A word is better recognized when presented in a grammatical sequence of words (e.g., within a sentence) than when presented in an ungrammatical sequence of words. This sentence superiority effect was recently demonstrated in a series of behavioral experiments where participants had better word identification rates for all positions in a sentence relative to an ungrammatical sequence of the same words. A previous EEG investigation of this phenomenon pointed toward an online, automatic processing of linguistic information starting at around 270 ms post-stimulus onset. Taken together, these results support the idea of an early role of the syntactic network in facilitating the recognition of a word presented in a sentence context, leading to the question of how this network is set in motion. To answer this question, we conducted an MEG experiment and source analysis of the brain areas implicated in syntactic processing. Source activations over time showed grammatical vs. ungrammatical differences first in the inferior frontal gyrus (325-400 ms), then the anterior middle temporal gyrus (475-525 ms), and finally in both in the inferior frontal gyrus and the posterior middle temporal gyrus (550-600 ms). We interpret the spatio-temporal dynamic of the sentence superiority effect in terms of an Interactive-Activation model with bottom-up activations and top-down influences operating between word and sentence-level representations.