Figure 1.

Overview of our primary analysis. Top left, fMRI sentence reading experimental protocol. Top right, A schematic illustrating how InferSent forms sentence-level representations from context-independent word-level semantic representations (GloVe). Bottom left, Baseline models, which formed sentences by combining GloVe vectors according to grammatical rules, concatenation, or averaging. Bottom right, The ridge regression approach used to map the sentence models to predict fMRI activation. In the diagram, this is illustrated for only the BoW(GloVe) vector; however, mappings were estimated in the same way for all models. The procedure used to evaluate model-based predictions is illustrated in Figure 4.