On Discriminating fMRI Representations of Abstract WordNet Taxonomic Categories

How abstract knowledge is organised is a key question in cognitive science, and has clear repercussions for the design of artifical lexical resources, but is poorly understood. We present fMRI results for an experiment where participants imagined situations associated with abstract words, when cued with a visual word stimulus. We use a multivariate-pattern analysis procedure to demonstrate that 7 WordNet style Taxonomic categories (e.g. 'Attribute', 'Event', 'SocialRole'), can be decoded from neural data at a level better than chance. This demonstrates that category distinctions in artificial lexical resources have some explanatory value for neural organisation. Secondly, we tested for similarity in the interrelationship of the taxonomic categories in our fMRI data and the associated interrelations in popular distributed semantic models (LSA,HAL,COALS). Although distributed models have been successfully applied to predict concrete noun fMRI data (e.g. Mitchell et al., 2008), no evidence of association was found for our abstract concepts. This suggests that development of new models/experimental strategies may be necessary to elucidate the organisation of abstract knowledge.