Content uploaded by Shannon Pruden
Author content
All content in this area was uploaded by Shannon Pruden on May 11, 2015
Content may be subject to copyright.
When [my elders] named any thing, and as they spoke turned
towards it, I saw and remembered that they called what they
would point out by the name they uttered. And that they meant
this thing and no other was plain from the motion of their body,
the natural language, as it were, of all nations, expressed by the
countenance, glances of the eye, gestures of the limbs, and
tones of the voice, indicating the affections of the mind, as it
pursues, possesses, rejects, or shuns. And thus by constantly
hearing words, as they occurred in various sentences, I col-
lected gradually for what they stood; and having broken in my
mouth to these signs, I thereby gave utterance to my will.
– St. Augustine
The average English-speaking 17-year-old knows more than
60,000 words. Since children start learning their first words
by about their first birthday, this comes to over ten new
words per day. These can be acquired without any training
or feedback; children can grasp much of a word’s meaning
after hearing it in the course of a passing conversation. Deaf
and blind children learn words, as do those who are ne-
glected and abused. In some cultures, parents make no ef-
forts to teach their children to talk, but these children none-
theless also learn words. There is nothing else – not a
computer simulation, and not a trained chimpanzee – that
has close to the word learning abilities of a normal 2-year-
old child.
Here is how they do it: Young children can parse utter-
ances (either spoken or signified) into distinct words. They
think of the world as containing entities, properties, events,
and processes; most important, they see the world as con-
taining objects. They know enough about the minds of oth-
ers to figure out what they are intending to refer to when
they use words. They can generalize; and so when they
learn that an object is called “bottle” and an action is called
“drinking,” they can extend the words to different objects
and actions. They can also make sense of pronouns and
proper names, which refer to distinct individuals, not to cat-
egories; and so they understand that “Fido” refers to a par-
ticular dog, not to dogs in general.
These capacities improve in the course of development.
Children become better at parsing utterances into words,
at dividing the world into candidate referents for these
words, and at figuring out what other people are thinking
about when they speak. Months after their first words, they
possess enough understanding of the language to learn
from linguistic context, exploiting the syntactic and seman-
tic properties of the utterances to which new words belong.
This enables the learning of many more words, including
those that could only be acquired through this sort of lin-
guistic scaffolding.
This is how children learn the meanings of words. Al-
though some of this might sound obvious, it is not the re-
ceived view in developmental psychology. Many scholars
would argue that some critical components are missing; an
adequate account requires special constraints, biases, or
principles that exist to make word learning possible. Others
would argue that I have attributed far too much to children;
BEHAVIORAL AND BRAIN SCIENCES (2001) 24, 1095–1103
Printed in the United States of America
©2001 Cambridge University Press 0140-525X/00 $12.50 1095
Précis of How Children Learn the
Meanings of Words
Paul Bloom
Department of Psychology
Yale University
New Haven, CT 06520
Paul.Bloom@yale.edu http://pantheon.yale.edu/~pb85
Abstract: Normal children learn tens of thousands of words, and do so quickly and efficiently, often in highly impoverished environ-
ments. In How Children Learn the Meanings of Words, I argue that word learning is the product of certain cognitive and linguistic abil-
ities that include the ability to acquire concepts, an appreciation of syntactic cues to meaning, and a rich understanding of the mental
states of other people. These capacities are powerful, early emerging, and to some extent uniquely human, but they are not special to
word learning. This proposal is an alternative to the view that word learning is the result of simple associative learning mechanisms, and
it rejects as well the notion that children possess constraints, either innate or learned, that are specifically earmarked for word learning.
This theory is extended to account for how children learn names for objects, substances, and abstract entities, pronouns and proper
names, verbs, determiners, prepositions, and number words. Several related topics are also discussed, including naïve essentialism, chil-
dren’s understanding of representational art, the nature of numerical and spatial reasoning, and the role of words in the shaping of men-
tal life.
Keywords: cognitive development; concepts; meaning; semantics; social cognition; syntax; theory of mind; word learning
paul bloom is Professor of Psychology at Yale Univer-
sity. He is the author of over eighty scientific publica-
tions in psychology, linguistics, and cognitive science,
and has written or edited four books. He serves as As-
sociate Editor for Developmental Psychology and Lan-
guage and Cognitive Processes. His research explores
the nature of language and thought, primarily from a de-
velopmental perspective.
simpler mechanisms suffice. Even those who are sympa-
thetic to this approach can justly complain that it is terribly
vague. What is meant by “object”? How important is chil-
dren’s understanding of other’s intentions? What sorts of
words are learned through syntactic support? Under what
circumstances are children capable of word learning? And
so on.
The goal of How Children Learn the Meanings of Words
(henceforth HCLMW) is to fill out the details of the above
account, to provide a substantive and explanatory account
of word learning. More specifically, HCLMW is an ex-
tended defense of the claim that word learning emerges
from the interaction of different capacities that humans
possess – social, conceptual, and linguistic. These capaci-
ties are powerful, emerge early, and are to some extent
uniquely human, but they are not special to word learning.
I’ll illustrate this perspective here by showing how it ad-
dresses some fundamental questions about words and how
they are learned.
1. What determines the sorts of words that
children first learn?
Whether a child is raised in the highlands of Papua New
Guinea or amongst the cafes of Harvard Square, first words
have a similar flavor. They include proper names for people
and animals, and common nouns such as ball and milk.
They include names for parts, like nose, modifiers, like hot,
words that refer to actions, such as up, and words that are
linked to social interactions, such as good-bye. To give a
feeling for this, at the age of 15 months, my son Max knew
the following words: airplane, apple, banana, belly-button,
book, bottle, bye bye, car, daddy, diaper, dog, eye, kitty,
light, mommy, and uh oh. This is a conservative list; these
are the words he used many times in different circum-
stances, without the need for any prompting.
1
At first blush, there is nothing to explain here. These are
the words he heard, after all. And, not coincidentally, they
correspond to notions that a child would be expected to
understand. You would not expect a 1-year-old to start off
with the words mortgage and conference. Children do not
often hear such words and do not understand what mort-
gages and conferences are. Bottles and diapers are often
discussed, and easily understood, and so names for these
things are the sorts of words that are first learned.
But it is not that simple. The character of children’s vo-
cabularies cannot be predicted solely on the basis of the
words they hear and the concepts they possess. For in-
stance, children’s vocabularies contain more names for ob-
jects, words such as dog, cup, and ball, than are present in
the speech that is directed to them. This is true across lan-
guages and cultures. And when taught a new word in the
presence of an object, children are prone to interpret it as
labeling the entire object, not a part of the object, a prop-
erty of the object, or the action that the object is taking part
in. These facts have led many scholars to conclude that chil-
dren possess a whole object bias, a preference to interpret
novel words as referring to object categories (e.g., Macna-
mara 1972; Markman 1989).
Where could such a bias come from? It might be a con-
straint on learning, either innate or learned, that applies
specifically to the interpretation of novel words. It might be
a bias of mentalistic interpretation; that is, children assume
that word-users are prone to refer to whole objects. Or, the
roots of this bias could be syntactic: children might assume
that nominals (or nouns, or count nouns – theories differ)
refer to kinds of whole objects.
Each of these proposals has its supporters, but they are
all too narrow in scope. The whole object bias is not limited
to words, or to nominals; it does not apply only to our in-
ferences about the thoughts of others. Spelke (e.g., Spelke
1994) has found that prelinguistic infants are strongly bi-
ased to parse the world into discrete bounded entities;
these entities (“Spelke objects”) correspond precisely to the
notion of “object” used in theories of language acquisition.
Similarly, there is a strong object bias in linguistic counting,
and, earlier, in children’s nonlinguistic enumeration of en-
tities in the world. In fact, the number data motivated De-
haene (1997) to present his own version of the whole object
bias: “The maxim ‘Number is a property of sets of discrete
physical objects’ is deeply embedded in [babies’] brains.” In
light of these findings from outside the domain of language,
the bias towards objects is best seen as a general stance
towards the world, one that is manifested in several areas,
including perception, numerical cognition... and word
learning.
Objects are special, but this does not mean that object
names are all that children can learn. In fact, such names
typically constitute a minority of a child’s vocabulary. (Max
was a bit unusual in this regard). Young children also pos-
sess words that refer to actions, properties, and substances,
as well as to parts, collections, holes, and other non-object
individuals. Children’s bias towards objects reflects a bias,
not a conceptual limitation. In fact, adults show precisely
the same bias; when placed in an experiment involving the
interpretation of a new word, you would also tend to favor
the object interpretation, even though you are fully capable
of reasoning about parts, properties, collections, and so on.
How do children overcome this bias? They can learn a
name for a non-object individual, such as nose or family,
with little trouble, so long as the word is used when a can-
didate whole object is not present, when pragmatic cues
dictate the word is not an object name, or when other indi-
viduals are made salient enough. Chapter 4 goes into some
detail as to what makes some individuals more cognitively
natural than others, for both children and adults (see also
Giralt & Bloom, 2000).
When children hear a new word that refers to a specific
object, such as a dog, they are faced with a choice. Does the
word refer to a kind, as with common nouns such as dog and
animal, or does it refer to an individual, as with a pronoun
such as he or a proper name such as Fido. This information
is not present in the input in any direct sense, and yet chil-
dren always seem to get it right.
2
Chapter 5 explains how
they do this. They attend to certain key differences between
these types of words; such as their syntax (“I saw the dog”
vs. “I saw Fido/him”), and the type of entities they refer to
(proper names and personal pronouns tend to refer to ani-
mate entities; common nouns and other pronouns have no
such restriction). Once a child knows whether a name refers
to a kind or to an individual, there is the further problem of
understanding the word’s precise meaning, a topic that is
pursued in Chapter 5 (for pronouns and proper names) and
Chapter 6 (for common nouns).
Some words will never show up in children’s early vo-
cabularies. This is in part for the banal reason that children
do not hear the words and do not possess the correspond-
Bloom: Précis of How Children Learn the Meanings of Words
1096 BEHAVIORAL AND BRAIN SCIENCES (2001) 24:6
ing concepts. But it is also because there exist logical de-
pendencies within language learning. No child will start off
learning the meanings of many, the, and some, even if they
hear these words and possess the corresponding notions.
This is because in order to understand what these words
mean, the child needs some inkling of the meanings of the
common nouns that surround them. One can learn dog just
by hearing it in isolation, but many requires a grasp of the
larger semantic context to which it belongs. (The extent to
which this is true of other types of words is a question we
will turn to below.)
It turns out that the nature of children’s first words has
little to do with the fact that they are children. Rather, it has
to do with the fact that they are people who don’t yet know
the language that everyone around them is using. Imagine
being transported, with all your intelligence and memories
intact, into the body of a 12-month-old raised in a foreign
country. Like a normal child, you would have a whole ob-
ject bias. You would be limited to learning only the words
you hear. And you would be subject to the logical depen-
dencies inherent within a language. So your first words
might well be: diaper, bottle, kitty, and so on. (Your sole ad-
vantage would be your conceptual sophistication, so if you
saw someone point to a modem and say the word for it, you
might, unlike a normal child, learn the word modem.) In
support of this, when adults are shown videos of parents in-
teracting with toddlers, in which the words that the parents
are using are replaced with beeps, adults tend to learn (and
fail to learn) the very same words that the toddlers do
(Gillette et al. 1999).
2. What determines the time-course of word
learning?
If you read a textbook summary of word learning, you will
learn the following: Children’s first words have bizarre,
non-adult meanings, and are learned slowly. Then, at about
16 months, or after learning about 50 words, there is a sud-
den acceleration in the rate of word learning. This is known
as a naming explosion, vocabulary explosion, vocabulary
spurt, or word spurt. Children now have a knack for word
learning; and learn five, ten, or even twenty new words a
day.
In Chapter 2 of HCLMW, I suggest that none of this is
true. I am not denying that young children use words in odd
ways. For instance, Bowerman (1978) notes that her daugh-
ter Eva used moon to talk about, among other things, a half
grapefruit, the dial of a dishwasher, and a hangnail. Clark
(1973) gives the example of a child who called a doorknob
apple. And Max would put objects on his head and happily
describe them as “Hat!” Adults don’t talk this way.
But such examples can be explained without positing any
qualitative difference between children and adults. Some
might be speech errors, slips of the tongue. Others might
not be errors at all. When a child calls a doorknob “apple,”
it could mean that the doorknob is like an apple. This is es-
pecially likely if children do not know the right word for
what they wish to talk about, and lack the linguistic re-
sources to use phrases such as “is like a. . . .” And children
are prone to mischief – when Max put a slice of green pep-
per on his head and called it a hat, he found the situation
hilarious, much more so than normal hat-naming. Further-
more, these odd usages are the exception. Large-scale stud-
ies of early language suggest that errors are rare and, when
they do occur, they can be explained in terms of immature
understanding, as when a child thinks that a cat should be
called “dog” (e.g., Huttenlocher & Smiley 1987).
Two other facts attest to the sophistication of children’s
first words. First, one sign that children have a mature un-
derstanding of the referential nature of words is when they
point at things and ask about their names. Two-year-
olds can ask “What’s that?”, but younger children can ask
the same question by pointing and saying something like
“Wha?,” “Tha,” or “Eh?”. (Max said “Doh?”). In Nelson’s
(1973) seminal study of 18 children’s first words, she found
that most had a word that was used in just this way before
they learned 50 words – and six of the children had one
among their first ten words.
Second, even very young children are surprisingly good
at learning words. In one study by Woodward et al. (1994),
13-month-olds were told the name of a novel object nine
times in a five minute session (“That’s a tukey. See, it’s a
tukey. Look, it’s a tukey . . .”). Another novel object was pre-
sent and was also commented on (“Oooo, look at that. Yeah,
see it? Wow, look at that. . . .”), but was not named. When
later asked to point out “the tukey,” they could do so better
than chance – even after a 24-hour delay.
What about the word spurt? Does such a thing actually
exist? There are countless studies that discuss this event,
but these define it as a point at which children learn words
at a certain rate, typically about 10 or more words in a 2– 3-
week period. By this definition, a spurt does exist and usu-
ally occurs in the second year of life. Given the size of the
adult vocabulary, its existence is a mathematical necessity –
in order to end up with tens of thousands of words, children
must go through a lengthy period in which they are learn-
ing more than three words per week.
The problem with this definition is that it has nothing to
do with a spurt (or burst or explosion) in any normal sense
of the term. Instead of a sudden transition in word learn-
ing, children might reach this point of word learning as the
result of a gradual increase in rate. Elman et al. (1996) sug-
gest that this is what happens in normal development: the
pace of vocabulary development exhibits a gradual linear
increase. The data I summarize in Chapter 2, as well as
more recent analyses (Ganger & Brent 2001), support the
conclusion that this is true for most if not all children.
This is more than a pedantic point about the proper use
of words such as “spurt”; it means that a common view of
word learning should be abandoned. It is not that children
start off learning words slowly and – boom! – they speed up
to an adult rate. All the theories that posit some special
event at this point in development – some nominal insight,
conceptual change, restructuring of phonological memory,
or shift to the use of syntactic cues – are explaining a phe-
nomenon that does not exist. The real story of the course of
word learning is more pedestrian: children start off slow
and gradually get faster. As far as we know, this continues
until puberty: A 2-year-old is learning words faster than a 1-
year-old, a 3-year-old faster than a 2-year-old... and a 9-
year-old is learning words faster than an 8-year-old (see An-
glin 1993). There is no evidence for a qualitative shift.
Why does the rate of word learning increase as children
get older? This is in part due to increasing experience with
the language, which opens up the space of possible words
that can be learned. A young child – like an adult learning
a second language – is restricted to words whose meanings
Bloom: Précis of How Children Learn the Meanings of Words
BEHAVIORAL AND BRAIN SCIENCES (2001) 24:6 1097
can be inferred through perceptual experience. As she
learns more of the language, she gains access to linguistic
information relevant to word learning. Later on, literacy ex-
poses children to more words, and it is likely that the gar-
gantuan vocabularies of some adults (over 100,000 words)
could not arise without the ability to read.
But some of the reasons for the increase in rate have to
do with developmental differences that would not apply for
an adult learning a second language. As children get older,
they get better at picking up words from context, at figur-
ing out what people are intending to say when they use
words, and at understanding the meanings of such words.
Like any other skill, word learning is more efficient after
several years of practice. And older children and adults are
better in general at tasks that involve demands on atten-
tional resources. For all of these reasons, 2-year-olds are
slower word learners than 5-year-olds, and 5-year-olds are
slower than 10-year-olds.
The rate of word learning has to stop increasing some-
time; adults are not learning hundreds of words a day. This
is not because of some “critical period” for word learning;
there is no evidence that children learn words quicker or
more efficiently than adults. It is because we run out of
words. Unless we learn a new language, our only opportu-
nities for word learning are proper names, archaic or tech-
nical terms, or new words that enter the language. Most 5-
year-olds are exposed to many new words in the course of a
day; most 25-year-olds are not.
3. How can we explain individual differences?
The above generalizations about the character and rate of
early word learning might make some people uncomfort-
able, since I am talking as if all children use the same words,
and develop in the same way. In fact, even if we restrict our
focus to normally developing children raised in Western
cultures, children do differ. Some learn words early and
quickly; there are 18-month-olds who produce hundreds of
words. And there are those who hardly speak at all until
their third birthdays, and yet end up normal language users.
Some adopt a referential style – first learning single words,
mostly nouns, and then combining them into sentences and
phrases. Other children are more expressive, starting off
with memorized routines, such as I want it, and using them
for social purposes.
The mere existence of individual differences tells us
nothing about word learning. There is variation in any as-
pect of human psychology and physiology that can be mea-
sured on a continuous scale. It would be a different story if
there were nothing but individual differences – if there
were no generalizations that one could make about word
learning – but nobody thinks that this is the case; the indi-
vidual differences that exist can be seen as constrained vari-
ations on a universal theme.
This is not to deny that the study of individual differences
can be of value. There is the obvious clinical question of
what sort of development is a cause for alarm. And studies
of individual differences might inform us about the nature
of word learning; one could explore the dimensions of vari-
ation as a way to determine the number and type of distinct
mechanisms that underpin normal word learning. For in-
stance, if children learn words through different capacities
than those involved in the learning of syntax, then one
might expect to find dissociations between the two types of
learning. There should be children who are good at word
learning but bad at syntax, and vice versa. Or, one might
find that the age of onset of certain syntactic milestones is
highly heritable, while this is not the case for milestones of
word learning – or vice versa.
Some of this work has generated interesting results (e.g.,
Fenson et al. 1994; Ganger et al. 1997), but I want to end
on a dour note. Nobody knows how to explain individual
differences. We do not yet know why they exist, and we do
not know what causes them. The usual suspects in devel-
opmental psychology – sex, birth order, social class – have
only small predictive power when it comes to explaining
variation in early word learning (see Fenson et al. 1994).
The only good predictor of children’s rate of vocabulary
learning is the vocabulary size of their parents, and this can
be explained in several ways – direct and indirect effects,
child-adult causation or adult-child causation. I do not think
we will learn more by collecting more data. The real prob-
lem here is that the dimensions of variation that we look at
are crude – we count words, measure the length of sen-
tences, compare the ratio of nouns to verbs, and so on. We
do this because we do not know what else to do. We will
only come to understand individual differences in the con-
text of a mature theory of how word learning works in gen-
eral.
4. What is the role of the input in word learning?
Children who are raised in situations in which nobody is try-
ing to teach them language nonetheless come to know the
meanings of words. Even in the happiest and most sup-
portive families, words are not always used to refer to what
children are attending to, and yet mapping errors are virtu-
ally non-existent. And there is abundant experimental evi-
dence that children can learn object names when they are
not attending to the object that is being named by the adult.
Children learn words, but they do not need to be taught
words.
On the other hand, many Western parents do try to teach
their children new words. They engage in “follow-in” label-
ing, in which they notice what their children are looking at
and name it. They have an implicit understanding that chil-
dren assume new nouns will be basic-level names, such as
dog or shoe, and so, when presenting children with nouns
that are not basic-level names, they use linguistic cues to
make it clear that the words have a different status. For in-
stance, when adults present part names to children, they
hardly ever simply point and say “Look at the ears.” Instead,
they typically begin by talking about the whole object (“This
is a rabbit . . .”) and then introduce the part name with a
possessive construction (“...and these are his ears”). Sim-
ilar linguistic support occurs for subordinates (“A pug is a
kind of dog”) and superordinates (“These are animals. Dogs
and cats are kinds of animals”).
It would be perverse to say that all of this careful behav-
ior on the part of adults is a waste of time. The argument
that runs through HCLMW (see especially Ch. 3) is that
children are remarkably good at figuring out the thoughts
of adults. Does it make sense to claim that adults are so
comically inept at figuring out the thoughts of their children
that they go through elaborate, useless, efforts? It is more
likely that parents know what they are doing. And in fact,
Bloom: Précis of How Children Learn the Meanings of Words
1098 BEHAVIORAL AND BRAIN SCIENCES (2001) 24:6
there is considerable evidence that children do learn words
best when the words are presented in just the circum-
stances that parents tend to teach them.
The natural conclusion here is that these naming pat-
terns on the part of adults really are useful, they just aren’t
necessary. Environments differ in how supportive they are,
and word learning is easier when speakers make the effort
to clarify their intent and exclude alternative interpreta-
tions. But children are good enough at word learning that
they can succeed without such support. (In this regard, as
in many others, first language learning by children is simi-
lar to second language learning by adults.) This leads to the
obvious prediction that children raised in environments in
which this support is present should learn words faster than
those in which it is absent. Although there is some anecdo-
tal evidence that children raised in societies without object
labeling are somewhat slower at word learning than those
raised in most Western societies (Lieven 1994), there is as
yet no systematic research into this issue.
5. What is fast mapping?
One striking fact about word learning is that young children
can grasp aspects of the meaning of a new word on the ba-
sis of a few incidental exposures, without any explicit train-
ing or feedback. The classic study was done by Carey and
Bartlett (1978). They casually asked 3- and 4-year-olds
to walk over to two trays, a blue one and an olive one, and
to “Bring me the chromium tray, not the blue one, the
chromium one.” All of the children retrieved the olive tray,
correctly inferring that the experimenter intended
“chromium” to refer to this new color. Even six weeks after
hearing the new word, children typically retained some un-
derstanding of its meaning, if only that it was a color term.
This process of quick initial learning has been dubbed “fast
mapping.”
Fast mapping is cited as evidence for the power, and
uniqueness, of children’s word learning. But does the same
ability show up when children are learning items other than
words? Markson and Bloom (1997) have pursued this ques-
tion in a series of studies designed to be similar to the orig-
inal Carey and Bartlett study. In one part of the study, three-
year-olds, four-year-olds and adults were exposed to ten
objects, and participated in a sequence of activities in which
they were asked to use some of the objects to measure other
objects. For one of the objects, they were told: “Let’s use
the koba to measure which is longer... We can put the
koba away now.” Subjects were not asked to repeat the
word, and there was no effort made to ascertain whether
they even noticed that a new word was being introduced.
During the test phase, subjects were presented with the
original array of ten items, and asked to recall which object
was the koba (“Is there a koba here? Can you show me a
koba?”). We found that even after a month, subjects tended
to remember which object was the “koba,” replicating and
extending the original fast mapping findings.
The second part of the study involved the learning of a
nonlexical fact. We told them: “Let’s use the thing my un-
cle gave to me to measure which is longer... We can put
the thing my uncle gave to me away now.” Then they were
tested with: “Is there something here that my uncle gave to
me? Can you show me something that my uncle gave to
me?” The results here are basically identical. All age groups
again performed significantly better than chance even after
a month.
Further studies found that children do equally well when
a fact is combined with a new word: “This came from a place
called Koba.” We also found that children retain the precise
word or fact that they have been taught. If taught “koba” or
“given by uncle,” and later asked “Where’s the modi?” or
“Where the one given by my sister?”, they do not point to
the object that was originally taught. Other studies suggest
that 2-year-olds can also fast map words and facts, and that,
as with older children and adults, there is no advantage for
words.
It is conceivable that there are two distinct capacities or
mechanisms explaining our results, one underlying the new
word task, the other underlying the nonlexical “Uncle” task.
But given two patterns of learning that are virtually identi-
cal, it is more parsimonious to see them as emerging from
one mechanism, and not two. This suggests that, as with the
object bias, fast mapping applies across domains, with
words as just one instance of this. Ongoing studies, dis-
cussed in Chapter 3, are exploring the nature and scope of
this learning mechanism.
6. What is the role of theory of mind in word
learning?
A good way to teach children what dog means to is to point
to a dog and say “Look at the dog!” Why does this work so
well? One proposal, advanced by Locke, Hume, and many
contemporary associationists, is that children learn words
through a sensitivity to the statistical co-occurrence be-
tween what they hear and what they see. Because of this,
the best way to teach a child an object name is to make sure
the child is observing the referent of the name at the mo-
ment you say it. In this way, the sound “dog” will become
reliably associated with the perception of dogs, and the
meaning will be quickly and accurately learned.
The alternative theory is presented most eloquently by
St. Augustine in the quote that begins this article. This is
that children learn words through their sensitivity to the
referential intentions of other people, through use of “the-
ory of mind.”
3
Because of this, the best way to teach a child
an object name is to make it as clear as possible that you are
intending to refer to the referent of that name; and the best
way to do this is to point and say the word. In this way, the
child can infer that the speaker means to pick out the dog
when using this new word, “dog,” and the meaning will be
quickly and accurately learned.
Much of HCLMW is a defense of this Augustinian the-
ory. Consider, first, the circumstances under which children
learn words. As discussed above, words are not always used
when the child is attending to what the word refers to. Some
of the time a child hears “dog,” she will be looking at a dog;
but some other times, she will be looking at her foot. If chil-
dren were simply attending to associations, one of two
things should happen. Either they should make silly mis-
takes, and think dog means foot, or they should be ex-
tremely cautious, and only guess the meaning of a word
when there have been enough instances of word use so that
the child could be confident that the mapping is the correct
one. But children do not make silly mistakes and they are
not cautious; despite the noisiness of the input, they never
think dog means foot, and they do not need multiple trials
Bloom: Précis of How Children Learn the Meanings of Words
BEHAVIORAL AND BRAIN SCIENCES (2001) 24:6 1099
to learn the word. This suggests that children attend to an
information source that is more robust and reliable than
spatio-temporal contiguity.
In particular, there is a rich body of research suggesting
that young children are exceptionally good at using men-
talistic cues – such as eye gaze and emotional expression –
to learn nouns and verbs, and can do so even when these
cues conflict with information provided by the statistics in
the scene. To take an example from a classic study by Bald-
win (1991), imagine a child who is staring at a novel object
when hearing a speaker say “Toma! There’s a toma!” One
might imagine, following Locke, that the child would natu-
rally connect the word with the object she’s looking at. And
she will – but only if the speaker is also looking at that ob-
ject. If the speaker is gazing in another direction, at another
object, the child will follow the speakers’ gaze and assume
that this different object is the referent of the word. It is
presumably though such a process that children avoid seri-
ous mapping errors.
Theory of mind does more for the child than solving the
mapping problem. In Chapter 3, I propose that some of
children’s basic assumptions about word learning, such as
the belief that words will be arbitrary and bi-directional
(“Sausserian”) signs, or that words do not have overlapping
reference (“mutual exclusivity”; Markman 1989), follow
from their understanding of the mental states of people
who use words. A further role of theory of mind concerns
children’s understanding and naming of human-made cre-
ations. One example of this, discussed in Chapter 7, con-
cerns representational art. When young children name pic-
tures, by themselves and by others, they name them based
on the intent of the artist, and do not rely on what the pic-
ture looks like. Even something as apparently simple as a 2-
year-old pointing to a scribble and calling it “Mommy” in-
volves an impressive act of mentalistic attribution.
To appreciate the importance of theory of mind, consider
what happens when it is missing. In HCLMW, I discuss
three populations who do not learn words – nonhuman pri-
mates, autistic children, and pre-linguistic children. I sug-
gest that in all of these cases, the problem in word learning
lies in the failure to appreciate the representational inten-
tions of other people. For babies and chimpanzees, this
deficit is so extreme that it entirely precludes word learn-
ing. The autism example is more nuanced, as autistic indi-
viduals differ in the ability to learn language. Many do not
speak at all, a few speak normally, and most fall in between.
Following Frith and Happé (1994), I argue that the extent
of an autistic individual’s problems in word learning is a di-
rect consequence of the extent of his or her impairment in
theory of mind. With only a mild impairment, word learn-
ing is largely successful.
4
With more of an impairment,
there are problems in word learning, including bizarre
mapping errors and confusions in pronoun usage. And with
a severe impairment in theory of mind, there is no word
learning at all.
7. What is the role of linguistic cues in word
learning?
Let’s return to children’s first words. These are learnable
through observation of the world and attention to the in-
tentional acts of word users, and include nouns such as ball
and verbs such as kiss. But a child’s first word could not be
a determiner, modal, conjunction, or preposition. It could
not be a verb such as dreaming, an adjective such as former,
or a noun such as tenure. The meanings of these words are
not accessible in the same way.
Children learn these words through the support of lan-
guage. This might happen by having someone explicitly tell
the child the meaning of a word, through definitions and ex-
amples. But such intentional teaching applies to a small
minority of words at best. Nobody has ever learned what
the means by having someone define it, because it is virtu-
ally impossible to define. Instead children learn its mean-
ing from understanding phrases such as “the big dog” and
inferring the semantic contribution that this morpheme
makes towards the meaning of the entire phrase (see Ch.
3). For content words, other sources help. For instance, a
child might hear a sentence like “I’m really annoyed that
you kicked the dog,” and could use the (nonlinguistic) fact
that the speaker is quivering with anger to guess that the
verb annoyed has a negative connotation, as well as the (lin-
guistic) fact that the verb takes a sentential complement to
infer that it is likely to refer to a mental state.
One can distinguish two sources of linguistic information
– semantic context and syntactic context. Semantic context
is provided by the meanings of the individual words and of
the utterance as a whole. A child hearing “Do you want me
to buy some lobster for dinner?” can infer from the seman-
tics that lobster refers to the sort of stuff that one might will-
ingly eat. Syntactic context is information provided by the
syntactic category of the word. A child hearing “I would like
you to clean your room” could infer from the syntax that
clean refers to an action that one entity does to another en-
tity.
5
In HCLMW, I have little to say about learning from se-
mantic context; there is little work on the topic, despite its
obvious importance. Such learning seems to be the result
of some general intelligence that has little to do with lan-
guage per se. The research focus of many developmental
psychologists has instead been on a far more encapsulated
(and tractable) issue – the role of syntactic cues in word
learning. This is the main topic of Chapter 8.
In the first study to explore the use of syntactic cues in
word learning, Brown (1957) showed 3- and 4-year-olds a
picture of an action performed upon a substance with an
object. Some children were told “Do you know what it
means to sib? In this picture, you can see sibbing” (verb syn-
tax), others were told “Do you know what a sib is? In this
picture, you can see a sib” (count noun syntax), and others
were told “Have you seen any sib? In this picture, you can
see sib” (mass noun syntax). The children were then shown
three pictures, one depicting the same action, another de-
picting the same object, and a third depicting the same sub-
stance. They were asked, according to what they were ini-
tially told, “show me another picture of sibbing” (verb
syntax), “another picture of a sib” (count noun syntax), or
“another picture of sib” (mass noun syntax). Brown found
that the preschoolers tended to construe the verb as refer-
ring to the action, the count noun as referring to the object,
and the mass noun as referring to the substance, and he
concluded: “young English-speaking children take the part-
of-speech membership of a new word as a clue to the mean-
ing of a new word” (p. 26).
Since this paper, there have been dozens of studies show-
ing that young children can use syntactic cues to determine
the aspects of the meanings of a word; some of this research
Bloom: Précis of How Children Learn the Meanings of Words
1100 BEHAVIORAL AND BRAIN SCIENCES (2001) 24:6
is summarized in the table below. (For the purpose of illus-
tration, English words are used in the examples, but the ex-
periments themselves would use nonsense words; for in-
stance, a child might hear “He feps the dog,” and then be
tested as to what she thinks “feps” means.)
The fact that children can use syntax to learn aspects of
the meanings of words in an experimental context raises the
question of precisely what role syntax plays in real-world
word learning.
6
We can quickly dismiss the idea that chil-
dren can learn words solely through attention to this infor-
mational source. Syntax is far too crude; it can help the
learner make broad distinctions, such as whether a word
refers to an individual versus to a property, or to an event
with one participant versus an event with two participants.
Word meanings are much more fine-grained. Children
have to learn the difference between cup and dog (both
count nouns), big and small (both adjectives), and two and
six (both quantifiers of precise numerosity). For this, syn-
tax is no help at all.
Is syntax ever necessary? Plainly, some words can be
learned without linguistic support. You can point to a rab-
bit and grunt “Rabbit!” and the word will be learned. But
Gleitman (1990) has argued that syntax plays a significant
role for the acquisition of certain other words, including
many verbs. While a noun like rabbit can be learned
through perceptual information and an attention to the ac-
tions of others, these information sources are inadequate
for learning a verb like thinking. Even if children know
what thinking is, the support of syntax is necessary in order
to establish the mapping from this notion to the English
word “thinking.” This proposal is supported by experiments
in which adults are exposed to much the same input as tod-
dlers are, and have to guess the meanings of words. Adults
find it harder to learn verbs such as thinking than nouns
such as rabbit; and linguistic support is a large help for the
verbs and no help at all for the nouns (Gillette et al. 1999).
It would be a mistake, however, to conclude that there
are two types of words: those learned through nonlinguis-
tic cues, and those learned through linguistic cues, syn-
tactic, and otherwise. Many words are acquired through
multiple mechanisms. Karen Wynn and I have looked at
number words (discussed in Ch. 9) as a case-study of such
learning. Number words are interesting because they are
learned in two distinct stages. In the first, children learn
that they correspond to numerosities of sets, but they do not
know which numerosities they correspond to. For instance,
a child might not know whether two applies to two entities
as opposed to five entities or a hundred entities. This lim-
ited understanding emerges from a sensitivity to linguistic
cues that tell the child that two is a number word – but
nothing else. In the second stage, which can occur much
later, children learn the precise meanings of the number
words – they learn that two means two – and they do so
through a developing sensitivity to extra-linguistic cues to
number word meaning.
8. What are word meanings?
The assumption throughout HCLMW is that to know the
meaning of a word is to have:
(1) a certain mental representation, or concept, that
(2) is associated with a certain form.
This is the sense of “knowing the meaning of a word” im-
plicit in most discussions of language development, both
scientific and informal. Saying, for instance, that a 2-year-
old has mixed up the meanings of cat and dog implies that
the child has the right concepts but has mapped them onto
the wrong forms. Although there are no word meanings
without corresponding concepts, there can be concepts that
are not associated with forms. A child might have the con-
cept of cat, but not yet know the word, and even proficient
adult users of a language might have concepts that they do
not have words for.
But what are concepts? Chapter 6 of HCLMW defends
an essentialist theory of concept learning and representa-
tion. This differs from certain alternative accounts which
posit that the mature understanding of dogs and chairs is
best captured either in terms of a weighted set of features
(as in prototype theories), or as a set of points in a multi-
dimensional feature space (as in exemplar theories). While
such representations might exist, I argue that any theory of
concepts is incomplete if it does not acknowledge the es-
sentialist bias that is central to our understanding of cate-
gories – the tacit belief that members of a category share
their properties by dint of deeper properties that the mem-
bers possess.
“Essentialist” is an epithet in some circles, but the sort of
essentialism that children naturally endorse is adaptive and
sensible. It is adaptive and sensible because it is true. Ob-
jects in the world are not randomly distributed with regard
to the properties they possess; instead they fall into cate-
gories. These categories, such as mountains, tigers, and
chairs, are the products of physical law, biological evolution,
and intentional design. Such categories tend to be induc-
tively rich, in the sense that once you know that something
is a member of such a category, you know other relevant
things about that entity. Examples of such categories are
tigers and chairs. An example of an unnatural category is:
“things that are not cows.” Such a category is inductively
poor; the only common property that members of this cat-
egory share is that they are not cows – and so it is useless
from a cognitive or scientific point of view (see also Bedford
1997).
Concepts are useful only insofar as they correspond to
the inductively rich categories. Perceptual similarity is an
excellent guide to the formation of such concepts, but it is
Bloom: Précis of How Children Learn the Meanings of Words
BEHAVIORAL AND BRAIN SCIENCES (2001) 24:6 1101
Example Syntactic Cue Usual Type
of Meaning
This is a cat Singular count noun Individual member
of a category
These are cats Plural count noun Multiple members
of a category
I see John Lexical NounPhrase Specific individual
Here is some Mass noun Non-individuated
water stuff
He sleeps Intransitive verb Action with one
participant
He kisses the Transitive verb Action with two
dog participants
This is a big Adjective Property
dog
The dog is on Preposition Spatial relationship
the table
not enough. To make sense of the world, to categorize en-
tities in a maximally useful way, the child has to grasp that
the category an object falls into is governed by deeper prop-
erties of the object (such as its internal structure), not solely
by what it looks like. This proposal does not entail that peo-
ple actually know what these deeper properties are. For in-
stance, to be an essentialist about tigers does not require
that you know the internal properties that make something
a tiger, just that you believe some such properties exist.
Hence, an essentialist can entertain the possibility that
something might resemble a tiger but not actually be a tiger,
or not resemble a tiger but be a tiger nonetheless. Some
scholars suggest that essentialism arose only as the result of
a scientific world-view (e.g., Fodor 1998a), but the opposite
is more likely: the belief that certain entities have essences
is what motivates scientific inquiry in the first place.
There is a large body of evidence suggesting that chil-
dren’s categorization, and their use of words, is governed by
an essentialist conceptual system. Children believe that if
two animals fall into the same category, they are likely to
share the same hidden properties, even if they don’t look
alike (e.g., Gelman & Markman 1986). They know that cat-
egory membership is not solely determined by appearance,
and so a porcupine that has been transformed to look like a
cactus is still a porcupine, a tiger that is put into a lion suit
is still a tiger (Keil 1989). They know that if you remove the
insides of a dog (its blood and bones) it is no longer a dog
and cannot do typical dog things such as bark and eat dog
food, but if you remove the outside of a dog (its fur), it re-
mains a dog, retaining these dog properties (Gelman &
Wellman 1991).
Most of the research into essentialism has focused on
children’s understanding of animals and animal names; in
large part because essentialism within philosophy is typi-
cally restricted to so-called “natural kinds.” But I argue in
HCLMW that the same sort of essentialism holds for hu-
man-made entities, for artifacts. Here the essence is not in-
ternal, rather, it concerns the intention of the creator. This
predicts that when children name and categorize artifacts
they should be sensitive to what the artifact was created to
be; something might not look like a typical clock, for in-
stance, but if its design reflects the intention for it to be a
clock, children will call it “a clock.” In Chapters 6 and 7, I
review the evidence in support of this extension of essen-
tialism (see also Gelman & Bloom 2000).
7
As noted above, the proposal here is that word meanings
are just those concepts that happen to have word forms as-
sociated with them. Under this view, there is no “lexical se-
mantics” separate from conceptual structure. If so, then the
constraints and biases that hold for word meanings should
not be lexically specific; they should instead follow from
pragmatic knowledge, from cognitive and perceptual bi-
ases, and from properties of concepts themselves. In
HCLMW, I suggest that two fundamental biases of word
learning follow in a direct fashion from how concepts are
understood.
The first is the bias to treat new words as referring to basic-
level categories. Brown (1958a) noted that some names are
more frequently used than others when talking about ob-
jects. For instance, we usually call Fido “a dog,” not “an an-
imal” or “a terrier.” And when children learn new words,
they most naturally treat them as falling into this interme-
diate level of abstraction, not too general (animal), and not
too specific (terrier). What makes these basic-level mean-
ings so special?
Brown (1958a) speculated that we describe things at the
basic-level, “so as to categorize them in a maximally useful
way” (p. 20). Subsequent work has led to more explicit for-
mulations of this insight. One promising analysis is from
Murphy and Lassaline (1997), who propose that the basic
level is an optimal compromise between informativeness
and distinctiveness: you can infer many unobserved prop-
erties once you know which basic-level category something
belongs to, and it is also relatively easy to make this catego-
rization. Hence, the focus on the basic-level reflects a con-
ceptual bias toward carving the world into inductively rich
categories; this bias, then, has its origin in conceptual struc-
ture (Markman 1989; but see Mandler & McDonough
1996, for a different perspective).
A second example is the shape bias: Many investigators
have found that when given a new count noun that refers to
a rigid object, children will typically extend that noun to
other rigid objects of the same shape, not those of the same
size, color, or texture.
8
It has been argued that this bias is
special to word learning, and exists because children note
the correlation between words used in a certain context and
words that are extended on the basis of shape (e.g., Landau
et al. 1988). The alternative, defended in HCLMW, is that
shape is important for object names because shape is im-
portant for object categories: if two objects are the same
shape, they are likely to belong to the same category.
If this is correct, then people should categorize the world
into shape-based categories even when not learning words.
When a child first sees a dog, for instance, she would be
more likely to attribute properties of that dog to other ob-
jects of the same shape than to other objects of the same
size, color, or texture. It should also follow that the shape
should be sensitive to top-down information. (For instance,
if two objects are the same shape, but this sameness of
shape is clearly due to some factor independent of category
membership, then children should not extend a new word
from one object to the other.) Much of Chapter 6 defends
the view that children’s generalization of words on the ba-
sis of shape emerges from the essentialist nature of con-
cepts, and is not special to the learning of words.
One final point about concepts: It is sometimes sug-
gested that the conceptual system is undifferentiated prior
to language learning, and that many concepts emerge
through exposure to words. As I review in Chapter 10, there
is surprisingly little support for this Whorfian claim, and
plenty of evidence for the alternative – much of word learn-
ing involves mapping words onto pre-existing concepts.
This is not to deny the banal fact that exposure to language
can affect our thoughts (if not, why would you be reading
this?); nor is it to deny that concepts might change over the
course of development, sometimes radically. But words are,
in the end, a tool for the expression of thought, not a tool
for creating it.
9. How does this theory differ from alternatives?
Make everything as simple as possible, but no simpler.
– Albert Einstein
HCLMW is largely a synthesis of theory and research done
by many people, and so I can be confident that at least some
Bloom: Précis of How Children Learn the Meanings of Words
1102 BEHAVIORAL AND BRAIN SCIENCES (2001) 24:6
people will agree with at least some of what I say. But the
broader conclusions that I draw are not consistent with the
dominant theories in the field. As I have discussed above,
many researchers believe that word learning, at least early
on, can be explained in simpler terms. Simple associative
mechanisms suffice. Children who hear the word “bottle,”
for instance, learn the word through foundational mecha-
nisms of learning that associate this word with what is per-
ceived at the time that it is used. Other scholars argue that
the capacities I have discussed above are not enough. In-
stead, children must possess special constraints, assump-
tions, or biases that are earmarked for word learning. Chil-
dren who hear the word “bottle” are guided by a constraint
on word meaning that leads them to assume that the word
refers to the whole bottle, and another constraint that
guides them to extend the word to other objects of the same
kind.
I discuss these proposals – the associationist view, and
the special constraints view – in some detail in HCLMW,
and I argue that they are mistaken. My arguments are not
abstract; I am not worried that the associationist approach
is too empiricist, or that the constraints approach is too na-
tivist. The problem with these views is that they fail to ade-
quately explain certain facts about the nature and course of
word learning.
What about other theories? The proposal in HCLMW
is consistent with the spirit of “cognitive,” “syntactic,” and
“socio-pragmatic” theories; disagreements lie in the relative
emphasis that is given to these different processes, and the
specific claims as to their nature. I have nothing critical to
say about connectionist theories of word learning except in-
sofar as such theories assume that children learn words
through associative learning, without theory of mind. (Un-
fortunately, all existing connectionist theories make this as-
sumption.) Finally, there is some recent excitement over
the claim that children exploit multiple converging cues in
the course of word learning. I agree this “multiple cues” ap-
proach is the right one; but I don’t know if anyone has ever
doubted it. Current theories along these lines are interest-
ing not because they insist on multiple cues, but because
they make substantive claims about the precise nature of
the cues and how they interact (e.g., Golinkoff et al. 1994).
HCLMW ends with a discussion of what special con-
straint theories and associationist theories have in common.
They share the view that word learning proceeds in a re-
flexive manner, through “dumb associative mechanisms” or
through the operation of encapsulated and specialized prin-
ciples. In HCLMW, I defend an alternative that could best
be called “Rationalist” – children learn words through the
exercise of reason. They figure out what people are intend-
ing to say when they use words, and they bring all of their
knowledge to bear when figuring out how a word should be
understood. Word learning sometimes looks automatic, but
only because children quickly become very good at this sort
of reasoning. And it sometimes looks dumb, but only be-
cause, as researchers, we often put children in situations
where the only right answer is the dumb one.
A careful examination of the facts of word learning re-
veals that the mechanisms underlying this process are rich
and varied, and, at this point of time, largely mysterious.
This might not be good news from the standpoint of re-
search into this area, and this Rationalist account might be
more complicated than the alternatives, but it is nonethe-
less the proposal that is most consistent with how children
learn the meanings of words.
ACKNOWLEDGMENTS
I would like to thank Frank Keil, Gil Diesendruck, and Karen
Wynn for providing helpful comments on an earlier draft of this
précis.
NOTES
1. As with most parents and researchers, I am adopting what
Lois Bloom has called “rich interpretation,” and so I should admit
that Max did not actually say any of these words – instead of
“kitty,” he said “gid-duh”; instead of “belly-button,” he said “beh-
buh,” and so on.
2. Children do sometime use words like Mommy and Daddy to
refer to people other than their parents. But this is not actually an
error, since such words are also used as nouns, as in “my mommy”
or “John’s daddy.”
3. The term “theory of mind” is chosen to make explicit the
connection with the autism literature, but no commitment is made
as to the precise nature of this capacity (such as whether it really
is a theory). Also, it is sometimes said that children only have the-
ory of mind when they can pass the notorious false belief task. I
think this is mistaken (see Bloom & German 2000), but in any
case, this is not how I’m using the term here; children could have
“theory of mind” even if they could not reason about beliefs that
are false.
4. But not entirely successful – even high functioning autistic
individuals use and understand words in subtly different ways
from normal children, especially with regard to names for repre-
sentations (Bloom et al. 2002).
5. Some scholars would reject this semantic/syntactic distinc-
tion, arguing that these examples are better viewed as falling on
opposite ends of a continuum. Even if one accepts that these types
of context are distinct, it is difficult in practice to change the syn-
tax without changing the semantics, and so there is controversy
over the interpretation of certain experiments that purport to
show that children use syntactic cues, as opposed to other linguis-
tic information, to infer aspects of word meaning.
6. It also raises the question of how children learn the syntac-
tic categories of words. The solution adopted in HCLMW is a vari-
ant of “semantic bootstrapping” (Pinker 1984); children use the
meanings of some words, such as object names, to learn their syn-
tactic category, and to learn the syntactic rules of their language.
These rules enable the syntactic categorization of other words,
and at this point, syntactic cues to word meaning can apply (see
also Bloom 1999).
7. More speculatively, it might be that certain puzzling facts
about how people use words such as water are due to an artifact-
like understanding of the category that includes considerations of
design and intent (Bloom 2002a; see also Chomsky 1995; Malt
1994).
8. These two biases are related, because members of a basic-
level object kind will tend to have the same shape. But the shape
bias is narrower in scope than the basic-level bias (applying only
to rigid objects). Also, a child who had only a shape bias would
tend to favor subordinate categories (terrier, rocking chair) over
basic-level categories (dog, chair).
Bloom: Précis of How Children Learn the Meanings of Words
BEHAVIORAL AND BRAIN SCIENCES (2001) 24:6 1103
Open Peer Commentary
Commentary submitted by the qualified professional readership of this
journal will be considered for publication in a later issue as Continuing
Commentary on this article. Integrative overviews and syntheses are es-
pecially encouraged.
Okay for content words, but what about
functional items?
Derek Bickerton
Department of Linguistics, University of Hawaii, Honolulu, HI 96822.
derek@hawaii.edu
Abstract: Though Bloom makes a good case that learning content-word
meanings requires no task-specific apparatus, he does not seriously ad-
dress problems inherent in learning the meanings of functional items. Ev-
idence from creole languages suggests that the latter process presupposes
at least some task-specific mechanisms, perhaps including a list of the lim-
ited number of semantic distinctions that can be expressed via functional
items, as well as default systems that may operate in cases of impoverished
input.
Bloom’s book (HCLMW) is detailed, lucid, and laden with data
supporting its claim that though the meaning of content words
cannot be learned through Lockean associationism, the process
requires no specially-dedicated apparatus, but rather a combina-
tion of strategies. However, functional items are words too, and
here the issue is less clear-cut. The comments that follow should
be interpreted less as criticism of the book itself than as a sugges-
tion of further work that Bloom and others will have to undertake
before we have a complete theory of word learning.
Bloom has very little to say about how functional items are
learned. True, Chapter 5 is entitled “Pronouns and Proper
Names,” but seventeen pages are devoted to “Proper Names” and
only four and a half to “Pronouns.” If difficulty of the problems
posed for acquisition was the criterion here, those proportions
would be reversed. But discussion is confined to first and second
person pronouns; third-person pronouns, acquired notoriously
later than these, are not mentioned. Other functional items are
treated in an equally cursory manner, if at all. For instance, the
only mention of prepositions (pp. 204–205) involves their possi-
ble function in learning differences between noun classes; there
is no discussion of how prepositions are learned in the first place.
As for modals, quantifiers, complementizers, and question words,
these are not mentioned anywhere.
It’s not that Bloom is unaware that there is a problem; he seems
not to know what the problem is. Discussing noun-determiner
combinations (pp. 115–17), he notes that determiners may not be
heard, or, if they are, that the phonological salience of the noun
will distinguish it from them. Such factors, he claims “explain why
it is harder to learn determiners than nouns” (p. 117). Actually,
they do not explain, nor does the fact that “Nobody ever points and
says ‘The!’ or ‘Of!’” (p. 115). What really makes meanings of de-
terminers and other functional items, bound or unbound, hard to
learn is that there is a very large, if not infinite, number of se-
mantic distinctions that they might in principle encode. In fact,
the number of semantic distinctions that can be encoded, across
the world’s languages, by functional items (call these Potentially
Encodable Distinctions, or PEDs) is finite and probably quite
small. One of my dearest wishes, uttered every time I see a grad-
uate student (but so far to no avail), is that someone would just
count and list PEDs. Though strictly a no-brainer, it might seem
at first sight a daunting task, but I would bet that by language No.
20 you would be getting diminishing returns, and by language No.
50 you would have the whole set or pretty close.
How do children learn the meanings of “a,” “the,” and other
functional items? Bloom’s remark that they “attend to NPs such as
‘the dog’ and ‘the cups’ and figure out the semantic role of the de-
terminer across all of these phrases” (p. 116) renames the prob-
lem rather than solving it. What stops them from concluding that
“the” and “a” embody a distinction between “rare” and “common,”
“pleasant” and “unpleasant,” or any other pair of meanings? But
Maratsos (1974) and much subsequent work shows that the mean-
ings of definite and indefinite articles are mastered early in ac-
quisition.
One possibility that Bloom probably won’t like is that PEDs are
innately specified in some way. The task would then be greatly
simplified; instead of having to “figure out” semantic distinctions
from a potentially infinite set, the child would simply select, from
a smorgasbord provided by universal grammar, the particular set
of PEDs that the target language had chosen. The task would be
made easier still by the fact that PEDs would be at least partly pre-
sorted according to the grammatical categories of the lexical heads
that each functional category modified. For instance, no child
would expect a specific-nonspecific distinction to be marked by a
modal or other verb modifier, or a realis-irrealis distinction (i.e.,
the distinction between events that have actually happened and
those that have not) to be marked by a determiner or other noun
modifier.
There is some empirical evidence in favor of such a proposal.
Creole languages derive from pidgins that have very few func-
tional items (see Roberts 1998 for historic records of the kind of
pidgin input that the first creole-speaking generation in Hawaii re-
ceived). Despite this, a wide range of functional items is created
by creoles, most probably by first-generation children (Bickerton
1984; 1999). These include definite and indefinite markers that
are both morphophonemically and semantically distinct from any
equivalent items in the creole’s substrate or superstrate languages.
Simplifying slightly for reasons of space, many unrelated creoles
have the following system: an overt article marking specific noun-
phrases already established in discourse (derived usually from a
superstrate demonstrative, “this” or “that” in English-influenced
creoles), an overt article marking specific noun-phrases not yet es-
tablished in discourse (derived almost always from the superstrate
numeral “one”), and a zero form marking nonspecific (generic,
undifferentiated or hypothetical) noun phrases, as well as mass
nouns. In English, we have to say either “I am going to buy shirts”
or “I am going to buy a shirt,” even though we may not know
whether we will buy one, or several, or none; in an English creole,
such as Guyanese, we would say “mi go bai shirt.” The only En-
glish equivalent is the cumbersome (and unidiomatic) “I am go-
ing shirt-buying.” Moreover, in contrast to English, a plural is im-
possible with a bare (determiner-less) noun, even when this is a
count noun: “di shirt-dem’, “the shirts,” but never “shirt-dem,”
“shirts.”
The foregoing suggests that there is a “default” system of de-
terminers in Universal Grammar (UG), and that children learning
English pick up articles quickly because the determiner system in
English, although different from the creole prototype, simply
reshuffles and re-labels the handful of PEDs used by the default
system. The obvious question then is, why do not children learn-
ing some article-less language (such as Japanese) try to impose ar-
ticles on them?
There is more than one possible answer. First, perhaps they do
try – to the best of my knowledge, no one has reviewed Japanese
acquisition data to find out – but give up as soon as they see that
Japanese employs a different set of PEDs. Second, perhaps more
likely, children can tell the difference between a true language
(with rich data and plentiful instantiations of PEDs) and a pidgin
(with severely impoverished data and few, if any, instantiations of
PEDs). In the former case, they keep looking in the input for
things that might mark PEDs; in the latter, they realize that this
would be a waste of time, and go with the default system.
Neither of these answers may be correct, there may not even be
a default system, but at least the proposal has more substance than
anything Bloom offers for functional items. It is significant that his
Commentary/Bloom: How Children Learn the Meanings of Words
1104 BEHAVIORAL AND BRAIN SCIENCES (2001) 24:6
book contains no mention of the acquisition problems faced by the
first generation of creole speakers (Bickerton 1999) or, even more
critically, by the creators of new sign languages (Kegl et al. 1999).
If he had looked at these, he might not, inter alia, have stated so
confidently that “syntax is an important informational source as to
the meanings of words” (p. 212). Like motherese, syntax is a handy
tool to have, but you can get along fine without it.
But Bloom throughout covers only “normal” acquisition cases,
where input is rich and well-formed. He should remember that if
hard cases make bad law, normal cases make bad science. Unlike
our colleagues in the physical sciences, we in the behavioral sci-
ences cannot put nature through a wringer with our experiments,
using them to create not just unusual cases but circumstances that
might never arise in the natural world. All the more reason for us
to exploit to the full the few “natural experiments,” flawed, in-
complete, and subject to interpretation though they be, that na-
ture has afforded us.
Concept modeling, essential properties, and
similarity spaces
Peter Gärdenfors
Lund University Cognitive Science, Kunghuset, S-222 22 Lund, Sweden.
Peter.Gardenfors@lucs.lu.se http://www.lucs.lu.se
Abstract: Bloom argues that concepts depend on psychological essential-
ism. He rejects the proposal that concepts are based on perceptual simi-
larity spaces because it cannot account for how we handle new properties
and does not fit with our intuitions about essences. I argue that by using a
broader notion of similarity space, it is possible to explain these features
of concepts.
In Chapter 6 of HCLMW, Bloom discusses the concepts expressed
by common nouns. He argues that such concepts are based on a
form of psychological essentialism (Medin 1989; Medin & Ortony
1989). As an alternative account he considers the “minimalist pro-
posal” that concepts are based on perceptual similarity spaces. He
rejects this account on two grounds: it cannot explain how we han-
dle new properties, and it does not fit with our intuitions about
essences.
Bloom seems to take for granted what is meant by “perceptual
similarity.” However, I believe that there is no purely perceptual
notion of similarity. I want to argue that by using a broader notion
of a similarity space, what I call a “conceptual space,” it is possible
to account for new properties and explain why our concepts seem
to be based on some form of essences.
Bloom’s first argument against the reduction of concepts to per-
ceptual similarity spaces is that abstract concepts like “stockbro-
ker,” “menu,” and “swearing” cannot be analysed in terms of per-
ceptual dimensions alone. I basically agree with this statement,
but it must be qualified. Firstly, it is not clear what exactly Bloom
means by “perceptual dimension” or “perceptual similarity.” It has
been a long time since empiricist philosophers believed that there
is a well-defined perceptual level of describing the world. There
is no such thing as a naked eye. As Popper argued, even a seem-
ingly innocent empirical statement like “This is a glass of water” is
loaded with theoretical (nonperceptual) assumptions concerning
the meaning of “glass” and “water.”
Secondly, and more importantly, our judgments of similarity are
seldom based on perceptual similarity alone. When we assess the
similarity of different phenomena, we take “hidden” factors into
account, such as forces and other causal variables. Also, the func-
tions of objects partly determine our similarity assessments. By
limiting his argument to perceptual similarity spaces, Bloom im-
poses a straitjacket on the role of similarity.
Bloom’s second argument concerns the role of “psychological”
essentialism in the formation of concepts. In modern psychologi-
cal theories (e.g., Osherson & Smith 1981; Rips 1995), “core”
properties are supposed to determine the meaning of concepts
while other “peripheral” properties (Kövescezs 1993) only have
diagnostic value. The core consists of those properties that are es-
sential to a concept, whereas the peripheral properties do not de-
termine whether something belongs to the concept or not, –
though they may be helpful in “identification procedures” (Osh-
erson & Smith 1981). Sometimes, the distinction is formulated in
terms of perception versus conception: perception delivers the
peripheral properties, which may serve as cues for conception
where the essential properties are represented.
On the other hand, essentialist theories have been criticized as
involving unnecessary metaphysical assumptions. A way out of the
conflict is provided by Medin (1989, p. 1477), who gives the fol-
lowing explanation of why people have a tendency to behave as if
psychological essentialism is valid:
If psychological essentialism is bad metaphysics, why should people act
as if things had essences? The reason is that it may prove to be good
epistemology. One could say that people adopt an essentialist heuristic,
namely, the hypothesis that things that look alike tend to share deeper
properties (similarities). Our perceptual and conceptual systems appear
to have evolved such that the essentialist heuristic is very often cor-
rect.
Bloom wants to explain how we can understand abstract concepts
by applying a “theory-theory.” This means that concepts should be
thought of as embedded in knowledge that contains theories of the
world (Murphy & Medin 1985). The “theory-theory” is closely re-
lated to theories of psychological essentialism: the theories express
essential properties of the concepts.
Following the ideas in my recent book on conceptual spaces
(Gärdenfors 2000), I propose that by exploiting similarity spaces
that involve non-perceptual (theoretical) dimensions, one can ac-
count for the “essentialism lite” that Bloom (pp. 167– 69) presents
as a part of his analysis of concepts. (For a similar position, see
Hampton 2001.)
A key idea is that the role core properties play in essentialist the-
ories of concepts can be handled by considering the salience of
different dimensions. A “core” (essential) property of a concept is
a property that belongs to a dimension with high salience, while a
“peripheral” property is associated with a dimension with lower
salience. Radical essentialism, then, corresponds to assigning ex-
treme salience to some “essential” dimensions when determining
the content of a concept. Furthermore, the salience weights of dif-
ferent dimensions vary, depending on the context. Hence, what
seems like a core property in one context may seem peripheral in
another. The problems philosophers and psychologists have had
in identifying the essential properties of concepts is a symptom of
the fact that there is no sharp border between core and peripheral
properties.
Once some dimensions other than the basic perceptual dimen-
sions are given higher salience weights, then there can be no sim-
ple perceptual account of similarity. For example, folk botany may
classify plants according to the color or shape of the flowers and
leaves, but after Linnaeus the number of pistils and stamens be-
came the most important dimensions. And these dimensions are
perceptually much less salient than the color or shape dimensions.
The general trend in the development of the concepts of a domain
is towards less and less dependence on perceptual similarity.
Most of the role that the theories are supposed to play in rep-
resentations of concepts can be taken over by those dimensions
that have the highest salience. In many cases these dimensions
may not be perceptual, but indeed correspond to what is con-
ceived of as theoretical entities in science. (Medin & Ortony
[1989] suggest that people’s beliefs in ultimate causes of category
membership are sometime so strong that such a causal variable
can function as an unidentified placeholder in the core of a con-
cept.) From this perspective, there is no fundamental conflict be-
tween a similarity-based account and a theory-theory based ac-
count. The upshot is that in theories too, there is similarity behind
the scene. Hahn and Chater (1997, p. 50) conclude: “Thus theory-
Commentary/Bloom: How Children Learn the Meanings of Words
BEHAVIORAL AND BRAIN SCIENCES (2001) 24:6 1105
based views demand a better account of similarity, rather than no
account of similarity in explaining concepts.”
Word meaning, cognitive development, and
social interaction
Alison F. Garton
School of Psychology, Edith Cowan University, Joondalup, Western Australia
6027, Australia.
a.garton@cowan.edu.au http://www.cowan.edu.au
Abstract: This review proposes that Bloom’s linkage of word meaning with
more general cognitive capacities could be extended through examination
of the social contexts in which children learn. Specifically, the child’s de-
veloping theory of mind can be viewed as part of the process by which chil-
dren learn word meanings through engagement in social interactions that
facilitate both language and strategic behaviours.
The publication of this book restores the study of word meaning
to a central place in the study of language, especially since it links
the learning of word meaning to more general emerging cognitive
abilities rather than postulating a nativist or an empiricist theory.
While the comprehensiveness of Bloom’s coverage cannot be
doubted, it is the exclusion of what is an important theoretical
stance that makes the book appear lop-sided. Pitting Bloom’s view
that word learning is a manifestation of other more general cog-
nitive abilities against a nativist (bias or constraint) view is akin to
Tomasello’s (2001) polarisation of the nativist view and the social
pragmatic approaches.
An argument that does not dichotomise the theoretical posi-
tions can be mounted (see also Baldwin & Moses 2001). Social in-
teraction has a fundamental role in the development of knowl-
edge, in particular, knowledge of symbolic systems such as spoken
language (including the meaning of words) and written language.
Vygotsky (1962) specified that social interaction created language
and that the integration of different tool (practical) and sign (sym-
bolic) systems by the child gave rise to new and increasingly so-
phisticated behaviour. This view highlights both the social and the
constructivist nature of the child’s developing knowledge.
Bruner (1983) developed the notion that language is best en-
couraged and developed in a functional way through social inter-
action. Bruner sought to demonstrate how children’s communica-
tive intentions were guided and enhanced through interaction
with a parent so that the product was not only an awareness of the
functions of language, but also the use of language to communi-
cate with others. Bruner examined various predictable and highly
structured everyday events, formats, to show how these could give
a child a framework within which to learn not just the meanings
of words but also the uses of language.
In research on the benefits of collaborative problem solving on
children’s cognitive abilities (e.g., Garton & Pratt 2001), if both
participants share a common conception of a problem and how to
solve it, then the degree to which the two participants can work
successfully on the task is enhanced. Such a shared task perspec-
tive can be achieved by talking. Social regulation via communica-
tion about role division and allocation, as well as planning and ex-
ecuting the task, facilitates problem solving particularly in less
competent children.
This facilitation has been linked theoretically to the importance
for children of being able to reflect on knowledge. Children’s the-
ory of mind is concerned with developing understanding of the na-
ture of knowledge and refers to the ability to understand that oth-
ers too know things, have beliefs and can think, based on
knowledge which may be true or false. One crucial means to such
understanding is through communication as evidenced in collab-
orative problem solving. However, it may be that communication
can only be successful if there is an existing propensity, or aware-
ness, in children to recognise the importance of the strategic
knowledge of the other partner in the interaction. Most research
neglects this aspect and fails to take account of the extent to which
children explicitly acknowledge and discuss any differences in lev-
els of understanding either before or during social interaction
A possible mechanism to draw together the research on com-
munication and awareness of the other in interaction, and the way
the meanings of words might be learned, is the child’s developing
theory of mind. Word meaning may indeed develop through more
general cognitive capacities or through the operation of biases or
constraints, but a promising lead is noted in the inclusion by
Bloom of the child’s capacity to infer the intentions of another as
a cognitive ability. This allows for the development of an argument
that links cognitive capacity, word learning, and theory of mind via
social interaction.
The proposal is that word learning occurs in, is facilitated by,
or perhaps is constrained by (Tomasello 2001), social contexts.
Specifically, in the case of young children learning language, these
contexts are highly predictable, regular, and structured and not
only involve language but other non-verbal gestures (such as
pointing and gazes) that provide support for the learning of what-
ever problem is the focus – be it language, jigsaw puzzles, or num-
ber. In each case, the participants, particularly the adult, often the
mother, ensures the child takes part and adequately (in her view)
responds to her encouragement.
In addition, in the case of learning the meanings of words, there
would need to be an understanding that the child is responding to
the aspects of the environment or problem to which the adult is
drawing attention and to which the child in turn should attend. In
general, the child, the less competent participant, sets the level of
skill or expertise and the pace for instruction and learning. The
adult, the more experienced participant, gauges the pre-existing
skills and the need for instruction and/or assistance, and then di-
vides the task or problem into manageable components. In this
way, the adult takes responsibility for the management and mon-
itoring of the interaction and for changing the definition of the
task by the child. Such interactions have been demonstrated to en-
hance cognitive functioning in young children (Garton et al., sub-
mitted) as well as being linked to language learning, including
word meaning (Garton & Pratt 1998).
Finally, it is worth noting that far from being a passive recipient
in such social interactions, the child is, through the communica-
tive and non-verbal cues, trying to make sense of what the adult
intends. In this, there is agreement with Bloom’s contention that
cognitive abilities play a role, but I argue that, more importantly,
the role of the other in social interaction contributes to the mani-
festation and deployment of these abilities in a supportive context
which itself then enhances language and cognitive development.
In accordance with Bloom’s contention, why should the learning
of word meanings be any different to other cognitive capacities
that are enhanced by social interaction?
Don’t preverbal infants map words onto
referents?
Lakshmi J. Gogate
Department of Psychiatry, SUNY Health Science Center at Brooklyn,
Brooklyn, NY 11203.
Lgogate@netmail.hscbklyn.edu
Abstract: Bloom provides a detailed account of children’s word learning
and comprehension. Yet, this book falls short of explaining the develop-
mental process of word learning. The studies reviewed do not explain how
infants begin to map words onto objects or the environment’s facilitative
role. Researchers must describe how several factors interact and explain
the relative importance of each during the development of word learning.
Bloom provides a wealth of detail on the complex issue of chil-
dren’s word learning and comprehension. Several aspects of this
book are very appealing. He presents a compelling rationale for
children’s learning of nouns before other lexical categories. Also
Commentary/Bloom: How Children Learn the Meanings of Words
1106 BEHAVIORAL AND BRAIN SCIENCES (2001) 24:6
important is that children learn the meanings of most lexical words
before grammatical words. In light of recent evidence, that new-
borns (Shi et al. 1999) and 6-month-olds (Shi & Werker 2001)
perceptually prefer lexical over grammar words, the present dis-
cussion is timely. Yet, this book falls short of explaining the devel-
opmental process of word learning.
The author acknowledges the different word learning capabili-
ties in children during development. “Prelinguistic infants are in-
teresting because they lack whatever capacities are necessary to
start talking, one-year-olds are interesting because they are word
learning novices,...and preschoolers are interesting because
they represent an illustrative midpoint between novice and ex-
pert” (p. 13). Therefore, one expects but finds no developmental
explanation of how prelinguistic or preverbal infants begin to learn
word meanings. Preverbal infants are interesting subjects of study
mainly because they can map word-like forms onto referents long
before they can produce them (Fenson et al. 1994). Anecdotal ev-
idence for mapping of speech onto objects is reported in 9-month-
olds’ use of idiosyncratic expressions that have a consistent mean-
ing (Blake & Fink 1987; Halliday 1975b). A discussion of how
preverbal infants begin to map words onto concepts is quintes-
sential to our understanding of how children might learn the
meanings of words. Yet, Bloom does not discuss a significant re-
cent body of experimental research on the development of early
word mapping in infants (Gogate & Bahrick 1998; Moore et al.
1999; Stager & Werker 1997; Tincoff & Jusczyk 1999; Werker et
al. 1998). The focus is mainly on empirical studies of word learn-
ing starting with two-year-olds, innate, intuitive, and essentialist
theories of children’s word and concept acquisition, and philo-
sophical theories of adults’ word learning. Children of two years
of age and up are adept at learning and understanding words.
These studies do not reveal how, during the first year, infants
might begin to map words onto objects. Focusing on the end-state
deters researchers from investigating the origins and the develop-
mental process of early word learning.
Several studies have attempted to explain how preverbal infants
learn to map novel utterances onto novel objects. Gogate and
Bahrick (1998) demonstrated that 7-month-olds learn the arbi-
trary relations between two vocalic syllables and two distinct ob-
jects only when the syllables are spoken in temporal synchrony
with the motions of objects. This finding was supported when 8-
month-olds did not learn the arbitrary relations between two mov-
ing objects and two word-like (CVC) syllables in the absence of
temporal synchrony (Werker et al. 1998). Fourteen-month-olds in
this study, however, mapped the words onto moving but not static
objects, suggesting that by this age infants still require dynamic
object presentations to learn word-object relations (also see
Moore et al. 1999). Preverbal infants can, however, map familiar
words onto objects in the absence of temporal synchrony. Six-
month-olds mapped the words daddy and mommy onto moving
but not temporally synchronous faces of their father and mother
(Tincoff & Jusczyk 1999). These findings support the following
general hypotheses about the process of infants’ learning of names
for objects. First, they suggest that during the development of lex-
ical comprehension, infants learn the “goes with” relation between
a word and its referent prior to learning that a word “stands for” a
referent (Golinkoff et al. 1994). Second, when infants first learn
novel word-object relations, they rely heavily on perceptual cues
such as intersensory redundancy in the form of temporal syn-
chrony between vocalizations and moving objects (7 and 8
months, Gogate & Bahrick 1998; Gogate, under review) and syl-
lable distinctiveness (Stager & Werker 1997; Werker et al. 1998).
Later, when infants are more adept at learning word-object rela-
tions, they do not need temporal synchrony, but still require ob-
ject motion (14-months, Werker et al. 1998). Still later, during the
second year, infants learn word-object relations under less obvi-
ous conditions, in the absence of object motion (Moore et al.
1999). They rely on social criteria such as adults’ eye gaze at the
referent (Baldwin et al. 1996) and referential intent (Tomasello &
Barton 1994) to learn the names for referents.
The author undermines the facilitative role of the environment,
in particular maternal communication, in infants’ early word map-
ping. He remarks that aspects of early word learning cannot be at-
tributed to parent’s or other’s interactions with infants (p. 45). But
cross-cultural findings show that mothers use ostensive naming
contexts to teach the names of objects and actions to their pre-
verbal and verbal infants (Messer 1978; Zukow-Goldring 1990;
1997). Notwithstanding the cultural variations in the use of os-
tensive naming (Schieffelin 1979), it sets the stage for infants’
word mapping early on. During ostensive naming, mothers use
temporal synchrony between the uttered word and the hand held
motions of an object. Recently, Gogate et al. (2000) provided evi-
dence for the dynamic nature of maternal multisensory naming of
objects and actions. Mothers of preverbal infants (5–8 mos.) of-
ten named objects and actions in temporal synchrony with the mo-
tions of objects, and simultaneously touched the infant with the
object. Mothers of lexically advanced infants (21– 30 mos.) used
temporal synchrony less often. They named the objects and ac-
tions while holding an object static. Although ostensive naming
alone cannot explain all of word learning, the changing nature of
maternal naming and the changing perceptual needs of the de-
veloping infant suggest an ongoing bi-directional relation between
the environment and the organism (Gogate et al. 2001). The crit-
ical issue is not whether one factor or another (e.g., maternal os-
tensive naming or infant’s understanding of others’ intentions)
best explains how words are learned (p. 192). Rather, researchers
must accurately describe how several factors interact at different
points in time and explain the relative importance of each during
the development of word learning in infants and children.
Innateness, abstract names, and syntactic
cues in
How Children Learn the Meanings
of Words
Heidi Harley
a
and Massimo Piattelli-Palmarini
b
a
Department of Linguistics, Douglass 200E, University of Arizona, Tucson,
AZ 85721-0028;
b
Cognitive Science Program, Communication 305,
University of Arizona, Tucson, AZ 85721.
hharley@u.arizona.edu
http://w3.arizona.edu/~ling/hh Massimo@u.arizona.edu
Abstract: Bloom masterfully captures the state-of-the-art in the study of
lexical acquisition. He also exposes the extent of our ignorance about the
learning of names for non-observables. HCLMW adopts an innatist posi-
tion without adopting modularity of mind; however, it seems likely that
modularity is needed to bridge the gap between object names and the rest
of the lexicon.
HCLMW is a tremendously useful and interesting book because
it tells us what is known about how children learn the meaning of
words, and, by extension, what isn’t. Bloom presents a thorough,
up-to-the-minute millennial overview of the state of the art, inte-
grated into an intellectually satisfying and elegant Big Picture. The
Big Picture masterfully painted by Bloom includes some familiar
elements and some surprising ones. Some elements of the story
that seemed familiar already, thanks to Bloom’s previous publica-
tions, included (a) the notion that children need a theory of mind
to learn words, (b) they integrate clues from many different do-
mains (including syntactic frames) to understand a novel word, (c)
the concepts that we tend to choose to name with words are those
that are inductively useful, and (d) word learning – especially
“fast-mapping” – is not accomplished associatively: word learning
is not statistical learning.
Surprising elements included the following: (a) Young children
are not especially good at learning the meanings of words – older
children and adults are better at it. (b) Learning the meanings of
words is not qualitatively different from learning facts about the
world, for instance, that a particular widget was given to someone
by their uncle. (c) We can precisely characterize the way in which
Commentary/Bloom: How Children Learn the Meanings of Words
BEHAVIORAL AND BRAIN SCIENCES (2001) 24:6 1107
a human’s “theory of mind” differs from that of other animals.
Other animals may understand goal-directed behavior, but they
do not understand the particular kind of goal-directed behavior
that is referential intent (HCLMW, p. 85, from Tomasello 1998b).
This last is an especially interesting claim, to which little space
is devoted. At first glance, it almost seems like begging the ques-
tion: children learn the meanings of words because children un-
derstand reference innately, which is what words are used for. The
reason that it’s not begging the question about word-learning per
se is that it is not only a nativist position, it is also a non-modular
one. We suggest that this jibes well with Bloom’s – in co-author-
ship with Steven Pinker in BBS – neo-Darwinian adaptationist ex-
planation of the emergence of language. Words are symbols par
excellence, but words are not the only things that can be symbols.
Bloom reports that a number of studies show that babies up to
about 18 months treat any communicative act, linguistic or non-
linguistic, as a potential Saussurean sign (pp. 77–78). Only after
18 months do children figure out that signs are typically only
phonological words. On the dust jacket of HCLMW, Gary Marcus
(an author with connectionist propensities) congratulates Bloom
for “steering between the extremes of nativism and empiricism.”
On this point, however, endorsing Tomasello’s view of referential
intent, Bloom is an unapologetic nativist, and rightly so. His na-
tivism, though, does not entail the existence of an innate language
faculty. Adaptationism is always at a loss in accounting for the
specificity of the language apparatus, leaning rather towards the
appeal to variants of general intelligence, “applied” to linguistic
communication. Bloom is by far the most linguistically sophisti-
cated of the adaptationists, but his proprietary dissociation of
strong innatism from language modularity and specificity, though
consistent with his neo-Darwinian positions, remains perplexing.
The available evidence from word learning appears to entail
Fodor’s Representational Theory of Mind. It also appears to ex-
clude Fodor’s unrepentant atomism, and the segregation of the
contents of the lexicon from all sorts of relevant knowledge about
the facts of life in general.
Because Bloom’s discussion is firmly grounded in experimental
reports, and also, at least implicitly, on his evolutionist-adapta-
tionist persuasions, most of the word-learning he discusses is the
paradigm case of object-naming. (As Guerts 2000 remarks, “How
Children Learn the Meanings of Nominal Expressions Denoting
Middle-Sized Observables” might have been a more accurate ti-
tle). He repeatedly emphasizes, however, that object-naming is in
fact not the paradigm case. Children’s vocabularies include many
non-object words, usually more than 50 per cent of their first
words (p. 91 – naps and parties and tomorrow are signal exam-
ples). The fact that object names are not the sine qua non of word
learning, combined with Bloom’s observation that older children
and adults are in fact the best word-learners, makes the reader
long for a whole new body of experimental literature dealing with
learning words for non-observables. It seems very likely that
nearly all the words that we learn over the course of a lifetime are
words for non-observables – it’s flat-out surprising to encounter a
new middle-sized observable as an adult. (When moving to Ari-
zona, it’s a shock to learn that there’s a middle-sized mammal you
have never even heard of: the coati.)
Bloom does his best to address this question, given the limited
literature, especially in his chapters on number words and syn-
tactic cues to word-learning. The evidence that children use syn-
tactic clues to detect word category, and that this leads them to
important conclusions about possible meanings, seems over-
whelming (both for object names, as in the English count/mass
distinction, and for non-nominal categories like verbs and adjec-
tives). However, it’s equally obvious that recognizing syntactic cat-
egory and consequently drawing broad conclusions about onto-
logical category (“this is a property,” “this is an action,” etc.), is a
far cry from really learning the meaning of a word. Bloom reports
some interesting debate about verb-learning using syntactic clues,
(pp. 202–203). For instance, Fisher et al. 1994 designed a verbal
wug-test where the action that was being named “daking” or “nad-
ing” was one that the children already knew a word for, like eating
or feeding. Looking at this experiment in light of Bloom’s earlier
discussion of the importance of the lexical contrast principle
(Ch. 2), it is particularly striking to learn that children did in fact
map these novel words to essentially the same concepts as those
denoted by the words they already knew. It seems possible that
lexical contrast operates in an interestingly different way in the
learning of non-object words.
Similarly, Bloom outlines the smallish body of literature on
learning adjectival and prepositional words. He notes, for exam-
ple, that the object-bias only applies to nouns; when exposed to
novel adjectives, children extended the words to superficial prop-
erties, not objects. Further, the shape-bias does not apply to ad-
jectives. One of the strange properties of adjectives is that, in set-
theoretic terms, they cut across “natural kinds” (a white cloud and
a white dog are both white); when used attributively, though, ad-
jectives pick out subkinds. Bloom suggests that children may be
sensitive to the difference between predication and modification.
A particularly interesting case to look at experimentally might be
that of relative clauses: the dog that is fep versus the dog is fep and
the fep dog – a modificational use of a predicative use of an ad-
jective. Similar questions arise with prepositional words, which re-
fer to actual or “metaphorical” spatial relations. Bloom reports a
study by Landau and Stecker 1990, comparing “This is a corp” to
“This is acorp the box,” where subjects concluded that a corp in
the first frame was an object, but a spatial relation in the second.
One wonders what subjects would have made of This is melting
acorp the box or The dog is eating this acorp the box – or The dog
acorp the box is panting. Finally, syntactic categories are famous
for not matching up cross-linguistically, or even intralinguistically
(Baker, in press; Hale & Keyser 1998). (What should we make of
the difference between Marie a chaud vs. Mary is hot ? or Mary
sings and Mary’s a singer?) Also, syntactic cues, as Bloom points
out, can only be used after a few closed-class terms have been
learned (p. 207), along with their syntactic function ( a, -s, is, -ing,
etc.). Yet, those closed-class items are among the hardest words to
learn (p. 116). There are clearly important things to be learned by
looking at children’s acquisition of such words, but it’s hardly been
touched, to judge from HCLMW.
Bloom’s book is a tour-de-force of synthesis and analysis. De-
cidedly, we think that, should one recommend one book, and one
only, on the acquisition of the lexicon, this would be it. However,
many central problems still remain open. Bloom has made a vir-
tuoso effort to cast a strongly mentalist and innatist theory of the
lexicon, stripped of modularity, and lexical atomism. If anyone can
succeed in doing that, then Bloom is the one. We have not been
stopped from wondering, however, if even he can do it.
Social attention need not equal social
intention: From attention to intention in early
word learning
Kathy Hirsh-Pasek,
a
Elizabeth Hennon,
a
Roberta M. Golinkoff,
b
Khara Pence,
b
Rachel Pulverman,
b
Jenny Sootsman,
b
Shannon Pruden
a
and Mandy Maguire
a
a
Department of Psychology, Temple University, Philadelphia, PA 19122;
b
Department of Educational Studies, University of Delaware, Newark, DE
19716.
Khirshpa@nimbus.temple.edu ehennon@temple.edu
{Roberta; kharap; rpulverm;jsoots}@udel.edu
shannimarie@yahoo.com mmaguire@astro.temple.edu
http://nimbus.temple.edu/~khirshpa/
Abstract: Bloom’s eloquent and comprehensive treatment of early word
learning holds that social intention is foundational for language develop-
ment. While we generally support his thesis, we call into question two of
his proposals: (1) that attention to social information in the environment
implies social intent, and (2) that infants are sensitive to social intent at the
very beginnings of word learning.
Commentary/Bloom: How Children Learn the Meanings of Words
1108 BEHAVIORAL AND BRAIN SCIENCES (2001) 24:6
Steve Pinker put it best when he described Paul Bloom’s book as
a “tour de force” on the subject of word learning. For centuries,
scholars pondered the nature of early word learning. In How Chil-
dren Learn the Meanings of Words (HCLMW), Bloom articulates
how this seemingly simple task “requires rich mental capacities –
conceptual, social and linguistic – that interact in complicated
ways” (Bloom 2000, p. 1). The book is a beautifully written, schol-
arly treatment of early language learning that draws on the au-
thor’s rich command of psychology, philosophy, and linguistics.
At the core of Bloom’s theory is a solution to the “mapping prob-
lem.” How do infants map the words that they hear onto the ob-
jects, actions, and events in their environment? How do infants
learn that the word “dog,” signifies the category of canines?
Philosophers, like Quine (1960), note that this is an intractable
problem. What allows a child to learn that the word “dog” refers
to the whole dog and not to dog parts? There have been many pro-
posed solutions for the indeterminacy of word mapping. Some
hold that infants operate with a set of constraints that limit the pos-
sibilities for word mapping (Golinkoff et al. 1994; Markman 1989).
Others argue that infants have no internal working hypotheses,
but rather associate the word heard with the most salient feature
in the environment (Plunkett 1997; Smith 2000).
Bloom aligns himself with yet a third alternative, social prag-
matics. Children are bathed in social environments, constantly in-
teracting with others. Infants, keenly aware of their social sur-
roundings, “read” the speaker’s intent. When a speaker says the
word “dog” while looking at the dog, the problem of reference fiz-
zles away. Attention to the speaker’s perspective narrows alterna-
tives for word mapping, allowing children to apprentice with mas-
ter word users (Baldwin & Tomasello 1998; Tomasello 1995b).
Bloom writes,
How do children make the connection between words and what they
refer to?... they do so through their understanding of the referential
intentions of others... children use their naïve psychology or theory of
mind to figure out what people are referring to when they use words.
(HCLMW, pp. 60–61, emphasis in the original)
Bloom contends that this ability to infer others’ intentions is avail-
able from the outset. Even by 6 months of age, infants follow an-
other’s gaze (Butterworth 1991; Carpenter et al. 1998). This fact,
Bloom suggests, is evidence that infants have a “broad” theory of
mind. It is this early ability to read intent that paves the way for
word learning:
it is impossible to explain how children learn the meaning of a word
without an understanding of... how children think about the minds of
others. (HCLMW, p. 2)
We generally agree with Bloom’s position. Speaker intent is a hall-
mark of word learning. We beg to differ, however, with two of
Bloom’s central tenets. First, attention to social information in the
environment need not imply social intent. Second, infants are not
sensitive to social intent at the very beginnings of word learning.
Social
attention
need not equal social intention.
Social prag-
matic theories regard attention to social cues as synonymous with
social intent. Yet, the two must be distinguished. Six-month-olds
who follow a speaker’s eye gaze are attracted to social cues, but
may not use them to infer speaker intent. Similarly, acts of joint
attention, central to early word learning, might be just that – two
people attending to an object. Indeed, in a joint attention task in-
fants fail to learn novel words when the task requires that they shift
attention from the object in their view to that which the mother
indicates (Dunham et al. 1993; Tomasello & Farrar 1986). They
are socially attentive to the mother, but do not use her intent as a
platform for word learning. It seems a stretch to assume that be-
cause social objects command infant attention, the infant is
blessed with the ability to attribute intent. Research in our labo-
ratories is exploring this distinction in children with autism. Re-
cent findings suggest that these children, who do learn some
words, are aware of attentional social information (Leekham et al.
2000) but cannot read social intent (Baron-Cohen 1995; Bloom
2000).
Infants are not initially sensitive to social intent: Development
proceeds from social attention to social intent.
Bloom contends
that infants are sensitive to social intent in word learning from the
outset. Evidence suggests otherwise. In one study, infants 10 to 24
months old heard novel words uttered in the presence of an unfa-
miliar interesting object and an unfamiliar boring one. Speakers
used eye gaze and pointing to indicate the intended referent. All
infants attended to the speaker’s gestures. Only 19- and 24-month-
olds, however, used speaker intent to map the word onto the bor-
ing object (Hollich et al. 2000). Ten-month-olds did what Bloom
argued “never happens” (p. 59). They mismapped – assuming that
the label referred to the most interesting object regardless of the
speaker’s intent (Hennon et al. 2001). Preliminary data suggests
that children with autism respond more like 10-month-olds than
like their older counterparts. These findings accord well with oth-
ers in the literature. Before 18 months of age, infants do not use
social intent in a word learning task (Baldwin & Tomasello 1998).
One way to reconcile these findings with Bloom’s is to suggest
that when infants begin to learn words, people serve as salient per-
ceptual objects that draw attention to, or highlight, word-to-world
mappings. At around 18 months of age – around the time of the
naming explosion – infants begin to note speaker intent and their
word learning strategies shift. The Emergentist Coalition Model
of word learning makes this prediction, offering a theoretical jus-
tification for the shift in word learning strategy (Hirsh-Pasek et al.
2000; Hollich et al. 2000). Under this scenario, children learn
some words or “arbitrary signs” (p. 17) without social intent, but
become more efficient word learners when social intent comes
on-line for language learning. The challenge for those of us in the
field becomes explaining how children who learn words like
“Fido” transform into those who learn words like “Fred.”
ACKNOWLEDGMENTS
This research was supported by NICHD grant Nos. 3U10HD25455-0552
to the first author, and NSF grants Nos. SBR9601306 and SBR961539 to
Hirsh-Pasek and Golinkoff. We would like to thank the following collabo-
rators for their input on this paper: Mandy Maguire, Shannon Pruden,
Khara Pence, Rachel Pulverman, Jenny Sootsman. Correspondence should
be sent to the first author: Kathy Hirsh-Pasek, Department of Psychology,
Temple University, 13
th
& Cecil B. Moore, Philadelphia, PA 19122.
Vocabulary and general intelligence
Arthur R. Jensen
School of Education, University of California, Berkeley, CA94720-1670.
nesnejanda@aol.com
Abstract: Acquisition of word meanings, or vocabulary, reflects general
mental ability (psychometric g) more than than do most abilities measured
in test batteries. Among diverse subtests, vocabulary is especially high on
indices of genetic influences. Bloom’s exposition of the psychological com-
plexities of understanding words, involving the primacy of concepts, the
theory of mind, and other processes, explains vocabulary’s predominant g
saturation.
The main message of Bloom’s extraordinarily detailed and prob-
ing analysis, which I found completely convincing, is the primacy
of conceptual thinking in the acquisition of vocabulary. The es-
sential direction of causality is concept to word rather than word
to concept, as is so commonly and mistakenly believed. The ac-
quisition process is as different psychologically from the stimulus-
response paradigm for paired-associate rote learning as one could
imagine. Vocabulary is acquired when words fill conceptual “slots”
that form in the course of mental development and seek to be
filled. The wide range of individual differences in vocabulary re-
flects differences in the number “slots” much more than differ-
ences in the amount of exposure to words. The causes of the avail-
able number of “slots” are still largely unknown but are certainly
related to chronological age and Spearman’s g factor. The proba-
Commentary/Bloom: How Children Learn the Meanings of Words
BEHAVIORAL AND BRAIN SCIENCES (2001) 24:6 1109
bility that an individual will know the meaning of a given word is
mainly a multiplicative function of that individual’s level of g and
the frequency of that word in the individual’s past experience.
I was first attracted to Bloom’s book by my prior interest in vo-
cabulary tests and in the fact that they are so strongly correlated
with IQ, while individual differences in the paired-associates
learning of words, nonsense syllables, or paralogs have such a rel-
atively weak correlation with IQ and even with vocabulary. Al-
though Bloom briefly alludes to this fact (p. 193 of HCLMW), I
think much more should have been made of it, as it not only sup-
ports his thesis but extends it to psychometric findings and to the
information processing theory of general intelligence. In combi-
nation with Bloom’s conclusions, the psychometric and behavior-
genetic facts about vocabulary should give pause to many psy-
chologists and educators.
For a starter, consider the following observations. Recall that a
factor in psychometrics is a source of variance (individual differ-
ences) common to a number of different tests; a test’s gfactor load-
ing is its correlation with the one factor that is common to all of a
number of diverse tests whose intercorrelations have been sub-
jected to a factor analysis. In the national standardization sample
of the Wechsler Intelligence Scales for Adults (WAIS), vocabulary
has the largest gloading (.87) among the eleven diverse subtests.
What may seem more surprising, however, is that when vocabulary
is factor analyzed among only the six completely nonverbal sub-
tests, its largest loading is only slightly lower (.82). Its loading
among just the verbal subtests is .92. Vocabulary therefore reflects
g much more than it reflects verbal ability residualized from g.
Bloom emphasizes, rightly, that the “theory of mind” plays an
important part in the child’s acquisition of word meanings. This
leads one to predict that, among nonverbal tests, vocabulary
should have its highest correlation with Picture Arrangement
(PA). Wechsler (1944, p. 88) described PA as a form of social in-
telligence, involving human, practical situations, and inferring
what the cartoon characters in a disarranged series of pictures are
trying to do. The correlation of PA with vocabulary is .65, a value
that is entirely predicted from these two tests’ gloadings. The
same results are found at every age level in the Wechsler Intelli-
gence Scale for Children (both in the American and Japanese ver-
sions) and in the British Intelligence Scales. A factor analysis of a
much larger number of diverse subtests, performed to examine
Carroll’s (in press) 3-stratum model of the factor structure of cog-
nitive abilities, allows a detailed analysis of Oral Vocabulary, show-
ing its percentages of variance in each of the three strata of an or-
thogonalized hierarchical factor analysis: (1) Verbal Ability (V),
(2) crystallized intelligence (Gc), and (3) general intelligence (g).
The test’s specificity and measurement error are the residual
(Res), that is, components of variance not common to other tests
in the battery. The averaged results of two batteries (of 29 and 16
subtests) are: g 561%, Gc 516%, V 58%, R 515%. A similar
hierarchical analysis of another large battery containing many
nonverbal tests (Gustafsson 1988) shows the factor composition of
vocabulary as g555%, Gc 534%, Res 511%.
Although the acquisition of vocabulary naturally depends on ex-
posure to words in some meaningful context, such exposure in-
teracts strongly with innate biological factors, as indicated by the
high degree of heritability of vocabulary tests. On a test of vocab-
ulary, monozygotic (MZ) twins are more alike than dizygotic (DZ)
twins (Newman et al. 1937). Vocabulary scores (with age partialed
out) are correlated .86 for MZ twins and .56 for DZ twins; the
broad heritability of vocabulary therefore, is estimated as 2(.86 2
.56) 5.60. In this same data set, the heritability of Binet IQ is .50,
of height, .57.
As a result of natural selection, advantageous traits typically
show genetic dominance, a component of the trait’s broad heri-
tability. Dominance can be detected by the effect of inbreeding
on the trait, a quantitative phenomenon known as inbreeding de-
pression (ID), which is manifested in the offspring of parents who
are closely related genetically, such as siblings or cousins (Jensen
1978; 1983). ID is measured as the percentage of depression of
the trait among inbred offspring compared with the mean of an
outbred group (i.e., offspring of genetically unrelated parents)
that is matched (or statistically controlled) for parental socioeco-
nomic and educational variables. The largest study of the effects
of inbreeding (cousin matings) on children’s mental test scores
found that among the eleven subtests of the Wechsler Intelligence
Scale for Children, vocabulary had the highest index of ID
(11.45%) as compared with the average of ID of 6.58% for the
other ten subtests (Schull & Neel 1965). The inbreeding effect on
the various subtests was correlated about 1.80 with the subtests’
gloadings (Jensen 1983).
Bloom’s book describes, more thoroughly than any other analy-
sis I have read, the purely psychological processes crucially in-
volved in children’s acquisition of word meanings. This fascinating
array of behavioral phenomena and its social-environmental con-
text accounts for why measures of vocabulary reflect so much of
the brain’s power that is represented psychometrically by the en-
compassing, and still causally unfathomed, gfactor (Jensen 1998).
The aim of reductively understanding the causal mechanisms of
individual differences in vocabulary is essentially the same as that
of discovering the physical basis of g, its predominant latent trait.
Good intentions and bad words
Frank C. Keil
Department of Psychology, Yale University, New Haven, CT 06520.
frank.keil@yale.edu http://www.yale.edu/psychology/
Abstract: Bloom makes a strong case that word meaning acquisition does
not require a dedicated word learning system. This conclusion, however,
does not argue against a dedicated language acquisition system for syntax,
morphology, and aspects of semantics. Critical questions are raised as to
why word meaning should be so different from other aspects of language
in the course of acquisition.
The failure of attempts to explain or model the acquisition of lan-
guage in terms of an all-purpose general learning system has been
taken as supporting arguments that humans have a domain specific
language acquisition system, variously thought of as a mental organ
(Chomsky 1965), a module (Fodor 1983) or a specialized instinct
(Pinker1994b). Children seem to be innately predisposed to prefer
certain families of language structures over others, preferences that
apparently cannot be explained by general principles of association,
problem solving, or massive parallel processing. In addition, it has
been difficult to expand the scope of this capacity beyond language
to a somewhat larger domain such as hierarchically organized rela-
tions, temporal strings, or symbolic relations. Language in particu-
lar seems to be matched to specific mental faculties.
While this view remains convincing to many more than forty
years after Chomsky proposed it, there is far less consensus on
what aspects of language are innately determined and how that
determination takes place. Paul Bloom’s book offers an extraordi-
nary re-examination of the claim that part of the language faculty
involves a capacity to acquire the meanings of words. At first blush,
it seems that learning words is surely near the center of the fac-
ulty. Children acquire words quickly and effortlessly; and they
seem to rule out all sorts of alternative meanings that, on associa-
tive grounds, should be quite compelling. There seems to be a
Word Acquisition Device (WAD), full of constraints that narrow
down the extraordinary range of possible meanings that could be
mapped onto words. Indeed, the impossibility of divining from
scratch the meanings that another person attaches to their words
is more striking on the surface than comparable arguments for
syntax. The ease of laying out the indefinitely large set of logically
possible alternative meanings for a word makes inescapable the
conclusion that something must be profoundly limiting the child’s
conjectures (Quine 1960).
Bloom’s book convincingly shows that there is no WAD as such.
Fast mapping of words works for non-words equally well. There
Commentary/Bloom: How Children Learn the Meanings of Words
1110 BEHAVIORAL AND BRAIN SCIENCES (2001) 24:6
is no explosion of word learning that might grant it a special sta-
tus. Finally, putative constraints on possible word meanings de-
volve into constraints on concepts, on pragmatics, and on how we
apprehend objects and events. Especially impressive in this mix is
the role of intentions. Bloom martials an extraordinary range of
evidence showing how another module, for a theory of mind, con-
strains the inferences that a young child will make about possible
word meanings. Whether it be word learning difficulties in autism
or ways in which a toddler overrides powerful associative sugges-
tions, it matters greatly to the success of word learning that we see
word users as intentional agents. Bloom’s analysis of the role of in-
tentions illustrates how several putative constraints on word
meaning might arise instead from broader principles concerning
how we interpret the beliefs and desires of others in communica-
tive contexts. There are no bad words, just bad concepts and bad
pragmatics.
Given this dismantling of the WAD, it is natural to wonder if
Bloom’s book should be regarded as the nose of the camel that will
eventually displace all need for a Language Acquisition Device
(LAD). I suspect there are some who will make just such an in-
ference, and others so devoted to the LAD that they will see
Bloom’s book as a threat to their enterprise as well. In both cases,
the inferences are misguided. Word meanings differ from natural
language syntax, and for that matter, from phonology and mor-
phology, in profound ways that make the demise of the WAD
largely irrelevant to the status of the LAD. If we are to believe
Fodor (1983), word meanings are not modules in that they are
clearly cognitively penetrable. It is an open question whether
modules are the only places where domain specific constraints
should exist (they might well apply in other cases; Keil et al., in
press); but the modular nature of syntax and the non-modular na-
ture of word meanings strongly suggest that the two are not likely
to be part of the same acquisition system. Even if syntax some-
times helps the learning of words, that assistance does not make
syntax and word meaning part of the same learning system. Syn-
tax also helps one learn irony, even though no one thinks of irony
as part of the LAD.
As someone who was once completely swayed by the idea of
specific constraints on word meanings as necessary for their ac-
quisition, I do now feel twinges of anxiety about other domains
where great learning success with relative ease suggests domain
specific constraints. If theory of mind works so well for words, why
not for syntax, or even number? One lesson from Bloom’s book is
that poverty of the stimulus arguments on their own should not be
taken as compelling evidence for a particular mental faculty, which
is probably why such arguments are the most frequently attacked
parts of nativist accounts (e.g., Elman et al. 1996). Such argu-
ments, however, are still extremely useful to questions of acquisi-
tion. They pose a problem that must be addressed in one way or
another. Bloom’s analysis shows us that, for word meanings, the
problem can be solved by converging sets of constraints, none of
which are specifically devoted to word meanings. For syntax, no
comparable set of converging constraints has been plausibly pro-
posed. The difference may involve the modularity of syntax. Mod-
ules of cognition and perception are so self contained that their
acquisition follows from their internal structure in ways not seen
in other sorts of domains. A WAD may also be made unnecessary
because of the close correspondence between word meanings and
concepts where, contra Fodor, there may be rich constraints on
concepts’ internal structure that make unnecessary specific ones
on words (Keil & Wilson 2000). A missing WAD, however, does
not preclude the existence of powerful constraints on the seman-
tics of language, such as relations governing scope, quantification,
and anaphora (Chierchia & McConnell-Ginet 1990). Word mean-
ing is special, not because it needs its own constraints, but because
it is so different from the rest of language.
ACKNOWLEDGMENT
Preparation of this paper was supported in part by NIH Grant R01-
HD23922 to Frank Keil.
How fast does a child learn a word?
Michael Maratsos
Institute of Child Development, University of Minnesota, Minneapolis, MN
55455.
marat001@umn.edu
Abstract: This discussion argues that for many word meanings, the child
has to assemble a new category, using relatively slow information-sifting
processes. This does not cause high semantic errors, because children
probably hold off using a word until much such sifting has occurred, rather
than producing the new word as soon as they have any information on it.
First, this is a really excellent book, one of the most worthwhile I
have read lately. Bloom provides pro tempore convincing argu-
ments on many central issues, such as how skillfully children judge
speaker’s attention and intention, the lack of a need for language-
specific heuristics in word learning, and (agreeing with Maratsos
& Deak 1995) the essentially fox-like nature of most word mean-
ing acquisition. All this is done with admirable clarity and open-
minded moderation. The book is a model treatment.
While acknowledging my general agreement with many of
Bloom’s central analyses, my discussion centers on a set of issues
where we probably disagree somewhat. These issues comprise of
four intertwined problems: (1) At the most general level, in the
child’s learning of a word, how common are slower processes sift-
ing information over different inputs of the word, as opposed to
immediate or near-immediate analysis from one or very few in-
puts? (2) When children acquire a word meaning, to what degree
do they just match with an already existing cognitive-perceptual
category, which I will call category recognition, versus having to
assemble a new conceptual structure from previously existing
nonlinguistic cognitive-perceptual elements, a process called here
category assembly? (3) Highly related to (1) and (2), do children
immediately use a word upon processing some initial input, or
does their own productive use occur only after processing a pos-
sibly large number of inputs? (4) Why are semantic errors so rel-
atively low in children’s uses?
Bloom is open on these problems, but his discussion favors the
position that children use a word quickly after hearing initial in-
put. Combined with the low rate of error, this means they must
have the means for quickly making a mostly accurate analysis.
Among the means for the latter, they probably do not have to as-
semble categories for new words; instead they mostly recognize
which already existent categories match the word used. (Docu-
mentation of Bloom’s preferences, which are not monolithic, ex-
ceeds the space available here.)
In many cases, relatively quick category recognition is probably
correct. That is, children plausibly have already formed some cog-
nitive-perceptual categories like “table” or “cat” prior to learning
words for such categories. This prior assignment is most likely for
what are called basic object concepts (Rosch 1977) and probably,
as Bloom suggests, for proper names. But category assembly,
which requires a slower process of sifting over many inputs and
analyzing or constructing a new assemblage of elements, is intu-
itively more likely for many word meanings; it seems particularly
likely for many verbs and other non-object words.
For example, we might think our word “eat” groups together
the “natural” set of ingestion situations and features. But Brown
(1998) documents how in Tzeltal, three more specific ingestion
meanings are encoded by the three most common ingestion
words. One means “to ingest crunchy-solid foods” (like carrots).
One means “to ingest less-than-crunchy-solid foods” (like ba-
nanas). And one means “to ingest tortillas and other bread sub-
stances.” These are not “intuitively obvious” nonlinguistic cate-
gories. To take another case, Bowerman (1996) discusses Korean
change-of-location meanings like “kohta” (to put an object elon-
gated – long or tall – in one dimension onto any base, e.g., a flower
into a vase, a hairpin into hair), or “puwlhita” (join a flat surface
to another flat surface) or “kkakta” (take off a covering layer with
a bladed instrument), and many others.
Commentary/Bloom: How Children Learn the Meanings of Words
BEHAVIORAL AND BRAIN SCIENCES (2001) 24:6 1111
It seems implausible – or at least worth not presupposing – that
children naturally nonlinguistically form a category of actions like
“put an object elongated in one dimension onto/into a base,” or
“eat breadlike substances.” It seems unlikely that they have always
pre-formed just the right cognitive-perceptual categories before
figuring out all word meanings that languages have. In such cases,
which are common among verb meanings compared across differ-
ent languages, they probably have to analyze how the word is used
over many different situations, in order to assemble the right col-
lection of cognitive-perceptual elements. There would be no likely
“initial best guess,” or pre-fabricated conceptual category.
Suppose such slower “assembly” processes are often the case.
What then becomes potentially puzzling, is that children make rel-
atively few semantic errors in their word use, valid both as a gen-
eral conclusion (Bloom) and as a conclusion about the Tzeltal and
Korean cases (Bowerman 1996; Brown 1998). For if there is a
somewhat protracted period of category assembly, so that each
word meaning development has to cover a lot of ground, would-
n’t children have to make many errors in the word use, before they
finally get to the right meaning? This paradox exists, however, only
if one assumes children immediately produce words immediately
upon processing some initial input. What if, instead, children only
produce a word after having gathered information about its use in
the course of many different uses? This would give time for slower
category assemblage processes to give a more accurate meaning,
before the child actually used the word. In fact, there is plenty of
evidence that children often work on a word meaning for some pe-
riod before they actually produce the word. To start with, current
estimates are that when children produce their first word, on av-
erage they comprehend around fifty. In general afterwards, many
more words are at least partly comprehended for some time be-
fore they are produced. So there is good reason to think children
often accumulate a good deal of information about a word before
they try it out themselves. This delayed production process would
give time for a constructed word meaning category to have be-
come mostly accurate before actual use occurred, explaining the
generally low degree of semantic error. Thus, slow acquisitional
processes can also result in relatively accurate initial productions.
In summary, it is very likely that children sometimes do get to
a mostly accurate meaning very quickly. But the kinds of word
meanings discussed above imply that slower category assembly
processes also play a part in acquisition. Part of the fox-like vari-
ety of the child’s procedures lies in the use of various kinds of faster
and slower processes. In the latter case, production delayed after
initial input processing plays a crucial role in making initial pro-
duction accuracy relatively high.
Fast-mapping children vs. slow-mapping
adults: Assumptions about words and
concepts in two literatures
Gregory L. Murphy
Psychology Department, New York University, New York, NY10003.
Gregory.Murphy@NYU.edu
Abstract: Research on children’s and adults’ concepts embodies very dif-
ferent assumptions of how concepts are structured, as reflected in their ex-
perimental designs. Developmental studies seem to assume that cate-
gories contain highly similar objects that can all be identified from one or
two examples. If concepts are more like those tested in adult experiments,
research on word learning may be misleading.
Paul Bloom’s thesis that much of word learning arises through
general-purpose concept learning is one that I have great sympa-
thy with (Murphy 1991). Children and adults can communicate to
a great degree, suggesting a basic commonality to their word
meanings and underlying conceptual structure. It is surprising,
then, that the adult and developmental literatures seem to have
very different ideas of what concepts are like.
A typical word-learning experiment in the developmental liter-
ature involves the naming of a single object, usually a very few
times. The child is then tested on a number of similar objects, of-
ten differing from the learning models in only one dimension.
Since Carey and Bartlett (1978) first demonstrated fast mapping,
similar experiments have been carried out dozens of times, many
of them described in Bloom’s book. Apparently, the assumption is
that children can learn the word’s meaning fairly well based on a
single example. Those familiar with this literature will probably
think “They DO learn the word based on a single example! That’s
what the fast-mapping results show.”
Before replying, let’s consider the typical concept-learning ex-
periment in the adult literature. Here, subjects are shown cate-
gory exemplars one at a time and are told that some of them are
called “Dax” and some “Kez.” Subjects start by guessing the iden-
tity of the first few items, but they eventually learn what the two
categories are and can name items correctly. The operative word
here is “eventually.” In fact, in a number of such experiments, sub-
jects take an extremely long time to learn the categories, even
when there are fairly few items involved. For example, in a classic
paper by Medin and Schaffer (1978, Experiment 2), subjects were
to learn to classify nine items into two categories. However, about
40% of the subjects failed to do so after 16 blocks of learning tri-
als. Lamberts (2000) didn’t stop his subjects, and they took an av-
erage of 38 blocks to learn these categories. Although this is an ex-
treme example, bad performance is rife in studies of adult concept
learning.
Clearly, taking 38 blocks to learn two categories is not fast map-
ping. Somehow, the bright children who could learn a category
based on a single example grew up into dull college sophomores
who could not learn two categories even when they saw all their
exemplars over and over again. It would be easy to blame the ed-
ucational system, but the categories and methods involved in these
experiments are so different that one cannot really compare them.
These differences reflect different assumptions, summarized be-
low.
Category structure.
To demonstrate learning in the fast-map-
ping experiment, a child is shown an object and then is tested on
another object that differs little from the original. The assumption
seems to be that members of the same category are very similar.
In contrast, in the adult literature, categories often have very weak
structure, with many overlapping features. In some cases, the cat-
egories seem perverse, with objects that differ in every dimension
in the same category (e.g., Medin & Schwanenflugel 1981).
Learning procedure.
In the fast-mapping learning phase, the
child sees a single exemplar and hears it labeled a few times. The
adult learner guesses the category, receives feedback, and contin-
ues until performance reaches a high criterion, usually perfect
classification of all the items in a block. These differences in pro-
cedure follow from the assumptions of category structure: If the
category is weakly structured, all the category members must be
tested to discover whether they have been learned.
Learning models.
These differences have led to very different
kinds of explanations of category learning. In the developmental
literature, the approach has been one of hypothesis-testing, fo-
cusing on biases and constraints on the child’s hypotheses. In the
adult literature, the focus has been on associative learning, in
which properties and exemplars are associated to categories, or at-
tention is gradually drawn to the critical dimensions that separate
the categories. If one reads enough papers of each sort, one must
come to the conclusion that both cannot be right. In fact, I sus-
pect that both are wrong.
There is clearly a sense in which fast mapping is true: If you use
a word in front of your 2-year-old, you may find him or her using
it just as you did. However, that correct performance is not a strin-
gent test. If you label a robin a bird, your 2-year-old will not be
correctly labeling ostriches, penguins, bats, and swans right after-
Commentary/Bloom: How Children Learn the Meanings of Words
1112 BEHAVIORAL AND BRAIN SCIENCES (2001) 24:6
wards. Indeed, a child may not be able to use this word with near-
perfect accuracy until he or she has seen pictures and read books
in which the entire variety of birds has been revealed: The child
may not “learn the word” until the age of 4, 6, or adulthood. On
the other hand, the very weak structure of many experimental cat-
egories is not justified either. Real categories often have a set of
typical members that could allow fast mapping as well as a set of
atypical members that would not (Rosch & Mervis 1975).
The two experimental paradigms I am contrasting seem to fa-
vor one aspect of the usual category structure at the cost of the
other: Category members are either almost uniform or have little
in common. What worries me is that neither of these situations
may tell us how children learn to name realistic categories that are
a mixture of the two.
What all this leads to is the conclusion that we need a better un-
derstanding of what real-world categories and word meanings are
like. When we consider statistics of how many words children
learn and how quickly, we may be overly liberal in attributing to
them knowledge of “the meaning,” based on correct uses for a few
central cases. If categories have more atypical exemplars and dif-
ficult cases, learning might not be complete for months or years.
I am not certain whether the long task of filling in the exceptions
and unpredictable cases would greatly change Bloom’s conclu-
sions about the word-learning process, but it is important to ex-
plore more realistic and perhaps messier conceptual structures
than has usually been done either with adults or children.
Why theories of word learning don’t always
work as theories of verb learning
Letitia R. Naigles
Department of Psychology, University of Connecticut, Storrs, CT 06269-
1020.
Naigles@uconnvm.uconn.edu
Abstract: Bloom’s theory of word learning has difficulty accounting for
children’s verb acquisition. There is no predominant preverbal event con-
cept, akin to the preverbal object concept, to direct children’s early event-
verb mappings. Children may take advantage of grammatical and linguis-
tic information in verb acquisition earlier than Bloom allows. A distinction
between lexical and grammatical learning is difficult to maintain for verb
acquisition.
Striking parallels can be drawn between the magical classes in
Harry Potter’s Hogwarts School of WitchCraft and Wizardry
(Rowling 1998) and the major types of word learning discussed in
Paul Bloom’s book How Children Learn the Meanings of Words.
The Hogwarts class called Transfiguration, which teaches students
how to transform teapots into tortoises, finds a match in Bloom’s
excellent discussion of how children learn names for natural kinds
versus artifacts; the class called Potions, which is about creating
substances from all sorts of raw material, finds a match in Bloom’s
insightful presentation of how children learn words for sub-
stances, and the class called Arithmancy shares its numerical fo-
cus with Bloom’s detailed chapter on children’s acquisition of
names for numbers. Where Bloom’s book suffers is just where the
parallels with Hogwarts break down: Hogwarts classes also in-
clude Charms, where “Alohomora” is taught to enable the open-
ing of locked places, and Defense against the Dark Arts, where
“Expelliarmus” is taught to force the disarming of one’s opponent.
Bloom’s book includes no comparable deep discussions of chil-
dren’s acquisition of the words for actions, events, and relations;
that is, the acquisition of verbs. And just as Harry Potter and his
friends would not have survived their adventures without their
spells from Charms and Defense Against the Dark Arts, I argue
here that Bloom’s theory of word learning suffers on account of
the subordinate position relegated to children’s acquisition of
verbs.
It is possible that Bloom’s neglect of verbs does not harm his
theory of word learning. If all word learning derives from the same
roots, then everything that Bloom describes for children’s noun
learning can be easily and simply extended to verbs. However,
there are ways in which verb learning is not like noun learning, so
that what Bloom has proposed for nouns does not easily transfer
to verbs.
Children’s early nouns tend to refer to objects and individuals,
such as dog, cup, ball, baby, and mama. This tendency – seen ro-
bustly across languages – has led some researchers to propose that
children are using a specifically linguistic principle (e.g., object
scope, whole object bias) to focus their initial attention on objects
as the referents of unfamiliar words. As part of his program to
avoid such specifically linguistic principles of lexical acquisition,
Bloom proposes instead that the roots of the object bias lie in in-
fant cognition. Numerous researchers have demonstrated a sus-
tained focus throughout preverbal infancy on objects and individ-
uals, and this focus could function in the word learning years to
make objects the most salient referents around. Thus, Bloom re-
solves Quine’s (1960) problem – that a given word in a given situ-
ation can have any number of possible meanings, including color,
texture, parts, or activities – by postulating that children target the
salient object in the situation as the default referent because of
their preverbal focus on objects.
The problem is that the same “plethora of possibilities” exists
for verb meanings, and there is no analogous “preverbal concept”
resolution. For example, the event of “a child brings a doll to
Grandma” can be referred to as carrying (if the manner of bring-
ing is key), bringing (if the manner is not key), giving (if the trans-
fer to Grandma is key), getting or taking (if receipt of the doll is
key), coming or arriving (if the doll’s motion is key), seeing (if the
child’s perceptions are key) and wanting (if Grandma’s desire is
key). Thus, different aspects of the event are highlighted, de-
pending on which verb is used. But then how might a child learn
any one of these verbs? The analogous resolution à la Bloom
would involve the proposal that one or two predominant prever-
bal event concepts would direct children’s attention to default
events for their initial referent mappings. But children’s early
verbs, which are often produced within their first 50-word vocab-
ularies (Bloom et al. 1993; Dromi 1987; Tardif 1996), are ex-
tremely heterogeneous in their semantic classes. They encompass
both general and specific meanings (come and draw), both causal
and noncausal meanings (break and go) and both externally-expe-
rienced and internally-felt meanings (eat, give, see, want). The va-
riety of children’s early verb meanings casts doubt on there being
a unitary supporting event concept that predates verb acquisition.
Bloom’s presentation of verb learning (HCLMW, pp. 201–203,
209–211) includes roles for general linguistic and specific syntac-
tic information (so-called syntactic bootstrapping; Gleitman 1990;
Naigles 1998) as clues to verb meaning. But his discussion has a
curiously uncertain quality to it. Bloom spends considerable space
debating the centrality of syntactic information in verb learning,
both in terms of its necessity (which he finally accepts) and its suf-
ficiency (which no one has ever claimed). Moreover, Bloom holds
that syntax could not be providing information for the earliest verb
learners because these children are still at the one-word stage of
language production. And of course, if children don’t know any
syntax (e.g., Tomasello 2000a), they certainly couldn’t be using it
to learn about verbs. However, Bloom overlooks the now-consid-
erable evidence that children’s configural analyses of linguistic in-
put, as well as their grammatical comprehension, are well in ad-
vance of their language production (e.g., Gomez & Gerken 1999;
Hirsh-Pasek & Golinkoff 1996; Santelmann & Jusczyk 1998).
Moreover, the finding that maternal utterance length is a stronger
predictor of toddlers’ subsequent spontaneous speech vocabulary
than child-mother joint attention or maternal word frequency,
shows that toddlers are processing entire utterances of maternal
speech even when they are only producing one- and two-word ut-
terances themselves (Hoff & Naigles 2002). Perhaps a dichotomy
Commentary/Bloom: How Children Learn the Meanings of Words
BEHAVIORAL AND BRAIN SCIENCES (2001) 24:6 1113
between the types of information that children use at the earliest
versus later stages of word learning is not needed.
It is possible that Bloom chose not to discuss verb learning in
depth because such a discussion invariably introduces – as it has
in this commentary – a consideration of children’s acquisition of
grammar as well. Bloom makes it clear that he regards these types
of language acquisition – the grammatical and the lexical – as dis-
tinct. But verbs do fit in both types; they are lexical items with dis-
tinct meanings as well as words with predictable relations to the
sentences in which they appear. Any theory of child word learn-
ing that desires to encompass all words and not just nouns must
grapple with this fact.
The name game updated
Katherine Nelson
Developmental Psychology, City University of New York Graduate Center,
New York, NY 10016.
knelson@gc.cuny.edu
Abstract: Bloom’s domain general theory remains strictly cognitive and
individualistic. By ignoring the contribution of social interaction and col-
lective construction of concepts, he fails to solve the word learning prob-
lem.
Paul Bloom’s book on learning the meaning of words (which, de-
spite the word “children” in the title applies across the lifespan)
covers a broad territory in a mostly balanced way, and especially
welcome is the move away from language-specific constraints on
word learning to the recognition of the necessary contributions of
domain general abilities. Also welcome is the declaration that
“children aren’t like scientists who have theories; they are like sci-
entists before they have theories: trying to make sense of some do-
main they know little about” (pp. 168–69). Highlighting meaning
as the central problem is another welcome move, contrasting with
most other current developmental theories. Meaning is defined in
terms of “narrow content,” the psychological (rather than social
and contextual) aspect, “relevant to explaining people’s intuitions
about reference and categorization” (p. 17). The problem here is
that “people’s” intuitions are not independent of social and con-
textual uses of words.
Although it contradicts many current positions the new theory
is not really new; it reformulates word learning theories from the
1970’s (L. Bloom 1973; Bowerman 1976; Nelson 1974). Bloom
recognizes and even relies on selected data from these studies, but
he ignores their theoretical proposals. Data are cited from Nelson
(1973) but the social interaction processes documented there, as
well as the individual differences in both parents’ and children’s
functional use of words and associated differences in form, are ig-
nored.
For Bloom all the action in word learning is in the heads of chil-
dren. The central focus of the theory is on children’s ability to in-
fer the intentionality of adults, but the influence of interactive so-
cial experience or knowledge is explicitly denied. Although Bloom
views the word learning problem as belonging to the domain of so-
cial cognition, his theory remains individualistic to the core, and
thereby fails to address the most critical problems.
Like most word learning theorists, Bloom is occupied with ob-
ject words and categories (acknowledging but not explaining other
kinds of words). His example of a novel concept learned through
language – “hat trick” in hockey – is explained in perceptual-action
terms. But as Searle (1995) points out, the concept of a game like
hockey relies on collective (not individual) intentionality; like
money, it is a social institution. You cannot learn the concept of
money without knowing about buying and selling, prices, com-
modities, banks, stores, governments, and so on. These socially es-
tablished entities depend upon the use of words to constitute their
reality. Young children who use the word money in reference to
coins and bills cannot yet have the concept of money or the con-
ventional meaning of the word “money.” How then can we account
for their appropriate use of words in the absence of knowledge of
their meaning, and how do children acquire meaning? Despite the
promise of the title and the extensive discussion of concepts,
Bloom fails to address these problems adequately. For Bloom,
essences are assumed to guide children to the meanings of words
for natural kinds and artifacts, but more abstract terms (like
money and game) defy this kind of easy solution, as Wittgenstein
(1953) emphasized. Even the simplest examples such as “chair”
are not as clearcut as Bloom and other “essentialists” assume
(Rosch 1978). Regarding the learning of abstract words, Bloom
states: “I think it will remain a mystery for a long time” (p. 94).
Taking children’s errors in word uses seriously, I proposed (Nel-
son 1974) that children focus on the function of things (from the
child’s perspective) in forming concepts and naming things. This
disposition, although akin to Bloom’s thinking about artifact
essences, often does not accord with the collective meaning of the
term, and thus children must tune their concepts and categories
to the language they are learning, even for everyday animate and
inanimate objects. They frequently use words without conven-
tional meaning, and continue to do so, especially for abstract
words like “money,” or “know” and “think.” How then do they ac-
quire meanings? Gradually, through using and hearing words used
in different contexts and observing the practices and discourses of
people in relation to those words they build and trim their con-
cepts to fit the uses of the language users around them (Nelson
1996).
By ignoring the fact, indeed the necessity, of this process,
Bloom distorts the relation between language and thought. He as-
serts that “Words are not necessary for thought. Structured and
abstract thought occurs without them” (p. 259). And, “Lan-
guage... is a tool for the communications of ideas... not a mech-
anism... to generate and appreciate these ideas in the first place”
(p. 258). Although he believes that we can learn new concepts
from language (such as “hat trick”), language has nothing to do
with generating or appreciating new concepts.
In this, Bloom greatly underestimates the significance of sym-
bols to culture and cognition, and the complexity of concepts that
people learn, use, and create, as well as the different cognitive
structures that cultures make manifest in their languages with
demonstrated effects on thinking and problem solving (Bower-
man & Levinson 2001). He does not accept that our words are ca-
pable of constructing worlds that would never exist, in individual
cognitive systems, or in communities, without the symbols that
they depend upon. It is not possible to address the problem of
word meaning without addressing the language and thought prob-
lem, and his rejection of any “contamination” of thought by lan-
guage renders his theory at best inadequate to the problem and,
at worst, incoherent.
An ideational account of early word learning:
A plausibility assessment
Rita Nolan
Department of Philosophy, State University of New York at Stony Brook,
Stony Brook, NY 11794.
rdnolan@ms.cc.sunysb.edu
Abstract: The theoretical framework of Bloom’s account of child word
learning is here assessed only for initial plausibility and neural plausibility.
The verdict on both dimensions is low, largely due to the size and charac-
ter of knowledge it is claimed that the child brings to the task. It is sug-
gested that elements of constructivist accounts could profitably be drawn
from to reduce this implausibility.
Bloom construes early word learning as a mapping task in which
the word maps onto a psychological entity that is a concept. His
test for successful mapping of referential terms is getting their ex-
tensions right; a concept’s role is to pick out the right category of
Commentary/Bloom: How Children Learn the Meanings of Words
1114 BEHAVIORAL AND BRAIN SCIENCES (2001) 24:6
things in order for the sole business of the language, communica-
tion, to proceed. The local linguistic context generally provides
only the language-specific word to be mapped onto the pre- and
non-linguistic concept, which plays much the same role as Locke’s
“ideas” did (Locke 1690/1975), minus his tabula rasa. To solve the
mapping problem, the child uses multiple strategies, of which the
central one is discerning the intentions of speakers. In the basket
of competencies available to the child, essentialism, the assump-
tion that many individuals are referred to by the same word be-
cause of a shared hidden essence, is also a significant asset. Draw-
ing upon a wealth of experimental results, Bloom applies this
explanation to defeat the alternative explanations of (1) empiricist
associationism and (2) constraints that are specifically linguistic.
Initial plausibility.
The account is low on initial plausibility.
First, it violates the requirement that the direction of explanation
be from the less understood to the better understood. Another
concern derives from the mutually-supporting character of the ba-
sic assumptions of Bloom’s explanation, for example, (1) an
ideational theory of word meaning, and (2) the independence of
human conceptualization and other higher cognitive competen-
cies from language. One might argue for the Ideational Thesis on
the basis of the Independence Thesis, and vice versa; but inde-
pendent confirmatory evidence of either assumption is desirable.
To these two assumptions Bloom adds the Essentialist assumption
and the Theory of Mind assumption. Taken together, these as-
sumptions yield a coherent explanation, to be sure, but no inde-
pendently compelling evidence is given for any one of them. In-
stead, the argument strategy of inference to the best explanation
is adopted, wherein it is sufficient to show that, of all available
candidate explanations, only this one accounts for all the cases.
But have all available competing explanations been considered?
Neural plausibility.
A different concern arises from the amount
of knowledge that is attributed to the prelinguistic infant, a con-
cern that raises issues that breach the boundaries of a psycholog-
ically autonomous theory. That amount of knowledge makes
Bloom’s proposed explanation neurally implausible. Human brain
plasticity and immature birth argue for human selection for adapt-
ability over adaptedness (Lorenz 1965). That we are born know-
ing so little that we cannot do much, enhances our abilities to
adapt to the environment through acquiring information from it –
which the adaptedness of other species, to their disadvantage,
prevents them from acquiring. Human brain plasticity, immature
birth, and high degree of adaptability make plausible the hypoth-
esis that humans have less in the way of species-specific (“off the
rack”) initial knowledge structures than are enjoyed by other
species, enabling humans to build knowledge structures that are
more useful to them because the structures are “tailor-made”
through interaction with the environment. From a design stance,
economy would be served if these structures were also general
rather than domain specific.
Constructivism.
The above considerations lead us to expect the
development of new information-bearing and other cognitive
neural structures in response to the organism’s interactions with
the environment, particularly structures that subtend major be-
haviorally manifested cognitive achievements like language de-
velopment. The neural constructivist proposal (Quartz & Se-
jnowski 1997) offers just that, itself building on the long tradition
of constructivism in psychology (see, e.g., Bruner1963; Bruner et
al. 1956; Piaget 1971; Vygotsky 1978). In Nolan (1994a), I
sketched an account of how new psychological cognitive struc-
tures might be constructed from structures already in place to-
gether with environmental input, and showed how that hypothe-
sis might explain experimental results then available on word
learning. And in Nolan (1994b), I proposed that a structure in
which words are used purely referentially, representing percep-
tual categories, may likely precede the development from it,
caused by environmentally presented problems, of true concep-
tual categories, understood as associated with predicates rather
than as merely referring expressions. Here are two examples, over-
simplified for brevity and intended only schematically, to illustrate
the sort of relations proposed among cognitive structures. The
word “dog” may initially be used as if it were the name of the fam-
ily pet, then to refer to each neighborhood dog, both uses being
based on perceptual categories alone; later, to solve problems with
the extension of the word and to resolve the roles played by its su-
perordinates and subordinates, that same word takes on predica-
tive content, and is transformed qualitatively into a conceptual cat-
egory. A second, much simpler example of the construction of a
new cognitive structure from an old one in response to environ-
mental complexities is the development from babbling to the
phones of the local language. In each case, the shift may not be
perceived as any simple act that the child does differently. These
issues cannot be stated within the framework of Bloom’s account.
The proposal of Quartz and Sejnowski comes at a time when the
interface between psychology and the neural sciences is in focus,
making the lack of consideration of a psychological constructivist
alternative to the three explanations considered by Bloom, strik-
ing. In the absence of consideration of all alternative explanations,
the strategy of argument to the best explanation fails.
The exchanges generated by Quartz and Sejnowski (1997) con-
firm that determining what constitutes a new, constructed, cogni-
tive neural structure requires cooperation from psychology, in the
form of identifying which cognitive achievements might signal the
emergence of a new cognitive structure, even though it may turn
out that neural construction is in some way continual. In collect-
ing and reviewing experimental results that are inconsistent with
both associationist and language-specific constraints, Bloom’s
book presents significant data for the cooperative project of psy-
chological and neurological constructivism. Furthermore, his ar-
guments (1) that the significant cognitive structures evidenced in
language development with respect to word learning are general
knowledge structures, and (2) that the cognitive strategies used by
the child in this task are multiple, cohere firmly with constructivist
explanatory hypotheses, neural and psychological.
Some cognitive tools for word learning: The
role of working memory and goal preference
Mihály Racsmány,
a
A
´gnes Lukács,
b
Csaba Pléh,
c
and Ildikó Király
d
a
Department of Psychology, University of Bristol, United Kingdom, and
Department of Psychology, University of Szeged, H-6722 Szeged, Peto˝fi 30,
Hungary;
b
Institute fir Linguistics, Hungarian Academy of Sciences and
Collegium Budapest, H-1068 Budapest, Benczùr U. 5, Hungary;
c
Center for
Cognitive Science, Budapest University of Technology and Economics, and
Collegium Budapest, H-1111 Müegyetèm RKP. 3, Hungary;
d
Research
Institute of Psychology, Hungarian Academy of Sciences, Budapest, H-1068
Szondi U. 83– 85, Hungary.
racsmany@edpsy.u-szeged.hu
alukacs@colbud.hu pleh@itm.bme.hu kiralyi@mtapi.hu
Abstract: We propose that Bloom’s focus on cognitive factors involved in
word learning still lacks a broader perspective. We emphasize the crucial
relevance of working memory in learning elements of language. Specifi-
cally, we demonstrate through our data that in impaired populations
knowledge of some linguistic elements can be dissociated according to the
subcomponent of working memory (visual or verbal) involved in a task.
Further, although Bloom’s concentration on theory of mind as a precon-
dition for word learning is certainly correct, theory of mind being a nec-
essary condition does not make it a sufficient one. On the basis of our stud-
ies we point out the importance of a theory of mind related goal preference
in acquiring spatial language. In general, we claim that more specific cog-
nitive preferences and constraints should be outlined in detail for the pre-
conditions of acquiring linguistic elements.
In Bloom’s detailed and well-argued concept of the acquisition of
word meaning, a rather ambitious general model is set up. Ac-
cording to this, both necessary and sufficient conditions for word
learning are of a general nature – he speaks about computing in-
tentions, concept formation, consideration of syntax, and “certain
Commentary/Bloom: How Children Learn the Meanings of Words
BEHAVIORAL AND BRAIN SCIENCES (2001) 24:6 1115
general learning and memory abilities” (HCLMW, p. 10). Rather
than evaluating the impressive general model, our intention is to
point out how one could elaborate on some of these assumed “gen-
eral abilities.”
Imagine a child who is equipped with a fully functioning theory
of mind, but lacking any storage system to temporarily store forms
of items. The child would certainly know that sounds have a dual
relationship to the world and to the mental state of the speaker,
but would not be able to consolidate the signs on account of the
lack of memory. In order to consolidate both form and meaning,
one has to have a working memory system among those general
memory abilities. This is of course a trivial aspect, but it is the cor-
nerstone of basic assumptions about how working memory is used
in learning new words. Individual differences in word learning and
vocabulary size and their relationship to working memory capac-
ity have been aptly documented (Baddeley et al. 1998; Gathercole
et al. 1992; 1997). Developmental pathology is of crucial interest
in this regard.
In one of our studies (Lukács et al. 2001) we obtained some in-
teresting dissociative relationships between vocabulary acquisi-
tion and working memory. As Table 1 shows, in Williams’ syn-
drome subjects of a wide age range (between 6 and 20 years) the
best predictor of picture naming performance for frequent words
was the age of the subject. In contrast, knowledge of rare words
was best predicted by verbal working memory measured in digit
span. This suggests that working memory is an important cogni-
tive resource for putting words into the mind.
Interestingly, we also observed that a clear dissociation shows
up between the acquisition of simple agreement related basic
grammatical morphemes like plural and accusative, and spatial
suffixes and postpositions (Racsmány et al., in preparation). In a
series of suffix elicitation tasks constructed by Pléh et al. (1996),
we contrasted the knowledge of spatial and non-spatial inflec-
tional forms in Williams’ syndrome children. As Table 2 shows,
there is a clear dissociation between using spatial and non-spatial
language (T-test for dependant samples: t 54,9, p ,0.01; suffixes
and postpositions taken together).
Working memory capacity was a strong predictor of perfor-
mance in morphological tasks. However, this time our analysis in-
volved not only verbal working memory, but spatial working mem-
ory capacities as well. In a multiple regression analysis, the two
modality-dependent working memory performances explained 93
percent of the variance of spatial morphological task performance
(F 548,66, p ,0,001, adjusted R
2
50,93). This result reveals
that in the process of spatial language learning, children have to
keep in mind phonological and spatial information at the same
time. Spatial and verbal working memory capacity give crucial
constraints for the rate and the level of spatial language learning.
There is another aspect where we would like to enrich the pic-
ture presented by Bloom. In present-day infant studies several
lines of research have demonstrated that the theory of mind com-
plex emphasized by Bloom goes through several preparatory
stages until it reaches its full articulation. Bloom clearly sees not
only the relevance of the theory of mind literature in explaining
the details of early vocabulary acquisition, but the importance of
preverbal preferences as well. We would like to point to one of
these constraints that seems to be very important in both early and
later stages: the notion of goal-directedness that is shown to be
central in early mental representation of action (see e.g., Csibra et
al. 1999). Goals are also of central importance in mental repre-
sentations underlying language (Jackendoff 1994), especially the
ones underlying spatial language (Landau & Jackendoff 1993).
Some of our data show that this goal-directedness of human cog-
nition is one of the easier cognitive templates of human language.
In Hungarian, where there is an obligatory distinction between
goal, static, and source relations, a study analyzing 12,000 utter-
ances in children between 1;5 and 2;5 of the MacWhinney (1995)
corpus showed the following percentages for spatial suffixes: goal
80%, static 13%, and source 7% (Pléh et al. 1996).
This clear preference for coding goal might well be explained
by input factors. To examine this, following the model of Landau
(1994), artificial spatial suffix and postposition learning situations
were created where children between ages 3 and 6 (n 5238) had
to learn the meaning of new spatial expressions (for the method
and results, see Király et al. 2001; Pléh et al. 1999) in different spa-
tial settings.
Table 3 shows that goal is easier to learn both in suffixes (F
2,92
5
8.77, p ,0.01) and in postpositions (F
2,142
59.64, p ,0.01), and
neither the spatial settings (like diagonal, under, etc.) nor age had
a significant effect.
Thus, goal as a cognitive template seems to be present at the
earliest ages, but surprisingly enough it is still used as a template
when primary language acquisition has already gone a long way.
Taking into consideration cognitive constraints such as working
memory capacity and goal preferences may help to better explain
individual differences and neuropsychological phenomena in lan-
guage acquisition. How would strong individual differences in vo-
cabulary size be possible among Williams’ syndrome children with
an unimpaired theory of mind? And how could autistic children
learn language with a serious theory of mind deficit? A possible
answer would be that simple cognitive mechanisms either speed
up or block the process of word acquisition. The limited capacity
of working memory supports language learning, works as a “lan-
guage learning device” (Baddeley et al. 1998), while goals as a cog-
nitive bootstrapping helps to unfold the meaning of words.
ACKNOWLEDGMENTS
Financial support for most of the research mentioned in the commentary
was provided by an OTKA grant (Hungarian National Science Founda-
Commentary/Bloom: How Children Learn the Meanings of Words
1116 BEHAVIORAL AND BRAIN SCIENCES (2001) 24:6
Table 1 (Racsmány et al.). Stepwise regression solutions for
different vocabulary measures in Williams’ syndrome
Dependent Regression
variable coefficient Equation First R
Rare words 0.65 1.8.2 14.2 span Span 0.65 F 5 8.59
Frequent words 0.75 29.18 10.51 age Age 0.75 F 5 15.26
Table 2 (Racsmány et al.). Percentage of correctly used spatial
and non-spatial morphological forms in Williams’ syndrome
Non-spatial grammatical
morphemes Spatial suffixes Spatial postpositions
87 46 49
Table 3 (Racsmány et al.). Learning artificial spatial markers
with different forms. Correct binary choice percents
Age 3;6 5;6
Goal Source Goal Source
Suffixes 70 52 60 42
Postpositions 64 48 74 29
tion) T 029514 to Csaba Pléh, by a grant of the McDonnell Foundation No.
4-27569 to Thomas Papathomas and Ilona Kovács of Rutgers University, and
by an NSF grant No. BCS01226151 to Ilona Kovács with Csaba Pléh as co-
principal investigator. A
´gnes Lukács and Csaba Pléh were junior and asso-
ciate fellows at the Collegium Budapest while working on this paper.
Empiricist word learning
Dan Ryder
a
and Oleg V. Favorov
b
a
Department of Philosophy, University of North Carolina at Chapel Hill,
Chapel Hill, NC 27510-3125;
b
School of Electrical Engineering and
Computer Science, University of Central Florida, Orlando, FL 32816.
dan@danryder.com http://www.danryder.com
favorov@cs.ucf.edu http://www.sinbad.info
Abstract: At first, Bloom’s theory appears inimical to empiricism, since he
credits very young children with highly sophisticated cognitive resources
(e.g., a theory of mind and a belief that real kinds have essences), and he
also attacks the empiricist’s favoured learning theory, namely, association-
ism. We suggest that, on the contrary, the empiricist can embrace much of
what Bloom says.
According to Bloom, the associationist believes “that the relation-
ship between... words and what they refer to is established not
through a process of reasoning and inference... but through a
sensitivity to covariation”(HCLMW, p. 58). Bloom attacks associ-
ationism (Ch. 3), arguing convincingly that a correlation between
a word and a type of object is not enough to make a child assign
that object as the word’s referent. The child must realize that
someone intends to refer to the object, which entails that the child
has a “theory of mind.” But does this mean that the child must use
a non-associationist “process of reasoning and inference” in order
to learn the meaning of the word? We think not. Though Bloom
establishes that word learning cannot result from a sensitivity to
covariation between a word and an object, the evidence does not
rule out the possibility that word learning could consist in the as-
sociation of a word with an intention to refer to an object. Associ-
ation occurs between representations, and the representations
need not be basic sensory ones. For example, the associationist
theory criticized by Bloom credits a child with the possession of
and the ability to apply theoretical representations, namely, of
words and object kinds. It is a separate (and open) question
whether the empiricist can account for the acquisition of such so-
phisticated concepts. (For an account that could explain such ac-
quisition, see Ryder & Favorov 2001.)
The empiricist supposes that word learning involves only gen-
eral inferential abilities. Although Bloom acknowledges the rele-
vance of such abilities (p. 211), and does not believe there is a
mechanism whose special purpose is to facilitate word learning
per se, he proposes a key role for mechanisms whose special pur-
poses are more broadly defined – for example, a theory of mind,
an understanding of principles of individuation, and an essential-
ist ability to form concepts. We think that the considerations
Bloom marshals in favour of his theory in fact support the central
involvement of a general purpose mechanism.
Consider the nature of the knowledge acquired through word
learning. Despite his objections to associationism, Bloom de-
scribes learning the meaning of a word as acquiring an “associa-
tion” (p. 17) or a “mapping” (p. 89) between a form and a concept
or mental representation. Actually, given his position that word
learning is a product of inference, it seems that what he should say
is that learning the meaning of a word is learning some kind of fact.
In the case of referring terms, one learns something about a word,
for example, “cat”: that it refers to instances of the kind cat. (At
the same time, one learns about the kind cat that its instances are
called “cat.”) This knowledge does not differ in format from any
other knowledge one might acquire about the environment.
Bloom’s commitment to an essentialist theory of many concepts
coheres well with the following picture of how a general purpose
mechanism could explain word learning. Essentialism is adaptive
(p. 153) because regularities in the world are organized around
“sources of correlation” (Ryder, submitted) or “substances” (Mil-
likan 1998; 2000). These are entities in the environment, for ex-
ample, real kinds and individuals, that have a set of correlated fea-
tures where this correlation occurs for a reason. Essentialism
encourages the conceptual tracking of sources of correlation. Con-
cepts allow one to re-identify kinds or individuals as the same
again, so that you can infer the presence of their currently unob-
servable features, and learn more about them in order to facilitate
re-identification in the future (Millikan 1998; 2000). This is what
explains successful learning and induction (Bloom, p. 153). Obvi-
ously, it is to an organism’s advantage to learn as many facts about
a source of correlation as possible, to permit more induction and
easier re-identification.
Now let us apply this idea to word learning. In order to believe
that “cat” refers to cats, one needs three concepts: a concept of the
word “cat,” a concept of reference, and a concept of cats. Bloom
is a psychological essentialist about the third concept; suppose that
this is the correct stance to adopt towards the other two as well –
after all, the word “cat” and intentions to refer are both kinds that
have a set of features that are correlated for an underlying reason
(social convention in one case, and human psychology in the
other). According to the hypothesis bruited above, our cognitive
system is set up to learn how to re-identify these kinds: here’s a
“cat” again, here’s a cat again, and here’s a case of reference again.
One of the regularities these kinds participate in, and thus one way
of re-identifying them, is that “cat” is used to refer to cats. Learn-
ing this fact will result from the general (essentialist driven) push
to learn to re-identify each of these sources of correlation. In do-
ing so, the general cognitive strategy will be to make use of what-
ever cues are available to help re-identify “cat”s, cats, and in-
stances of reference. (Note that the learner need not possess any
deep knowledge of the essences or individuative principles un-
derlying the things it learns about, a possibility that Bloom ac-
knowledges for some things [e.g., particular kinds, p. 168] but,
puzzlingly, not others [e.g., individuals vs. kinds].)
Now, it emerges very clearly from Bloom’s book how resource-
ful children are in learning words. It seems that if there is a valid
cue available – whether artifact function, shape, substance, spatio-
temporal history, others’ intentions, object coherence, familiarity,
importance, or syntactic context – at some point in her develop-
ment, a child will make use of it. (Most of the cues Bloom rules
out would actually constitute very poor evidence for what a word
means.) This is just what one would expect if the above general
purpose learning story were true.
All that is left to worry the empiricist is the apparent belief in
essentialism that drives the whole process. We have proposed that
there is no such belief, but rather that this cognitive habit emerges
from a simple circuit that forms the basic building block for the
entire cortical network (Ryder & Favorov 2001; Favorov & Ryder,
submitted).
The other way to learn the meaning of a word
Sam Scott
Cognitive Science Department, Carleton University, Ottawa, ON K1S 5B6,
Canada.
sscott@ccs.carleton.ca
http://www.chat.carleton.ca/~sscott2/sam
Abstract: Bloom’s book can be viewed as a long argument for an anti-
Whorfian conclusion. According to Bloom, word learning is usually a
process of mapping new words to pre-existing concepts. But an exception
to this generalization – the learning of words from linguistic context –
poses a problem for Bloom’s anti-Whorfian argument.
Bloom’s book is a well-supported, well-articulated, and very wel-
come argument against Whorfian linguistic relativism. But there
Commentary/Bloom: How Children Learn the Meanings of Words
BEHAVIORAL AND BRAIN SCIENCES (2001) 24:6 1117
may still be an important gap in the argument. According to
Bloom, word learning is usually a process of mapping new words
to pre-existing concepts,
1
and most of his specifically anti-Whor-
fian comments assume that this is the case. But there seems to be
another way to learn the meaning of a word. Simply hearing a word
in context, without a definition or obvious referent present, is in
many cases enough to at least begin the process. The interesting
thing about this type of word learning is that the concept appears
to be acquired along with (or perhaps even because of) the word
being learned.
Towards the end of his book, Bloom (2000, pp. 250–54) explic-
itly lays out some examples of the different ways to learn the mean-
ing of a word. The first two are fairly uncontroversial examples of
cases in which new words are mapped onto existing concepts (the
“fendle” and “grommet” examples). At the other extreme, Bloom
considers the hypothetical case of someone learning the meaning
of the hockey term “hat trick” by being told that “a hat trick is when
someone scores three goals in a row” (p. 254). With this defini-
tional statement, the person learning the word is explicitly being
handed a concept as well. This process of simultaneous word
learning and concept acquisition can only occur because of the lin-
guistic abilities of the teacher and the learner. Bloom correctly
points out that this is not an embarrassment for the anti-Whorfian,
since: “It is not the form of the language that causes the concepts
to emerge; it is the content that the language conveys.” (p. 254, his
emphasis).
The problem with the “hat trick” example is that it is far too sim-
ple. The word “hat trick” seems to have a pretty good definition,
and that definition is what is being provided to the person learn-
ing the word. So it is plausible that the learner is not acquiring a
new simple concept hat_trick,
2
but is forming a compositional
concept something like three goals in a row, which then gets
mapped to the term “hat trick.” In this case, it is quite clear that
there is nothing Whorfian going on. In fact, the learner has not ac-
quired any genuinely new conceptual content at all, just a conve-
nient way of referring to a particular combination of his or her ex-
isting concepts.
But there are at least two reasons why this process of composi-
tional concept acquisition (if that is indeed what is going on in this
case) could not be very common. First of all, there is no empirical
support for the idea that lots of concepts have a definitional struc-
ture, and as a result, most researchers have abandoned defini-
tional theories of concepts altogether (Laurence & Margolis
1999). If it turns out that we need to sneak definitions back in for
linguistically acquired concepts, it will be a bit of an embarrass-
ment (also see Scott 2002).
Secondly, there are many concepts that seem to be acquired
through language, but in a non-associative and non-definitional
way. As an example, consider a child learning a word like “debate”
by hearing it used repeatedly in conversation, but without an im-
mediately obvious referent or definition to pair it with. I think
Bloom would probably agree that a concept (not just a word) is be-
ing acquired in this example, since he comments in connection
with the “debate” example that “nobody would deny that children
can learn at least some words from hearing them used in conver-
sation,” and adds that “the precise nature of this learning process
is a mystery” (Bloom 2000, p. 194, my emphasis).
Despite the lack of empirical evidence about this type of word
learning, it is pretty clear what we have to say about it. Given
the large number of words that must be learned this way, and
given the lack of an obvious definition being provided, this type
of exposure to new words must cause the child to instantiate a
new, non-definitional concept place-holder (Bloom might prefer
“essence-holder”) to map to the new word (see also Millikan
1998). Prototypes or beliefs about the new concept can be formed
and stored later, but the concept must be there first, or it can play
no role in the child’s mental life. Of course it is possible, for in-
stance, that the child has a fully formed debate concept just wait-
ing for an appropriate word (Fodor 1998a). But this seems fairly
unlikely, and in any case, the child would have no way of knowing
immediately that the word “debate” should be used to name this
pre-existing concept.
So it seems that large numbers of non-definitional concepts
must be at least partially formed by exposure to conversational lan-
guage alone. If so, then this type of linguistic concept formation
needs to be specifically addressed to keep the anti-Whorfian ar-
gument afloat. My only suggestion in this regard is that we do not
need to cling to the idea that concepts are always (or even often)
prior to language. All we need is a demonstration that linguistic
ability itself does not give rise to or meaningfully constrain the
ability to form concepts. Bloom’s book stands as an impressive col-
lection of empirical evidence that points towards this conclusion.
But we are not quite there yet.
ACKNOWLEDGMENTS
Thanks to Paul Bloom for helpful discussion, and to Rob Stainton for read-
ing an early draft.
NOTES
1. Following Laurence and Margolis (1999), a concept is a sub-propo-
sitional mental representation. I believe that this definition is mostly com-
patible with Bloom’s usage.
2. Words in capitals refer to concepts; words in double quotes refer to
lexical items.
Children request teaching when asking for
names of objects
Sidney Strauss and Margalit Ziv
School of Education, Tel Aviv University, Tel Aviv 69978, Israel.
sidst@post.tau.ac.il mziv@post.tau.ac.il
Abstract: We propose that in addition to children’s requests for word
names being a reflection of an understanding of the referential nature of
words, they may also be requests for adult’s teaching. These possible re-
quests for teaching among toddlers, along with other indications, suggest
that teaching may be a natural cognition that may be related to the devel-
opment of theory of mind.
In analyzing the process of learning word meanings by children,
Bloom assigns a central role to children’s developing theory of
mind. According to this account, very young children have the
ability to use mentalistic cues and figure out adults’ referential in-
tentions. As Bloom recognizes, in the normal course of learning
the meaning of words, young children not only infer others’ in-
tentions when others speak, gesture, and so on, but they also ac-
tively request adults to name objects for them. In so doing, tod-
dlers may be initiating teaching moments. We now explore this
possibility.
It has been proposed (Strauss et al., in press a) that teaching,
like language, can be regarded as a natural cognition. The combi-
nation of several points may support this notion. (1) Teaching
without theory of mind may be found among non-primates (Caro
& Hauser 1992; however, teaching with theory of mind may be
species-specific to humans (Premack & Premack 1996). (2) Teach-
ing is universal among human beings. (3) The vast majority of hu-
mans spontaneously engage in teaching even though they have not
been taught how to teach, but merely exposed to it. (4) Teaching
is remarkably complex. It involves myriad and multifaceted men-
tal processes and assumptions about others’ minds and how learn-
ing takes place in their minds. (5) The visible part of teaching is
the external behaviors teachers exhibit. The assumptions and
complex mental processes made by teachers while teaching are in-
visible, and they cannot be inferred from the visible part of teach-
ing. (6) Already during their preschool years, young children show
attempts to teach (Ashley & Tomasello 1998; Strauss et al., in press
b; Wood et al.1995). In concert, these reasons suggest that teach-
ing may be a natural cognition.
Frye & Ziv (in press) suggest that interpreting teaching in the-
Commentary/Bloom: How Children Learn the Meanings of Words
1118 BEHAVIORAL AND BRAIN SCIENCES (2001) 24:6
ory of mind terms highlights its two main components: the differ-
ence in states of knowledge of two parties as a prerequisite for
teaching, and the intention to reduce the difference in knowledge
by enhancing the knowledge, or understanding of the learner. Al-
ready at the age of 3 years, children can recognize that in order for
teaching to occur there is a need for a knowledge gap between
teacher and learner, or, in other words, that a knowledgeable per-
son accompanied by an ignorant learner are the prerequisites for
teaching. Several recent studies have investigated preschoolers’
teaching strategies and suggested that children’s teaching was re-
lated to their theory of mind understanding (Astington & Pelletier
1996; Wood et al. 1995). It wouldn’t be surprising to hear a 3-year-
old child say: “Dad, teach me (or show me) how to put this toy to-
gether.”
There may be a possibility that a year earlier, toddlers at the age
of 2 begin to realize that they can produce object naming on the
part of others and, thus, have some understanding of the prereq-
uisites of teaching. They may, for example, have some sensitivity
to their own lack of knowledge and to the adult’s different knowl-
edge status that enables satisfying their request. According to this
interpretation, the child may also have implicit knowledge that
his/her request may result, on the adult’s part, in an intentional
reference to the specific object the child herself pointed, or re-
ferred to.
Another possibility is in line with what Bloom claims about the
implicit reasoning stages underlying children’s inferences about
thoughts of others. Here the implicit reasoning process involved
in a request for intentional teaching of objects’ names may be the
following:
1. Objects have names, or words that refer to them (based on
previous experience/knowledge about words).
2. I don’t know the word referring to this object.
3. Adults know the word referring to this object (213 – aware-
ness of the knowledge gap).
4. If I point to this object the adult will pay attention to it, too
(joint attention, social referencing).
5. If I ask “What’s that?”, the adult will name it (initiating some-
one else’s intentional teaching).
This analysis suggests that the origins of understanding teach-
ing as a natural cognition, specifically, beginning to appreciate the
knowledge gap between the self and others, should be empirically
investigated already in toddlers, and that requests for objects’
names may provide a natural context for this exploration. Fur-
thermore, exploring toddlers’ emerging awareness of their own
and others’ knowledge may contribute to the understanding of the
early developmental stages of what develops during the preschool
years to children’s theory of mind.
Could we please lose the mapping metaphor,
please?
Michael Tomasello
Department of Psychology, Max Planck Institute for Evolutionary
Anthropology, D-04103 Leipzig, Germany.
tomas@eva.mpg.de
http://www.eva.mpg.de/psycho/
Abstract: Although Bloom gives more credit to social cognition (mind
reading) than do most other theorists of word learning, he does not go far
enough. He still relies fundamentally on a learning process of association
(or mapping), neglecting the joint attentional and cultural learning skills
from which linguistic communication emerges at one year of age.
There are many things to like about this book. Most importantly,
Bloom argues and presents evidence that learning words requires
no special learning principles or specifically linguistic “con-
straints” but that, instead, it requires only general learning
processes along with an understanding of other people’s minds.
Coming from a confirmed linguistic nativist, this proposal will
hopefully be the end of the misguided search for a priori, specifi-
cally linguistic word learning constraints and principles (see Nel-
son 1988 and Tomasello 1992 for earlier arguments along these
lines).
However, the welcome focus on mind-reading (social cogni-
tion) as fundamental to the word learning process is not taken far
enough. Bloom still retains the mapping metaphor – basically just
associationism – as fundamental to word learning (as do the ma-
jority of word learning theorists). Here is Bloom’s summary (2000,
p. 17):
To know the meaning of a word is to have
1. a certain mental representation or concept
2. that is associated with a certain form
Under this view, two things are involved in knowing the meaning
of a word – having the concept, and mapping the concept onto the
right form.
But what is the nature of this association or mapping? Where is
the understanding of minds here? As with many word learning
theorists, Bloom’s appeal to social “cues” from other persons or
their minds is only on the surface – to help identify particular ref-
erents in particular circumstances. However, as I have argued in
many places, a word is an intentional phenomenon through and
through and this must be reflected in the fundamentals of the
learning process itself, not just in “cues” (Tomasello 1992; 1995a;
1999; 2000b; 2001).
To see this most clearly, let us ask the question of why children
start learning language at the specific age that they do, that is, at
around one year of age. Could it be due to newly emerging skills
of associative learning or mapping? No, because human infants
are very good at associative learning from several months of age
onwards (Haith & Benson 1997). Could it be due to word learn-
ing constraints or principles that emerge at one year of age? The
problem here is that there is no independent way to observe or
measure constraints; they are only inferred from the child’s lin-
guistic behavior, after the fact (this is why Smith [2000] refers to
word learning constraints and principles as “skyhooks”). Could it
be due to newly emerging skills of speech perception? Not really,
because, although infants are gaining speech skills during this age
period, they are clearly able to perceive and recognize particular
isolated words from at least six months of age onwards (Juszyk
1999). Other hypotheses that Bloom considers and rejects in-
clude: syntactic knowledge, changes in parental speech, the mo-
tor control required for speech, phonological knowledge, mem-
ory, conceptual abilities, and theory of mind. His conclusion is that
“In the end, nobody knows why word learning starts at about 12
months and not at six months or three years” (p. 45).
But I know why. Word learning begins when it does because it
depends on a more fundamental social-cognitive skill, namely,
the ability to share attention with other human beings – which
emerges in nonlinguistic form near the end of the first year of life.
Thus, many different studies have found that infants begin to de-
velop joint attentional skills at around 9 to 12 months of age, in-
cluding such things as following the gaze direction and gestures
of adults, imitating adult actions on objects, and directing adult
attention to outside objects using various kinds of gestural signals
(see Tomasello 1995b, for a review). Children also show their first
signs of comprehending language at around the same age, with
the first linguistic productions coming soon after (Fenson et al.
1994).
Most importantly, Carpenter et al. (1998) found that children’s
initial comprehension and production of language correlated
highly with their skills of joint attentional engagement with their
mothers (i.e., their ability to engage in relatively extended bouts
of attention directing and sharing). Indeed, they found that
roughly half of the variability in the sizes of infants’ word com-
prehension and production vocabularies was predicted by the
amount of time (and style with which) infants spent in joint atten-
tional interaction with their mothers during a 10-minute observa-
tion session (see Tomasello 1988, for a review of similar findings
Commentary/Bloom: How Children Learn the Meanings of Words
BEHAVIORAL AND BRAIN SCIENCES (2001) 24:6 1119
at older ages). The reason that linguistic skills are so highly corre-
lated with joint attentional skills is simply that language is itself
one type – albeit a very special type – of joint attentional skill. A
language is a set of historically evolved social conventions by
means of which intentional agents attempt to manipulate one an-
other’s attention.
Appeals to association or mapping do not help us to understand
how children learn to use these social conventions. For instance,
if we suppose that the child “maps” a novel word onto an external
object, it is difficult to explain how she can then learn different
words for the same object, such as Fido, dog, animal, pet, or pest.
Perhaps one might say, as Bloom does in some places, that the
child “maps” the word onto a concept. But not all word learning
involves pre-existing concepts; much recent work shows that chil-
dren sometimes learn concepts as a result of being exposed to
words (Bowerman & Choi 2001). What would the mapping pro-
cess look like in these cases?
In my opinion, the process is best conceived as one of estab-
lishing joint attention, in which the child must understand not just
the adult’s intentions to some outside entity but rather his inten-
tions toward her attention to some outside entity; that is, the child
must understand the adult’s communicative intentions and then
engage in a process of cultural learning in which she aligns herself
with those intentions (Tomasello 1998a; 1999; 2001). A dog (or 6-
month-old infant) may associate or map the sound “dinner” onto
the object “food,” but this does not constitute an intersubjectively
understood linguistic symbol used to direct and share attention
with other persons – so it is not word learning.
Thus, suppose that I am a one-year-old child encountering an
adult making funny noises at me. What am I to make of this odd
behavior? Perhaps nothing – perhaps it is just noise. But given that
I have previous experience interpreting the adult’s nonlinguisti-
cally expressed communicative intentions – in her behaviors such
as pointing to and showing me things and events – I might decide
that she is making these funny noises in an attempt to direct my
attention to something. If I am lucky, I may figure out precisely
what entity or event in the world she is attempting to direct my at-
tention to. How is this mapping?
What we have here are not two things – a word and an object –
being associated or mapped, but one person using a symbol (sig-
nifier) to indicate for another person some entity, situation, or ac-
tivity (signified). Until word learning theorists understand this
fundamental point and incorporate it into their theories – rather
than effacing it with associationistic metaphors – their theories
will continue to confuse processes of intention reading and cul-
tural learning with those of association and mapping, and they will
not be able to explain, among other things, why language emerges
at the age that it does.
Could we please lose the mapping metaphor, please?
Words, grammar, and number concepts:
Evidence from development and aphasia
Rosemary Varley
a
and Michael Siegal
b
a
Department of Human Communication Sciences, University of Sheffield,
S10 2TAUnited Kingdom;
b
Department of Psychology, Western Bank,
University of Sheffield, United Kingdom S10 2TP.
R.A.Varley@sheffield.ac.uk M.Siegal@sheffield.ac.uk
Abstract: Bloom’s book underscores the importance of specifying the role
of words and grammar in cognition. We propose that the cognitive power
of language lies in the lexicon rather than grammar. We suggest ways in
which studies involving children and patients with aphasia can provide in-
sights into the basis of abstract cognition in the domain of number and
mathematics.
Writing in the tradition of Brown (1973) and Macnamara (1982),
Bloom provides a thorough review of research on how children
learn the meanings of words. The rich texture of his book testifies
to the diversity and depth of research in this area, and its implica-
tions for our understanding of how the mind of the young child
works. Bloom claims that there is no mechanism that is uniquely
dedicated to children’s word acquisition, but that it is built on abil-
ities that exist for other purposes such as theory of mind. He pro-
ceeds to address the impact of words and concepts in domains
such as numerical reasoning, and draws upon the evidence from
a range of conditions such as aphasia and deafness to examine the
interplay between language and cognition. In this commentary,
we examine the issues of words, grammar, and concepts and how
they might reconfigure the human mind.
In the cognitive sciences, there are increasingly frequent claims
that certain forms of reasoning can only be performed though ac-
cess to the resources of the language faculty. In the recent re-
search of Spelke and her colleagues, a link between language and
cognition (e.g., in the form of numerical and spatial reasoning) has
been demonstrated. For example, the ability to combine sources
of visuo-spatial information has been reported to depend on lan-
guage (Hermer-Vasquez et al. 1999). Moreover, exact arithmetic
addition calculations have been shown to be associated with a lan-
guage representational format, whereas estimations of magnitude
are language independent (Dehaene et al. 1999).
However, there remains a critical need to establish the relative
contributions of the components of language – the grammar and
the lexicon – to cognitive operations. Some take the view that
grammar is crucial in many sophisticated cognitive capacities such
as theory of mind reasoning (e.g., Carruthers 1996; de Villiers &
de Villiers 2000), whereas others maintain it is the lexicon – the
pairing of concepts with linguistic forms – that configures some
aspects of human cognition. Bloom describes two competing
claims in the domain of numerical cognition: that of Chomsky
(1988), who maintains that grammar provides a rule-based blue-
print for the potentially recursive combination of individual units
with potentially infinite outputs, and the alternative claim that
number words create the potential to develop a mathematical fac-
ulty that extends beyond the numerical capacity apparent in pre-
verbal infants and some non-human species (Sulkowski & Hauser
2001; Wynn 1998).
Evidence from aphasia provides important insights on the role of
language in cognition, although the evidence is limited to the role
of language in a mature cognitive system rather than in the initial
configuring of the system. The relation of grammar to cognition can
be determined from the performances of people with severe agram-
matic aphasia on behavioural tasks, while the role of lexical knowl-
edge can be established through cases of global aphasia where the
system of word forms and meanings is itself profoundly impaired.
Studies on theory of mind and causal reasoning in severe agram-
matic aphasia have shown that the cognitive power of language lies
not in the grammar (Varley 2002; Varley & Siegal 2000; Varley et al.
2001) as reasoning is retained in such instances. These studies
prompt a shift in the language and thought debate from the relation
of thought to an undifferentiated language faculty, to the more spe-
cific relation of the role of the lexicon in thinking.
Bloom sets out an agenda for future investigation of the nu-
merical and mathematical abilities of people with aphasia. The
challenge is to demonstrate, first, the extent to which the number
faculty is retained in the absence of grammar, much as is the case
for theory of mind and causal reasoning, and, second, to deter-
mine whether patients with number word processing problems
are capable of dealing with numerical problems beyond the abil-
ity to estimate and discriminate small numbers that lie within the
capacity of preverbal infants. Bloom’s hypothesis of number words
creating a capacity for mathematics is strictly developmental, with
progression from small numerosities, to the acquisition of number
words, leading in turn to increased mathematical understanding.
In this respect, number concepts once acquired can be mapped
to different surface symbolic representations. They can take the
form of number words or other forms of numerical notation (such
as Arabic or Roman numerals). In the established system, having
Commentary/Bloom: How Children Learn the Meanings of Words
1120 BEHAVIORAL AND BRAIN SCIENCES (2001) 24:6
acquired numerical symbols, the language scaffolding (i.e., the
number words) may be removed – as occurs in cases of aphasia –
but the capacity for reasoning may be retained in the form of
numerals that sustain calculation. A similar situation might be
observed in the case of a dissociation in the domain of music be-
tween the ability to name a note and to understand its symbolic
musical value (Luria et al. 1965). Just as Bloom concludes that
people without words are capable of rich mental lives because of
non-linguistic conceptual structures, so might already established
number concepts sustain mathematical reasoning despite im-
pairment of surface word forms.
Bloom’s survey rightly emphasizes the role of language in the
acquisition of abstract concepts such as numbers. While many ob-
ject and artifact names correspond to things with vivid perceptual
features, others are abstract and have little existence outside of
the language faculty. “Bees” differ from “beliefs,” and “dogs” from
“democracy.” The investigation of numbers and other abstract
concepts in aphasia may provide a window on the role of word
forms and concepts in cognition, and the sustainability of such
concepts without the associated language form. Bloom shifts the
language and thought debate from grammar to the lexicon, but
also provides elegant illustrations of how grammatical structure
provides one ingredient among the cues that support word learn-
ing and related conceptual development. However, once estab-
lished, the role of grammar as a facilitator of cognition may di-
minish or “switch off.” In our view, it is the facility to construct and
manipulate surface symbolic representations such as words and
numerals that characterizes the huge adaptive advantage of hu-
man cognition and cultural transmission. An issue that cries out
for investigation is to determine whether the loss of such repre-
sentational systems in global aphasia necessarily accompanies im-
pairment in numerical reasoning.
Word extension: Akey to early word learning
and domain-specificity
Sandra R. Waxman
Department of Psychology, Northwestern University, Evanston, IL 60208.
s-waxman@northwestern.edu
http://www.psych.northwestern.edu/People/Waxman.htm
Abstract: Bloom provides a masterful synthesis of recent advances in
word-learning, placing them within the framework of abiding theoretical
issues. I will augment and challenge his approach by underscoring the sig-
nificance of word extension for questions concerning (a) the origin and
evolution of infants’ expectations, and (b) domain-specificity in word-
learning.
1. Overview.
Word learning, perhaps more than any other de-
velopmental achievement, stands at the crossroad of conceptual
and linguistic organization. Consequently, the very best work en-
gages key theoretical issues from each of these realms. This vital-
ity on the theoretical front has been matched, pace for pace, by
impressive advances on the empirical front. A major challenge in
describing how children learn the meanings of words is to synthe-
size these empirical advances while placing them squarely within
the context of guiding theoretical issues. Bloom has met this chal-
lenge masterfully. I will amplify and also challenge his approach
by underscoring the significance of word extension in theories of
acquisition.
2. The significance of extension.
Many assume that word-
learning is successful when a learner applies a new word to the
same object named by her caregiver. This phenomenon, some-
times known as “fast mapping,” is essential. But successful word-
learning requires more. Rather than (merely) mapping words to
the objects on which they were introduced, learners must extend
novel words systematically (Waxman 1999; in press; Waxman &
Booth 2001a).
There is an interesting wrinkle here: Different kinds of words
highlight different aspects of experience. Consequently, each
grammatical form supports a unique pattern of extension. In En-
glish, count nouns are typically extended on the basis of object cat-
egories; adjectives on the basis of object properties (e.g., color, tex-
ture); and proper nouns are not extended beyond the named
individual. This is more than a descriptive fact; the relation be-
tween grammatical form and meaning has important logical con-
sequences. Patterns of extension are tied to the semantics of gram-
matical forms. In particular, any count noun provides a logical
means for tracing the identity of a designated individual within a
category or kind (Macnamara 1986).
a. The origin and evolution of infants’ expectations.
How do
learners solve this wrinkle? Which of these expectations, if any, do
infants recruit at the onset of acquisition, and how do these evolve
over development? I have proposed (1) that infants embark upon
the task of word-learning equipped with a broad, universally-
shared expectation linking novel content words (independent of
grammatical form) to a wide range of commonalities among
named objects; (2) that this initially general expectation is subse-
quently fine-tuned as infants detect the specific correlations be-
tween the surface cues for particular grammatical forms and their
associated meanings in the native language; and (3) that as infants
develop this more specific set of expectations, they first distin-
guish (count) nouns (as compared to other grammatical forms)
and map them specifically to category-based commonalities; other
expectations (e.g., linking adjectives to properties) emerge later
(Waxman 1999b; in press).
Infants’ broad initial expectation has been demonstrated by 9 to
12 months (Balaban & Waxman 1997; Waxman & Booth 2001c;
2002; Waxman & Markow 1995), early enough to guide the
process of word-learning from the very outset. Following Brown
(1958b), I have argued that at this early point, words serve as in-
vitations to form rich inductive categories: providing a common
name for a diverse set of objects directs infants’ attention to com-
monalities among them, and initiates in infants a search for
deeper, perhaps non-obvious, commonalities among named enti-
ties.
A more specific expectation, linking nouns to object categories,
emerges a few months later, at approximately 14 months (Waxman
1999a; Waxman & Booth 2001b). This developmental unfolding
accords well with current theories of acquisition which are predi-
cated on an assumption that learners must first be able to map
nouns to objects and categories of objects (Gentner 1982; Gleit-
man 1990; Grimshaw 1994; Huttenlocher & Smiley 1987; Marat-
sos 1998; Pinker 1984; Talmy 1985; Waxman 1999b). Indeed, the
argument is that the acquisition of these other grammatical forms
is grounded in the prior acquisition of nouns.
b. Word-learning: Domain-general or domain-specific principles?
There is a broad consensus that language acquisition is supported
by powerful principles of organization within the learner. Consid-
erable debate remains as to whether these principles are dedi-
cated specifically to language or whether they are domain-general
cognitive abilities. Bloom and his colleagues have made a strong
claim for the domain-general view (Bloom & Markson 2001;
Markson & Bloom 1997). Booth and I have challenged this claim
on both logical and empirical grounds (Waxman & Booth 2000;
2001a). Word extension figures importantly in this discussion,
which began with an examination of “fast-mapping.” Evidence re-
vealed that children successfully map not only novel words (e.g.,
“This is a koba”) but also novel facts (e.g., “My uncle gave this to
me”). This indicates that fast mapping is not specific to word learn-
ing. It does not, however, constitute “evidence against a dedicated
system for word-learning” as originally claimed (Markson &
Bloom 1997).
The problem is the “slippage” between the scope of the evi-
dence (that fast-mapping is not specific to word-learning) and the
more sweeping claim (that word-learning is the product of do-
main-general cognitive abilities). Although fast-mapping figures
in the acquisition of both words and facts, this does not mean that
Commentary/Bloom: How Children Learn the Meanings of Words
BEHAVIORAL AND BRAIN SCIENCES (2001) 24:6 1121
these two domains rest upon the same set of underlying princi-
ples. By analogy, discovering that one ingredient (whipped egg-
whites) is involved in two recipes (lemon meringue and cheese
soufflé) does not mean that the recipes share any other ingredi-
ents. The question is whether, in addition to their shared compo-
nents, there are also distinct principles invoked in acquiring words
versus facts.
We have documented at least one such distinction – in chil-
dren’s patterns of extension. We taught children either a novel fact
or word for an object, and then examined their extensions beyond
this object. The difference was dramatic. Children spontaneously
and systematically extended nouns beyond the designated object
to other members of the same object category, but revealed no sys-
tematicity in extending novel facts. This sharp contrast illustrates
that there are indeed principles invoked in the acquisition of
words that are not invoked for facts. Acquiring the appropriate ex-
tension for a word is actually a very different matter than doing so
for a fact. The extension for any novel content word can be deter-
mined (roughly) by its grammatical form. These distinct patterns
of extension do not rely upon previous knowledge about the par-
ticular word or the particular object to which it has been applied.
In contrast, determining the extension of a fact depends crucially
on knowledge about the kind of fact (e.g., enduring or transient)
and the kind of object (e.g., animate or inanimate).
3. Conclusion.
Word extension provides a window through
which to view the evolution of word-learning and its consequences
for language and conceptual development.
A multiplicity of constraints: How children
learn word meaning
Chris Westbury and Elena Nicoladis
Department of Psychology, University of Alberta, Edmonton, Alberta T6G
2E5, Canada.
chrisw@ualberta.ca
http://www.ualberta.ca/~chrisw/Westbury.html elenan@ualberta.ca
http://www.ualberta.ca/~elenan/
Abstract: This book is an excellent and accessible overview of the position
that children learn the meanings of words by applying a variety of nonlin-
guistic cognitive tools to the problem. We take issue with Bloom’s empha-
sis on Theory of Mind as an explanatory mechanism for language learning;
and with his claim that only unitary objects are nameable.
Psychological experiments reveal that human infants are capable
of remarkable cognitive feats within the first few days of life, but
most of these pass unnoticed outside of the lab. Most of us discern
little behavior in young babies that we would not expect a skilled
animal trainer to coax from a pigeon or a rat. Even the first tenta-
tive signs of language can be dismissed as mere parroting. At some
point in the first year of life, it becomes clear that the healthy in-
fant has a mind of his or her own. The child begins to participate
actively in the symbolic world, demonstrating an ability not sim-
ply to repeat words in appropriate contexts, but also to understand
what they mean.
How infants are able to master this complex task with such ap-
parent ease is the subject of Paul Bloom’s book – at least, its al-
leged subject. In his own final-chapter summary, Bloom baldly
states that “Nobody knows how children learn the meaning of
words” (p. 262). Though it is true that no one yet knows how chil-
dren learn words, Bloom makes a wonderful guide to how Bloom
believes they do. What he believes is worth knowing. A unifying
theme is the idea that word-learning is not a unitary phenomenon
at all, but rather a dynamic set of processes involving a multitude
of disparate abilities. Bloom implies that acquisition of word
meaning will ultimately be explained in terms of the same kind of
complex set of biologically-based, evolved mechanisms by which
cognitive scientists are coming to explain other cognitive abilities.
One way Bloom addresses this complexity is by presenting data
suggesting that there are different ways to learn different parts of
speech. He devotes individual chapters to discussing acquisition
of the meanings of concrete nouns, pronouns and proper nouns,
verbs, visual representations, abstract nouns, and number words.
He is careful to point out that by acquiring new skills and words,
the child’s mind becomes different. It is not the same mind that
acquires the first words and that acquires derivational morphol-
ogy. There is not one single problem of “how children learn the
meanings of words” but many distinct problems.
One criticism we have is that, in the early chapters, Bloom is
slippery about when language starts, in large part because he
sometimes uses the term “language” to refer to production and
sometimes to comprehension, though the starting points of com-
prehension and production are quite distinct. This slipperiness
could be particularly confusing to readers unfamiliar with lan-
guage acquisition.
Although we are sympathetic to Bloom’s general approach, we
are not convinced by the emphasis he lays on Theory of Mind
(ToM) as an explanatory mechanism for language learning. There
are two reasons for our skepticism. One is that much of what
Bloom wishes to attribute to ToM might more appropriately and
more specifically be attributed to pragmatics. Although, clearly,
pragmatic understanding and ToM are closely related, they are not
identical. In attributing so much to ToM, Bloom abandons the
fine-grained functional distinctions he has made so much of else-
where in his book. A second problem is that ToM stands as much
in need of explanation as the learning of word meaning. It is dis-
satisfying to explain one mystery in terms of another. We also feel
it is unnecessary in this case. Although it is certainly true that nei-
ther ToM nor pragmatics are yet fully understood, a large litera-
ture has subjected them both to the same kind of fine-grained
functional decomposition that Bloom is undertaking for learning
of word meanings. A presentation of some of the functional com-
ponents contributing to the development of ToM, pragmatics, and
the learning of word meaning might have given readers a better
appreciation of how they all may be underlain by similar functional
primitives. We take issue in particular with the claim that “Theory
of mind is the only area, other than language itself, in which [chim-
panzees] are manifestly inferior to human beings” (HCLMW,
p. 85). It is very clear that, for example, context-shifting under in-
ternal control and the related idea of inhibition of response show
a similar – and relevant – disparity (Diamond 1988; 1991).
A brief analysis of functional primitives contributing to ToM
would have allowed some reference to the underlying neurology
and to its possible evolutionary significance. Many of the identified
functional primitives believed to contribute to the development of
these “abstract” cognitive skills are known to be largely under con-
trol of pre-frontal cortex. This brain region has been uniquely se-
lected in human beings (Deacon 1997) and is widely implicated in
imaging studies in adults of access to word meaning (Binder & Price
2000). If Bloom’s readers were better able to appreciate the wide
range of functions built upon pre-frontally-controlled functional
primitives, some of the mystery of why access to word meaning has
come to be possible at all might have been reduced.
Another area in which we felt Bloom’s overview fell short, was
in his claim (Ch. 4) that the objects underlying what can be named
are unitary. Bloom writes that “a man on horseback is not an ob-
ject because no children could ever learn a word that refers to such
an object” (p. 93). The literature on children’s compound word
learning contains many findings that contradict this claim. Clark
(1981) gave an example of the compound apple juice chair, used
by a child to refer to the chair that was close to the apple juice. An-
other child referred to whiskers as nose-beard (Becker 1994). A
child in Nicoladis and Yin (2001) referred to a girl holding a crab
as the crab girl. Contrary to Bloom’s claims, novel compound
nouns appear in children’s vocabulary very early in development
and denote a variety of relationships between objects (Clark 1981).
In the context of a book about such a large and difficult topic,
these criticisms are mere quibbles. We enjoyed this work, and rec-
ommend it to any reader interested in its topic. It requires little
Commentary/Bloom: How Children Learn the Meanings of Words
1122 BEHAVIORAL AND BRAIN SCIENCES (2001) 24:6
familiarity with linguistics or psychology of language, and could be
used for a high-level undergraduate class on children’s language
acquisition. It is a good summary of a great deal of recent research
in the development of meaning. Even researchers familiar with
the body of work reviewed will find it handy to have as a reference.
It will certainly remain on our bookshelves.
Rational statistical inference: A critical
component for word learning
Fei Xu
a
and Joshua B. Tenenbaum
b
a
Department of Psychology, Northeastern University, Boston, MA 02115;
b
Department of Brain and Cognitive Science, Massachusetts Institute of
Technology, Cambridge, MA 02139.
fxu@neu.edu
jbt@psych.stanford.edu
http://www.psych.neu.edu/People/faculty.shtml
http://www-psych.stanford.edu/~jbt/
Abstract: In order to account for how children can generalize words be-
yond a very limited set of labeled examples, Bloom’s proposal of word
learning requires two extensions: a better understanding of the “general
learning and memory abilities” involved, and a principled framework for
integrating multiple conflicting constraints on word meaning. We propose
a framework based on Bayesian statistical inference that meets both of
those needs.
Bloom argues that four cognitive capacities are necessary and suf-
ficient to account for how children learn the meanings of words
(HCLMW, p. 10): “an ability to infer the intentions of others, an
ability to acquire concepts, an appreciation of syntactic structure,
and certain general learning and memory abilities.” We agree with
many of Bloom’s arguments for the necessity of these components,
but perhaps more cognitive structure needs to be elucidated in or-
der to explain how children solve the inductive problem of word
learning: how to generalize a word from a very limited number of
examples to an infinite set of possible referents. Specifically, more
work is required on two fronts: a better understanding of the “gen-
eral learning and memory abilities” involved, and an integrative
framework for explaining how different sources of constraint in-
teract to guide children’s inductions of word meaning.
To see what is required, consider the following experimental
word learning situation. A puppet is introduced to a 4-year-old
child as having his own names for things. The puppet picks out an
object (e.g., a terrier) and says, “This is a blicket.” The child is then
asked to give the puppet all and only the blickets from a large set
of objects. In Bloom’s account, several sources of knowledge act
to constrain this inductive inference: linguistic knowledge – the
child knows that the puppet just used a count noun; an under-
standing of the world – the child carves up the world in terms of
objects, which belong to different natural kinds; and a theory of
mind – the puppet looks at the object while naming it so that the
reference of the speaker is clear. Yet, these three components are
not sufficient for the child to figure out the meaning of the word
“blicket” here. In particular, how does the child know whether the
word refers to all animals, all dogs, or just all terriers? Each of
these possible meanings is consistent with all of the above con-
straints, and indeed, each corresponds to a word in English that
has to be learned somehow.
One possible answer that Bloom considers is a basic-level bias:
children could assume that words refer to basic-level concepts. In
this case, the child would assume that “blicket” refers to all and
only dogs. However, two problems arise. First, the status of the ba-
sic-level bias is controversial. Callanan et al. (1994) found that the
basic-level bias only showed up for familiar words, but not for un-
familiar words among children. Second, one of Bloom’s arguments
against the special constraints view is that such constraints (e.g.,
Markman’s [1989] mutual exclusivity) in fact limit the kinds of
words a child could learn, and the child would have to override the
constraints later on. The proposed basic-level bias has exactly that
problem, in limiting children to learning only words for basic-level
concepts.
The induction problem becomes more interesting when multi-
ple examples are presented. Suppose the puppet picks out three
blickets, each one a terrier – is this a different situation from the
one in which only a single blicket is presented? Or suppose the
puppet picks out three blickets, one a terrier, the second a poodle,
and the third a German Shepherd, or perhaps it picks out three
blickets, one a terrier, the second one a pig, and the third a peli-
can – what would the learner do when asked to pick out the other
blickets? Each scenario seems to suggest a different generaliza-
tion as a function of the number and distribution of examples. We
found that both adults and children were sensitive to these factors
(Tenenbaum & Xu 2000; Xu & Tenenbaum 2001). Given one ex-
ample, their generalizations tended to be graded with a soft
threshold around the basic level (e.g., dog). Given three examples,
they tended to choose all and only those objects matching at the
level of the most specific concept including all three examples:
when all three examples were terriers, they tended to choose only
the other terriers (subordinate level matches), and likewise when
the examples were spanned only at the basic or superordinate
(e.g., animal) levels. This sharpening and focusing of generaliza-
tion from one example to three examples cannot be explained by
appealing to constraints imposed by learners’ linguistic knowl-
edge, conceptual knowledge, or theory of mind, because the ad-
dition of two new examples does not introduce additional infor-
mation along these dimensions.
An alternative is to appeal to the fourth component of Bloom’s
proposal: “general learning and memory abilities.” We endorse
this move, but much work remains in specifying how these learn-
ing abilities could account for children’s abilities to infer how far
and in what ways to generalize the meaning of a word from very
few examples. To address generalization, these capacities must go
beyond the ability merely to remember words and facts (Ch. 2 of
HCLMW). Neither can we appeal simply to standard general
learning theories, such as those based on connectionism (e.g.,
O’Reilly & Munakata 2000; Rumelhart & McClelland 1986),
which typically require many more than a few training examples,
and both positive and negative examples, before they can gener-
alize meaningfully to new objects.
We propose that word learning, and concept learning and rea-
soning more generally, is guided by domain-general mechanisms
that operate according to the principles of rational statistical in-
ference (Tenenbaum & Xu 2000; Tenenbaum & Griffiths 2001).
These principles can explain why, when given three examples but
not when given only a single example, learners tend to generalize
up to, but no further than, the level of the most specific concept
spanning the examples. If the examples represent a random sam-
ple from the extension of the word, then it would be a suspicious
coincidence if the first three blickets encountered were all terri-
ers when the word “blicket” actually referred to all dogs. This line
of reasoning licenses the inference that “blicket” refers only to ter-
riers. However, given just one example, it would not be so sur-
prising to see a terrier labeled as “blicket” regardless of whether
the word referred to just terriers, all dogs, or even all animals,
which explains why learners do not restrict generalization to just
terriers given only a single example.
We conjecture that this sort of statistical reasoning acts as an im-
portant inductive principle in many aspects of word learning. For
instance, while children often show a “shape bias” – a tendency to
extend novel words to other objects of the same shape (Landau et
al. 1988) – this bias might be overcome by presenting several ex-
amples with rather different shapes but practically identical tex-
tures (Mintz & Gleitman 1998). Then more frequent generaliza-
tions based on texture might be expected, based on the reasoning
that it would be a suspicious coincidence for all examples of a very
general shape-based category to have the same texture if that fea-
ture was not essential to the word’s meaning.
These cases also show the need for some additional cognitive
Commentary/Bloom: How Children Learn the Meanings of Words
BEHAVIORAL AND BRAIN SCIENCES (2001) 24:6 1123
structure beyond the components of syntax, concepts, theory of
mind, and “general learning” that Bloom claims are necessary and
sufficient to account for word learning. In particular, if the statis-
tical force of multiple examples can override a fundamental con-
ceptual bias towards basic-level or shape-based categories, then
the learner must have some principled framework for integrating
these different sources of constraint and adjudicating among
them. We have suggested (Tenenbaum & Xu 2000) that the ma-
chinery of Bayesian inference might provide an appropriate inte-
grative framework, by specifying how a rational learner ought to
combine prior knowledge with input statistics. In brief, the
learner’s existing conceptual structure (e.g., the concepts of ob-
ject, terrier, dog, animal) and her syntactic knowledge (e.g., count
vs. mass noun) might jointly determine the prior probability of dif-
ferent candidate word meanings, while the likelihood of the ob-
served data, given each hypothesis, would be a function of both
the particular objects labeled and theory-of-mind factors such as
joint attention and communicative intent that determine how la-
bels are taken to correspond to objects. Likelihoods and priors are
then combined according to Bayes’ rule to yield the learner’s pos-
terior probability that each hypothesis corresponds to the true
meaning of the word given all the available constraints. This
framework explains why conceptual knowledge, such as the basic-
level or shape biases, might carry more weight when only one ex-
ample of a word has been encountered, than when multiple ex-
amples have been seen; these biases enter into the prior, which is
independent of the number or distribution of examples encoun-
tered, while the force of the examples enters into the likelihood,
which exerts increasing influence as the number of examples in-
creases.
In sum, the inductive challenge of word learning – and chil-
dren’s success at meeting this challenge – seems to require a
Bayesian inference framework, over and above the cognitive ma-
chinery that Bloom has claimed is necessary and sufficient. The
Bayesian inference machinery alone is not sufficient to solve the
problem, but it provides a rational framework in which the power
of “general learning” abilities can be explained, and the conflict-
ing constraints of multiple sources of information can be com-
bined and reconciled.
Author’s Response
Controversies in the study of word learning
Paul Bloom
Department of Psychology, Yale University, New Haven, CT 06520-8205.
Paul.Bloom@Yale.edu http://pantheon.yale.edu/~pb85/
Abstract: How Children Learn the Meanings of Words
(HCLMW) defends the theory that words are learned through so-
phisticated and early-emerging cognitive abilities that have
evolved for other purposes; there is no dedicated mental mecha-
nism that is special to word learning. The commentators raise a
number of challenges to this theory: Does it correctly character-
ize the nature and development of early abilities? Does it attribute
too much to children, or too little? Does it only apply to nouns, or
can it also explain the acquisition of words such as verbs and de-
terminers? More general issues are discussed as well, including
the role of the input, the relationship between words and con-
cepts, and debates over nativism, adaptationism, and modularity.
You could do worse than study word learning. It is a fasci-
nating topic, one that has obvious clinical and educational
relevance and bears directly on major issues in psychology,
philosophy, and linguistics. It is a relatively easy area to
study – researchers typically use simple methods such as
observation of parent-child interactions, and experiments
in which children and adults are taught new words and then
tested as to what they think the words mean. And it is an
easy field to get into. Anyone can become fully conversant
in the major theories and core findings by reading a few
books and a couple of dozen articles.
This last part is not entirely a good thing. For whatever
reason, we know little about how children learn the mean-
ings of words. It is a wide-open field, quite different from
domains such as object recognition and syntactic process-
ing, where there are articulated theories of the relevant psy-
chological mechanisms and representations, and where the
experimental literature is devoted to distinguishing subtle
predictions that these different theories make.
My hunch is that word learning is intrinsically harder to
understand than object recognition and syntactic process-
ing. This is due to its nonmodular nature. If the theory pro-
posed in How Children Learn the Meanings of Words
(HCLMW) is right, then a complete theory of word learn-
ing awaits a complete theory of the mind in general, one
that includes an explanation of how we form concepts and
how we think about the minds of other people.
Still, we can make some progress. HCLMW is an attempt
to explain precisely what goes into the act of learning a
word, leaving open for future research the precise nature of
these learning mechanisms. There are also claims about re-
lated issues, such as children’s understanding of artwork,
the role of language in our appreciation of the infinite na-
ture of number, and the development of object cognition.
The physicist Wolfgang Pauli once sneered at another
scientist’s work by saying “This isn’t right. It isn’t even
wrong.” The commentators were smart, constructive, and
gracious – and most of them paid me the compliment of ar-
guing that at least some of HCLMW really is wrong. There
are those who focus on the proposed mechanisms of word
learning, suggesting that I have attributed too much to the
child, or too little. Some are concerned that I do not do jus-
tice to words other than names for objects. Others hold dif-
ferent views about the role of input, or alternative perspec-
tives about the relationship between language and thought.
I will discuss these in turn, and conclude with some com-
ments about the proper place of word learning in a theory
of language and mind.
R1. Mechanisms of word learning
R1.1. General learning capacities
Word learning is the product of several aspects of the mind
working together. Those emphasized in HCLMW are the-
ory of mind, conceptual structures, and syntactic knowl-
edge, but others include phonological abilities (to learn
words, a person must be able to identity them in the input),
perceptual capacities (a creature with no sense organs
would lack the necessary information to learn words), and
general learning and memory capacities.
Some of the commentators focus on learning and mem-
ory. Racsmány et al. discuss the role of working memory
in word learning, focusing particularly on individual differ-
ences. Jensen approaches intellectual abilities, also from
an individual difference perspective, and he expands on the
observation that word learning is intimately related to more
general cognitive capacities. (In this regard, it may be in-
Response/Bloom: How Children Learn the Meanings of Words
1124 BEHAVIORAL AND BRAIN SCIENCES (2001) 24:6
terestingly different from the learning of the rest of lan-
guage: phonology, morphology, and syntax). These are ex-
actly the sorts of issues that any complete theory of word
learning is eventually going to have to address.
Xu & Tenenbaum propose a Bayesian inference frame-
work, and argue that it is consistent with much of what we
know about word learning, including the bias towards ba-
sic-level categorization, the shape bias, and children’s ap-
propriate use of multiple cues in different contexts.
There is a certain ambiguity in their presentation, how-
ever. They sometimes talk as if they are presenting a distinct
and novel theory of word learning, something above and be-
yond what others have proposed. But for this to be con-
vincing, they need to show that their approach can tell us
something about word learning that we do not already
know, or that does not follow from every other approach. As
of yet, the data that they collect do not meet this criterion.
It is no surprise, for instance, that someone who hears a
word used to describe a terrier, a pig, and a pelican will
think the word means “animal,” and not “dog” – “animal” is
the only interpretation consistent with the data.
A milder proposal, which they also endorse, is more plau-
sible: Word learning is a Bayesian process, and thus all ad-
equate theories must propose learning mechanisms that are
consistent with this. We are all Bayesians, whether we know
it or not.
R1.2. Theory of mind
Many researchers see word learning as the process of es-
tablishing a mapping between a form and a meaning.
Tomasello objects to this. He worries that it is incompati-
ble with what we know about the role of theory of mind in
word learning. According to Tomasello, word learning is
best understood as one person using a symbol to indicate
something for another person – “mappings” is an associa-
tionist metaphor that we would do well to dispense with.
But the notion of a form-meaning mapping is not in-
compatible with theory of mind, is not associationist, and
not a metaphor. It is a banal truth. You have not learned the
meaning of a word until you have stored in memory a map-
ping between a phonological form and a conceptual repre-
sentation. (If you do not like the word “mapping,” you can
replace it with “association,” “relationship,” or “link” – it
doesn’t matter.) I share Tomasello’s enthusiasm for theory
of mind, but it shouldn’t lead us into solipsism. To learn the
English word “dog,” it is not enough to be smart about the
minds of other people. You also need to know something
about English and store this knowledge in memory.
Part of the problem might be that although Tomasello
starts off by correctly quoting me as endorsing a mapping
between a word and a concept, he then moves to attacking
something I do not defend – a mapping between a word and
an actual object in the world. This is a different issue, but,
still, I’ll take the bait: What precisely is the problem with
word-object mappings? Hirsh-Pasek et al. and Gogate
note that when we learn to talk, part of what we learn is
which words refer to which things in the world. Proper
names are the simplest example of this. To learn a dog is
named Fido is to establish a mapping between a word
(Fido), and an object (a dog). This might be done through
grasping a speaker’s intention, but nonetheless, what you
learn is not only about someone’s intention. It is about a dog.
Hirsh-Pasek et al. are mildly skeptical about the role of
theory of mind. They note that the ability to attend to social
cues does not entail rich social cognition. A baby might fol-
low her mother’s gaze because of an innate mechanism, or
through a past history of reinforcement, without any attri-
bution of mental states. They also present some intriguing
new data showing that 10-month-olds fail to use intentional
cues when exposed to a new word in a preferential looking
task.
These findings are presented as a challenge to the theory
in HCLMW, but I don’t see why. One of my central claims
is that you need theory of mind to learn words. Hirsh-
Pasek et al. find that 10-month-olds do not use theory of
mind when exposed to new words. This is just as I would
expect – since 10-month-olds do not learn words.
Children start to learn words by about their first birth-
day, and the theory in HCLMW predicts that at this point
in development, they should show the relevant theory of
mind capacities. And they do; see, for example, Johnson et
al. 1998; Kuhlmeier et al., under review; Moses et al. 2001.
More generally, I agree with Tomasello that the emer-
gence of word learning at about the child’s first birthday is
the result of the emergence of the requisite social capaci-
ties. None of the data presented by Hirsh-Pasek et al. is
inconsistent with this position.
Not everyone agrees that babies do not learn words. The
sharpest challenge to this view is from Gogate, who sug-
gests that more of HCLMW should have been devoted to
discussing research with 6-month-olds. She asks, “Don’t
preverbal infants map words onto referents?”
We can answer that question with a question: If prever-
bal infants are mapping words onto referents, why are they
preverbal? Why is it that 6-month-olds do not speak, and,
except in the most careful laboratory circumstances, show
no understanding of even the most common words? My
guess is that they do not learn words. They might associate
sounds with objects, in much the same way that nonhuman
animals can, but they have no conception of reference, no
understanding that words are symbols that can refer to the
external world. The experiments that Gogate summarizes
are of great interest, but they bear on babies’ conceptual,
phonological, and associative capacities – not word learn-
ing.
R1.3. Conceptual capacities
Gärdenfors agrees that perceptual similarity spaces are in-
adequate as representations of even seemingly simple con-
cepts, and that some version of psychological essentialism
is true. But he suggests that the right approach is not to give
up on similarity spaces, but instead to allow for spaces that
involve nonperceptual, or theoretical, dimensions. For in-
stance, while gold and fool’s gold might be similar on more
perceptual dimensions, they are dissimilar on the more ab-
stract dimension of “atomic structure.” One doesn’t need
“essences” or “theories” as part of our psychological theory,
then – you can replace them with dimensions, and retain a
similarity-space analysis of concepts.
The specifics of this proposal (see Gärdenfors 2000) de-
serve more discussion than is possible here, and Gärden-
fors is surely correct that we do need a theory of similarity
that extends beyond the perceptual – most of the time
when we say that two things are similar, or dissimilar, we are
not talking about physical features. (Consider what it is to
talk about the similarity of politicians, universities, or jour-
Response/Bloom: How Children Learn the Meanings of Words
BEHAVIORAL AND BRAIN SCIENCES (2001) 24:6 1125
nals.) At the same time, though, there is reason to be skep-
tical about the prospects of explaining concepts as entities
clustered together in a multi-dimensional similarity space.
The problem is that nobody has ever found the right di-
mensions.
For instance, it would be nice to show that we think of cer-
tain people as stockbrokers because we see them as similar
in certain regards. And Gärdenfors is correct that the di-
mensions in which they are similar do not have to be per-
ceptual. But they do have to be more basic than the concept
that is represented. You have not made much progress if you
say that all stockbrokers are similar along the dimension of
stockbroker-ness (see also Fodor 1998b) The failure to actu-
ally reduce any real-world concepts to simpler components
is a serious problem with the similarity space approach, at
least as a complete theory of conceptual representation.
R2. Alternative approaches to word learning
Nolan says that babies cannot be learning words through
capacities such as theory of mind, because babies do not
have such capacities. Based on considerations such as our
immature birth, our neural plasticity, and economy consid-
erations from a design stance, she endorses Lorenz’s view
that there was human selection for adaptability over adapt-
edness. Other animals have innate special capacities; hu-
mans do not.
This may be right. Of course, there are arguments, based
on considerations of learnability, and economy considera-
tions from a design stance, that favor the idea that humans
are actually not so different from other animals. These ar-
guments may be right too.
This type of debate has its place, but when it comes to a
matter as empirical as what babies know, a better approach
might be to actually look at research into what babies know.
Such research shows that they know surprisingly much. It
is not an exaggeration to say that the major result from de-
velopmental psychology in the twentieth century concerns
the extraordinary abilities that human possess before they
speak their first words. Much of this work, from dozens of
laboratories around the world, is reviewed in HCLMW. If
Nolan thinks there is a different way to interpret these
data, she should say what it is.
Nolan also objects that I never consider an excellent al-
ternative theory: the “neural constructivist proposal” of
Quartz and Sejnowski (1997). This is true. I also do not con-
sider nativism, connectionism, dualism, empiricism, modu-
larity, or evolutionary psychology. This is because none of
these are theories of word learning. They are broad theo-
ries of the mind, and while they might motivate specific the-
ories of word learning, they do not themselves make any
concrete claims or predictions in this domain.
Nolan goes on to summarize her own specific theory of
learning, motivated by neural constructivism. She suggests,
for instance, that a word such as “dog” starts off being used
very narrowly, as the name for the family pet, and gets in-
creasingly broader in scope. This is the sort of claim that de-
serves serious consideration, and is discussed in Chapters 4
and 5 of HCLMW. In fact, when young children first hear
a word like “dog,” they do not narrowly restrict it; they ex-
tend it to other members of the same basic-level category
(see Waxman).
Ryder & Favorov endorse an empiricist view of word
learning – and see me as a fellow empiricist. This is in part
because at one point I say that learning a word involves
forming “an association” between a form and a concept (see
the reply to Tomasello for more on that point), and in part
because HCLMW endorses the view that word learning in-
volves general learning processes. But word learning, ac-
cording to the analysis in HCLMW, involves specific abili-
ties as well, such as theory of mind and dedicated systems
of object parsing. Since the empiricist is committed to the
existence of “only general inferential abilities,” this theory
is not empiricist.
R3. Do we need special mechanisms of word
learning?
R3.1. Word extension
Waxman reviews her research showing that children start
off with a broad expectation that words map onto objects,
and later converge on narrower category-specific expecta-
tions, such that count nouns refer to objects. She then sug-
gests that children compute the extension of words in a
unique manner, one specific to lexical items. To support this
claim, she summarizes a study that compares the extension
of a word, koba, to the extension of a fact, “the one that my
uncle gave me” (Waxman & Booth 2000). The word, being
a count noun, was extended to other objects of the same
kind as the target object. The fact was not.
As discussed in Bloom and Markson (2001), this is un-
surprising. It is well known that count nouns refer to kinds,
while a fact such as “the one my uncle gave me” pertains to
an individual, not a kind. Of course they were generalized
differently. If Waxman and Booth chose different stimuli,
they would have obtained different results. It is easy to find
a word that behaves like an individual-based fact – consider
a proper name like “Fred.” And it is easy to get children and
adults to generalize on the basis of kind without using a new
word – if you ask them to “find the one that it is the same
kind as that one,” they behave just the same way as they do
if you tell them “find another koba” (Diesendruck & Bloom,
in press). Words just are not that special.
While there is a lot about word learning that could con-
ceivably be language-specific, determining the extension of
a common noun is an unlikely candidate for a dedicated
process. Waxman’s own experiments show that young chil-
dren generalize names on the basis of conceptual structure
– the category that the word maps onto – and not on the
basis of some unspecified word-specific principle. More
generally, one would expect that people would extend
words such as “dog,” “mortgage,” and “game” on the basis
of what they think are dogs, mortgages, and games. What
else could it be?
R3.2. Pragmatics
Westbury & Nicoladis raise the interesting suggestion
that much of what HCLMW attributes to theory of mind is
better seen as falling into the language-specific domain of
pragmatics.
One objection to this is based on parsimony. We know
there is such a thing as theory of mind – reasoning about
the mental lives of other people – and we know that chil-
dren who are starting to learn words possess this capacity to
at least some degree. So the burden of proof is on someone
Response/Bloom: How Children Learn the Meanings of Words
1126 BEHAVIORAL AND BRAIN SCIENCES (2001) 24:6
who claims that something else is needed. This burden gets
stronger given that, as I argued in Chapter 3, the same the-
ory of mind capacities that are involved in nonlinguistic
tasks suffice for word learning (see also Bloom 2002b).
More generally, there may be no such domain as prag-
matics. There is a certain class of emotions that we call
“road rage,” but nobody views road rage as a distinct emo-
tion – it is just the name we give to the anger that is evoked
in a certain situation. Similarly, pragmatics might be the
name we give to the exercise of theory of mind in the do-
main of language and communication. This does not mean
that the study of pragmatics is uninteresting, just that it
should be thought of as an instance of theory of mind, not
an alternative to it.
R4. How can we best explain the learning of
words that are not object names?
Several commentators noticed that HCLMW is preoccu-
pied with objects and their names. This is for several rea-
sons. Object names are very common in children’s early lan-
guage, and children are strongly biased to treat new words
as object names (see also Waxman). Object names are
plausibly viewed as the entry point to the semantics of lan-
guage – when children start to refer, to communicate about
the external world, they do so by talking about objects.
And, finally, it is where the data is. As Harley & Piattelli-
Palmarini point out, there is disappointingly little litera-
ture on the learning of words for nonobservables such as
“party and “nap,” even though such words dominate adult
vocabularies.
At the same time, however, HCLMW includes an ex-
haustive review of all aspects of word meaning acquisition,
including names for parts and collections, adjectives, verbs,
and prepositions. A chapter is devoted to number words
(with implications for quantifiers in general) and another
devoted to names for representations. And the theory itself
applies to the learning of all of these words, although there
are differences in the extent to which different learning
mechanisms apply (syntax is particularly important for
number words, for instance, while theory of mind is highly
relevant for second-person pronouns).
Naigles, on the other hand, doubts that the same sorts
of conceptual constraints that are present for object names
also exist for verbs. She is presumably not denying that
such constraints exist – after all, if children interpret an
event in a trillion ways, each equally salient, verb learning
would be impossible. Naigles’ own research designs (e.g.,
Naigles 1990) assume, reasonably, that children are highly
constrained in their interpretations of events. She presents
children with a scene that can (from an adults’ perspective)
be interpreted in two ways and demonstrates that syntac-
tic cues reduce the potential interpretations of the verb
from two to one. Presumably, then, her proposal is that the
pre-linguistic conceptual biases are weaker for events than
for entities. When the child sees Fido, the whole object
“pops out” at them; but when the child sees grandma get-
ting a doll, there are two or three or even more interpreta-
tions that come to mind, not just one. In support of this,
Naigles points out (a) that the same event can be described
in many different ways: a child bringing a doll to grandma,
can be carrying, bringing, coming, arriving, and so on, and
(b) that children’s first verbs are semantically diverse, in-
cluding general meanings (come), specific ones (draw), and
so on.
The problem with this argument is that both observations
apply just as well to objects and nouns. The family pet can
be Fido, a dog, an animal, a pet, and so on. And children’s
first nouns are diverse: general (toy), specific (cup), and so
on. So what’s the difference? In any case, if there were an
event/object difference of the sort Naigles proposes, it
would suggest, at best, that syntax plays more of a role in
verb learning than in noun learning, which is fully compat-
ible with the theory in HCLMW. It does not suggest that
verb learning differs in kind from noun learning.
Finally, I had suggested that 1-year-olds learn verbs be-
fore they have sufficient grasp of syntactic structure. Nai-
gles disagrees – but the studies she cites are unpersuasive,
since none of them actually explore syntactic knowledge in
1-year-olds. Note also that my general point was that chil-
dren can learn at least some verbs without syntactic cues.
Does Naigles doubt this? If a child is stuck with multiple in-
terpretations, syntax can help pick out the right one – but
so can other sources of information, such as pragmatic cues
and repeated exposures across different situations (see
Pinker 1994a).
On a more agreeable note, Naigles notes the rich paral-
lels between the theory outlined in HCLMW and Harry
Potter’s Hogwarts School of WitchCraft and Wizardry. I
hope the word gets out; it might help sales of the paperback.
Bickerton discusses the acquisition of functional items,
such as pronouns and determiners, and argues, based on
cross-linguistic comparisons and the sorts of functional
items created by Creole speakers, that there is a limited
class of meanings that such words can encode. For instance,
two determiners might differ on the semantic notion of def-
initeness (avs. the) or individual/stuff (much vs. many), but
you are not going to find determiners that express a contrast
between rare versus common, or pleasant versus unpleas-
ant. This is an important point, one that should have re-
ceived more emphasis in HCLMW.
Bickerton makes two other claims. The first is that func-
tional items should be particularly hard to learn. For in-
stance, he suggests that pronouns pose a harder acquisition
problem than proper names. But this seems backwards. If
Bickerton is right, and children only have to consider a
small set of candidate meanings (fewer than 50, say) when
they hear a new pronoun, this should make learning pro-
nouns much easier than learning proper names. (By any ac-
count, proper names can have a lot of meanings – specific
people, cities, books, etc.) In general, innate constraints
make learning easier, not harder.
Bickerton also worries that I won’t like it if there are in-
nate constraints that specifically apply to the meanings of
functional items, presumably because it would conflict with
my theory. But I like it just fine. My beef is with a dedicated
module for word learning, not with representational con-
straints on word meanings. It seems perfectly plausible that
children start with expectations as to the meanings encoded
by nouns, or by verbs, or by determiners (see also Keil). On
the other hand, Bickerton is too quick to assume that such
expectations would have to be specifically linguistic. While
it is conceivable that universal grammar provides a list of se-
mantic features that correspond, say, to any functional mod-
ifier of a noun, it might also be that these semantic features
emerge from a more general intuitive metaphysics. For in-
stance, one versus many, individual versus stuff, and so on,
Response/Bloom: How Children Learn the Meanings of Words
BEHAVIORAL AND BRAIN SCIENCES (2001) 24:6 1127
may be natural ways to conceptualize the sorts of things that
Noun Phrases (NPs) refer to, and so any linguistic elements
that semantically modify nouns would be constrained in this
regard.
Bickerton ends with a methodological suggestion: one
can get a lot of insight about word learning by looking at un-
usual cases in which children are exposed to particularly im-
poverished input. Agreed. But Bickerton’s own examples
are puzzling – he thinks it is a serious omission that
HCLMW does not discuss first-generation Creole speakers
and deaf children who create their own sign languages. But
while such research is fascinating, and tells us much about
the sorts of conceptual and linguistic capacities that exist in
the absence of exposure to a pre-existing language, it does
not bear directly on the question of how children learn the
meanings of words. This is because, as Bickerton himself
repeatedly argues, such children don’t learn words; they
create them.
R5. What is the role of the environment?
A few of the commentators, with considerable enthusiasm,
have made the following claims:
People often speak to children
Children need to hear a word in order to learn it
Children learn words in a social context
It would make for a more exciting response if I were to
disagree with any of this. But plainly, for children to learn
words, they do have to hear them in situations where there
is good evidence as to their meaning. Indeed, in many cul-
tures, people try to teach words to children, and children
benefit from this teaching. Adults do, too. If you wanted to
learn French, you would be better off with friendly people
who are naming objects for you than having to pick up the
language by listening to strangers in a bar.
But while these facts constrain any account of word
learning, they do not themselves constitute a theory – we
need to explain how children use the available information
to learn words.
This is apparently controversial. Nelson sees the ap-
proach in HCLMW as too individualistic, too focused on
the “heads of children.” But while I agree with much of
what she says – words are used in a social context, words
have rich meanings – it was never clear what her alter-
native is. Gogate sees the proper goal of researchers as
describing how different informational cues interact at dif-
ferent points in time and characterizing their relative im-
portance over the course of development. This is a useful
step, but in the end we still need to know what it is, in the
child’s mind, that explains how the cues are understood and
utilized. (Once we know this, by the way, we will be a bet-
ter position to answer the descriptive questions that inter-
est Gogate.)
One specific question that was raised in HCLMW is how
powerful children’s learning mechanisms are. Are children
good enough at learning that they can succeed in an un-
sympathetic environment, one where nobody is making
any effort to teach them words? My tentative answer,
based on the very limited cross-cultural evidence, is that
they can. For this reason, perhaps, Gogate suggests that I
underestimate the role of maternal communication. To
clarify the issue, she describes some research of her own
showing that mothers are sensitive to their children’s level
of linguistic knowledge when they name objects for them.
This is a nice finding, but it is not, however, about word
learning in children. It is a finding about theory of mind in
mothers.
I don’t doubt that children benefit from this sort of fine-
tuning by their parents – in Chapter 3, I review evidence
showing that they do. But this does not show that the
process is necessary, or even important. As an analogy, some
parents will teach their children to play checkers. It would
be surprising if parents did not modify their teaching be-
havior based on the child’s age and ability, and it would also
be surprising if children did not benefit from these efforts.
Would you infer from this that the only way a child could
learn checkers is by being taught the game by attentive and
sensitive parents?
Oddly, word learning through social interaction is some-
times seen as an alternative to attributing rich capacities to
children (e.g., Nelson). But, as Garton and Strauss & Ziv
discuss, participation in social interaction itself requires
powerful theory of mind capacities. The discussion of the-
ory of mind in HCLMW focused mostly on children’s abil-
ity to ascertain the intentions of the speaker, but both of
these commentators point out that children do much more
than this. For instance, they try to elicit names for things.
Even 2-year-olds will point and ask, “What’s that?”, an ap-
parently simple act that exhibits children’s rich tacit knowl-
edge of the mental states of themselves and others.
R6. What is the relationship between words and
concepts?
Just as with claims about the importance of the input, the
position that language affects thought is often put forth with
some vigor, as if it were a bold conjecture. It isn’t. Nobody
doubts the importance of language for an individual or a so-
ciety, or the power of language as a mechanism for com-
municating ideas, for persuasion, threat, seduction, reli-
gious conversion, and so on. How could you convey a notion
such as “the speed of light is 186,000 miles per second,”
without language? In fact, Chapter 9 defends the stronger
claim that you cannot even understand such a notion with-
out language – in particular, you need the generative sys-
tem that underlies the production of number words (see
also Varley & Siegal).
So what’s the fuss about language and thought? There are
actually three controversial and interesting questions:
1. How much of thought exists prior to – and indepen-
dently of – language?
While there is evidence that some forms of thought, such
as precise numerical reasoning, do require a natural lan-
guage, one can have rich and structured thought without
words. The evidence from this comes from studies of ba-
bies, of nonhuman primates, and from adults with aphasia
(see again Varley & Siegal). The only commentator to ob-
ject to this conclusion is Nelson, who calls it a “distortion,”
but never says why.
A second question is:
2. Is thought shaped only by the content of what is com-
municated, or does the form (the grammar) also matter?
This is the classic debate over the Sapir-Whorf hypothe-
sis. Whorf famously claimed, for instance, that English
Response/Bloom: How Children Learn the Meanings of Words
1128 BEHAVIORAL AND BRAIN SCIENCES (2001) 24:6
speakers think differently about time and space than Hopi
speakers because of their linguistic experience. Impor-
tantly, Whorf focused on the form of language – his claim
would be a lot less interesting if he were simply stating that
Hopi adults tell their children one thing about time and
space and Americans tell them another. Instead, he was say-
ing that the structural properties of the different languages
have this effect.
There has been some recent interest in this proposal, and
some new empirical work. In the end, as argued in Chap-
ter 10, I think that the evidence for it in general is still quite
weak (see also Bloom & Keil 2001). But consider now a po-
tentially related question:
3. To what extent are concepts shaped by patterns of
naming?
To see the issue here, imagine a child who already has a
mature understanding of what a dog is, and hears Fido be-
ing called “a dog.” Assuming the child makes the proper
mapping, he or she could go on to use the word exactly
right; calling all and only dogs “dogs.” To the extent that this
is an accurate conception of what happens in word learn-
ing, naming plays no role in the shaping of concepts.
In contrast, consider Scott’s example of “debate.” A child
might hear the word for the first time but only have a vague
idea as to what a debate is. Over the course of time, by hear-
ing the word used to refer to diverse events and reasoning
about what these events have in common, he or she learns
more about what debates are. As Block (1986) points out,
this sort of process is particularly relevant for the acquisi-
tion of more abstract and theory-laden notions – your un-
derstanding of genes, mortgages, and phonemes is likely to
emerge as the result of hearing how people use the words
“gene,” “mortgage,” and “phoneme.” This is what Marat-
sos describes as “category assembly,” where a concept is
shaped through exposure to language, as opposed to “cate-
gory recognition,” where a pre-existing category is triggered
by exposure to a word.
Both category-assembly and category-recognition apply
for the learning of any given word, though they differ in
their relative importance. Maratsos gives the example of
the weird Tzeltal word that means “to ingest crunchy
foods.” Maratsos is correct that there is a lot of category as-
sembly here; the word is unlikely to immediately map onto
the right concept. But there is also category recognition –
upon hearing the word for the first time, it is likely that the
child has some notion of the appropriate concept, perhaps
that it has something to do with eating. Murphy gives the
example of the weird English word “bird,” and notes that
while children might start off with a pretty good rough idea
as to the boundaries of the category (category-recognition),
their concepts of bird will later be expanded, for instance,
by hearing penguins described as “birds” – a clear case of
category assembly.
Scott wonders if the existence of category assembly
shows that Whorf was right. It does not. Recall that the in-
terest of the Whorfian view is that language has its effects
through the form of a language, and so it would really be
striking evidence for this position if the child’s concept of
bird was affected by the fact that the English word “bird”
has one syllable, or if her crunchy-eating notion was af-
fected by the fact that Tzeltal has classifiers. But this does
not happen. When a child learn that penguins are birds by
hearing them called “birds,” this is contentful learning of
the most straightforward type, such as learning the nation-
ality of an actor by hearing him described as “Canadian” or
the status of a food by hearing it called “a fruit.”
After all, if the effect of names were based on their form,
not their content, you would expect children to have prob-
lems with different words that sound the same. The same
sound is used to describe the flying creature and the object
for hitting baseballs, and a child who assumed that same
sound 5same category, would form a very general category
of bat that contains both objects. But children do not do
this. In fact, the literature (e.g., Davidson & Gelman 1990)
reviewed in Chapter 10 shows that children only use names
to enlarge and restrict category membership when they
have independent semantic support to do so. It is not
enough to hear a penguin called “bird,” the child needs
good reason to believe that it actually is a bird – a member
of the same category as robins and sparrows – before the
category is enlarged.
Maratsos and Murphy discuss the relative contributions
of category-recognition and category-assembly, and both
note that HCLMW emphasizes the process of category-
recognition. This is because of the observational and ex-
perimental evidence suggesting that when children learn
their first words – cup, drink, Fido, and so on – they get the
meanings pretty much correct from the very start. Serious
errors never occur, suggesting that children do not need the
shaping power of multiple exposures. Category-assembly
might be needed for the fine-tuning of word meaning
(learning that penguins are birds, and so on), but is not the
dominant process of word learning.
Maratsos has a clever reply to this: He suggests that chil-
dren really do need multiple exposures in early word learn-
ing. They make so few mistakes because they are cautious,
and tend not to use a word until they have sufficient infor-
mational support that they have converged on the right hy-
pothesis. I think the experimental data suggest that this is
not the case for object name learning – but it is a more open
question whether this sort of caution exists for verbs and
names for abstract entities; this is an issue that warrants fur-
ther research.
Murphy wonders why children who are so good at learn-
ing words taught to them by developmental psychologists
turn into dull college undergraduates who are so bad at
learning words taught to them by cognitive psychologists.
He has several explanations for this, all of which make
sense, and concludes by stressing the need to use more re-
alistic conceptual structures in our experiments, which also
makes sense. My problem is with his suggestion that the ap-
propriate word learning experiments would have to explore
“the long task of filling in the exceptions and unpredictable
cases” – in other words, that there should be more focus on
situations that require category assembly. Murphy notes
that there is a sense in which a child does not actually learn
the word “bird” until it is used with near-perfect accuracy,
which may have to wait until adulthood. But this is not the
usual conception of word learning. The perspective
adopted in HCLMW is that children learn “bird” when they
have a rough idea as to what the word refers to. In the end,
this might be just a definitional issue, one that reflects dif-
ferent research interests, but from my perspective, the
process of working out the hard cases and atypical instances
is an important part of conceptual development – but it is
not word learning.
Response/Bloom: How Children Learn the Meanings of Words
BEHAVIORAL AND BRAIN SCIENCES (2001) 24:6 1129
R7. The big picture: Word learning, nativism,
modularity, and natural selection
Keil observes that, while the learning of word meanings may
be the result of general cognitive processes, the development
of syntax, phonology, and morphology may not be. In fact,
there are reasons to expect them to be different. These other
aspects of language are likely to be modular; the conditions
under which you can extract a NP from a complement are
specified in a way that does not make reference to the rest of
mental life, and the learning of the language-specific condi-
tions for NP extraction might be similarly encapsulated. In
contrast, word meanings are nonmodular; the meaning of the
word “mortgage” is intimately tied to your understanding of
other things as well, and the same holds true for how we
come to learn this English word.
Keil is also correct that the position in HCLMW does not
preclude the existence of powerful constraints on the se-
mantics of language. The learning of word meanings might
be the product of general processes, but the representation