Verified by 
In a new article appearing in the Journal of Cognitive Neuroscience, researchers at the Massachusetts Institute of Technology, led by Cory Shain and Hope Kean, explore how the human brain shows signatures for linguistic composition. This refers to the ability we have to put together individual units of meaning into larger units, typically hierarchically organized.
The research is an excellent example of innovative neuroscience methods being paired with thoughtful and critical linguistic hypotheses to deliver novel insights on the question: “Which parts of the brain care about different core components of language?”
Shain and colleagues discovered an intricate profile of distributed sensitivity to both syntax and semantics throughout a broad frontotemporal network. This spanned from the posterior and middle lateral temporal cortex to the inferior frontal cortex.
The main take-home message here is that language network regions do not display binary sensitivity to either words, grammar, or larger meanings built from multiple words; rather, they vary in their sensitivity to at least two of these features.
They make the important conclusion that the “internal specialization” for language “is primarily a matter of degree rather than kind.” In addition, they practice something that should be standard across the field: they provide a Table with full definitions of working concepts and a Figure plotting a full schema of possible neural response profiles based on various hypothetical effects.
The authors show that major language network nodes display different levels of sensitivity to syntax, "combinatorial semantics" (referring to the types of complex meanings we can only derive from multi-word phrases, as opposed to individual words), and lexicality (referring to our ability to judge whether a string of letter forms a real word or not), rather than one fixed region being selective for a specific component.
Cortical Mosaics for Linguistic Structure
The authors found that posterior temporal and inferior frontal language sites showed a length effect in so-called "Jabberwocky" sentences (i.e., sentences in which all the "content words," like nouns, verbs, adjectives, and adverbs, have been replaced with fake words) for linguistic structural context, and all language areas showed at least some effect of lexicality (i.e., a difference for real words vs. fake words).
This is in line with recent ideas about a "cortical mosaic" architecture for linguistic structure within overlapping portions of posterior temporal and inferior frontal cortices for processing demands that bias syntactic and semantic computations, whereby, for example, effects of composition can be found within a narrow strip of tissue within the broader lexicality-sensitive cortical sites (a spatial mosaic), or where different demands of sentence-level inferential semantics can be detected over closely overlapping temporal windows within a small area of cortex (a spatiotemporal mosaic).
Shain and colleagues also found that the posterior temporal cortex uniquely exhibits equal sensitivity to syntactic structure and combinatorial semantics, with or without lexical content. This potentially points to this region as a site heavily involved in linguistic structure-building.
This provides strong evidence that the brain’s language hubs are not monolithic in their interest in specific components of linguistic structure—meaning it's not syntax over here and semantics over there, as has been claimed by some of the field’s most influential researchers such as Friederici, Duffau, and Hagoort. Instead, it points to a kind of mosaic-like architecture whereby syntax and semantics are closely intermingled, but different regions exhibit various biases for processing loads and strategies.
Someone Must Extinguish Thy Flame: Finding Signatures of Syntax
There is a qualification to be made about the hypothesis presented by many authors before Shain and colleagues, and replicated in their study, that “syntactic hubs” would be unveiled via “identical responses” for real sentences and Jabberwocky sentences. Lexico-semantic modulations inevitably tax syntactic processes, given the centrality of lexical syntax in driving and initiating phrase structure formation and, subsequently, long-distance dependency formation.
As such, there are some core syntactic processes certainly captured in both real sentences and Jabberwocky sentences, such as hierarchical depth. But even here, as Shain and colleagues mention, there are certain properties of lexical meaning that can be inferred by specific properties of pseudowords, and every phrase structure is built out of properties such as lexical selection criteria (pseudowords do not select for anything).
Even if a brain region did in fact execute the same level of local neuronal cortical processing for real sentences and Jabberwocky sentences with respect to building inferences of recursive hierarchical structure, the lexico-semantic differences would naturally interfere here, given the close regulation of semantics by syntax.
In other words, why would we expect to find a “pure” syntactic response profile when syntax is itself anything but pure? It is in the game of regulating semantic structure via a complex interface of 'morphophonological' transformations.
In addition, while one of the advantages of fMRI is its spatial resolution, the authors keep to large macroanatomical areas (i.e. the entire anterior temporal surface, the entire posterior temporal surface, etc.), which may be neurobiologically implausible in terms of making spatiotemporal claims of cortical computation and the possibility of distinct response profiles that might, nevertheless, bias much more heavily towards combinatorial semantic and/or syntactic information.
The authors also cite Fedorenko’s intracranial study (using electrodes recorded directly from the surface of the brain in epilepsy patients), and Nelson and colleague’s intracranial study, using this to motivate their analysis of nodal structure-building. They do not cite more recent intracranial work failing to replicate much of Nelson and colleagues’ major effects of structure (including discussion of why this might be the case).
The authors’ finding that “no language region appears to be a hub for abstract (i.e., content-independent) combinatorics” speaks to the idea of unbounded hierarchical structure-generation being uniquely tied to its elements of selection, as per many models of minimalist syntax, as opposed to models that claim, for instance, that the human brain has a domain-general structure-building capacity that applies over distinct representational domains (e.g., language, mathematics, music, etc).
Overall, it may not be the case, as some have claimed (e.g., Dehaene, Fitch, Bornkessel-Schlesewsky) that there are dedicated neural circuits for abstract combinatorics. Instead, human language might be built off a constellation of domain-general neural codes but with unique integration of these codes alongside domain-specific stored representations (as proposed in the recent ROSE model for the brain’s neurocomputational architecture for language, in which lexical units are intrinsically maintained in their base code throughout higher-order syntactic operations), partly accounting for the kind of “difference of degree but not kind” perspective unveiled by Shain and colleagues. Still, there may be types of neural signatures that fall outside of the poor temporal resolution afforded by fMRI that can detect more rapid shifts between distinct levels of linguistic structure.
