Linguistic Relativity and Word Acquisition: A Computational Approach

Linguistic Relativity and Word Acquisition: A Computational Approach
Eliana Colunga (ECOLUNGA@CS.INDIANA.EDU)
Michael Gasser (GASSER@CS.INDIANA.EDU)
Computer Science Department
Indiana University
Bloomington, IN 47405

Abstract

Language plays a pervasive role in our day-to-day experience
and is likely to have an effect on other non-linguistic aspects
of life. At the same time, language is itself constrained by the
world. In this paper we study this interaction using Playpen, a
connectionist model of the acquisition of word meaning. We
argue that the interaction between linguistic and non-linguistic
categories depends on the pattern of correlations in the world
and on their relation to the correlations defined by words. We
then discuss three kinds of possible interactions and present
simulations of each using Playpen, a neural-network model of
the acquisition of word meaning.
Introduction
Language plays a pervasive role in our day-to-day experience.
Thus it is no surprise that people wonder to what extent
language in general and the particular language one speaks
affect the rest of our cognitive abilities. Does language affect
thought; that is, do linguistic categories influence general
cognitive categories? More generally, how do linguistic and
non-linguistic categories interact?
The Linguistic Relativity Hypothesis, associated most
closely with Benjamin Lee Whorf (Whorf, 1956), concerns
the first question. The claim, in its strongest form, is that linguistic
categories exert a direct influence on general cognitive
categories. SinceWhorf, many researchers have attempted to
find evidence for this influence (see Lucy (1996) for a review),
but there is as yet no agreement that the evidence has
been found.
People have usually studied the effect of language on cognition
by looking for differences in adult speakers of different
languages. Negative results (Rosch, 1973; Kay & McDaniel,
1978) are met with the arguments that the experiments are
biased towards Indo-European languages, are dealing with a
part of perception not subject to linguistic effects, or involve
irrelevant language distinctions which should not be expected
to have an effect in the first place. Positive results (Carroll &
Casagrande, 1958; Kay & Kempton, 1984; Bloom, 1981) are
generally explained away as effects of culture, biased stimuli
or the linguistic nature of the task being used.
More recently, linguistic relativity has been studied in the
context of learning, and the picture there looks more promising
for linguistic relativity. It has been shown that the order in
which children learn certain words, as well as their patterns
of overgeneralization, depend on the language being learned
(Brown, 1994; Bowerman, 1996), evidence for an effect of
language on the rest of cognition. This is further supported
by converging evidence from studies showing how learning
labels in the laboratory can affect children’s performance in
tasks like word generalization and analogical problem solving
(Jones&Smith, 1993;Gentner, Rattermann,Markman,&
Kotovsky, 1995). Other developmental studies also show parallels
between linguistic and non-linguistic performance in
various domains (Jones, Smith, Landau, & Gershkoff-Stowe,
1992; Smith & Sera, 1992).
We believe that learning is the right place to look for relativistic
effects, but we also believe that the empirical work
on development must be supplemented with a computational
account, one which looks at how the demands of linguistic
and non-linguistic tasks may lead to long-term effects. In the
paper we present the beginnings of such an account. In the
next section we discuss the role of correlations in the learning
of linguistic and non-linguistic categories. Next we present
a computational model and discuss the results of three simulations
demonstrating possible kinds of relativistic effects.
Finally we consider the implications of the model for future
research on linguistic relativity.
Linguistic and Non-Linguistic Correlations
We propose that the way linguistic and non-linguistic categories
interact depends on the nature of the correlations in
the world and the way these relate to the correlations in the
language. During their first year, babies experience the world
without any of the biases that are built into language. It
is by now clear that they learn a great deal about how the
world works during this time (Baillargeon, 1994; Spelke,
Breinlinger, Macomber, & Jacobson, 1992). One aspect of
this learning is the discovery of correlations between features
along different dimensions (Younger, 1990). These nonlinguistic
correlations define what we will call non-linguistic
categories. In its second year, the child begins to learn language,
which introduces its own categories, defined in terms
of the correlations between words and non-linguistic dimensions.
The linguistic categories may agree or disagree with
the non-linguistic categories. Figure 1 shows some of the
ways in which the two sorts of categories can be related to
one another.
Given these sorts of correlations, linguistic and nonlinguistic
cognition could interact at several levels. First,
some words could be rendered easier to learn than others.
In fact, some categories of words are consistently learned
before others, an effect which cannot be explained by frequency.
Across languages nouns are learned before verbs
(Nelson, Hampson, & Shaw, 1993) and instrument verbs are
learned before other verbs (Huttenlocher, Smiley, & Charwords
􀀀􀀀􀀀
􀀀􀀀􀀀


􀀀􀀀
􀀀􀀀


􀀀􀀀􀀀
􀀀􀀀􀀀


􀀀􀀀
􀀀􀀀


􀀀􀀀
􀀀􀀀


Figure 1: Possible correlations between linguistic and nonlinguistic
categories. Non-linguistic features, represented by circles,
may be associated with words, represented by squares, in such a
way that the words agree with non-linguistic categories (squiggly
pattern), disagree with non-linguistic categories (solid pattern), subdivide
non-linguistic categories (cross-hatched pattern), or combine
non-linguistic categories patterns (diagonal hatched pattern).
ney, 1983; Behrend, 1990). These orderings have been explained
in terms of the traditional view that categories are
formed around strong correlational structure (Rosch, Mervis,
Gray, Johnson, & Boyes-Braem, 1976; Kersten & Billman,
1997). This strength-of-correlations account should also hold
for the correlations between words and other perceptual inputs.
Words that agree with previously learned categories
should be easier to learn than words that disagree with them.
Another place where an effect of language on cognition
could be found is in the highlighting (or downplaying) of
dimensions that are relevant (or irrelevant) to the language
being learned. This effect on attention could be shown in
both linguistic and non-linguistic tasks. Hunt and Agnoli
(1991) suggest that learning a language that makes a distinction
could make speakers of that language more sensitive to
that distinction in non-linguistic tasks, a direct effect of language
on perception.
An example of a linguistic task where effects of language
on attention can be observed is the development of the shape
bias. Children at around 18 months of age tend to generalize
words to novel objects of the same shape as the exemplar,
rather than to novel objects that share size, color or material
with the exemplar. Because concrete nouns in most languages
are organized mainly in terms of shape, this attentional bias
helps them generalize correctly. The shape bias appears only
after the child has learned roughly 50 nouns, most of them
naming categories of things that are similar in shape (Jones
et al., 1992). This suggests that it is the learning of words
that drives the learning of the bias. The way in which the
linguistic categories correlate with perceptual dimensions apparently
causes the learner to attend to particular dimensions,
at least in the context of linguistic tasks.
A more dramatic effect of language would be on the nature
of the non-linguistic categories themselves. Non-linguistic
categories are built up out of correlations between perceptual
dimensions. Linguistic categories may agree with these
non-linguistic correlations if words correlate with correlating
perceptual dimensions. Alternatively, the non-linguistic correlations
may be irrelevant for the linguistic categories. The
child has both linguistic and non-linguistic tasks to perform.
In performing the non-linguistic tasks, the child can rely on
non-linguistic correlations, but if the language agrees with
these correlations, she can rely on linguistic correlations as
well. On the other hand, if the non-linguistic correlations
have nothing to do with the language, non-linguistic tasks can
only be performed using these correlations. This implies that
the strength of the non-linguistic correlations might vary with
the languages. While we know of no direct evidence for this
possibility, the strong version of the Linguistic Relativity Hypothesis
predicts this sort of effect may be found.
In what follows, we provide illustrations in the model of
all three sorts of interactions, the influence of the match between
linguistic and non-linguistic categories on the relative
ease of words, the influence of linguistic categories on attention
to perceptual dimensions, and the influence of linguistic
categories on the way in which non-linguistic categories are
represented.
The Model
Playpen (Gasser & Colunga, 1997) is a connectionist model
of the acquisition of word meaning. For the purposes of this
paper, the following features of the model are relevant:
1. The network is a generalization of a continuous Hopfield
network. Units are updated randomly until the network
settles.
2. Network units have relative phase angles in addition to activation,
and feature binding is handled through the synchronization
of unit phase angles. Units affect each other’s
phase angles via the weights on the connections joining
them.
3. Units are of two types. Micro-object units (MOUs) represent
object features. Micro-relation units (MRUs) represent
relations between the features of separate objects.
Relation words take the form of MRUs. Each MOU has a
single phase angle; each MRU has two phase angles, one
for each of the objects it relates.
4. Connection weights are adjusted via the contrastive Hebbian
learning rule (Movellan, 1990).
5. Non-linguistic features and relation words interact through
one or more intermediate layers of MRUs.
Three characteristics of Playpen make it especially well
suited to the study of the interaction between language and
perception. First, linguistic meaning and non-linguistic concepts
are not rigidly distinguished. This is important because,
if we are to enter the linguistic relativism debate without any
biases, we should not assume from the start that linguistic
and non-linguistic concepts are independent. The model allows
correlations to develop in the layers of MRUs separating
words and non-linguistic perception as learning takes place,
and these correlations can have more or less of a linguistic
character. Second, the model is designed to deal with relational
knowledge. Languages vary more in the way in which
they express relational information than in the way they express
information about objects, so it is more likely that effects
of language will be found in relational words (Gentner
& Boroditsky, 1998)While we do not model these properties
of relation words, they argue for focusing on relation words
as a possible site of relativistic effects, and modeling these
effects would require a system capable of handling relations.
In particular, a model must be able to learn relational correlations
(Gasser & Colunga, 1998). Third, the model can
be “run” in both the “comprehension” and the “production”
directions, allowing for the possibility of mutual effects of
language and perception on one another.
Experiments
The three simulations we describe here were based on a set
of pre-defined correlations among non-linguistic dimensions
and correlations between the non-linguistic dimensions and
words. There were two non-linguistic dimensions, and relations
within each dimension correlated with relations within
the other. That is, a pair of objects with particular values on
one dimension tended to have particular values on the other.
The relational correlations are shown in Figure 2a.
CD
DD
DC
AA
AB
BA
CC
BB
A
Dim 1 Dim 2
a
A
b
Hard Language Word 2
Easy Language Word 1
D
C C
B B
D
Figure 2: Correlations used in experiments. A, B, C and D represent
possible micro-relations between features on Dimensions 1 and
2. (a) The micro-relations are associated with each other across the
dimensions in the two clusters shown. (b) Possible pairings of relations
on the two dimensions are associated with one or the other of
two words. In the Easy language, the words agree with the nonlinguistic
correlations; in the Hard language, the words correlate
only with micro-relations on Dimension 1.
We defined two “languages,” an Easy language, which
agrees with the non-linguistic correlations, and a Hard language,
which disagrees with the non-linguistic correlations,
as shown in Figure 2b. Each language consists of two relational
words. For the Easy language, the categories in the
world agree with the categories promoted by the words. That
is, each of the two correlational clusters existing in the world
is associated with one of the words in the language. For the
Hard language, thewords cut across the two prelinguistic correlational
clusters in such a way that the word describing a
pair of values along the two dimensions is determined by the
value along Dimension 1 only. For example, according to the
pattern of correlations between dimensions in Figure 2, the
pairs of values labeled A-B and B-A should be in the same
category but they are assigned to different words in the Hard
language. That is, the value along Dimension 2 is not predictive
of the linguistic category.
The architecture of the networks used in these simulations
is shown in Figure 3.
Dimension 1 Dimension 2
micro-relation
micro-object
unit
RUs
DIMENSIONS
PERCEPTUAL
HIDDEN
WORDS
unit
Figure 3: Network architecture. Micro-object units are represented
by squares, micro-relation units by diamonds. Arrows indicate complete
connectivity between layers. Each Hidden MRU is associated
with a pair of Perceptual Dimension MOUs. A possible pattern
across the network is shown. Darkness indicates activation, and arrow
direction indicates relative phase angle.
The networks are trained and tested on two different tasks.
For Non-linguistic Pattern Completion, they are presented
with a pattern on one of the Perceptual Dimensions and expected
to produce an appropriate pattern on the other. (Note
that there are always two possibilities for the appropriate pattern.)
The network can learn to solve this task using the connections
joining the Perceptual Dimension and Hidden Relation
layers or the connections between the two Hidden Relation
layers. For Production, the networks are presented with
a pattern on the Perceptual Dimensions and expected to output
a word.
Experiment 1 - Difficulty of languages
The goal of this experiment is to see how the different correlational
patterns both between dimensions and with the words
affect the difficulty of learning the two languages. The networks
were first trained in a Pre-linguistic Phase on Nonlinguistic
Pattern Completion alone for 30 repetitions of the
relevant training patterns (epochs). Next, during a Linguistic
Phase, Pattern Completion training was discontinued, and the
networks were trained on Production for seven epochs. We
predict that the Easy language will be learned faster than the
Hard language during the Production phase because the Easy
language categories agreed with the non-linguistic categories.
During the Pre-linguistic Phase, the networks mastered the
Pattern Completion task by learning weights between the two
Hidden layers representing the non-linguistic correlations.
Results for the Linguistic Phase are shown in Figure 4.
The datawere submitted to a 2(Language) 7(Epoch) analFigure
4: Results for Experiment 1. The Hard language is harder
to learn than the Easy language.
ysis of variance for a mixed design. This analysis revealed a
main effect of epoch, indicating that the networks get better
as they receive more training. More importantly, as predicted,
there is a main effect of language (p < :001). Thus, the Easy
language is learned faster than the Hard language, although
by the end of the training the two networks have comparable
performance. No interactions between language and epoch
were found.
The results make sense for two reasons. Given the correlations
between the Perceptual Dimensions and the two language
situations, a network learning the Easy language could
choose to attend to Dimension 1, to Dimension 2, to both
dimensions or even to different dimensions for different values
along the dimensions. In contrast, to learn the Hard language
the network needs to attend to Dimension 1 and ignore
Dimension 2. Since the space of possible good solutions is
larger for the Easy than for the Hard language, the words in
the Hard language to be harder to learn than the words in the
Easy language.
A second reason for the ease of the Easy language concerns
the effect of the non-linguistic correlations on language
learning. In the case of the Easy language, learning the right
association between one perceptual input and its corresponding
word should improve the chances of producing the right
word for the other instances of that word. This is because
of the previously existing correlations. At the beginning of
Production training, any associations between a Hidden unit
and the Word layer indirectly affect the other Hidden units
involved in non-linguistic clusters with that Hidden unit. In
the case of the Easy language, the correlations help since linguistic
and non-linguistic categories agree; in the case of the
Hard language, the correlations fail to help solve the Production
task.
This experiment demonstrated howwords can differ in ease
of learning to the extent that they agree with non-linguistic
categories. That is, given a particular set of perceptual dimensions,
for example, the set of dimensions that is learned
relatively early because of its salience or importance to the
child, words will differ in the degree to which those dimensions
define them. And this difference will lead to differences
in ease of learning. The comparison holds within languages
as well as across languages. If this is true, this would explain
the facilitated learning of instrument verbs over other
verbs. The instrument together with the action form a tight
correlational cluster that is likely to be there prelinguistically.
In contrast, for a more abstract verb, for example, enter, a
child would have to concentrate on the one thing that matters
(path) and ignore the other aspects of the situation to which
the word applies. This is also consistent with findings that
in Tzeltal, a Mayan language with an apparently complicated
system for expressing spatial relations, context-specific spatial
relation words are learned earlier than the more abstract
spatial prepositions (Brown, 1994).
Experiment 2 - Highlighting dimensions
The goal of the second simulation is to verify that the networks
trained on the Hard language do in fact pay more attention
to the relevant than to the irrelevant dimension. To test
this we presented the trained networks with novel perceptual
input patterns. We predict that the networks trained on the
Hard language will produce the word which is consistent with
the relevant dimension (Dimension 1), while those trained on
the Easy language should show no such preference. For example,
in Figure 2, if the network is given the values for A in
Dim1 and C in Dim1, the networks trained on the Hard language
should tend to outputWord 1, because only the pattern
on Dim1 (A) counts. In the same situation networks trained
on the Easy language could output either Word 1 (consistent
with AA and AB) or Word 2 (consistent with CC or DC).
Networkswere first trained in the Pre-linguistic Phase, then
in the Linguistic Phase for 7 epochs of training on the Production
patterns. We were only concerned with the performance
following this training. To compare the performance of the
networks, we subtracted the number of words agreeing with
Dimension 2 from the number of words agreeing with Dimension
1. Thus a positive result indicates a preference for
Dimension 1, a negative result a preference for Dimension 2.
The results are shown in Figure 5. A T-test revealed that,
as expected, the networks trained on the Easy language had
a different preference pattern from those trained on the Hard
language. In fact, the networks trained on the Easy language
showed no preference for either word while the networks
trained on the Hard language showed a significant preference
for the words consistent with Dimension 1.
Experiment 3 - Effect of language on non-linguistic
categories
The goal of Experiment 3 is to determine whether the difference
in the two languages can have an effect on the way in
which the network learns the correlations between the Perceptual
Dimensions. During pre-linguistic training in Experiments
1 and 2, the networks readily learned the weights
on the connections joining the two Hidden layers representing
these correlations. Since each hidden unit is associated
with a pair of values along one of the Perceptual Dimensions,
these weights are easily interpreted. Pre-linguistic training
results in positive weights on each of the connections joining
Hidden-layerMRUs representing pairs of perceptual features
which correlate and negative weights on the other connections.
As can be seen from Figure 6a, there are eight correlating
pairs; hence eight of the weights joining the Hidden
Figure 5: Results for Experiment 2. The networks trained on the
Hard language responded with the word consistent with Dimension
1 97% of the time, while networks trained on the Easy language
showed no preference.
layers are positive, while the other eight weights are negative.
For example, the weight on the connection joining the hidden
units representing A on Dimension 1 and B on Dimension 2
is positive, while the weight for A on Dimension 1 and C on
Dimension 2 is negative.
In Experiment 3, rather than training the networks on the
Non-Linguistic Pattern Completion task before training on
the Production task, we trained them on the two tasks simultaneously
by alternating between the two tasks.
For the Hard network, the two tasks must be solved using
completely different weights. To learn to produce the correct
word, the network must rely on the connections from
the Dimension 1 Hidden layer to the Words layer. To learn
to perform the Pattern Completion task, it needs to learn the
inter-Hidden-layer correlation weights.
For the Easy network, on the other hand, because the linguistic
and non-linguistic correlations agree, the two tasks
can make use of the same weights. In particular, there are two
ways in which the network could learn to solve the Pattern
Completion task. It could make use of the inter-Hidden-layer
correlation weights, as we expect in the Hard network. Alternately,
it could rely on the Hidden-to-Words connections, using
the word as a bridge between the two dimensions. These
two paths are shown in Figure 6. Because the Easy network
can perform the Pattern Completion task without the inter-
Hidden-layer weights if it has the Hidden-to-Words weights,
and because it needs the Hidden-to-Words weights anyway to
solve the Production task, we predict the inter-Hidden-layer
correlation weights will be smaller in the Easy than in the
Hard network.
We trained 10 networks each on the Easy and Hard set
of patterns, alternating Production and Pattern Completion
tasks. For this experiment, we are interested only in the inter-
Hidden-layer weights that resulted during training, not in the
performance of the networks on the tasks. After four epochs
of training, we compared the correlation weights for the A
and B input patterns, that is, the A-A, A-B, B-A, and B-B
inter-Hidden-layer weights, for the Easy and Hard networks.
As we expected, the weights in the Hard network were significantly
larger (p < :02) than the weights in the Easy network.
This shows that the nature of the linguistic categories can directly
influence the weights representing the non-linguistic
correlations.
In this experiment we showed how the kind of language
being learned can affect the way the same information is
learned. More importantly, the same task was solved with
or without linguistic knowledge depending on the correlation
patterns between the words and the world. This points
out a flaw in one of the most frequent complaints about relativism
experiments, namely that whenever a cross-linguistic
difference is found, the task is declared to be linguistic in nature.
In our illustration, both networks solve the same task
using different parts of the architecture. There was no behavioral
difference between the two networks in either of the
two tasks they were trained on and yet their representation of
the knowledge necessary to solve the tasks was different. We
think this is a direct effect of the structure of the languages
being learned by the networks on cognition. This suggests
that it is not the task that makes the process linguistic or nonlinguistic,
and to a certain extent, that it could be the structure
of the language that does. The fact that we found no behavioral
differences reflecting the weight differences in the networks
should not be discouraging. Brain scan studies could
be performed on people to look for effects analogous to the
weight differences in the networks. Also, preliminary results
show language effects during the course of learning suggesting
that that is a good place to start looking for evidence for
relativism.
a b
Figure 6: Two paths for performing Non-Linguistic Pattern Completion.
(a) The network uses the between-Hidden-layer connections
representing the correlations between the Perceptual Dimensions.
This is possible with both the Easy and Hard networks. (b) The network
uses the Hidden-to-Words connections. This is possible only
with the Easy Network.
Conclusions
In this paper we have argued that linguistic relativity can be
best studied in terms of the correlations between different perceptual
dimensions, the correlations between linguistic categories
and perceptual dimensions, and the way in which these
correlations interact during the learning of language and of
non-linguistic tasks. We focused on three specific relativistic
effects and showed how each of these could be simulated
with a simple neural-network model of word learning. We
believe that such a model is crucial to the relativity debate.
Without an explicit account of how the learning of linguistic
and non-linguistic categories depends on different kinds of
correlations, it will remain unclear precisely what linguistic
relativity might mean for cognition.
References
Baillargeon, R. (1994). How do infants learn about the physical
world?. Current Directions in Psychological Science,
3, 133–140.
Behrend, D. (1990). The development of verb concepts: children’s
use of verbs to label familiar and novel events.
Child Development, 61, 681–696.
Bloom, A. (1981). The linguistic shaping of thought: a study
in the impact of language on thinking in China and the
West. Lawrence Erlbaum, Hillsdale, NJ.
Bowerman, M. (1996). Learning how to structure space for
language: a crosslinguistic perspective. In Bloom, P.,
Peterson, M. A., Nadel, L., & Garrett, M. F. (Eds.),
Language and Space, pp. 385–436. MIT Press, Cambridge,
MA.
Brown, P. (1994). The ins and ons of Tzeltal locative expressions:
the semantics of static descriptions of location.
Linguistics, 32, 743–790.
Carroll, J. & Casagrande, J. (1958). The function of language
classifications in behavior. In Maccoby, E., Newcomb,
T., & Hartley, E. (Eds.), Readings in social psychology,
pp. 18–31. Henry Holt, New York.
Gasser, M. & Colunga, E. (1997). Playpen: toward an architecture
for modeling the development of spatial cognition.
Tech. rep. 195, Indiana University, Cognitive
Science Program, Bloomington, IN.
Gasser, M. & Colunga, E. (1998). Where do relations come
from?. Tech. rep. 221, Indiana University, Cognitive
Science Program, Bloomington, IN.
Gentner, D. & Boroditsky, L. (1998). Language acquisition
and conceptual development, chap. Individuation, Relativity
and Early Word Learning. Cambridge University
Press, England.
Gentner, D., Rattermann, M. J., Markman, A., & Kotovsky,
L. (1995). Two forces in the development of relational
similarity. In Simon, T. J. & Halford, G. S. (Eds.),
Developing Cognitive Competence: New approaches
to process modeling, pp. 263–313. Lawrence Erlbaum
Associate, Hillsdale, NJ.
Hunt, E. & Agnoli, F. (1991). The Whorfian hypothesis: a
cognitive psychology perspective. Psychological Review,
98, 377–389.
Huttenlocher, J., Smiley, P., & Charney, R. (1983). Emergence
of action categories in the child: evidence from
verb meanings. Psychological Review, 90, 72–93.
Jones, S. S. & Smith, L. B. (1993). The place of perceptions
in children’s concepts. Cognitive Development, 8, 113–
140.
Jones, S., Smith, L., Landau, B., & Gershkoff-Stowe, L.
(1992). The origins of the shape bias. Boston. Boston
Child Language Conference.
Kay, P. & Kempton, W. (1984). What is the Sapir-Whorf
hypothesis?. American Anthropologist, 86, 65–79.
Kay, P. & McDaniel, C. (1978). The linguistic significance
of the meanings of basic color terms. Language, 54,
610–646.
Kersten, A. W. & Billman, D. (1997). Event category learning.
Journal of Experimental Psychology: Learning,
Memory and Cognition, 23(3), 658–681.
Lucy, J. A. (1996). The scope of linguistic relativity: an analysis
and review of empirical research. In Gumperz, J. J.
& Levinson, S. C. (Eds.), Rethinking Linguistic Relativity,
pp. 37–69. Cambridge University Press, Cambridge.
Movellan, J. (1990). Contrastive Hebbian learning in the continuous
Hopfield model. In Touretzky, D., Elman, J.,
Sejnowski, T., & Hinton, G. (Eds.), Proceedings of the
1990ConnectionistModels Summer School, pp. 10–17.
Morgan Kaufmann, San Mateo, CA.
Nelson, K., Hampson, J., & Shaw, L. K. (1993). Nouns
in early lexicons: evidence, explanations and implications.
Journal of Child Language, 20, 61–84.
Rosch, E. H., Mervis, C. B., Gray, W. D., Johnson, D. M.,
& Boyes-Braem, P. (1976). Basic objects in natural
categories. Cognitive Psychology, 8, 382–439.
Rosch, E. (1973). Natural categories. Cognitive Psychology,
7, 573–605.
Smith, L. B. & Sera,M. D. (1992). A developmental analysis
of the polar structure of dimensions. Cognitive Psychology,
24(1), 99–142.
Spelke, E. S., Breinlinger, K., Macomber, J., & Jacobson, K.
(1992). Origins of knowledge. Psychological Review,
99, 605–632.
Whorf, B. L. (1956). Language, Thought and Reality: Selected
Writings of Benjamin Lee Whorf. MIT Press,
Cambridge,MA.
Younger, B. (1990). Infants’ detection of correlations among
feature categories. Child Development, 61, 614–620.