Research Article Language Is Not Just for Talking Redundant Labels Facilitate Learning of Novel Categories Gary Lupyan,1,2 David H. Rakison,1 and James L. McClelland3 1Department of Psychology, Carnegie Mellon University 2Center for the Neural Basis of Cognition, Pittsburgh, Pennsylvania and 3Department of Psychology, Stanford University ABSTRACT���In addition to having communicative func- tions, verbal labels may play a role in shaping concepts. Two experiments assessed whether the presence of labels affected category formation. Subjects learned to catego- rize ������aliens������ as those to be approached or those to be avoided. After accuracy feedback on each response was provided, a nonsense label was either presented or not. Providing nonsense category labels facilitated category learning even though the labels were redundant and all subjects had equivalent experience with supervised cate- gorization of the stimuli. A follow-up study investigated differences between learning verbal and nonverbal asso- ciations and showed that learning a nonverbal association did not facilitate categorization. The findings show that labels make category distinctions more concrete and bear directly on the language-and-thought debate. The ability to form categories is ubiquitous in the animal kingdom, yet it is only humans who habitually use names for theirs.1 Learning to name things allows for linguistic commu- nication, and it has been argued that words stabilize abstract ideas in working memory and make them available for in- spection (Clark, 1997 James, 1890 Rumelhart, Smolensky, McClelland, & Hinton, 1986 Vygotsky, 1962). Although a great deal of research has examined the communicative function of words (see Roy, 2005, for a novel approach) and the role of categorization in perception (Goldstone, 1994 Goldstone & Barsalou, 1998), to date few studies have addressed the role that category names themselves play in the learning of categories. Learning that two objects are both called ������dax������ may commu- nicate to the learner that they share commonalities, which in turn may cause the objects to be grouped into a common category (Waxman & Markow, 1995). The presence of words may there- fore turn an unsupervised learning task into a supervised one (Cabrera & Billman, 1996). But do words do more than alert the learner to group together similarly named objects? Although nonhuman animals categorize nonlinguistically, humans have the potential benefit of labels���category names���as they decide to which category an object belongs. The crucial question, then, is how the presence of labels affects human categorization. Investigations of the role of names in categorization are re- lated to investigations of how categorization affects perception. Numerous studies have shown that perception of initially equidistant items is affected by categorization most notably, categorizing items into different groups makes them more dis- similar, and this effect has been shown to be the result of rep- resentational change rather than of a bias in similarity ratings due to knowing that two objects have different labels and thus belong to different categories (Goldstone, Lippa, & Shiffrin, 2001). Though this line of research is relevant to the present study, the question addressed by the present study is whether category labels facilitate category formation when categoriza- tion experience is controlled. The specific question of how labels affect category learning has been most thoroughly investigated in two contexts. First, work on children���s language acquisition has revealed that in- fants as young as 9 months of age more readily individuate la- beled than unlabeled objects (Xu, 2002), and that by 12 months of age, infants have an expectation that words refer to object categories (Waxman & Hall, 1993 Waxman & Markow, 1995). Waxman and Markow argued that words serve as ������invitations to form categories������ and that superordinate labels, such as ������vehi- cle,������ lead children to form the appropriate category. Labels that are correlated with regularities in the world (e.g., shape is pre- dictive of solid, but not nonsolid, categories) have been shown to improve the learning of these regularities (Yoshida & Smith, 2005). Address correspondence to Gary Lupyan, 342C Baker Hall, Car- negie Mellon University, Pittsburgh, PA 15213, e-mail: glupyan@ cnbc.cmu.edu. 1 We define a category as a group of stimuli that evoke a common response. That response may be verbal (e.g., calling some colors ������blue������ despite sub- stantial variation in their hue) or nonverbal (e.g., performing the same action with different-looking objects or clustering all the blue items together during a sort). The stimuli within a category typically bear some perceptual or functional relationship to each other. PSYCHOLOGICAL SCIENCE Volume 18���Number 12 1077 Copyright r 2007 Association for Psychological Science

In experiments such as those conducted on infants by Wax- man and her colleagues, the labels are semantically empty. A different sort of labeling influence can be seen in older children. For instance, learning to associate small, medium, and large groupings with the labels ������baby,������ ������mommy,������ and ������daddy,������ re- spectively, facilitates relational judgments, enabling children to transfer the size relation to novel stimuli in this case, the effect of the labels is tied to their semantics (Kotovsky & Gentner, 1996 Ratterman & Gentner, 1998). A second line of research on how labels affect category learning builds on James���s (1890) law of dissociation by varying concomitants���the idea that associating A with B on one occa- sion and with C on another leads to A becoming dissociated from both B and C, and thereby becoming a more abstract object. Extending James���s reasoning, Miller and Dollard (1941) hy- pothesized that associating different responses with otherwise similar stimuli should increase the perceived difference be- tween the stimuli. Miller and Dollard saw object names as a kind of motor response, and so hypothesized that associating un- differentiated stimuli with different names increases the differences between the stimuli and facilitates placing them into separate categories. This hypothesis was tested with mixed re- sults because the relevant experiments failed to control for stimulus familiarity (e.g., Arnoult, 1953 Battig, 1956 Rossman & Goss, 1951), and it was unclear whether increased discrim- inability or facilitated categorization arose from the learned associations between stimuli and labels or merely from addi- tional experience with the stimuli (Gibson & Gibson, 1955 Robinson, 1955). It is important for us to be clear about what we mean by the term label. We use the term to refer to anything that is (a) con- sistently correlated with a category and (b) used to refer to a category. The category can comprise objects, sounds, actions, spatial relations, and so on. In principle, any cue can serve as a label, and what counts as a label is a function of experience and environment. Individuals proficient in a sign language treat motor gestures as labels, whereas the subjects in the study re- ported here, being hearing college students, have had an im- mense amount of experience treating words (both oral and written) as labels. It is often through words that people come to know what cat- egories are relevant. For instance, calling certain objects ������chairs������ suggests that chairs are a useful and relevant category. The question addressed by the present work was not whether labels facilitate category formation because they point out the relevant categories, but rather whether labeled categories are easier to acquire than unlabeled categories because they have a name���even when categorization can be performed without relying on labels. Thus, the two experiments we report in this article are the first to directly test the idea that labels make category differences ������more concreted������ (James, 1890, p. 333). Our experiments address two main questions: Are labeled cat- egories easier to learn than unlabeled categories even when the labels are entirely redundant, contributing no additional infor- mation? How does associating stimuli with verbal labels com- pare with learning a nonlinguistic category association? EXPERIMENT 1 In the first experiment, we measured performance of subjects learning to associate novel objects with behavioral responses. Some subjects performed this task while learning names for the stimulus categories, and others did not learn names. We ex- pected that if it is easier to learn named than unnamed cate- gories, performance would be superior in the former condition. Method Subjects Forty-eight Carnegie Mellon University undergraduates (ages 18���24) participated in the experiment for course credit. The subjects were randomly assigned to label and no-label groups. Data from 4 subjects were excluded because they did not follow instructions. Data for the test phase of the experiment were not available for 2 subjects because of experimenter error. Materials The stimuli were a subset of the YUFO stimulus set (Gauthier, James, Curby, & Tarr, 2003). Items in one category (shown on the left in Fig. 1) had flatter bases and a subtle ridge on their ������heads.������ Items in the other category (shown on the right in Fig. 1) had more rounded bases and smoother heads. Subjects��� responses on a questionnaire following the category training Fig. 1. The two categories learned by the subjects. The stimuli on the left have flatter bases and a subtle ridge on their ������head������ the stimuli on the right have more rounded bases and a smoother head. 1078 Volume 18���Number 12 Redundant Labels Facilitate Categorization