Abstract
An influential position in lexical semantics holds that semantic representations for words can be derived through analysis of patterns of lexical co-occurrence in large language corpora. Firth (1957) famously summarised this principle as "you shall know a word by the company it keeps". We explored whether the same principle could be applied to non-verbal patterns of object co-occurrence in natural scenes. We performed latent semantic analysis (LSA) on a set of photographed scenes in which all of the objects present had been manually labelled. This resulted in a representation of objects in a high-dimensional space in which similarity between two objects indicated the degree to which they appeared in similar scenes. These representations revealed similarities among objects belonging to the same taxonomic category (e.g., items of clothing) as well as cross-category associations (e.g., between fruits and kitchen utensils). We also compared representations generated from this scene dataset with two established methods for elucidating semantic representations: (a) a published database of semantic features generated verbally by participants and (b) LSA applied to a linguistic corpus in the usual fashion. Statistical comparisons of the three methods indicated significant association between the structures revealed by each method, with the scene dataset displaying greater convergence with feature-based representations than did LSA applied to linguistic data. The results indicate that information about the conceptual significance of objects can be extracted from their patterns of co-occurrence in natural environments, opening the possibility for such data to be incorporated into existing models of conceptual representation.