Learning semantic features for fMRI data from definitional text

Abstract

(Mitchell et al., 2008) showed that it was possible to use a text corpus to learn the value of hypothesized semantic features characterizing the meaning of a concrete noun. The authors also demonstrated that those features could be used to decompose the spatial pattern of fMRI-measured brain activation in response to a stimulus containing that noun and a picture of it. In this paper we introduce a method for learning such semantic features automatically from a text corpus, without needing to hypothesize them or provide any proxies for their presence on the text. We show that those features are effective in a more demanding classification task than that in (Mitchell et al., 2008) and describe their qualitative relationship to the features proposed in that paper.