On the Acquisition of Semantic Categories

This paper, written for my Language and the Mind course, is a summary of the current state of knowledge regarding the acquisition of semantic categories–that is, what we know about how you learn the meaning of words.

The difficulty entailed by the acquisition of semantic categories is nowhere stated so eloquently as by Willard Van Ortman Quine, in the problem he calls “the indeterminacy of translation” (1960, p. 27). Quine imagines a linguist attempting to learn a hitherto unknown language. One expects that the utterances the linguist will first comprehend are those “keyed to present events” (Quine, 1960, p. 29). For instance, the linguist might hear “gavagai” from his native guide as a rabbit hops by and naturally assume the word means “rabbit.” But it is impossible, by any amount of analysis, observation, or querying for the linguist to conclusively determine that “gavagai” corresponds to “rabbit,” rather than, for instance, “Let’s go hunting” or “Look, there are some separate and distinct parts of a rabbit” (Quine, 1960, p. 52). Despite this difficulty, children of all cultures are able to acquire a lexicon and use it with “native” fluency; even 1-year-old children require fewer than ten exposures to a word in order to identify its meaning, and 2- and 3-year-olds can often approximate a word’s meaning after a single exposure (Koenig & Woodward, 2007, p. 618). How is it, then, that a human infant, unlike an infant of any other species, is capable of “associating together the most diversified sounds and ideas,” as Charles Darwin articulates it (1871, p. 86)?

Conservative estimates place the lexical knowledge of an average adult at around 60,000 entries, not including idioms, proper nouns, and decomposable words; considering that a person’s first utterances occur around one year of age, and a person’s vocabulary is almost entirely in place by the end of high school, this means that a person learns an average of ten words per day in their first 17 years of life (Bloom, 2000, p. 6). This is an impressive figure, made more remarkable by our findings that children can perform “fast-mapping,” the learning of a word (and recollection of it weeks later) after a single exposure, and in the absence of any explicit naming. Any theory that proposes to explain how we acquire semantic categories must account for a child’s capacity for rapid learning.

By examining the current literature on language and concept acquisition, we find that the presence of preferred levels of categorization, coupled with the ability to infer a speaker’s referential intent, are sufficient to allow a person (in particular, a child) to acquire semantic categories. Furthermore, as we will see, these categories are essentially the same in both children and adults, facilitating efficient learning via induction down taxonomies in both groups.

To begin, let us explicitly consider what is involved in knowing a word’s meaning. In broad terms, the meaning of a lexical item is the class of objects or actions to which the item refers; if one knows that star refers to the objects shining in the sky, one may be said to know the word’s meaning. As Paul Bloom points out, “we are comfortable translating a word from an ancient Greek text into the English word star, even though the ancient Greeks believed that stars were holes in the sky” (Bloom, 2000, p. 19). From this we conclude that even if two people have very different beliefs about the referent of a term, they may still be said to possess the same meaning for the word.

Generally, then, we can think of a word’s meaning as being a reflection of the conceptual category that it delimits. We may possess more concepts than we do words (that is, there is not a surjective function mapping from concepts to words), and there is no one-to-one relationship between concepts and words (there is not an injective mapping), but conceptual and semantic structures are closely related (Vigliocco & Vinson, 2007). For this reason, much of the literature in concepts and categorization treats a language’s semantics as interchangeable with the conceptual categories underlying its words (Vigliocco & Vinson, 2007, p. 196).

To possess a semantic category, however, a person need not be capable of distinguishing members of that category in the outer world. That is, one may use robin without being able to distinguish a robin from a sparrow or any other non-robin (Bloom, 2000, p. 19). Further evidence that conceptual categories are not identical with word knowledge comes from neuropsychology, where numerous findings indicate that very localized brain damage may bring about impairments in knowledge of specific semantic categories without affecting performance in non-verbal tasks involving those categories (Vigliocco & Vinson, 2007, p. 196). Thus, a word’s meaning may not always be identifiable in terms of an internal category representation; it may additionally (or instead) have the property of sameness of reference. Perhaps the best approach to word meaning, then, is a middle path. There is the psychological, internal aspect of meaning (called the narrow content), and there is the sociocultural, contextual aspect (called the broad content), and the sum of these determines the meaning of a word (Bloom, 2000, p. 21).

With this in mind, we will consider word meanings as expressions of underlying conceptual categories, implying that acquiring semantic categories is identical with acquiring a lexicon. This is a simplification, but for our purposes it is an acceptable one, so long as one notes that it is not necessarily the whole story.

Having outlined what we mean when we discuss semantic categories, what is their use in one’s mind—that is, why is it beneficial that one acquire them? The primary reason is that semantic categories, and especially taxonomic categories, serve as useful tools in learning about the world. Taxonomic categories facilitate generalization and induction (Carroll, 2008, p. 114; Locke, 1694, book 3, chap. 3). For instance, having learned that salamanders are a type of animal, one can correctly infer that salamanders breathe, reproduce, and grow, among many other things. There is evolutionary (selective) value in making such generalizations, too. A toddler that cannot generalize from knowing that this stove at this time burned her when she touched it to knowing that any stove at any time has the potential to burn her will do quite poorly in the world. Instead of learning “a distinct name for every particular thing,” and a complete set of features for each thing, one learns categories (Locke, 1694, book 3, chap. 3, sec. 4). Category knowledge, then, “though founded in particular things, enlarges itself by general views; to which things reduced into sorts, under general names, are properly subservient,” and one may generalize properties from one class of things to another (subordinate) one (Locke, 1694, book 3, chap. 3, sec. 4).

To be of any value though, one’s generalizations must be useful. As Jorge Luis Borges satirizes, one’s understanding of the world is not improved by dividing animals into the categories of those that have just broken a flower vase and those that have not (Borges, 1999, p. 231). How, then, does a child decide which features of an object shall determine category membership? That is, how is it that children form useful semantic categories, as opposed to the trivial, Borgesian categories that are possible?

Part of the answer here is that in working with semantic categories, children show a strong preference for what is known as the basic level. This privileged basic level is a “middle level of specificity” between the most general (e.g., a physical object) and the most narrow categorizations (e.g., an antique African mahogany secretary desk) (Murphy, 2004, p. 210). Roger Brown speculated in the 1950s that we describe things at the basic level ‘‘so as to categorize them in a maximally useful way” (Bloom, 2000, p. 148), and this suspicion has since been repeatedly confirmed. The basic level is a compromise between the assurance of accuracy obtained by categorizing objects at a very general level and the predictive power found in categorizing them at a very specific level (Murphy, 2004, p. 210).

A literature review performed by Gregory Murphy and Mary Lassaline (1997) found an overwhelming amount of experimental data indicating that people decide on an object’s basic level as a compromise between informative qualities and ease of identification, and that they show a strong bias toward the basic level when performing naming tasks. Knowledge of an object’s basic-level category is informative in that it provides a great deal of information by induction. At the same time, the categorization is relatively easy to distinguish; it is, for instance, much easier to tell a dog from a cat than it is to tell a Labrador from a Golden Retriever. Thus, “animals which have just broken a flower vase” could not be a basic level category, as it is both difficult to distinguish (at least for any animal one has not been observing for some length of time), and it provides only a trivial amount of information by induction. A variety of studies indicate that some such knowledge of inductive generalizations is likely to be stored in the mind when one learns about a category, but novel implications for naturally existing categories may also be computed in the course of processing a sentence (Carroll, 2008, p. 114).

Numerous studies confirm that children are essentially experts at understanding the basic level at as early as 2.5 years of age, but they do not develop full competency with superordinate levels until age 4, nor with subordinate levels until age 5.5 (Mervis & Crisafi, 1982; Furrer & Younger, 2005; Hammer, Diesendruck, Weinshall, & Hochstein, 2009). In one particular study, though, Jing Liu, Roberta Michnick Golinkoff, and Kimberly Sak (2001) find that 3- to 5-year-olds are able to form superordinate categories with relative ease under the right conditions. They require multiple objects of the novel superordinate category to be presented, presumably because only then is the basic level “blocked” from consideration (Liu et al., 2001, p. 1691). Additionally, a dramatic increase is seen in the ability of 3-year-olds to acquire superordinate categories when 3-D objects are presented instead of images (Liu et al., 2001). With over two hundred children in this well-controlled study, we may consider these results to be reliable. It would be beneficial, however, for previous studies concluding that 3-year-olds are unable to apply superordinate categories to be repeated to confirm the effect of 3-D objects versus 2-D representations.

When learning their first language, then, children form semantic categories most readily at the basic level. Further basic-level bias exists, however, in the realm of property induction. Sandra R. Waxman, Elizabeth B. Lynch, K. Lyman Casey, and Leslie Baer (1997) demonstrate that when children learn that a property applies to a certain object, they are most likely to infer that the property applies both to the item’s basic-level category and its subordinate category; they do not assume the newly learned property applies only to the subordinate category to which the object belongs, and they do not assume it holds for all of the object’s superordinate category (Waxman et al., 1997, p. 1077). For instance, if children learn that an Irish Setter wags its tail when it is pleased, they will also assume, in the absence of contrary evidence, that dogs in general wag their tails. Some caution is required in interpreting these results; Waxman et al. (1997) used a relatively small sample of twelve children aged 3 years 6 months to 4 years 2 months, and they replaced two children who incorrectly applied the data in all three of the test sets. Despite these limitations, though, the study’s conclusions are accepted elsewhere in the literature on categorization (Gelman, 2003, p. 56)

Thus, we see why we do not in fact find children (or adults) using Borgesian, nonsense categories. Qualities separating one category from another are not arbitrary, but are instead pragmatic, inductively useful characteristics. Members of a category share some properties such that in virtue of knowing that an object belongs to that category, one knows “further facts about it that are not true” of things outside the category (Bloom, 2000, p. 148). Since children tend to assume novel names apply at the basic level, and since they tend to apply novel information to the basic level, the Borgesian categories are as unlikely empirically as one might intuitively feel they should be.

It is important to note that the mind may not in fact express a strict version of the conceptual hierarchy implied by the superordinate/basic/subordinate-level description of category knowledge; instead, a great deal of recent research suggests what are known collectively as spreading activation models of the lexicon. These models claim that the lexicon is organized less hierarchically and more as a web of interconnecting entries, with related categories activating one another when they are accessed (Carroll, 2008, pp. 115-117). In this case, the taxonomic hierarchies implied in the traditional view of the lexicon are still useful, but primarily in considering category acquisition and the use of induction, where their predictive value is well supported.

In order to acquire semantic categories, it is crucial that a child be able to deduce a speaker’s referential intent—after all, hearing a word and matching it to an object or action are the primary means by which a child learns word meanings (Koenig & Woodward, 2007, p. 618). For a child who understands, though, that words tend to bear some relationship with the objects and events around her, and that speakers may offer clues as to their intended referent, matching word with meaning becomes a relatively straightforward process: simply search the environment for referential candidates, and see which ones, if any, match the speaker’s cues (Koenig & Woodward, 2007, pp. 618-619). Children are influenced by a speaker’s cues at as early as 18 months of age, and by age 2, they are able to apply “a very flexible social understanding” of an adult’s intentions (Akhtar & Tomasello, 2000; Tomasello, 2001, p. 141). Furthermore, in a well-controlled study using nearly fifty subjects of the target age, Michael Tomasello and Katharina Haberl (2003) found that even 12-month-olds are able to keep track of what another person has and has not seen in a situation, correctly inferring the object of an experimenter’s excitement upon returning to a room to be the thing previously unseen by her.

In actual practice, however, children simplify the search for a referent even further, explaining in part how they are capable of feats such as fast mapping. Since the late 1980s, we have known that when presented with two objects, one for which they know a name and the other for which they do not, children choose to apply a new category name to the novel object with much higher frequency than the familiar one (Diesendruck & Markson, 2001). Diesendruck and Markson (2001) confirm this result, and find that it applies not only to names, but also to new pieces of information. In one of their studies, a 3-year-old is given two novel objects, along with some piece of information applying to one of them (e.g., “My uncle gave me this.”). When asked later to give the experimenter the object to which some other novel fact applies (e.g., “I got this for Christmas”), they choose the other object, the one for which they previously knew no information, at a rate significantly higher than chance; children avoid applying either a novel name or a novel fact to a single object (Diesendruck & Markson, 2001, p. 633). Successive experiments, designed to rule out competing explanations for the phenomena, lead to the conclusion that children assume names are universal conventions and infer referential intent by evaluating what a speaker knows about the situation (Diesendruck & Markson, 2001). These results may be considered definitive; a large number of participants were tested across multiple trials, and the tests were well controlled with regard to the participants, the objects used for naming, and the novel pieces of information that were applied to the objects.

Thus, when presented with a situation filled with familiar objects (as in, for instance, a room in a child’s home) and one unfamiliar object, a child will most often apply novel information to the unfamiliar object. Likewise, “if a speaker says two different things, regardless of whether it is two labels or two facts, she or he probably has two different referential intentions in mind” (Diesendruck & Markson, 2001, p. 634). These judgments on the child’s part are, as we might imagine, quite consistent with experience, and they indicate a powerful feature of a child’s word-learning ability: once a small handful of names or facts have been acquired, building on them becomes a simple matter; a child applies the rule that for new terms, or new facts not clearly applying to familiar objects, a new referent should be found to which they can apply.

In the preceding experiments, children were explicitly presented with novel names. However, children do not require this sort of explicit object naming, common in Western cultures, to acquire semantic categories. Paul Bloom notes that the linguistic practices of some cultures, most notably those of the Kaluli people of Papua New Guinea, are conspicuously devoid of parents’ labeling of objects for children (2000, p. 8). Furthermore, when Kaluli children name objects for adults, they are greeted with indifference (Bloom, 2000, p. 8). Despite this seeming handicap, Kaluli children become perfectly fluent, suggesting that the direct naming present in Diesendruck and Markson’s experiments (2001) might be replaced with indirect discussion of the novel names and information without changing the results significantly.

Having considered how children decide on an object’s category, as well as how they apply novel information about an object, we turn to the question of whether significant differences exist between the semantic categories of infants, older children, and adults. The question is difficult to answer partly because of the challenges inherent in studying young children. With infants, it is sometimes problematic to ensure that subjects focus on the environmental features that the experimenter wishes to test—for instance, when displaying a series of images on a screen for an infant in order to measure his reaction to each image, the child might actually be interested only in the screen itself (Murphy, 2004, p. 289). For this reason, among others, it is more common for toddlers and pre-school age children to be used in studying children’s semantic categories (Bloom, 2000, p. 12).

When studying the categories of preschoolers, more direct, language-driven methods can be used than the standard dishabituation and paired preference tests used with infants. In this case, though, it can be difficult to discern whether the cause of a failure lies with the child’s understanding of a taxonomy or her understanding of the test question. After all, it is a strange question indeed for a child to be asked a question like “If all trees have lignin, do banyan trees have lignin?” (Murphy, 2004, p. 326). The child will likely recognize that the experimenter knows much more about lignin than he, and one does not often ask questions to which one already knows the answer. Thus, the fact that older children perform significantly better on these sorts of explicit naming tasks may indicate not that they understand category hierarchies or induction using taxonomies better, but simply that they better understand the sort of metalinguistic word game in play here (Murphy, 2004, p. 326). Even if a child can play the word game, she might simply misunderstand the question. For instance, in a classic study by Carol Smith on children’s understanding of hierarchies, the following question was posed: “A tody is a kind of bird. Does a tody have to be a robin?” To this a child responded, “Yes, or could be a robin or blue jay or sparrow” (Smith, 1979, p. 448). While this is technically a wrong answer, it does not necessarily arise from a misunderstanding of how the semantic category applies.

With these difficulties present in studying children’s categories, it is not surprising that in the first years of serious study of children’s categories, it was commonly assumed that the distinctions used by children in creating categories were fundamentally different than those used by adults, and that children’s concepts must somehow change over time to become adult-like. For instance, studies by Jean Piaget in the 1960s conclude that children performing sorting tasks often do so on the basis of thematic relations, wherein pairs of objects might be related in some way without any overall organization, instead of on the basis of shared properties as in a taxonomy (Murphy, 2004, p. 319). More recent work, however, suggests that these findings are artifacts of the particular experimental methods used. Jerry Fodor points out informally that if children really were to categorize objects thematically under normal circumstances, as Piaget and others conclude, we would have tremendous difficulty distinguishing whether a child was talking about a dog or a leash, a thematically related pair (Murphy, 2004, p. 320); thus, even apart from further research, we have strong evidence that children must be able to form taxonomic semantic categories.

More formally, Simone Nguyen and Gregory Murphy (2003) test 3-, 4-, and 7-year-olds, as well as adults, on their ability to categorize objects by three different methods: using script categories (e.g., breakfast foods), taxonomic categories (e.g., fruits), and evaluative categories (e.g., junk food). From five experiments using a very large number of subjects, they conclude that from age 3 on, children are able to correctly use both taxonomic and script categories, and that the two appear to develop simultaneously; children are as competent using one as they are with the other (Nguyen & Murphy, 2003, p. 1802). Thus, early results à la Piaget reflect only children’s ability to use script categories, not their full category competence. With this in mind, Nguyen and Murphy say, “[the] Piagetian view that children do not have taxonomic concepts has been largely discarded in recent years” (2003, p. 1799). Since children exhibit adult-level proficiency with taxonomic categories as early as 3 years of age, the work of Piaget and others after him fails to demonstrate that children’s concepts are fundamentally different from those of adults.

Another piece of evidence is sometimes cited in support of this claim, however. A child’s first words are often said to “blur the semantic distinctions between objects, properties, and actions” (Bloom, 2000, p. 36). For instance, a child may use the verb “to fly” to refer to birds, or call a window pane water (Dromi, 1987; Carroll, 2008, p. 268). Bloom suggests, however, that these might not be true speech errors, that they may arise not from a misunderstanding of the lexical and syntactic categories involved, but instead from the child doing the best he can with the lexical and syntactic knowledge he has (2000, p. 37). For instance, a child who points at a cookie jar and says “Cookie!” may not do so because she thinks cookie refers to all cookie-related objects; instead, the background reasoning taking place may be along the lines of “I know about cookies, and this is clearly not a cookie, but it’s related to cookies, so the best thing to call it is probably cookie” (Bloom, 2000, p. 37).

Therefore, in the absence of further, truly specific research, we conclude that there is no reason to believe that children’s first words (and thus the categories behind them) are essentially different from those of adults. Furthermore, having considered the two major arguments against children having concepts similar to adults, we conclude that the semantic categories of children and adults are of a similar character, and that semantic category acquisition proceeds similarly in both groups.

It is worth noting that there is some disagreement over the usefulness of the sorts of category testing we discuss above. Arthur Markman and Brian Ross (2003) suggest that the laboratory tasks that subjects (both children and adults) are generally tested on, wherein the subjects’ sole, explicit goal is to learn classifications, are a poor measure of real category acquisition. The majority of actual taxonomy acquisition, they note, takes place for a pragmatic purpose; what people remember about objects is specific to the way in which they interact with them, so there is likely “no single category-learning mechanism” (Markman & Ross, 2003, p. 595). This view, it should be noted, is strongly supported by more recent research, as revealed in the literature review by Gregory Ashby and Todd Maddox (2005), and most recently in the experiments on natural category use performed by Aaron Hoffman and Bob Rehder (2010).

Markman and Ross believe people know much more about categories than is generally revealed in testing; we know that category knowledge allows one to interpret novel noun phrases through context, and allows one to be particularly sensitive to the characteristics that set one category apart from its neighbors, but these features are seldom tested in the laboratory (2003, p. 596). Thus, in examining the extant literature on classification, the pair concludes that mismatches between current models of category knowledge—prototype theory, feature-based models, and so on—and observations of natural category use are a consequence of the overwhelming focus on pure classification in laboratory tests (Markman & Ross, 2003). This oversight, they say, stems from the very common, very mistaken assumption that there is indeed some universal means of category acquisition applicable to all types of things (Markman & Ross, 2003, pp. 595-596; Ashby & Spiering, 2004).

What does this imply for the research we have discussed? Clearly there is some (useful, well-founded) dissent over the methods employed in these “traditional”-style studies. In all likelihood, however, the result is that the studies conclude too little. That is, what has been shown by such studies is true, and accurate with regard to people’s ability to acquire lexical categories, but it is also likely that people are capable of more than the studies acknowledge.

Having discussed category acquisition as it acts in the mind, we turn now to semantic categories as they exist in the brain. The neurolinguistic literature is in clear consensus regarding the brain regions involved in novel category acquisition. The prefrontal cortex and basal ganglia are primarily activated in such tasks, though different types of learning activate other, distinct brain regions (Ashby & Spiering, 2004). For instance, in their literature review, Gregory Ashby and Brian Spiering (2004) find that different areas are activated when acquiring categories by rule-based tasks compared to information integration or prototype distortion tasks.

A review of the literature on the topic of brain storage of lexical categories by Karalyn Patterson, Peter J. Nestor, and Timothy T. Rogers (2007) notes that virtually all contemporary theories of semantic storage agree on one point: the majority of a person’s category knowledge is related to perception and action, and is therefore stored in the brain in areas that either overlap with or are identical to the regions responsible for perceiving and acting. Some of the most poignant research supporting this conclusion involves words and sentences referring to motion. Processing these units of language has been shown to activate primary motor regions, indicating that the same portions of the brain responsible for controlling action are also engaged in understanding language related to it (Vigliocco & Vinson, 2007, p. 196). Thus, contemporary theories of semantic category storage hold that most knowledge is widely distributed across the brain.

Contemporary theories of semantic storage may be classified as wholly distributed or as distributed with a central “hub” (Patterson et al., 2007, p. 976). In the distributed-only view, access to a certain lexical item occurs by accessing information about its shape, color, actions, and so on independently, as a task requires them.  In the distributed-plus-hub view, however, knowledge of the different features of a category all connect to a shared hub in the anterior temporal lobes. This hub, then, facilitates access to the features in a similar way regardless of the task requiring access (Patterson et al., 2007, p. 977).

Patterson et al. conclude that the distributed-plus-hub view is most plausible in light of evidence from various disorders. The presence of amodal semantic impairment (i.e., impaired knowledge of a certain lexical category affecting all sensory modalities) in patients with a variety of acquired disorders seems to indicate the presence of a relatively localized region through which a lexical item is accessed (Patterson et al., 2007, p. 978). Currently, however, there is no broad consensus in the field; the existence of a semantic categorical hub is still disputed, so while we may conclude with certainty that storing knowledge of most word meanings involves widely distributed portions of the brain, further research is required to definitively conclude whether it also involves a shared region facilitating future access.

The question of the brain’s representation of stored category knowledge, then, is not solved. It is still an area of active study. The same may be said of the category models and methodology examined by Markman and Ross (2003). While their critique has been well received in the field, with over 150 new studies citing their paper since its publication, a great deal of prior research could be repeated using methods revised according to their suggestions. In some cases, significant changes to prior studies’ conclusions may be expected from doing so. Of special interest in this case are studies designed to discover how early a child is able to usefully apply category knowledge, such as the experiments of Jing Liu et al. (2001).

On the other hand, Quine’s problem of reference may safely be considered solved: the acquisition of semantic categories with correct reference, we find, is facilitated by being aware of a speaker’s intent and sharing with the speaker a bias for the basic level of categorization. As methods of categorization are essentially the same between children and adults, communicative intent is not nearly so opaque as Quine suggests. Still open to discussion, however, is the degree to which social versus environmental cues affect the conclusions a person draws about referential intent, as is the best model of category structure in the mind.

Thus, the study of semantic category acquisition is by no means a closed field. Future research in the preceding areas will continue to sharpen our understanding of the categorization performed by both children and adults, allowing finer measures of what language learners know of categories and when they acquire that knowledge.

References

Akhtar, N., & Tomasello, M. (2000). The social nature of words and word learning. In Becoming a word learner: a debate on lexical acquisition (pp. 115-135). New York: Oxford University Press.

Ashby, F. G., & Maddox, W. T. (2005). Human category learning. Annual Review of Psychology, 56, 149-78.

Ashby, F. G., & Spiering, B. J. (2004). The neurobiology of category learning. Behavioral and Cognitive Neuroscience Reviews, 3(2), 101-113.

Bloom, P. (2000). How children learn the meanings of words. Cambridge, MA: MIT Press.

Borges, J. L. (1999). Selected non-fictions (Vol. 1). (E. Weinberger, Ed., E. Allen, S. J. Levine, & E. Weinberger, Trans.) New York: Viking.

Carroll, D. W. (2008). Psychology of language (5th Edition ed.). Belmont, CA: Thomson/Wadsworth.

Darwin, C. R. (1871). The descent of man, and selection in relation to sex (Vol. 1). New York: D. Appleton and Company.

Diesendruck, G., & Markson, L. (2001). Children’s avoidance of lexical overlap: A pragmatic account. Developmental psychology, 37(5), 630-641.

Dromi, E. (1987). Early lexical development. New York: Cambridge University Press.

Furrer, S. D., & Younger, B. A. (2005). Beyond the distributional input? A developmental investigation of asymmetry in infants’ categorization of cats and dogs. Developmental Science, 8(6), 544-550.

Gelman, S. A. (2003). The essential child: Origins of essentialism in everyday thought. New York: Oxford University Press.

Gómez, R. L. (2007). Statistical learning in infant language development. In M. G. Gaskell (Ed.), Oxfod handbook of psycholinguistics (pp. 601-16). New York: Oxford University Press.

Hammer, R., Diesendruck, G., Weinshall, D., & Hochstein, S. (2009). The development of category learning strategies: What makes the difference? Cognition, 112(1), 105-119.

Hoffman, A. B., & Rehder, B. (2010). The costs of supervised classification: The effect of learning task on conceptual flexibility. Journal of Experimental Psychology, 139(2), 319-340.

Jackendoff, R. (2002). Foundations of language: Brain, meaning, grammar, evolution. New York: Oxford University Press.

Koenig, M. A., & Woodward, A. (2007). Word learning. In M. G. Gaskell (Ed.), Oxford handbook of psycholinguistics (pp. 617-26). New York: Oxford University Press.

Liu, J., Golinkoff, R. M., & Sak, K. (2001, November/December). One cow does not an animal make: Young children can extend novel words at the superordinate level. Child Development, 72(6), 1674-1694.

Locke, J. (1694). An essay concerning humane understanding (2nd Edition ed.).

Markman, A. B., & Ross, B. H. (2003). Category use and category learning. Psychological bulletin, 129(4), 592-613.

Mervis, C. B., & Crisafi, M. A. (1982). Order of acquisition of subordinate-, basic-, and superordinate-level categories. Child Development, 53(1), 258-266.

Moss, H. E., Tyler, L. K., & Taylor, K. I. (2007). Conceptual structure. In M. G. Gaskell (Ed.), Oxford handbook of psycholinguistics (pp. 217-234). New York: Oxford University Press.

Murphy, G. L. (2004). The big book of concepts. Cambridge, MA: MIT Press.

Murphy, G. L., & Lassaline, M. E. (1997). Hierarchical structure in concepts and the basic level of categorization. In K. Lamberts, & D. R. Shanks (Eds.), Knowledge, concepts, and categories (pp. 93-131). Cambridge, MA: MIT Press.

Nguyen, S. P., & Murphy, G. L. (2003). An apple is more than just a fruit: Cross-classification in children’s concepts. Child Development, 74(6), 1783-1806.

Patterson, K., Nestor, P. J., & Rogers, T. T. (2007). Where do you know what you know? The representation of semantic knowledge in the human brain. Nature Reviews Neuroscience, 8, 976-987.

Quine, W. V. (1960). Word and Object. Cambridge, MA: MIT Press.

Smith, C. L. (1979). Children’s understanding of natural language hierarchies. Journal of Experimental Child Psychology, 27(3), 437-458.

Tomasello, M. (2001). Perceiving intentions and learning words in the second year of life. In M. Bowerman, & S. C. Levinson (Eds.), Language acquisition and conceptual development (pp. 132-158). Cambridge: Cambridge University Press.

Tomasello, M., & Haberl, K. (2003). Understanding attention: 12- and 18-month-olds know what is new for other persons. Developmental Psychology, 39(5), 906-912.

Vigliocco, G., & Vinson, D. P. (2007). Semantic representation. In M. G. Gaskell (Ed.), Oxford handbook of psycholinguistics (pp. 195-215). New York: Oxford University Press.

Waxman, S. R., Lynch, E. B., Casey, K. L., & Baer, L. (1997). Setters and samoyeds: The emergence of subordinate level categories as a basis for inductive inference in preschool-age children. Developmental Psychology, 33(6), 1074-1090.

Xu, F. (2007). Concept formation and language development: count nouns and object kinds. In M. G. Gaskell (Ed.), Oxford handbook of psycholinguistics (pp. 617-26). New York: Oxford University Press.