A Cross-linguistic Study of the Development of Sentence InterpretationStrategies

Elizabeth BatesUniversity of California, San Diego

Brian MacWhinneyCarnegie-Mellon University

Cristina CaselliConsiglio Nazionale della Ricerche, Rome

Antonella DevescoviUniversity of Trieste

Francesco Natale and Valeria VenzaConsiglio Nazionale della Ricerche, Rome

bates, elizabeth; macwhinney, brian; caselli, cristina; devescovi, antonella; natale,francesco; and venza, valeria. A Cross-linguistic Study of the Development of SentenceInterpretation Strategies. Child Development, 1984, 55, 341-354. Studies of the developmentof sentence comprehension strategies in English have indicated that, at first, children tend to relyon pragmatic and semantic strategies, whereas, later on, they rely primarily on word order todetermine the basic grammatical relations. However, before making strong conclusionsregarding the role of semantics in comprehension, it is necessary to distinguish between (1)extragrammatical knowledge of world events and (2) abstract semantic distinctions that may be anintegral part of the parsing process. An earlier study of American and Italian adults indicated thatuse of such abstract semantic strategies may be a core strategy for parsing in Italian. In thepresent study, we compared sentence interpretation in American and Italian children betweenthe ages of 2 and 5. From the earliest stages, children showed sensitivity to the relative informa-tion value of the various cues in their native language; Italians relied primarily on semantic cues,whereas American children relied on word order. In general, the data did not support claimsregarding the existence of universal hypotheses about language structure. There was also evi-dence that the failure of these young children to make full use of certain interpretive cuesresulted from their inability to appreciate the discourse functions of these cues.

Although most researchers in the field ofchild language are interested in universalprocesses, the bulk of our knowledge of ac-quisition derives from studies of childrenacquiring English. As a result of these lim-itations in the data base, we run the risk ofelevating idiosyncratic facts about English tothe status of language universals. To see whythis is true, let us consider two recent pro-posals regarding universals in language ac-quisition.

Case Marking versus Word Order

The first of these two acquisitional uni-versals has been proposed most recently by

Pinker (1981, p. 78) in this form: "For case-inflected languages, children will utter sen-tences in the dominant word order, and willuse the dominant word order as a cue incomprehending sentences, before they havemastered their language's morphology."Similar proposals were offered on the basisof production data for both English (Brown,1973) and German (Stern & Stern, 1907).The principle also fits well with dataon sentence comprehension in English(Keeney & Wolfe, 1972). However, data fromother languages show that, in its strongestform, the developmental priority of wordorder over morphology cannot be a languageuniversal. For example, Hakuta (1982) has

This research was carried out with support from National Science Foundation linguisticsgrant BNS7905755 to Brian MacWhinney and Elizabeth Bates. We gratefully acknowledge thesupport of the Istituto di Psicologia, Consiglio Nazionale de la Ricerche in Rome, Italy. Requestsfor reprints should be sent to Brian MacWhinney, Department of Psychology, Carnegie-MellonUniversity, Pittsburgh, Pennsylvania 15213.

[Child Development, 1984, 55, 341-354. © 1984 by the Society for Research in Child Development, Inc.All rights reserved. 0009-3920/84/5502-0003$01.00]

342 Child Development

shown that young Japanese children acquireorder and inflectional cues simultaneously andcannot use one in the absence of the other. InPolish, Weist (in press) found that 2-year-oldscan make reliable use of both case markingand word order in sentence interpretation. Stillstronger evidence against the universalistperspective comes from a study of sentenceinterpretation in four languages by Slobin andBever (1982) and an investigation ofHungarian sentence comprehension byMacWhinney and Pleh (Note 1). In Slobin andBever's data for Turkish and MacWhinney andPleh's data for Hungarian, it was clear thatchildren had attained almost perfect use of thesuffixes marking the nominative/accusativedistinction by 2Vz years of age. However, thestatistically predominant order cue (subject-object-verb [SOV] in both of the languages)had no effect on comprehension until after theage of 4. Slobin and Bever (1982) concludethat there is no universal predisposition toprefer word order as a cue to sentenceinterpretation; rather, children use thosepatterns that are most regular and consisten

Word Order versus Lexical SemanticsThe second proposed universal in lan-

guage acquisition holds that there is a de-velopmental primacy of semantic strategiesover word order strategies. These two types ofstrategies converge on a common inter-pretation for the vast majority of naturallyoccurring noun-verb—noun (NVN) sentencesin English. However, there is a potentialconflict between strategies when listeners areasked to interpret such improbable sentencesas "The dog pats the mother."

Based on the performance ofAmerican children with probable andimprobable sentences in the active and thepassive voice, Bever (1970) proposed that 3-year-olds rely primarily on semantic strategies,whereas 4-year-olds make primary use of wordorder. In a further study of sentencecomprehension in 45 American children 2-5years of age, Strohner and Nelson (1974)concluded that, even at 2 and 3, childrencontrol a standard subject-verb-object (SVO)expectation (leading to correct performance onprobable or simple reversible actives).However, they argue that, until 4-5 years, thisword-order expectation is completelyovercome by semantic strategies. Because ofthis, the 2-3-year-olds overwhelmingly chosethe more probable actor in sentences such as"The baby feeds the

mother," whereas 4-5-year-olds shift to choiceof the first noun regardless of lexical relations.Note that the major conflict between theStrohner and Nelson study and the Bever studyinvolves the age of the first appearance ofword order strategies. Both studies find thatword order ultimately predominates oversemantic strategies and that semantic strategiesare stronger than word order strategies at first.

Chapman and Kohn (1978) suggestthat sentence interpretation strategies dependstrongly on the kind of semantic informationthat is set into competition with word order.Given lexical items such as "baby," "mother,"and "feed," the canonical relationship amongterms is so familiar and so powerful thatchildren might well act out the expected eventwithout any linguistic input at all. By contrast,pairs of sentences such as "The ball hit theboy" and "The boy hit the ball" are not subjectto the same amount of bias. In the world of apreschool child, it must seem equally plausiblefor balls to hit people as it is for people to hitballs.

We will refer to the contrast noted byChapman and Kohn (1978) as the contrastbetween strategies based on "lexicalsemantics" and those based on "event prob-abilities." If children have a general "agent-action" heuristic in sentence interpretation,then we should always expect the more humanand/or animate of two nouns to be chosen assentence subject. We say that a strategy of thissort is based on "lexical semantics," sincefeatures such as [+ animate] and [+ human] arecoded directly in the lexicon. A parsing systemthat depends on lexical semantics does notneed to reconstruct the likelihood of an event.It simply looks up the major class featuressuch as [+ animate] or [+ human] on the nounand creates a construal of the sentence on thisbasis. Alternatively, children may processsentences by relying on "local" event prob-abilities for specific combinations of objectsand actions. Thus, they might be quite con-sistent in interpreting strings with "Mommy,""baby," and "feed" (e.g., "The Mommy feedsthe baby," "Feeds baby Mommy," or "Thebaby was fed by the Mommy"), just becausethey know that mothers feed babies and thatbabies do not feed mothers. Chapman andKohn provided strong evidence supporting theimportance of this distinction betweenstrategies based on lexical semantics and thosebased on event probabilities. They controlledthe degree of event probability in theirstimulus

Bates et al. 343sentences and found no evidence for a generalsemantic strategy like "agent-action" at anyage level. Children were influenced by eventprobability for familiar items, but they reliedon word order (or performed randomly) foritems without strong a priori relationships. Thefinding that children rely on eventprobabilities in comprehension is, of course,nothing new (Bloom, 1974; Clark, 1973,1980). What is new here is the realization thatformal semantic strategies are fundamentallydifferent from strategies based on eventproabilities and that only the latter are clearlypresent in English-speaking preschoolers.

The Chapman and Kohn (1978)findings place limits on the generality of theprogression from semantic strategies to SVOthat had been noted by Bever (1970) andStrohner and Nelson (1974). Nevertheless, itdoes seem to be true in all three studies that,up to the age of 4, English-speaking childrenare swayed more by meaning relations than byword order. Only after that point is theirreliance on SVO order powerful enough toovercome predictable meaning relationships.

When we turn to the cross-linguisticdata on the acquisition of sentence com-prehension strategies, a markedly differentpicture begins to emerge. In French, Sinclairand Bronckart (1972) found that reliance onSVO word order increased from 2 to 7 years ofage. But interpretations based on eventprobabilities were also strong at all ages. Theyused sentences with the three possibleorderings of two nouns and a verb (noun-verb-noun [NVN], noun-noun-verb [NNV], andverb-noun-noun [VNN]) with the verb in theinfinitive. On NVN sentences with aninanimate first noun, such as "the bottle topush the boy," the most frequent interpretationwas patient-action-agent. Unlike English,where SVO word order eventually completelydominates over any alternative strategy, at nopoint in the Sinclair and Bronckart study didan SVO word order strategy "defeat" aprobable event strategy when the twostrategies were placed into direct competition.Thus, there is reason to believe that, even inlanguages as closely related as English andFrench, children may be learning quitedifferent ways of parsing sentences.Unfortunately, Sinclair and Bronckart did notinclude in their stimuli sentences with ananimacy contrast that did not also representhigh-probability activities such as opening abottle. In other words, there were no sentences

like "The dog the pencil pushes" in theirexperiment. Inclusion of such sentences wouldallow us to distinguish application of eventprobabilities from more formal strategies basedon lexical semantics.

A recent study by Hakuta (1982)provides us with clearer evidence regarding thestatus of lexical semantic strategies inlanguages other than English. Hakuta's stimuliwere created by random assignment of nounssuch as "tiger," "goat," "chair," and "banana"to verbs such as "push" and "chase." He foundthat, when word order and lexical semanticswere placed in direct competition in Japanese,lexical semantics dominated quite clearly.Hakuta's stimuli were designed in a way thatrequired use of formal" semantic strategies, notjust strategies based on probable events. Whenhis experiment is compared with that ofChapman and Kohn (1978) in English, thereseems to be evidence for marked differencesbetween Japanese and English in regard to theuse of formal lexical strategies in sentencecomprehension. Unfortunately, Hakuta did notreport the results for the competition betweenanimacy and word order in full detail(apparently because the role of animacy was soabsolute), nor did he present comparisonresults for Japanese adults. However, hisresults certainly suggest the possibility that, insome languages, formal semantic strategiesmay be acquired well before word orderstrategies.

It is important to collect adult data inexperiments like these for several reasons.First, the discrepancy between the studies inEnglish and those in other languages mightwell reflect differences in the target languages.Certainly, we cannot assume that syntacticstrategies will prevail in all adult samples,since adults (including English-speakingadults) have been shown to use semanticinformation in sentence interpretation undersome conditions (e.g., Gleitman & Gleitman,1970, in a study on the interpretation ofnominal compounds). Second, it has beenargued that comprehension experiments of thissort have no ecological validity; when childrenare presented with semi-grammatical oranomalous stimuli, they may respond in bizarreways that have nothing to do with their day-to-day use of the grammar (Bridges, 1980; Cromer,1976). However, if we can show that adults fromdifferent languages respond to such stimuli inconsistent and language-specific ways, then itseems fair to conclude that interesting facts aboutthe target language may influence the behavior ofchildren under similar circumstances. This is

344 Child Developmentparticularly true if the language-specific out-comes of laboratory experiments relate in atheoretically coherent way to facts about lan-guage use in natural circumstances.

English, Italian, and the Competition ModelTo provide a concrete instance of the

ways in which a variety of cues can interact insentence interpretation, let us take a look attwo related languages—English and Italian.Both languages use SVO word order withoutcase inflections to indicate basic semantic-syntactic relations. However, in Italian, allpossible word orders can and do occur undercertain conditions, including oddities like "Thelasagna ate Giovanni" (OVS). In deviationsfrom SVO, assignment of semantic-syntacticroles must be based on some combination ofcues other than word order: number of personagreement with the verb, the presence of aclitic object pronoun, con-trastive stress tosuspend "default" interpretation, or a variety ofsemantic-pragmatic cues, in the verbal andnonverbal context, that make it obvious whodid what to whom (Bates, 1976; Duranti &Ochs, 1979). This means that Italian listenersmust be prepared to integrate semantic-pragmatic information directly into the parsingprocess.

By contrast, word order is rigidly pre-served in English, although some variations arepossible in informal conversation. Specifically,informal English permits OSV topicalizations(e.g., "Egg creams I like") and VOStopicalizations (e.g., "Really gets on mynerves, that guy"). Although these nonca-nonical orders are rare, they do occur inspontaneous speech and therefore should beavailable as schemata for the interpretation ofnonstandard orders in comprehension. Bycomparison, SOV and VSO are not permissiblein modern English under any circumstances(e.g., "I egg creams like" or "Really gets on,that guy, my nerves"). Furthermore, insentences such as "I petted the cat Mary liked,"the relative clause is an OSV fragment. Inimperatives such as "Kick the ball, Harry," wesee the use of a VOS order. This means that,for simple declarative sentences with anycombination of two nouns and a verb (NVN,NNV, or VNN), English-speaking listeners canalways find a model of that word order in theirlanguage (i.e., SVO, OSV, or VOS) that yieldsa single consistent interpretation of who didwhat to whom. In Italian, on the other hand, allpossible word orders exist, but none yields asingle, consistently correct interpretation ofwho was the actor and who was acted upon.

The effect of these informallanguage differences on sentence processingout of context has been demonstrated byBates, McNew, MacWhinney, Devescovi,and Smith (1982). Middle-class Italian andAmerican adults were asked to interpret 81grammatical and semigrammatical sentenceswith orthogonal variations of word order,lexical animacy, contrastive stress, and top-icalization. The major result was that Italiansand Americans used completely oppositestrategies to interpret the same sentences.That is, given a sentence such as "the pencilkicks the cow," Italians choose the cow assubject while Americans choose the pencil.The main effect of animacy in Italian ac-counted for 42% of the variance, comparedwith 3% in English; the main effect of wordorder in English accounted for 51% of thevariance, compared with 4% in Italian. Onemight argue that Italians were simply "stuckfor an answer," since word order is ambigu-ous in their language, and they had to carryout special extragrammatical computations tointerpret these sentences. However, regard-less of word order, reaction times for the Ital-ians were very fast if an animacy contrast waspresent, suggesting that the semantic dis-tinction between animate and inanimatenouns is a readily available part of the Italianparsing process. There was also a tendency inboth languages to choose the topicalizedelement as the actor. Contrastive stress hadan effect in that reactions were faster andclearer for NVN items with default stress (thepragmatically neutral, "best" combination)and for VNN and NNV items with contrastivestress (the pragmatically marked, "best" vari-ations). However, the effect of these prag-matic cues was significantly greater in Ita-lian. In fact, in the absence of semantic orpragmatic cues, Italians gave extremely slowand inconsistent responses—even on the ca-nonical word order.

Elsewhere (Bates & MacWhinney,1982), we have developed a model of thefunctional control of sentence processingcalled the "competition model." Early di-vergences between children exposed to dif-ferent language structures are directly pre-dicted by this functionalist approach togrammatical development. Thus, we wouldexpect English-speaking 2-year-olds to relyextensively on word order in processing andItalian-speaking 2-year-olds to rely exten-sively on semantic and pragmatic strategies.

According to the competition model, de-cisions in sentence interpretation are madeby evaluating the relative weights of the cuespresent in the stimulus. Similar models for

Bates et al. 345comprehension are currently espoused byAnderson (1983) and Thibadeau, Just, andCarpenter (1982). These models assign norole to universal predispositions or to under-lying separations of autonomous processingcomponents (Forster, 1979). In such compe-tition models, semantic contrasts are inte-grated into the parsing system on an equalfooting with traditional grammatical cues.Proceeding from a model of this type, ourdevelopmental model (Bates & MacWhin-ney, 1982; MacWhinney, 1978; MacWhin-ney, Bates, & Kliegl, in press) claims that theage of acquisition of cues to sentence inter-pretation is determined by the relative "cuevalidity" (Brunswik, 1956) of those structuresin the target language. A functionalist/competition model of this type is proposed asa theoretical baseline for research in this area.To the degree that evidence can be gatheredfor universals, predispositions, or separationsof components, we may be required to makeour developmental model more complex. Ingeneral, we believe that it will be necessaryto account for both the creativity and theplasticity of the acquisitional processes usedby children while they are discoveringgrammar.

Method

Subjects.—Eighty children participatedin the study: 40 were English-speaking chil-dren attending middle-class preschools inthe Denver area; 40 were Italian-speakingchildren attending middle-class preschoolsaround Rome. Within each language group,there were 10 children (five males and fivefemales) at each of four age levels: 2Vz, 31A,4J/2, 5Vz. The children's ages varied within 2weeks of the half-year mark. These fourgroups will be referred to as 2-year-olds, 3-year-olds, 4-year-olds, and 5-year-olds,although the actual age range is narrowly re-stricted.

Materials.—Each child received a totalof 54 grammatical or semigrammaticalsentences to enact. There were twosentences at each level of a 3 x 3 x 3 design,representing orthogonal combinations ofword order (NVN, VNN, NNV), animacy(animate/animate, animate/inanimate, in-animate/animate), and contrastive stress (de-fault stress only, contrastive stress on firstnoun, contrastive stress on second noun).Each sentence consisted of two singularcommon nouns with definite articles and atransitive action verb in the present tense. Allverbs were given in the third-person singu-lar. The animate nouns were taken from a setof 18 names for familiar animals (dog, goat,

monkey, etc.). The inanimate nouns weretaken from a set of nine simple objects (rock,ball, pencil, etc.). The verbs were selectedfrom a group of nine verbs representing sim-ple actions (eats, bites, grabs, etc.). A total of40 different protocols were constructed, eachwith a unique random assignment of animatenouns, inanimate nouns, and verbs to appro-priate conditions. This procedure insuredthat specific event probabilities resultingfrom particular lexical combinations (e.g.,lions are inherently fiercer than sheep) werecompletely homogenized across children andconditions. The between-sentence random-ization procedure was exactly the same as theone described for American and Italian adultsin Bates et al. (1982). The 40 protocols weretranslated into matching English and Italianforms, so that each protocol could be ran-domly assigned to one Italian and one Ameri-can child.

Experimental objects included smallplastic versions of the 18 animals (approxi-mately 2 inches in height), plus exemplars ofthe nine inanimate objects (which wereroughly equivalent in size to the toy animals).A plastic cheese tray with a transparent plas-tic dome was used to present the objects tothe child while each sentence was readaloud.

Procedure.—All children were testedindividually in their preschools, except for asvibset of the 2-year-olds in each languagegroup who had to be tested individually inthe home. Italian children were tested by na-tive speakers of Italian; Americans weretested by native speakers of English. Onlyfemale experimenters were used.

The experimenter began the testing ses-sion by introducing the experimental toys tothe children, asking them to name each one,and repeating the object name (whether ornot the children had named it correctly).Then she explained that they were going toplay a game in which the children must actout (i.e., "Show me with the toys") everysentence uttered by the experimenter. Beforeeach stimulus sentence was read, the two as-sociated objects for that item were placed onthe plastic tray under the clear plastic dome.This insured that children would watch theobjects "on stage" without touching themuntil the entire sentence was delivered. Tominimize bias introduced by place-directedresponses (e.g., picking the animal on theright, or picking the nearest object), toys wereplaced on the tray in random positions fol-lowing a clockwise circular pattern acrosstrials (i.e., horizontal from the children's

346 Child Development

perspective on one trial, diagonal on another,vertical on another, etc.). Each sentence wasread aloud to the children, using a stan-dardized and exaggerated intonational pat-tern to underscore the contrast between de-fault and contrastive stress (as in Bates et al.,1982). Then the clear plastic dome was re-moved, and the children were told to "showme." If a child had clearly forgotten thesentence, or asked for a repetition, thesentence was read a second time without re-placing the dome. This procedure greatly re-duced the number of idiosyncratic responses(e.g., the children acting as agents on both toyobjects) and place-oriented errors that havebeen reported in many studies using theenactment method (for a review, see Bridges,1980).

Scoring.—While one experimenter pre-sented stimuli, a second wrote down thechildren's behaviors with the toys on eachtrial. With a few exceptions, the first objectpicked up and moved by the children wasconsidered to be the object chosen as agent.Exceptions included cases where the chil-dren clearly indicated having changed theirminds by setting the first one down andpicking up the second in a clear agentive mo-tion (e.g., shoving it heavily against the firstwhen instructed to make one object hit orpush another). There were also some idio-syncratic exceptions for cases in which thefirst object moved was quite clearly intendedas the patient. An example includes a case inwhich a child responded to "eat" or "kiss" byholding the body of the patient toy up againstthe agent animal's mouth.

The level of uninterpretable responsesobtained in this experiment was consistentlylower than that reported by other in-vestigators, perhaps because of some of themethodological changes we introduced (e.g.,use of the plastic dome during sentence pre-sentation). Even in the youngest age groups,no child fell below 50% interpretable be-haviors across trials. For all but two children,at least 86% of the responses were interpret-able.

ResultsThe overall design of the experiment

permitted a 2 x 4 x 3 x 3 x 3 analysis ofvariance (language x age x word order xanimacy x stress) with language and age asbetween-subjects variables, and order, ani-macy, and stress as repeated measures.Although we conducted a full five-factorialANOVA and two four-factorial ANOVAs foreach language, the results are most clearly

understood by looking at the eight three-factorial ANOVAs for the four age groups inthe two languages. Each of these analyseswas based on data from 10 subjects. Theseeight analyses permitted us (a) to determinethe age level in which each effect firstreached significance and (b) to calculate theamount of variance accounted for (sum ofsquares effect/sum of squares total from theANOVA) by different grammatical cues andcombinations of cues at each point in de-velopment.

Figures 1, 2, and 3 illustrate the inter-actions between order and animacy at eachage level. Figure 4 displays the percentage ofvariance accounted for in each age and lan-guage group. In all these figures, corre-sponding values for American and Italianadults from Bates etal. (1982) are included forcomparison with the developmental findings.

Two-year-olds.—Among the youngestAmerican children, there is a significant maineffect of word order, F(2 ,18) = 4 .06,p < .05. Figure 1 displays the percentage ofchoice of the first noun in NVN sentences ateach age. Figures 2 and 3 display choices forVNN and NNV sentences, respectively.Comparing Figure 1 with Figures 2 and 3, wesee that the word order effect is primarilybecause of a first noun strategy with NVNitems. Performance on the other two wordorders is essentially random when summedacross conditions. There is also a significantword order x animacy interaction, F(4,36) =3.42, p < .05, reflecting an effect of animacyprimarily in NNV (Figure 3), where there is atendency toward an OSV interpretation whenboth nouns are animate. As we shall see, thisis the only point in the entire study in whichAmerican children show some sign of thestrong OSV and VOS strategies that Bates etal. (1982) report for American adults. The factthat an OSV pattern occurs only in theyoungest group and only with reversiblesentences is an anomaly that we must con-sider in more detail later in this paper.

Among the youngest Italian children, theonly significant effect is a main effect of ani-macy, F(2,18) = 5.04, p < .05. This effect is inthe predicted direction: children are morelikely to choose the animate noun on non-reversible items, regardless of word order. Inshort, from the very first age group there isalready a language-specific difference be-tween American and Italian children.American children rely primarily on wordorder (although there is an interaction withanimacy in NNV sentences); Italian children

Bates et al. 347rely primarily on animacy as a lexical seman-tic cue. Whereas the most important effect inEnglish at this time is word order, which ac-counts for 3.5% of the variance, the most im-portant effect in Italian is animacy, whichaccounts for 17.5% of the variance (see Figure4).

Three-year-olds.—By this age, the wordorder main effect among American childrenis quite a bit larger than it wa,s at age 2 (17.3%vs. 3.5% of the variance ),F (2,18) = 26.37, p <.001. The word order x animacy interactionthat was significant at age 2 is no longersignificant at this age.

Things look quite different for the 3-year-old Italians. There are now significantmain effects for both word order, F(2,18) =14.15,p < .001, and animacy, F(2,18) = 13.28,p < .001, but the animacy effect is muchlarger (34% vs. 5.5% of the total variance—seeFigure 4). The word order effect is in thesame direction as the one we find for Ameri-can children at every age—a first-noun strat-egy with NVN items, with performancearound chance on the other two orders when

summed across conditions. Finally, there is asignificant order x stress interaction, F(4,36)= 3.59, p < .05, among 3-year-old Italians,with stress having its effect primarily in theinterpretation of noncanonical (VNN andNNV) orders. However, this stress effect runsin the opposite direction from stress effectsreported by Bates et al. (1982) for Italianadults. In the adult study, listeners tended toavoid choosing the stressed noun as subject(presumably on the assumption that the sub-ject is usually old information, whereas con-trastive stress indicates new information).Children, instead, tended to choose thestressed noun as subject. We suggest that thisis essentially a nonlinguistic strategy; "If youdon't know what else to do, pick the one thatthe experimenter says loudest." Such aninterpretation is compatible with a report onstress effects in older American children byMacWhinney and Price (1980). In that study,second graders were asked to point to thepicture that corresponded to a sentence.They tended to point to the picture that con-tained the object that was stressed in thesentence. Seventh graders, like the adults in

348 Child DevelopmentBates et al. (1982), pointed at the picture thatdid not contain the stressed item.

Four-year-olds.—Among the American4-year-olds, all three main effects reachedsignificance: word order, F(2,18) = 29.03,p < .001; animacy, F(2,18) = 16.57, p < .001;and stress, F(2,18) = 3.78, p < .05. Thesethree factors account for 30.7%, 6.4%, and1.0% of the variance, respectively. The inter-action of word order with animacy was alsosignificant, F(4,36) = 5.87, p < .01, account-ing for 5.0% of the variance. This interactionreflects a tendency for animacy to have itseffect primarily on NNV items (see Figure 3).Children tended to pick the first noun onNNVs when the animate was first or when theitems were reversible. They tended to pickthe second noun on NNVs when the secondnoun was animate. By contrast, performanceon VNN items was essentially random at thisand all other age levels in English. Thisinteraction is entirely different from the onefound at 2. At 2, children picked the secondnoun as the subject in reversible sentences.At 4, they have shifted to a first-noun strategy

in NNV and pick the first noun in reversiblesentences. This shift in strategies is reflectedin a significant three-way interaction of wordorder x animacy x age, F(12,144) = 2.95,p < .01, in the four-way ANOVA for English.This complex pattern of responses on NNVsentences seems to be evidence for a de-velopmental shift between word order strat-egies. Rather than using the cue of the rela-tive position of the noun before the verb asthey did at 2, 4-year-olds are using the abso-lute cue of position at the beginning of thesentence. Such a transition could be repre-sented in formal terms as the movement froma loose system of positional patterns (Mac-Whinney, 1982) to a more organized parsingsystem (Marcus, 1980).

Among the 4-year-old Italians, there aresignificant main effects of animacy, F(2,18) =13.00, p < .001, and word order, F(2,18) =25.46, p < .001, along with a significant wordorder x animacy interaction, F(4,36) = 7.10,p < .001. Animacy accounts for 31% of thevariance, word order for 9.0%, and the interac-tion for 4.0%. The pattern that emerges here is

Bates et al. 349the most surprising finding in the study.Although word order remains a secondarystrategy (i.e., it accounts for less variance thananimacy), there is nevertheless more evi-dence here for overgeneralization from SVOthan we have in any English-speaking group.In particular, Italian children seem to havedeveloped a strategy of taking the first nounas the agent. As can be seen in Figures 1-3,Italian 4-year-olds actually make more use ofword order than Italian adults. First of all,they tend to choose the first noun on VNN andNNV orders (although this tendency is muchsmaller when there is conflicting informationfrom animacy). Since this pattern does notcorrespond to any clear SOV or VSO bias inthe adult model, it seems fair to conclude thatchildren are overgeneralizing from the ca-nonical SVO word order. Unlike their agemates in English, the Italians seem to haveovergeneralized in a way that takes the firstnoun as the agent even when that noun is notsentence-initial. A second piece of evidencefor overgeneralization from SVO order iscontained in the relative effects of animacy

and word order. On improbable NVN itemssuch as "The pencil kicks the cow," wordorder "wins" over animacy in this age groupwhereas animacy clearly "wins" over wordorder in the adult sample. This diminution ofrole of animacy in NVN sentences thatemerges at 4 is reflected in a significant age xanimacy x word order interaction, F(12,144)= 1.87, p < .05, in the four-way ANOVA forItalian. Between the ages of 3Vz and 4Vz, theseItalian children seem to have developed anovergeneralization from SVO order at least asstrong as that observed by Bever (1970) andreplicated by Maratsos (1974). Of course, Be-ver's overgeneralization data were based onperformance with English passives, andthere were no passive sentences in our study.But it is interesting that analogous patterns ofovergeneralization occur in both languagegroups, at the same age level. This is particu-larly striking since a first-noun strategy onimprobable NVN items runs directly counterto the Italian adult pattern.

Finally, there is no main effect of stress

350 Child Developmentin these 4-year-old Italian children, nor arethere any interactions with stress. The factthat the 3-year-old stress strategy has disap-peared in this group supports our view thatthe earlier stress effect was an immature,nonlinguistic strategy at odds with adult useof the same information. The disappearanceof this use of stress between 3 and 4 is re-flected in a significant age x stress x wordorder interaction, F(12,144) = 2.05, p < .05, inthe four-way ANOVA for Italian.

Five-year-olds.—In both languagegroups, performance was essentially thesame at 5 as it was at 4. In English, all threemain effects again reached significance: wordorder, F(2,18) = 18.78, p < .01; animacy,F(2,18) = 4.65, p < .05; and stress, F(2,18) =3.74, p < .05. There was also a significantorder x animacy interaction, F(2,18) = 5.73, p< .001.

The only change compared with the pre-ceding age level is that the tendency to gen-eralize a first-noun strategy to NNV items hasbegun to drop (Figure 3). This may mean that5-year-olds are beginning to adopt the

second-noun, OSV strategy that is so clearlyevidenced by American adults.

In Italian, we also find the same patternsthat operated at 4: a strong animacy effect,F(2,18) = 19.05, p < .001 (30.5% of the vari-ance); a weaker word order effect, F(2,18) =13.71, p < .001 (9.7% of the variance); and asignificant order X animacy interaction,F(4,36) = 5.16, p < .01 (reflecting that ordereffects are greatest when there is no conflictwith animacy). The only way the 5-year-oldsdiffer from the 4-year-olds is in the increaseduse of the first-noun strategy on reversibleNNV items. This indicates that the strategy ofselecting the first noun as agent is still strong,but has become coordinated with the ani-macy strategy.

To summarize, from the earliest agegroup we have evidence of differentiationbetween American and Italian children in thedirection of the respective adult models.Word order accounts for more variance thanany other cue in English, at every age. Ani-macy is the largest effect in Italian, at everyage. These facts are reflected in the fact that,

6O

Bates et al. 351in the five-way ANOVA, the interactionsof language X word order and language xanimacy are both clearly significant. How-ever, a more surprising finding is the over-generalization of SVO by 4- and 5-year-oldItalians—a much larger reliance on wordorder than we find in educated Italian adults.What we have here is a constant reliance onanimacy at every age, tempered by a tempo-rary resurgence of word order in the 4-5-year-old range. We also have evidence thatchildren in these two languages are for-mulating word order strategies of severaldifferent types. It seems that American 2-year-olds use simple positional relations inprocessing of NNV. Later, English-speakingchildren develop initial noun strategies. InItalian, however, children formulate theirword order rules in terms of the first nounbeing the agent, particularly when that nounis animate.

The main conclusion from this experi-ment is that children are sensitive from thebeginning to the information value of cues intheir particular language. Moreover, thissensitivity seems to be the most importantsingle factor accounting for the developmentof sentence comprehension. This result isparticularly clear in Figure 4, where the ab-solute magnitude of word order and animacyeffects is plotted without regard for direction.Word order is the most important cue tosentence interpretation in English, at everyage level. Animacy is the major cue tosentence interpretation in Italian, again atevery age level.

We should stress that the animacy ma-nipulation in this experiment requires mas-tery of an abstract semantic contrast betweenanimate and inanimate. There are no local,high-probability relationships in the chil-dren's experience to connect pencils andcows, erasers and zebras. From this point ofview, our results are compatible with Chap-man and Kohn (1978), who report that Ameri-can children rely exclusively on word order,unless items involve highly probable re-lationships such as those connecting "baby,""feed," and "Mommy." From the same pointof view, the consistent behavior of Italianchildren is all the more impressive. Theirperformance in this experiment must neces-sarily involve use of an abstract semanticcontrast that has very little effect on Americanchildren. Any attempt to view this contrast asbased on nonlinguistic strategies must fail,since there is no evidence for the use of anysuch strategy even at the youngest ages in soclosely related a language as English.

Although the results of this study do notfit well with approaches that attempt to em-phasize the separation between syntax andsemantics, they are entirely compatible withthe functionalist approach to grammar ar-ticulated in Bates and MacWhinney (1982).According to that approach, languages andchildren can integrate abstract semantic cuesdirectly into the parsing process. The resultsof the current study are also quite compatiblewith models that account for order of acquisi-tion in terms of cue validity such as the onepresented in MacWhinney (1978) and Mac-Whinney et al. (in press).

Remaining IssuesThere remain three results of this study

that defy a straightforward explanation interms of cue validity alone: (1) the non-adult-like use of stress in Italian 3-year-olds,(2) the overgeneralization of SVO order thatemerges in Italian 4-year-olds, and (3) thefailure of English-speaking 4- and 5-year-oldsto adopt the second noun strategy used byEnglish-speaking adults. We believe a singleaccount can provide an explanation of each ofthese three phenomena, along with a varietyof related phenomena in the language de-velopment literature.

We begin by noting that, below the age of7, children seem to have difficulty with thoseaspects of language that serve an inter-sentential, discourse function in the adultgrammar. For example, Karmiloff-Smith(1982) has noted certain reorganizations thattake place around 7 years of age in the use ofpronouns and determiners by French chil-dren. She argues that children begin with asentence-level grammar that is designed toconvey information about individual events.Hence, they may appear to use pronouns anddeterminers correctly, but only in the serviceof "local" functions. For example, the use ofpronouns and articles in her preschool sam-ple was constrained primarily by "exophoricfactors"—that is, whether the referent wasactually in the room or not. After 7 years ofage, children seem to be sensitive to thecohesion of texts and narratives, so that pro-nouns and determiners come under the con-trol of new, anaphoric constraints. From thispoint of view, there is one thing that all thedeviations and omissions by the children inour sample have in common: they are avoid-ing those aspects of word order variation andcontrastive stress that are under discoursecontrol in the adult languages.

To trace out the impact of the emerging

352 Child Development

attention to discourse function on the use ofstress in Italian, we repeated our experimentwith a sample of 10 7V2-year-old Italian chil-dren and analyzed the results in anotherrepeated-measures analysis of variance(order X animacy X stress). The order, ani-macy, and order x animacy effects wereidentical to those obtained with S'/a-year-oldsin the main experiment; the overgeneraliza-tion of SVO that we have reported here doesnot disappear by 7Vz. However, among theseolder children, stress had begun to functionas in the adult speaker with the stressed nounbeing chosen as the subject. This exactlyreplicates the findings of MacWhinney andPrice (1980) who had also argued that adultuse of contrastive stress involves rules thatoperate above the level of the individualsentence, since they involve intersententialcoordinations within a discourse grammar.Although preschool children do use stress ina "local" way to mark emphasis at thesentence level, complete mastery of the dis-course functions of stress may take years todevelop.

The second result that seems at first todefy a straightforward explanation in terms ofcue validity is the overgeneralization of SVOby Italian children. We believe that thisovergeneralization occurs because thesechildren do not yet understand thepragmatic/discourse function of word ordervariations in their language—or at least theydo not understand them well enough to applyconsistent discourse strategies, out of con-text, in an experimental setting. When Italianadults were presented with the same stimuli,they used discourse as well as semantic cuesto disambiguate word order. When there wasno contextual information of any kind, adultsperformed at nearly random levels on thenoncanonical orders and were not even par-ticularly consistent on the NVNs (presum-ably because OVS interpretations still"sounded right" on some items). PerhapsItalian adults do not overgeneralize SVO theway children do because they know better,even at the price of random performance. IfItalian children do not have this additionaldiscourse knowledge, there is nothing toblock them from the simplifying assumption(based on the distribution of SVO orders inthe adult language) that the first noun is usu-ally the subject.

Overgeneralization of SVO in sentenceproduction during this period has also beenreported for Italian by Bates (1976) and forSlavic by Radulovics (1975) and Slobin(1966). It is also interesting that the onset

time for overgeneralization of order in com-prehension coincides with Bever's (1970) re-port on the emergence of the SVO word orderstrategy in American children. Of course, Be-ver's data involved overgeneralization of afirst-noun strategy to the English passive, andwe have no passive constructions in this ex-periment. Nevertheless, it is at least possible,in the light of the English-Italian comparison,that children are able to reflect on basic wordorder in a new way somewhere in the periodfrom 4 to 7 years of age.

Third, the English word order data alsopoint out the fragmentary, nonintegratednature of early comprehension. Although theAmerican children controlled SVO from 2years of age onward, we had very little evi-dence in this study for the kind of second-noun strategies we find in American adults.The adult second-noun strategy for NNVsentences may be based on use of the OSVstructure of object relative clauses. There wasa similar tendency to OSV interpretation inthe 2-year-olds, but it makes little sense toview this pattern as the acquisition and sub-sequent loss of a pattern for interpreting OSVtopicalizations. Rather, it appears thatEnglish-speaking children first acquire SVOword order in terms of its component pieces:SV order and VO order. In the NNVsequence, only the SV pattern is satisfied. Byage 3j children have already begun to go be-yond this fragmentary approach to sentenceprocessing.

More surprising than the disappearanceof a second-noun strategy is that Americanchildren show very little evidence for anovergeneralized first-noun strategy. Theonly place where such generalization occursis on reversible NNV items at 4 years—atendency that has already begun to diminishby 5. In fact, there was less overgeneraliza-tion from SVO to other word orders in En-glish than there was in Italian—even thoughAmerican children rely more on word orderoverall. It is possible that, after 3, childrenhave knowledge of both a fragmentary SVpattern and a general first-noun strategy,which cancel each other out in NNV orders.This interpretation is bolstered by the factthat 4-5-year-olds were very likely to choosethe second noun as the actor on IAV con-structions. Evidently, in this case, they aresupported by the convergence betweentopicalization structure, OSV structure inobject relative clauses, and animacy. Notealso that the only evidence for English over-generalization of SVO in Bever's (1970)original studies occurred on passive

Bates et al. 353sentences. Since passives do follow an NVNorder, they would presumably not be "canceledout" by children's sensitivity to adult rightand/or left topicalization.

To summarize, we have found that,from the age of 2Yz, child speakers of Englishand Italian are remarkably attuned to the inter-pretive cues provided by their respectivelanguages. Thus, the major determinant of thedevelopment of sentence processing strategiesappears to be cue validity. Yet, there arecertain cues that seem to be inadequatelyprocessed by the young child. These lead toerrors and overgeneralizations that can beexplained if we assume that, before the age of6 or 7, children do not have control of thediscourse functions that "make sense" out ofpragmatic word order variations.

1. MacWhinney, B., & Pleh, Cs. The develop-ment of sentence comprenension in Hungar-ian. Unpublished manuscript, Carnegie-Melon University, 1982.

Anderson, J. The architecture of cognition. Cam-bridge, Mass.: Harvard University Press,1983.

Bates, E. Language and context: Studies in theacquisition of pragmatics. New York:Academic Press, 1976.

Bates, E., & MacWhinney, B. Functionalist ap-proaches to grammar. In L. Gleitman & E.Wanner (Eds.), Language acquisition: Thestate of the art. New York: Cambridge Uni-versity Press, 1982.

Bates, E., McNew, S., MacWhinney, B., Deve-scovi, A., & Smith, S. Functional constraintson sentence processing: A cross-linguisticstudy. Cognition, 1982, 11, 245-299.

Bever, T. G. The cognitive basis for linguisticstructures. In J. R. Hayes (Ed.), Cognition andthe development of language. New York:Wiley, 1970.

Bloom, L. Talking, understanding, and thinking.In R. L. Schiefelbusch & L. L. Lloyd (Eds.),Language perspectives: Acquisition, re-tardation and intervention. Baltimore: Uni-versity Park Press, 1974.

Bridges, A. SVO comprehension strategies re-considered: The evidence of individual pat-terns of response. Journal of Child Language,1980, 7, 89-104.

Brown, R. A first language: The early stages.Cambridge, Mass.: Harvard University Press,1973.

Brunswik, E. Perception and the representativedesign of psychology experiments. Berkeley:University of California Press, 1956.

Chapman, R. S., & Kohn, L. L. Comprehensionstrategies in two- and three-year-olds: Ani-mate agents or probable events? Journal ofSpeech and Hearing Research, 1978, 21,746-761.

Clark, E. Non-linguistic strategies and the acqui-sition of word meanings. Cognition, 1973, 2,161-182.

Clark, E. V. Here's the top: Nonlinguistic strate-gies in the acquisition of orientational terms.Child Development, 1980, 51, 329-338.

Cromer, R. Developmental strategies for lan-guage. In V. Hamilton & M. Vernon (Eds.),The development of cognitive processes. NewYork: Academic Press, 1976.

Duranti, A., & Ochs, E. Left-dislocation in Italianconversation. In T. Givon (Ed.), Syntax andsemantics. New York: Academic Press, 1979.

Forster, K. I. Levels of processing and the struc-ture of the language processor. In W. T.Cooper & C. T. Walker (Eds.), Sentence pro-cessing: Psycholinguistic studies presented toMerrill Garrett. Hillsdale, N.J.: Erlbaum,1979.

Gleitman, L., & Gleitman, H. Phrase and para-phrase. New York: Norton, 1970.

Hakuta, K. Interaction between particles andword order in the comprehension of simplesentences in Japanese children. Develop-mental Psychology, 1982, 18, 62-76.

Karmiloff-Smith, A. Language as a formal problemspace. In W. Deutsch (Ed.), Child language:Beyond description. New York: Springer,1982.

Keeney, T., & Wolfe, J. The acquisition of agree-ment in English. Journal of Verbal Learningand Verbal Behavior, 1972, 11, 698-705.

MacWhinney, B. The acquisition of mor-phophonology. Monographs of the Society forResearch in Child Development, 1978,43(1-2, Serial No. 174).

MacWhinney, B. Basic syntactic processes. In S.Kuczaj (Ed.), Language acquisition: Syntaxand semantics. Hillsdale, N.J.: Erlbaum,1982.

MacWhinney, B., Bates, E., & Kliegl, R. The im-pact of cue validity on sentence interpretationin English, German, and Italian. Journal ofVerbal Learning and Verbal Behavior, inpress.

MacWhinney, B., & Price, D. The development ofthe comprehension of topic-comment mark-ing. In D. Ingram, C. C. Peng, & P. Dale(Eds.), Proceedings of the First InternationalCongress for the Study of Child Language.Lanham, Md.: University Press of America,1980.

354 Child DevelopmentMaratsos, M. P. Children who get worse