponse of a

hearer (or speaker) is that any particular linguistic form, at

various times, elicits a variety of responses in the same per-

son. Therefore, the meaning of any form is not given by single

response, or, indeed, by a collection of responses at some par-

ticular time, but by the potential distribution of responses 3:2
that form. (1965, p. I+1). _

 

Meaning acquisition, in an association framework, depends upon the
establishment and strengthening of the links between the stimulus word
and its verbal re$ponses. This procedure has been typically explained
in terms of the laws of association; the most important of these being
ccntiguity and frequency. The more often two words, or a word and an
object, are perceived together (Spatially or temporally) the stronger
will be the link between them.

In terms of method, the association paradigm asks a subject to
respond to a stimulus word with another word or words. There are four
major types of association tasks: (1) in a discrete-free association
task the subject responds with the first word that "pops into his mind";
(2) in continuous-free association the subject is asked to respond with
associates until either a desired number of associates have been pro—
duced or until some fixed time limit has expired; (3) a discrete—con-

trolled association task asks the subject for one response , but that

 

 

response must be in some pre-defined category (e. g. respond with the

opposite of the stimulus word); and (H) a continuous—controlled associ—
ation task is similar to a discrete—controlled task except more than one
associate to the stimulus word is required in the former case.

A distinction ought to be made between these approaches to meaning
and the more familiar ones which use association values. Association
values are numbers assigned to stimulus items (e.g. words, nonsense syl-
lables) which reflect how many different responses the stimulus word has
elicited in a group of subjects participating in one of the four types of
association paradigrs (q.v. Woodworth and Schlosberg, 19 5H; Underwood and
Schulz, 1960). The higher the association value, the more responses elic—
ited. With association values, a comparison between stimulus words is
made in terms of the size or strength of the association elicited. In one
situation comparisons are made in terms of similar specific responses elic—
ited by each stimulus word. In the other situation comparisons are made
in terms of the numeric association values. The former is a comparison of
meaning while the latter is a comparison of meaningfulness.

To lay a proper foundation for the model of continuous free associ-
ation behavior presented in chapter 3, it is necessary to examine some re-
lationships central to the study of verbal learning and behavior. As noted
above, association strength is a construct which accounts for observed dif-
ferences in the strength of the stimulus (S) —- response (R) bond. Response
strength is typically measured by reaction time and/or response frequency
or communality (q.v. Vbodcmrth 8 Schlosberg, 1951+). That is, in an associ—
ation task, those responses linked to the stimulus word more strongly will
be emitted more quickly and more frequently (when a discrete free associ—

ation task is administered to the same subject with the same stimulus word

 

 

10

several times). The ingredients which produce or affect associative
strength are the subject of some disagreement-—depending, in the main,
upon the theoretical position one takes. The frequency of the S—R pair-
ing, the recency of the pairing, the closeness of the stimulus and re-
sponse objects, and the type and schedule of the pairing reinforcement
are put forth by different investigators as key ingredients of associate
strength (q.v. McGeoch 8 Irion, 1952).

The nature of an association task (but not the nature of association,
Er: _s_e_) implies directionality. The stimulus is linked to the response
because the S elicits the R or because the S comes before the R. This
suggests that forward association (S—R) is the normal state of affairs
and backward association (R-S) is an unusual state which must be dis—
counted if the notion of directionality is to be maintained. Backward
associations have been shown to exist (e.g. Murdock, 1958) and a great
deal of energy has been devoted to "explaining away" the phenomenon,
though no one has done so to everyone's satisfaction.

A different approach was taken by Asch and Ebenholtz who report a
series of studies which support the principle of associative symmetry:
"when an association is formed between two distinct terms, a and b, it
is established simultaneously and with equal strength between b and a
[italics omittedJ" (1967, p. L#81). Their studies strongly indicate that
backward associations are typically weaker than the corresponding forward
associations because of an experimental artifact: in learning S—R pairs,
the subject experiences (evokes, pronounces) the R member of the pair more
so than the 3 member. This uneven experience makes the R member more
available than the S member. When both members of the pair are made equal-

ly avaliable to the subject as a possible response, the strength of the

  

_l
I
lI
1

  

 

 

ll

S-R and R—S associations are very nearly equal (q.v. Asch 8 Ebenholtz,
1967; Horowitz, Brown, 8 Weissbluth, 1961+; Horowitz, Norman, 8 Day, 1966).

Item availability (I-AV) and response strength are related concepts.
They are not equivalent, however, because response strength reflects a
long term, more stable, relationship between verbal units while I-AV can
be changed much more easily [see below].

Underwood and Schulz (1960) present a two stage analysis of verbal
learning: the response learning stage and the association stage. In
the first stage a response is learned by integrating it into a whole unit
(e. g. treating a word as a word rather than a collection of letters) and
by making the response avaliable. (Tip of the tongue phenomenon mnight be
considered as an example of integrated, but not available verbal Lnnits) .
In the associative phase, the integrated, available response is paired with
a stimulus item.

In summarizing their research, Underwood and Schulz proposed the
"spew hypothesis" which states that, "the order of emission of verbal
units [in a continuous free association task] is directly related to fre-
quency of experience with those units" (1960, p. 86). They reason that
more frequently experienced items will be more available and, therefore ,
will start entering into an association before less frequently experienced
items. While there is snpport for the spew hypothesis from other investi—
gators (e.g. Noble, 1963; Osgood 6 Anderson, 1957; Jakobovits, 1966),
other studies show that frequency alone is not a sufficient determinant
of I-AV. Woodmr'th and Schlosberg (195”), Horowitz and his associates
(1964, 1966), and Asch and Ebenholtz (1967) indicate that recency of ex-
perience and mode of exPerience (e.g. does subject produce the item from

memory or read it) are also major components of I-AV.

 

 

 

 

 

.’

 

a:

i

I‘:-

v
F

i,

 

é“.

.- ,

\\

.

K 1

- "v
a"

 

 

 

A: ‘
l‘ A ‘
x H‘
\o ‘
‘A
.\.~ A
\,‘ \
V‘

   

 

 

12

I-AV is an important variable in a theory of association behavior.
A researcher can only study behavior's of the subject. In a free associ-
ation task this behavior is mainly the associates given in response to a
stimulus. In studies of verbal learning, the verbal units are often un-
known or unfamiliar (especially when the units are not words but are non-
sense syllables, or strings of numbers, etc. ). In these studies the sub—
ject must go through both parts of the response learning phase -- integrat—
ing the unit and making it available -- before an association can be given.
However, in free recall or association tasks the subject produces responses
from memory which must already be integrated. Therefore, in a free associ-
ation task the role of I-AV is more directly related to overt subject be-
havior, than in studies of verbal learning, and I-AV is more directly a
determinant of the recall of verbal units than is associative strength
(q.v. Asch 8 Lindner, 1963).

Relationships Between Association and Mediation Approaches

The difference between the association and mediation approaches to
meaning is not as great as might be inferred from the preceding paragraphs.
Classical conditioning underlies both. The relationship between mediated
meaning acquisition and classical conditioning was shown in an interesting
study by Stats and Stats (1957) . Subjects were slum a nonsense syllable
paired with several different words. The words, chosen from the Semantic
Atlas (Jenkins, Russell 8 Suci, 1958), were very similar in their affective
meaning components. Semantic differential ratings of the nonsense sylla-
bles after the pairings showed a shift in the affective meaning of the non-
sense syllable toward that of the words . Additional support of the role
of classical conditioning in meaning acquisition was found by Pollio (1963)
and Stats and Stats (1958).

 

I I

_ Fi-

       

13

= Because of the apparent haphazard nature of contiguity (i.e. any
typical or atypical word-word, word-object, or object—object pairing
strengthens the association bond), and because certain responses to
stimuli could not be adequately described by the association laws , there
has been a strong interest in mediational interpretations of these phe-
nomena (q.v. Cofer 8 Foley, 1992; Jenkins, 1963). These writers suggest

’ that free associates are determnined not only by contiguity, frequency
and the other laws of association, but also by various mediation paradigns.
For example, "dark" might be an associate of "heavy" because of the medi—

? ating response, "light". That is, "dark" can be thought of as being an
associate of "light", and "heavy" can also be considered related to
"light". Thus, in a free-association task the stimulus word "heavy"

' mnight elicit the response "dark" because of the previously formed rela-
tionship, heavy—light-dark. This type of mediation paradigm might help
explain certain oddities in free-association behavior. It is known
(q.v. McNeill, 1966), for example, that adults frequently give opposites
of the stimulus word in free-association tasks . Opposites , however,
occur less frequently together than other types of word pairs. In gram-

1 matical English sentences, "good" would be more frequently paired with a
noun (e.g. boy) than with its opposite, "bad". The fact that "g "
strongly elicits "bad" as an associate indicates that the simple laws
of association are not sufficient as they are based on frequent pairings
of words. Mediation has been proposed to explain the elicitation of
opposites (q.v. Ervin, 1961; Jenkins, 1963). In fact, the notion of

mediated contiguity makes it possible to abandon the more restricted

 

.“ .'

 

1'4

concept of primary stimulus generalization (q.v. Deese, 1965; Cofer 8
Foley, 190.2) and adopt the more general principle of mediated stimulus
generalization.

In practice, Osgood, Suci and Tannenbaum (1957, p. 20) consider
the semantic differential related to a controlled association task.
Bousfield (1961) views the semantic differential as a controlled asso-
ciation task in which the subject chooses appropriate adjectives rather
than emnitting free responses. Deese (1965) argues that the semantic
differential ratings are derivable from associational structures. Stats
and Stats (1959) state that the same operation of word-word pairings
strengthens the interword association and distance from the origin of
semantic space —- a mediation measure of meaningfulness related to associ-
ation values. Pollio concludes a series of enperiments dealing with both
association and mediation responses to a stimulus word by taking,

the position that both classes of events imply, or at least

suggest, certain relations among words and that these re—

lations can be described by a single structural conceptual-

ization encompassing both classes of events (1966, p. 11).

Empirical relationships have been reported between the two approach-
es to meaning. Stats and Stats (1959) had subjects rate 10 words on a
good—bad semantic differential scale and later rate the first 20 asso—
ciates of each of these 10 words. Averaging over subjects and associates
they found a rank order correlation of +.90 between the ratings of the ten
words and the average of their first 20 associates. Jenkins and Russell
(1956) report a correlation of +.7l between an association measure of
meaningfulness and distance from the origin of semantic space. Wimer
(1963) and Howe (1965) obtained correlations (r = +.36, +.Sl) between the

same two measures.

 

 

15

There is research reported which relates the spew hypothesis to
measures derived from a mediation approach to meaning. For children,
Pollio (1969) found the correlation between a word's location in seman—
tic space and the location of its first free associate. The correlation
was significant (p. < .01) separately for each dimension in 3—dimensional
semantic space (r: +.6|+, +.69, and +.lH4 for the evaluative, potency and
activity dimensions respectively).

For adults, similar results were found except the correlation be—
tween the potency scores did not reach as high a level of significance.
According to the studies reported above (q.v. Stats and Stats, 1959;
Pollio, 1964) associates of a word ought to lie near that word in semantic
space. One would expect frequent word—word pairings to have more of an
effect (in terms of acquiring detachable portions of responses) than in—
frequent ones. Therefore, we would expect first associates to be closer
in semantic space to a stimulus word than later associates. DeBurger and
Donahoe (1965) found that succeeding associates are less similar in mean—
ing (i.e. farther away in semantic space) to the stimulus word. In a re-
lated study, Portncy (1961) reported that reinforcing the first associate
of a word had greater effect of the word's evaluative meaning than rein—
forcing the third associate of the word. In continuous free—association
behavior, Pollio (1966) found that responses given in rapid succession
to each other formed a cluster whose average distance between them in
semantic space was less than the distance between responses which were not
temporally clustered by the respondant.

These studies in general SLpport the theoretical position noted at
the beginning of this section; viz: that several of the association mea-

sures and mediation measures are related.

 

 

 

16

slum

Psychological theories of meaning have evolved from the earlier
mentalistic approaches and the strict behaviorism of Watson and his
followers to the more liberal behavioral approaches today. These ap—
proaches typically consider meaning to be related to processes which
occur within a person. In the main, these processes are thought of
as being habit, bonds, some form of mediated response, or some com—
bination of these.

Both mediation and association approaches to meaning, including
the theory underlying each and their methods of measurement, are sub—
ject to some criticism. This does not vitiate their importance to cur»—
rent thoughts in the psychology of language. They are , by far, the ma—
jor theories Lnnderlying most of the thinking and research in this area.
This pervasiveness outweighs the criticism in terms of their importance
to this study.

The research findings presented do not exhaustively survey the rel—
evant literature. Such a task would be larger than the sc0pe of this
thesis. Rather, an attempt was made to indicate those variables and re—
lationships relevant to continuous free-association behavior which will
be major considerations in the model presented in chapter 3.

This is an appropriate place to restate the goal of this study. Sim-
ply stated, it is to specify a model of verbal behavior which ultimately
will identify the theoretical relationships between association and medi-
ation principles of meaning. Also, such a model should predict empirical

relationship between both measures of meaning.

 

   

17

There are several ways to organize this effort, and at this time
it is impossible to forsee which will ultimately have the greatest pay-
off. The strategy here, is to generate a model based upon association
principles -- Specifically a model of continuous free association behav-
ior. A major conclusion of the next chapter is that information process—
ing models are very useful in the behavioral sciences. This type of model
clearly specifies procedures which hopefully will produce relationships of
interest among the variables. Thus, if a model is to exhibit relationships
between free association structures and mediation measures of meaning (as
in the above studies) then the model must account for the generation of
free associates. The model presented in this study will be a first ap-
proximation to this goal.

Before the variables described in this chapter can be organized with—
in a model of individual continuous free association behavior it will be
rnecessary to discuss information processing models, their construction,
their relative merits, and their relationship to computer simulation of

cognitive processes. Such are the topics of chapter 2.

 

 

CHAPTER II

This chapter examines the primary method of inquiry to be used:
Information Processing Models (IPMS). There are two major divisions to
this examination. First, types and roles of models will be discussed -—
leading to a general presentation and evaluation of IPMS and their re-
lationship to computer simulation of cognitive processes. Next, several
examples of IPMS will be presented. These are Simulations of verbal be-
havior or language processing. The implications of this method of inquiry
and of these examples will be discussed vis-a-vis the subject matter of

this study.

Models and Simulation

 

Confounding any discussion of models in scientific inquiry are the
numerous philosophic and psychological distinctions between models and
theories, between various types of models, and between judgnents of the
relative value of the different kinds of models. Models have been dis—
tinguished fromn theories by separating the structure from the content of
the plnenomenon of interest (q.v. Kaplan, 196”, pp. 269 — 265; Rudner, 1966,
p. 29). Rather than unduely magnify the importance of this distinction to
this discussion, the position here is the same as that taken by Newell and

Simon:

18

 

 

19

. we shall use the terms 'model' and 'theory' sub-
stantially as synonyms. The term 'model' tends to be ap—

plied to those theories that are relatively detailed and

that permit prediction of the behavior of the system

througn time, but the line between theories that are called

'models' and other theories is too vague to be of mnuch use.

(1963a, p. 365)

While there are other uses for models in science (e.g. the null
hypothesis model is used as a straw man for comparative purposes, or mod-
els used for control purposes or approximations-w (q.v. Ackoff, 1967) the
point-of-view taken here is that models have a value directly related to
their heuristic role or deductive fertility.

Why should a scientist ever concern himself with a model?

In one rather obvious sense, the point of employing a model

belongs to the context of discovery rather than to that of

validation; for models function as heuristic devices in

science. (Rudner, 1966, p. 25)

Models have been classified in various ways (e. g. Ackoff, 1967,

p. 101+; Tatsuoka, 1968; Kaplan, 1969, pp. 273-275). For purposes of
discussion the classification scheme of Springer, Herlihy and Beggs
(1965) will be adepted. They classify models into one of three general
categories: abstract models, symbolic models, and physical models. Ab-
Stract models are mental images (q.v. Boulding, 1956) of reality. Sym-
bolic models are either verbal or mathematical. And, physical models are
iconic (physically isomorphic) or analogic (functionally isomorphic). Of
these, the model builders in the social sciences are symbolic models most
frequently. This may be due in part to custom (most models a theoretician
has eXperienced are symbolic), or practical considerations (physical mod-
els -- if applicable - are difficult to construct), or esthetic evalua-
tions (mental models are not rigorous enough). Of the two types of
symbolic models, the verbal are more cannon while the mathematical are
more in vogue (due to the difference in perceived rigor and the affinity

of some researchers to be "scientific").

 

 

20

Information ProcessinLMode ls:

The task of this thesis is to construct and evaluate a model.

This model is symbolic in format but is neither verbal nor mathematical
in the common uses of these terms. It will be an Information Processing
Model, an IPM.

Evaluating the information processing approach in psychology
Reitman describes it as,

one way of looking at psychological activity. It deals with

processes and functions; it emphasizes whatever it is that any

particular behaviors ge_t_ done; it is also concerned with the

fine structure of behavnor. The accomplishments resulting

from thinking, problem solving, and psychological activity

generally can be accounted for only if we study them in great

detail. When we do so, we discover that even simple behaviors
appear to be made Lp of a great many steps integrated into

complex sequences . . . . In other words, this approach allows

us to view man as dynamic systems analyzing, seeking, and

doing things, as purposive organisms manipulating objects

and information to achieve ends. (1969, p. 1193)

The information processing approach is applicable to content areas
other than psychology. In fact, its generality makes it applicable to
non-human systems (e.g. communication networks within a formal organ—
ization, and processing within a general purpose digital computer).

Hart (1967) presents one way to specify the essentials of the inform—
ation processing approach. Models employing this approach are characterL
ized by their components, structure, and primitive processes. There are
five basic types of components: (1) a set of containers or storage loca—
tions; (2) a set of possible contents of the containers -- where the con-
tents can be (or stand for) a word, number, person, nation, process, etc.;
(3) a set of links which connect the containers; (H) a set of labels which
nane the containers and links; and if the model is empirical, (5) a pro—

perty set may be attached to any of the containers or links. One form of

 

 

 

I.
‘ . n.-
_ ,h . .
‘_ 'l

,

..'

fi 0

. . g-»

«Hos‘hfi' '

  

{I ‘

21

a property set is a set of ordered pairs. The first member of the pair
Specifies the dimension of the container or link (e.g. color) while the
second member of the pair gives the valLe of that dimension (e. g. red).

Structure of IPMS depends upon the organization of containers (re—
gardless of content) and links. Links may be uni- or bi—directional.
Usually not all containers will be linked with each other and the dif—
ferent resulting organizations (e.g. rings, linear) structnme IPMS.

IPMS can easily represent hierarchical systems. If a group of contain—
ers and links are grouped together under one name, then that name labels
the contents of a hierarchical container. Hierarchies (level n+1) of
subsystems (level n) can be created. Property sets, links and structure
among hierarchical containers can be specified. The importance of hierL
archical systems should not be mninimnized -- especially when dealing with
complex phenorena (q.v. Simon, 1965)

For complex phenomena there may be, and usually are, several

levels of explanation; we do not explain the phenomena at

once in terms of the Simplest mechanisms, but reduce them to

these simplest mechanisms through several stages of explana-

tion. We explain digestion by reducing it to chemical events;

we explain chemical reactions in terms of atomic processes;

we explain the atomic processes in terms of the interactions

of subatomic particles. Every flea has its little fleas, and

the scientist's view accepts no level of explanation as 'ulti-

mate.‘ (Newell 8 Simon, 1961, pp. 155-156)

Primitive processes in IPMS function on both hierarchical and non—
hierarchical (atomic) levels. These processes can affect the structure
or the state of the system. Structural processes can add or delete con-
tainers, links, hierarchical conponents, and change the directionality
of links. State processes may modify the contents of containers , nanes
of containers or links and the elements of property sets. Processes

may be stated in conditional form. This plus the fact that the contents

 

 

 

22

of a container may name a process to be executed, makes IPMS very pow-
erful. One example of the power of IPMS is their ability, in theory,
to calculate anything computable (q.v. Davis, 1958).

An example of an IPM would be helpful. The example is taken from
an article by Gregg and Simon (1967) which will be discussed more fully
later. The model was designed to represent the behavior of a subject
(_S_) in a simple concept learning task. The concept to be learned is
chosen by the experimenter (E) in advance and can be any one of the 2N
possible concepts (where N is the number of dimensions -- each dimension
has two values). In the experimental procedure E presents S with a series
of stimulus instances. A stimulus instance contains a sarple of the 2N
possible concepts (e.g. if the dimensions were size, number, color, and
shape then a stimulus instance might be five large red circles). The E
responds to the instance by stating whether or not it contains an example
of the concept chosen by E, but unknown by _S_. E appropriately reinforces
_S_'s response. A concept is learned when E makes a predetermined number of
correct responses in a row.

As presented in Table 1 the IPM consists of seven processes. In
terms of the description of general IPMS, this model can be considered as
composed of seven hierarchical containers, the contents of each represents
a set of processes. The use of conditional processes, the linkage struc-
ture among the processes and the possible use of property sets (e.g. the
number of correct learning trials may be kept in a property set) should

be noted.

Evaluatiog of IPMs:

It is important to evaluation IPMS vis-a—vis the other symbolic

 

 

Table 1.

   

23

Processes in an IPM of Simple Concept Learning

 

 

Name of
Process

Process

 

EO

Do 1:3, £9, 81, 132.

If reinforcement = "right" then increase the number of
correct learning trials in a row by 1. Call this
number "tally."

If reinforcement = "wrong" then set tally equal to 0.

If tally equals the preset criterion defining the attain—
ment of the concept , halt.

If tally is less than the criterion do 82 then E0.

 

31

If the E's current hypothesis of the correct hypothesis
is a member of the stimulus instance respond "positive";
otherwise re3pond "negative."

 

E2

Compare the _S_'s response with the correst response. If
the E's response is correct, reinforce "right"; other-
wise reinforce "wrong."

 

82

If reinforcement was "wrong" adopt a new hypothesis from SS.

 

E3

Generate a stimulus instance by sampling randomly from each
pair of the N dimensions .

 

39

If the concept adopted by E is present in the stimulus
instance then the correct response the _S_ can give is
"positive"; otherwise the correct response the S can
give is "negative."

 

85

 

Generate a new hypothesis of the correct concept by sam—
pling at random from the list of 2N possible hypotheses.

 

Note. -- Adapted from Greg 8 Simon (1967, p. 253-259).

 

 

2”

models to see if the level of explanation and insight afforded by the
former compare favorably with those afforded by the latter.

To make comparisons involves the application of criteria. As noted
before, the term "model" is used in a manner simnilar to the term "theory."
The proper criteria to be used to evaluate theories are a major topic in
the philosophy of science. The ones adapted here are falsifiability, use—
fulness, precision, and parsimony. The first is the sine qua non of the-
ories according to Popper (1961) —— theories must, in principle, be capa-
ble of being proved false. The second is important because a major pur—
pose in the construction of IPMS is the heuristic role of the model --
criteria used to evaluate a model should not be determined without a con-
cern for the purpose of the model. Another aspect of usefulness is ap-
plicability in terms of the model's practicality and generality. The
third criterion, precision, also has two aspects. A theory is precise if
it is Stated clearly and rigorously, and a theory's precision is inversely
related to the size of its error of prediction. The last criterion, par—
simony, is adopted because of the esthetic value placed on explaining more
and more with less and less. Parsimony is related to falsifiability. The
less parsimonious a theory the more difficult it is to be falsified (e.g.
if fine number of degrees of freedom in a fineory equals or is greater than
fine number of empirical observations, the theory is not falsifiable because
fine paraneters will cover all instances of possible observations).

It is important to compare the three different types of symbolic mod-
els. Comparing IPMS to verbal models Kaplan believes that a generalized
form of IPM is, "far more effective than philosophical dialectics in free—

ing behavioral science from fine stultifications of both mechanistic

 

 

25

materialismnand mentalistic idealismfl (1964, p. 292). In Choosing be-
tween mafinematical models and verbal models there is a general prefer-
ence for fine former because of its increased clarity, rigor and deduct-
ive fertility. To these Arrow adds the greaternpossibility of mathe—
1matical models, "to tap the great resources of modern theoretical stat-
istics as an aid in empirical verification" (1956, p. 31).

The evaluation of mathematical models versus IPMS is most signif-
icant because of the generally held belief that mathematical models are
to be preferred.over general verbal models. The relationship of mathe-
matical models to IPMS is one of inclusion. Mathematical models can
be considered as special cases of IPMS, and for that reason IPMS are
more general. Mathematical models rarely deal with explicit processes
and therefore are less valuable to the researdher~who is interested in
processes, Er se. In fine concept learning model presented above, the
processes were clearly stated and hypothesize how a person learns a con-
cept. An analagous mathematical model predicted not processes but em—
pirical measures of concept learning behavior (e.g. number of errors be-
fore fine concept is learned). In addition, mathematical models are
limited.by the complexity of the phenomenon of interest.

If the mathematics is known to the model builder or can be

discovered by him, he will be able to determine the implications

of his model. If the mathematical techniques for solving cer-
tain equation systems are not known or available to the model

builderg he is in no better'a.position than if he had only a

natural language model. The effect of finis last condition is

to constrain fine model builder to consider only that class of

models for which he knows solutions are available. Unfortun-

ately this constraint may have a spurious effect on the model
builder: e.g., he may oversimplify a complex situation. In
general, many of fine mathematical models of human behavior

are elegant and simple. Sometimes, the constraints of the

mathematical medium force unfortunate compromises upon fine

model and reduce its ability to predict. (Feigenbaum 8
Feldman, 1963, p. 271)

 

 

 

v3, ‘- ‘r-‘v

 

.CWS’TT-gs-m .. _

 

 

 

 

26

For example, mathematical techniques for dealing with non—linear sys—
tems have not been highly developed (or are unlcnown by many social sci—
entists). Many of fine phenomena of interest to social scientists will
probably not be explained best by linear descriptions (q.v. Lindsay,
1963a). On the other hand fine effect of complexity and non—linearity
upon IPMS is not thought of as significant (q.v. Feigenbaum 8 Feldman,
1963, p. 271).
It would be helpful to compare a mathematical model with an IPM.
Gregg and Simon (1967) made such a comparison between their IPM of sim-
ple concept learning and Bower and Trabasso's (1961+) mathematical model
of fine same phenonenon.
Gregg and Simon's IPM of simple concept learning is outlined in
Table 1. Bower and Trabasso's model consists of fine following two state-
ments and the analytic deductions from these statements.
1. On each trial the subject is in one of two states, K or K.
If he is in state K (he 'knows' fine correct concept), he
will always make the correct response. If he is in state
K (he 'does not know' the correct concept), he will make an
incorrect response with probability p.

2. After each correct response, fine subject remains in__his pre—
v10us state. After an error, he shifts from state K to state
K wifin probability II. (Gregg 8 Simon, 1967, p. 21+?)

Several criteria were applied in making fine comparison between fine
two models. The most relevant of finese are generality, rigor, parsimony,
usefulness, and validation procedure.

In terms of generality, Gregg and Simon compellingly argue that IPMS
are more general. Starting with a Bayesian position, finey Show that the
a Ester-{Lori (i.e. after the evidence is in) credibility of a fineory (or

model, or hypofinesis) is a joint fnmnction of the likelihood or accuracy

 

 

s

mlq

 

. .

 

 

 

27

of the fineory and fine a priori plausibility of fine theory. That is, more
believable theories depend not only on the accuracy of their predictions,
but also upon the perceived reasonableness of the theory before testing.
If this position is accepted, then Gregg and Simon might say that the
statement of reasonableness of the theory can often take the form of an
IPM. It was from fine reasonableness argument that Bower and T‘rabasso
developed fineir mafinematical model. Since the mafinematical model was de-
veloped from the crude IPM (the reasonableness argument) it can be con-
sidered a special case of fine IPM. In fact, Gregg and Simon Show that the
mafinematical model is a special case of a family of related IPMS. Each
member of the family is different from each ofiner and the difference may
be of theoretical import to concept learning tasks (for example, a dif-
ferent IPM would change process SS in Table l to allow for sampling of
only those hypotheses still supportable by the current stimulus instance).
However, the same mathematical model is derivable from each of fine related
IPMS. Therefore, fine IPMS are more general.

In terms of rigor, IPMS can be stated as rigorously as desired. It
should be remembered, however, finat neither mathematical models nor IPMS
are as rigorous as cormonly believed.

The guarantees of unarbiguity are usually overrated both for
mathematics and for programs [an operationalized IPM]. The suc—
cessive waves of rigorization that have swept through fine mathe-
matical world testify that what is unambiguous in one generation
is not in the rnext. Similarly, the fact that most programs never
are fully debugged indicates a similar failing in programs.

(Newell 6 Simon, 1963a, p. 379)

In terms of parsimony, the mathematical model has two free parameters
( H , 2) while the IPMS have none. Thus, the IPMS are easier to falsify.

And in terms of usefulness, fine following finree points are noted in favor

 

r‘r—I “‘WA'F -. '

a
v.

"I

.W

fine ..
\
V“ ~

 

28

of the IPM: (1) having a family of IPMS for each mathematical model
implies that rejection of an IPM is not fatal; (2) the IPM separates
subject processes (those named with an "S" in Table 1) from experimenter
processes (those naned with an "E"). This separation allows fine research—
er to test the effects of fine subject's behavior in different experimental
situations. As stated, the mathematical model cannot make such investi—
gations; (3) the IPMS generate more useful data (e. g. fine model can trace
and "repo " the subject's actual hypotheses and responses —— which can
later be compared with those of real subjects).

And finally, fine validation of fine mathematical model involves
"proving the null hypothesis" (i.e. finere is no difference between the
model's performance and the performance of human SS). Since this is not
considered to be statistically permissible Bower and T‘rabasso place,

their main reliance on finding 'critical' experiments that sep-

arate alternative hypotheses radically. But . . . the variant

predictions in fine critical experiments come . . . not from fine

stodnastic theory but from the informal, and only partially

stated, process models finat stand behind the theory. (Gregg

8 Simon, 1967, p. 270)

Therefore, in terms of these criteria the IPM is to be preferred.
Gregg and Simncnn also corpare the two types of models empirically (in terms
of prediction) and statistically (in terms of error variance). Again their
conclusion favors fine IPM. In fineir article the choice of the specific
models used mignt have unfairly stressed the value of IPMS. With different
models, sore of their arguments might not have been appropriate or as tel-
ling. However, Lindsay (1963b) and Abelson (1961+) arrive at similar con—
clusions with different models.

Certainly, IPMS cannot be perfect. Wlnat then are some disadvantages?

Newell and Simnon (1963a) identify a major disadvantage of these models,

 

29

yi_z. , the absence of a deductive formal system for making inferences from
the model. A second consideration has to do with the amount of time need—
ed to rigorously state an IPM. Usually such rigorous statements take the
form of a computer program. This implies that the model builder must take
fine time to learn the programming language and an inordinate amount of
time to debug the program. These, disadvantages, however, do not outweigh
fine benefits of IPMS.

Reviewing this comparison of models it is argued that IPMS are to be
preferred over ofiner symbolic models, especially in those situations in
which (1) fine actual behaviors are important to understand, and (2) a

major value placed on fine model is its heuristic insight.

Construction of IPMS:

 

Before concluding fine presentation of general IPMS it seems appro—
priate to discuss some factors related to fineir construction. Building
IPMS, like other models, depends upon the definition of the task, delimit-
ation of fine system's boundaries, adoption of the level of analysis, iden—
tification of the processes and relationships, etc. These considerations
have been discussed elsewhere (e.g. Ackoff, 1967) and will not be present—
ed here. The purpose of the next several paragraphs is to identify some
more specific problems concerning IPM construction.

Carroll and Farace (1968) make an interesting distinction between
theory—rich and data-rich models (with what finey call heuristic models
lying between these extremes). Theory—rich models are constructed by rep-
resenting fineoretical relationships within the model, while data-rich mod—
els use information obtained empirically (e.g. in giving values to para-

meters). While IPMS may be theory-rich or heuristic, they usually are not

 

 

n.

 

30

data—rich (with emphasis on the word "rich"). Data-rich IPMS are anal—
ogous to Simple predictive mafinematical models (e.g. multiple regression)
whose validity criterion is accuracy of prediction rather than accuracy
plus the irnsight obtained from the modeling of processes.

Related to fine fineory-rich, data-rich dimension is the role of prob—
ability models. Stochastic processes have a proper role in IPMS of human
behavior (e.g. in the generation of environmental noise, experimenter pro—
duced stimuli, or experimental situations). They also have an undesirable
role in these IPMS —- when they are used because fine deterministic proces—
ses cannot be hypothesized. For example, if the model cannot predict
(based on sore criteria) which fork in a strange road a motorist will
choose, the model chooses randomly. This is considered a weakness in the
model because (a) humans do not act randomly, or (b) it is better for sci—
ence if scientists act as if humans do not behave randomly.

Thirdly, full tee of the value of IPM; requires fine construction of
a family of related models. These models (which may differ in structure,
content, or process) investigate differences in assumptions (often stated
as processes) and differences in environmental conditions. The family of
models are not very difficult to form -- each usually involves some change
in the first model. Therefore, wifin little added expense, the heuristic
payoff has increased sizeably (q.v. Newell 8 Simon, 1961, p. 175).

Finally, a comment is in order concerning fine trade-off between mod—
el—building and practicality. Science can be considered as a series of
successive approximations. First stabs into model building will necessar-
ily be gross. Measurement precision will be low, relevant factors will be

omitted (due to ignorance or a desire for simplicity), and irrelevant

 

31

factors included. It is usually the later approximations which can be

judged with criteria other than "future possibilities."

Computer Simulation:

 

The remaining part of finis section is devoted to computer simulation.
Corputer simulation, here, is thought of as the typical mefinod for opera-
tionalizing IPMS. The term "Simulation" has been used to cover a broad
range of activities (q.v. Crawford, 1966; Hermann, 1967; Abelson, 1968).
In finis paper fine term will, in general, refer to the Simulation of cog-
nitive processes .

First, a brief comment about artificial intelligence.

It is often argued that a careful line must be drawn be—
tween the attempt to accomplish with machines the same
tasks that humans perform, ancT the attempt to simulate

the processes humans actually use to accomplish finese
tasks. (Newell 8 Simon, 1963b, p. 279)

 

Machines and computer programs which are designed to accomplish by any
means, task which up till then only humans could accomplish are within
the realm of artificial intelligence. Machines and programs designed

to accomplish tasks humans can accomplish in a (hypothesized) manner used
by humans are instances of Simulation. In practice, finis distinction does
not hold up well. Many of the techniques and principles applicable to
artificial intelligence are need in computer simulation, and vice versa.
Also, many researdners jump back and forth between these two areas , behav—
ing similarly in bofin. In theory fine distinction is not tenable. First
of all, if it proves mrfinwhile to maintain fine distinction, then at most
it seems to be one of levels of eXplanation. If behavior is simulated at
a given level of a hierarchical IPM, then at a more atomic level, fine pro-

cesses are determined by artificial intelligence mechanisms.

 

32

For example, Feldman (196 3) believes finat humans function in fine
binary choice experiment by hypothesizing a rule which fits the past
presentations of stimuli. Therefore, in his simulations of finis behav-
ior his model uses hypotheses. But the mechanisms used to generate
these hypotheses (at fine more atomic level) are not purported to rep—
resent fine human processes of hypofinesis generation.

If behavior is simulated with a non-hierarchical IPM, fine distinc—
tion between autonata and human beings becomes important. Since brain
processes and computer processes at the atomic levels are finought to be
fundamentally different (q.v. Newell 8 Simon, 1961) all Simulations are
based upon artificial intelligence mechanisms.

More importantly, fine value of the distinction itself can be ques-
tioned. In terms of producing heuristic models of behavior,

any automaton, whether it is intended to simulate human behav-

ior or just do man-like finings, is by definition a model of

behavior. If a machine accomplishes the sane result that a

person does, then the machine is manifestly a model of human

behavior (Green, 1961, p. 86).

Therefore, at least for the purposes of this study, no unnecessary
distinction will be made between fine artificial intelligence and the simu—
lation literature. Bofin sources will be used when applicable.

Mnile the actual simulation is the typical operationalization of the
IPM, fine computer program is fine theory. Or, as Frijda (1967) points out,
fine program only represents the theory because fine program includes pro—
cesses which fine researcher does not believe to be true or useful (even

if the mochl works satisfactorily) such as random processing.

@erationalizing the Simulation:
'Dnere are two steps involved in operationalizing an IPM: Program—

ming the model and running fine program. Both of finese may contribute to,

 

33

or hinder a mochl—builder. Programming an IPM clearly increases fine
clarity of fine model's statement. Below, Lindsay, stresses the effect
of simulation on psychological theories . His remarks are certainly ap—
plicable to most areas of the behavioral sciences.

It has long been a feature of psychological theorizing that
would-be fineories suffer from chronic vagueness . The result
is a fineory which can be stretched to fit anything. The
genesis of finis difficulty lies in the fact finat the fineorist
knows what he is saying and so does his audience. Hence, it
is often possible to put together assumptions whidn logically,
will not fit, or to make deductions which , logically, do not
follow. These unfortunate juxtapositionings may go unnoticed
by an intelligent fineorist and his informed listeners , who
can readily and unwittingly supply the missing pieces , ignore
fine excesses, and beg fine answer which they know is there
even if it is not. 'Ihe computer, though, is a very stupid
audience. From one point of view, it may prove more valuable
now while it is stupid than later when it is not; for today
it will not tolerate vagueness. Mnen a theorist with an idea
sits down to convey his idea to a machine he almost invari—
ably finds finat he must first sharpen it up. And when fine
machine attempts to simulate the idea, the theorist almost
invariably finds it will not do what it is supposed to do.
(1963b, pp. 50-51)

The desire for clarity, however, may force premature closure on the
form and extent of the model. In addition finere seems to be a Whorfian
nature to computing languages. Different languages process information
differently. Once a language is chosen (or forced onto a researcher be-
cause of its availability) the researcher must translate fine IPM into the
programming language and accept its implied assumptions. This is even
true of those languages Specially created for the simulation of cognitive
processes (q.v. Newell 8 Simon, 1963a, p. I+25).

The second step, running the program is beneficial for complex mod—
els which would be impractical to simulate by hand (i.e. follow the steps
of the IPM or program without using fine computer). Furthermore, it is

only finrough running fine program that inconsistencies become evident.

 

3n

Producing a correctly working program is a long iterative process. The
final program is frequently quite different from the first program.
Thirdly, running the program, perhaps under different conditions, allows
for the testing of fine IPM and fine evaluation of specific subprocesses.
And finally, the actual running of the program may open new paths of
study. For exanple, fine concept of "insight" may take on a new respect—
ability when simple deterministic processes within a computer program pro-
duce "insightful" behaviors (q.v. Newell, Shaw, 8 Simon, 1958).

There are several disadvantages with simulating fine model on fine
computer —- principally time and money. In addition there are the con-
straints imposed by fine size of the available computer. Are finere enough
storage locations for the model, or must it be distorted to fit? Process—
ing speed of the computer is a related factor.

A program can operate only in terms of what it krnows. This

krnowledge can come from only two sources. It can come from

assumption -- from the programmer's stipulation finat such and

such will be the case. Alternatively, it can come from exe—

cuting processes finat assure that fine particular case is such

and such -— either by direct modification of the data structures

or by testing. Now the latter source -— executing processes -—

takes time and space; it is expensive. The former source costs

nofining: assumed information does not have to be stored or

generated. Therefore the temptation in creating efficient pro—

grams is always to minimize the amount of generated information,

and hence to maximize fine amount of stipulated information. It

is fine latter finat underlies most of fine rigidities. Something

has been assured fixed in order to get on wifin the programming,

and fine concealed limitation finally shows itself. (Newell, 1962,

p. l+20)

A more detailed examination of the relationship between IPle and
computer simulation is possible. The components, structure, and primitive
processes of IPM; can be compared wifin fine components, structural arrange—
ments, and primitive processes permitted in computer languages. The com-

parison between computer languages and IPP's is, however, beyond fine scope

 

35

of finis paper, and has been discussed elsewhere (e.g. Reitman, 1965;
Newell, Shaw, 8 Simon, 1958; Newell 8 Simon, 1961). The main conclu-
sions can be identified 'from finese and ofiner papers. (1) There is a
class of computer languages particularly suited for Simulations of cog—
nitive processes (q.v. Green, 1963, p. 89—99). (2) There is a reason—
able correspondence at several levels between IPMS and computer languages
(e. g. Gladun, 1966) -— though some language processes must necessarily be
for housekeeping purposes and do not pretend to correspond with behaviors
(q.v. Baker, 1967; Frijda, 1967).

At a grosser level fine organization of programs and IPMS have major
similarities. Baker (1967) identified the two major approaches used in
simulation programs, fine basic premise approach and the surface approach.
The basic premise type of simulation program starts with a minimal set of
rules and derives the observable data from these rules. The surface type
of program starts with observable behaviors (data) and does not stipulate
an overall mechanism. Thus the basic premise —— surface distinction in
simulation prograns parallels fine theory—rich —— data—rich classification

of models noted earlier.

Testing fine Simulation:

Whenever a model is built it should be tested. All types of models
can be inspected to see how closely finey meet fine criteria desired by the
philosophers of science —- e.g. falsifiability, parsimony, etc. Computer
Simulations as models have special problems of validation. The positions
taken here are fine same as finose presented by Hermann (1967): (1) Com-
puter Simulation models are never corpletely validated. Rafiner, models

have during fineir growth different degrees of validity. (2) There is no

. u. .a 1 .6 u so
‘ . . . .c ....v v.. .. s L... . 8 ~.-. .11 ha . at \-\.

 

 

 

36

one correct validity procedure for all models. The proper procedures
are a function of fine purpose of fine model. (3) Dependence upon one type
of validity criterion is not as valuable as using several criteria.

Hermann identifies five types of validation procedures useful in
judging the correspondence between IPMS and fineir behavioral referents.

(1) The level of internal validity or reliability is ascertained finrough
test-retest procedures: What is fine Size of the variability among the
outcomes of several executions of the model —- with each execution having
fine same initial conditions? (2) Face validity depends upon the model's
output "looking good" to the modeler. The dimensions for testing the
goodness of fine look should be Specified in advance of the observation.

(3) Variable-parameter validating procedure compares the values of fine
model's constants and variables with those in the analogous real Situation.
One aspect of variable—parameter validity is sensitivity testing. What
are the differences in output caused by differennt initial values of the
variables or parameters? (1+) Event validity is a function of the accu—
racy of the model's predictions. (5) Hypothesis validity includes empir—
ical relationships anong variables similar to those represented in fine mod—
e1.

There are several types of techniques recomended for validating sim-
ulations of cognitive processes. The most general of these is T‘uring's
Test (1963). This test asks an observer to distinguish between computer
output and human behavior (usually in written form). The more the observ—
er errs in identifying fine two reports, the more the model as a simulation
of fine behavior is validated. T‘uring's Test takes many forms and can be

applied at different levels of analysis. For example, it can be applied

 

 

 

 

 

iv

 

A \
«Nu

m

l.

 

 

 

37

to fine grossest from of behavior sucln as the final product of the
model and man (as in testing artificial intelligence models), and it
can be applied to finose lower level processes which produce these
macrobehaviors.

Other validating procedures are protocol matching and statistical
testing. Protocol matching is a form of Turing's test. It entails fine
comparison of a person's step-by—step examination of his own thought
processes with a trace of fine computer processes. Though a useful pro-
cedure, there are several problems with protocol matching (e.g. Dennett,
1968). These will not be discussed since finis procedure is applicable
to finose models of specific individuals refiner than those models of
generalized individuals —- and the model presented in the next chapter
is of fine generalized type. In addition, this procedure assumes a con—
sciously functioning subject, whereas association behavior is not gen—
erally finought of as being consciously planned.

Statistical testing may be considered to be more appropriate for
models of generalized individuals. Most statistical comparisons will
involve proving the null hypothesis (i.e. showing finat the model and the
modeled produce the same output), which is a questionable procedure.
Also, if the model produces numerical values, it is difficult to estimate
fine number of degrees of freedom within fine model (suppose, for example,
a multiple regression type of model predicts a score representing the
average score; since beta weights rather than individual scores went into
making that prediction, how many degrees of freedom, comparable to indi—
viduals, should be used to test fine average?). A more workable alterna-
tive is to compare the output from a family of models using one's judg-

ment (a weak form of Turing's Test) as fine criterion.

I r)

—w—

   

38

§wr

In summary, finis section of chapter 2 argues that an IPM is a very
useful way of presenting a theory. This type of model compares favors
ably wifin other symbolic models used in the behavioral sciences , fine
verbal and fine mathematical. The comparison is especially favorable
whenever fine goal of fine theory construction includes insignt as well as
predictability. Secondly, computer simulation is seen as an operational—
ized IPM. The benefits and limitations of fine conversion from an IPM to
a running computer program were presented. Finally, fine problems of mod—
el validation was discussed. Because IPMS have more to offer a researcher,
they are harder to validate than ofiner types of models . Models of behav-
ior are first of all behavioral science and secondly models. If they do

not explain behavior, the fact that they are consistent is of little import.

IPMS of Verbal Behavior

 

There are computer simulation models which are not directly concerned
wifin verbal behavior, but do have something to contribute to a model of
free association behavior. Presumably knowledge of models dealing with
automatic language translation and linguistics (q.v. Garvin, 1963), seman-
tic nets (q.v. Quillian, 1967), and answering in English (q.v. Green,

23 31; , 1963) will aid in fine construction of a model of free association
behavior because all deal with natural language. Also, those IPMS of

information storage and retrieval (q.v. Garvin, 1963) might suggest solu—
tions to fine problem of storing words in memory and later retrieving them
in a free-association task. These models will be considered only second-
arily, however, to simplify the task of constructing an IPM of free asso—

ciation behavior.

 

 

£3357“. ant-37.. f

 

 

 

39

As primary resources, several related models of verbal learning
and verbal behavior will be used. These models can be modified to
handle natural language units, they present approaches to fine storage and
retrieval problem, and fineir structure is such finat with fine addition of
some processes finey can be adapted to model free association behavior.

Any complex IPM is difficult to talk about. To describe, in any
depfin, its structures and processes is usually prohibited because of its
size and complexity. For example,

fine description of a recent version of Newell, Shaw, and Simon's

General Problem Solving program covers more finan one—hundred

pages, and even so contains only fine main details of the system.

Furthermore, the discussion assumes a knowledge of an earlier

basic paper on GPS and a knowledge of Information Processing

Language-V, the computer language in which it is written. Final-

ly, fine appendix, whidn simply nanes the routines and structures

employed takes another twenty—five pages. Unless one is famil-

iar with similar systems, a thorough grasp of the dynamic proper-

ties of so complex a model almost certainly presupposes experience

with fine running program and its output. (Reitman, 1965, p. 21+)
The models of verbal learning and verbal behavior described below have,
for fine most part, complexities on the order of finat of GPS. Therefore,
fineir description must of necessity be terse.

Five related models will be discussed: EPAM I, EPAM II, EPAM III,
WEPAM, and SAL I—III. These models simulate subjects in either a paired-
associate (P-A) paradigm, or a serial anticipation paradigm of nonsense
syllable learning.

In fine P—A Situation fine subject is presented wifin a list of stim-
ulus—response pairs. For eadn pair, fine stimulus item is presented
to the subject whose job it is to give fine correct response. After the

subject responds or after a fixed interval of time, fine correct response

is presented. Usually fine list of pairs is presented until the subject

‘ *fi'i... .— 1 , "
. . .l J ‘
Se .1. 1.. mg; nm'm
u-. 3“;
4 _. é. "
:s. r. "
I

h

 

 

140

learns fine list to some criterion. Each time the list is presented,

fine order of the pairs on fine list is randomized. In the serial anti-
cipation paradigm, the subject is presented with one list of items. Each
item (except fine last) serves as fine stimulus for the next item, and
(except for fine first) serves as a response to fine previous item. When
fine list is presented several times, fine order of the items on the list
is not changed.

'Ihe five models below describe finose processes a human subject goes
througn in such experimental situations. The interpretation of the func-
tioning of these models provides one possible set of explanations of some
psychological phenonena related to learning (e.g. forgetting, retroactive

inhibition) .

w:
The first and simplest of these models is EPAM I (Elementary Perceiver
and Manorizer). EPAM I was developed by Feiganbaun and Simon (1962b) to
account for the serial position effect. In serial anticipation learning,
if the total nunber (or percentage) of errors is plotted as the ordinate
against fine serial position of fine items on fine list, a typical bowed
curve results: more errors are made on items in fine middle of the list
than at either end, and fewer errors are made at the beginning of fine list
than at fine end of the list. This curve represents the serial position
effect. As described by Feigenbaum (1959, p. l+6447), EPAM I consists of
four macroprocesses.
MO: Serial Mechanism.
The central processing mechanism operates serially and is capable of

doing only one thing at a time. Thns, if many finings demand processing

 

‘41

activity from fine central processing mechanism, finey must share the total
processing time available. This means finat fine total time required to
memorize a collection of items, when finere is no interaction among them,
will be the sum of fine individual items.

M1: Unit Processing Time.

The fixation of an item on a serial list requires the execution of
a seqtence of information microprocesses that, for a given set of experi—
mental conditions, requires substantial processing time per item.

M2: Immediate Memozy.

There exists in fine central processing mechanism an immediate memory
of very limited size capable of storing information temporarily; and all
access to an item by the microprocesses must be finrough the immediate
memory.

M3: Anchor Points.

Items in a list which have unique features associated with them will
be treated as 'anchor points' in fine learning process. Anchor points will
be given attention (and finus learned) first; items immediately adjacent to
anchor points will be attended to second; items adjacent to these, next; and
so on, until all of fine items are learned.

3% of MO - M3

M0 establishes a serial processor, capable of doing only one fining at
a time; this creates a need for deciding fine order in which items will be
processed, i._e an attention focus, and M3 establishes a mechanism for de-
termining this order. M2 provides a tarpomry storage, while the processes

in M1 are permanently fixating the item.

 

 

#2

To account for fine serial position effect, Feigernbaun and Simon
(1962b) posit that fine anchor points at the beginning of the learning
task are the first and last items on the list. The subject focnses on
one of finese items (choosing at random between than) and memorizes it.
Once an item is memorized fine effective beginning or end of fine list is
changed to fine first and last unknown items —— which finen become fine anchor
points. This process is outlined in Figure 1.

EPAM I is a powerful, yet simple model. Its simulated data agree
closely with empirical data and, as a theory, is to be preferred on par-
simonious grounds over fine more complex explanations of the phenomenon
(q.v. Feigenbaum 8 Simon, 1962b). On heuristic grounds, EPAM I can be
judged quite favorably. For example, if certain items in fine middle of
the serial list are altered fine characteristic curve is changed. EPAM I
accounts for this (with fine sane processes as fine serial position effect)
by stipulating finat these items become additional initial anchor points.
Thus, "one would also expect that ofiner items could be made unique by
printing them in red . . . , or by making some items much easier to learn . . .
or by explicit instructions ..., etc." (G.A. Miller, 1963, p. 325). The
model also lends support to other learning phenomena such as one—trial
learning —- an item in EPAM I is eifiner learned or it isn't. EPAM I is
limited because it cannot learn a list in which an item appears more finan
once -- a task SLbjects can do with difficulty.

The valte of EPAM I, w, to finis study is negligible. Its import-
ance lies in fine fact finat fine EPAM I macroprocesses oversee the more spe-

cific microprocesses of EPAM II and EPAM III.

 

l+3

IDENTIFY INITIAL
ANCHOR POINTS.

[CHOOSE ONE OF THE ANCHOR L

 

 

POINTS AR RANDOM

l

 

 

 

MEMORIZE ITEM AT THE ANCHOR
POINT CHOSEN. ELIMINATE
THAT ITEM AS AN ANCHOR POINT.

 

 

 

 

ARE THERE ANY UHLEARHEO\ yes
ITEMS ON THE LIST? /

 

no

r
EXIT.
LIST LEARNED.

 

 

 

 

 

SET THE ITEM(S) NEXT
TO THE ONE JUST MEM-
ORIZED AS ANCHOR

 

POINT(S).
1

Figure I. Item Selection and Learning in EPAM I.

 

.' w

_ a. w
an.

$1 :4 “farm or m“..- ,.. s .-

-~ .0 ..

 

 

nu

EPAM II:

EPAM II is mudn more complicated than EPAM I because it posits one
plausible set of microprocesses at the information processing level of
explanation which seem to account for several verbal learning phenomena.

A part of the EPAM II program simulates fine experimenter and experimental
conditions in a verbal learning situation (e.g. simulation of the memory
drum controls fine amount of time the subject has to respond). In this dis-
cussion, the major focus will be upon fine microprocesses used in the Simu—
lated subject.

The inputs to EPAM II are binary coded nonsense syllables. If the
program makes a response, it is also with binary coded nonsense syllables.
Coded nonsense syllables are used because fine routines within the program
which convert fine nonsense syllables to coded nonsense syllables (and vice
versa) have not been developed. They are not central to the goals of EPAM
II. The coding of nonsense syllables is done letter by letter. Each let-
ter is represented by ten bits -— five of which are redundant. Feigenbaum
(1959) callsthe binary coded external stimulus (the nonsense syllable) a
"stimulus input code" or "code."

There are two major sets of microprocesses in EPAM II. Performance
processes function to produce the response associated with the stimulus .
Learning processes are more complex. They work,

to discriminate each code from fine others already learned, so finat

differential response can be made; second, to associate information

about a 'respcnse' syllable with the information about a 'stimulus'
syllable so finat the response can be retreived if the stimulns is

presented. (Feigenbaum, 1963, p. 301)

Figure 2 presents an overview of the performance processes. The code

is sorted through a discrimination net to a terminal. A discrimination

 

1+5

 

[INPUT STIMULUS CODE._]

 

 

 

DISCRIMINATE CODE TO
FIND IMAGE.

 

FIND ASSOCIATED
CUE CODE.

 

 

 

T
DISCRIMINATE CUE
CODE TO FIND RESPONSE
IMAGE.

 

 

 

 

 

' T
[OUTPUT RESPONSE COOL]

 

Figure 2. EPAM 11 Performance Processes.
(Adapted from Feigenbaum. 1963. p. 300).

 

46

net to a terminal. A discrimination net is a tree of binary testing

nodes which examine bits of the code to identify characteristics of a
letter in a given position (e.g. is the third letter closed?). Terminals
are either empty or contain images which are more permanent representations
of the stimulus (the code e>dsts in "immediate memory" and if the memory
drum turns, the code is lost; an image, on the other hand, is never lost).
If the code matdies the image (i.e. the code is recognized) a cue code is
sought. A cue code is a subset of the response code which, at the time

it was stored at the terminal, was minimally sufficient to retrieve the
stored response image.

In the learning microprocesses, discrimination learning functions to
correctly identify stimulus and response items. _'I'he discrimination net
is modified (i.e. learning takes place) whenever identification is incor-
rect.

To understand how the discrimination and memorization processes
work, let us examine in detail a concrete example from the learn—
ing of nonsense syllables. Suppose that the first stimulus—res—
ponse associate pair on a list has been learned. (ignore for the
moment the question of how the association link is actually form-
ed). Suppose that the first syllable pair was DAX—JIR. 'Ihe dis-
crimination net at this point has the simple two-branch structure
sham in Fig. 3. Because syllables differ in their first letter,
Test 1 will probably be a test of some characteristic on which
the letters D and J differ. No more tests are necessary at this
point.

Notice that the image of JIR which is stored is a full image.
Full images must be stored —— to provide the information for
reco 'zin the stimulus. How much stimulus image information
18 required the learning system determines for itself as it grms
its discrimination net, and makes errors which it diagioses as
inadequate discrimination.

To pursue our sinple example, suppose that the next syllable
pair to be learned is PIE-JUK. There are no storage terminals

in the net, as it stands, for the two new items. In other words,
the net does not have the discriminative capabilityto contain
more than two items. ‘Ihe input code for PIB is sorted by the net

 

Figure 3. Discrimination Net after the Learning of the First Two
Items. (Adapted from Feigenbaum, 1963, p. 303).

 

Figure 4. Discrimination Net of Fig. 3 after the Learning of
Stimulus Item, PIB.
(Adapted from Feigenbaum, 1963. p. 304).

 

us

interpreter. Assume that Test 1 sorts it down the plus branch
of Fig. 3. As there are differences between the incumbent image
(with first letter D) and the new code (with first letter P) an
attempt to store an image of PIB at this terminal would destroy
the information previously stored there.

Clearly what is needed is the ability to discrimdnate further.
A match for differences between the incumbent image and the
challenging code is performed. When a difference is found, a
new test node is created to discriminate upon this difference.
The new test is placed in the net at the point of failure to
discriminate, an image of the new item is created, and both
images —— incumbent and new —- are stored in terminals along
their appropriate branches of the new test. and the conflict is
resolved. The net as it ncw stands is shown in Fig. 1+. Test

2 is seen to discriminate on some difference between the letters
P and D.

The input code for JUK is no» sorted by the net interpreter.
Since Test 1 cannot detect the difference between the input
codes for JUK and JIR (under our previous assumption), JUK is
sorted to the terminal containing the image of JIR. The match
for differences takes place. Of course, there are no first-
letter differences. But there are differences between the in—
cumbent image and the new code in the second and third letters.

Notic1n ' c1ng Order. In which letter should the matching process
next scan for differences? In a serial machine like EPAM, this
seaming must take place in some order. This order need not be
arbitrarily determined and fixed. It can be made variable and
adaptive. To this end EPAM has a noticing order f__or letters _o_f
syllables, which prescribes at any moment a letter-scanning se-
quence for the matching process. Because it is observed that
subjects generally consider end letters before middle letters ,
the noticing order is initialized as follows: first letter,
third letter, second letter. When a particular letter being
scanned yields a difference, this letter is promoted Lp one
position on the noticing order. Hence, letter positions rela-
tively rich in differences quickly get priority in the scanning.
In our example, because no first—letter differences were found
between the image of JIR and code for JUK, the third letters

are scanned and a difference is found (between R and K). A test
is created to capitalize on this third—letter difference and the
net is grown as before. The result is shown in Fig, 5. The
noticing order is updated; third letter, promoted up one, is at
the head. (Feigenbaum, 1963, pp. 302-3014)

 

 

Association learning functions to pair the correct response to its

stimulus. When an image is placed in an empty terminal a cue of the cor-

rect response is also placed in the same terminal. Thus the response is

 

 

v‘ V
ma.” «.4 r‘ 7"

 

 

 

 

 

 

 

 

Figure 5. Discrimination Net of Fig. 4 after the Learning of the
Response Item. JUK.
(Adapted from Feigenbaun, 1963, p. 304).

 

   

 

50

associated with the stimulus. The one is determined by trial and error
to be that minimal subset of the response code, which when sorted through
the discrimination net will retrieve the correct response. It is impor-
tant to remember that the cue is the minimal satisfactory subset. Be-
cause as learning takes place the structure of the net changes and a giv—
en cue may no longer be sufficient to retrieve the correct response. At
this later time the cue may not contain sufficient information to be test—
ed at a test node (e.g. the cue code may be the first letter of the re—
sponse, while the testing node is checking the third letter position).
When this happens one of the two branches below the test node is chosen
randomly. One of three possibilities now exists. (1) The cue can by
chance be sorted to the correct terminal and the correct response will be
given (thougi there is no guarantee that this will happen the next time
the stimulus item is presented). (2) The cue code can be sorted to an
empty terminal and no response be given. (3) The one code can be sorted
to a non-empty, but incorrect terminal, and an incorrect response is made.
In both the second and third cases, additional learning processes are
brought into play (when the correct response becomes available in the mem—
ory drum) and the cue is modified to insure that when it is sorted through
the net as it now exists, it is minimally sufficient to retrieve the cor—
rect response. If an empty terminal was found the learning processes be—
gin to build a response image in that terminal.

The basic structme and processes of EPAM II have been presented.
Before evaluating the model a more extended example of its functioning
would be helpful. The example below is adapted from Feigenbaum (1959,
pp. 86—96). 'Ihougi this example is of serial anticipation learning, it

should be evident that the same microprocesses can effect the simulation

 

51

of P—A learning. In the example, the experimental situation consists of
one list of six nonsense syllables: KAG, LUK, RIL, PEM, ROM, TIL. Ini-
tially the noticing order (N.O.) is first letter, third letter, second
letter, and the mem'mum number of test nodes added to the net each time
it is grown is three (it is efficient to detect several differences each
time the net is grown and add more than one test node at a time). The
learning criterion is one perfect trial. Macroprocesses oversee the
learning and determine the order in which stimulus—response items within
the list are learned. rIhe processing of the example is outlined in Table
2 and Figures 6-10 which summarizes the "learning" of the list by EPAM II.

The effectiveness of EPAM II as a model is indicated by the verbal
behaviors it simulates.

Study of the behavior of EPAM in an initial set of about a hun-
dred simulated experiments shows that a variety of 'classical'
121?} learning phenomena are present. Referring to traditional

ls, these include serial position effect, stimulus and re-
sponse generalization, effect of intra—list similarity, types of
intra—list and interlist errors, oscillation, retroactive in—
hibition, proactive effect on learning rate (but unfortuaately not
proactive inhibition), and log-linear discriminative reaction
time. Further experiments, especially those involving inhibition
phenomena and transfer phenomena are now in progress. (Feigenbaum
8 Simon, 1963a, p. 335)
To illustrate how the model simulates one of these phenomena, there is
an instance of stimulus generalization in the example presented in Table
2. Stimulus generalization is the name given when stimulus B's response
is given to stimulus A (stimulte A and stimulus B are similar on some
dimension). In Trial u of the example TIL is erroneoxely given in re—
sponse to PEM. This is an instance of stirmzlus generalization. The test
nodes do not discriminate between PEM and ROM (operationally they can be

defined as similar stimuli because both meet the same testing conditions).

 

52

:3.» 93.5.8 8. men? mg pm moor.“
nmom ...JH .moo o5. 5?..qu umeOmm<
.ubc o B. omoom gm.

 

mono: poop roam. . 54...: puma Iii HS.
-80 we “a Ema: .23,“ mfi 8. .51 «8: 20m
mo ommfi Show E. 2.3m opoEfiHOmE woos 2mm
.msla wfiqmpc8 3353 m8: .5 N
0p M3 mo mwmfi “how B. common coco v54
.N. am new 3mg 98.93 “SHE GSA owmquHOmE M3 92
I .o.z E 586
mos Tn 53.93 ”roar.“ urn. .wfl 5H3
@805 mm: a 938.8 8. mood
Imam ..IH 85m 6:319? 33.089».
688: Pam mono: pump Hmcowfio
low 02 .pmc can. @585 cough m..n
x3 mo meH can. "x3 3ng
dome m..n mwdommwn 0c :96: III: HE.
Jump wmfi pm $0 8an on we coma. 08c 28
5.2 of co 8am new? we. 8. mm ago; as
mason E. .wpmmu. 98.on “Emu mam mm. woos 4.2 H
.3 «E. .389 9953 Pawn.“ 5H3 96: M3
.m puma 8m mason .o.z 8:6 ”eel opquEomE Soc 9Q
8902 wcwmmmooum ngovcH 3:09QO 93m HE

 

 

 

 

 

 

 

mficsfl nowadays. show 5 wfimmoooa HH 25m .8 verso é

. N OHAMH

 

53

coop mommmooua manhood
650% me 35:53.
598 cm 38 mums.“ cH
.omcoawwn #0858 93 mo
counrooamm mfi. mmpcwugw
on pcmwflmsm .393

co w...” x3 spams ompmwo
uOmmm Hulcm moo o5. .H

.pmc mfi. CH promo: mmwcmno dmcoflh

loom. o: 853 ogom mun ocm oofinmoa
saga ma zom "20m anagrams

.pmc any 8 spoon. mo 58. amfima

umfi m as. new @389 we .o.z

. mpwqueflomflo you or 3mm“ 93va

BE .88m .58 an? Erase
win new mo ...ng "Em wagons

.5.
.5
was

THUG

 

.m 8&3 wow

[:39 mgmwupmn ocm oposoona
mun To .80 en 8. omens ma .o.z
can omquHmpmo Qofioa “one: o5 .80..“
:89? .38“. 99.53” umufl mum you
at no c8. «5 pm 38o. mgmown
Help gnflwwm mamas. no: 2e, $5
5.5mm .7: ma moo omcommmn m um
sensor “and ma. 35.28 33083,
.o.z 8. 8n
mummy 9393 35. mom mmooc 3mz
.umc 0v omoom Pam muooc 3o: mg
.fl ocm “WM mpgmoom poccmo ma. new
NH .98 mo oowm ”Ema :38 somehow
me do. no mmEH as. organ
.3.»
com mums: c0053 upméfiomwo and
new mummy $th Rafi 98 ~85.
.pmc on. broom mam moooc camp“ 38
$3. swim 8 e389 we .o.z
.vfloz “on on 953 9953 emery och
...WIHM mango Hungry 0o. outflow
we 20m mo omMEH "20m 31....ng

08C
08C
08C

AHM

AHB
20m
am
AHM

 

mopoz

 

 

mafimmmoonm .3535

 

 

omconmem

 

935nm

 

Hog

 

 

AfiLSOV

. a manna.

 

51+

 

 

 

 

 

 

 

 

 

 

 

 

 

AH. Som
tom Em N.
am AHM
flux V54
.3 am 8m .88 x3 3
I .a gorgeous ow #fifioflhfidm ...iu .du.
is 8. smegma ma sum 80 “Eaters mpmfiomme. do. 28
Imam mun moo .8889? .omwcmsoc: mun .52 sex am
.vlam. gasp...“ 8. NM 658 we aw "Em mpmcefioma 88 fix a
83... no: 88 85.5 9 lsmwfififi we “52 dm v13
Hog .mov 2WD mob. .H 650% mW .3...» "HR muméegmwn Mono: 92
. m magma mom a -
.zomlrmm
03H 8. a Hard. fi mfip
smoocm 90c mm: 985. .m in: AB.
.ummcmfifi we pm: .3. :8
2n. .88 “a 856 so 80: Em
5m new no ds,/mac. .m we .slmm fins eflfim 25 ~98: .5 m
Imp corona murmofl .Eon owlolampwn on. pcofloflmmdom 45 x3
imam poo mooo .iH .H poo ma 1.9 "£97sz mumUOmmmx USA 92
.omsm ma
:8 8 mmcoammm .N
.uoc page 06. Low
chwWEmom on H53 “Z.”
cm 3.» mom moo NEE—mam
882 mfimmmooa €35 wmcoammm 92% amps
Torrcoov .m magma

 

 

 

 

kag,l--

 

 

 

KEY to Figures 6-10.
Circles represent testing nodes
identifying the node and the

position of the letter being
tested.

Rectangles are terminals which
may be empty. contain an
image. or an image plus a

‘ response cue code.

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Figure 6. Discrimination Net after Trial 1.
(Adapted from Feigenbaum, l959. p. 90).

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

+ -
+ - + -
e o
+ - + -
luk,r-- Q
+ -
ril

 

 

 

 

 

 

Figure 7. Discrimination Net after Trial 2.
(Adapted from Feigenbaum. 1959. p. 9l).

 

v.-'

+

3

 

 

 

 

56

 

 

 

 

 

 

 

10k ’7'-

 

 

 

 

 

 

 

 

 

til

 

 

 

 

 

 

 

 

ril

 

 

 

 

 

 

 

 

 

 

 

 

 

 

m’t-I

 

Figure 8. Discrimination Net after Trial 3.

(Adapted from Feigenbaum, l959. p.93).

 

 

 

 

 

 

 

Figure 9. Changes in Discrimination Net after Trial 5.
(Adapted from Feigenbaum, 1959. p. 94).

 

58

 

 

 

 

 

 

 

 

 

 

luk,r-l - Q

 

 

 

 

 

 

 

+ - + -
- , e
4. -

 

Mgt'1 -

 

Figure l0. Discrimination Net after Entire List is Learned.
(Adapted from Feigenbaum, l959. p.96).

 

 

 

59

In addition to these phenomena, the model also contributes to a
theory of forgetting (q.v. Feigenbaum 8 Simon, 1961). In EPAM II items
are never permanently lost from memory; they are misplaced. 'Ihus, what
appears to be forgetting is, in fact, the effect of later learning (i.e.
retroactive inhibition). See Trial 6 in the preceeding example for an
instance of forgetting.

EPAM III:

Certain deficiencies in EPAM II (such as learning a list containing
two identical nonsense syllables which differed according to some external
property; e. g. color) lead Feigenbaum and Simon (1962a) to construct EPAM
III. EPAM III is a much more complex IPM than EPAM II. It uses a hier—
archical discrimination net (not present in EPAM II) which sorts and learns
letters, nonsense syllables, and stimulus—response pairs similarily. It
efficiently makes use of early learning by treating previously learned
letters as syllables and previously learned syllables and stimulus-response
pairs . By means of a property set attached to letters , syllables and
pairs , EPAM III can discriminate between items alphabetically similar but
having different contexts (e.g. coming at the beginning of the list),
modes of production (e.g. oral, visual), and different external character-
istics (e.g. pica type).

The discrimination net in EPAM III is composed of three related parts.
The letters' portion of the net is very similar to that of the entire EPAM
II net. It is composed of test nodes (which test some attribute of the
encoded letter) and terminals which may be empty or contain an image of a
letter. ‘Ihe tests in the letters' portion of the net are, in effect, bi—
nary tests checking if the encoded letters meet or do not meet the crite-

rim being checked.

 

60

The syllables' portion of the net is composed of attribute test—
ing nodes, subobject nodes and terminals. The attribute testing nodes are
n—ary branching. 'Ihey test individual syllables (e.g. what is the color
of this syllable?). The terminals contain images of a syllable. The
image of a syllable consists of cue tokens (analogous to cue codes in
EPAM II) for each letter in the syllable.

'Ihe stimulus-response pairs' portion of the net consists of n—ary
branching tests (which identify subobjects of the pair and attributes of
the pair) and terminals which contain complete or partial images of pairs.
An image of a pair consists of the cue tokens for each syllable member of
the pair.

Since some nodes in the syllables' and pairs' portions of the net
test for lower level components (i.e. letters and syllables respectively)
these nodes are called subobject nodes. Figure 11 presents a partial dis-—
crimination net of EPAM III. In Figure ll (and througiout the remaining
discussion of EPAM III) 0 represents an object (i.e. letter, syllable, or
pair); 0' is the terminal 0 is sorted to; and 0" is the image stored in
terminal 0'. 0" may or may not correspond to 0, it is simply the image in
0'. 0' is also the cue token of 0". It names the terminal containing 0".
0" consists of a list of the cue tokens of the subobjects of an object plus
a property set of its attributes.

Suppose from previous learning trials, part of a discrimination net
exists as shown in Figure 11. In the experimental situation the list to
be learned contains the pair CAT-m6 (which has already been completely
learned). When CAT appears in the window of the simulated memory drum,
the simulated subject should respond with 113G (before DOG appears in the

AS a .32 .2»: se.... e335 .
szzamsgzauscsfisfitxa233< ;.23:

 

.3330 05 3 £353.:

.3 v33... c3225. 2» a a a I

 

 

 

   

    
                     

 

 

 

 

  

 

5;: van: 3 9. .55.... .c u /
a use; no: 33 258 . /
2:? 05 :23 8.0.: m. 8.
:3: mo cotton .2393 e I 9 a
s9.» 33:8 53 22 .o
ax can my moguearm HuSz
a... 6 I e .
1 .G8 a M .:‘~ \
5 I
.5 m a \
.auoo-h<uv \
:.Eee_....... e
\ . I - o
\ —ZH: .m” .9 pg ’
I. /
Es. ..-m 35:: s. ,
e. .5 5:3 9. e. a at”... e e e a
wmxue hmm~u
mu 2:: ma 2:: .<
open—Em

even ¢-m

 

 

Sumac

 

.532 t

 

 

62

memory drum window). The performance and classification processes would
proceed thusly. CAT is a syllable and will not be recognized until an
image is found. . In the syllables' portion of the discrimination net
(center of Figure 11) the first subobject testing node asks for the first
letter of the syllable. Since CAT is not yet recognized this question
cannot be answered. The letter C is sorted througi the letters' portion
of the net (left side of Figure 11) reaching terminal C' which contains
the image C". (Just as in EPAM II, if learning were not complete at this
point, C migxt have been sorted to an empty terminal or one whose image
did not match the object). C is recognized and branch C' is taken in the
syllables' portion of the net. The second subobject testing node there
requires the identification of the second letter of the syllable, A. A
is sorted through the letters' portion of the net until it is recognized.
Branch A' is taken and CAT is sorted to terminal CAT' which contains an
image (i.e. CAT is recognized). Since learning is complete the terminal
sorted to is, in fact, the correct one. But, in general, there is no re—
quirement that the correct terminal be reached for a stimulus to be re—
cognized.

Once (AT is recognized an attempt is made to make a response. A
dummy stimulus-response pair, CAT'—_, is formed (the response is not
yet in the memory drum window). This dummy pair is sorted into the pairs'
portion of the net (rigit side of Figure 11). The subobject node there
requires that the stimulus member be recognized —- which it is. The dummy
pair is sorted down the CAT' branch which terminates at terminal (CAT-DOG)' .
Again, because earlier learning had been complete an image of the reSponse

exists at that terminal and is the correct one. Within that image is the

- V"

 

 

.rp - a» 11:5

.
ww-
nVo

fl

AV

:1.

.II
.4.
vvfin

o-
J
A

“a-
45'.

Hi 43"“
W-

 

63

cue token DOG' . This identifies the syllable terminal which contains cue
tokens D' , 0' and G' which, in turn, identify terminals in the letters'
portion of the net. Stored along with the images of the letters is in-
formation needed to make an actual response DOG. It should be noted that
once the response image DOG" was found, further sorting through the dis-
crimination net was not necessary since cue tokens in EPAM III name a
terminal and do not themselves have to be sorted through the net -— as

is the case with the cue codes in EPAM II. This is an example of early
learning affecting the rate of later learning.

Assuming that the subject had enough time (i.e. the memory drum did
not turn) to make a response, that response is compared with the correct
response (when the drum finally turns) and if they do not match the net is
augmented. This takes the form of adding test nodes to the net, changing
or modifying existing images, and adding a new image to a terminal.

To better understand some of the processing involved in classification
and learning (image building and discrimination) in EPAM III, an example
of P-A learning is outlined in Table 3 and Figures 12-16. Initially the
discrimination net is as described in Figure ll. The only pairs on the
list to be learned which will be discussed are CAT-DOG and CAB—MAN. CAT—
mG has already been learned but neither CAB nor MAN has affected the net
in any way.

If there is no interference from other pairs on the list, CAT-DOG
and CAB—MAN have been correctly learned after Trial l4. Trials 1 and 2
of this example contain instances of stimulus generalization (CAB elicits
DOG). While the alternating responses to CAT illustrates oscillation, a

familiar phenomena in verbal learning.

Citric: .w..-..~LO....¢< CDAwQL eas Equine t3 m.~L.BX..,~ E< -m. ...maBN

 

 

       

.nnuaao mH m-m eases may .m

mammmoga stoapfioom
pom wannaagm mH MEG. 39: H Hanan.

. mH .asH .mH .mH .mH

 

   

 

  

 

  

 

     

61+

 

     

 

 

  
           

 

 

 

 

 

:H ampmaHaHmomHu mm: zaz moch .N .HH .mm .m .Na .m .m .3 .N .H "mampm zaz
.poc :H «was mmwcmgu oz a
683583 ts Hos mH as .H .H: firs cH mm 258 m5
.pmc CH moms mmwamso oz .5 .3 dog sun mm Sammy won
.pw: cH mum: mmwcmcu oz « .AH Hmng cH mm «ammo sac
.mH magma mom !. ommsmfio mH $2 a
.50 mesh NE. .3
.pm: On ounce cam oopmcH
IEdHomfio ma e853 CH oncommum .m .mm .1. Wm .m 52 .m T.” "mmopm z<z
.pmc 5 come wmmcmso oz «
..B<u Hmcaeump op cashew mH m<o .N .mm .mm .am .HN .NON .mH .H ”mampm mac
.5: 5” mpg wwwsmso 02 a "mampm GB
.90: 5” come newsman. 02 a
Hound. mHfi
whommn HH mnsmHm mxHH mxooH pmz .H .mN .mm .mm .am .HN .ou .mH .H "wampm sac
mmpoz 3H am 3. 08.35 553
mfimmmoonm HMEQEH 5" EMPH

 

 

was $3083 seams 5 saw no 393. 5.

 

 

 

 

66

1

IS A RESPONSE IN es ENTER S-R ASSOCIATIVE
THE DRUM HINDOH? PHASE OF F20.

 

no
3

2 .
o - HAS A RESPONSE ELICITED no ‘DISCRIMINATE _
av THE STIMULUS? TNE RESPONSE _‘1‘

yes

 

 

 

 

 

 

4 5

F22. CHECK THE ELICITED RESPONSE HAS THE ELICITED es ‘
LETTER BV LETTER AGAINST THE RESPONSE CORRECT?
RESPONSE IN THE DRUM HINDOH. ~ °

 

 

 

 

 

 

 

 

 

 

 

 

 

 

no
6
F0. LEARN S-R ASSOCIATION.
SORTW NU u ERNINAL
5' IN THE NET. SAVE s'.
7 OISORININATE TNE RESPONSE.
HAS THE MEMORY \res . J
DRUM TURNED? /
no
a
SORT RESPONSE R TO TERNINAL R' .
SAVE R' . 9 yes
CONSTRUCT AN 54! OBJECT.

 

PUT 5' ON THE S-R OBJECT AS ‘ DOES THE TERMINAL (S-R)‘
ITS FIRST COMPONENT. GIVE HAVE AN IMAGE?
THE S-R OBJECT A VALUE l
IDENTIFYING IT AS AN S-R "0
OBJECT.
SORT THE S-R OBJECT TO A
TERMINAL. (S—R)‘.

 

 

 

 

 

 

 

 

 

IO
33
A DISCRIMINATE THE S-R OBJECT. (IRON-l Em
1N6 ITS IMAGE IN THE S-R TERMINAL. 1 f

 

 

 

 

Figure 12. Flowchart of F20, the Paired-Associate Learning Routine
of EPAM III.
(Adapted from Hynn, 1966. p. 53).

\fil.

 

 

67

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

TI 12
DOES THE 54! IMAGE HAVE no ADD 5' TO THE S-R IMAGE
A STIMULUS COMPONENT? . As ITS FIRST COMPONENT.
yes
13 , 14
DOES THE S-R IMAGE HAVE no ' ADD R' AS THE SECOND COM-
A RESPONSE COMPONENT? PONENT OF THE 542 IMAGE.
yes
15 Is
DISCRIMINATE THE I . HAS THE MEMORY es
S-R OBJECT. *\ DRUM TURNED?
no
17
REPLACE THE RESPONSE COMPONENT OF THE
M IMAGE HITH R'. DESTROYING THE
PREVIOUS ASSOCIATION. IF THE RESPONSE
COMPONENT HAS DIFFERENT FROM R'.
18 33
DISCRIMINATE THE 1 . r—im,

 

STIMULUS SVLLABLE:J

 

Figure 12. Flowchart of F20. the Paired-Associate Learning Routine
of EPAM 11! (continued).

 

68

ENTER RESPONSE PRWICI’IM PHASE 0F F20.

19

"F21. RESPOND AND REMEMBER THE RESPONSE,

s-R OBJECT A VALUE INDICATING
IT IS AN s-R OBJECT.

 

 

 

 

 

 

 

 

21 20
OOES S' ALREADY HAVE AN IMAGE? SORT STIMULUS 5 TO
(IS THE STIMULUS RECOGNIZED?) TERMINAL 5'.

yes no

22

DISCRIMINATE THE STIMUS, ADDING
IMAGE OF S IN TERMINAL 5'.

 

 

2

ADD STIMUS TOKEN 5' TO THE S-R OBJECT.
SORT THE S-R OBJECT TO TERMINAL. IN NET.

 

 
    

 

  

EXIT HITH NO
RESPONSE .

 

 

 

 

 

 

 

 

25

DOES TERMINAL (S-R)‘ \ "0

MREADV HAVE AN IMGE? / 27 ' yes
yes HAS THE MEMRV

DRIH TURNED?

26 r

OISCRIMINATE THE M OBJECT L
GROHING ITS IMAGE IN THE NET. r-

 

 

 

Figure 12. flowchart of F20. the Paired-Associate Learning Routine
of SP”! III (continued).

 

 

DOES '

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

   
      
  
    

 

 

 
 

 

 

28
( C ) MES THE 54! IMAGE HAVE no
- '\ A RESPONSE COMPONENT?
yes
29
DOES THE RESPONSE COMPONENT
NAME A TERMINAL IN THE NET? '
yes no
so .
SORT THE RESPONSE SUBOBJECT
, OF THE S-R IMAGE To A TERM-
INAL IN THE NET.
32
OUTPUT RESPONSE AND Is AN IMAGE IN
TALLV TIME. TERMINAL R'?
33

EXIT HITH NO
RESPONSE.

Figure l2. Flowchart of F20. the Paired-Associate Learning Routine
of EPAM III (continued).

 

 

"1"?"
-&v}y——.—

 

7O

 

  

WHAT IS
THE TYPE

OF
OBJECT?

 
      
    
  
 

  

letter

K8

 
 

S-R pair

  
 
 
  

   

syllable
WHAT 15

FIRST
MEMBER OF
S-R PAIR

   

 

 
   
 

WHAT IS

 
   
   
 

 

FIRST
K8 LETTER OF
SYLLABLE K9
I
HIM' ,’ CAT'
I
I
’ (CAT-DOG)‘
S a CAT'
R . DOG'

 

   

HHAT IS
SECOND

LETTER OF
SYLLABLE

 
 

K8

 
   

K9

Discrimination Net of Fig. ll agter Trial l.
6

Figure l3.
Adapted from Hynn, 1966. p. 7 .

 

letter

K9

K9

CAT'

Figure 14.

71

     
     
   

  

HHAT IS
THE TYPE

OF
OBJECT?

syllable

 

   
   
      

WHAT IS
FIRST
LETTER OF
SYLLABLE

 

VHAT Is
SECOND

LETTER OF
SYLLABLE KB

A' \ U'

 

   
       
   

WHAT IS
THIRD
LETTER OF
SYLLABLE

S-R pair

 
 
   
     
   

WHAT IS

FIRST
MEMBER OF

K9 S-R PAIR K8
I
/

HIM' ,1 CAT'
I
,/ (CAT-MAN)‘

 

’ s a CAT'
R=mw

Discrimination Net Of Fig. l3 after Trial 2.
(Adapted from Hynn, 1966, p. 67).

 

 

 

 

   

 

HHAT IS
FIRST
MEMBER OF K9
S-R PAIR

 

)l

 

 

 

' B12551]

 

Figure l5. Discrimination Net of Fig. 14 after Trial 3.
(Adapted from Hynn, 1966. p. 67).

 

 

 

 

S-R pair
NHAT IS
FIRST
MEMBER OF
NIM', S-R PAIR K9
// O
I’ CAT KB
I
(CAT-DOG)‘ (CAB-MAN)‘
S - CAT' 5 - CAB'
R - 006' R - MAN'

 

 

 

Figure l6. Discrimination Net of Fig. l5 after Trial 4.
(Adapted from Mynn. l966. p. 67.

 

 

' 7 men. _.

 

 

73

EPAM III accounts for most of the same phenomena as EPAM II:
stimulus and response generalization, oscillation and retroactive in-
hibition, forgetting (fincugh in EPAM III information can be permanently
lost from the discrimination net -- q.v. step 17, Fig. 12 —- which does
not occur in EPAM II), effects of similarity among list items, and others.
Neither model accounts for proactive inhibition , backward association , and
free recall. EPAM III (but not EPAM II) contributes to an understanding
of the role of meaningfulness and familiarity in verbal learning (q.v.
Simon 8 Feigenbaum, 1961+). And, EPAM III can learn lists in which the
same item occurs more than mce. For the purposes of this paper it is not
necessary to evaluate the EPAM models further. A more thorough examination
is available in Feigenbaum and Simon (1962a) and Wynn (1966). ‘Ihe reason
for identifying the phenomena "accounted for" by these models is simply to
lend weight to their credibility as models.
new:

The weaknesses of the EPAM models noted above, along with some the-
oretical and empirical considerations, led Wynn (1966) to develop a mod-
ified version of EPAM, called WEPAM (Wynn's Elementary Perceiver and
Memorizer) . WEPAM uses as its main structure a discrimination net similar
to that of EPAM III. The net differs from earlier models in several
important ways:

a) multiple representations (images) of the same objects in

memory, b) multiple retrieval pathways —- both divergent and

convergent -- to these images , c) multiple responses associated

with each stimulus, and d) processes by whid'I the retrieval

structure in early stages of learning incorporates redmdant

information which is in part later eliminated in the interests

of more efficient retrieval. (Wynn, 1966, p. 138)

In addition, WEPAM employs a position testing node as well as the attri-
bute and subobject nodes used in EPAM III. Position nodes identify the

 

   

 

 

71+

location of a letter in a syllable. Figure 17 shows a WEPAM net early in
P-A learning. In the pairs' portion of the net, there is an instance of
a stimulus item (VEC) associated with two possible responses (IAJ, GIW).
In the syllables' portion of the net there are three position testing
nodes (P, L, G). Figure 18 shms the syllables' portion of the same net
later in the learning situation. In this Figure there are examples of
multiple paths to the same object (e. g. DAX, GIW) and multiple representa-
tions of the same object (e.g. VAF, IGP).

Three techniques employed in WEPAM produce these multiple paths and
multiple images. First of all a noticing order (as in EPAM II) is used
separately for syllables, S—R pairs , and letters. Secondly, the net can
be made more efficient by bypassing redundant nodes which test, one after
the other, the same attribute or the same position. And, thirdly branch
recruitment is possible. Branch recruitment is a process whereby a branch
in the net is duplicated at the node below. This may occur depending upon
initial parameter conditions, the current N.O. , and time remaining before
the simulated memory drum turns.

The general learning processes of the WEPAM model will not be discus—
sed because of their close similarity to those in EPAM III. It is import:—
ant, however, to further study WEPAM'S use of multiple responses stored
with stimulus items in the pairs' portion of the net because of its obvious
similarity with what is needed in a model of free association behavior.

In its most complex form, the image stored in the terminal node in
the pairs' portion of the net consists of (l) a property set identifying
d'Iaracteristics of the S—R pair, (2) a stimulus token with its (3) associ-
ated respcnse list, and (H) a response error list. When a dummy S—R pair

(i.e. S'—_) is sorted to a terminal in the pairs' portion of the net, an

 

'C.‘ '* my,” T‘L‘-_

 

J

 

75

 

 

.uu>

 

 

_zHon a
.i<>u m

 

 

.i<>

.HHHH .o .ooaH .ccsz sore oooaauav
.umz z<aw3 HoochH a we coHuioa .HH mesoHi

 

E .30: axe: o: 32 5H5:
masocuen some xios massum LcHaucaHLH
.mHLoHH>m a cH LouumH oi.o :oHuHmoa
HHHHcouH Ho. 4. av mono: Logos .o. 2 >5 uoH
-HHuoam :oHHHmoa :H miouuwH meucouH oH- H
mweoc .Hw: Ho coHuLoa .moHanHzm :H .>ux

a .o g
a .u—

34

 

 

 

55
.a _>

.4 .>

..IWI... .
.fl. . .

oHaHHHHo Hos zaau
ma uz<m

zomhmoa
LH3 m-m LouuoH .mamHHmH

 

 

 

 

 

 

 

 

 

 

.22 .a .82 .5»: .8... Bass
.acHHLooH Looas Looca HH .aHu Ho Ho: coHHacHEHEooHo Ho coHoLoa .ooHHHHHHw .mH oaaoHu

 

  

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

roe .H I ,
as. .23 a: sea
L
a as gs;
a u .<
.>
s. HH .H. .a .u a ._
. N . < .< .J . .
. . a.
6 n
7 u H a .u
.>
._ a
.1- m .3
.n— .5 .U
s. . m
.w .u .6
.<
.3 P . ..O
i 321;.“ ll
3:2. «A 233—

 

 

 

[@l

 

/|\ 1.4 11 ulus c ...I- (N III she ass 5\U\

 

 

 

 

77

an attempt is made to match 8' with S". If they match the first response
on the response list that is not associated with an error flag becomes the
first response camdidate. If all responses in the list are marked with an
error flag, the last one examined becomes the candidate.

When the memory dnmn turns and presents the correct response, WEPAM
may use one of several possible procedures to (1) add a response token to
the response list, (2) augment an existing response token on the list,

(3) change the value of a responses' error flag, and (1+) move a response
higher on the response list. All of these techniques can have a profound
effect on later learning trials.

The WEPAM model simulates some of the same phenomena of verbal learn—
ing as the EPAM models do: types and frequency of response errors, stim-
ulus and response generalization, oscillation and forgetting. Forgetting
does not involve a permanent loss of information (as in EPAM III); rather
information may be temporarily misplaced due to retroactive inhibition (as
in EPAM II) or permanently misplaced (i.e. a node is stranded) due to by-
passing of nodes.

In WEPAM most of the explained phenomena occur because of micropro—
cesses similar to, but not exactly the same as those in the EPAM models.
WEPAM also simulates the effects of overlearning (errors appear after items
seem to be learned), backward association (R—S), and stimulus redintegnation
(an incomplete stimulus abject may elicit the correct response). These
phenomena are not accounted for by the EPAM models. WEPAM as well as EPAM
fails in simulating proactive inhibition.

SAL I-III:
A second IPM based on an EPAM—like discrimination net and processes

was developed by Hintzman (1968). He called his model SAL (for Stimulus

no a. 3...... :3} ~- '~, 3

 

 

 

78

and Association Learner). It is not necessary to discuss SAL at great
length here , because of its structural and functional similarity with
the EPAM and WEPAM models. Therefore, what follows is concerned mainly

with important differences between the SAL models and the earlier ones .

Despite the common assumption of the discrimination net ,
SAL differs from EPAM in several respects. First, learning
in SAL is a stochastic process, while in EPAM it is deter-
ministic. If an investigator gives EPAM a list to learn,
erases the memory and then presents the same list again , he
will obtain two identical (or nearly identical) protocols
. . . . 'Ihe SAL model, in contrast, is governed by stochastic
processes , and can generate any number of unique protocols for
the learning of a given list. . . . It should be mentioned
here that stochastic processes are used in SAL only to facil-
itate the derivation of prediction. They are intended as
statements of ignorance , rather than assertions that learning
is basically probabilistic.

Second, in SAL all processes which are not necessary in
order to do rmning simulations of PA learning have been elim-
inated. 'Macroprocesses,‘ such as those in EPAM concerned with
allocation of processing effort, have been greatly simplified.
Also, SAL does not make Lee of 'stimulie images' or of a scan
for differences between the image and the presented stimulus as
does EPAM. It is hoped that, since there are fewer postulated
processes in SAL, it will be easier to identify specific pro—
cesses or combinations of processes with specific resulting
predictions . 'Ihus , it should be easier to understand why the
model makes a correct or incorrect prediction, and to make
appropriate changes when needed.

'Ihird, SAL uses the discrimination net only for stimulus dis—
crimination learning, while EPAM uses it for both stiImIlus learn-
ing and response integration. Accordingly , the 'task environ-
ment' of SAL consists only of lists of trigram-digit pairs,
where the responses are already well known, and only the stimuli
are unfamiliar. The purpose of this restriction is simple. It
is felt that if stimulus discrimination learning is to be under-
stood, it should be isolated from possible confounding processes ,
such as those cmcerned with response integration, and so on.
(Hintzman, 1968, pp. 124-125)

SAL exists as three hig'Ily related IPMS. The later versions are more
complex than the earlier ones and include, for the most part, all processes

of iiie earlier mes. In SAL I discrimination learning differs from that in

 

 

...L

A.~I

 

 

79

EPAM in two major ways. First, there is no N.0. which "learns" (i.e.
changes in N.O. based upon experience). Rather the N.O. is fixed —-
always being the first, then the second, then the third letter of the
stimulus trigram. Second, there are probabilistic processes. When an
error occurs in discrimination learning a new test node is added to the
net with probability a. If an error occurs and no new test node has

been added (probability l—a), then the correct response replaces the old
response with probability _13. The probabilities a and b are parameters of
the model and are initially set by the experimenter before running the
simulation.

SAL I simulates stimulus generalization, oscillation, perserverance
(same incorrect response given to the same stimulus item over several
trials), effect of stimulus similarity and other phenomena. The model
fails to handle retroactive inhibition. Also since the model does not
attempt to simulate response processing there are no failures to respond
to a stimulus object.

SAL II was developed mainly to handle retroactive inhibition. In
SAL II learning can occur after a correct response (as well as after an
incorrect response) with probability 3. g is a parameter between 0 and a.
After a correct response a new test node is added to the net with proba—
bility 9. Below this test node is an empty terminal. In later trials, if
the stimulus is sorted to a blank terminal, the model responds with any
item on the list (randomly) and stores the reinforced response in the
empty terminal. SAL II simulates the effects of overlearning and retro—

active inhibition.

 

 

 

 

80

In SAL III more than one response may be associated with a stimulus
item (as in WEPAM). This modification of the model was thought to be
necessary if the model were to account for proactive inhibition and mod-
ified free recall. In SAL III responses to a stimulus item are stored in
a ptsh down stack (PIE). In a PIE new responses to be associated with a
stimulus item are added to the tOp of the stack; older items are pushed
down one level. Response items at the top of a stack are more available
as responses than items lower in the stack. Thus, the PDS is a simple
method for making response availability a function of recency. Incorpor-
ated into SAL III is a short-term memory process which functions to move
all items in a FIB up one level with an a priori probability g. By this
procedure newer items in short—term memory (i.e. at the top of a PDS) can
become permanently forgotten. SAL III can simulate the difference usually
found between two methods of measuring retention: recall and recognition.
That is , the model presents a mechanism which produces higher recognition
scores than recall scores. Also, the model can simulate proactive inhi-
bition and can explain some of the empirical research in this area.

As an overall assessment, SAL and EPAM models,

account for oscillation , stimulus generalization , retroactive

interference, and the effects of stimulus similarity on list

difficulty. EPAM contains assumptions not present in SAL,

which make it applicable to problems of serial learning,

response integration , and presentation rate , and which allow

it to predict negative transfer. At the same time, SAL is

able to simulate some phenomena that present versions of

EPAM cannot, mainly through the use of overlearning assump-

tions (SAL II) and the storage of multiple associations (SAL

III). Althougi all subprocesses in SAL are all-or-ncne, it

is cmsistent with a number of facts (such as the effects of

overlearning on retention) which have always seemed to prove

that an incremental habit strength notion was needed.
(Hintzman, 1968, p. 157)

   

 

81

Saver:

In this section of chapter 2, five IPMS were presented. All of
these models siImIlate a subject (or group of subjects) and the experi—
mental conditions in a verbal learning situation. EPAM I consists of
fomr macroprocesses. It simulates the serial position effect and the
von Restorff. effect (i.e. changes in serial position curve due to unusual
items in the list) by hypothesizing that the list item learned on a trial ,4..-
is chosen from a subset of items located at anchor points . EPAM II adds I
one plausible set of microprocesses to the above structure. EPAM II (and
the EPAM III, WEPAM, and SAL models) uses as its main structure the dis-
crimination net. Coded input stimulus items are sorted through the net
until a terminal is reached. The terminal may be empty (and no response
is made) or it may contain an image. The stored image is matched against
the input stimulus. If they match the stimulus is said to be recognized.
Once recognized an associated cue to the correct response (if one exists)
is sorted through the net to find and produce a coded response. Whenever
an error occurs and processing time remains, learning processes are brought
in to change the structure and/Or content of the net .

EPAM III extends EPAM II by including within the net nodes for letters
and stimulus-response pairs as well as for syllables. WEPAM is an EPAM
III-like structure which builds multiple representations of objects in nets
and multiple retrieval pathways to these objects . None of the EPAM models
permits this. SAL is an EPAM II-like structure. It is concerned solely
with stimulus learning. SAL is the only one of these models which incor—

porates stochastic processes to a major extent .

 

82

While the models differ from one another in their specific set of
microprocesses , they all use the discrimination net as the primary form
of memory organization. Taken together, the number of phenomena "explain-
ed" by these models is extraordinary. This in itself, supports the posi-
tion that further examination of this type of model is warranted. There
is a second reason for developing an EPAM-like model of free—association
behavior. Namely, the simplest most straightforward position to take is I
that all verbal behavior phenomena can ultimately be explained with one :1
theory or one IPM. And 50, first models of free association behavior
ought to try to fit into existing models of verbal behavior. Finally, it
should be noted that EPAM can simulate internal mediated responses (q.v.
Feigenbaum 8 Simon, 1963b) and both WEPAM and SAL III incorporate the use
of multiple responses associated with a stimulus item. These processes
may turn out to be necessary in models of free association behavior. Such

models are presented in the next chapter.

 

 

 

 

CHAPTER III

In this chapter a family of related information processing models
(IPI’B) is described. These IPMs offer an approach to the cognitive pro-
cesses within a respondant engaged in a typical continuous free assoc-
iaticn (C—F-A) task. There are two subdivisions to this chapter. In the
first, the scope of the problem is discussed. This includes major assmrp—
tions characteristics of C—F—A behavior, and requirements of the models.

The second division describes the models —- their structure and operation.

3% of the Problem

As described in chapter 1 a C—F—A task requires that a subject (S)
give a series of responses to a specified stimulus item. The stimulus
item can be a word, nonsense syllable, dysyllable, etc. The usual require-
ments governing the responses are that the stimulus item cannot be given
and no item can be given more than once. The task is completed when either
a given number of responses is given or a set amount of time has elapsed.
E usually records the actual responses and the number of responses given.
He may also record the time intervals between all responses.

A final IPM of C—F-A behavior should produce associates to a stimulus
item from a cognitive memory structure. To account for the response learn-
ing phase, the association phase, and the response giving phase a model
must specify the operations which build and modify a memory, and identify
procedures for the association and retrieval of responses from the memory.

83

 

       

 

 

8‘4

In addition, the model should include deterministic procedures which
hypothesize the mechanisms accounting for those variables found to be
viable to an understanding of C—F—A behavior. The main variables to be
included in a final model are those underlying response strength, response
integration, response availability and response elicitation. For reasons
given earlier, this first model is mainly concerned with item availability
(I—AV).

The last part of chapter 2 describes the EPAM, WEPAM, and SAL models

 

of verbal learning. Upon examination these models were found to be very
general and highly heuristic. The discrimination net underlying the models
and the procedures described for the modification of the net can be applied
to a C—F—A model. To review briefly, these models posit operations which,
at face value, seem to simulate or account for mediation and association
of verbal units (through response cue codes); hierarchical associations

(by means of the letters'-syllables'-pairs' portion of the EPAM III and
WEPAM nets); and multiple responses associated with stimulus items (in
WEPAM and SAL III). These operations are needed in a C—F—A model. It is
therefore reasonable to base a first C—F—A model on the EPAM—type discrim-
ination net memory.

There are two reasons why the EPAM—WEPAM—SAL models should not be
directly employed as a complete C—F—A model. A major wealciess in the ver-
bal learning models is the lack of parallel processing. It is difficult
to clearly define and separate serial and parallel operations (q.v. Minsky
8 Papert, 1969). The notion of parallel operations may seem to be in con-
flict with the character of general purpose machines which operate sequen—

tially. The conflict is due to an erroneous identification of a machine

 

 

 

85

with its physical properties. Rather machines must be considered as a
combination of hard and software . Though a machine moves through a pro-
gram (IPM) sequentially, it is possible to simulate parallel operations by
means of a hierarchical program structure (e.g. at a given time to process
A then B then C, but treat that block of operations as if they occurred
simultaneOLBly).

Wynn's (1966) review of the area indicates that humans operate in a
parallel mode for at least some of the cognitive processes -- including
sensation , perception , attention and association. Of course the hierar—
chical organization of these processes may Operate serially. While some
IPMS attempt to include parallel processing (e.g. Reitman, Grove, 8 Shoup,
1961+; Selfridge 8 Neisser, 1963) none of the IPMS previously described in-
clude them. This, "failure to provide for parallel processing in any re-
spects is probably [their] most serious weakness" (Wynn, 1966, p. 210).

The second reason for not using the EPAM—type models as complete IPMS
of C-F-A has to do with some important differences between verbal learning
experiments and association experiments: (1) In verbal learning tasks the
stimulus and responses must be learned and integrated as part of the task.
An association experiment deals with the elicitation of previously learned
and integrated verbal units. Thus, _E_ can treat the S in a learning experL
iment as if he had a limited Imam memory; in C—F—A the S's memory is not
limited as all his past learning can be teed. Related to this is E's at-
tempt to amtrol the learning environment and stimuli in the former case
while he is unable to control or even guess them in the latter case. (2)
In verbal learning situations it is important to differentiate between cor—
rect and incorrect responses; while in free association the distinction is

not applicable. (3) In the typical verbal learning study the _S__ must give

 

 

 

 

86

one response to each stiTmIlus. In C-F-A many responses are called for.
(H) The time limit for responding is relatively short in a learning para—
digm, while it is absent or much longer in the association paradigm.

A central assumption underlying any theory is consistency. This mod-
el of C-F-A behavior is no exception. For this model it is necessary to
assume that the procedures which generate free associates for one individ—
ual are identical with the procedures operating within another individual.

For this model this means that all individuals have a similar memory struc—

 

ture —- that of a generalized discrimination net. Observed differences be-
tween individuals engaged in C-F—A behavior must be attributed to differ—
ences in the content of the net, relationships among the components of the
net, and different values of the various parameters or property sets attach—
ed to parts of the net. Once these idiosyncrasies have been determined for
an individual it should be possible to treat his C-F—A behavior the same
as that of other individuals. The corollary to this assumption is that the
model is consistent within an individual over time. These assumptions are
not particularly unreasonable and there should be little surprise that there
is some related evidence supporting them (e.g. Jenkins, 1960; Cofer, 1958).

In sum there is a series of EPAM—type IPMS which can account for encod—
ing—decoding behaviors; memory structure and growth; and those operations
which associate stimulus and response items in a simulated learning experi—
ment. This type of model must be modified to account for parallel process-
ing, the essential characteristics of a C-F-A experiment, and the stable
cognitive functions within a S engaged in a C—F—A experiment.

The next section of this chapter presents such a model. Because this

model is a first attempt and because parts of the model will be simulated

 

 

    
  

‘ e

, I
l. Q
I a

_. A—___-q—.--—..——.—

 
  
   

   
 

87

by hand rather than by machine, certain simplifications are needed. Spec—
ifically, the model will be solely concerned with the retrieval of stored
words from a generalized discrimination net. For the most part an EPAM-
WEPAM—SAL type of model will be used to handle all parts of C-F—A other
than response retrievals. The model can be thought of as being similar
to a counterpart of the SAL IPMS which specify stimulus and association
learning and assume response learning and integration. That is , the C—F—A
model will be concerned with re8ponse retrieval while the earlier models
will assume responsibility for stimulus and response learning, their
association and basic net structure.

The adoption of the earlier models to handle these chores is not
totally applicable. It is assumed (and not documented bGICW) that they
can be simply modified to (1) deal with natural language rather than non-
sense syllables, and (2) generate values for variables - such as I—AV --
needed by the model. Correspondingly, the model will have parallel opera—

tians described for the response retrieval phase only.

An IPM of C- F-A Behalior

 

The C-F—A model described in this section is exclusively concerned
with the retrieval and evocation of responses to a stimulus object from a
verbal memory. The model does not describe specifically had such a memory
is built. It does, however, require that the memory be of a certain form.
The organization of the discrimination net (memory) will be discussed first,
and then the routines which control retrieval and evocation will be describ—
ed. Table ll lists the more frequent abbreviations used in the ensuing de—

scription.

 

 

 

 

 

88

Table 1+. SuTma_ry of Abbreviations Used to Describe the C—F—A Model

 

 

 

Abbreviation Description

CDI‘ Current Date-Time. Number of siJmIlated time
units from beginning of processing.

DT Date—Time. Any specified simulated time period.

EN Exit Number of responses. A parameter. If NR
exceeds EN processing stops.

EI‘ Exit Time units. A parameter. If CDT exceeds EI‘
processing stops.

IRT InterResponse Time. One IRT is attached to each
active list of responses. IRT counts elapsed DI‘S
between responses from the same list (cf. Y).

MS Memory Size. A parameter which specifies had
many items can be put into short term memory.

NM Number of Markers. NM equals the number of SMs
plus the number of RMs active during the CDT.

NMM Number of Markers Maximum. A parameter. If NM
equals NMM no additional markers can be initiated.

NR Number of Responses evoked.

PDS Push Dam Stack.

RM Response Marker.

SM Stimulus Marker.

STM Short Term Memory. STM can only hold MS items.
If additional item is added to the top of STM an
item is dropped ("forgotten") from the bottom.

TM Time Marker. TM tracks the processing in the Time
Executive Routine.

Vi A series of 3 parameters. .They specify the increase
in an item's I—AV due to different types of processrng.

Y A parameter. Whenever any IRT equals Y time units
another response from the list associated with the
IR'I‘ starts its processing.

9i A series of 2 parameters. The 9's are thresholds

 

against which the I—AV of responses are compared.

 

 

 

89

The discrimination net is quite similar to the nets presented in
chapter 2. There are two divisions to the net: the letters' portion
and the units' portion. The pairs' portion of the net is omitted be-
cause the C-F—A model represents S—R associations in the units' portion.

The letters' portion is identical to that part of the EPAM III net -—
consisting of attribute testing nodes, empty and filled terminals, and
branches (including K8 and K9). The units' portion consists of attribute
testing nodes, subobject nodes, and terminals. To that extent it is sim-
ilar to the other nets. The major difference occurs at the terminals.

If the only type of verbal unit stored in memory is an English word, then
each word may occur many times throughout the net (as in WEPAM) but only
once as a first item in a terminal. Terminals may contain (and usually will
contain) more than one word. The first word in a terminal can be considered
loosely as a stimulus word. All other words in the terminal may be thought
of as cue codes for potential responses to the first word.

Words are stored in a terminal in a push dam stack (PDS) similar to
those in SAL. Most recent associates to the first word are higher in the
stack than less recent associates. An example of a terminal in the units'
portion of the net is given in Table 5. That terminal is described in an
annotated form of IPL-V (Newell, 9: iii. , 196l+) -- a programming language par-
ticularly suited for this type of model. The middle portion of table 5 con—
tains the attached property set for the first object. The values in the
property set are tested when the first object, DOG, is sorted through the
net or is being used as a possible response. The response list in Table 5
indicates that IDS is associated with five possible responses. The responses
are stored in a PDS of finite size. That is, the PDS simulates forgetting

of an association from a long term memory. Whenever a more recent associate

 

 

.I-

 

 

 

Table 5 .

A Terminal in the Units' Part of the Discrimination Net——
An Example of Coding in IPL—V.

 

 

27

90

91

92

 

(name of terminal) 90 (attached property set)

815 (stimilus object DOG)
91 (attached response list)

(property set) 0

D13 (what is mode of first object?)

V22 (it is printed.)

D63 (what is U]? of first object?)

V69 (83 time units.)

D79 (what is I—AV of first object?)

V8” (22)

D91 (had many time units are in IRT for response list
associated with first object?)

V92 (IRT has not yet been set for this response list.)

(response list) 92 (attached use list)

R13 (cue code of most recent response, CAT)

R17 (cue code of next most recent response, PUPPY)
R13 (cue code of next response, CAT)

R914 (cue code of next response, ANIMAL)

R29 (cue code of least recent response, HORSE)

(use list) 0

U13 (has R13 been used as a response?)
D2 (no.)

U17 (has R17 been used as a response?)
D2 (no.)

U13 (has R13 been used as a response?)
D2 (no.)

man (has R99 been used as a response?)
D2 (no.)

U29 (has R29 been used as a response?)
D2 (no.)

 

 

 

 
 
   

‘
\

 
 

‘ .
‘ ‘

  
 
    
 
 

 

 

91

to the first object is learned and added to the top of the response list
(i.e. where CAT is no») all lower items are pushed down one space, and
if all spaces were taken in the PDS, the lease recent response would be
ptrshed off of the bottom and lost.

In order for a word to be recognized in the memory its coded form
must be sorted through the net until it reaclnes a terminal with a first
object. (It is plausible to assume that sorting occurs with few errors
because responses given in C—F—A experiments are highly learned and well
integrated). If the word being sorted is a stiImIlus word then the terminal
reached may contain cue codes for possible responses to that stimulus. In
Table 5 R13 represents a cue code for a possible response to stimulus word
DOG. In order for tlne ote code to be recognized, it is sorted through the
net LmItil it reaches a terminal containing CAT as the first object. The
use of cue codes in C—F—A is much closer to the cue codes of EPAM II than
the cue tokens of EPAM III because they do not name a terminal —- rather
they must be sorted through the net.

Thus, every word is stored in a terminal as a first object. A stim—
ulus word and a response word must find their first objects if they are to
be recognized. If a stiImilus is sorted to the proper terminal, then the
cue codes stored at that terminal become available as possible responses.
Many words are also stored in the response lists of different terminals in
the form of cue codes.

When sorting an object througln the net it is necessary to distinguish
between potential stimuli and potential responses. Both are being sorted
to a terminal whose first image (hopefully) matches the coded object.

Once recognized, havever, the two kinds of objects are treated differently.

Potential stimuli will not have the opportimity of being evoked as responses;

 

 

 

they initiate their response list as a time-ordered set of potential re—
sponses. Potential responses, when recognized, are immediately processed
in the response giving phase of C—F—A. In order to identify objects being
sorted through the net, an SM is used to mark the position of a potential
stimulus and an RM is med to mark the position of a potential response.

At this time, the processing of the C-F-A model is controlled by
six routines. Figure 19 share the relationships among these routines.

It may be helpful to consider the C—F—A model as a type of board game with
three different kinds of markers ("men") moving around the board. There
is one TM which keeps track of the Time Executive Routine. The TM only
moves within this routine. There are zero or more SMs and RMs subject to
the constraint that the number of SMs plus the number of RMs cannot be
greater than NMM. SMs keep track of potential stimuli and RMs follafl the
processing of potential responses. SMs and RM; move about through all six
routines.

Generally speaking, the C—F-A model takes an encoded stimulus object
and sorts it to a terminal where it is compared with the first object for
recognition. Upon recognition, the associated words in that terminal are
all popped out -— most recent first. A specific number of time units, Y,
must elapse between the emission of consecutive potential responses from
each response list. Each potential response emitted is a cue code which
must be sorted througln the net Lnntil recognized. When it is recognized the
property set attached to that terminal is examined to see if the response's
I-AV is sufficient for a response to be evoked. Evocation depends upon I—AV
elapsed time, the contents of STM, the values of the thresholds (9i) , and

other factors. Some potential responses which cannot be evoked are "strong"

 

 

 

 

 

TIME EXECUTIVE ROUTIer

 

 

I
[_MACROPROCESSING ROUTINE

/

 

 

 

 

 

 

 

 

 

 

 

STIMULUS RESPONSE
SORTING GIVING
ROUTINE ROUTINE
NET SORTING
ROUTINE

FINDING TERMINAL
ROUTINE

Figure 19. Interrelations Among Routines in C-F-A.

 

 

91+

enough to be treated as mediated stimuli starting the entire process
over again. That is, while there is only one nominal stimulus in an ex-
periment, it is likely that there will be several functional stimuli.

The Time Executive Routine (T-O) controls all other routines and
functions mainly to start and stop a simulated C—F—A experiment and
control the parallel processing. Figure 20 gives the fladchart of T—O.
Starting a new experiment the TM moves through steps T—l, T-2, and T—3.

An experiment is stopped at T—16 when either condition in T—7 is met or
no RMs or SMs exist. Throughout the bulk of the experiment, the TM con—-
trols the movement of the RMs and SMs by cycling through T—H, T-S and

T—6. T-5 is the essence of parallel processing in C—F—A. All markers
(RMs and SMs) have a DI‘ attached to them. All routines except T—O and the
Macroprocessing Routine consist of timed processes. T—5 moves all markers
one time Lmit. Within each time unit markers are moved in the order of
their attached DTs .-- earliest first.

T-O also increments the IRIS of the active response lists. Whenever
an IRT exceeds Y a new RM is created (T-l2) for processing the next unused
potential response in that response list.

The Macroprocessing Routine (M-O) controls all other routines (except
T—O). While both M-0 and T-O control all of the timed routines it was de—
sirable and necessary to separate this control into two routines. It was
desirable to have one routine dealing solely with parallel processing. It
was necessary to separate the routines as they are hierarchically organized:
no marker leaving T—O (at T—3 and T—13) ever returns to T-O. This makes it
impossible to construct T—O and M-0 as one routine.

M-O links the Stimulus Sorting Routine and the Response Giving Routine
It also counts the number of responses given in the experiment. Figure 21
gives a flowchart of M—O.

 

 

 

 

95

 

 

 

 

O:
:22sz 2

no).
:3 e: :2
-cH-H

.huzz wimp mzo
>m emu m>~ku<
zu<u mm<m¢uz~

auh

   

 

 

 

 

          

 

  

«hum ¢<4u achm <
3 «O 55.59 3

 

 

   

~mmmz¢amu¢ ac
hm“; maze zo mumzcawwz
ommaza >z< uxuzk m¢<

 

 
 

was

0:

     

u Ho.o:Hu=o¢ o>Hu=uoxm uEHH Ho usaguoni .o~ ogamHi

 

 

 

.enz

k< xx has
cut 44<u

 

 

 

 

 

 

hum one 92¢ hou zu<hh<
.mumzoamux ac emu; maze
«on :8 3m: < mh<nhuz~

 

 

«Huh

owe
.H .H> is .m:
5.5 “was:
-52.. mm 5.5.
225;:qu

—-h

 

 

 

 

 

 

.zm o» hog =u<hh<
.Sm z< mh<uxu
.humano m=4=:~hm
amoouzm z~<hmc

 

 

 

     

.tha mtuh mzo
>m ecu um<u¢uz~

on»

  

 

.>Hzooz<¢ rush wzoz<
wmcozu .ha 4<uuhzmau
m<= xmx¢¢x uzo z<zh mxo:
.hmg—u u>c£ ha hwuH
Inz<u :hH: m¢u3¢<x .huz:
u:_h mzo maux¢<x 44¢ ”>9:

 

 

mo

 

 

Nah

 

 

. H-..
E E 5..
a-.. s3
3

 

 

 

 

~m¢ux¢¢x u>~hu<
>z< uzmzb m¢<

«a»

 

 

 

 

 

.<..h_-u .3 0523 2338323: mo €228:

 

. G 0.53..—

 

 

 

 

ptx 3:... 1.53
35.. 322mm.—
.Ems H2 2:

 

 

 

 

 

 

 

.7: .5 H! .28
c4. ...20
.338 9.35

3:23: .6“.
Snags—cocoa:

 

   

 

 

 

 

 

T:

 

     

 

 

 

 

 

   

 

    

   

 

i.\|§s.¢s. can: 5.5K
/ e: 2.: 2 55: S
2-..
use
as
so
a-..
ti ..E U
es a 2.... a 5
moses. we: so .58..
a sea .2 3:... H-.. 3 2 «use 8 .2 may: 2
were. .2. z-..
2-.. H-..
iii—3.9.35, 8 .7... :55.
gm... .5 2.: .5: a-.. a .n 3 :5
85982 a: means. .2 we: a as: 9:53
a: 8.. H... .a 33:... N-.. 3.5.5 as
manage:

 

 

 

no:

 

 

 

 

 

 

97

‘Ihe other four routines of C—F—A contain timed processes . In con-
structing the nodel it was necessary to decide upon the nunber of tine
units each process should take. The following arbitrary, but reasonable
rules were followed. (1) Exits from routines and calls to other rou-
tines take no time as they serve as links between processes. (2) Deci—
sions, in general, take nore time than simple processes. (3) Highly
practiced processing should occur faster than less frequently practiced
processing. Net sorting should be a major component of all uses of a
verbal nemory while giving free associates is only one use of this memory.
Therefore, the Net Sorting Routine should process markers much faster than
the other tined routines. Figures 22-25 outline the timed processes. The
number of time units required for processing is indicated at the lower
rigut corner of each component in those Figures. In order for a particu-
lar canponent to be processed, all time units for that conpcnent must be
completed.

Figure 22 gives the flowchart of the Stimulus Sorting Routine (S—O).
It is a simple routine designed to find the terminal whose first image
matches the encoded stimulus object provided that the terminal contains
at least one item in its response list. Since the goal of a C—F—A experi-
ment is to give responses to stimuli, it is of little use to recognize a
stimulus object which has no associated responses. Exactly how a terminal
could be constructed whid'x cmtains a first image but no associated res-
ponse list is beyond this nodal. One possibility is that processing time
"ran out" when the terminal was constructed.

'Ihe Response Giving Routine (R-O) is outlined in Figure 23. When an
RM is sorted to a terminal (i.e. a response is recogxized) the I—AV of the

re8ponse (first image) found in the property set attached to the terminal

7% _ “ i1

 

inane mo 2.333. action 33.5.3 mo £2930: .NN 953...

m\-r 8.58%

 

 

 

 

 

 

 

 

 

 

 

 

 

r 45:5.“—
653 ms. \ 2.
no... T-m

O I! . . . . 0 .p02 ._.< 2m he

a 7. 2 ..m 5.. c a-.. .35
9 on... .35 .25.. ..Gz
SEE... < .68 92...! _ . m-m . . .38....» 5 .58
. Vim pan

«0;. o:

 

 

 

 

 

 

 

2. we... .52.... e: 38 885.. 3.5.“: .. 2
31 «um:

n i n Ir
E E... 85982 \ as...
2. was... 38.3.. e... .23 2.: K55... < 2528 2......

 

 

 

 

 

 

99

is decreased (q.v. R—8). In actuality it makes sense to consider the
I—AVs of all first itens decreasing at each new CDI'. In practice it is
easier to renember when the value of I—AV for a particular first item was
last changed (its DI‘ -— see Table 5), and decrease the I—AV according to
the ratio DT/CDI‘. This is the procedure used in the ARGUS model of think-
ing (Reitnan, 1965; Reitman, Grove 8 Shoup, 1961+).

Once computed the new I-AV is conpared against a series of thresh-
olds, 6i. While there does not seem to be any direct evidence about this,
it seens reasonable that a greater I—AV is needed for an overt response
than for an internal ("unconscious") mediated response. 9i is the thresh—
old pareneter of I—AV which must be at least equalled if an overt res—
ponse is to be given. Similarly, this model assumes that an internal res-
ponse will occur only with an I—AV of sufficient strength (at least equal
to 92). Responses greater than or equal to 91 will be evoked if the res—
ponse is neither identical with the stimulus word nor given earlier as a
response. Both of these requirements are conditions of a typical C—F—A
experiment. A person, however, does not always remenber the words he has
already given as response. In the C—F—A nodel the STM which holds the
responses given is a PBS of finite length. 'Ihus, it is possible for the
sane response to be given nore than once.

Before continuing the description of the C—F—A model it is interest-
ing to compare the methocb used in WEPAM and C—F—A for dloosing among pos-
sible responses in a response list. Both nodels have an ordered list of
possible respcnse. They are ordered so that itens at the top of the list
are examined first. In WEPAM associated with each item in the list is an
error marker indicating whether or not that particular response had been

given in error previously. WEPAM produces as a response the first item

 

'1oo

 

.Eubb >m

 

 

 

   
     

 

  

.<-u-u mo meeuaom ucw>ww omeoammm mo ugoguzo—u .mm ogaowu

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

uu<z. .m... .o , <
><-. mm<w¢cmo
...«z. .m...
< mz.<.zou
:u.== smzu<me o:
.«z.z.u. < m. _
s-. u
..m..
, mmzoamug m.=.
go. .z.o. ...z.
z< a2... amaze.
..-z E 2.. 5.. e. 2 7.. ...s.
o-z ..<u _ N-“
...z =u=c.:. 2. exam
m-.
o:
n o: . m ...-. zasceze m...¢<:
.5. E 2.2!... 2 ..9. 588... .53
.au oz< .mzoamuz mos was m.=e zo “macaw“. sumszs
zu<eh< .aums .. m-“ .zs z<:. was: ”aux. m.
xe¢x .umzoamue .-g
ammsza ex“: ug<h
v.“

 

.Acm::_ueouv <-u-u mo oewpaom mew>vu mucoamom mo peasozopu .m~ cease;

 

   
   

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

     

N
. .
N .zem .o as. o. .m ma<mw.um¢.n
wu<=. em... ao< .o ><-. wm<u¢uz.
N.-. N.-.
e
o ..mzoamme heu>o
z< m< ma<=.
m.-. em... muaflwe“
N
..au as . .m..z= N. >. ma.:.
.1 ma<=. hm... hm... .o ><-. um<u¢uz. o:
m .o s .925 ...m 8 .8 65E. M
.2m z< o. z: m.:» maz<=u Npuuamo m:.=:..m
c.-. a.-. mu .. m. .o gen 2.
, mu<=. .mg.. m.
a.-.
N n no»
.m».z= > m N .
E was: .55... o: ..No A a: 2 mafia“ ”menu—Wham“
.o s.-. um¢m¢uz. N.-.
awn“ OF C no»
. N ‘---|L
NSA.$.2

    

 

  

O:

 

 

 

102

in the list not previously given in error. In C-F-A all items in the res—
ponse list are first of all discriminated (since criteria for evaluating
them as possible responses cannot be considered until the responses are
recognized) and then the corresponding RMs are moved through R—O to see
if they meet the necessary requirements for evocation. Responses are
popped up from a response list under the control of its IRT (q.v. T-12).
Exactly which response in the list should be popped up is controlled by
marking those responses already used (R—u). This mark is stored in the
terminal in the use list. This marking of used responses corresponds
directly with WEPAM's error marks.

Potential responses can be either candidates for evocation, mediated
response, or merely processed responses (whenever I-AV is less than 92).
The I-AV of each of these potential responses is raised according to Vi
before the RM leaves R—O (see R-12, R—l8, R—19). The I—AVs of responses
which are candidates for evocation are raised more than the I—AVs of med—
iated responses, which, in turn, are raised more than the I—AVs of proces-
sed respcmses. The work of Hormitz and his associates (1961+, 1966) strong-
ly supports this ordered raising of the I—AVs.

The Net Sorting Routine (N-O) is presented in Figure 2H. Understand—
ing NO is rather straightforward. First of all, it identifies the loca-
tion of a marker and then it takes appropriate action. The recursive
nature of this routine is evident at N-16 where the routine use: itself to
discriminate the letters of a word. It should be noted that the time units
needed for processing in N-O are given in tenths of time units.

Figure 25 outlines the Finding Terminal Routine (F—O). F—O is called
on two occasions: when N—O sorts a marker to an empty terminal and when

S-O finds a terminal with no associated response list. In a C—F—A experiment

 

.<-u-u mo m:.u=oz u:.u.om an: we «Lusuzopu .em o.=u.u

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

    

 
 

.. .mooz ...z.
o. =uz<¢m
s m.:. :o..o.
N. 3 z-..
.Naoz we. .< _
28...... .c 3....”
m.:. «a. 23°. .. .
0 mac: .xuz
z... =uz<¢. < m<= o. o. zuz<..
6H: .i. . . wox 32h
muz . . no». 072
3
m. N.
Nunez ”.2. .< ..
aN.mN. N.=.....< .
m... ..o .35. < “mega...”
..=.mmmuu=m u><= ...smo muss o: . 0-. N.<.
o-. «<3
.-z ”.2

en:

  

~mooz azuhmuh
uh=m~¢Fh< z<
h< aux¢<x m—

        
 

 

..-. .e ...«(x .s.
o-. ...u
i .32....

>hgxm-zoz < =z~u
T 2. ,

    

~4<z~zuub >hexu
z< h< zuxx<x m~

.p-a h<
hm: xmhzu

   

 

 

 

 

 

 

 

 

 

 

lOU

..smsc..=ou. <-.-. .0 o....o. c:...om ..z .0 ...guzo.. ..N 0.3...

 

         
     

N. No. ...eo. m<z ...a.o.=m
.1. .<z.ze.. .2. :..= .z<z N+ ..x.
..oz ...s.o.=. ..z. z... ...smomsm o.o

mpuz

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

      

 

.z.a<.. 1.2... < ...z. m. ..-z
3... 0.. 3.2
o . .
. ...z zes... ..zm.
. ...oz .x.z -z.. z.== ..-z .<
c. =.z<.. o-z o.z. .. .=.z.
.-x ..<. _ .2. ....... <.
. ON-z .ooz ..z. .< ...m..
...a.o.=. .z. ..<.
. ..-z
.2.
.. ...cz ...z N. N...... o. .z<>....
o. :uz¢¢., o. ...z ...a.o.=. .<
.-x ..<. a..... ...a.o.=. ..
..-z
N. ...s..§.. <
m» .< 5...... ..

 

~72

 

 

 

 

 

 

 

105
N

inn-U mo 9.333. 35.5: 2.2.5.... ..c «.223: .3 0.50....

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

8...... m\ 2...... 2... N . 8.8..
2... .5 Al .3 88s.... .... 9...... .o ..m
..s. 2...... 2. ... 88.. 2...... ms .85.... .... ...s.
3.. .2... e... .8... .5. i
8. . 2-. 2-.
N .38....
9...... -
8.... ...s.
a .-.,.
a-.-
. o: . no»
....<z .2.
...hmmewgwwfi... .85.... a... .s 8...... 88... _ N k ,
.5 2...... .8... .... .o ..m ..S .... .2... 8.... ..o . ..Sz 2......
...S ...o .... 3... 3.2 8...... .82 ..s 88.. .2. . 2.88.... ...s.
.... ...e... ..8... m-..
. E
«a... .2. 2.
n _ A a 23.8.. 8......
...on...wfl.......<mwmm...... N ...,...e. 8...... .c 83-? 5...... .... .o
.8... ..o ..a 8 35...... ...z .55.... . 3.. .8 5.3.5...
1 . «mi .5. .8... .5.
7. N-ml

 

 

 

106

empty ”terminals do not make sense as they do in a verbal learning study.
In learning studies an enpty terminal (due to inadequate learning) pro—
duces no response on a given trial. In a C—F—A situation the notion of
trials is irrelevant and reSponses are usually given (perhaps because
there is not a short time limit on response giving). Secondly, stimulus
items with no associated responses are of little use in C—F-A. In these
two circumstances F—O attenpts to find a successful terminal by choosing
among terminals close to the non-successful terminal.

This then is the basic C—F—A nodel. It consists of six related
routines: tine executive, macroprocessing, stimulus sorting, response
giving, net sorting, and terminal finding. For reasons noted in the
previous chapter a family of related IPMS is desirable. Major extensions
of the C-F-A nodel are described in chapter 5. Outlined below are pos-
sible minor extensions of the basic nodel -- all of which constitute the
family of IPMS.

Obviously, the first route to be taken is to change the values of
the various parameters. It seems reasonable that various judicious choices
of the parameters will cause the nodel to behave differently: perhaps to
such an extreme difference that the model could be considered to have
changed in kind. Sensitivity testing as described in chapter 2 is the
memod to be taken to explore these nodels.

Other nodels in the family could include one or nore of the follow—
ing: (1) Include with each letter terminal in the discrimination net
some measure of its I-AV and use these values to adjust the recognition
of letters . This alternative, however, does not seem to be too worth-
while since the hig‘u frequency of exposme to letters (as carpared with

words) should tend to raise their I-AV over any reasonable thresholds

 

'n....'-._..__~.

 

 

 

 

107

(q.v. Cofer, 1961). (2) Have the response list PDS function similar to
those in SAL III wherein over time the items in the stacks rise, losing
newer items because of short term memory. (3) Have the net built on
pronunciation rather than printing. Use of the oral—Aurel node implies

a phoneme—units division of the net. In this case, I—AV for each phoneme
might be useful. (u) Change R—O eliminating the need for 92. This means
that all RMs marking responses whose I—AVs are not greater than or equal
to 61 will be converted to SMs. This should greatly increase the activity

of markers througuout the experiment.

m

‘Ihis chapter specified some of the requirements of a general C—F—A
nodel. Some of these requirenents stemned from basic needs of nost IPMS
of cognitive processes -— such a parallel processing, and sone require-
ments arose from the differences between verbal learning and C—F—A ex-
perinental situations. The model described takes as input a coded stim—
ulus word and sorts it through a conpleted discrimination net to a ter-
minal. Associated responses in that terminal are popped-up one at a tine,
sorted through fire net for recognition and then tested for possible evoc-
ation. Recognized responses can be treated in one of three ways: (1) they
can be evoked; (2) they can be treated as internal stimulus objects; or
(3) they can be considered for either of the above -— but neither occurs.
The choice anong these possibilities depends in the main on the relative
location of a response in its response list, its DI‘, and its I—AV. The
last two of these are values attached to each first item in a terminal.

Other characteristics of the model include a parallel processing routine

 

 

 

 

108

(T—O) and an error correcting routine (F—O) which finds "successful"
terminals. The Chapter also briefly mentions some possible other'members
of the family of C—FLA.models.

Before the model can be critically appraised its functioning needs

to be tested. Chapter u gives it one of the necessary tests.

nu.- ~ mu...‘nml -

 

 

CHAPTERIV

This chapter describes a contrived experimental situation includ—
ing a simplified simulated S who operates according to the rules of the
C-F-A nodel presented in the previous chapter. In actuality the C-F—A
nodel operates on a hypothetical verbal memory under set parametric con-
ditions. The states of the system at various times (DI‘s) throughout the
simulation are outlined. The first part of this chapter discusses the
problems involved in simulating the experiment with the C-F-A nodel. The

remainder of the chapter describes the simulation itself.

Problems of the Simulation

 

Two problems with this simulation make its value less than what
could be theoretically desired. First there is the problem of the ver—
bal memory. The C—F—A model does not, at this time, build one. Since
one is needed to test the nodel it must be obtained from other sources.

A primary alternative would normally be either tabled listing of stimulus
words and their response lists (such as Deese, 1965), or the responses
given by real SS in a real C-F-A experiment.

Both of these procedures, however, implicitly assume that the res—
ponses evoked by SS in a C-F-A experiment are identical with the internal
response lists -- not in format or coding, but in terns of the actual

words and their order in the response list. There are several reasons to

109

 

 

 

 

110

believe the assumption to be false. For example, the directions given
the S may stipulate that no response may be given more than once. If
the assumption were true then the internal response lists in the S's ver—
bal memory would not contain any word more than once. Also, the work on
item availability (I-AV) and response strength indicates that only highly
available and strongly associated items in a response list will be evoked
rather than all of the items. Thirdly, Jung (1966) declares that the ac—
tual responses given in word association tests depend to a large extent
upon non-associative factors such as subject set.

Since the C—F—A model does not build a verbal memory and since lit—
tle faith can be placed in the results of C—F—A experiments, this simu-
lation will use an g hog hypothetical memory. This should cause little
concern as the purpose of this simulation is not to test all levels of
validity (q.v. Hermann, 1967) but to test the lowest levels. The pri—
mary object here is to see if the model evokes any responses.

The second problem with this simulation is the fact that hand rath-
er than computer techniques will be used. Simulating even a simple mod-
el by hand is arduous, and the difficulty is greatly increased when the
model functions in a parallel mode. With a computer, many simulations can
be easily run with different initial conditions of the memory and of the
parameters. This would allow for a fuller range of validity testing. For
this reason , the simulation should not be considered to be a formal test
of the C-F-A model (cf. Feigenbaum, 1959, p. 85).

A proper question at this time would ask what the results of a "good"
or "validating" simulation should look like. For all of the reasons noted
earlier, this simulation can be judged a success if paramountly (1) non-

trivial responses are evoked from the model after being presented with a

 

 

stimulus word, and, to a lesser extent, (2) if the model and the res-

ponses manifest face validity in terms of the theories and research
known in this area. The second of these conditions will be taken up in

the next chapter. The first of the conditions is considered below.

A Simulation of the C—F—A Model

 

 

The hypothetical discrimination net is shown in Figure 26 and the
identification of the attribute testing nodes in the letters' part of
the net is given in Table 6. The net is similar to those generated by
EPAM III (with the exception of the pairs' portion). For simplicity,
attribute testing nodes in the units' portion are minimal; limited here
to D—2. Also, it should be clear that empty terminals and branches not
relevant to this simulation have been omitted from Figure 26. The K8 and
K9 branches are used here in the same manner as in EPAM III and WEPAM:

K8 is taken when the value found does not as yet have a branch grown from
the node, K9 is taken when the dimension tested at the node is irrelevant
to the object.

The terminals in the units' part of the net only show their first
objects. The sixteen first objects and their associated response lists
are shown in Table 7. Again it must be remembered that this organization
of words is not claimed to be necessarily realistic. All simulations of
cognitive processing of verbal materials have physical limitations on the
size of their memories. These sizes, fl'lough they differ, are all consider-
ably smaller than the most conservative estimates of the size of the human
verbal memory. If a small subset of a human memory were obtainable, it is

reasonable to expect to find some words appearing as both first objects

 

112

>a¢2m >zzam

wage: no
«uhkun :he
mzh mu H<zz

mpno

 

 

~38 “a
schema azN
axe ma h<=z

ano

  

~o¢oz ac

   
  

«whens sz~
we» ma p¢==
m-a
. m2 <
m:.w
Amman“ quul.

   

|§

m x x

wage: mo
«uhhun :hv
uzh mu h<zz

ppoo

onn mm<m

wx m

    
 

mg <

 
   

wage: no
mmhhmn ozN
mzh ma h<:3

one
u.4.o
same: no

«mhhmn hmp
uzh mu htzz

  
  

   
    

   

wage: no page: mo
mmhhmn cut «when; can
mzh mm h<:3 uzh mu h<=3

 

.wFQmeu cm com: umz covuacwewgummo .mm exam?“

555 E

 

   
  

h z

  
 
  

    

  

cpuo

   

~o¢oi me
«when; :hc
mxh mm h<:3

 

E

    

wage: no

    
  

 

  
  

«mush «ween; eke
axe we e<=z
m-a z
x H 20‘
eggs: as s¢$~z<
“whens azN
< p u e: s 2::
«-9
E m. 52.8 ..3 a
use; as we,» h<== so.

 

—.a

 

 

 

113
21

.Aeozcwueouv opae~xm cw com: umz cowuacmewsomwo .oN ugamru
h PW e a. we. we.
a c x c x c z c c
Fe-o oeo Nu o
x H A
Q «In
um a : o~-o p~-o
: c z
a p r Ea
¢¢-o owua
A x

e x
aolwa.._
: :

mnno

cal
-II
11'
r-

I
O

 

 

11H

 

 

 

Table 6. Identification of Attribute Testing Nodes in
Letters ' Part of Net .
Node Type Test
Tl Vertical Line?
T2 Vertical midpoint?
T3 Horizontal line?
Tu Horizontal top?
TS Horizontal middle?
T'6 Horizontal bottom?
T7 Straight line?
T8 Diagonal line?
T9 Diagonal midpoint?
T'lO Diagonal down from right?
T11 Enclosed space?
Tl? Curved line?
T'l3 Concavity below?
T‘lu Cmcavity above?
T'lS Open to the left?
T16 Open below?

 

 

 

 

 

115

Table 7. Organization of Hypothetical Memory.

 

 

ANIMAL
dog

horse
rabbit
lion

rabbit

tiger

HORSE
animal

LEG

LION

tiger
tiger
animal

dog
dog
RABBIT
bunny
easter
tail
foot
SHOE

 

‘7. ‘T_’A—‘:j

 

 

116

and in response lists. Also some words would be found which appear in
either the first position or in the response list. In the sense of pre-
ceeding lines, the hypothetical memory used in this simulation does not
seem to be troublesome. Parenthetically, it should be pointed out that
the associative principle of frequency is represented directly in this
verbal memory (e.g. see PUPPY).

Throughout the simulation, words will have to be frequently recog—
nized by sorting them (in actuality, their markers) through the discrim-
ination net mder the control of the Net Sorting Routine, N—O. This is
a complex operation to follow, and as it is serially connected to the
other aspects of the simulation it would be helpful to rm this sub—sim-
ulation separately. The results of this sub-simulation will be directly
incorporated into the full simulation.

Table 8 outlines the sub-simulation for the discrimination of the
word LEE. It starts with a marker at D-1 and a marker (RM or SM) at N—l.
Under the control of N-O, the marker in the net is sorted to the L and E
terminals in the letters' portion and ultimately to the E terminal in
the units' portion. The number of time units needed to recognize LEG is
23.3 time units. In terms of the rest of the C-F-A routines 2” time mite
will elapse from the calling of N-O to the exit from N—O. Table 9 gives
the results of discriminating the other 15 first objects in memory.

The other initial conditions which must be specified are the values
of the parameters and the values of the dimensions stored in the property
set attached to each terminal (q.v. Table 5). Table 10 gives the values
of the parameters. According to these values the simulation will end when
either five responses are given or 1000 time units have elapsed. The ca—

pacity of the short term memory is seven words. This effectively prevents

 

   

 

 

117

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

m2 ...... .22 .22 .92 .M2
.memp wH m¥ a:OMJ: NNN :OMJ: .mHIZ
5. 28.23 5.88 on m 8% 8cm .32 .32 5.2 .72
.523 9.. m2 :2: m: dd .82 .22 .22
am now 50:95 0c mi... 893. 82.3mm

28 To .2.FI.3.2 .212 .72
8.0. ..m: .22 +8.2 $95. a 208.3an
.82 .312 .35 .22
.:.c .32 .32 .m-2 .72 £2 .le
22 id .32 «T2
.72 62 a2 42
as: no $3.3 SH To .32 ...T2 .22 £2
E83 no :oMpmcefirsomwo arm .72 .32 .32 .22
:2 Ta mTfimT2 ...Tz .le
mm 5.. 283 e Engage
.Hmd .2d .32 472
.53 .32 .Bd T2 .72 .muz .72
.mmeocdcfiwmp 8.33% 88 he .8 mm id .22 62
Puma .3953 cm woooc wcmpwwp HQ .512 .mnz .mnz .le
.Empw mo mm min .312
nmfifl chum no :03??ng fiwmm .272 .22 .32 J2
2 Ta 472
T2 .72 6.2 .32
m «.2 472 $2
72 6.2 .72 .72

o to

 

mopoz

 

525 2.3
MO m£#COH CHV
was emmaflm

 

oowhom med. mo ocm
em #02 coavmcgwuowfio
CH swim: mo cofifimom

 

65262 mfiEom $2
5 .2952 no ugh:

 

 

an: .82 8325802 53H 53 mo 29.9.2 5.

. m manna.

 

 

 

118

Table 9 . Time Units Required to Recognize All
Items in Hypothetical Memory

 

 

 

 

 

 

 

 

Items Time Units
ANIMAL 27
ARM 214
BUNNY 28
CAT 18
Doc 18
EARS 28
EASTER 28
FOOT 21
HORSE 17
LEG 2n
LION 27
PUPPY 27
RABBIT 1n
SHOE 19
TAIL 27
TIGER 29
Table 10 . Value of Parameters for Simulation
Parameters Value
EN 5
ET 1000
MS 7
mm 7
V1 I-AV + . 5(100—I-AV)
V2 I—AV + .25(100-I-AV)
V3 I-AV + .1(100-I-AV)
Y 25
91 25
02 10

 

 

 

 

119

STM from playing any major role in this simulation as EN is less than MS
(see Table 1+ for a list of abbreviations). The maximum total number of
markers (SMs plus RMs) which can be active at any UI‘ is seven.

The possible values of I-AV range from 0—100. The I—AV of a word
is raised during processing to Vi where the vane of 3 depends upon what
happened to the RM. The most a word's I—AV can be raised is to decrease
its difference from 100 by 50%. The least it can be raised is by 10% of
its distance from the maximum.

The value of Y specifies that an IRT‘ equal to 25 will cause a new
RM to be initiated for the next mused response on that response list.
This occurs, of course, when contraindications are not present.

Finally, the chosen values of theta mean that a word will be con—
sidered as a possible candidate for evocation if its I-AV is greater than
or equal to 25. If a word's I-AV is less than that but at least 10 its
RMwillbe changed toan SM.

Because the C-F-A models does rot produce a memory, it does rot build
up values for each word's I-AV and DI‘ (last date-time the I-AV was changed).
This simulation starts at CDT equal to 201. The first 200 time units were
required to actually build the verbal memory. It is rot knam at this time
if that amount of time is reasonable. The first two columns of Table 11
give the initial values of each word's I-AV and 171'. They were chosen ran—
domly as no other information is available.

Table 12 presents the simulation. It traces all major CDTs from 201
to 5H1 when the simulation ends because five responses were given to the
stimulus word ANIMAL. The Table is set up so the movement of the markers
can be followed. It would be impossible to describe what happened in the

simulation with a non-tabular format. As an overview of the simulation,

 

 

120

 

 

 

 

Table 11. I-AV and DI‘ of First Items in Memory.
At the Beginning At the End of the
of the Simulation. Simulation.

Item I-AV* DI‘** I-AV UI‘
ANIMAL 13 131+ 1'4 500
ARM 76 1 76 l
BUNNY 1+6 77 17 1+17
CAT 53 173 66 306
DOG 22 83 16 271+
EARS 1‘ 138 11 l+93
EASTER 68 118 39 ”AS
FOOT 88 187 66 519
HORSE 9’4 170 75 331
LEE 38 105 38 105
LION 52 82 1+9 538
PUPPY 83 131 83 131
RABBIT 77 73 62 539
SHOE 3 83 3 83
TAIL 5 1'40 12 l+67
TIGER 53 182 70 1+81

 

 

 

 

 

*Choosen randomly with replacement from 1-100.
“(loosen randomly with replacement from 1-200.

 

 

121

 

 

 

Table 12 . An Outline of C-F—A Response Giving
End of Position
CDT Marker of Marker“ Notes"
201 T—l Initialization completed.
'r-2 Stimulus object ANIMAL obtained.
202 M1 M1 created. CDI‘ attached to M1. M1 is
N—2#l an SM. Looking for ANIMAL terminal.
228 Ml s-2#1 ANIMAL recognized.
231+ Ml R-l#l Change M1 to an RM. Initialize IRT=0
for response list associated with ANIMAL
2H1 Ml R—MZ Ml marks DOG. Mark IDG under ANIMAL as
used. Update M1.
2‘42 Ml N-2#l Looking for DOG terminal.
259 Ml R—6#1 DOG recognized.
M2 M-U. IRT for ANIMAL met. Initiate a new RM.
Set IRI'=0. Attach CDT to RM.
2614 M1 R-SIM I-AV for m6 = 22 (83/268)=7.
M2 R-3#3
266 M1 R-9#2
M2 R-lH'IZ Mark CAT under ANIMAL as used. Update
M2. M2 marks CAT.
267 Ml R—9#3 I-AV of DOG = 7 is not i 91.
M2 N—2#l Looking for CAT terminal.
270 Ml R-17#3 I-AV of IDG = 7 is not 3 92.
272 Ml R-19#2 I-AV of DOG increased by V3 = 16.
27a Ml R-1u#2 DT of poo updated.
275 Ml M-12 Ml removed.
28% M2 R—6#1 CAT recognized.
M3 M—H IRI' for ANIMAL net. Initiate a new RM.
Set IRTr-O. Attadn CDI‘ to RM.
290 M2 R-8#'+ I-AV for CAT = 53 (173/290) = 32.
M3 R-3#3
292 M2 R—9#2
M3 R-1+#2 Mark HORSE under ANIMAL as used. Update
M3. M3 marks HORSE.
293 M2 R-9#3 I—AV- of CAT = 32 is a 91.
M3 N-2#l Looking for HORSE terminal.
300 M2 R-llfl CAT evoked as a reapcnse.
302 M2 R-12#2 I-AV of CAT increased by Vl = 66.
309 M2 R-13#2 CAT adchd to top of STM.
306 M2 R-lun DT of CAT updated.

 

 

 

 

*The first letter indicates the routine; the second item marks the sub-
caxponent of that routine; the third item, if it exists, specifies the
number of time units the marker has been in that subccmpcnent.

M'Notes are anly given when interesting. Whenever a marker enters the NEI‘
Sorting Routine, its position is not followed until it exists .

 

 

 

 

 

122

 

 

 

 

 

 

Table 12 . (continued)
End of Position
CUI‘ Marker of Marker Notes
307 M2 M—7 NR=1
R-l#l Since R—l is still an entry point for
ANIMAL stimulus.
309 M2 R-1#3
M3 R—6#l HORSE recognized.
M14 M—H IRT‘ for ANIMAL net. Initiate a new RM.
Set IRIsO. Attach CUI‘ to RM.
31% M2 R—H#2 Mark RABBIT under ANIMAL as used.
Update M2. M2 marks RABBIT.
M3 R-8#3
M14 R-2#2 Mark R-l as no longer an entry point
for ANIMAL stimulus.
315 M2 N-2#l Looking for RABBIT terminal.
M3 R-BH I—AV for HORSE = 9M (170/315) = 51.
nu R—3#1
318 M3 R—9#3 I-AV of HORSE = 51 is 3 91.
Ml} R—lfill
319 M3 R-10#l
nu R—MIZ Mark LION under ANIMAL as used. Update
MN. m marks LION.
320 M3 R-10#2
mu N-2#l looking for LION terminal.
325 M3 R-lliM HORSE evoked as a response.
327 M3 R-l2#2 I-AV of HORSE increased by V:L = 75
328 M2 R—6#l RABBIT recognized.
M3 R-13#l
329 M2 R—6#2
M3 R-13#2 HORSE added to top of STM.
331 M2 R-8#l
M3 R-ll+#2 DT of HORSE updated.
332 M2 R-8#2
M3 M—7 NR=2.
M—l2 M3 removed $ R-l is not an entry point
for ANIMAL stimulus.
33% M2 R—BM I-AV for RABBIT = 77 (73/334) = 17.
-- T-12 IRI' for ANIMAL net. No mused responses .
IRT dropped.
337 M2 R-9#3 I-AV of RABBIT = 17 is not ; 91.
3% M2 R-17i3 I—AV of RABBIT = 17 is z 9 .
31m M2 R—lBfiu MZ is dnangd to an SM. M2 is updated.
I-AV of RABBIT increased by V2 = 38.
3'46 M2 R-11M2 UT of RABBIT npdated.
m R-6#l L101 recognized.

 

 

 

 

 

123

 

 

 

Table 12 . (continued)
End of Position
CUI‘ Marker of Marker Notes
31+? M2 N-2#1 Looking for RABBIT terminal.
mu R—6#2
352 M“ R-8#|+ I-AV for LION = 52 (82/352) = 12.
355 MH R-9#3 I—AV of LION = 12 is not >.. 91.
358 MR R—l7#3 I-AV of LION = 12 is z 92.
360 M2 s-2#1 RABBIT recognized.
Ml} R-18#2
362 M2 S—2#3
mu R-18#lt M16 is dnanged to an SM. MM is updated.
I—AV of LION increased by V2 = 31+.
361+ M2 S-3#2
Mu R—lu#2 DI‘ of LION updated.
365 M2 S-3#3 RABBIT as a stimulus has responses.
Mu N-2#l looking for LION terminal.
366 M2 R—l#l M2 is dnanged to an RM. Initiate IRI‘=0
for response list associated with RABBIT.
373 M2 R—‘MZ Mark BUNNY unthr RABBIT as teed. Update
M2. M2 marks BUNNY.
371+ M2 N—2#1 Looking for BUNNY terminal.
391 MR S-2#l LION recognized.
M5 M-U: IRT for RABBIT met. Initiate a new RM.
Set IRI' = 0. Attach CDT to RM.
396 M14 S-3#3 LION as a stimulus has responses.
M5 R-3#2
397 mu R—1#l MN is dnanged to an RM. Initiate IRI‘=0
‘ for response list associated with LION.
399 M14 R-l#3
M5 R—ll#2 Mark EASTER under RABBIT as used. Update
M5. M5 marks EASTER.
#00 MM R—3#1
M5 N—2#l looking for EASTER terminal.
l+01 M2 R-6#l BUNNY recognized.
M4 R-3#2
I+014 M2 R-Wl
M14 R-l4#2 Mark TIGER under LION as used. Update
Mu. Ml} marks TIGER.
#05 M2 R-8#2
MM N-2#1 Looking for TIGER terminal.
W7 M2 R-Bflt I-AV for BUNNY = l46(77/u07) = 8.
l$15 M2 R-19#2 I-AV of BINNY increased by V3 = 17.
lt16 M2 R-l'Ml
b5 M—u IRI‘ for RABBIT met. Initiate a new RM.
Set IRI' = 0. Attadn CDT to RM.
ltl7 M2 R—1M2 131‘ of mm updated.
1% R-lfll

 

 

 

 

 

 

 

 

Table 12 . (continued)
W Posifion
CDI‘ Marker of Marker Notes
ln18 M2 M-l2 M2 removed.
M6 R-1#2
I+22 M6 R-3#3
M7 M—u IRI‘ for LION net. Initiate a new RM.
Set IRI' = 0. Attach CDT to RM.
uzu M6 R-|+#2 Mark TAIL under RABBIT as used. Update
M6. M6 marks TAIL.
M7 R-l#2
#25 M6 N—2#l Looking for TAIL terminal.
M7 R-l.#2
I427 M5 R-6#1 EASTER recognized.
M7 R-3#2
ln30 M5 R-8#l
M7 R-‘HIZ Mark TTGER urnder LION as used. Update
. M7. M7 marks TIGER.
I+31 M5 R—8#2
M7 N—2#1 Looking for TIGER terminal.
1433 Min R—6#1 TIGER recognized.
M5 R—8#H I-AV for EASTER = 68(118/u33) = 19.
0.36 Min R-8#1
M5 R—9#3 I—AV of EASTER = 19 is not a 9 .
l$39 Ml-l R-BIM I-AV for TIGER = 53(182/939) =122.
M5 R—l7#3 I-AV of EASTER = 19 is >82.
Hill M1 M—u IRT for RABBIT met. Initiate an RM.
SetIRI‘= 0. AttachCDI‘toRM.
nu R—9#2
M5 R-l8#l
1.042 Ml R—lfll
Mu R-9#3 I—AV of TIGER = 22 is notzel.
M5 R-18#3
#143 Ml. R—1#2
Mil R—l7#l
M5 R-18#ln M5 is changed to an SM. M5 is updated.
I—AV of EASTER is increased by V2 = 39.
ans M1 R-3#l
Mu R—17#3 I-AV of TIGER = 22 iszez.
M5 R-lu#2 UT of EASTER tpdated.
HHS M1 R—3#2
nu R—18#1
M5 N—2#1 Looking for EASTER terminal.
1+1}? Ml R—3#3
M2 M—li IRI‘ for LION met. Initiate an RM. Set
IRI‘ = 0. Attach CDT to RM. -
Mu R-18#2
“#9 MI lid-W2 Mark EARS under RABBIT as teed. Update

 

 

 

Ml. MlnnarksEARS.

 

n72

n73

 

 

 

 

 

 

Table 12 . (continued)
of Position
CUI' Marker of Marker Notes
M2 R-1#2 .
MM R-18#ll MA is changed to an SM. M14 is updated.
I-AV of TIGER is increased by V2 = 1+1.
l+50 Ml N-2#1 Looking for EARS terminal.
M2 R-l#3
MM R—llHVl
.451 M2 R-2#l
Mu R-l'4#2 UT of TIGER updated.
M6 R-6#l TAIL recognized.
l$52 M2 R-2#2 Mark R—l as no longer an entry point
for LION stimulns.
M14 N-2#1 Looking for TIGER terminal .
M6 R-6#3
#57 M2 R-1+#2 Mark ANIMAL under LION as used. Update
M2. M2 marks ANIMAL.
M6 R-8#‘+ I-AV for TAIL = SOHO/#57) = 2.
L$58 M2 N-2#l Looking for ANIMAL terminal.
M6 R-9#1
1459 M6 R—9#2
M7 R—6#l TIGER recognized.
l$60 M6 R—9#3 I-AV of TAIL = 2 is not a 91.
M7 R—6#2
l+63 M6 R—l7#3 I—AV of TAIL = 2 is not a 92.
M7 R—8#2
l+65 M6 R-19#2 I-AV of TAIL increased by V = 12 .
M7 R-8#ln I-AV for TIGER = urcuanmasi = no.
ass M3 M-u IRT for RABBIT net. Initiate an RM.
Set IRT‘ = 0. Attach CDI‘ to RM.
M6 R—llwl
M7 R—9#1
1+6? M3 R-l#1
M6 R—lu#2 UT of TAIL updated.
M7 R-9#2
l+68 M3 R-l#2
M5 M-12 Marker removed.
M7 R-9#3 I-AV of TTGER = I+0 is z 91.
l+71 M3 R—2#2 Mark R—l as no longer an entry point
for RABBIT stimulus.
M7 R-10#3 TIGER is neither the stimulus object
nor in STM.
I+72 M3 R—3#1
M7 R—llfl
— T—12 IRT for LION met. No unused reSponses.
IRI' dropped.
l+73 M3 R-3#2
M5 S-2#1 EASTER recognized.
M7 R—11#2

 

 

 

 

 

 

 

W

“86

 

 

 

 

 

 

 

Table 12. (continued)
End of Pdsition
CUI' Marker of Marker- Notes
Q75 M3 R—Qill
M5 S-2#3
M7 R-llilu TIGER evoked as a response.
I+76 M3 R-lt#2 Mark FOOT under RABBIT as used.
Update M3. M3 marks FOOT.
M5 S-3#1
M7 R-12#l
ln77 Ml R-6#l EARS recognized.
M3 N-2#1 looking for FOOT terminal.
M5 S-3#2
M7 R—12#2 I-AV of TIGER increased by Vl = 70.
|+78 Ml R—6#2
MS S-3#3 EASTER as a stimulus does not have any
responses associated with it.
M7 R-13#l
l+79 Ml R-6#3
M5 F-1#1
M7 R-13#2 TIGER added to top of STM.
l+80 Ml R-8#1
M14 S-2#l TIGER recognized.
M5 F-l#2
M7 R-lWl
l+81 Ml R-8#2
Mu S-2#2
M5 F-Zill
M7 R-1u#2 Dr of TIGER updated
I+82 Ml R-8#3
Mu S—2#3
M5 F-2#2
M7 M—7 NR = 3.
M-l2 M7 removed as R-l is not an entry point
for LICN stir-.1115.
H83 Ml R-8#l§ I-AV for EARS = I“138/883) = 1.
MM S-3#l
M5 F—2#3 Yes. EARS terminal is off of marked
terminal.
#88 Ml R-9#1
M2 R-6#1 ANIMAL recognized.
M14 S-3#2
M5 F-3#1
#86 Ml R-9#3 I-AV of EARS = l is not % 91.
M2 R—6#3
MM R-lfl Change M“ to an RM. Initiate IRT = 0
for response list associated with TIGER.
MS P—ufil

 

 

 

w

 

 

127

 

 

 

 

 

 

 

 

 

Table 12. (continued)
End of Position
CDI‘ Marker of Marker Notes
I‘88 M1 R-17#2
M2 R-8#2
Mu R—1#3
M5 F-lt#2 Yes. EARS is successful as a stimulus
since it has responses associated
with it.
.489 Ml R-17#3 I-AV of EARS = l is not a 82.
M2 R—8#3
Mn R—3#1
M5 F—7#1
.490 M1 R—19#1
M2 R-8#ll I—AV for ANIMAL = 13(13u/u90) = ‘4.
Mn R—3#2
M5 F-7#2 Terminal EARS is dnosen as stimulnB.
Attach M5 to EARS. Attach CDI‘ to M5.
l#91 Ml R-19#2 I-AV of EARS is increased by V3 : 11.
M2 R—9#1
Mu R-3#3
M5 S-5#1
-— T—12 IRI‘ for RABBIT net. No unused responses.
IRT dropped.
l+93 Ml R-l‘4#2 UT of EARS updated.
M2 R—9#3 I-AV of ANIMAL = 1+ is not a 0 .
Mu R-l-t#2 Mark LION under TIGER used. date Mu.
M5 S-5#3 Mu marks LION.
HEM Ml M-12 Marker renoved.
M2 R-l7#1
Mu N—2#1 Looking for LION terminal.
M5 R-1#l Change M5 to an RM. Initiate IRT = 0
for response list associated with EARS.
l+96 M2 R-17#3 I-AV of ANIMAL = 1+ is not a 92.
M5 R-1#3
l697 M2 R—19#l
M3 R—6#1 FOOT recognized.
M5 R—2#l
l$98 M2 R-19#2 I-AV of ANIMAL increased by V3 = lit.
M3 R—6#2
M5 R-2#2 Mark R-l as no longer an entry point for
EARS stinulus.
500 M2 R—lu#2 UT of ANIMAL updated.
M3 R-8#1
M5 R—3#2
501 M2 M—12 Marker removed.
M3 R—8#2
M5 R-3#3

 

 

 

 

 

128

 

 

 

Table 12 . (continued)
End of Position
CUT Marker of Marker Notes
503 M3 R-8#1+ I-AV for FOOT = 88(187/503) = .
M5 R-IMZ Mark RABBIT under EARS used. Update
M5. M5 marks RABBIT.
50‘ M3 R-9#l
M5 N-2#1 Looking for RABBIT terminal.
506 M3 R-9#3 I-AV of FOOT is >, 8 .
511 Ml M-Q Set IRI‘ = 0. Att CDT to RM.
M3 R-11#2
513 Ml R—l#2
M3 R-ll#|+ FOOT evoked as a response .
515 Ml R-3#1
M3 R-12#2 I-AV of IOOT is increased by Vl = 66.
517 Ml R-3#3
M3 R-13#2 FOOT added to top of STM.
M5 R—6#1 RABBIT recognized.
519 Ml R-u#2 Mark ANIMAL under TIGER used. Update
Ml. Ml marks ANIMAL.
M3 R-1u#2 131‘ of FOOT updated.
M5 R—6#3
-- T-l2 IRT for EAIG net. No unused reSponses.
IRT dropped.
520 Ml N-2#l Looking for ANIMAL terminal.
M3 M—7 NR = H
M“ R-6#l LION recognized.
M5 R-8#l
523 MM R-8#l
M5 R-8#|+ I—AV for RABBIT = 38(3146/523) =
526 M4 R-8#3 I—AV for LION = 31+(364/526) = 2142
M5 R-9#3 I-AV of RABBIT = 25 is 3 81.
529 MR R—9#3 I-AV of LION = 24 is not a. 81.
532 M”: R-l7#3 I-AV of LION = 21+ is a 82.
M5 R-ll#3
533 MM R-l8#l
M5 R-ll#u RABBIT evoked as a response .
535 Ml} R-l8#3
M5 R—l2#2 I-AV of RABBIT is increased byV
536 M2 Mu IRI' for TIGER met. Initiate a "3.: RM.
SetIRT=O. AttadnCDI'toRM.
Mu R—18#u m is changed to an SM. MM is updated.
I-AV of LION increased by V =
M5 R-13#1 2
538 M2 R-1#2
Mu R-1u#2 UT of LION updated.
M5 R—lufil

 

 

 

 

 

 

 

129

 

 

 

 

Table 12 . (continued)
End of Position
CDT Marker of Marker Notes
539 M2 R— 1#3
Ml} N-2#l Looking for LION terminal.
M5 R-lu#2 DT of RABBIT is updated.
5140 M2 R- 3#2
MS M—7 NR = 5. Set stop flag. M5 removed.
5H1 -- ------ STOP.

 

 

 

 

 

 

 

 

 

 

 

 

130

Table 13 lists the five responses given to ANIMAL, the DI' it was given,

and fine first object in fine nenory it was associated with.

Table 13. Responses Evoked During Simulation.

 

 

 

 

 

 

131' Response Stimulus

300 (‘AT ANIMAL

325 HORSE ANIMAL

l‘75 TIGER LION

513 FOOT RABBIT

533 RABBIT EARS
m

In this chapter fine C-F-A nodel was tested by means of a hand simu-
lation. Since the nodel does not include a net building routine a hypo—
thetical memory was constructed. The menory contains 16 first objects
each of which has up to five associated reaponses. The sinulation was to
stop when either five overt respanses were made or 1000 time units had
elapsed. The sinulaticn took 3141 time units before the five responses
were evoked. This net the major criterion set for the nodel -- that res-

ponses are evoked. The next dnapter evaluates the simulation more closely.

 

ITEM

Omte

with

(Hi

We

lat

655

 

 

CHAPTERV

This chapter is organized into finree parts. In the first fine C—F—A
model and the execution of fine hand simulation is evaluated. Major
strengths and weaknesses of the nodel are identified along with an assess—
ment of different dimensions of the nodel's validity. The second section
contains approaches to needed extensions of the nodel if it is to be more
completely tested. Mainly finis section is concerned with fine problems of
a net bm'lding routine. The last section is more speculative as it deals

with nore distant future explorations with fine nodel .

An Evaluatian of the C-F-A Model

Before examining fine C-F-A nodel in detail, it would be wise to re—
vian some of the'mefinods and criteria for assessing its validity. Kaplan
(196”), Hermann (1967) and others do not consider validity to have only
two values: valid and invalid. Rather validity is a matter of degree,

depending in part upon fine goals of the nodel and fine state of its develop-

nent. The C—F—A nodel has as its primary goals understanding of the re-
latianships anong its conponents (q.v. Dnbin, 1969) and insight into free
association behavior. In this early stage of its developnent prediction
is not of major concern.

The pattern model [understanding] may nore easily fit expla—

naticns in early stays of inquiry, and fine deductive nodel
[prediction] explanations in later stages (Kaplan, 1961;, p. 332).

131

‘1"“"‘"“"‘l

 

 

 

{BIBS

hibit

 

 

 

132

Hermann's (1967) five levels of validation were discussed in chap-
ter mo. His lowest level is an assessnent of test-retest reliability.
Since fine simulation of fine C-F-A nodel was run only once, there is no
neasure of this reliability. However, fine C-F-A nodel is alnost purely
deterministic in nature. The sole exception is fine random process which
may occur in fine Time Executive Routine at T-5 . In the actual simulation
it was never necessary to execute finis random part of T—S. Thus, there
is no reason to believe finat it will be utilized by all or nost future
simulations. Furthernore, fine nere execution of that random component
does not guarantee finat fine outcones of fine simulation will be altered in
any important way. While never tested, it is reasonable to expect fine test—
retest reliability of fine C—F-A nodel to be hign. That is, it should ex-
hibit very similar behaviors and outputs when it operates under identical
initial conditions (menon and paraneters ) .

Hermann' s fourfin and fiffin levels are nore appropriate for deductive-
predictive nodels finan for fine C-F-A nodel. One aspect of his third level
is sensitivity testing. Such testing requires multiple executions of the
simulation with different initial conditions . Unlike the discussion above
about fine lowest validity level, it is much nore difficult here to estimate
the results of nultiple runs. Sone of finese considerations will be included
in fine ensuing, nore general evaluation of the nodel.

The ofiner part of his finird level requires a conparison of the nodel
and fine nodeled. This plne his second level, face validity, are similar
to Kaplan's (196%) norm of correspondence. Before applying this type of
criterion to fine C—F—A nodel several ofiner criteria should be discussed

briefly .

 

intem

sibly ,
for i:

inam

testr

sinpl

ingi

appee

not
SCie
Ply
51 g
a In

W

Qits
In

Out

   

 

133

Kaplan's second norm is finat of coherence. Coherent nodels are
internally consistent. They fit existing fineories , are simple , and pos-
sibly, are esthetically pleasing. One sensible way to test a simulation
for internal consistency is to see if it executes without a terminal error
in any one run. The one sinulation executed did not terminate wifin an
error. To test for contradictory outcones requires fine sort of sensitivity
testing nentioned previously. Simplicity may mean one of two things . A
sinple theory may be one which is not structurally complex, or it may be
one finat is parsinoneous in terns of its free paraneters. Models attenpt-
ing to explain cognitive processes must be structurally corplex. EPAM I
appears to be an exception to finis, but that may be a function of fine
scope of fine phenomenon it is explaining. The C-F-A nodel has ten free
paraneters. Wifinout further sinulation finere is no way to tell whether or
not finere are too many free paraneters.

The pragnatic norm is Kaplan's third criterion. Valid nodels need
not be practical in an everyday sense. Rafiner they should be useful to
science itself. They not generate interesting questions as well as sup-
ply sore answers. The C—F-A nodel is proposed as a neans of obtaining in-
signt into the relationship between free association behavior and, (when
a nore advanced version is corpleted) meaning. How well C-F-A neets the
pragnatic norm remains to be seen.

Finally, it is instructive to review fine criteria proposed to eval-
uate IPMs specifically. The C-F-A nodel simulates a general individual.
This rules out protocol matdning. Statistical and enpirical conparisons
between fine nodel's output and an average person's output is also ruled
out due to fine inpcssibility of having fine nodel and fine average person

 

I in!

         
  
 

‘.

. ...,. as; V ;.. ..‘tfiffifn s‘I

   

  
  

   

..n... —. .-

,._ __ “— -_--.‘-r-r

 

13”

start with identical conditions. The only procedure finat seens workable
is a loose version of Turing's test, whidn is similar to face validity.
Detailed process simulation does not usually lend itself to
significance tests. Connon sense inpression of similarity
seems fine only basis for judgment. There is nothing wrong
wifin this use of comon sense. (Frijda, 1967, p. 65)
In sum, fine major criterion applicable for evaluation is face validity--
or equivalently fine norm of correspondence or Turing' 3 test . This criterion F
is not applicable to fine possible full range of corparisons. Mainly, this

is because fine nodel uses a hypofinetical nenory. Since the dnaracteristics

" ml .19 ‘1';

of associates given in a C-F—A task depend upon fine structure and content
of fine verbal nenory, it is not sensible to corpare fine nodel's output
wifin finat of an individual (or generalized individual) rigorously. (No—
tive finat fine EPAM-type nodels are not so limited as fine verbal learning
experiment defines fine verbal nenory of interest.)

The remainder of finis section presents an evaluation of the nodel
in terns of a gross examination of its output and a nore detailed look at
fine principles and plnenonena "represented" or possibly "accounted for" in
sone way wifinin fine nodel's processes. What follows is organized around
four related topics: (1) the output of fine simulation outlined in fine last
chapter, (2) fine structure of the C-F-A nenory, (3) the functioning of the
C—F—A nodel especially in terns of sore principles of verbal behavior, and
(H) fine najor strengths and weaknesses of fine nodel.
Sinnlation Output

The single nost significant result of fine simulation is fine fact finat
responses were evoked. The fact finat parts of fine C-F—A nodel were de—
signed to evoke responses (q.v. R-ll in Figure 23) in no way diminishes
fine inportance of finis result. The nodel as described is too corplex for

 

 

 

 

 

 

 

135

one to ascertain before execution whefiner or not responses will be evoked
under a given set of conditions. True, one successful execution does not
guarantee ofiners , wifin different initial conditions , but it does lend
weignt, img _f_a_c_l_:g, to an optimistic expectation of future runs. In addi-
tion, when fine execution ceased five unpredictable responses were evoked.

A nodel which evoked predictable associates might upon presentation
of fine stimulne word ANIMAL, respond wifin DOG, CAT, HORSE, RABBIT, and
LION (q.v. Table 7). This type of nodel places the burden of free-assoc—
iation behavior upon fine processes which build fine menory net, instead of
fine processes which retrieve responses from the net. As stated in chapter
four finere are reasons to believe finat humans do not evoke all responses
directly associated in nenory with a stimulus object. Therefore, nodels
based upon retrieval of items fron nenory are to be preferred. The C—F—A
nodel is of this type. Also, when fine sinulation stopped at DIES“,
there were two active markers being processed. Wifinout further processing
finere is no way to tell which additional responses (if any) would have
been evoked.

A second interesting characteristic of the execution is the time each
response was evoked. Table 13 sunmarizes finose tines. What is apparent
is from an inspection of finese ms, is fine fact finat finey are not regular
and finat inter-response intervals vary greatly. There are two groups of
responses (1) CAT, HORSE, and (2) TIER, FOOT, RABBIT. The inter—response
interval between finese groups is 150 tine units while the interval wifinin
fine groups rnever greater finan HO tine units. 'Dnis tenporal grouping of
free associates has been stuch'ed by Pollio (1966) who fournd that humans
tenporally group fineir associates , and fine average sennantic distance be-
twaen groups was greater finan fine distance wifinin groups . In the C—F—A

_wnp.
I 1 "B
.
.5
b

 

 

 

 

 

 

 

136

m1 fine temporal clustering seems to be due primarily to an internal
nediation of stirrulus objects . Anofiner contributer to fine inter-response
latency is fine additional processing required wlnenever an unsuccessful
terminal is reached. At DT=I+78, fine potential stimulus EASTER was elim-
inated because it did not have responses associated wifin it. Routine

F-O was called and EARS was substituted for EASTER. The processing of
F—O possibly increased fine number of time units between evoked responses
four and five.

While ANIMAL was fine only nominal stinulns in fine simulation, there
were four functional stinuli (q.v. Underwood, 1963) which nediated overt
responses . In addition ofiner items in fine menon effected fine processing
of possible candidates for evocation (e.g. EASTER) and in sone sense served
as internal mediators .

Mediation in fine C-F-A nodel is nore conplex finan finat in the EPAM—
type nodels. Those nodels and fine C—F—A nodel mediate responses by dis-
criminating some coded representation (e.g. cue codes) of the response
finrough fine net. The , every response has beenn mediated. In addition to
this response nediation, only fine C—F-A nodel includes a form of stimulus
nediation. Stimlne nediation occurs whenever a word' s I-AV is less than
91 but greater finan 92. Response nediation is similar to Osgood's rm's
(see chapter one), in fine sense finat they bofin function for all inputs.
Stimlus nediation in fine C-F—A nodel does not always occur and is closer
in its operation to fine finree stage sinple chain nediation paradigm (q.v.
Jenkins, 1963).

A fourth major result of fine simulation is fine permanent alternation
of parts of fine nenory as a function of processing. Newell, Shaw and
Sinon (1958) call finis type of alteration a form of learning. Table 11

 

 

137

presents fine I-AV and DI‘ of all first items in the nenory before and
after fine simulation. There was only one noninal stimulus and five
evoked responses, but fine I—AV and UT of 12 of the 16 itens in neme
were changed.

This type of learning by the nodel helps produce nore successful
nediated reactions. Tho exanples of this occurred in fine simulation.
The first exanple concerns TIGER which was evoked as a response at DI=I+75.
'Dne TIER that was evoked was fine second TIGER under LION, not fine first

 

(see Table 7). At UI‘=I+0!+ the first TIER started being processed. Its
I-AV was too small for it to be evoked. At DIEM-$9 its I-AV was raised
by V2 to 1&1. The second TIER was started into processing at DT=H30.
Its I—AV was compared wifin 61 at DI‘=I+68 which was after fine I—AV was
raised to 1+1. Hence, TIER became a candidate for evocation on its second
attenpt. The other exanple is similar to the first. RABBIT as a response
to ANIMAL did not have an I—AV sufficient for evocation. But in its pro—
cessing the I-AV was raised above fine finreshold (at mean). When RABBIT
as a response to EARS was processed it could be evoked as a response be—
cause of its raised I—AV.

A fifth attribute of fine sinulation whidn ought to be pointed out
is finat parallel processing actually occurred. The existence of T-S in
fine Time Executive Routine does not guarantee parallel processing. It
nerely stipulates finat if fiere is nore than one nnarker active at a CDI‘,
finen finey shall be processed in a parallel node. The simulation starts
wifin only one active marker, fie SM for ANIMAL. The entire simulation
could have occurred with only oe nnarker active at each CDT. That wasn't

fine case. The actual anount of tine different nunbers of markers were

 

 

 

138

active is given in Table 1“. Tie Table shows finat in the simulation from
one finrough seven markers were active. There was no instance in which
nore finan seven markers were active in any CDI‘. If a situation occurred
in which an eignfin marker were needed, fine nodel would have prevented its
initiation at T-ll because fine parameter NMM was set equal to seven. NMM
serves to limit fine anount of paallel processing and is in line with evi-
dence reviaved by Miller (19 56) and ofiners on fine limitation of human in-

formation processing.

Table 14

Time Units Different Nunbers
of Markers Were Active

 

 

 

Nunber of Markers Tine Units
1 65
2 100
3 76
1+ 1W
5 18
6 31+
7 2

 

 

Finally, it is interesting to notice that all six C-Fe-A routines were
used in the simulation, and finat fine interrelations annong fine routines
functioned as expected. In real O-F-A situations §_s rarely fail to give
responses or stop responding in fie middle of an experinent. They operate
as if under pressure to give a response. The F-O Routine in fine C—F—A
nodel operates to sinulate this behavior. F—O was used only once during
execution which was unexpected. There was no reason to believe before
beginning fine simulation that F—O would be called at all. At DIE-H79 the

routine was called when stinnlus EASTER was discriminated to a terminal

 

 

 

 

 

139

which contained no associated responses . What a hunan _S does when con—
fronted with a stimulus word for which he has no real (as opposed to overt)
associates is unknown. But fine hunnan does give associates. So does the
C-F—A nodel. F-O operates by finding a terminal close to fine unsatis—
factory one which has associated responses. In fine simulation that ter-
minal was headed by fine stimulus word EARS. I

Considering fine structure of fine discrimination net, F-O operates
a forced stimulns generalization. In fine EPAM—type of nodels stinmlus
generalization occurs because of inconplete previous learning. In fine
C—F—A nodel stinnulus generalization operates because of F-O. EASTER is
fine stimulus, but a response to EARS is given. The nodel can be thought
of as being "under pressure" to respond when no response is available.

This pressure makes the difference between fine "T" in EASTER and fine "8"
in EARS uninportant (see Figure 26, D-lO). A form of stimulne general-
ization follows.

In nost instances of C—F—A or pooled discrete free association exper—
iments finere are itenns in the list of associates finat seem conpletely out
of context. For exanple, Deese (1965) notes finat an associate of BUTTER-—
FLY is WINTER. In fine simulation, fine fourfin associate to ANIMAL is FOOT
which, in turn, precedes RABBIT. In fine C-F-A nodel it is fine F-O Routine
and fine nediation of stimulus items which produce finese difficult to eXplain
Wee»-

! Mennory Structure

Because the C-F-A m1 is heavily based upon fine EPAM-type of nodel
little needs to be said about fine structure of the nenory net. Most of the
previously described strengfins and weaknesses of finose nodels are applicable

 

 

140

 

to C-F-A. A few variations need to be pointed to . None of the earlier
nodels need to allow for a conplex hierarchy of connections within fine
nenory. Models of verbal learning do not require it. But because, "fine

associations a subject fornns are probably nunerous and hierarchically

 

organized" (G.A. Miller, 1963, p. 328), fine C—F—A nodel requires a nore
conplex nennory structure. A cursory exannination of fine hypofietical nem-
ory presented in Table 7 reveals such a hierarchical nenory. Whenever
net building is added to this nodel it met be able to produce an arrange—
nent ofmenory items sinuilarto finat in the_a_d_h_9_cnenory.

WEPAM, SAL III and C-F-A all associate none than one response with
a stinulus item. There are sone differences however. In SAL III only
fine top (nost recent) response is available as a possible response. The
reSponse list also functions as a stochastic short term nenory —- finere is
a probability finat fie topnost response on any list will be pushed up and
"forgotten". The C-F—A nodel operates with a central short term menory.
It retains all associates, finougn they may nnot be available enough to be
evoked. Whereas , SAL III activates one response, C-F-A activates an entire
response list.

In WEPAM, the topnost response not previously given in error becones
fine candidate for evocation. The error indicator is stored with fine re-
Sponse list. If finis form of marker were used in a free association nodel
it would be very inefficient. All of fine nunerous instances of a response
througnout fie nennory would have to be similarly marked and sinnultaneonely
updated whenever necessary. In fine C-F-A nodel fine I—AV of a word is stored
onlyonce finrougnoutfinenennoryatfineternnfinalinwhichthewordis fine

first item. WEPAM functions as if fine responses in a response list are

 

 

 

1‘41

 

known since fine error indicators are decked wifinout discrinninating the
response. In fine C-F-A nodel it is inpcssible to make any inspection of
a response or its associated property set until it has been recognized by
neans of discrimination.

Functioningof Model

For reasons given in fine first chapter, I-AV is fine variable of najor
interest in finis nodel. It differs from associative strengfin in three ways.
First, I—AV is nore directly applicable to the response learning-giving
phase, while associative strength is also relevant to fie associative phase
of verbal learning. Second, I-AV is considered to be nore sensitive to
change, while associative strengfin is nore stable. And, third, I—AV is
determined mainly by frequency of experience , recency of eanerience and
node of experience of fine verbal unit. On fine other hand associative
strengfin may depend upon fie reinforoement history of an S—R pair as well
as fie sinnpler principles of association.

The C-F-A nodel does not specify how each item's I-AV should be
originally estimated. That would be (a proper function of a net building
routine. However, it is instructive to examine how I-AV is handled in the
current nodel. An assunption inplied wifinin fine nodel is finat fine major
factors which affect I-AV do not becone Operative during fine building of
fie verbal nennory, but, refiner during fie response giving phase. That
is , fine valne of an itann' s I-AV is nannipulated during response processing
refiner finan during item learning.

Recency, frequency, and node of experience are fine najor factors
which affect an item's I—AV (Rosenzweig 8 Postman, 1957; Horowitz, Norman,
6 Day, 1966). Figure 23 outlines fie response giving phase of C—F-A.
Recency affects I-AV in two ways. Most directly, recently processed itens

 

 

 

 

 

142

 

in nemory are not as affected by fine effects of fine passage of time on
I-AV as older itens are. R-8 in Fignue 23 stipulates finat all itens'

I—AVs decrease as a function of elapsed tine . Itens with higher DI‘s

(nore recent) have fineir I—AVs decreased least. The second way recency
affects I-AV is less direct. The nost recently acquired item in a re-
sponse list is always the item at the tap. A PDS Operates by "popping—up"
fine topmost item first. It is fine recent itens from each response list
whidn becone fie first candidates for evocation. Since association experi-
ments end before a S's entire verbal nenory is depleted, potential responses ..
left unprocessed are finose furfiner down on fine response lists . Therefore, “
nore recent itenns are more likely to have fineir I-AVs exanined (q.v. R—9

or R-17), Ecreased (R—8) or raised (R—12, R-18, R-19).

Frequency is represented in fine verbal nennory directly, by allowing
fine sane item to appear in the sane (or different) response lists as often
as needed. (Again, exactly what procedure is followed depends upon fine
unspecified net building routine) . Frequency affects I-AV by neans of
successive processing of fine sane item. The two exanplcs described ear»-
lier in this chapter illustrates finis point: The first T'IGER in the re—
sponse list of LION did not have an I-AV sufficiently high for it to be
evoked as an overt response. But fine processing of fine first TIER raised
its I-AV permitting fine second TIER to be evoked. The second exanple is
similar to this dealing with RABBIT (under ANIMAL and later under EARS)
instead of TIGER.

Horowitz , Nornan and Day (19 66) experimentally manipulated I—AV . They
found finan an item' s I-AV is raised nost when fine itenn is overtly produced
from nnelnory. Tie I—AV is also raised wl'en fine item is seen, but fie increase
is not as. great as when it is produced from nenory. The C-F-A nodel raises

 

 

 

 

 

1143

the I—AV of item's in a manner consonant with these findings. The nodel of
fine item serves to raise its I-AV at least a little (see Figure 23, R-19) .
This internal processing corresponds to fine "seen" condition of the experi-
nnent. The nodel raises an item's I-AV the greatest anount when the item

is evoked from nenory (R-12). And, in a condition not paralleled by fine ex—
periment, fine nodel can raise an item's I-AV an internnediate anount whenever
fine I-AV is less finan 91 but greater finan 92 (R-18). This middle condition
seems reasonable as items neeting finis condition are treated as mediating
stinnuli: nediating stinuli are additionally processed but not evoked,

while itenns whose I-AV are not greater finan 92 are not additionally processed.

In sun, I-AV is quite specifically treated in fine C-F-A nodel. Each
first item in fine nnenory has a nuneric value of I-AV assigned to it. In
fine response giving phase , fine I-AV functions in a manner which is con-
sistent wifin existing fineory and experimentation: I-AV is a nnnaj or factor
in determining which of fine potential responses will be evoked, and the
valLe of an item's I-AV depends upon recency, frequency and node of pro-
cessing. The EPAM—WEPAM—SAL nodels, on fine ofiner hand, do not represent
or treat I—AV in any manner whatsoever.

The variable employed by finese earlier nodels to govern response
giving is relative (or absolute) associative strength . In fineory , the
strength of association depence upon fine freqnency of association and the
nunber of associates of fine stimulus word. In addition, each succeeding
presentation of fie S-R pair contributes less to fine strengfin of the
bond between finem (Deese, 1965). This is trne under fine classical con-
ditioning paradign in which contiguity of fine S-R pair is so important, nnd
it is true in fine-operant conditioning paradignn in which reinforoenent
strengfin and scledule is inportant.

 

 

In EPAM II associative strength is represented indirectly by the
degree of conpleteness of a response one code. Cue codes which are well
learning (i.e. conplete) function as if fine response they seek is strongly
associated wifin fie stimulus . Connplete oue codes always retrieve the
correct response. The degree of conpleteness of a one code @pends upon
fine nature of fie discrimination net, not fine nunber of trials. If the

net is built in a manner which forces a one code to pass nunerous tests 7""

 

finen finat one code will be connplete before other one codes. Since the

i \ . '
.

cue tokenns of EPAM III nane a terminal and do not have to be sorted
finrougn fine net, associative strength is not represented in finat nodel
in a manner which permits variation in fine level of the variable. Rather
fine value of associative strength is eifiner zero (before the one token is
added) or at a maximum (at fine trial the one token is added to the terminal).

Wynn's (1966) nodel is similar to EPAM III in accounting for absolute
associative strengfin. It does a better job wifin relative associative
strength since nore finan oe resPonse is stored at a terminal. The order
of responses in fine response list and fine presence or absennce of the error
mark (see chapter 2) determines fie relative strength of responses to a
stimulne.

SAL III also reflects relative associative strength because of the
possible multiple responses. However, fine concept of absolute strength
is irrelevant to fine SAL nodel as response learning is assured to be com—
plete and perfect.

In fine C-F-A nnocel a precise interpretation of associative strength
depends upon fine, as yet unspecified, net building routine. That routine
will be discussed in greater lengfin later in finis chapter. It is appro—

priate to speculate about its effects on associative strength at finis time.

 

 

 

 

 

1‘45

In building his nenory for a future free association task a S is
not confronted wifin repeated pairs or lists of fine sane objects. It is
finis repetition in fine verbal learning studies wwhich allows associative
strength to be built up by reinforcenent and/or contiguity. In free
association it does not seem likely finat a person will experience the
pairs of words often enougn to permit an inorenental fineory of associative
strength to function faster finan the negative effects of time. An incre—
nnental theory specifies finat each succeeding occurrence of a S-R pair
contributes (less) to fine strength of fineir association. Elapsed tine
between successive pairings Operates to decrease the strength of asso-
ciation. One possible way a person can build up a nennory of associates
is by neans of One-trial learning and contiguity. That is, whenever a
pair of worth is experienced togefiner it is associated conpletely (if it
is associated) and synnnnetrically. This approach eliminates fine need for
several presentations of S-R pairs and for reinforcenent. It places a
great deal of reliance on oe-trial learning (e.g. Rock 8 Heiner, 1959;
Estes, 196“) and associative synnetry (Asch 8 Ebenholtz, 1967; Horowitz,
Norman 8 Day, 1966). This approach would also reverse the order of the
two stages of verbal learnings (Underwood 8 Schulz, 1960) described in
chapter one.

Thus , fie net building routine would operate by finding contiguous
pairs of words. If fiey neet sore criterion they are associated conn—
pletely (associative strengfin is at a nandnunn) and sunnnnetrically (each
word is bofin a stimulus to and a response of fine other nenber of the pair).
It is during processing in fine response giving phase that availability be-
cones important. Availability is part of the second stage of verbal

 

 

 

1‘46

learning (after the stages are reversed). Asch 8 Lindner (1963) found
some support for this reversal of the two stages of learning.

Strengths and Weaknesses

 

The C—F—A model succeeds in that it is the first IPM of free asso—
ciation behavior. It posits specific deterministic processes which Operate
on the principles of association. The model operates in a parallel node
to evoke a string of response to a stinnulus word. Some of the internal
processing depends upon a word's I-AV. The C-F—A model is the first of the
related IPMS to do this. Neither EPAM, WEPAM, nor SAL incorporated I—AV
or parallel processing in any direct fashion. Nor do the earlier models
employ a type of mediation found in this model.

What are the major limitations of the model? There are two important
limitations which serve to prevent an adequate assessment of the validity
of the model. The first is the lack of a net building routine. Without
such a routine the model must work on an a_d 1392 memory which (thougn satis—
factory for testing fine operation of the model) makes any direct compar-
isons between model and human output specious. If the verbal memory were
created by an adequate net building routine, it would be possible to assign
to each word in the memory some measure of its connotative meaning. (One ,
type of routine could build up such values as the net was constructed.
Another version of the routine would not include these values, but if fine
memory were reasonable, fine values could be obtained from normative data.)
With such a routine it wwould be possible to attempt to replicate by simula—
tion some of the experiments described in the first chapter. A net building
routine is essential if fine model is to be tested in terms of event validity
(q.v. Hermann, 1967). The next sections of finis chapter contain approaches

toward fine solution of these problems.

 

1‘47

The other limitation related to validity assessment has less to do
with the nodel & s_e. For practical reasons it was impossible to run
more than one simulation of the nodel. That one simulation was a success
insofar as the operation of the nodel is concerned. From fine simulation
it was possible to see the effects of the interactions of fine components
of fine various routines. It could be determined that responses were evoked,
parallel processing occurred and a form of nediation took place. On the

other hand, one simulation is not sufficient to permit an adequate testing

 

of fine nodel's sensitivity to different initial conditions. It is impor—
tant to know fine effect upon output of different nenory content and struc—
ture and of different values of fine paraneters.

In addition to these major limitations, several lesser difficulties
are apparent from an examination of fine simulation. When fine menory was
constructed it was asstm'ed finat 200 simulated time units would be suffi-
cient for fine task. Since it took over 300 tine units to evoke five re—
sponses, it now appears that 200 tine units is far too few for net build-
ing -- especially if a net of reasonable size is constructed.

Givm a large nenory and fine additional nunber of time units needed
to construct it, then fine formula given in R-8 (Figure 23) for reducing
an item's I—AV as a function of elapsed time needs to be dnanged. It is
possible wifin a large enough renory finat many time units will elapse be-
tween successive Lses of one item. If that is so, then fine denominator
increases much nore rapidly finan the numerator in the formula given in
R—8. It will be possible, under such conditions, for an item's I-AV to be
always reduced to a level below any of fine thresholds -— leading to a nod-

el which does not evoke any responses.

 

This problem is a weakness of the particular formula used in R-8.
But other formulae are possible and reasonable. All should, however,
decrease I-AV as a function of elapsed time. This is important if re-
cency is to play a role in the determination of the value of I-AV.

A definitive evaluation of the C—F—A nodel at this time is not pos-
sible. Wlnat is possible is to begin to assess fine validity of the nodel
at different levels. This section of the dnapter offered such an assess-
ment. The nodel is at a level of development cannon to IPMS.

For a simulation of even noderate conplexity, it is such a

considerable achievement to get a 'dry run' version working

that investigators often do not pitch their levels of aspi-

ration much beyond that point. That a nodel may work well

on sinple illnstrative data carried finrougn a few represen-

tative steps , however, does not at all guarantee that it

will behave properly when run full-scale wifin a large body

of data. (Abelson, 1968, p. 307).

Nothing has been proven by fine nodel or fine simulation, but the mo—
del does exist at some higher level of credibility. Some insight has
been gained and now some patience is needed to continue the investigation

with different versions of “fine nodel and with additional simulations .
Some extensions of fine model are described in the next section.

Extensions of the C-F—A Model

While finere are many possible extensions to finis nodel, obviously
all of them cannot be discussed here. The nost appropriate extensions
to explore are finose whidn are needed immediately if fine nodel is to be
developed further. In finis case, a net building routine is central to a
better C-F-A nodel.

An initial net building routine should be limited to assign learning .

At this time it looks as if it would be easier to construct a nenory from

 

 

 

 

.llllll.lu I | l

 

word pairs or sentences finan from word—object pairs. Sign learning
(word—object) is very inportant to a realistic C—F—A nodel. Probably
TABLE is associated with CHAIR because of non-verbal co—occurrences of the
objects finemselvas instead of verbal co-occurrences in fine spoken or writ-
ten language. Some important preliminary work has been done (e.g. Minsky,
1963; Evans, 1968) which points fine way toward a sign learning net build-
ing routine. The oonplexity of these approadnes is beyond the scope of
fine current nodel.

Consequently, let fine input to the nodel be a series of word pairs.

If a word pair is perceived finen eadn item will be learned oonpletely

 

and each word will be associated synmetrically. Thus, when a word pair

is peroeived each word would be conpared with existing first itens in the

nemory. If the word did not exist it would be discrinninated through the

net and added. In addition each word would be added to the top of fine

ofiner word's response list -~ associating finem. Since the response giving

routine effectively manipulates fine value of a word's I-AV, it is not

necessary for the net building routine to treat I—AV in any conplioated

fashion. One possibility would have an item's I-AV raised by one (up to

some limit) each time that item is processed. ‘
Suppose the input to the net building routine were English sentences.

A simple expedient would be to treat fine sentence as a collection of all

possible word pairs. Irrespective of fineir sequential order, all words

would be associated with each ofiner. It would be desirable to include

sore effect of contiguity in this by forming better (stronger, nore likely)

associations between words closer together in fine sentence, but the model

a currently envisaged has noway to Cb this.

 

 

 

 

 

150

If the sentence were to be treated as an entirety refiner than a
collection of word pairs finen parts of a vast body of knowledge about
modern linguistics, psydnolinguistics, and conputer understanding of
natural language becomes pertinent. Some of this literature is appli-
cable if it sheds light on fine problem of selecting and associating words
from a sentence. To critically review or summarize these areas here is
impossible. i-bwever, one interesting contribution will be discussed.

It was chosenn not only because it contributes to the theory of a net

 

building routine, but also became it is a relatively nodern, working IPM.

The nodel is Raphael's (1968) SIR -- Semanitc Information Retrieval.
SIR's nenory is basically unstructured, consisting of words with asso-
ciated property lists (much like fine C-F—A nodel). Property lists con-
tain ofiner words and fine relationship between fine first word and each of
the ofiners. The SIR nodel atten'pts to "understand" natural English.

Given some input sentences fine nodel detemnines fine relationships between
the words. At present fine relationships it can process ,are set-inclusion,
part—whole, numeric quantity, set nenbership, ownership, and spatial
arrangenents.

Wifin a developed nenory, SIR answers some questions posed to it, dem—
onstrating its "understanding" of English. Suppose SIR were given as
input four sentences: (1) Every boy is a person. (2) There are two hands
on each person. (3) John is a boy. And, ('4) each hand has five fingers.
Through a limited analysis of syntax fine nodel finds subset—superset rela—
tions (e.g. boy—person), subpar't relations (e.g. hand-finger) and ofiner re—
lations anong fine content words of each sentence. When queried, "How many
fingers does Jdnn have?", SIR responds, "The answer is 10". (q.v. Raphael,
1968, pp. 65-66)

\
.5 a. 4 .—_.- -4... M ‘fi .. .~ " ' “
~ ‘ . ' .a., ’ n ‘ .

 

 

151

 

To employ some of SIR's principles in a C-F-A nodel would require
an analysis of the input sentence into the syntactical or logical rela-
tions anong the words . A small nurrber of important dyadic relations
(or a hierarchy of relations coupled with a limited amount of process-
ing tine) would be used. The C-F—A nodel could finen associate word pairs
found to be related. Still later, fine nodel could store fine type of fine
relationship between members of a word pair and use finis information dur— m.
ing fine response giving phase. Responses given in a C-F-A task differ in 3
their relationship to fine stimulns word. It is known that these relation- ‘1
ships differ as to fineir relative frequency of occurrence and fineir asso—
ciated response latencies (e.g. Karwoski 8 Schachter, 19%). By compar-
ing the performances of versions of fine extended ’C-F-A nodel it might be
possible to determine if fine enpirical findings were due to an input-
storage process , an output routine, or some ofiner situation.

Procedures for net building and handling sentences as input are
the nost important of possible future extensions to C-P-A. Other weak—
rnesses of fine current nodel need to be corrected. The nodel as now de-
scribed is very inefficient in net organization and discrimination learn-
ing. The nodel must learn how to lean. Certain rearrangements in the
net structure ought to occur as a function of processing. The restructur-
ing would make later retrievals nore efficient than earlier ones . Wynn
(1966) inplements several of finese efficiences in his nodel. WEPAM permits
different pafins througn fine net to fine sane terminal. It also builds loops
to bypass earlier nodes in fine discrimination net when finose nodes are
redundant or "get in fine way." For exanple, fine letters' portion of the
C—F-A model's hypofinetical nennory (Figure 26) is constructed by discrim—

inating attributes of letters as finey appear temporally at the beginning

 

 

 

. nwd'w,‘ -

o ‘nnp—I

   

-' new“. ”#1"? .3! T i’.’ "1'3 ‘ '1 fly“;
' :4‘2‘. ~. ~~ ' ,

R— _..._;.-..._;_. —___—,_____- _.._.

152

of the eXperiment. If a letter occurred late in the texrporal sequence,
its terminal would be deep in the net. In Figure 26 the terminal for
"E" is none testing nodes into fine net. Since "LI" appears frequently
in English it is very inefficient for all processing to pass through the
preceding nodes. Wynn's methoch would allow for a more direct access to
"B" as testing nodes and paths are changed as a function of processing.

Anofiner need for fine nodel is for it to handle context. It is
known (q.v. Howes 8 Osgood, 1951+) that different free associates are
given when the stimulus word is preceded by other words (verbal context).
Also it is comm for people to nodify fineir verbal behavior depending
upon where finey are (e.g. in a church or a locker roon) orwho they are
with (ones parents or ones peers). The current C—P-A nodel is fineoretically
equipped to deal with fine problem of context. Whenever a response is asso-
ciated with a stinulus item, its property set could be augrented to con-
tain sore coding of fine context. Appropriate attribute testing nodes
included in fine nemory net would test for the presence of desired or un-
desired contexts and thereby, nodify the output.

This section dealt mainly with those few najor additions needed if
fine C—F—A nodel were to build its om nennory as a function of its verbal
experience. Especially difficult will be handling sentences in a manner
which utilizes the syntactic relations anong fine words. The problem of
making verbal associations from the physical world is very inportant but
not considered. The ability to abstract verbal relationships from the
physical world is a najor need for any conprehensive nodel of net build-
ing for free association. These problens of net bnnilding would completely
overwhelm any first attempt at a C—P—A model and therefore, the mission

of a net building routine was deliberate.

 

.I
I'lli.

 

 

153

Sons Future Enqplorations with fine C-F-A Model

If work wifin fine C-F-A uochl is to be continued, finen initial
efforts need to deal wifin sensitivity-parameter testing and net build—
ing. Also, itappearsfinatitisnecessarytocodefineIPMsoitcan
beprocessedbymadninerafinerthanbyhand. Thesetypesofthings are
reasonhbly straigntforward in concept, if not in practice , and sons of
finem have been disossed earlier. This section of fine chapter is con-
cerned with more distant explwations of an extended C—F-A nodel . A
caution here seens necessary. There is no 3 pri_9_r3 reason to believe
that the best way to proceed with future study of C-P-A behavior is
by studying a nodel refiner finan fine subject matter pg; s_g (q.v. Kaplan,
19m, 1). 279).

The phenonena of one-trial learning has been discussed earlier. It
was relevant to fine presentation of fine BPAM—WEPAM—SAL nodels and the
extended C-F-A nodel. Both verbal (e.g. Rock 8 Heiner, 1959) and mafin-
ematical (e.g. Estes, 1961+) arguments have been used to present and defend
the one-trial position. A critique of one-trial learning showed that
both approaches were either non-supportable or indistinguishable from an
incremental fineory position of learning (Postman, 1963). that can be
concluded? The IPMs treat one-trial learning as a useful or needed
concept, while a critical review concludes otherwise. Part of fine prob-
lanmignt be due to fine different types of nodels. The incremental fineory
is verbal and it does not contain encplicit internal processes (thougn it
does contain explicit internal variables such as habit strength). Conse-
quently, Pcsfinan not base his tests of fine one-trial fineory on overt
responses made to specified stimfli. It is particularly difficult to

 

 

151+

separate acquisition from response giving since there is no way to
ascertain acquisition wifinout studying the responses made.

An IPM is not so limited. This type of model can separate acqui—
sition from output mechanisms. Thus, an IPM (such as C-F-A) can employ
one—trial learning but have a response pattern identical wifin those pro-
duced by an incremental fineory. It is fine processes between the acqui-
sition and response giving phases whidn permit finis.

This illustrates fine difficulty encountered whenever an IPM and
anofiner type of model are compared. It also illustrates some of the
value if IPMS. In terms of one-trial learning researchers can build
various processors which operate between acquisition and response giv-
ing (the C-P-A model is one possibility). Througn testing it may be pos-
sible to settle some of fine differences between a one-trial and incre-
mental position. It is also likely, that work with an IPM may offer
anofiner possibility, gig, finat differences between fine two theories re—
flect two different abstractions from a more corplex model.

Similarly , fine C-F-A model and the EPAM—WEPAM—SAL models contribute
to a more complex version of fine contiguity versns reinforcement contro-
versy (q.v. McGeoch 8 Irion, 1952, p. lt6f.). The 1% require only tem-
poral or Spatial contiguity for acquisition , but need sore effects of
frequency (including reinforcement in fine learning models) before the
changes in fine memory net are developed sufficiently for further pro-
cessing -- such as response giving. In fine C—F—A model fine further pro—
cessing raises an item's I-AV above fine minimal finreshold. In the ver-
bal models of learning (e.g. Guthrie, 1952) or in fine more formalized
theories (e.g. Hull, 19%) it is difficult to separate the internal

 

   

155

processes and fine temporal sequence of finese processes (q.v. Jenkins,
1965, p. 27). IBIS clearly outline such processes and, finereby, offer
approaches toward a more general formulation of the problem.

There are several areas into which a more general C-F-A model could
explore. One possibility is verbal satiation . Verbal satiation names
fine phenonenon of a loss or change in meaning of a word as a result of its
continued repetition. In fine C-F-A model, the experimental situation ...
would consist of a word associated with itself several times. This results in
in fine most recent responses to fine word being fine word itself.

Without simulating finis condition it is impossible to specify exactly
what would occur. However, two possibilities seem likely. First of all,
finere should be an increase in elapsed time between fine presentation of
the word as a stimulus and fine first evoked response. In fine Response
Giving Routine (Figure 23) potential responses are examined serially from
fine most recent to fine least recent. The recent potential responses are
identical with fine stimulus word. Verbal satiation studies do not allow
the stimulus word to be given as a response. In C—F-A, R-lO prevents the
stimulus word from being given as a response. Thus, all of fine reSponses
which are identical with fine stimulus word must be processed before any
ofiner word becomes a candidate for evocation .

The second possibility derives from fine first. Suppose fine first
response to CAT is DOG. In fine satiation condition, many copies of CAT
are placed before DOG on fine response list. As noted above, finere should
be some elapsed time before DOG is fine current candidate for evocation.
The more time elapsed, fine greater DOG's I-AV will be reduced (q.v. R—8).

In such a situation it is possible finat DOG is no longer available as a

 

   

 

156

I‘e'Sponse and may serve at most as an internally mediated stimulus word.
The exPectation finen would be for an idiosyncratic response to CAT .

The literature of verbal satiation is not in agreement (of. Lambert
8 Jakobovits, 1960; Jakobovits 8 Lambert, 1961). In fact Yelen and
Schulz (1963) could not find much support for the existence of verbal
satiation. The C-P-A model is not equipped to deal wifin a word' 8 loss in
meaning measured by semantic differential rating scales (as in fine above
three studies). If loss in meaning is measured by increased latency of
response and lack of commonality of response , finen studies by Wertheimer
and Gillis (1958), and Smith and Raygor (1956) are applicable. These
studies show finat men satiation occurs less cannon associates are given.
One hypothesis derived from fine predicted behavior of fine C-F—A model is
finat more internal processing (caused by storing many stimulus words as
potential responses) produces a greater chance for idiosyncratic overt
responses. Fillenbaun (1963) found that when _S_s repeated fine stimulus
word for four seconds finey had less loss of meaning (measured by common—
ality of response) finan words repeated for one or three minutes . It
should be noted, however, finat fine difference between the one and three
minute conditions was small and not in the direction predicted.

Another area in which fine model ougnt to explore more fully is mean-
ing. Initially it was hoped that this model could relate C-F-A behavior
wifin a mediational approadn to word meaning. It turned out that the scOpe
of finis problem was greater finan expected and could not be dealt with be-
fore a model which produced free associates was develOped. The current

model employs mediation in its processing. It also gives fine meaning of

a word -- either defined interverbally or relationally . The model does

 

 

 

 

 

840!“ "
-
-
0 ‘ N
we-
_\

 

 

 

157

not carbine finese approaches in order to permnit a mediational measure
of a word's meaning. In terms of initial plans, this is a serious fail-
ing of this study.

It is still not clear how fine C-F—A model could be modified to in—
corporate such a measure. Suppose , for exanple, the approach chosen was
fiat of Osgood, Suci and Tannenbaum (1957) as presented in chapter one.

In their fineory the mediators are not letters of a word (as in EPAM and
C-F-A) , nor are finey indicators of a semantic or linguistic relationship
(as in Reitman's Argus (1965) or SIR), nor are they words used to trans-
late between languages (as in EPAM III). Rafiner they are "light-weight"
components of fine response to a word. In assign learning wifin C-F—A

this would entail part of fine response to one word mediating the response
to fine second word. A variant of this procedure may occur in fine current
model. Suppose Ri are fine ordered set of potential response to a stimulus
word 31- If an unknown word 82 is paired with $1 finen each word will be
the topmost potential re5ponse for fine ofiner word. If an interverbal
meaning for $2 is asked for, it is likely finat 81 and its responses will
be used. The C-F-A model does not employ conponents of the response to
81; it uses 81 in full. Consequently, it appears that fine principle of
internal mediation is incorporated directly in the C—F—A model . What is
still missing is a mefinod for obtaining a quantitative measure of a word's
meaning as a function of fine mediation .

T\wo alternative mefinods for obtaining finese measures seem worth de-
veloping. In one case, an internal semantic space is hypofinesized. For
each dimension of the space fine net building routine determines a word' 8
location relative to finat dimension. This is the heart of Osgood's theory.

The input would be definitional or descriptive messages about an unknown

 

mm in terms of words already in the memory. The known words are located

in semantic space. By means of an as yet undefined processor each dimen—
sion of each old word in the definitional input would contribute to fine
location of fine new word. (Parenthetically, it should be noted finat "lo—
cation" is used here figuratively. The structure of fine discrimination
net need not be changed from finat of the current model. All that is needed
is for fine relative locations on each dimension to be added to fine property
set for a terminal.) Once this difficult part is corpleted it is concep—
tually easy to include in a net, attribute nodes which test for values of
finese locations. In addition, fine routines which retrieve response could
have a series of thresholcb testing fine positional indicators. Only finose
words "near" another word could serve as a mediator for that word. Thus,
I—AV would determine whether or not a response will be evoked, and relative
location in semantic space would be fine new method (of. Figure 23, R—l7)
for controlling mediation.

The other alternative takes a different tack. It does not assume
internal processing (at fine time assign learning occurs) produces fine
measures of location. Instead, it assumes finat the measures are a function
of fine measuring instrument. In this situation, the C—F—A model would
require a routine to respond to a semantic differential rating scale. Sup-
pose DOG were being rated on a "good-bad" scale. Using Quillian's (1967,
1968) procedure, markers would start at fine three terminals: DOG, GOOD,
and BAD. Each response to eadn stimulus, eadn response to each response,
etc. would be examined until fine markers crossed pafins. Some weighting
scheme would determine where DOG ougnt to be rated on the scale as a func-
tion of elapsed time to intersection or number of terminals examined before

fine two paths met.

 

 

 

 

 

If these two alternatives could be developed it would be very inter-

esting to explore their consequences. Osgood and his associates (1957) do
not distinguish between finese two possible ways of obtaining measures of
meaning. (This distinction is similar to the one made previously about
one-trial and incremental theories of learning.) The first method Open-
ationalizes meaning as a representational mediated reaction. The second
method does not require mediation of that sort at acquisition time . The
measures of meaning and a chaining type of mediation occnms in the response
giving or test taking phases .

There are ofiner tOpics deserving exploration with fine C-F-A model .
Adults in a C-F-A experiment do not usually give obscene words as response.
Often they give rhyming reaponses and Opposites. Processes within fine cur-
rent model may contribute to an understanding of finese phenomena. Obscene
words could be handled in two ways; eifiner by treating finem similar to
stimulus words (q.v. Figure 23, R—lO) or by adding to each word's property
set a role marker indicating fine situations in which the word is permitted
to be spoken. Rhyming responses of fine CAT—HAT sort may be due to an error
in decoding fine stimulus word. CAT could be sorted to the HAT terminal.
The C—F-A model (and EPAM and WEPAM) does not check to see if fine terminal
reached matches fine input stimulus . A procedure could be included in the
Stimulus Sorting Routine which only treats an object as a stimulus if the
terminal sorted to has a first image identical with fine object. If this
condition is not met, fine terminal must be a response terminal. This type
of procedure will produce a form of response generalization. Mnen fine
error in discrimination occurs at fine first letter of fine word, rhyming

responses are possible.

 

 

 

 

160

Finally there is the problem of opposites: since they do not often
occur contiguously it is difficult to explain how one can evoke fine ofiner.
If responses are mediated by worth with low I—AV (as in fine current model)
opposites will occur. For example, if in building the net HOT and WATER
are associated together and COLD and WATER are similarly associated, finen
over time it is possible that the two adjectives will have a higher I-AV
than WATER because finey are experienced more frequently. When presented
wifin HOT, the model's most recent associate might be WATER which is not
strong enough to be evoked. Acting as a stimulus, WATER, evokes COLD as
a response. This is fine position taken by Pbrowitz, Brown and Weissbluth
(1961+) who showed that this interpretation based upon I-AV is not equiv—
alent with a simple chaining paradigm.

m

This last part of chapter five pointed fine way for some future explo—
rations with fine cm'rent and extended C—F—A models. One value of the C—F-A
model (and other IPMS of cognitive processes) is its capability to tempo—
rally separate internal processes. This viewpoint may contribute to a more
fundamental understanding of fine phenonena of verbal behavior. Some of the
phemmena discussed in this section are satiation, meaning, and evocation
of opposites.

The purpose of this chapter was to evaluate fine model and its simu—
ulation in terms of strengths and weaknesses. There are several major
weaknesses of fine model. Of considerable importance is the lack of any
net building routine. This forced fine use of an a_d h_o_c memory and made

it impossible to tat several aspects of fine model's validity. In addition,

 

 

 

 

 

151

because the model was not coded for computer processing, it was only
possible to execute one hand simulation. Several executions are needed,
however, if sensitivity and parameter testing is to be conducted. The
last major weakness of the model is its inability to produce measures of
a word' 8 meaning based upon a representational mediation paradigm.

On fine positive side, C-F—A is fine first working model of free asso-
ciation behavior. Lending weight to its face validity are fine facts that
(1) it operates upon a hierarchically organized verbal memory, (2) in a
parallel mode, and (3) evokes unpredictable reSponses. (1+) Item avail-
ability is treated directly in finis model (but not in fine earlier ones).
The treatment of I—AV corresponds closely with what is known about the
variable. In addition the model (5) employs a form of stimulus mediation
which is important to its processing. Finally, (6) the model learns.

The contents of the net are changed as a function of earlier processing
and these changes affect later outcomes .

Later in the chapter possible net building routines were considered
briefly along with the problems of handling sentence input for assign
learning. It was noted that one of fine advanntages of an IPM of cognitive
behavior is its ability to temporally separate different processes . A
possibly important role for these IPMS is to make finese explorations in
order to shed light on existing fineoretical controversies -— such as one—
trial learning. More Specific directions. for future exploration were also
mentioned.

Finally, it must be stressed finat fine C-F-A model is a first try,

a partially justified guess. As Popper (1962) emphasizes scientific

knowledge progresses by finese conjectures and by criticisms of them.

 

162

Science gains if fine model is refuted and it also gains if it can not,
as yet, be refuted. Both conjectures and refutations are central to the
undertaking. At fine conclusion of this study finere is, at best, an interim

model of C-F-A behavior, and a preliminary evaluation of it. That is a
reasonable beginning.

.‘n ’I.

-!;J- . u.“-

 

 

WES

 

"'\"(7v' -. -

 

 

 

 

 

 

REFERENCES

Abelson, R.P. Mathematical models of fine distribution of attitudes
under controversy. In N.O. Fredericksen 8 H. Gulliksen (Eds.),
Contributions to mathematical psycholog. New York: Holt , 1961: .
Pp. 191—160.

Abelson, R.P. Simulation of social behavior. In G. Lindzey 8
E. Aronson (Eds.), The handbook of social psychology. (2nd ed.)
Vol. 2. Researdn methods. Readingji'assa Addison-Wesley, 1968.
Pp. 279-356.

 

Ackoff, R.L. Scientific method: Optimizing applied research decisions .
New York: Wiley, 1967. f

Alston, W.P. Philos% of language. Bnglewood Cliffs, N.J.:
Pmntim- , 0

Arrow, l<.J. Mafinematical models in the social sciences. General Systems,
1956, _l_, 29—97.

 

Asch, S.E., 8 Ebenholtz, S.M. The principle of associative symmetry.
In L.A. Jakobovitz 6 M.S. Miron (Eds.), Readin s in the s cholo
of lan . Englewood Cliffs, N.J.: Prentice—Hall, I957.
““‘Eu'égEPp. 990—51 .

Asch, S.E., 6 Lindner, M. A note on "strength of association."
J. Psychol., 1963, §§_, 199-209.

 

Baker, F. B. The internal organization of computer models of cognitive
behaVior'. Behav. Sci., 1967, 3.3, 156-161.

 

Berlo, D.K. The process of communication. New York: Holt, 1960.

 

Berlyne, D.E. Mediating responses: A note on Fodor's criticisms.
J. verb. Learn. verb. Behav., 1966, _5_, l#08--l+ll.

Bloomfield, L. w. New York: Holt, 1933.
Boulding, K.E. The lm_nag§' . Ann Arbor, Mich.: Univ. of Mich. Press, 1956.
Bousfield, W.A. The problem of meaning in verbal learning. In C.N. Cofer

(Ed. ) , Verbal learniniand verbal behavior. New York: McGraa-Hill ,
1961. Pp. 81.910

 

 

Boner, G.H., 8 Trabasso, T.R. Concept identification. In R.C. Atkinson
(Ed.), Studies in mathematical gpsycholog. Stanford: Stanford
Univ. Press, 1961;. Pp. 32-99.

 

Brown, R. Words and things. New York: Free Press, 1958.

 

Carroll, J .B. Languagg and thong} Englewood Cliffs, N.J.: Prentice-
Hall, 1961+.

 

Carroll, T.W., 8 Farace, R.V. Systems analysis, computer simulation,
and survey research. East Lansing, Mich: Computerinstitute for
Social Science Research, 1968.

 

 

Cofer, C.N. Comparison of word associates obtained by the methods of [a
discrete single word and continued association. Psychol. Rep. , ,~
1958, 3, 507-510. 1

: '3. 1: .-.~

.‘I
w .

Cofer, C.N. Comments on Professor Deese's paper. In C.N. Cofer (Ed.),
Verbal learning and behavior. New York: McGraw—Hill, 1961.
Pp. 31-38.

 

Cofer, C.N. , 8 Foley, J .P. Jr. Mediated generalization and the inter-
pretation of verbal behavior. I. Prolegomena. Psychol. Rev. ,
1992, i9, 513-540.

Crawford, M.P. Dimensions of simulation. Amer. Psychol. , 1966, 2_l_,
788-796.

Creelman, M.B. The experimental investigation of meaning: A review
of the literature. New York: Springer, 1966.

 

Davis, M. Computability and vunsolvability. New York: McGraw-Hill, 1958.

 

DeBurger, R.A. , 8 Donahoe, J. Relationship between fine meaning of
verbal stimuli and their associative responses. J. verb. Learn.
verb. Behav., 1965, H, 25—31.

 

Deese, J. On the structure of associative meaning. Psychol. Rev. ,
1962, 99, 151-175.

 

Deese, J. The structure of assgciations in language and thought.
Baltimore: Johns HOpkins Press, 1965.

Bennett, D.C. Machine traces and protocol statements. Behav. Sci. ,
1968, _1_11, 155—161.

Dubin, R. Theory building. New York: The Free Press, 1969.

 

Ervin, 8.1!. Changes with age on the verbal determinants of word
associations. Amer. J. Psychol., 1961, 2:, 361-372.

Erwin-Tripp, S., 8 Slobin, D. Psycholinguistics. Ann. Rev. Psychol. ,
1966, ll: 935-u7u.

 

 

165

Estes, W.K. All-or—none processes in learning and retention. Amer.
Psychol., 1961+, _1_9, 16—25. ‘—

Evans , T.G. A program for the solution of a class of geometric—analogy
intelligence-test questions. In M. Minsky (Ed. ) , Semantic infer»
nation processing. Cambridge, Mass.: M.I.T., 1968. Pp. 271—353.

 

Feigenbaum, E.A. An information processing fineory of verbal learning.
Santa Mama: The RAND Corp., 1959. P-1817.

Feigenbaum, E.A. The simulation of verbal learning behavior. In E.A.
Feigenbaum 8 J. Feldman (Eds.), Congguters and finougnt. New York:
McGraw—Hill, 1963. Pp. 297-309.

 

Feigenbaum, E.A. , 8 Feldman, J. Introduction. In E.A. Feigenbaum 8
J. Feldmnan (Eds.), _Oggguters and thought. New York: McGraw—Hill,
1963. Pp. 1-8, 269- .

Feigenbaum, E.A., 8 Simon, H.A. For ttin in an association memo .
Santa Monica, Calif.: The RAND gory, 1961. P—73II.

Feigenbaum, E.A. , 8 Simon, H.A. , Generalization of an elementary perv
ceiving and memorizin machine. Santa Monica, Calif.: The
RAND Corp. , 1962. _25'6'5'13- '."'(a"5

 

Feigenbaum, E.A. , 8 Sim, H.A. A theory of fine serial position effect.
Brit. J. Psychol., 1962, _5_3_, 307-320. (b)

 

Feigenbaum, E.A. , 8 Simon, H.A. Brief notes on the EPAM theory of
verbal learning. In C.N. Cofer 8 B.S. Musgrave (Eds.), Verbal

behavior and learnin : Problems and processes. New York: McGraw—
PHIL, 1963. 5. 335E335. (a)

Feigenbaum, E.A. , 8 Simon, H.A. Performance of a reading task by an
elementary perceiving and memorizing program. Behav. Sci. , 1963,
8, 72-76. (b)

 

 

Feldman, J. Simulation of behavior in fine binary choice experiment.
In E.A. Feigenbaum 8 J. Feldman (Eds.), Computers and finought.
New York: McGraw-Hill, 1963. Pp. 329—3M6.

Fillenbaum, S. Verbal satiation and clnanges in meaning of related items.
Jo WI‘b. Learn. wzlbo WVO, 1963, 2., 263’2710

Fodor, J.A. Could meaning be an r ? J. verb. Learn. verb. Behav.,
1965, it, 73—81.

 

Frijda, N. H. Problems of computer simulation. Behav. Sci., 1967, _1_2_,
59-67.

 

Garskof, B. , 8 Houston, J. Measurement of verbal relatedness: An
idiographic approach. Psychol. Rev., 1963, 3, 277-288.

 

 

166

Garvin, P.L. (Ed.), Natural languagg and fine computer. New York:
McGraw-Hill, 1963.

 

Gladun, V.P. Memory organization for list processing. Cybernetics,
1966, 3, 26-29.

 

Green, B.F. Jr. Computer models for cognitive processes. Psychometrika,
1961, _2_6_, 85—91.

 

Green, B.F. Jr. , Digital corputers in research: An igtroduction for
behavioral wand social scientists. New York: McGrav—Hill, 1963.

Green, B.F. Jr. , et a1. Baseball: An automatic question answerer. gn-
In E.A. Feigenbaum 8 J. Feldman (Eds.), Computers and thought. ’
New York: McGranI—Hill, 1963. Pp. 207-216.

.nELM
.
l‘

Gregg, L.W. , 8 Simon, H.A. Process models and stochastic theories of
simple concept formation. J. math. Psychol., 1967, 3, 2116-276.

Guthrie, E.R. The psycholog of learniing. Rev. Ed. New York: Harper,
1952.

Hart, R.D. Summary of hierarchical system descriptions. Unpublished
manuscript, Michigan State University, 1967.

Hermann, C. Validation problems in games and simulations with special
reference to models of international politics . Behav. Sci. ,
1967, _1_g, 216-231.

Hintzman, D. L. Explorations with a discrimination net model for paired-
associate learning. J. mafin. Psychol., 1968, 6_, 123—162.

 

Horowitz, L.M., Brown, Z.M., 8 Weissbluth, S. Availability and fine
direction of associations. J. enp. Psychol., 196“, 6g, 5141—599.

Horowitz, L.M. , Norman, S.A. , 8 Day, R.S. Availability and associative
symmetry. ngchol. Rev., 1966, 13’ 1-15.

 

Howe , E. S. Uncertainty and ofiner associative correlates of Osgood' s

D“. J. verb. Learn. verb. Behav., 1965, 3, L198-509.

Lknwes, D. , 8 Osgood, C.E. On fine combination of associative probabilities
in linguistic contexts. Amer. J. Psychol., 19514, 61, 2u1-258.

 

 

Hull, C.L. Principles of behavior. New York: Appleton-Century, 1993.

Jakobovits, L.A. Comparative psycholinguistios in the study of cultures.
mt. Jo PSYdIOle, 1966, i, 15-370

Jakobovits, L.A. , 8 Lambert, W.E. Mediated satiation in verbal transfer.
J. eg. Psychol., 1962, 61, 396-351.

 

I} F; {-3

‘,,.,._—.. was s .6

 

Jeri

Jen'-

Jer

Je:

Je

 

167

 

Jenkins, J .J . Commonality of association as an indicator of more
general patterns of verbal behavior. In T.A. Sebeok (Ed.),
Style in language. Cambridge, Mass.: M.I.T., 1960. Pp. 307-329.

 

Jenkins, J .J . Mediated associations: Paradigns and situations. In
C.N. Cofer 8 B.S. Musgrave (Eds.), Verbal behavior and learning
Problems and processes. New York: McGraw-Hill, 1963. Pp. 710-295.

Jenkins, J.J. The learning fineory anproach. In C.E. Osgood 8 T.A.
Sebeok (Eds.), Psycholin ‘stics: A survey of theogy and research
roblems. Blooniington, 3.: Indiana University Press, 1965.
E)'_"2""1>p. 0-314.
1%

Jenkins, J .J . , 8 Russell, W.A. Annual technical report: Basic studies
on individual and P behavior. Contract N0. N8 onr—66216
between Office of aval Research and University of Minnesota, 1956.

Jenkins, J .J . , Russell, W.A. , 8 Suci, G.J. An atlas of semantic profiles
for 360 words. Amer. J. Psychol., 1958, 21, 688-699.

 

Jung, J. Experimental studies of factors affecting word associations.
Psychol. 6111;, 1966, §_6_, 125-133.

 

Kaplan, A. The conduct of ‘ ui : Methodology for behavioral science .
San Francisco: Chan er, 19 1+. P

 

Karwoski, T.F. , 8 Schachter, J. Psychological studies in semantics: III.
Reaction time for similarity and difference. J. soc. Psychol. ,
19118, 38, 103-120.

Lambert, W.E. , 8 Jakobovits, L.A. Verbal satiation and changes in the
intensity of meaning. J. enp. Psychol., 1960, .59.: 376-383.

 

Lindsay, R.K. Inferential memory as the basis of machines which
understand natural language. In E.A. Feigenbaum 8 J. Feldman (Eds.),
Computers and finouggt: New York: McGraw-Hill, 1963. Pp. 217-233.
(a3

Lindsay, R.l<. Information processing fineory. In W. S. Fields 8 W. Abbott
(Eds. ), Information storage and neural control. Springfield, Ill. :
C. C. Thomas, 1963. Pp. 393.56. (b)

 

Mackay, D.M. Towards an information-flow model of human behavior. In
W. Buckley (Ed.), l‘bdern s tems research for fine behavioral scientist.
Chicago: Aldine, 1968 5. 559-568.

McGeoch, J. A., 8 Irion, A. L. The psycholo ogy of human learning. (2nd. ed.)
New York: David McKay,1952.

McNeill, D. A study of word association. J. verb. Learn. verb. Behav.,
1966, §_, 598-557.

 

 

168

Miller, G.A. The magical number seven, plus or minus two. Psychol.
Re_v., 1956, Q, 81-97. .

Miller, G.A. Comments on Professor Postman's paper. In C.N.Cofer 8
B.S. Musgrave (Eds. ), Verbal behavioral and learning: Problems and
processes. New York: McGraw—EII, 1963. Pp. 321-929.

 

Miller, G.A., Galanter, E., 8 Pribram, K.H. Plans and the structure of
behavior. New York: Holt, 1960.

 

Miller, J .G. The individual as an information processing system. In
W.S. Fields 8 W. Abbott (Eds.), Information stora and neural control.
Springfield, 111.: C.C. Thomas, 1963. Pp. 351-325.

Minsky, M. Steps toward artificual intelligence. In E.A. Fiegenbaum 8
J. Feldman (Eds.), Computers and thougnt. New York: McGraw—Hill,
1963. Pp. HOB-H50.

 

Minsky, M. , 8 Papert, S. Perceptions: An infioduction to comeutational
geometg. Cambridge, Mass.: 'M.I.T., 1969.

Morris, C., Sigs, langu_ag_e, and behavior. Engler Cliffs, N.J.:
Prentics—Hall, 19%.

Mowrer, O.H. The psychologist looks at language. Amer. Psydnol. ,
1959, 3, 660-699.

Murdock, B.B. Backward associations in transfer and learning. J. en_<p.
Psychol., 1958, 35, lll-llu.

Newell, A. Some problems of basic organization in problem-solving
programs. In M.S. Yovits, G.T. Jacobi, 8 G.D. Goldstein (Eds.),
Self-organizing systems 1962. Washington: Spartan, 1962. Pp. 393-
l+23.

Newell, A., Shaw, J.G., 8 Simon, H.A. Elements of a theory of human
problem solving. Psychol. Rev., 1958, E, 151-166.

Newell, A. , 8 Simon, H.A. The simulation of human thought. In University
of Pittsburgn, Department of Psychology (Ed. ), Current trends in
psychological theory: A bicentennial ro . Pittsburgh: Unl—
versity of Pittsburgn Press, 1961, Pp. 152-179.

Newell, A., 8 Simon, H.A. Computers in psychology. In R.D. Luce,

R.R. Bush, 8 E. Galanter (Eds.), Handbook of mathematical s -
dnolog. Vol. 1. New York: Wiley, 1963. Pp. 361—l+28.(a)

Newell, A., 8 Simon, H.A. GPS, a program finat simulates human thought.

In E.A. Feigenbaum 8 J. Feldman (Eds.), ngputers and thouat.
New York: McGraw—Hill, 1963. Pp. 279—293. b

 

169

Newe11,A., e_t_ _a_nl_., Information processmg manual — V. (2nd. Ed.).
Englewood Cliffs, N. J. Prentice—Hall, 1969.

 

Noble, C.E. An analysis of meaning. Psychol. Rev., 1952, _5_9_, 421-1130.

Noble, C.E. Meaningfulness and familiarity. In C. N. Cofer 8 B. S.
Musgrave (Eds. ), Verbal behavior and learning: Problems and pro—
cesses. New York: McGraw—H111,l963, Pp. 76- 119.

 

Osgood, C.E. The nature and measurement of meaning. Psychol. Bull. ,
1952, 19.: 197-237.

Osgood, C. E. A behavioristic analysis of perception and language as
cognitive phenorena. In Cont roaches to co tion.
Cambridge. Harvard University Press, 1957. Pp. 75-118.

Osgood, C.E. Corments on Professor Bousfield's paper. In C.N. Cofer
(Ed. ), Verbal learnin and verbal behavior. New York: McGraw—
Hill, 1961. Pp 91-106.

 

Osgood, C.E. On understanding and creating sentences. Amer. Psychol. ,
1963, pg, 735-751.

Osgood, C. E. Meaning cannot be rm" J. verb. Learn. verb. Behav.,
1966, 5, L+02— L107. '

Osgood, C. E. , 8 Anderson, L. Certain relations among experienced
contingencies, associative structure, and contingencies in endoded
messages. Amer. J. Psychol., 1957, 7_0, l111-1420.

Osgood, C.E., Suci, G.J., 8 Tannenbaum, P.H. The measurement of meaninJ,
Urbana: . University of Illinois Press, 1957.

 

 

 

Piaget, J. The lan_gn_1_a_g_e and finougnt of the child. New York: World,
195

Pollio, H.R. Word association as a function of conditioned meaning.

J. exp. Psychol., 1963, pp, 959—960.

Pollio, H.R. Sone semantic relations among word-associates. Amer. J.
Psychol., 1969, _7_7_, 299-256.

Pollio, n.R. The structural basis of word a_s_sociation behavior. The
Hague: Mouton, 1966.

 

Popper, K.R. The logic of scigtific discovery. New York: Science
Editions, 1961.

Popper, K. R. Conjectures and refutations: The m of scientific
lcnowledgg. New York: Basic Books, 1962.

 

 

   

Portnoy, M. Conditioning of verbal behavior and its effects upon
meaning: An experimental study of positive and negative rein-
forcement applied to word associations of varying strengths and
its consequent effects upon meaning. Diss. Abstr., 1961, 22,
9911. "'

 

Postman, L. One-trial learning. In C.N. Ccfer 8 B.S. Musgrave (Eds.),
Verbal behavior and learning: Problems and processes . New York:

Quillian, M.R. Word concepts: A theory and simulation of some basic
semantic capabilities. Behav. Sci., 1967, l_2, L#104130.

 

Quillian, M.R. Semantic memory. In M. Minsky (Ed.), Semanitc informa— F”
t_i_cn processing. Carbridge, Mass.: M.I.T., 1968.—m -

 

Raphael, B. SIR: Semantic information retrieval. In M. Minsky (Ed.),
Semantic information processing. Cambridge, Mass.: M.I.T., 1968. 1
Pp. 33.13”.

 

Razran, G.H.S. A quantitative study of meaning by a conditioned salivary
technique (semantic conditioning). Seience, 1939, _9__0_, 89-90.

Reitman, W.R. Information-processing models in psychology . Science ,
1969, 199, 1192-1198.

Reifinan , W. R. Cognition and thought: An information-processing approach .
New York: Wifey, 1965.

Reim’ W.R., Grove, ROB.’ 8 SIIO‘JP’ ROG. ”$15: An infOIflEtiOD-pm-
cessing model of thinking. Behav. Sci., 19614, 9, 270-281.

 

Rock, I., 8 Heimer, W. Further evidence of one-trial associative learning.
Amer. J. Psychol., 1959, 1;, 608-611.

 

Rosenzweig, M.R., s Postman L. Frequency of usage and the Perception
cf'words. Science, 1958, 321, 263-266.

Rudner, R.S. Philosophy of social science. Englewcod Cliffs, N.J.:
Prentice-Hall, 1966.

Selfridge, O.G., 8 Neisser, U. Pattern recognition by machine. In E.A.
Feigenbaum 8 J. Feldman (Eds. ) , Computers and thought. New York:

Simon, H.A. The architecture of complexity. General Systems, 1965, _1_0_,
63-76.

Simon, H.A. , 8 Feigenbaum, E.A. An information-processing theory of some
effects of similarity, familiarization, and mean1ngfulness in verbal
learning. J. verb. Learn. verb. Behav., 1961:, 3, 385-396.

 

   

Smith, D.E.P., 8 Raygor, A.L. Verbal satiation and personality.
J. abnorm. soc. Psychol., 1956, .3, 323-326.

 

Springer, O.H., Herlihy, R.E., 8 Beggs, R.I. ‘Mathematics for manage-
ment series. Vol. 2. Advanced methods and modeIs. Homewccd, I11.:
Irw1n, 1965.

Staats, A.W., 8 Staats, C.K. Attitude established by classical condi-
timingo Jo 313mm. P§YChOl., 1958, £29 37-140.

 

 

 

Staats, A.W., 8 Stats, C.K. Meaning and m: Correlated but separate.
Psychol. Rev., 1959, _6_6_, 136-11m.

Staats, C.K. , 8 Stats, A.W. Meaning established by classical condi-
tioning. J. exp. P§ychcl., 1957, _5_l_l_, 711-80.

Tatsmka, M.M. Mathematical models in the behavioral and social sciences.
In D.K. Whitla (Ed.), W of measurement and assessment in
behavioral sciences. Reading, Mass.: Addison-Wesley, 1968.

Pp. 3.390

 

Turing, A.M. Computing machinery and intelligence. In E.A. Feigenbaum,
8 J. Feldman (Eds.), Conputers and tl'ought. New York: McGraw-
Hill, 1963. Pp. 11-35.

Underwood, B.J. Stimulus selection in verbal learning. In C.N. Ccfer 8
B.S. Musgrave (Eds.), Verbal behavior and learning: Problems and
msses. New York: P‘chraw-Hill, 1963. Pp. 33-‘48.

Underwood, B.J., 8 Schulz, R.W. Meaningfinlness and verbal learnipg.
Hnila.: Lippincott, 1960.

Vygotsky, L.S. Thought and language. Cambridge, Mass.: M.I.T., 1962.

Wertheimer, M. , 8 Gillis, W.M. Satiation and the rate of lapse of verbal
meaning. J. gen. Psychol., 1958, 5_9, 79-85.

Wimer, C. Pan analysis of semantic stimulus factors in paired-associate
learning. J. verb. Learn. verb. Behav., 1968, 1, 397-907.

Woodworfin, R.S., 8 Sdnlcsberg, H. Experimental psEhclcfl. (Rev. Ed.)
New York: Holt, 1959.

Wynn, W.H. An informatim—hp‘rocessinffi model of certain asEcts cf p_51_i red-
associate earnin . p s c dissertation, Unnlver51ty
of Texas, 1966.

Yelen, D.R., 8 Schulz, R.W. Verbal satiation? J. verb. Learn. verb.
WVO, 1963, 9;, 372-377.