A Computational Model of Semantic Memory Categorization: Identification of a Concept’s Semantic Level from Feature Sharedness

Abstract

Recent studies have shown that members of superordinate concepts share less features than members of basic-level concepts. An artificial neural network model was implemented to evaluate whether feature sharedness could distinguish between these two types of concepts and whether lesioning the network would particularly affect less shared features and superordinate categorization. The model was successful in the semantic categorization test, supporting the idea that superordinate and basic-level concepts can be distinguished on the basis of feature sharedness. In contrast, lesion results proved that the model structure was not adequate to evaluate the relation between feature sharedness, processing requirements, and patient performance. Limitations and future directions for modeling semantic memory and for semantic computing are discussed.