Age of Acquisition Effects in Word Reading and Other Tasksingilizceveingilizce.com/ingilizce/Age of Acquisition... · 2013-02-07 · AGE OF ACQUISITION 3 the conditions under which

Journal of Memory and Language 47,1–29 (2002)doi:10.1006/jmla.2001.2834

Age of Acquisition Effects in Word Reading and Other Tasks

Jason D. Zevin

University of Southern California

and

Mark S. Seidenberg

University of Wisconsin–Madison

Recent studies have suggested that age of acquisition (AoA) has an impact on skilled reading independent ofontrolleddies sug-r an AoAties. Anerms ofd, AoAon is notames for

M ance

MeGradevMcsiomewh

ASeicon

factors such as frequency. This result raises questions about previous studies in which AoA was not cand about current theories in which it is not addressed. Analyses of the materials used in previous stugest that the observed AoA effects may have been due to other factors. We also found little evidence foeffect in computational models of reading that used words that exhibit normal spelling–sound regulariAoA effect was observed, however, in a model in which early and late learned words did not overlap in torthography or phonology. The results suggest that with other correlated properties of stimuli controleeffects occur when what is learned about early patterns does not carry over to later ones. This conditicharacteristic of learning spelling–sound mappings but may be relevant to tasks such as learning the nobjects. © 2002 Elsevier Science (USA)

Key Words:age of acquisition; reading; connectionist modeling; linguistic development.

any studies of word reading have examinedof this early learning on adult perform

atie&

aa

s

e in

d-n-ng,n-s-leg-il- aced

tero

how stimulus properties such as frequenlength, spelling–sound consistency, and imaability affect performance (for reviews, seBalota, 1994; Seidenberg, 1995). Over the pseveral years, another factor, age of acquisi(AoA), has drawn considerable attention (Ghand & Barry, 1998, 1999, a, b; Morrison Ellis, 1995). The basic idea is that the agewhich a word is learned in acquiring spoken laguage affects the performance of skilled reers. People learn words such as TOP SYRUP before words such as TAX anSYRAH. As operationalized in recent studiethe AoA hypothesis is that there will be an effe

This research was supported by National Institute

1

is&om).vehat

ntal Health (NIMH) Grant PO1-MH47566, NICHDnt RO1-MH 29891, and an NIMH research scientiselopment award to Mark S. Seidenberg. We thank JClelland and Matt Lambon Ralph for helpful discusns. The models described in this article were implented using software developed by Michael Harmom we also thank.ddress correspondence and reprint requests to Markdenberg, Department of Psychology, University of Wissin, Madison, WI 53706. E-mail: [email protected].

0749-596X/02 $35.00

y,e-est

onr-

atn-d-ndd,

ct

when other factors such as frequency of usagadult language are controlled.

The existence of an AoA effect on word reaing would be consistent with evidence concering other types of age-dependent learni(Doupe & Kuhl, 1999; Quartz & Sejnowski1997). In many cognitive domains, early learing results in a reduction in plasticity that limitthe ability to acquire new information. Phonological acquisition provides a classic examp(Werker & Tees, 1984): Learning the phonoloical structure of one’s language limits the abity to learn new phonetic contrasts (e.g., insecond language). Similarly, there is evidenthat the ability to learn the morphology ansyntax of a language drops monotonically afapproximately 7 years of age, although it controversial (see Flege, Yeni-Komshian,Liu, 1999). Lexical acquisition is not thought tbe highly age dependent (Markson & Bloo1997; McCandliss, Posner, & Givon, 1997Still, it is possible that early learned words haan advantage over later learned words and tthis would carry over to how they are read.

f

tay--,

S.-

© 2002 Elsevier Science (USA)All rights reserved.

E

l bpve

e etonnnd

- &&0

Thn-&n,e

iteoend

skcAnu

iboA9A

toede in

3;n

d

t

ofOnhti-

luse-ousevi- iserehy

onus incyntAthatrlierctsifynd

ndaoutnd

f- ofthemorla-esn-re-n-andndlly,od- forthe ef-nistage

2 ZEVIN AND S

The ages at which people learned particuwords are unknown, of course, but they canestimated from other measures. For examGilhooly and Logie (1980) collected subjectiratings of AoA, familiarity, imageability, andconcreteness for nearly 2000 words. Thnorms have been widely used in studies offects of AoA on several tasks including tachisscopic identification (Lyons, Teer, & Rubestein, 1978), word naming (Brown & Watso1987; Coltheart, Laxon, & Keating, 1988), aobject naming (Carrol & White, 1973; Ellis &Morrison, 1998) and with neurologically impaired patients (Hirsh & Ellis, 1994; HodgsonEllis, 1998; Lambon Ralph, Graham, Ellis Hodges, 1998). The Gilhooly and Logie (198data were obtained from 36 adult subjects. AoA ratings also correlate significantly with idependent measures of AoA (Gilhooly Gilhooly, 1980; Lyons et al. 1978; MorrisoEllis, & Chappell, 1997), suggesting that thprovide reliable information.

Given estimates of the frequencies wwhich words occur in adult usage and whwords were acquired, it seems natural to csider whether the two factors have independeffects on skilled performance. Morrison aEllis (1995) orthogonally manipulated AoA anfrequency in naming and lexical decision taand found a strong AoA effect with frequencontrolled but no frequency effect with Aocontrolled. They also observed that AoA afrequency had been confounded in previostudies, raising the possibility that effects attruted to frequency might have been due to ASubsequent studies (Gerhand & Barry, 191999a,b) replicated Morrison and Ellis’s Aoeffect with frequency controlled, but contrarythe earlier results, significant effects of frquency were observed with AoA controlleNonetheless, the finding that AoA affects pformance independent of frequency seemspresent a challenge for models of word read(e.g., Coltheart, Curtis, Atkins, & Haller, 199Plaut, McClelland, Seidenberg, & Patterso1996; Seidenberg & McClelland, 1989) that not explicitly take this factor into account.

The research described below was motiva

IDENBERG

are

le,

sef--

-,

)e

y

hnn-ntd

sy

ds-.

8,

-.r-tog

,o

ed

led us to examine more closely whether ageacquisition has an effect on skilled reading. the empirical side, the concern was that it migbe difficult to isolate effects of age of acquistion because it is correlated with many stimuproperties including frequency. Below we prsent analyses of the materials used in previstudies and other data suggesting that the dence for an effect of AoA on skilled readingweak at best. On the theoretical side, we winterested in developing a better account of wage of acquisition could have an effect skilled reading or other tasks. Many previostudies have employed a bottom-up strategywhich AoA is treated as a factor, like frequenor length, that might account for independevariance in adult performance. However, Aoneeds to be understood in terms of a theory addresses why some words are learned eathan others and how early experience affelater performance. Such a theory would clarthe relationship between the AoA measure aother factors that affect word learning askilled performance and would provide stronger basis for generating predictions abthe role of age of acquisition in reading aother tasks.

After examining existing studies of AoA efects in reading, we describe investigationsthese effects using a computational model of mapping from orthography to phonology (Har& Seidenberg, 1999). Modeling was useful fseveral reasons. First, it allows direct maniputions of the frequency and timing of exposurto words using stimuli that are exactly cotrolled with respect to properties (such as fquecy and length) that are normally highly cofounded. Second, such models embody explicit theory of reading acquisition and skilleprocessing in which the roles of frequency atiming of exposure can be examined. Finaprevious analyses of the behavior of such mels suggest a possible computational basisage of acquisition effects. In some models,“entrenchment” of early learned items has anfect on later performance (Ellis & LamboRalph, 2000; Munro, 1986). Thus, connectionmodels are consistent with the existence of

Q

cee c

nad-rd

c

iai

o

t

ft

n

wee

gllt

h

ngisi-euchrn-es,

lit-ofop-m-et-rly

us ise- isc-seer

stduli

ofndt-s-

inentry,s:ts

therndi- re-let-ton-

orais-rdiketedrs

AGE OF AC

the conditions under which such effects ocand how they relate to the conditions that govreading. We focused on the mapping betworthography and phonology because it playsimportant role in the naming and lexical desion tasks that have been used to study AoAfects in reading.

To foreshadow the results, the simulatioyielded two complementary findings. Simultions using a large corpus of English woryielded no effects of AoA on skilled performance. There was an initial advantage for wothat were presented more often early in traininbut there was no residual effect on skilled peformance. This occurred because the corretions between orthography and phonology thexist across words in English reduce the effeof early exposure to individual items. These rsults, taken with the analyses of previous bhavioral studies, suggest that age of acquisiteffects in word reading are likely to be minimwith other properties that are correlated wAoA controlled. However, a significant age oacquisition effect was observed in a simulatiin which early and late learned words were chsen so that they overlapped little in terms of othographic or phonological structure. This arficial condition, which is not characteristic oreading acquisition, yielded an advantageearly learned words in skilled performance wiother factors controlled.

The simulations suggest that the occurreof age of acquisition effects depends on the ture of the learning task, specifically whethwhat is learned about one pattern carries oveothers with which it shares structure. Thus,observed the effect in a simulation using matals that explicitly eliminated the overlap btween early and later learned patterns, but when the stimulus patterns exhibited the relarities in the correspondences between speland sound that are characteristic of the Engwriting system. This analysis also extends to simulations reported by Ellis and LamboRalph (2000), Smith, Cottrell, and Anderso(2001), and Monaghan and Ellis (2002), wobserved robust age of acquisition effects usmaterials and tasks that differ from reading

UISITION 3

urrnenani-ef-

s-s

s

both the modeling and the analysis of existibehavioral studies suggest that age of acqution has little impact on skilled reading. At thsame time, the modeling also suggests that seffects may occur for other tasks, such as leaing the names associated with objects or facfor which the learning of one pattern carries tle information about others. The full range effects can be explained in terms of basic prerties of learning in connectionist networks eploying distributed representations. Such nworks provide deeper insight about how ea

e-e-onl

thfno-r-i-forh

cena-err toe

ri--

notu-ingishhenno

ingin

experience affects later performance.

PREVIOUS STUDIES

Two strategies have been used in previostudies of AoA effects in word reading. Oneto conduct experiments in which AoA and frquency are manipulated factorially. The otherto use multiple regression to show that AoA acounts for unique variance in predicting reponlatencies or proportions of errors. We considthese in turn.

Morrison and Ellis (1995) conducted the firexperiments factorially manipulating AoA anfrequency in word reading tasks. Their stimwere equated across conditions in terms mean Kuc˘era and Francis (1967) frequency aother variables (e.g., imageability, length in leters, the N measure [Coltheart, Davelaar, Jonason, & Besner, 1977]) but varied significantly terms of rated AoA. This study and subsequones using similar methods (Gerhand & Bar1998, 1999a,b; Monaghan & Ellis, in presTurner, Valentine, & Ellis, 1998) yielded effecof AoA with such stimuli.

These studies raise concerns about whestimulus frequencies were equated across cotions as the designs of these experimentsquired. Properties of words such as length in ters are objective and, therefore, easy manipulate or control across conditions. In cotrast, the frequency counts derived from corpsuch as Kuc˘era and Francis (1967) are stattics—estimates of a variable (how often a wois used) whose actual values are unknown. Lother statistics, frequency counts are associawith measurement error arising from facto

SE

vie

etie

n,nf-ass

n,sXs

-iti

)saslyon:nd

r-it aess-e-eatg

anf-

ntted

esr-n

4 ZEVIN AND

texts used in generating the corpus, and indiual differences in language experience. Thsources of error can complicate the interprtion of frequency effects in behavioral stud(Gernsbacher, 1984).

One problem is that the widely used Browcorpus (from which the Kuc˘era & Francis, 1967norms are derived) is relatively small, which itroduces considerable error in the estimatesindividual words, particulary in the lower frequency range. Table 1 provides frequency dfor the stimuli used in previous age of acquition studies derived from Kuc˘era and Franci(1967) and two other sources: theEducator’sWord Frequency Guide(WFG [Zeno,1995]) andCELEX (Baayen, Piepenbrock, & van Rij1993) databases. Whereas the Brown corpuabout 1 million words, the WFG and CELEcorpora both are more than 16 million wordThe data also include a measure of rated famiarity (Gilhooly & Logie, 1980), which Gernsbacher (1984) showed provides a more sens

available for most inconsistent items in Monaghan and El†p , .10, *p , .05, **p , .01, *** p , .001.

IDENBERG

d-sea-s

-or

tai-

is

.il-

ve

frequency words. Morrison and Ellis (1995equated their early and late AoA stimuli in termof Kuc̆era and Francis (1967) frequency, butTable 1 indicates, the items differ significanton the other measures in the expected directiEarly acquired words are also more frequent afamiliar. The early and late stimuli in the Gehand and Barry (1998, 1999a,b) studies exhibsimilar pattern; there are numerical differencbetween the early and late stimuli on all meaures, and they are significant using log WFG frquency and familiarity. The materials in thTurner et al. (1998) study also differ such thearly words were higher in frequency (loCELEX, log WFG) and rated familiarity thanwere late words. In a recent study, Monaghand Ellis (2002) examined age of acquisition efects for words with consistent or inconsistespelling–sound correspondences. They equathe stimuli with respect to frequency estimatderived from both the Brown and CELEX copora. The stimuli in the inconsistent conditio

ly

s onh

measure of frequency differences among lowerexhibit small differences in the direction of ear

TABLE 1

Properties of the Stimuli Used in Previous Studies of Effects of Age of Acquisition and Frequency

Study Condition KF log(KF) CELEX log(CELEX) WFG log(WFG) FAM

Morrison & Ellis (1995) Early 23 2.63 512 5.78 477 5.62 5.62Late 24 2.63 301 4.82 107 3.32 4.10Difference 21 0.00 211 0.96** 370** 2.30*** 1.52***

Gerhand & Barry Early 105 3.01 1986 5.91 2164 5.41 5.35(1998,1999a,b) Late 75 3.15 881 5.50 306 3.61 4.62

Difference 30 20.14 1105 0.41 1858† 1.80* 0.73**

Turner et al. (1998) Early 52 3.24 555 5.51 2184 6.90 5.69Late 50 2.86 309 4.63 1274 6.13 4.97Difference 2 0.38 246 0.88** 910 0.77* 0.72***

Monaghan Ellis (in press): Early 35 2.63 654 5.56 411 5.20 NAInconsistent words Late 25 2.30 420 4.88 141 3.36 NA

Difference 10 0.33 234 0.68 270* 1.84** NA

Monaghan Ellis (in press): Early 33 2.14 672 4.97 469 4.31 4.97Consistent words Late 29 2.07 496 4.93 199 3.76 4.55

Difference 4 0.07 176 0.03 270 0.65 0.42

Note. In all cases, stimuli were matched using Kuc˘era and Francis (1967). Turner et al. (1998) also matched their itemspoken frequencies from Baayen et al. (1993). WFG, Zeno (1995); KF, Kuc˘era & Francis (1967); CELEX, written Englisfrequencies from Baayen et al. (1993); FAM, familiarity from Gilhooly and Logie (1980); NA,5 familiarity ratings not

lis (2002).

CQ

ec

maa

a

dembe

usue oteres

llis yteid

s ofionei-

s of30lota06tti, la-30la-re-

FGcere-rnif-

t isvi-re- ofof

edn-

AGE OF A

words being higher in frequency on all thrmeasures; using the WFG norms, the differenare statistically reliable. For the consistent itethe differences between the conditionssmaller and nonsignificant on all three meures.The consistent condition is the only oneTable 1 in which an age of acquisition effect wnot obtained.

These cases are similar to the ones studieGernsbacher (1984), who showed that sevapparently conflicting findings in the conteporary word recognition literature could traced to the relative insensitivity of the Kuc˘eraand Francis (1967) frequency norms; stimthat apparently were equated on this meadiffered in terms of rated familiarity. In thstudies in Table 1, stimuli that were equatedthe Kuc̆era and Francis norms differed in rafamiliarity and/or another measure of fquency based on a larger corpus. The incontent word condition in the Monaghan and E(2002) study is the least clear case insofar astimuli did not differ reliably on two frequencmeasures but did on a third. It should be nohowever, that the WFG norms appear to prov

isci-f-he

re-thate-e-ndb- ef-5)ichm- &

an

-

m

*p , .05, **p , .01, *** p , .001.

UISITION 5

eess,res-ins

byral-

lire

nd-is-sthe

d,es-

ents the correlations among several measurefrequency and the naming and lexical decislatencies in three large-scale studies. The Sdenberg and Waters (1989) data set consistmean-naming latencies for 3000 words from undergraduate subjects; the Spieler and Ba(1997) data are naming latencies for 29words from 31 subjects; and the Balota, Piloand Cortese (2001) data are lexical decisiontencies for 2905 words from 60 subjects (young adults and 30 older adults). The corretions between estimated frequencies and sponse latencies are highest for the Wnorms, which also account for unique varianwhen entered into a simultaneous multiple gression with the other norms. Below we retuto methodological issues about the use of dferent frequency norms. Here the main pointhat the early and late acquired stimuli in preous studies were not closely matched in fquency and, thus, did not provide strong teststhe role of age of acquisition independent this factor.1

Some of the studies in Table 1 also includconditions in which age of acquisition was cotrolled and frequency varied, which yieldedmixed pattern of results. Morrison and Ell(1995) found a frequency effect in lexical desion but not in naming; age of acquisition efects, in contrast, were found in both tasks. Tfact that there was an AoA effect but not a fquency effect in the naming task suggested the AoA effect could not be due wholly to a frquency confound. However, this pattern of rsults did not replicate in a study by Gerhand aBarry (1998) using the same stimuli; they oserved both frequency and age of acquisitionfects in naming. The Morrison and Ellis (199data also exhibited an atypical pattern in whlexical decision latencies were faster than naing latencies for the same words (cf. Balota

d

Chumbley, 1984; Forster & Chambers, 1973).

1 Another bit of evidence that the age of acquisition effectreported by Monaghan and Ellis (2002) was due to differ-ences in frequency was reported by Strain, Patterson, andSeidenberg (2002), who found that using frequency countsderived from either the CELEX or WFG database as a co-variate in the analyses of variance eliminated the age of ac-

TABLE 2

Various Frequency Measures as Predictors of Naming Lexical Decision Latency in Large-Scale Studies

Study Measure r Unique variance

Spieler & Balota WFG 2.35 2.39***(1997) FAM 2.32 0.82*

NAMING CELEX 2.29 0.12KF 2.27 0.03

Seidenberg & WFG 2.23 0.72*Waters (1989) FAM 2.21 0.22

NAMING CELEX 2.21 0.11KF 2.18 0.27

Balota et al. WFG 2.63 3.97***(2001) FAM 2.62 3.86***

LEXICAL CELEX 2.58 0.22DECISION KF 2.51 0.80**

Note. WFG, word frequency from Zeno (1995); FAM, familiarity from Gilhooly & Logie (1980); CELEX, fre-quency from Baayan et al. (1993); KF, frequency froKuc̆era & Francis (1967).

quisition effect in the Monaghan and Ellis data.

SE

nb

e

io

st-no

e-in

ea

nsfr&neata

ntaaTn

e0

toor-t ofand

ineseuntse,

ef-cegesby

ge-ntoted

e-r-

arunde-p-thes

arethersekel.,a

ksple

6 ZEVIN AND

In summary, the factorial studies leave opewindow of uncertainty as to whether the oserved effects were due to differences in agacquisition or frequency.

The second methdology employed in tharea involves using multiple regression to islate unique variance in response latencies aciated with AoA (Brown & Watson, 1987; Buler & Hains, 1979; Lyons et al., 1978; Morriso& Ellis, 2000). These studies reported effectsAoA independent of other stimulus propertiincluding imageability, familiarity, and frequency. We conducted a similar analysis usthe data from the three large-scale studiesword naming and lexical decision mentionabove (Balota et al., 2001, Seidenberg & Wters, 1989; Spieler & Balota, 1997) and fousimilar results. For 528 of the words in thestudies, there are data concerning bothquency (Zeno, 1995) and AoA (GilhoolyLogie, 1980). For all three data sets, AoA afrequency were significantly correlated with rsponse latencies (Table 3). For the SpielerBalota (1997) and Balota et al. (2001) daboth factors account for unique variance.

It is important to avoid making a “correlatiois causation” error in interpreting these dahowever, because both AoA and frequencycorrelated with other stimulus properties.illustrate, Table 4 provides the correlatioamong AoA, frequency, Coltheart’sN, length inletters, familiarity, imageability, and concretness also from the (Gilhooly & Logie, 198

forom

ie

y

†p , .10, **p , .01, *** p , .001.

IDENBERG

a-

of

s-

so-

fs

gofd-

dee-

d-nd,

,reos

-,-

tions make it difficult to isolate effects dueage of acquisition per se. Some additional infmation is provided by assessing the amoununique variance associated with frequencyage of acquisition after the other measuresTable 4 have been partialed out (Table 5). Thresults indicate that whereas frequency accofor a small but significant amount of variancthe age of acquisition measure does not.2 Thesedata suggest that, rather than there being anfect of age of acquisition on skilled performanindependent of other stimulus factors, the aat which words are learned are determinedfactors such as frequency, length, and imaability. Thus, after these factors are taken iaccount, there is no residual effect associawith the age of acquisition measure.

The results in Table 5 differ from those rported by Brown and Watson (1987) and Morison et al. (1997), who conducted similanalyses using smaller sets of words and fosignificant effects of age of acquisition indpendent of frequency. The differing results apear to be related to differences betweenWFG norms and the Brown and CELEX normused in earlier studies. The WFG normsbased on a larger sample of texts thanBrown norms and the sample is more divethan either the Brown or CELEX samples. Lithe American Heritage norms (Carroll et a1971), the WFG sample includes texts frombroad range of reading levels, including boofor school-age children. Each text in the samwas assigned a grade level based on a readity formula. Frequency data are providedeach word at each grade level, ranging fr

s

2 The amount of unique variance attributed to eithervariable is surprisingly small. One relevant factor may bethat effects of lexical frequency are reduced or eliminatedby exposure to neighboring words. Words that have manyneighbors (e.g. consistent ones) do not show strong fre-quency effects in naming. Another is that naming is lesssensitive to frequency effects than are other tasks becauseit only measures time to initiate the response; frequencyeffects can also show up in things such as duration of thewhole utterance (Balota & Abrams, 1995) and duration ofonsets that contain continuants (Kawamoto, Kello, Jones,

TABLE 3

Frequency and AoA as Predictors of Response Latenc

Study Measure r Unique variance

Spieler & Balota WFG 2.28 2.59***(1997) AoA .28 2.35***

Seidenberg & WFG 2.19 1.52**Waters (1989) AoA .17 0.64†

Balota et al. WFG 2.49 9.20***(2001) AoA .44 5.15***

Note. AoA, age of acquisition; WFG, word frequencfrom Zeno (1995).

& Bame, 1998).

iin5

tseaw ath

ithhemthe

atonutr-

rdsoreerbe-velusions

gie,

** p , .01.*** p , .001.

sented above, we used the sum of thesequencies. The fact that the WFG frequenccorrelate more highly with response latencthan the other norms (Table 2) and yieldresidual effect of age of acquisition (Tablemay be related to the inclusion of this broadrange of texts.

To examine this issue further, we conducregression analyses using different subsetthe WFG corpus. Specifically, we examinhow much variance the WFG and AoA meures accounted for when the data from logrades were excluded (Table 6). The resultsall three of the large-scale behavioral studieshibit a consistent pattern: As more of the dfrom lower grade levels are excluded,

theon.bynd ofe-

e-

ofis,r al

e ase

e

e

*p , .05, **p , .01, *** p , .001.

fre-eseso)

er

ed ofds-erforex-taecy

decreases while the amount associated wAoA increases. In two of the three studies, tAoA effect reaches significance with data frothe younger grades excluded, although amount of variance account for is very small.

One interpretation of these results is ththere is a small effect of age of acquisition skilled performance which the WFG norms (bnot Brown or CELEX) pick up because the copus included texts for younger readers. Wothat are learned earlier may tend to be used moften in texts that are appropriate for youngreaders. Table 7 presents the correlations tween rated age of acquisition and grade-lefrequency for the 528 words used in previoanalyses; there are strong negative correlatwhich decline gradually with age. Thus it coube argued that the WFG frequency data for lower grades covertly encode age of acquisitiOn this view, skilled performance is affected two independent factors, age of acquisition afrequency of usage in adult language, bothwhich are captured by the cumulative WFG frquency measure.

There is a different explanation for these rsults, however: Unlike the Brown or CELEXcorpora, the WFG norms provide estimatesthe cumulative frequencies of words, thathow often they have been encountered ovelong period of time (e.g., since an individuabegan to read). Cumulative frequency may bbetter predictor of adult performance becau

AGE OF ACQUISITION 7

TABLE 4

Correlations among Six Standard Lexical Measures and AoA

Variable AoA WFG IM FAM CON LEN

WFG 2.5141***IM 2.5861*** .1073*FAM 2.6740*** .7203*** .2026***CON 2.3840*** .0056 .8082*** 2.0099LEN .1984*** 2.0666 2.1483*** 2.0605 2.1717***N 2.1976*** .1417** .1195** .1245** .1215** 2.7142***

Note. AoA, age of acquisition; WFG, log Zeno (1995) frequency; IM, imageability; FAM, familiarity (Gilhooly & Lo1980); CON, concreteness; LEN, number of letters; N, Coltheart’s N.

*p , .05.

TABLE 5

Unique Variance Accounted for by Frequency and AoAIndependent of Other Lexical Variables

Study Measure Unique varianc

Spieler & WFG 1.27**Balota (1997) AoA 0.29

Seidenberg & WFG 0.69*Waters (1989) AoA 0.01

Balota, et al. WFG 2.94***(2001) AoA 0.34

Note. AoA, age of acquisition from Gilhooly and Logi(1980); WFG, frequency from Zeno (1995).

it affects how lexical information is repre-

oho

anarehuc

o

iel

eerse-

dly-alle-rexed

(1967) but which contribute to cumulative fre-

me

†p , .10, *p , .05, **p , .01, *** p , .001.

sented in memory (as, for example, in the cnectionist models discussed below). On tview, age of acquisition norms account fvariance in skilled performance because thindex how frequently words were usedyounger ages, information that the Brown aCELEX norms do not include. Thus there iseffect of cumulative frequency on skilled peformance rather than separate effects of agacquisition and adult frequency of usage. TWFG norms provide a reliable estimate of cmulative frequency, leaving no residual effeof age of acquisition.3

In summary, the data in Table 1 and the crelational analyses suggest that the age ofquisition effects observed in previous studmay have been due to confounds with “adufrequency (measured by Kuc˘era & Francis and

3 It is important to recognize that the grade-level frquency data in the WFG norms are not literally data cocerning the grades (or ages) at which the texts were rRather, they reflect the assignment of texts to grade leusing a formula that weighs factors such as number of woper sentence and number of syllables per word. On measure,Charlotte’s Weband The Old Man and the Seaareboth assigned to the 4th grade reading level, for examThus, the data from the lower grade levels reflect texts tare likely to be read by children at a given age but also teof approxinmately similar structural complexity that aread at older ages. On our view (supported by the modepresented below), these norms are relevant because provide estimates of the cumulative frequency, rather ththe exact timing, of exposures to words.

n-isreytdn-ofe-t

r-

WFG). One difficulty in developing a well-controlled AoA experiment arises from thstrong correlations between AoA and othlexical variables presented in Table 4. Thecorrelations make it difficult to design factorial experiments in which AoA is varied for asufficient number of items with these another factors controlled. The regression anases suggest that AoA may account for a smamount of variance in skilled performance bcause it is correlated with how often words aread at younger ages, data that are not indeby “adult” norms such as Kuc˘era and Francis

8 ZEVIN AND SEIDENBERG

TABLE 6

Unique Variance Accounted for by AoA with Different Subsections of the WFG Norms Used as Predictors

WFG subsection

Study Predictor 2–131 3–131 4–131 5–131 6–131 7–131 8–131 9–131

SB AoA .36 .41 .44† .47† .50† .54† .56† .57†Frequency 1.26** 1.17** 1.01** .85* .84* .86* .78* .67*

SW AoA .04 .04 .06 .07 .08 .10 .10 .12Frequency .98* .97* .89* .83* .87* .95* .91* .91*

BCP AoA .39† .46† .52* .58* .63* .68* .72* .68*Frequency 2.43*** 2.22*** 2.04*** 1.92*** 1.97*** 2.10*** 2.11*** 2.18**

Note. WFG 5 Zeno (1995) frequency counts 2–13 5 Grade levels 2 (2nd grade) to 131 (University) in the WFG norms;SB 5 Spieler and Balota (1997); SW 5 Seidenberg and Waters (1989); BCP 5 Balota et al. (2001)

todi- ascy

smsg.d”nt.”

t aree-

alar

ac-s

t”

by

quency of exposure.

THEORETICAL ISSUES

The above discussion addressed somethodological issues that arise in attemptingisolate age of acquisition effects. The data incate a need to consider what statistics suchestimated age of acquisition and frequenmeasure and how they relate to the mechanithat underlie lexical acquisition and processinThe concept “age at which a word is acquireseems clear enough and intuitively differefrom “frequency of usage in adult languageHowever, whereas frequency norms reflecproperty of words (namely, how often they aused), age of acquisition norms reflect somthing different, a behavioral event (learningword by a certain age). This event is very simi

e-n-

ead.velsrds

this

ple.hatxts

relingthey

to a task such as naming aloud: one behavioran

hrtol.

utmrm

n

cteaep e

inthud

na

rw

f ce

er

in-s,

in-vem-

theo

oyinge-y

m-t et). tova-ngi-d,ls

at-inghlyheonbyer-re-to

hed-esheld

l

Note. All correlations significant,p , .001.

concerns how long it took to learn a word, tother how long it takes to pronounce a woThis point is particularly clear with respect “objective” measures of AoA (Morrison et a1997) obtained by determining the ages which children can name pictured objects. Jas studies of word reading have examined factors that make some words easier to nathan others, age of acquisition can be considewith respect to the factors that cause sowords to be learned earlier than others.

Among these factors is frequency. In matheories, the frequency with which a stimuluspracticed or experienced affects how early awell it is learned as well as skilled performanIf the age at which a word is learned is affecby how often it is experienced, then empiricestimates of AoA may covertly encode frquency of occurrence during the acquisition riod. Moreover, we have also seen that ageacquisition ratings are correlated with gradlevel frequency data from the WFG norms,cluding data from higher grades well past ages at which the words were acquired. Thage of acquisition norms appear to be relatefrequency of occurrence over a multiyear timspan beginning with initial acquisition.

Seen in this light, word frequency, as stadardly operationalized using norms such Kuc̆era and Francis (1967), provides the maining chronological data concerning hooften words are experienced in adulthooThese observations suggest that both age oquisition and “adult” frequency norms reflehow often words are encountered but at diffent points in a developmental continuum raning from initial acquisition to adulthood. ThWFG norms take matters one step further, p

ed.

,atsthe

eede

yisnde.dl-e-of--es, toe

-s

e-

d.ac-tr-g-

o-

countered at multiple points along this contuum, as well as cumulative frequency. Thuage of acquisition and frequency seem moretrinsically related than recent discussions hasuggested. In effect, studies like the ones sumarized in Table 1 attempted to dissociate effects of frequency of exposure during twwidely spaced time spans.

Connectionist Modeling

Connectionist models of reading that empldistributed representations and gradual learnfrom experience provide a theoretical framwork for examining effects of the frequencand timing of learning experiences on perforance (e.g., Harm & Seidenberg, 1999; Plaual., 1996; Seidenberg & McClelland, 1989Such models illustrate three points relevantthe AoA hypothesis. First, frequency has persive effects on network performance, includihow quickly a word is learned (“age of acquistion”) and level of skilled performance. Seconthese effects are intrinsically related. Modesuch as Seidenberg and McClelland’s (1989)tempted to provide a unified account of readacquisition and skilled performance in whicthe same computational principles appthroughout the developmental continuum. Teffects of frequency on learning a word and skilled performance are both realized changes to the weights governing network pformance. Thus the behavior of the system flects the cumulative effects of exposure words over time. Finally, the magnitudes of teffects of frequency of exposure differ depening on the state of the network, which changover time as knowledge is acquired. As tmodel picks up on the similarities that ho

AGE OF ACQUISITION 9

TABLE 7

Correlation between AoA and WFG Frequency at Different Grade Levels

Grade level

1 2 3 4 5 6 7 8 9 10 11 12 13 Tota

2.68 2.67 2.63 2.60 2.53 2.50 2.47 2.45 2.43 2.38 2.35 2.31 2.17 2.51

-across words, and as the weights assume values

E

ee

thd-i

ter

)ed

cnc

ro

e

tivaitamu6itnn

et

utyMhaa

e

ete

inhc

ene.ithn--

d-h

essedlyuplso

onfor of

ecter-ion

ey

edof

here-ceated

nulte-n-er- ofan-dsentille-on. isx-

10 ZEVIN AND S

that allow output to be produced accurat(i.e., minimize error), the effects of pattern frquency decline.

Some properties of these networks favoridea that there will be an advantage for worthat are learned earlier in training (Ellis & Lambon Ralph, 2000). (We assume for the remader of this discussion that stimuli are equaalong other dimensions.) Consider a netwosuch as Seidenberg and McClelland’s (1989which weights are initially set to random valuand output units take values of 1 or 0. The ajustments to the weights that occur using bapropagation with a logistic activation functioare proportional to the activation of the unit acording to the terma(1 2 a), wherea is the ac-tivation value. The adjustments are therefolargest when the activations are in the middlethe logistic function (around .5), as occurs whthe network is initialized with small randomweights. The adjustments become smaller asweights assume values that cause unit acttions to approximate more closely the target vues of 1 or 0. Thus, there is a loss of plasticassociated with learning the early trained pterns. In effect, early trained patterns becoentrenched in the weights (for an early discsion of this phenomenon, see Munro, 198Both Ellis and Lambon Ralph (2000) and Smet al. (2001) emphasized these aspects ofwork behavior in explaining age of acquisitioeffects.

There is, however, another factor to considthe effects of similarities across training paterns. The mapping between spelling and soin English exhibits considerable systematiciReading models such as Seidenberg andClelland’s (1989) employed representations tallowed the weights to encode these regulties. Thus, what is learned about one word cries over to other words with which it sharstructure. This property modulates the effectsexposure to a given word. Until the model bgins to encode the systematic aspects ofmapping, performance on a pattern is highly dpendent on how often it is trained. By latertraining, the weights reflect the structure of tentire training set, changing its behavior. On

IDENBERG

ly-

es

n-dkins-

k-

-

ef

n

hea-l-yt-e

s-).het-

r:-nd.c-atri-r-

sof-he-

ee

little impact, creating a discrepancy betwefrequency of training and network performancFurthermore, new words can be learned wlittle training if they share structure with knowwords. In the limit, a new word can be pronounced correctly with no training, as in nonword generalization. Thus, there is an initial avantage for words that are trained with higfrequency, but as the model learns, there is land less of a disadvantage for later trainitems. In effect, the entrenchment of earlearned words is reduced as the model pickson patterns that hold across words (see aMarchman & Bates, 1994).

In summary, the entrenchment phenomenin connectionist networks provides a basis age of acquisition effects, but other propertiesthe task and materials to be learned will affwhether there is the long-lasting effect on pformance suggested by the age of acquisithypothesis.

Using this theoretical framework, the issuof AoA effects in reading can be clarified bconsidering two factors,cumulative frequencyand frequency trajectory. Cumulative fre-quency refers to how often a word is presentto the network from the beginning to the endtraining. This is a simplified analog of howoften people have encountered a word to tpoint at which performance is assessed. Fquency trajectory refers to how experienwith a word is distributed over time. Thus,given cumulative frequency can be associawith different trajectories.

The AoA hypothesis, then, is the predictiothat frequency trajectory has an effect on adperformance independent of cumulative frquency. Specifically, if the cumulative frequecies of words (as well as other stimulus propties) are equated, then words for which mostthe training occurs early should show an advtage over words with other trajectories. Worthat are trained more often early in developmwill, in general, be learned earlier than wwords that are mainly trained later; thus, frquency has an effect on age of acquisitiHowever, the age of acquisition hypothesisthat there will be a further effect of this early e

x

c

u

,utl

ao

la

c

ofs ofntksove.sed

theicgsd, ar-isla-ceion

inla-ic

s,thee-elfe a

AGE OF AC

A measure such as Kuc˘era and Francis (1967frequency provides a poor estimate of cumutive frequency. Given the nature of the teused to generate the corpus, it tends to undetimate the frequencies of many low-frequenwords including ones that are experiencmainly during childhood. The WFG normprobably provide better information about cmulative frequency, but this is difficult to independently assess. Age of acquisition normscontrast, provide imperfect information abofrequency trajectory because some words are learned early (e.g., BOTTLE, CUP) are aused frequently later in life, whereas othe(e.g., TEDDY, BOOTIE) are not.

Because the actual cumulative frequencand frequency trajectories of different words anot known, and because frequency norms rated AoA provide imperfect estimates, we tothe approach of using simulation modeling explore the phenomena. Simulation also lowed control over stimulus properties that anormally confounded. Thus, we could creaconditions in which it was certain that cumutive frequency and stimulus properties weclosely matched while manipulating frequentrajectory, providing a strong test of the age

o

ree

o

un

o

iedet-n of toheicalcaltlese-erwasal-ly.r- inwutp,

-

acquisition hypothesis.

SIMULATION 1

In the first simulation, a model was trained a large corpus of words using the standard tenique of probabilistically presenting words duing training as a function of their estimated frquencies of occurrence (Seidenberg McClelland, 1989). The critical data concernsubset of items for which we manipulated fquency trajectory while keeping cumulative frquency constant. Some of these words wmore frequent early in training compared later (Early condition), whereas other words flowed the complementary trajectory (Late codition). By the end of training, however, cumlative frequencies of words in the two conditiowere the same. In addition, the same wordspeared in both the Early and Late conditioacross different runs of the model.

This model differs from previous models

QUISITION 11

)la-tsres-y

eds-

-int

hatsors

iesrendk

toal-rete-

reyof

nch-r-e-&a

e--re

tol-n--s

ap-ns

f

The task was closely related to the problemlearning the spelling–sound correspondenceEnglish, information that plays an importarole in the naming and lexical decision tasused in the behavioral studies discussed abThe input and output representations were baon English orthography and phonology, and training corpus, a large set of monosyllabwords, instantiated the quasiregular mappinbetween the two (Seidenberg & McClellan1989). Previous simulations have used moretificial tasks and stimuli that did not capture thrich structure (discussed further below). Simution 1, therefore, provides more direct evidenconcerning the occurrence of age of acquisiteffects in reading.

Methods

Architecture. The basic architecture shown Fig. 1 was used in all simulations. For Simutions 1 and 2, models with 100 orthograph(input) units, 250 phonological (output) unitand 100 hidden units were used. In addition,phonological layer had 20 hidden units that mdiated connections between this layer and its(cleanup units [Hinton & Shallice, 1991]). Thcleanup units differentiate this model fromsimple feed-forward net such as the one studby Seidenberg and McClelland (1989). The nwork is given an input pattern, and activatiospreads through the network over a seriestime steps. Each unit propagates activationthe other units to which it is connected. Tfeedback connections between the phonologand cleanup units create a type of dynamisystem called an attractor network that setinto a stable pattern over time (for additional dtails, see Harm & Seidenberg, 1999). A furthfeature of the model was that each time step discretized into a series of moments, which lows a unit’s activation to ramp up gradualThus, the learning algorithm (continuous recurent back-propagation) changes the weightsways that improve accuracy but also hoquickly the network produces the correct outp(for discussions, see Harm, 1998; Bisho1995).

Corpus and training. The training corpus con

D S

reto-i-rlyol-ofedofs,nds,

henea

cythelyfre-r-

ck-orn-

msofsst

chl atre-sinad-hether-

FIG. 1. Model architecture used in all simulations.

words. Of these words, 108 were critical itemwhose frequencies were manipulated, astailed below. The remaining 2,783 words (bacground items) were assigned frequencies tafrom the Marcus, Santorini, and Marcinkiewic(1993) norms, which are based on 43 million tkens from theWall Street Journal.4

The critical items were divided into two lisof 54. Sets of 4 items were created by exchaing onsets and rimes. The lists were counteranced such that, for example, FOIST and MIoccurred on one list and FIST and MOIST the other. Thus, each list contained each onand rime in the quadruple, but in different cobinations. The model was run 10 times with dferent initial random weights (between 0.1 a20.1), analogous to replications with differesubjects. Each list occurred five times in eatrajectory. Thus, the same items occurredboth the Early and Late conditions across simlations. The data presented below are averaacross the 10 runs of the model.

The Early and Late trajectories were dsigned to provide a strong test of the effects

teu-mste

a--ut1srem

4 The Wall Street Journalcorpus has been used extesively in sentence processing research, and at the timbegan this research it was the largest available corpuEnglish. The lexical sample is somewhat skewed insofawords such as STOCK, MARGIN, and INFLATION aroverrepresented compared to other corpora. In our simtions, the norms were used only to ensure that the bground items in the training set were presented with a dibution of frequencies similar to that seen in natulanguage. When the goal is to examine the effects of quency on individual words, other norms such as the W(Zeno, 1995) are preferable.

EIDENBERG

sde-k-kenzo-

tsng-

bal-STonset

m-if-ndntch

inu-ges

e-of

early exposure on later performance; they wenot intended to capture the observed trajecries for individual words, which are more varable. The frequencies of the words in the Eaand Late conditions were manipulated as flows. Training consisted of 10 epochs100,000 trials each. Early items were assigna frequency of 1000 for the first 3 epochs100,000 training trials. For the next 4 epochthe frequency was adjusted to 500, 100, 50, a10 in succession. Finally, for the last 3 epochthe frequency was set to 1. The trajectory in tLate condition was the complement of the oin the Early condition. Late items started atfrequency of 1 for the first 3 epochs; frequenwas adjusted to 10, 50, 100 and 500 overnext 4 epochs; and the frequency finalreached 1000 for the last 3 epochs. Thesequencies are within the range of the raw Macus et al. (1993) frequencies used for the baground items. As with the frequencies used fthe noncritical words, these assigned frequecies were square root transformed, and itewere sampled probabilistically. This methodcompressing the frequency distribution allowthe model to learn very low-frequency itemafter a relatively small number of trials (Plauet al., 1996). The actual frequencies with whithe critical items were presented to the modeeach epoch are given in Fig. 2. The mean fquency for Early items in the first epoch wa41, and the mean frequency for Late itemsthis same epoch was 4. Frequencies werejusted over time such that in the last epoch, tLate items had a mean frequency of 40 andEarly items had a mean frequency of 4. Impotantly, by the end of training, the Early and Lawords had been trained equally often; the cmulative frequencies averaged across itewere 198 for Early words and 196 for the Lawords,t(107), 1.

On each training trial, a word was probbilistically selected for training and its orthographic pattern was activated on the inpunits. Activation propagated forward for 1time ticks. On the 12th time tick, error wacomputed and the weights of the model weadjusted accordingly. The learning algorith

n-e wes ofr aseula-ack-stri-ralfre-

12 ZEVIN AN

computes error on the basis of the differenceFG

ds

t

r

m

eict

g

vr

e

g

eo

ell

edas

dnss

-,

yn-

fass;is

- ofisg,ed,e-cen.e-

hesame level.

FIG. 2. Frequency trajectories of critical items in Simula

-

between the desired and observed outputs given time tick as well as the state of the moat earlier time ticks. In this way, each adjument of the weights leads to incrementamore accurate and faster computation of desired output.

Results and Discussion

The model’s performance was assessusing both accuracy and sum squared er(SSE) measures. The model’s output for a wowas scored as correct if the output for eaphoneme was closer to the correct phonethan any other by Euclidean distance. TheSSEmeasure was the sum of the squared differenbetween the computed output and the targThe two measures are highly related; corrwords produce lower error scores than docorrect words. However, among the correwords, differences inSSEreflect the relativedifficulty of generating a response (see, e.Seidenberg & McClelland, 1989). Thus, thmodel’s performance can continue to improafter it has learned to produce the correctsponse, as in human performance.

At the end of training, the model produccorrect output for 98% of the training seNearly all errors were on low-frequency stranwords such as COUP, PLAID, and RHEUMwhich are thought to require input from the othography → semantics → phonology pathwaythat was not implemented here (Harm & Seidberg, 2002, Plaut et al., 1996; Strain, Patters

tion 1.

& Seidenberg, 1995).

QUISITION 13

at aelt-

llyhe

edrord

che

ceset.ctn-

For the smaller set of critical words, thmodel learned to produce correct output for aitems within the first epoch. Mean sumsquared error for these items was calculatafter each epoch. As shown in Fig. 3, there wa small effect of frequency early in trainingthat rapidly disappeared. This was confirmeby t tests on the difference between the meain the Early and Late conditions. Error scorewere significantly lower for Early words compared to Late words after the first epocht(107) 5 4.24, p , .001, and this effect re-mained significant after 5 epochs,t(107) 52.09,p , .05. By epoch 6, when the frequenctrajectories began to cross, the effect was nosignificant,t(107) 5 1.12,p . .10. At the endof training, when the cumulative frequency othe two groups was closely matched, there walso no reliable difference between conditionin fact the means were identical at .50. At thpoint, all critical items were still pronouncedcorrectly.

The first simulation indicates that with stimulus properties equated, there is an effectfrequency trajectory early in training, but theffect rapidly recedes. By the end of traininwhen the cumulative frequencies are equatthere is no residual effect. Early in training, bfore much learning has occurred, performanis better on words that are trained more ofteThis is simply a frequency effect during thearly phase. As training continues, performance in the two conditions converges to t

-

AGE OF AC

.,eee-

dt.e,r-

n-n,FIG. 3. Performance over time for critical items in Simu

srla

tr

T

hT

lacb

s”e

t

r

ee

olscen

tct

ti-eirthe-rdscyla-rei-atact

a-ce ofn fre-ce

re-n-n-

s inerk- al.rit-foras- in

che-r-cu-rly

-n pre-rdelle-ofu-

rlynd

14 ZEVIN AND S

SIMULATION 2

Simulation 2 was a replication of the firsimulation that addressed two concerns. Fieffects of the frequency trajectory manipution might have been difficult to detect becauthe critical stimuli all contained spelling paterns with consistent spelling–sound corspondences. In addition, the stimuli were costructed in quadruples such as FIST–MOISMIST–FOIST, ensuring that every word–bodoccurred at least twice with the same pronunation. In the type of network studied herlearning of one item with a given spellingsound pattern (e.g., FIST) carries over to otitems containing the same pattern (e.g., MISreducing the effects of exposure to the itemself (a neighborhood effect). The net result wthat all of the critical words were learned retively rapidly; there was an effect of frequenof exposure early in training, but it was oserved on the sum squared error measure,on how rapidly the model learned (i.e., ageacquisition). Therefore, we created a new of critical stimuli containing only “strangewords (Seidenberg, Waters, Barnes. & Tanhaus, 1984) that have atypical spellings aspelling–sound correspondences. Because have few close neighbors, these words shlarger effects of frequency both in behaviostudies (e.g. Seidenberg et al., 1984) andconnectionist models (e.g., Seidenberg McClelland, 1989). Therefore, we expectedsee effects of frequency trajectory both on SSEand on how quickly these words were learne

A second issue concerns the processes gave rise to the Fig. 3 data. One possibilitythat these data reflect two complementary agacquisition effects. Thus far, we have followthe behavioral research in emphasizing the psible effect of early high-frequency exposure skilled performance. However, there might abe a complementary effect of high-frequenexposure late in training. Thus, the similar levof performance in the Early and Late conditioat the end of training might derive from twsources: an AoA effect and a recency effec(Lewis, 1999, found evidence for both in a fanaming task). Therefore, we added a con

EIDENBERG

tst,-

se-e-n-–

yci-e,–er),

it-as-y-notofet

n-ndheyowal in& to

d.thatis ofdos-noylss

o

erol

jectory. For this condition, a subset of the crical items from Simulation 1 were assigned thnormal frequencies and included among background stimuli. After running the simulation, we isolated a large subset of these womeeting the conditions that (a) their frequentrajectories were very flat and (b) their cumutive frequencies were similar to what they wein Simulation 1. Thus, the Flat trajectory condtion acts as a baseline against which the dfrom Simulation 1 can be compared. An effeof either the Early or Late trajectory in Simultion 1 would be indicated by better performanthan in the Flat trajectory condition at the endtraining. Finally, the flat trajectory conditiowas also used to assess whether cumulativequency has an effect on network performanindependent of trajectory, by comparing the sults for two subsets of stimuli from the flat codition whose cumulative frequencies were cosiderably different.

Methods

The same model and corpus were used aSimulation 1. The critical items from the earlisimulation were included among the bacground items and assigned their Marcus et(1993) frequencies, and a different set of 48 cical items was selected. The main criterion the critical items was that their bodies not be signed the same pronunciation in other wordsthe training list; thus, they included words suas BEIGE, PHLEGM, and SCOURGE. Thstimuli were divided into two lists, with the assignment of lists to training condition countebalanced across two simulations. The mean mulative number of presentations for both Eaand Late words was 183.

Stimuli in the Flat trajectory condition consisted of 95 of the critical stimuli in Simulatio1. These items were selected because, whensented throughout training at their standaMarcus et al. (1993) frequencies, they are wmatched to the critical items for cumulative frquency. The mean cumulative frequency these words was 200, comparable to the cumlative frequencies for these words in the Eaand Late conditions in Simulation 1 (198 a

Q

ees

iasgr

ifacsi-c

caeav-.7

es

hhhrerhrtrio

e

nuoa

el

-

nc-cyndherlye

t-elrlyesg

ese

lycyn-r-

ot

or

AGE OF AC

Results

After 10 epochs, the model generated corrphonological codes for 98% of the training sPerformance on the critical items was assesin terms of SSE,accuracy, and how quicklywords were learned (i.e., age of acquisitionmodel time). Because the models were initiized with different random weights, and becauwords were selected probabilistically durintraining, individual runs of the model diffeslightly from one another in terms of performance including when during training individuawords were learned. Analogous individual dferences are seen in children. For each item,of acquisition was defined as the point at whi75% of the models generated correct responThis criterion is similar to one used in the Morrson et al. (1997) study in which the age at whichildren acquired a word was defined as the aat which 75% of the children could name a pitured object accurately. By this measure, theerage “age” at which Early items were acquirwas approximately 2.09 epochs, whereas theerage age for Late items was approximately 6epochs. This difference is significant,t(34) 512.14. Note that epochs are defined with respto the total number of training trials on all itemincluding the 2843 background words, not tnumber of exposures to individual words. Tmean numbers of trials to learn words in tEarly and Late conditions were 296 and 250,spectively. These data indicate that the Eawords were acquired more rapidly than were tLate words, as expected. It took fewer exposuto learn the Late words because they benefifrom prior learning of other words. Even fostrange words, then, there is generalizatbased on exposure to other words.

Accuracy over the course of training is dpicted in Fig. 4A. As in the previous simulationthe advantage for the Early items dissipatedthe cumulative frequency of the Late items coverged on that for the Early items. Mean accracy for both conditions was 85% at the endtraining. This level of accuracy is somewhlower than that for the consistent words in Simulation 1. This finding is consistent with thview that performance on the most difficu

UISITION 15

ctt.ed

nl-e

-l-gehes.

hge-v-d

thography→ semantics→ phonology. How-ever, the error rate did not differ in the two frequency trajectory conditions,t(47) , 1. Thus,although the frequency trajectory manipulatioaffected the “age” at which items were aquired, it had no residual effect on accurawhen the cumulative frequencies of Early aLate items converged. Figure 4B shows tchange in sum squared error over time for Eaand Late items, which is very similar to thaccuracy graph.

One further aspect of the data is worth noing. Toward the end of training, the modbegan to exhibit some unlearning of the Eawords, as indicated by the slowly rising scorin this condition for both measures. Protectinearly acquired words from unlearning requirintermittent reexposure to these items over tim(Hetherington & Seidenberg, 1989). The Eartrajectory entailed a steep decline in frequentoward the end of training. This property, takewith the probabilistic nature of sampling, resulted in too few exposures to maintain pe

0

ct

eee-

lyeesed

n

-,as--ft-

t

formance at the maximum level. We did n

FIG. 4. Performance over time for Simulation 2: (A) err

t

t

n

ra-t

u

cot

pmm

mo

bo

aeb

pac

eoot

orx-or-

era-rvee

s atilee-chr-ined

ze–he ondi-tsct---

ithse

2)ions

in Simulation 1.

16 ZEVIN AND S

systematically examine performance afterepochs because it was at this point that theconditions converged on the same cumulatfrequencies. We do know, however, that a smnumber of additional training trials on the critcal items is sufficient to stop the slow erosionperformance seen in Fig. 4. This behavior ofmodel is broadly consistent with human peformance; knowledge acquired during chilhood may degrade over time through lackuse but can be revived with modest additioexperience.

We now consider the results for the Flat tjectory condition. This condition addresses concern that the results of Simulation 1 mighave derived from two complementary AoA efects: one due to high frequency of exposearly in training and one due to high frequenof exposure late in training. If this were correthen performance at the end of training in bthe Early and Late conditions should be bethan that in the Flat condition in which frequecies changed very little across epochs. Thissult was not observed. Figure 5 summarizes formance in the Flat condition and on the saitems in the Early and Late conditions froSimulation 1.

Results in the Flat condition closely resebled those obtained in the Early condition. Bconditions exhibited a small advantage earlytraining compared to the Late condition, but the end of training all conditions converged the same level of performance. The mean SSEinthe Flat condition was .48, compared to .48 .49 in the Early and Late conditions, resptively. No effect of frequency trajectory was oserved,F(1,93) , 1. The early advantage in thFlat condition reflects the fact that the items ha mean frequency of 20 presentations 100,000, which was higher than that in the Lcondition over these epochs. However, the mulative frequency of Flat items (200) was nsignificantly different from those of Early anLate items F(1,93) , 1.

Data concerning the role of cumulative frquency are presented in Figure 6, which shthe sum squared errors for the highest and lest frequency 25 items. The mean cumula

hehtf-recyt,thtern-re-er-e

-th inyn

ndc--

eaderteu-

otd

-wsw-

ive

(544 for the highest frequency words and 60 fthe lowest). Cumulative frequency has the epected effect on performance which is better fhigh-frequency words (.46) than for low-frequency words (.55),t(47) 5 3.22, p , .005.Note that these means are substantially lowthan the means for the critical items in Simultion 2. This suggests that the failures to obseAoA effects were not due to floor effects on thcritical items.

Discussion

Results from the Early and Late conditionwere consistent with Simulation 1. There waslarger difference between these conditions unwell into training, which reflects the fact that thcritical words have few neighbors and, therfore, performance does not benefit as mufrom training on other words. However, peformance in the two training conditions agaconverged as the cumulative frequencies evenout. Thus, the results of Simulation 1 generalito stimuli that have less consistent spellingsound mappings. Performance on words in tFlat condition converged to the same level asthese same words in the Early and Late contions in Simulation 1, indicating that the resulfor the Early and Late conditions did not refletwo complementary types of facilitation. Finally, there was an effect of cumulative frequency in the Flat condition: At the end of training, performance was better on the words whigher cumulative frequencies than on tho

EIDENBERG

10woivealli-ofher-

d-ofal FIG. 5. Performance in the Flat condition (Simulation compared to the same items in the Early and Late condit

ferwith lower ones.

Q

uc

qrul

thath

oer

o-a

d

dr-

ge

a-byrda-

t,lyr. ofin-ms, in

FIG. 6. Performance on high and low cumulative frequency items within the Flat condition.

These results suggest that whereas cumtive frequency has an impact on performanfrequency trajectory does not. The age of acsition hypothesis tested in previous behavioexperiments was that there would be a resideffect of early word learning on skilled aduperformance. However, although words in Early condition were learned more rapidly thwords in the Late condition, performances in

AGE OF AC

two conditions were nearly identical by the en

eel

dr

insnnri-dcod

dteg

liit-rge inantrlyatlss.

isorial

2d

helytsto-r-

d

of training.

SIMULATION 3

To this point, the results suggest that whcumulative frequencies and stimulus propties are equated across conditions, there istle, if any, effect of frequency trajectory. Whamatters is how often a word is encounterenot the pattern of encounters over time. Hewe consider another factor that may have cotributed to these results: the fact that the traing corpus consisted of words that exhibit sytematic relationships between orthography aphonology. What the model learns about oword carries over to other words that shastructure with it, reducing the effects of lexcal frequency (Seidenberg & McClellan1989) and, thus, the effects of any frequentrajectory manipulation. These neighborhoeffects were larger for the consistent worused in Simulation 1 than for the strange itemused in Simulation 2; the consistent worwere learned more rapidly and yielded betasymptotic performance than the stran

UISITION 17

la-e,ui-alal

tened

nr-it-t,en---dee

,ydsssre

lative frequencies were very similar in the twcases. Although the strange words have fewclose neighbors, their orthographic– phonlogical correspondences are not arbitrary;word such as BEIGE is not pronounce“glorp”; it overlaps with the more distantneighbors BINGE, BARGE, WEIGH, andmany other words among the backgrounstimuli. Thus, the systematic aspects of the othography→ phonology mapping might havereduced trajectory effects even for the stranwords.

Suggestive evidence is provided by simultions of age of acquisition effects presented Ellis and Lambon Ralph (2000). Feed-forwamodels were trained to produce a transformtion of arbitrary bit vectors. In their training seoutput vectors were generated by randomchanging 10% of the bits in the input vectoEllis and Lambon Ralph observed strong ageacquisition effects such that items that were troduced early had an advantage over late iteeven when the later items were much highercumulative frequency. The nature of the stimumeant that learning on any given trial carried ltle information relevant to other items. Undethis condition, there was a residual advantafor mappings that became entrenched earlytraining. Ellis and Lambon Ralph provided thorough discussion of why this entrenchmeoccurs. In essence, learning that occurs for eatrained items involves large weight changes threduce the model’s sensitivity to error signagenerated by the presentation of later itemSmith et al. (2001) provided a similar analysof the results of their simulation, which was alsconstructed so that what was learned on one tdid not carry over to other trials.

Together, the results of Simulations 1 andand of the Ellis and Lambon Ralph (2000) anSmith et al. (2001) simulations suggest that tnature of the input–output mapping, specificalwhether what is learned on one trial predicanything about other trials, may be crucial producing AoA effects. To investigate this hypothesis, we devised a training regime delibeately unlike the orthography → phonologytranslation in English. Items for the Early an

-

u-Late trajectory conditions in Simulation 3 were

E

hpndlit

0

w omnh

/p.eatr ithtio 2nnsthte

thletshr n

m pnoo

usnge-t 40lyal

fre-me.er

asre- tou-asof

ofor-e2velto5.5e,helereerbyialofeds.norm-tos.

t toes

18 ZEVIN AND S

constructed such that Early and Late items minimal orthographic or phonological overlaIn addition, we did not include any backgrouitems; thus, what the model learned depensolely on the properties of the critical stimuThese conditions are more comparable to ones studied by Ellis and Lambon Ralph (20and Smith et al. (2001)5.

Methods

The training set consisted of 68 words. Tlists were created out of different inventoriesletters and phonemes. One list included itesuch as COB, COG, COP, HOG, HOP, aTOG, whereas the other contained items sucBAD, BAN, BANE, PANE, PAN, and PAT.Some phonemes occurred in both lists (e.g.,but in different positions in different lists (e.gonset and coda). The model’s phonological rresentation (Harm & Seidenberg, 1999) trethese as separate phonemes; thus, whalearned about onset /p/ does not carry ovecoda /p/. The simulation was run twice, wlists assigned once to each trajectory condi(Early, Late). In contrast to Simulations 1 andno other words were presented during trainiThus, the model could learn regularities amothe items within a training condition, but theregularities did not extend to the items in other list, and performance was not modulaby exposure to any noncritical items.

Due to the smaller size of the training set,models in Simulations 3 and 4 used a scadown architecture with 29 orthographic uni40 hidden units, and 10 cleanup units. Tphonological layer was kept the same. Fquency trajectories for items in Simulationsand 4 were similar to those in Simulations 1 a2. However, because no “background” itewere present, the range between lowest (910,000) and highest (290 per 10,000) frequewords is more dramatic. This is because hfrequently an item is presented depends on b

asatinn-

5 The simulations in this article were actually completebefore we were aware of the Ellis and Lambon Ralp(2000), Smith et al. (2001), and Monaghan and Ellis (200simulations.

IDENBERG

ad.ded.he0)

ofs

d as

/),,p-ts isto

n,

g.g

eed

ed-,e

e-3dser

cywth

of

other items in the training set. In the previosimulations, nearly 3000 words were beitrained, so that even items with very high frquencies were seen only, on average, aboutimes per 100,000 trials. In this simulation, on68 items were trained, resulting in higher refrequencies, although the log-compressed quencies used to select items were the saAlso because of the smaller training set, fewtraining trials were required. The model wtrained for 10 epochs of 10,000 trials each,sulting in 100,000 training trials, as opposed1 million in Simulations 1 and 2. The mean cmulative frequency of Early words (1474) wnot different from the cumulative frequency Late words (1467),t(67) , 1.

Results and Discussion

Figure 7 presents the accuracy and meanSSEdata over the course of training. By the endtraining, the model had learned to produce crect output for all words. Whereas all of thEarly items were learned within the firstepochs, the Late items did not reach this leuntil much later. The mean number of trialslearn the Early words was 1.3 epochs versusfor the Late items, a highly reliable differenct(67) 5 49.10. Again, these numbers reflect tpoint in training as a function of all trials for alitems. Because so many of the Early items wlearned within the first epoch, the mean numbof exposures before learning was computedexamining the model’s performance at 1000 trintervals. By this measure, the mean numberexposures to a given item before it was learnwas 242 for Early items and 270 for Late itemNote that this is different from Simulation 2, iwhich feweractual exposures were required fthe learning of the Late items. In the present siulation, knowledge of the Early items seemedimpede rather than aid learning of the Late itemThe contrast provides a reminder of the extenwhich learning spelling–sound correspondencnormally depend on exposure to neighbors.

In contrast to previous simulations, there wa small but reliable advantage for words thwere presented frequently early in training Simulation 3, even after the cumulative freque

dh2)

f1

er

lanprpeo erciniv

iec

nist.u-s,

there was an advantage for the early trained

ureter-er-dys

s).heingndo-

ed, let-osi-hatnot,,

.d

eteu-e,

A.ndus,tedu-

gh

ro-

ulation 3: (A) error rate and (B) sum squared error.

As shown in Fig. 7B, there was an advantageEarly words that was maintained through epochs of training. A t test on the mean SSEatthe end of training revealed that error was rably greater for Late words (1.13) than for Eawords (0.74),t(67) 5 10.08,p , .001.

The critical difference between the simutions concerns the nature of the stimuli athus, the mapping between input and outcodes. Simulations 1 and 2 used a large coof words that exhibit the correlations betwespelling and sound characteristic of English thography. These correlations modulate thefects of frequency of exposure to a given woyielding no residual effect of frequency trajetory on skilled performance. This result obtawhen other stimulus properties and cumulatfrequencies are controlled.

In Simulation 3, the normal consistenciesthe mapping between spelling and sound wnot maintained because we eliminated the baground items and created nonoverlapping stim

QUISITION 19

in a training list carried over to other words othe same list but not to words on the other lGiven this sharp dissociation between the stimlus characteristics of Early and Late word

AGE OF AC

FIG. 7. Performance over time for critical items in Sim

o

or0

li-ly

-d,utusnr-f-

d,-se

nrek-u-

items.

SIMULATION 4

Simulation 3 strongly suggests that the natof the mapping between input and output demines whether frequency trajectory affects pformance. However, this simulation differefrom the earlier ones in a number of other wa(e.g., number of units, size of training corpuTherefore, we ran a final simulation using tsame procedures as in Simulation 3 but usstimuli that, like the ones in Simulations 1 a2, contain overlapping orthographic and phonlogical patterns.

Methods

The same items from Simulation 3 were usbut rather than segregate items such that noter or phoneme was repeated in the same ption between lists, we organized the lists so tno letter or phoneme occurred on one list but on the other. For example, HUB, HUG, LUCKPAT, and MAD were on List 1, whereas HUCKLOG, LUG, MATE, and PAD were on List 2Cumulative frequencies of Early (1474) anLate (1467) words were matched,t(67) 5 1.12,p . .20.

Results and Discussion

As in Simulations 2 and 3, Early items werlearned quickly (1.7 epochs), whereas Lawords required more training to be named accrately (3.7 epochs). This difference is reliablt(67) 5 9.80,p , .001. This is reflected in thechange in accuracy over time, shown in Fig. 8Also note that accuracy on both the Early aLate items reached 100% by the 6th epoch; thalthough frequency trajectory had the expeceffect on AoA, it had no residual effect on accracy. The model’s ability to generalize fromEarly items to Late items meant that even thouit took muchlonger in terms of training epochsfor the Late items to be learned, they were p

-

rdduced correctly after many fewer trials per word;

oew

hn

ila

i

m

n

u

a

20 ZEVIN AND S

the mean number of exposures to produce crect output was 262 for Early items and 52 fLate items. As shown in Fig. 8B, sum squarerror on the Late words decreased more slothan that on the Early words, but performancethe two conditions eventually converged. TSSEwas not different between Early (1.13) aLate (1.13) items at the end of training,t(67) ,1. As in Simulations 1 and 2, there was no resual effect of frequency trajectory when cumutive frequencies were matched. Error declinmuch more rapidly for the Late words in thsimulation (Fig. 8A) than in Simulation 3 (Fig7A). This is because learning on the Early itetransferred to performance on Late itemwhereas in Simulation 3 learning on Early aLate items was independent.

Because this simulation was identical every other respect to Simulation 3, the resindicate that the factor relevant to producingfrequency trajectory effect in Simulation 3 w

v

oabtoca rth oaneaitrrinx

is-ce

hatim-h-fre-iesly.iveitr-

.e-or-hn-alheoutator

FIG. 8. Performance over time for critical items in Simu-lation 4: (A) error rate and (B) sum squared error.

words.

GENERAL DISCUSSION

Studies of age of acquisition effects haraised important questions about the effectsearly experience on later learning. An effectage of acquisition on skilled reading would cinto question the results of many previous havioral studies and models in which this facwas not investigated. The potential theoretiimportance of this phenomenon, as well methodological and theoretical concerns, ledto examine it further. Examination of the mateals used in previous studies suggested that did not provide strong evidence for an effectage of acquisition independent of other meures of frequency with which AoA was cofounded. The regression analyses provided dence that age of acquisition ratings maccount for a small amount of variance skilled performance with other factors statiscally controlled, but via the fact that they acorrelated with how often words are used padulthood. Thus there was no effect of AoA dependent of cumulative frequency as inde

eoff

lle-r

als

usi-eyfs--vi-yni-ee--

ed

The results of Simulations 1 and 2 are constent with these conclusions and provide evidenconcerning the computational mechanisms tgive rise to the behavioral phenomena. The sulations provide a strong test of the AoA hypotesis because the cumulative frequencies andquency trajectories were known, and propertof early and late stimuli were equated exactThe training corpus was a large representatsample of monosyllabic words, which exhibthe statistical regularities characteristic of the othography→ phonology mapping in EnglishThere was an initial advantage for words prsented more frequently early in training but nresidual effect of early learning on skilled peformance. This was true for both words withighly consistent spelling–sound correspodences (Simulation 1) and words with atypicspellings and pronunciations (Simulation 2). Tadvantage for early trained words is washedas the model picks up on the similarities thhold across words. This occurs more rapidly f

EIDENBERG

or-rdlyined

d--

eds.s

s,d

inlts aste

words such as LAST, whose component spelling

Q

oaIr

aed

reani

dh

te

i-larb

inInmrlen

ionAp

thiotiv a

eicr

laon

-e-

here,rnsrlyg arns.ve ofpo-hisails

a-k. ofhe

nsitstoof iner-

erethetheut.t-es,hehe offre-nd

e-er-se- theyre- bylsoseon-h

thee-ts.

AGE OF AC

patterns are pronounced consistently acrmany words, than for strange words suchBEIGE, which has fewer close neighbors.both cases, however, early and late trained woconverged to the same level of performancethe number of exposures evened out. This behior can be traced to basic properties of conntionist models (Seidenberg & McClellan1989). Knowledge in these models is encodedweights on connections among units, whichflect the cumulative effects of exposure towords. Changes to the weights that occur wheword is trained also benefit words with whichoverlaps. This leaves little room for early worto maintain an advantage because the weigthat support them also facilitate learning lalearned words.

Simulations 3 and 4 provide further evdence consistent with this analysis. In Simution 3, we removed the overlap between eaand late trained words and observed a reliaage of acquisition effect: There was an advatage for early trained words that was matained throughout the course of training.this case, learning of the late items was ipeded by the model’s knowledge of the ealearned words. Finally, in Simulation 4, wreintroduced the overlap between early alate trained words, and the age of acquisiteffect was eliminated, further demonstratithat the critical factor that gave rise to the Aoeffects in Simulation 3 was the lack of overlaamong the early and late patterns.

In summary, both the behavioral data and simulations are consistent with the conclusthat, whereas there is an effect of cumulafrequency on reading performance, there isindependent effect of the age at which wordslearned.

Conditions That Create Age of AcquisitionEffects

In the remainder of this article, we considother types of conditions and tasks for whage of acquisition effects are likely to be moprominent. Our Simulation 3 and the simutions previously reported by Ellis and LambRalph (2000), Smith et al. (2001), and Mo

UISITION 21

sss

ndsasv-c-,in-

lla

tstsr

-lylen--

-y

dng

ene

nore

rhe-n-

quisition effects will occur under some circumstances. Although these simulations differ in dtail, they share an important property: Given tnature of the stimuli and network architectuwhat was learned about early trained pattedid not carry over to later trained patterns. Eatrained patterns became entrenched, yieldinpersistent advantage over later trained patteOur main point is that the conditions that girise to these effects are not characteristicreading an alphabetic orthography but are tentially relevant to other tasks. To see tclearly, it is necessary to examine some detof the simulations.

The Ellis and Lambon Ralph (2000) simultions involved a simple feed-forward networThe input and output layers each consisted100 units, and there were 50 hidden units. Tinput stimuli consisted of random bit pattercreated by activating a random 20% of the unon the input layer. The model was trained copy the input onto the output, but with 10% the bit values changed (randomly determinedadvance). Two aspects of the simulations undlie the strong age of acquisition effects that wobserved. One has to do with the nature of patterns that were trained and the other with nature of the mapping between input and outp

The important property of the training paterns is that, unlike words in natural languagthey did not exhibit a rich internal structure. Tstatistical structure of the lexicon reflects tfact that there are constraints on the orderingletters and phonemes and differences in the quencies with which these elements occur aco-occur. Much of this structure ultimately drives from constraints imposed by speech pception and production; for example, certain quences of phonemes are ruled out becausecannot easily be articulated, and the relative fquencies of patterns are determined in partease of articulation. These constraints are areflected in alphabetic writing systems becauthey are codes for representing speech. In ctrast, the stimuli in the Ellis and Lambon Ralp(2000) simulation were constructed so that probability that any given unit was on was indpendent of the probabilities for all other uni

SE

helleu

tieww bthromthe

anizapomodrnin

pedinfo

irndetpbuit

einen

onifngplns drue

anernli

oc-putthhe

edtondg– ac--

dyncean

ef-utls,heas tois-isi-id.regeinrnsn

ui--w-to

edrecallingrlapoed

ntew-ryer-s.st

22 ZEVIN AND

pattern does not carry information about otpatterns. Using an architecture with a smanumber of hidden units than input or outpunits promotes the discovery of subregularithat hold across patterns (as occurs, e.g.,words). If these regularities do not exist, hoever, then the model can learn the task onlymemorizing individual patterns, even though mapping is prima faciehighly consistent. Undethese conditions, early trained patterns becentrenched; the large initial weight changes favor these patterns are difficult for later trainpatterns to overcome.

The nature of the mapping between input output codes also promoted pattern memortion in these simulations. The fact that the mping between input and output involved randchanges to 10% of the bits meant that the mcould not generalize from early trained patteto later trained ones accurately. The mappbetween input and output codes contained atial regularity (90% of the input bits mapponto the corresponding output bit), but the consistent elements were random and, thereunlearnable except by memorization.

The Smith et al. (2001) simulation was simlar in that the stimuli were random bit pattethat were not internally structured. Their mowas also trained to copy the input to the outhrough a smaller number of hidden units without the random changes to 10% of the bLike Ellis and Lambon Ralph’s model, Smithal.’s model performed the task by memorizthe training patterns and again exhibited trenchment of early learned patterns.

The Monaghan and Ellis (2002) simulatialso conforms to this analysis, although it dfers from the other simulations in interestiways. The simulation again involved a simfeed-forward network. Unlike the simulatiodiscussed above, the training patterns weresigned to capture some aspects of lexical stture. The input and output layers were dividinto three slots, analogous to a consonvowel-consonant (CVC) syllablic structurWithin each slot, there were 10 bit patte(phonemes) that were repeated across stimu

ui-

IDENBERG

rrtsith-y

e

eatd

da--

elsg

ar-

-re,

-sl

utt

s.tg-

-

e

e-c-dt-

.s inn

ously activated; what was learned about one currence of a pattern over the whole set of inunits could carry over to other patterns wiwhich it overlapped (i.e., those containing tsame phonemes).

Monaghan and Ellis (2002) also manipulatthe consistency of the mapping from input output. In a behavioral experiment, they fouthat whereas words with inconsistent spellinsound correspondences produced an age ofquisition effect, words with consistent correspondences did not. The stimuli in this stuwere discussed earlier; there is some evidethat the effect was due to frequency rather thage of acquisition. In the simulation of these fects, the consistency of the mapping from inplayer to output was varied. On 80% of the triathe model was trained to copy the input; on tother 20%, the input of the consonants wcopied but the vowel was randomly assignedone of the other 9 possible vowels. The constent patterns did not produce an age of acqution effect, whereas the inconsistent patterns d

The results for the consistent condition alike those we observed in Simulation 1: no aof acquisition effect when the stimuli overlap structure. The results for inconsistent patteappear to conflict with the results of Simulatio2, in which we did not observe an age of acqsition effect for words with atypical (inconsistent) spelling–sound correspondences. Hoever, the differing results are traceable properties of the stimuli. Our model was trainon a large set of words; the critical stimuli wea subset of “strange” words that contain atypispelling–sound correspondences. The modeindicates that these words nonetheless ovesufficiently with other words in the corpus twash out the initial advantage for early trainitems.

Monaghan and Ellis’s (2002) inconsistestimuli were wordlike patterns in which thvowel was randomly mapped onto other voels for 20% of the items. Given the arbitranature of these mappings, the model could pform the task only by memorizing the patternAs in other conditions in which patterns mube memorized, there was a strong age of acq

Q

is

l

a

or/a

rbd

g

y,s

hdh

g

ni

h 2hseh

omo

hanrk

lowial

1rn-el

ngof

rdnsteedursosszeeofupnsed

idnd

ctsl-

arec-ingeg-c-

e.is

gen-uro-p-cs.s

csn-l,

the

AGE OF AC

gree of arbitrariness is not seen in Englwords, even strange ones. Although vowgraphemes in English map onto multipphonemes, the range of possibilities is costrained. No vowel grapheme maps onto possible vowels (Venezky, 1970); typically thirregular pronunciation is a small number phonetic features away from the regular pnunciation. Thus, HAVE is irregular, but /aelike /eI/, is a front unrounded vowel, not more distant vowel such as /oυ/. This generalpattern is also observed with other irregulapronounced vowels; for example, EA may pronounced as in BEAD, BREAD, anBREAK, all of which contain mid-to-highfront unrounded vowels (/i/, /«/, and /eI/, re-spectively). A word such as BEIGE is stranin the sense that it lacks immediate neighbobut the EI → /eI/ mapping is supported bother words in the lexicon (WEIGH, EIGHTHEIR). Finally, although vowel graphememap onto multiple phonemes in English, tpronunciations are typically cued by surrouning letters. The regularities that exist over tunits termed rimes (or word bodies) have bestudied extensively, but there are partial relarities involving other parts of words as we(Kessler & Treiman, 2001). In Monaghan aEllis’s (2002) stimuli, the alternative pronuncations of vowels were assigned independenof context.

These examples illustrate only some aspeof the statistical structure of words in EnglisThe important point is that the characteristicsthe stimuli in the Monaghan and Ellis (200simulation were quite different, even though tsimulation was intended to be relevant to contency effects in English. Their stimuli produclarge age of acquisition effects because tlacked the redundancy of English words.

In summary, all of the simulations of ageacquisition effects are consistent with the saconclusion: residual effects of age of acquistion skilled performance depend on the naturethe mapping between codes, specifically whetwhat is learned about early learned patterns cries over to later patterns. When the stimuli atask afford this type of learning, the netwo

UISITION 23

helen-ll

ef

o-,

lye

ers,

e-e

enu-lld-tly

cts.

of)eis-dey

fenoferr-d

encodes regularities across patterns that althe model to generalize, washing out the initadvantage for early trained words. Simulationsand 2 provide the most direct evidence conceing such effects in reading insofar as the modwas trained on a large corpus of words exhibitithe spelling–sound mappings characteristicEnglish. When the stimuli and task do not affothis type of learning (the Ellis and LamboRalph [2000] and Smith et al. [2001] simulationand Monaghan & Ellis’s [2002] inconsistencondition), the network is forced to memorizpatterns, yielding an advantage for early trainones. In this light, it is interesting to consider oSimulation 3 in which the Early and Late itemoverlapped among themselves but not acrlists. In this case, the model could generalifrom one Early item to another and from onLate item to another, but the orthogonal naturethe lists was such that the Late items as a growere learned suboptimally; the representatiodeveloped to support the Early items impedacquisition of the Late items.

It should be noted that our simulations dnot address all aspects of lexical processing aso cannot be taken as showing that AoA effecannot occur. The simulations involved knowedge of orthographic→ phonological corre-spondences, and we have argued that theyconsistent with behavioral studies of age of aquisition effects that used tasks, such as namand lexical decision, in which this knowledgplays an important role. The simulations sugest that the age at which this knowledge is aquired has little impact on skilled performancThe original age of acquisition hypothes(Brown & Watson, 1987; Morrison & Ellis,1995), however, concerned the effect of the aat which words are acquired in spoken laguage, an aspect of lexical learning that osimulations did not address. Acquiring a spken word vocabulary involves learning mapings between phonology and semantiSkilled reading often involves computationfrom orthography to phonology to semanti(for behavioral evidence, see Van Orden, Johston, & Hale, 1988; for a computational modesee Harm & Seidenberg, submitted). Hence,

S

a

os

ubthaanatthnceo

ithT

nvldirga-ths

dcc-

oh

ure

a

dsted

agethedis-iesma-or-evi-the allni-edveon,ter-is

both

rea-

tenap-nd

su-en

uir-u-an-

seheontheatialon-e-yji,an-dngr-nd

24 ZEVIN AND

mantics mappings could have a residual impon the orthography→ phonology→ semanticscomputation. None of the simulations of ageacquisition effects, including our own, addrethis possibility.

This issue needs to be examined in futresearch. However, two points shouldnoted. First, we have presented evidencethe results of existing behavioral studies cbe explained in terms of the impact of lexicfactors such as frequency, imageability, alength on word reading. Thus, it is not clewhether there is an age of acquisition effecbe explained further. Second, properties ofphonology→ semantics mapping make it ulikely to be the source of effects of age of aquisition on reading. The mapping betwethese codes is largely arbitrary for monomphemic words; words that overlap with thsound of the word CAT do not overlap within meaning. Thus, what is learned aboutphonology → semantics mapping for CAdoes not carry information that facilitatelearning the mapping for SAT or FAT. Givethe computational analysis presented abothis might seem like a condition that woupromote a strong age of acquisition effectspoken language acquisition, which in tucould affect reading via the shared phonolo→ semantics pathway. However, other charteristics of the phonology→ semantics mapping need to be taken into account. First,mapping between phonology and semanticnot entirely arbitrary; there are partial regulaities among many monomorphemic wor(e.g., correlations between the phonologicharacteristics of words and their grammaticlass [Kelly, 1992]). More important, inflectional and derivational morphemes make csistent (but quasiregular) contributions to tmeanings of many words (SeidenbergGonnerman, 2000). Second, both phonoloand semantics are themselves highly strtured; the words of a language occupystricted regions of the much larger spacepossible phonological forms or meanings. Aof these properties will facilitate the learninof mappings between phonology and sem

EIDENBERG

ct

fs

reeatnldrtoe

--nr-ete

s

e,

nnyc-

eis

r-salal

n-e

&gyc--

ofllgn-

reducing effects of the ages at which worare learned, as in the simulations presenabove.

Which Types of Knowledge Yield Age ofAcquisition Effects?

On our account, the key issue regarding of acquisition effects concerns the nature of stimuli and task being learned. The research cussed in this article, like the behavioral studdiscussed above, focused on the use of infortion concerning orthographic–phonological crespondences in English. The analyses of prous studies, the theoretical analysis of problem, and the results of the simulationssuggest that AoA effects are likely to be mimal in this domain. However, the modeling lto the identification of other conditions that girise to age of acquisition effects. The questithen, is whether these conditions are characistic of other types of human learning. Thissue needs to be considered further using behavioral and modeling approaches.

One obvious question is whether there aage of acquisition effects in reading nonalphbetic writing systems such as Chinese. WritChinese exhibits less consistency in the mping between written symbols (characters) atheir pronunciations. Chinese words are ually taught as arbitrary associations betwewritten words and meanings, a process reqing several years for the mastery of a few thosand characters. There may be a lasting advtage for early learned words in Chinebecause of the more arbitrary nature of tmapping. This unresolved empirical questineeds to be addressed carefully. Many ofearly learned words are nonarbitrary in ththey contain characters that provide partcues to pronunciation. The same need to ctrol for other correlated properties (e.g., frquency) will also arise. This is illustrated brecent studies of AoA effects in reading Kanthe Chinese characters that are part of Japese writing. Yamazaki, Ellis, Morrison, anLambon Ralph (1997) reported data indicatian AoA effect on Kanji naming; however, futher analyses by Yamada, Takashima, a

itcohh

v

n

s9tca

ngug

st eace wchsh

nnia

ry0c

inn

p

e-o a.g.,

nticgeo-re-

orm-hy

ngg.inin-

ver,of

-ordngsarlyre-a-

a as-hil-sonrn-ednotheinging

ion,nd

. Onreesentnallan-pre-er,ow

AGE OF AC

were at work. They found that the ease wwhich naive students could learn the pronunations of the characters in question was alsstrong predictor of naming latency. Thus, teffect seems to be due to stimulus factors otthan age of acquisition.

Age of acquisition effects have been obserin several tasks other than reading. Manythese studies are also subject to the methodoical concerns we have raised, but the findiare suggestive. One task that probably yiegenuine AoA effects is learning the names aciated with faces. Moore and Valentine (19studied this using faces rated for both subjecfrequency and AoA. The earlier acquired fawere named more quickly than the later quired faces, with subjective frequency cotrolled. Moore and Valentine (1999) also fouthat AoA effects in face naming were stronthan those in name reading. Lewis (1999) fosimilar effects with faces from long-runninsoap operas, where more objective controlthe time at which individuals came in and oupublic awareness were possible. WherMoore and Valentine attributed the effects to of acquisition, Lewis interpreted them as effeof cumulative frequency. Although further rsearch is needed, the effects are consistentthe theory presented here. Unlike words, faname pairs provide a strong test of the AoA pothesis because the earlier acquired itemnot vary predictably along other dimensions tmake them easier to learn or recognize. Asfrom partial phonological regularities in namgender (Cassidy, Kelly, & Sharoni 1998) avarious national/ethnic regularities (e.g., orarely meets an Italian named Wong), matchnames to faces is essentially an arbitrary mping in that what is learned early does not caover to later items.

Recent studies of Dutch by BrysbaeLange, and Van Wijnendaele (2000) and Brbaert, Van Wijnendaele, and De Deyne (20also yielded results consistent with our acount. They found larger effects of AoADutch on associate generation and semaclassification tasks than on word naming. Woassociations have an arbitrary learned com

QUISITION 25

hi-a

eer

edoflog-gsldsso-8)iveesc-n-dernd

ofofasgets-ith

e–y- doatideede

ngp-

rry

t,s-0)-

ticrdo-

as BREAD–BUTTER and HUSBAND–WIFEcannot be due simply to overlap in meaning bcause other pairs that overlap in meaning tsimilar degree are not highly associated (eBREAD–CAKE, HUSBAND–MAN). More-over, both associate generation and semaclassification tasks involve using knowledabout word meanings, not merely orthgraphic–phonological correspondences. Thelationship between form (orthographyphonology) and meaning is much less systeatic than the relationship between orthograpand phonology; words that overlap in spellitend to overlap in sound but not in meaninThus, the age of acquisition effects observedthese tasks may be related to the use of thisformation. Further research is needed, howeto determine more definitively whether ageacquisition has an effect on the orthography→semantics or phonology→ semantics mappings. Furthermore, any task that uses wmeanings is open to difficulties in establishithe chain of causality: Are early AoA wordeasy because they are early, or are they ebecause they are easy? This problem willquire some ingenious methodological innovtions before it can be solved.

Finally, consider the problem of learningsecond language. It is well known that somepects of language learning are easier for cdren than for adults (Flege et al., 1999; John& Newport, 1989). The second language leaing situation is one in which what is learnearly in experience (the first language) is highly predictive of what is to be learned in tlater phase (the second language). Assumthat both languages make use of overlappneural structures (for an interesting discusssee Perani et al., 1998), it follows that secolanguage learning should be disadvantagedthis view, so-called “sensitive period” effects aactually extreme cases of AoA effects—failurto learn in later life that reflect the entrenchmof early learned patterns—and not maturatiochanges in the neural substrate supporting guage acquisition, as has been classically sumed (Lenneberg, 1967; Neville & Baveli2000). Further progress in understanding h

E

a

26 ZEVIN AND S

life will be facilitated by examining tasks inwhich such effects are likely to be most poweful and by further exploring the computation

i,i-abiskhs

gpumt&inalitoinnnrmst Te

athnth ecc

ie usfo

mechanisms underlying these tasks.

CONCLUSIONS

The purpose of our research was to examage of acquisition effects on skilled readingtopic with potentially broad theoretical implcations that has been the focus of considerresearch. Ironically, the main conclusion to drawn from our research is that age of acqution effects are likely to occur, but for tasother than reading an alphabetic orthograpAge of acquisition effects reflect a loss of platicity associated with success in masterintask, a phenomenon that occurs in many tyof learning and species. The zebra finch’s scess in acquiring its characteristic song iposes significant constraints on its ability acquire additional vocal behavior (Doupe Kuhl, 1999). Similarly, the child’s success acquiring the phonological inventory or syntof a language may constrain the child’s abito learn other languages (Johnson & Newp1989; Werker & Tees, 1984). Issues concernthe nature and limits of plasticity in differedomains and their neural and computatiobases are central ones in cognitive neuscience. Connectionist models provide a coputational framework for understanding platicity in terms of the nature of the material be learned and how what is to be learned isfected by what has already been learned. entrenchment phenomenon discussed abovone outcome that occurs in such networks,we have taken a step toward specifying conditions that give rise to it. Under other coditions, other outcomes are observed. In reading case studied here, later learning iscilitated by prior knowledge rather than rstricted by it. In the catastrophic interferencase (McCloskey & Cohen, 1989), later sucess in learning results in forgetting of earlmaterial. Gaining a deeper understandingthe principles that govern the entire set of ocomes, and how they relate to the various tathat humans perform, is an important goal

IDENBERG

r-l

nea

bleei-

sy.-

aesc--

o

xyrt,g

talo--

-oaf-he is

nde-e

fa--e-roft-ksr

APPENDIX: STIMULI FOR ALLSIMULATIONS

Simulation 1

List 1 List 2 List 1 List 2

bail beast hatch hoardbay beet hound huntbelt bill maze mainbench bin moist matchbent bit mope mistblimp bleat pare parboard bound pinch pipebroil brag pool pursecap cab quit quenchcar care seem siftcheat cart serve sightclip chimp skirt skitcog clam slam slipcore coat street standcrass cool stuck stickcurse crab stunt strayface fail swift swervefeast fat tab tagfill felt tart tapfine fin tight teemfist flirt tin tentflit flog toil torefloat foist train tropegrab grace trick truck

Simulation 2

List 1 List 2

ache aislebeige boughbroad broochcaste chaisechic choirclique coupdraught ewefriend gaffegauge ghoulhearth heirhymn monkmonth myrrhpear phlegmpint plaidplaque psalmqueue realmrheum roguescheme scourgesew shoesieve skisponge swordtouch vaguevalse veldt

y.),

sfre-

r-

-s.

-sk.-

-

t:re-

y.

).c-ce

).ng

n

d la-

.is-

D.ic

f’s

n

AGE OF AC

Simulation 3

List 1 List 2

bad cobban cogbane copbat cubbate flogbid flopbide hogbin hopbit hubbite huckfad hugfade logfan luckfat lugfate plopfin pluckfine plugfit rollmad rugmade slobman slopmane sopmat stopmate stubmid stuck

Simulation 4

List 1 List 2

bad banbane batbate bidbide binbit bitecob cogcop cubfad fadefan fatfate finfine fitflog flophog hophub huckhug logluck lugmad mademan manemat matemid mitmite padpan panepat pinpine pitplopy pluck

QUISITION 27

REFERENCES

Baayen, R. H., Piepenbrock, R., & van Rijn, H. (1993).TheCELEX lexical database [CD-ROM]. Philadelphia: Uni-versity of Pennsylvania, Linguistic Data Consortium.

Balota, D. A. (1994). Visual word recognition: The journefrom features to meaning. In M. A. Gernsbacher (EdHandbook of psycholinguistics(pp. 303–356). SanDiego: Academic Press.

Balota, D. A., & Chumbley, J. (1984). Are lexical decisiona good measure of lexical access? The role of word quency in the neglected decision stage. Journal of Ex-perimental Psychology: Human Perception and Perfomance, 10,340–357.

Balota, D. A., Pilotti, M., & Cortese, M. J. (2001). Subjective frequency estimates for 2938 monosyllabic wordMemory & Cognition,29,639–647.

Balota, D. A., & Abrams, R. A. (1995). Mental chronometry: Beyond onset latencies in the lexical decision taJournal of Experimental Psychology: Learning, Memory, and Cognition, 21,1289–1302.

Bishop, C. M. (1995). Neural networks for pattern recognition. Oxford, UK: Clarendon.

Brown, G. D. A., & Watson, F. L. (1987). First in, first ouWord learning age and spoken word frequency as pdictors of word familiarity and word naming latencMemory & Cognition, 15,208–216.

Brysbaert, M., Lange, M., & Van Wijnendaele, I. (2000The effects of age-of-acquisition and frequency-of-ocurrence in visual word recognition: Further evidenfrom the Dutch language. European Journal of Cogni-tive Psychology, 12,65–85.

Brysbaert, M., Van Wijnendaele, I., & De Deyne, S. (2000Age-of-acquisition effects in semantic processitasks. Acta Psychologica, 104,215–226.

Butler, B., & Hains, S. (1979). Individual differences iword recognition latency. Memory & Cognition, 7,68–76.

Carroll, J. B., & White, M. N. (1973). Word frequency anage of acquisition as determiners of picture-namingtency. Quarterly Journal of Experimental Psychology,25,85–95.

Cassidy, K. W., Kelly, M. H., & Sharoni, L. (1998). Inferringgender from name phonology. Journal of ExperimentalPsychology: General, 128,362–381.

Coltheart, M., Curtis, B., Atkins, P., & Haller, M. (1993)Models of reading aloud: Dual-route and parallel-dtributed-processing approaches. Psychological Review,100,589–608.

Coltheart, M., Davelaar, E., Jonasson, K., & Besner,(1977). Access to the internal lexicon. In S. Dorn(Ed.), Attention and performance VI(pp. 535–555).Hillsdale, NJ: Erlbaum.

Coltheart, V., Laxon, V. J., & Keating, C. (1988). Effects oword imageability and age of acquisition on childrenreading. British Journal of Psychology, 79,1–12.

Doupe, A. J., & Kuhl, P. K. (1999). Birdsong and humaspeech: Common themes and mechanisms. Annual Re-

E

f-

i- oc:

e

al

c:

n

re

sr

,its

l

dl

ne

tIo

to

d

e

f-of a

.ola-

-y

n

ia:

-

i-

ll

:

i-

r-al

s?-

-us

28 ZEVIN AND S

Ellis, A., & Morrison, C. (1998). Real age-of-acquisition efects in lexical retrieval. Journal of Experimental Psychology: Learning, Memory, and Cognition, 24,515–523.

Ellis, A. W., & Lambon Ralph, M. A. (2000). Age of acqusition effects in adult lexical processing reflect lossplasticity in maturing systems: Insights from connetionist networks. Journal of Experimental PsychologyLearning, Memory, and Cognition, 26,1103–1123.

Flege, J. E., Yeni-Komshian, G. H., & Liu, S. (1999). Agconstraints on second-language acquisition. Journal ofMemory and Language, 41,78–104.

Forster, K. I., & Chambers, S. M. (1973). Lexical access naming time. Journal of Verbal Learning and VerbaBehavior, 12,627–635.

Gerhand, S., & Barry, C. (1998). Word frequency effectsoral reading are not merely age-of-acquisition effein disguise. Journal of Experimental PsychologyLearning, Memory, and Cognition, 24,267–283.

Gerhand, S., & Barry, C. (1999a). Age-of-acquisition afrequency effects in speeded word naming. Cognition,73,B27–B36.

Gerhand, S., & Barry, C. (1999b). Age of acquisition, wofrequency, and the role of phonology in the lexical dcision task. Memory & Cognition, 27,592–602.

Gernsbacher, M. A. (1984). Resolving 20 years of incontent interactions between lexical familiarity and othography, concreteness, and polysemy. Journal of Ex-perimental Psychology: General, 113,256–281.

Gilhooly, K. J., & Gilhooly, M. L. (1980). The validity ofage-of-acquisition ratings. British Journal of Psychol-ogy, 71,105–110.

Gilhooly, K. J., & Logie, R. H. (1980). Age-of-acquisitionimagery, concreteness, familiarity, and ambigumeasures for 1,944 words. Behavior Research Methodand Instruments, 12,395–427.

Harm, M. W. (1998). Division of labor in a computationamodel of visual word recognition. Unpublished doc-toral dissertation, University of Southern California.

Harm, M. W., & Seidenberg, M. S. (1999). Phonology, reaing, and dyslexia: Insights from connectionist modePsychological Review, 163,491–528.

Harm, M. W., & Seidenberg, M. S. (submitted). Computithe meanings of words in reading: Division of labor btween visual and phonological processes.

Hetherington, P., & Seidenberg, M. S. (1989). Is there “castrophic interference” in connectionist networks? Proceedings of the 11th annual conference of the Cnitive Science Society(pp. 26–33). Hillsdale, NJ: Erl-baum.

Hinton, G. E., & Shallice, T. (1991). Lesioning an attracnetwork: Investigations of acquired dyslexia. Psycho-logical Review, 98,74–95.

Hirsh, K. W., & Ellis, A. W. (1994). Age of acquisition anlexical processing in aphasia: A case study. CognitiveNeuropsychology, 11,435–458.

Hodgson, C., & Ellis, A. W. (1998). Last in, first to go: Agof acquisition and naming in the elderly. Brain and

IDENBERG

-

f-

nd

ints

d

d-

is--

y

-s.

g-

a-ng-

r

Johnson, J. S., & Newport, E. L. (1989). Critical period efects in second language learning: The influence maturational state on the acquisition of English assecond language. Cognitive Psychology, 21,60–99.

Kawamoto, A. H., Kello, C. T., Jones, R. J., & Bame, K(1998). Initial phoneme versus whole word criterion tinitiate pronunciation: Evidence based on response tency and initial phoneme duration. Journal of Experi-mental Psychology: Learning, Memory, and Cognition,24,862–885.

Kelly, M. H. (1992). Using sound to solve syntactic problems: The role of phonology in grammatical categorassignments. Psychological Review, 99,349–364.

Kessler, B., & Treiman, R. (2001). Relationships betweesounds and letters in English monosyllables. Journal ofMemory and Language, 24,592–617.

Kucera, H., & Francis, W. N. (1967). Computational analy-sis of present-day American English. Providence, RI:Brown University Press.

Lambon Ralph, M. A., Graham, K. S., Ellis, A. W., &Hodges, J. R. (1998). Naming in semantic dementWhat matters? Neuropsychologia, 36,775–784.

Lenneberg, E. H. (1967). Biological foundations of lan-guage. New York: Wiley.

Lewis, M. B. (1999). Are age-of-acquisition effects cumulative-frequency effects in disguise? A reply toMoore, Valentine, and Turner (1999). Cognition, 72,311–316.

Lyons, A., Teer, P., & Rubenstein, H. (1978). Age-at-acqusition and word recognition. Journal of Psycholinguis-tic Research, 7, 179–187.

Marchman, V. A., & Bates, E. (1994). Continuity in lexicaand morphological development: A test of the criticamass hypothesis. Journal of Child Language, 21,339–366.

Marcus, M., Santorini, B., & Marcinkiewicz, M. A.(1993). Building a large annotated corpus of EnglishThe Penn Treebank.Computational Linguistics, 19,313–330.

Markson, L., & Bloom, P. (1997). Evidence against a dedcated system for word learning in children. Nature,385,813–815.

McCandliss, B. D., Posner, M. I., & Givon, T. (1997). Brainplasticity in learning visual words. Cognitive Psychol-ogy, 33,88–110.

McCloskey, M., & Cohen, N. J. (1989). Catastrophic inteference in connectionist networks: The sequentilearning problem. In G. H. Bower (Ed.),The psychol-ogy of learning and motivation(Vol. 23, pp. 109–164).New York: Academic Press.

Monaghan, J., & Ellis, A. W. (2002). What exactly interactwith spelling–sound consistency in word namingJournal of Experimental Psychology: Learning, Memory, and Cognition, 28,183–206.

Moore, V., & Valentine, T. (1998). The effect of age of acquisition on speed and accuracy of naming famofaces. Quarterly Journal of Experimental Psychology,51A, 485–513.

Moore, V., & Valentine, T. (1999). The effects of age of

AGE OF ACQUISITION 29

acquisition in processing famous faces: Exploring thelocus and proposing a mechanism. In Proceedings ofthe twenty-first annual conference of the Cognitive

l-

in

-ley:

e aley

neIne:

ic S

,8io

atmin

to

A

inde

bn

Seidenberg, M. S., & Waters, G. S. (1989). Word recogni-tion and naming: A mega study. Bulletin of the Psycho-nomic Society, 27,489.

en-or

.T.

d-r-

5).

d

eo-n--

t- on

).e

-

echion

ctesea.

in

yce

Science Society(pp. 416–421). Mahwah, NJ: Erbaum.

Morrison, C. A., & Ellis, A. W. (2000). Real age of acquistion effects in word naming and lexical decisioBritish Journal of Psychology, 91,167–180.

Morrison, C. M., & Ellis, A. W. (1995). Roles of word frequency and age of acquisition in word naming and ical decision. Journal of Experimental PsychologLearning, Memory, and Cognition, 21,116–153.

Morrison, C. M., Ellis, A. W., & Chappell, T. D. (1997). Agof acquisition norms for a large set of object namestheir relation to adult estimates and other variabQuarterly Journal of Experimental PsychologHuman Experimental Psychology, 50A, 528–559.

Munro, P. W. (1986). State-dependent factors influence ral plasticity: A partial account of the critical period. J. L. McClelland, D. E. Rumelhart, & the PDP Rsearch Group (Eds.),Parallel distributed processingExplorations in the microstructure of cognition, Vol. 2(pp. 471–502). Cambridge, MA: MIT Press.

Neville, H. J., & Bavelier, D. (2000). Specificity and plastity in neurocognitive development in humans. In M.Gazzaniga (Ed.),The new cognitive neurosciences(pp.83–98). Cambridge, MA: MIT Press.

Perani, D., Paulesu, E., Galles, N., Dupoux, E., DehaeneBettinardi, V., Cappa, S., Fazio, F., & Mehler, J. (199The bilingual brain: Proficiency and age of acquisitof the second language. Brain, 121,1841–1852.

Plaut, D. C., McClelland, J. L., Seidenberg, M. S., & Pterson, K. E. (1996). Understanding normal and ipaired word reading: Computational principlesquasi-regular domains.Psychological Review, 103,56–115.

Quartz, S., & Sejnowski, T. J. (1997). The neural basiscognitive development: A constructivist manifesBrain and Behavioral Sciences, 20,537–596.

Seidenberg, M. S. (1995). Visual word recognition:overview. In P. Eimas & J. L. Miller (Eds.),Handbookof perception and cognition: Language. New York:Academic Press.

Seidenberg, M. S., & Gonnerman, L. M. (2000). Explainderivational morphology as the convergence of coTrends in Cognitive Sciences, 4, 353–361.

Seidenberg, M. S., & McClelland, J. L. (1989). A distriuted, developmental model of word recognition anaming. Psychological Review, 96,523–568.

-.

x-

nds.

:

u-

-

-.

S.,).n

--

of.

n

gs.

-d

Seidenberg, M. S., Waters, G. S., Barnes, M. A., & Tanhaus, M. K. (1984). When does irregular spelling pronunciation influence word recognition? Journal ofVerbal Learning and Verbal Behavior, 23,383–404.

Smith, M. A., Cottrell, G. W., & Anderson, K. L. (2001)The early word catches the weights. In T. K. Leen,G. Dietterich, & V. Tresp (Eds.),Advances in neural in-formation processing systems(Vol. 13, pp. 52–58).Cambridge, MA: MIT Press.

Spieler, D. H., & Balota, D. A. (1997). Connectionist moels of word naming: An examination of item level peformance. Psychological Science, 8, 411–416.

Strain, E., Patterson, K., & Seidenberg, M. S. (199Semantic effects in single-word naming.Journal ofExperimental Psychology: Learning, Memory, anCognition, 21, 1140–1154.

Strain, E., Patterson, K., & Seidenberg, M. S. (2002). Thries of word naming interact with spelling-sound cosistency. Journal of Experimental Psychology: Learning, Memory, and Cognition,28,207–214.

Turner, J. E., Valentine, T., & Ellis, A. W. (1998). Contrasing effects of age of acquisition and word frequencyauditory and visual lexical decision. Memory & Cogni-tion, 26,1282–1291.

Van Orden, G. C., Johnston, J. C., & Hale, B. L. (1988Word identification in reading proceeds from thspelling to sound to meaning.Journal of ExperimentalPsychology: Memory, Language, and Cognition, 14,371–386.

Venezky, R. L. (1970). The structure of English orthography. The Hague, Netherlands: Mouton.

Werker, J. F., & Tees, R. C. (1984). Cross-language speperception: Evidence for perceptual reorganizatduring the first year of life. Infant Behavior & Develop-ment, 7, 49–63.

Yamada, J., Takashima, H., & Yamazaki, H. (1998). Effeof ease-of-acquisition on naming latency for JapanKanji: A reanalysis of Yamazaki et al.’s (1997) datPsychological Reports, 83,991–1002.

Yamazaki, M., Ellis, A. W., Morrison, C. M., & LambonRalph, M. A. (1997). Two age of acquisition effects the reading of Japanese Kanji. British Journal of Psy-chology, 88,407–411.

Zeno, S. (Ed). (1995). The educator’s word frequencguide. Brewster, NJ: Touchstone Applied ScienAssociates.