Categorical structure in early semantic networks of nouns

Categorical Structure in Early Semantic Networks of Nouns Thomas Hills (thills@indiana.edu) Department of Psychological and Brain Sciences, 1101 E. Tenth Street Bloomington, IN, 47405 USA Mounir Maouene (mmaouen@ensat.ac.ma) UFR : Artificial Intelligence and Bioinformatics, ENSAT, Tangier, Morocco Josita Maouene (jcmaoune@indiana.edu) 1 Adam Sheya (aasheya@indiana.edu) 1 Linda B. Smith (smith4@indiana.edu) 1 define different superordinate kinds. In particular, Rogers and McClelland (2004) demonstrated how patterns of coherent co-variation across features could create superordinate categories. The idea of co-variation—of systems of correlations in overlapping feature patterns— provides one possible way to address the criticism that feature correlations are too-unconstrained to explain human category structure (e.g. Ahn, Kalish, Medin & Gelman, Although previous studies have explored how feature- correlations structure specific categories, no prior studies have focused on how this structures the system of noun categories children learn. The purpose of the present study is to provide such a description. A descriptive study seems pre-requisite to the examination of any claims about what feature correlations—of themselves or in concert with other processes—can do by way of creating children's category knowledge. The present study specifically examined 130 nouns that are among the first nouns children learn and the structure of the features (derived from adult feature generation studies) associated with those categories. The analyses concentrate on perceptual and functional features-- features of things that should be evident in even young children's experiences. Perceptual and functional features are also of interest because of several disputes of the possibly different roles that the two kinds of features might play in category organization. Perceptual features are typically defined as static visual features such as color, shape, and part structure (e.g., having legs or not); functional features typically encompass roles (e.g., used to drink from), behaviors (breathes or flies), and transient properties (e.g., can be opened). There have been suggestions that different kinds of categories differ in the relative importance of these two kinds of features with perceptual features perhaps more important for animals and functional features more important for artifacts, although there is considerable dispute (see De Renzi & Lucchelli, 1994; Komatsu 1992). In the literature on cognitive development, the debate centers around the relative importance of the two classes of Abstract Despite what we know about children’s ability to categorize, it is not clear to what extent information in the environment is capable of facilitating higher-order category knowledge, nor to what extent different kinds of object features play different kinds of roles. As a start we built a network of 130 early-learned nouns with 1394 perceptual and functional features as given by adult judgments. Then we analyzed the basic structural properties of the network. These revealed a small world structure and a high degree of feature overlap in local clusters. To identify the local clusters, we used a clique percolation algorithm to parse the network in terms of the statistical properties of feature overlap. This enabled us to identify clusters of items with a strong resemblance to common categories, such as animals, foods, and vehicles. Perceptual and functional features were found to play different roles in the categorization, with functional information being less redundant but more specific than perceptual information. Keywords: early semantic network, clusters, perceptual and functional features, percolation algorithm, feature correlations. Theories of human category structure are often based on feature-category correlations. Traditional theories of categorization posited necessary and sufficient defining features that determine category membership (reviewed in Murphy, 2002), but other important theories posit probabilistic feature correlations with category members typically being similar across clusters of correlated properties (Rosch et al., 1976). It is generally accepted that people learn category-feature correlations (McRae, Cree, Seidenberg & McNorman, 2005; Younger & Cohen, 1990) and there is supporting evidence both from developmental (Rakison & Poulin-Dubois, 2002) and category-specific deficit studies (for example, Caramazza & Shelton, 1998; Tyler, Moss, Durrant-Peatfield & Levy, 2002, but see Warrington & Shallice, 1984). Computational studies suggest further that the latent structure available in a system of many categories with many feature-category correlations may be sufficient to