Overregularization: Fact and Fiction
Gary F. Marcus & Steven Pinker
Massachusetts Institute of Technology1

Draft 
8/9/90


Running Head: Overregularization




How do children learn the English past tense system?  The past tense of  
regular verbs is  formed  by adding "ed" to the stem,  e.g.. "talk" + "ed" 
becomes "talked", "review" + "ed" becomes "reviewed", etc. But many 
verbs, the irregulars, do not follow this rule. Some irregular verbs undergo 
a vowel change, e.g. "break" becomes "broke",  in others the stem is 
completely replaced: e.g. "go" becomes "went" in the past. Children often 
make "overregularization" errors, while learning the past tense system, in 
which "ed" is added to an irregular stem, producing a form such as 
"breaked" or "goed". Introductory textbooks often present 
overregularization as the paradigm example of rule learning, because the 
child productively creates new forms that she has never heard using a rule. 
Slobin (1971) claimed 
	"Then as soon as the child learns only one or two regular past tense 
forms -- like helped and walked -- he immediately replaces the 
correct irregular past tense forms with their incorrect over-
generalizations from the regular forms. Thus children actually say it 
came off, it broke, and he did it before they say it comed off, it breaked, 
and he doed it.  Even though the correct forms may have been 
practiced for several months, they are driven out of the the child's 
speech by overregularization, and may not return for years. " 
[emphasis added]
In the classical view of past tense acquisition, the child goes through three 
stages.  First, the child learns the past tense of a few verbs by rote, for 
example, the child hears "talked" or "went" and uses that form. The child 
repeats only forms she has heard; the child  simply imitates the parent, 
and cannot create new forms. In the next stage,  the child hears enough 
regular past tense forms to induce the  add "ed" rule.  The child applies this 
new rule not just to regular verbs, but to irregular verbs as well, and thus 
creates incorrect forms such as "goed" or "breaked." Finally,  the child 
eliminates overegularization errors, and uses the past tense as the adult 
does. However, no detailed account of the transition from the second stage 
to final stage has been proposed.
	How does the child learn to eliminate these overregularization 
errors?  A simple hypothesis is  that parents correct their children's errors.  
However, research on parental correction, or negative evidence, by Morgan 
& Travis (1990)  shows children can learn the past tense without parental  
corection. Analyzing the transcripts of three children (Adam, Eve, and 
Sarah), Morgan and Travis reported 
	"Expansions and Clarification Questions occurred more often 
following ill-formed utterances in Adam's and Eve's input, but not in 
Sarah's.  However, these corrective responses formed only a small 
proportion of all adult responses following Adam's and Eve's 
grammatical errors. Moreover, corrective responses appear to drop 
out of the children's input while they continue to make 
overgeneralization errors... the contention that language input 
generally incorporates negative information appears to be 
unfounded."
	One way to explain how a child could unlearn an incorrect form 
without parental feedback is the uniqueness principle (Wexler (1980) and 
Pinker (1984, 19xx)) which states that "... the need for negative evidence in 
language acquisition can be eliminated if the child knows that when he or 
she is faced with a set of alternative structures fulfilling the same function, 
only one of the structures is correct unless there is  direct evidence that 
more than one is necessary. " (pinker 1984:113).
For example, suppose a child hypothesizes that "blue" is a noun. Even 
without negative evidence the child will learn that "blue" is actually an 
adjective, which will block the hypothesis that "blue" is a noun. So the 
uniqueness principle explains how a child might abandon a incorrect 
hypothesis without negative evidence.  
 In past tense acquisition, we might expect the correct irregular form to 
preempt the incorrect overregularized form. But this does not occur;  
Kuczaj (1977) showed that a child may use both  the correct  and incorrect 
forms, even within a single conversation.  For example,  a child might say 
both "I broke it" and "I breaked it". Rumelhart & McClelland (1986) claimed
	"Correct use of irregular forms is never completely absent, and the 
same child may be observed to use the correct past  of an irregular, 
the base+ed form, and the past+ed form, within the same 
conversation."
So uniqueness does not seem to apply to the past tense; this  raises questions 
for the uniqueness principle, and presents a difficult learnability problem, 
since neither uniqueness nor negative evidence can explain how a child 
learns to stop overregularizing.
	Perhaps because of this learnability problem, some researchers have 
proposed an alternate model of past tense acquisition, which does not 
involve rules at all. Rumelhart and McClelland (1986) argue that rather 
than learning a system of rules, a child has a network of nodes and 
connections between those nodes. Learning involves adjusting the weights 
of these connections. This model has been extensively critiqued by Pinker 
and Prince (1988) and Lachter & Bever (1988)). 
	However, there is little quantitative, empirical data to support either 
the PDP model or the classical three stage model of overregularization. 
Through the examination of transcripts of children, we will try to answer 
these questions: When does overregularization first occur, what 
mechanisms govern overregularization, and how and when does a child 
learn to stop overregularizing?

Study I: How frequent is overregularization?

 How frequent is overregularization?  Do children overregularize virtually 
always, as Slobin (1971) has suggested? Half the time? Or only rarely? There 
has been little quantitative documentation of overregularization errors; but 
we must know this information to construct a proper model of 
overregularization and past tense acquistion. One possbility is that 
overregularization errors are very common regularization errors, as Slobin 
(1971) argued. This  would suggest that the errors are caused by a rule, 
which the child applies to all past tense verbs, regular and irregular. The 
major burden of such a model would be to explain how the child learns to 
stop applying the rule to irregulars. A second possibility is that the child 
alternates between the two forms, as Kuczaj (1977) claimed. If this is the 
case, we must solve the learnability problem discussed above. Namely, how 
does the child eliminate an incorrect past tense form from the lexicon? A 
third possibility is that  regularization errors are quite rare. If this is the 
case, errors might be the result occasional processing difficulties rather 
than from a fundamental difference in the language of the child and the 
adult. 
We studied transcripts of spontaneous speech of 9 children in depth (as well 
as briefer transcripts of a number of other children) to determine the 
frequency of overregularization errors.

Method
	Subjects. We studied nine children and almost 13,000 utterances. The 
children ranged in age from 1;3 to 5;0. All data was taken from CHILDES, a 
on-line set of transcripts managed by Macwhinney & Snow (1985).  Table 1 
shows the children, their ages, and the frequency of the transcripts.
 

----------------
Insert Table about Here

Child		Age		Experimenter  	sample freq.
Abe 		2;6 to 5;0  	Kuczaj		2 hours a week
Adam 	2;3 to 4;10  	Brown 
Eve 		1;6 to 2;3	Brown 
Sarah 	2;3 to 5;1	Brown
Allison	1;9 to 3;2 	Bloom  
Nat		2;8		Bohannon
Naomi 	1;3 to 4;9 	Sachs 
Nathaniel 	2;3 to 3;9 	Snow 
April 	1;10 to 2;11	Higginson 

---------------

	Procedures. 
For Abe, the data was gathered by Stan Kuzcaj (1976).  Kuzcaj recorded the 
number of times Abe used each of sixty-six irregular verbs (listed in the 
appendix), the number of times the child produced the present stem of the 
verb with "ed" appended (e.g. "goed" or "breaked") and the number of times 
the child produced  doubly marked pasts (past+ed forms), in which  "ed" 
was added to the irregular past form, e.g. "wented" or "broked"  for each 
month from age 2;6-5;0.  
	Adam, Eve, and Sarah were counted by hand by Hollander . {insert 
her explanation here}.
	For the other 5 children, we ran all analyses on a Sparcstation 4, 
running under unix. We tabulated all irregular past tense utterances 
using a modification of the "freq" program from the MacWhinney and 
Snow programs. This program counts the number of times each word is 
used in a particular transcript session, for a particular speaker. We  
counted the number of occurrences for each word listed in the appendix, 
which includes correct irregular past forms, as well as stem+ed and 
past+ed forms. 
	This procedure was repeated for all of the normal English speaking 
children in the CHILDES database. Thus in addition to the 5 individual 
children, XX databases with small samples from many children were 
counted. Also, Abe's parents use of the past tense was calculated in the 
same way.
	To calculate total overregularization, we simply divided the total past 
tense errors, for the 9 children and the conglomerate data from the other 
databases, by their total past tense uses. We also did this for each child, 
individually.   We calculated the percentage of all past tense utterances that 
included overregularizations.
	Limitations:
	 We gathered these data using simple pattern matching techniques, 
rather than spending thousands of hours  counting by hand. However we 
have had to sacrifice some accuracy.  There are several places we may have 
made small numbers of errors.
	We have not analyzed words individually to make sure they are used 
as past tense verbs. So participle forms are included (e.g.. "the window was 
broked"). The word "seed" presented particular problems since most of its 
uses are as a noun rather than as an overregularized past of  "see." 
Fortunately most of the words we searched did not have this problem.
	A hand analysis of the data of Abe Kuczaj revealed totals that were 
only different by X% percent from the machine generated versions. Since 
this is only a small percentage, we decided that the error was acceptable.
	We have counted repetitions such as "I broked it . I broked it" as two 
OR errors. (We believe that the two utterances are independent since some 
children have been observed to say both a correct and incorrect form in a 
single utterance as in "I broke it. I breaked it.")
	We did not include no-change verbs (hit, cut, put, etc) in our study 
since (a) Kuczaj did not, and (b) it is often impossible to tell from written 
transcripts whether a no-change is used in the past or the present. (Eg. "I 
cut my finger").

	Results.  Overregularization occurred in  only 2.5% of all utterances. 
Table 2 shows that no child OR-d more than 8% of the time, except Abe 
Kuczaj (23%).  

----------
Insert table with each child, or%, here
Child		%OR
Abe
Adam
Eve
Sarah
Allison
Nat
Naomi
Nathaniel
April
-----------
If the child was always using the same rule, overregularization would have 
been much more frequent.  Instead, overregularization is rare; this 
suggests we should explore mechanisms that cause occasional processing 
errors rather than fundamentally different representations of the past 
tense.
We will next try to characterize when the errors are made.
what else can I say here???


Study II: Are all verbs overregularized equally?
	
	Does the child stop overregularizing all verbs at the same time? 
There are three possibilities. First, a child might stop overregularizing all 
verbs at the same time. This would suggest that the child either learns or 
develops the uniqueness principle, and then all the overregularized 
irregular past tense forms are forced out by the correct forms. A second  
possibility is that each verb  follows the same learning curve,  but at the 
same time.  For example, "break" and "ring" might follow the same 
general curve, but "break" might be learned before "ring."  This  would 
suggest that the child  must decide whether each verb is regular or 
irregular. Once the child decides a verb is irregular, the child would cease 
(perhaps gradually) to overregularize that verb. A third possibility is that 
each verb follows a different pattern. This would suggest either that 
overregularization errors are random or that overregularization errors are 
constrained by lexical properties if individual verbs. 
	In addition to examining the learning curves of overregularization, 
we must determine whether  all verbs are regularized with the same 
frequency, or if  some verbs are regularized more frequently than others.  
Again, there are several possibilities. The simplest is that all verbs are 
regularized with the same frequency, ie break, string, ring, and go all are 
overregularized in 2.5% of their uses. This would suggest that  
overregularization is not sensitive to the lexical properties of individual 
verbs.  A second possibility is that different verbs are OR-ed at different 
percentages. E.g. "break" might be 20% while "go" was only 1%. This might 
indicate  that properties of particular verbs affect the likelihood of 
overregularization. One alternate explanation would be that  the child adds 
"ed" to a verb at random intervals. If so we would expect to find a normal 
distribution of errors, with most verbs being regularized at 2.5% but some 
would be regularized more, while others would be overregularized less 
often. This study will explore these two questions: the time course of 
overregularization, and its lexical specificity.

Method
 We graphed the overregularization rate of each verb Abe uttered  for each  
month from 2;6 to 5;0, based on data was taken from Kuzcaj (1976). Abe was 
studied because he was the only child with enough data available to study 
the hypothesis. Also, we  prepared two dependent measures to test the 
tendency of  a particular verbs to overegularize. First, we calculated Abe's 
Overregularization By Verb.   We used Abe's percent overregularization for 
each verb as the dependent measure. (E.g. Abe overregularized break  16 of 
48  past uses = 33%). Second, we calculated the Average Overregularization 
By Verb, Other 8 Children: the individual OR percentages of the other eight 
children were averaged, to form an  overall percentage  for each verb. (E.g. 
the average error for break was 12%).

	Results. Examination of Abe showed that different verbs follow 
different patterns at different times.
------------------------------------
insert two histograms:
 one for Abe, one for All Children
------------------------------------

A Chi-Square (???)  test shows that different verbs are overregularized at 
different percentages, etc.
	Paragraph on criteria for learning, results.
	We found that different verbs were regularized at different 
percentages, and different verbs are learned  correctly at different times. 
This rules out  the possibility that the child learns  to stop overregularizing 
all at once.   In the next study we explore various factors that might 
contribute to overregularization, including a verb's phonology, a verb's 
relation to other verbs which follow a similar stem/past pattern, or its 
frequency of use affect the likelihood of its overregularization.

STUDY III:  What lexical factors affect overregularization?
	 In this study we explore several possible mechanisms that might 
affect the rates at which different verbs are overregularized. These include 
measures of phonological association between stem and past tense form, 
models of distributed memory structure, and positive parental evidence. We 
also consider  the role of processing load in overregularization. In the next 
section, we describe each hypothesis in depth.
Method
	Correlation Studies
As in the last study, we used Abe's overregularization percentages for each 
verb, and the average of the other 8 children, to test the tendency of 
particular verbs to regularize. To test each of the first four hypotheses, we 
correlated these two measures with some other measure, as described 
below.
(1) Number of Phonemes Changed:  If the child's language representation 
were an associative network, phonological information should affect 
overregularization. We suppose a memory model in which the similarity of 
a pair of items controls its ease of recall. Such a model might consist of 
micro-rules where for example a vowel change such as o->e  would be 
harder than no change at all.
We predict that the more different a present tense form is from the past 
tense form, the harder it is to map the stem form to the past tense form, and 
thus a child is the more likely  to overregularize that verb. E.g. bring-
brought should be overregularized more than see-saw.   
We counted the number of phonemes changed between stem and past forms 
for each verb.  Dipthongs were treated as a single phoneme. For example in 
"see-saw" the first phoneme "s" is the same, but the second phoneme is 
different, "ee" versus "aw". So "see/saw" counts as 1 phoneme different, 
whereas "bring/brought" counts as two phonemes different. We tested the 
correlation between the frequency of overregularization and the number of 
phonemes changed.

(2) Number of Phonemes Preserved:  The converse of the last hypothesis is 
that the more similar the stem form and past tense form are the easier it 
should be for the child to associate stem and past tense forms of the same 
verb, if we have a model of memory such as the model sketched for the 
phonemes changed hypothesis, above.  We measured this  phonological 
similarity by calculating the number of phonemes in which a given stem 
form and past tense form differed. (Here diphthongs were counted as two 
phonemes.) An example of this measure is that  feel-felt  should be harder 
than throw-threw because feel/felt counts as 1, while throw/threw  counts 
as 2.  At first glance hypotheses 1 and 2 might appear to be the same, but 
actually thay do make different predictions in some cases. For example  
see/saw and swim/swam have the same amount of phonological change , 
but swim/swam has more material preserved.  We tested the correlation 
between overregularization rates and our measure of phonemes preserved 
between stem form and past tense form.

(3) Cluster strength: In an associative memory model such as the one 
proposed by  Rumelhart and McClleland, formation of the past tense 
depends on phonological regularities. If  the mapping between a present 
tense and past tense forms is similar to the mapping for another pair, the 
mappings will reinforce each other. The more neighbors following the 
same stem/past pattern a verb has,  and the more frequent those neighbors 
are, the stronger the stem/past association should be. For example, 
ring/rang  and drink/drank should make sing/sang easier, because similar 
words share nodes and links between those nodes. 
Several measures of cluster strength were used.   In the first, we used the 
verb taxonomy found in the appendix to Pinker & Prince (1988).Verbs are 
divided intoa heirarchy of  lines, subclasses, classes, and superclasses, 
depending on the the pattern of change from stem form to past tense form. 
The more similar two verbs are in terms of their change from present to 
past, the closer they will be on this heairarchy.   Four scores were 
calculated for each verb. For the first method, frequencies were taken from 
Francis & Kucera (198x). The frequency of all the verbs on the same line 
were summed (except the verb itself). Second,  the frequency of all verbs in 
the same subclass, but different line were summed. Third, the frequency of 
all verbs in the same class, but different subclass were formed. Fourth, the 
frequency of verbs in the same superclass, but not same class, were 
summed. Regressions were calculated with each of these scores, with all 
combinations of predictors, for both of the verb-specific measures of 
overregularization. 
	Ullman (1990) prepared a similar taxonomy. (awaiting his 
explanation for this section). 
	Finally, Ullman (1990) recalculated the second method, using the 
children's' own correct past tense frequencies, rather than the frequencies 
from Francis and Kucera (1986).  With each of these measures of cluster 
strength, we tested the correlation between overregularization and 
phonemes changed.

(4) Parental frequency:  Overregularization might result from a failure to 
retrieve the proper past tense form, thus we predict that increasing the 
strength of a memory trace should decrease the chance of an 
overregularization error.  The more often parents use a past tense form, the 
stronger the child's association between that past  tense form and the 
corresponding present tense form, and the more likely the child is to 
remember that form, and thus the child is less likely to overregularize it.  
We tested the correlation between Abe's overregularization rates with  the 
frequency with which Abe's parents said a given past tense form in our 
transcripts.  We also used Abe's parents as paradigmatic parental speech, 
and thus tested the  correlation between their past tense frequencies amd 
the average overregularization rate of the other 8 children.  Finally, we took 
the past tense counts  from Francis and Kucera's (198?) corpus that is based 
on 1,000,000 words of written text as a measure  of frequency of input. We 
correlated that measure with overregularization rates for both Abe and the 
average of all children.

	 Processing Hypotheses. 
 	We hypothesized that if an utterance is long or complex, processing 
is more likely, due to limited resources, to break down at some step and 
make the child overregularize.  Consider learning to drive a car. While 
none of the component actions, such as signaling, steering, or accelerating 
are difficult by themselves, their combined cognitive demands often result 
in an error in one of the components, such as causing the novice driver  to 
neglect to signal a turn.  So we hypothesized that greater cognitive demands 
are more likely to cause a  child to overregularize. To measure the cognitive 
demands the child faces, we measured three things: sentence length 
counted in words, clause length counted in words, and complexity  
measured as the number of verbs in a sentence.   To  prepare these 
measures , we began by extracting the complete text of every past tense 
utterance from Abe.  From this corpus we extracted utterances where Abe 
used both a correct form and incorrect form in a single session. For 
example, if Abe said both "I fell" and "the doll falled down" in one session 
then  we extracted both utterances, but if the verb was not used in both a 
correct past tense form and an incorrect tense form, we removed it from the 
corpus.  We prepared all processing measures on the basis of this corpus 
which contained all and only within session correct/incorrect past tense 
forms.   We did not remove repeated utterances such as "I broke it. I broke 
it". That uttterance would be counted as two distinct sentences in all three 
measures.
	We then measured the length of each utterance in three ways. First. 
we counted the number of words in each sentence. Second, we divided up 
run-on sentences into  logical sentences. For example, "it belonged on here 
hey you put the thing" we divided into "it belonged on here" and "hey you 
put the thing".  Third, we divided sentences into clauses.  We countedands 
that were used to coordinate sentences  in the beginning of the second 
sentence. For example "Mommy made a big turtle and you made a little 
turtle also"  was divided into "mommy made a big turtle" and "and you 
made a little turtle also."
	To calculate complexity, we counted the number of verbs in each 
sentence by dividing the transcripts into sentences using the second method 
above.  Sentencesa with one verb received a complexity score of one; 
sentences with two verbs such as "I think that daddy came home" received 
a score of two, and so on. Tabulation was done by hand
	For the three measures of length we compared the mean length of 
the correct and overregularized utterances. We determined whether 
sentences containing overregularizations were longer than sentences with 
correct past tense forms using three measures of length,  and whether 
sentences with overregularizations were more complex than sentences 
containing correct past tense uses.
	Results. Table N shows the results for each hypothesis. Phonological 
similarity and difference seemed to have no effect on overregularization. 
None of the 3 phonological measures predicted overregularization either for 
Abe or the aggregate measure of the other children. The number of 
Phonemes changed  from the present form to the past tense form did not 
predict overregularization (rphonemes_Abe = -.20, NS)  rphonemes_kids=-
.08, NS. Similarly, phonemes shared  between present and past form failed 
to predict OR (rphonemes_Abe = .11, NS rphonemes_kids= .07, NS).
	Finally, none of the measures of cluster strength showed a 
singificant correlation either with Abe's overregularization rates or with 
the group average overregularization rates.  The measure calculated from 
the Pinker and Prince appendix failed to predict overregularization, both 
for Abe (r= .07) and for the other children r=.20,NS)  These cluster strength 
scores were calculated using Francis and Kucera frequencies. We also  
found no correlation using the child's own frequencies rather than those 
derived from Francis and Kucera (1982).  Finally, the revised clustering 
prepared by Ullman also failed to predict overregularization (#s here).
The only measure that predicted overregularization was Parental 
frequency   (rabe  = -.37, pabe < .01, rallkids = -.40, pallkids <.005).   The 
direction of the correlation indicates that the more the parents used past 
tense form of a particular verb, the less likely the child was to 
overregularize that verb. These results are consistent with a model in 
which overregularization is caused by memory failure. A child's memory 
trace should be stronger the more often the parent uses an irregular verb in 
the past tense, and thus the child is less likely to have retrieval failure. We 
argue that overegularization is caused by memory failure, and so 
overregularization decreases with increasing parental frequency.
We tested whether overregularizations are speech production errors  
resulting from decreased available processing capacity, by comparing the 
lengths of sentences with correct and overregularized versions of the same 
irregular verbs. No differences were found (mean length sentence with 
correct irregular past = 8.6 words; mean length with overregularized verb = 
7.5) Other measures of sentence complexity also could not predict 
overregularization.   Overregularized forms did not appear more frequently 
in complex sentences, measured in number of clauses (MEANS? t(30) =  
1.38, p = .177)).
In sum, the only factor which significantly predicted tendency  towards 
overregularization was the frequency with which a child hears a given past 
tense form. The results suggest that (1) the lexical factor of frequency of use 
does affect overregularization, so overregularization is not random and (2) 
memory plays an important role in overregularization.
General Discussion
To construct our model of past tense acquisition, we must  first consider 
when children mark past tense.   Kuczaj (1977) has argued that children 
mark past tense obligatorily after the time when  they make their first 
overregularization errors.   Kuczaj concluded that "once the child has 
gained stable control of the regular past tense rule, he will not allow a 
generic verb form to express  "pastness," which eliminates errors such as 
go, eat, and find, but results in errors like goed, eated, and finded, as well 
as wented, ated, and founded." (Kuczaj 1977 p. 593).
	The other preliminary to constructing our model of past tense 
acqusition is to consider how adults process the past tense. Pinker (19??),  
argued that adults have separate systems for regular and irregular past 
tense formation; irregular past tense forms are stored as a list, while 
regular past tense forms are generated by a rule.  Regular past tense forms 
are not storerd in memory, they created on-line. The  regular rule  is a 
default; it is used unless it is preempted by the retrieval of an irregular past 
tense form.
	The low frequency of children's overregularization errors, their 
negative correlation with parental frequency, and the absence of 
correlations with any other factor we have investigated suggest a simple 
model of the cause of overregularization errors. (1) The child marks  past 
tense obligatorily. (2) If the child (or adult) can recall an irregular  past 
tense form, he or she uses it. But if (3) the child fails to retrieve an irregular 
form of a verb stored in memory, then (4) the child  reverts to the default 
regular rule, and produces an overregularized form if the verb is irregular.  
The only difference between children and adults is that   children are more 
likely to forget an irregular  past tense form than adults.  
Overregularization occurs when a child forgets the past tense form of an 
irregular.  Stemberger (personal comm.) has shown that adults also make 
overregularization errors, albeit very rarely. These are probably also caused 
by a temporary failure to retrieve the correct past form. 
	Overregularized forms are not stored in memory,  rather they are 
created on-line when the child cannot recall the irregular past, thus 
overregularized forms need not be explicitly removed from the child's 
language.  We are no longer faced with a difficult learnability problem.  The 
uniqueness principle is not required for acquisition of the past tense; 
because the child never stores two forms.  To recover from these errors, the 
child  must simply improve retrieval of the past tense form. The child can 
learn the correct ofrm through positive evidence. Since the incorrect form is 
never stored, negative evidence is not required to remove it.
 	Our model is consistent with the results reported by Ullman & Pinker 
(in prep.) "Squishy verbs" are verbs  such as dive /?dove/?dived which have 
two acceptable past forms. In a questionnaire study Ullman & Pinker  
asked subjects  to rate the goodness of  two alternate past tense forms for a 
number of "squishy verbs". For example, a subject would rate both the 
regular past tense form "dived" and the irregular past temse form "dove".   
Ullman & Pinker found that  the goodness of the regular past tense form 
was inversely proportional to the goodness  of the irregular past tense form. 
Verbs were only squishy when the irregular form was infrequent.  This is 
consistent with our model of the child, which predicts that the child is more 
likely to produce a regular form when the irregular form is infrequent. 
Thus we  predict that the better the irregular past tense form, the weaker 
the regular form will be.
	Our model is also consistent with two facts about language change. 
First,  we predict that low frequency irregulars are overregularized more 
frequently. In the limit, we predict that low frequency irregulars should 
become regularized all the time, and thus become a regular verb. This can 
be seen with verbs like chide, the past of which used to be chide..  Because 
the past was low frequency, work has now become regularized; the past is 
now worked. (Other examples include stride/strode  now strided; for most 
speakers,  cleave/clove now cleaved, and so on.) 
	Seondly, new verbs in a language are nearly always regular. For 
example, the past of to fax  is faxed.  Berko (1958)  showed that children 
productively apply the regular rule to novel verbs.  Our model predicts that 
a child regularizes a novel form because he or she can't retrieve an 
appropriate irregular past, so uses the regular rule to form a novel past.
	To sum up, when the child cannot recall an irregular past tense 
form, the child marks the past tense form using the default regular rule. 
As  a child improves at retrieving past tense forms, overregularization 
errors become less frequent.  Learning does not depend  on negative 
evidence or on eliminating an incorrect form using the uniqueness 
principle, and no radical restructuring needs to be posited to explain how 
children learn the past tense system in English.



1Acknowledgements, funding, etc.

