IQ, Learning, Development, and Conceptual ChangeT Л а Λ *1 1 Λ 4 (Λ -kA C f 1 Λ Λ -r“\ f ' ' ' Л 1 * Л Т Т С l í Х Т л т * T V V / w i - /Л уАLila G le itm an , S u sa n C arey , E lissa N e w p o rt , a n dE lizab e th S p e lk e , ed ito rs

Names for Things: A Study in Human Learning, John M acnam ara, 1982

Conceptual Change in Childhood, Susan Carey, 1985

"Gavagai!" or the Future History of the Anim al Language Controversy, David Prem ack, 1986

Systems That Learn: A n Introduction to Learning Theory for Cognitive and Computer Scientists, Daniel N. O sherson, 1986

From Simple Input to Complex Grammar, James h . M organ, 1986

Concepts, Kinds, and Cognitive Development, Frank C. Keil, 1989

Learnability and Cognition: The Acquisition o f Argum ent Structure, Steven Pinker, 1989

M ind Bugs: The Origins of Procedural Misconception, K urt VanLehn, 1990

Categorization and N am ing in Children: Problems o f Induction, Ellen M. M arkm an, 1990

The Child's Theory of M ind , H enry M. W ellman, 1990

The Organization of Learning, Charles R. Gallistel, 1990

Understanding the Representational M ind, Josef Pem er, 1991

A n Odyssey in Learning and Perception, Eleanor J. Gibson, 1991

Beyond Modularity: A Developmental Perspective on Cognitive Science, A nnette Karmiloff-Smith, 1992

Beyond Modularity

A Developmental Perspective on Cognitive Science

Annette Karmiloff-Smith

A Bradford Book The MIT Press Cam bridge, M assachusetts London, England

Chapter 1

Taking Development Seriously

Nature has contrived to have it both ways, to get the best out of fast dumb systems and slow contemplative ones, by simply refusing to choose between them. (Fodor 1985, p. 4)

Have you noticed how quite a large num ber of developm ental psy­chologists are loath to attribute any innate predispositions to the h u ­m an infant? Yet they w ould no t hesitate to do so w ith respect to the ant, the spider, the bee, or the chim panzee. W hy w ould N ature have endow ed every species except the hum an w ith som e domain-specific predispositions? Yet, if it tu rns ou t that all species have such p redis­positions, that m ost can m aintain a goal in the face of changing envi­ronm ental conditions, and that m ost have the capacity for learning on the basis of interaction w ith conspecifics and the physical environ­m ent, w hat is special about h u m an cognition? Is it sim ply that the content of know ledge differs betw een species? Is it language that makes hum ans special? O r are there qualitatively different processes at w ork in the hum an m ind? Does h u m an cognitive change affect all dom ains of know ledge sim ultaneously, or does developm ent occur in a dom ain- specific fashion? Are cross-species differences relevant only to adult cognition, or do hum ans differ from other species from birth?

This book sets out to address such questions and to dem onstrate that one can attribute various innate predispositions to the hum an neonate w ithout negating the roles of the physical and sociocultural environm ents and w ithou t jeopardizing the deep-seated conviction that we are special— creative, cognitively flexible, and capable of con­scious reflection, novel invention, and occasional inordinate stupidity.

Is the Initial Architecture of the Infant M ind Modular?

Fodor's 1983 book The M odularity o f M ind (which I later criticize) m ade a significant im pact on developm ental theorizing by suggesting how

2 Chapter 1

the nativist thesis and the domain-specificity of cognition are relevant to constraints on the architecture of the hum an mind. For Fodor, the notion of "architecture" refers to the organization of relatively fixed and highly constrained innate specifications: the invariant features of the hum an inform ation-processing system . Unlike B runer (1974-75) and Piaget (1952b), w ho argue for dom ain-general developm ent, Fo­dor holds that the m ind is m ade up of genetically specified, indepen­dently functioning, special-purpose "m odules" or inpu t system s.1 Like Fodor, I shall use the terms "m odule" and "in p u t system " as synonym s. Each functionally distinct m odule has its ow n dedicated processes and proprietary inputs.

According to Fodor, inform ation from the external environm ent passes first th rough a system of sensory transducers, w hich transform the data into form ats that each special-purpose inpu t system can process. Each in p u t system , in turn , ou tputs data in a com m on format suitable for central, dom ain-general processing. The m odules are deem ed to be hard-w ired (not assem bled from m ore primitive proc­esses), of fixed neural architecture, dom ain specific, fast, autonom ous, m andatory, autom atic, stim ulus driven, giving rise to shallow outputs, and insensitive to central cognitive goals.

A further characteristic of m odules is that they are informationally encapsulated (or, as Pylyshyn [1980] p u tit, "cognitively impenetrable"). O ther parts of the m ind can neither influence nor have access to the internal w orkings of a m odule, only to its outputs. M odules have access only to inform ation from stages of processing at low er levels, no t to inform ation from top-dow n processes. In o ther w ords, w hat the m ind know s and believes cannot affect the w orkings of a m odule.

For Fodor, the essential fact about m odules is their informational encapsulation. H e is neu tra l about w hether they are resource encap­sulated (i.e., w h e th er different m odules share, say, inference algorithm s2). In defense of inform ational encapsulation, Fodor cites the example of perceptual illusions such as the Muller-Lyer illusion (figure 1.1). In that illusion, even w hen subjects have m easured the

< >

> <

Figure 1.1The M uller-Lyer illusion.

Taking Development Seriously 3

two lines and th u s have explicit know ledge of their equal length, they cannot preven t them selves from seeing one of the lines as longer than the other, depend ing on the direction of the arrow heads at their extremities. Tihe subject's explicit know ledge about equal line length, available in w ha t Fodor calls the "central system ," is no t available to the perceptual system 's com putation of relative lengths. In other w ords, the m odule for perceptual processing is self-contained and has no access to the inform ation elsew here in the m ind. Gallistel (1990) gives a similar definition w hen discussing the cognitive architecture of o ther species. For instance, a lthough the rat can represent nongeo­metric data (such as color, smell, and texture) and can use them for various purposes, the ra t's system for determ ining position and head ­ing in space can m ake use of geom etric data only. It is im penetrable to inform ation from nongeom etric sources, even w hen such data are highly relevant to the ra t's curren t goal.

For Fodor, it is the co-occurrence of all the properties discussed above that defines a m odule or an in p u t system . Alone, particular properties do n o t necessarily entail m odularity. For instance, rapid autom atic processing can also take place outside in p u t system s. A nderson (1980) provides examples of this from skill learning.3 He show s that, w hen learning a new skill, subjects initially focus consciously on com ponent parts, bu t that once skill learning is com plete the parts becom e com ­piled into a procedure w hich is executed rapidly, automatically, and unconsciously. Such task-specific expertise should n o t be confounded w ith the Fodorian concept of a m odule, w hich includes h ard wiring, fixed neural architecture, m andatory stim ulus-driven processing, in ­formational encapsulation, and insensitivity to central cognitive goals.

Each m odule is like a special-purpose com puter w ith a proprietary database. By "proprietary" Fodor m eans that a m odule can process only certain types of data and that it autom atically ignores other, potentially com peting input. A m odule com putes in a bottom -up fash­ion a constrained class of specific inputs; that is, it focuses on entities that are relevant to its particular processing capacities only. A nd it does so w henever relevant data p resen t them selves—that is, an inpu t system cannot refrain from processing. This enhances autom aticity and speed of com putation by ensuring that the organism is insensitive to m any potential classes of inform ation from other in p u t system s and to top-dow n expectations from central processing.

Inpu t system s, then , are the parts of the hum an m ind that are inflexible and unintelligent. They are the stupidity in the m achine— bu t they are just w hat a young organism m ight need to get initial cognition off the g round speedily and efficiently.

4 Chapter 1

I argue that developm ent involves a process of going beyond m od­ularity. For Fodor, how ever, developm ent do esn 't really exist.4 Rather, Fodor posits a built-in dichotom y betw een w hat is com puted blindly by the in p u t system s an d w hat the organism "believes. " It is in "central processing" that the h u m an belief system is built up , by deriving top- dow n hypo theses abou t w hat the world is like from the interface betw een the o u tpu ts of in p u t system s and w hat is already stored in long-term m em ory. In contrast w ith in p u t system s, Fodor considers central processing to be influenced by w hat the system already knows, and therefore to be relatively unencapsulated , slow, nonm andatory, controlled, often conscious, and influenced by global cognitive goals. Central processing receives ou tpu ts from each inpu t system in a com­m on representational form at, a language of though t (Fodor 1976). Central processing, then , is general-purpose. It is devoted to the fixation of belief, the building up of encyclopedic know ledge, and the planning of intelligent action, in contrast to the special-purpose, do­main-specific com putations of modules.

While endorsing the im portance of several aspects of Fodor's thesis for understand ing the architecture of the hum an m ind, I shall provide a view that differs from the notion that m odules are prespecified in detail, and shall question the strictness of the dichotom y that Fodor draw s betw een m odules and central processing.51 shall also challenge Fodor's contention that the outputs of in p u t system s are automatically encoded into a single com m on language of thought.

Prespecified Modules versus a Process of Modularization

Fodor's detailed account of the encapsulation of m odules focuses predom inantly on their role in on-line processing. There is little dis­cussion of ontogenetic change, except to allow for the creation of new m odules (such as a reading m odule). Fodor takes it as dem onstrated that m odules for spoken language and visual perception are innately specified. By contrast, I w ish to draw a distinction betw een the notion of prespecified m odules and that of a process of modularization (which,I speculate, occurs repeatedly as the product of developm ent). Here I differ from Fodor's strict nativist conception. I hypothesize that if the hum an m ind ends up w ith any m odular structure, then , even in the case of language, the m ind becom es m odularized as development pro­ceeds. M y position takes account of the plasticity of early brain devel­opm ent (Neville 1991; Johnson, in press). It is plausible that a fairly limited am ount of innately specified, domain-specific predispositions (which are no t strictly m odular) w ould be sufficient to constrain the classes of inpu ts that the infant m ind com putes. It can thus be hy ­

Taking Development Seriously 5

pothesized that, w ith time, bra in circuits are progressively selected for different dom ain-specific com putations; in certain cases, relatively encapsulated m odules w ould be formed. Thus, w hen I use the term "innately specified" in this book, I do not m ean to im ply anything like a genetic b lueprin t for prespecified m odules, present a t b irth .6 Rather, as will be clear, I argue for innately specified predispositions that are m ore epigenetic than Fodor's nativism . The view that I adop t th rough­out the book is that N ature specifies initial biases or predispositions that channel attention to relevant environm ental inputs, w hich in turn affect subsequent brain developm ent.7

The thesis that developm ent involves a process of gradual m odu­larization ra ther th an prespecified m odules rem ains speculation a t this po int in time. It will not, therefore, be developed fu rther in the book. However, it does m erit m ention in this introductory chapter to delin­eate the extent to w hich I find Fodor's views useful for thinking about the hum an m ind an d the extent to w hich I call for certain modifica­tions. Together w ith quite a num ber of cognitive developm entalists, I think Fodor's thesis has poin ted to w here a dom ain-general view of developm ent such as Piaget's is likely to be w rong. H owever, I shall argue for a m ore dynam ic view of developm ent than Fodor's m odu­larity of mind.

The choice betw een prespecified m odules and m odularization is an empirical one. O nly fu ture research using on-line brain-activation studies w ith neonates and young infants can distinguish betw een the two hypotheses. If Fodor's thesis of prespecified m odularity is correct, such studies should show that, from the very outset, specific brain circuits are activated in response to domain-specific inputs. By con­trast, if the m odularization thesis is correct, activation levels should initially be relatively d istributed across the brain, and only w ith time (and this could be a short or relatively long time during infancy) w ould specific circuits always be activated in response to domain-specific in p u ts .8 The m odularization thesis allows us to speculate that, al­though there are m aturationally constrained atten tion biases and domain-specific predispositions that channel the in fan t's early devel­opm ent, this endow m ent interacts richly w ith, and is in re tu rn affected by, the environm ental input.

W hatever its shortcom ings, Fodor's m odularity thesis has offered cognitive science m uch food for thought. N onetheless, I aim to chal­lenge Fodor's dism issal of the relevance of a developm ental perspec­tive on cognitive science. Developm ent, in m y view, is the key to understand ing the adu lt m ind. M oreover, I question Fodor's oft-cited claim that "the limits of m odularity are also likely to be the limits of w ha t w e are going to be able to understand about the m ind" (1983,

6 Chapter 1

p. 126). I shall argue th a t cognitive scientists can go beyond m odularity to study the m ore creative aspects of hum an cognition. But m y con­tention is that such an endeavor will be greatly enhanced by a devel­opm ental perspective on the problem .

'What Constitutes a Domain?

Irrespective of w he ther they agree w ith Fodor's strict m odularity the­sis, m any psychologists now consider developm ent to be "dom ain specific." M uch depends, of course, on w hat one understands by "dom ain," and it is im portan t no t to confuse "dom ain" w ith "m odule." From the po in t of view of the child 's m ind, a dom ain is the set of representations sustain ing a specific area of knowledge: language, num ber, physics, an d so forth. A m odule is an inform ation-processing unit that encapsulates that know ledge and the com putations on it. Thus, considering developm ent dom ain specific does not necessarily im ply m odularity. In o ther w ords, the storing and processing of in­form ation m ay be dom ain specific w ithou t being encapsulated, hard ­wired, or m andatory.

Fodor's discussion of m odularity is defined over very broad do­mains, such as language. He talks, for instance, of the "language m odule" and the "perceptual m odule." O thers tend to draw finer distinctions w ith in a dom ain—e.g., the syntactic m odule, the sem antic m odule, an d the phonological m odule. Still o thers (Marslen-W ilson and Tyler 1987) reject the notion of on-line m odularity of processing altogether. T hroughout the book, I shall argue for dom ain specificity of developm ent ra ther th an m odularity in the strict Fodorian sense. I shall retain the term "dom ain" to cover language, physics, m athe­matics, and so forth. I will also distinguish "m icrodom ains" such as gravity w ith in the dom ain of physics and p ro n o u n acquisition w ithin the dom ain of language. These m icrodom ains can be though t of as subsets w ithin particular dom ains.

The need for this finer distinction of w hat constitutes a dom ain stem s from the fact that I will p u t forw ard a phase m odel of develop­m ent, ra ther than a stage model. In a stage m odel, such as Piaget's, overarching changes occur m ore or less sim ultaneously across differ­ent dom ains. O ne alternative view is that broad changes occur w ithin a dom ain—for example, that a particular type of change occurs first w ith respect to language an d later w ith respect to physics. The m odel discussed in this book differs from both of these conceptions. It in ­vokes recurrent phase changes at different times across different micro- dom ains a n d repeatedly w ith in each dom ain. Take the case of the dom ain of language as an example. In the m icrodom ain of pronoun

Taking Development Seriously 7

acquisition, a sequence of changes X-Y-Z (e. g ., from im plicit to explicit to verbal justification) m ight be com plete in a child by age 7, w hereas in the m icrodom ain of understand ing w hat a w ord is the sam e se­quence m ight already be com plete by age 5. I shall thus distinguish the broad dom ains (language, m athem atics, and so forth) from the m icrodom ains (e.g. p ronouns and counting) that they subsum e. W henever I refer to dom ain-general or domain-specific theories, these are situated at the level of b road dom ains.

Development from a Domain-General Perspective

Fodor's nativist thesis is in sharp contrast w ith dom ain-general theo­ries of learning, such as Piaget's constructivist epistem ology, w hich w ere once popular in the developm ent literature.9 Piagetian theory argues th a t neither processing n o r storage is dom ain specific. Of course, implicitly at least, Piagetians m ust acknowledge that there are different sensory transducers for vision, audition, touch, and so forth. They do no t accept, how ever, that the transducers transform data into innately specified, dom ain-specific form ats for m odular processing. For Piagetians, developm ent involves the construction of dom ain- general changes in representational structures operating over all as­pects of the cognitive system in a similar way.

A t this juncture I shall risk outraging som e of m y form er colleagues at G eneva University by suggesting that Piaget and behaviorism have m uch in com m on. W hat, link Piaget and Skinner? An aberration, to be sure! Yet I arrive at this liaison dangereuse betw een such unlikely bedfellows by opposing the dom ain-general view w ith the dom ain- specific explanation of developm ent.

N either the Piagetian no r the behaviorist theory grants the infant any innate structures or dom ain-specific know ledge. Each grants only som e dom ain-general, biologically specified processes: for the Piaget­ians, a set of sensory reflexes and three functional processes (assimi­lation, accom m odation, and equilibration); for the behaviorists, inherited physiological sensory system s and a complex set of laws of association. These dom ain-general learning processes are held to ap ­ply across all areas of linguistic and nonlinguistic cognition. Piaget and the behaviorists th u s concur on a num ber of conceptions about the initial state of the in fan t m ind. The behaviorists saw the infant as a tabula rasa w ith no built-in know ledge (Skinner 1953); P iaget's view of the young infant as assailed by "undifferentiated and chaotic" inputs (Piaget 1955a) is substantially the same.

N eedless to say, there are fundam ental differences betw een these two schools. P iagetians view the child as an active inform ation con­

8 Chapter '1

structor, behaviorists as a passive inform ation storer. Piagetians con­ceive of developm ent as involving fundam ental stage-like changes in logical structure, w hereas behaviorists invoke a progressive accum u­lation of know ledge. H owever, in the light of the presen t state of the art in developm ental theorizing, Piagetians and behaviorists have m uch in com m on in their view of the neonate 's "know ledge-em pty" m ind and their claims that dom ain-general learning explains subse­quen t developm ent across all aspects of language and cognition.

Development from a Domain-Specißc Perspective

The nativist/m odularity thesis projects a very different picture of the young infant. R ather than being assailed by incom prehensible, chaotic data from m any com peting sources, the neonate is seen as p repro­gram m ed to m ake sense of specific inform ation sources. C ontrary to the Piagetian or the behaviorist infant, the nativist infant is off to a very good start. This doesn 't, of course, m ean that no thing changes during infancy an d beyond; the infant has m uch to learn. But the nativist/m odularity stance posits that subsequent learning is guided by innately specified, dom ain-specific principles, and that these prin­ciples determ ine the entities on w hich subsequent learning takes place (Gelman 1990b; Spelke 1991).

The dom ain specificity of cognitive system s is also suggested by developm ental neuropsychology and by the existence of children in w hom one or m ore dom ains are spared or im paired. For example, autism m ay involve a single deficit in reasoning about m ental states (theory of m ind), w ith the rest of cognition relatively unim paired. Williams Syndrom e, by contrast, p resen ts a very uneven cognitive profile in w hich language, face recognition, and theory of m ind seem relatively spared, w hereas num ber and spatial cognition are severely retarded. A nd there are num erous cases of idiots-savants in w hom only one dom ain (such as draw ing or calendrical calculation) functions at a h igh level, w hile capacities are very low over the rest of the cognitive system . By contrast, D ow n Syndrom e is suggestive of a m ore across-the-board, dom ain-general deficit in cognitive processing.

A dult bra in dam age points to dom ain specificity, also. It is rem ark­ably difficult to find convincing examples in the neuropsychological literature of an across-the-board, dom ain-general d isorder (Marshall1984), a lthough a case m ight be m ade for an overall deficit in planning in patients w ith prefrontal dam age (Shallice 1988). But in m any in­stances, disorders of h igher cognitive functions consequent upon brain dam age are typically dom ain specific— that is, they affect only

Taking Development Seriously 9

face recognition, num ber, language, or som e other facility, leaving the other system s relatively intact.

So if adults m anifest dom ain-specific dam age, and if it can be show n that infants come into the w orld w ith som e domain-specific pred is­positions, d oesn 't th a t m ean that the nativists have w on the debate over the developm entalists still ensconced on the theoretical shores of Lake Geneva (Piaget's form er bastion of anti-nativism and anti­m odularity)? N ot necessarily, because it is im portan t to bear in m ind that the greater the am oun t of domain-specific properties of the infant m ind, the less creative an d flexible the subsequent system will be (Chomsky 1988). W hereas the fixed constraints provide an initial adap ­tive advantage, there is a tradeoff betw een the efficiency and auto- maticity of the in fan t's in p u t system s, on the one hand , and their relative inflexibility, on the other. This leads me to a crucial point: The more complex the picture we ultimately build of the innate capacities of the infant mind, the more important it becomes for us to explain the flexibility of subsequent cognitive development. It is tow ard such an end—exploring the flexibility and creativity of the hum an m ind beyond the initial state— that my w ork in language acquisition and cognitive develop­m ent has been concentrated, in an attem pt to determ ine both the domain-specific and the dom ain-general contributions to develop­m ent. It is im plausible that developm ent will tu rn ou t to be entirely dom ain specific or dom ain general. A nd although I will need to invoke som e built-in constraints, developm ent clearly involves a m ore dy­nam ic process of interaction betw een m ind and environm ent than the strict nativist stance presupposes.

Reconciling Nativism and Piaget's Constructivism

W hat theory of developm ent could encom pass the dynam ics of a rich process of interaction betw een m ind and environm ent? A t first blush, a theory w ith a central focus on epigenesis and constructivism , like Piaget's, w ould seem the m ost appropriate. The notion of constructiv­ism in Piaget's theory10 is the equivalent at the cognitive level of the notion of epigenesis at the level of gene expression. For Piaget bo th gene expression an d cognitive developm ent are em ergent products of a self-organizing system th a t is directly affected by its interaction w ith the environm ent. This general aspect of Piaget's theory, if m ore for­m alized, m ay well tu rn ou t to be appropriate for fu ture explorations of the notion of progressive m odularization discussed above. H ow ­ever, m uch of the rest of P iaget's theory has come un d er a great deal of criticism. A grow ing num ber of cognitive developm entalists11 have become disenchanted w ith Piaget's account of the infant as a purely

10 Chapter 1

sensorim otor organism . For Piaget the new born has no dom ain-spe­cific know ledge, m erely sensory reflexes and the three dom ain-general processes of assim ilation, accom m odation, an d equilibration. By con­trast, the infancy research that I shall discuss in the following chapters suggests that there is considerably m ore to the initial functional ar­chitecture of the brain than Piaget's theory posits. Yet the exclusive focus of nativists like Fodor and Chom sky on biologically specified m odules leaves little room for rich epigenetic-constructivist processes. Moreover, Fodor's concentration on inpu t system s— he has far less to say about either o u tp u t system s or central processing—do esn 't help us to understand the w ay in w hich children tu rn ou t to be active participants in the construction of their ow n knowledge.

A lthough for C hom sky (1988) and Spelke (1991) a nativist stance precludes constructivism , I argue that nativism and Piaget's epigenetic constructivism are no t necessarily incompatible—w ith certain provi­sos. First, to Piaget's view one m ust add som e innate, knowledge- im pregnated pred ispositions12 that w ould give the epigenetic process a head start in each dom ain. This does not im ply m erely adding a little m ore dom ain-general structure than Piaget supposed . Rather, it m eans add ing dom ain-specific biases to the initial endow m ent. But the second proviso for the m arriage of constructivism and nativism is that the initial base involve less detailed specifications than some nativists p resuppose and a m ore progressive process of modularization (as opposed to prespecified m odules). Fodor does not, for instance, discuss the cases in w hich one of his prespecified m odules cannot receive its proprie tary in p u t (e.g., auditory in p u t to a language m odule in the case of the congenitally deaf). We know that in such cases the brain selectively adap ts to receive other (e.g., visuom anual) nonam ditory inputs, w hich it processes linguistically (Changeux 1985; Neville 1991; Poizner et al. 1987). M any cases of early brain dam age indicate that there is far m ore plasticity in the brain than Fodor's strict m odu­larity view w ould im ply. The brain is not p restructu red w ith ready­m ade representations; it is channeled to progressively develop rep re­sentations via interaction w ith bo th the external environm ent and its ow n internal environm ent. A nd, as I stressed above, it is im portant not to equate innateness w ith presence at birth or w ith the notion of a static genetic b lueprin t for m aturation. W hatever innate com ponent w e invoke, it becom es p art of our biological potential only through interaction w ith the environm ent; it is latent until it receives input (Johnson 1988; Johnson, in press; M arler 1991; O yam a 1985; Thelen 1989). A nd that in p u t affects developm ent in return.

The proposed reconciliation of nativism and constructivism will allow us to adhere to Piaget's epigenetic-constructivist view of the

Taking Development Seriously 11

developm ental process, bu t to drop his insistence on dom ain gener­ality in favor of a m ore dom ain-specific approach. Furtherm ore, the Piagetian focus on o u tp u t system s (i.e., on the in fan t's and the child 's action on the environm ent) is an im portan t addition to the nativ ist's accent on in p u t system s. But Piaget's strong anti-nativism an d his argum ents for across-the-board stages no longer constitute a viable developm ental fram ew ork.13

The need to invoke dom ain specificity will be apparen t th roughout the book. For example, it will becom e clear in chapter 2 that dom ain- general sensorim otor developm ent alone cannot explain the acquisi­tion of language. Syntax does no t sim ply derive from exploratory problem solving w ith toys, as som e Piagetians claim. Lining u p objects does not form the basis for w ord order. Trying to fit one toy inside ano ther has noth ing to do w ith em bedded clauses. General sensori­m otor activity alone cannot account for specifically linguistic con­straints; if it could, th en it w ould be difficult to see w hy chim panzees, w hich m anifest rich sensorim otor and representational abilities, do no t acquire anyth ing rem otely resem bling hum an language despite extensive training (Premack 1986).

Despite these criticisms of P iaget's view of early infancy and my rejection of his stage view of developm ent, I hope by the end of the book to have persuaded you that im portant aspects of Piaget's epis- tem ology should be salvaged and that there is far m ore to cognitive developm ent than the unfolding of a genetically specified program . If we are to understand the h u m an m ind, our focus m ust stretch well beyond the innate specifications. Infants and young children are active constructors of their ow n cognition. This involves both dom ain- specific constraints and dom ain-general processes.

In sum , there seem s to be som ething right about both Fodor's and Piaget's approaches to h u m an cognition. My ow n solution to this potential dilem m a has been to take an epistemological stance that encom passes aspects of bo th nativism and constructivism .

The Notion of Constraints on Development

N ow adays, m any discussions in developm ental psychology concern constraints on developm en t.14 But dom ain-general and dom ain-spe­cific theories treat th e notion of constraints differently. For the dom ain- general theorist, the w ord "constraints" carries a negative connotation; it is taken as referring to factors w hich curtail a child 's com petence. By contrast, for the dom ain-specific theorist "constraints" takes on a positive connotation: Domain-specific constraints potentiate learning by limiting the hypothesis space entertained. They enable the infant

12 Chapter 1

to accept as in p u t only those data w hich it is initially able to com pute in specific w ays. The dom ain specificity of processing provides the infant w ith a lim ited yet organized (nonchaotic) system from the out­set, and not solely at the tail end of the Piagetian sensorim otor period .15

New Paradigms for Studying Young Infants

Piaget's p ioneering experim ental w ork on developm ent was focused on older children. For his exploration of infancy, Piaget had to rely solely on observation of his ow n three children. There w ere no para­digm s available th en for the experim ental study of early infancy. Since the m id 1960s, how ever, m ethodological innovations have opened up exciting new experim ental possibilities. Experim ents now focus on the different in p u t system s th rough w hich new borns and young infants com pute data relevant to a variety of cognitive dom ains. A nd, al­though I do n o t share Fodor's pessim ism that w e shall never under­stand central sy stem s,16 he is right that in p u t system s are m uch m ore am enable to strict experim ental research, particularly in infancy.

Let me digress for a m om ent to look briefly at the new paradigm s for infancy research, since they will crop u p th roughout the book. These paradigm s have been used by researchers in terested in the infan t's sensitivity to data relevant to language, physics, num ber, hum an in tention , tw o-dim ensional notation, and so forth. They are thus im portan t for all the chapters in this book.

The new experim ental approaches were devised to su rm ount prob­lems arising from Piagetian-inspired research w hich required infants to dem onstrate their abilities by m anual search. N eonates and young infants cannot engage in m anual search. W hat they do well is suck and look (and, alas for parents, cry). These capacities form the basis of the new m ethodologies. There are three m ain infancy techniques; two fall un d er the habituation/dishabituation paradigm , and the third uses preferential looking or listening.

In the habituation/dishabituation paradigm , the infan t is presented repeatedly w ith the sam e stim ulus set until it show s lack of interest by starting to a ttend for shorter times. Then a different stim ulus set is presented. If the in fan t show s renew ed in terest by attending for a longer time, it can be concluded that the new stim ulus is apprehended (perceived, understood) by the infant as different from the earlier one. The stim ulus set can be visual, auditory, or tactile, depend ing on the experim ent. A n in fan t's in terest in an event (e.g., seeing a circle after a series of squares of different sizes and colors) typically m anifests itself as prolonged attention. By clever m anipulation of variables of

Taking Development Seriously 13

shape, color, size, an d so forth, the researcher can hom e in on the nature of the difference to w hich the infant is sensitive. Say the new ­born show s decreasing in terest in squares despite constant variations in size and color, bu t sudden ly show s renew ed in terest on the first p resentation of a circle; th en one can conclude that shape discrimi­nation is p resen t at b irth and does no t have to be learned. By contrast, if the new born continues to show lack of in terest on presen tation of the circle, one can conclude that the circle is app rehended as being equivalent to the set of squares—i.e., that shape discrim ination is a later achievem ent (although in fact, as Slater [1990] has show n, it is p resen t at birth). The sam e logic is u sed to test discrim inations of other types of stimuli.

"Interest" is m easured either by greater am plitude of sucking or by longer length of looking. In the form er case, the infant is given a non­nutritive pacifier w hich is attached to an apparatus that m easures sucking am plitude. As the infan t habituates to the original stim ulus, its sucking am plitude decreases. If the new stim ulus is apprehended as different, the in fan t's sucking am plitude increases; if not, it p lateaus or decreases further. As will be discussed in chapter 2, such a tech­nique has been used to explore the infant's preference for listening to its m other tongue over o ther linguistic input, as well as its capacity for categorical perception of various speech sounds. Thus, if the infant is p resen ted w ith a set of "va" sounds, and then after habituation w ith a set of "ba" sounds, increased sucking am plitude dem onstrates the infan t's sensitivity to the difference betw een the sounds (i.e., to voice- onset time). Such techniques help us to explore the effects of environ­m ental in p u t on innate predispositions. For a child in a Spanish­speaking environm ent, for instance, sensitivity to the distinction betw een "va" and "ba" m ay be p resen t early in infancy but d isappear once the patterns of the in p u t language have been learned, because spoken Spanish does no t differentiate betw een "va" and "ba".

The technique for m easuring looking time is based on the sam e principle as the one m easuring sucking am plitude. The infan t is re­peatedly exposed to a visual stim ulus. Each time the stim ulus is p re­sented, the infant will look a t it for a shorter length of tim e, until it habituates. After hab ituation to a given stim ulus set, the infant's length of looking at a new stim ulus is recorded as a m easure of its renew ed in terest or its boredom . Again, subtle m anipulation of vari­ables can determ ine the features to w hich the in fan t is particularly sensitive. The use of this technique will be discussed in chapter 3. For example, infants show surprise (look longer) at a display of a ball that seem s to stop in m id-air w ithou t support, o r at a display of an object

14 Chapteţ· 1

that appears to have passed th rough a solid surface— that is, they are sensitive to violations of certain law s of physics.

M easuring looking time is som ew hat m ore subjective than m easur­ing sucking am plitude. Thus, looking time m ust be recorded by ob­servers unaw are of the particular display being view ed by the infant on any trial. But, as Spelke (1985) has po in ted o u t /7 the interpretation of test-trial looking an d sucking patterns in experim ents of this kind depends on the finding, now obtained in hu n d red s of laboratories th roughout the w orld, that habituation to one stim ulus set is followed by longer looking (or longer sucking) for the test display. In other w ords, the in terpretation rests on the fact th a t infants extract a com­m on feature across the set of stimuli in the habituation display, and differentiate that from a specific feature of the test display.

A third infancy paradigm involves preferential looking or listening. Here habituation an d dishabituation are no t m easured; rather, the infant is p resen ted w ith two stim ulus displays sim ultaneously and m easurem ent is taken of w hich display the infant prefers to look at. Again, m easurem ents are determ ined by observers w ho cannot see the displays visible to the infant. C hapter 4 illustrates uses of this technique to m easure infants ' capacity to m atch the num ber of audi­tory stim uli (e.g., th ree drum beats) to the num ber of objects in either of two visual displays, one containing two objects an d the o ther con­taining three.

A lthough the infancy data discussed throughout the book are truly im pressive, certain questions about the habituation and preferential techniques rem ain open. Does the violation of a physical principle have to be extrem e, or are infants just as sensitive to subtler violations? W hat conclusions can legitim ately be draw n from the dem onstration that the infan t is sensitive to a novel stim ulus: that domain-specific attention biases and principles are built into the infant m ind, or m erely that we have trained infants to discrim inate in the course of the actual experim ent? Any particular experim ent w ould rem ain inconclusive on this issue. H ow ever, if results from different experim ents dem onstrate that new borns or 4-m onth-olds can m ake discrim inations for one set of stimuli b u t cannot do so for another, then it cannot be claimed that discrim ination is solely the result of task-specific learning. Rather, discrim ination is constrained by w hether or not the infant can already show sensitivity to the particular characteristics of the stimuli. This allows tentative conclusions regarding innate specifications and those involved in subsequen t learning—tentative since m any other inter­pretations are possible.

I discuss the infancy research in som e detail in the first part of each of chapters 2 th rough 6. But every time, I go on to show that devel­

Taking Development Seriously 15

opm ent com prises m uch m ore than the domain-specific constraints. In particular, it involves "representational redescription," a process that increases the flexibility of the know ledge stored in the mind.

Beyond Domain-Specific Constraints: The Process of Representational Redescription

How does inform ation get stored in the child's m ind? I argue that there are several different ways. O ne is via innate specification as the result of evolutionary processes. Innately specified predispositions can either be specific or nonspecific (Johnson arid Bolhuis 1991). In both cases, environm ental in p u t is necessary. W hen the innate com­ponent is specified in detail, it is likely that the environm ent acts simply as a trigger for the organism to select one param eter or circuit over others (Changeux 1985; Chom sky 1981; Piatelli-Palmarini 1989).18 By contrast, w hen the innate predisposition is specified m erely as a bias or a skeletal outline, th en it is likely th a t the environm ent acts as m uch m ore than a trigger—that it actually influences the subsequent structure of the brain via a rich epigenetic interaction betw een the m ind and the physical/sociocultural environm ent. The skeletal outline involves attention biases tow ard particular inputs and a certain num ­ber of principled predispositions constraining the com putation of those inputs. Note that I am hypothesizing that the hum an m ind has both a certain am ount of detailed specification and som e very skeletal domain-specific predispositions, depend ing on the domain.

There are several o ther w ays in which new inform ation gets stored in the child 's mind. O ne is w hen the child fails to reach a goal and has to take into account inform ation from the physical environm ent. A nother is generated by the child 's having to represent inform ation provided directly by a linguistic statem ent from, say, an adult. These are both external sources of c h a n g e /9 An internal source of change is illustrated by the above-m entioned process of m odularization in such a way that inpu t and o u tp u t processing becomes less influenced by other processes in the brain. This causes know ledge to become more encapsulated and less accessible to other system s. But ano ther essen­tial facet of cognitive change goes in the opposite direction, w ith knowledge becoming progressively m ore accessible.

My claim is that a specifically hum an w ay to gain know ledge is for the m ind to exploit internally the inform ation that it has already stored (both innate and acquired), by redescribing its representations or, m ore precisely, by iteratively re-representing in different represen ta­tional form ats w hat its internal representations represent. I will deal w ith this in detail in a m om ent.

16 Chapter '1

Finally, there is a form of know ledge change that is m ore obviously restricted to the h u m an species: explicit theory change, w hich involves conscious construction and exploration of analogies, though t experi­m ents and real experim ents, typical of older children and adults (Carey 1985; Klahr 1992; K uhn et al. 1988). But I will argue that this more obvious characteristic of hum an cognition is possible only on the basis of prior represen tational redescription, w hich tu rn s implicit inform a­tion into explicit know ledge.

To convey a m ore tangible feel for the theoretical discussion on w hich I am about to em bark, let me start w ith a couple of examples— one having to do w ith learning to play the piano an d one having to do w ith learning to solve Rubik's Cube.20

W hen one is learning to play the piano, initially there is a period during w hich a sequence of separate notes is laboriously practiced. This is followed by a period during w hich chunks of several notes are played together as blocks, until finally the w hole piece can be played m ore or less autom atically.21 It is som ething like this that I shall sub­sequently call "reaching behavioral m astery." But the autom aticity is constrained by the fact that the learner can neither start in the m iddle of the piece nor play variations on a them e (Herm elin and O 'C onnor 1989). The perform ance is generated by procedural representations w hich are sim ply ru n off in their entirety. There is little flexibility. At best the learner starts to be able to play the whole piece softer, louder, slower, or faster. It is only later that one can in terrup t the piece and start at, say, the th ird bar w ithou t having to go back to the beginning and repeat the en tire procedure from the outset. I hypothesize that this cannot be done on the basis of the autom atized procedural rep ­resentations. Rather, I posit, it involves a process of representational redescription such that the know ledge of the different notes and chords (rather than sim ply their run-off sequence) becom es available as m anipulable data. It is only after a period of behavioral m astery that the p ianist can generate variations on a them e, change sequential order of bars, in troduce insertions from other pieces, and so forth. This differentiates, for instance, jazz im provisation from strict adher­ence to sheet m usic. The end result is representational flexibility and control, w hich allows for creativity. Also im portan t is the fact that the earlier proceduralized capacity is not lost: for certain goals, the pianist can call on the autom atic skill; for o thers, he or she calls on the m ore explicit represen ta tions that allow for flexibility and creativity. (Of course, the playing of som e pianists rem ains at the procedural level.)

In contrast w ith the beginning pian ist's initial conscious attention to particular notes, w hich gradually becom es proceduralized, I found that I had to "sw itch off" m y consciousness to solve R ubik's Cube. In

Taking Development Seriously 17

other w ords, I had to stop trying to analyze w hat I w as doing until I could actually do it! In the early course of learning to solve the problem , I developed a sort of proprioceptive solution w hich I could perform very rapidly b u t w hich I had m uch m ore difficulty repeating at a slow er pace. M y "know ledge" at that stage was em bedded in the procedural representations sustain ing the rapid execution. But I d id no t stop there. After reiterating a solution m any times, I found that I started to recognize certain states of the cube and then knew w hether or not I was on the path to m y solution. But I still could not in te rrup t my solution and proceed from just any starting state. W ith m ore time still, I found that I could predict w ha t the next few m oves w ould be before actually executing them . Finally I came to a point w here I could explain the solution to m y daughter. She, however, did not use m y explicit instructions b u t w en t th rough the sam e progression from procedural to explicit know ledge that I had experienced (only faster). This m ove­m ent from implicit inform ation em bedded in an efficient problem ­solving procedure, to rendering the know ledge progressively m ore explicit, is a them e that will recur th roughout the book. This is p re­cisely w ha t I th ink developm ent is about: Children are n o t satisfied w ith success in learning to talk or to solve problem s; they w an t to u nderstand how they do these things. A nd in seeking such u n d er­standing, they become little theorists.

D evelopm ent and learning, then, seem to take tw o com plem entary directions. O n the one h an d , they involve the gradual process of proceduralization (that is, rendering behavior m ore autom atic and less accessible). O n the o ther hand , they involve a process of "explicita­tion" and increasing accessibility (that is, representing explicitly in ­form ation that is implicit in the procedural representations sustaining the structure of behavior). Both are relevant to cognitive change, but the m ain focus of this book will be the process of representational explicitation—w hich, I posit, occurs in a variety of linguistic an d cog­nitive dom ains th roughou t developm ent.

The RR Model

For a num ber of years I have been building a m odel that incorporates a reiterative process of representational redescription. I call this the RR m odel. I will first m ake som e general points and then provide a sum m ary of the model.

The RR m odel attem pts to account for the way in w hich children 's representations become progressively m ore m anipulable and flexible, for the em ergence of conscious access to know ledge, and for children 's theory building. It involves a cyclical process by w hich inform ation

18 Chapter 1

already p resen t in the organism 's independently functioning, special- purpose representations, is m ade progressively available, via redes- criptive processes, to o ther parts of the cognitive system. In other w ords, representational redescription is a process by w hich implicit inform ation in the m ind subsequently becomes explicit knowledge to the m ind, first w ith in a dom ain and then som etim es across domains.

The process of representational redescription is posited to occur spontaneously as part of an internal drive tow ard the creation of intra­dom ain and inter-dom ain relationships. A lthough I shall stress the endogenous natu re of representational redescription, clearly the proc­ess m ay at tim es also be triggered by external influences.

The actual process of representational redescription is dom ain gen­eral, but it is affected by the form ánd the level of explicitness of the representations supporting particular domain-specific know ledge at a given time. W hen I state that representational redescription is dom ain general, I do n o t m ean to im ply that it involves a sim ultaneous change across dom ains. Rather, I m ean that, w ithin each dom ain, the process of representational redescription is the same. To reiterate: the RR m odel is a phase m odel, as opposed to a stage model. Stage models such as Piaget's are age-related and involve fundam ental changes across the entire cognitive system . R epresentational redescription, by contrast, is hypothesized to occur recurrently w ithin microdomains throughout developm ent, as well as in adulthood for some kinds of new learning.

I will deal w ith the RR m odel and the process of representational redescription again in chapters 7 and 8. But it is essential to outline the m odel here in order to situate theoretically the empirical research in the following chapters on children as linguists, physicists, m athe­maticians, psychologists, and notators. At this stage the account may seem rather abstract, bu t hang in there. I prom ise that it will become m ore tangible once I deal w ith the specific dom ains in chapters 2 through 6. I also hope that the piano and Rubik's cube analogies will help sustain the discussion.

Let us now look at the RR m odel in a little detail. Development, I argue, involves th ree recurrent phases. D uring the first phase the child focuses predom inantly on inform ation from the external environm ent. This initial learning is data driven. During phase 1, for any m icrodo­main, the child focuses on external data to create "representational adjunctions." R epresentational adjunctions, I hypothesize, neither alter existing stable representations nor are brought into relation w ith them . Once new representations are stable, they are simply added, dom ain specifically, to the existing stock, w ith m inim al effect on w hat is already stored. In o ther w ords, independently stored representa­

Taking Development Seriously 19

tional adjunctions do no t yet entail w hat I m ean by representational change. Phase 1 culm inates in consistently successful perform ance on w hatever m icrodom ain has reached that level. This is w ha t I term "behavioral mastery. "

Behavioral m astery does no t necessarily imply that the underly ing representations are like the adult's. Successful perform ance can be generated by a sequence of independen tly stored representations that will ultim ately have to be linked into a m ore coherent system . The sam e perform ance (say, correctly producing a particular linguistic form, or m anaging to balance blocks on a narrow support) can be generated at various ages by very different representations. Later (phase-3) behavior m ay appear identical to phase-1 behavior. We thus need to draw a distinction, as illustrated in figure 1.2, betw een behav­ioral change (which som etim es gives rise to a U -shaped developm ental curve) an d representational change, because behavioral m astery is not tan tam ount to the end po in t of the developm ental progression in a given m icrodom ain.

Phase 1 is followed by an internally driven phase during w hich the child no longer focuses on the external data. Rather, system -internal dynam ics take over such th a t internal representations becom e the focus of change. In phase 2, the current state of the child 's rep resen­tations of know ledge in a m icrodom ain predom inates over inform a­tion from the incom ing data. The tem porary disregard for features of the external environm ent during phase 2 can lead to new errors and

100

80 -

60 -

40 -

20 -

3 4 65 7 8 9 1 0

age

Figure 1.2Behavioral change (EJ) versus representational change (♦).

20 Chapter 1

inflexibilities. This can, bu t does not necessarily, give rise to a decrease in successful behavior—a U -shaped developm ental curve. As figure 1.2 show s, how ever, this is deterioration at the behavioral level, not a t the representational level.

Finally, during phase 3, internal representations and external data are reconciled, an d a balance is achieved betw een the quests for in­ternal and external control. In the case of language, for example, a new m apping is m ade betw een inpu t and o u tpu t representations in order to restore correct usage.

But w hat abou t the form at of the internal representations that sus­tain these reiterated phases? The RR m odel argues for a t least four levels at w hich know ledge is represented and re-represented. I have term ed them Im plicit (I), Explicit-1 (El), Explicit-2 (E2), and Explicit-3 (E3). These different form s of representation do n o t constitute age- related stages of developm ental change. Rather, they are parts of a reiterative cycle th a t occurs again and again w ithin different micro­dom ains and th roughou t the developm ental span.

The RR m odel postu lates different representational form ats at dif­ferent levels. A t level I, representations are in the form of procedures for analyzing an d responding to stimuli in the external environm ent. A num ber of constrain ts operate on the representational adjunctions that are form ed a t th is level:

Inform ation is encoded in procedural form.

The procedure-like encodings are sequentially specified.

N ew representa tions are independently stored.

Level-I represen ta tions are bracketed, and hence no intra-dom ainor inter-dom ain representational links can yet be formed.

Inform ation em bedded in level-I representations is therefore no t avail­able to o ther operators in the cognitive system . Thus, if tw o procedures contain identical inform ation, this potential inter-representational com m onality is n o t yet represented in the child 's m ind. A procedure as a whole is available as data to o ther operators; however, its component parts are not. It takes developm ental time and representational redes­cription (see discussion of level E l below) for com ponent parts to become accessible to potential intra-dom ain links, a process which ultim ately leads (see discussion of levels E2 and E3) to in ter-represen­tational flexibility and creative problem -solving capacities. But at this first level, the potential representational links and the inform ation em bedded in procedures rem ain implicit. This gives rise to the ability to com pute specific in p u ts in preferential ways and to respond rapidly

Taking Development Seriously 21

and effectively to the environm ent. But the behavior generated from level-I representations is relatively inflexible.

The RR m odel posits a subsequen t reiterative process of represen­tational redescription.22 This involves levels E l, E2, and E3. Level-El representations are the results of redescription, into a new com pressed format, of the procedurally encoded representations at level I. The redescriptions are abstractions in a higher-level language, an d unlike level-I representations they are not bracketed (that is, the com ponent parts are open to potential intra-dom ain and inter-dom ain represen­tational links).

The E l representations are reduced descriptions that lose m any of the details of the procedurally encoded inform ation. As a nice example of w hat I have in m ind here, consider the details of the grated image delivered to the perceptual system of a person w ho sees a zebra (M andler, in press). A redescription of this into "striped anim al" (ei­ther linguistic or image-like) has lost m any of the perceptual details. I w ould add th a t the redescrip tion allows the cognitive system to u n ­derstand the analogy betw een an actual zebra and the road sign for a zebra crossing (a European crossw alk w ith broad, regular, black and yellow stripes), a lthough the zebra and the road sign deliver very different inpu ts to the perceptual system . A species w ithou t represen­tational redescriptions w ould n o t m ake the analogy betw een the zebra and the "zebra crossing" sign. The redescribed represen tation is, on the one hand , sim pler and less special purpose but, on the other, m ore cognitively flexible (because it is transportable to o ther goals). Unlike perceptual representations, conceptual redescriptions are productive; they make possible the invention of new term s (e.g. "zebrin ," the antibody w hich stains certain classes of cells in striped patterns).

N ote that the original level-I representations rem ain intact in the child 's m ind and can continue to be called for particular cognitive goals w hich require speed and autom aticity. The redescribed repre­sentations are used for o ther goals w here explicit know ledge is required.

A lthough the process of representational redescription can occur on line, I posit that it also takes place w ithou t ongoing analysis of incom ­ing data or production of o u tp u t. Thus, change m ay occur outside norm al inpu t/ou tpu t relations, i.e. sim ply as the p roduct of system- internal dynam ics, w hen there are no external pressures of any kind. I will come back to this po in t in a m om ent.

As representations are redescribed in to the E l form at, w e w itness the beginnings of a flexible cognitive system up o n w hich the child's nascent theories can subsequently be built. Level E l involves explicitly defined representations th a t can be m anipulated and related to other

22 Chapter I

redescribed representations. Level-El. representations thus go beyond the constraints im posed at level I, w here procedural-like representa­tions are sim ply used in response to external stimuli. Once know ledge previously em bedded in procedures is explicitly defined, the potential relationships betw een procedural com ponents can then be m arked and represen ted internally. I examine several examples of this below, particularly in chapters 2 and 3. M oreover, once redescription has taken place and explicit representations become m anipulable, the child can then in troduce violations to her data-driven, veridical descriptions of the world—violations w hich allow for p re tend play, false belief, and the use of counterfactuals. This I explore in detail in chapter 5.

It is im portan t to stress that although E l representations are avail­able as data to the system , they are not necessarily available to con­scious access and verbal report. T hroughout the book w e shall examine examples of the form ation of explicit representations w hich are not yet accessible to conscious reflection and verbal report, bu t w hich are clearly beyond the procedural level. In general, developm entalists have not d istinguished betw een implicitly stored know ledge and E l representations in w hich know ledge is explicitly represented bu t is not yet consciously accessible. Rather, they have d raw n a dichotom y betw een an undefined notion of som ething implicit in behavior (as if inform ation w ere no t represen ted in any form) and consciously acces­sible know ledge that can be stated in verbal form. The RR m odel postulates that the h u m an representational system is far m ore complex than a m ere dichotom y w ould suggest. I argue that there are m ore than two kinds of representation. Levels exist betw een implicitly stored procedural inform ation and verbally statable declarative know l­edge. It is particularly via a developm ental perspective that one can p inpoint this m ultiplicity of levels of representational formats.

The RR m odel posits that only at levels beyond E l are conscious access and verbal repo rt possible. A t level E2, it is hypothesized, representations are available to conscious access b u t not to verbal report (which is possible only at level E3). A lthough for som e theorists consciousness is reduced to verbal reportability, the RR m odel claims that E2 representa tions are accessible to consciousness bu t that they are in a sim ilar represen tational code as the E l representations of w hich they are redescriptions. Thus, for example, E l spatial represen­tations are recoded into consciously accessible E2 spatial represen ta­tions. We often draw diagram s of problem s w e cannot verbalize. The end result of these various redescriptions is the existence in the m ind of m ultiple representa tions of similar know ledge at different levels of detail and explicitness.

Taking Development Seriously 23

At level E3, know ledge is recoded into a cross-system code. This com m on form at is hypothesized to be close enough to natural lan­guage for easy translation into statable, com m unicable form. It is possible that som e know ledge learned directly in linguistic form is im m ediately stored at level E3.23 C hildren learn a lot from verbal interaction w ith o thers. H ow ever, know ledge may be stored in lin­guistic code but not yet be linked to similar know ledge stored in other codes. O ften linguistic know ledge (e.g., a m athem atical principle gov­erning subtraction) does no t constrain nonlinguistic know ledge (e.g., an algorithm used for actually doing subtraction24) until both have been redescribed into a sim ilar form at so that inter-representational constraints can operate.

In the following chapters, I distinguish three levels of represen ta­tional format: I, E l, and E2/3. For the p resen t purposes, I do not distinguish betw een levels E2 and E3, both of w hich involve conscious access. N o research has th u s far been directly focused on the E2 level (conscious access w ithou t verbal report); m ost if no t all metacognitive studies focus on verbal report (i.e., level E3). However, as m entioned above, I do no t w ish to foreclose the possibility of consciously acces­sible spatial, kinesthetic, an d other non-linguistically-encoded representations.

There are thus m ultiple levels at w hich the sam e know ledge is re­represented . This no tion of m ultiple encoding is im portant; develop­m ent does not seem to be a drive for economy. The m ind m ay indeed tu rn out to be a very red u n d an t store of know ledge and processes.

Before I conclude m y account of the RR model, it is im portan t to draw a distinction betw een the process of representational redescrip­tion and the ways in w hich it m ight be realized in a model. The process involves recoding inform ation that is stored in one representational form at or code into a different one. Thus, a spatial representation m ight be recoded into linguistic form at, or a proprioceptive represen­tation into spatial format. Each redescription, or re-representation, is a m ore condensed o r com pressed version of the previous level. We have just seen that the RR m odel postulates at least four hierarchically organized levels at w hich the process of representational redescription occurs. Now , empirical data m ight refute the existence of this hierar­chy (i.e., refute the RR model) w hile leaving the process of represen ta­tional redescription unchallenged. Indeed, as figure 1.3 illustrates, there are several alternative m odels in w hich the process of represen­tational redescription m ight be realized. First, as the RR m odel pre­sum es, it could involve the passage from implicit representations to a level of explicitly defined representations which are not available to conscious access (level E l), and finally to a format w hich is available

24 Chapter 1

I f o r m a t I f o r m a t

E l f o r m a t

E2 f o r m a t

•УE3 f o r m a t

E l f o r m a t

E2 f o r m a t E3 f o r m a t

I f o r m a t

E2 f o r m a tE l f o r m a t E3 fo rm a t

Figure 1.3Possible m odels of RR.

to conscious access (level E2) and verbal report (level E3). A n alter­native view w ould be that implicit representations are redescribed directly into either the E l, the E2, or the E3 format. Thus, inform ation m ight be directly recoded into linguistic form, ra ther than via the El level (as the RR m odel posits).

M odels can also differ w ith respect to constraints on the process of representational redescription. For example, a m odel m ight postulate that redescription in to one or two different form ats occurs autom ati­cally every time new in p u t is com puted and stored. By contrast, the RR m odel posits that in m ost instances a period of behavioral m astery m ust be reached before redescription occurs. Again, if it were show n that redescription occurs before behavioral m astery, the m odel w ould require m odification bu t the general concept of the process of rep re­sentational redescription w ould rem ain. The RR m odel argues for three recurren t phases leading to behavioral m astery and beyond. Again, w ere it show n that such phases did not exist, the process of redescription w ould no t necessarily be refuted. O n the other hand , if the process of represen tational redescription were to lose its plausi­bility (i.e., if all represen ta tions in the m ind w ere of equivalent status, or if totally distinct constraints were operative on procedural versus

'faking Development Seriously 25

declarative know ledge, ra ther than each level involving redescription of the previous one), then dearly the m odel w ould lose plausibility, too.

Let m e again stress the concept of reiterative developm ental phases. At any given time the child m ay have only level-I representations available in one m icrodom ain, b u t m ay have E l representations avail­able in ano ther m icrodom ain and E2/3 representations in yet another. This obviously holds across dom ains, too. It is hypothesized that there are no overarching dom ain-general changes in representational format at any given age. There is no such thing as a "phase E2 child". The child 's representations are in E2 form at w ith respect to a given m icrodom ain.

The actual process of representational redescription is considered dom ain general, b u t it operates w ithin each specific dom ain at differ­en t m om ents and is constrained by the contents and level of explicit­ness of representations in each m icrodom ain. Again, w ere each level of representational redescription to tu rn o u t to occur across the board at identical ages (e.g., level I u p to age 2, level E l from age 2 to age 4, and E2/3 from age 5 on), w hich I deem m ost unlikely, then the m odel w ould be refu ted and the process w ould have a different theoretical status.

The m odel also posits that representational change within phases involves adding representational adjunctions. Here negative feedback (failure, incom pletion, inadequacy, m ism atch betw een in p u t and out­put, etc.) plays an im portan t role, leading progressively to behavioral m astery.25 But in the transition between phases, it is hypothesized that positive feedback is essential to the onset of representational redescrip­tion. In o ther w ords, it is representations that have reached a stable state (the child having reached behavioral m astery) that are redescribed.

This success-based view of cognitive change contrasts w ith Piaget's view. For Piaget, a system in a state of stability w ould n o t sponta­neously im prove itself. Rather, the Piagetian process of equilibration takes place w hen the system is in a state of disequilibrium . The RR m odel also runs counter to the behaviorist view that change occurs as the result of failure or external reinforcem ent. Rather, for the RR model certain types of change take place after the child is successful (i.e., already producing the correct linguistic ou tpu t, or already having consistently reached a problem -solving goal). R epresentational redes­cription is a process of "appropriating" stable states to extract the inform ation they contain, w hich can th en be used m ore flexibly for o ther purposes.

26 Chapter 1

I do not, of course, deny the role of cognitive conflict in generating other types of change (through, for instance, the m ism atch betw een theory-driven expectations and actual outcomes). W hat I am stressing here is the additional— and , I hypothesize, crucial—role of internal system stability as the basis for generating representational redescrip­tion. A nd it is from the repeated process of representational redes­cription, ra ther th an sim ply from interaction w ith the external environm ent, that cognitive flexibility and consciousness ultim ately emerge.

The Importance of a Developmental Perspective on Cognitive Science

If our focus is o n cognitive flexibility and conscious access to know l­edge, w hy no t explore the data from adult psychology? Surely adults are far m ore cognitively flexible than children, so w hat justifies a developm ental perspective? Not, rest assured, the fact that data from children are cute. O ne only has to glance at the developm ental liter­ature to notice that a sizable num ber of researchers are absorbed w ith the ages at w hich children reach cognitive m ilestones. But others— and I count m yself am ong them — use the study of developm ent as a theoretical tool for exploring the hum an m ind from a cognitive science perspective. We are no t really in terested in children p er se.26

A developm ental perspective is essential to the analysis of hum an cognition, because understand ing the built-in architecture of the h u ­m an m ind, the constrain ts on learning, and how knowledge changes progressively over time can provide subtle clues to its final representa­tional form at in the adu lt m ind. The w ork of Spelke (1990), w hich I discuss in chapter 3, has been particularly influential in pointing to the im portance of a developm ental perspective on cognitive science.27 For example, the processes for segm enting visual arrays into objects are overlaid, in adults, by other processes for recognizing object cat­egories. But by focusing on how very young infants segm ent visual arrays into objects before they are able to categorize certain object kinds, Spelke is able to generate new hypotheses about how the adult visual system m ay function .28

Furtherm ore, distinctions such as declarative/procedural, con­scious/unconscious, an d controlled/autom atic, w hich are often used to explain adu lt processing, tu rn out to involve far m ore than a di­chotom y w hen explored w ithin a developm ental context. But in as­sum ing a developm ental perspective we m ust take the notion "developm ental" seriously. Paradoxically, studies on neonates and infants are often n o t developm ental at all. Like studies on adults, they frequently focus n o t on change b u t on real-time processing w ithin

Taking Development Seriously 27

steady-state system s. It is of course essential to determ ine the initial state of the hum an m ind— and for certain abilities the initial state is not necessarily presen t at birth b u t is p resen t only after the necessary neurological structures have reached m aturation (M ehler and Fox1985). The notion "developm ental" goes beyond the specification of the initial state, however. A nd a developm ental perspective does not apply m erely to the details of on-line, steady-state processing in chil­dren. Also, it does no t sim ply m ean a focus on learning in children of different ages ra ther than the adult. W hen m aking theoretical use of developm ent w ithin a cognitive science perspective, the specific age at w hich children can successfully perform a task is, to som e extent, irrelevant.

The fundam ental im plication of a developm ental perspective in­volves behavioral and representational change over time. I shall often use a later phase in a developm ental sequence to understand the status of representations underly ing earlier behavior—particularly in the in­teresting cases w here child and adu lt behaviors are practically iden­tical. This notion of representational change over time will be m y focus th roughout this book. It is for all these reasons that I m aintain that a developm ental perspective has m uch to offer cognitive science's ef­forts to m ore fully understand the adult m ind.

The Importance of a Cognitive Science Perspective on Development

Cognitive science focuses on cognition as a form of com putation, and on the m ind as a complex system that receives, stores, transform s, retrieves, and transm its inform ation. To do so it uses a variety of disciplines: psychology, philosophy, anthropology, ethology, linguis­tics, com puter science, and neuroscience. I have pointed to the im ­portance of a developm ental perspective on cognitive science. But w hat about the converse? W hat difference does it m ake w hether or n o t w e study developm ental psychology from a cognitive science perspective?

C onsider this analogy. C om puter scientists use com puters in two ra ther different ways: as a practical tool and as a theoretical tool (Rutkowska 1987). W hen com puters are u sed to solve practical prob­lem s such as designing robots and expert system s, the focus is on successful behavior; how the com puter does its job is irrelevant (A. Clark 1987,1989). Thus, the in troduction of a "kludge" (som ething that rem ains unexplained bu t th a t w orks for a particular task) poses no problem . But w hen m odelers use com puters as theoretical tools for sim ulating m ental processes and testing psychological theories, the focus shifts to questions abou t appropriate architectures and mech-

28 Chapter 1

anism s and abou t the na tu re of representations. H ow the com puter does its job then becom es a central concern.

Similarly, developm ental psychologists fall, grosso modo, into two categories: those w ho see the study of children as an end in itself and those w ho use it as a theoretical tool to understand the w orkings of the hum an m ind in general. In the form er case, as m entioned above, m any developm entalists focus on behavior—e.g ., on the particular age at w hich the child can do X. Decades of developm ental research w ere w asted, in m y view, because the focus w as entirely on low ering the age at w hich children could perform a task successfully, w ithout concern for how they processed the inform ation. I once began an article (Karmiloff-Smith 1981, p . 151) as follows: ''The enticing yet aw ful fact about child developm ent is that children develop! Awful, because it has provoked a p lethora of studies, totally unm otivated theoretically, accepted for publication in certain types of journal because the results are 'significant'— significant statistically, since it is indeed easy to obtain differential effects betw een, say, 5 and 7 year olds, bu t questionable as to their significance scientifically." Fortunately, how ever, the study of children is u sed w ith in a cognitive science perspective also—i.e., as a theoretical m eans of understand ing the hum an m ind in general. In such w ork, the focus is on the initial architecture, the processing m echanism s, and the na tu re of internal representational change.

M any recent books an d articles have focused on w hat cognitive science and inform ation-processing m odels m ight offer the study of developm ent (Bechtel an d A braham sen 1991; A. Clark 1989; Klahr et al. 1987; Klahr 1992; McTear 1987). In this book, m y aim is to h ighlight w hy a developm ental perspective is essential to cognitive science.

The Plan of the Book

The first part of each of the following five chapters— on the child as a linguist, a physicist, a m athem atician, a psychologist, and a notator— concentrates on the initial state of the infant m ind and on subsequent dom ain-specific learning in infancy and early childhood. Each chapter then goes on to explore empirical data on older children 's problem solving and theory building, w ith particular focus on cognitive flexi­bility and m etacognition.

I m ight have devoted a separate chapter to the child as a concept former, since there has been extensive research on this topic.29 H ow ­ever, conceptual developm ent is relevant to each of chapters 2-6: how children categorize objects in the physical w orld, how they m athe- m atize that w orld, how they conceive of hum an agents versus physical

Taking Development Seriously 29

objects, and how they encode that know ledge linguistically and in external notations such as d raw ing and m aps. C oncept form ation will thus perm eate each chapter ra th er than be treated separately.

In chapters 7 and 8 1 take ano ther look at the reconciliation betw een nativism and Piaget's constructivism , and discuss the need for m ore formal developm ental m odels. H ere I com pare aspects of the RR m odel w ith connectionist sim ulations of developm ent. A t all tim es, I place particular em phasis on the status of representations sustaining differ­ent capacities and on the m ultiple levels at w hich know ledge is stored and accessible. I end the book by taking a final look at the RR m odel and speculating on the status of representations in nonhum ans, w hich—how ever complex their behaviors—never become linguists, physicists, m athem aticians, psychologists, or notators.

Chapter 2

The Child as a Linguist

. . . young children know something about language that the spider does not know about web-weaving. (Gleitman et al. 1972, p. 160)

“What's that?"(Mother: "A typewriter.“)“No, you're the typewriter, that's a typewrite."

(Yara, 4 years)

W hat m akes us specifically hum an: the complexity of our language? our problem -solving strategies? You m ay be shocked by my suggestion that, in som e very deep sense, language and some aspects of hum an problem solving are no m ore o r less complex than the behaviors of o ther species. Complexity as such is n o t the issue. Spiders weave complex w ebs, bees transm it complex inform ation about sources and quality of nectar, ants interact in complex colonies, beavers build complex dam s, chim panzees have complex problem -solving strate­gies, and hum ans use complex language. A nd there are hum ans w ho acquire fluent language even though they are unable to solve certain problem s that the nonlinguistic chim panzee can solve. So it is n o t a lack of general problem -solving skills that stops the chim panzee from acquiring language. Som ething about the capacity to acquire language m ust be innately specified in hum ans. A lthough language is specific to hum ans, there is also a dom ain-general difference betw een hum an and nonhum an intelligence. Unlike the spider, w hich stops at w eb weaving, the hum an child—and , I m aintain, only the hum an child— has the potential to take its ow n representations as objects of cognitive attention. Norm ally developing children not only becom e efficient users of language; they also spontaneously become little gram m arians. By contrast, the constraints on spiders, ants, beavers, an d probably even chim panzees are such th a t they do not have the potential to analyze their ow n knowledge.

32 Chapter 2

This cross-species difference is beautifully captured by the quotation from the 4-year-old a t the beginning of this chapter. A "typewrite"! W hy do esn 't the child sim ply accept the label provided by the adult and use the correct w ord , "typew riter"? W hy has she bothered to w ork out that the formal function of the suffix "-er" is agentive—i.e., that one can often take verb stem s and add "-er" to form a w ord for a hum an agent {baker, dancer, teacher), so w hy not typew riter? The child 's ability cannot be explained aw ay solely on the basis of statistical regularities in the inpu t. The latter m ight give rise to a sporadic error in outpu t, such as the use of "typew rite" to refer to the object, or to an occasional m iscom prehension that "typew riter" refers to a hum an agent. But statistical regularities cannot explain w hy the child bothers to go beyond the in p u t/o u tp u t relations and reflect m etalinguistically on the w ord.

T hroughout the book, I shall argue that w hat is special about h u ­m ans is the fact that they spontaneously go beyond successful behav­ior. In the case of language, as in other areas of cognition, norm ally developing children are n o t content w ith using the right w ords and structures; they go beyond expert usage to exploit the linguistic know l­edge that they have already stored. I argue that this is possible via the repeated process of representational redescription discussed in chap­ter 1. M etalinguistic reflection requires flexible and m anipulable lin­guistic representations.

In every dom ain th a t w e explore, tw o major theoretical positions divide developm entalists into ra ther rigidly opposing camps: either acquisition is dom ain general or it is dom ain specific. The m odularity/ nativist view of language acquisition is that it is dom ain specific; i.e., that innately specified linguistic structures constrain the child 's proc­essing of linguistic in p u t.1 The strictly dom ain-general view considers language to be m erely a special case of other, dom ain-general struc­tures and processes.2

I shall argue that language acquisition is both dom ain specific and dom ain general; i.e ., that som e initial domain-specific constraints channel the progressive build ing u p of domain-specific linguistic rep­resentations b u t that, once redescribed, these representations become available to dom ain-general processes. This results in m ultiple repre­sentations of sim ilar linguistic inform ation, b u t in different represen­tational form ats. In o ther w ords, I shall agree w ith aspects of the nativist thesis as far as the very early stages of language acquisition are concerned, b u t w ith the proviso that we invoke a less static notion of a fully prespecified linguistic m odule, in favor of progressive m od­ularization. O verw helm ing data now exist to substantiate the hypo th­esis that, from the outset, infants process linguistic data in

The Child as a Linguist 33

linguistically constrained ways. These attention biases serve to build up linguistically relevant representations, not solely dom ain-general sensorim otor ones. H owever, unlike dyed-in-the-w ool nativists, I m aintain that that is not all there is to language acquisition. The innate specification m akes infants especially attentive to linguistic inpu t and sets the boundaries w ithin w hich language acquisition can take place; however, a m ore constructivist position opens up possibilities for representational flexibility, w hich ultim ately leads to m etalinguistic aw areness. Let us first look at the ra ther different w ays in which dom ain-general and dom ain-specific theorists view early language acquisition.

Language Acquisition as a Domain-General Process: The Piagetian Infant

If you w ere a disciple of Piaget, how w ould you explain the onset of language? First, you w ould no t g rant the neonate any innately speci­fied linguistic structures or m echanism s w hich are preferentially attentive to linguistic input. Indeed, Piagetians m aintain that both syntax and sem antics are solely products of the general structural or­ganization of sensorim otor intelligence. The culm ination of the sen­sorim otor period is the first time, according to the theory, that the infant is capable of symbolic representation. In explaining the timing of the onset of language a t around 18 m onths, Piagetians m ake no appeal to possible m atuxational constraints. Rather, they m aintain that language does no t appear earlier because it is an integral part of the onset of the symbolic (or semiotic) function, w hich includes not only language b u t also deferred im itation, p re tend play, and m ental im­agery. For Piagetians, language is not an independently developing capacity. They explain the late onset of language by pointing to the time it takes for sensorim otor action schemes to become progressively coordinated and internalized so as to m ake symbolic representation possible.

But can one really deny the young infan t's capacity for symbolic representation? To do so one w ould have to ignore the cogent argu­m ents of M andler (1983, 1988, in press), w hich are based on the now- extensive data suggesting the existence of symbolic representation early in infancy. How, M andler asks, could a young infant recall an action to be im itated after as long as 24 hours (Meltzoff 1988, 1990) w ithout the benefit of accessible know ledge represented in long-term memory? Likewise, how could an infant of 6-9 m onths recall the exact size of an object and precisely w here it w as located behind a screen (Ashm ead and Perlm utter 1980; Baillargéon 1986) if it could not rep ­resent them in an accessible form? In fact, the data which have accu-

34 Chapter 2

m ulated since the early 1980s call into question the very notion of a purely sensorim otor stage of hum an developm ent prior to language.

But a Piagetian disciple w ould have to ignore or re in terpret the new infancy data th a t I am abou t to discuss, and continue to argue that language is part of the semiotic function, available to the child only w ith the culm ination of sensorim otor intelligence. Indeed, Piagetians seek precursors of all aspects of language in the child 's sensorim otor interaction w ith the environm ent. Linguistic recursivity, for instance, is n o t traced to any dom ain-specific constraint. Rather, the Piagetians' explanation lies in a dom ain-general recursive process em erging from the infan t's earlier em bedding of sensorim otor action schem es such as seeing and grasping. This em bedding, they argue, is the product of postnatal circular reactions such as reiterated sucking (Sinclair 1971). Piagetians explain the em ergence of w ord order and difficulties there­w ith purely in term s of prior understand ing of the order of sensori­m otor actions. Playing w ith containers—em bedding objects one into another—is considered a necessary precursor to the em bedding of clauses. Piagetians see the cognitive concepts of agent, action, and patient as the prerequisites of early sentence structures (e.g. subject, verb, and object). N otions such as no u n phrase, verb phrase, subject, and clause are labeled adultom orphism s and said no t to be available to the young child 's linguistic com putations before the acquisition of elaborate cognitive structures. Indeed, some Piagetians m aintain that " the stages of cognitive developm ent determ ine the na tu re and the form of the linguistic structures that children are able to produce and und erstan d " (Ferreiro and Sinclair 1971) and, m ore recently, that "basic language com petence [is] constructed by the child subsequent to and on the m odel of the child 's fundam ental achievem ent during the pre-verbal practical intelligence period" (Sinclair 1987).

But w hat if, from a Piagetian stance, you favored C hom sky's struc­turalism w hile negating the nativist im plications of his theory? Indeed, Piagetians tend to hold onto Chom sky's (1965) now-obsolete account in term s of deep an d surface linguistic structures and different trans­form ations, only they see these structures and transform ations as special cases of p rio r cognitive structures and operations (Sinclair 1987). Piagetian psycholinguistics rem ain m ore com patible w ith C hom sky's earlier transform ational m odel than w ith his later m odel involving innately specified linguistic principles and param eters (Chom sky 1981), w hich they w ould have difficulty fitting into their cognitively based th ink ing about early language developm ent.3 Yet, paradoxically, C hom sky 's m ore recent theory, w hich is based not on rules bu t on principles, could be m ore readily in tegrated into an epi­genetic view than his earlier rule-based transform ational view.

The Child as a Linguist 35

Could the sam e m echanism s used for parsing visual scenes also account for specifically linguistic principles that determ ine semantic- syntactic relations? Vision an d language seem to adhere to their ow n domain-specific principles, a t least in adults. This does n o t necessarily im ply that the principles have to be innately specified in detail, though they m ay be. W hat it does suggest is that the infant w ould have to start o u t w ith innately specified linguistic predispositions an d a tten ­tion biases so as to constrain the class of inputs that it com putes in ways relevant to not violating specific linguistic principles, and not sim ply engage in a data-driven exercise of pattern m atching on the basis of the external in p u t and nonlinguistic, dom ain-general cognitive structures.4

M oreover, Piagetians are h ard pressed to explain the natural con­straints on children 's inferential pow ers. Were dom ain-general, cog­nitively based generalizations operative, then, given the in p u t data, the child w ould m ake m any inappropriate linguistic generalizations. But such generalizations are n o t m ade. Inferences that children do and d o n 't m ake in language acquisition are governed by specifically linguistic principles w hich constrain the class of inputs open to such generalizations. Dom ain specificity seem s to w in out over dom ain generality in the early stages of language acquisition. Yet Piagetians continue to explain all linguistic notions as deriving from cognitive ones and to see syntax and sem antics as generalizations from senso­rim otor and conceptual representations.

From the Piagetian stance, then , you m ight predict that linguistic retardation w ould necessarily accom pany severe cognitive retarda­tion. But such a prediction w ould tu rn out to be w rong. Indeed, studies of children w ith internal hydrocephaly and spina bifida (A nderson and Spain 1977; Crom er 1991; H adenius et al. 1962; Swischer and Pinsker 1971; Tew 1979) and of those w ith Williams Syndrom e (Bellugi et al. 1988; U dw in et al. 1987) show that complex syntax and léxico- m orphology (correct gram m ar, eloquent vocabulary, etc.) m ay coexist w ith very severe general cognitive im pairm ents.

In sum , as a Piagetian you w ould reduce linguistic universals to general cognitive universals and endorse Sinclair's recent statem ent (1987) that "language com petence and the w ay it develops in the child [is] an integral part of a general cognitive com petence."5

Language Acquisition as a Domain-Specißc Process: The Nativist Infant

H ow different your thinking w ould be if you w ere a domain-specific theorist! W ith a grow ing num ber of developm ental psycholinguists, you w ould argue that young children focus specifically on language

36 Chapter 2

as a problem space in its ow n right and not as part of dom ain-general input. Those taking a dom ain-specific view of language acquisition expect the neonate to possess a num ber of linguistically relevant a t­tention biases. They attribute the tim ing of the onset of language to innately specified m aturational constraints, ra ther than viewing it as the final outcom e of dom ain-general sensorim otor developm ent. For m any nativists, language is m odular (i.e., totally indep en d en t of other aspects of cognition).6 For o thers, it is dom ain specific ra ther than strictly m odular. In bo th cases, children 's learning of their native tongue is though t of as an innately guided process.

Thus, if you took a dom ain-specific stance on language acquisition, you w ould seek in the neonate and in the early infan t specifically linguistic precursors to the onset of language at 18 m onths. A nd your efforts w ould be rew arded.

A t least th ree problem s face the language-learning infant7: how to segm ent the speech stream into m eaningful linguistic units, how to analyze the w orld into objects and events relevant to linguistic encod­ing,8 and how to handle the m âpping betw een the un its and the objects and events at both the lexical and the syntactic level. The nativist argues that these problem s could n o t be surm ounted w ithout prior linguistically relevant processes that constrain the w ay in w hich the child com putes linguistic in p u t com pared to o ther auditory input. There m ust therefore be som e innate com ponent to the acquisition of language—but, to reiterate, this does n o t m ean that there has to be a ready-m ade m odule. A ttention biases and som e innate predisposi­tions could lead the child to focus on linguistically relevant in p u t and, w ith time, to build u p linguistic representations that are dom ain- specific. Since w e process language very rapidly, the system m ight w ith time close itself off from other influences—i.e., becom e relatively m odularized.

Let us now focus on research aim ed at uncovering the linguistic constraints on the neo n a te 's and the young in fan t's early language, and explore how such very young children build u p and store lin­guistically relevant representations. Recent research suggests that the infant's m ind com putes a constrained class of specifically linguistic inpu ts such that, in their in terpretation of sound waves, infants make a distinction betw een linguistically relevant and other, nonlinguistic auditory input. According to M ehler et al. (1986), 4-day-old infants are already sensitive to certain characteristics of their native tongue. Using the non-nutritive sucking habituation technique described above in chapter 1, M ehler tested French babies' sensitivity to the difference betw een French and Russian inpu t from the sam e bilingual speaker. Previous studies h ad already show n that 12-hour-old infants

The Child as a Linguist 37

differentiate betw een linguistically relevant inpu t and other nonlin­guistic acoustic inpu t. But M ehler's new research show ed that at birth infants do not yet react to differences betw een languages. Thus, the stimuli received during the 9 m onths in u tero do not provide suffi­ciently differentiated in p u t for the child to show preferential attention to its native tongue at birth. But only 4 days after birth—i.e ., after exceedingly little exposure, the infants studied by M ehler et al. show ed sensitivity to the different prosodic patterns of French and Russian.

It is no t m erely to overall phonological or prosodic patterns that young infants are sensitive. They also a ttend to features w hich will ultim ately have syntactic value, and they do so extremely early. Jus- czyk et al. (1989) stud ied infants raised in an English-speaking envi­ronm ent and found that at 4 m onths the infants w ere sensitive to cues that correlate w ith clause boundaries of bo th English and Polish input. By 6 m onths, how ever, the infants had lost their sensitivity to Polish clause boundaries, b u t they continued to dem onstrate sensitivity to clause boundaries in their native tongue. In o ther w ords, the archi­tecture of the infant m ind is such that it is sensitive at the outset to the clausal structure of any hum an language. Thus, som e fairly general features about the prosodic (and perhaps the syntactic) structure of hum an languages appear to be built into the system or to be learned exceedingly early on the basis of som e linguistic predispositions. These early sensitivities channel the in fan t's com putation of all sub­sequen t in p u t and serve to progressively select the appropriate struc­tures for the child 's native tongue and stabilize them .

Such data suggest that som e specifically linguistic predispositions and attention biases allow the in fan t to learn any hum an language, and that, in interaction w ith the particular environm ental in p u t from the child 's native tongue(s), particular pathw ays for representing and processing language are selected. By puberty the o ther pathw ays are lost, and by then the processing of language in a native-like way has becom e relatively m odularized.

F urther grist for the anti-dom ain-general mill comes from w ork show ing that stabilization of phonologically relevant perceptual cate­gories does n o t require the prio r establishm ent of sensorim otor pro­gram s (M ehler and Bertoncini 1988). Experim ents have also been devised to dem onstrate that infants are sensitive to the difference betw een relative pitch, w hich is linguistically relevant, an d absolute pitch (e.g., male versus female voice), w hich is socially relevant; to rhythm ic aspects of linguistic input; to vowel duration; to linguistic stress; to the contour of rising and falling intonation; and to subtle phonem ic distinctions.9

38 Chapter 2

Studies have also suggested that, well before they can talk, young infants are already sensitive to w ord boundaries (Gleitman et al. 1988) and to clause boundaries w ithin w hich gram m atical rules apply (Hirsh-Pasek et al. 1987). Using a preferential-listening procedure sim ­ilar to the preferential-looking procedure described above in chapter 1, H irsh-Pasek et al. had 7-10-m onth-old babies listen to tw o types of acoustic input. From a recording of a m other speaking to her child, m atched sam ples w ere constructed by inserting pauses either at nor­mal clause boundaries or at w ithin-clause locations. Already at 7 m onths, the babies oriented longer to the sam ples segm ented at the clause boundary th an to those in w hich the pauses violated such natural linguistic boundaries. In other w ords, the young infant already analyzes auditory in p u t according to phrase-boundary m arkers—that is, in a linguistically relevant w ay that will later su p p o rt the represen­tation of syntactic structure.

Interestingly, there have studies of infants w ho received no linguis­tic in p u t at first: congenitally deaf children born of hearing parents w ho did not know sign language. The exciting finding w as that, even though they lacked the benefit of the linguistic m odel available to hearing children an d to deaf children of signing deaf paren ts, these children nonetheless invented a visuom anual system that displayed several of the constraints of natural language (Goldin-M eadow and Feldm an 1979; Feldm an et al. 1978). Of course, their visuom anual system d id not develop in to a full-fledged sign language. M oving from linguistic predispositions to language-specific constraints (French, En­glish, Am erican Sign Language, Spanish, Polish, etc.) requires input. But such case studies once again point to the im portance of a dom ain- specific, innately guided process that can get language acquisition off the ground even in the absence of a model.

A nd w hen a linguistic m odel is available, young children are clearly attentive, no t to som e dom ain-general input, bu t to domain-specific inform ation relevant to language. M any data exist w hich dem onstrate the child 's early analysis of language as a formal domain-specific problem space (Bloom 1970; Karmiloff-Smith 1979a; Valian, 1986, 1990). O ne of the exam ples I find particularly telling comes from the w ork of Petitto (1987),10 w ho studied children 's acquisition of the personal p ronouns "you" and "I" in American Sign Language (ASL). Petitto 's subjects w ere congenitally deaf bu t w ere grow ing up in a norm al linguistic environm ent since their paren ts w ere native deaf signers. In ASL, personal p ronouns are am ong the few signs that resem ble natural gestures. "I" is encoded by pointing to oneself, "you" by pointing tow ard the addressee. Now, if sensorim otor action schemes w ere the necessary bases of dom ain-general acquisition of

The Child as a Linguist 39

language, the signs for "I" an d "you" should appear very early in ASL acquisition as a natural extension of gestures. Moreover, they should not show the deictic errors typical of spoken language, such as chil­dren 's tem porary m istake of using "you" to refer to them selves (Chiat 1986; Tanz 1980). N either of these dom ain-general predictions holds for ASL acquisition. Petitto 's data show that nonlinguistic pointing is p resent well before the syntactic use of p ronouns appears in ASL and that the linguistic use of p ronouns appears at the same tim e in ASL as in spoken language. This is no t to deny that o ther signs of a more lexical natu re m ay appear early in ASL (Bonvillian et al. 1983; M eier and N ew port 1990).11 But the po in t is that p ronouns are no t an exten­sion of gesture; rather, they are an integral part of the domain-specific developm ent of language as a system.

Im portant, too, is the fact that deaf children acquiring ASL as their native tongue actually do m ake subsequent errors in their use of the personal pronouns. They start by using p ronouns correctly. However, subsequently in developm ent they tem porarily use the ASL sign for "you" (a po int to the addressee) to refer to them selves, or they p ro­visionally replace personal p ronouns by the use of proper nam es. A nd children do this despite their earlier m astery and despite the seem ing transparency betw een the sem antic and the syntactic relations. But the deaf child ignores the indexical aspect of the signs (the pointing gestures that correspond to sem antic inform ation) and focuses on the formal aspect of the signs (personal p ronouns as a formal linguistic system).

The RR m odel w ould explain this developm ent in term s of changing representations. Initially the child focuses on the inpu t data an d stores two independen t level-I representations for "you" and "I." Subse­quently, once the child is p roducing consistently efficient ou tpu t, the level-I representations are redescribed such that the linguistic com­ponen ts m arking personal pronom inal reference are explicitly defined in E l form at. Then links across the two entries' com m on com ponents can be draw n, such that new representations can form a subsystem of personal pronouns.

Late-occurring errors constitu te a striking illustration of how do ­main-specific m echanism s of data abstraction constrain the child to analyze just those aspects of the inpu t th a t are relevant to the formal linguistic system . By contrast, different domain-specific m echanism s in terpre t identical in p u t in nonlinguistic ways (e.g., po in ting in ter­p re ted as a social gesture ra th er than as an arbitrary linguistic sign).

A nother example of the dom ain specificity of data abstraction and production m echanism s com es from neuropsychological studies of brain dam age in deaf adu lt signers (Poizner et al. 1987). These patients

40 Chapter 2

can be show n to be capable of im itating a m ovem ent m anually and yet be incapable of p roducing the sam e form w hen this is being used in a linguistic context. In o ther w ords, the ou tpu t does not call on dom ain-general processes; m anually realized linguistic signs seem to be stored and processed separately from m anual gestures. It therefore seem s that the linguistic system becomes dom ain specific and rela­tively m odularized over the course of learning. It is these dom ain- specific processes that are im paired, bu t such dam age m ay have no effect on the capacity to produce nonlinguistic m anual gestures of similar form and complexity.

A t any time, th en , identical in p u t m ay be open to different in ter­pretations, depend ing on the particular domain-specific focus of the child (or adult). As far as language is concerned, from the outset and throughout the acquisition process, domain-specific constraints spec­ifying how to abstract and represen t linguistically relevant data seem to be operative for bo th sem antics and syntax.

The Infant's and the Young Child's Sensitivity to Semantic Constraints

H ow do children w ork o u t the m apping betw een concepts and the lexicon of their native tongue? Once again, w e shall see that p re­existing constraints narrow children 's hypotheses about the possible m eanings of w ords (Carey 1982; E. Clark 1987; Dockrell and Campbell 1986; Gleitm an 1990; Hall 1991; M arkm an 1987, 1989; M errim an and Bowman 1989). In this w ay a virtually insoluble induction problem is avoided. G leitm an (1990) provides a particularly illum inating discus­sion of the general issue of constraints on w ord learning. She poses Q uine's (1960) problem as follows: Given a linguistic o u tp u t and a situation to w hich it refers, how could an intelligent adult, let alone an 18-month-old toddler, settle on the m eaning of a new w ord in the face of the m ultitude of interpretative options available? For instance, w hen seeing an adu lt po in t tow ard a cat and say "Look, a cat", how can the child decide w hether the speaker uses "cat" to m ean the whole anim al, the cat's w hiskers, the color of the cat's fur, the m at on w hich it is standing, the bow l of w ater from w hich it is drinking, the action of the cat's licking its fur, the noise of its purring, the ribbon around its neck, the fact that the speaker likes anim als, or the background details of the scene (and so on, ad infinitum )? We shall see that the sam e potential induction problem s occur no t only for nouns b u t also for verbs. H ow can one infer from details in the external environm ent alone the linguistic distinction betw een, say, "look" and "see", or that betw een "chase" an d "flee"? These induction problem s w ould arise only, G leitm an argues, if the learner's sole am m unition w ere unaided

The Child as a Linguist 41

observation-based in terpretations of the scene being described linguis­tically. But this is no t the case. G leitm an proposes that the in fan t's perceptual an d conceptual processing of events and objects in the environm ent are constrained to specific levels of abstraction and tax­onomy. The child does not approach the w ord-learning task through m ere observation. Rather, the child 's hypothesis space w ith respect to the possible m eanings of the w ords in her language is subject to principled constraints. These result from domain-specific biases on m appings betw een objects/events and w ords, as well as from sensi­tivity to distinctions w ithin the linguistic system itself. Let m e deal w ith each of these in turn.

The first involves the interaction betw een linguistic constraints and those deriving from the child 's in terpretations of the physical w orld (via visual or, in the case of the blind, haptic perception). Carey (1982)i2 form ulates the problem succinctly by asking: W hen a child hears a w ord, to w hat ontological types does she assum e the w ord refers— w hole objects, features of objects, substances, or what? Do children build u p w ord m eanings solely on the basis of a com position of se­m antic features (round, furry, green, sharp , etc.), com ponent by com­ponen t (E. Clark 1973; Baron 1973), or are there constraints on possible w ord m eanings that bias the w ay in w hich the child in terprets the linguistic input? There have been several attem pts to answ er these questions, b u t the one m ost relevant to the p resen t discussion comes from a constraints view of early infancy and later w ord learning.

M andler (1988 and in press) has provided the m ost thoroughly w orked ou t speculations abou t the w ay in w hich young infants build representations that are suitable for subsequent linguistic encoding. According to M andler, young infants engage in a process of perceptual analysis w hich goes beyond their rapid and autom atic com putation of perceptual input. Perceptual analysis results in the form ation of perceptual prim itives such as s e l f - m o t i o n / c a u s e d m o t i o n / p a t h /

s u p p o r t / a g e n t . These prim itives guide the w ay in w hich infants parse events into separate entities that are supported or contained an d w hich m ove from sources to goals along specific kinds of paths according to w hether the m ovem ent is anim ate or inanim ate. M andler argues that these perceptual prim itives are redescribed into an accessible image- schematic form at, thereby providing a level of represen tation in ter­m ediate betw een perception and language. A nd it is these accessible im age schem as that facilitate sem antic developm ent (i.e., the m apping betw een language and conceptual categories). Image schem as are nonpropositional, analog representations of spatial relations and m ovem ents; that is, they are conceptual structures m apped from spa­tial structure.

42 Chapter 2

The redescription of perceptual prim itives into image-schematic rep­resentations, and of the latter into language, indicates how the RR m odel outlined in chapter 1 can be applied to very early infancy. I have stressed the fact th a t representational redescription can occur outside input/ou tput relations. M andler extends the RR m odel to on-line proc­essing, suggesting that redescription also takes place as the child is actively engaged in analyzing perceptual in p u t and redescribing it into the m ore accessible form at of im age schemas. As w ith the RR model, M andler postulates that the form ation of image schemas requires an innately specified m echanism of analysis, not necessarily innately specified content.

The redescription into language of image schemas conceptualizing spatial relations suggests a tighter relationship betw een language and cognition for sem antics than in the case of syntax.

H ow do young children learn the m eanings of w ords in their lan­guage? Clearly they a ttend to the environm ent in w hich adults and others use such w ords and explain their m eaning. But is this enough? W hen the adu lt po in ts to som e object and says "That's an X", such ostensive definitions typically underdeterm ine a w ord 's m eaning se­verely. To surm ount this problem , children m ust bring to the word- learning situation a lim ited num ber of hypotheses about possible types of w ord m eaning. M arkm an and h er colleagues (H orton and M arkm an 1980; M arkm an 1980; M arkm an and Wachtel 198813) have show n that as of 3 years of age (and perhaps as early as 18 m onths, coinciding w ith the vocabulary bu rst [Bloom et al. 1985; Dromi 1987; McShane 1979; N elson 1973]), children seem to abide by three assum ptions about the m apping betw een w ords and their referents: the whole- object assum ption, the taxonomic assum ption, and the assum ption of m utual exclusivity. First, 3-5-year-olds assum e that a new label refers to an object as a w hole, and n o t to its substance, constituent parts, color, texture, size, shape, etc. Second, children extend a new ly ac­quired label to objects of the sam e taxonomic kind, ra ther than to objects that are related them atically to the original one. If a child hears "See the dax" and sees an adu lt pointing to an object, the child m aps "dax" onto the w hole object ra ther th an onto one of its parts, although nothing in the adu lt utterance indicates that. Also, children tend to assum e that new w ords refer to a basic category level (e.g., dog) rather th an to a superordinate or subordinate kind (e.g., anim al or poodle). The third assum ption calls for m utual exclusivity, such that, on hear­ing a new label (e.g., "viper"), children tend to apply it to an object for w hich they do no t yet have a label, given that o ther objects present in the array (say dogs, cats, etc.) are ones for w hich they already have

The Child as a Linguist 43

a label. This m eans that the child can learn a new w ord w ithout relying on any poin ting on the part of the speaker.

M arkm an poin ts ou t that these biases are no t determ inistic but probabilistic.14 They can be overridden w hen there is sufficient other inform ation to suggest an alternative in terpretation. The m utual-ex- clusivity assum ption, for example, leads children to expect that each object has only one label. Thus, on hearing "Look at its nice fu r" , as an adu lt points to a cat, a child w ho already know s the w ord for cat can use the m utual-exclusivity assum ption to override the whole- object assum ption and acquire a new w ord for a feature of the referent (fur). Likewise, on hearing "T hat's a nice animal" the child can override the basic-category-level assum ption and learn the superordinate label (animal).

Hall (1991) h as recently show n that similar biases constrain the way in w hich children come to understand w hat he calls "restricted" versus "unrestricted" m eanings of w ords. For example, w hereas the w ord "person" continues to refer to som eone th roughout his or h er life span and in any situation, w ords like "youth" or "passenger" refer only at certain tim es and in certain circum stances; they have restricted m ean­ings. M oreover, som eone can be sim ultaneously both a person and a passenger. These are intricate facts about w ord m eaning w ith w hich the child has to come to grips. H all's research show s that even adults are implicitly aw are of these difficulties. They tend to teach young children unrestricted w ords via pure ostensión, w hereas restricted m eanings are taugh t via a com bination of ostensión and direct expla­nations, providing clues to the learner about how to restrict these special m eanings. It is the absence of these extra clues that biases the child to take pu re ostensión as referring to whole kinds of objects at a m iddle category level ra ther th an to properties, parts, superordinate, or subordinate levels.

These various default assum ptions or biases w ork to guide chil­d ren 's initial hypotheses about noun m eanings, helping them disre­gard countless possible b u t incorrect inductions. But they can obviously lead to errors, too. Thus, the biases m ust be strong, yet flexible enough to be overridden by other, m ore pertinent inform ation. Clearly we need to draw a distinction betw een the probabilistic biases operative in the w orking o u t of possible sem antic m appings for w ord m eanings and the m ore determ inistic constraints of syntax.

The Infant's and the Young Child's Sensitivity to Syntactic Constraints

Those holding the view that the acquisition of language stem s from nonlinguistic cognitive constraints w ould be unlikely to entertain the

44 Chapter 2

idea that infants are sensitive to purely syntactic constraints on lin­guistic input. Yet Katz et al. (1974) show ed that 17-month-olds can use syntactic inform ation to distinguish betw een a noun referring to a class of objects and one functioning as a proper nam e. A nd this capacity w as apparen t well before the infants w ere using determ iners in their ow n outpu t. Thus, w hen infants heard "a dax" they chose another doll similar to the one the experim enter already nam ed as "a dax", w hereas w hen they heard "Dax" they chose the individual doll to w hom the experim enter had given the proper nam e "Dax." These data indicate that language is a problem space per se for infants well before they are producing m uch language them selves. In other w ords, infants m ake use of m orphosyntactic subtleties w ithin the linguistic system itself to w ork o u t m eaning.

But w hat about m ore complex aspects of syntax? Are infants sen­sitive to w ord order in linguistic strings and to differences betw een transitive and intransitive verb structures? Hirsh-Pasek et al. (1985) used the preferential-looking paradigm described above in chapter 1 to probe infants' sensitivity to w ord order. Infants w ith relatively little linguistic o u tpu t w ere show n tw o anim ated scenes on tw o screens. While w atching the displays, they heard from a h idden speaker a sentence that m atched only one of the tw o scenes. Significantly longer looking at the video that m atches the speech ou tpu t dem onstrates infants' sensitivity to the linguistic distinctions being encoded. Hirsh- Pasek et al. show ed that 17-month-old infants w ho w ere not yet producing anything like sentences could nonetheless distinguish be­tw een sentences like "Big Bird is tickling Cookie M onster" and "Cookie M onster is tickling Big Bird". If they w ere m erely relying on the w ords in the utterances, irrespective of w ord order, their looking time should have been random betw een the tw o displays. But this w as not the case. They looked significantly longer at the display that m atched the linguistic ou tput, dem onstrating that at this young age w ord order was already linguistically relevant to them .

A lthough infants are sensitive to w ord order, one should not con­found this w ith sensitivity to serial order. The order to w hich they are sensitive in language is dependen t on structure (i.e., the relative order of n o u n phrases an d verb phrases) (Chomsky 1987; Crain and Fodor, in press), n o t on the order of single w ords. D om ain-neutral theories argue that infants w ork o u t how their language functions on the basis of rules that order conceptual categories or real-world events. But this is not so. There is no conceptual reason why, for instance, pronouns and proper nam es cannot be modified by prenom inai adjectives. W hat, conceptually, w ould preclude children from pointing to two individuals and, using deictic pronouns, saying "Big he, little she"? But

The Child as a Linguist 45

according to studies by Bloom (1990), children never violate this spe­cifically linguistic constraint of English. Rather, both production and com prehension experim ents show that, in com puting language, chil­d ren order abstract linguistic categories, n o t conceptual categories. Chil­d ren analyze p ronouns as n o u n phrases, not as single w ords. N oun phrases cannot be m odified by prenom inai adjectives. That violates a constraint on English. Now, if children generated general cognitive hypotheses (i.e., no t specifically linguistic ones) in order to u n d er­stand adu lt language, surely they w ould opt for the sim plest hypo th ­esis (order of elem ents) instead of the co gn itive ly m ore complex hypothesis (order of structu re-dependen t phrases). But it is the lin­guistically relevant, dom ain-specific hypothesis that young children use.

H irsh-Pasek et al. (1988) explored children 's understand ing of an even m ore complex linguistic distinction. They w ere in terested in children 's processing of constraints on causative verbs. C hildren heard ou tpu ts such as "Big Bird is tu rn ing Cookie M onster" (or "Big Bird is tu rn ing w ith Cookie M onster") and saw two scenes on the video screens: one depicting Big Bird m aking Cookie M onster physically tu rn around, the o ther w ith the anim als bo th turning next to one another. The preferential-looking procedure was again used to assess w hether children looked longer at the scene m atching the verbal out­put. O ther trials used verbs w hich these young children w ere unlikely to have ever heard before, such as "flexing", in both transitive and intransitive sentences. A lthough stable effects w ere not established consistently at 24 m onths, by 27 m onths (long before such distinctions figure in their output) children looked significantly longer at the dis­play m atching the linguistic ou tpu t. These results allow us to conclude that shortly after the second birthday a child know s that only a tran­sitive verb expresses the presence of a causal agent and that causal agency cannot be in an oblique argum ent position (the with clause). Further, the child understands that the with clause excludes a transitive reading. It is difficult to see how children so young could have learned such subtle linguistic distinctions solely on the basis of dom ain-general sensorim otor actions.

The Need for Both Semantic and Syntactic Bootstrapping

Gleitm an (1990) draw s a distinction betw een sem antic bootstrapping (the use of sem antics to w ork o u t syntax) and syntactic bootstrapping (the use of syntax to predict semantics). While m ost developm ental psycholinguists have focused on one or the other of these processes,

46 Chapter 2

Gleitm an argues th a t language developm ent involves both. Both make critical use of canonical relations betw een syntax and semantics.

The sem antic bootstrapping hypothesis involves w ord-to-w orld m appings by w hich the child searches the observable environm ent for possible referential candidates (see Pinker 1984, 1987 for detailed accounts). G leitm an calls this the "observational learning hypothesis" and agrees that part of the child 's acquisition of verb m eanings takes place via this route. The distinctions betw een som e closely related verbs (e.g. break/tear, shatter/crum ble) m ust be w orked out by obser­vational learning, because these verbs do not differ in their syntactic fram es (Filmore 1968).

However, sem antic bootstrapping, although necessary, is no t suf­ficient to explain the norm al child 's acquisition of m any verb m ean­ings.15 In support of this claim, Landau and Gleitm an (1985) analyze the problem s besetting the congenitally blind child 's distinction be­tw een the verbs "see" and "look". How, they ask, could blind children use observational learning to guide their hypotheses about the m ean­ings of these tw o verbs? In fact, the problem obtains for sighted children, too. Evidence for w ord m eaning does not simply lie in the external environm ent of physical objects and actions. Rather, the ev­idence resides in the design of language itself, in the different subca­tegorization fram es in w hich these verbs can be u sed .16 Some verbs take three argum ent structures, som e two. Some verbs encode paths and goals expressed in prepositional phrases, others do not. The use of particular verbs w ith particular subcategorization frames depends on the perspective taken by the speaker. Do you in terpret an action betw een tw o people as one of "giving to" or "taking from ", as one of "fleeing from" or "chasing after", and so forth? The sam e event can be described in very different w ays. To work out the linguistic m ean­ings, the young child m ust be sensitive to these intralinguistic differ­ences. The structure of subcategorization frames helps children figure out speakers' in ten tions as well as the differences in verb m eanings used to describe potentially equivalent extralinguistic contexts.

The m ajor th rust of G leitm an's argum ent is that, blind or sighted, children cannot rely on observational learning alone. Rather, the child m ust also bring to the language learning situation relatively sophisti­cated presuppositions about the structure of language itself. C aretak­ers do n o t provide a runn ing com m entary on events and scenes in the world. A nd even if they did, ostensive definitions are u nderdeter­m ined. In any case, ad u lt o u tpu t also refers to things not happening in the here and now. For example, a father m ight say to his infant: "W hen you 're done w ith eating your dinner, we'll look at Sesame Street. Then D addy'll get you undressed, and take you u p to the bath

The Child as a Linguist 47

before M um m y gets back from w ork. O h, dear, look w hat you 've done now. You've d ropped it all over the floor. I'll get the broom . And, w e're late, listen, I can already h ea r M um m y's car." Simple word-to- w orld m appings w ould lead to u m pteen erroneous hypotheses about w ord m eanings in general, an d verb m eanings in particular.

G leitm an takes a different stance. She m aintains that children must- be using sentence-to-world m appings in trying to w ork out the sem antic distinctions betw een such closely related verbs as look/see, listen/hear, fall/drop, h ide/disappear, and chase/flee. The differences are rarely observable from the extralinguistic contexts in w hich they are used, bu t they can be inferred from the intralinguistic contexts in w hich they are u sed because of their different subcategorization frames. O ne says "I hid the ball" b u t n o t "I d isappeared the ball"; "I fled from the m an" but no t "The m an chased from m e"; "I looked at the ball" b u t not "I saw at the ball". It is the fact that similar m eanings can be expressed via verbs w ith different subcategorization frames that narrow s the interpretative options. In this way, the syntax functions, to use Gleit­m an 's w ords, "like a k ind of m ental zoom lens" for fixing on ju st the in terpretation am ong m any possible ones that the speaker is express­ing. Again, it is difficult to see how dom ain-general data-abstraction m echanism s could alone give rise to an understand ing of such subtle linguistic distinctions.

Beyond Infancy and Early Childhood

The extraordinary feat of language acquisition takes place effortlessly in a short span of time. By the tim e a child is 3 or 4 years old, she is speaking an d understand ing ra ther fluently. So is that all there is to language— a set of constraints for attending to, processing, an d rep­resenting linguistically relevant input; biases that constrain the way in w hich the child represen ts objects and events in the world; and the subsequent processes of sem antic and syntactic bootstrapping? Does acquisition involve no th ing m ore than reaching behavioral m astery of each aspect of the linguistic system ? Let us explore these questions by jum ping a couple of years and im agining our infant as having become a fluent speaker of h er native tongue.

The R R Model and Becoming a Little Linguist

The RR m odel, outlined in chapter 1, argues that norm al developm ent involves considerably m ore than reaching behavioral m astery. M an­dler has posited the form ation of im age-schematic representations that m ediate betw een perception an d language, and has used the process

48 Chapter 2

of representational redescription to account for the passage from one representational form at to another. The RR m odel further postulates that the linguistic representations them selves also undergo subse­quent redescription, such that they become linguistic objects of atten­tion outside their on-line use in com prehension and production. In other w ords, young children go beyond behavioral m astery, beyond fluent ou tpu t and successful com m unication, to exploit the linguistic know ledge they have already stored. It is this that ultim ately allows them to become little linguists.

The linguistic representations built up during infancy and early childhood serve young children for com prehending and producing their native tongue. But these initial linguistic representations are not, I argue, available as data for m etalinguistic reflection. They are stored and ru n as procedures for effective com prehension and production. They are, to use the m etaphor from chapter 1, inform ation in the m ind and not yet know ledge to the mind.

To become flexible an d m anipulable as data (level-El represen ta­tions) and thus ultim ately accessible to m etalinguistic reflection as well as to cross-dom ain relationships w ith other aspects of cognition (level- E2/3 representations), the know ledge em bedded implicitly in linguistic procedures (level-I representations) has to be re-represented.

It is, of course, easy to determ ine w hen a child has verbally statable m etalinguistic know ledge. But the RR m odel postulates a first level of redescription w hich is no t available for verbal report and for which m ore subtle empirical clues m ust be sought. The fact that such redes­cription does take place can be gleaned from late-occurring errors and self-repairs. Let us briefly consider three examples.

The first is from the acquisition of French. In French, the w ord "m es" is a plural possessive adjective (my 4- plural marker). "Ma voiture" m eans "m y car"; "m es voitures" m eans "m y cars". But, in contrast w ith English, in spoken French the plural m arker is heard on the possessive adjective ("mes"), no t on the noun ("voitures"). So the little w ord "m es" conveys a lot of inform ation in spoken French. My experim ents show ed th a t 4-year-olds use this term easily in situations w here possession and plurality have to be expressed ("mes voitures" implying "all my cars"). They have efficiently functioning level-I rep­resentations. By contrast, 6-year-olds spell out redundantly the m ean­ing com ponents of the w ord "m es". They use explicit m arkers for each of the implicit features in "m es", producing ou tpu ts such as "toutes les m iennes de voitures", w here totality is expressed by "toutes", plurality by "les", and possession by "m iennes". M etalinguistic ques­tioning at this age show s that the reasons for this explicit overm arking of features (level-El representations) are not available to conscious

The Child as a Linguist 49

access. That requires yet ano ther level (E2/3) of redescription. The overm arking subsequently disappears; older children again use "m es", b u t they can also explain the various m eaning com ponents of the possessive determ iner system (Karmiloff-Smith 1979a, 1986).

A similar example comes from N ew port's (1981) studies of the ac­quisition of Am erican Sign Language. In ASL signs have m orpholog­ical structure, bu t initially children use holistic signs (level-I representations). Deaf parents w ho are non-native signers (i.e., w ho acquired sign language late in life) cannot analyze the signs into their m orphological com ponent parts. By contrast, children acquiring ASL as a native language analyze its m orphological structure. They express that know ledge via late-occurring errors in their ou tpu t after they have been using the sign correctly for som e time. The errors involve separate staccato m ovem ents isolating tw o separate morphological m arkers, instead of the norm ally flowing holistic sign. It is som ething like the equivalent in spoken language of pronouncing the w ord "typew riter" correctly at first, an d then subsequently pronouncing it as "type - w rite - er". This extraction of com ponent parts from the initial holistic signs is again suggestive of representational redescrip­tion (level-El representations). N othing in N ew port's data suggests that children are consciously aw are of the segm ented form of their new productions. In o ther w ords, the representations are no t yet in E2/3 form at. The overm arking subsequently disappears; older children again use signs that look like the ones they used w hen younger. However, the RR m odel posits that the later identical o u tp u t stem s from representations m ore explicit than the procedural ones that u n ­derlie the initial productions.

N ote th a t in neither the ASL nor the French examples can the children get the com ponent m orphological inform ation directly from environm ental inpu t, because parents do not spell out the separate m orphological m arking in their productions. The "errors" in the French an d ASL exam ples suggest that the child analyzes the level-I representations and extracts the implicit inform ation that they contain. Since the original procedures rem ain intact and are produced concur­rently w ith the explicit overm arking, I argue that this analysis is carried ou t on redescriptions (El format) of the procedures. A nd it is these redescribed representations that are the basis for norm al children 's subsequent building of theories about language and for their re­sponses to m etalinguistic tasks17 (level-E2/3 representations). In other w ords, the external environm ent serves as input to linguistic attention biases to form and store linguistically relevant representations, bu t redescriptions of internal representations serve as the basis for further

50 Chapter 2

developm ent and for children 's spontaneous folk theories about how language functions as a system .

The th ird exam ple is from spontaneous self-repairs and their relation to subsequent m etalinguistic aw areness. H ere is a m etalinguistic ex­planation from a 10-year-old. The context w as two pens, one eraser, one earring, and the child 's ow n watch. The experim enter h id the child 's w atch an d th en asked "W hat did I do?". The exchange was as follows:

Child: You h id the watch.Exp: W hy did you say "the watch"?Child: W ell. . . "my w atch" because it belongs to m e, b u t I said, "you hid the w atch" because there are no other w atches there. If yo u 'd pu t yours out, I w ould have h ad to say "you hid m y w atch", because it could have been confusing, bu t this w ay it's better for m e to say "you hid the w atch" so som eone d oesn 't th ink yours w as there too.

This is an eloquent exam ple of how children can produce elaborate verbal statem ents once they have access to that part of their linguistic knowledge. (Note that correct usage of "the", "m y", etc. occurs m uch earlier, around 4-5 years of age.)

Now, if one w ere to consider only the difference betw een young children 's correct usage and older children 's m etalinguistic state­m ents, one w ould m erely postulate tw o levels of representation: the implicit level-I representations sustaining correct usage and the level- E2/3 representations sustaining the verbal explanations. To posit the existence of E l representations betw een the two, one needs to find other kinds of data. Spontaneous self-repairs tu rn ed out to be the clue I was seeking. Take the h id ing gam e outlined above. D uring testing, children often m ake self-repairs. They som etim es m ake lexical repairs: "You h id the pe . . . no, the w atch." A t o ther tim es they make referential repairs: "You h id the blue pe . . . the red pen ." But they also m ake w hat I call "system ic repairs": "You hid m y w at . . . the w atch ." (Note that this is precisely equivalent, at the repair level, to the m etalinguistic statem ent above.) Such repairs are n o t corrections of errors; "m y w atch" identifies the referent unam biguously. Rather, they denote children 's sensitivity to the force of different determ iners, which are no longer independen tly stored bu t are part of a linguistic subsystem . Such subsystem s, I argue, are built u p from the extraction of com m on features after representational redescription. Younger chil­dren do not m ake these self-repairs, bu t this is precisely w hat children of around 6 display in such circum stances. In o ther w ords, although they are unable to provide verbal explanations of their linguistic

The Child as a Linguist 51

know ledge about the relationship betw een "the" and "m y" in refer­ential com m unication, their self-repairs bear w itness to the fact that som ething has changed in their internal representations since the period of correct usage.

I w ould now like to take you th rough a little m ore detail of som e of my psycholinguistic experim ents aim ed at testing aspects of the RR m odel. The data dem onstrate the progression from behavioral m as­tery, to subsequent representational change, and finally to children 's consciously accessible theories about how language functions as a system . We will start w ith children 's use and thoughts about w hat counts as a "w ord", th en look a t how they build theories about the functioning of little w ords like "a" and "the" in sentences, and finally go beyond the sentence to ex tended discourse.

From Behavioral M astery to Metalinguistic Knowledge about Words

H ow do young children segm ent the continuous speech stream into appropriate formal w ord boundaries? There is no simple physical basis in the in p u t to cue children about how to isolate w ords (Tunm er et al. 1983). Of course, if children w ere pure behaviorists, this w ould pose a serious problem , and segm entation errors w ould pervade their out­pu t. Yet, a lthough segm entation errors occur at the very earliest stages of language acquisition (Peters 1983), they are rare once m orphology and functors appear in the child 's ou tpu t. Moreover, w hen segm en­tation errors do occur (e.g., "a nadult" or "u n léléphant"), they do not persist. C hildren do no t learn language by passively soaking up the inpu t w ith all its inheren t problem s. C hildren actively construct rep ­resentations at formal w ord boundaries on the basis of linguistically relevant constraints and of abstractions—not copies—of the linguistic input. Indeed, once young children are beyond the very initial stage of language acquisition and are consistently producing bo th open- class and closed-class w ords in new , nonform ulaic contexts, there can be no question that at som e level these are represented internally as words. That is, w hereas 3-year-olds represent and process formal w ord boundaries as such, they seem to know little if anything explicit about w hat counts as a w ord.

N um erous stud ies18 have show n that it is no t until about age 6, and for some tasks even later, th a t children know explicitly that bo th open- class w ords (e.g. "boy", "chair", "silence", "run", "think") an d closed- class w ords ("the", "any", "to", "in", "w hen", "of") are zvords. W hen asked to count w ords in a sentence, young children frequently neglect to count the closed-class item s. W hen asked directly if "table" is a

52 Chapter 2

w ord, they agree; b u t w h en asked if "the" is a w ord, they answ er in the negative. Yet 3-year-olds can correctly perceive and produce w ords like the.

The RR m odel posits that З-year-olds' representations of formal w ord boundaries are in level-I form at. By age 6 a child 's statable know ledge that the counts as a w ord is, according to the m odel, in the E2/3 form at. But w ha t hap p en s betw een these two ages?

The RR m odel predicts that there m ust exist a level of representation betw een that underly ing the correct segm entation of the speech stream into w ords like the, in w hich formal w ord boundaries are represented as p art of on-line inpu t/ou tpu t procedures, and the level of representation th a t allows for direct off-line m etalinguistic reflection about the fact that the is a w ord. This m iddle level is the E l represen­tational form at. It involves a redescription of inform ation into a format that is accessible to certain tasks outside norm al inpu t/ou tpu t relations bu t no t yet to m etalinguistic explanation.

I set out to test this prediction (Karmiloff-Smith, G rant, Jones, and Cuckle 1991). Previous studies in w hich children were asked w hether X is a w ord, or to count the num ber of w ords in a sentence, d id not engage norm al language processing and required a totally off-line stance. Such tasks therefore dem and a high level of explictness (E2/3 representations). If w e are to capture som ething betw een the totally on-line use of w ord representations and the full metalinguistically accessible know ledge in off-line tasks, w e need to devise a w ay of engaging children 's norm al language processing w hile getting them to access that know ledge for partially off-line reflection. The following technique did just that: C hildren of ages 3 -7 w ere given a series of partially on-line tasks of a sim ilar design. They listened to a story in w hich the narrator paused repeatedly on open-class or closed-class w ords. D epending on the task, the child was asked to repeat " the last w ord" " the last sen tence," or " the last thing" that the storyteller had said each time she stopped . N o explanation was given as to w hat counted as a w ord, a sentence, or a thing. The design of our task did not preclude the types of errors found in previous research, including responding w ith m ore th an one w ord (e.g. "on the floor" instead of "floor" or "knock over" instead of "over"), responding w ith single syllables ("lence" instead of "silence", "kind" instead of "kindness", "thing" instead of "nothing"), or m aking segm entation errors ("isa" instead of "a"; "kover" from "knock over").

This partially on-line technique engages norm al language process­ing and causes an in terrup tion of the construction of a representation of the speech inpu t. N ote, how ever, that the task also has an off-line m etalinguistic com ponen t.19 The child m ust know w hat the term word

The Child as a Linguist 53

m eans and differentiate this from instructions to repeat the last sentence or the last thing. To access and reproduce the last w ord, the child m ust focus on her representation of the acoustic input, m ake a decision as to w hich segm ent of it constitutes the last word, and repeat that segment.

In another experim ent, w e com pared a group of subjects' data from the on-line w ord task w ith their responses to off-line direct questioning about w hether closed-class and open-class items are words.20 For the latter, we sim ply asked children to help a teddy bear find out w hat counts as a w ord and read out one by one a list of w ords, asking "W hat do you think about X? Tell Teddy if X is a w ord."

We hypothesized that the off-line task w ould require level-E2/3 representations, w hereas the partially on-line task w ould require the level-El format. We thus predicted that 3- and 4-year-olds w ould fail both types of task because their representations of w ords are still in procedurally encoded level-I form at, that children around 5 w ould succeed on the partially on-line task but be less successful on the fully off-line m etalinguistic task, and that children of age 6 or 7 would succeed on both tasks, because by then they have m ultiple levels of representation w ith respect to the concept word.

These predictions w ere borne out. A num ber of the youngest sub­jects could do neither task well, suggesting that their representations of formal w ord boundaries w ere still implicit in the level-I format. But our results show that som e children as young as 4У2, and the majority from age 5 on, treat both open-class and closed-class w ords as words, and that they differentiate word and sentence w hen the task has an on­line com ponent engaging norm al language processing. These children were significantly worse, how ever, on the off-line task that involved E2/3 representations. O n that task, although young children accepted exemplars from the open-class category as words, they rejected several exemplars from the closed-class category. Only the older subjects were very successful at bo th tasks.

In general, then, the older child 's level-E2/3 theory is one that has changed from rejecting w ords like "the" b u t accepting w ords like "chair" (because they denote som ething in the extralinguistic context) to considering "chair", "the", etc. as all equivalent in their status as words, by virtue of the fact that they are part of a system w hose elem ents combine in principled ways. The latter—an intralinguistic account—w as found to be available only at around age 6 and beyond.

The developm ental progression highlighted by this study is im por­tant. First, as of age 3, w hen their o u tpu t is m ore or less devoid of segm entation errors, we m ust g rant that children represent formal w ord boundaries for bo th open-class and closed-class w ords. How­

54 Chapter 2

ever, these representations are inaccessible for purposes outside input/ ou tpu t relations. They are, according to the RR m odel, in the level-I format. Second, som ething occurs internally betw een ages 3 and 5 such that by around age 4V2 children can access the represented knowl­edge and succeed on o u r partially on-line task. The RR m odel posits that this is possible because the level-I representations have been redescribed into an accessible E l format. A nd, third, som ething m ust again occur internally beyond age 5 or 6 to explain w hy, by then, children can engage in m ore consciously accessible theory construc­tion about w hat words are and can access such know ledge in off-line tasks. This, I m aintain, requires a further redescription into the E2/3 format.

The RR m odel posits that this developm ental progression can be explained only by invoking, not one representation of linguistic know ledge, to w hich one either has or does not have access, but several re-representations of the sam e knowledge, allowing for in ­creasing accessibility.

From Behavioral M astery to M etalinguistic Knowledge of the Article System

Nom inal determ iners such as articles exist in one form or another in all languages, b u t their obligatory contexts differ m arkedly from one language to the next. English m arks the indefinite/definite contrast ("a"/"the") and uses tw o different surface forms to express the indef­inite article ("a") and the num eral "one". A lthough French also m arks the indefinite/definite contrast by different articles, in that language the num eral and the indefinite are realized by a single form ("un" or the fem inine counterpart "une"). Russian m arks the indefinite ("adna"), b u t the definite has no surface realization; for definite ref­erence the noun is used w ithout a determ iner. Swedish places the indefinite before a n o u n as a separate w ord ("et hus"—a house), but the definite m arker is suffixed to the no u n ("huset"—the house). And so on. C hildren have to be sensitive to nom inal m arking in general and m ust also learn about the particular syntactic realization of the nom inal system in their ow n language.

Recall how infants show sensitivity to distinctions conveyed by articles, well before they are part of their ow n output? O ne reason concerns the phonological and prosodic patterns of language. As we saw earlier in this chapter, 4-day-old infants are already sensitive to the phonological patterns of their native tongue. A nd well before they are producing articles, they can use the presence or absence of articles to decide w hether a n o u n is a p roper nam e (no article; e.g ., "Dax") or a com m on noun (e.g., "a dax"). Gerken (1987) has also dem onstrated

The Child as a Linguist 55

this early sensitivity to syntactic cues.21 She asked very young children w ho w ere no t yet producing articles to imitate short sentences in which either articles or nonce filler syllables of equivalent length and stress were also placed before nouns. If children were sim ply constrained by length, phonological or prosodic cues, their im itations should be equivalent for bo th types of sentence. It tu rned out, however, that they selectively om itted articles, w hereas they im itated the nonce syllables. This suggests that, in their com prehension of the sentences to be im itated, these young children w ere processing articles syntacti­cally—i.e., differently from the fillers, w hich they probably processed phonologically.

It tu rns out, too, that articles appear early in production (Brown 1973; Karmiloff-Smith 1979a; M aratsos 1976; Tanz 1980; W arden 1976), despite the fact that early on they seem to carry far less m eaning than nouns and verbs. So w hat is the status of these early representations of articles? Let us look at this question via an experim ent that specif­ically explored children 's understand ing of the contrast betw een def­inite and indefinite articles.

Im agine a very sim ple experim ental setup that I used som e years ago, in w hich tw o dolls—a boy and a girl—have playroom s w here various objects are displayed. In one situation, the girl doll has three cars, one book, and one ball, and the boy doll has one car, one pencil, and three balls. The crucial difference betw een the tw o dolls' posses­sions is th a t for som e trials the girl doll has several Y's and the boy doll one Y, w hereas for o thers the boy doll has several X's and the girl doll one X. Take the context illustrated in figure 2.1 as an example. Now, if you w ere to hear m e say "Lend m e the car" you could infer that I m ust be talking to the boy because he 's the only one w ith one car. Likewise, w ere I to say "Lend m e a car" you could infer that I am m ore likely to be talking to the girl because she has several cars. This w as the child 's task: to guess w hich addressee I was speaking to in a series of trials w ith varying contexts.22

Now, this is a task at w hich children of ages 3 and 4 succeed easily. M uch developm ental research stops at the po int w hen children 's per­formance is at ceiling. But m y research strategy has always been to explore developm ent beyond behavioral m astery in an attem pt to uncover subsequen t changes in internal representations. As of age 3, children are alm ost 100% successful at the m apping betw een each of the indefinite and definite articles and one of their deictic functions. This is a healthy start as far as behavioral m astery of one of the article's functions is concerned. But w ha t can w e say about w hat children "know " about the definite and indefinite articles? M ore precisely, w hat can w e say about young children 's internal representations of these

56 Chapter 2

Girl doll 's p layroom Boy doll's playroom

Child seated here

Figure 2.1Definite-indefinite discrim ination. (After Karmiloff-Smith 1979a; u sed w ith perm ission of Cam bridge U niversity Press.)

linguistic forms? N othing very substantial beyond a guess. It is not until w e take a truly developm ental perspective—until we know som e­thing about children 's subsequent developm ent—that w e can infer the status of the early representations sustaining such efficient understanding.

It turns out that later, around the age of 5 or 6, French speakers— although they continue to be successful at interpreting the definite article—start to m ake m istakes w ith respect to the indefinite article. They tem porarily in terpret "prête-m oi une voiture" (with no stress on "une") to m ean "lend m e one car" ra ther than the indefinite "lend me a car". They pick the doll w ith a single car instead of the one w ith several.23 This late-occurring failure is an im portant clue to represen­tational change. It points to the fact that the 5-year-old has become sensitive to the dual function of the indefinite article in French, and not just to the distinction betw een the definite and the indefinite article. The phonological form "une" (or its m asculine counterpart, "un") is a hom ophone w hich, as m entioned above, conveys both indefinite reference (English "a") and the num eral function (English "one").

The RR m odel accounts for this developm ental progression as fol­lows. A lthough 3-year-olds reach behavioral m astery for each of these

The Child as a Linguist 57

functions in separate contexts and m ake no errors, they do so because they have tw o independen tly stored procedures for producing the sam e phonological form for indefinite reference and for the num eral. Subsequent representational redescription of each of these procedures into the m ore explicit E l form at m akes it possible to link the com m on phonological form across the two representations of form-function pairs. But, since there is a fairly strong one-form /one-function con­straint during language acquisition (Slobin 1985), 5-year-olds tem po­rarily m ark the tw o m eanings by two different forms in production. They produce "une voiture" for "a car" and "une de voiture" for "one car".24 In com prehension, as pointed out above, they start to make m istakes as to w hich of the tw o functions (num eral or indefinite reference) is in tended .

W ith these new facts in m ind, are we getting any closer to being able to say anything about the child 's internal representations of the indefinite and definite articles? Since the errors and repairs w ith re­spect to the dual function of the indefinite article occur later than behavioral m astery, som ething m ust have happened internally be­tw een ages 3 and 5 to explain this. The RR argum ent is that w hen 3- year-olds can first correctly u n d ers tan d or produce simple functions of the definite and indefinite articles (such as deictic functions, w hich point to the curren t extralinguistic context), they do so by calling on two independen tly stored level-I representations w hich m ap a p h o n ­ological form onto a specific functional context. In o ther w ords, these young children know how to in terp re t the definite article "the" to refer deictically to a context w here a singleton (e.g., one car) is focused upon. A nd they also know how to in terpre t the indefinite article "a" to refer to a context w here the speaker is referring to any one of several examples of a group of similar item s, or to use an identical form in French to refer to the num eral. W hat the 3-year-old does no t "know " is that there is a functional relationship betw een these efficiently func­tioning procedures—that the articles together form part of a linguistic subsystem . In o ther w ords, the RR m odel posits that now here in the very young child 's internal representations is there any explicit ind i­cation of the com m on functional links betw een the articles. If such relationships w ere explicitly represented, then these specific errors and repairs should occur at any time, no t solely after behavioral m as­tery. This suggests that the know ledge em bedded in the efficiently functioning bu t independen tly stored representations of very young children is no t yet encoded in the E l form at, and that links across the procedural representations for the different functions are as yet only implicit in level-I representations.

58 Chapter 2

W hat happens, then , after behavioral m astery? The RR m odel pos­tulates that once behavioral m astery has been achieved (i.e., once part of the system functions efficiently and a pattern of internal stability has been reached), the level-I representations undergo a process of redescription. The original level-I representations rem ain intact and can still be called for certain purposes, bu t the redescribed know ledge em bedded in them is now also available as explicit internal data in the El format. Thus, the French-speaking child 's internal representations now explicitly m ark the relationship betw een identical forms—i.e., the fact that, say, the phonological form "un" paired to the nonspecific reference function is the sam e as the phonological form "un" paired to the num eral function. Thus, children begin to represen t determ iners internally as part of a linguistic subsystem rather than as indepen­dently stored form -function pairs. It is this new ly form ed represen­tational link that explains the sudden occurrence of errors of interpretation of the indefinite article in 5-year-olds—errors that are no t apparen t in the successful perform ance of 3- and 4-year-olds be­cause their independen tly stored representations do not explicitly represen t the link betw een the different functions of the articles. For the 3-year-old (and, I hypothesize, for certain fluent-speaking yet otherw ise severely retarded children), the representational link is po­tential, or implicit, in the fact that, from the observer's external view­point only, each independen tly stored procedural representation contains analogous inform ation. H owever, it is only as of age 5, after behavioral m astery in this part of the linguistic system , that the rela­tionship betw een the representations is explicitly stored.

Storing representations in the E l form at does not m ean that the know ledge is available to conscious access and verbal report. The child still has a w ay to go before he or she can consciously access that linguistic know ledge for verbal reporting. For the articles, this tends to occur around age 7 or 8. To gain a sense of young children 's passage from explicitly represen ted know ledge to consciously accessible and verbally statable m etalinguistic know ledge, let us take a peek at some m ore data. To simplify, let us take the sam e linguistic category of nom inal determ iners ("a", "the", "m y", "som e", etc.).

W hat if a child is asked to actually give a verbal explanation, ra ther than m erely in terp re t and use the constraints on articles? Let us re tu rn to the sim ple experim ental context of the boy doll and the girl doll, and their respective playroom s, illustrated in figure 2.1. The child has correctly guessed w hich doll the experim enter is addressing, d epend­ing on w hether the o u tp u t contained the definite or the indefinite article. H ow do children explain their correct guesses w hen your questions involve accessing know ledge represen ted at level E2/3?

The Child as a Linguist 59

Well, the youngest subjects, a lthough they m ust have used the con­trast betw een the articles to m ake their correct guess, explain this on the basis of real-world know ledge, saying som ething along the lines of "You m ust have been talking to the boy, because boys like cars" (irrespective of the fact that the girl doll has m ore cars than the boy doll). Later in developm ent, children explain their correct guesses by referring to contextual features— for example, "You w ere speaking to the boy, because he 's got one car". It is really ra ther late in develop­m ent, around age 8 or 9, th a t children m ake explicit reference to the linguistic clue that all children m ust have in fact used w hen m aking their correct guess: "You m ust be talking to the boy, because you said 'lend m e the (stressed) car'." A round age 10, children even provide inform ation about the linguistic subsystem from w hich the referential clue was taken, as in the following explanation: "It's got to be the boy, because you said 'th e '; if y o u 'd been talking to the girl, yo u 'd have had to say 'lend m e a car' or, m aybe, 'one of your cars'."

Let m e reiterate that all children m aking successful guesses used the linguistic clues. These m ust therefore be represen ted internally, but only in the I or the E l form at. It takes several years before children can consciously access their representations of such linguistic know l­edge and report on them verbally. By then, I argue, their representa­tions of this linguistic category are also in the E2/3 format.

In the developm ental literature, w hen children cannot report on som e aspect of their cognition it is often im plied that the know ledge is som ehow absent (i.e., no t represen ted a t all). The RR m odel pos­tulates som ething different: that the know ledge is represented inter­nally, b u t still in the I or the E l format, neither of w hich is accessible to verbal report. The end state is such that the sam e inform ation is re­represented at several different levels of explicitness. This allows for different levels to be accessed for different goals: from level I (for rapid inpu t/ou tpu t com putations) to level E2/3 (for explicit m etalinguistic tasks).

I should also m ention that children do n o t reach behavioral m astery for all the functions of the articles by age 3. For m any other functions (including the anaphoric function of the definite article, such as the use of the expression "the m an" after one has introduced "a m an" into the discourse), behavioral m astery is reached considerably later in developm ent.25 A nd those functions also later undergo the sam e three steps a t various ages—behavioral m astery, representational re­description, verbally statable theory about how that part of the lin­guistic system functions.

It is clear that developm ent involves far m ore than the infan t's initial sensitivity to the presence or absence of articles or the toddler's fluent

60 Chapter 2

usage. In order for a child to becom e a potential linguist, the child 's representations have to undergo m ultiple levels of redescription.

Beyond the Word and the Sentence

So is that it? Is the developm ental picture of language acquisition one in w hich the child starts w ith innately specified attention biases and data-abstraction m echanism s, reaches behavioral mastery, proceeds from there through several levels of representational redescription, and finally becomes able to form ulate verbally com m unicable theories about how the system functions? Does all linguistic know ledge take this route? Clearly not.

Am ong the m any developm ents to occur in children 's later language is the passage from the sentential functions of various linguistic m ark­ers to their discourse functions. A n earlier study of children's p roduc­tion of spans of discourse in narrative had show n that initially children m erely juxtapose a sequence of correct sentences, m aking only m ini­mal use of discourse constraints (Karmiloff-Smith 1980, 1985). H ow ­ever, w ith developm ent, children structure their narratives as a single unit ra ther than as a m ere juxtaposition of sentences, and they then adhere to w hat I called the "them atic subject constraint" (see examples below).

Spontaneous repairs tu rn ed ou t to be very informative about chil­dren 's developing capacity for discourse organization and about their adherence to the them atic subject constraint. The following are typical examples of such repairs taken from the data of a task in w hich children generated stories from a sequence of pictures. (It is im portant to note in these examples that the p ronoun repaired to a noun phrase is not am biguous w ith respect to the in tended referent, because the story has only one female protagonist.)

There's a boy and a girl. H e 's trying to fish. A nd to get h er bucket, he hits the girl and she star . . . he hits the girl w ho starts crying.

This boy and girl are o u t playing. H e 's gonna catch som e fish but she . . . bu t the girl w o n 't lend him h er bucket. So he just takes it an d the girl gets real sad.

These and m any other examples suggest that as of age 6 or 7 children operate u n d er the "them atic subject constraint," a discourse constraint which stipulates that pronom inalization in subject position be re­served for the them atic subject of the total discourse (in this case, the boy). By contrast, subsidiary characters tend to be referred to w ith full noun phrases (or p roper nam es, or stressed pronouns), despite the

The Child as a Linguist 61

fact that the different sexes of the protagonists w ould obviate any potential am biguity w ith pronom inal reference. As can be seen above, the p ronoun is repaired even though it is perfectly clear that "she" refers to the girl; the girl is the only female referent in the sequence of pictures. O ther research has show n sim ilar constraints to obtain in the discourse production of bo th adults and children (Reichmann 1978; Tyler 1981, 1983; Tyler and M arslen-W ilson 1978, 1981).

We have seen in this chapter that older children often have level- E2/3 representations w hich allow them to explain m etalinguistically a num ber of aspects of the w ay language functions at the sentential level. Do children (or adults, for that m atter) have metalinguistic know ledge of the discourse constraints on the very sam e m arkers? In o ther w ords, are w e again to w itness behavioral m astery followed by representational redescription and finally by conscious access?

Some recent research indicates that neither children nor adults can provide m etalinguistic explanations of discourse constraints (Karmi­loff-Smith et al., in press). They cannot explain w hy speakers use p ronouns or full n o u n phrases in particular discourse contexts. Even as adults, w e clearly do n o t have access to all aspects of the linguistic system that w e use. Certain aspects of spoken language are inacces­sible to m etalinguistic reflection w hereas o thers, as w e saw earlier in this chapter, are available for spontaneous theory-construction and conscious access. The rules governing discourse constraints do not seem to reach the E2/3 form at, and it rem ains an open question w hether they are redescribed into the E l format. Two related linguistic facts seem to be operative. First, there is a difference betw een the local, sentential function of linguistic m arkers and their m ore global discourse function. Take as an example the pronoun "she". At the local, sentential level, "she" provides inform ation about features such as fem inine, singular, and pro-form —i.e., the referent is female, is alone, and either is in the p resen t deictic space, has just been referred to linguistically, or can be taken for granted through shared know ledge betw een the interlocutors. C hildren and adults have metalinguistic access to these features. According to the RR m odel, they m ust be represented in the E2/3 form at. But in an extended span of discourse involving m ore th an one referent, use of the pronoun "she" also provides inform ation beyond these features. It encodes the role of one referent (e.g., the m ain protagonist) relative to o thers in the overall story structure. In o ther w ords, it reflects the speaker's m ental m odel of the span of discourse as a w hole. W hen in subject position, the pronoun can usually be taken to refer by default to the m ain protag­onist. Reference to a subsidiary protagonist is usually m arked linguis­tically by use of a full n o u n phrase, a p roper nam e, or a stressed

62 Chapter 2

pronoun in subject position. There is a complex interplay betw een nouns, and zero anaphora, m arked differentially as discourse unfolds rapidly in real time. It is this discourse function of the p ronouns and noun phrases to w hich neither child nor adult subjects have m etalin­guistic access. The only w ay the linguist can have access to constraints on the dynam ics of discourse functions is by freezing the fast-fading m essage of on-line spoken text into a static w ritten form that leaves a trace in a different representational form at (Karmiloff-Smith 1985).

The RR m odel focuses on know ledge grow th outside norm al input/ o u tpu t relations. But discourse constraints operate only on line. The discourse function or m eaning of a particular use of a p ronoun to m ark the them atic subject is relevant only w hile that discourse is being uttered. In o ther w ords, discourse constraints are relevant only to rapid on-line com putations of the o u tp u t system. Decisions as to w hether to use a p ro n o u n or a full no u n phrase at this particular po int in this stretch of discourse are not stored in long-term memory. Thus, such on-line com putations are probably rarely if ever redescribed and so cannot be available to m etalinguistic reflection.

From the Nativist Infant to the Constructivist Linguist

This chapter began by exploring dom ain-general versus dom ain-spe­cific perspectives on language acquisition. The bulk of w ork on neo­nates and very young infants suggests that the domain-specific solution is likely to be correct. H um an babies a ttend preferentially to language over o ther auditory input, require only a few days of inpu t to differentiate certain characteristics of their native tongue from other languages, an d are sensitive very early on to m any abstract structure- dependen t features of language. Certain children w ith otherw ise se­vere cognitive retardation acquire language late bu t easily, w hereas, despite rich representational capacities, the m ost intelligent of chim­panzees can at best be taugh t—through incredibly extensive training— strings of m anually encoded lexical item s (Gardner and G ardner 1969) or a sim ple form of language-like logic (Premack 1986). This involves a one-to-one m apping betw een concepts and arbitrary symbols, p rob­ably th rough the use of dom ain-general m echanism s. But this is not language (Premack 1986; Seidenberg 1985). The symbols are no t signs w ithin a structured system . A list of lexical item s, how ever long, does not constitute nam ing and bears little or no relationship to the lin­guistic com petence of even a 2- or 3-year-old.

M uch of the recent infancy w ork seems to m ove in the direction of C hom sky's claim that the abstract structure of language is innately specified in hum ans in som e detail. Clearly, w e m ust invoke some

The Child as a Linguist 63

innately specified attention biases and linguistic predispositions. It may tu rn out that the innate principles sustaining language are more detailed than those sustain ing other dom ains, such as num ber. N one­theless, le t's not foreclose the possibility of an epigenetic process that gradually creates the dom ain specificity of language. W hatever the level of detail of the innate linguistic specification, there m ust be some predispositions for language; that is w hy other species can never learn a structured linguistic system . But the innate specification does not alone explain language acquisition. We saw that the m apping betw een innate predispositions and the in p u t of the child 's native tongue re­quires complex sem antic and syntactic bootstrapping. For norm al de­velopm ent, this is still only part of the picture. To understand how our linguistic representations become flexible and m anipulable (i.e., open to m etalinguistic reflection), w e need to invoke several levels of representational redescription beyond the sem antic and syntactic bootstrapping that leads to behavioral m astery. This, in m y view, also differentiates hum an capacities from those of o ther species. Thus, even if the chim panzee w ere to have an innately specified linguistic base, I speculate that it w ould still not go as far as the norm al hum an child. It w ould never w onder w hy "typew riter" isn 't used to refer to people. It w ould sim ply repeat the linguistic labels that it was given. But children do not sim ply reach efficient usage; they subsequently develop explicit representations w hich allow them to reflect on the com ponent parts of w ords and to progressively build linguistic theo­ries. A lthough th is holds for som e aspects of language, it does not hold across the board. There are facets of syntax and discourse cohe­sion that are never available to m etalinguistic report, even in adults.