Voxel selection masks

Our hypothesis concerned the role of the GSN in representing sensory-motor information about concepts. Therefore, we created a voxel-selection mask based on the activation-likelihood estimation (ALE) meta-analysis by Binder et al. (2009), encompassing the cortical areas that were reliably associated with “general” semantic processing (Fig. 3A). The map from Binder et al. (2009) was thresholded at p < 0.05 and converted into a binary mask. To investigate which portions of the GSN contributed the most to decoding accuracy, we performed the analysis separately on each of its five regions: lateral temporoparietal, medial parietal, medial temporal, lateral prefrontal, and medial prefrontal.

A, Masks used for voxel selection. B, C, Decoding accuracy for each encoding model in each of the two masks. Error bars represent the 99% CI. *p < 0.0005, **p < 10⁻¹⁰.

As a control, the models were also evaluated in a mask corresponding to the WFN obtained from the contrast pseudowords > rest in the present dataset, thresholded at p < 0.05 (corrected). This mask included visual, somatosensory, and motor/premotor areas, as well as the thalamus, and had minimal overlap with the GSN mask (Fig. 3A). Because these regions are more strongly activated during bottom-up perceptual processing than during top-down processing (Goebel et al., 1998; O'Craven and Kanwisher, 2000), we expected their activation patterns to encode information primarily about word form rather than semantic content. Voxels displaying a low temporal signal-to-noise ratio (<200) were excluded from both masks.