Neural patterns in linguistic cortices discriminate the content of verbal working memory

Verbal working memory (WM) is characterized by the presence of psycholinguistic effects, whereby items associated with richer linguistic representations are usually better recalled, such as words vs. nonwords (lexicality effect). This effect is accounted for by language-based models, assuming a direct and obligatory involvement of lexical linguistic knowledge at all stages of WM processing or by redintegration models considering that lexical linguistic knowledge only intervenes during post-memory trace reconstructive processes. We contrasted these two accounts in functional neuroimaging experiment by assessing to what extent and at what WM stage word and nonword memoranda can be distinguished based on their multivariate neural patterns in linguistic cortices. fMRI scans were obtained from 28 healthy young adult participants. The participants were invited to encode lists composed of word or nonword items presented at a very fast rate (2 items/s) and to maintain the items over a 6-second delay or not, followed by a probe recognition phase. Multivariate voxel pattern analyses successfully decoded word and nonword stimuli during the encoding phase in all conditions, as well as during the maintenance phase but only during the active maintenance condition. This study supports language-based WM models assuming continuous support of linguistic knowledge during all WM stages.