Letting clusters and paths emerge from early semantic hypernetwork structure of features and their nouns
College of Liberal Arts and Sciences
Social and Behavioral Sciences
Recently, Hills and colleagues (2009 a,b) have described the potential categorical information contained in the features of early-learned basic level nouns by examining the binary graph-theoretic properties of developing noun-feature networks with a deterministic method: the clique percolation. The networks were built from the overlap of words normatively acquired by children at three different ages: 20 months --21 nouns--, 25 months --56 nouns-- and 30 months -- 130 nouns-- and their perceptual and functional features from adult feature generation norms (1394 token features). The resulting networks had small-world structure, indicative of a high degree of feature overlap in local clusters. The results also suggested that overlapping features among these early-learned nouns created higher-order groupings common to adult taxonomic designations as well as ad hoc categories. However, these methods are limited as they are only descriptive and yield minimal semantic information such as the degree of connectivity of local structures, whether a link of similarity (unspecified) exist or not between two neighbors or identify cliques of connectivity with unspecified semantics. To account for these limitations, we present a different type of representation, the hypernetwork (Berge, 1956) that keeps the semantic in the system, and a different formalism the Formal concept analysis (FCA), a non deterministic method, that builds relationships of containment (Wille, 1984). Furhte, machine-learning algoritms automatically cluster and build inclusions for the features and their nouns ( 21, 56 and 130 object names) at the 3 different ages mentionned above. We compare our results to the percolation method results obtained by Hills and colleagues. The power of the system lies in its automaticity and its ability to form many intermediate clusters at all stages of the growth of the network as well as showing the emerging paths from that structure.
Workshop on infant language development
Maouene, Josita; Canada, Kelsie; and Maouene, Mounir, "Letting clusters and paths emerge from early semantic hypernetwork structure of features and their nouns" (2013). Faculty Scholarly Dissemination Grants. 1139.