The spatiotemporal dynamics of semantic integration in the human brain

The study addresses how the human brain integrates lexical information across words to derive sentence meaning, and whether distinct or overlapping cortical regions support semantic integration processes. Prior work shows engagement of posterior temporal, prefrontal, and parietal cortices in semantic processing, but consensus is lacking due to limited spatiotemporal resolution and confounds between lexical retrieval and coherence evaluation. The authors focus on inferential naming to definitions (as opposed to picture naming) to disentangle integrative lexical access (reference to common objects) from semantic coherence in non-referential sentences. They also probe how strongly a sentence constrains lexical search space (semantic narrowing). The central aim is to map the spatiotemporal dynamics and regional specialization of semantic integration using intracranial EEG with high temporal and spatial precision.

The paper situates its work within debates on whether language and semantic processing share substrates or are dissociable, and whether inferior frontal, inferior parietal, and posterior temporal cortices have specialized subregions for distinct semantic processes. Prior studies using ERPs and fMRI have limited capacity to resolve rapid, distributed computations underlying lexical access and coherence. Research often emphasizes picture naming over naming-to-definition, leaving inferential semantics underexplored. Evidence points to involvement of posterior temporal lobe, prefrontal, and parietal regions in semantics; some studies suggest overlap between language and semantic systems, while others argue for dissociability. MEG and iEEG work implicate pSTS in early phrase construction and IFG/IFS in later evaluative processes. The default-mode and hippocampal networks have been linked to contextual integration and memory processes relevant to semantics. The authors build on these findings to separate effects of referential success, coherence, and lexical narrowing with intracranial resolution.

Participants: 58 native English-speaking epilepsy patients (11 male; 18–41 years; 2 left-handed) with no history of language deficits underwent intracranial recordings (approved protocol HSC-MS-06-0385). Electrode implantation: Primarily stereotactic depth electrodes (sEEG; 56 patients) and subdural grids (SDE; 2 patients). A total of 13,298 contacts were implanted; 9388 included after artifact and seizure-onset exclusions. Localization via pre-op MRI/post-op CT coregistration; electrodes projected onto cortical surface models. Recording: Signals digitized at 2 kHz; re-referenced to common average; trials with artifacts/incorrect responses excluded. Stimuli and design: Orthographic sentence definitions (mean 6.5 words; range 3–12) presented via rapid serial visual presentation (RSVP) at 500 ms/word, preceded by 1000 ms fixation and followed by a 1 s blank. Patients had 2 s to respond. Two trial types: Referential (definitions that refer to a common object) and Non-Referential (no common lexical item). Within Non-Referential, sentences were either semantically coherent (weak violation: conceptually coherent but lacking a lexicalized term) or semantically incoherent (strong violation of theta roles/selectional restrictions). Example coherent non-referential: “A person at the circus who makes you commute”; incoherent: “A place where oceans shop.” For Non-Referential trials, patients responded “nonsense.” Within Referential trials, a norming study (n=80) quantified semantic narrowing (likelihood the referent could be inferred before the final word), yielding two subtypes: strong narrowing (high predictability before the final word) and limited narrowing (low predictability until the final word). Behavior: Reaction times measured from final word offset to articulation onset. Experimental blocks typically included 42 Referential and 42 Non-Referential trials (most patients completed two blocks). Signal processing: Broadband gamma activity (BGA; 70–150 Hz) extracted after line-noise removal (band-stop Butterworth filters), bandpass Hilbert transform, and smoothing (Savitzky-Golay FIR; 3rd order; 251 ms frame). BGA computed as percent change from baseline (−500 to −100 ms before first word). Significant activations assessed with one-tailed t-tests at each time-point and FDR correction (q<0.05). Low-frequency hippocampal theta power also examined. Group-level mapping: Surface-based mixed-effects multilevel analysis (SB-MEMA) to generate population maps, accounting for sparse sampling and intersubject variability, with geodesic Gaussian smoothing (3 mm FWHM). Familywise error rates determined via Monte Carlo simulations (5000 iterations); results restricted to regions with ≥3 patients’ coverage (unless noted). Conjunction SB-MEMA used to identify regions sensitive to Reference (referential vs non-referential), Coherence (coherent vs incoherent within non-referential), and Narrowing (limited vs strong within referential), regardless of effect direction. Analyses focused on the language-dominant left hemisphere, with checks for specificity versus right hemisphere. Regional parcellations followed HCP atlas definitions (e.g., MPC, vmPFC/OFC, MFG/IFS, pSTS/pMTG, hippocampus/PHG).

Behavior: Mean articulation RT was 1765 ms (SD 680 ms) after final word offset. Referential trials were faster than Non-Referential (paired t-test; t(1,83)=1.86, p=0.032). Spatiotemporal sentence processing: Population BGA maps showed increasing activation over sentence duration across a distributed network. Early activation: IFG, medial parietal cortex (MPC), anterior temporal lobe (ATL), and posterior middle temporal gyrus (pMTG). Late activation: ventromedial PFC (vmPFC), posterior cingulate, and orbitofrontal cortex (OFC). Ventral temporal cortex, inferior lateral temporo-occipital cortex, and inferior frontal sulcus (IFS) were active throughout, with clear increases at the final word. Reference (referential vs non-referential, time-locked to final word): Greater BGA for referential trials in middle frontal gyrus (MFG) and middle IFS (500–700 ms: β=0.08 (SD 0.03), p<0.001; 700–900 ms: β=0.10 (SD 0.05), p<0.001), MPC and parahippocampal cortex (500–700 ms: β=0.11 (SD 0.06), p<0.001; 700–900 ms: β=0.13 (SD 0.08), p<0.001), vmPFC (500–700 ms: β=0.07 (SD 0.01), p<0.001; 700–900 ms: β=0.08 (SD 0.04), p<0.001), and OFC (500–700 ms: β=0.09 (SD 0.04), p<0.001; 700–900 ms: β=0.11 (SD 0.05), p<0.001). Non-referential trials showed higher BGA in posterior superior temporal cortex (500–700 ms: β=0.07 (SD 0.01), p<0.001) and anterior IFG (aIFG; 500–700 ms: β=0.08 (SD 0.03), p<0.001). Coherence (within non-referential): Incoherent > coherent: medial frontal cortex (300–500 ms: β=0.10 (SD 0.04), p=0.001) and superior medial parietal cortex (300–500 ms: β=0.09 (SD 0.03), p<0.001). Coherent > incoherent: IFS (300–500 ms: β=0.08 (SD 0.03), p=0.001), aIFG (700–900 ms: β=0.11 (SD 0.05), p=0.001), angular gyrus (700–900 ms: β=0.06 (SD 0.01), p<0.001), pMTG (700–900 ms: β=0.08 (SD 0.03), p=0.002), and OFC (700–900 ms: β=0.07 (SD 0.01), p=0.001). Narrowing (within referential): Limited narrowing (harder lexical access) > strong narrowing: pSTS (onset ~250 ms; 300–500 ms: β=0.10 (SD 0.04), p=0.001; 500–700 ms: β=0.08 (SD 0.03), p<0.001), MPC (500–700 ms: β=0.12 (SD 0.07), p=0.003), IFS (300–500 ms: β=0.11 (SD 0.05), p=0.008; 500–700 ms: β=0.12 (SD 0.07), p=0.004), ATL (500–700 ms: β=0.06 (SD 0.01), p=0.004), and OFC (500–700 ms: β=0.12 (SD 0.07), p<0.001). Articulation RTs did not differ between narrowing conditions (strong: 1886±708 ms; limited: 1904±699 ms; t(1,34)=-0.09, p=0.46). Medial temporal dynamics: Early hippocampal theta increases for referential trials (100–400 ms) followed by parahippocampal BGA increases (300–1000 ms), consistent with memory-related contributions to inferential semantics. Overall: IFS was sensitive to all semantic contrasts with early effects; OFC showed later sensitivity across processes. pSTS/pMTG contributed to reference, coherence, and narrowing with a temporal progression from posterior temporal to frontal regions.

Findings reveal a mosaic-like cortical organization supporting semantic integration, with distinct yet partially overlapping contributions across inferior frontal, posterior temporal, and medial networks. IFS shows broad sensitivity to reference, coherence, and narrowing, with early engagement and later subregional differentiation, suggesting a lexico-semantic hub for control and integration. OFC, engaged later, also responds to all semantic processes, indicating roles in semantic saliency and higher-order integration. Posterior temporal regions (pSTS/pMTG) contribute to integrative lexical access and coherence, with early sensitivity to difficult reference (limited narrowing) and later coherence effects, aligning with MEG evidence for N400 dynamics propagating from STS to frontal cortex. Medial parietal cortex, vmPFC, and medial temporal lobe structures (hippocampus/PHG) are engaged during successful inferential reference, consistent with default/memory network involvement in situation model construction, context integration, and representational search. The temporal cascade suggests early composition and lexical search in posterior temporal sites, followed by frontal semantic control/evaluation and engagement of medial networks for referential inference and memory integration. Results refine accounts of semantic cognition, supporting a controlled semantic cognition framework wherein representational and control components are distributed across temporal, frontal, and medial parietal/temporal structures.

Using extensive intracranial recordings during orthographic sentence comprehension, the study delineates complementary cortical mosaics for semantic integration across posterior temporal and inferior frontal cortices. IFS and OFC emerged as hubs responsive to integrative lexical access (reference), semantic coherence, and lexical search difficulty (narrowing), with distinct temporal profiles. Posterior temporal regions and medial networks (MPC, vmPFC, hippocampus/PHG) contribute differentially over time to constructing referents and coherent meanings. The work advances a spatiotemporally resolved model of semantic integration and suggests that semantic processes are interwoven across multiple cortical territories rather than strictly localized. Future research should further disambiguate lexical access from other semantic processes at finer temporal scales, examine right-hemisphere contributions with improved coverage, and test whether similar mosaic organization applies to syntactic processing and language disorders.

- Electrode coverage was driven by clinical needs, yielding sparse and uneven sampling (notably reduced right-hemisphere coverage), limiting inferences about hemispheric asymmetries. - The narrowing manipulation did not strictly isolate lexical access; limited narrowing sentences still afforded varying degrees of facilitation. - Word-by-word dynamics of reference resolution were not mapped; analyses focused on confident final-word windows. - Assumption that patients constrained lexical search mid-sentence could not be ensured; narrowing positions varied across sentences. - No assessment of vividness ratings between referential and non-referential trials, a potential confound for future study. - Group SB-MEMA maps require sufficient patient coverage and apply thresholds that may miss smaller effects; directionality in conjunction maps not encoded. - Data are not publicly shareable due to HIPAA constraints (available on request), limiting external reanalysis.