Decoding the meaning of unconsciously processed words using fMRI-based MVPA

Abstract

Does the human brain elicit patterns of activity associated with the meaning of words in the absence of conscious awareness? Do such non-conscious semantic representations generalize across languages? This study aimed to address these questions using fMRI-based multivariate pattern analysis (MVPA) in a masked word paradigm. Animal and non-animal words were visually presented in two different languages (i.e. Spanish and Basque). Words were presented very briefly and were masked. On each trial, participants identified the semantic category and provided a visibility rating of the word. A support vector machine (SVM) was used to decode word category from multivoxel patterns of BOLD responses in seven canonical semantic regions of a left-lateralized network that were prespecified based on a previous meta-analysis. We show that the semantic category of non-conscious words (i.e. associated with null visual experience and chance-level discrimination performance) can be significantly decoded from BOLD response patterns. For Spanish, such discriminative patterns of BOLD responses were consistently found in inferior parietal lobe, dorsomedial prefrontal cortex, inferior frontal gyrus and posterior cingulate gyrus. While for Basque, these were found in ventromedial temporal lobe and posterior cingulate gyrus. All of the areas identified have previously been associated with semantic processing in studies involving animals-tools and animals-artifacts contrasts. In conscious trials, such patterns were found to be distributed over all seven regions of the semantic network in both Spanish and Basque. However, we found no evidence of across-language generalization. These results demonstrate that even in the absence of conscious awareness and lack of behavioural sensitivity to the words, putative semantic brain areas carry information related to the meanings of the words. The generalization of semantic representations across languages, however, may require deeper conscious semantic access.

Keywords: Machine learning; Multivoxel pattern analysis; Semantic processing; Unconscious; Visual masking.