The Emergence of Semantic Categories from Distributed Featural Representations

Abstract

This paper presents a computational model of semantic memory, trained with behaviourally inspired vectors.The results are consistent with the conceptual structure account (Tyler, Moss, Durrant-Peatfield & Levy, 2000), which claims that concepts can be understood, and the effects of random damage predicted, based on (i) the number of correlations its features make, and (ii) the distinctiveness of those correlated features; the former indicating category membership, the latter distinguishing between concepts.The model shows a changing direction of domain-specific deficits as damage accumulates (animals concepts lost first, then objects upon severe lesioning).Also, the pattern of error differs between domains; animals tend to be confused for other members of the same category, whilst object errors disperse more widely across categories and domain.Recent neuropsychological evidence demonstrates a similar pattern for semantically impaired patients.For both patients and the model, this can be attributed to the timing of featural loss: distinctive features are lost earlier than shared features.The model demonstrates that the relative timing of feature loss differs between domains, resulting in the emergence of domain-specific effects.