A key requirement from cognitive models of language comprehension is that they specify the distinct computational mechanisms that are engaged in language processing and the division of linguistic labor across them. Here, I address this requirement from a cognitive neuroscience perspective by employing functional MRI to study the neural implementation of comprehension processes. My experimental approach, unprecedented in studies of language, combines available methods to simultaneously achieve (i) increased functional resolution, via localization of functional brain regions at the single-participant level; (ii) ecological validity, through data-driven, model-free paradigms using naturalistic stimuli; and (iii) statistical rigor, by explicit comparison of functional profiles across regions.

Using this approach, I first contrast two cortical networks engaged in comprehension: one, the “high-level language network”, is selectively recruited by linguistic processing but not by other cognitive functions; another, the “multiple-demand network”, is recruited across diverse cognitive tasks, both linguistic and non-linguistic. I show that, during naturalistic cognition, each network shows high synchronization amongst its constituent regions, whereas regions across the two networks are functionally dissociated. Thus, these two systems likely play distinct roles in comprehension, which I then characterize by demonstrating that the language network closely tracks linguistic input whereas the multiple-demand network does not. This finding critically constrains the possible contributions of the multiple-demand system to comprehension.

Next, I focus on the high-level language network and examine two current hypotheses about its internal structure. In one study, I find that activity elicited by syntactic processing is not localized to focal language regions but is instead distributed throughout the network, suggesting that syntax is cognitively inseparable from other aspects of language. In another study, I estimate the timescales over which different language regions integrate linguistic information and find that they share a common profile of temporal integration. Therefore, the topographic division of linguistic labor across this network is not organized along distinct integration timescales.

Collectively, these results account for crucial inconsistencies in the literature and challenge common theoretical views. By characterizing the fundamental functional architecture of comprehension mechanisms, these results provide novel insights into the ontology of linguistic mechanisms that give rise to human language.