Investigating the semantics of abstract concepts: Evidence from a property generation game

Concrete nouns are the dominant stimuli in semantic priming, property generation, and other literatures that contribute to our knowledge about semantic representation. In contrast, we know very little about abstract concepts, even though a substantial proportion of English nouns (over 40%) are rated as abstract. I take initial steps toward resolving disagreements in the literature about the nature of abstract concept representations by collecting and coding the largest set of property generation norms for abstract concepts to date. Study participants participated in a game in which they described abstract and concrete concepts. The data were investigated using corpus-based measures of association strength (measures of the degree to which concept words and description words occur in the same contexts), contextual similarity (measures of the degree to which they occur in similar contexts), and semantic density. Consistent with dual coding theory and other theories in which language plays a more fundamental role for abstract concepts, corpus-based measures of semantic density and cue-target similarity more accurately predicted lexical decision times and priming latencies for abstract (vs. concrete) concepts. However, measures of association strength, such as pointwise mutual information, were higher between cues and their descriptions for concrete (vs. abstract) cues. Further study suggested that measures of association strength were much better at detecting relationships that were relevant to concrete words, such as perceptual relationships, than were measures of contextual similarity. Computational simulations of the feature generation task suggested that a combination of both types of measures resulted in better fits to human data than did either considered in isolation. In conjunction with additional evidence from regression and simulation, these results suggest that measures of association strength and contextual similarity tap different psychologically relevant constructs, and that the semantic system makes use of both kinds of information.