The internal architecture of language
Only published papers are included; for preprints, see Papers.
Last updated: May 2024
Syntactic/combinatorial linguistic processing is distributed across the language network
Many claims exist that syntactic, or more general combinatorial, linguistic processing draws on a particular brain region within the language network. However, the idea of a focal syntactic ‘hub’ or ‘hubs’ is not empirically supported. Instead, robust sensitivity to syntactic structure is present throughout the language network, including its frontal and (anterior and posterior) temporal components.
This ubiquitous sensitivity to syntactic structure across the language network can be found with controlled experimental paradigms.
And during naturalistic language comprehension.
The processing of word meanings, combinatorial linguistic processing, and the processing of sublexical regularities are not spatially segregated within the language network.
A common claim about the architecture of the language system is that the processing of word meanings draws on brain areas that are distinct from brain areas that support combinatorial processing (syntactic structure building and compositional meaning construction). Across numerous studies (using both fMRI and intracranial recordings), we have shown that this is not the case.
All areas of the language network support both lexico-semantic processing and combinatorial processing, although the Shain, Kean et al. (2024, JOCN) paper suggests that the posterior temporal language area only supports syntactic but not combinatorial semantic processing.
This paper has an extensive introduction that reviews several literatures that are relevant to the relationship between lexico-semantic and syntactic/combinatorial processing and provides additional evidence against a dissociation between them.
This paper presents evidence against a dissociation between lexical access and phrase-structure building during language production.
This paper presents evidence that the language network is sensitive to linguistic structure below the word level, showing stronger responses to nonwords that are more phonotactically probable compared to those that are less probable.
The language network has a relatively short temporal receptive window
Brain regions and networks differ in the span of context over which they integrate information, or what is commonly referred to as “temporal receptive windows (TRWs)”.
These papers demonstrate that the language network is not sensitive to discourse-level structure. The Blank & Fedorenko (2020, NeuroImage) paper also shows that the TRW of the language network is somewhere between a word and a sentence, and that it is similar across the different language areas.
This paper demonstrates that as long as words within a span of a few words can be combined into complex meaning representations (even if their order does not obey the word order constraints of the language), the language network responds as strongly as it does to well-formed sentences, which suggests that local word-order-independent composition drives the activity in the language network (but stay tuned: the picture is more complicated!).
Functional heterogeneity within the core language network
In spite of broad similarities among the language areas in their functional response profiles (as discussed above), some evidence is starting to suggest differences among the areas as well as functional heterogeneity that does not map onto areal divisions.
This paper shows that the posterior temporal language area does not appear to support compositional semantics: only lexical-semantic processing and syntactic structure building (Shain*, Kean*, et al., 2024, JoCN).
And this paper uses intracranial recordings and shows that language-responsive electrodes show three distinct temporal profiles, which appear to correspond to different-sized temporal receptive windows. Importantly, these electrodes are interleaved across the language network, which makes it challenging to see these differences with methods like fMRI (Regev*, Casto*, et al., 2024, Nat Hum Beh).
The “extended language network”: Differences between the “core” left-hemisphere language network and other language-responsive areas
As discussed above, the core left-hemisphere (LH) language areas are generally characterized by highly similar functional profiles. However, one area that was originally included as part of the LH language network (e.g., in Fedorenko et al., 2010, J Neurophys)—the language area within the angular gyrus—consistently separates from the rest of the LH language areas in task-based experiments and in patterns of functional correlations during naturalistic-cognition paradigms. Also, a number of brain areas outside left frontal and temporal cortex show responses to language, including the right hemisphere (RH) homotopic areas, areas on the ventral temporal surface and along the cortical midline, areas in the cerebellum, and some subcortical areas (e.g., Lipkin et al., 2022, Nat Sci Data). We are actively investigating the organization of this “extended language network” and the differences between the “core” LH areas and some of these other areas.
The papers below provide evidence for a) strong functional integration among the core LH language areas, and b) a tri-partite organization of the bilateral language network into LH areas, RH areas, and bilateral areas in the angular gyrus (areas within each of these sub-networks are strongly integrated, and areas that straddle sub-network boundaries are much more weakly correlated).
These papers use functional correlations during rest and story comprehension.
This paper uses effect sizes in the language localizer paradigm.
This paper uses dynamic functional correlations.
