Overregularization in Language Acquisition

Author(s): Gary F. Marcus, Steven Pinker, Michael Ullman, Michelle Hollander, T. John
Rosen, Fei Xu and Harald Clahsen
Source: Monographs of the Society for Research in Child Development, Vol. 57, No. 4,
Overregularization in Language Acquisition (1992), pp. i+iii+v-vi+1-178
Published by: Wiley on behalf of the Society for Research in Child Development
Stable URL: http://www.jstor.org/stable/1166115
Accessed: 24-11-2017 06:50 UTC

JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide
range of content in a trusted digital archive. We use information technology and tools to increase productivity and
facilitate new forms of scholarship. For more information about JSTOR, please contact [email protected].

Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at
http://about.jstor.org/terms

Society for Research in Child Development, Wiley are collaborating with JSTOR to
digitize, preserve and extend access to Monographs of the Society for Research in Child
Development

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
 MONOGRAPHS OF THE
SOCIETY FOR RESEARCH IN
CHILD DEVELOPMENT
Serial No. 228, Vol. 57, No. 4, 1992

OVERREGU LARIZATION
IN LANGUAGE
ACQUISITION

Gary F. Marcus
Steven Pinker
Michael Ullman
Michelle Hollander

T. John Rosen
Fei Xu

WITH COMMENTARY BY
Harald Clahsen

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
 MONOGRAPHS OF THE SOCIETY FOR RESEARCH IN CHILD DEVELOPMENT
Serial No. 228, Vol. 57, No. 4, 1992

CONTENTS

ABSTRACT v

I. INTRODUCTION 1

II. THEORETICAL BACKGROUND 6

III. METHOD 24

IV. THE RATE OF OVERREGULARIZATION 34

V. THE RELATION OF OVERREGULARIZATION TO CHANGES IN

THE NUMBER AND PROPORTION OF REGULAR VERBS IN
PARENTS' SPEECH AND CHILDREN'S VOCABULARY 70

VI. THE RELATION OF OVERREGULARIZATION TO

THE DEVELOPMENT OF TENSE MARKING OF
REGULAR AND IRREGULAR VERBS 100

VII. FACTORS CAUSING DIFFERENCES

IN OVERREGULARIZATION RATES AMONG VERBS 116

VIII. SUMMARY AND CONCLUSIONS 129

APPENDIX 145

REFERENCES 157

ACKNOWLEDGMENTS 165

COMMENTARY

OVERREGULARIZATION IN THE ACQUISITION OF INFLECTIONAL

MORPHOLOGY: A COMPARISON OF ENGLISH AND GERMAN
Harald Clahsen 166

CONTRIBUTORS 179

STATEMENT OF
EDITORIAL POLICY 181

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
 ABSTRACT

MARCUS, GARY F.; PINKER, STEVEN; ULLMAN, MICHAEL; HOLLANDER, MI-

CHELLE; ROSEN, T. JOHN; and Xu, FEI. Overregularization in Language
Acquisition. With Commentary by HARALD CLAHSEN. Monographs of the
Society for Research in Child Development, 1992, 57(4, Serial No. 228).

Children extend regular grammatical patterns to irregular words, re-

sulting in overregularizations like comed, often after a period of correct
performance ("U-shaped development"). The errors seem paradigmatic of
rule use, hence bear on central issues in the psychology of rules: how cre-
ative rule application interacts with memorized exceptions in development,
how overgeneral rules are unlearned in the absence of parental feedback,
and whether cognitive processes involve explicit rules or parallel distributed
processing (connectionist) networks. We remedy the lack of quantitative
data on overregularization by analyzing 11,521 irregular past tense utter-
ances in the spontaneous speech of 83 children. Our findings are as follows.
(1) Overregularization errors are relatively rare (median 2.5% of irregular
past tense forms), suggesting that there is no qualitative defect in children's
grammars that must be unlearned. (2) Overregularization occurs at a
roughly constant low rate from the 2s into the school-age years, affecting
most irregular verbs. (3) Although overregularization errors never predom-
inate, one aspect of their purported U-shaped development was confirmed
quantitatively: an extended period of correct performance precedes the first
error. (4) Overregularization does not correlate with increases in the num-
ber or proportion of regular verbs in parental speech, children's speech,
or children's vocabularies. Thus, the traditional account in which memory
operates before rules cannot be replaced by a connectionist alternative in
which a single network displays rotelike or rulelike behavior in response to
changes in input statistics. (5) Overregularizations first appear when chil-
dren begin to mark regular verbs for tense reliably (i.e., when they stop
saying Yesterday I walk). (6) The more often a parent uses an irregular form,
the less often the child overregularizes it. (7) Verbs are protected from

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
 overregularization by similar-sounding irregulars, but they are not attracted
to overregularization by similar-sounding regulars, suggesting that irregular
patterns are stored in an associative memory with connectionist properties,
but that regulars are not. We propose a simple explanation. Children, like
adults, mark tense using memory (for irregulars) and an affixation rule that
can generate a regular past tense form for any verb. Retrieval of an irregu-
lar blocks the rule, but children's memory traces are not strong enough to
guarantee perfect retrieval. When retrieval fails, the rule is applied, and
overregularization results.

VI

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
 1. INTRODUCTION

Overregularizations like comed and foots are among the most conspicu
ous grammatical errors in child language, and they have been comme
on for as long as language development has been studied (Bateman,
Brown, 1973; Brown & Bellugi, 1964; Bybee & Slobin, 1982a; Carl
1947; Carroll, 1961; Cazden, 1968; Chamberlain, 1906; Ervin, 1964; Erv
& Miller, 1963; Guillaume, 1927; Kuczaj, 1977a, 1981; Menyuk, 1
Miller & Ervin, 1964; Slobin, 1971, 1978; Smith, 1933; see also Edwards
1970). These errors are made possible by the fact that English has two wa
of creating inflected forms. Most verbs add the suffix -ed to their stems
form the past tense, but about 180 exceptional or "irregular" verbs f
their past tenses in idiosyncratic ways such as a vowel change (come-c
replacement of a final consonant or rhyme (make-made, teach-taught), su
tution of another form (go-went), or no change at all (cut-cut). Overregu
zation errors consist of applying the regular pattern to an irregular
Since children do not hear these forms from their parents, the errors re
the operation of a creative process, presumably corresponding to a m
operation implementing the -ed-suffixation rule posited by grammarians.
Past tense overregularization is just one kind of error in one peripher
aspect of one component of the grammar of one language. Nonethele
has assumed a surprising prominence in cognitive science over the
several decades. Overregularization has been offered as the quintessen
demonstration of the creative essence of human language (Chomsky,
and of the necessity of explaining cognitive processes by rules and repres
tations rather than by rote and reinforcement (e.g., Brown & Bellugi, 196
Lenneberg, 1964; McNeill, 1966; Slobin, 1971; Smith, Langston, & Nisb
in press).
Overregularization has also become famous because of its interest
developmental course, first noted by Ervin and Miller (1963; see also
den, 1968; Miller & Ervin, 1964; Pinker & Prince, 1988). The first over
larization errors seem to appear after a period in which, when child
mark tense on irregular verbs at all, they do so correctly. Overregulariza

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
MONOGRAPHS

thus represents a decline in performance in overt tense marking, resultin

in a U-shaped curve if the proportion of irregular past tense forms that
are correct is plotted against age. The nonmonotonicity has been taken a
evidence for successive "reorganizations" of the child's linguistic system
reflecting a tendency to ferret out generalizations and to prefer them t
exceptional forms, which must later be reintegrated in some fashion. As an
example of cognitive reorganization triggered by a fondness for regularity,
overregularization has often been considered paradigmatic of language de
velopment (Bowerman, 1982a, 1982b; Slobin, 1973) and cognitive develop-
ment (Bever, 1982; Strauss, 1982) in general. It has been used as a metapho
for psychological processes ranging from solving balance-beam problems
(Karmiloff-Smith & Inhelder, 1974-1975) to learning computer-user inter
faces (Grudin & Norman, 1991) to developing expertise in medical decisio
making (Patel & Groen, 1991).
Recently, explaining overregularization has become relevant to the very
foundations of cognitive science. Virtually all discussions have implicitly
assumed that overregularizations can be modeled in principle only by som
explicit representation of a rule in the child's head. Rumelhart and McCle
land (1986, 1987) showed this assumption to be false. They devised a com
puter simulation of an associative network that acquired hundreds of regu
lar and irregular verbs and generalized properly to dozens of new verb
that it had not been trained on. More strikingly, the model appeared to g
through a U-shaped developmental sequence, first producing irregular ver
forms correctly and later overregularizing them, and it also seemed to mani-
fest several other effects previously known to characterize children's behav-
ior. But the model had no explicit representation of words, rules, or a
distinction between regular and irregular systems; it simply mapped from
features standing for the sounds of the verb stem to features standing for
the sounds of the past tense form. The Rumelhart-McClelland model is
a prominent representative of the parallel distributed processing (PDP)
connectionist, or neural networks approach, in which cognitive processe
are modeled as densely interconnected networks of simple neuronlike units.
Its apparent success is commonly seen as a dramatic piece of support for
the PDP approach in general and as posing a severe challenge to rule-base
approaches to language and cognition (see, e.g., McClelland, Rumelhart,
& Hinton, 1986; Rumelhart & McClelland, 1986; Smolensky, 1988). One
reviewer (Sampson, 1987), writing in the Times Literary Supplement, called
the potential implications of the Rumelhart-McClelland model for the study
of language "awesome" because "to continue teaching [linguistics] in the
orthodox style would be like keeping alchemy alive."
Lachter and Bever (1988) and Pinker and Prince (1988) have chal-
lenged the psychological reality of the Rumelhart-McClelland model on a
variety of grounds. But they praised it for the unprecedented precision of

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
 MARCUS ET AL.

its quantitative predictions about child language data. Indeed, Rumelhart

and McClelland were far ahead of the field of developmental psycholinguis-
tics in the quality and quantity of data required for proper tests of their
hypotheses. As such they have underscored the absence of systematic quan-
titative reports of the developmental course of overregularization, its distri-
bution across children, verbs, and time, its relation to the size of the child's
vocabulary, and the lexical factors that cause some verbs to be overregular-
ized more than others. This data gap has left many fundamental questions
unanswered.

The phenomenon of overregularization itself is not in doubt, nor is th

creative nature of the psychological processes that cause it, which re
itself in other ways. Children frequently inflect their own invented v
such as speeched (Chamberlain, 1906), by-ed (i.e., "went by"; Miller & Ervi
1964); eat lunched (Kuczaj, 1977a), broomed (Clark, 1982), and gra
(Pinker, Lebeaux, & Frost, 1987). In many experiments, beginning w
Berko's (1958) classic "wug test," children are given a made-up verb an
then asked to use it in a past tense context, for example, Here is a man w
likes to rick. He did the same thing yesterday. Yesterday he - . The childr
readily produce appropriate forms such as ricked and, when provided
an existing irregular stem, frequently overregularize it (Anisfeld & Tucke
1967; Berko, 1958; Bryant & Anisfeld, 1969; Cox, 1989; Derwing & Bak
1979; Kim, Marcus, Hollander, & Pinker, 1991; Kuczaj, 1978; Marchm
1988; Miller & Ervin, 1964; Pinker et al., 1987). In other experiment
children have been found to judge overregularizations (Kuczaj, 1978)
made-up forms resembling regularly inflected forms (Anisfeld, Barlo
Frail, 1968; Anisfeld & Gordon, 1968) as acceptable.
Beyond the mere fact that children do more than memorize their
ents' verb forms, however, not much is known about the details of overr
larization. Nowhere in the literature can one find solid answers to such basi
questions as, How often do children overregularize? (One percent of
opportunities? Fifty? One hundred?) At what age do they start? At w
age do they stop? What is the developmental curve for irregular verbs, an
does it really look like a U? Are all verbs overregularized, or only s
and what factors account for differences among verbs? How is overregula
ization related to other events in children's language development, suc
their vocabulary growth, and the syntax of tense marking? Given how of
overregularization has been used to explain other things, it is surpris
how little about the phenomenon itself has been documented.
This Monograph is an attempt to fill these gaps. Using the large se
transcripts of children's spontaneous speech recently made available by
Child Language Data Exchange System (ChiLDES; MacWhinney & Snow
1985, 1990), together with previously published tallies of children's vocabu
lary and new unpublished data, we document the process of overregula

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
 MONOGRAPHS

tion in quantitative detail. With these data in hand, we will be in a po

to evaluate explanations of the psychological processes causing overreg
ization.

Presumably children are not designed to overregularize per se; their

errors are a temporary side effect of a mechanism designed to learn th
language. Thus, our investigation of overregularization is organized by
research program that addresses the central question of language acquis
tion research: how children acquire the ability to produce and understan
an infinite number of sentences of their native language from the fini
sample of speech available in conversations with adults. The "learnability
approach" to language development (see Pinker, 1979, 1984; Wexler
Culicover, 1980) seeks to characterize these learning mechanisms; it trea
language acquisition as a difficult software engineering problem and a
tempts to understand how children solve it. The approach dictates the start-
ing point for this Monograph (indeed, for any systematic study of language
development). One must begin with an analysis of what is learned, wh
information is available to learn it, and what kind of computational mecha-
nisms are capable of carrying out the learning task successfully. Such mecha
nisms define hypotheses about the psychology of the child, which can then
be tested and refined against developmental and linguistic data. (As we
shall see, overregularization is a prime example of one of the fundament
problems in understanding language learnability: how children avoid o
unlearn errors in the absence of parental corrections.) Learnability is not
theory but a research problem, and as such it in no way presupposes th
nature of the child's language learning mechanisms (e.g., whether they a
symbolic rules or connectionist networks; whether they are specific to lan-
guage or widely applicable across many cognitive domains). What it does
pick out the central empirical fact to be explained-that children, at som
point, become capable speakers of a language-and it seeks an explanatio
in terms of explicit computational operations, not metaphors or impression-
istic descriptions. As such, it attempts to characterize the innate mechanisms
responsible for human linguistic abilities. This does not single out learnabil-
ity research as a "nativist" approach because all explicit theories involvin
learning must at some point specify the innate mechanisms that do th
learning if they are to avoid an infinite regress.
Chapter II, the first substantive chapter of the Monograph, then, lay
out the logic of linguistic irregularity and overregularization and propos
a simple psychological hypothesis of how children could learn the irregul
system and why they might produce errors while doing so. We outlin
behavioral predictions relevant to this hypothesis and the current state
the evidence concerning them.
After describing the subjects and methods in Chapter III, we presen
the vital statistics of overregularization in Chapter IV. We begin by estimat-

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
 MARCUS ET AL.

ing the overall rate of overregularization, a figure of obvious importance.

As we shall see, one's explanation of children's behavior would be very
different depending on whether they make overregularization errors almost
none of the time or almost all the time. We then examine whether the rate
of overregularization substantially varies across time, children, verbs, and
combinations of these sampling units. For example, we test whether an
overall steady overregularization curve for a child might actually be a com-
posite of curves for individual verbs each overregularized intensively for a
brief period. We also compare the rate of overregularization to the rate of
inflecting regular verbs, and we define and search for U-shaped develop-
mental sequences.
In Chapter V, we examine a hypothesis that underlies the surprising
ability of the Rumelhart-McClelland model to display U-shaped develop-
ment. This hypothesis is that overregularization is triggered by an increase
in the proportion of regular verb forms that the child processes during
development. We first discuss the complex issue of how exactly one should
compare the behavior of network simulations and children. On the basis of
this discussion, we correlate children's overregularization rates with changes
in the number of regular verbs and the proportion of regular verbs among
all verbs in the children's speech, the speech they hear from their parents,
and the children's vocabularies. Measuring vocabulary size is a notoriously
difficult problem in language development, and we present a novel tech-
nique, which we also compare to other estimates in the literature.
In Chapter VI, we examine evidence that points to the development of
the entire productive tense marking system, rather than the balance of
regular and irregular verbs among the vocabulary items feeding it, as the
immediate cause of the onset of overregularization.
Chapter VII, the final empirical chapter, tests hypotheses about the
causes of overregularization by focusing on inherent properties of different
verbs that might cause them to be overregularized more or less often. We
consider factors related to memory strength (the frequency of an item in
parental speech and its similarity to related items in parental speech), the
relatedness of a stem to its past form, and the complexity of the mapping
from stem to irregular form. This chapter includes a test of a second hy-
pothesis inspired by the Rumelhart-McClelland model, that regular and
irregular marking are computed in a single pattern associator.
In the concluding chapter, we summarize the findings, integrate them
within a simple theory, and discuss related theoretical issues.

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
 II. THEORETICAL BACKGROUND

At first glance, the development of overregularization has a straightfor

ward explanation. This traditional explanation, employed both within ps
cholinguistics and in the other branches of cognitive science that have u
overregularization as a metaphor, relies on a dissociation between two p
chological processes: rote memory and rule deployment. The rote proce
memorizes verb forms one by one, and the child can use it at the outset
language development: children hear their parents say broke, so they s
broke. But the English regular past tense rule is not present at the outset o
language development, and the child cannot deploy it until he or she h
learned it, presumably by abstracting the regular pattern from a set
regular forms accumulated over time from parental speech and juxtapo
as past and stem forms of the same verb (walk-walked, use-used, play-playe
etc.). The young child sticks to correct forms because they are availab
from rote memory and there is no machinery capable of overregularizi
them yet; the older child, possessing the rule, can apply it to irregular stem
resulting in overregularization errors.
The explanation is inadequate. The problem is that it does not poin
to any difference between the rule-possessing child and the rule-possess
adult. But there is a difference: children say comed, and adults do not.
children say comed because they possess a regular rule, why do adults, w
also possess the regular rule, not say it too? In fact, the standard explanatio
in terms of a progression between a rote-only child and a rote-plus-r
child does not make any predictions about overregularization appearing
all. It predicts that the younger, ruleless child would fail a wug test (where
the child has to provide an inflected form for a novel word like to wu
whereas the older, rule-possessing child would pass one. For similar reas
it predicts that the young, ruleless child would inflect familiar irregu
verbs correctly (e.g., broke) while leaving new ones unmarked (e.g., stic
whereas older children would inflect familiar irregular verbs correctly (e.g.

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
 MARCUS ET AL.

broke) and overregularize new ones (e.g., sticked). But it does not predict that
children who first use a given verb properly (e.g., broke) would later err on
that very verb (e.g., breaked), which they do (Cazden, 1968; Ervin & Miller,
1963; the phenomenon is documented in more detail in Chap. IV, Sec.
IIA). Possessing a rule that can be applied to novel verbs and applying a
rule to existing verbs that already have a past tense form and that do not
allow the rule are two different things, and the traditional account does not
distinguish between them.
For similar reasons, the rule-rote distinction sheds little light on the
other arm of the U, the fact that breaked ultimately gives way to exclusive
use of broke. Once children have developed a rule and are overapplying it,
how and why do they curtail its use?
Clearly, understanding the developmental course of overregularization
requires more than distinguishing between rule and rote. We must also
examine how the regular rule interacts with irregular items. Specifically, to
explain differences in the ways children and adults treat irregular verbs, we
must first understand how adults treat irregular verbs. Then we can ask
what children might be doing that adults are not. We must also examine
the information about irregular verbs available from conversation with
adults that the child's language learning process might feed on. With knowl-
edge of what is learned and what information is available to learn it, we can
develop a theory of how the learning works and why it might result in
errors in intermediate stages.
In this chapter, we review the current state of knowledge on these
issues (deferring discussion of the Rumelhart-McClelland model to Chap.
V) and develop a hypothesis that appears adequate to account for both
the development and the cessation of overregularization. The hypothesis is
extremely simple, uniting a standard proposal from formal linguistics (that,
in the human language system, irregular memorized items block the appli-
cation of regular rules) and a standard proposal from cognitive psychology
(that retrieval of items from memory is probabilistic and sensitive to fre-
quency of exposure). According to this hypothesis, children's language sys-
tem, like that of adults, is designed so that retrieval of an irregular form
suppresses overregularization, but retrieval is imperfect, and when it fails,
the regular rule applies as a default, leading to overregularization errors.
The hypothesis will lead us to an equally simple empirical prediction: that,
at all ages, evidence that the child's system is designed to suppress overregu-
larization should be available in the form of low overall rates of overregular-
ization in comparison to utterances containing the correct irregular form.
We will then compare this prediction to competing ones in the literature.
This will set the stage for our investigation of the basic quantitative facts of
overregularization in the following chapters.

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
MONOGRAPHS

I. BLOCKING OF REGULARIZATION IN ADULTS

Let us begin with a deceptively simple question. Why are overregu

izations "errors" at all? This is intended not as a prescriptive question
what one ought to say but as a descriptive question about what adult
say. Adults who speak the most common dialect of American Englis
not say breaked, and they judge it as sounding deviant or childlike (P
& Prince, 1988; Ullman & Pinker, 1990, 1992).
The answer is surely not that overregularizations are defective in term
of communicative function; the meaning of breaked is perfectly clea
fact, any child who is willing to overregularize has a communicative adva
tage over adults. No-change verbs in the adult language like cut and se
ambiguous between past and nonpast tenses: On Wednesday I cut the g
could mean last Wednesday, next Wednesday, or every Wednesday;
Wednesday I cutted the grass could mean only a preceding Wednesday.
Another unsatisfactory answer is that adults do not say breaked
cutted because they have never heard other adults say them. It is unsatisf
tory because it assumes that adults would fail a wug test, and clearly
do not; people do not stick conservatively to the past tense forms that th
heard their parents use or that they currently hear other adults use. Adu
are not reluctant to create or accept past tense forms of verbs they n
encountered before, such as John plipped or Yeltsin has finally out-Gorbac
Gorbachev (Kim, Pinker, Prince, & Prasada, 1991; Prasada & Pinker, in
press). Indeed, new verbs enter the language frequently, such as snarf (re-
trieve a computer file), scarf (devour), frob (randomly try out adjustments),
and mung (render inoperable), and their past tense forms do not require
separate introductions. In other words, one cannot explain why adults avoid
breaked by saying that they have never heard anyone else say breaked because
adults have never heard anyone else say snarfed either, but they do not avoid
snarfed.

A. The Blocking Principle

The problem with overregularizations is not that they have never been
heard before but that the irregular counterpart has been heard. Clearly,
there is a psychological mechanism that causes the experience of hearing
an irregular form to block the subsequent application of a regular process
to that item. Thus, several linguists have defined this phenomenon in terms
of a blocking principle (Aronoff, 1976; Kiparsky, 1982; see also Pinker, 1984,
who called it the unique entry principle): an idiosyncratic form listed in the
mental lexicon as corresponding to a particular grammatical modification
of a word (past tense, in this case) blocks the application of a general rule

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
 MARCUS ET AL.

that would effect the same grammatical modification. Thus, broke, listed as
an idiosyncratic past tense form of break, blocks application of the regular
past tense rule, preempting breaked; geese, listed as a plural of goose, blocks
gooses; better, listed as a comparative of the adjective good, blocks gooder. Of
course, some verbs do have two commonly heard past tense forms, for
example, dived and dove; in these cases, the regular version, which is idiosyn-
cratic precisely because it flouts the blocking principle, would be recorded
from the speech input just as if it were an irregular (Pinker, 1984; Ullman
& Pinker, 1990).
Now that we have a mechanism that causes overregularizations to be
treated as deviant by adults, we can state the developmental problem more
precisely. Children do not apply blocking in the same circumstances that
adults do. The question now is, Why not? One obvious possibility is that
children might have to learn the blocking principle and that overregular-
izations appear before they have done so. In the following section, we show
that this hypothesis is probably wrong and that the source of children's
overregularization errors must be sought elsewhere. Before doing so, how-
ever, we add a few remarks about the status of the blocking principle.

B. The Psycholinguistic Status of the Blocking Principle

Blocking is notjust a restatement of the fact that overregularizations are

deviant sounding, nor is it a general prohibition against synonyms. Rather, it
is a principle specifically governing the relations among the inflected ver-
sions of a given stem. Every word in a language implicitly defines a matrix
or "paradigm" of its grammatically modified forms (first person singular
present, first person plural present, first person singular past, and so on;
see Carstairs, 1987). Blocking dictates that any cell of the matrix for a given
word that is filled with an unpredictable specific entry from the input may
not also be filled by applying a general rule for that particular combination
of grammatical features. It can therefore be interpreted as a psychological
mechanism that suppresses the operation of the regular rule in these cir-
cumstances.1

Blocking thus differs from a pragmatic prohibition against synonymy,

such as Clark's (1987, 1990) "principle of contrast" stating that "every two

1 The blocking principle, in turn, might be reduced to Kiparsky's (1982) more general
"elsewhere condition": if the conditions licensing one morphological process constitute a
proper subset of the conditions licensing another morphological process, then whenever
the conditions of both processes are met by some item, only the process with the more
restrictive conditions may apply; the other process applies "elsewhere." In this case, an
irregular form like broke applies in the conditions "past tense of the verb break," and the
regular rule applies in the more general condition "past tense of a verb," so the latter is
preempted.

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
MONOGRAPHS

forms contrast in meaning." First, blocking rules out only competing mor-
phological variants and is agnostic about synonyms in general (e.g., it doe
not rule out couch and sofa; for recent arguments that languages tolerat
true synonyms, see Gathercole, 1987, 1989; for counterarguments, see
Clark, 1988, 1990).
More important, there are empirical reasons why blocking per se, not
contrast, is needed to explain the fact that irregular verbs cannot be regular-
ized. Neither violations of blocking nor conformity to it can be predicted
from the availability of contrasts in meaning. Members of variant pairs such
as leaped-leapt, dived-dove, and sneaked-snuck do not differ reliably in meaning
(or do so with near-imperceptible subtlety); they seem to be recorded by
the learner in response to the brute fact of hearing them (later we discuss
one reason why they arise to begin with). Conversely, verbs like put, make,
give, take, have, come, go, and throw each have dozens of extremely distinct
meanings, especially in combination with particles such as out, up, off, in,
and away. Nonetheless, without exception, speakers are not tempted to use
the regularized versions of these verbs to express the contrasting meanings;
taked the punishment stoically, gived away the answer, haved a ball, maked out, and
so on are always impossible. This shows that different past tense forms do
not necessarily evoke different meanings and that different meanings do
not systematically call for different past tense forms (for discussion, see
Kim, Pinker, et al., 1991; Pinker & Prince, 1988; and Ullman & Pinker,
1990, 1992).
In addition, blocking just rules out certain kinds of forms in a specific
level of grammatical representation; it does not guarantee that an alterna-
tive way of expressing a given notion exists, and thus it differs from a
communicative principle like contrast that is designed to provide a speaker
with exactly one way to express each notion. In fact, there are special cir-
cumstances where blocking applies but the irregular is ruled out for other
reasons and the speaker is left with an expressibility gap. In particular,
blocking applies at a specific level of the mental representation of word
forms, not to entire words; roughly, it applies at the level of inflected stems.
An entire word may consist of a stem plus a (noninflectional) prefix. For
example, speakers of English intuitively perceive forgo as containing the
stem go. Hence, they cannot say forgoed because went, the listed past tense
form of the stem go, blocks the application of the regular rule to go. How-
ever, speakers' sense of the familiarity or naturalness of an irregular form is
a phenomenon that takes place at the level of the entire word, not the stem.2

2 By "irregular form" here we mean "completely irregular," including the meaning

of the combination of the prefix and stem, as in forgo, whose meaning cannot be predicted
from the combination of for- and go. This differentiates it from compositionally prefixed
forms whose stems preserve their meanings, such as retake or overdo.

10

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
 MARCUS ET AL.

Forgo is a relatively uncommon word, and uses in the past tense are even
rarer, so forwent, while not sounding grammatically impossible like forgoed,
nonetheless sounds stilted and unnatural (Pinker & Prince, 1988; Ullman
& Pinker, 1992). The verb thus has no perfectly usable past tense form.
The regular past is ruled out by blocking (applying at the level of the stem);
the irregular past is tainted by unfamiliarity (applying at the level of the
entire word). This shows that blocking is a mechanism that operates on
specific representations of the speaker's grammatical system.

II. BLOCKING AS A MEANS OF RECOVERING FROM CHILDHOOD ERRORS

IN THE ABSENCE OF PARENTAL NEGATIVE FEEDBACK

Logically speaking, given the blocking principle, children could learn

that forms such as breaked are not English; they would just have to hea
their parents say broke. Conversely, given the knowledge that forms lik
breaked are not English, children could learn the blocking principle. (As w
saw, not hearing one's parents say breaked is not enough because parents
never say wugged either and it is admissible.) Are there any empirical reasons
to choose between these possibilities? In this section, we suggest that it
unlikely that children acquire blocking from evidence about which form
are ungrammatical; it is far more likely that they determine which form
are ungrammatical using blocking. The problem stems from the concept
"the knowledge that forms like breaked are not English" or, more generally,
"evidence about which forms are ungrammatical."

A. Negative Evidence and Overregularization

A significant problem in explaining language acquisition is that children

do not receive "negative evidence": feedback from parents indicating, fo
any string of words the child may utter, whether it is a grammatical sen
tence. Children are not corrected or misunderstood more often when they
speak ungrammatically (Brown & Hanlon, 1970), and although it is occa-
sionally suggested that there is weak statistical information about grammati-
cality in the differential likelihood of parents' repetitions, expansions, o
topic changes, there is considerable doubt as to whether such information
exists and whether it is necessary or even useful to the child (see Bowerman,
1987; Gordon, 1990; Grimshaw & Pinker, 1989; Pinker, 1989; the literatur
and issues are analyzed in detail by Marcus, 1992). A lack of negative evi-
dence means that, if the child ever develops a linguistic system that gener-
ates a superset of the target language (all the grammatical forms in th
target language plus some ungrammatical forms not in the target language),
the parental input cannot tell the child that anything is wrong (Gold, 1967;

11

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
MONOGRAPHS

Osherson, Stob, & Weinstein, 1985; Pinker, 1979). To explain how the adu
grammar is attained, then, one must explain either how children avoid
generating supersets or, if they do, how they expunge their errors.
Overregularization errors in particular pose this problem. A child wh
is producing breaked and broke is speaking a superset of adult English in this
domain, which allows only broke. And overregularization errors in particular
do not reliably occasion negative parental feedback. Kuczaj (1977a, p. 599
noted that, in his investigation, the children (especially his son Abe, wh
was the main subject) were not systematically corrected for overregulariza-
tion errors. The following typical chunk of dialogue, which we have foun
in transcripts of conversations between Kuczaj and Abe (MacWhinney &
Snow, 1985), illustrates his observation:

Father. Where is that big piece of paper I gave you yesterday?

Abe. Remember? I writed on it.
Father. Oh that's right don't you have any paper down here buddy?

Moreover, it seems unlikely that children attend to corrections, requests fo

clarification, recastings, and so on when they do occur. For example,
describing systematic overregularization of participles by his 41/2-year-old
daughter, Zwicky (1970) reports that "six subsequent months of freque
corrections by her parents had no noticeable effect." The following d
logue, from Cazden (1972), gives the flavor of such attempts:

Child. My teacher holded the baby rabbits and we patted them.

Adult. Did you say your teacher held the baby rabbits?
Child. Yes.
Adult. What did you say she did?
Child. She holded the baby rabbits and we patted them.
Adult. Did you say she held them tightly?
Child. No, she holded them loosely.

More precisely, Morgan and Travis (1989) report a quantitative study

on the availability of negative evidence about overregularizations and its
relation to children's recovery from such errors. They tabulated the number
of overregularization errors and errors in wh- questions, and their cor-
rect alternative forms, in the speech of the children known as Adam, Eve,
and Sarah (Brown, 1973) and cross-classified them in terms of whether the
utterances were followed by parental expansions, exact imitations, partial
imitations, clarification questions, confirmation questions, attempts to move
the conversation on, and no response. No consistent contingency was found
between errors and parental responses: for Adam expansions and clarifica-
tion questions were more likely to follow his ungrammatical sentences; for

12

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
 MARCUS ET AL.

Eve it was expansions and partial imitations that occurred more frequently
following her ungrammatical sentences; and for Sarah all five categories of
parental response showed the opposite pattern, occurring more frequently
after well-formed utterances. Unless a child can figure out the kind of parent
he or she has (e.g., a grammatical sentence expander or an ungrammatical
sentence expander), such feedback is useless (see Marcus, 1992). Moreover,
Morgan and Travis showed that, as children get older, even this inconsistent
feedback signal disappears. Since the errors continue after the parent has
stopped supplying potential feedback, unlearning the errors must depend
on some other information source.

B. How Blocking Obviates the Need for Negative Evidence

Presumably, this information source is a constraint endogenous to

the child's language system. Blocking is just the right kind of constrain
children do not have to receive direct information that breaked is ungram-
matical; they can infer it from hearing broke. In sum, because the inpu
information needed to learn blocking-that forms like breaked are
ungrammatical-is not available to children, it is unlikely to have been
learned from the input, but it may be part of the machinery that does the
learning.3
If children's language systems incorporated a mechanism implement-
ing blocking (Pinker, 1984) or its equivalent (e.g., MacWhinney, 1978), we
would have a straightforward explanation of how they recover from over-
regularizations in the absence of negative evidence, converging on the adult
state. Each time an irregular past tense form is heard in parental speech,
the child can record it in the lexicon, and the regular rule is thereafter
blocked from applying to it. Thus, blocking is consistent with a develop-
mental sequence progressing from overregularization to correct perfor-

S Some such assumption has proved unavoidable to many researchers devising explicit
models of language learning, working in a variety of frameworks. For example, Anderson
(1983) required a constraint similar to blocking when applying his symbol-processing
model of cognition called ACT to the acquisition of morphology (although the principle
was not specific to language acquisition but constrained his production-system architecture
across all the tasks it was given). In their very different connectionist model, Rumelhart
and McClelland (1986) also implemented machinery that carries out a version of blocking.
As we shall see, in their supervised learning paradigm, a special input pathway was de-
signed that presented a single "correct" past tense form representing the parental input,
and an error-correcting learning procedure acted to suppress the tendency of the model
to produce an output that deviated from it. Then a separate mechanism, the whole-string
binding network, which contains explicit representations of correct and overregularized
forms, sets them in competition with one another to select the form that best approximates
the output of the network.

13

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
MONOGRAPHS

mance (the adult state) without negative evidence. Blocking would prevent
children from ever generating a superset of English; they would progres
from breaked to broke, never saying both at the same stage.

III. EVIDENCE RAISING DOUBTS ABOUT BLOCKING IN CHILDREN'S

LANGUAGE DEVELOPMENT

Unfortunately, when we examine the full developmental sequence

inflection in children, we immediately run into a problem. Blocking of ove
regularizations explains monotonic improvement, not U-shaped devel
ment. That is, the principle tells us only how children might get out of
overregularization stage. It does not explain how they got into it; indeed
seems to predict that the child should never get into it in the first place. A
child who respects blocking should never allow the regular rule to apply
an irregular form; the irregular would win the competition from the start
and no U-shaped sequence should be seen. If children really go from co
rect irregulars to overregularizations back to correct irregulars, the hypot
esis that blocking is inherent to children's language system is cast into dou
At first one might try to explain the full sequence as follows. Supp
blocking is just one manifestation of a more general principle stating o
that one of two competing forms must be eliminated, not that the fo
witnessed in the input eliminates the one generated by rule. Such principle
can be found in Clark's (1987) principle of contrast, Wexler and Culicov
(1980) uniqueness principle, and Slobin's (1973) operating principle of o
to-one mapping.4 If the competition between regularized and irregul
forms is two-way rather than one-way, then two successive replacemen
each respecting the principle, would define a U-shaped developmental
quence. In the first arm of the U, the listed form is eliminated by t
newly acquired rule (perhaps falling under the general phenomenon o
"imperialism" of newly acquired inflectional rules discussed by Slobin, 1973
and MacWhinney, 1978). In the second, blocking per se, which requir
that the form attested in the input always win out over the one genera
by a rule, applies. Crucially, in this hypothesis, the child still avoids a
perset of English at all stages: broke is replaced by breaked, which is in tur
replaced by broke; at no time may the two forms co-occur.
But this solution is inadequate. Aside from the fact that the imperialism
phenomenon (an across-the-board preference for a newly acquired rule
itself unexplained, the empirical picture of U-shaped development that

4 Clark's principle of contrast acts more like blocking per se when it is combined w
her principle of conventionality, which forces the child to use the form that is conventio
in the speech community whenever competing alternatives arise.

14

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
 MARCUS ET AL.

assumes, in which the child actually loses early irregular forms, does not
seem to be accurate given the (admittedly scanty) data available at present.
All the authors that have actually recorded overregularization errors note
that these errors coexist with, rather than replace, the early irregulars (Caz-
den, 1968; Ervin & Miller, 1963; Kuczaj, 1977a, 1981). This is exactly what
all the principles dictating unique forms proscribe. Maratsos (1987) points
to such coexistence, which can last for months or years, as a reason to doubt
the existence of such principles. Children, it seems, do generate a superset
of English (they use break and broke simultaneously), so if blocking is the
explanation of how children eventually abandon this superset, it leaves it a
puzzle why they adopted it to begin with and why they retain it for so long.5

IV. BLOCKING AND RETRIEVAL FAILURE: A SIMPLE EXPLANATION OF

OVERREGULARIZATION ERRORS

We appear to be at an impasse. The mechanism required to explain

adult language (blocking) seems to be systematically flouted by the child
during a period in which correct and overregularized forms coexist. But
the information necessary to learn blocking--negative evidence through
parental feedback-does not exist. As a result, the processes causing the
appearance and disappearance of overregularization have been shrouded
in mystery, and it is tempting to treat it as a qualitative developmental stage
that the child enters and exits, driven by some fundamental reorganization.
But positing two qualitatively different kinds of machinery, one for chil-
dren, one for adults, is not exactly parsimonious, especially since no account
has been proposed of how the unknown changing machinery leads to the
development of overregularization. Is there a simpler alternative?
One of them is proposed in MacWhinney (1978, pp. 6-7) and Pinker
(1984, pp. 194-195). Say that children possess a correct irregular in lexical
long-term memory and represent it as the past of the corresponding stem,
but that either the content of the memory entry for the irregular or the
link to the stem, or both, is not accessible 100% of the time. If an irregular
past tense lexical entry is not retrieved, it obviously cannot block regulariza-
tion. If the child intends to mark tense and possesses a process for inflecting

"5 Another possible solution, originally proposed by Kuczaj (1977a, 1981), is that chil-
dren may fail to realize that a given irregular form corresponds to the past tense version
of some stem. Rather, they may treat the irregular past as an independent verb, and
blocking would not apply. Errors would cease when the two verbs were united, presumably
when the child noticed that they were semantically identical except for pastness and (in
most cases) were phonologically similar as well. We discuss this solution in Chap. VII,
Sec. II.

15

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
MONOGRAPHS

arbitrary stems, the output will be an overregularization.6 Similarly, if the

content of a stored past tense form is retrieved, but without its "past tense"
link or feature, overregularizations consisting of an affixed past stem would
result, like broked or wented.7
The blocking-and-retrieval-failure hypothesis is appealing because it
can be deduced from the very logic of irregularity, supplemented only by
an uncontroversial fact about human memory known since Ebbinghaus.
What is the past tense form of the verb to shend, meaning "to shame"? If
you answered shended, then you have overregularized; the correct form is
shent (Bybee & Slobin, 1982b). Of course, this "error" is not surprising.
Irregular forms are not predictable (that is what "irregular" means), so the
only way you could have produced shent was if you had previously heard it
and remembered it. But you have heard it zero times, hence cannot have
remembered it. Now, if in two years you were asked the question again and
overregularized it once more, it would still not be surprising, because you
would have heard it only once. Since memory storage and retrieval are
probabilistic, with a higher probability of retrieval for items that have been
presented to the learner more often, hearing an irregular a small number
of times should be only somewhat better than not hearing it at all. Thus,
low-frequency irregulars are inherently prone to overregularization (Mac-
Whinney, 1978; Pinker, 1984; Slobin, 1971).
Children, by definition, have not lived as long as adults. Among the life
experiences that one accumulates through the years is hearing the past tense
forms of irregular verbs. Many verbs for a given child will be like shent for
an adult: never heard; heard but not attended to; heard and attended to,

6 The principal difference between MacWhinney's (1978) and Pinker's (1984) exposi-
tions is that Pinker takes the blocking principle, as it was explicated and justified by
linguists to explain adult knowledge, and simply attributes it to the child, whereas Mac-
Whinney introduced it as a specific new claim about the child's morphological acquisition
system. Note that MacWhinney has since disavowed the claims that regularizations are
produced by a rule and that they are suppressed by rote forms across the board. He
has proposed instead that regularizations are produced by a propertywise generalization
process like that of the Rumelhart-McClelland model and that overregularizations are
suppressed by specific, individually learned inhibitions of each overregularized form by
its irregular counterpart (MacWhinney, 1987, pp. 285, 295; see also MacWhinney & Lein-
bach, 1991; and Chap. VIII, Sec. III).
7 Presumably, the fourth logically possible retrieval pattern can occur as well: the
existence of a stored past tense form linked to a stem is registered, but its content cannot
be recovered, a state like that studied in memory phenomena such as tip of the tongue
(Brown & McNeill, 1966), feeling of knowing (Hart, 1965), and deep dyslexia (Coltheart,
Patterson, & Marshall, 1980). In such cases, we would not see overregularization, because
the activation of the irregular memory entry blocks the rule; nor would we see the correct
irregular, because its content is temporarily unavailable. What would surface is the un-
marked stem.

16

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
 MARCUS ET AL.

but not enough times to be able to recall on demand reliably. A child should
overregularize these verbs, even with a grammatical system identical to
adults'. If children's memory retrieval is noisier than adults', they should
make these errors even more often, holding number of exposures constant.
But regardless of whether there are quantitative differences between chil-
dren and adults, there need be no qualitative differences; blocking of a
regular rule by retrieval of an irregular stored in memory and a memory
retrieval rate of less than 100% are sufficient to account for the phe-
nomenon.

The retrieval failure hypothesis predicts tha

confined to childhood, and, indeed, it is not. Adu
regularization errors in their spontaneous spe
Chap. IV, Sec. IIB). Errors occur even more of
adults must utter past tense forms under time pr
of irregulars, the errors are more likely with
(Bybee & Slobin, 1982a). Even unpressured lan
of overregularization on low-frequency verbs
admit regular past tense forms as more or less na
American speech, for example, dreamt-dreamed a
Pinker & Prince, 1988; Ullman & Pinker, 19
Ullman and Pinker (1990, 1992) have found that these "doublets" have
lower average nonpast stem frequencies than verbs that are exclusively ir-
regular. Moreover, within doublets, the frequency of the irregular past
tense form correlates significantly with experimental subjects' ratings of the
naturalness of the irregular versus the regularized past tense forms: low
frequency forms like slew, slunk, trod, rent, and strove were likely to be pre-
ferred in regularized versions. Finally, the hypothesis is consistent with the
fact that irregulars in general tend to be high in frequency in English and
that lower-frequency irregular verbs in earlier stages of the language (e.g.,
geld-gelt, cleave-clove, abide-abode) were likely to become regular over time
(Bybee, 1985).8 Low-frequency past tense forms are always in danger of not
being uniformly memorized in some generation; if so, the verbs, if they
remain in the language at all, will become regular. Verbs that survive as
irregulars are thus more likely to be high in frequency.
In sum, the hypothesis of blocking with occasional retrieval failure aug-
ments the traditional rule-rote distinction to explain the time course of
overregularization errors as follows. Very young children have not yet

8 Bybee examined 33 surviving verbs from three classes of strong verbs in Old English.
Fifteen have come through in Modern English as irregular verbs; 18 have become regular.
The surviving irregulars have a mean Francis and Kucera (1982) frequency of 515 over
all their inflectional forms, 137 in the past tense; the regularized verbs have a mean
frequency of 21 over all forms, five in the past tense.

17

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
MONOGRAPHS

learned that English obligatorily marks tense and that an -ed suffixation rule
is available to do so. However, they can memorize past forms from their
parents and occasionally use them; all past tense forms recorded from chi
dren will thus be correct ones. Overregularizations appear when the regul
rule is acquired. But the previously acquired irregulars do not go anywhere,
nor are they ever incapable of blocking overregularization: they just hav
to be retrieved to be able to do the blocking, and they are retrieved probabi-
listically. The cure for overregularization is living longer, hearing the irreg-
ulars more often, and consolidating them in memory, improving retrievabil-
ity. Crucially, this account serves to demystify overregularization, requiring
no qualitative difference between children and adults during the overregu
larization period.

V. PREDICTIONS ABOUT THE RATE OF OVERREGULARIZATION

One crucial datum about overregularization is its actual rate as a pr

portion of the child's opportunities to make such errors. If overregular
tion occurred at a rate of 0%, then blocking alone would explain everythin
Obviously, the rate is not 0%. But, surprisingly, there are few hard dat
to what it is, and this allows for an interesting empirical test.

A. Predictions of the Blocking-and-Retrieval-Failure Hypothesis

The blocking-and-retrieval-failure hypothesis predicts that the chi

linguistic system is at all times designed to suppress regularization of verb
remembered to be irregular. This suppression of regularization cannot
perfect because the child's memory is not perfect, but it is as good as
child's memory retrieval process. If we assume that children's memory
words, although imperfect, is quite good (the child is, after all, successfull
using thousands of words and acquiring them at a rate of approximate
one per waking hour; Miller, 1977), then overregularization should be
exception, not the rule, representing the occasional breakdown of a sys
that is built to suppress the error. The overregularization rate, therefo
while not being 0%, should be as close to 0% as the child's rate of successfu
memory retrieval permits. Minimally, an observed overregularization r
that is systematically less than 50% and not attributable to any factor
founded with irregular forms would serve as evidence that the child's
guage system is biased against overregularization in favor of an irregu
form when it is available. Blocking effects exactly that bias, and the lo
the rate turns out to be (assuming that it is less than 50%), the less need w
would have for any explanation other than blocking and retrieval failu

18

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
 MARCUS ET AL.

Children's overall overregularization rate, of course, is a weighted aver-

age of rates for different verbs, which themselves are predicted to range
from 100% (for a verb never heard or attended to in the past tense, like
shent for adults) to 0% (for an overlearned verb). But the best-learned verbs
are the ones children hear their parents use most frequently, and the verbs
that parents use most frequently are also likely to be the ones that the child
uses most frequently and hence should be better represented in samples of
children's speech than the rarer, hence more poorly memorized, hence
more overregularization-prone verbs. Thus, the design of the child's linguis-
tic system to block regularization of irregular verbs should be apparent in
pervasively low rates of overregularization in children's spontaneous speech
samples.

B. Competing Predictions

Strictly speaking, there are no alternative theories of the learning pro-

cess in the literature that can be said to make competing "predictions" about
children's overregularization rate. (Rumelhart and McClelland do present
an alternative theory, but we save it for Chaps. V-VII, because its critical
predictions do not concern the rate of overregularization.) The relation
between data and hypothesis has generally gone in the other direction:
researchers have made rough assumptions about what the overregulariza-
tion rate is and have drawn conclusions about the nature of the learning
process from them. But because few of the researchers cite actual data
on the overregularization rate, and because their idealizations of the data
contradict one another, the researchers are in effect making competing
predictions about what the rate should turn out to be in a large-scale quanti-
tative study such as the one we are about to report. In this section, we treat
these idealizations as predictions of the approaches they have been taken to
support and compare them with the predictions of the blocking-and-
retrieval-failure hypothesis.
A rate of 100% is by far the most common idealization of the empirical
picture in the literature. Recall that it underlies the one-to-one mapping
hypothesis, which explains the two phases of overregularization in terms of
subsequent replacement of irregulars by overregularizations and overregu-
larizations by irregulars. It also inspired the picture of the child as the
exception-hating rule monger that was imported into other branches of
developmental psychology. As such, the assumption of a 100% overregular-
ization rate can be found everywhere from textbooks to technical articles to
the popular press. Here are examples, respectively, of each:

Interestingly, when the general rule is learned, children will often

stop using the previously learned irregular form and instead produce

19

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
MONOGRAPHS

a regularized version. Thus a child who had been using the correct past
tense of sing-namely, sang-may start using the form singed ... in-
stead. Before the correct irregular form is again learned, some children
have been known to produce forms that combine the irregular and the
regular form. In the case of sing, this would be sanged. Thus children
may proceed through as many as five steps in the acquisition of some
inflections ... No inflection, Adult Form, Overregularization, Transi-
tion, Adult Form. [Reich, 1986, p. 148]

At this time [overregularization] the irregular forms that the child

had used earlier fade out in favor of the overregularized forms. When
the irregular forms later reassert themselves, they have a new status:
they are no longer isolates operating independently from their unin-
flected counterparts and from regular inflected forms; rather, they are
integrated into a system, as exceptions to it. [Bowerman, 1982a, p. 321]

The errors [children] do make are actually logical overgeneraliza-

tions of rules. Instead of "He went," for instance, they may say "He
goed"-to them, a perfectly reasonable past tense of "go." They reject
irregularities. When they have learned phrases by rote like "It broke"
and "two mice," they will toss them out once they become aware of
past tenses and plurals, notes University of California psychologist Dan
Slobin. Suddenly, they start using "It breaked" and "two mouses." Says
Jill de Villiers, only half joking: "Leave children alone and they'd tidy
up the English language." [Gelman, 1986, p. 85]

We find a subtly different idealization in the following, superficially

similar statement of the facts, from Eve Clark's article arguing for the princi-
ple of contrast:

Children are pattern-makers. And when they begin to acquire the

inflections that mark tense, for instance, they typically take irregular
verbs such as break, bring, and go, and treat them as if they belonged to the
regular paradigm of walk, open, and jump. So the past tense of break
is produced as breaked, bring as bringed, and go as goed. [Clark, 1987,
p. 19; emphasis added]

Note that Clark is not claiming that irregular past forms are tossed out or
even that they fade; she says only that irregulars are treated as regular verbs.
Crucially, while they are overregularizing, most children do not successfully
inflect regular verbs with -ed 100% of the time. Brown (1973) reports that
none of the three children he studied were producing regulars more than
90% of the time for six consecutive hours of samples during the time they
began to overregularize, although they achieved this criterion several
months later. We examine these data in detail in Chapter VI; for now

20

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
 MARCUS ET AL.

assume that children inflect regular verbs during the overregularization

years with a success rate of 75%. In that case, Clark's statement would be
consistent with children overregularizing 75% of the time (or whatever the
actual rate of inflecting regular verbs turns out to be). Indeed, it is broadly
consistent with children continuing to produce correct irregulars some pro-
portion of the 25% of uses that are not overregularizations. That, in turn,
would be literally consistent with the finding discussed earlier that children
use both overregularized and correct versions of a verb during the overreg-
ularization stage, as long as the overregularization rate was always capped
by the successful suffixation rate for regular verbs.
However, many writers have interpreted the finding of coexistence as
showing a more extreme deviation from complete overregularization. For
example, in reference to an aspect of the development of verb argument
structure that did not consist of complete replacement of a correct form
by an error, Bowerman (1982a, p. 342) amends her earlier summary of
overregularization as follows:

However, it has become clearer in recent years that overregulariza-

tion is not the all-or-none phenomenon it was once taken to be: Irregu-
lar forms rarely drop out, but rather continue to compete with their
overregularized counterparts throughout the period of error mark-
ing. . . . The relative strength of the irregular and overregularized
forms in this competition reflects a complex interplay of factors, such
as how long the irregular forms have been part of the child's repertoire
before their role in a broader system is perceived, how frequently they
have been said or heard, whether the "irregular" forms are truly maver-
icks or belong to minor patterns of their own, and whether the child
routinely activates a newly grasped systematicity in the course of sen-
tence construction or perceives it only more passively.

No clear value for the assumed overregularization rate is given, but "relative
strength of irregular and overregularized forms" implies in context that for
some verbs overregularizations predominate and that for others the correct
form does. There is no reason to expect that overregularizations or correct
irregulars are systematically preferred, and given the large number of differ-
ent biasing factors at work over a large number of verbs and children, we
might expect on statistical grounds that they would roughly even out. We
can therefore read Bowerman as basing her theorizing here on the assump-
tion that the overall overregularization rate does not deviate from 50% by
an amount large enough to attribute to some pervasive single cause.
In a review article on language development, Maratsos (1983, p. 763)
is more explicit about his assumptions about children's overregularization
rate:

21

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
MONOGRAPHS

Kuczaj (1977b) examined longitudinal samples for one subject, and

6-hour cross-sectional samples for 13 children of different levels of
competence. Again, no evidence indicates that -ed overregularizations
drove out irregular past forms. Instead, even within individual verbs,
overregularized and irregular forms alternated, often for periods of
months to years. Overregularization, in general, ranged in frequency
from .20 to .60 of the children's uses.

It is not clear from Maratsos's chapter what the figures of "from .20 to .60
of the children's uses" (attributed to an unpublished talk by Kuczaj) refer
to, as we shall see in Chapter IV. For now, it suffices to note that Maratsos
is suggesting that the typical rate of overregularization is not far from 50%
(the literal implication of "alternation"), perhaps 40% (the midpoint of the
range he cites). It is also clear that these are the data he considers as a
refutation of blocking. In his review of Pinker (1984), he writes,

Yet empirically, children do not act as though they have such a

solution [blocking], but may alternate between the overregularized -ed
form and the irregular form for a period of months to years, using
both broke and breaked. .... It is clear that their analysis and resolution
of such alternatives is a long-drawn-out tabulational process, not one
which quickly seizes upon one or two properties of the language as
heard. [Maratsos, 1987, p. 19]

Thus, the literature contains a range of idealizations of children's typi-

cal overregularization rate, and theories based on them, ranging from
100%, to the rate of regular affixation (say, 75%-95%), to some aggregate
rate reflecting systemwide indifference (perhaps not too far from 50%, or
some range of values that includes it such as 20%-60%). A rate of exactly
0% is of course never entertained; nor is a rate near 0%, or even greater
than 0% but pervasively less than 50%, representing a systematic preference
for irregulars. Since we have derived this very prediction from the
blocking-and-retrieval-failure hypothesis, there are now clear competing
predictions. Thus, the first empirical tests in this Monograph will involve
quantitative analyses of large samples of children's speech in an attempt to
estimate children's characteristic overregularization rate. The estimate will
be compared with the various predictions and assumptions in the literature
and thus will test the corresponding ideas about the psychology of overregu-
larization.

VI. SUMMARY

The mere fact that children develop a regular rule cannot explain wh
they overapply it to irregular verbs. Adults block the application of regular

22

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
 MARCUS ET AL.

rules to idiosyncratic memorized items; we suggest that children do, too, but
their retrieval of idiosyncratic items from memory, especially low-frequency
ones, is probabilistic, and overregularizations occur when it fails. The hy-
pothesis explains how children unlearn their errors in the absence of nega-
tive parental feedback and does not make the dubious prediction that over-
regularizations replace correct irregulars in children's speech. It predicts
that overregularizations should generally be rare in children's speech rela-
tive to correct irregular past tense forms, contrary to most current views
about language development.

23

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
 Ill. METHOD

I. SUBJECTS

To assess overall overregularization rates we wanted to examine as large

and diverse a set of children as possible. Using the 1990 version of the
ChiLDES data base and documentation (MacWhinney, 1990), we selected
all the unimpaired English-speaking children who met the following crite
ria. (1) The children were known to speak Standard American English. (2
The transcripts were in CHAT format, the standard for ChiLDES transcripts,
allowing efficient and accurate computer searches using the CLAN software
package (MacWhinney, 1990). (3) The investigator did not include warning
that the transcripts were in a preliminary state and possibly dangerous t
use. (4) Information was available about the subjects and the circumstance
in which their speech was recorded. (5) The transcripts contained at leas
10 irregular past tense forms per child (regardless of whether they wer
correct or overregularizations). Applying these criteria yielded a sample o
83 children, who produced 11,521 utterances containing past tense forms of
irregular verbs. Table 1 shows the children, their ages, and their recordin
schedules. Among these 83 children we focus on the 10 represented by
longitudinal samples and on the 15 with single samples from Hall, Nagy,
and Linn (1984). The remaining 58 came from multichild data bases, an
we used their data as a replication of the basic findings on overregularization
rate obtained from the main sample.
Sarah was a child from a working-class background whose parents had
high school degrees and whose father worked as a clerk; all the other ind
vidual children were from professional families. The children in the grou
data bases are all from middle-class, although not necessarily professional
backgrounds. Ten of the children were black (Adam and nine of the chil
dren in the Hall et al. sample). There were 5 boys and 5 girls among th
children with individual data bases, 10 and 5 from Hall et al. (1984), 8 an
6 from Gathercole (1979), 14 and 10 from Gleason (1990), and 10 and 10
from Warren-Leubecker (1982).

24

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
 MARCUS ET AL.

TABLE 1

CHILDREN STUDIED

Total
Child Age Source Samples Sampling Frequency
Abe ........... 2-6-5-0 Kuczaj (1976) 210 Weekly
Adam .......... 2-3-5-2 Brown (1973) 55 2-3/month
Allison ......... 1-5-2-10 Bloom (1973) 6 Occasionally
April .......... 1-10-2-11 Higginson (1985) 6 Occasionally
Eve ............ 1-6-2-3 Brown (1973) 20 2-3/month
Naomi ......... 1-3-4-9 Sachs (1983) 93 Weekly to monthly
Nat............ 2-8 Bohannon & Marquis (1977) 21 Within 1 month
Nathaniel ....... 2-3-3-9 Snow (unpublished)a 30 Weekly
Peter .......... 1-3-3-1 Bloom (1973) 20 Monthly
Sarah .......... 2-3-5-1 Brown (1973) 139 Weekly
15 children ..... 4-6-5-0 Hall et al. (1984) 30 2 days/child
24 children ..... 2-1-5-2 Gleason (1980) 72 3 samples/child
20 children ..... 1-6-6-2 Warren-Leubecker (1982) 20 1 sample/child
14 children ..... 2-9-6-6 Gathercole (1979) 16 1-4 samples/child
a See MacWhinney and Snow (1985).

To answer questions about longitudinal development and voc

size, we need to focus on samples that began before the onset of
larization and continued long enough for performance in tense
to approach adult levels. Brown's (1973) Adam, Eve, and Sarah m
criterion; overregularizations are absent from their early transcr
their later transcripts extend to Brown's "Stage V," in which most in
are supplied in their correct forms more than 90% of the time.
There are also extensive longitudinal transcripts for Kuczaj's
(Kuczaj, 1976, 1977a, 1978), but they begin later than those for A
and Sarah, and Abe was already overregularizing in the first, so
about his onset of overregularization must remain unanswered.
we examine other aspects of his development and take advanta
prodigious numbers of overregularizations and correct irregula
transcripts when focusing on individual verbs and their develo
courses.

Finally, when examining the effects of lexical facto

regularization rates for individual verbs from all 19 ch
them with various properties of the verbs. Such analys
Individual children often supply too few errors to p
of overregularization rates and the wide range of pr
needed for correlational analyses, but aggregate dat
playing averaging artifacts. Therefore, in the lexica
verging results from 19 individual children who overr
possible, an aggregate measure that combines their ove

25

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
MONOGRAPHS

II. PROCEDURE

We tabulated all past tense forms of irregular verbs, correct and ove
regularized, for all the children (for an exhaustive list of irregular verbs in
present-day American English, sorted into subclasses, see the appendix
Pinker & Prince, 1988). Irregular verbs such as dream and dive admitting of
a common regular alternative in adult speech were excluded. Overregular
ations included stem + ed forms like eated and past + ed forms like ated. We
did not search for other possible kinds of error involving the regular suffix
such as sweepened or brecked (for broke); such errors, in any case, are quite
rare. The actual tabulation procedure differed slightly for the differen
children.

For Abe, the data were gathered and tabulated by Kuczaj (1976), the
boy's father. For each month from age 2-5 to age 5-0, Kuczaj recorded the
number of times Abe used each of 70 irregular verbs, the number of times
he produced the present stem of the verb with -ed appended (such as goed
and breaked ["stem + ed" errors]), the number of times he produced doubly
marked pasts in which -ed was added to the irregular past form (such as
wented and broked ["past + ed" errors]), and the number of times he used the
stem form in obligatory past tense contexts (such as Yesterday we go out).
Kuczaj's tables also contain one occurrence of beed, but Kuczaj did not tally
other uses of be, was, or were, presumably because they could also be forms
of the auxiliary be; this would also account for why have-has-had and do-did
are absent. Our analyses of Abe thus omit all three verbs. Kuczaj reported
his data in two ways: in a table listing the number of correct and incorrect
forms, summed over verbs, for each month (his table 18) and in an appendix
listing how many times each individual verb, in each of its possible forms,
was used each month (his app. G). Occasionally, there were discrepancies
between these two tables that we could not resolve from the text of the
thesis, so we relied on tallies of the raw data from Kuczaj's investigatio
provided to us by Michael Maratsos. Data from the months 2-5-2-7 and f
four no-change verbs were absent from these tallies; for them we used t
thesis tables exclusively.
For Adam, Eve, and Sarah, verb usages were tallied on a DEC Microva
II running UNIX. Individual transcript files were combined into a sing
master file for each child. The FREQ program in the CLAN software packag
(MacWhinney, 1990; MacWhinney & Snow, 1990) counts the number o
times every word is used in a particular transcript session, for a particu
speaker, and it was run on the individual transcript files and the combin
files. The combined frequency list for each child was then edited to include
only words that the child may have used as a verb, including words th
occur only infrequently as a verb (e.g., fish, color, ground, milk). For all su
items (i.e., all words that are not exclusively verbs in the child's vocabulary)

26

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
 MARCUS ET AL.

the UNIX utility FGREP, which finds matches of alphanumeric string patterns,
extracted all the transcript lines they occurred in. Each of the resulting lines
was checked by hand and excluded if the matched word turned out not to
be used as a verb. If a word appeared in a single-word utterance, it was
excluded; thus, sleep or put appearing alone were not counted as verbs, but
Adam sleep or put Mommy were included. Verbs repeated in successive senten-
ces such as I falled down. I falled down, Momma were counted separately since
children are capable of saying both a correct and an overregularized version
of a single verb in successive utterances, as in Abe's Daddy comed and said
"hey, what are you doing laying down?" And then a doctor came.
For all instances of Adam, Eve, and Sarah's no-change verbs such as
cut and put, past tense usages were distinguished from present and infinitival
usages by hand; when the transcript did not provide information regarding
the verb's tense, it was inferred from the context. Contractions such as
gimme, gonna, I'm, it's, and doesn't were excluded, as were participles such as
broken or gone and the quasi-modal used to. A very small number of mimicked
utterances at early ages, regular participles, and irregular participles that
are identical to past tense forms may have been included. Intentionally
included were verbs that were not very clearly uttered but were clear
enough for the transcriber to have made a reasonable guess; so were some
slight phonetic variations such as -in for -ing, particularly for Sarah, whose
samples were transcribed more narrowly than the others. Brown (1973)
notes, however, that all three children's speech was carefully transcribed
with regard to the presence or absence of phonetic material corresponding
to inflections.

For Adam, Eve, and Sarah, have, be, and do were included only when
they were used as main verbs (i.e., possessional have, have to unless tran-
scribed as hafta, copula be, and pro-verb do), never when they were used as
auxiliaries (i.e., perfect auxiliary have, progressive and passive auxiliary be,
do used to form questions, negations, or emphatics). This is consistent with
the criteria used by Brown (1973; Brown, Cazden, & de Villiers, 1971),
Bybee and Slobin (1982a), and Slobin (1971).
For the other children, and for the Gleason, Warren-Leubecker, and
Gathercole collections, we used a Sun Microsystems Sparcstation 4 running
under UNIX to tabulate all irregular past tense utterances. Using the FREQ
program, we extracted the number of occurrences for each uniquely irregu-
lar verb listed in the Pinker and Prince (1988) appendix, together with all
forms ending in -ed, with the exception of the no-change verbs, read (which
is orthographically a no-change verb), and all forms of do, be, and have (i.e.,
neither auxiliary nor main verb usages were counted). We then isolated all
the overregularization errors from this list by removing regular verbs and
other part of speech categories. Because we did not check this large collec-
tion by hand, we were unable to exclude overregularized participles such

27

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
MONOGRAPHS

as in the window was broked. Since we did exclude correct participle form
they were distinct from the past tense forms (about 60 irregular verbs h
this property; see Pinker & Prince, 1988), this can result in an overestima
of overregularization rates, although it would be small. The word seed
sented particular problems since most of its uses are as a noun rather
as an overregularization of see; these were eliminated by hand. Repeti
were counted separately, as with the Brown children.9
The verb get is complicated. When adults speaking the standard di
use got in a stative possessional sense (as in I've got an ice cream cone), it is
perfect participle of get (meaning "obtain"), accompanied by some for
the auxiliary have. The meaning is possessional because of the semanti
perfect aspect in English: if the state resulting from obtaining something
the past currently holds, you possess it now (Bybee, 1985). However
children do not attend to the auxiliary, it would be natural for the
reconstrue got as a present tense form meaning "possess," and ther
numerous forms like Look, I got an ice cream cone that suggest that they
often use got as a present tense verb. For these usages we would erroneou
credit the child with the correct past form of get. Kuczaj (1976) noted
problem and used the context to distinguish present from past usage
got in Abe's speech. For Adam, Eve, and Sarah, we excluded all forms
were clearly present tense statives. For the other children, our obser
overregularization rates for get are probably underestimates, and sinc
is a frequent verb, the overall overregularization rates across verb to
will be, too. As we shall see when the verb is excluded, however, the degr
of underestimation is small.

To verify the accuracy of the machine-generated tabulations for

children other than Adam, Eve, Sarah, and Abe, we compared an exha
tive hand tabulation for these four children with machine-generated t
like those used for the other children. The hand-generated totals w
calculated by (1) extracting all utterances containing irregulars and all utt
ances with forms containing -ed, not including have, be, do, and the
change verbs, and (2) checking these utterances to remove all particip
nouns, and all other nonpast forms. The mean discrepancy between
two estimates of the overall overregularization rates for the four chil

9 Because the ChiLDES transcripts contain typographical errors and inconsistenci

a handful of past tense forms may have gone undetected by the automatic search pr
dure. For example, Michael Maratsos has called our attention to an overregulariz
rendered as fell'd in April's transcripts. We ascertained that such isolated misses are
by checking for all words spelled with 'd in the transcripts of Adam, Eve, Sarah
Abe; none were overregularizations. To be consistent, we did not add April's fell'd to
overregularization total, because the transcripts from various children contain a smal
unknown number of mistranscribed correct irregulars that our searches also missed,
as gotalright and stucki'm.

28

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
 MARCUS ET AL.

was less than 0.2 percentage points (maximum 1.2 percentage points), with
the machine giving a higher estimate for two of the children and the human
giving a higher estimate for the other two. We also compared the overregu-
larization rates derived by the two methods across verb types. The two sets
of estimates correlated fairly well: r = .98, .77, .91, and .82 for the four
children, respectively, and the overregularization rates averaged over types
were higher when estimated by machine for two of the children, lower
for the other two (never by more than 2 percentage points). Most of the
discrepancy can be attributed to small samples for some verbs, where dis-
agreement over a single sentence can greatly affect the overregularization
rate for that verb (e.g., if a verb was used twice, once correctly and once
incorrectly, its estimated overregularization rate could change from 0% to
100% if one of the sentences is omitted or misclassified). For the children
other than Adam, excluding verbs used fewer than 10 times brings the
correlations between verbs' overregularization rates calculated by hand and
by machine to .98 or greater. Thus, in several analyses in this Monograph,
when it is important to exclude verbs that were produced too few times to
yield reliable estimates of their overregularization rates, we use a criterion
of a minimum of 10 tokens per verb per child.
For several analyses it is necessary to tabulate the frequency of use of
past tense forms in the speech of the adults that habitually talked to the
children. In all cases the same criteria for counting verbs were used for
adults and children. Although we did not distinguish among the different
adults talking to a given child (e.g., the mother, father, or psycholinguists,
who came to be treated like family members), we did keep separate tallies
for the sets of adults talking to each child, so each child has his or her own
adult speech data.

III. CALCULATING OVERREGULARIZATION RATES

We defined the child's "overregularization rate" as the proportion of

tokens of irregular past tense forms that are overregularizations:

(No. of overregularization tokens)

[(No. of overregularization tokens)
+ (No. of correct irregular past tokens)]

Overregularization tokens included both stem + ed and past + ed forms. Vir-

tually all children's past tense forms are in past tense contexts (Brown, 1973;
Kuczaj, 1976), so we need not take into account the semantic correctness of
the tense marking. Overregularization rates were calculated over tokens for
a given verb for a given child and over tokens of all verbs for a given child.

29

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
MONOGRAPHS

A. Rationale for Calculation of Overregularization Rates

Note that this measure of the overregularization rate excludes no-mark-

ing errors (stems in obligatory past tense contexts). This is because no-mark-
ing errors could be caused by a failure to try to mark the past tense at a
(in which case the stem would be used because in English it is the most
frequent and grammatically nonspecific form), and the issues at hand per
tain to how tense is marked, on those occasions when the child decides to
mark it at all, not to whether the child decides to mark tense, a logically
independent question. For example, if overregularization were measured
as a proportion of all forms in past tense contexts (by adding no-marking
errors to the denominator), it would appear that young children's overregu-
larization rates were low, but that could just be because they were usually
not trying to mark tense at all, not that they were successfully suppressing
errors. And if overregularizations were lumped together with no-marking
errors to yield an overall error rate (by adding no-marking errors to the
numerator), then overregularization errors themselves would be hidden
from view; a given error rate could correspond to nothing but overregulari-
zation errors, nothing but bare stem errors, or anything in between. In
particular, tests of a regression in development (the left arm of the U-
shaped developmental sequence) would be impossible. Cazden (1968, p.
437) describes the sequence as "no use, followed by infrequent but invari-
ably correct use, followed only later by evidence of productivity"; the distinc-
tion between "use" and "correct use" presupposes that correct forms are
being considered as a proportion of total marked forms (irregulars plus
overregularizations). Thus, Rumelhart and McClelland (1986), using the
same rationale that we do, plot a quantity corresponding to our overregular-
ization rate in their demonstration of how their model mimics U-shaped
development, as shown in the graph that we reproduce below as Figure
17. That is, Rumelhart and McClelland assume that the past/nonpast tense
distinction has been mastered independently of the computations per-
formed by their model in deriving past tense forms, so they simply feed it
correct stem-past pairs from the start.
Of course, the decision to count only overtly tense-marked forms be-
cause they are clear examples where the child has decided to mark tense
does not imply that nonmarked stem forms exclusively represent the ab-
sence of a decision to mark tense. Rather, no-marking errors are highly
ambiguous and can arise in at least four ways. First, as mentioned, they can
represent a failure to attempt to mark past tense. Second, while attempting
to mark tense, the child may retrieve the information that an irregular past
form of a given verb exists but fail to retrieve its phonetic content (see n. 7
above); the past feature or pointer would block the regularization, but the
child would not have the irregular form at hand either. Third, a child may

30

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
 MARCUS ET AL.

fail both to retrieve the irregular and to block the regular and might thus
feed the stem into the regular suffixation process, which ordinarily results
in an overregularization. But the regular suffixation process does not suc-
ceed 100% of the time for regular verbs, so it may occasionally fail here
too, yielding the unchanged stem. Finally, some irregular verbs actually
show no change between stem and past (e.g., hit, cut, put, set), and the child
might analogize the no-change pattern to similar irregular verbs.
During the early stages of language development, where stem forms
are the great majority of forms uttered (Brown, 1973; Cazden, 1968; see
also Chap. VI below), it is reasonable to assume that the absence of intention
to mark is the usual or exclusive cause. Although tense and agreement are
marked obligatorily in languages like English, the child cannot be born
knowing this. Thus, it is not surprising that there might be an initial period
in which children have not yet developed any mechanism for systematic
marking of any specific inflectional feature. Indeed, bare stems constitute
the great majority of young children's verb forms in English, not only in
past tense contexts, but also in progressive and third person singular con-
texts (Brown, 1973). Thus, the simplest hypothesis is that most of these
early stems represent failures to mark tense at all, not attempts that fail for
the three possible reasons listed above.
During the later stages, when the child frequently marks tense on irreg-
ular verbs, it is plausible that some of the stem errors represent attempts to
mark tense that end up unsuccessful. Luckily, we shall see in Chapter VI
that, during the time when overregularization errors occur, stem errors
become rare, so little hinges on whether or not one looks at an index that
includes them. Where relevant, we discuss these possible effects.

IV. ESTIMATES OF NO-MARKING ERRORS (STEMS IN

OBLIGATORY PAST TENSE CONTEXTS)

Although no-marking errors are not counted in the overregularization

rate, we will require them in a number of other calculations, such as how
often the past tense is overtly marked at all or how successful regular past
tense marking is. Unfortunately, automatic computer searches cannot dis-
criminate stems used in obligatory past tense contexts from stems used
properly as infinitives, imperatives, and present tense forms. Thus, the rele-
vant data must be tabulated by hand by a linguistically sophisticated scorer.
This heroic task was beyond the scope of this study but has already been
done for Adam, Eve, and Sarah by Courtney Cazden (Brown, 1973; Brown
et al., 1971; Cazden, 1966, 1968), summarized in unpublished tables that
she and Roger Brown have generously provided us. It has also been done

31

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
MONOGRAPHS

for Abe by Stan Kuczaj, summarized in an appendix to his thesis (Kuczaj,

1976).
For certain analyses in Chapters IV and VI, we will need to combine
data on the form of marking (correct irregular vs. overregularized) with
data on the presence of marking (unmarked stem vs. correct irregular). For
Abe this is straightforward because all token counts are provided in a single
table in Kuczaj (1976). For Adam, Eve, and Sarah this cannot be done
because Cazden did not count overregularizations at all (neither as irregular
errors, as correct irregulars, nor as correct regulars) and counted fewer
correct irregular past tokens than we did. The reason was that she needed
a constant definition of obligatory past context across stems and correct forms,
and many tokens occurred in contexts that were not unambiguously obliga-
tory for past tense and thus could not be included.10 In order to combine
the data sets we adopted the assumption that overregularizations were not
any more likely than correct irregulars to occur in "obligatory" contexts as
opposed to "nonobligatory" contexts; the distinction, after all, is defined in
terms of the knowledge of the observer, not the child. Therefore, we calcu-
lated what proportion of total irregular tokens (by our counts) Cazden listed
and multiplied this proportion by the number of our overregularization
tokens, yielding an estimate of the number of overregularization tokens
occurring in obligatory past tense contexts (i.e., an estimate of the number
of overregularizations that Cazden would have found had she looked for
them among the contexts in which she counted correct irregulars and un-
marked stems). These token estimates could then be meaningfully com-
bined with Cazden's counts of irregular and stem tokens.

V. OTHER ERRORS RELATED TO OVERREGULARIZATION

Note finally that our study excludes two kinds of errors that are some-
times lumped with overregularization but that are logically distinct from it
In languages with richer inflectional systems than English, children of
inflect a stem with an affix that is incorrect for some feature of the word
such as gender or animacy, or incorrect for some feature of the conte
such as case or definiteness (Pinker, 1984; Slobin, 1973). Pluralizing m
nouns (e.g., waters) or extending the third person singular suffix to oth
persons or numbers (e.g., we walks) would be examples in English. Suc
errors are best characterized as underdifferentiation-an insensitivity to som

10 Brown et al. (1971, p. 2) define obligatory contexts as including "adverbs like yester
day, marginal notations [comments by the transcriber], expansions [by adults], continu
of tense, etc."; attempts by the child to imitate adult sentences containing tensed for
were also counted.

32

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
 MARCUS ET AL.

systematic distinction relevant to inflection-rather than as overregulariza-

tion, which involves idiosyncratic lexical exceptions to a systematic inflec-
tional process.
Second, we are excluding overapplications of irregular patterns to inap-
propriate irregular verbs such as tooken or brang. Because irregular patterns,
even when used productively, have qualitatively different properties than
the regular suffix (reviewed in Chap. 8, Secs. IB, IV), they are logically
distinct from overregularizations and should be tallied separately. Theories
differ as to whether regular and irregular patterns are in fact handled by
the same kind of mechanism; according to the particular hypothesis we are
considering in this Monograph, they would represent competition between
two irregular forms for one memory slot, not a competition between a
stored irregular and the regular rule (for discussion, see Ullman & Pinker,
1990, 1992). These issues can be sidestepped in the present investigation;
Bybee and Slobin (1982a) point out that over-irregularizations are ex-
tremely rare in preschool children's language, and there are very few in
our samples. For Adam, Eve, and Sarah, Cazden (1966) reports only beat-
bate and hit-heet for Adam and beat-bet for Sarah, to which we can add bite-bat
for Adam. Sweepened was the only such form we happened to come across
for Abe, although, because there is no mechanical way of searching for such
errors, there may be others. Overextensions of irregular patterns to regular
verbs appear to be even rarer; we are aware only of trick-truck from Adam.

33

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
 IV. THE RATE OF OVERREGULARIZATION

In this chapter, we estimate children's typical overregularization rat

an attempt to resolve the empirical inconsistencies in the literature
test possible explanations of the nature of overregularization.
As discussed in Chapter II, the most frequent characterization in
literature is that once children begin overregularizing they do so al
time, replacing correct irregulars altogether. Other characterization
children varying freely between overregularized and correct irreg
forms, either at a rate near the rate of tense-marking regular ver
irregulars are not being discriminated from regulars) or at a rate that is
markedly far from values that can be interpreted as representing
temwide indifference (if neither the irregular form nor the overregular
form of irregular verbs is systematically favored, but a panoply of fact
settles a competition between them for each verb). In either case, a pref
ence for the irregular would emerge slowly before culminating in
performance. Such a stage of free variation would call into questio
kind of blocking or uniqueness principle (Maratsos, 1987) and would
unsolved the learnability problem of how children eliminate the inc
forms.

The third possible empirical pattern is that overregularization er

are rare relative to correct irregulars. If so, there would be no qual
difference between children and adults. Both would discriminate between
regularized and irregular forms, presumably because application of their
regularization mechanism to listed irregulars is blocked. Children's occa-
sional overregularizations, like adults' speech errors and adults' uncertaint
about low-frequency irregulars like smote, could be attributed to probabilistic
imperfect retrieval from rote memory. Strictly speaking, it would suffice to
show that the rate of regularizing irregular verbs is reliably less than 50%
to demonstrate that children possess some mechanism that acts to give th
irregular past tense form of a verb priority over the regularized form, con-
trary to the predictions of all hypotheses other than blocking. The claim
that that mechanism in fact is blocking (which actively suppresses overregu-

34

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
 MARCUS ET AL.

larization), as opposed to some weak statistical preference for the irregular

for some other undiscovered reason, would be more convincing the lower
the overregularization rate is, the more pervasive such a low rate turns out
to be, and the more clearly it can be shown that irregularity per se, not
some variable confounded with it, is associated with low rates of using the
regular suffix.
Of course there is no single quantity that we can call "children's over-
regularization rate." Any estimate must aggregate over children, ages, and
verbs, and the estimates will necessarily vary depending on which subsets
one chooses to include. Therefore, after calculating some initial estimates
of the overall rate, we will break down the data in a variety of ways, to ensure
against various kinds of averaging artifact and sampling error. Ideally, the
range of different estimates should not stray too far from one of the pre-
dicted values (near 100%, near the regular tense-marking rate, near 50%,
near 0%) as one focuses on subsamples.

I. OVERALL OVERREGULARIZATION RATE

We first calculated overall overregularization rates. The median over

regularization rate across the 25 children with individual transcripts wa
2.5%. Table 2 shows the relevant data for the 25 individual children, who
distribution of overregularization rates is plotted in the histogram in Figure
1. The distribution is roughly exponential, with most children at the ex
treme low end; only two children, April and Abe, overregularized mor
than 10% of the time (13.0% and 24.0%, respectively). The average overreg
ularization rate across the 25 children was 4.2%. The three group data bas
replicate this figure, with rates ranging from 1.0% to 6.5% and a mean o
3.7%. After we have examined some of the factors that affect overregulari-
zation rates, we will be in a better position to speculate on why Abe's rat
was so much higher than those of the other children (see Sec. VII below
and also Chap. VI); some of the difference, we shall see, may be artifactua
But even Abe was very far from overregularizing 100% or even 50% of th
time, and his is the highest rate we see. Thus, the global data suggest tha
overregularization is a relatively rare phenomenon; if they legitimately r
flect children's tendencies, it would suggest that children's language systems,
like adults', are strongly biased to suppress overregularization, contrary t
common belief.

II. CHANGE IN THE OVERREGULARIZATION RATE OVER TIME

Of course, the surprisingly low rates obtained may be an averaging

artifact: each child could go through a circumscribed U-shaped period of

35

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
MONOGRAPHS

TABLE 2

OVERREGULARIZATION RATES FOR INDIVIDUAL CHILDREN

Child Correct Stem + ed Past + ed Total Overreg Rate

Abe ......................... 1,786 465 99 2,350 .240

Adam ....................... 2,444 44 4 2,492 .019
Allison........................ 31 2 0 33 .061
April......................... 47 6 1 54 .130
Eve ......................... 283 23 1 307 .078
Naomi ....................... 378 34 2 414 .087
Nat ......................... 52 0 0 52 0
Nathan ...................... 243 11 3 257 .054
Peter ........................ 853 17 4 874 .024
Sarah ...... ... ................ 1,717 61 4 1,782 .036
Hall et al. (1984):
ANC .................... . 79 2 0 81 .025
BOM ...................... 112 1 0 113 .009
BRD ...................... 128 2 0 130 .015
CHJ ....................... 151 4 0 155 .026
DED ..................... 106 5 0 111 .045
GAT ...................... 159 10 0 169 .059
JOB ....................... 130 0 0 130 0
JUB ....................... 132 8 0 140 .057
KIF........................ 100 0 0 100 0
MAA...................... 105 2 0 107 .019
MIM ...................... 77 0 0 77 0
TOS........................ 84 0 0 84 0
TRH ..................... 47 3 0 50 .060
VOH ......... ............. 64 1 0 65 .015
ZOR ........................ 98 0 0 98 0

Mean of individual children = .042

Aggregate data bases:

Gleason (1980) .............. 472 32 1 505 .065
Gathercole (1979) ............ 454 16 1 471 .036
Warren-Leubecker (1982) ..... 317 3 0 320 .009
Mean of aggregate data bases = .037

constant or indiscriminate overregulariz

many more months of near-perfect perf
were a span of 33 months in which overre
of the months but went up to 96% in the ot
ization rate for the span would be only 2
Figure 2. (All our developmental graphs p
forms that are correct [100% minus the ove
the proportion overregularized, so that re
as U's, not inverted U's.)
Figures 3-6 plot monthly overregulariz
to four transcripts) for the children whose
lated longitudinally (Adam, Eve, Sarah, an
Appendix Tables Al-A4. For all four child
early in the span sampled and lasts for th

36

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
 16

14

12

T 10
_0

-o
%0"8

6-
z
Z

0 10 20 30 40 50 60 70 80 90 100

Overregularization Rate
FIG. 1.-Histogram of overregulari

Hypothetical C
100

" 90
80
_ 80
o

70

u 60

50

40

30

20

10

2-0 2-6 3-0 3-6 4-0 4-6 5-0

Age
FIG. 2.-Hypothetical developmental sequence that would yield low overregulariza-
tion rates as an averaging artifact.

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
 Adam
100

2 90
0o 80
S70
u 60
- 50
c 40
30
20

i 10

2-0 2-6 3-0 3-6 4-0 4-6 5-0

Age
FIG. 3.-Percentage of Adam's irregular past tense forms that are correct (100%
minus the overregularization rate).

Eve
100

O 90
S80
70
60
50o
c 40

a 30
20
2 lO
10
0 01.
1-6 2-0
Age
FIG. 4.-Percentage of Eve's irregular past tense forms that are correct (100% minus
the overregularization rate).

38

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
 Sarah
100

90
o 80
u 70
u 60

i-50
S40
30

? 20
N 10
2-0 2-6 3-0 3-6 4-0 4-6 5-0
Age
FIG. 5.-Percentage of Sarah's irregular past tense forms that are correct (100%
minus the overregularization rate).

Abe
100

? 90
o 80
u 70

60
50

S40
30
20

S10
0 '
2-0 2-6 3-0 3-6 4-0 4-6 5-0
Age
FIG. 6.-Percentage of Abe's irregular past tense forms that are correct (100% minus
the overregularization rate).

39

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
MONOGRAPHS

Eve, whose samples end at 2-3, we see fairly steady overregularization from
some time in the third year at least into the sixth year. Only for Abe is ther
any hint of a reduction in overregularization rate late in development.
The graphs demonstrate that low rates characterize the entire period
of overregularization. Adam's highest monthly rate of overregularizing i
only 6.8%; Eve's is 23.1% and Sarah's 15.8%. Even Abe, an outlier among
the 25 children, displayed an overregularization rate of only 47.6% in hi
most extreme month (see Fig. 6); this is the highest rate we find among 109
monthly estimates from the four children who were examined longitudi
nally."I

A. Is There a U-Shaped Developmental Sequence?

The ubiquitous claim that children pass through a U-shaped develop-

mental sequence in acquiring irregular verbs (in particular, that their per-
formance declines at some point) is based on observations by Cazden (1968),
Ervin and Miller (1963), and Miller and Ervin (1964) that a few irregular
forms were often used correctly by the children before they started to over-
regularize. U-shaped development has never been documented quantita-
tively, however, and Marchman (1988) and Stemberger (1989) have ques-
tioned whether it exists. In this subsection, we define the phenomenon
precisely and test for it in our data.
Many measures of children's performance that one plots against time
will show some kind of dip, for a variety of reasons, and many others will
not. Until one specifies with precision what a U-shaped sequence is supposed
to refer to, one cannot test whether such a sequence exists or what causes
it. We will follow the definitions of Cazden (1968) and Rumelhart and
McClelland (1986), who were referring to a transition from a period in
which past tense forms are marked correctly whenever they are marked at
all to a period in which some overregularization errors occur as well (see

" Note that it would not even be sound to conclude that Abe went through a stage
at which his overregularization rate was 47.6%, because of the statistical phenomenon of
regression to the mean. This particular month was chosen post hoc because of its high
rate. Any monthly estimate reflects a sum of the child's true overregularization rate (a
psychological tendency that could be measured as how frequently the child would overreg-
ularize over a very large number of comparable opportunities at that age) and an error
component (from sampling and other sources of noise independent of the true rate). By
deliberately selecting the month with the maximum observed overregularization rate, we
are simultaneously selecting for samples with high true overregularization rates and for
samples with high noise in the direction of overregularization. The true rate, therefore,
would be expected to be lower than the sampled rate in such months. For similar reasons
it would not be sound to point to Abe's data from 4-8 and conclude that he went through
a stage in which his overregularization rate was zero.

40

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
 MARCUS ET AL.

Chap. III, Sec. IV). The previous section showed that when overregulariz-
ations appear, they neither predominate nor alternate with the correct ir-
regulars but are always a minority. However, one can still test whether a
period of extremely accurate performance precedes the first overregu-
larizations-in other words, whether children at some point in development
get worse (even if they never get very bad).
Our sample includes nine children with extended longitudinal data.
For two of them, Abe and Nathaniel, the first transcript contains an overreg-
ularization. If these transcripts happen to coincide exactly with the boys'
very first uses of irregular past tense forms, the transcripts show that they
did not undergo U-shaped development. But it is more likely that the tran-
scripts began well after irregulars of one or both kinds were being pro-
duced, in which case the transcripts are uninformative. Note that, for the
same reason, any set of transcripts that does not begin early enough in the
child's language development is apt to underestimate the length of any
overregularization-free sequence in which irregular pasts are used.
For the other seven children, one does see correct irregulars in the
transcripts before the appearance of the first overregularization (not sur-
prisingly, since three of them-Adam, Eve, and Sarah-were the ones stud-
ied by Cazden). However, the impression of a developmental change could
be a sampling artifact. Since we now know that children's overall overregu-
larization rate is low, their tendency to overregularize could be unchanged
throughout development, but their early samples might simply be too small
to contain any examples of overregularization. Imagine drawing playing
cards from a deck with replacement, looking for a black king (whose fre-
quency in the deck is approximately equal to children's overregularization
rate). One might have to draw a large number of cards before the first one
appears, even if the deck is complete and properly shuffled.
It is not legitimate to test for a change in rate by comparing overregu-
larization rates before and after the first overregularization because the post
hoc nature of the dividing line will inflate the chances of obtaining a spuri-
ous difference. One stringent test can be conducted as follows. If the child's
overregularization rate is p, then, under the null hypothesis of no change
in this rate over time, the chance that the first irregular verb form in the
sample will be correct is 1 - p. If the likelihood of a child overregularizing
an utterance is unaffected by whether the child overregularized the previ-
ous past tense utterance (an assumption we will examine below), then the
chance that the first two utterances will both be correct is (1 - p)2, the
chance that the first three will be correct is (1 - p)3, and so on. One can
test whether there is an improbable run of consecutive correct irregular
past tenses at the beginning of a child's records by calculating (1 - p)",
where n is the number of irregulars in the transcripts preceding the tran-
script containing the first overregularization. (Correct forms preceding the

41

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
MONOGRAPHS

TABLE 3

TESTS OF U-SHAPED DEVELOPMENT: CORRECT IRREGULARS PRECEDING THE SAMPLE

WITH THE FIRST OVERREGULARIZATION

FIRST
OVERREGULAR-
IZATION CONSECUTIVE
CORRECT IN OVERREGULARIZATION
CHILD Age Sample PRECEDING SAMPLES RATE PROBABILITY

Adam ...... 2-11 18 381 .01926 .0006

Allison ...... 2-10 6 18 .06061 .3245
April ....... 2-1 2 3 .12963 .6593
Eve ........ 1-8 5 7 .07818 .5656
Naomi ...... 1-11 20 15 .0870 .2555
Peter ....... 2-6 14 275 .024027 .0012
Sarah....... 2-10 33 231 .03648 .0002

first overregularization
including them would corr
the child had brought the
of that very session.)
Table 3 shows the result
of obtaining a string of co
the early overregularizatio
is very small (.0012 or les
no conclusions about U-sh
regularization, although c
appear substantially befo
overregularization rates.
single sample in a meta-a
bility of observing such lo
the null hypothesis. Follow
individual probabilities t
root of the number of c
score back to probability
their sentences, avoids an
of obtaining the data unde
rate is .000045.

These analyses crucially depend on the assumption that the probab

of overregularizing is independent of whether the preceding irregula
tense utterance was correct or an overregularization. We tested the validi
of the analyses in two ways. First, we estimated the validity of the assump
directly on a sample of data, namely, Abe from 3-3 to 3-4. His condi
probability of producing a correct irregular past (as opposed to an overre
larization) given that his preceding irregular past tense form was co

42

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
 MARCUS ET AL.

was .58; his conditional probability of producing a correct irregular past

given that his preceding irregular past tense form was an overregularization
was .62. The difference is negligible and not in the direction that would
compromise the probability analyses. Second, we tested the robustness of
the conclusions under a pessimistic assumption of one kind of violation of
independence: that every other correct verb form in a pre-overregular-
ization string was completely determined by a perseveration effect from the
preceding correct verb form. Under this assumption, each string of correct
forms would be effectively half as long as originally measured, so we divided
the number of initial consecutive error-free pasts in Table 3 by two. The
probabilities of obtaining strings of correct forms of those lengths or
greater, under the hypothesis of no change in underlying overregulariza-
tion rates, remain less than .05 for Adam, Peter, and Sarah individually and
less than .06 for the aggregate sample in the meta-analysis.
In sum, there is quantitative evidence that children's first overregulari-
zation follows an extended period in which their overtly tensed irregular
verbs are all correct, and this effect can be demonstrated very strongly for
some children. In this sense, children do get systematically worse as they
get older.
Other senses of "U-shaped development."-Some discussions in the lit-
erature claim either that there is no such thing as U-shaped development
of irregular tense marking or that connectionist models developed after
Rumelhart and McClelland's can account for it better than their model did.
These discussions, however, refer to very different phenomena than the
one discussed by Ervin and Miller, Cazden, and Rumelhart and McClelland.
We review these other senses of "U-shaped development" and point out the
extent to which the corresponding empirical claim is supported in our data.
In the network model described by Plunkett and Marchman (1990),
early acquired verbs were permanently resistant to overregularization; the
so-called onset of overregularization in their model pertained to its perfor-
mance on newly acquired verbs.12 In actuality, once children begin to over-
regularize, they produce errors for many of the verbs that earlier they had
produced correctly (i.e., the sequence does not consist of correct perfor-
mance for some irregulars early and overregularization only for newly ac-
quired verbs, with the early correct ones eternally protected). For Adam,
15 of his 23 overregularized types (65%) had been produced correctly at
least once before. For Eve and Sarah, the respective figures were 3 of 9
(33%) and 15 of 26 (58%).
Stemberger (1989) suggested that Rumelhart and McClelland (1986)

12 "Interestingly, however, the initial 20 verbs are likely to continue to be mapped

correctly by the network, even in the presence of the other erroneous mappings" (Plunkett
& Marchman, 1990, sec. 4, p. 32).

43

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
MONOGRAPHS

were misled by the psycholinguistics literature into trying to model a non

istent phenomenon. He plotted data from Abe showing no signs of
shaped development. However, Stemberger is simply discussing a diff
quantity than the one Ervin and Miller, Cazden, and Rumelhart and McCle
land focused on. What Stemberger plotted was not overregularization
but a different measure, overall correct performance, defined as one min

[(No. of overregularization tokens) + (No. of stems in past contexts)

[(No. of overregularization tokens) + (No. of stems in past contexts)
+ (No. of correct irregular past tense tokens)]

The data in Appendix Tables Al-A4 confirm that the overall rate of co
performance does not decline notably with age for any of the children. B
it is not clear that anyone in the psycholinguistics literature has ever clai
that it did, so the point of Stemberger's objection is unclear.
Plunkett and Marchman (1990, 1991) refer to yet another sense of
term "U-shaped development" in arguing for the psychological realit
their connectionist models. In Plunkett and Marchman (1991), the lear
curves all start out at levels of performance far less than 100% and t
increase; the authors call the small wiggles in this overall increasing c
"U-shaped development." Although all the children we examine show
ups and downs in their monthly measures of overregularization rates, the
are many explanations of these blips, of which sampling error is the
plest.13
Finally, in Plunkett and Marchman (1990), any verb that is used once
correctly and then once incorrectly is characterized as undergoing U-shaped
development, which is misleading for a different reason: any stationary
stochastic process (e.g., a string of coin flips) will produce local sequences
with such patterns. Neither the fact that children produce such sequences
nor the ability of their model to do so is surprising.

B. When Does Overregularization Cease?

There are no signs of overregularization going away or even decreasing

in Adam's, Eve's, or Sarah's samples, which last through the early 5s for the
former and the latter. Perhaps because Abe's overregularization rate is
higher to begin with, one can discern a trend of gradual overall improve-

13 Kruschke (1990, p. 61) makes the point a slightly different way: "The recent work
of Plunkett and Marchman [1991] does not exhibit U-shaped learning, contrary to their
claims. They showed that acquisition fluctuated depending on the particular training
sequence, but they failed to mention that on average their model showed monotonic, not
U-shaped, acquisition."

44

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
 MARCUS ET AL.

ment (the slope of the best-fitting straight line is + 6.0 percentage points
per year), superimposed on seemingly random fluctuations that leave many
late months with higher overregularization rates than early ones. Clearly,
overregularization diminishes extremely gradually. On the basis of a per-
sonal communication from Slobin, Kuczaj (1977a) notes that it is still present
in school-age children of 9 or 10.
Two studies provide us with estimates of the overregularization rate in
older children. Moe, Hopkins, and Rush (1982) report a sample of 10,530
irregular past tense utterances from over 329 first graders. The overregu-
larization rate in their data is 2.8%. Carlton (1947) reports 2,196 past tense
tokens among the speech of 96 fourth graders she recorded. These included
13 overregularizations. If approximately 75% of past tense tokens in speech
are irregular (see Chap. V), the overregularization rate for this group is
0.8%.
Do overregularization errors ever completely disappear? Joseph Stem-
berger has kindly provided us with the full set of past tense overregular-
izations in his corpus of 7,500 adult speech errors. The list includes 25 past
tense overregularizations (18 of the stem, 7 of the irregular past form);
Stemberger (1989) suggests that the rate of adult speech errors might be
about one error per 1,000 sentences. If we assume that all sentences contain
verbs, that about 10% of verbs in casual speech are in the past tense (Adams,
1938; Smith, 1935), and that 75% of adults' verb tokens are irregular
(Slobin, 1971; see Chap. V), we get a very crude estimate of adults' overreg-
ularization rate of .00004-three orders of magnitude lower than pre-
schoolers' and two orders lower than fourth graders'.
So although in one sense both children and adults overregularize, there
is also a dramatic difference in their rates of doing so. Perhaps the differ-
ence is just a consequence of hearing more tokens of each irregular verb as
one lives longer, with more exposures leading to more reliably accessible
memory traces. For example, a negative exponential learning curve with a
time constant of one order of magnitude of improvement in retrieval proba-
bility for every 5 years' worth of irregular past tense tokens could handle
the reported overregularization rates from the preschool years through
adulthood comfortably. In the absence of more plentiful and finer-grained
data, it is premature to claim that there is no qualitative difference between
children and adults, but current evidence does not demand that there be a
difference.

III. DIFFERENCES IN OVERREGULARIZATION RATE AMONG VERBS

Another possibly misleading effect of averaging could result from com-

bining data on individual verbs: perhaps a few very commonly used verbs

45

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
MONOGRAPHS

are never or rarely overregularized, but most verbs are overregularized

most of the time or at least indiscriminately. Because these verbs may differ
from one child to another, the question can be addressed only by examining
overregularization rates for different verbs in individual children. The rele-
vant data for Adam, Eve, Sarah, Abe, and the other children are presented
in Appendix Tables A5-A9. Aggregate measures of overregularization
rate, described in Chapter VII, are presented in Appendix Table A10.
It is important to note, however, that sampling error can make esti-
mates of overregularization rates for particular verbs of particular children
extremely misleading. In the extreme case, if a verb is used only once, its
observed overregularization rate can only be 0% or 100%, regardless of the
child's actual overregularization tendency. If it is used twice, the observed
rate can be 0%, 50%, or 100%, and so on. Even with a low overall regulariza-
tion rate and slightly larger samples, high estimates of overregularization
rates for a given verb frequently will arise by chance. Histograms of the
overregularization rates for all a child's verbs show spikes at values corre-
sponding to ratios of small integers such as 0, 1/4, 1/2, 3/4, and 1. But if we
restrict attention to only those irregular verbs that were used 10 times or
more in the past tense (the cutoff suggested in Chap. III), we find that
Adam did not overregularize any of his 32 verbs at a rate higher than 10%,
Eve overregularized fall 8 out of 10 times but did not overregularize any of
the other 9 more than 20%, and Sarah overregularized throw 7 out of 10
times but did not overregularize any of the other 26 more than 33%. Histo-
grams of Adam, Eve, and Sarah's overregularization rates for verbs used a
minimum of 10 times are displayed in Figures 7-9.
Even for Abe, the extreme overregularizer, most commonly used verbs
fall at the low end of the distribution of overregularization rates (mod
10%-20%, median 30%, mean 32%). This is shown in Figure 10, a histo-
gram of overregularization rates for Abe's verbs used 10 times or more.
It is clear that the overall low overregularization rate is not an artifact
of averaging a few verbs that are never overregularized but used extremely
frequently with a majority of verbs that are usually overregularized but
used less often. Rather, low overregularization rates characterize most of
children's commonly used verbs. Nonetheless, it is also quite clear (especially
for Abe) that different verbs are overregularized at different rates, and
estimates of overall overregularization rates must be qualified by examining
different verbs. For Adam, Eve, Sarah, and Abe, chi-square tests amply
show that overregularization rates differ among verbs used more than 10
times (all p's < .001). It is statistically permissible to perform such chi-square
tests for four other children in our sample (GAT, Naomi, Nathan, and
Peter); the tests are significant for all of them. Chapter VII is devoted to
investigating properties of verbs that make them more or less likely to be
overregularized. The most important, not surprisingly, is frequency: verbs

46

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
 MARCUS ET AL.

that parents use in the past tense less often are more likely to be overregular-
ized by their children.
It is important to note that the correlation between verb overregulariza-
tion rate and verb frequency necessarily interacts with any cutoff designed
to eliminate small samples. The verbs that parents use less often are also
the verbs that their children use less often (see Chap. VII). Therefore,
excluding verbs likely to give rise to sampling errors will also tend to exclude
verbs with higher overregularization rates. Appendix Tables A5-A8 con-
firm that many infrequently used verbs yield high estimated overregulariza-
tion rates. In fact, computing the unweighted mean of the overregulariza-
tion rates of all verb types yields figures of 22% for Adam, 17% for Eve,
18% for Sarah, and 47% for Abe; the mean of the type means for all 25
children is 10.5%, the median 11%. One might wonder whether the rela-
tively high type mean for Abe when it is calculated over all his verb types
casts doubt on the conclusion that his grammatical system suppresses over-
regularization. Such doubts are not warranted. Means calculated over types
do not reflect the functioning of the child's grammatical system because a
verb used once is weighted exactly as strongly as a verb used a thousand
times. Rather, what the typewise mean is capturing is the distribution of
frequencies of the child's irregular verb vocabulary: the more low-frequency
verbs a child uses, the higher the mean. In fact, although the theory invok-
ing blocking and retrieval failure predicts low overregularization rates over-
all, it also predicts high overregularization rates for low-frequency verbs:
100% for verbs that have never been heard or attended to in the past tense
form (like shend for adults), lower but still high rates for verbs that have
been heard occasionally, and so on (see Chap. II). In a sense, a child's
seldom-used verbs are misleading for two reasons: we cannot be confident
about the actual overregularization rate of an irregular verb used once or
twice, and a child cannot be confident that a verb is irregular if he or she
had heard it used only once or twice.
Irregular verbs are, quite generally, overregularized at low rates, but
the verbs themselves can be very different from one another. This suggests
that the overall low overregularization rate reflects some phenomenon that
suppresses overregularization globally across the irregulars (modulo fre-
quency) and is unlikely to be attributable to any general bias of the child to
favor correct irregulars because of some lexical property that many of them
happen to possess. For example, it might be argued that children's gram-
matical systems allowed both overregularizations and correct irregulars but
that they had a response bias in favor of producing irregulars because they
tend to be shorter and simpler than overregularizations and that this is the
cause of the measured overregularization rate being less than 50%. Such a
counterexplanation can easily be ruled out. Among the irregular verbs that
our subjects used, 14 have correct past tense forms that are the same length

47

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
 Adam

30

25

2 20

w15
E
z

10

0 10 20 30 40 50 60 70 80 90 100

Overregularization Rate
FIG. 7.-Histogram of overregularizat
per verb).
7
Eve

cn

04

E3-
z

0 10 20 30 40 50 60 70 80 90 100
Overregularization Rate
FIG. 8.-Histogram of overregularization
verb).
48

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
 Sarah

20-

15-
0

0>

5-

0-
0 10 20 30 40 50 60 70 80 90 100
Overregularization Rate
FIG. 9.-Histogram of overregularization rates of Sarah's verbs (10 or more tok
per verb).
7
Abe

6-

5-

4-

E
z

31-
0

0-
0 10 20 30 40 50 60 70 80 90 100

Overregularization Rate
FIG. 10.-Histogram of overregularization rates of Abe's verbs (10 or more tokens
per verb).
49

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
MONOGRAPHS

or shorter than their overregularized counterparts, usually with a simila

phonological structure: be, buy, do, feel, go, hear, keep, leave, lose, mean, say
sleep, sweep, and tell. If simplicity, not irregularity, were responsible for the
overall low overregularization rates, then these verbs, lacking that advan-
tage, should show overregularization rates much closer to 50%. Instead, th
overregularization rates for tokens of these verbs were 0.7% for Adam,
9.3% for Eve, 2.8% for Sarah, and 19.9% for Abe. The mean overregular-
ization rates across these verb types were 18.3% for Adam, 5.25% for Eve,
5.0% for Sarah, 36.9% for Abe, and about 8% for the aggregate rates from
all 19 children who overregularized at least once (see Chap. VII). Medians
were far lower, usually 0%. All these figures are close to or lower than those
obtained for the full set of irregular verbs. More generally, the data suggest
that the only property of irregular past tense forms that is likely to account
for children's strong tendency to prefer them to overregularizations is irreg-
ularity itself.

IV. CHANGES IN OVERREGULARIZATION RATES OVER TIME

AMONG DIFFERENT VERBS

The most stringent test of the hypothesis that overregularization is

probabilistic and relatively rare event would look at the fate of individu
irregular verbs for individual children as they grow older. This is indepe
dent of the overall level of overregularization for different verbs that
have just examined, just as the waveform of a sound wave is independen
of its amplitude and its DC component. For example, it is possible tha
each child goes through a stage for each verb during which the verb
overregularized exclusively (see Fig. 11) or as often as it is produced cor
rectly. If these stages are fairly brief and circumscribed, the steady low rat
of regularization could be an averaging artifact of a sequence of deep na
row U's, one for each verb. If so, or if the verbs all follow some other set
of out-of-phase developmental curves, the protracted period of overregu-
larization would reflect a failure to apply blocking to different verbs at
different times.

There are other possible interactions among children, verb, and age
that would be noteworthy. One can determine whether any verbs cease to
be overregularized altogether before the end of the period or, alternatively,
whether a child begins to stop overregularizing all verbs at the same time;
the latter finding would suggest that the child learns or develops the
blocking principle only at that point. Another possibility is that various verbs
follow different and largely unsystematic patterns, perhaps because, as we
have suggested, overregularization is a quasi-random performance deficit.
Clearly, low rates of overregularization are not an artifact of a sequence

50

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
 MARCUS ET AL.

Hypothetical Child

90

0 80
70
0
u 60

S50
' 40
0.
S30 I
20

210

0
Young Old
Age
FIG. 11.-Hypothetical developmental sequence for different verbs that would yield
low overregularization rates as an averaging artifact.

of transient overregularization stages, one for each verb: individual verbs

can be overregularized across large spans of time. The very first verb that
Adam overregularized was feel at 2-11; he also overregularized feel in his
last sample at 5-2. Similarly, throw was overregularized at 3-4 and at 4-4,
make at 3-5 and 5-2, and fall at 3-5 and 4-10. Sarah's first overregularization,
heared at 2-10, appeared again at 4-11; winned and maked also made appear-
ances in the samples separated by a year or more. Even the 9 months' worth
of samples from Eve contain falled at 1-10 and again at 2-2. More than half
of Abe's overregularized verbs (44) were overregularized over a span of 1
year or more; 32 were overregularized over a span of 2 or more years.
Unfortunately, when we turn away from the simple question of whether
overregularizations of a given word reappear across long time spans and
try to trace each one of a child's irregular verbs over time, we run up against
severe sampling limitations. As mentioned, small samples yield inaccurate
estimates of overregularization rates, and many of the samples of tokens of
a given verb for a given child in a given month were very small. Thus,
developmental curves for individual verbs with low token frequencies for
the child can oscillate wildly among a few discrete values, revealing little
about changes in the underlying true rates. With these caveats in mind, we
now examine curves for individual verbs for Abe, the most prolific overreg-
ularizer.

51

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
 MONOGRAPHS

For most of the 70 irregular verbs that Abe used, the curves can best
be described as chaotic and highly variable from verb to verb. They are
most conveniently summarized by pointing out four rough patterns, shown
in Figures 12-15. Some verbs, like eat (Fig. 12), are overregularized in the
earlier transcripts but appear to be completely mastered before the end of
the sampling period. (Other verbs with this pattern include cut, fall, go, make,
think, and throw.) A second class of verbs, such as say (Fig. 13), is rarely
overregularized at any point. (Find, forget, see, and tell are others.) A few,
such as draw (Fig. 14), are overregularized throughout the sample (build is
similar); such verbs were used only rarely and might be overregularized
because they are also rare in parental speech, although sampling error can-
not be ruled out. But many verbs, such as win (Fig. 15), bite, break, blow, buy,
catch, come, feel, get, know, put, and shoot, show no interpretable trend, oscillat-
ing between samples with high and low measured overregularization.
In sum, apart from haphazard variation (possibly due to sampling er-
ror) and overall low or high rates (to be discussed in Chap. VII), the only
meaningful temporal pattern for individual verbs seen in Abe's data is that
a few appear to be mastered in the late 4s, thereafter resisting overregular-
ization completely.

V. HOW HIGH CAN ESTIMATES OF

CHILDREN'S OVERREGULARIZATION RATE GO?

Given that the overregularization rate varies across verbs and, to a lesser
extent, across children and ages, there is no absolute "overregularization
rate"; the estimates will change somewhat with methodological decisions of
which verbs, children, and ages to include. The low estimates in Section I
above included all the samples meeting the criteria described in Chapter
III, without any exclusions based on how the data came out. Still, it is
conceivable that these criteria adventitiously included some samples of ques-
tionable representativeness, relevance, or accuracy. In this section, we will
test the robustness of the low estimated overregularization rate, by eliminat-
ing all samples that could be challenged, for any plausible reason, as poten-
tially deflating the measured overregularization rates. If these worst-case
estimates continue to be low after steps are taken to bias them as high as
possible, one can have more confidence that there is some real process
causing them to be low and that the effect is not an artifact of the adoption
of one selection criterion or another.
First, for some of the children the samples are small, and we may have
caught them in conversations where few irregular verbs were used in the
past tense, or perhaps only a small number of well-mastered ones. Hence,

52

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
 MARCUS ET AL.

we can eliminate all children who contributed less than 100 irregular verb
tokens to the overall sample.
Second, because of possible U-shaped developmental sequences, some
of the youngest children sampled may not have entered the stage at which
overregularization had begun. If so, it would be misleading to count their
samples from the pre-overregularization period (with rates of 0%, by defi-
nition) because the psychological process of interest was not yet occurring.
Moreover, even if overregularization had begun, it may not have reached
its peak. So, following the suggestion of one of the referees of an earlier
draft, one might question the inclusion of Allison, April, Nat, and Peter,
none of whom were recorded any later than 3-1. These children could not
have been excluded a priori because Abe had his worst month of overregu-
larizing at 2-10 and Eve overregularized at 1-8. But let us assume that
these four children were on the verge of overregularizing at high rates and
exclude their data. We will continue to include Eve, even though she falls
within the age range of the children excluded here for being too young;
this decision renders overregularization estimates higher than they would
be if she were excluded. The rationale might be that her language develop-
ment was precocious and hence her true overregularization tendency was
indeed manifested in the available samples.
Third, we might be sampling overregularization-shy children from the
opposite arm of the U, where the errors have already disappeared. The
Hall et al. (1984) children might be suspect for these reasons. (In fact, data
from the children we analyzed in longitudinal detail-Adam, Sarah, and
Abe-do not themselves call for such an exclusion because, while Abe
showed an overall decline in the age range corresponding to the Hall et al.
children, he was still overregularizing at rates from 0% to 22%, and Adam's
and Sarah's overregularization rates increased.)
Turning now to the different samples within longitudinal records, we
can consider the fact that there was a demonstrated overregularization-free
period in the early months of Adam, Peter, and Sarah. So here we will
exclude all correct forms in the transcripts preceding the transcript in which
the first overregularization is found.
Fifth, perhaps not all verbs deserve to be included. Get could be prob-
lematic for children other than Abe, for whom Kuczaj (1976) carefully
coded all forms as the past of get versus the present of got. For Adam, Eve,
and Sarah, recall that we checked contexts and excluded present tense
forms, but some cases counted as past forms could have been present forms.
For the other children the computer counted all instances. We will report
estimates that exclude all instances of got for children other than Abe.
Sixth, one might argue that have, be, and do (included only for Adam,
Eve, and Sarah) are special. Even though we did not include any auxiliaries
in our analyses, these verbs obviously share their morphemes with the auxil-

53

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
 Abe: Eat
56 13 1 2 3

1001 6 3 2 8 1 5
o 90 1
LM
o 80
0
( 70"
C 60
i 5050
cc 40
0.
0
o 30
L_ 20
10
2 2

2-0 2-6 3-0 3-6 4-0 4-6 5-0

Age
FIG. 12.-Example of a verb that resists overregularization as the child gets older

Abe: Say
5 15 13 17 12 10 10 5 4 10 9 8 3 11

1001
9 18 21 16 5 13 s
12 62 8
2 7

90
C)
21 8
m 80
0

70

0 60

50

c 40
o 30
"20
No

10

0 .....
2-0 2-6 3-0 3-6 4-0 4-6 5-0
Age
FIG. 13.-Example of a verb that is rarely overregularized at any age

54

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
 Abe: Draw

100

"i 90
0 . 80
70

0 60

I- 50
40
e 40
0 30

20

10

0
2-0 2-6 3-0 3-6 4-0 4-6 5-0
Age
FIG. 14.-Example of a verb that is overregularized throughout development

Abe: Win
2 2 2 4

100

- 90
o
S80
70

0 60
50
a)
e 40

o 30
20

10
3 1 4

2-0 2-6 3-0 3-6 4-0 4-6 5-0

Age
FIG. 15.-Example of a verb with a chaotic developmental pattern

55

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
MONOGRAPHS

iaries, and they are irregular not only in the past but in the third perso
singular present. Conceivably, the child might treat nonauxiliary version
of have, be, and do similarly to the auxiliary versions and differently from
other verbs. Actually, when Stromswold (1990) examined overregulariza-
tion rates for auxiliary and nonauxiliary forms of these verbs in a large
sample of ChiLDES children overlapping with the one we use here, she
found that the nonauxiliary versions-that is, the ones that we have been
counting-were overregularized at rates comparable to those of the other
verbs that we report here. Counting correct forms from the children wh
made overregularization errors during the developmental spans in which
they made them (a procedure that yields slightly higher estimates than the
ones we report), Stromswold obtained overregularization estimates of 2.5
for have, 4.3% for be, and 7.5% for do (see also App. Tables A5-A10). But
again we will report the effects of excluding them.
Table 4 summarizes the results of various exclusions on the estimated

overregularization rates. The results are clear. Even combining all th

exclusions in an effort to avoid every possible bias toward low estima
rates, we are only able to push the estimate up to a mean across child
of 9.8% and a median across children of 8.7%, ranging from 3.6% to A
24.0% (which we will focus on below). These worst-case estimates are s
far closer to 0% than they are to 100%, 50%, or, for the bulk of the child
even 25%. They are low enough that it seems difficult to avoid the con
sion that the child's grammatical system contains some mechanism th
while allowing overregularizations to occur, is strongly biased against them
in favor of the correct irregular counterpart.14

VI. COMPARISON OF OVERREGULARIZATION WITH REGULARIZATION

The fact that the absolute rate of overregularization is quite low do

not disprove the suggestion (e.g., in Clark, 1987) that children fail to d
criminate regular from irregular verbs during the period in which they ar
overregularizing. Conceivably, children inflect regular verbs only 2.5%
the time during this period, leaving the regulars uninflected the remainin
97.5% of the time. If so, irregular past forms would not be blocking t
regularization of irregular verbs, as the blocking-and-retrieval-failure
pothesis claims; instead, the low rate of overregularization would presu
ably mean that regular suffixation was an extremely error-prone proc

"14 Moreover, the shape of the developmental curves for Adam, Eve, and Sarah
virtually unchanged, and the tests of U-shaped development reported in Sec. IIA of C
IV all remain significant.

56

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
 MARCUS ET AL.

TABLE 4

UPPER-BOUND ESTIMATES OF CHILDREN'S OVERREGULARIZATION RATES

A. ESTIMATE OF OVERREGULARIZATION RATE

Exclusion Mean Median Range

Samples with fewer than 100 past tokens .............. .05 .03 .00-.24
Selected young children (Allison, April, Nat, Peter) ..... .04 .025 .00-.24
Hall et al. (1984) children ........................... .07 .06 .00-.24
All three subsets ................................... .09 .07 .02-.24
Adam, Eve, Sarah, and Peter before first overregular-
ization ........................................ . 05 .03 .00-.24
Combined with excluded children ..................... .09 .07 .02-.24
got (other than Abe) ............................... ..05 .04 .00-.24
have, be, do (Adam, Eve, Sarah) ...................... .05 .03 .00-.24
All four verbs ...... ... ... ................... ...... . .05 .04 .00-.24
All exclusions combined ............ ............... . 10 .09 .04-.24

B. UPPER BOUNDS FOR INDIVIDUAL CHILDREN, FOLLOWING ALL EXCLUSION

Child Correct Stem + ed Past + ed Total Overreg Rate

Abe ......... 1,786 465 99 2,350 .240
Adam ....... 1,276 43 4 1,323 .036
Eve ......... 175 21 1 197 .112
Naomi....... 281 33 2 316 .111
Nathan ...... 214 11 3 228 .061
Sarah ....... 963 61 4 1,028 .063

failing most of the times it is in

tinue to be used, together wit
regularization process failed.
The data necessary to evaluat
the rates of inflecting regular v
estimates are presented in ful
show that children are not fai
verbs. The rates of correctly aff
the months including and follo
summing over tokens, are 73%
97% for Abe. These are way hi
rates, and the difference can b
data are available for the four
are the months for which Caz
contexts, which do not extend
nently, we can compare the rate
tory verbs with the rate of pr

57

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
MONOGRAPHS

of all obligatory contexts, or

[No. of overregularization tokens (in obligatory contexts)]

{(No. of overregularization tokens) + (No. of irregular past tokens)'
+ [No. of stem tokens (in obligatory contexts)]}

where our counts of the number of overregularization tokens are scale

to make them commensurable with Cazden's tallies, using the procedur
explained in Chapter III. These overregularization rates were 0.6% fo
Adam, 4.5% for Eve, 1.1% for Sarah, and 23.2% for Abe, far lower than
the rate of regular marking. Moreover, these rates were lower in all 65
individual months.

Although children clearly discriminate irregular from regular verbs in

general, it is not clear that they discriminate them in the circumstances that
lead to overregularization errors. We have suggested that overregulariza-
tion is caused by a failure to retrieve an irregular past tense form, leadin
to the application of the regular rule to the stem. If retrieval is all or none,
then, at the moment the child attempts to retrieve the irregular form but
fails, the verb should be indistinguishable to the child from a regular form
and should be overregularized by the same process that supplied the affi
of a correct regular past tense form. Not all retrieval failures should lea
to overregularization, just as not all regular verbs in obligatory past tens
contexts are successfully marked for past tense. But among usages of irregu-
lar verbs that are not correct irregular past forms, the proportion that are
overregularizations (as opposed to being left unmarked), which equals

[No. of overregularization tokens (in obligatory contexts)]

{(No. of overregularization tokens)
+ [No. of stem tokens (in obligatory past contexts)]}

should be identical to the proportion of regular verbs that are successfull

affixed (as opposed to being left unmarked) in past tense contexts.
Stemberger (1989) compared these proportions on two of Abe's sam-
ples and found that the overregularization rate calculated as a proportion
of nonirregular past tense attempts (overregularizations plus stems) was far
lower than the rate of inflecting regular verbs correctly. (The rest of Abe's
samples contain very few unmarked verbs in past tense contexts, either
regular or irregular, so both marking rates are close to 100%; see App.
Table A4.) We performed an analogous test on Adam, Eve, and Sarah
again after adjusting the number of overregularizations to be commensur
ble with Cazden's counts. Replicating Stemberger, we find that the overa
rates of overregularization given nonuse of the irregular during the chi
dren's periods of overregularization (13.3% for Adam, 15.9% for Eve, and

58

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
 MARCUS ET AL.

10.4% for Sarah) were less than their rates of inflecting regular verbs
(66%-97%) and that this was true for 32 of the 33 individual months
available.

This analysis shows that when children fail to use an irregular past
tense form in a past tense context, they do not treat it identically to a regular
verb; they overregularize the irregular verb far less often than they correctly
suffix a regular verb. In other words, irregular verbs tend to suppress appli-
cation of the regular process even when they are not supplied in the correct
irregular past tense form. Assuming that there is no consistent reason why
children should decide to mark tense on irregular verbs less often than on
regular ones, there are two plausible explanations for this difference. One
is that the child is analogizing the no-change pattern seen in verbs like cut
and set to other irregular verbs. Later, we review abundant evidence that
verbs that end in t or d are particularly susceptible to this interference.
Second, as mentioned in note 7 above, retrieval of an irregular may not be
all or none. On some occasions, the child may not retrieve the content of
an irregular entry (or may not retrieve enough of it to allow the form to be
articulated) but might retrieve the information that a past tense entry exists
(i.e., the child might retrieve only the pointer or tag to the entry). Since it
is the existence of a past tense entry that ordinarily blocks regularization
(rather than the particular phonetic content of any irregular form), such
piecemeal retrieval would be sufficient to prevent an overregularization
without actually supplying the correct irregular form.

VII. COMPARISON WITH PREVIOUS ESTIMATES IN SPONTANEOUS SPEECH

Given the stereotype that children go through a stage in which they

always overregularize or even overregularize in free variation with correct
irregular forms, our finding of a consistently low rate of overregularization
across children, ages, and commonly used verbs comes as a surprise. Why
has the low rate not been noted before?
One reason is that few investigators have counted the number of correct
irregular past tense uses; typically, only a list of errors is reported. A second
reason is that most summaries in the secondary literature, such as book
chapters and textbooks, do not even try to provide empirical citations to
document the overregularization rates they assume when discussing theo-
retical implications, as the quotations in Chapter II show.
Indeed, one source that does, Maratsos (1983), provides estimates that
bear no obvious relation to the study he cites. Recall that Maratsos claimed
that "overregularization, in general, ranged in frequency from .20 to .60 of
the children's uses," based on an unpublished talk by Kuczaj (1977b). Marat-
sos is referring to the study Kuczaj reported in his thesis (Kuczaj, 1976) and

59

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
MONOGRAPHS

in a published report (Kuczaj, 1977a), and these sources list a much low
range of overregularization rates calculated over uses: .01-.39 for diffe
children and .12-.24 for different subclasses of verbs.'5 Breaking down
rates according to individual children's performance with different s
classes (many obviously with tiny sample sizes) gives a wider range,
the range runs from 0.00 to 1.00, not from .20 to .60. Maratsos (pers
communication, October 1991) has explained to us that he in fact der
these estimates from Kuczaj's unpublished data by averaging overregu
ization rates over verb types for each month, not "uses," as his paper
the range of monthly typewise means ran from .16 to .69, with the bulk
the months falling roughly within the .20-.60 range that he reports.
discussed in Section III above, averages over types should be systemati
higher than the true overregularization rate because they overweight
rare and hence difficult verbs.

Let us now examine the actual data on overregularization rates in pre

schoolers in the original studies that report them. Bybee and Slobin (1982
report figures from a pooled sample of 31 children; the rates are liste
separately for different subclasses of irregular verbs. However, these figure
are also averages over verb types, ranging from .1 to .8 for the differen
subclasses. The numbers of verb types and tokens in each class are not
reported, so actual overregularization rates cannot be computed from th
data. Fortunately, Slobin (1971) published actual token frequencies of ind
vidual verbs summed over 24 of these 31 children (those from the Miller
Ervin, 1964, samples). From his table one can calculate that the pooled
children overregularized at a rate of 10.2%, far closer to our estimates.
Recently, Valian (1991) reported overregularization rates in a cross-
sectional sample of 21 children, divided into four groups based on mea
length of utterance (MLU), a measure that roughly correlates with gram
matical development in English-speaking children (Brown, 1973). Th
mean ages of the four groups ranged from 2-0 to 2-7, and the mean MLU
ranged from 1.77 to 4.22, the range spanned by Adam and Sarah betwee
2-3 and 3-8 and by Eve between 1-6 and 2-3 (Brown, 1973). Valian include
the main verb versions of have, be, and do and excluded their auxiliary
versions. Using the individual subject data that Valian has kindly provide
to us, we find that children in the first two groups (mean age 2-0 and
2-5) produced 67 correct irregular pasts and no overregularizations. Th
children in the third group (mean age 2-5, MLU 3.39) produced 109 correc

"15 These figures, taken from Kuczaj's (1977a) tables, refer to overregularizations as a
proportion of total usages of irregular verbs in past tense contexts, including unmarke
stems. Since most of the children's errors were overregularizations, not stems, the figure
for overregularization rates as a proportion of overtly marked verbs are similar, rangin
from .01 to .40.

60

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
 MARCUS ET AL.

irregular past tense forms and 9 overregularizations and had a mean over-
regularization rate of 8.4%. Of the three children in the fourth group (mean
age 2-7, MLU 4.22), one child produced 15 overregularizations of a single
verb, but the other 32 irregular past tokens were correct; the children's
mean was 13.2%. The mean overregularization rate of the children in the
study was 5.1%.
The ChiLDES archive contains a number of transcript sets that we
excluded for the methodological reasons listed in Chapter III. However, it
is of some interest to check whether they are consistent with the samples that
we did report. Among the transcript sets that are from Standard American
English-speaking children and in CHAT format, there were three sets of
transcripts from individual children and six group transcripts, ranging in
size from 5 to 1,596 irregular past tense tokens (median 54). The mean
overregularization rate of the three individual children was 12.5% (range
5.1%-16.7%). The six group data bases showed overregularization rates
ranging from 0% to 10.8%, with a mean of 3.25%.
The only partial anomaly in the empirical literature consists of the data
reported by Kuczaj (1976, 1977a). Qualitatively, those data are consistent
with our finding of a pervasive bias against overregularizing: neither the
mean overregularization rate for the cross-sectional sample nor that for Abe
comes close to the rates of 50% or higher that had been assumed in the
literature. But there are quantitative differences. As we have seen, Abe is
the outlier from among the 25 English-speaking children we studied in the
ChiLDES data base. Kuczaj's cross-sectional sample (1977a) of 14 children
from 2-6 to 5-6 includes six with overregularization rates comparable to
those reported here (1.1%-8.33%), but for the other eight the rate ranged
from 26.1% to 40.2%, with an average over all the children of 20.9%. Thus,
neither Abe nor most of the 14 children in Kuczaj's cross-sectional sample
appear to come from the distribution of unselected children we analyzed
(most of whom had very similar backgrounds to Kuczaj's subjects). But Abe
and the cross-sectional sample of children appear to come from the same
distribution. What Abe and the sample have in common is that they were
both studied by Kuczaj, as part of the only investigation we discuss that
recorded children's speech expressly for the purpose of studying verbal
inflection. Could there be some systematic factor in Kuczaj's methods that
would push estimated overregularization rates upward?
According to his dissertation, Kuczaj (1976, p. 5) selected his 14 chil-
dren from a larger sample of 23, according to the criteria of "(1) clarity of
speech, (2) willingness to play games (in particular, to imitate model senten-
ces when asked), and (3) willingness of the parents to aid the investigator by
participating as experimenters (that is, by playing games with their child)."
Among the ways that the parents aided Kuczaj were supplying the child
with erroneous feedback, choosing the times to turn the tape recorder on

61

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
MONOGRAPHS

and off, and asking the child to talk about specific events that took place in
the past and about specific hypothetical events. (This was another attemp
to elicit past tense forms because in English, as in many languages, the
inflection called "past tense" is also used to express hypothetical events
certain contexts, as in I wish I knew the answer or If I won the lottery.) As Kucz
points out, "The recorded conversations may then not be representativ
samples of normal conversation" (p. 8).
Kuczaj does not mention whether, in recording Abe, he tended to us
the same manipulations that he instructed the parents of his cross-sectional
sample to engage in. The transcripts themselves suggest that he may hav
We tested for a conversational style involving a high degree of past tens
elicitations in Abe's transcripts. Using the CLAN program COMBO, we
searched for occurrences in Abe's parents' speech of three representativ
sequences of words that would tend to call for past tense forms in subs
quent child utterances: what did, what'd, and happened. Abe's parents use
these forms in 1.9% of their utterances, more frequently than 21 of th
other 24 sets of adults in our samples, whose mean was 1.1%.16
It is unclear whether Kuczaj's selection criteria (or the parent's criteri
for when to begin recording) adventitiously correlated with tendencies t
overregularize. However, the instructions to parents to elicit past tens
forms from their children could have elevated overregularization rates, for
reasons we examine in the following section.

A. Discourse Patterns That Can Elevate Estimates of Overregularization Rates

There are three reasons to think that efforts to elicit past tense forms
from children might cause them to overregularize more than they would in
their spontaneous speech.
First, if children have any tendency to avoid forms in their spontaneous
speech that they are uncertain of (in particular, the past tense form), then
when the parent, rather than the child, chooses the verb, by asking a leading
question, errors may be more likely. One example might be the followin
dialogue between Abe and his mother (from his 98th sample):

Mother. And what did you choose to do?

Abe. I choosed to make cookies.

Second, by eliciting descriptions of hypothetical and past tense events

the adult may be creating a discourse situation in which the child feels more

16 Indeed, two of the children whose parents exceeded Abe's in past tense elicitations,
Allison and Nat, were considerably younger, and younger children tend to be prompted
more often in conversation than older children (see Marcus, 1992).

62

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
 MARCUS ET AL.

compelled on pragmatic grounds to mark tense than he would be otherwise.

As we discuss in Chapter VI, Abe and Kuczaj's (1977a) cross-sectional sam-
ple had unusually high levels of compliance with the requirement of English
syntax that past tense be marked in main clauses describing past or hypo-
thetical events (i.e., they seldom used the stem form of verbs in obligatory
past tense contexts), and overregularization tends to be accompanied by
such high levels of past tense marking. The following bits of dialogue-
from Abe's 34th and 144th transcripts, respectively-give a flavor of this
possibility for past and hypothetical contexts:

Father. Abe do you remember what all you did yesterday? See if
you can tell me what all you did yesterday.
Abe. We played the new games we fixed the wagon and we opened
gifts.
Father. What did you do outside?
Abe. Hide. [Played hide and seek.]
Father. You hide? you hid?
Abe. Uhhuh and I count.
Father. You counted?
Abe. Uhhuh.
Mother. Abe was it ... and he counted and came and looked for
US.

Father. What happened then?

Abe. I finded Renee.

Father. What would have happened if they couldn't have found any
water?

Abe. They gotted a hose.

Third, because of the syntax of English, questions without auxiliaries

require insertion of do and the verb in stem form. If the child's representa-
tion of the stem form is primed by its appearance in a leading question, the
stem could become unusually available for the regular inflection process
and relatively less liable to being blocked by the irregular (for evidence that
stem forms and regularly inflected forms prime each other, see Fowler,
Napps, & Feldman, 1985; and Stanners, Neiser, Hernon, & Hall, 1979).17
In contrast, spontaneous use of irregular past tense forms might involve
activation of the entire lexical entry of a verb and the feature "past tense,"
directly indexing the stored past tense form, rather than first activating the

"17 Furthermore, providing the stem may have actively suppressed retrieval of the
irregular version. Presenting an adult subject with a subset of a category of remembered
words can impede retrieval of the rest (Slamecka, 1969). We thank Endel Tulving for
pointing this out to us.

63

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
MONOGRAPHS

stem form. The effect might underlie exchanges like this one in Abe's 150th
transcript:

Father. What did you hear?

Abe. I heared something like the TV.

There is suggestive evidence that higher levels of past tense elicitation

in general may indeed increase overregularization rates. The number of
times Abe heard what did, what'd, and happened per month correlated posi-
tively with his overregularization rate for that month, r(31) = .35, p < .05.
(Across children, the overregularization rate correlates positively, although
nonsignificantly, with the number of these forms per parental utterance,
r[23] = .18, p = .20.) To see if these correlations might reflect a causal
relation, we extracted all of Abe's stem overregularizations and correct past
tense forms (excluding get, read, and the no-change verbs).18 The preceding
parental utterance was then scored manually either as a clear past tense
elicitation (e.g., Tell me what you did, What did you [verb]?, and so on) or as
some other sentence type. Overregularizations were significantly more likely
after past tense elicitations than other kinds of sentences (27% vs. 19%,
X2[1]= 9.64, p < .005).
We also tested for one of the specific mechanisms by which certain kinds
of past tense elicitations might increase overregularization: the priming of
the child's representation of the stem form by its immediately preceding
use in a parental question. Abe was significantly more likely to produce a
stem overregularization of a particular verb (compared to the correct irreg-
ular form) when his parents used the stem form of that verb in the conversa-
tional turn immediately preceding the overregularization than when they
did not use the stem form in that turn (32% vs. 21%, X2[1] = 5.60, p < .05).
To see if these discourse influences operate generally in children, we
ran similar tests on the much smaller set of overregularizations from Adam,
Eve, and Sarah. We collected all the stem overregularizations in their tran-
scripts and paired each one with the nearest correct irregular past tense
form in the transcript. All three children showed a greater overregulariza-
tion rate following past tense elicitations (9% vs. 2% for Adam, 44% vs. 15%
for Eve, and 25% vs. 18% for Sarah); the difference was significant by a
chi-square test for Eve individually and in a meta-analysis combining the
probabilities of the chi-square tests for the four children (p < .01). However,
we were unable to confirm the operation of a specific mechanism contribut-
ing to this effect that we tested for, namely, the priming of the stem by an

18 Only stem overregularizations, not past overregularizations, were extracted, be-

cause this analysis was conducted simultaneously with the one described in the following
paragraph, which requires stem overregularizations.

64

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
 MARCUS ET AL.

immediate prior occurrence in the parent's prompt. Tests of this effect were
inconclusive: Adam's and Eve's samples contain more overregularizations
of a verb after the parent had just used its stem; Sarah's contained fewer.
In sum, there are reasons to believe that the higher overregularization
rates among the children Kuczaj studied, including Abe, are partly arti-
factual. The data are, as Kuczaj noted, not "representative samples of nor-
mal conversation"; rather, they appear to be part way between naturalistic
speech and experimentally elicited speech. Our analyses of discourse effects
on overregularization, together with our review of data from actual elicita-
tion experiments in the next section, suggest that Kuczaj's methodology
might have led to systematic overestimates of children's spontaneous rate
of overregularization.
Note that we are not suggesting that the entire difference between
Kuczaj's subjects and those in the rest of the literature is artifactual; in
Chapter VI, we will examine the possible effects of high levels of marking
tense on overregularization, a difference between Kuczaj's subjects and the
other children that is of some theoretical interest. Note as well that these
analyses are an attempt to resolve the discrepancy between Kuczaj's data
and the others in the literature and should not be interpreted as a criticism
of Kuczaj's methodology in general. Three of the four hypotheses he was
testing in Kuczaj (1977a) had nothing to do with overall overregularization
rates; rather, they compared the two different kinds of overregularization
errors (eated and ated) and the relative overregularization rates of verbs with
different properties. For these purposes there were good reasons to have
sought samples with large numbers of past tense forms, just as we focus on
Abe in our investigation of the time course of individual verbs in Section
IV above.19 It is only when treating Kuczaj's data as estimates of children's
overall overregularization rates in spontaneous speech that the nonrepre-
sentativeness of the samples must be taken into account.

19 Kuczaj's tests of his fourth hypothesis, however, may have been affected by his
methodology, although of course he was not aware of the effects we now point out. Kuczaj
disagreed with Brown's (1973) claim that irregular past tense marking was acquired before
regular past tense marking, which Brown had based on differences in the relative ages at
which the two kinds of marking were supplied in obligatory past tense contexts more than
90% of the time. Kuczaj noted that the finding might depend on whether overregular-
izations had been counted as incorrect forms or ignored when the percentages were tallied,
which had not been made clear in Brown (1973). (In fact, they were ignored; Brown et
al., 1971.) Kuczaj counted the overregularizations twice: as incorrect irregular past tense
forms and as correct regular past tense forms. The resulting percentages of correct use
in obligatory contexts showed an advantage to regular marking, contrary to Brown's claim.
Thus, Kuczaj's "nonreplication" was partly just due to differences in definitions, but it was
also due to how these differences interact with his special recording circumstances. Be-
cause these circumstances may have led to unusually high overregularization rates, Kuc-
zaj's estimates of irregular marking were thereby depressed and his estimates of regular
marking inflated.

65

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
MONOGRAPHS

VIII. PREVIOUS ESTIMATES OF OVERREGULARIZATION FROM ELICITED

PRODUCTION EXPERIMENTS

Several experimental studies have elicited past tense forms in sentenc

completion tasks using existing irregular English verbs. For example, in the
experiment by Kuczaj (1978), the overregularization rates for the group
of 3-4-year-olds, 5-6-year-olds, and 7-8-year-olds were, respectively, 29%
49% (42% stem overregularizations, 7% past+ ed overregularizations), an
1%. Bybee and Slobin (1982a) found that their third graders (8-6-10-1)
overregularized between 2% and 55% of the time, depending on the ver
subclass. Marchman (1988) found the following overregularization rates f
her different age groups: 4-year-olds, 32%; 5-year-olds, 33%; 6-year-old
22%; 7-year-olds, 10%; and 9-year-olds, 5% (calculated from her table
based on 76% of the test items being irregular, as mentioned in her text
Note that in all such studies the overregularization rates are generally le
than 50%; once again, there is virtually no evidence that young children
overregularize exclusively or in free variation with correct irregular forms.
Of course the overregularization rates in elicited production tasks ar
still far higher than those obtained from spontaneous speech, but the tw
kinds of estimates are not comparable. Bybee and Slobin (1982a), Prasada
Pinker, and Snyder (1990), and Stemberger and MacWhinney (1986) foun
that adults, when put under time pressure, are prone to making overreg
larization errors at even higher rates than children (from 6% to 31% of the
time in the Bybee-Slobin study, depending on the subclass), presumab
because of a greater likelihood of retrieval failure. It is plausible that many
children feel that they are under pressure in experiments even if it is n
explicitly stated. Furthermore, if children ever fall into a strategy of treating
each experimental item as a pure sound, rather than as a word they kno
it essentially becomes a novel form, and regularization is the most accessible
option.
But most important, in all such tasks children are being supplied with
the stem itself seconds before they are asked to supply the past form (e.g.,
This is a girl who knows how to swing. She did the same thing yesterday. She ).
This contrasts with naturalistic settings in which children produce a past
form for an irregular in response to a mental representation of the verb's
meaning plus the feature for past tense; the phonetic form of the stem need
never be activated. Thus, experimental elicitations of irregular past tense
forms using the stem as a prompt, like parents' leading questions containing
the stem, are likely to prime the child's representations of the stem form
and possibly suppress the irregular past, leading to an increased likelihood
of overregularization, an effect we were able to document for Abe in the
preceding section.

66

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
 MARCUS ET AL.

IX. PREVIOUS ESTIMATES OF OVERREGULARIZATION FROM JUDGMENT

AND CORRECTION EXPERIMENTS

Another source of data that might be thought to show that childr

are indifferent to the past tense forms of their irregular verbs comes from
Kuczaj's (1978) judgment task. In one experiment, children aged 3-9 yea
had to judge whether any member of a group of puppets "said someth
silly." One puppet produced a sentence with a correct irregular past ten
a second produced an overregularization, and, for verbs other than n
change verbs, a third produced a past+ ed form. In a second experimen
children of the same ages produced past tenses for irregular verbs supp
in the future tense (discussed in the preceding section), then judged a p
pet's version of the verb (always different from the child's version), a
then judged a second puppet's version (the third possible kind of past te
form). Finally, children were offered a forced choice among the three ver-
sions and asked which of the three they thought their mother would u
In many conditions, overregularizations were judged as acceptable a lar
proportion of the time, as high as 89% for stem overregularizations for the
youngest children in the first experiment.
However, here too the data are not comparable to overregularizatio
rates from spontaneous speech. Grammaticality judgment is a signal det
tion task, and it is fallacious to assume that every time a child accepts
fails to correct a given form, the child's grammar deems it well form
Rather, just as with all yes-no data, the perceived payoffs for hits, misses,
false alarms, and correct rejections affect rates of saying yes. For child
in an experimental setting this could involve a variety of demand character
istics such as the perceived politeness of rejecting or correcting anothe
creature's language more than a given proportion of the time. In the la
guage of signal detection theory, this defines a "criterion" or bias for sayin
yes that is superimposed on their "sensitivity" in internally representi
grammatical and ungrammatical utterances as different, which we can
sume is a probabilistic process. Lacking direct manipulations of bias, t
best one can do in determining whether children have knowledge of irregu-
lar pasts is to compare their yes rate for correct irregulars to that for inco
rect overregularizations; if the former are higher, children must be discrim
nating between them (see Grimshaw & Rosen, 1990). Kuczaj's data provi
15 opportunities to make such comparisons: three age groups in experiment
1, each of which was asked to judge overregularizations of no-change irreg-
ulars and to judge overregularizations of other irregulars; and three a
groups in experiment 2, each of which was asked tojudge stem overregular-
izations, to judge past overregularizations, and to choose their favorite from
among the three. (The children's choice of their mother's favorite form was

67

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
MONOGRAPHS

almost always identical to their own choice and thus is not an independent
data set.) Of these 15 comparisons, only one involved a failure to discrimi-
nate irregulars from overregularizations: the middle age group (5-7) in
experiment 2 preferred past overregularizations over irregulars or stem
overregularizations in the forced choice task (although they did not even
produce many such forms in the elicited production task, as noted above).
In other words, the judgment data confirm that children systematically fa-
vor irregular forms as the preferred past tense version of irregular verbs
(see also Lachter & Bever, 1988).
More recently, Cox (1989, p. 204) told children that a puppet "was
learning to talk but was having trouble with some of his words," and the
child was asked to help him say the correct words. Twelve sentences, each
with an overregularized noun or verb, were provided. Children were not
asked to judge the sentences, and there were no correct irregulars among
the experimental stimuli, so we cannot assess children's discrimination abili-
ties from the data. Correction performance was surprisingly poor: none of
the six sentences with verbs was corrected by more than 16% of the children
around the age of 5, and none of the six sentences with nouns was corrected
by more than 28%, except, inexplicably, tooths. Since, as Cox herself notes,
the children who failed to correct an overregularization did not necessarily
use it themselves, a response bias against correcting the puppet too often is
a likelier explanation than an absence of knowledge, especially since she
told the children that the puppet was having trouble with only "some" of
his words but presented no sentences that were actually correct.
Finally, the consistent findings that overregularization rates are low and
that irregulars are preferred to them help explain the otherwise paradoxical
phenomenon that children who have been observed to overregularize will
vehemently correct their parents when they mimic the children's errors
(Bever, 1975; Slobin, 1978; see also Lachter & Bever, 1988). Similarly, Ervin
and Miller (1963) noted that their subjects often corrected their own over-
regularizations; we do not know of any reports of children correcting one
of their irregular past tense utterances to an overregularization.

X. SUMMARY

Overregularization percentages in the single digits are characteristic

most children, ages, and commonly used verbs. This suggests that, dur
the ages at which they are overregularizing, children are neither failing
discriminate irregular verbs from regulars nor freely alternating betw
overregularizations of irregulars and the corresponding correct past ten
forms but instead show a pervasive strong bias for the correct form. S

68

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
 MARCUS ET AL.

a bias is predicted by the hypothesis that children block regularization when-

ever they retrieve an irregular form from memory (MacWhinney, 1978;
Pinker, 1984) but is not predicted by any other hypothesis in the literature.
Finally, what has been called "U-shaped development" corresponds to the
following two events: before the first overregularization, there is a measur-
able extended period during which all irregular past tense forms are cor-
rect, and overregularization tails off gradually during the school-age years.

69

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
V. THE RELATION OF OVERREGULARIZATION TO CHANGES IN
THE NUMBER AND PROPORTION OF REGULAR VERBS IN
PARENTS' SPEECH AND CHILDREN'S VOCABULARY

In this chapter, we review the operation of the Rumelhart-McCle

model and how it models children's U-shaped developmental sequen
Then we examine the empirical assumptions justifying the sequence
puts that Rumelhart and McClelland trained the model on. We test whet
these assumptions are reasonable using the available data from the sp
neous speech of Adam, Eve, and Sarah, the three children we have
focusing on who displayed a U-shaped sequence in their longitudinal
opment.

I. WHY THE RUMELHART-McCLELLAND MODEL OVERREGULARIZES

The core of the Rumelhart-McClelland model is a pattern associator

network that takes a phonological representation of the stem as input and
computes a phonological representation of the past tense form as output.
The pattern associator consists of two layers of nodes-a set of input units
that are turned on in patterns that represent the sound of the verb stem
and a set of output units that are turned on in patterns that represent the
sound of the verb's past tense form-and weighted connections between
every input unit and every output unit. Each unit corresponds to a sequence
of phonological features, such as a high vowel between two stop consonants
or a back vowel followed by a nasal consonant at the end of a word. The
word itself is represented solely by the set of feature sequences it contains.
When a set of input nodes is activated, each node sends its activation level,
multiplied by the link weight, to the output nodes it is connected to. Each
output node sums its weighted inputs, compares the result to a threshold,
and probabilistically turns on if the threshold is exceeded. The output form
is the word most compatible with the set of activated output nodes.
During a learning phase, the network compares its own version of the

70

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
 MARCUS ET AL.

past tense form with the correct version provided by a "teacher," and it
adjusts the strengths of the connections and the thresholds so as to reduce
the difference between the actual state of each output node and the correct
state. By this process of recording and superimposing contingencies be-
tween bits of sounds of stems (e.g., the distinctive features of endings such
as -op or -ing) and bits of sounds of past tense forms (e.g., the features of
-opped and -ang), the model improves its performance over time, and it can
generalize to new forms on the basis of their featural overlap with old ones.
The model contains nothing corresponding to a word or rule and thus
makes no qualitative distinction between regular and irregular mappings;
both are effected by connections between stem sounds and past sounds.
Rumelhart and McClelland's explanation of the sequence of overregu-
larization flowed from the ways in which parallel distributed processing
(PDP) models generalize. In building such a model, there are numerous
ways to bias it toward conservative recording of individual input items,
toward liberal overgeneralization according to frequent patterns, or some
combination. The challenge was to duplicate the child's transition from
conservatism to overgeneralization in a single model. Rumelhart and
McClelland proposed a simple and ingenious hypothesis. Not only are irreg-
ular verbs high in frequency, but the reverse is true as well: the verbs highest
in frequency are irregular. For example, the top 10 verbs in Kucera and
Francis's (1967) frequency list are all irregular. If children acquire verbs
in order of decreasing frequency, they will develop a vocabulary with an
increasing proportion of regular verbs as they begin to run out of the high-
frequency irregulars and encounter more and more regular verbs. In partic-
ular, Rumelhart and McClelland assumed that, at some point in develop-
ment, the child shows "explosive" vocabulary growth, which would result in
a sudden influx of a large number of regular verbs. Because the regular
pattern will be exemplified by many different verbs, the learning procedure
will strengthen many links between stem features and the features defining
the -ed ending. The effects of these newly modified link weights could over-
whelm the existing weights on the links between idiosyncratic features of
irregular stems and the idiosyncratic features of their pasts, resulting in
overregularization. As the irregulars continue to be processed, the discrep-
ancies between the overregularized and teacher-supplied correct forms will
be registered, and the crucial idiosyncratic links will be strengthened over
time, eventually allowing the irregular forms to reappear.
Given these assumptions, Rumelhart and McClelland were able to
model the developmental sequence with one additional assumption: the
vocabulary explosion occurs after the child has just acquired his 10th verb.
Their 10-verb decision results in two training phases. First, the model is
presented with the 10 highest-frequency verbs (excluding do and be, which
can also be auxiliaries), of which only two (20%) happen to be regular,

71

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
MONOGRAPHS

10 times apiece. Then the model is presented with that list plus the 41
next-most-frequent verbs, constituting a set in which 80% of the verbs are
now regular, 190 times apiece. In phase 1, the model learned the 10 verb
successfully; when phase 2 begins on the eleventh cycle and it is suddenl
swamped with regulars, the model overregularizes the irregulars. The re
covery process begins immediately, reaching asymptote shortly before th
200th epoch (see Figs. 16 and 17, taken from Rumelhart & McClelland,
1986).
The Rumelhart-McClelland model challenges the traditional account
of overregularization, which depended on separate rote and rule mecha
nisms, in favor of a single mechanism that begins to overregularize because
of an influx of newly acquired regular verbs, a presumed consequence of
vocabulary growth spurt. Let us call this explanation of the cause of overreg-
ularization the vocabulary balance hypothesis; it is also a feature of the more
recent network simulations by Plunkett and Marchman (1990).

II. PINKER AND PRINCE'S CRITIQUE

Pinker and Prince (1988; see also Prince & Pinker, 1988) examined
Rumelhart and McClelland's assumptions about development. Rumelhart
and McClelland cited Brown (1973) in support of their assumption of a
vocabulary spurt near the onset of overregularization, but Brown did no
discuss vocabulary acquisition at all. According to standard sources (see,
e.g., Ingram, 1989) children's "word spurt" usually occurs at 1-6, about
year too early to account for the onset of overregularization for most chil-
dren, which occurred at a mean age of 2-5 for the seven children examined
in Chapter IV. Pinker and Prince examined Brown's (n.d.) vocabulary lists
for Adam, Eve, and Sarah, which were drawn from five evenly spaced sam-
ples spanning the overregularization sequence, plus a fourth child in the
one-word stage. They found neither an explosive growth in vocabulary near
the onset of overregularization nor, more significant, an increase in the
percentage of the child's vocabulary samples that was regular: the propor
tion regular stayed around 50% before, during, and after the onset. Pinke
and Prince also cited data (partly from Slobin, 1971) suggesting that th
proportion of regular verb tokens among all verb tokens in parental speech
to children is about 20%-30% during overregularization, nowhere near
the 80% proportion that Rumelhart and McClelland used to override th
irregular patterns. They argued that an endogenous transition from rote
to rule is still required to account for the data, as in the traditional account.
Pinker and Prince's (1988) data showing that the proportion of regular
verb types in children's longitudinal samples' stays at around 50% seem
paradoxical at first: if there are only 180 irregular verbs and thousands o

72

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
0
1.0

Q)
Regular

0.9 -
r12

0.8 Irregular

Q)
0.7

0.6

0.5
0 40 80 120 160 200

Trials
FIG. 16.-Performance of the Rumelhart-McClelland model on regular an
verbs as a function of training epochs. The dip in the curve for irregular ver
the 10th epoch corresponds to the onset of overregularization. From Rum
McClelland (1986).

tO

1.0

0.8

Qk

0.6
C.

0.4

0.2

0.0 11 I 1
0 40 80 120 160 200
Trials
FIG. 17.-Tendency of the Rumelhart-McClelland model to overregularize
verbs as a function of training epochs. Overregularization tendency is measu
ratio of the strength of the correct irregular response to the sum of the streng
correct and the overregularized responses. Points below the line correspond to
to overregularize. From Rumelhart and McClelland (1986).

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
MONOGRAPHS

regulars, isn't an increase in the percentage of regular verbs a mathematic

certainty after the 180th irregular is acquired, and a statistical near certa
well before that? The answer is that Pinker and Prince's type estimates w
from fairly small samples (about 700 utterances per child per stage)
hence were not pure estimates of type frequency but something combinin
type and token frequency: types with higher token frequency were m
likely to have been sampled. Because the token frequency of irregula
much higher than that of most of the regulars, it is possible that wh
children learn lower-frequency regular verbs, they may not displace
earlier acquired irregulars. Permit, understand, remember, misbehave, and s
may compete among themselves for air time in children's speech, lea
general-duty verbs like come, go, take, put, eat, and so on to occupy a cons
proportion of verb slots in conversation throughout development.
Because Pinker and Prince's data reflected both type and token f
quencies, the force of their critique is uncertain. To evaluate the vocabular
balance hypothesis, then, one must first establish whether it is the propo
tion of regular types, tokens, or some other index that is relevant. T
issues must be addressed. What is the psychological event that corresp
to an episode of network learning, according to Rumelhart and McC
land's theory? And what kinds of changes in the schedule of learning
sodes cause overregularization in pattern associator networks?

III. WHAT IS A LEARNING EPISODE?

Rumelhart and McClelland make the following assumptions about

real-world events that correspond to a learning episode:

The [simulation] run was intended to capture approximately th

experience with past tenses of a young child picking up English fr
everyday conversation. Our conception of the nature of this experienc
is simply that the child learns first about the present and past tenses
the highest frequency verbs; later on, learning occurs for a much larg
ensemble of verbs, including a much larger proportion of regula
forms.

Although the child would be hearing present and past tenses of all
kinds of verbs throughout development, we assume that he or she is
only able to learn past tenses for verbs already mastered fairly well in
the present tense. This is because the real learning environment does
not, in fact, present the child with present-tense/past-tense pairs.
Rather, it presents the child with past-tense words in sentences oc-
curring in real-world context. The child would therefore have to gener-
ate the appropriate present tense form internally with the aid of the
entire sentence and context, and this, we suppose, requires that the

74

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
 MARCUS ET AL.

child already know the present tense of the word. [Rumelhart &
McClelland, 1987, p. 222]

The assumption here is that an episode of learning consists of hearing

a past tense form, using the context to recover its corresponding stem from
the mental lexicon,20 feeding the stem into the internal pattern associator,
comparing the output with the past tense form actually heard, and adjusting
the weights in response to discrepancies. A stem-past pair would be fed into
the model only when an adult used the past and the child possessed the
stem in his or her vocabulary (and knew it was related to the past form).
This means that the proportion of regulars fed into the past tense
learning system would be determined by the proportion of occasions that
the parent used a regular past tense that the child already possessed in stem
form. However, it is impossible to tell from transcripts exactly when this
conjunction of parent's use and child's knowledge occurs. Instead, there are
three ways to estimate the relevant proportions indirectly, each with differ-
ent assumptions.
For all three, it is useful to assume that all verbs have an approximately
constant distribution of uses in different tenses, so we can collapse across
tenses and increase sample sizes, reducing the danger of underestimating
the number of regular verbs. Counting all verbs, not just past tense forms,
also reduces the size of a possible confound: if one were to find a positive
correlation between the increase in the number of regular past tense forms
and the tendency to overregularize, it could reflect the effects of a newly
acquired regularization process on the ease of generating past tense forms
of regular verbs rather than the effects of having many regular verbs on
the development of an ability to regularize.21 In fact, we have found that
irregular verbs take up a somewhat larger proportion of past tense tokens
than of total verb tokens (about 85% vs. 65%-75%), but this difference only
strengthens the conclusions that we will be making on the basis of all verb
tokens.

20 Rumelhart and McClelland refer to the input form as the "present," but the present
tense form would include an irrelevant -s affix for the third person singular; "stem" is
actually what they had in mind.
21 Lachter and Bever (1988) point out that the direction of a cause and effect relation
between possessing regular verbs and overregularization cannot completely be resolved,
even if one were to examine possession of regular verbs per se rather than mastery of the
past tense form of such verbs. The problem, they note, is that once a child has developed
the ability to generate and analyze regular past tense forms productively, he or she no
longer has to memorize such forms from parental speech. As a result, the learning events
and memory space that beforehand would have been dedicated to acquiring regular past
tense forms can now be applied to the learning of brand new verbs, and such verbs should
be acquired at a more rapid rate. We will not pursue this possibility further, although it
would be an important consideration if one were to find the relevant correlation.

75

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
MONOGRAPHS

The first method assumes that children produce regular and irregular
verbs in approximately the same proportions that they process regular an
irregular past tense forms in their parents' speech (i.e., it assumes that if a
child uses a verb, he or she knows it, and that children in conversation with
parents will use different verbs in roughly the same proportion as their
parents). Under this assumption, the proportion of verb tokens that are regular
in the child's speech indirectly estimates the proportion of regular learning
episodes.
Second, since the occurrence of a parental token is necessary for a
learning episode to take place, if we assume that children know a constant
proportion of the verb tokens their parents address to them, we can mea-
sure the proportion of verb tokens that are regular in the parent's speech.
In practice, Rumelhart and McClelland ignored token frequency en-
tirely in assembling the training set for their model: every verb was fed in
the same number of times, once per epoch. This assumes that a third mea-
sure, the proportion of regular verb types among all verb types, is the
relevant factor-although it is inconsistent with their psychological interpre-
tation of a learning episode, which would be driven by parental tokens.
Rather, the teaching schedule they actually modeled is more consistent with
some kind of off-line learning, fed by a preprocessor: the child takes a pass
through his entire verb lexicon, feeding each stem-past pair into the pattern
associator once per scan. If we entertain this interpretation of a learning
episode, which corresponds literally to Rumelhart and McClelland's learn-
ing schedule, rather than the token-driven interpretation they discuss, we
can test the vocabulary balance hypothesis by trying to estimate the proportion
of verb types in the child's vocabulary that are regular. Note that it is not valid to
use the percentage of a child's vocabulary that is regular as a surrogate for
the number of on-line learning episodes even if it turns out that the propor-
tion of regular verbs among parental tokens is constant. That is, one cannot
assume that the proportion of regular learning episodes is determined by
the proportion of the child's vocabulary that is regular, because this larger
regular vocabulary could correspond to a larger number of regular types
that the parent is cycling through a constant number of regular tokens
in his or her speech, leaving the proportion of regular learning episodes
constant.

IV. HOW DO CHANGES IN LEARNING EPISODES LEAD TO

OVERREGULARIZATION?

Assuming that we know what a learning episode is, what kind of

changes in the distribution of learning episodes lead to overregularization,
according to the vocabulary balance hypothesis?

76

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
 MARCUS ET AL.

A. Type versus Token Frequency

First, it is clear that both type and token frequency have important
consequences. The Rumelhart-McClelland model overregularized because
it changed its connection strengths with each input pair in a direction that
reduced the discrepancy between computed and input past forms. After
the first epoch, in which the model was suddenly bombarded with regulars,
about 80% of the changes that the model made were designed to make it
more likely to generate regular forms, because 80% of the inputs were new
regulars. Many of the changed connections involved links from phonologi-
cal features that were also shared with irregulars (since most irregulars are
phonologically similar in some way to at least some regulars). Because the
network did not have enough specific feature units to register each verb on
its own set of units, the overlap was high enough that each irregular was
represented by many units whose links had just been adjusted to help pro-
duce the regular ending, and overregularization resulted. This effect would
obviously depend strongly on the number of regular types, because the
wider the range of regular forms that are fed in, the greater the probability
that a given phonological feature of an irregular verb will be shared by
some regular verb and hence develop stronger links to the incorrect regular
pattern. But this effect can also be mitigated by token frequency: if, say,
each irregular had been repeated four times for each regular (reflecting
the real-world higher token frequency of irregulars), the links that joined
features unique to the irregulars to their corresponding irregular past forms
would have been strengthened several times during the epoch to reduce
the errors with such forms, and overregularization would be less likely.
Indeed, one of the noteworthy properties of the Rumelhart-McClelland
model is its distributed phonological representation of words, with no units
dedicated to words per se (see Pinker & Prince, 1988), so there is no physical
basis for a distinction between types and tokens in the model at all. Only
feature-to-feature mappings, whether they be from a single word or a set
of similar words, are represented. The actual behavior of the model will
depend on the number of regular types, the phonological range of the
regular types, their degree of overlap with irregulars, the token frequencies
of both irregulars and regulars, and other factors. In any case, it is clear
that overregularization of an irregular does depend on the ratio of regular
to irregular tokens and hence is relevant to testing the vocabulary balance
hypothesis.

B. Proportion versus Changes in Proportion

Second, the percentage of regular learning episodes at a given time is

not the relevant factor in predicting overregularization. In the Rumelhart-

77

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
MONOGRAPHS

McClelland model, unlike in children, the process of recovery from overreg-

ularization begins immediately after its onset (see Fig. 17), correct irregular
forms predominate within a few epochs, and at asymptote they are pro-
duced most of the time, all with a constant level of 80% regular learning
episodes. This is an obvious property of any model that is designed to
perform correctly at asymptote: even with the most unfavorable proportion
of regular episodes, the irregulars must eventually reassert themselves.
Overregularization is a short-term consequence of the increase in the per-
centage of regular episodes with development. Although a properly de-
signed model could learn to overcome any particular level of dominance of
regulars, this adjustment cannot take place instantaneously, and influxes of
regulars will cause temporary overregularization, before the crucial links
between nodes unique to an irregular and its idiosyncratic past have been
sufficiently strengthened. For this reason, the difference between the rapid
recovery of the Rumelhart-McClelland model and the protracted period of
overregularization of children does not speak against the model. It is possi-
ble that as children learn more and more words, new regulars are constantly
washing over them; no sooner do they adjust their irregulars to the leveling
effect of one wave of regulars than a new wave comes in. Thus, the proper
test of the vocabulary balance hypothesis involves a correlation between the
most recent increase in regular learning episodes and the current rate of
overregularization; this is why Rumelhart and McClelland appeal to a pe-
riod of explosive growth in vocabulary to trigger overregularization.22

C. Proportion of Regular Verbs versus Number of Regular Verbs

Third, for the analyses in which types are being examined, it may not
be the proportion of verb types that are regular that is the relevant pre-
dictor. The problem is that the competition in pattern associators is not
between the regular pattern and a single irregular pattern shared among
all the irregulars. Rather, irregulars are different from each other, not just
from the regulars. Imagine that at one stage there are six different irregu-
lars, each with a different change (e.g., go-went, come-came, hit-hit, etc.), and
six regulars. At the next stage, there are 12 different irregulars, each with
a different change, and 12 regulars. The proportion of regulars in the
sample remains the same, but the ratio of regulars to any particular irregular

22 In recent experiments on the behavior of connectionist networks at learning inflec-

tional mappings in sets of artificial verbs with different training schedules and vocabulary
mixtures, Plunkett and Marchman (1990, 1991) have confirmed that both token frequen-
cies and rate of vocabulary increase have direct effects on the tendency of standard con-
nectionist models to produce outputs analogous to overregularizations.

78

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
 MARCUS ET AL.

pattern has doubled. Therefore, in this scenario it is the number of regulars,

not the proportion of regulars, that would predict overregularization.
The preceding scenario is not fully accurate, however: the ratio of
learning episodes for the regular pattern to a given irregular pattern would
be identical to the number of regulars only if every irregular were totally
idiosyncratic. But virtually all the irregulars share their patterns of change
with other irregulars, so the calculation is too extreme. Consider a scenario
that is extreme in the other direction: the six irregulars in phase 1 fall into
three classes (e.g., sing-sang, ring-rang,feed-fed, breed-bred, wear-wore, tear-tore),
and the new irregulars in phase 2 fall into the same classes (e.g., spring-
sprang, lead-led, swear-swore). Here, the ratio of regulars to any irregular
vowel change pattern is 3: 1 in both phases, and we would expect overregu-
larization to be less likely; only a change in the proportion of regulars would
clearly induce it.
In reality, the situation is likely to be somewhere between these ex-
tremes because while English irregulars do fall into a restricted number of
kinds of change, we would expect the number of patterns in a child's vocab-
ulary, not just the number of irregulars per pattern, to increase somewhat
with development. Therefore, it is not clear whether overregularization
rates should be correlated with the proportion of total types that are regular
(appropriate if all new irregulars fall into old patterns and hence protect
old irregulars) or with the number of types that are regular (appropriate if
each new irregular is unique), and we will examine both correlations.

D. Summary of Tests

We will examine whether there are increases over time in the propor-
tion of verbs that are regular verbs among the child's tokens, the parents'
tokens, and the child's types and, if there are, whether such increases are
related to the child's tendency to overregularize. The relations will be tested
at two levels of temporal detail.
First, we will compare the monthly rate of increase in each vocabulary
factor for the months before the first overregularization with the monthly
rate of increase for the months during which overregularization is taking
place (i.e., the first month containing an overregularization and all the
months after it). This is an objectively specifiable dividing line but requires
some further comments. For Eve, there are a small number of tokens before
the first overregularization, and recall that there is no statistical evidence
that she in fact underwent a transition between stages (see Chap. IV, Sec.
IVA). For Sarah, the dividing line could be questioned. Brown (n.d.; see
Pinker & Prince, 1988) noted that Sarah produced a past tense of hear at
2-10 that was literally pronounced as an overregularization (heared, with a

79

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
MONOGRAPHS

schwa instead of the r), but he worried whether it could have been an od
pronunciation of heard. The first completely unambiguous overregulariza
tion does not occur in her transcripts for another 5 months. However, th
hypothetical pattern of distortion that Brown considered is not indepen
dently motivated by Sarah's other mispronunciations: for example, at 3-
she pronounced hurted as hahted, suggesting that if heard was intended but
mispronounced, it would have surfaced as hahd, not hea-ad (Alan Prince
personal communication, January 10, 1989). Furthermore, excluding thi
datum has the effect of weakening the evidence for the vocabulary balan
hypothesis. Therefore, we will not second-guess the transcription, and w
will count heared as Sarah's first overregularization. In any case, the firs
recorded overregularizations of all three children should be taken, not a
the literal moment of arrival of the ability to overregularize, but as a partly
arbitrary dichotomization of the developmental span into a period when
overregularizations are likely to occur and a preceding span when overregu-
larizations are unlikely to occur.
A more precise test comes from correlating the monthly rate of increase
in a vocabulary measure between month t and month t + 1 with the overreg-
ularization rate at time t + 1. This analysis combines the factors that differ-
entiate the pre-overregularization stage from the overregularization stag
and the factors that cause overregularization to be more frequent in on
month than another during the overregularization stage. According to th
Rumelhart-McClelland model, there is no qualitative difference between the
two. Because longitudinal transcripts from a given child are not indepen
dent sampling units, the correlation coefficients are best treated as descrip-
tive statistics; the meaning of tests of significance is not clear. Fortunately,
the issue will seldom arise, as the sign of the correlation coefficients we find
will usually obviate the need for significance testing.

V. PARENTAL TOKENS

The proportion of regular verb tokens among all adult verb tokens
plotted for Adam, Eve, and Sarah in Figures 18, 19, and 20. As mention
in Chapter III, the data come not only from the child's parents but al
from the other adults speaking to the child in the transcripts. For all three
children, a bit more than one-quarter of the parental verb tokens we
regular, and this did not change over the course of development. As me
tioned, the reason the proportion is constant is that most of the high
frequency verbs that are indispensable for casual conversation are irregular,
and they do not move aside to make way for the more numerous but lower-
frequency regulars.
The proportions are similar before and after overregularization begin

80

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
 MARCUS ET AL.

for Adam, 30% before, 29% during; for Eve, 29% before, 25% during; for
Sarah, 25% before, 26% during. More important, the rate of change in
the proportions was not systematically larger during the overregularization
period: for Adam, 1.2 percentage points per month before, - 0.1 during;
for Eve, 0.9 before, - 0.7 during; for Sarah, 0.2 before and during. The
correlations between the rate of monthly change in proportion regular and
the child's overregularization rate are in the wrong direction for Adam
(-.40) and Eve (-.29) and close to zero (.03) for Sarah.

VI. CHILD'S TOKENS

The proportion of the child's tokens that consists of regular verb

plotted on the same axes as their adults' proportions, also in Figures 1
The proportion oscillates between 26% and 45% for Adam, actually dec
for Eve from around as high as 54% (possibly sampling error) to a st
state between 23% and 30%, and shows some early dips to the teens
Sarah before mainly oscillating within the 20%-40% range. The reas
for regular verbs being used a steady minority of the time are no doubt t
same as for the adults.

The proportion of tokens that are regular does not systematically in

crease after overregularization begins: 37% versus 34% for Adam; 50%
versus 27% for Eve; and 25% versus 30% for Sarah. The monthly rates o
change for these figures are actually higher before overregularization th
during: respectively, 2.4 and - 0.3 percentage points for Adam; 7.9 and
- 3.8 for Eve; and 4.2 and - 0.6 for Sarah. The correlations between
monthly change in percent tokens regular and the child's overregularization
rate are -.03 for Adam, -.22 for Eve, -.10 for Sarah.

VII. CHILD'S TYPES

As mentioned, measures based on the child's types are the least psycho-
logically realistic measures to focus on, because they are meaningful only i
rule learning is an off-line pass through the child's vocabulary, which
not correspond to Rumelhart and McClelland's psychological assumptio
But the measures are worth examining both because they represent
form of the vocabulary balance hypothesis that is most likely to be consist
with some developmental trend (since the proportion of regular types mus
increase with development) and because it was in fact a change in regu
types that defined the training sequence given to the Rumelhart-McClellan
model.

81

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
 Adam
c 100
& 90
a 80

c 70
60
I--
50
5O
040

a)
S20 Adam's Adults
10
eO

S 20 2-6 3-0 3-6 4-0 4-6 5-0

Age
FIG. 18.-Percentage of verb tokens that are regular for Adam and the adults convers-
ing with Adam.

Eve
I, 100
o 90-
- 80-
C 70

0 60
1-
50-
40

0j 301 Eve
320
2 Eve's Adults
2 10
eO
0. 0 1-6
0 2-0
Age
FIG. 19.-Percentage of verb tokens that are regular for Eve and the adults conversing
with Eve.

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
 MARCUS ET AL.

Sarah
cc 100
" 90
a 80-

) 70
a)

S60
50
Sarah
S40-

30

r20-
2 10
CL O
200 2-6 3-0 3-6 4.0 4-6 5-0
Age
FIG. 20.-Percentage of verb tokens that are regular for Sarah and the adults convers-
ing with Sarah.

Measuring the proportion of children's vocabulary that consists of regu-

lar verbs is an extremely difficult problem, for it faces the notorious pitfalls
involved in estimating children's vocabulary size in general (for extensive
discussion, see, e.g., Lorge & Chall, 1963; Miller, 1977; Moe et al., 1982;
Seashore & Eckerson, 1940; Templin, 1957). The source of the problem
here is that we are confined to the actual words that children used in sam-
ples. Obviously, the child will use only a small fraction of his or her total
vocabulary in any given sample. Since high-frequency verbs are more likely
to appear than low-frequency verbs, the number of low-frequency verbs will
be systematically underestimated. And since there are more low-frequency
regulars than low-frequency irregulars, counting types per sample will sys-
tematically underestimate the proportion of regulars (this was a problem
with Pinker and Prince's 1988 estimates). There is no completely adequate
solution to the problem of measuring children's vocabulary, but there are
various estimates that can be examined, and at the very least the direction
of changes in the proportion that is regular can be compared for months
associated with different levels of overregularization, and the resulting con-
clusions can be compared across different methods in an attempt to arrive
at converging conclusions.

83

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
MONOGRAPHS

A. Method 1: Cumulative Vocabulary

One measure that is designed to be generous to low-frequency forms

is the child's cumulative vocabulary totals. That is, one assumes that the child
never forgets. If a word is used in a given month, it is credited to the child's
vocabulary from then on. Figures 21-26 show the children's cumulative
vocabulary growth for regular and irregular verbs, and the proportion of
cumulative vocabulary that is regular; overregularization rates are included
in these graphs for ease of comparison of their developmental courses.
Table 5 shows the rate of change of the number and proportion of regular
verbs in the vocabularies of Adam, Eve, and Sarah before and during the
period marked by the first overregularization; the static vocabulary figures
for the month before the first overregularization and the last month of the
transcripts are also provided as reference points. By mathematical necessity
each child possesses a larger cumulative regular vocabulary later in develop-
ment than earlier, and, as expected, the regulars take up a larger proportion
of the child's total verb vocabulary later. However, as the decelerating vocab-
ulary curves (most visible for Adam) suggest, for both the number of regular
verbs and the proportion of verb vocabulary that is regular, the rates of
increase are much larger for the stages before than during overregulariza-
tion, contrary to the vocabulary balance hypothesis. Similarly, the monthly
changes in the number of regular verbs and in the proportion of regular
verbs among all verbs correlate negatively with overregularization rate for
all three children.

Unfortunately, the decelerating vocabulary curve and concomitant neg-

ative correlation with overregularization may be a direct consequence, even
an artifact, of the cumulative measure for vocabulary. Cumulative vocabu-
lary is equivalent to sampling without replacement. Imagine that parents
use 500 regular verbs, with equal token frequencies, when speaking to their
children, at all ages. Imagine that every month children attend to and ac-
quire 10% of the verbs they hear and produce every word they have ac-
quired at least once. At the end of month 1 their cumulative vocabulary is
50 words. At the end of month 2 it is not 100 words but only 95-the 50
they learned in month 1, plus 10% of the 450 words that they had not
previously acquired in month 1 (i.e., 45 words; the other five they attended
to do not count because they had already been acquired). In month 3 their
cumulative vocabulary will be 136 words, reflecting the addition of only 40
new words (10% of the 405 remaining), and in month 4 it will be 172 words
(36 new ones). In other words, new words will be acquired at a faster clip
early in development than later, even with a constant learning rate and a
constant number of words in the environment. If there is indeed a relatively
limited set of regular verbs for the child to acquire during the preschool

84

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
 MARCUS ET AL.

years, this is one reason why the vocabulary balance hypothesis could be
false.

However, these sampling considerations could also mean that the decel-
erating growth curve is an artifact. Imagine that the child possesses a con-
stant 500 words throughout development, but manages to produce only 50
during a month's worth of samples. Following the same arithmetic as de-
scribed above, the fact that cumulative vocabulary is a form of sampling
without replacement means that we as investigators will spuriously credit
the child with having "acquired" fewer and fewer new words with each
succeeding month. It is very difficult to tell to what extent the curves in
Figures 21-25 represent a genuine sampling effect in vocabulary learning,
a sampling artifact in measuring vocabulary from production data, or both.
Thus, it is important to supplement these direct estimates with some indirect
measure that is free of this possible bias.

B. Method 2: Jackknife Estimates Using "Mark-Recapture" Patterns

There is a family of techniques commonly used in biology and demog-

raphy for estimating population sizes from multiple samples. The simplest
version is commonly known as "mark-recapture" (see Seber, 1986). Here is
an idealized example of how one might estimate the number of squirrels in
a forest. Trap 50 squirrels, paint their tails orange, release them, allow
enough time for them to diffuse through the forest, trap 50 squirrels again,
and see how many have orange tails. If there are 10 such recaptures, then
the first trapping session must have represented 10/50 or one-fifth of the
total population. Since 50 were trapped initially, the forest population must
be 250.

This logic can be applied to vocabulary estimation as follows. A verb is

a squirrel, a transcript is a trapping session, and a verb that appears in two
successive transcripts has been recaptured. The proportion of verbs at t2
that also appeared at t1, multiplied by the number of verb types recorded
at t1, is an estimate of the vocabulary size. Note that this procedure avoids
the possibly artifactual deceleration in vocabulary acquisition inherent in
cumulative measures. If a child had a static vocabulary of 500 words, 50 of
which were recorded in each sample, then the second sample would consist
of five of the words that appeared in the first sample (1/10 x 50) and 45
new words (1/10 x 450), and the recapture rate of 1/10 (5/50), multiplied
by the first sample size (50), would yield the correct figure of 500. This
would be true of every pair of successive samples. Note as well that, if new
verbs are acquired between t1 and t2, the estimate will be an unbiased esti-
mate of the vocabulary at t2. Imagine an idealized case in which 100 baby
squirrels were born between capture and recapture. Orange-tailed squirrels

85

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
 Adam
300

250

25 Cumulative Regular Verbs

$ 200

S 150

E 100-
z
z
50 50- Cumulative Irregular Verbs

2-0 2-6 3-0 3-6 4-0 4-6 5-0

Age
FIG. 21.-Adam's cumulative regular and irregular verb vocabulary

Adam
100

90 Overregularization
80
70
c 60 Cumulative Vocabulary Regular
5 50
"- 40
30
20
10

2-0 2-6 3-0 3-6 4-0 4-6 5-0

Age
FIG. 22.-Proportion of Adam's cumulative verb vocabulary that is regular and his
overregularization rate (subtracted from 100%).

86

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
 Eve
300

250
()
,.0

, 200
o
o 150- Cumulative Regular Verbs
E 100

50Cumulative Irregular Verbs

01 t
1-6 2-0
Age
FIG. 23.-Eve's cumulative regular and irregular verb vocabulary

Eve
100

90

80 Overregularization
S70
c 60

O
,) 50 - Cumulative Vocabula
. 40
30-
20
10

0 1-6 ' ' 2-0 '

Age
FIG. 24.-Proportion of Eve's cumulative verb vocabulary that is regular and her
overregularization rate (subtracted from 100%).

87

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
 Sarah
300

250
')
Cumulative Regular Verbs
4) 200
.0

150
U)

E 100
z
Cumulative Irregular Verbs
50

2-0 2-6 3-0 3-6 4-0 4-6 5-0

Age
FIG. 25.-Sarah's cumulative regular and irregular verb vocabulary

Sarah
100

90- Overregularization
80

4)
70

60 Cumulative Vocabulary Regular

8
0) 50
Ue
4)
4) 40

30

20
10

0
2-0 2-6 3-0 3-6 4-0 4-6 5-0
Age
FIG. 26.-Proportion of Sarah's cumulative verb vocabulary that is regular and her
overregularization rate (subtracted from 100%).

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
 MARCUS ET AL.

TABLE 5

GROWTH OF CUMULATIVE REGULAR AND IRREGULAR VOCABULARY

Adam Eve Sarah

Average monthly rate of increase

before first overregularization:
Number of regular verbs .......... 11.6 12.0 10.2
Proportion of verbs regular ........ 2.4 3.6 2.9
Level at month before first
overregularization:
Number of regular verbs .......... 117 34 68
Number of irregular verbs......... 68 28 53
Total ........................... 185 62 121

Percentage regular ............... 63 55 56

Average monthly rate of increase
during overregularization:
Number of regular verbs .......... 6.6 11.3 7.2
Proportion of verbs regular ........ .5 1.1 .6
Level at end of transcripts:
Number of regular verbs .......... 283 124 269
Number of irregular verbs......... 99 70 99
Total ........................... 382 194 368

Percentage regular ............... 74 64 73

are now recaptured with probability 50/350, so the rec

1/7. Seven times the capture sample size of 50 is 350,
size at recapture.
Unfortunately, unequal token frequencies for diff
systematic underestimates if one were to use this pro
Imagine that some squirrels are more trap shy than
case, 40% of the squirrels might be "shy," where "shy
probability of blundering into a trap (1/5, in our exa
One expects the first sample to capture 30 of the 15
only 10 of the 100 shy squirrels. Sixteen percent (40/
in the forest now have orange tails, but the second trapp
capture only seven of them: six bold ones (1/5 x 30) an
x 10). Since the second sample consists of 40 squirre
+ [1/10 x 100]), the 7/40 recapture rate, multiplied b
the first sample, yields an estimate of 229, 21 less than t
children's vocabulary, verbs with lower token frequen
and because many of them are regular, one would obta
estimates of total verb vocabulary, regular verb vocabu
of vocabulary that is regular, and the rate of increase in

89

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
MONOGRAPHS

(although comparisons of higher- versus lower-growth months might still

be roughly accurate).
Biostatisticians have dealt with the trap-shyness problem by applying
the "generalized Jackknife" estimator to the mark-recapture methodology,
first developed by Burnham and Overton (1978, 1979) and extensively in-
vestigated by Otis, Burnham, White, and Anderson (1978). Instead of two
trapping sessions, there are k of them. The numbers of squirrels that have
been captured only once, twice, three times, and so on are tallied. The
effects of unequal capturability can thus be estimated by taking into account
the distribution of multiple recaptures. While the simple capture-recapture
estimate assumes a uniform distribution of capturability, with the extra in-
formation of number of recaptures one can assume that individual capture
probability is a random variable from an arbitrary distribution. Otis et al.
(1978) found empirically that the Jackknife estimator produces accurate
estimates if many individuals are caught a relatively large number of
times-that is, if the multiple recapture rate across samples is high.
In our case, this procedure consists of comparing sets of five consecu-
tive transcripts and counting how many verbs were used in each of the five
transcripts, how many in only four, and so on. (To make the estimates for
the three children comparable, for Sarah we use five consecutive pairs of
transcripts because her speech was sampled for 1 hour once a week, whereas
Adam's and Eve's was sampled for 2 hours once every 2 weeks.) As required,
many verbs did appear in multiple transcripts from such sets. This set of
numbers is fed into the Jackknife algorithm, providing an estimate of vocab-
ulary size for that period. With nonoverlapping sets of five consecutive
transcripts, we obtain independent estimates for different ages. The decel-
eration in cumulative estimates is eliminated, as is the underestimation in-
herent in simple mark-recapture estimates.
The estimates span periods of 2/2 months rather than a single month,
which has both disadvantages and advantages. The growth estimates are
temporally coarser, and there is no vocabulary estimate at all for Eve that
corresponds exclusively to the period before her first overregularization.
However, a larger temporal window may catch effects of vocabulary growth
that act over longer time spans than the 1-month window used so far.
The estimator is not free of complications. Because the kind of context
in which the recording takes place is similar in all recording sessions, those
verbs most appropriate to those contexts will be recorded more often. (In
the ecology literature, it has also been noted that achieving equal capturabil-
ity is impossible, even with randomized capture locations on each sampling
occasion; see Chao, 1987.) Furthermore there is a free parameter that must
be decided on in calculating the estimates: the "order" of the estimate,
corresponding to the maximum number of recaptures (out of five, in our
case) that are counted in the calculations. Higher-order estimates have lower

90

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
 MARCUS ET AL.

bias but higher variance; there is a complex procedure for selecting the
optimal order for a given estimate. For simplicity's sake, we will uniformly
report estimates of order 4. We have found these generally to be the highest
of the estimates of different order, especially for regular verbs; hence, they
are fairest to the vocabulary balance hypothesis. But in any case we also
found that the growth curves for different order estimates are almost per-
fectly parallel, so the correlations we calculate are not notably affected by
this choice. With these considerations in mind we can cautiously compare
periods of high regular vocabulary growth with periods of low regular vo-
cabulary growth, even if the magnitudes of particular increases and totals
are not to be taken as perfectly accurate estimates.
Figures 27-32 show that the vocabulary estimates obtained from this
method are higher for young children and somewhat lower for older chil-
dren than the cumulative estimates, eliminating the severe deceleration that
was inherently unfavorable to the regular vocabulary balance hypothesis.
They also do not display the unlikely constant 50/50 regular-irregular ratio
that would correspond to interpreting Pinker and Prince's figures as esti-
mates of types.
Overregularization rates are included in the graphs for ease of compar-
ison of their developmental courses. Table 6 shows the rate of growth of
estimated vocabulary per interval during the period preceding the first
overregularization and the period beginning with it; numbers and propor-
tions of regular verbs at the end of these two intervals are shown as well.
The number of irregular types shows a very small increase with time; the
number of regulars shows a larger one. As Table 6 shows, for Adam, the
rate of increase in the proportion of verbs that are regular is larger during
the overregularization stage than before it (1.1 and - 0.2 percentage points,
respectively, per five-sample interval). This, however, is the only comparison
from among all those we have performed that is in a direction consistent
with the vocabulary balance hypothesis. For Adam, the rate of increase in
the proportion of verbs that are regular from one interval to the next shows
no correlation with the overregularization rate at the end of the interval
(r = .004); there was virtually no difference in the number of regular verbs
acquired per interval during the overregularization stage versus before it
(9.7 and 9.1 new regular verbs per interval, respectively); and the sample
correlation coefficient for the relation between the size of the increase be-
tween intervals and the overregularization rate for the second interval is
negative (r = -.08). Sarah acquired regular verbs at a faster rate before
her overregularization period than during it (42.2 and 10.2 additional verbs
per interval, respectively), and these increases correlated negatively with
their ensuing overregularization rates (r = -.26). Similarly, the proportion
of her vocabulary that was regular increased much more rapidly before
than during her overregularization stage (8.3 and 0.4 percentage points per

91

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
 Adam
300

250

E 200
4e- Q> Estimated Regular Verbs

E 100
50Estimated Irregular Verbs

50

2-0 26 3-0 3-6 4-0 4-6 5-0

Age
FIG. 27.-Adam's Jackknife-estimated regular and irregular verb vocabulary

Adam
100
90 - Overregularization
80

S70
60D
60 Estimated Vocabulary Regular

? 50
S40
30
20
10

2-0 2-6 3-0 3-6 4-0 4-6 5-0

Age
FIG. 28.-Proportion of Adam's Jackknife-estimated verb vocabulary that is regular
and his overregularization rate (subtracted from 100%).

92

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
 Eve
300

250

. 200
Estimated Regular Verbs
150

E 100

0-
1-6 2-0
Age
FIG. 29.-Eve's Jackknife-estimated regular and irregular verb vocabulary

Eve
100

90
80~ Overregularization
S70
C1
S 60
S50 SEstimated Vocabulary Regular
S40
30
20
10

1-6 2-0
Age
FIG. 30.-Proportion of Eve's Jackknife-estimated verb vocabulary that is regular and
her overregularization rate (subtracted from 100%).

93

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
 Sarah
300

250 Estimated Regular Verbs

3 200

o 150
Estimated Irregular Verbs
"E 100
z

50

02-0 2-6 3-0 3-6 4-0 4-6 5-0

Age
FIG. 31.-Sarah's Jackknife-estimated regular and irregular verb vocabulary

Sarah
100

90- Overregularization
80

G 70
60-
50 Estimated Vocabulary Regular
L- 40
l30
20
10

0
2-0 2-6 3-0 3-6 4-0 46 5-:0
Age
FIG. 32.-Proportion of Sarah's Jackknife-estimated verb vocabulary that is regular
and her overregularization rate (subtracted from 100%).

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
 MARCUS ET AL.

TABLE 6

GROWTH OF REGULAR AND IRREGULAR VOCABULARY,

JACKKNIFE ESTIMATES

Adam Eve Sarah

Average rate of increase before first

overregularization:
Number of regular verbs ........ 9.1 a 42.2
Proportion of verbs regular ...... -.2 a 8.3
Level at month before first
overregularization:
Number of regular verbs ........ 136 a 124
Number of irregular verbs ....... 82 a 69
Total ......................... 218 a 193

Percentage regular.............. 62 a 64
Average rate of increase during
overregularization:
Number of regular verbs ........ 9.7 9.5 7.8
Proportion regular. ............. 1.1 .6 .4
Level at end of transcripts:
Number of regular verbs ........ 213 106 209
Number of irregular verbs ....... 87 78 96
Total ......................... 300 184 305

Percentage regular. ............. 71 58 69

a Eve had too few samples before her first overregularization to yield Jackknif

interval, respectively) and correlated negatively (-.17) w

tion rate. For Eve, the temporal coarseness of the Jackk
vents a before-and-after comparison, but both the inc
regular verbs and the increase in proportion of verb
regular correlate negatively with overregularization ra
relevant interval.

C. Comparison to Previous Estimates of Vocabulary Size and Composition

With all the hazards of vocabulary estimates it would be reassuring to

compare ours with previous estimates in the literature that report type and
token counts from much larger samples of children's speech.
E. Horn (1925) examined the vocabularies of 80 children from 1 to 6,
plus 270,000 tokens from kindergartners and first graders. The combined
lists yielded 5,000 words, from which he selected 1,084 that recurred a
certain minimum number of times across samples. Of these, 233 are un-
ambiguously verbs, 81 irregular and 152 regular, or 65% of types regular.
M. D. Horn (1927) amassed 489,555 tokens of the speech of kindergarten

95

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
MONOGRAPHS

children, a set composed of 7,097 types. She reports the 1,003 most fre-
quently used words. The list contains 97 regular verbs and 74 irregular
verbs, or 57% of types regular. These figures are somewhat smaller than
the cumulative and Jackknife proportions that we find for Adam and Sarah
at the end of their transcripts (which range from 69% to 74% regular), but
that is to be expected because both Horns excluded lower-frequency words
which in turn systematically underestimates the number of regular verbs
(When we look at single samples of Adam's and Sarah's speech near the
end of their transcripts, which introduces a similar bias, we ourselves get
lower figures, near 58%.) However, M. D. Horn provided data for th
proportion of verb tokens that are regular, which should be unaffected by
this sampling bias. The token figures are similar to our estimates: 24,581
regular tokens and 80,370 irregular tokens, or 23% regular.
Even better estimates come from Moe et al.'s (1982) figures for first-
grade children (mean age 6-9) because they report all verb tokens, not just
the most frequent ones. Their lists include 418 regular verb types and 108
irregular types, corresponding to 79% of verb types being regular and 33%
of verb tokens. These figures can be compared with the estimates shown in
Tables 5 and 6 for Adam and Sarah at the end of their transcripts: Adam
(5-2), 74% cumulative types regular, 71% Jackknife-estimated types regu-
lar, 33% tokens regular; Sarah (5-1), 73% cumulative types regular, 69%
Jackknife-estimated types regular, 20% tokens regular.

D. Comparison to the Analyses of Marchman and Bates

Before we summarize the conclusions of this chapter, we briefly exam-

ine contrary claims made in a recent paper by Marchman and Bates (1991)
They analyze data on children's vocabulary size and overregularizations,
concluding, "Verb vocabulary size is highly predictive of... the onset of
overregularization errors" (p. 7). In particular, they interpret their data as
support for a "critical mass hypothesis" according to which the "vocabulary
size is related to ... the subsequent onset of overregularization errors when
verb vocabularies become sufficiently large" (p. 6). They take this as evi-
dence supporting the connectionist models of Plunkett and Marchman
(1990, 1991), which also, they suggest, show critical mass effects. In this
section, we attempt to resolve the discrepancy between these conclusions
and ours.

Marchman and Bates used different methods of vocabulary estim

from ours, different measures of overregularization, and different predi
tions about the relevant correlations. We examine each in turn before dis-
cussing their actual findings.
Method of estimation of vocabulary size and overregularizations. -The vocab-

96

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
 MARCUS ET AL.

ulary estimation technique employed a checklist given to 1,130 mothers of

children between the ages of 1-4 and 2-6. Each mother filled out the check-
list once, so there are no longitudinal data. One section of the checklist
contained 680 word stems, including 46 irregular verb stems and 57 regular
verb stems. The other consisted of 21 irregular past tense forms and 30
overregularized versions of these verbs (stem overregularizations of all
verbs, plus past overregularizations of some of them).23
As Marchman and Bates point out, it is desirable to use a variety of
methods of vocabulary acquisition, given the limitations inherent to any
single technique such as the ones we have noted and dealt with in this
chapter. The main advantage they see for parental checklists is that they
provide "more accurate estimates of vocabulary size" (1991, p. 3), presum-
ably owing to the potentially larger corpus of speech drawn from (namely,
however much of a child's speech a parent can remember). Unfortunately,
this putative advantage is vitiated by the fact that Marchman and Bates
provided only a small subset of English irregular verbs on their list. For
example, Adam at 2-6 and Eve at 2-3 had already produced within their
transcripts (a small subset of their total speech) more regular verbs, and
more irregular verbs, than Marchman and Bates included on their checklists
(see Figs. 21 and 23), and this was true for Sarah by the time of her first
overregularization (see Fig. 25). Thus, the verbs that Marchman and Bates's
parents had the opportunity to check off constitute a subset of the verbs
that we know (from our transcript data) that children of the relevant ages
use. Therefore, Marchman and Bates's vocabulary figures, contrary to their
claims, are not more accurate than our transcript counts; we know that they
must suffer from a greater degree of underestimation.
A more fundamental objection, however, is that the checklist data per-
tain exclusively to types-Marchman and Bates estimate how many verb
types were overregularized at least once but have no data on the overregu-
larization rate either within or across verbs. As we shall see when we return
to their empirical claims, this is a crucial limitation, confounding their main
conclusion.

The logic of the prediction.-Marchman and Bates, citing the simulations

of Plunkett and Marchman (1990, 1991), predict that overregularizati
should be triggered by the attainment of some absolute size of verb vocabu
lary. As a result, they tested the correlation between vocabulary size and th
number of verbs overregularized. This contrasts with our tests, which t
for correlations between the recent change in vocabulary size (i.e., the r
of acquisition of new vocabulary) and overregularization.
Examination of the Plunkett and Marchman papers, however, rais

"23 These figures correspond to the number of verbs actually listed in their append
their text mentions 20 irregular past tense forms and 31 overregularized forms.

97

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
MONOGRAPHS

questions about the basis for Marchman and Bates's prediction. First
Plunkett and Marchman repeatedly conclude that the rate of vocabulary
expansion (the factor that we, but not Marchman and Bates, correlated
with overregularizations) is a critical variable in their models' tendency to
overregularize (e.g., Plunkett & Marchman, 1990, pp. 12-13, 18). Indeed,
they note that their model can attain 100% correct performance on any
absolute level of vocabulary size (p. 11). Second, when Plunkett and
Marchman do discuss a putative effect of a critical mass in vocabulary, the
effect in question is "recovery from erroneous performance" (p. 11), exactly
the opposite of the decrement for irregular verbs that Marchman and Bates
are trying to explain. (Later, they discuss generalization to new stems, but
this is still different from overregularization of previously learned ones.)
Third, Plunkett and Marchman's claim for a critical mass effect in their
model (i.e., recovery after vocabulary size reaches 50, constant across differ-
ent rates of vocabulary increase) bears no obvious relation to the simulation
data they present. They cite no numbers, instead inviting the reader to
inspect their figures 1 and 2, but neither figure shows any discontinuity in
performance for irregular verbs at the point corresponding to the 50th verb
acquired. Moreover, their figure 1 shows clear differences in the recovery
profile depending on the rate of vocabulary increase.
The empirical claims.-Marchman and Bates report two relevant empiri-
cal findings. The first is that overregularizations are rarely reported for
children with reported verb vocabularies less than 15-30 words. They take
this as evidence for a critical mass effect, but the reasoning is unsound. If
children's vocabulary increases with age and they begin to overregularize at
a certain point after vocabulary acquisition begins, by logical necessity any
finite data set will yield numbers corresponding to the average and mini-
mum sizes of their vocabulary at the time they begin overregularizing. In
particular, the minimum will simply correspond to the vocabulary size of
the most precocious overregularizer in the sample. This does not constitute
evidence for a critical mass effect, that is, that across variation in other
factors such as rate and composition of vocabulary, a given child will begin
to overregularize when he or she attains a specific vocabulary size.
The second finding is that vocabulary size correlates well with overregu-
larizations, holding linear effects of chronological age constant in a multiple
regression. But this finding, too, is close to being a statistical necessity, not
an empirical discovery. Recall that Marchman and Bates's data on overregu-
larization consist of the number of irregular types that are overregularized
at least once; they have no data on the overregularization rate (the probabil-
ity that the child will use an overregularization as compared to the correct
irregular past). Clearly, a child cannot overregularize a verb that is not in
his or her vocabulary. If the child has a vocabulary of five irregular verbs,
he or she cannot be recorded as having overregularized six verbs. If the

98

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
 MARCUS ET AL.

child has a vocabulary of 50 irregular verbs, he or she can, in principle,

overregularize 50 verbs but no more, and so on. Thus, even if children
overregularize at the same rate regardless of vocabulary size, the number
of overregularized types recorded will almost certainly be larger for chil-
dren with larger vocabulary sizes simply because larger vocabularies yield
more logical opportunities to overregularize types. Thus, Marchman and
Bates's correlation is an artifact that provides no support for the claim that
larger vocabularies predict a greater likelihood of overregularizing.
In sum, while there may be promise in the use of parental checklists as
a corroborating method of estimating vocabulary size, they have no a priori
advantages over direct counts from spontaneous speech transcripts, and as
Marchman and Bates have deployed them, they have a number of disadvan-
tages. Moreover, Marchman and Bates's conclusion-that children, like con-
nectionist models, overregularize because they attain a critical vocabulary
size-must be rejected. First, connectionist models have not been shown to
overregularize because they attain a critical vocabulary size; rather, their
overregularization is mainly caused by high rates of input of regular verbs,
which is the motivation for our own analyses discussed in this chapter.
Second, children have not been shown to overregularize because they attain
a critical vocabulary size; the correlations that Marchman and Bates report
are artifacts.

VIII. SUMMARY

Our estimates of children's types, adults' types, and children's token

provide virtually no support for the hypothesis that overregularization
triggered by increases in the number or proportion of regular verbs av
able to the child. Regular verbs remain a roughly constant proportion
adults' and children's conversational tokens, and never dominate. Regul
types-which in any case do not correspond to on-line learning episod
necessarily increase with development, both absolutely and as a proport
of total verb vocabulary, but the sizes of these increases do not correla
positively with children's tendency to overregularize, which is what the vo-
cabulary balance hypothesis requires. It appears that something endog
nous to children's grammatical systems, and not a change in either the
environments or their vocabularies, causes overregularization errors t
begin.

99

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
 VI. THE RELATION OF OVERREGULARIZATION TO
THE DEVELOPMENT OF TENSE MARKING OF
REGULAR AND IRREGULAR VERBS

If vocabulary development does not predict the onset of overregulariza-

tion, does anything? Pinker and Prince (1988) suggested that children
lag in overregularizing might be due to the development of the regu
tense-marking process itself. If, consistent with traditional assumptions and
contrary to the assumptions behind the Rumelhart-McClelland model, ch
dren have an ability to memorize stems and pasts independently of t
ability to generate the past from the stem, and if the specifics of Engl
tense marking take time to learn, then before they are learned correct irre
ular pasts could be produced. Overregularizations would be absent becau
the child's regular past tense-marking machine is "off," not because it
"on" but starved of regular inputs.
There is nothing in this hypothesis that is incompatible with the
Rumelhart-McClelland model being an accurate model of the child's p
tense-marking machine. One would, of course, have to give up the id
that there is no rote storage outside the past tense-marking machine a
that U-shaped development can be explained entirely by the internal work-
ings of the machine as it processes a changing input mixture. The seque
would be explained by a transition from a stage where only the rote lexicon
was working to a stage at which a connectionist model began to proce
stems and pasts as inputs. MacWhinney and Leinbach (1991) endorse (b
do not implement) such a two-module system in their connectionist model o
past tense inflection. Indeed, Rumelhart and McClelland themselves brie
entertained such a possibility, and there are numerous suggestions in t
connectionist literature for how to implement the equivalent of gates
on-off switches for multimodule network models (see, e.g., Jacobs, Jordan,
& Barto, 1991). Thus, the analyses we discuss in this chapter do not spe
to how past tense marking is computed (i.e., whether it is computed
traditional symbol processing or by parallel distributed processing), as lo
as it can be computed independently of rote storage of lexical items.

100

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
 MARCUS ET AL.

To see if overregularization begins only when the productive regular

process is first activated, we need the results of wug tests administered to
children at various points before and after they overregularize. Obviously,
these data do not exist, and it is not practical to hope for them. One- and
2-year-old children make poor experimental subjects in elicited language
production tasks, especially with made-up words introduced on the spot.
Fortunately, there are data from spontaneous speech that indirectly bear
on the hypothesis.
In principle, a ruleless child with excellent rote memory could produce
regular and irregular pasts with equal facility, as long as they were available
in the input to be memorized. In practice, however, the young child pro-
duces correct irregular past tense forms more reliably in obligatory past
tense contexts than regular past tense forms (Brown, 1973; de Villiers &
de Villiers, 1973; see also n. 19 above). Presumably, irregulars are easier
to memorize and produce because of their higher token frequency, their
phonological simplicity (all irregulars have monosyllabic roots, but many
regulars do not; Pinker & Prince, 1988), their greater phonological salience
(vowel changes might be more perceptible than a word-final t or d, especially
as part of a consonant cluster), or some combination of these factors. How-
ever, once the regular process is acquired, the irregular advantage should
be nullified: any verb, whether its past tense form is remembered or not,
can be supplied with a regular past tense form at that very moment. There-
fore, the initiation of the regular process should be visible in spontaneous
speech in a reduction or elimination of the difference between the ability
to supply the past tense forms of regular and irregular verbs in obligatory
past tense contexts. One can then see whether this equalization occurs near
the onset of overregularization.
A related signature of the acquisition of a productive regular process
can be sought in the absolute rate of marking tense on regular verbs. In
English syntax, all main clauses must be marked for tense. Thus, we can
think of syntactic processes as issuing a subroutine call to morphology, de-
manding a tensed form of the particular verb to be used. English, of course,
executes this subroutine in different ways: the various irregular forms and
the regular process. The beauty of the regular process is that it can always
keep the syntax satisfied: any verb, familiar or unfamiliar, can be suffixed
to mark tense and hence can be used in a main clause (except possibly in
the unusual circumstances discussed in Chap. II, Sec. IB). However, for a
child who has not yet acquired the regular suffixation process, if the regular
past form of a verb had not been previously memorized, the desire to mark
tense on it would have to go unsatisfied. Mastery of the regular process,
then, should be visible in attainment of high absolute rates of marking the
past tense on regular verbs in obligatory past tense contexts. Again, one

101

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
MONOGRAPHS

can see whether attainment of such levels occurs near the time of the first
overregularization, and whether degree of successful regular marking cor-
relates with overregularization across time and children. The prediction
depends on the assumption that the syntactic requirement of tense marking
has itself been acquired (if children do not know that tense is obligatory,
they could leave regular verbs in past tense contexts unmarked even if they
knew how to mark them), so if children do not reliably mark the tense of
either regular verbs or irregular verbs (which do not depend on a produc-
tive process) in obligatory contexts, this prediction cannot be tested. But if
regular verbs are reliably marked, we have indirect evidence for a produc-
tive regular process; if irregular verbs are reliably marked but regular verbs
are not, we have indirect evidence that the regular process has not yet
developed. (The acquisition of obligatory marking of irregular verbs will be
discussed in a separate section.)
Note that these predictions about levels of irregular and regular mark-
ing appear to contrast with those of the Rumelhart-McClelland model under
the assumption that the model is responsible for early correct performance
with irregulars, with no separate lexical storage. During the early period of
learning, before the influx of regular verbs triggers overregularization, the
model rapidly attained successful performance with both irregular and reg-
ular verbs (80%-85% of correct features generated), with, if anything, a
slight advantage for the regulars (see their fig. 4, reproduced here as Fig.
16). Furthermore, Rumelhart and McClelland point out that their model is
not just doing the equivalent of rote memorization before it overregularizes
but is showing "substantial generalization." Newly presented regular verbs
are inflected with 75% accuracy (chance is 50%) on their very first exposure,
based on the training during the period in which irregulars are being pro-
duced correctly.

I. THE DEVELOPMENT OF REGULAR TENSE MARKING AND

OVERREGULARIZATION

In this section, we examine whether there is an association between

overregularization and the development of the productive regular affixa-
tion process. We begin with Adam, Eve, and Sarah, who show initial
overregularization-free months in their longitudinal samples, allowing tests
of whether the onset of overregularization coincides with development of
high levels of marking regular verbs in obligatory past tense contexts (high
both in absolute terms and in terms of a rate comparable to the rate of
marking the tense of irregular verbs).

102

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
 MARCUS ET AL.

A. The Brown Children

We present Cazden's data on rates of supplying correct past tense forms

of irregular and regular verbs in obligatory past tense contexts in Figures
33-38 and in Appendix Tables A1-A3.24
As Brown (1973) pointed out, curves plotting use of a morpheme in
obligatory contexts as a function of age are invariably noisy. In particular,
early points often represent a tiny number of instances, because the gram-
matical structures that allow obligatory contexts to be recognized themselves
develop with age, and this can result in severe sampling error. For example,
the early spikes where Sarah anomalously marks irregular and regular verbs
100% of the time represent two out of two and three out of three tokens,
respectively.
Nonetheless, several general qualitative features of the data are notice-
able. First, for all three children, the first month with an overregularization
displayed a high level of marking of irregular past tense: 84% for Adam,
93% for Eve, and 70% for Sarah (although the neighboring months are
higher for Sarah and much lower for Eve). Second, regular marking is low
before the first overregularization but displays rapid increases to high levels,
overlapping those seen for irregular verbs, shortly afterward. Adam's first
overregularization occurred during a 3-month period in which regular
marking increased from 0% to 100%; Eve's occurred during a 7-month
period in which she went from 0% to 95%; Sarah's occurred during a
4-month interval in which she went from 0% to 78%. (Moreover, for Adam
and Sarah there appears to be a brief decline in regular marking coinciding
with a reversion to several months of no overregularizations immediately
after the first recorded one, followed by a rise to levels close to 100% around
the time when overregularization resumes.)
Table 7 confirms that rates of regular marking are low and smaller
than those for irregular marking for the period before the first overregular-
ization, and that the difference is narrowed or eliminated for the period
beginning with the first overregularization. The increase in both kinds of
marking and the relatively larger increase for regular marking can be cap-
tured in a 2 x 2 ANOVA whose factors are regular versus irregular verbs
and before versus during overregularization and whose dependent variable

"24 The data are grouped into months to be commensurable with our overregulariza-
tion data. Sarah's recording sessions were half as long but twice as frequent as those of
Adam and Eve; to facilitate comparisons among the three children, Cazden pooled succes-
sive pairs of Sarah's samples. Sometimes these pairs straddled a boundary between months
of chronological age, the units we have been using. When this happened, we divided the
tokens in that pair of samples by two, assigning half to the month preceding the boundary,
half to the month following it.

103

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
MONOGRAPHS

TABLE 7

PROPORTION OF VERB TOKENS IN OBLIGATORY PAST TENSE CONTEXTS THAT WERE

MARKED FOR TENSE BEFORE THE FIRST MONTH WITH AN OVERREGULARIZATION
AND IN THE PERIOD BEGINNING WITH THE FIRST MONTH

IRREGULAR REGULAR

No. of No. of
Proportion Obligatory Proportion Obligatory
Marked Contexts Marked Contexts

Adam, before ........ .74* 369 .08* 123

Adam, during........ .91* 559 .73* 173
Eve, before .......... .18* 11 .11* 19
Eve, during.......... .62* 259 .66* 118
Sarah, before ........ .65* 131 .44 26
Sarah, during ........ .90* 508 .85* 134
* Proportions significantly different from .5.

is the proportion of verb tokens marked for

Verbs were marked more reliably after over
fore, F(1,2) = 199.80, p < .005, and, margina
for regular than for irregular verbs (for the in
of regular/irregular and before/during, F[ 1,2]
tion between tense marking and overregula
continuous measures: there are positive corre
the proportion of regular verbs marked in o
and the overregularization rate: r(14) = .33
r(22) = .44 for Sarah. Correlations with irregula
although smaller, for all three children: .29,
How might one examine the second predic
overregularization is associated with the acqu
past tense-marking process, namely, that c
"obligatorily" during the overregularization
"obligatory" means nothing less than 100%
children attained. Brown (1973) employed th
was being marked "obligatorily" if it was sup
contexts in six successive hours of speech. Ho
in obligatory marking for any particular m
order of acquisition of 14 heterogeneous morphemes; the reason he

"25 Sarah's correlation with regular marking would be statistically significant if one
were to treat months as independent sampling units; the other correlation coefficients
would not be. Such significance tests, however, are meaningless because monthly transcript
sets are not independent samples from a population, as noted earlier.

104

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
 MARCUS ET AL.

adopted the 90% over 6 hours criterion was because it could be used to
derive an objective rank ordering of the age of acquisition of the mor-
phemes. For our purposes, which do not involve ordering heterogeneous
morphemes, the criterion is in some ways too strict (because extended use
of a morpheme in, say, 75%-85% of obligatory contexts clearly indicates
relevant knowledge) and in others too lax (because a span with a lucky
streak of correct usages could be counted as mastery, only to return to lower
levels because of regression to the mean). Brown notes that the develop-
mental curves show such patterns in several instances.
One convenient benchmark is whether a child marks tense more often
than not in obligatory contexts (i.e., if significantly more than 50% of verb
tokens are marked). The 50% figure is, of course, partly arbitrary, but it
does literally reflect a systematic preference that tense be marked. More-
over, because children initially mark most morphemes at rates far less than
50% (most notably, the regular past tense), it is unlikely to represent the
base rate for the proportion of times a past tense is called for on pragmatic
communicative grounds, and attainment of a level of marking greater than
50% represents movement toward the adult state where tense marking is
truly obligatory.
In Table 7, marking rates that are significantly greater than .5 by a
two-tailed binomial test (computed by the approximation to the normal
distribution) are indicated with an asterisk. All three children, before they
began to overregularize, left regular verbs unmarked more often than they
marked them (Adam and Sarah marked irregular verbs more often than
they left them unmarked). After overregularization had begun, all three
children marked both irregular verbs and regular verbs more often than
they left them unmarked. In sum, for Adam and Sarah, and to a lesser
extent Eve, there is evidence that the onset of overregularization is tempo-
rally associated with attainment of high rates of marking tense on regular
verbs, comparable to rates of tensing irregular verbs.
The data from Adam, Eve, and Sarah help resolve a paradox noted
long ago in the developmental literature. Ervin (1964) remarked that for
some of her subjects there were no regular past tense forms preceding the
first overregularizations. Among the nine children with extended longitudi-
nal transcripts in our sample, we found this to be true for Naomi and April
as well: the first regular verbs marked for past tense appeared a few days
after their first overregularization. The puzzle arises because no specific
rule of morphology can be innate. Therefore the child must acquire any
particular rule on the basis of individual regularly inflected forms memo-
rized from parental speech. All children, then, should be capable of produc-
ing at least some regular past tense forms before their first overregulariza-
tion. The paradox disappears when we see that Adam, Eve, and Sarah used
regular past forms at very low rates before their first overregularization,

105

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
MONOGRAPHS

presumably because memorizing them or retrieving them from memory i

difficult (although not impossible) owing to their relatively low frequency
and salience. For some children, then, the rates can be as low as zero, or at
least low enough that no examples turn up in their transcripts. At some
point the child extracts a suffixation process capable of generating regulars
freely, apparently from a small amount of evidence if the frequency of
producing regular forms reflects the number serving as the basis for the
rule, and uses it simultaneously to inflect regular verbs and occasionally to
overregularize irregular verbs.

B. The Kuczaj Children

Let us turn to Abe. As mentioned, before-and-after comparisons are

not possible for him, because he overregularized in his first sample. How-
ever, Kuczaj (1977a) noted that in the period ending at 2-6, Abe's overregu
larization rate and his rate of marking regular and irregular verbs in obliga-
tory contexts were both lower than they were in subsequent months. Figures
39 and 40, taken from the appendices to Kuczaj (1976) (reproduced in our
App. Table A4), show the data in full.26 High rates of regular marking
(often higher than the rate of irregular marking) characterize the entire
period, starting at 76%, increasing in 4 months to 98%, and staying close
to that level thereafter. Because Abe's data, unlike those from Adam, Eve,
and Sarah, include the later phase where overregularization begins to di-
minish (the right-hand arm of the U), the correlation coefficient between
regular marking and overregularization calculated over the entire period is
low (.09). The reason is that although overregularization is lower in some o
the very early months, when regular marking is also low, overregularization
becomes low again in the latest months, when regular marking, of course
remains high. When the late reduction of overregularization is held constant
by partialing out age, the correlation between overregularization and regu-
lar marking is far higher, .44.
As Kuczaj (1977a) points out, his cross-sectional sample shows a similar
distribution to Abe's longitudinal sequence. All 14 children overregularize,
and the sample includes the late span when overregularizations decrease
the three oldest children in the sample overregularize the least. All the
children marked regular verbs in obligatory contexts at high rates (84%-
100%) except one, the child who was youngest and had the lowest MLU,

26 These percentages were taken directly from Kuczaj (1976) rather than being calcu-
lated from the transcripts or raw data tables, so in some cases they are based on slightly
different token counts for number of correct irregulars than the ones we have been using
(i.e., the data listed in the second column of App. Table A4).

106

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
 MARCUS ET AL.

and this child also had the lowest overregularization rate except for the
three oldest children. Statistically, the children in the cross-sectional sample
behave similarly to Abe's months; although the simple correlation coeffi-
cient between overregularization and regular marking is low (-.09), the
partial correlation with age held constant is high (r = .47, p < .05).
Summary of developmental patterns.--Several general relations appear to
hold consistently among the children we have examined. Before the first
overregularization, regular verbs in obligatory past tense contexts are left
unmarked more often than they are marked, and they are marked at lower
rates than irregular verbs. In the period beginning with the month of the
first overregularization, there is no consistent difference between the rates
of marking regular and irregular verbs; both are marked more often than
they are left unmarked, often at very high rates. During developmental
spans before overregularization begins to diminish, rates of regular marking
correlate positively with the rate of overregularization.
The pattern is consistent with the hypothesis that the immediate trigger
for overregularization is the acquisition of the process responsible for regu-
lar tense marking, which is independent of early use of correct irregular
forms. Once productive regular marking has been acquired, children can
now inflect the stem productively on any occasion that they try to mark the
past tense of an irregular verb but fail to retrieve its past form. They do so
for the same reason that they can now inflect regular stems at high rates,
despite the initial disadvantage in memorizability that the regulars had
faced.

Of course with these correlational data we do not have strong evidence

that a common underlying acquisition event causes the correlation between
overregularization and regular marking, as opposed to both phenomena
simply increasing with age for independent reasons. At this point the most
that can be said is that we have found consistent relations in the predicted
direction between overregularization and the level of regular marking in
obligatory contexts, both of them showing their major developmental in-
creases in the same general window, before the later diminishment of over-
regularization decouples them. The consistent correlations in the predicted
direction contrast with the results of tests of the vocabulary balance hypothe-
sis in the preceding chapter, where all the predictor measures of vocabulary
either remained constant or correlated in the wrong direction with overreg-
ularization.

Note, finally, that although children systematically mark tense by the

time they begin to overregularize, the rates vary from child to child. It
possible that some of the individual differences among children in overreg-
ularization rates are related to these differences. Recall that Abe and Kuc-
zaj's cross-sectional sample showed unusually high overregularization rates

107

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
 Adam
100

90 ' Overregularization
80

2 70
&-

S60 h
- 50 - Irregular
o 40
a. 30-
20
10

2-0 2-6 3-0 3-6 4-0 4-6 5-0

Age
FIG. 33.-Proportion of Adam's irregular verbs in obligatory past tense contexts that
were correctly marked for tense. Overregularizations do not enter into these data but are
shown in the curve at the top (subtracted from 100%).

Adam
100 -

90 i Overregularization
-80

o 60-
"c 50-
0 40
"Q Regular
. 30
20
10

20 26 2m6
2-0 30 36 3-0
40 503-6 4-0 4-6 5-0
Age
FIG. 34.--Proportion of Adam's regular verbs in obligatory past tense contexts that
were correctly marked for tense and his overregularization rate (subtracted from 100%).

108

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
 Eve
100 Overregularization
90

. 80
2 70 "
60 0 '
50 :
S40
o. 30 0 Irregular
20
10
0-
1-6 2-0
Age
FIG. 35.-Proportion of Eve's irregular verbs in obligatory past tense contexts that
were correctly marked for tense. Overregularizations do not enter into these data but are
shown in the curve at the top (subtracted from 100%).

Eve
Overregularization
100
90

S80
&e 70

o 60
E50
2 40
S30 Regular
20
10

1-6 2-0

Age
FIG. 36.-Proportion of Eve's regular verbs in obligatory past tense contexts that were
correctly marked for tense and her overregularization rate (subtracted from 100%).

109

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
 Sarah
Overregularization
100-lt
90 as

80 p ..
70O
,60 ".i Irregular
S50
o 40
& 30
20-
10

0
2-0 2-6 3-0 3-6 4-0 4-6 5-0

Age
FIG. 37.-Proportion of Sarah's irregular verbs in obligatory past tense contexts that
were correctly marked for tense. Overregularizations do not enter into these data but are
shown in the curve at the top (subtracted from 100%).

Sarah
100"
90
80 - Overregularization
0

2 70

o 60
S 50 Regular
o 40
4) 30
20
10

2-0 2-6 3-0 3-6 4-0 4-6 5-0

Age
FIG. 38.-Proportion of Sarah's regular verbs in obligatory past tense contexts that
were correctly marked for tense and her overregularization rate (subtracted from 100%).

110

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
 Abe
100 ,*?u. * CIIIII
90 .
S80
2 70

0 60- Overregularization
" 50
40 Irregular
S30
20
10
0
020 2-6 3-0 3-6 4-0 4-6 5-0
Age
FIG. 39.-Proportion of Abe's irregular verbs in obligatory past tense contexts that
were correctly marked for tense. Overregularizations do not enter into these data but are
shown in the curve at the top (subtracted from 100%).

Abe
100----- - ms
Regular
90
S80
. 70
S60 Overregularization
S50
o 40
& 30
20
10
0
020 2-6 3-0 3-6 4-0 4-6 5-0

Age
FIG. 40.-Proportion of Abe's regular verbs in obligatory past tense contexts that
were correctly marked for tense and his overregularization rate (subtracted from 100%).

111

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
MONOGRAPHS

compared to the other children we examined. Although part of the explana-

tion may lie in the nonrepresentative circumstances in which Abe's conversa-
tions were recorded (see Chap. IV, Sec. VII), this artifact seems unlikely to
account for the entire difference. Also unusual about Abe and Kuczaj's
cross-sectional sample is the high rate of tense marking of both irregular
and regular verbs after Abe's first samples: Abe marked 97% of his regula
verbs, and the 14 children marked 93.9% of theirs, compared to Adam'
72% marking of regulars during the period studied by Cazden in which
he overregularized, Eve's 66%, and Sarah's 85%.27 For example, Abe's ob
served overregularization rate is 24%. Imagine that this figure comes from
his failing to retrieve the past form of an irregular verb on 24.7 out of every
100 occasions on which he tries to mark it for tense, and then applying the
regular suffixation process to the stem on 97% of those occasions. Wha
would happen if Abe suffixed verbs only as successfully as Adam did? He
would have regularized the irregulars only 72% of the time, resulting in
17.8 overregularizations and 6.9 stem forms. (In fact, the number would b
less because Adam overregularized his nonmarked irregulars in past tense
contexts a smaller proportion of the time than he marked his regular verbs.)
We would measure his overregularization rate as 17.8/(17.8 + 75.3), o
19.1%, 5 percentage points closer to the rest of the children.28

"27 Some of the differences among the Brown children may simply reflect how old
they were when Cazden stopped tallying their marking rates and hence the proportion
of early transitional samples that contributed to these means. Note as well that part of th
difference between the Brown children and the Kuczaj children can be attributed to the
different methods that Kuczaj and Cazden used in counting correct past tense forms.
Cazden apparently counted a correct past tense form as "supplied" only if it was in on
of the obligatory contexts for the past tense that she used to identify no-marking errors
other correct pasts were ignored (hence the difference between the number of supplied
irregular forms she tallied and the number of correct irregulars we found, as summarized
in App. Tables A1-A3). Kuczaj appeared to have counted all overt past tense forms as i
their very appearance was prima facie evidence that each one was obligatory in its context
This difference in method is unlikely to be responsible for much of the difference, how-
ever, because, for most of Abe's months and Kuczaj's children, the estimates of tens
marking were 100% or close to it; therefore, subtracting some correct past tense forms
would make little or no difference.

28 Another way of replacing Abe's rate of successful regularization with Adam's wou
be to use their rates of regularizing an irregular verb when the correct irregular past fo
was not produced (i.e., overregularizations as a proportion of overregularizations p
stem forms in past tense contexts; see Chap. IV, Sec. VI) rather than their rates of suffixin
regular verbs in past tense contexts. In that calculation, the discrepancy would be reduc
almost entirely. The comparison would be suspect, however, because the estimate of Ab
rate of regularizing irregular tokens is not independent of his overall overregularizati
rate, the figure we would be adjusting.

112

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
 MARCUS ET AL.

II. THE DEVELOPMENT OF IRREGULAR TENSE MARKING

AND OVERREGULARIZATION

So far, we have examined evidence on the relation between overregu

larization and the marking of regular verbs in obligatory contexts. Th
data suggest that by the time overregularization first appears, children are
already marking irregular verbs more often than not in obligatory past tens
contexts (see Table 7 and Figs. 33-37 and 39; Abe's first month may be
counterexample). Presumably, these early relatively high rates of irregu
marking reflect partial acquisition of the syntactic requirement in Engli
that verbs be marked for tense in main clauses (see Stromswold, 1990)
together with knowledge of the semantics of the past tense itself. Given
general motive to mark past tense on irregular verbs, the regular affixation
process, once it is acquired, can serve as the means to do so when retriev
fails. Indeed, once the Brown children began to overregularize, they nev
did so at a rate higher than their rate of nonmarking of irregular verbs in
obligatory contexts prior to overregularization. That is, when overregula
izations appear, they occupy some of the space in children's past tense
conversations formerly occupied by bare stems. This is consistent with t
hypothesis that overregularization steps in when retrieval fails; if over
regularizations occurred at much higher rates than immediately precedin
nonmarking rates, they would seem to represent some radical "reorganiz
tion" or global "regression," but they do not. In Kuczaj's subjects, we d
not have estimates of the rate of irregular marking prior to the onset
overregularization, but even in the samples available we see that the rat
of nonmarking in the youngest-age samples are far higher than overreg
larization rates thereafter. Indeed, acquisition of regular affixation seem
to top off the reservoir of obligatory past tense contexts; as Kuczaj (1977
p. 593) notes, "Apparently once the child has gained stable control of th
regular past tense rule, he will not allow a generic verb form to expres
'pastness,' which eliminates errors such as go, eat and find, but results
errors like goed, eated, and finded."
Although the reliable marking of tense on irregular verbs accompanie
overregularization, it does not appear to be tightly linked in time to its onse
or level.29 In individual months, children can use irregular verbs correct
in obligatory contexts more often than not without overregularizing (e.g
Adam and Sarah) and can overregularize while leaving irregulars unmark

"29 Note that this is a modification of a conclusion stated in an earlier draft of thi
Monograph circulated as a technical report, which was based on the ages at which Brow
90% for 6 hours requirement had been met, before Cazden's full data set was ma
available to us.

113

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
MONOGRAPHS

more often than not (Abe's first month). Moreover, although the rate of
irregular marking in obligatory contexts tends to correlate over months with
overregularization rates in all children, the correlations were always lower
than the corresponding ones between overregularization and the rate of
regular marking. Nor is this entirely surprising. Although marking past
tense reliably should increase the number of overregularization errors (and
hence could correlate with the appearance of such errors for sampling rea-
sons if the total number of irregular past forms were low, which it is not
for Adam and Sarah), it does not necessarily increase the rate of overregular-
ization calculated over all irregular verbs marked for past tense. That is
because if a child neglected to mark irregular verbs for tense, we would see
both few overregularizations and few correct irregular past tense forms
(with unmarked stems taking their place), all things being equal.
Nonetheless, we can speculate on two conceivable reasons why past
tense overregularization does tend to occur in the presence of high rates of
irregular past tense marking. One is that the acquisition of the obligatory
tense requirement in English might be an impetus for the acquisition of the
regular suffixation process itself, which is a means for the tense requirement
to be satisfied for all verbs, memorized or not. (We return to this possibility
in Chap. VIII, Sec. IC, where we consider how a regular rule might be
learned.) Second, because tense is stored as part of the lexical entry of
irregular verbs, they might tend to be activated by normal lexical retrieval
processes driven by the intent to communicate certain kinds of semantic
content. In comparison, activation of the regular operation, a grammatical
process, might be invoked more often for grammatical reasons (to satisfy
the tense requirement). That is, broke automatically comes to mind when one
is thinking simultaneously of breaking and of pastness, whereas thinking of
fixing and pastness summons only the word fix, plus a call to the regular
tense-marking operation. If lexical retrieval is more automatic or reliable
than on-line application of regular marking, then once a child's language
system is driven by the grammatical requirement to mark tense, there is a
motive to apply regular marking when retrieval of an irregular fails, and
overregularizations will result. In contrast, without the requirement, tense
might be marked only when the irregular form was lexically activated for
semantic reasons. Whether these speculations can be supported by indepen-
dent data, or whether they are necessary at all, we leave as a question for
future research.

III. SUMMARY

It appears that acquisition of the regular suffixation process of English

is the proximal trigger for overregularization errors: overregularizations

114

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
 MARCUS ET AL.

first appear during a window in which the child goes from leaving most
regular verbs unmarked in past tense contexts to marking them more often
than not, and at rates comparable to those for irregular verbs. An overall
tendency to mark past tense on irregular verbs, combined with imperfect
retrieval of irregular forms from memory and with a regular process that
is capable of applying to any stem, even if closely tied to an irregular past,
sets the stage for overregularization errors.

115

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
 VII. FACTORS CAUSING DIFFERENCES
IN OVERREGULARIZATION RATES AMONG VERBS

Although overregularization rates are low in general, not all verbs ar

overregularized at the same rate, and some verbs are overregularized b
some children at some stages more often than they are produced correctly.
By examining what it is about a verb that makes it more or less likely to be
overregularized, one can test many hypotheses about the psychology o
overregularization.
In this chapter, we correlate a variety of lexical factors with overregular-
ization rates across a large set of irregular verbs. To minimize averagin
artifacts, we calculated these correlations for each of the 19 ChiLDES chil-
dren with individual transcripts that overregularized at least once. When
we report averages for these correlations and tests of the average against a
null hypothesis of zero, we first transformed each child's correlation coeffi-
cient (r) to Fisher's z; the reported mean correlations were obtained by
averaging the z scores and transforming the mean z back to a correlation
coefficient.

In addition, when suitable, we computed a single measure that aggre-

gates overregularization tendencies for verbs across all 25 children. Pooling
tokens across children is inappropriate, as the children with the largest
samples would dominate the means. However, averaging together each
child's overregularization rate for a given verb is also not appropriate be-
cause such means are in danger of being artifactually influenced by individ-
ual children with extreme values and idiosyncratic subsets of verbs. For
example, if a child who is fond of using the verb hear also happened to be
a high overregularizer across the board, then the mean overregularization
rate of hear across children would be artifactually high. This problem is
accentuated for small samples of a given verb for a given child, which can
contribute extreme high or low overregularization rates to the mean across
children. To minimize these artifacts, for each child we considered only
verbs that he or she used at least 10 times in the past tense. Furthermore
we standardized each child's set of overregularization rates for different

116

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
 MARCUS ET AL.

verbs (i.e., for each child we converted the overregularization rate for each
verb to a z score, using the mean and standard deviation of the overregulari-
zation rates for different verbs used by that child). The z scores for a given
verb were then averaged across children. These scores, representing the
relative overregularization rate for each verb in the aggregate sample, are
summarized in Appendix Table A9. For the convenience of the reader we
have also linearly rescaled these mean z scores back to proportions that can
be more intuitively interpreted as "average overregularization rates." For
each verb, we multiplied its z score by the mean across children of the
standard deviations that had been used to calculate the z scores (i.e., the
mean of the 25 standard deviations, one for each child, of the overregulari-
zation rates for that child's different verbs), added it to the mean overregu-
larization rate for that verb (including zeros from the children who never
overregularized), and added an additional constant of .02 to raise the least-
overregularized verb from a negative value to zero. Note that these figures
are primarily for ease of comparing relative overregularization rates among
verbs and do not literally correspond to a mean across children.

I. FREQUENCY

As we discussed in Chapter II, if overregularization results from a

failure to retrieve a listed past tense form, forms with greater memory
strength should be more resistant to overregularization (MacWhinney,
1978; Pinker, 1984; Slobin, 1971). The more often a parent uses a pas
tense form, the stronger the memory trace for that form should be, an
the stronger the association between it and the corresponding stem form
Thus, the adult frequency of an irregular verb in its past tense form should
be negatively correlated with its overregularization rate for children. Bybee
and Slobin (1982a) found a significant negative rank-order correlation ove
verbs between their preschool children's overregularization rates and the
frequencies of the verbs in the speech of the preschool children's caretakers.
They found a similar effect for some subclasses of verbs for the past tense
forms experimentally elicited from their third-grade subjects. MacWhinney
(1978) documents similar findings in children acquiring other languages.
We sought to replicate this effect with our larger sample of children and to
examine it within individual children, to ensure that it is not an averagin
artifact.

Three measures of adult frequency were used: (1) for each verb used
in the past tense by a child, the frequency of the past tense form of tha
verb in the speech of the adults talking to that child; (2) an aggregate
parental frequency measure computed by averaging the 19 sets of adul
frequencies, one set from each child; and (3) the past tense counts from

117

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
MONOGRAPHS

Francis and Kucera's (1982) corpus of a million words of written text. (We
include correlations with the Francis and Kucera data base because it is the
most commonly used source of frequency information in psycholinguistics;
it should predict children's behavior less well then parental frequency
counts, of course, because it is from written English addressed to adults.)
Frequencies and mean frequencies were converted to logs because the fre-
quencies could range over several orders of magnitude (especially for
Francis-Kucera figures) and because we expected that a frequency differ-
ence of 1 versus 10 would have a greater effect than a frequency difference
of 1,001 versus 1,010. For the purpose of calculating correlations with these
logs, frequencies and mean frequencies of zero were arbitrarily assigned
the value zero.

Frequency had a very clear effect. The aggregate overregularizatio

rates across 19 children significantly correlate with the aggregate parent
frequency counts, r(37) = -.37, p < .05, and nonsignificantly correla
with Francis-Kucera frequencies, r(37) = -.14, p > .10. Of the 19 children
all had overregularization rates across verbs that correlated negatively with
the log aggregate parental measure (range -.12 to -.61), 18 correlate
negatively with the log frequency of their own parents, and 16 had negative
correlations with log Francis-Kucera frequency. These sets of individu
children's correlation coefficients have means that are significantly less than
zero: mean r = -.33, t(18) = 8.89, p < .001, for the aggregate parent
frequency; mean r = -.34, t(18) = 6.80, p < .001, for each child's own
parents; mean r = -.16, t(18) = 4.45, p < .001, for Francis-Kucera fr
quency.

II. PHONOLOGICAL SIMILARITY BETWEEN STEM AND PAST

An important fact about irregular verbs is that their past tense form
although unpredictable in other regards, generally preserve most of
phonological composition of their stems. Go-went and be-was are exceptio
for the other irregulars such as come-came, feel-felt, and bring-brought, th
and stem overlap to an extent that would be uncanny if the pair cons
of two arbitrary words linked only as memorized paired associates. Pi
and Prince (1988) point out that theories of irregular morphology sh
explain this fact, but that the Rumelhart-McClelland model failed to d

A. Number of Changes from the Stem Form to the Past Form

In some theories of generative grammar (e.g., Chomsky & Halle, 1

Halle & Mohanan, 1985), irregular pasts are generated by applying to

118

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
 MARCUS ET AL.

stem one or more rules that replace a circumscribed substring of phonologi-

cal segments. It is in the very nature of rules that any segment not changed
by the rule is left untouched and hence will automatically appear in the past
tense form. This is how the similarity between the members of an irregular
pair is explained.
MacKay (1976) suggested that these rules are applied by speakers on
line when they produce irregular past forms and that each application con-
sumes a determinate portion of processing resources. When comparing the
response times for adult subjects to produce past tense forms when given
their stems, he found that verbs with "simple" vowel changes were produced
most quickly, followed by regular verbs, followed by verbs with "complex"
vowel changes, followed by verbs with both a vowel change and the -t suffix.
If MacKay's hypothesis that the psychological complexity of irregulars is
predicted by the number of rule applications in the grammatical derivation,
as proposed by irregular-rule theories, such effects may also affect chil-
dren's overregularizations. Irregulars with more changes require more rule
applications and hence may be harder to produce; when the derivation
breaks down, the regular rule steps in.
This hypothesis can be tested by correlating the number of phoneme
changes that must be executed to derive the past from the stem with the
overregularization rate for that stem. What counts as a "change" will of
course depend on one's theory of possible phonological operations, but a
reasonable if crude first approximation would be to count each single vowel
substitution, consonant substitution, consonant addition, or consonant dele-
tion as one change. For example, in see-saw, one phoneme, the vowel, must
be replaced; in sweep-swept, a vowel must be replaced and a consonant suf-
fixed, for a total of two. Hence, on average, a verb like sweep should be
overregularized at a higher rate than a verb like see. In these calculations,
we treated each diphthong as a single phoneme.
We did not find a consistent positive correlation between number of
phonemes changed and overregularization rate: for the aggregate rates
across the 19 children, r(34) = -.10.30 For 14 of the 19 children, correla-
tions were also negative, as was the mean of the correlations (-.08).

B. Degree of Phonological Overlap between Stem and Past Forms

It is possible that there really are no irregular rules, and that the com-
monalities between stem and past (a heterogeneous set of patterns that

30 For Adam, Eve, and Sarah, no-change verbs were included; have, be, and do were
excluded because it is not clear how irregular-rule theories would treat them (all three
have irregular present tense forms, and be has two past forms, was and were, neither
related to the infinitive).

119

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
MONOGRAPHS

range from simple vowel changes to such severe distortions as bring-brought)

are to be accounted for by other means. For example, these pairs may have
been generated in earlier stages of the language by genuine rules, now
defunct; the pairs that were produced by these rules, because they shared
phonological material, were easy for learners to memorize and have prefer-
entially survived over the centuries in Darwinian fashion (see Bybee &
Slobin, 1982a, 1982b; Lieber, 1980; Pinker & Prince, 1988).
One extreme mechanism by which stem-past similarity might operate
is by affecting the likelihood with which children recognize that a stem form
and its past tense counterpart are alternative versions of the same verb, as
opposed to two independent verbs. In fact, it has been proposed that failure
to unite two inflected forms within a single inflectional paradigm for a given
verb is an important general cause of overregularization, accounting for
why it takes place despite the blocking principle (Bybee & Slobin, 1982a;
Clark, 1987, 1990; Kuczaj, 1981; Pinker, 1984; Pinker & Prince, 1988; see
also n. 5 above). Two phenomena suggested this hypothesis. First, Kuczaj
(1981) and Pinker and Prince (1988) noted that children productively inflect
irregular stems: ate, ated, and ating coexist with eat, eated, and eating, as if
they were two verbs. Second, Bybee and Slobin (1982a) noted that irregular
verbs that end in a vowel that changes from stem to past (fly-flew, see-saw,
know-knew, blow-blew, etc.) are particularly prone to being overregularized.
If phonological overlap between stem and past is a critical cue for two forms
to be lumped together as versions of the same verb, the meager common
portion among verbs in this class (e.g., initial s- for see and saw) would make
it harder for the child to recognize that they were forms of a single verb.
The hypothesis that a failure to unify two stems within a common verb
paradigm is the major cause of overregularization appears to be too strong.
Empirically, errors like ating and thoughting are uncommon, and tend to
appear later in development than simple overregularizations (Kuczaj,
1977a, 1981). Theoretically, the cues that would tell the child that the two
forms were versions of the same verb are present throughout development,
leaving it a puzzle that the child takes so long to notice them. Moreover,
the linguistically valid cues for common membership militate strongly
against the child's ever considering the past stem to be an independent verb.
Kuczaj (1981) notes that the child correctly treats the past stem as indicating
pastness: was wenting occurs; is wenting does not. But according to cross-
linguistic research in lexical semantics, tense is an extremely unnatural,
perhaps nonexistent, inherent semantic component of verbs (Bybee, 1985;
Pinker, 1989; Talmy, 1985). Languages do not like independent verbs that
mean "do X" and "do X in the past." If children's hypotheses mesh with
what is linguistically possible, they should not posit such verbs.
Weaker versions of the hypothesis are plausible, however. Perhaps pho-
nological dissimilarity between stem and past does not have so large an

120

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
 MARCUS ET AL.

effect that it prevents children from mentally linking them over the long
term, accounting for the very existence of overregularization, but does have
a smaller effect over the short term, causing slight delays in mentally unify-
ing past and stem for verbs with dissimilar pasts compared to verbs with
similar pasts. Moreover, if links between stem and past are not immediately
acquired at full strength, stem-past similarity could have an influence on
children's ability to retrieve the past form after some kind of link had been
formed. This could come about if on some occasions the child retrieves the
past only after first mentally activating the stem. Common phonological
material between stem and past might improve the speed or probability
with which a past form is activated, if, for example, both forms are linked
via pointers to a common representation of their phonological constituents,
or if they actually overlap in their underlying phonological representations
at some level. That is, it might be easier to retrieve strung given string than
brought given bring.
If phonological similarity has an effect on overregularization, it would
be the degree of phonological overlap, not the degree of phonological
change, that should be related to overregularization proneness. The degree
of phonological overlap for irregular pairs was quantified as follows. We
began by counting the number of phonemes preserved from stem to past.
For example, forget-forgot has five phonemes preserved, whereas catch-caught
has only one. Counting shared phonemes alone would be highly misleading,
however, because speakers certainly represent words in a format that is
more structured than a simple list of segments. In particular, the
consonant-vowel (CV) skeleton underlying a form is thought to be a distinct
level of representation in phonology (see Kaye, 1990), and this would sa-
liently capture the relatedness of see and saw or throw and threw even if the
sheer number of shared segments was slight-the lack of a consonant at
the end of a verb is itself a feature of similarity. Thus, we also counted the
segments that changed content but preserved their positions in the verb's
CV skeleton (with half the weight of phonemes that preserved both their
content and their skeleton position), so that see and saw were counted as
more similar than say and said. Similarly, consonants that preserved their
content but changed position in the CV skeleton were given a quarter as
much weight as consonants that preserved both content and position, so
that I was given less weight in feel-felt than in steal-stole. (As it turns out,
nothing hinges on this particular way of measuring stem-past similarity;
similar results occur if all phonemes are weighted equally or if shared pho-
nological features, rather than shared phonemes, are counted.)
The prediction is that the amount of preserved phonological structure
should correlate negatively with overregularization rates; that is, the more
similar the stem and past, the less likely the child is to overregularize the
verb. At first it might appear that the prediction has already been tested in

121

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
MONOGRAPHS

the preceding section: every phoneme that is changed is a phoneme that

not preserved, so the number of phonemes changed (previously tested)
should correlate negatively with the amount of overlapping phonologica
structure (to be tested here). However, the correlation is only -.55, leavin
enough independent variance to test for a negative correlation with overreg-
ularization rates even though we just saw that the phonemes-changed me
sure had no effect. In fact, the correlation is positive for the aggregat
overregularization rate (.10), for the mean correlation for these children
(.07), and for 14 of the 19 children individually.
One might note that the irregular with the least phonological materia
preserved, go-went, is high in frequency. This suggests that any difficulty i
learning more dissimilar past pairs might be compensated for by the higher
frequency of such items, masking an effect of dissimilarity per se. Indeed,
the correlation between aggregate parental frequency and phonological
overlap is -.31 (p < .005). An effect of phonological overlap could b
unmasked in a multiple regression analysis using frequency and phonolog
cal overlap as predictors. Such an analysis fails to reveal an effect of phono-
logical overlap: the partial correlation, holding constant the child's own
parental frequency, is in the nonpredicted direction for the aggregate mea-
sure over children and for 10 of the 19 individual children. (Similar none
fects are found when frequency is partialed out of the correlation betwee
overregularization and the number of phonemes changed, which is not
surprising because the correlation between the two predictors is only -.05
In sum, we have failed to find clear evidence that the degree of stem-pas
similarity or dissimilarity affects the overregularization rate of irregula
verbs.

The failure of any measure of relative stem-past similarity to explain

differing overregularization rates is unexpected given the pervasiveness of
such similarity in the English irregular verb system. Perhaps a modicum of
phonological similarity, together with the obvious semantic similarity be-
tween stem and past, may be sufficient for children to recognize the relat-
edness of most verbs and their irregular pasts. Thereafter, irregular pasts
are retrieved directly in response to a call to the entire lexical paradigm and
the "past" feature, bypassing the stem form itself (as was suggested in Chap.
IV, Secs. VIIA and VIII).

III. PROTECTION BY FAMILIES OF SIMILAR IRREGULAR PAIRS

Not only are irregular verb stems similar to their past tense forms, but
they also tend to be similar to other irregular verb stems that have compara-
ble past tense forms. Irregular verbs fall into clusters such as sting-stung,
swing-swung, string-strung, and so on. The minor rules posited by some theo-

122

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
 MARCUS ET AL.

ries of generative morphology (e.g., a rule changing i to A; see Halle &

Mohanan, 1985) are meant to explain this second kind of similarity as well.
The form of irregulars is by definition unpredictable on phonological
grounds, so the rules must be tagged as applicable only to a fixed list of
words, but if the number of rules is smaller than the number of words, the
existence of similarity clusters is explained.
However, Pinker and Prince (1988) pointed out problems for such the-
ories. The irregular clusters are held together by far more common features
than just the segment changed by the putative rule: string, sting, and swing
share not only an i, but also a velar nasal as their final consonant and an s
as part of an initial consonant cluster. Trying to capture these hypersimilarities
by adding them to the rule as context terms (e.g., "Change i to A in the
context C- ng") fails in both directions. It falsely includes many forms
like bring-brought and sing-sang, and it fails to include verbs that are clearly
related to the cluster by family resemblance, such as stick (final consonant
velar but nonnasal) and spin (final consonant nasal but nonvelar; see also
Bybee & Slobin, 1982a).
In some ways, the Rumelhart-McClelland model handles these imper-
fect partial similarities well: after being trained on 82 irregulars, some of
the model's outputs for new irregulars that it had not previously encoun-
tered were correct, such as wept, clung, and bid, despite the complex and
highly probabilistic nature of the patterns that such generalizations repre-
sent. Furthermore, the model proved to be highly sensitive to the subregu-
larity that no-change verbs all end in a t or a d, overgeneralizing it to regulars
and to other irregulars that end in t or d. This can be attributed to the fact
that the model records the relative frequencies of many different mappings
between substrings of stems and substrings of pasts, superimposing them
across the different verbs that exemplified them.
Pinker and Prince (1988), while disagreeing with Rumelhart and
McClelland's suggestion that both regular and irregular forms are gener-
ated in a single associative network, noted that their model might offer
insights as to how irregular verbs are stored. If the traditional notion of
rote memory for irregular storage is thought of not as an unstructured list
of slots but as also involving some kind of associative network in which
recurring similarities are recorded and superimposed, the hypersimilar
family resemblance classes can be explained because they contain sets of
verbs that are easier to memorize than unrelated singletons, and that are
prone to occasional generalizations (e.g., brung, bote) by analogy. In this
interpretation, Rumelhart and McClelland would be providing a better
model of the irregular rote component of the inflectional system. Related
suggestions had been made prior to Rumelhart and McClelland (1986) by
Bybee and Slobin (1982a) and MacWhinney (1978). MacWhinney suggests
that some children utter some productive forms by a mechanism distinct

123

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
MONOGRAPHS

from both rote lexical storage and rule application, namely, rhyme-drive
analogies. Bybee and Slobin suggested that speakers form schemas for recog-
nizing typical phonological patterns of irregular past tense forms. Children
learn to associate past forms with their stems more easily if they confor
to a past tense schema, and they are more likely to select stored forms that
conform to a schema when producing past tenses.
If belonging to a family of similar irregulars undergoing simila
changes strengthens the memory trace of a given irregular form, it should
be more resistant to overregularization than more isolated irregulars, hold-
ing frequency constant. The prediction that partial regularity blocks over
regularization was first suggested by Slobin (1971) and has been furthe
tested by Bybee and Slobin (1982a) and Kuczaj (1977a, 1978). The mos
robust effect is that verbs that end in t or d are less likely to have -ed added
and are more likely to be uttered in no-change form, than verbs withou
those endings. This is true for both no-change irregulars, leading to im-
proved performance, and other kinds of irregulars, leading to no-chang
errors; both kinds are protected from overregularization (for reviews, se
Pinker & Prince, 1988; and Rumelhart & McClelland, 1986). (We see the
effect in our data as well: of the 11 verbs ending in t or d listed in App
Table A10, 10 of them were overregularized at aggregate rates lower tha
the mean across verbs.) As mentioned, the Rumelhart-McClelland model
duplicated this phenomenon. However, Pinker and Prince (1988) point ou
that the effect is potentially so overdetermined that identifying the psycho-
logically active cause or causes is nearly impossible. The no-change class
large (the largest among the irregular verbs), shows an exceptionless hyper-
similarity (all its verbs end in t or d), involves a single kind of change (none)
shares its verb-final consonant with the regulars, and when regularized r
sults in a phonological pattern (adjacent identical stop consonants) that th
phonology of English tries to avoid. Thus, the existence of an effect o
family strength should be confirmed with other materials.
Bybee and Slobin also showed that children overregularized subclasses
with different kinds of vowel changes at different rates. They attributed th
differences to different degrees of stem-past similarity, as discussed in the
preceding section. But as we have seen, the effects of stem-past similarit
are difficult to demonstrate, if they exist at all. Furthermore, Pinker an
Prince (1988) showed that stem-stem similarity may be the more relevan
factor: the overregularization rates for the different vowel change classe
correlate well with the number of English irregular verbs sharing the vowel
changes that the class members undergo. They suggested that this explain
why the Rumelhart-McClelland model mimicked the ranking of overregu
larization rates for these subclasses, at least in one stage. However, eve
here the existence of a stem family effect was not perfectly clear: Bybe
and Slobin's subclasses, as interpreted by Rumelhart and McClelland, wer

124

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
 MARCUS ET AL.

heterogeneous and contained many possible contaminants, such as the in-

clusion of the unusual go-went in the blow/grow/know subclass (for discussion,
see also Egedi & Sproat, 1991).
A better test of the family strength effect would eschew the necessarily
imperfectly constructed subclasses in favor of a direct measure of the
strength of the family members for each irregular verb. Because there are
many different possible models of associative memory (differing in how
widely their associations spread in phonological space), none of them simu-
lated perfectly by any given family strength measure, we computed three
different measures, of differing degrees of inclusiveness, all of them based
on the principal dimensions of similarity within families of irregulars
(Pinker & Prince, 1988). The first measure was based on rhymes: for each
verb we summed the frequency (not the log frequency) of the past tense
forms of each of the other irregular verbs whose stems and past tense forms
rhyme with those of the verb in question. For example, for sting-stung we
would add the frequencies of clung, flung, swung, and so on. The verb's own
frequency was not included; although it surely affects the strength of the
family it belongs to, we wanted to see if we could find independent support
for a family strength effect, unconfounded by the frequency effect already
documented. The second measure was based on the final consonant cluster:
we summed the frequencies of the past tense forms of all the irregular verbs
that shared the final consonant cluster with the verb stem in question and
that underwent the same change from stem to past (vowel change, conso-
nant change, consonant addition, and so on; see the subclasses in the appen-
dix to Pinker & Prince, 1988). For example, for stick-stuck we would add the
frequencies of struck and snuck, even though their stems, strike and sneak,
do not rhyme with stick. The third and most inclusive measure added the
frequencies of the irregular verbs that shared a final consonant with the
verb stem in question and that underwent the same change from stem to
past. For example, for stick-stuck we would add the frequencies not only of
struck and snuck but also of stunk and slunk. Verbs ending in a vowel were
treated as if they shared a final consonant.
Because many irregular neighbors like slunk would be far-fetched can-
didates for children's lexicons, we actually selected members of irregular
word families, and took their frequencies, from the adults' speech in the
transcripts of the child in question, using the 19 children with individual
transcripts that overregularized at least once. The independent variables
took on very different values for each child, so aggregate measures are not
appropriate.
The prediction we are testing is as follows. The higher the frequencies
of an irregular verb's family members, the less likely the verb is to be over-
regularized. The prediction appears to be borne out. For the family of
rhymes, the correlation coefficient between family strength and overregu-

125

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
MONOGRAPHS

larization rate was negative for 17 of the 19 children. The mean of th

correlations, -.07, is significantly different from zero, t(18) = 2.23, p
.05. For the family of verbs sharing a final consonant cluster and a pa
tense change, the correlation was negative for all 19 children, with a me
of -.11, t(18) = 9.99, p < .001. For the most inclusive family, sharing a
final consonant and a past tense change, the correlations were also negati
for all 19 children, with a mean of -.11, t(18) = 10.03, p < .001.
To ensure that this effect cannot be attributed to a confound with the
frequency of each verb, we held the log parental frequency constant in a
partial correlation analysis. For all three family sizes, 16 of the 19 children
had negative partial correlation coefficients, and the mean partial correla-
tion coefficient remained negative and significantly less than zero (for
rhyme, mean r = -.08, t[18] = 2.22, p < .05; for final consonant cluster,
mean r = -.08, t[18] = 4.005, p = .001; for final consonant, mean r =
-.08, t[18] = 4.01, p = .001).
We conclude that there is a small, although reliable, effect whereby
verbs are protected from overregularization to the extent that they are
phonologically similar to other verbs (weighted by their frequencies) dis-
playing the same irregular pattern.

IV. ATTRACTION TO FAMILIES OF SIMILAR REGULAR VERBS

The preceding analysis confirms the hypothesis of Slobin, Bybee, and

Kuczaj that partial regularity blocks overregularization. It is consistent both
with the Rumelhart-McClelland model and with Pinker and Prince's aug-
mentation of the traditional rote-rule model in which the rote component
has some associative-memory-like properties. A test that might distinguish
the latter two models is whether families of similar regular verbs pull an
irregular toward overregularization in the same way that families of similar
irregulars pull it away. Since in the traditional rote-rule model regular past
tense forms need not be stored because they can be generated by a rule,
under the simplest hypothesis regular past tense forms should not attract
irregulars. Of course, storage of regulars is possible within such theories-
what could prevent it, given that it allows the learning of individual irregular
items?-but such storage is not necessary except under certain circum-
stances in which the existence or form of the regular cannot be predicted,
such as in the child before the regular rule is learned, and doublets such as
dived and dove where both members must be stored. Thus, in general, we
should not find strong effects of storage of regulars in a rule theory, and
in no case should the ability to generalize to new forms depend on the
previous storage of similar old ones. That this property does in fact distin-
guish rule-based theories from the Rumelhart-McClelland model was shown

126

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
 MARCUS ET AL.

by the behavior of the model on newly presented regular verbs. At asymp-

tote, the model erred on 33% of the regular verbs it was tested on, produc-
ing no output at all for six that were dissimilar from those in its training
set, such as jump and pump (Pinker & Prince, 1988; Prasada & Pinker, in
press). In contrast, adults easily regularize highly unusual sounding novel
forms such as ploamph or keelth (Prasada & Pinker, in press), and between
the ages of 2 and 5 Abe left virtually no regular verb in past tense contexts
unmarked, including his own unusual inventions eat lunched, bonked, borned,
axed, fisted, and poonked (Kuczaj, 1977a).
To test whether families of regular verbs pull similar irregulars toward
overregularization, we first extracted the 1,826 regular verbs rhyming with
the irregulars that were listed in an on-line version of Webster's Seventh Colle-
giate Dictionary, which contains phonological representations of 8,217 verbs.
Just as for the irregular families, three different size nets were cast. In the
first, the sum of the frequencies of the regulars in a parent's speech that
rhymed with each of the child's irregulars was computed (e.g., winked would
contribute to the regular attractor strength of stink). In the second, verbs
with the same final consonant cluster were grouped, so that yanked and
honked as well as winked join the family of stink. In the third, verbs with the
same final consonant were grouped, so that hiked and harked would also
belong to stink's attracting family. Verbs that ended in a vowel were treated
as if they had a common final consonant.
If families of regular verbs pull similar irregulars toward overregulari-
zation, correlations between regular family strength and overregularization
rate should be positive. In none of our tests do we find statistically significant
correlations in that direction. For families of rhymes, the correlation was
positive for 7 of the 19 children, with a mean of -.01. For families sharing
a final consonant cluster, correlation coefficients were positive for 11 of the
19 children, with a mean of .08, not significantly different from zero, t(18)
= 1.14, p > .10. For families sharing a final consonant, the correlation was
positive for 11 of the 19 children, with a mean (.06) not significantly differ-
ent from zero, t < 1. When the verb's own frequency is partialed out, the
correlations with the smallest family strength measure (rhymes) are positive
for only 8 of the 19 children, and the mean correlation across children was
negative. For the middle-sized family (shared final consonant cluster), the
correlations were positive for 12 of the children; the mean correlation was
.09, not significantly different from zero, t(18) = 1.45, p > .10. For the
largest (final consonant) family, the correlations with family strength were
positive for 12 of the 19 children, with a mean of .06, not significantly
different from zero, t(18) = 1.11, p > .10. Furthermore, it is not the case
that the negative correlations come from children with smaller (hence nois-
ier) samples; of the seven children who went in the nonpredicted direction,
six were among the 14 children whose samples contained more than 100

127

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
MONOGRAPHS

irregular past tense tokens, and one of them was Abe, who had the richest
data base of overregularizations.
Each of these measures was computed a second time, excluding polysyl-
labic verbs. Pinker and Prince (1988) point out that all irregulars are mono
syllabic except for forms like forget, understand, and overthrow that contain a
prefixed monosyllabic irregular. Therefore, if a verb is nonprefixed and
polysyllabic, it is guaranteed to be regular. If a child became sensitive to
this contingency, he or she could sequester such verbs from the mechanism
giving rise to family strength effects, and our including them in our esti
mates could dilute our estimates of such effects. But the results were virtu-
ally identical in all cases.

V. SUMMARY

Correlations between children's overregularization rates and vari

properties of different irregular verbs yield the following conclusions
expected, irregular verbs with lower-frequency past tense forms are m
likely to be overregularized, underscoring the important role of the mem
strength of the irregular past tense form in the overregularization proces
The degree of phonological similarity or dissimilarity between a stem
its past tense form appears to have no influence on overregularizati
suggesting that the errors are not primarily caused by difficulties in exec
ing phonological changes in generating irregular forms or in uniting stem
and pasts as part of a single verb paradigm during learning. We fou
evidence that irregular verbs are protected from overregularization by fam
lies of similar irregulars, although we failed to find evidence that they
drawn toward overregularization by families of similar regulars. This
ports Rumelhart and McClelland's assumption that irregular patterns
stored in an associative memory, but fails to support their assumption tha
the regular pattern is stored in the same system.

128

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
 VIII. SUMMARY AND CONCLUSIONS

The facts of overregularization can be summed up simply. After

period in which all the child's past tense forms of irregular verbs are corre
the child begins to overregularize. Overregularization then occurs at a l
rate throughout the preschool and early school years, affecting all irregula
verbs, to an extent that depends on the verb's rarity in parental speech. Its
overall rate appears to be independent of changes in the mixture of regular
and irregular verbs in the child's speech, the child's parents' speech, or
child's vocabulary. Instead, it seems to depend on the acquisition of t
tense-marking system as a whole: development of the ability to mark regul
verbs reliably for tense appears to be the immediate harbinger of overregu
larization, and reliable marking of irregular verbs for tense accompanies it.
Aside from frequency, verbs' proneness to overregularization depends t
small extent on the strength of the verb's phonological neighborhood: clus-
ters of similar irregular verbs protect one another from overregularization
In contrast, clusters of similar regular verbs do not appear to pull an irreg
lar toward overregularization.
These facts can be accounted for by a simple theory. The child stor
irregular past tense forms in a rote memory system, in which the strength
of a memory trace is monotonically related to the frequency with which it
is encountered. In addition, this memory system has some of the propertie
of an associative network: stem-past pairs displaying similar relations re
force each other. (This same property occasionally leads to irregular gen
alizations such as brang and wope.) Regular past tense forms, in contrast, ar
generated by a mental concatenation operation that attaches a suffix t
stem. Because this rule can always be applied on line, regularly inflecte
forms need not, in general, be stored (although they can be under cert
circumstances, such as before children have learned the rule). And beca
it simply adds an affix to the end of a stem with unspecified properties, th

129

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
MONOGRAPHS

similarity of a given stem to previously encountered ones plays no role.3

The two systems interact in a simple way: the retrieval of a stored irregular
entry blocks the application of the regular rule.
The fact that overregularizations are a small minority of irregular past
tense utterances at all stages shows that the blocking process is active in the
child as soon as there is evidence for two modes of inflection at all. When
overregularizations do occur, they are straightforwardly explained as a fail-
ure to retrieve the irregular past form (or, for past+ed errors, its "past"
feature) in real time. This tendency is related to the frequency of the form
in an obvious way that is an immediate consequence of the logic of irregular-
ity and the fallibility of human memory. In the extreme case, an irregular
form that has been attended to with zero frequency (e.g., shend for adults
and many irregular verbs for young children) will have no memory trace;
hence, it will be retrieved from memory with zero probability and will always
be overregularized if the form is to be tense marked at all. An irregular
form that has been heard once has a weak memory trace and hence a
probability of being retrieved that is greater than zero but less than one.
Irregulars that have been heard more times have correspondingly stronger
memory traces and lesser overregularization probabilities; irregulars that
have been heard thousands of times will be successfully retrieved virtually
always. The learnability problem of recovering from errors is solved by a
blocking principle that operates throughout development, fed by irregular
forms whose potency increases with increasing exposures during devel-
opment.
Of course there are aspects of overregularization that remain to be
explained. There is considerable unexplained variance in exactly which
verbs are overregularized at which rates and ages, and among different
children's overregularization rates, and the temporal relation between oblig-
atory tense marking and overregularization is as yet unclear. At present, we
believe that we have run up against the limitations of the available data. Just
as prior characterizations of overregularization were unclear or misleading
owing to the limitations of paper-based diaries and transcript samples, cur-
rent computer data bases of ChiLDES size, although they have enabled
tremendous progress, are still not up to the enormous task facing us in
trying to understand details of language development. An hour of speech
a week is still a paltry sampling of the richness of the language acquisition
process, especially in the critical third year of life. We suspect that child
language data bases with sampling frequencies of an order of magnitude

31 The variation in the phonetic form of the regular affix, whereby d surfaces as t
following unvoiced consonants and id following coronal stops, can be attributed to general
phonological processes operating throughout English, not to the regular process itself (see
Pinker & Prince, 1988).

130

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
 MARCUS ET AL.

greater than current ones would resolve many of the uncertainties in the
current conclusions.

I. OBSERVATIONS ON THE NATURE AND LEARNABILITY OF THE

TENSE-MARKING SYSTEM

Our suggested explanation of the overregularization process has thre

parts: decision to mark tense, imperfect memory retrieval, and possession of
a regular process with universal applicability. Although the second postulate
reflects a possibly uninteresting memory limitation of the child, the other
two reflect quite remarkable linguistic accomplishments.

A. Observations on Tense Marking

As Pinker (1982, 1984) and Slobin (1982) point out, obligatory gram
matical constraints pose difficult learning problems for the child. The fa
that an inflection is obligatory means that there are no pragmatic cues t
the semantic features that the inflection is encoding; parents must expre
the pastness of a past event, regardless of how relevant it is in the conversa-
tional context. Moreover, once the child has somehow figured out that pa
tense inflection encodes past tense, if he or she mistakenly assumes that
is optional, no parental input short of negative evidence can contradict th
assumption. Pinker and Slobin thus suggest that the child is innately pre
pared to consider obligatory inflectional tense marking as a possible con
straint in the language to be acquired; such a hypothesis is easily discon-
firmed in systems where inflection for tense is in fact optional or no
available. Our seemingly homely explanation for overregularization in
terms of retrieval failure depends on the child having solved these daunting
learnability problems (for an explicit hypothesis as to how the child solve
them, see Pinker, 1984; and for evidence that children solve them quickly
see Stromswold, 1990; the problems are of course finessed in network simu-
lations that are fed correct stem-past pairs in isolation). If the 2-3-year-old
child did not consider tense marking to be obligatory, the failure to retrieve
an irregular past in a past tense context would not automatically lead t
overregularization; the child could simply leave the verb unmarked.

B. Observations on Regular Rules

A second noteworthy linguistic achievement of the late 2-year-old is

possession of a process that is capable of yielding an inflected output form
for any verb, no matter how strongly linked with an idiosyncratic irregular

131

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
MONOGRAPHS

form, and regardless of whether a family of similar regular forms is avail-

able to serve as an analogy-supporting model. As Pinker and Prince (1988
point out, a rule that simply concatenates an affix with a stem, characterized
in terms of a variable standing for any stem rather than particular patterns
of the typical phonological contents of stems, easily provides this capability.
A set of associations between stem phonology and past tense phonology
tied in varying degrees to the patterns it has been trained on. While such
models could, in principle, approximate the unlimited applicability of a rule
by training it on a set of regular stems that span enough of the phonological
space of English to cover all cases, in practice this ability is compromised by
the necessity of curbing links to the regular ending in order to avoid applica
tion to the irregulars. Presumably, it is for this reason that the Rumelhart-
McClelland model failed to display the appropriate generalization abilitie
for novel regulars (Prasada & Pinker, in press). In managing to come ou
with a past tense form close to 100% of the time when called for, despit
less than perfect memory retrieval, the children we have studied (most nota-
bly Abe) clearly have mastered an inflectional process of very wide applica-
bility. If this regular process were not capable of applying to arbitrary irreg
ular stems (and, in fact, 90% of Abe's irregulars were regularized at leas
once, together with many unusual creative forms like poonked), or if it de-
pended on the existence of similar high-frequency regulars (which it doe
not seem to), then when faced with irregular retrieval failure, the child
would be left with no choice but to utter the stem, even if his or her language
system had called for a past tense form.
We envision regular rules as mental symbol concatenation operations
similar in operation to the core rules of syntax (see Pinker, 1991; Pinker
Prince, 1988, 1991; Prince & Pinker, 1988). This hypothesis is about a kin
of mental machinery available to all language learners but put to use in
different ways in different language systems. In English, it is used in
condition-free rule within a component of inflectional morphology that
takes inputs from the rest of the morphological system and delivers it
output to the syntactic component. This computational characterization may
differ from many properties of rules informally called "regular" in tradi
tional descriptive grammars. First, a regular rule in our sense need not b
the only productive morphological phenomenon in a language; analogizin
of irregular patterns (e.g., brung) can add new forms as well. Second, it need
not apply to the majority of words in a given part-of-speech category; there
is nothing in our characterization of regularity that directly pertains to the
relative number of words involved, except that, as a fully productive an
general process, a regular rule is potentially applicable to an unlimited num-
ber of stems whereas irregular analogizing becomes less and less probable
as one departs from a subclass prototype. As such, a regular rule may be
the only available device for borrowings, neologisms, conversions, onomato-

132

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
 MARCUS ET AL.

poeia, and other forms that cannot be analogized to existing irregulars (see
Kim, Pinker, et al., 1991; Prasada & Pinker, in press). Thus, a regular
pattern is liable to assume majority status if such forms enter a language in
large numbers, as has happened in English since the Middle English period
(Pyles & Algeo, 1982). Third, a regular rule need not be totally insensitive
to properties of the stem it affixes to: multiple rules may exist, each affixing
an allomorph that applies to stems with a different property, although the
conditions would be well defined (e.g., ending in a vowel vs. a consonant)
and applicability to a stem within the class would be all or none, not propor-
tional to the stem's global similarity to previously encountered stems. The
most common hallmark of a regular process in our sense is default status,
the ability to apply freely to any word not already linked to an irregular
form, regardless of whether it is covered by memorization or analogy (e.g.,
We rhumba'd all night). The computational nature of the rule, as concatena-
tion of an affix to a variable standing for the stem, renders it uniquely
capable of playing this default role.

C. How Might a Regular Rule Be Learned?

We have said nothing about how any particular regular rule, such as
the English past tense rule, is acquired. Extraction of the suffixation pattern
itself can be achieved by examining the phonological differences between
forms like walk and walked, which can be accomplished by a variety of pat-
tern extraction algorithms (see, e.g., MacWhinney, 1978; Pinker, 1984;
Pinker & Prince, 1988; Rumelhart & McClelland, 1986). A greater challenge
is to show how the child decides whether to internalize a pattern as a regular
rule or as a list of (possibly analogy-supporting) individual items such as
bend-bent. How might this be done? Given the results of Chapter V, children
do not appear to depend on a regular pattern applying to a majority of
tokens or to be influenced by either the relative or the absolute number of
types. Indeed, children begin to overregularize with something on the order
of 100 verbs in their vocabularies, and they mark few or none of them for
past tense before they begin to overregularize the irregulars, raising the
possibility that the child does not need to process large amounts of input
data to seize on the regular pattern. We lack evidence that would allow us
to identify which cues children actually use to acquire a regular rule, but
we can list several logical possibilities, each one a visible consequence of the
regular nature of a rule and hence a potential cue to regularity (for further
discussion, see Pinker & Prince, 1991). We can also check to see if any of
these cues is clearly ruled out by existing evidence on the linguistic input
available to children.

First, Pinker and Prince (1988) suggested that the crucial cue might be

133

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
MONOGRAPHS

the ability of a morphological process to apply successfully to several kinds

of stem, each belonging to a different competing irregular pattern. For
example, need-needed exists despite bleed-bled and feed-fed; blink-blinked exists
despite drink-drank and sink-sank; seep-seeped exists despite sleep-slept and
sweep-swept. Indeed, all the rhymes of irregular patterns in English (except
perhaps for verbs ending in -ing) can also be found in verb roots that are
regular. This property is diagnostic of the fact that the regular rule doe
not carve out a set of phonological territories in the interstices of those
claimed by irregular families but can apply to anything at all. The phonolog-
ical omnipotence of the regular rule is even more apparent in regular-
irregular homophones, such as lie-lay (recline) and lie-lied (fib),fit-fit (intran-
sitive) and fit-fitted (transitive), and meet-met and mete-meted. (Of course, the
homophones themselves are surely too rare to be the cues used by children.)
Second, the mere heterogeneity of the stem patterns that are heard to
be regularly inflected (regardless of whether the inflection trumps some
competing irregular pattern characteristic of such stems) may tell the child
that the inflection is the product of a rule with either well-defined conditions
or none at all. The Rumelhart-McClelland model is driven by a process
roughly of this ilk because the applicability of the regular suffix in the model
depends on the variety of input patterns that are linked to the output nodes
representing that suffix, and these links are strengthened with exposure to
the relevant patterns. But the learning process we consider here is somewhat
different. As mentioned, the range of generalization of the regular pattern
in the Rumelhart-McClelland model is computed over the same representa-
tion that is used to represent the words to begin with. This representation
has to have enough information to distinguish regulars from any irregular
and for the full phonetic form of the past tense form to be reconstructed.
This allows the potential conditions for the regular pattern to be sliced too
finely; the regular pattern can be tied to highly specific combinations of
stem patterns, with no across-the-board generalization to all nonirregular
stems. If, instead, the phonological content of words is represented in one,
detailed kind of representation, but a small number of candidate conditions
of a rule are represented in a separate, much sparser one, then a morpho-
logical pattern can be recognized as fully applicable as soon as it is seen to
appear with every one of the stem conditions in this smaller, better-defined
set (for discussion, see Pinker, 1984; Pinker & Prince, 1988, 1991; and
Prasada & Pinker, in press).
Third, a very reliable cue for regularity is the ability of a process to
apply to verbs that are derived from other categories such as nouns, adjec-
tives, and names. These verbs do not have verb roots and hence have no
base lexical entry to which an irregular memorized form can be attached;
only a fully general regular rule can apply to them, acting as a default (see
Kim, Marcus, et al., 1991; Kim, Pinker, et al., 1991; Kiparsky, 1982; Pinker

134

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
 MARCUS ET AL.

& Prince, 1988). For example, ring the city (i.e., "form a ring around") has
the past tense form ringed, not rang, because this sense of to ring has a noun
root, not a verb that could be collapsed with the one underlying ring-rang.
Hearing such forms could, in principle, tell a learner that the -ed affix is
not stored with verb roots and hence must be a regular operation.
Fourth, even a single word with a highly unusual sound pattern, such
as out-Gorbachev'd or rhumba'd, provides information about the generality
and possible default status of a morphological pattern.
Fifth, the syntactic requirement that tense be marked obligatorily may
impel the child's learning mechanism to seek a pattern that can provide the
needed form under a wide range of circumstances. The fact that Adam,
Eve, Sarah, and Abe were marking irregular verbs for tense a majority of
the time near the period in which they were first controlling the regular
suffix lends some credence to this suggestion. Of course obligatoriness itself
cannot distinguish between regular rule and irregular storage plus analogy,
but it might cause the child to promote the most general process to rule
status.

Finally, aspects of the phonological properties of regular

might provide cues about its regularity. For example, the English
inflectional process consists of a suffixation, which is the same ki
cess used for the third person singular inflection (-s), the progres
tion (-ing), the plural, and the possessive. This could signal to
that, in English, suffixation is the process used quite generally
inflection whereas the mutations seen in irregular verbs do no
any larger system (see Wurzel, 1989). Moreover, a suffix that
change on a stem, or some change that is consistent across all st
be classified as regular because it would suggest a process that
verb as an opaque variable rather than in terms of its phonolog
(Kiparsky, 1982). In English, these properties hold of regular
and of course are confounded: adding a -d suffix leaves virtua
unchanged (the exceptions are the irregular verbs flee, say, hea
and do).
As mentioned, this is a list of hypothetical cues, which we provide to
allay possible suspicions that regularity in our sense is unlearnable or learn-
able only from input sequences that our data already rule out. To decide
which of these cues are in fact used by the child's learning mechanisms, it
is necessary to determine which of them reliably accompany regular rules
across different languages, which are in fact available in parental speech to
children of the relevant ages, and which ones are controlled by children
simultaneously with their treatment of the pattern as a regular process.
The actual set of regular verbs that children use to induce regular past
tense inflection is, of course, difficult to determine from transcript data,
which record a small fraction of the speech heard by children and provide

135

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
MONOGRAPHS

no information about what portion of that speech is actually used by

child. In Appendix Table Al , we attempt to bracket the true list by provi
ing one list that is likely to be too big and another that is likely to b
small. The maximal list contains all the regular verbs used at least on
the entire set of transcripts, in any tense, by Adam, Eve, Sarah, or the a
conversing with them. This list approximates the pool of verbs potent
occurring in past tense forms in parental speech to which the children m
be attending before they deploy the regular inflection productively. W
not know, however, that each child heard each verb in this list, or that e
child heard each verb before acquiring the regular inflection, or that
child had actually attended to the verbs he or she did hear. Thus, we
provide a more conservative estimate, consisting only of the regular v
that each child actually used (in any tense) in his or her transcripts prece
the first transcript with an overregularization.
The maximal list is consistent with several of the possibilities for
to regularity entertained in this section. It contains many verbs that rhy
with irregulars (e.g., ache, blind, blink, call, care, cheat, claw, cry, dare, die,
exercise, fold, fry, guide, hatch, heat, invite, hand, kick, kid, land, lick, leak,
live, mind, pee, peel, peep, pick, pin, play, pray, pretend, rake, reach, row, rem
repair, scare, scoot, sew, ski, smell, sneeze, snow, spell, spray, squeeze, stare,
steam, surprise, sway, tease, tick, tie, trick, trim, trust, turn, use, wet, yell). In
most of the irregular clusters of rhymes have regular counterparts in
list. In addition, there are a number of likely denominal verbs (inclu
bomb, bubble, chain, clip, color, comb, dust, end, fan, fish, glue, hammer, iron,
mail, paint, paste, pee, pump, rope, sail, screw, shovel, skate, ski, smoke, stamp, st
steam, tape, thread, unbuckle, unbutton, unchain, unplug, unscrew, and wee
and onomatopoeic verbs (including bang, bark, bash, blast, bop, burp,
crack, crash, fizz, growl, howl, jabber, meow, pop, smash, swish, zip, and z
which may have the same status as denominals (see Pinker & Prince, 1
Verbs with noncanonical and complex phonologies (especially verbs th
not conform to canonical pattern for English basic words, namely, monos
lables or polysyllables with initial stress) are represented by allow, at
belong, decorate, disappear, disturb, erase, excuse, meow, organize, prepare, pre
recognize, remember, and urinate, among others.
Interestingly, even the minimal lists (regular verbs used by each c
before overregularizing) exemplify most of these potential cues. The read
can confirm that each child controlled many verbs rhyming with a variet
of irregulars. Adam and Sarah had acquired several denominal verbs (
paint, plug, rain, rope, and screw and comb, paint, rain, smoke, and wee-w
respectively) and onomatopoeic verbs (crack, growl, knock, squeal, and squ
and knock and bump, respectively). Adam and Sarah also had phonolog
noncanonical verbs (attach, exercise, and remember for Adam; organiz
Sarah). Even Eve, whose minimal list comes from only four transcripts, u

136

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
 MARCUS ET AL.

eight verbs rhyming with six kinds of irregulars (cry, kick, lick, pick, play,
show, turn, use), one likely deadjectival verb (empty), and an onomatopoeic
form (crack). Thus, the data suggest that each of the possible cues to regular-
ity entertained in this section is available to children; whether they are used,
necessary, or sufficient, however, remains unknown.

II. RELEVANCE TO HISTORICAL CHANGE

Children regularize irregular verbs, especially the lower-frequen

ones, and the English language, over the centuries, has been regulariz
irregular verbs, especially the lower-frequency ones (Bybee, 1985). Th
must be some connection. Perhaps anticipating a trend, Newsweek (Gelm
1986) attributed to Jill de Villiers the half-joke, "Leave children alone
they'd tidy up the English language."
But the remark can be turned around. Children, in fact, are left alone
(Brown & Hanlon, 1970; Marcus, 1992; Morgan & Travis, 1989), bu
does not really matter whether they are or not, for time is on their side.
they have to do is wait, and they will be adults in full custody of the lang
that is passed on to the next generation. If children really had a distaste f
irregular verbs, nothing could have stopped them from tidying up the
glish language long ago, yet we still have over 100 irregular verbs, most qu
secure, some of them transmitted as irregulars in a generation-to-generati
chain of successful memorizations linking us to prehistoric peoples th
sands of years in the past (Pyles & Algeo, 1982). Clearly, children are
the relentless rule engines of earlier accounts but are quite happy to le
irregulars, and why not? If one has the mental tools to acquire on the orde
of 50,000 words, each representing an arbitrary sound-meaning pairin
and the ability to link abstract features like tense to words with sp
markings, memorizing and retaining another few hundred words with
features built in should be no strain. And blocking their regularized co
terparts is simply a consequence of the fact that stem and irregular
part of an organized grammatical system, with designated slots for ea
feature-marked variant of a word.

As Bybee and Slobin (1982b) point out, it is adults who bear most of
the blame for the permanent regularization of irregulars in a language
They note that one cannot attribute historical change to children's overregu-
larization errors unless children prefer these errors and continue to do so
into adulthood, which, for virtually all irregular verbs in a given generation,
they do not. Given what we know about overregularization in children and
adults (e.g., Ullman & Pinker, 1990, 1992), the following scenario seems
more plausible. A weak irregular memory entry in adults can lead to occa
sional blocking failures, hence regularizations, for the same reason that

137

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
MONOGRAPHS

children overregularize. Presumably, this reduces the frequency of th

regular past tense form in the parent generation's speech further, and co
bined with an overall decline in all tense forms for the verb, it may
to the point where one generation of children rarely hears it and h
never ceases to overregularize it, at which point it has changed to a regul

III. RELEVANCE TO CONNECTIONIST

MODELING OF LANGUAGE ACQUISITION

Although some of the questions treated in this Monograph were fir

raised in Pinker and Prince's (1988) analysis of Rumelhart and McClellan
connectionist simulation of language acquisition, this Monograph is not
tended as an extension of their critique, and the points we make here
meant to shed light on the nature and causes of overregularization in g
eral. They are not intended either to support or to refute the entire set
connectionist models, or the entire set of models based on grammati
rules, but pertain to many issues that are independent of that distincti
such as the relevant rarity of overregularization, its dependence on t
frequency of irregular past forms, and its relation to tense marking
general.
In only three places did we explicitly test predictions of the Rumelhart-
McClelland model. The first examined their suggestion that children's U-
shaped development could be explained without a distinction between early
rote storage of lexical items and later deployment of a productive morpho-
logical process. Rather, they suggested, a single pattern associator could
display that transition given the assumption that irregulars are higher in
frequency than regulars and that the child undergoes a vocabulary explo-
sion causing the sudden acquisition of a large number of regulars and a
concomitant change in the number of regular items submitted to the past
tense learning device. Our results fail to find evidence either for a change
in the input to the past tense system or for a vocabulary explosion at the
right point in development that was assumed to cause it. In contrast, the
data suggest that children use correct irregulars before overregularizing
because they lack the productive mechanism generating regular forms alto-
gether. These discoveries remove one of the more dramatic phenomena
that had been adduced as evidence for the Rumelhart-McClelland model
(e.g., Sampson, 1987; Smolensky, 1988). Of our other two tests, one sup-
ported the Rumelhart-McClelland model's prediction that irregular verbs
are stored in a memory system that records patterns of mapping between
stems and pasts and applies them in graded fashion to similar stems. The
other test failed to support the model's prediction that the patterns shown
by regular verbs are stored and generalized in this way.

138

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
 MARCUS ET AL.

A natural question arises. Do the incorrect predictions of the

Rumelhart-McClelland model apply to connectionist models of the acquisi-
tion of inflection in general or only to their particular implementation?
Theorists arguing for connectionism as a general solution to problems in
cognitive science have, virtually unanimously, provided one answer: the
problems are specific to the Rumelhart-McClelland model, and are primar-
ily technological. The Rumelhart-McClelland model, an early effort in the
recent revival of neural net modeling, was a feedforward network without
a hidden layer; that is, each input node was connected directly to each
output node, and activation passed from input to output in a single step.
Gasser and Lee (1991), MacWhinney and Leinbach (1991), McClelland
(1988), and Plunkett and Marchman (1991, 1990) have claimed that most
of its problems are eliminated when higher-tech connectionist machinery is
substituted. The usual suggested replacement is a feedforward network with
one hidden layer of nodes separating input and output (e.g., MacWhinney
& Leinbach, 1991; Plunkett & Marchman, 1991), its connection weights
modified by the error back-propagation learning algorithm (Rumelhart,
Hinton, & Williams, 1986).
The claim that addition of a hidden layer is sufficient to remedy the
Rumelhart-McClelland model's problems, especially its failed predictions
about the dependence of overregularization on vocabulary balance, has
never been demonstrated, however, and it is probably wrong. Egedi and
Sproat (1991) have tested a parallel distributed processing (PDP) model
that was trained on a sequence of English verbs similar to that fed to the
Rumelhart-McClelland model, and they submitted their model to similar
kinds of evaluation. The model enjoyed the advantages of a hidden layer
whose weights were trained by the back-propagation algorithm, a more
realistic phonological representation, and a more powerful mechanism con-
verting the output node activations to pronounceable words. Nonetheless,
the behavior of the Egedi-Sproat model was virtually identical in the rele-
vant respects to that of the Rumelhart-McClelland model. It displayed U-
shaped development in its acquisition of irregulars at a point at which it
was suddenly flooded with regulars, but it showed no such trend with a
constant input mixture. And it failed to produce a coherent output form
for large numbers of new regular verbs on which it had been tested if those
verbs differed phonologically from the common patterns in the training set.
Plunkett and Marchman (1991, 1990) claimed that their hidden-layer
model, trained on a variety of schedules with artificial verbs, showed U-
shaped development with a constant mixture of regular and irregular verbs
as input. But they examined no psychological data and defined "U-shaped
development" as any wiggle in a developmental curve rather than the ex-
tended initial period of correct performance in children that psychologists
had pointed to and that we confirmed quantitatively. Even this claim was

139

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
MONOGRAPHS

abandoned in a subsequently written paper (Plunkett & Marchman, 1

which relied on a changing vocabulary balance as much as the Rumel
McClelland model did; moreover, in that paper, they switched to a t
and a fourth definition of "U-shaped development" (alternation bet
correct and irregular tokens, and immunity of early learned irregulars f
overregularization; see Chap. IV, Sec. IIA), again leaving the original
nomenon unexplained within their framework.
It is not surprising that the mere technological improvement of addin
a hidden layer does not change the psychological fidelity of models with
basic design of the Rumelhart-McClelland model. The main point of P
and Prince (1988) was not that connectionist models are incapable of m
ing psychological phenomena but that many of the key theoretical comm
ments in explaining such phenomena have little to do with whether a mo
is implemented in one or another kind of computational hardware
involve more basic questions of representation and organization, such
What is the format of representation for words? How many subsystems
used to map from stem to past? What kind of computation does each
do? What are its inputs and outputs? The problem with many connectioni
models, they argued, was inherent, not in connectionist machinery it
but in the attempts of connectionist theorists to bypass these issues
attempt to use a single, all-purpose learning device for all linguisti
cognitive functions. In particular, Pinker and Prince questioned Rume
and McClelland's commitment (also adopted by most subsequent con
tionist modelers) to a system that lacked distinct representations for wor
that did not distinguish between regular rules and irregular storag
computed both within a single device, and that fed only phonological
possibly semantic) information into the past tense system, ignoring morp
logical and syntactic structure. Obviously, adding a hidden layer of n
to the Rumelhart-McClelland model has nothing to do with these issu
The modeling efforts of MacWhinney and Leinbach (1991) demo
strate this point nicely. The authors do not attempt to model child
U-shaped development; they attribute it to a dissociation between rote sto
age of lexical entries and pattern extraction, as in the more traditi
accounts of MacWhinney (1978), Pinker (1984), and Pinker and P
(1988), and consider the lexical entries to belong to a separate system
they do not attempt to model. But they fail to acknowledge that the com
ment to traditional linguistic distinctions exemplified in their model
even deeper. MacWhinney and Leinbach's representation for the phon
ical content of the stem has a separate subset of hardware nodes for
position in a word, rather than representing order implicitly in a pattern
activation across a large set of context-sensitive nodes, as in the Rumelhar
McClelland model. More important, the architecture of their model, w
they arrived at after a period of trial and error, is clearly tailored

140

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
 MARCUS ET AL.

qualitative architectural distinction between irregular storage and a stem-

independent regular suffixation process. The main route of their model is
a four-layer (two-hidden-layer) pattern associator with 200 nodes in each
hidden layer, sufficient for memorizing the 180 irregular verbs in English
(perhaps even by using some of the hidden nodes as "grandmother"cells, a
direct implementation of the traditional notion of "rote memory"; see Sca-
lettar & Zee, 1988). But they supplemented this conventional connectionist
pattern associator with a distinct second route, which connected each input
node directly with the output node coding the same phoneme and position,
with an innate weight of 1.0 (the maximum), bypassing the intermediate
layers and omitting all connections between one segment in the input and
segments with different positions or different contents in the output. The
bypass route is a second, innate, nonassociative copying mechanism, tailor
made for regular verbs because the regular map simply copies the stem
without internal modification and regardless of its content. Without calling
attention to it, MacWhinney and Leinbach have designed a model with two
distinct, innate pathways, one suited for irregular storage and analogy, the
other for regular suffixation of the stem (for further discussion, see Prasada
& Pinker, in press).
In sum, we see no evidence that adding a hidden layer to a model with
the basic design of the Rumelhart-McClelland model alleviates its empirical
problems, in particular, the two that we examined in this Monograph. At this
point, the models that leave its basic design untouched inherit its problems,
and another model with better overall performance abandoned the radical
assumption of a single associative map and adopted the traditional tenet of
separate mechanisms for lexical entries, irregular links, and a nonassociative
copying process ideal for the regulars.
Although we have shown that existing connectionist models restricted
to a single associative network do not account for the facts of children's level,
onset, and lexical distribution of overregularization, we are not claiming
that no connectionist model is capable of doing so. The set of possible
connectionist models encompasses a wide variety of relative propensities for
rote memory versus generalization and includes models in which the bal-
ance between these tendencies changes during a training run. These ten-
dencies are influenced by a variety of design parameters left open to the
network creator, such as the coding scheme for the input features, their
degree of probabilistic blurring, the topology of the network (e.g., the num-
ber of hidden layers and the number of nodes in each one), the learning
rate, the momentum factor, the temperature, the training schedule, and
others. Conceivably, someone might find a combination of design parame-
ters that allows some unitary connectionist network to display childlike rote
and regularization modes of behavior at different points in a realistic train-
ing sequence.

141

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
MONOGRAPHS

But even if such behavioral mimicry were achieved, this hypothetical

model would still have to be evaluated against the full set of data relevan
to the psychology of morphology (see Pinker, 1991; Pinker & Prince, 1991
and it is unlikely that a generic model tweaked with just the right parameter
settings to display past tense overregularization would be consistent wit
the larger picture. For example, the English plural has very different vocab-
ulary statistics from the past tense (only a few of the high-frequency Englis
nouns are irregular, so irregular forms cannot plausibly dominate early
vocabulary), but it appears to develop with the same U-shaped sequence
and low overregularization rate and submits to the same grammatical treat-
ment in terms of lexical storage and a default regular rule (Cazden, 1968
Kim, Marcus, et al., 1991; Marcus, in preparation; Pinker & Prince, 1988)
Moreover, there is a family of facts of a very different kind that indepen-
dently favors a dual-mechanism (rule-rote) model over any unitary network,
no matter how well it is designed to display rulelike and rotelike modes.
These facts are summarized in the final section.

IV. EVIDENCE THAT CHILDREN RESPECT QUALITATIVE DISTINCTIONS

BETWEEN REGULAR AND IRREGULAR PROCESSES

Linguistic research has shown that regular and irregular inflected

forms differ qualitatively, in terms of their sensitivity to qualitative gramma
cal distinctions and their relation to other grammatical processes. In at least
three cases, children have been shown to be sensitive to these distinctions
in their patterns of regularization and overregularization.
First, Gordon (1985) noted that English compounds can contain irregu-
lar plurals (e.g., mice-infested) but not regular plurals (e.g., *rats-infested). The
phenomenon has been explained in terms of Kiparsky's (1982) level-
ordering theory of morphology, but the kernel of the explanation can be
captured simply: the process forming such compounds takes as input stems
stored in the lexicon, not complex forms created by an inflectional rule; if
irregulars are stored in the lexicon, as we have suggested, they automatically
can feed compounding. In an elicitation experiment, Gordon found that
children produced novel compounds containing irregular plurals (mice-
eater) but never novel compounds containing regular plurals (*rats-eater).
Nor did they form compounds out of their own overregularizations-
children who said mouses would nonetheless avoid saying *mouses-eater.
Second, Stromswold (1990) looked at the three irregular verbs that
are morphologically identical to auxiliaries: do (as in do something or do it),
possessional have, and copula be. Each of these verbs is identical to an auxil-
iary not only in the stem form but in every single one of its irregularly
inflected versions. Do, whether used as a main verb or as the auxiliary in

142

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
 MARCUS ET AL.

negations, inversions, and emphatics, has the irregular forms does and did.
The perfect auxiliary have, like possessional have, maps onto the irregular
forms has and had. Be, whether serving as a copula verb, a progressive
auxiliary, or a passive auxiliary, has the irregular forms am, is, are, was, were,
and been. Moreover, the semantic relations are the same in all cases: the
relation between have a book and had a book is the same as the relation
between have eaten and had eaten; I am tired is to I was tired as I am resting is
to I was resting. Clearly, there are too many of these parallelisms to be
coincidental, and a parsimonious assumption is that the irregular forms of
the main verb and of the auxiliary versions are stored in the same mecha-
nism because, if they were not, at least one divergence among these 10
comparisons would be expected. But their susceptibility to overregulariza-
tion is qualitatively different: in a sample of 40,000 child sentences con-
taining these verbs, Stromswold found that the main verb versions are over-
regularized at rates comparable to those we have found here, whereas the
auxiliary versions of the same verbs were never overregularized. This ex-
treme conservatism was predicted, on the grounds of learnability considera-
tions and other developmental data, by the theory in Pinker (1984), ac-
cording to which the child's language acquisition mechanisms recognize
auxiliary verbs as belonging to a distinguished set of elements because of
certain semantic and phonological properties, and as a result sequester their
lexical representations from regular morphological operations. In other
words, children appear to be capable of treating two verb forms identically
in terms of irregular inflectional patterns but qualitatively differently in
terms of regular inflectional patterns.
Third, Kim, Marcus, et al. (1991) tested children's sensitivity to the
constraint mentioned in Section IC of Chapter VIII that verbs derived from
nouns and adjectives are regular, even if homophonous to an irregular
verb. For example, to high-stick (i.e., "hit with a stick held high") has the past
tense high-sticked, not high-stuck. Recall that the effect is a consequence of
the fact that possessing an irregular past tense form is a property that holds
of verb roots, not verbs in general (unless they have an irregular verb root
as their head); verbs formed from nouns have a noun as their head, not an
irregular verb root; hence, there is no source for an irregular form, and
the default regular rule reliably applies. So in order to predict a verb's past
tense form, its phonological properties do not suffice (nor do its semantic
features; see Kim, Pinker, et al., 1991); its morphological structure must be
input to the relevant mechanisms. The effect has been demonstrated to be
extremely robust in naive adults faced with inflecting novel denominal verbs
(Kim, Pinker, et al., 1991), and Kim, Marcus, et al. (1991) showed it to
hold in 3-8-year-old children as well: children regularize denominal verbs
homophonous with irregulars, such as to fly in the sense of "to cover a piece
of paper with flies," more often than they overregularize the irregular verb

143

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
MONOGRAPHS

itself, even when the irregular verb is used with nonstandard meanings and
hence is equally unfamiliar.
These three phenomena suggest that it is necessary to attribute chil-
dren's regularizations (including overregularizations) to a different mecha
nism than their irregulars, on the basis of the qualitatively different inputs
and outputs the two patterns implicate: the production of irregulars but
not overregularizations feeds compound formation; the production of ir-
regulars but not overregularizations is fed by auxiliary verbs; the production
of regularizations of verbs but not their irregular forms is fed by verbs
derived from nouns. We have shown that exactly this kind of distinction
between memory and rule mechanisms provides a simple account for a
huge variety of facts about the rate, onset, and lexical properties of overreg-
ularization.

144

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
 APPENDIX

TABLE Al

ADAM'S RATES OF OVERREGULARIZATION AND PAST TENSE MARKING

Irreg Irreg Reg Reg

Overrg Overrg Corr Stems Irreg Corr Stems Reg
Corr Stem Past Overrg Oblig Oblig Mark Oblig Oblig Mark
Irreg + ed + ed Rate Cntexts Cntexts Rate Cntexts Cntexts Rate
Age (1) (2) (3) (4) (5) (6) (7) (8) (9) (10)
2-3 ...... 40 0 0 .00 15 15 .50 0 5 0
2-4 ...... 14 0 0 .00 4 11 .27 4 24 .14
2-5 ...... 51 0 0 .00 38 12 .76 1 27 .04
2-6 ...... 90 0 0 .00 54 8 .87 1 16 .06
2-7 ...... 58 0 0 .00 60 11 .85 0 11 0
2-8 ...... 61 0 0 .00 44 16 .73 4 13 .24
2-9 ...... 31 0 0 .00 26 21 .55 0 12 0
2-10 ..... 36 0 0 .00 31 3 .91 0 5 0
2-11 ..... 39 2 0 .05 31 6 .84 9 11 .45
3-0 ...... 74 0 0 .00 47 18 .72 16 11 .59
3-1 ...... 115 0 0 .00 64 2 .97 8 0 1.00
3-2 ...... 103 0 0 .00 51 10 .84 5 7 .42
3-3 ...... 181 0 0 .00 108 10 .92 51 10 .84
3-4 ...... 132 3 0 .02 80 6 .93 15 7 .68
3-5 ...... 124 3 0 .02 95 1 .99 14 2 .88
3-6 ...... 108 1 0 .01 60 0 1.00 14 0 1.00
3-7 ...... 93 1 1 .02
3-8 ...... 133 2 0 .01
3-9 ...... 85 3 0 .03
3-10 ..... 31 1 0 .03
3-11 ..... 76 1 0 .01
4-0 ...... 57 0 0 .00
4-1 ...... 43 1 0 .02
4-2 ...... 59 1 0 .02
4-3 ...... 42 1 0 .02
4-4 ...... 79 4 0 .05
4-5 ...... 47 1 0 .02
4-7 ...... 151 10 1 .07
4-8 ...... 40 2 0 .05
4-9 ...... 55 3 0 .05
4-10 ..... 68 1 2 .04
4-11 ..... 79 0 0 .00
5-2 ...... 51 3 0 .06

NOTE.--The columns correspond to (1) numbe

(e.g., breaked) overregularization tokens; (3) nu
overregularization (proportion of irregular past
irregular past tense forms in obligatory past tense
past tense contexts; (7) irregular
rat past marking
larizations, where an irregular verb was supplied
obligatory past tense contexts; (9) number of st
regular past marking rate (proportion of obligat
tense). The last six columns are based on unpubl
1968) and Brown (1973).

145

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
 TABLE A2

EVE'S RATES OF OVERREGULARIZATION AND PAST TENSE MARKING

Irreg Irreg Reg Reg

Overrg Overrg Corr Stems Irreg Corr Stems Reg
Corr Stem Past Overrg Oblig Oblig Mark Oblig Oblig Mark
Irreg + ed + ed Rate Cntexts Cntexts Rate Cntexts Cntexts Rate
Age (1) (2) (3) (4) (5) (6) (7) (8) (9) (10)
1-6 ...... 2 0 0 0 1 5 .17 2 2 .50
1-7 ...... 5 0 0 0 1 4 .20 0 15 .00
1-8 ...... 13 1 1 .13 13 1 .93 2 8 .20
1-9 ...... 52 0 0 .00 15 13 .54 4 12 .25
1-10 ..... 24 3 0 .11 15 12 .56 7 5 .58
1-11 ..... 34 2 0 .06 25 18 .58 6 5 .55
2-0 ...... 25 0 0 .00 12 19 .39 12 6 .67
2-1 ...... 30 9 0 .23 20 24 .45 25 2 .93
2-2 ...... 60 7 0 .10 51 7 .88 20 1 .95
2-3 ...... 40 1 0 .02 10 4 .71 2 1 .67

NOTE.-For descriptions of columns an

TABLE A3

SARAH'S RATES OF OVERREGULARIZATION AND PAST TENSE MARKING

Irreg Irreg Reg Reg

Overrg Overrg Corr Stems Irreg Corr Stems Reg
Corr Stem Past Overrg Oblig Oblig Mark Oblig Oblig Mark
Irreg + ed + ed Rate Cntexts Cntexts Rate Cntexts Cntexts Rate
Age (1) (2) (3) (4) (5) (6) (7) (8) (9) (10)
2-3 ...... 8 0 0 0 1.5 0 1 0 .5 0
2-4 ...... 19 0 0 0 4.5 8 .36 0 2 0
2-5 ...... 21 0 0 0 8.5 3 .74 1 0 1.00
2-6 ...... 26 0 0 0 16 14 .53 2 1.5 .57
2-7 ...... 57 0 0 0 20.5 13 .61 0 4.5 0
2-8 ...... 20 0 0 0 10 2 .83 0 .5 0
2-9 ...... 80 0 0 0 24 6 .80 9 6 .60
2-10 ..... 25 1 0 .04 7 3 .70 10 4 .71
2-11 ..... 27 0 0 0 10 3.5 .74 1.5 .5 .75
3-0 ...... 46 0 0 0 12 5.5 .69 3.5 1 .78
3-1 ...... 34 0 0 0 35 0 1 4 2.5 .62
3-2 ...... 52 0 0 0 24 2 .92 3.5 2 .64
3-3 ...... 39 1 0 .02 24.5 2 .92 2.5 2 .56
3-4 ...... 20 0 0 0 19.5 1 .95 1 1 .5
3-5 ...... 44 1 0 .02 17 6 .74 3 0 1.00
3-6 ...... 35 1 0 .03 16.5 4.5 .79 2.5 1 .71
3-7 ...... 61 0 0 0 34 4 .89 5.5 .5 .92
3-8 ...... 63 1 0 .02 38.5 4.5 .90 8 2.5 .76
3-9 ...... 27 0 0 0 23.5 0 1 2.5 0 1.00
3-10 ..... 64 3 0 .04 45.5 1 .98 21.5 1 .96
3-11 ..... 67 2 0 .03 42.5 5.5 .89 6 0 1.00
4-0 ...... 30 0 1 .03 21.5 4.5 .83 4 1 .80
4-1 ...... 67 1 0 .01 45 1 .98 13 1 .93
4-2 ...... 122 9 1 .08 50 4 .93 22 0 1.00
4-3 ...... 41 1 0 .02
4-4 ...... 76 1 0 .01
4-5 ...... 54 2 0 .04
4-6 ...... 70 0 0 0
4-7 ...... 71 5 1 .08
4-8 ...... 64 12 0 .16
4-9 ...... 59 7 1 .12
4-10 ..... 56 4 0 .07
4-11 ..... 78 4 0 .05
5-0 ...... 58 2 0 .03
5-1 ...... 36 3 0 .08

NOTE.-For descriptions of columns and data, se

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
 TABLE A4

ABE'S RATES OF OVERREGULARIZATION AND PAST TENSE MARKING

Irreg Reg
Corr Overrg Overrg Overrg Mark Mark
Irreg Stem + ed Past + ed Rate Rate Rate
Age (1) (2) (3) (4) (5) (6)
2-5 ....... 9 2 2 .31 .31 .76
2-6....... 24 1 0 .04 .58 .63
2-7....... 27 16 2 .40 .69 .87
2-8 ....... 52 12 1 .20 .81 .80
2-9 ....... 102 36 6 .29 1 .98
2-10...... 44 34 6 .48 1 1
2-11 ...... 76 31 11 .36 .96 1
3-0....... 24 16 2 .43 1 1
3-1....... 64 19 11 .32 .99 1
3-2....... 82 8 5 .14 .99 1
3-3....... 91 36 9 .33 .97 1
3-4 ....... 107 38 4 .28 1 1
3-5....... 89 31 1 .26 1 1
3-6....... 74 12 1 .15 1 1
3-7....... 58 4 0 .06 1 1
3-8....... 88 16 2 .17 1 1
3-9....... 80 22 2 .23 1 1
3-10 ...... 69 14 2 .19 1 1
3-11 ...... 75 21 3 .24 1 1
4-0....... 13 2 6 .38 1 1
4-1....... 67 19 7 .28 1 1
4-2....... 42 7 4 .21 1 1
4-3....... 39 14 10 .38 1 1
4-4....... 16 10 0 .38 1 1
4-5....... 29 3 0 .09 1 1
4-6....... 68 3 0 .04 1 1
4-7....... 59 3 1 .06 1 1
4-8....... 38 0 0 0 1 1
4-9....... 26 0 0 0 1 1
4-10...... 62 16 0 .21 1 1
4-11...... 47 6 1 .13 1 1
5-0....... 45 13 0 .22 1 1

NoTE.-The columns correspond to

(e.g., breaked) overregularization to
overregularization (proportion of i
marking rate (correct irregular pasts
past tense contexts); and (6) regular
regular pasts and stem forms in ob
(1976; see also Kuczaj, 1977a).

147

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
 TABLE A5

ADAM'S IRREGULAR VERBS

Stem Past Overreg Stem Past Overreg

Verb Correct + ed + ed Rate Verb Correct + ed + ed Rate

be .......... 137 0 0 0 light........ 0 0 0 ..

beat......... 1 2 0 .67 lose......... 50 0 0 0
bend ....... 1 0 0 0 make........ 141 7 0 .05
bite ........ 12 0 0 0 mean........ 0 0 0
bleed ....... 0 0 0 . meet ........ 1 0 0 0
blow ....... 0 1 0 1 put.........69 0 0 0
break....... 97 0 0 0 reada ......

bring........ 17 0 0 0 rend ....... 0 0 0

build ....... 0 0 0 ... ride......... 5 0 0 0
buy ......... 10 0 0 0 ring ....... 0 0 0
catch ....... 73 2 0 .03 rise......... 0 0 0
choose ...... 0 0 0 ... run......... 9 0 1 .10
come ....... 51 1 0 .01 say ........ 129 0 0 0
cut ......... 11 1 0 .08 see ........ 104 0 0 0
dig ........ 0 0 0 ... send ........ 0 0 0
do ......... 93 1 0 .01 shake....... O 0 0
draw ....... 1 4 0 .80 shoot ....... 35 0 0 0
drink........ 0 0 0 ... shrink ...... 0 0 0
drive ....... 1 1 0 .50 shut ........ 0 0 0
eat ......... 21 0 0 0 sing ........ 0 0 0
fall ........ 182 2 0 .02 sit ......... 1 0 0 0
feed......... 0 0 0 ... sleep ........ 0 1 0 1
feel ........ 0 3 0 1 slide ....... 0 0 0
fight....... 0 0 0 ... spend ...... 0 0 0
find......... 60 0 0 0 spin .... .. O0 0 0 .
fly ......... 1 0 0 0 spit......... 0 0 0
forget ...... 27 0 0 0 stand....... 0 1 0 1

freeze ...... 0 0 0 steal ....... 4 0 2 .33

get ........ 586 0 0 0 stick ....... 2 4 0 .67
give ...... . 19 0 0 0 strike ....... 0 0 0 .
go ......... 125 0 0 0 string ...... 0 0 0
grind ...... 0 0 0 . sweep ...... 0 0 0
grow ........ 1 2 0 .67 swim ....... 0 0 0 .
hang........ 0 0 0 .. . swing ...... 0 0
have........ 56 0 0 0 take......... 62 1 1 .03
hear ........ 53 0 0 0 teach ........ 0 1 0 1
hide .. ..... 0 0 0 ... tear......... 1 0 0 0
hit ......... 35 0 0 0 tell ......... 59 0 0 0
hold ....... 0 1 0 1 think........ 33 0 0 0
hurt ....... 28 0 0 0 throw ...... 3 3 0 .50
keep ....... 0 0 0 .. . wake........ 1 1 0 .50
know....... 1 0 0 0 wear....... 0 0 0 .
leave ....... 15 0 0 0 win......... 0 4 0 1
lend .... .. 0 0 0 ... wind........ 0 0
let. ........ 3 0 0 0 write ....... 17 0 0 0
lie. ..... .. 0 0 0 ...

a Analyzed only for Abe.

148

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
 TABLE A6

EVE'S IRREGULAR VERBS

Stem Past Overreg Stem Past Overreg

Verb Correct + ed + ed Rate Verb Correct + ed + ed Rate

be ......... 22 0 0 0
light .. .... O 0 0
beat......... 0 0 0 .. . lose........ 3 0 0 0
bend ........ 0 0 0 ... make ....... 7 0 0 0
bite ........ 7 0 0 0
mean........ 0 0 .
bleed ........ 0 0 0 ... meet ....... 0 0 0
blow ........ 0 0 0 put ........ 38 0 0 0
break ....... 15 0 0 0 reada.......
bring........ 2 0 0 0 rend ........ 0 0 0
build ........ 0 0 0 ... ride......... 0 0 0
buy ......... 3 0 0 0 ring ....... 0 0 0 ...
catch ........ 6 0 0 0 rise ........ 0 0 0
choose ...... 0 0 0 ... run........ 0 0 0
come ....... 23 4 0 .15 say ........ 10 0 0 0
cut ........ 0 0 . see ........ 3 1 0 .25
dig ........ O 0 0 send ........ 1 0 0 0
do ......... 7 2 0 .22 shake....... 0 0 0
draw........ 0 0 0 ... shoot....... 0 0 0
drink........ 0 1 0 1 shrink ....... 0 0 0
drive....... 0 0 ... shut ....... 1 0 0 0
eat .......... 0 0 0 sing ....... O 0 0
fall ........ 2 8 0 .80 sit ......... 4 0 0 0
feed ........ O 0 0 sleep........ 0 0 0
feel ......... 0 0 0 slide ....... 0 0 0
fight ....... 0 0 0 spend ...... 0 0 0
find ........ 1 0 0 0 spin ....... 0 0 0
fly.......... 0 0 0 ... spit ........ 0 0 0
forget ...... 18 0 0 0 stand ....... 0 0 0
freeze ...... 0 0 0 steal ....... 0 0 0
get ........ 55 0 0 0 stick ... ..... 1 0 0 0
give ....... 3 0 0 0 strike........ 0 0 0
go ......... 20 5 0 .20 string ...... 0 0 0
grind ...... 0 0 0 sweep ...... 0 0 0
grow ........ 0 0 0 swim ....... 0 0 0
hang........ O 0 0 swing ...... 0 0 0
have........ 20 0 0 0 take......... 0 0
hear ....... 0 0 0 .. teach ....... 0 0 0
hide ....... 0 0 0 ...
tear......... 0 0 1 1
hit ......... 1 0 0 0 tell ........ 2 0 0 0
hold ....... 0 0 think........ 1 0 0 0
hurt ........ 5 0 0 0 throw ....... 0 1 0 1
keep ........ 0 0 0 .. wake........ 0 0
know........ 0 0 0 ... wear........ 0 1 0 1
leave ........ 1 0 0 0 win ........ 0 0 0 ...
lend ....... 0 0 0 wind........ 0 0
let ......... 0 0 ... write ....... 1 0 0 0
lie.......... 0 0 0

a Analyzed only for Abe.

149

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
 TABLE A7

SARAH'S IRREGULAR VERBS

Stem Past Overreg Stem Past Overreg

Verb Correct + ed + ed Rate Verb Correct + ed + ed Rate

be ......... 182 0 0 0 light ....... 2 0 0 0

beat......... 0 0 0 lose ........ 19 1 0 .05
bend ....... 0 1 0 1 make....... 52 8 0 .13
bite......... 11 0 0 0 mean....... 1 0 0 0
bleed ....... 0 0 ... meet ....... 2 0 0 0
blow ....... 1 0 0 0
put......... 55 1 0 .02
break ....... 96 0 0 0 reada .......
bring........ 9 0 0 0 rend ....... 0 0 0 .
build ....... 0 0 0 ... ride. ...... 0 0 0 .
buy ......... 18 3 0 .14 ring ....... 0 0 0 ...
catch ........ 31 1 1 .06
rise .... ...... 0 0 0
choose ...... 0 0 0 . run......... 4 2 0 .33
come ........ 15 3 1 .21 say ........ 101 0 0 0
cut ........ 6 0 0 0 see ......... 59 0 0 0
dig ........ . 0 0 0 ... send ....... 0 0 0
do ......... 127 0 0 0 shake....... 2 0 0 0
draw........ 0 0 0 shoot.......O 0 0
drink........ 4 0 0 0 shrink...... 1 0 0 0
drive....... 0 0 ... shut ........ 3 0 0 0
eat ......... 14 0 0 0 sing ........ 1 0 0 0
fall ........ 10 1 0 .09 sit ......... 0 1 0 1
feed........ 0 0 0 ... sleep ....... 2 0 0 0
feel ........ 0 0 0 slide ....... 0 3 0 1
fight ........ 0 2 0 1 spend ...... 0 0 0
find......... 27 0 0 0 spin ....... 2 0 0 0
fly ......... 2 1 2 .6 spit ........ 0 0 0 .
forget ....... 30 0 0 0 stand........ 0 0 0
freeze ...... 1 0 0 0 steal ....... 0 2 0 1
get ........ 434 0 0 0 stick ... ..... 1 0 0 0
give ....... 33 0 0 0 strike ........ 3 0 0 0
go ......... 66 6 0 .08 string........ 0 0 0
grind ...... 0 0 0 . sweep ...... 0 0 0
grow ....... 1 1 0 .50 swim ....... 0 1 0 1
hang........ 0 1 0 1 swing ...... 0 0 0
have........ 137 0 0 0 take........ 35 2 0 .05
hear ....... 10 5 0 .29 teach ....... 5 0 0 0
hide ....... 2 1 0 .33 tear......... 0 0 0
hit ......... 27 0 0 0 tell ......... 25 1 0 .04
hold ....... 0 0 0 ... think ........ 10 0 0 0
hurt ....... 9 1 0 .10 throw ...... 3 7 0 .70
keep ....... 1 0 0 0 wake ....... 3 2 0 .40
know........ 5 0 0 0 wear....... 1 0 0 0
leave........ 8 0 0 0
win ........ 4 3 0 .43
lend ........ 1 0 0 0 wind....... 1 0 0 0
let. ....... . 1 0 0 0 write ....... 1 0 0 0
lie......... . 0 0 0 ...

a Analyzed only for Abe.

150

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
 TABLE A8

ABE'S IRREGULAR VERBS

Stem Past Overreg Stem Past Overreg

Verb Correct + ed + ed Rate Verb Correct + ed + ed Rate

bea . ....... light........ 0 0 0

beat........ 0 0 0 ... lose ........ 8 1 0 .11
bend ....... 1 1 0 .50 make ....... 150 24 3 .15
bite ........ 3 4 0 .57 mean....... 3 1 0 .25
bleed ....... 0 0 0 ... meet ......... 0 0 0
blow ....... 4 9 0 .69 put......... 77 11 0 .12
break ....... 38 4 12 .30 read ........ 1 1 0 .50
bring........ 2 5 1 .75 rend .......... 0 0 0
build ....... 1 8 0 .89 ride......... 2 1 0 .33
buy ........ 7 5 0 .42 ring ........ 0 0 0 ...
catch ....... 22 5 1 .21 rise......... 0 0 0
choose ...... 1 1 0 .50 run ........ 4 6 0 .60
come ....... 20 52 4 .74 say ........ 282 3 0 .01
cut ........ 15 12 0 .44 see ........ 166 4 3 .04
dig ........ 2 3 0 .60 send ....... 8 1 0 .11
doa ........ shake........ 2 0 0 0
draw ....... 1 6 0 .86
shoot........ 10 2 1 .23
drink........ 5 5 0 .50 shrink ......... 0 0 0
drive........ 0 2 0 1 shut ....... 0 0 0
eat ........ 82 18 2 .20 sing ........ 2 1 0 .33
fall........ 72 54 3 .44 sit ......... 1 1 0 .50
feed......... 0 1 0 1 sleep ........ 1 2 0 .67
feel ......... 5 11 0 .69 slide ....... 0 0 0
fight......... 1 2 0 .67 spend ...... 0 1 0 1
find......... 143 1 2 .02 spin ........ 0 0 0
fly ......... 5 4 0 .44 spit......... 0 2 1 1
forget ...... 67 0 0 0 stand....... 1 2 0 .67
freeze ...... 1 0 0 0 steal ....... 3 0 0 0
get ........ 194 4 50 .22 stick ....... 1 1 0 .50
give ....... 5 2 0 .29 strike ....... 0 0 0 ...
go ......... 117 60 4 .35 string ..... 0 1 0 1
grind 0...... 0 0 . sweep ...... 0 2 0 1
grow ....... 2 6 0 .75 swim ....... 0 2 0 1
hang........ 0 4 0 1 swing ...... 0 3 0 1
havea ...... take ........ 21 6 3 .30
hear ....... 13 27 0 .68 teach ....... 3 0 0 0
hide ......... 2 0 0 0 tear......... 0 0 0
hit ......... 16 8 0 .33 tell ........ 52 7 0 .12
hold ....... 0 4 0 1 think........ 75 9 9 .19
hurt ....... 16 2 0 .11 throw ...... 5 12 0 .71
keep ......... 1 0 0 0 wake....... 3 1 0 .25
know........ 11 6 0 .35 wear........ 0 3 0 1
leave ....... 13 2 0 .13 win ........ 16 9 0 .36
lend ....... 0 0 0 . . wind........ 0 0 0
let ......... 0 0 ...
write ........ 1 7 0 .88
lie..... .... 0 0 0 ...

a Analyzed only for Adam, Eve, and Sarah.

151

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
 04 . : ? " .: . . . . 0 . 0 ": "0000 "
N

z -? .. . .o -? . . .o .o
S 0.0 ...0 .....0 . - ... . 0 . 0 .... .0 0 -

S. . . . . . . . .

. . 0 . .0. . . .. -

. .c..k.b. ..;. . . .. .. . . . . .
0 . . ... . -.

- - Z. C.. 0 . . . ......0
0 0
- --C " -Q 0 Q "- -%Q0 %Q ' " -0 -0 0 0 "

152

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
 :O : - :00 : : .0 :
.*..... . . . ....... . .*.*

o . . . . *00

..0 . .0 . . . . . . .

0 0 00~ 00
000 0~ 00....
0 o .......

.-. . . . . . . . . . .

...0 . ....

0 . . .0

0.....0 0 ... ...

o o o o

0. 0 . .. 00 0 . .0...000 .
0 ? .0.
? ? ? ?. .
. . .
...O.
. . ... . ..
. . . ... . .
. . . ?

0 . 0 00. . . . .. ... 0 . .o

* 00 .

* 0 . .00 . . 0I

0 . ; . .

?C~ 91

153

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
 O 0 ' ' 0'' o0 ' " 0o
N

0 . . . . .

Y o o oo oo

S. .

S oo
o , .- . .o .oo
. o
. C
...

1 o . .ooo. .

< o o o o o
. . 6
H

(d ?O

Z "
E - - ? ? . ? ' ? ?

13... >

C. C. . . . .. . . .

Z o. ? oo

154

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
 TABLE A10

AGGREGATE MEASURES OF VERBS'

OVERREGULARIZATION RATES AND FREQUENCIES IN PARENTAL SPEECH
Mean Mean z Mean Mean z
Standardized Converted Standardized Converted
Overregu- Back to Log Overregu- Back to Log
larization Overregu- Mean larization Overregu- Mean
Rate larization Parental Rate larization Parental
Verb (z score) Rate Frequency Verb (z score) Rate Frequency

be ........ - .48 .03 5.81 light....... ... ... 0

beat ...... ... ... .70 lose ...... -.40 .04 2.39
bend ...... .... . . . ... .34 make ..... .30 .16 3.26
bite........ -.55 .01 3.10 mean ..... ... ... 1.45
bleed ...... .... . meet ...... ... ... 1.39
blow ...... 1.66 .40 1.20 put ....... . -.55 .01 6.27
break ..... .20 .14 2.32 read ...... ... ... 4.08
bring ..... -.43 .03 2.11 rend ........
build ..... ... ... .29 ride ...... ... ... .45
buy........ .16 .14 2.19 ring ...... .....
catch ..... .04 .12 1.58 rise........ ... ... 0
choose. ....... . run ...... 2.51 .54 2.08
come...... .35 .17 2.88 say ....... -.31 .06 3.95
cut ....... 1.81 .42 4.09 see ....... -.56 .01 3.11
dig .... ... . ...... send ...... ... ... .92
do ........ - .25 .07 6.39 shake ..... ... ... .92
draw ..... .. ... .75 shoot ...... -.42 .04 1.28
drink ..... .80 .25 1.10 shrink. ....
drive ..... . . . ... 1.34 shut ...... ... ... 3.02
eat....... -.50 .02 2.24 sing....... ... ... .51
fall ....... .70 .23 2.65 sit ....... ... .. . 1.58
feed ...... ... .... 0 sleep...... ... ... .92
feel........ 1.64 .39 .61 slide....... .. ... 0
fight ...... ... ... 1.25 spend ..... ... ... 1.10
find ...... - .20 .07 2.25 spin .........
fly ...... . . ... . .62 spit ........
forget ..... -.60 .01 2.30 stand ..... .. . ... -1.10
freeze ...... . ... 0 steal...... ... ... .59
get . ....... -.19 .08 4.80 stick ...... - .06 .10 2.32
give ...... - .47 .03 2.70 strike ..... . . . .
go ....... .27 .16 3.69 string ..... ..
grind ..... ........ sweep ..... ... ... -.69
grow ...... ... ... 0 swim...... ... ... -1.39
hang ..... ... ... 1.39 swing .....
have ...... -.48 .03 5.10 take ...... .30 .16 3.08
hear ...... .66 .22 2.15 teach ..... ... ... 1.16
hide ...... ... ... .69 tear ...... ... ... .92
hit ....... -.30 .06 3.93 tell ....... -.58 .01 2.97
hold ...... ... ... .59 think ..... -.51 .02 3.60
hurt ...... -.40 .04 4.07 throw ..... 3.06 .64 1.44
keep ...... ... . .. 1.22 wake...... ... ... 1.10
know ..... .14 .13 2.14 wear...... ... . ... .85
leave ..... - .63 .00 2.75 win ...... .17 .14 1.64
lend ...... ....... wind....... ... ... 0
let ....... ... ... 5.19 write...... -.43 .03 1.23
lie ....... ...... .98

155

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
 TABLE All

REGULAR VERBS AVAILABLE TO CHILDREN

A. REGULAR VERBS USED AT LEAST ONCE BY

ADAM, EVE, SARAH, OR THE ADULTS CONVERSING WITH THEM

ache, act, add, aim, allow, answer, ask, attach, back, bake, balance, bang, bark, bash, believe,
belong, blast, bless, blind, blink, bob, boil, bomb, bop, borrow, bother, bounce, bow, bowl, box,
breathe, brush, bubble, bump, burn, burp, bury, bust, call, camp, care, carry, carve, chain, chan
chase, cheat, check, chew, chirp, choke, chop, clap, claw, clean, climb, clip, close, color, comb, co
cool, copy, cough, count, cover, crack, crash, crawl, criss-cross, cross, crush, cry, curl, dance, dar
dash, decorate, dial, die, dip, dirty, disappear, disturb, dodge, drag, dress, dribble, drill, drip,
drool, drop, drown, dry, dump, dust, empty, end, erase, excuse, exercise, faint, fan, fasten, fetch,
file, fill, finish, fish, fix, fizz, flash, flip, float, fold, follow, fool, frighten, fry, glue, grab, growl,
guess, guide, hammer, hand, handle, happen, hatch, hate, heat, help, holler, hook, hop, hope,
howl, hug, hunt, hurry, hush, invite, iron, itch, jabber, joke, juggle, jump, kick, kid, kill, kiss,
knock, land, last, laugh, leak, lean, learn, lick, lift, like, listen, live, load, lock, look, love, mail,
manage, march, marry, mash, match, matter, measure, melt, meow, mess, mind, miss, mix, mock,
move, murder, name, nap, need, obey, open, organize, own, pack, paint, pardon, park, pass, paste
pat, pay, pee, pee-pee, peek, peel, peep, perch, pick, pin, pinch, place, plant, play, plug, point,
poke, polish, pop, pour, pout, practice, pray, prepare, press, pretend, print, promise, pull, pump,
punch, punish, push, race, rain, raise, rake, reach, reattach, recognize, remember, remind, repair
rest, rinse, rip, rock, roll, rope, row, rub, ruin, rustle, sail, salute, save, scare, scoop, scoot, score,
scratch, scream, screw, scribble, seem, serve, sew, share, sharp, sharpen, shave, shop, shout, shove
shovel, show, shrug, sigh, sip, skate, ski, skip, slap, slip, smack, smart, smash, smell, smile, smoke
smooth, snap, sneeze, sniff, snow, snuggle, soak, sound, spank, spell, spill, splash, spoil, spray,
squash, squeak, squeal, squeeze, squirt, squish, squoosh, stab, stamp, staple, stare, start, starve, sta
steam, steer, step, stir, stop, straighten, stretch, study, suck, surprise, swallow, sway, swish, switch,
talk, tangle, tape, taste, tease, thank, thread, tick, tickle, tie, tighten, tip, touch, trap, trick, trim,
trip, trust, try, tuck, turn, twist, type, unbuckle, unbutton, unchain, unplug, unscrew, untangle,
untie, urinate, use, visit, wag, wait, walk, want, wash, watch, wave, wee-wee, weigh, wet, whirl,
whisper, whistle, wiggle, wink, wipe, wish, wobble, wonder, work, worry, wrap, wreck, yawn, yel
zip, zoom

B. REGULAR VERBS USED BY ADAM

IN THE TRANSCRIPTS PRECEDING HIS FIRST OVERREGULARIZATION

ask, attach, bake, belong, bless, bow, burn, call, care, carry, change, check, chew, climb, close,
cook, copy, count, crack, crawl, cross, cry, dance, drip, drop, dry, dump, excuse, exercise, fasten,
finish, fish, fix, fold, frighten, growl, hand, happen, help, hop, hug, hurry, jabber, jump, kick, kiss,
knock, laugh, like, listen, live, look, love, march, mess, miss, mix, move, need, open, pack, paint,
park, pay, peek, pick, pinch, play, plug, point, pour, press, pretend, pull, push, rain, reach,
remember, rock, roll, rope, save, screw, shine, shop, show, skip, slip, smoke, snow, spank, spill,
squeal, squeak, squeeze, stay, step, stir, stop, talk, taste, thank, tickle, tie, touch, try, turn, use,
wait, walk, want, wash, watch, whisper, wipe, wonder, work

C. REGULAR VERBS USED BY EVE

IN THE TRANSCRIPTS PRECEDING HER FIRST OVERREGULARIZATION

answer, bang, change, climb, crack, cry, dance, drop, empty, fix, happen, help, jump, kick, lick,
like, look, move, open, pardon, pick, play, show, spill, step, taste, thank, touch, turn, use, wait,
want, watch, wipe

D. REGULAR VERBS USED BY SARAH

IN THE TRANSCRIPTS PRECEDING HER FIRST OVERREGULARIZATION

bounce, bump, call, clean, close, comb, count, cover, cross, cry, dance, die, dress, dump, fix,
happen, help, hug, jump, knock, laugh, lift, like, look, love, need, open, organize, paint, park, pat,
peek, pick, pinch, play, poke, pour, pray, press, pull, push, rain, rip, rock, roll, shop, show, smoke,
spank, stay, step, stop, talk, taste, tease, thank, tie, touch, turn, use, wait, walk, want, wash,
watch, wee-wee, work, yawn

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
 REFERENCES

Adams, S. (1938). Analysis of verb forms in the speech of young children, and their
relation to the language learning process. Journal of Experimental Education, 7,
141-144.

Anderson, J. R. (1983). The architecture of cognition. Cambridge, MA: Harvard

Press.

Anisfeld, M., Barlow, J., & Frail, C. M. (1968). Distinctive features in the pluralization
rules of English speakers. Language and Speech, 11, 31-37.
Anisfeld, M., & Gordon, M. (1968). On the psychophonological structure of English in-
flectional rules. Journal of Verbal Learning and Verbal Behavior, 7, 973-979.
Anisfeld, M., & Tucker, G. R. (1967). The English pluralization rules of six-year-old
children. Child Development, 38, 1201-1217.
Aronoff, M. (1976). Word formation in generative grammar. Cambridge, MA: MIT Press.
Bateman, W. G. (1916). The language status of three children at the same ages. Pedagogical
Seminary, 23, 211-240.
Berko, J. (1958). The child's learning of English morphology. Word, 14, 150-177.
Bever, T. G. (1975). Psychologically real grammar emerges because of its role in language
acquisition. In D. Dato (Ed.), Developmental psycholinguistics: Theory and applications
(Georgetown University Round Table on Languages and Linguistics). Washington,
DC: Georgetown University Press.
Bever, T. G. (Ed.). (1982). Regression in mental development: Basic processes and mechanisms.
Hillsdale, NJ: Erlbaum.
Bloom, L. (1973). One word at a time: The use of single word utterances before syntax. The
Hague: Mouton.
Bohannon, J. N., & Marquis, A. L. (1977). Children's control of adult speech. Child Develop-
ment, 48, 1002-1008.
Bowerman, M. (1982a). Reorganizational processes in lexical and syntactic development.
In E. Wanner & L. R. Gleitman (Eds.), Language acquisition: The state of the art. New
York: Cambridge University Press.
Bowerman, M. (1982b). Starting to talk worse: Clues to language acquisition from chil-
dren's late errors. In S. Strauss (Ed.), U-shaped behavioral growth. New York: Academic.
Bowerman, M. (1987). Commentary: Mechanisms of language acquisition. In B. Mac-
Whinney (Ed.), Mechanisms of language acquisition. Hillsdale, NJ: Erlbaum.
Brown, R. (1973). A first language: The early stages. Cambridge, MA: Harvard University
Press.

Brown, R. (n.d.). Grammars for Adam, Eve and Sarah at Stages I-V. Unpublished manu-
scripts, Harvard University, Department of Psychology.
Brown, R., & Bellugi, U. (1964). Three processes in the child's acquisition of syntax. In
E. Lenneberg (Ed.), New directions in the study of language. Cambridge, MA: MIT Press.

157

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
MONOGRAPHS

Brown, R., Cazden, C., & de Villiers, J. (1971). Directions for scoring fourteen English mor
phemes obligatory in some contexts. Unpublished manuscript, Harvard University, Depart-
ment of Psychology.
Brown, R., & Hanlon, C. (1970). Derivational complexity and order of acquisition in child
speech. In J. R. Hayes (Ed.), Cognition and the development of language. New York
Wiley.
Brown, R., & McNeill, D. (1966). The "tip of the tongue" phenomenon. Journal of Verbal
Learning and Verbal Behavior, 5, 325-337.
Bryant, B., & Anisfeld, M. (1969). Feedback versus no-feedback in testing children's
knowledge of English pluralization rules. Journal of Experimental Child Psychology, 8,
250-255.

Burnham, K. P., & Overton, W. S. (1978). Estimation of the size of a closed p

when capture probabilities vary among animals. Biometrika, 65, 625-633.
Burnham, K. P., & Overton, W. S. (1979). Robust estimation of population siz
capture probabilities vary among animals. Ecology, 60, 927-936.
Bybee, J. L. (1985). Morphology: A study of the relation between meaning and form. Phil
Benjamins.
Bybee, J. L., & Slobin, D. I. (1982a). Rules and schemes in the development and use of
the English past tense. Language, 58, 265-289.
Bybee, J. L., & Slobin, D. I. (1982b). Why small children cannot change language on their
own: Suggestions from the English past tense. In A. Ahlqvist (Ed.), Papers from the 5th
International Conference on Historical Linguistics (Current Issues in Linguistic Theory,
Vol. 21, Amsterdam Studies in the Theory and History of Linguistic Science IV).
Philadelphia: Benjamins.
Carlton, L. E. (1947). Anomalous preterite and past participle forms in the oral language
of average fourth grade children. American Speech, 22, 40-45.
Carroll, J. B. (1961). Language development in children. In S. Saporta (Ed.), Psycholinguis-
tics: A book of readings. New York: Holt, Rinehart & Winston.
Carstairs, A. (1987). Allomorphy in inflexion. London: Croom Helm.
Cazden, C. B. (1966). Verb inflections. Unpublished manuscript, Harvard University, Grad-
uate School of Education.

Cazden, C. B. (1968). The acquisition of noun and verb inflections. Child Development
433-448.

Cazden, C. B. (1972). Child language and education. New York: Holt, Rinehart &
Chamberlain, A. F. (1906). Preterite forms, etc., in the language of English
children. Modern Language Notes, 21, 42-44.
Chao, A. (1987). Estimating the population size for capture-recapture data wi
capturability. Biometrics, 43, 783-791.
Chomsky, N. (1959). A review of B. F. Skinner's Verbal behavior. Language, 3, 2
Chomsky, N., & Halle, M. (1968). The sound pattern of English. New York: Harp
Clark, E. V. (1982). The young word maker: A case study of innovation in t
lexicon. In E. Wanner & L. R. Gleitman (Eds.), Language acquisition: The st
art. New York: Cambridge University Press.
Clark, E. V. (1987). The principle of contrast: A constraint on language acqu
B. MacWhinney (Ed.), Mechanisms of language acquisition. Hillsdale, NJ: Erl
Clark, E. V. (1988). On the logic of contrast. Journal of Child Language, 15, 317
Clark, E. V. (1990). On the pragmatics of contrast. Journal of Child Language, 1
Coltheart, M., Patterson, K. E., & Marshall, J. C. (Eds.). (1980). Deep dyslexi
Routledge & Kegan Paul.
Cox, M. V. (1989). Children's over-regularization of nouns and verbs. Journ
Language, 16, 203-206.

158

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
 MARCUS ET AL.

Derwing, B., & Baker, W. (1979). Recent research on the acquisition of English morphol-
ogy. In P. Fletcher & M. Garman (Eds.), Language acquisition. New York: Cambridge
University Press.
de Villiers, J. G., & de Villiers, P. A. (1973). A cross-sectional study of the acquisition
of grammatical morphemes in child speech. Journal of Psycholinguistic Research, 2,
267-278.

Edwards, M. L. (1970). An annotated bibliography on the acquisition of Englis

morphology. Ohio State University Working Papers in Linguistics, 4, 149-164.
Egedi, D. M., & Sproat, R. W. (1991). Connectionist networks and natural langua
phology. Unpublished manuscript, AT&T Bell Laboratories, Linguistics R
Department, Murray Hill, NJ.
Ervin, S. M. (1964). Imitation and structural change in children's language.
Lenneberg (Ed.), New directions in the study of language. Cambridge, M
Press.

Ervin, S. M., & Miller, W. R. (1963). Language development. In H. W. Stevenson (Ed.),

Child psychology: The sixty-second yearbook of the National Society for the Study of Education
(Pt. 1). Chicago: University of Chicago Press.
Fowler, C. A., Napps, S. E., & Feldman, L. (1985). Relations among regular and irregular
morphologically related words in the lexicon as related by repetition priming. Memory
and Cognition, 13, 241-255.
Francis, N., & Kucera, H. (1982). Frequency analysis of English usage: Lexicon and grammar.
Boston: Houghton Mifflin.
Gasser, M., & Lee, C.-D. (1991). A short-term memory architecture for the learning of
morphophonemic rules. In R. Lippmann, J. Moody, & D. Touretzky (Eds.), Advances
in neural information processing systems 3. San Mateo, CA: Morgan Kaufmann.
Gathercole, V. (1979). Birdies like birdseed the bester than buns: A study of relational comparatives
and their acquisition. Unpublished doctoral dissertation, University of Kansas.
Gathercole, V. (1987). The contrastive hypothesis for the acquisition of word meaning: A
reconsideration of the theory. Journal of Child Language, 14, 493-531.
Gathercole, V. (1989). Contrast: A semantic constraint? Journal of Child Language, 16,
685-702.

Gelman, D. (1986, December 15). The mouths of babes. Newsweek, pp. 84-86.
Gleason, J. B. (1980). The acquisition of social speech and politeness formu
H. A. R. Giles (Ed.), Language: Social psychological perspectives. Oxford: Pergam
Gold, E. M. (1967). Language identification in the limit. Information and Cont
447-474.

Gordon, P. (1985). Level-ordering in lexical development. Cognition, 21, 73-93.

Gordon, P. (1990). Learnability and feedback. Developmental Psychology, 26, 217-
Grimshaw, J., & Pinker, S. (1989). Positive and negative evidence in language ac
[Commentary on D. Lightfoot's "The child's trigger experience: 'Degree-0' le
ity"]. Behavioral and Brain Sciences, 12, 341.
Grimshaw, J., & Rosen, S. T. (1990). Knowledge and obedience: The developmen
of the binding theory. Linguistic Inquiry, 21, 187-222.
Grudin, J., & Norman, D. A. (1991). Language evolution and human-computer
tion. Proceedings of the 1991 annual meeting of the Cognitive Science Society. Hills
Erlbaum.

Guillaume, P. (1927). Le d6veloppement des l66ments formels dans le langage de l'enfant.

Journal de Psychologie, 24, 203-229. (Translated as "The development of formal el
ments in the child's speech." In C. A. Ferguson & D. I. Slobin [Eds.], Studies of chil
language development. New York: Holt, Rinehart & Winston, 1983)
Hall, W. S., Nagy, W. E., & Linn, R. (1984). Spoken words. Hillsdale, NJ: Erlbaum.

159

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
 MONOGRAPHS

Halle, M., & Mohanan, K. P. (1985). Segmental phonology of modern English. Linguist
Inquiry, 16, 57-116.
Hart, J. T. (1965). Memory and the feeling-of-knowing experience. Journal of Education
Psychology, 56, 208-216.
Higginson, R. (1985). Fixing-assimilation in language acquisition. Unpublished doctoral di
sertation, Washington State University, Department of Anthropology.
Horn, E. (1925). Appropriate materials for instruction in reading. In G. M. Whipple (Ed.
Twenty-fourth yearbook of the National Society for the Study of Education (Pt. 1) (Report of
the National Committee on Reading). Bloomington, IL: Public School Publishing Co
Horn, M. D. (1927). The thousand and three words most frequently used by kindergarte
children. Childhood Education, 3, 118-122.
Ingram, D. (1989). First language acquisition: Method, description, and explanation. New York
Cambridge University Press.
Jacobs, R. A., Jordan, M. I., & Barto, A. G. (1991). Task decomposition through competi-
tion in a modular connectionist architecture: The what and where vision tasks. Cogni
tive Science, 15, 219-250.
Karmiloff-Smith, A., & Inhelder, B. (1974-1975). If you want to get ahead, get a theory
Cognition, 3, 195-212.
Kaye, J. (1990). Phonology: A cognitive view. Hillsdale, NJ: Erlbaum.
Kim, J. J., Marcus, G. F., Hollander, M., & Pinker, S. (1991). Children's inflection is
sensitive to morphological structure. Papers and Reports in Child Language, 30, 39-46
Kim, J. J., Pinker, S., Prince, A., & Prasada, S. (1991). Why no mere mortal has ever flow
out to center field. Cognitive Science, 15, 173-218.
Kiparsky, P. (1982). Explanation in phonology. Dordrecht: Foris.
Kruschke, J. K. (1990). ALCOVE: A connectionist model of category learning. Researc
Report No. 19. Bloomington: Cognitive Science Program, Indiana University.
Kucera, H., & Francis, N. (1967). Computational analysis of present-day American Englis
Providence, RI: Brown University Press.
Kuczaj, S. A. (1976). "-ing," "-s," & "-ed": A study of the acquisition of certain verb inflection
Unpublished doctoral dissertation, University of Minnesota, Department of Psy
chology.
Kuczaj, S. A. (1977a). The acquisition of regular and irregular past tense forms. Journal
of Verbal Learning and Verbal Behavior, 16, 589-600.
Kuczaj, S. A. (1977b, March). Old and new forms, old and new meanings: The form-function
hypothesis revisited. Paper presented at the biennial meeting of the Society for Research
in Child Development, New Orleans.
Kuczaj, S. A. (1978). Children's judgments of grammatical and ungrammatical irregular
past tense verbs. Child Development, 49, 319-326.
Kuczaj, S. A. (1981). More on children's initial failure to relate specific acquisitions.Journal
of Child Language, 8, 485-487.
Lachter, J., & Bever, T. G. (1988). The relation between linguistic structure and associative
theories of language learning-a constructive critique of some connectionist learning
models. Cognition, 28, 195-247.
Lenneberg, E. H. (1964). A biological perspective on language. In E. H. Lenneberg (Ed.),
New directions in the study of language. Cambridge, MA: MIT Press.
Lieber, R. (1980). On the organization of the lexicon. Unpublished doctoral dissertation,
Massachusetts Institute of Technology, Department of Linguistics and Philosophy.
Lorge, I., & Chall, J. (1963). Estimating the size of vocabularies of children and adults:
An analysis of methodological issues. Journal of Experimental Education, 32, 147-157.
MacKay, D. (1976). On the retrieval and lexical structure of verbs. Journal of Verbal Learn-
ing and Verbal Behavior, 15, 169-182.

160

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
 MARCUS ET AL.

MacWhinney, B. (1978). Processing a first language: The acquisition of morphophonol-

ogy. Monographs of the Society for Research in Child Development, 43(1-2, Serial No. 174).
MacWhinney, B. (1987). The competition model. In B. MacWhinney (Ed.), Mechanisms of
language acquisition. Hillsdale, NJ: Erlbaum.
MacWhinney, B. (1990). The CHILDES Project: Computational Tools for Analyzing Talk (Ver-
sion 0.88). Pittsburgh, PA: Carnegie-Mellon University, Department of Psychology.
MacWhinney, B., & Leinbach, J. (1991). Implementations are not conceptualizations: Re-
vising the verb learning model. Cognition, 40, 121-157.
MacWhinney, B., & Snow, C. E. (1985). The Child Language Data Exchange System.
Journal of Child Language, 12, 271-296.
MacWhinney, B., & Snow, C. (1990). The Child Language Data Exchange System: An
update. Journal of Child Language, 17, 457-472.
Maratsos, M. P. (1983). Some current issues in the study of the acquisition of grammar.
In J. H. Flavell & E. M. Markman (Eds.), P. H. Mussen (Series Ed.), Handbook of child
psychology: Vol. 3. Cognitive development. New York: Wiley.
Maratsos, M. P. (1987, January 2). Gift of tongues. Times Literary Supplement, p. 19.
Marchman, V. (1988). Rules and regularities in the acquisition of the English past tense.
Center for Research on Language Newsletter (University of California, San Diego), 2(4).
Marchman, V., & Bates, E. (1991). Vocabulary size and composition as predictors of morphological
development (Tech. Rep. No. 9103). San Diego: University of California, Center for
Research in Language.
Marcus, G. F. (1992). Negative evidence in language acquisition. Manuscript submitted for
publication.
Marcus, G. F. (in preparation). Children's overregularization of English plurals: A quantitative
analysis.
McClelland, J. L. (1988). Connectionist models and psychological evidence. Journal of
Memory and Language, 27, 107-123.
McClelland, J. L., Rumelhart, D. E., & Hinton, G. E. (1986). The appeal of parallel
distributed processing. In D. E. Rumelhart, J. L. McClelland, & the PDP Research
Group, Parallel distributed processing: Explorations in the microstructure of cognition: Vol.
1. Foundations. Cambridge, MA: Bradford Books/MIT Press.
McNeill, D. (1966). Developmental psycholinguistics. In F. Smith & G. Miller (Eds.), The
genesis of language. Cambridge, MA: MIT Press.
Menyuk, P. (1963). A preliminary evaluation of grammatical capacity in children. Journal
of Verbal Learning and Verbal Behavior, 2, 429-439.
Miller, G. A. (1977). Spontaneous apprentices-children and language. New York: Seabury.
Miller, W. R., & Ervin, S. M. (1964). The development of grammar in child language. In
U. Bellugi & R. Brown (Eds.), The acquisition of language. Monographs of the Society for
Research in Child Development, 29(1, Serial No. 92).
Moe, A. J., Hopkins, C. J., & Rush, R. T. (1982). The vocabulary of first-grade children.
Springfield, IL: Thomas.
Morgan, J. L., & Travis, L. L. (1989). Limits on negative information in language learning.
Journal of Child Language, 16, 531-552.
Osherson, D. N., Stob, M., & Weinstein, S. (1985). Systems that learn. Cambridge, MA: MIT
Press.

Otis, D. L., Burnham, K. P., White, G. C., & Anderson, D. R. (1978). Statistical inference
from capture data on closed animal populations. Wildlife Monographs, 63, 1-135.
Patel, V. L., & Groen, G. J. (1991). The general and specific nature of medical expertise:
A critical look. In A. Ericsson & J. Smith (Eds.), Toward a general theory of expertise:
Prospects and limits. New York: Cambridge University Press.
Pinker, S. (1979). Formal models of language learning. Cognition, 7, 217-283.

161

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
MONOGRAPHS

Pinker, S. (1982). A theory of the acquisition of lexical interpretive grammars. In J.

Bresnan (Ed.), The mental representation of grammatical relations. Cambridge, MA: MIT
Press.

Pinker, S. (1984). Language learnability and language development. Cambridge, MA: Harvard
University Press.
Pinker, S. (1989). Learnability and cognition: The acquisition of argument structure. Cambridge,
MA: MIT Press.

Pinker, S. (1991). Rules of language. Science, 253, 530-535.

Pinker, S., Lebeaux, D. S., & Frost, L. A. (1987). Productivity and constr
acquisition of the passive. Cognition, 26, 195-267.
Pinker, S., & Prince, A. (1988). On language and connectionism: Analysis o
distributed processing model of language acquisition. Cognition, 28, 73-1
Pinker, S., & Prince, A. (1991). Regular and irregular morphology and the p
status of rules of grammar. In Proceedings of the 17th annual meeting of t
Linguistics Society. Berkeley, CA: Berkeley Linguistics Society.
Plunkett, K., & Marchman, V. (1990). From rote learning to system building (T
9020). San Diego: University of California, Center for Research in Lang
Plunkett, K., & Marchman, V. (1991). U-shaped learning and frequency effect
layered perceptron: Implications for child language acquisition. Cogniti
102.

Prasada, S., & Pinker, S. (in press). Generalizations of regular and irregular m
patterns. Language and Cognitive Processes.
Prasada, S., Pinker, S., & Snyder, W. (1990, November). Some evidence that irr
are retrieved from memory but regular forms are rule-generated. Paper pres
annual meeting of the Psychonomic Society, New Orleans.
Prince, A., & Pinker, S. (1988). Rules and connections in human languag
Neuroscience, 11, 195-202.
Pyles, T., & Algeo, J. (1982). The origins and development of the English langu
New York: Harcourt Brace Jovanovich.
Reich, P. A. (1986). Language development. Englewood Cliffs, NJ: Prentice-Ha
Rosenthal, R. (1984). Meta-analytic procedures for social research. Newbury Par
Rumelhart, D. E., Hinton, G. E., & Williams, D. (1986). Learning internal repr
by error propagation. In D. E. Rumelhart, J. L. McClelland, & the PD
Group, Parallel distributed processing: Explorations in the microstructure of co
1. Foundations. Cambridge, MA: Bradford Books/MIT Press.
Rumelhart, D. E., & McClelland, J. L. (1986). On learning the past tenses of En
In J. L. McClelland, D. E. Rumelhart, & the PDP Research Group, Parallel d
processing: Explorations in the microstructure of cognition: Vol. 2. Psychological an
models. Cambridge, MA: Bradford Books/MIT Press.
Rumelhart, D. E., & McClelland, J. L. (1987). Learning the past tenses of En
Implicit rules or parallel distributed processing? In B. MacWhinney (Ed.),
of language acquisition. Hillsdale, NJ: Erlbaum.
Sachs, J. (1983). Talking about there and then: The emergence of displaced
parent-child discourse. In K. E. Nelson (Ed.), Children's language (Vol. 4)
NJ: Erlbaum.
Sampson, G. (1987, June 12). A turning point in linguistics. Times Literary Su
643.

Scalettar, R., & Zee, A. (1988). Emergence of grandmother memory in feed forward
networks: Learning with noise and forgetfulness. In D. Waltz & J. A. Feldman (Eds.),
Connectionist models and their implications: Readings from cognitive science. Norwood, NJ:
Ablex.

162

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
 MARCUS ET AL.

Seashore, R. H., & Eckerson, L. D. (1940). The measurement of individual differences in

general English vocabularies. Journal of Educational Psychology, 31, 14-38.
Seber, G. A. F. (1986). A review of estimating animal abundance. Biometrics, 42, 267-
292.

Slamecka, N. J. (1969). Testing for associative storage in multitrial free recall. Journal of
Experimental Psychology, 81, 557-560.
Slobin, D. I. (1971). On the learning of morphological rules: A reply to Palermo and
Eberhart. In D. I. Slobin (Ed.), The ontogenesis of grammar: A theoretical symposium. New
York: Academic.

Slobin, D. I. (1973). Cognitive prerequisites for the development of grammar. In C. F

son & D. I. Slobin (Eds.), Studies of child language development. New York: Holt, R
hart & Winston.

Slobin, D. I. (1978). A case study of early language awareness. In A. Sinclair, R. J. Jarvell

& W. J. M. Levelt (Eds.), The child's conception of language. New York: Springer.
Slobin, D. I. (1982). Universal and particular in the acquisition of language. In E. Wann
& L. R. Gleitman (Eds.), Language acquisition: The state of the art. Cambridge: Cam-
bridge University Press.
Smith, E. E., Langston, C., & Nisbett, R. E. (in press). The case for rules in reasonin
Cognitive Science.
Smith, M. E. (1933). Grammatical errors in the speech of preschool children. Child Develo
ment, 4, 183-190.
Smith, M. E. (1935). A study of some factors influencing the development of the sentenc
in preschool children. Journal of Genetic Psychology, 46, 182-212.
Smolensky, P. (1988). The proper treatment of connectionism. Behavioral and Brain S
ences, 11, 1-74.
Stanners, R. F., Neiser, J. J., Hernon, W. P., & Hall, R. (1979). Memory representati
for morphologically related words. Journal of Verbal Learning and Verbal Behavior, 1
399-413.

Stemberger, J. P. (1982). Syntactic errors in speech. Journal of Psycholinguistic Rese

313-345.

Stemberger, J. P. (1989). The acquisition of morphology: Analysis of a symbolic model of

acquisition. Unpublished manuscript, University of Minnesota, Department
guistics.
Stemberger, J. P., & MacWhinney, B. (1986). Frequency and the lexical storage of regu-
larly inflected forms. Memory and Cognition, 14, 17-26.
Strauss, S. (Ed.). (1982). U-shaped behavioral growth. New York: Academic.
Stromswold, K. J. (1990). Learnability and the acquisition of auxiliaries. Unpublished doctoral
dissertation, Massachusetts Institute of Technology, Department of Brain and Cogni-
tive Sciences.

Talmy, L. (1985). Lexicalization patterns: Semantic structure in lexical forms. In T. Sho-

pen (Ed.), Language typology and syntactic description: Vol. 3. Grammatical categories and
the lexicon. New York: Cambridge University Press.
Templin, M. (1957). Certain language skills in children. Minneapolis: University of Minnesota
Press.

Ullman, M., & Pinker, S. (1990, October). Why do some verbs not have a single past tense form?
Paper presented at the fifteenth annual Boston University Conference on Language
Development.
Ullman, M., & Pinker, S. (1992). The English past tense system: Native speaker judgments of
regular and irregular forms and the effects of frequency and similarity. Unpublished manu-
script, Massachusetts Institute of Technology, Department of Brain and Cognitive
Sciences.

163

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
MONOGRAPHS

Valian, V. (1991). Syntactic subjects in the early speech of American and Italian children
Cognition, 40, 21-81.
Warren-Leubecker, A. (1982). Sex differences in speech to children. Unpublished master'
thesis, Georgia Institute of Technology.
Wexler, K., & Culicover, P. (1980). Formal principles of language acquisition. Cambridg
MA: MIT Press.

Wurzel, W. U. (1989). Inflectional morphology and naturalness. Dordrecht: Kluw

Zwicky, A. (1970). A double regularity in the acquisition of English verb m
Papers in Linguistics, 3, 411-418.

164

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
 ACKNOWLEDGMENTS

We thank John J. Kim, Sandeep Prasada, Alan Prince, and Kari

Stromswold for many helpful discussions and suggestions; Joan Bybee, E
Clark, Michael Maratsos, and Joseph Stemberger for comments on an ea
lier draft; Roger Brown, Courtney Cazden, Jill de Villiers, Beth Levin, M
chael Maratsos, Dan Slobin, Joseph Stemberger, Karin Stromswold, and
Virginia Valian for sharing data and other information; and Elena D'Agu
tino, Marie Coppola, Hartley Kuhn, and Elliza McGrand for assistance wi
coding, manuscript preparation, and figures. Parts of this Monograph we
presented at the 1989 and 1990 Boston University Conference on Langua
Development. Supported by National Institutes of Health grant HD 1838
and National Science Foundation grant BNS 91-09766 to Pinker and by
grant from the Alfred P. Sloan Foundation to the Massachusetts Institu
of Technology Center for Cognitive Science. Marcus and Ullman are sup
ported by National Defense Science and Engineering graduate fellowship
Xu was supported by a Dana Foundation award from Smith College. Corr
spondence should be sent to Marcus ([email protected]) at E10-109,
MIT, Cambridge MA 02139, or Pinker ([email protected]) at E10-01
MIT, Cambridge MA 02139.

165

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
 COMMENTARY

OVERREGULARIZATION IN THE ACQUISITION OF INFLECTIONAL

MORPHOLOGY: A COMPARISON OF ENGLISH AND GERMAN

Harald Clahsen

In this Commentary, I will summarize results of studies of two inflec-

tional subsystems of German, noun plurals and the formation of participles,
and compare the results with those of Marcus et al. on English past tense
formation. It will be argued that the evidence from both languages con-
verges on one major point: that children's inflectional systems involve quali-
tative differences between regular and irregular morphology. Moreover,
analysis of German indicates that this distinction cannot be reduced to fre-
quency differences.
The main discovery of Marcus et al. is, I think, that overregularization
errors such as *go-ed are much less frequent than previously thought-that
at all ages such errors account for only a small minority of children's irregu-
lar verbs (2.5%). Marcus et al.'s findings are based on sophisticated statistical
and linguistic analyses of a massive amount of data. Adopting the dual-
mechanism model from Pinker and Prince (1992), they propose a straight-
forward explanation of overregularization in language acquisition. Children
(like adults) possess two distinct psychological mechanisms for inflection, a
symbol-manipulating rule system and an associative memory system. Only
when the child fails to retrieve an irregular form from his or her memory
is the regular rule applied, resulting in occasional errors.
However, there is a problem with this conclusion. Marcus et al. make

The research reported here is supported by a grant from the German Science Foun-
dation (grant C197/5-1). Correspondence should be addressed to Harald Clahsen, Allge-
meine Sprachwissenschaft, Universitat Disseldorf, Universitatsstr. 1, 4000 Diisseldorf 1,
FRG; e-mail: [email protected].

166

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
 MARCUS ET AL.

general claims about the occurrence and distribution of overregularizations

in language acquisition, but their sole evidence comes from past tense for-
mation in English child language. Thus, it is unclear whether their results
also hold for other inflectional systems and for languages other than En-
glish.
English inflection has a number of peculiarities. Relative to other lan-
guages, it is inflectionally poor. For example, it has only one productive
noun plural (-s) and a single productive past tense inflection (-ed), whereas
some languages can have several different noun plurals or tense affixes.
Moreover, there are inflectional systems with far greater irregularity. For
example, Bybee (1991) argued that the German plural system has no regu-
lar default. Finally, the regular-versus-irregular distinction is confounded
with frequency in English: the regular past tense ending applies to the
vast majority of verbs, including virtually all the lower-frequency Latinate
vocabulary. Thus, the affix -ed may occur in overregularizations, not because
of a unique psychological process, but simply because children have heard
it used with so many different English verbs. Similarly, the plural affix -s is
used with nearly all nouns in English, and it is therefore not surprising that
English-speaking children use -s in overregularizations (e.g., *mouse-s, *ox-es,
etc.; cf. Gordon, 1985). Again, Bybee (1991) argued that there are no quali-
tative differences between regular and irregular morphology, that the ob-
served differences are due instead to type frequency, that is, the number
of distinct lexical items involved. This view is opposed to Marcus et al.'s
explanation of overregularizations. Evidence from languages with different
vocabulary statistics than English is necessary to decide this controversy.
The data on German provide such evidence.

The Acquisition of English Past Tense Formation

The research reported in this Monograph is part of a larger ongoing

project carried out by Steven Pinker and his colleagues at the Massachusetts
Institute of Technology and other places. In this project, one subsystem
of human language, inflectional morphology, is studied in detail from the
perspectives of a variety of disciplines.
The main body of this Monograph presents a detailed description of the
development of overregularizations in English child language. Overregular-
ization errors are shown to be rare at all ages, and their occurrence is corre-
lated over time with obligatory past tense marking of regular verbs; that is,
overregularizations appear when the child ceases using bare stems to refer
to past events. For months beforehand, all overtly marked irregular past
forms are correct. Second, Marcus et al. found a frequency effect in the distri-
bution of overregularizations: children make these errors more often for

167

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
MONOGRAPHS

verbs that their parents use in the past tense less frequently. Third, Marcus
et al. observed effects of similarity in the distribution of overregularizations
children make fewer overregularization errors for verbs that fall into fami-
lies with more numerous and higher-frequency members (see also Bybee
Slobin, 1982).
Concerning the distribution of errors, it is striking that, of the various
types of potential errors in past tense formation, only one type is productiv
in the data reported by Marcus et al.: overapplications of the regular pas
tense affix -ed to irregular stems (*go-ed, etc.). In contrast, overapplications
of irregular patterns are extremely rare. In the data from Adam, Eve, an
Sarah, for example, there were only four instances in which irregular pa
terns are applied to inappropriate irregular verbs (beat-*bate, beat-*bet, hit-
"*heet, bite-*bat). Overextensions of irregular patterns to regular verbs are
even rarer; Marcus et al. found only one instance of such an irregularization
error, trick-*truck from Adam.'
Fletcher (1991) reported data from a case study of a British child show-
ing that the distribution of past-time forms contrasts in an interesting way
to the one found by Marcus et al. for American children. Fletcher observed
that the British child has two different types of affixes to refer to past
time events, -ed (like the American children) and -en forms, as in *maden
*broughten, *getten, etc. These -en forms appear without auxiliaries and, fre
quently (about 15%-20%), over an extended period of time (roughly from
age 2-5 to age 3-5). Fletcher also noted that there was no identifiable differ-
ence in meaning between the past tense forms and the -en forms. He pro
posed a rather complicated phonological conditioning of the occurrence
of -en. However, the list of verbs to which -en is assigned suggests that
morphological explanation might be more appropriate: -en may be affixed
to irregular verbs but not to regular ones. Thus, forms such as *builden
*cutten, *putten, etc. are documented in the data reported by Fletcher, but
forms such as *walken, *loven, and *saven are not. It will be shown below
that a similar distinction is made by German-speaking children.
Summarizing briefly, I think that there is evidence that both American
and British children distinguish between regular and irregular inflection.
Marcus et al.'s data indicate that the child has two qualitatively different
types of past tense forms: (i) irregular forms that are not productively over-
extended and (ii) the affix -ed, which may occur both with regular and
(inappropriately) with irregular verbs. Fletcher's data suggest that British
children have a separate affix that is restricted to irregular verbs.

SMarcus et al. do not systematically investigate overapplications of irregular patterns.

However, Marcus (personal communication) noted that, in a follow-up investigation, he
and his colleagues found that such errors are indeed far rarer than overregularization
errors.

168

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
 MARCUS ET AL.

Marcus et al. argue that two different psychological mechanisms under-

lie the observed regular-irregular distinctions: rule-based processes allowing
productive on-line concatenation of affixes to stems and memory-based re-
trieval of stored items. They also convincingly argue that both mechanisms
are involved in children's past tense formations in English. Irregular verb
forms can be memorized from parental speech as soon as words of any
kind can be learned-thus the correct irregulars in early development. The
occurrence of overregularizations, however, must await the abstraction of
the English rule from a set of word pairs. Once older children possess the
rule, they apply it in past tense sentences whenever they fail to retrieve
the irregular, resulting in occasional errors. Given that irregular forms are
memorized items, they should be affected by properties of associative mem-
ory such as frequency and similarity, whereas regular forms should not be
so affected. The frequency and similarity effects observed in the distribution
of overregularization errors support these predictions.
However, it could be the case that the observed regular-irregular dis-
tinctions simply reflect peculiarities of past tense formation in English. In
order to rule out this possibility, it is crucial that the findings regarding
English be replicated in different languages and in different inflectional
systems. In the following, I will therefore summarize the results of my
and my colleagues' studies of the acquisition of two inflectional systems of
German: noun plurals and participle formation.

Investigating German Inflectional Morphology

In our project of inflection in German, my colleagues and I are ad-

dressing questions similar to the ones Steven Pinker and his colleagues ad-
dress with respect to English. Our major empirical goal is to discover
regular-versus-irregular distinctions in nominal and verbal morphology of
German. Like Pinker and his colleagues, we include a variety of data: (i)
linguistic analyses of adult language; (ii) adult experimental data; and (iii)
different kinds of language acquisition data, namely, language-unimpaired
monolingual children (LI), specifically-language-impaired children (SLI),
and adult second language learners (L2). Results from our project are pre-
sented in Clahsen, Rothweiler, Woest, and Marcus (1992), Rothweiler and
Clahsen (1991), and Clahsen (1991). In the following, I will first briefly
describe plural formation in German. Then I will summarize our results on
the acquisition of noun plurals in German child language.

Noun Plurals in Adult German

There are five plural affixes in German, -(e)n, -s, -e, -er, and -0, three
of which also allow a variant with an umlaut, for example, der Apfel-die Apfel

169

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
MONOGRAPHS

(the apple/apples), die Kuh-die Kiihe (the cow/cows), and der Wald-die Wiilder
(the forest/forests). The use of these plural affixes (except -s) with specif
nouns is highly arbitrary. There exist preferred combinations of nouns and
plural allomorphs that are determined by gender and/or the morphopho
nological characteristics of the noun; however, the list of exceptions is quite
long (cf., among others, K6pcke, 1988; Mugdan, 1977).
The plural -s is special in several ways. On the one hand, -s is extremely
infrequent. It applies to fewer than 30 common nouns and is thus far less
frequent than all the other plural inflections. On the other hand, there ar
various reasons suggesting that -s is the default plural affix in adult German.
First, the use of -s is not restricted by properties of the stem/root to which
it is assigned. Second, the plural -s is used for names (die Miillers, "the
Mullers"), for newly created expressions such as clippings (Sozis [the clipped
form of Sozialisten], "socialists"), for acronyms (GmbHs [the acronym mean-
ing incorporation], "Inc."), and with borrowed and foreign words (Kiosks
"newsstands"). Third, the plural -s is the sole plural marker that does no
occur inside compounds (*Autosberg, "*cars heap"; *Sozistreffen but Sozialis-
tentreffen [clipped and full forms, respectively], "socialists' meeting").
In Clahsen et al. (1992), we proposed an analysis of German noun
plurals in the framework of lexical morphology (Kiparsky, 1982, 1985).
(In this model, differences between regular and irregular inflection are
handled by assigning morphological [and phonological] rules to a set of
ordered levels that operate successively in the derivation of an inflecte
word: irregular inflection takes place at level 1, compounding at level 2
and regular inflection at level 3.) We argued (in contrast to Bybee, 1991
that the German plural system has a regular (default) affix, namely, the
plural -s, which, according to the level-ordering model, is assigned at level 3.
This ensures that the plural affix -s occurs with nouns that do not yet have
a marked lexical entry for plural. Moreover, because the default plural -s
on level 3, it is unavailable to rules of compounding that apply at level 2
The irregular plural allomorphs are on levels 1 or 2 and are therefore
available for compound formation. In this way, the distribution of plural
inside compounds can be explained (for further details, see Clahsen et al.
1992).

Noun Plurals in German Child Language

Previous studies found that the plural is marked early and that, from
age 3 on, plural markings are used in about 90% of the obligatory
contexts (Park, 1978). It was also found that overregularizations are rar
and that the plural -n is most often used in overregularizations (Schaner
Wolles, 1988; Veit, 1988). This tentatively suggests that German-speaking

170

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
 MARCUS ET AL.

children develop a regular plural form. An additional criterion for distin-

guishing between regular and irregular plural forms in adult German is the
distribution of plural markings inside compounds. With respect to child
German, however, the use of plurals inside compounds has not been investi-
gated in previous studies.
To fill this gap, Clahsen et al. (1992) analyzed plural formation and
plurals inside compounds in the longitudinal data of 20 German-speaking
monolingual children, 19 SLI children, and the extensive corpora from a
language-unimpaired child, Simone. We found that, in the area of noun
plurals, SLI children do not show any specific deficits. This contrasts with
other areas of inflectional morphology, particularly subject-verb agreement,
case marking, and gender, in which German-speaking SLI subjects are se-
verely impaired (cf. Clahsen, 1991).
In our data, overregularizations of plural allomorphs are rare (3%-
10%) throughout the whole period of observation. In overregularizations,
the plural affixes -n and -s are used significantly more often than all other
plural affixes: 17 children use -n in overregularizations and three children
-s. The children's overregularizations of -s or -n are also qualitatively differ-
ent from other errors in plural marking: -s or -n overregularizations are
not sensitive to properties of the stems/roots at which they occur, and -n or
-s may replace any other plural affix. Moreover, -n and -s are used in in-
vented words. In contrast, the other plural allomorphs of German are not
used productively in overregularizations. These observations indicate that
the children have two kinds of plural markings: (i) irregular forms that are
restricted in use and (ii) regular affixes that may replace any other plural
allomorph. The allomorphs -n and, for three children, -s seem to be regular
default plural forms.
With respect to the use of plural markers within compounds, we ob-
served that the plural allomorphs -e, -er, and -n occur in nominal compounds
(e.g., bild-er-buch, "picture book"; hund-e-hiitte, "kennel") and that the plural
-s does not occur within compounds (auto-bahn, "highway"; oma-zeitung,
"grandmother newspaper"). Such compounds are correct adult-like forms.
The most important observation is that the plural -n can be left out
inside compounds, even in cases in which it is required in German; consider
the following examples:

(1 a) *blume-vase (= blume-n-vase, "vase"),

(1b) *bauer-hof (= bauer-n-hof, "farm"),
(Ic) *tanne-baum (= tanne-n-baum, "Christmas tree").

Concerning the omissions of -n inside compounds, three observations are

relevant. First, the only affix omitted from compounds is -n. Second, -n
left out only when it is a true plural marker; for example, in compound

171

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
 MONOGRAPHS

such as rasenmiiher (lawn mower) in which -n is the final consonant of th

root rasen, -n is not left out by the children. Third, what is most striking is
that the children who omitted the plural -n in compounds were generall
the same children who overregularized with -n. We found a statisticall
significant correlation (r = .69, p < .005) between the rates at which diffe
ent children overregularized -n and the rates at which those children omit-
ted -n from compounds, indicating that, the more often children overregu-
larized, the more often they omitted -n in compounds.
We suggested a morphological analysis (in terms of lexical morphology
to account for the observed correlation between plurals inside compound
and overregularizations: that the plural allomorphs -n and -s are regula
affixes (i.e., on level 3) and that the other allomorphs are irregular (i.e., o
level 1). This means that, in case the child does not find a plural entry o
level 1 for a particular noun, that noun either will not be inflected for plura
or will receive a plural marking on level 3 (i.e., with -n or -s). This accounts
for the overregularizations of -n and -s. Because compounding takes plac
at level 2 and -n and -s are level 3 affixes, they are left out in compounds
(For details, see Clahsen et al., 1992.)
Our findings concerning the omission of -n in compounds provide an
example where children make an "error" with respect to the adult languag
but the "error" conforms to subtle principles of grammatical organizatio
(i.e., level ordering). In adult German, the -n plural occurs frequently in
compounds. Thus, there is no adult model for omitting -n inside com-
pounds, and the children go systematically against a well-established inpu
pattern. However, in the level-ordering model, the "error" automatically
follows from the children's misinterpreting -n as a default plural. This indi-
cates that children not only absorb the properties of the adult language but
at the same time construct their own grammars and in doing so obey general
principles of grammatical organization.
A comparison of the results concerning German noun plurals with
those concerning past tense and plural formation in English shows paralle
with respect to the regular-irregular distinction. Similar to what Marcus e
al. (among others) observed for past tense formation in English child lan
guage, we found that, in the noun plurals of German-speaking children,
the default affix occurs in errors while the irregulars do not. Moreover, the
default affix is omitted in nominal compounds, both in child and in adul
German; this conforms to what Gordon (1985) found for English. In con-
trast to English, however, the default plural of German cannot be deter-
mined on the basis of frequency. Thus, the observed qualitative parallel
between English and German are hard to explain in a unitary inflectiona
architecture influenced only by vocabulary statistics. On the other hand
morphological theory (e.g., the level-ordering model) can account for the
regular-irregular distinctions observed in both languages.

172

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
 MARCUS ET AL.

Participle Formation in Adult German

In adult and in child German, simple past tense forms are rarely used
in spoken discourse. Instead, events in the past are expressed through what
is traditionally called the present perfect. This consists of a finite auxiliary
(haben, "to have," or sein, "to be") and nonfinite past participle forms-a
situation similar to that found in English. Therefore, the closest equivalent
to English past tense formation is German participle formation.
As can be seen in the following examples, three patterns of participle
inflection are distinguished in traditional grammars: (i) weak inflection,
which involves -t affixation without stem/root changes (2a); (ii) strong inflec-
tion, that is, -n affixation plus stem changes (2b, 2d); and (iii) mixed inflec-
tion, or -t affixation plus stem changes (2c). In a manner similar to that in
English, ablaut and other changes of the verbal root/stem in German occur
in both simple past tense forms and past participles. Only "weak" verbs
never involve root/stem changes (cf. [2a]). Finally, prefixation with ge- is
prosodically determined: ge- occurs in participles when the verbal root/stem
has stress on the first syllable. Because this is most often the case in German
verbs, the ge- prefix is very frequent:

(2a) kdaufen = gekauft

to buy = bought
(2b) vertre'iben = vertrieb = vertrieben
to expel = expelled = expelled
(2c) rennen = rannte = gerannt
to run = ran = run

(2d) gehen = ging = gegangen

to go = went = gone

Wunderlich (1991) proposed that the distribution of -t and -n in partici-

ples can be accounted for in terms of two rules:

(3a) n; ROOT[ + PART] - [+ PART];

(3b) t; ROOT[ ] - [ + PART].

The -n affixation rule in (3a) is a property-specific ru

only to roots that are lexically marked with the featur
trast to (3a), the left-hand side of the -t affixation rule (3
any restrictions on possible roots. This means that (3b)
any verbal root that is not yet linked to the feature [

173

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
MONOGRAPHS

the rule in (3b) has default status.2 In order to yield the correct distribution,
the order in which the rules in (3) are applied has to be fixed. Otherwise
nothing would rule out participles such as *gegangt in which the default affi
has been (incorrectly) assigned to a strong verb. This order of applicatio
independently follows from the "elsewhere condition" of lexical morphol
ogy (Kiparsky, 1982, 1985). The elsewhere condition is a general constrain
on morphological processes and ensures that property-specific rules have
priority over default rules, which are applied "elsewhere." Thus, (3a) is
applied on all stems/roots with the specific feature [+ PART]; in the re
maining cases, the general rule (3b) applies.
Participle formation of novel or invented verbs confirms the propose
distinction between (3a) and (3b). Consider, for example, the invented ver
faben. There is only one grammatical way of forming a past participle ou
of such a novel verb, the regular gefab-t; stem/root changes (*gefub-t) an
-n affixation (*gefab-en) are ungrammatical. This shows that, when idiosyn-
cratic information on participle roots is not lexically specified as in the case
of novel verbs, the default rule applies.
Crucially, the linguistic differences between -t and -n affixation in Ger-
man past participle formation do not coincide with quantitative differences
in vocabulary statistics. In German, verbs affixed with the participle affix -t
do not outnumber the irregular verbs requiring the participle affix -n.3 This
is in contrast to past tense formation in English, where the default status of
the -ed suffix is confounded with its high frequency.4 With respect to acquisi-
tion, the comparison between the two languages may therefore allow th
qualitative and quantitative differences between regular and irregular sys
tems to be teased apart. Such a comparison will be made in the next section.

2 For the small number of verbs in the so-called mixed class we require an additiona
specification to ensure that, in forming past participles, -t is assigned to the irregular pa
tense root. Otherwise, we would get ungrammatical participles such as *gerennt instead o
gerannt. It is not clear whether this has to be specified in a separate rule (cf. Wunderlich
1991).
S According to Ruoff (1981), 1,000 verb types account for 96% of all verb tokens in
German. The token frequencies are as follows: 47% strong verb tokens, 32% verb tokens o
the mixed class, and 17% weak verb tokens. The type frequencies among these 1,000 ver
are as follows: 502 strong verb types, 50 verb types of the mixed class, and 448 weak ver
types. Thus, in terms of type and token frequencies, verbs requiring the participle affi
-t are similar to verbs requiring the participle affix -n. A second relevant source is Meie
(1964), who counted the 8,000 most frequent word forms in German texts. Among th
1,200 most frequent word forms are 23 strong past participles with the -n affix, 3 partic
ples of the mixed class with the -t affix, and 8 participles of weak verbs with the -t affix
Again, there is no frequency preference for the (regular) -t affix.
4 In Francis & Kucera (1982), there are 26,201 past tense tokens and 2,347 types.
Regular past tense forms make up 59% of the tokens and 91% of the types; irregula
forms are then 41% of the tokens and 9% of the types. Thus, regular forms are clearly
dominant in English, in both type and token frequency.

174

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
 MARCUS ET AL.

Participles in German Child Language

In her research summary of the early diary studies of the acquisition

of German, Mills (1985) quoted some instances of overregularizations of -t
in participles, but she provided little data.
More recently, Rothweiler and Clahsen (1991) investigated spontane-
ous speech samples from 22 monolingual German-speaking children stud-
ied longitudinally, 19 SLI and 3 language unimpaired. As with noun plurals,
we found no qualitative differences between the SLI children and the nor-
mal controls in the area of participle inflection; that is, we found the same
types of errors and similar overregularization rates.
Our most important finding is that only the default -t affixation rule is
overregularized by the children, whereas the property-specific -n affixation
rule is not overextended. The overregularization rates for the -t suffix range
from 0 to 0.18 with a mean of 0.04 (and a standard deviation of 0.05) for
the SLI children and a mean of 0.10 (and a standard deviation of 0.06) for
the language-unimpaired children.5 Notice that, with respect to English past
tense formation, Marcus et al. found similar overregularization rates for the
-ed suffix, ranging from 0 to 0.24, with a mean of 0.04. This is within the
same distribution as the overregularization rates for the participle affix -t
in German, suggesting that children use similar mechanisms to acquire the
two inflectional systems.
The participle affix -t may occur on all kinds of verbs, on weak verbs
and verbs of the mixed class as well as on strong verbs, yielding overregulari-
zation errors such as *genehmt instead of genommen (taken) and *gegebt in-
stead of gegeben (given). In contrast to that, the participle affix -n is restricted
to strong verbs yielding correct adult-like participles. In all the corpora
from the language-unimpaired children, we found only one instance in
which -n is (incorrectly) assigned to a weak verb (*geschlachten instead of
geschlachtet, "slaughtered"); in the data from the SLI children, three in-
stances of this kind were found. Thus, although the irregular participle
affix -n was not productively extended to regular verbs, the regular partici-
ple affix -t was (comparatively often) overapplied to strong verbs.
A parallel (asymmetrical) distribution can be observed in the stem er-
rors. In the data from both the SLI children and the normal controls, there
are no examples of a participle in which an irregular stem pattern has been
extended to a weak verb. For example, in forming a participle of the regular
verb lieben (to love), the child might produce *geloben instead of the correct
form geliebt, in analogy to similar-sounding strong verbs such as fliegen-
geflogen (to fly-flown), bieten-geboten (to offer-offered), fliehen-geflohen (to

"5 We calculated overregularization rates as the number of cases with -t assigned to

strong verbs divided by the total number of past participles of strong verbs.

175

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
MONOGRAPHS

flee-fled), etc. Such errors, however, do not occur in the data. The only
kinds of stem errors we found were regular stems replacing irregular ones,
for example, *gebind instead of gebunden (bounded) and *gewinn instead of
gewonnen (won).
The observed asymmetries in the distribution of errors suggest that
children qualitatively distinguish between regular and irregular inflection.
Specifically, the German data indicate that only default rules such as the -t
affixation rule are overregularized; in contrast, irregularizations, that is,
overextensions of irregular patterns (the affix -n and irregular stems), to
regular verbs are not productive. The same holds for the data on English
past tense formation investigated by Marcus et al. Strikingly, these parallels
between the results obtained from English and those obtained from German
were found despite differences between the two languages in terms of vo-
cabulary frequencies. This shows that the regular-irregular distinction can-
not be reduced to frequency differences.

Conclusion

We found dissociations between regular and irregular inflectional pro-

cesses in three areas of morphology: English past tense formation, German
noun plurals, and participle formation in German. Regular patterns are
overextended to irregulars in children's inflection errors, but not vice versa.
Moreover, linguistic processes such as compounding are sensitive to the
regular-irregular distinction. For example, English and German com-
pounds may contain irregular plurals, but not regulars. The acquisition data
on German and English indicate that this also holds for child language.
These observations have a straightforward linguistic explanation: chil-
dren and adults employ two kinds of morphological processes: (i) specific
processes that are applied only under restricted circumstances and (ii) de-
fault processes that can be applied with no restrictions. The default-specific
distinction is represented in Kiparsky's model of lexical morphology in
terms of different layers in the lexicon and in terms of different kinds of
morphological rules. Specific processes can be restricted either to lexical
items (i.e., suppletion rules) or to properties of stems/roots (e.g., -n affixa-
tion). Suppletion processes are specific to lexical items. Property-specific
rules are sensitive to properties of the stems/roots to which they are as-
signed, for example, -n affixation in German participles (cf. [3a]). Typically,
specific processes are not overapplied. Moreover, the results show that de-
fault processes are positioned strictly after all other processes. Thus, specif-
ically marked forms (e.g., suppletives) do not undergo the default process,
and default rules cannot enter into other processes, such as compounding.

176

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
 MARCUS ET AL.

In contrast to lexical morphology as well as other versions of morpho-

logical theory (cf. Anderson, 1985; Halle, 1990) in which default and spe-
cific inflectional processes are typically described in terms of symbolic rules,
in the dual-mechanism model adopted by Marcus et al. the regular-irregular
distinction is interpreted in terms of two qualitatively different psychological
mechanisms: regular inflections are based on symbolic rules (like in lexical
morphology), but irregular inflections are claimed to be rote-based retrieval
processes of stored items. Marcus et al. provide evidence for the dual-
mechanism approach from past tense formation in English child language,
for example, frequency and similarity effects in irregulars but not in reg-
ulars.

In the German acquisition data, however, we do not (yet) have evidence

that irregular inflection should be handled by a distinct associative memory
system. We found that children (like adults) possess productive inflectional
rules that do not crucially depend on vocabulary frequency, but the possibil-
ity that even irregular inflectional processes are based on symbolic rules
cannot be excluded. Moreover, in its present form, the dual-mechanism
approach leaves some theoretical questions open. For example, what is the
status of property-specific processes such as -n affixation in German partici-
ples? Such processes differ from default rules because they apply only un-
der restricted circumstances; on the other hand, -n affixation is clearly a
productive concatenation of an affix to a stem rather than a rote-based
process. Thus, the dual-mechanism model needs to distinguish between
different types of symbolic rules, and it requires additional principles that
ensure that the rules are applied in the right order. Marcus et al.'s blocking
principle seems to be too narrow in this regard because it is restricted to
idiosyncratic forms listed in the lexicon and thus would not apply to
property-specific rules such as -n affixation. Kiparsky's elsewhere condition
would be more appropriate.
In sum, there is strong evidence that grammatical inflection in the lan-
guage of children and adults involves productive affixation rules. However,
the additional claim of the dual-mechanism model that irregular inflection
is governed by mechanisms of associative memory rather than symbolic
rules requires further theoretical and empirical support before we can
transfer it to other inflectional systems and to languages other than English.

References

Anderson, S. (1985). Phonology in the twentieth century: Theories of rules and theories of represen-
tation. Chicago: University of Chicago Press.
Bybee, J. (1991). Natural morphology: The organization of paradigms and language ac-
quisition. In T. Huebner & C. Ferguson (Eds.), Crosscurrents in second language acquisi-
tion and linguistic theories. Amsterdam: Benjamins.

177

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms
MONOGRAPHS

Bybee, J., & Slobin, D. I. (1982). Rules and schemas in the development and use of English
past tense. Language, 58, 265-289.
Clahsen, H. (1991). Child language and developmental dysphasia: Linguistic studies of the acquisi-
tion of German. Amsterdam: Benjamins.
Clahsen, H., Rothweiler, M., Woest, A., & Marcus, G. F. (1992). Regular and irregular
inflection in the acquisition of German noun plurals. Manuscript submitted for publication.
Fletcher, P. (1991, September). The significance for the cognition hypothesis of an unusual past
time form in an English child. Paper presented at the Symposium for Rick Cromer,
Cambridge.
Francis, N., & Kucera, H. (1982). Frequency analysis of English usage: Lexicon and grammar.
Boston: Houghton Mifflin.
Gordon, P. (1985). Level-ordering in lexical development. Cognition, 21, 73-93.
Halle, M. (1990). An approach to morphology. North-Eastern Linguistic Society, 20, 150-184.
Kiparsky, P. (1982). From cyclic phonology to lexical phonology. In H. V. D. Hulst & N.
Smith (Eds.), The structure of phonological representations (Pt. 1). Dordrecht: Foris.
Kiparsky, P. (1985). Some consequences of lexical phonology. In C. Ewen & J. Anderson
(Eds.), Phonology yearbook (Vol. 2). Cambridge: Cambridge University Press.
K6pcke, K.-M. (1988). Schemas in German plural formation. Lingua, 74, 303-335.
Meier, H. (1964). Deutsche Sprachstatistik. Hildesheim: Olms.
Mills, A. (1985). The acquisition of German. In D. Slobin (Ed.), The cross-linguistic study of
language acquisition. Hillsdale, NJ: Erlbaum.
Mugdan, J. (1977). Flexionsmorphologie und Psycholinguistik. TiAbingen: Narr.
Park, T.-Z. (1978). Plurals in child speech. Journal of Child Language, 5, 237-250.
Pinker, S., & Prince, A. (1992). Regular and irregular morphology and the psychological
status of rules of grammar. In Proceedings of the 17th Annual Meeting of the Berkeley
Linguistics Society. Berkeley, CA: Berkeley Linguistics Society.
Rothweiler, M., & Clahsen, H. (1991, October). Past participles in German child language: A
comparative study of normal and disordered language development. Paper presented at the
Boston University Conference on Language Development.
Ruoff, A. (1981). Hdufigkeitsw6rterbuch gesprochener Sprache. Tiubingen: Niemeyer.
Schaner-Wolles, C. (1988). Plural vs. Komparativerwerb im Deutschen: Von der Diskre-
panz zwischen konzeptueller und morphologischer Entwicklung. In H. Gfinter (Ed.),
Experimentelle Studien zur Flexionsmorphologie. Hamburg: Buske.
Veit, S. (1988). Das Verstandnis von Plural- und Komparativformen bei (entwicklungs)
dysgrammatischen Kindern im Vorschulalter. In G. Kegel et al. (Eds.), Sprechwis-
senschaft und Psycholinguistik. Opladen: Westdeutscher.
Wunderlich, D. (1991). Eine Unterspezifikationsanalyse der deutschen Verbmorphologie.
Unpublished manuscript, University of Diisseldorf.

178

This content downloaded from 103.231.241.108 on Fri, 24 Nov 2017 06:50:12 UTC
All use subject to http://about.jstor.org/terms