Using fMRI activation to conceptual stimuli to evaluate methods for extracting conceptual representations from corpora

We present a series of methods for deriving conceptual representations from corpora and investigate the usefulness of the fMRI data and machine learning methodology of Mitchell et al. (2008) as a basis for evaluating the different models. Within this framework, the quality of a semantic model is quantified by its ability to predict the fMRI activation associated with conceptual stimuli. Mitchell et al. used a manually-acquired set of verbs as the basis for their semantic model; in this paper, we also consider automatically acquired feature-norm-like semantic representations. These models make different assumptions about the kinds of information available in corpora that is relevant to representing conceptual knowledge. Our results indicate that automatically-acquired representations can make equally powerful predictions about the brain activity associated with the stimuli.