The application of multivariate approaches to neuroimaging data analysis is providing cognitive neuroscientists with a new perspective on the neural substrate of conceptual knowledge. In this paper we show how the combined use of decoding models and of representational similarity analysis (RSA) can enhance our ability to investigate the inter-categorical distinctions as well as the intra-categorical similarities of neural semantic representations. By means of a linear decoding model, we have been able to predict the category of the words subjects were seeing while undergoing a functional magnetic resonance images (fMRI) acquisition. Moreover, RSA in anatomically defined region of interest (ROIs) revealed a significant correlation with length of words and real item size in primary and secondary visual areas (V1 and V2), while a semantic distance effect was significant in inferotemporal areas (BA37 and BA20). Together, these findings illustrate the possibility to decode the distinctive neural patterns of semantic categories and to investigate the peculiar aspects of the neural representations of each single category. We have in fact been able to show a significant correlation between cognitive and neural semantic distance and to describe the gradient of information coding that characterizes the ventral path: from purely perceptual to purely conceptual. These results would not have been possible without a double exploration of the same dataset by means of decoding models and RSA