Page 11

Chapter 2
Theories of Language Acquisition
In chapter 1 we argued that Gold's (1967) machine language
learner was handicapped because it possessed no biased learning
mechanisms and it learned language in isolation without the
benefits of contextual cues. Although the learner needs both of
these factors, different theories of acquisition have been developed
that rely heavily either on the internal biases of the learner
(Chomsky 1965; Wexler and Culicover 1980; Light-foot 1989;
Bickerton 1984) or on how the learner might acquire grammar
through attention to contextual cues in the input language (Snow
1986; Bates and MacWhinney 1987). Although these theories are
perceived as quite distinct in the literature, there are common
threads that link them together. Upon analysis, these common
assumptions provide clues to the kinds of sensitivities that learners
might share when they approach the language-learning task. In this
way, theory can be used to guide the understanding of process.
This chapter is divided into four sections. In an attempt to make
language acquisition theory accessible to individuals who work in
other areas, we devote section 2.1 to introducing the uniqueness of
language and the phenomenon that current theories of acquisition
must explain, and we pose three questions that motivate this
review. Section 2.2 is further divided into two parts, corresponding
to two hypothetical families of theories of language acquisition.
Explication of each type of theory is followed by an evaluation. In
section 2.3 we collapse the dichotomies that appear to divide the
two families of theories, and in section 2.4 we revisit the three
questions. Our aim is to reach a compromise between what have
traditionally been called the nativistic and interactionist theories of
language acquisition. Although we lean slightly toward the
nativistic side of the debate (for reasons to be outlined here and in
the rest of the book), we believe that language acquisition is best
described through the dynamic
 Page 12
interplay of a number of forces (the coalition model; see chapter 7).
Thus, in this chapter we seek to outline the assumptions that
theories of language acquisition share and the way in which the
studies to be presented in this book begin to address these
assumptions.
2.1
Language Acquisition and Theory in Linguistics
Some researchers have urged their colleagues to "hang loose!"
(Miller 1981) with respect to commitments to any particular theory
of grammar. Yet the first step in developing a theory of language
acquisition must be to specify what it is that the child must acquire.
Without a vision of the adult state, the questions that researchers
ask might lack precision and relevance. Thus, the researcher who
wishes to conduct a theory-based research program must be guided
by descriptions of the adult state, and such descriptions are theory
laden. The safest course, as Pinker (1984) has cogently argued, is
to study aspects of language that all theories of grammarregardless
of what else they doseem to agree on. For example, most theories
seem to have the "same small set of grammatical categories and
relations, phrase structure rules, inflectional paradigms, lexical
entries, lexical rules, [and] grammatical features" (p. 25). Studying
topics such as these fairly ensures that the researcher will avoid
"parochialism" or "impending obsolescence'' (p. 25) since,
whatever theory is currently in vogue, they are topics it will have to
ideal with.
In this book we have followed Pinker's (1984) advice with respect
to commitments to theories of language. Indeed, we have carried
his advice a step further and have also not committed ourselves to
any particular theory of language acquisition. Instead of making
such a commitment, we focus on questions that all theories must
address. Thus, this review is organized around three questions:
1. What is present when grammatical learning begins?
2. What mechanisms are used in the course of acquisition?
3. What types of input drive the language-learning system forward?
The first question is concerned with what the child brings to the
task of language learning. This includes cognitive, social, and
linguistic knowledge that may either be innate or originate in prior
learning. On the surface, theories identified as "nativistic" seem to
presuppose far more at the start than theories that emphasize how
the child makes use of the linguistic and nonlinguistic environment.
 Page 13
The second question focuses on the actual processes that children
use to acquire language. For example, do children use domain-
general learning procedures that are common to, say, perception
and problem solving, or do they use domain-specific learning
procedures that are peculiar to language processing?
The third question asks what types of inputs drive the language-
learning system forward. Do children rely primarily on increasing
knowledge of social conventions as in the pragmatic account, or do
they perform increasingly sophisticated structural analyses of the
linguistic stream? Does the environment play a relatively large or a
relatively small role in assisting children to uncover the rules of
their native tongue?
Whether explicitly or implicitly, all theories of language
acquisition presuppose their own answers to these questions.
Before we look at the various answers, however, the problem of
language acquisition, or what must be learned in the first place,
needs to be described in more detail. In what follows we outline a
general view of the nature of language and what is necessary to
acquire a language. In so doing, we provide readers whose own
research may be outside the area of language acquisition with a
brief introduction to these issues (for a fuller treatment see Fletcher
and Macwhinney 1995 and Ingram 1989, for example).
Among the insights about language that are widely accepted by
both linguists and acquisition theorists are that all languages
comprise (1) units at a number of levels that encode meaning (e.g.,
words, phrases, and clauses) and (2) relations or ways of arranging
those units to express events and relationships. Languages are also
hierarchically organized. Units at one level (e.g., phrases) are
embedded in units at a higher level (e.g., clauses). Thus, any theory
of acquisition must account for how children become sensitive to
the types of units that exist and to the relations or arrangements of
these units in their native language. Only with these sensitivities
could children work toward achieving adult competency.
All theories of grammar and of acquisition presuppose that words
are the basic building blocks of language, despite the fact that
words are composed differently in different languages. For
example, in languages like Chinese and English, words carry
relatively few grammatical markers and may consist of a single
morpheme. In a sentence such as "You saw me," each word serves
a separate grammatical function. To say the same thing in a
language such as Imbabura Quechua, a speaker uses a single word
containing much grammatical information:
 Page 14
riku- wa- rka- nki
see first person object past second person subject
Thus, a single word in Imbabura Quechua can be composed of a
rich combination of affixes and stems that carry much of the
linguistic burden that is expressed by separate words in languages
like English. From a psychological perspective, one might
speculate that languages that contain mostly separate wordlike
units are easier for children to acquire than the so-called
agglutinative languages that "paste" together different grammatical
elements into a single word. The actual data on children acquiring
these languages, however, suggest otherwise. Regardless of the
properties of the language, the milestones of acquisition seem to
occur at about the same ages around the world (Lenneberg 1967;
Slobin 1985a, b).
Despite differences in the way languages use words to transmit
grammatical content, the words in all languages fall on a
continuum that ranges from what are referred to as "open-class"
words to what are referred to as "closed-class" words.
Prototypically, open-class words bear content (e.g., "chair,''
"psychology," "think") and closed-class or "function" words (e.g.,
pronouns, articles, and prepositions) contain no content, in the
traditional sense, but rather serve as operators to signal different
meaning relations and grammatical units. Closed-class items are
essential for specifying the architecture of the phrases within
sentences. In general, across languages, children acquire open-class
words, which label the objects and actions in their environment,
before they acquire the closed-class words, which carry more
abstract meanings (Brown 1973; but see Nelson 1973 for some
strategic differences in children's early vocabulary).
Across all languages, the open-class words may be further
subdivided into a finite set of "form classes" such as "nouns,"
"verbs," "adjectives," and "adverbs." Not all languages contain all
these categories, but all choose from this limited set (see Greenberg
1963; Comrie 1981). In all languages where acquisition has been
studied, the preponderance of children's first words appear to come
from the class of nouns (Gentner 1983; but see Bloom, Tinker, and
Margulis, in press; Nelson, Hampson, and Shaw 1993), although
one need not claim that children are aware of these adult
designations for grammatical categories. Indeed, nouns even appear
first in languages such as Korean that are characterized by much
noun ellipsis and that have verbs in sentence-final position (Au,
Dapretto, and Song 1994). These form class units in turn provide
the core for larger
 Page 15
constituents such as noun phrases (e.g., "the big red book") and
verb phrases (e.g., "jumped happily"). All grammatical theories
attempt to account for the composition of phrases and clauses.
In sum, to learn a language, a child must first find the units that
compose that language. Although languages seem to draw
universally from the same finite set of units, the child faces a
somewhat different task in uncovering what these units are
depending on the language to be learned. Nonetheless, all children
in the second year of life show evidence of having found the units
used for grammar building by the way in which they begin with the
smallest unit (words) and gradually create phrase- and then clause-
sized units.
With regard to the relations between words, languages also share a
number of common features. Of these, the most important is that
languages are "structure dependent" (Chomsky 1972) rather than
"serial-order dependent." This means that " . . . knowledge of
language depends on the structural relationships in the sentence
rather than on the sequence of items'' (Cook 1989, 2). This fact has
a number of implications for language acquisition. For example,
even though the subject and the verb may be separated by a
variable number of words, children must be able to find these
elements in order to create agreement between them (e.g., "John,
the boy next door, works for . . ."). Thus, children must be able to
find the units in their particular language and detect which units are
structurally dependent on what other units in that language. This in
turn involves recognizing the hierarchical structure of the language
and learning what grammatical devices (e.g., word order,
inflectional markings, or tones) the language uses to create various
types of syntactic units.
Structure dependency is best illustrated by the simple example of
questions in English. For instance, the structure of "Will John
come?" implies that questions are formed by inverting the second
word in the parallel declarative sentence, in this case "John will
come." However, such a rule quickly runs into trouble since it
would generate ungrammatical questions like "Sister John's will
come?" This simple example demonstrates that inverting the word
in any particular position in a sentence could never work; what
counts is the type of word involved.
Syntactically speaking, "will" is an auxiliary. Perhaps the rule of
question formation involves movement of the auxiliary. But this
generalization is not specific enough. Consider a sentence with a
relative clause, such as "The man who will come is John." "Will"
and "is" are both auxiliaries. Which one should be moved? Moving
the first auxiliary, "will," yields the
 Page 16
ungrammatical question "Will the man who come is John?"
Moving the second auxiliary, "is," yields the acceptable question
"Is the man who will come John?" Jumping to the conclusion that
moving the second auxiliary is the correct rule also runs up against
any number of counterexamples. The answer is that in English, the
auxiliary from the main clauseand not from the relative clauseis the
one that is moved. Thus, any movement in a sentence must take
syntactic categories (such as main and relative clause, subject of
the sentence, object of the sentence) into account and not the order
of the words. Rules of question formation in English (or any other
language) must be described with respect to the grammatical
structure of the sentence, not the serial order of the words.
How children become aware of the syntactic categories over which
the rules are written in language is a continuing topic of debate in
theories of language learning (see Pinker 1984). Although children
generally don't learn how to identify these categories consciously
until perhaps a 7th-grade grammar class, their linguistic behavior
(such as question formation, subject-verb agreement, and numerous
other phenomena) illustrates that syntactic categories must be
unconsciously present at least by the time they are using four- and
five-word sentences. The question for acquisition researchers is
when and how such grammatical concepts enter the child's
linguistic armamentarium. At least by age 3, children do not appear
to violate rules for structure dependency (Crain and Nakayama
1987).
In sum, the question of how learners acquire the units and relations
of their native language is the question of how they learn language.
Acquisition involves becoming sensitive to linguistic structures,
not just to the serial order of the words. Theories of acquisition
must therefore account for how learners acquire linguistic units and
linguistic rules.
2.2
Two Varieties of Language Acquisition Theories:
Sketches of a Field
Despite the fact that theories of language acquisition are relatively
immature, the current theories provide a starting point from which
to consider how language learning gets off the ground. For the
purposes of this discussion, theories of language acquisition will be
classified into two families defined broadly by their commitment to
what the child brings to the task, the processes used to acquire
language, and the input considered central for acquisition. As in
any exercise of this sort, "shoehorning" a theory into a family may
do some violence to the details of that theory.
 Page 17
Table 2.1 Distinction among the major theories
THEORY TYPE
Inside-out Outside-in
Initial structure Linguistic Cognitive or social
Mechanism Domain-specific Domain-general
Source of structure Innate Learning procedures

What will be gained in exchange for some imprecision, however,

will be an efficient way to present some of the major theoretical
cuts in the field. By way of preview, the distinctions between these
families of theories are presented schematically in table 2.1.
Outside-in theories. One family of theories, which we characterize
as the "Outside-in" group, contends that language structure exists
outside the child, in the environment. Children attend to the salient
objects, events, and actions around them and construct language
from rather meager linguistic beginnings. Indeed, early grammar is
sometimes posited to be a transparent mapping of salient perceptual
events onto linguistic form. The hypotheses the child entertains
about what might constitute language-relevant data are derived and
constrained either by the social environment or by the child's
inherent cognitive capabilities, not by any specific linguistic
knowledge. In this scenario, therefore, the language-learning
problem is not privileged, but rather is one that is solved by
domain-general learning proceduresby the same procedures that
allow the child to analyze the environment into ongoing events
composed of actions and objects. Thus, as table 2.1 indicates, all
Outside-in theories must of necessity focus on the processes by
which language is acquired since they do not grant that the child
has been endowed with any language structure a priori. These
processes must, in turn, be very powerful given that the Outside-in
theories do not constrain them or the grammar ultimately acquired
in any way. Thus, in general, Outside-in theories identify language
learning as a totally bottom-up process, no different from learning
that takes place in other domains. There are two identifiable
subtypes of Outside-in theoriessocial-interactional and
cognitivediffering mainly on which type of input is presumed to
contribute most to language learning.
The social-interactional theories hold that the social interactions in
which the child participates provide the route into language
acquisition by
 Page 18
highlighting just those aspects of events that will be translated into
linguistic forms. The cognitive theories, emphasize the role played
by the child's prior understanding of events and relations in the
nonlinguistic world as well as the child's cognitive processing
capabilities. Children use language to label the cognitive categories
(e.g., agent, action) they have constructed. They then use
distributional evidence or pattern detection of a general sort to
cross-classify cognitive categories into linguistic ones like "noun
phrase" and "subject of the sentence."
Inside-out theories. The second family of theories, which we
characterize as the "Inside-out" theories, contends that language
acquisition occupies its own separate niche or module in the brain
and has its own unique mechanisms (Fodor 1975; Chomsky 1981).
Language acquisition is the process of finding in the linguistic
environment instantiations of the considerable innate linguistic
knowledge the child already has. In the sense that the child works
to find counterparts for a priori categories, the process of language
acquisition is Inside-out in character. It is a process of discovery
instead of construction.
There are also two subtypes of Inside-out theories. One subtype,
the structure-oriented theories, places primary emphasis on the
content of the grammar to be acquired. For example, such theories
are concerned with how children select between grammatical
options to "set" the "parameters" that correspond to their native
language (Hyams 1986; Lightfoot 1989) and how they give
evidence of the operation of universal linguistic principles such as
the "binding" principles that apply to the interpretation of nouns,
pronouns, and anaphors (e.g., Wexler and Chien 1985). In general,
these theories presuppose rich linguistic structure that primes the
child's entry into the linguistic system.
The second subtype of Inside-out theories, the process-oriented
theories, also grants that the child must be innately endowed with
domain-specific linguistic knowledge, such as a version of
Chomsky's (1981) Principles-and-Parameters Theory or Bresnan's
(1978) Lexical-Functional Grammar. The difference between these
theories and the structure-oriented ones lies in their emphasis on
uncovering the mechanisms the child uses to acquire language
(Landau and Gleitman 1985; Gleitman 1990; Pinker 1984, 1989).
Given the emphasis on mechanisms, some of these theories
concern themselves with the acquisition sequence from its earliest
point.
As is evident even from this cursory review, the Outside-in and
Inside-out theories appear to offer very different assumptions about
what the child brings to the task, what mechanisms guide
acquisition, and what
 Page 19
inputs the child uses. However, we will show that under the two
theories learners actually approach the task with more in common
than a superficial analysis suggests. Indeed, the two theories and
their subtypes all seem to posit a learner who comes to the task
with fairly sophisticated linguistic knowledge. In section 2.4 what
are often considered dichotomies between the theories will be
shown to be hyperbolic statements that actually obscure the
theories' commonalities of position on the three questions listed
above. Theories of language acquisition, we will argue, differ more
in degree than in kind.
Outside-in Theories
Most of the Outside-in theories, be they social or cognitive in
character, were developed as a direct response to Chomsky's (1965)
nativistic and domain-specific view of language acquisition.
Outside-in theories have been attractive to researchers for three
reasons. First, they emphasize development, change, and culture
more than the older nativistic theories. Second, because the social
or cognitive categories that form the substrate for acquisition
appear prior to language itself, researchers can investigate
emerging language-relevant phenomena prior to the time when
children begin to use multiword speech. Third, the Outside-in view
(it is claimed) allows language learning to be accounted for by the
same sorts of domain-general mechanisms that account for other
types of learning. That is, according to outside-in theories, a child
works to learn language structure; it does not come for free.
Social-Interactional Theories
The intellectual roots of researchers in the social-interactional
group trace back to the speech-act theorists (Austin 1962; Searle
1969), who emphasize the functional uses of language. The social-
interactionists believe (to varying degrees) that children construct
their native language through social commerce. Nelson (1985) puts
this position succinctly:
Language learning takes place within the framework of social
interaction, and the nature of the particular kinds of interaction
experienced determines not only the function and content of the
language to be acquired but which segments will be learned first and
how these segments will subsequently be put together or broken down
for reassembly. (p. 109)
The sense of this quotation, embedded in a discussion of individual
differences in language development, is that depending on the kind
of cultural, social, and linguistic environments children are exposed
to, they will learn
 Page 20
not only different "function and content" but different-sized
linguistic units and different grammatical rules for their
recombination. For example, depending on the kinds of speech
children hear directed to them, they may first learn unanalyzed
"gestalts" (e.g., social expressions like "What's that?" uttered as a
single unit) instead of learning single words that are then freely
recombined (see also Peters 1985). The source of linguistic
knowledge here is Outside-in: there is little discussion of what
linguistic knowledge the child brings to the task other than the
ability to participate in and profit from these social encounters.
An early proponent of a strong version of the social-interactional
position was Bruner (1975), who argued that the structure of social
interaction could be mapped transparently onto linguistic structure.
On this view, when infants engage in give-and-take games with
their caregivers in which they alternately play the role of "agent of
action" and "recipient of action," they are essentially learning
language. As Bruner later wrote in representing that position, " . . .
something intrinsic in the nature and organization of concepts . . .
predisposes language to be the way it isincluding the grammar"
(1983a, 27). Bruner (1983a) subsequently recanted this position
and concluded that " . . . systems of language . . . are autonomous
problem spaces, that however much their conquest may be aided by
non-linguistic knowledge or external support from others, they
must be mastered on their own" (p. 28). Despite his later retraction,
a portion of the study of language acquisition was redirected by
Bruner (among others) to the study of the ways in which the social
environment supportsand in stronger versions, accounts forsuch
acquisition.
Under these formulations, repeated interactions with the
environment provide the material for the child's construction of
increasingly detailed scripts that serve as the substrate for initial
language learning (Nelson 1985). Children can memorize sentences
and phrases embedded in "formats" (Bruner 1983a, b) or "routines"
(Snow 1986; Nelson 1985) and can later attend to the relationships
between the stored sentences, deriving from them a set of sentence
"rules." As children advance, the social environment might provide
recasts and corrections of their grammatically incorrect utterances,
facilitating and refining their early productions (Hirsh-Pasek,
Treiman, and Schneiderman 1984; Hirsh-Pasek et al. 1986; Nelson
1977; Sokolov and Snow 1994; Moerk 1983; Bohannon and
Stanowitz 1988; Furrow, Nelson, and Benedict 1979). Under this
view, social interactions provide a cultural classroom for language
learning. Children are presumed to bring little specifically
linguistic knowledge to
 Page 21
the task of language learning. Rather, they must have some way of
interpreting the social interactions around them, of setting up and
interpreting joint attentional episodes, of understanding some
global relations between the input sentences and represented
meaning, and of detecting the corrections that caregivers provide in
what are often subtle and inconsistent responses to their sentences
(see Bohannon and Stanowitz 1988, for a discussion of how this
might work). Thus, the child begins language learning as a social
partner. Social processes like joint attention, script construction,
"fine-tuning" of caregivers' utterances to the level of the child's,
and correction from the environment form the foundation for the
abstraction of linguistic units and the continuing revision of
linguistic rules. The motivation that drives the system forward is
often an instrumental one: the child is eager to fulfill various social
functions, many of which are best carried out through linguistic
means. It is the child's interpretation of the social environment that
propels the language-learning system onward.
The work of some social-interactional theorists does not appear to
answer the question posed here, namely, how children break into
the grammar of their native language. Rather, it seems to center
more on issues surrounding the uses to which language is put. As
Bruner (1983b) writes:
Whatever else language is, it is a systematic way of communicating to
others, of affecting their and our own behavior, of sharing attention. . . .
Let us not be dazzled by the grammarian's questions. Pragmatic ones
are just as dazzling and mysterious. How indeed do we ever learn to get
things done with words? (p. 120)
On the other hand, some social-interactional theorists' work is
motivated by a desire to understand how social interaction supports
language acquisition per se. As Snow (1989) writes:
No one denies that social interaction is a prerequisite to normal
development, in language as in other domains. The question arises
about special features of social interactionwhether there are aspects of
the ways in which adults or older children interact with infants and
young children that are crucial or helpful to language development, and
if so, what those features are and how they help. (p. 2)
In short, some social-interactional theorists, like Snow, continue to
focus on what it is about social interaction that might facilitate the
learning of grammar. Others appear to be concerned with the
pragmatic aspects of language and not the syntactic (e.g., Berko
Gleason 1993).
 Page 22
Cognitive Theories
Because children appear to interpret their environment in terms of
cognitive categories made up of agents, patients, actions, locations,
and so on, cognitive theorists such as Schlesinger (1971, 1988) and
Braine (1976) argue that these cognitive categories can provide a
solid foundation for language learning. Unlike those who have
argued from a more social perspective, these theorists (who do not
negate the role of social input) have also developed models of
language learning.
For example, in describing his semantic assimilation model for
language acquisition, Schlesinger (1988) argues that children begin
the task of language learning with early relational categories that
are
semantic case-like categories (such as agent, patient, location) rather
than syntactic ones. [These semantic categories] . . . reflect the way the
child interprets the world around him. . . . The child interprets the
environment in terms of semantic relations and learns how these are
expressed in his native language by means of word order, inflection and
function words. (p. 122)
Evidence from studies of early event perception and memory adds
credence to Schlesinger's claims. Children may indeed come to the
language-learning task equipped with language-relevant categories
like agency (Golinkoff 1981), animacy (Golinkoff et al. 1984),
causality (Cohen and Oakes 1993), and possibly path or location
(Mandler 1988, 1992).
The developers of the most detailed of these models, Bates and
MacWhinney (1987, 1989), endorse a domain-general view of
language learning in which minimal language structure is given
from the start. Building on Maratsos and Chalkley's (1980) analysis
of how children discover form classes through unconstrained
distributional analysis, MacWhinney and Bates (1989) write:
[T]he universal properties of grammar are only indirectly innate, being
based on interactions among innate categories and processes that are
not specific to language. In other words, we believe in the innateness of
language, but we are skeptical about the degree of domain-specificity
that is required to account for the structure and acquisition of natural
languages. (p. 10)
According to Bates and MacWhinney's "Competition Model,"
based on connectionist-type learning mechanisms, the child looks
for form-function mappings through the use of such constructs as
"cue validity" and "cue strength." "Cue validity" describes the
extent to which a particular cue for how a language works is
available (i.e., present in the surface structure) and reliable (i.e.,
leads to the same outcome when it is available). Cue validity can
therefore be assessed by looking at what grammati-
 Page 23
cal devices a language uses to mark certain meanings. "Cue
strength" is how much weight the learner gives to units of linguistic
information. A particular cue will be weighted more heavily if it
has high cue validity. Thus, for English, preverbal position tends to
be a highly reliable and often available cue for agency. It will
correspondingly be assigned greater cue strength than it would in a
language like Italian, where word order is less rigidly constrained
and semantic roles are marked in other ways.
The Competition Model has undergone extensive additions and
revisions, although the assumptions underlying it have not
changed:
[L]anguage acquisition becomes a problem of pattern detection that
may or may not require the application of innate linguistic knowledge.
We suspect that more general principles of pattern detection and
distributional learning are sufficient for the task . . . (Bates and
MacWhinney 1989, 26)
Thus, children note distributional evidence in the input and piece
together the grammar of their language. The claim is that far less
linguistic knowledge needs to be in place at the start of the
language-learning process than would be supposed by the nativistic
or Inside-out theories to be discussed below. General-purpose
cognitive mechanisms such as induction and hypothesis testing can
be employed in the service of language acquisition, and these
processes are sufficient to guarantee successful grammatical
learning. Theories like Bates and MacWhinney's are being
extended in connectionist treatments in which computer
simulations are used to model a distributional theory of language
acquisition (Plunkett 1995).
Summary of the Outside-in Position
To the three questions posed above, then, Outside-in theorists
would offer the following answers:
1. What is present when grammatical learning begins? The child
brings to the language-learning task a rich social and/or cognitive
system that includes cognitive categories such as agent and
recipient. The child also brings the ability to attend to and organize
linguistic data, although this ability does not emanate from any
language-specific module apart from general cognition.
2. What mechanisms are used in the course of acquisition?
Although some theories do not explicitly mention mechanisms, the
thrust of the Outside-in theories is that children are equipped with
domain-general learning procedures, not particularly constrained in
any way, which enable them to analyze events into cognitive
categories and to detect the covarying language forms. For
grammatical learning the chief mechanism
 Page 24
appears to be distributional analysis, since children form ever
wider, more abstract concepts that ultimately overlap with adult
linguistic concepts.
3. What types of input drive the language-learning system forward?
In the Outside-in theories, burgeoning social knowledge and/or
cognitive skills provide the impetus for continued language
learning. At the earliest stages, language is simply a transparent
overlay onto social or cognitive categories.
Critique of the Outside-in Position
Four criticisms have been leveled at the Outside-in approaches to
language acquisition (see Golinkoff and Gordon 1983; Snow and
Gilbreath 1983; Pinker 1984); some of these have been prefigured
in the above review. The account offered by these theories appears
to have the advantage of parsimony in that domain-general
abilitiespresumably the same abilities that service, say, perception
or number knowledgeare seen as supporting language acquisition
(Piatelli-Palmarini 1980). Yet these general accounts, even in their
more advanced instantiations, fall far short of explaining some of
the basic facts of language acquisition.
The first problem is that although the Outside-in theories profess to
make minimal nativistic assumptions about what the child brings to
the language-learning task, they presuppose much in the way of
nonlinguistic and linguistic knowledge at the start of language
acquisition without often explicating the source or process by
which that knowledge was gained. For example, Bates and
MacWhinneys' Competition Model requires the "early availability"
of the following concepts (from MacWhinney 1987, 259):
Major item type: nominal, verbal, operator
Nominal status: common, proper, pronoun, and dummy
Number: singular, plural, and dual
Individuation: mass, count, and collection
Tense: present, past, and future
This list represents only part of the linguistic material that the
model explicitly presumes to be present at the start of language
learning. Likewise, Schlesinger (1988) proposes that
the child indeed does have access to relational categories underlying
sentences, since these are at first not abstract, formally defined
categories . . . but categories that reflect the relationship in situations
referred to by the sentence he hears: the agent of an action, the location
of an entity, and so on. If anything, it is these cognitively based
categories that must be assumed to be innate. (p. 123)
 Page 25
Thus, the Outside-in approach presumes a host of possibly innate
and abstract givens that include ways of (1) interpreting the
environment in terms of categories like agents and actions; (2)
segmenting the speech stream into linguistically relevant units; (3)
attending to the order and frequency of these units; (4) memorizing
the units for purposes of comparing and computing their frequency
in order to resolve competition (Bates and MacWhinney 1987); and
(5) creating categories of ever-expanding power such that they
approximate grammatical categories like ''subject of the sentence"
(Bates and MacWhinney 1987; Braine 1976; Bowerman 1973;
Schlesinger 1988).
To sum up, the first problem faced by Outside-in theories is that
they contain many hidden assumptions about the linguistic
knowledge that the child brings to the task. Indeed, these linguistic
assumptions are necessary for their arguments to go through. As
Kelly and Martin (1994) state with respect to theories that rely on
distributional analyses:
[R]eference to the statistical structure of the environment will not be
sufficient to characterize learning since the nature of the environment is
partly defined by how the learner categorizes items within it.
Furthermore, these categories may in fact be intrinsically linguistic in
nature. Hence we are back to a basic problem in language learning:
What notions about the nature of language do children bring to the task
at hand? Reference to statistical regularities has not eliminated this
problem at all. (pp. 115-116).
The second problem faced by some Outside-in theories is that they
treat linguistic knowledge as though it is reducible to knowledge in
other domains. That is, Outside-in theorists have tended to treat
language as though it is a transparent mapping between cognitive
or social categories and linguistic forms. But this view faces
several obstacles and questions. If there is any such transparency,
children quickly go beyond it when they master the complexity of
their language. Furthermore, if the transparency assumption were
true, why wouldn't all languages be the same? Why wouldn't all
languages rely on the same order of elements if it was more
"natural" to use the sequence doer action result? Humans
presumably construct these categories as a result of universal
cognitive processes.
The fatal flaw in this assumption is that cognitive and social
categories do not map transparently onto linguistic forms; nor is
linguistic knowledge the same as social or cognitive knowledge.
One argument to this effect is made by Roeper (1987), who points
out that the cognitive category of "agent" is not the same as the
linguistic category of "agent." For example, one can say "the
robber of the bank," but not "the thief of the
 Page 26
bank." Robbers and thieves are both agents, but the words "robber"
and "thief" cannot fill the same grammatical slots in the phrases
given. Thus, knowing the cognitive agent would not be sufficient
for expressing this idea linguistically.
The third problem faced by Outside-in theories is the need to
explain how the child starts with one kind of linguistic system
based on cognitive and social categories and transforms it into the
adult linguistic system based on abstract syntactic categories. As
Bloom, Miller, and Hood (1975) first pointed out:
Braine's (1974) formulation, and pragmatic explanations in general,
establish a serious discontinuity with later development. Such claims
fail to contribute to either a) how the child eventually learns
grammatical structure, or b) the systematic (semantic-syntactic)
regularities that are manifest among the earliest multiword utterances.
(p. 33)
Gleitman (1981) humorously refers to this problem of discontinuity
as the tadpole-to-frog problem. What we have called Outside-in
theories begin with cognitive categoriesthe tadpoleand must
transform them into more abstract linguistic categoriesthe frog. For
example, if children begin with the category of agent and not the
category of sentence subject, they must somehow transform the
former into the latter. Since sentence subjects can be agents,
patients of action, locations, and so on, there is no one-to-one
mapping between cognitive and linguistic categories. Yet the
Outside-in theories cannot adequately explain how this
transformation occurs. They are unable to offer workable solutions
to this problem of discontinuity for two related reasons: some of
the Outside-in theories conflate language and communication and
therefore do not see it as a problem; others seem to rely on
unarticulated assumptions about the grammar that the child must
ultimately acquire (Golinkoff and Gordon 1983).
One response to this problem of discontinuity is to note that there
may be no discontinuity problem at all. Perhaps syntax (and not
just semantic categories) is there from the beginning. There is
reason to believe that children can use syntactic information; that
they can even use abstract syntactic categories from a very early
age; and that if there is some shift from a semantically or
communicatively based system to a syntactic one, it occurs quite
early in the language acquisition process. Evidence from children
as young as 17 and 18 months of age, for example, suggests that
children are keenly aware of syntactic structure and that syntax is
not reducible to semantics (see Levy 1988; Pinker 1984; Bloom
1990; Choi and
 Page 27
Bowerman 1991). Further, in an exquisite demonstration of the
impact of syntax, Choi and Bowerman (1991) show that young
children (17 months of age) use syntax in interaction with semantic
information to carve up the domain of spatial relations. In English,
for example, the notion PATH (into or out of) is marked in the
same way whether it results from a spontaneous or a caused motion
(e.g., "John rolled the ball OUT OF the box" vs. "The ball rolled
OUT OF the box"). In Korean different words are used to represent
PATH depending on whether the movement is caused or
spontaneous. Thus, the same semantic space is carved up
differently by different languages, a fact that seems to be
internalized by children who are but a year-and-a-half old. Syntax
here is influencing semantic organization; form is influencing
function. (This result is perhaps most striking given that Bowerman
(1973) was one of the early proponents of an Outside-in,
semantically based approach to language acquisition.) In short,
there may not be a tadpole-to-frog shift on this view.
The fourth problem is perhaps the most toublesome for the
Outside-in theories. It falls under what has been referred to as the
"poverty-of-the-stimulus" argument. In Chomsky's (1988) words,
"As in the case of language, the environment is far too
impoverished and indeterminate to provide this system to the child,
in its full richness and applicability" (p. 153). Three examples will
illustrate what the environment fails to contain. The first example
of how the linguistic stimulus (sentences that the child hears, in
isolation from their context) cannot provide sufficient data for
distributional analysis is what are often called "empty categories"
(Chomsky 1981). These are categories that are presumed to exist at
some abstract, underlying level but that are not realized in the
surface structure of the sentence. Psycholinguistic research with
adults (e.g., Bever and McElree 1988; see Fodor 1989, for a
review) has demonstrated that online sentence processing reflects
the existence of these categories. In imperative sentences (e.g.,
"Get dressed!''), "you" is implied but is not present in the surface
form. Children (who hear many imperative sentences) do not
produce sentences like "Get dressed himself!" although they do
produce sentences like "Get dressed yourself!" Since they cannot
be performing a distributional analysis on a "you" that is not
present in the speech they hear, children must, in some way, be
aware of its existence at a deeper level. No amount of pattern
detection will supply learners with the clair-voyance needed to
retrieve the nonexistent "you"; some other mechanism must
account for their sensitivity to this category.
 Page 28
Structure dependency provides a second example of how the
stimulus presents insufficient data for pattern analysis. The
sentence cited earlier, "The man who will come is John," makes
this point in that its surface form gives little indication of how it is
hierarchically organized (but see Lederer and Kelly 1991). Yet this
sentence is composed of a main clause ("The man is John") and a
subordinate noun phrase ("who will come''), and only the "is" from
the main clause may be fronted to form a question. Thus, only the
question "Is the man who will come John?" is grammatical; the
question "Will the man who come is John?" is not. No amount of
"ground up" surface structure analysis will allow learners to
stumble on the fact that not all noun phrases are treated equally; the
grammatical role the noun phrase plays in the sentence determines
what alterations it may undergo. Without this structural knowledge,
however, children will make many incorrect generalizations about
the grammar of their language.
A third piece of evidence in favor of the poverty-of-the-stimulus
argument is that reliance on the surface cues provided in the input
often leads to the wrong generalizations. Consider the following
examples provided by Gibson (1992) in his critique of Bates and
MacWhinney's Competition Model:
The chicken cooked the dinner. (chicken = agent)
The chicken was cooked. (chicken = theme)
The chicken cooked. (chicken = theme)
Under the Competition Model, cue validity and cue strength predict
that preverbal position here should be identified with agency. It is
not. Children do not seem to make the overgeneralization errors
that would be expected if they were analyzing the input in ways
suggested by the Outside-in theories. In fact, there are no surface
cues in the impoverished input that can assist the child in these
cases.
In sum, the following four limitations make the Outside-in theories
less than plausible accounts of how children learn language: (1) the
problem of hidden assumptions embedded within the theories; (2)
the conflation of language and communication that leads theorists
toward reductionist accounts of language (e.g., linguistic
knowledge as reducible to cognitive or social knowledge); (3) the
problem of discontinuity in grammatical development; and (4) the
poverty-of-the-stimulus argument. Although almost all Outside-in
theories suffer from one or more of these limitations, one is more
or less free from all of them.
 Page 29
Bloom's Version of a Cognitive, Outside-in Theory
Bloom (1970, 1973, 1991a, 1993) is an Outside-in theorist who has
consistently appreciated that linguistic knowledge is not reducible
to other kinds of knowledge: " . . . linguistic development is neither
isomorphic with nor a necessary result of cognitive development"
(Bloom, Lightbown, and Hood 1975, 30); " . . . how they [children]
have learned to think about the objects, events, and relations in
their experience is something apart from how they have learned to
represent such information in linguistic messages" (p. 29). Thus,
Bloom saw language as a "separate problem space" early on and
does not view language acquisition as a simple one-to-one mapping
between cognitive knowledge and linguistic forms.
Yet Bloom is an Outside-in theorist because to the question of what
the child brings to language acquisition, her answer is completely
"bottom-up." That is, she does not believe that the child begins
with any a priori linguistic categories (Bloom, Lightbown, and
Hood 1975; Bloom 1991a, personal communication); instead, she
believes that the child applies domain-general cognitive principles
(as opposed to principles specific to language) to the task of
language learning. She writes:
The processes involved in this learning [grammatical learning] are the
same as for learning anything else (physical knowledge, biology, etc.)
and consist of 1) detection of novelty, identity, and equivalence . . . and
2) abstraction, coordination, and integration, for connections between
categories. These processes themselves no doubt have a biological
basis (categorization, for instance) but depend upon such garden variety
context factors as salience, frequency, and the like. (personal
communication)
Further, although Bloom believes that children express notions like
"agent" and "location" in their earliest multiword utterances, she
avoids the problem of discontinuity by claiming that the child is
able to make use of syntactic categories from the first multiword
utterance (although not before; see Bloom 1973). Counter to
Outside-in theorists who argue for the separate acquisition of
semantics and syntax (see Schlesinger 1988; Braine 1976;
Bowerman 1973; Bates and MacWhinney 1987; Brown 1973),
Bloom maintains that children learn semantics and syntax together,
as opposed to deferring syntactic learning to some unspecified
future time (Bloom 1970; Bloom, Lightbown, and Hood 1975;
Bloom, Miller, and Hood 1975). The basis for this claim is that
Bloom found sentences such as "Lamb go here" and "Put lamb
here'' among children's earliest utterances. In "Lamb go here,"
"lamb" serves as sentence subject and as patient (or object
affected); in "Put lamb here," the unstated "you" is the
 Page 30
sentence subject and "lamb" is the grammatical direct object and
patient. Such sentences, produced by the same child, revealed too
much linguistic sophistication to be reducible exclusively to
semantic relations. If the child had access only to semantic
relations, the same noun playing the same semantic role should not
vary in its syntactic function. On the other hand, if children are
sensitive to syntactic requirementsin this case, the argument
structure required by particular verbs"lamb" may play the same
semantic role but have two different syntactic functions depending
on the verb. Bloom's data revealed that children consistently made
complex mappings between meaning and form, respecting the
syntactic subcategorization requirements of verbs that were highly
similar in meaning.
Given increasing evidence that children do indeed treat language as
a formal objectas opposed to a system isomorphic with cognitive or
social knowledgemany researchers have returned to Bloom's
position, crediting young children with knowledge of both form
classes (e.g., noun, verb) and syntactic categories (e.g., subject of
the sentence) (Pinker 1984; Bowerman 1988; Ihns and Leonard
1988; Valian 1986; Levy 1988). Some researchers, such as Pinker
(1984), indeed go further than Bloom does. According to Bloom's
view, although children are engaged early on in formal linguistic
analysis, their original categories are not identical to adult
categories but smaller in scope. Nonetheless, Bloom's emphasis on
the child's ability to engage in formal linguistic analyses provides a
bridge between the Outside-in and Inside-out theories. A
foundational assumption for Inside-out theories is that young
children conduct linguistic analyses and use syntactic as opposed to
only semantic categories. Inside-out theories differ from Bloom's
Outside-in account, however, on the source of the child's ability to
perform linguistic analyses and on the nature of the learning
mechanisms.
Inside-Out Theories
A key difference between the Inside-out and Outside-in theories is
that the Inside-out theories begin with the premise that the
complexities of language, as well as its species-specificity, could
not be accounted for without innate constraints on the "language
faculty." As Karmiloff-Smith (1989) writes:
I fail to see how it could be biologically tenable that nature would have
endowed every species except the human with a considerable amount
of biologically specified knowledge. (p. 273)
 Page 31
Inside-out theories grant the child domain-specific linguistic
knowledge (hence the term "language faculty") and emphasize
grammar discovery rather than grammar construction.
Compared with the Outside-in theories, the inside-out theories
deemphasize the role that the environment plays in language
acquisition. This is a direct result of the stance these theories take
on what constitutes language. Since on the Chomskyan (1975,
1981) view, there are linguistic phenomena such as empty
categories that are not "visible" (or "audible") in the linguistic
input, linguistic knowledge cannot in principle be wholly
discovered through inspection of the available input. Hence, the
Inside-out theorists often allude to the notion "poverty of the
stimulus."
The existence of grammatical elements like empty categories is
central to the Inside-out arguments. On this view, the child could
never learn about something that is absent from the surface
structure of the sentence and present only at an abstract level. If
one grants these linguistic facts, no corpus of utterances, however
vast, could ever advance the child's grammatical learning. The
child must be genetically "prepared" to encounter these empty
categories. Indeed, there is mounting evidence that both child and
adult processors are sensitive to some types of empty categories
(such as traces left behind by various movement processes) (see
Goodluck 1991).
The Chomskyan Basis of Inside-Out Theories
Current Inside-out theories represent an advance over previous
formulations based on Chomsky 1965. In the older versions, the
child's mind was likened to a black box whose contents defied
inspection. Into this box, commonly known as the "Language
Acquisition Device," were dumped all the linguistic generalizations
that could not arguably be derived from linguistic experience
(Chomsky 1972). When the number of linguistic rules and
transformations became unwieldyboth in the grammar and by
extension in the child's Language Acquisition DeviceChomsky
(1981) developed the "Principles-and-Parameters Theory" of what
he (1988) refers to as the "language faculty," which consists of
a set of invariant principles and a set of parameters that are set by some
linguistic environment, just as certain receptors are 'set' on exposure to
a horizontal line. (Lightfoot 1991, 647).
Thus, this theory replaced a grammar in which rules and
transformations had proliferated with an elegant, more streamlined
grammar. Multiple and
 Page 32
contradictory rules were replaced by a finite set of principles and
parameters; multiple transformations by a single transformation
("Move &x97;). By "principles," Chomsky means something very
different than Slobin (1985a,b), namely, universal statements about
how language works. One such principle, the Projection Principle
(Chomsky 1981), states that "[t]he properties of lexical entries
project onto the syntax of the sentence" (Cook 1989, 11). For
example, nouns "project'' into noun phrases; verbs into verb
phrases. Thus, once learners know the form class of a lexical item
(e.g., "chair" is a noun), they will also know how it may behave
syntactically. Knowledge of the Projection Principle allows
children to recognize that individual lexical items are part of larger
constituents that compose sentences.
With the construct of "parameters," Chomsky captures the ways
different languages handle various grammatical phenomena. One
suggested parameter called "pro-drop" captures whether or not a
language allows sentences without subjects. For example, English
does not permit subjectless sentences and even requires the use of
expletives or dummy elements to avoid this (e.g., "It's raining").
However, Italian does allow subjectless sentences (e.g., "It's
raining" is expressed simply by the verb "Piove"). To say that
children begin the language-learning process in possession of such
a parameter implies that they need to hear (in theory) no more than
a few key sentences to set the switch of this parameter to match
their target language: either "subjectless sentences are allowed"
(Italian) or "subjectless sentences are not allowed" (English).
Thus, the development of Principles-and-Parameters Theory
represented a clear advance for Inside-out theories of language
acquisition. First, it acknowledged the many important ways in
which languages differ while at the same time making the Inside-
out acquisition account seem more universal. Second, by positing a
"menu" for the setting of parameter switches, it introduced a
powerful metaphor. According to Chomsky (1988), " . . . universal
grammar is an account of the initial state of the language faculty
before any experience" (p. 61). All children need to do in language
acquisition is set a few switches, corresponding to the language
they are immersed inan act whose repercussions will extend
throughout the grammatical systemand acquire the vocabulary
items particular to that language (Cook, 1989). Third, since there
was no agreement on what these parameters were, a whole new
program of linguistic research was initiated. Once again,
acquisition research was central to this enterprise. If the proposed
parameters were the correct ones, they should be able to account
for acquisition.
 Page 33
Types of Inside-Out Theories: Structure- Versus Process-Oriented
Theories
The basis for separating the Inside-out theorists into two groups
rests on presuppositions about how the process of language
acquisition gets underway. The structure-oriented theorists begin
their story after the initial stages of language learning,
presupposing answers to our first question, "What does the child
bring to the language-learning task?" Also, until recently (Lightfoot
1989), theorists in this camp have been relatively less concerned
with finding answers to our second question, "What mechanisms
does the child use to acquire language?" The process-oriented
theorists, on the other hand, are dedicated to understanding the
origins of grammatical acquisition and the mechanisms that make it
possible. Thus, although both subgroups presuppose that nature
prepares the child for the task of learning language, the process-
oriented group provides more complete answers to the questions
that motivate this review.
Structure-Oriented Theories
The prototypical structure-oriented theorist is Chomsky himself
(1988), who has articulated the strongest version of this position.
Along with other structure-oriented theorists (e.g., Wexler 1982;
Hyams 1986, Lightfoot 1989), he assumes that whatever the exact
array of principles and parameters turns out to be, nature has given
these to the child. Thus, to the question of what the child brings to
the task of language acquisition, structure-oriented theorists answer
that the child has been endowed with considerable, explicit,
domain-specific, linguistic knowledge. The social or cognitive
framework for language acquisition so important to the Outside-in
theories is reduced to a given by Chomsky (1988), who writes that
the child
approaches language with an intuitive understanding of such concepts
as physical object, human intention, volition, causation, goal, and so
on. These constitute a framework for thought and language and are
common to the languages of the world . . . (p. 32)
For structure-oriented theorists the interesting questions have to do
with children's manifestation of the principles and parameters of
universal grammar; entry into the domain-specific system is a
given. These theorists presuppose that the child comes to the
language-learning task with the ability to segment the linguistic
stream, find word classes and grammatical categories, conduct
phrase structure analyses, and set parameters. What interests these
theorists is (1) exactly what principles and parameters children
begin with and (2) what constitutes the "primary linguistic data"
(Chomsky 1988) for setting the switches of the parameters.
 Page 34
According to Chomsky (1988), parameters can be set, and the basis
for principles revealed, by exposure to only a few sentences of the
language to be learned. This position is sometimes referred to as
the "instantaneous" view of language acquisition. When the actual
data produced by children contradicted this extreme position (see
e.g., Wexler and Chien 1985, on the late appearance of Principle B
of the binding theory), the explanation endorsed by some structure-
oriented theorists was biological maturation. Thus, with regard to
the question of the mechanisms responsible for language
acquisition, some Inside-out theorists assume that biological
maturation, parallel to the maturation of bodily organs, is primarily
responsible. Further, structure-oriented theorists eschew traditional
notions of learning as explanations for language growth. Chomsky
(1988) has written that learning a language is not really something
that the child does. Rather, it is something that "happens" to the
child in an appropriate environment. The analogy that Chomsky
(e.g., 1988) consistently uses is visual perception. Just as one
would never argue that an organism had "learned" to develop sense
receptors on the retina for the perception of horizontal and vertical
lines, so one should not argue that language is "learned."
Finally, for the structure-oriented group, the input that drives the
system or "triggers" parameter setting is certainly linguistic input.
Since not all sentences are usable by the child, the language faculty
selects relevant data from the linguistic input (see Lightfoot 1989
for discussion of what constitutes the "primary linguistic data"). In
addition to having a critical role in the setting of the parameters, the
environment is treated as playing a role analogous to the one that
nutrition plays in the onset of puberty: an impoverished
environment may restrict or suppress language growth in the same
way that poor nutrition may delay the onset of puberty (Chomsky
1988).
Process-Oriented Theories
The differences between process-oriented and structure-oriented
theories are slight. Structure-oriented theorists tend to focus on the
emergence of particular linguistic structures as a way to validate
aspects of grammatical theory. Process-oriented theorists, however,
emphasize the means through which children discover their
grammar by focusing on the mapping between function and form.
Not surprisingly, structure-oriented theorists are often linguists and
process-oriented theorists tend to be psychologists. Like the
structure-oriented theorists, the process-oriented group grants that
nature has endowed the child with
 Page 35
domain-specific linguistic knowledge that makes language learning
possible. These Inside-out theorists, however, focus on how the
child breaks into language. Thus, the fact that the child is seen as
being prepared to learn certain linguistic facts does not remove the
mystery of how the child actually accomplishes this feat. As the
name we have given them reveals, process-oriented theorists, such
as Gleitman and her colleagues (1990; Gleitman and Wanner 1988;
Gleitman and Gillette 1995) and Pinker (1984, 1989) and his
colleagues, are concerned with how initial linguistic representations
are formed and how acquisition continues once children have
produced their first words.
Pinker (1984) offers the most explicit overall theory of language
acquisition to date. To the question of what the child brings to the
task of language acquisition, Pinker answers that " . . . the child is
assumed to know, prior to acquiring a language, the overall
structure of the grammar, the formal nature of the different sorts of
rules it contains, and the primitives from which those rules may be
composed" (p. 31). That is, Pinker believes the child is equipped
with word classes like noun and verb, syntactic categories like
subject and object, and other grammatical knowledge.
Gleitman (1990; Gleitman and Wanner 1988), too, begins from the
premise that a child operating without any constraints on how to
interpret the linguistic input (or the nonlinguistic world, for that
matter) would wallow forever in a hypothesis space too large to
chance upon the correct interpretation of a word, let alone a
sentence. For this reason the child needs considerable language-
specific knowledge to even begin the task of language
learningknowledge about how to parse sentences, knowledge of
potential sentence structures, and so on. The task that children face
is somehow to derive instantiations of this knowledge from their
particular language. One glimpse into the learning problem from
this perspective is provided by word learning. Without constraints
on the myriad hypotheses a child might generate, the word "cat,"
uttered in some nonlinguistic and linguistic context (e.g., the
caregiver points to a cat lying on a pillow and says, "See the cat?"),
could be interpreted in any number of ways. It could refer to the
cat's tail, which happens to be swishing, the union of the cat and
the pillow it is lying on, or perhaps some property of the cat like its
fur. Clearly, the child must be predisposed to interpret words in
specific ways, compatible with the way already accomplished
speakers seem to use them. Markman (1989) and Golinkoff,
Mervis, and Hirsh-Pasek (1994) discuss a set of principles for
lexical acquisition that would do this work.
 Page 36
The acquisition of lexical items is only the beginning of the story,
however: the child must cross-classify these words into word
classes. Inside-out theories grant the child the universal set of word
classes. However, as Pinker (1984) points out, it is not enough to
say that children are born with word classes such as "noun" and
"verb"; they also need to be able to find instantiations of these
classes in the input. To solve this problem, Pinker (1984) posits a
mechanism that he calls "semantic bootstrapping," based on related
proposals by Grimshaw (1981) and Macnamara (1972). On this
proposal, children who encounter labels for persons, places, or
things are prepared to link these back to their linguistic counterpart,
namely, the word class "noun." The semantic bootstrapping
proposal goes well beyond this example. In Pinker's (1987) words:
[T]he claim of the Semantic Bootstrapping Hypothesis is that the child
uses the presence of semantic entities such as "thing," "causal agent,"
"true in past," and "predicate-argument relation" to infer that the input
contains tokens of the corresponding syntactic substantive universals
such as "noun,'' "subject," "auxiliary," "dominates," and so on. In the
theory outlined in Pinker (1984) this knowledge is used by several sets
of procedures to build rules for the target language. (p. 407)
Note that Pinker's account presupposes that (1) children can parse
incoming sentences; (2) they can figure out which word in the
string corresponds to the object label; and (3) they (innately)
possess "linking rules" for joining the word for the referent to the
class "noun" and further to "sentence subject."
Gleitman (1990) points out that each of these issues constitutes a
significant research problem in its own right. For example, upon
seeing a rabbit jump and hearing "See the rabbit jump!", how is the
child to know that the phonological segment "rabbit" (assuming it
has been discriminated from the stream of words around it) maps to
the object in motion? In this regard, Pinker may be overestimating
the transparency of the mapping between language and
environmental events (Gleitman 1990). Despite their
disagreements, however, Pinker and Gleitman would answer the
first question in somewhat similar ways. Children come to the
language-learning task with a linguistic endowment that
predisposes them to make certain hypotheses rather than others
about words and about grammar.
With regard to our second question on the nature of the
mechanisms for language learning, both Pinker and Gleitman
believe that children use domain-specific and domain-general
processes to analyze the linguistic input. From a domain-general
perspective, both theorists would agree that children must apply
general learning devices (e.g., pattern detection)
 Page 37
to uncover the details of the way their language works. On the
domain-specific side, children are predisposed (because of their
innate grammatical endowment) to come up with certain
hypotheses rather than others as they analyze the linguistic stream.
Indeed, theorizing far more explicitly than previous researchers,
Pinker (1984) posits a set of six procedures to describe how the
child conducts syntactic analyses. According to the "structure-
dependent distributional learning" principle, for example, children
use what they already know about grammar to assist them in
learning about words whose semantic and syntactic features are not
immediately transparent. So, a child who hears the sentence "John
told the truth" can tell from the fact that the article "the'' precedes
the word "truth" that it is probably a noun. This allows the child to
continue to parse the sentence without having to stop.
Once children have worked out something of the grammar, it
would be very useful for them to exploit the frames in which words
appear to figure out the nuances of their meanings. The clearest
case is that of blind children, who although they cannot observe
environmental contexts of use, nonetheless seem perfectly able to
distinguish between such verbs as "look" and "see." To account for
this phenomenon, Landau and Gleitman (1985) propose the
mechanism called "syntactic bootstrapping." For example, if a noun
occurs immediately before and after a verb (as with "push"), the
verb may be causative. If a noun occurs only immediately before
(as with "run"), the verb is likely to be noncausative. Obviously, in
order for syntactic bootstrapping to be possible, there must be
correlations between syntax and meaning that the child must be
able to exploit at a relatively early age. A number of current
research projects suggest that these correlations do exist and that
adults and children exploit them (see Fisher, Gleitman, and
Gleitman 1991; Naigles 1990; Fisher 1994; Gleitman and Gillette
1995).
Regarding the third question (What inputs drive the language-
learning system forward?), there is again considerable agreement
between Gleitman and Pinker. The input that propels language
learning is linguistic in nature, although the role of the
extralinguistic environment is far from minimal. Unlike the
structure-oriented theorists, Pinker and Gleitman do not talk of the
environment as a "trigger" but instead focus on the nature of the
information available in the language input, as well as in the child's
observation and interpretation of events. On both accounts the
semantic system can be used to help learners bootstrap their way
into the syntactic system at the same time that the syntactic system
helps learners narrow
 Page 38
down particular facets of the semantic interpretation. In this sense,
both Pinker's and Gleitman's positions allow for a dialectic process.
Nevertheless, although both authors acknowledge strong
contributions from semantics and syntax (Pinker 1990; Gleitman
1990), Pinker puts relatively more weight on the semantic input in
the dialectic whereas Gleitman puts relatively more weight on the
syntactic input.
In sum, the process-oriented Inside-out theorists, although granting
that nature has prepared the child for language learning, are
distinguished from the structure-oriented Inside-out theorists by
their emphasis on the question "how"; that is, they ask, "How do
children actually operate on the linguistic stream to uncover the
categories nature has prepared them to search for?" In their
attention to mechanism, the process-oriented theorists seem to
resemble Bloom (1991a), an Outside-in theorist who grants that
children actively construct grammar. Where Pinker and Gleitman
differ from Bloom is on their insistence that children do not begin
the task with exclusively domain-general learning procedures or
structures; both Pinker and Gleitman permit nature to narrow the
child's hypothesis space considerably and to endow the child with
domain-specific processes that facilitate grammar acquisition.
Critique of the Inside-Out Position
Several criticisms can be leveled at the Inside-out theories,
especially at the structure-oriented branch. First, the relegation of
linguistic insights to innate constraints built into the grammar itself
is open to question. Specifically, Inside-out theories are vulnerable
to the criticism that claims of innateness make a theory
unfalsifiable because, as Fodor and Crain (1987) argue, "A theory
which admits substantive innate principles can all too easily add
another one to the list, in response to a new observation about
natural languages" (p. 46).
The next three criticisms stem from the first. Given that much
linguistic knowledge is innate, Inside-out theories (particularly the
structure-oriented branch) often appear (1) to be
nondevelopmental; (2) to reduce the role of the environment to
serving as a trigger; and (3) to rely on the nonmechanism of
maturation when structures do not appear instantaneously.
With respect to whether acquisition has a developmental curve, it is
only in recent years (in response to data showing that children do
not get it all right from the start) that structure-oriented Inside-out
theorists have begun to grapple with the actual course of
acquisition. Bates and MacWhinney (1987) point out that this
literature is still characterized by
 Page 39
too much of an emphasis on language structures emerging cleanly
in a single trial, never to be used incorrectly again. Yet the course
of acquisition is not a smooth, ascending curve. Children have been
known to use correct structures alongside incorrect structures for
years (see Kuczaj 1977). Infant perception may be a good analogue
for how development occurs in other domain-specific, encapsulated
areas where children start out with innate knowledge. As Spelke
(1990; Spelke et al. 1992) argues, infants appear to begin the
process of perception with the expectation that objects have
solidity, take up space, and so on. However, infants do not operate
according to the Gestalt principles that older children and adults
operate with in perception; for example, they do not employ
"completion"the notion that objects that are partially occluded
continue behind their occluder. Thus, innate knowledge provides
the starting point, or the skeleton around which later learning is
organized.
In emphasizing that input sentences "trigger" linguistic progress,
the metaphor of the child as a passive reactorironically, as in the
mechanistic modelhas been reasserted. Inside-out theorists of the
structure-oriented type minimize psychological mechanisms such
as imitation, comparison, storage, and concept formation that the
child undoubtedly uses to acquire language. As Bloom (1991a)
writes:
A developmental perspective assumes that children play an active part
in acquiring language. They are in effect, "the agents of their own
development" . . . . Most learnability-theoretic research has an
essentially mechanistic world view, in which language depends on
inborn, specifically linguistic constraints triggered by relevant instances
in the input. Developmental research and theory has an organismic
 world view, with an emphasis on change, integration, and process. (pp.
13 14).
The lack of a developmental approach seriously impedes the
structure-oriented Inside-out theories from being able to explain
when particular structures are triggered by the input available from
the environment. Since for the Inside-out theories the environment
can be presumed to be a constant, it must be something within the
organism that allows the triggering to take place at time 2 when it
did not occur at time 1.
This point brings us to the next criticism. For the structure-oriented
Inside-out theorist the role of the environment in language learning
is likened to the role that nutrition plays in human growth: it is
necessaryand no more. Other than containing the linguistic input,
and the extralinguistic situations that covary with the input, the
environment serves a minimal function in the process of language
acquisition. Inside-out
 Page 40
theorists (e.g., Lightfoot 1989) are just beginning to grapple with
what it means to say that the environment serves as a "trigger" for
the acquisition of linguistic structures, even though research on
how children use the environment to service acquisition has been
going on for years (Snow and Tomasello 1989). Linguistic and
extralinguistic features of the environment appear to be
coincidentally arranged in a coalition that assists the child in
learning language.
The last criticism is that some of the Inside-out theorists either
posit no mechanisms for learning language or else rely on
maturation when the principle or structure of interest does not
appear in a timely fashion. Chomsky, again anchoring one end of
the continuum, repeatedly likens the process of acquisition to what
takes place in the development of the retina. Other than having
visual experiences, the organism has no control over the increase in
the density of the receptors on the retina. For Chomsky, language is
analogous. Referring to it as the product of "learning" is a
misnomer; instead, like vision, it is something that just happens to
the child. When it doesn't happen fast enough, biological
maturation is sometimes invoked as an explanation (Borer and
Wexler 1987). And yet as Weinberg (1987) writes:
By assuming maturation [as an account for the sequenced appearance
of various grammatical structures] we are led to a theory that merely
catalogues points in time when different principles come into play
without searching for any understanding of why particular principles
come in any given order. (p. 25)
Summary
Even a cursory review of the literature like this one highlights basic
differences among the leading theories of language acquisition
(recall table 2.1). Among the most obvious is that Outside-in
theories are likely to be domain general, with an emphasis on the
learning process, whereas Inside-out theories are likely to be
domain specific, with an emphasis on both internal structure and
processes. In Outside-in theories, the learning processes evolve
from general cognitive or social bases and are sensitive to external
input that allows children to construct grammar from rather meager
beginnings. Notice, however, that these processes must somehow
be sensitive to language-relevant information in the environment;
otherwise, children could randomly abstract units and relations that
would lead them to generate incorrect grammatical hypotheses. By
contrast, Inside-out theories assume the child to be endowed with a
strong internally consistent structural foundation (a universal
grammar) that drives the
 Page 41
learning process. Processes are thus "informed" of what kinds of
things to look for, and learning can be thought of as a process of
discovery rather than one of construction.
Given these basic distinctions, other attested differences follow.
First, in Inside-out theories, language learning rests on the child's
possession of linguistic structure; in Outside-in theories, it rests on
cognitive and social structures. Second, Inside-out theories use
domain-specific processes to construct language; Outside-in
theories use domain-general processes. Third, because of the
Inside-out theories' strong reliance on an a priori internal grammar,
language learning is thought to rely more on innate knowledge in
these theories than in Outside-in theories. Fourth, compared to
Outside-in theories, language learning is considered to be relatively
instantaneous in Inside-out theories because as some argue, much
less needs to be learned and what must be learned need only be
triggered by the appropriate environmental stimuli (at least for
those who hold that the source of linguistic structure is innate).
These distinctions provide a sketch of the major families of
language acquisition theories. But are these theories really as
polarized as the literature suggests? Are the sketches offered in this
chapter accurate portrayals of how each group of theorists believes
that language is learned? We think not. Closer inspection of the
theories reveals that many of the proposed dichotomies are
"hyperboles." That is, these dichotomies are a product of
exaggerated statements that differ more in degree than in kind. If it
is true that theories of language acquisition have more in common
than has generally been assumed, then it logically follows that
points of consensus may emerge with regard to the three questions
around which this chapter is organized.
2.3
Collapsing Dichotomies:
How Theories of Language Acquisition Rely On Hyperbole
In this section we refine our look at the theories by discussing why
characterizations of differences among them may be more
hyperbolic than real. We argue that the dichotomies that have
historically characterized theories of language acquisition are better
cast in terms of continua than in terms of polar extremes. By
collapsing theories onto continua, we do not intend to downplay the
significance of prior work in language acquisition. Rather, adopting
this treatment allows us to show the perhaps unanticipated
commonalities among the theories that guide our research. Indeed,
 Page 42
in section 2.4, we discuss what might be common answers to the
three questions guiding this chapter. The research reported in this
book empirically investigates this common groundnamely, in our
view, the basic sensitivities that form the foundation for language
acquisition.
The first of the three hyperboles that characterize theories of
language acquisition is this:
Hyperbole 1
Outside-in theories account for grammatical development in terms
of cognitive and social categories. Inside-out theories account for
grammatical development in terms of the discovery of facts about
an a priori grammatical structure.
Like all hyperboles, this one contains some truth and much
exaggeration. Panel (a) of figure 2.1 exposes the hyperbole by
characterizing the major families of theories with respect to the
kinds of categories that children are thought to bring to the
language-learning task. That is, Outside-in theories do postulate
that children start with cognitive and social categories like agents
and actions that somehow turn into linguistic categories like nouns
and verbs. In stark contrast, Inside-out theories provide the child
with a rich linguistic structure at the outset, although the nature of
this endowment depends on the particular theory.
What makes this dichotomy collapse is that neither group of
theorists relies exclusively on the categories shown on its side of
the opposition. That is, even Chomsky (1988), the most extreme of
the Inside-out theorists, acknowledges the role played by social
understandings and cognitive processes in the acquisition of
language. For example, Chomsky assumes that the child has a
knowledge of "theta roles," which are the action role categories
(e.g., agent, recipient) of the Outside-in theorists. On the other
hand, the Outside-in theorists use linguistic scaffolding to get their
models off the ground. For example, Schlesinger's (1971, 1988)
model presupposes a child who is sensitive to inflectional markings
in the input. Thus, both families of theories grant that the child has
at least some linguistic sensitivities at the start (if not a full
grammar) and is capable of conceptualizing the environment in
terms of language-relevant cognitive and social categories.
Given that each of these families of theories requires linguistic,
cognitive, and social categories, the dichotomy between theories
that rely on linguistic structures and those that rely on cognitive or
social structures collapses into a continuum (see figure 2.1, panel
(a)). Anchoring the
 Figure 2.1
Collapsing the hyperbolic dichotomies that characterize theories of language acquisition re
 Page 44
continuum at one end would be a theory that relied totally on
linguistic categories, apart from any categories constructed through
environmental interaction. Notice that there is no theory that falls at
that end, except perhaps for the incorrectly stereotyped view of
Chomsky's theory. Anchoring the other end of the continuum
would be a theory that allowed for no linguistic structure at all at
the outset of language learning. No modern-day theory is to be
found here either, given the Outside-in theorists' claims about what
is presumed to be present as children start learning language
(MacWhinney 1987; Schlesinger 1971, 1988). The two families of
theories, although still not totally overlapping in their distribution,
are now seen to have much more in common than previously
thought. Indeed, they differ more in degree than in kind.
The same picture emerges when we examine the learning
mechanisms implicated in language learning.
Hyperbole 2
Outside-in theories rely on domain-general learning processes;
Inside-out theories rely on domain-specific learning processes.
This hyperbole implies that Outside-in theories allow the child to
construct language via the same domain-general learning
procedures first used to construct cognitive and social knowledge.
Thus, as language-learning mechanisms these theories tend to rely
on general learning procedures, identical to those used to acquire
knowledge in other domains. On the other hand, Inside-out theories
are said to rely exclusively on processes tailored to detect and form
linguistic generalizations. Panel (b) of figure 2.1 illustrates the
kinds of mechanisms each family of theories claims that children
use to build language.
For a number of reasons, however, turning attention to processes
does not sharpen the line between the theories. First, as Sternberg
(1988) writes:
Processes need a knowledge base on which to operate, and the extent of
the knowledge base in large part determines what processes can operate
under which circumstances, as well as how effectively they can
operate. (p. 277)
The Outside-in theorists' domain-general learning procedures
operate on (minimally) a knowledge base that includes the
identification of linguistic elements; otherwise, the child could not
form linguistic generalizations. Second, the Outside-in theories
include procedures or mechanisms that are themselves domain
specific. For example, in the accounts of Schlesinger (1979) and
MacWhinney and Bates (1987), children "look for" patterns of
word order, inflections, or tense markers. More recent in-
 Page 45
stantiations of the Competition Model posit a learner who can
identify a word as having occurred immediately before (or after) a
verb. To accomplish this task, a learner must have some way of
identifying a verb (as opposed to words from other classes) in the
input stream (see Gibson 1992). In order for acquisition to proceed
under these theories, then, children must, a priori, be able to
identify linguistic units. Note the effect that this has on the
hyperbolic dichotomy we have drawn between the theories (panel
(b) of figure 2.1, and table 2.1). Domain-general learning in the
Outside-in theories has become open to the linguistic categories
present in the Inside-out, domain-specific theories. Once we
incorporate linguistic unit identification into the mechanisms for
language learning, these mechanisms begin to look increasingly
domain specific. Building linguistic sensitivities into the process
rather than into the knowledge baseas the Outside-in theorists have
doneis but a backdoor way to acknowledge that the child must use
some domain-specific knowledge to acquire language. The result is
to blur the line that separates the families of theories with regard to
the source of linguistic structure and process. It becomes
impossible to distinguish whether some implicit knowledge of
language guides the operation of the processes or whether the
activation of language-relevant processes guides the construction of
linguistic structure. Thus, on this analysis, the two families of
theories begin to resemble one another. Both contain some
reference to domain-specific learning procedureswhether couched
in terms of the knowledge base (as in the Inside-out theories) or in
terms of the processes employed (as in the Outside-in theories).
Also, both contain references to domain-general learning processes,
as when children extract a particular morphological pattern by
noting its frequency in the input. It is therefore false to conclude
that Outside-in theories rely exclusively or even primarily on
domain-general mechanisms. Rather, both families of theories fall
on a continuum with regard to domain-specific claims (see figure
2.1, panel (b)). In the following passage Gelman and Greeno
(1989) capture the sense of why both families of theories include
domain-specific knowledge of language, whether they place that
knowledge in structure or in process:
If we grant learners some domain-specific principles, we provide them
with a way to define a range of relevant inputs, the ones that support
learning about that domain. Because principles embody constraints on
the kinds of inputs that can be processed as data that are relevant to that
domain, they therefore can direct attention to those aspects of the
environment that need to be selected and attended to. (p. 130)
 Page 46
In other words, without some structural knowledge of what the
processes are to operate on, the organism's hypothesis space would
be far too wide to form the correct linguistic generalizations.
In sum, all theories of language acquisition require a learner who
has access to both domain-specific and domain-general learning
procedures. Domain-specific procedures are necessary if the child
is to become sensitive to peculiarly linguistic information such as
"hidden" elements or empty categories (see Fodor 1989). Domain-
general procedures such as categorization, abstraction, and
weighting of cues can also play an important role in acquiring
linguistic knowledge (MacWhinney 1987).
This hyperbole exposed, a third follows:
Hyperbole 3
Outside-in theories avoid claims for innateness whereas Inside-out
theories are replete with such claims.
This hyperbole is represented in panel (c) of figure 2.1. It is clear
that Inside-out theories assume that children enter the language-
learning task in possession of a veritable armamentarium of
linguistic principles, parameters, and categories, depending on the
theory. The collapsing of the hyperbolic dichotomy here hinges on
whether the Outside-in theorists truly avoid such presuppositions.
If we grant, as in the above discussion, that even Outside-in
theorists build in some domain-specific linguistic constraints so
that the processes responsible for language acquisition will (1)
"look for" appropriate input in the environment and (2) derive a
limited class of linguistic generalizations, the innateness dichotomy
also collapses into a continuum (again, see panel (c) of figure 2.1).
The thorny nature of the innateness issue for many developmental
questions (see Aslin 1981), along with the fact that it is difficult to
evaluate these claims empirically, leads us to avoid that quagmire.
Instead, we think this argument is best cast in terms of the
availability of information by the time that language acquisition
begins.
As pointed out earlier, although they claim to be minimalists, all
Outside-in theorists (save L. Bloom (1991a)) posit that certain
language-relevant information must be available by the time
language learning begins. It is often the case that the origins of this
knowledge are glossed over and its existence simply stated as a
given. It may be that these theorists believe this knowledge is a
product of the domain-general processes that (they would argue)
construct linguistic structure. However, as we pointed out in
discussing hyperbole 2, the fact that this infor-
 Page 47
mation is written into the procedures for learning rather than into
the structure makes the information no less available at the start of
language acquisition.
The question that should be asked is, ''How much language-specific
knowledge and what kinds of domain-specific processes are needed
at the outset of language learning to ensure that grammatical
learning takes place?" Otherwise put, how much does the child
need to know to be able to learn about syntax? Indeed, this question
is at the heart of the current attempts to define the conditions for
language learnability. Once the innateness argument is cast in these
less dogmatic terms and redefined as availability, it is clear that this
dichotomy, too, collapses into a continuum. Again, no theory
occupies the extremes. Even those from the structure-oriented
Inside-out camp assume that some learning takes place in the form
of parameter setting; and even the staunchest Outside-in theorist
posits that some linguistic information is available to learners as
they begin their journey toward linguistic competence. Panel (c) of
figure 2.1 illustrates the new continuum.
By way of analogy, no good archaeologist would jump to a
conclusion after uncovering one piece of pottery. Rather, equipped
with procedures designed to unearth the relevant data, the
archaeologist, like the child, would look for confirming evidence
and for patterns in the uncovered shards. To continue the analogy,
the archaeologist knows to "look for" certain types of clues like
particular markings or carvings on earthenware, just as the child
knows to "look for" evidence like word order and inflections in the
linguistic input. Both the archaeologist and the child are guided by
a constrained set of hypotheses. This does not, however, negate the
need for learning, nor does it argue that learning should be
instantaneous.
In sum, we have reexamined the earlier analysis of the two families
of theories. We derived three hyperboles often used to characterize
these theories in the literature. In all cases, these hyperbolic
differences have exposed the artificial dichotomies on which these
theories have traditionally been distinguished, and have revealed
that the theories are much better described in terms of continua.
2.4
Language Acquisition Theories:
Common Ground
Given that there is more agreement between these families of
theories than would be assumed from a surface analysis, in this
section we consider
 Page 48
whether some consensus can be reached on the answers to the three
questions raised earlier. In doing so, we also present our own view
of the language acquisition process.
Question 1: What is present when language learning begins?
Whether the theory is Outside-in or Inside-out, the child is viewed
as being sensitive to and able to detect a number of linguistic units
and potential arrangements of those units. The Inside-out theories
capture this knowledge in structure and the Outside-in theories in
the processes the child brings to bear on linguistic input. The result,
however, is the same: children must have access to linguistic units
(nouns, verbs, phrases, and clauses) and must be sensitive to the
possible relations between those units if they are to proceed in the
learning of grammar.
Question 2: What are the mechanisms used in the course of
language acquisition? Minimally, all theorists seem to agree that
children must be sensitive, a priori, to language-relevant data.
Children come equipped with strategies that enable them to "look
for" language units like clauses, phrases, and words. Unless these
units are detected, the abstraction of grammatical patterns is
impossible. Even those arguing in favor of initial categories
composed of agents and actions must at some point account for
how these agents and actions become the epiphenomenal linguistic
categories that are functionally equivalent to nouns and verbs. For
the purposes of a consensus view, then, one need only argue that
there must be some processes capable of detecting sentencelike
units, phrasal units, word units, and form classes (e.g., nounlike
units and verblike units), and some processes that can come to
distinguish open-class words from the smaller set of closed-class
words (e.g., inflectional markings).
These processes must also "look for" organizational patterns in the
data to detect linguistic relations. All theories of acquisition require
that children attend to the order of units, be it the order of words, of
inflectional markers relative to words (Levy 1988), or of phrases
relative to the larger constituents in which they are housed. The
theories also need to account for the child's knowledge that all
nouns (agents and themes) or clauses are not alikefor the
knowledge that language is hierarchical and that there can be
embedded clauses and main clauses.
In answer to the second question, then, our opinion is aligned with
a kind of process-oriented, Inside-out view (see figure 2.1) in
which children come to the learning task with some sensitivities to
properties in the input that are informed by internal grammatical
knowledge. It is essential not only that children learn about the
units and patterns in the input but also
 Page 49
that they be able to recover the hidden grammatical elements and to
recognize dependencies within the input that they hear. We believe
that there are multiple and overlapping cues for grammatical
structure in the input; infants must be primed to detect these inputs
and extract these correlations. In chapter 7 we develop this idea
further as a mechanism we call "guided distributional learning."
Indeed, there is ample evidence from the literature on animal
learning that animals come prepared to form certain kinds of
associations or to pick up certain kinds of information from the
world around them. For example, Gallistel et al. (1991) describe
how migratory songbirds, as nestlings, use specialized
computational mechanisms to store an image of the stellar
constellations, which they use later to guide their migration. In the
following passage, the word "infant" could easily be substituted for
"bird", and the context understood to be about language learning
rather than the learning of stellar configurations:
[W]e have a domain-specific learning mechanism that determines what
in the bird's environment will be attended to at a certain stage in its
development, how that input will be processed, and the structure of the
knowledge that is to be derived from that input. (p. 18)
As the wall between process and structure crumbles, so too does
the distinction between the first two questions. The mechanisms
used in the course of development will be those processes used to
abstract language-relevant data from the inputthose processes that
lie at the intersection of the various theories. Sometimes these
processes will be guided by structural constraints and operate in a
language-specific way: for example, "Look for inflectional
markings." At other times they will be guided by more domain-
general learning procedures: for example, "Attend to ordering
relationships." Thus, children perform guided distributional
learningfinding regularities in the input that they have been primed
to notice. Guided distributional learning, therefore, entails an
intertwining of the domain-specific and domain-general processes.
Question 3: What types of input drive the system forward? With
regard to this question, we propose that consensus among the
various theories be identified not by examining the intersection of
the possible input sources they presume, but by examining the
union of these sources. This coalition of input sources was
introduced in chapter 1 and will be developed further in chapter 7.
As the theories were reviewed above, it became evident that many
theorists concentrate on one type of input as more central than
 Page 50
others for ensuring grammatical development. The very fact that
input sources as diverse as environmental events, social interaction,
prosody, and syntactic patterns are all viewed as criterial for
different theories of grammatical learning should give one pause.
Perhaps a more eclectic stance will better explain how children
come to abstract or to discover the units and relations embedded
within the grammar. Indeed, the recent proliferation of
bootstrapping theories leads one to suspect that children live in a
benevolent world in which these input sources covary reliably with
one another (Shatz 1978; Landau and Gleitman 1985; Gleitman
1990; Pinker 1984, 1987; Grimshaw 1981). In English, for
example, the highly stressed noun is usually found in last position
in the utterance (Goldfield 1993), is commonly found preceded by
an article ("the" or "a"), and is often the focus of joint attention in a
social context (Messer 1983). It is this coalition view that
represents the eclectic position that we adopt as our answer to the
question of what types of input drive the system forward.
2.5
Summary:
Early Sensitivities to Language Input
In sum, then, this review of current acquisition theories provides a
sense of the foundation that children have for later language
learning. By the time that language (grammatical) learning begins,
children are sensitive to certain language-relevant information in
the input. Using previously undiscovered consensus among the
theories as a guide, we propose that very early on, children are
sensitive to information in the input that will allow them to "look
for" such basic units of language structure as words, phrases, and
clauses (chapter 4). Children are also sensitive to basic relations
such as agency and location, to order (e.g., of words (chapter 5)
and inflections), and at some point to the syntactic configurations
in which words appear (chapter 6).
With early attention to these basic units and relations, children can
begin the journey to full linguistic competence by noting the
distribution of units and markers within the language input that
they hear. The process by which this occurs is the guided
distributional learning mentioned above and discussed further in
chapter 7. In fact, it is only when these foundations are in place that
children will be sufficiently developed to learn language in ways
suggested by some of the theories.
Given the linguistic presumptions noted above, we must ask, "How
do children come to discover these basic units and relations in the
input?"
 Page 51
Here we suggest that the processes that children bring to the task
are sensitive to language-relevant information from a number of
sources concurrently. Units of language emerge amid a kind of
multimedia display in which the notes of prosody, context,
syntactic position, and so on, form a chord or coalition that children
can reliably identify as a nounlike or verblike unit. These smaller
units then act consistently within sentence patterns that also covary
with certain types of environmental events and
prosodic/intonational patterns. In short, children try to use all the
pieces of information at their disposal, although not necessarily at
the same time, in abstracting the units and relations of grammar.
This multimedia display will be explored in greater detail in
chapter 7 when we outline a framework of language
comprehension.
Infant and toddler sensitivity to many pieces of the multimedia
program has already been discovered. For example, during the last
decade researchers have made significant progress in uncovering
some of the ways in which children segment the speech stream (see
Hirsh-Pasek and Golinkoff 1993, for a review). A number of
studies have also investigated infant sensitivity to semantic
information in the input (e.g., Bowerman 1988; Bates, Bretherton,
and Snyder 1988; Mandler 1992). Less research, however, is
available that focuses on the infant's sensitivity to linguistic or
syntactic organization. This is largely because such information has
been very difficult to obtain from preverbal infants. In the research
that follows, we address this gap and report the development of a
new method for studying early language comprehension. This
method provides a way of looking at language processing in very
young children (beginning at 13 months) and a way of permitting
children to use the coalition of sources at their disposal. After
describing the method in some detail, we demonstrate how it has
been used to unveil some of the early processes that children bring
to the language-learning task.
In sum, if the account we have developed about the origins of
grammar is plausible, then unlike Gold's learner, the real learner is
a biased one who lives in a real social world. To learn more about
the kinds of language processes that young children bring to the
task of grammatical learning, we must begin to develop methods of
investigating early language sensitivities, that can be used with
very young children who are just beginning to learn language, and
that allow children to capitalize on their sensitivities to the
coalition of language-relevant inputs. In what follows, we present a
method of language comprehension that meets these criteria.