Neurobiology of Language

Here, we test the hypothesis that early sustained exposure to complex bilingual environments can positively affect reading development by altering structural interhemispheric connectivity via the corpus callosum (CC). Interhemispheric connectivity has been shown to be inefficient in dyslexia, but also to support compensatory pathways when genetic risk for reading difficulties is present, by enabling the preserved right hemisphere to support a dysfunctional left hemisphere. Mediation models were conducted on children aged 9-10 years (with a 2-year follow-up assessment) from the Adolescent Brain Cognitive Development database (N>10,000). Polygenic scores (PGS) for dyslexia and cognitive performance and continuous bilingualism indices were used as predictors, with reading aloud as the outcome. Bilingualism showed a positive effect on reading partially mediated by the anterior CC, independently of overall brain size. In contrast, genetic predispositions to reading difficulties influenced reading primarily through overall brain size rather than CC connectivity specifically. These two pathways were independent, suggesting that bilingual experience and genetic risk operate through distinct neuroanatomical mechanisms. These findings suggest that recurrent early exposure to complex bilingual environments may shape the brains structural connectivity toward a more balanced and integrated bilateral frontal organisation. The results highlight potential brain compensatory pathways induced by environmental experiences that may support more efficient reading development and mitigate risks for developmental dyslexia.

Two decades of research have provided evidence for gray matter volume (GMV) differences in developmental dyslexia (or reading disability, RD) in the left perisylvian cortex. However, there are concerns about result inconsistencies, likely attributable to small sample sizes, lenient statistical thresholds, and insufficient accounting for demographic variables and global GMV (Ramus et al., 2018). To address these concerns, we conducted a Discovery and Replication Study (N=262) using data from the Adolescent Brain Cognitive Development Study. We found GMV differences between the RD and Control Groups did not replicate across the Discovery and Replication Studies using voxel-based morphometry (VBM) in Statistical Parametric Mapping (SPM), and that a more conservative threshold yielded far fewer results. We then conducted Reproducibility Studies and first found that when using surface-based morphometry in FreeSurfer instead of VBM, the Discovery and the Replication Study results again failed to converge. Second, we combined all groups in a factorial VBM/SPM analysis and the interaction analysis provided quantitative confirmation for diverging between-group difference results across the two studies. Third, we tested for the role of covariates of no interest and found that when total GMV is not controlled for, this divergence dissipates and group differences in RD (main effect of Reading Ability) are amplified. In conclusion, replication of GMV differences in RD is low, even when using large, well-matched groups, and analyses approaches play a modulating role. As such, results from prior studies using lenient statistical thresholds and not accounting for total GMV should therefore be viewed with caution.

Although many individuals with chronic aphasia respond to language therapy, there remains a need for adjunctive interventions that can enhance treatment response. Approaches targeting multiple modalities, such as gesture cueing, and neuromodulation techniques, such as transcranial magnetic stimulation, have shown promise for supporting language recovery. The present pilot study investigated whether enhancing activation of the action semantic network could facilitate verb production in individuals with chronic aphasia. Participants were recruited as a convenience sample and completed a within-subject design in which all individuals received each condition. Two non-linguistic methods of activating the action semantic network were evaluated: (1) pantomimed gesture cues to prime action concepts and (2) intermittent theta-burst stimulation to the left posterior middle temporal gyrus (pMTG), an intact action-semantic network node in our participants. We examined individual and combined effects of gesture priming and stimulation to test whether a combined approach would yield additive or interactive benefits. Using a Bayesian generalized linear mixed-effects model, we observed a moderate interaction between gesture priming and stimulation site. Contrary to predictions, combining gesture priming with pMTG stimulation did not produce additional benefits over either intervention alone. Instead, pMTG stimulation attenuated the priming advantage observed under vertex stimulation, and gesture priming attenuated the advantage observed with pMTG stimulation alone. Posterior estimates provided substantial preliminary evidence for this interaction in our pilot sample size. These findings suggest that combined activation of the action semantic network through gesture and neuromodulation approaches may not benefit verb retrieval above and beyond each approach alone.

Children with developmental language disorder (DLD) have persistent language learning difficulties and often perform poorly on pseudoword repetition, a task that probes phonological, memory, and speech-motor processes that support vocabulary acquisition. Research on the neural basis of pseudoword repetition in DLD is limited. We used whole-brain functional MRI (fMRI) to examine pseudoword repetition and repetition-based learning in 46 children with DLD (ages 10-15 years) and 71 age-matched children with typical language development. During scanning, children heard and repeated pseudowords paired with visual referents, allowing us to track learning-related changes in neural activity across repetitions. Repeated pseudoword production yielded comparable behavioural learning across groups, with faster productions by later repetitions. Post-scan, form-referent recognition was comparable across groups, whereas pseudoword repetition accuracy was lower in DLD. Pseudoword repetition engaged a distributed neural network, including inferior frontal cortex bilaterally (greater on the left), premotor and sensorimotor cortex, and posterior temporal and occipital regions. Group differences emerged primarily in regions where activity was task negative (i.e., below baseline or deactivated): lateral occipito-parietal cortex (posterior angular gyrus), medial parieto-occipital cortex (retrosplenial), and right posterior cingulate cortex. Learning-related decreases in activity were similar across groups, but region-of-interest analyses showed reduced leftward lateralisation of activity in inferior frontal gyrus in DLD. These findings suggest weaker disengagement of the default mode network during a linguistically demanding task in DLD. Although repetition-based pseudoword learning recruited similar neural mechanisms in both groups, these mechanisms may operate less efficiently in DLD, alongside reduced hemispheric specialisation in inferior frontal cortex. HighlightsO_LISimilar repetition-related neural attenuation across groups during pseudoword learning. C_LIO_LIReduced default-mode network suppression during pseudoword repetition in DLD. C_LIO_LIReduced left-hemisphere specialisation of inferior frontal cortex in DLD. C_LIO_LIRepetition-based learning in DLD supported by less efficient neural networks. C_LI

The present study examined whether thematic reversal anomalies are processed similarly across subject and object experiencer constructions in Malayalam. Event-related brain potentials (ERPs) were recorded as 30 first-language speakers of Malayalam read transitive sentences with the two types of experiencer verbs, in which the thematic role assignment for the preceding arguments was either correct or reverse. The reversal anomaly became apparent only at the position of the experiencer verb. A linear mixed-models analysis confirmed a biphasic N400-P600 effect at the verb for both verb types when the argument roles were reverse. Thus, our results suggest a uniform processing strategy for TRAs irrespective of the type of experiencer verb involved. However, the N400 amplitude was larger for the object experiencer verb compared to subject experiencer verbs. We suggest that the quantitative difference observed for object experiencer verbs is due to the inverse linking of grammatical function and thematic roles associated with these verbs. In other words, verb-specific linking properties modulate the processing of TRAs involving object experiencer verbs. We argue that this modulation occurs because the parser recalibrates cue weighting when the expected form-to-meaning mappings are overridden by the inverse linking properties of object experiencer verbs.

Humans comprehend language incrementally, updating the representation of sentence meaning with each incoming word. These updates are guided by the distance between each perceived word and prior expectations--the prediction error. The alignment between large language models (LLMs) and cortical activity inspires the hypothesis that the cortical computation of prediction error is Surface-based, driven by statistical patterns of word form co-occurrence. In contrast, psycholinguistic models propose that prediction error computation is Meaning-based, driven by word semantics. We used polysemic words with ambiguous semantics to distinguish these models: ambiguity would introduce uncertainty into meaning representations and hence the prediction error, if Meaning-based, but would not affect the prediction error, if Surface-based. We examined how ambiguity influenced prediction error signatures in self-paced reading times and magnetoencephalographic (MEG) neural responses during sentence processing. While an LLM-based proxy of prediction error robustly predicted reading times and neural responses to unambiguous words, it failed to predict either under ambiguity. That is, prediction error computation was altered by uncertainty in word meaning, which supports the Meaning-based model and corroborates the essential role of word meaning in predictive language processing. Our findings highlight an important limitation of LLMs as in silico models of the human language faculty.

PurposeTo examine speech iconicity for shape in aphasia, we compared iconicity ratings from people with aphasia to those from neurologically intact individuals and evaluated how iconicity relates to phonological and semantic processing profiles in aphasia. MethodEleven people with aphasia and 11 age- and gender-matched neurologically intact participants rated how rounded or pointed 50 auditory pseudowords sounded using a 5-point scale. Ratings from participants with aphasia were compared to predicted iconicity ratings derived from reference ratings from prior work and to ratings from neurologically intact participants. For each participant with aphasia, correlations between individual ratings and predicted ratings were related to measures of phonological and semantic processing. ResultsRatings from people with aphasia were significantly correlated with both the predicted ratings and the ratings from neurologically intact participants. The strength of the correlation between individual ratings and predicted ratings did not differ significantly between groups, although there was a trend toward weaker correlations in the aphasia group. There were indications that greater language impairment was associated with greater disruption of iconicity ratings; in particular, deficits in phonological segmentation and semantic processing were associated with reduced sensitivity to shape iconicity. ConclusionThese findings suggest that sensitivity to shape iconicity is preserved in individuals with aphasia to varying degrees. The specific nature of language impairment appears to play an important role in determining iconicity processing in aphasia.

People frequently gesture while speaking, even when listeners cannot see them--for instance, during phone calls or behind barriers. Congenitally blind individuals also gesture, indicating that gestures serve functions beyond visual communication. Previous models of gesture production (e.g., Kita & Ozyurek, 2003; Rauscher et al., 1996) suggest that gestures facilitate speech, but they rely heavily on behavioural data and provide limited insight into temporal dynamics. This study used magnetoencephalography (MEG), a neuroimaging technique with high temporal resolution, to investigate when gestures influence speech. Twenty-three native Japanese speakers took part in a storytelling task under two conditions: Gesture-Required (gesture use instructed) and Gesture-Prohibited (hands kept still). Participants described cartoon clips across multiple sessions (30 trials x 3 sessions per condition). Using speech onset as the reference point, we compared root mean square (RMS) values within a -0.25 to 0 second window. RMS values were higher in the Gesture-Prohibited condition, with increased activity in the bilateral anterior temporal lobes (Left ATL: p = .049; Right ATL: p = .027), but not in motor regions (p = .29). These findings suggest that gestures reduce neural load in language-related regions before articulation. Co-speech gestures may support speech planning by facilitating lexical retrieval or semantic structuring. The lack of motor region effects indicates that this influence is linguistic rather than motoric. This study provides direct direct neurophysiological evidence of the timing of gesture-speech interaction, supporting models that view gestures as an integral part of speech production.

Impaired reading, i.e., alexia, is common after left hemisphere stroke. The most common deficit in alexia is a difficulty reading aloud pronounceable novel words, also called pseudowords. While semantic and phonological processes both subserve reading real words, pseudoword reading deficits in alexia are typically ascribed to phonological deficits alone. Some theories, however, suggest that pseudoword reading relies in part on lexical-semantic knowledge, such that semantic deficits could also contribute to poor pseudoword reading in alexia. Leveraging a large sample of left-hemisphere stroke survivors, we examine the cognitive and neural substrates of pseudoword reading accuracy and two error types: lexicalization errors, where a pseudoword is incorrectly read as a real word, and nonword errors, where a pseudoword is read as an incorrect nonword. 76 left-hemisphere stroke survivors read 60 pseudowords aloud, and performed two pseudoword repetition tasks to assess phonological processing and two picture naming tasks to assess mappings between lexical semantics and phonology. Regression models assessed how pseudoword repetition and naming related to overall accuracy and rates of lexicalization and nonword errors in pseudoword reading. Voxel-based and connectome lesion-symptom mapping localized the neural territory responsible for these errors. Both pseudoword repetition and naming independently related to pseudoword reading accuracy. Pseudoword repetition but not naming deficits predicted higher rates of lexicalization errors, while naming but not pseudoword repetition deficits predicted higher rates of nonword errors. Greater nonword error rate also predicted smaller imageability effects in real word reading (t(71)=-3.2, p=0.002). Lexicalization errors were associated with lesions to and disconnections of the left putamen and basal ganglia. Nonword errors were associated with lesions to the superior and middle temporal gyri, as well as broad temporo-parietal disconnections, overlapping with previous lesion-mapping results implicating these regions in semantic contributions to word reading. These results suggest that lexicalization errors result from impaired planning and execution of novel motor plans, causing a reliance on the well-learned motor plans associated with lexical items. In contrast, greater rates of nonword errors, relative to lexicalization errors, occur when semantic contributions to reading are impaired. Overall, these findings demonstrate that semantic processes are involved in reading pseudowords, at least in stroke alexia. These findings support connectionist accounts of reading in which damage in the direct orthography to phonology route for reading leads to reliance on semantic representations, even for pseudowords, suggesting a reinterpretation of pseudoword reading as a pure measure of phonological reading deficits.

Language impairments in aphasia are characterized by various representational disruptions that may be reflected in discourse production. This research examines the capacity of transformer-based language models, particularly GPT-2, to serve as a computational framework for analyzing variations in aphasic narrative speech. A longitudinal dataset of narrative speech samples collected at six time points from individuals with aphasia (N = 47) was utilized as part of an intervention study. All transcripts were processed via the GPT-2 language model to obtain activation values from each of the 12 transformer layers. Statistically significant differences in activation magnitude across aphasia subtypes were found at every layer (all p < .001), with the most pronounced effects in the deeper layers. Pairwise Tukey HSD tests revealed consistent distinctions between Brocas aphasia and both Anomic and Wernickes aphasia, suggesting a shared activation profile between the latter two. Longitudinal tests revealed significant changes over time, especially in the final three layers (10-12). These findings suggest that transformer-based activation patterns reflect meaningful variation in aphasic discourse and could complement current diagnostic tools. Overall, GPT-2 provides a scalable tool to model representational dynamics in aphasia and enhance the clinical interpretability of deep language models.

Prior studies have reported inconsistent results for neuroanatomical differences between early bilinguals and monolinguals. These studies primarily measured gray matter volume (GMV), involved small samples, and prioritized adults. Few studies of early bilinguals have measured cortical thickness (CT), which offers more anatomical specificity. It remains unclear whether results derived from differing metrics and approaches (e.g., vertex-versus parcel-wise analyses) converge. Using data from the Adolescent Brain Cognitive DevelopmentSM (ABCD) Study, we compared neuroanatomy between large groups of early cultural Spanish-English bilingual and English monolingual children (N = 1,209) matched on age, pubertal status, sex, handedness, socioeconomic status (SES), and nonverbal reasoning. Whole-brain voxel-based morphometry revealed areas of greater and of lesser GMV in bilinguals than monolinguals across all lobes. Vertex-wise CT analyses similarly identified widespread differences, with bilinguals showing areas of both thicker and thinner cortex. We contextualized these findings with parcel-wise CT analyses (average CT values), utilizing two atlases of differing spatial granularity. Parcel-wise results showed good correspondence with vertex-wise findings when implementing the more fine-grained atlas (Destrieux), but use of the coarser atlas (Desikan-Killiany) provided results that led to different conclusions. Finally, we tested for interaction effects between bilingualism and SES on CT and found several regions where differences between bilinguals and monolinguals in CT were modulated by SES. Together, these findings indicate that early bilingualism is associated with extensive neuroanatomical differences relative to monolinguals during childhood, and that these results can vary as a function of neuroanatomical metric, analysis approach, atlas granularity, and SES. Research HighlightsEarly Spanish-English bilingual and monolingual children differ in gray matter volume and cortical thickness across multiple brain regions. Cortical thickness differences between bilinguals and monolinguals cannot be firmly attributed to adaptations associated with language or executive control. Socioeconomic status modulates cortical differences between early bilinguals and monolinguals, revealing unique thickness patterns for those with lower versus higher SES backgrounds. Parcel-wise between-group cortical thickness results are affected by atlas choice and can influence the interpretation of the findings.

Verbal fluency is a behavioral task that requires the generation of words from a semantic category (category fluency) or words beginning with a specific letter (letter fluency). Although word production engages a frontal-temporal-parietal network, no studies have tested how lesions to temporal and parietal lobe areas that represent semantic and phonological knowledge dampen neural responses in the left pars triangularis and the left pars opercularis, two adjacent regions in the left inferior frontal gyrus implicated in word search and retrieval. Here, 52 patients with temporal lobe lesions underwent clinical functional MRI while performing the category and letter fluency tasks. We investigated where lesion presence was inversely related to the magnitude of task-specific neural responses in pars triangularis and pars opercularis using a technique referred to as voxel-based lesion activity mapping (VLAM). We found that lesions to the left anterior superior temporal gyrus, left temporal pole, left hippocampus, left insula, and underlying inferior fronto-occipital fasciculus were associated with reduced neural responses in the left pars triangularis during the category fluency task. Lesion damage to the right hippocampus was associated with reduced neural responses in the left pars opercularis during category fluency. By contrast, lesions to the left posterior superior temporal gyrus, left supramarginal gyrus, left parietal operculum, and the inferior fronto-occipital fasciculus and left arcuate fasciculus were associated with reduced neural responses in the left pars triangularis and the left pars opercularis during the letter fluency task. These results suggest that anatomically dissociable brain networks interact with the left inferior frontal gyrus when different search strategies constrain the retrieval of word representations.

Developmental stuttering is a complex neurodevelopmental disorder characterized by disfluent speech. At the individual level, the behavioral manifestations of stuttering vary considerably, likely reflecting heterogeneity in underlying neural mechanisms. In this study, we examined individual-specific differences in the brains of children who stutter (CWS), by implementing normative modeling, a framework that quantifies how an individual deviates from an age- and sex-matched reference population. We applied this approach to identify individual-specific structural brain atypicalities using gray and white matter volumes. These volumes were derived from MRI scans from a large mixed-longitudinal dataset of 235 and 240 scans from CWS and fluent controls respectively, aged between 3 and 12 years. Individual deviation maps capturing these atypicalities were then used to cluster CWS into subtypes based on similarities in their neuroanatomical profiles. This analysis identified four neural subtypes with distinct neuroanatomical atypicalities relative to fluent controls. The key findings were a basal ganglia-thalamo-cerebellar subtype associated with higher stuttering severity and lower rates of recovery, and a white matter subtype characterized by mild severity and a higher likelihood of recovery. The remaining two subtypes showed cerebellar differences alongside alterations in brain regions involved in sensorimotor integration. Moreover, cerebellar volume atypicalities were present in all four subtypes, indicating that cerebellar alterations were present across otherwise distinct neural profiles and may represent a shared neuroanatomical feature of stuttering. These findings indicate that examining individual-specific neural differences and subtyping based on patterns of neural atypicalities provides valuable insight into the heterogeneity of developmental stuttering and represents a promising direction for improving our understanding of the disorder.

Neural tracking of slow temporal modulations in speech supports extraction of prosodic and syllabic structure critical for speech comprehension, yet whether these automatic cortical tracking mechanisms are altered in children with developmental language disorder (DLD) remains unclear. We recorded MEG while children with and without DLD listened to a story, and quantified source-level lagged speech-brain coherence and frequency-specific cortical functional connectivity across bilateral cortical regions. Children with DLD showed significantly reduced coherence in the 0.9-2.5 Hz range associated with prosodic information in the story, spanning bilateral auditory and speech-related cortex. In the 2.5-5 Hz range, linked to syllabic-rate modulations in the story, group differences were right-lateralised. No reliable differences were observed at higher modulation rates (5-9 Hz or 12-40 Hz). These coherence reductions were accompanied by altered functional connectivity between cortical regions across all frequency bands, indicating disrupted large-scale coordination within speech-processing networks in DLD.

Cortical tracking of acoustic features is essential for the neural processing of continuous stimuli such as speech and music. For example, it has been shown that children with dyslexia show atypical cortical tracking. This tracking may therefore reflect a fundamental auditory temporal processing mechanism supporting literacy more generally. In the current pre-registered study, we tested the hypothesis that cortical tracking of speech and music predicts reading ability in healthy young adults (N = 32), evaluated through a lexical decision task. Participants first completed an online session in which they performed a lexical decision task to assess their reading skills. This was followed by an electroencephalography (EEG) session, in which participants listened to a naturalistic short story and a music track. Using mutual information, we showed that neural activity aligned to both speech and music across a wide range of frequencies. Interestingly, cortical tracking was stronger for speech at very low frequencies, while it was stronger for music at higher frequencies. Critically, cortical tracking predicted reaction times in the lexical decision task in a frequency-dependent manner: stronger delta-band tracking (~1-3 Hz) for both speech and music was associated with faster reaction times, whereas stronger alpha-band tracking (~12 Hz) for speech was associated with slower reaction times. These findings remained significant even when controlling for stimulus type, age, musical experience and reading enjoyment. These results suggest that cortical tracking of speech and music reflect a domain-general temporal processing mechanism that is associated with reading ability beyond stimulus-specific features, and beyond development. These findings advance the neurobiological underpinnings of literacy and could potentially be leveraged for developing new reading interventions.

Human sentence comprehension proceeds word-by-word, with prior research proposing two central sources of cognitive demand during incremental processing: forward-looking disambiguation of the incoming information stream, and backward-looking retrieval of information associated with previous words from working memory. Recent work has shown that Transformer-based language models successfully generate predictions about sentence processing load in human psycho- and neurolinguistic data by operationalizing disambiguation cost as next-token surprisal, and memory retrieval cost as normalized attention entropy (NAE). Such models, however, remain difficult to interpret as it is not well understood what factors play causally into the decision to assign a cost value to a given word in such artificial neural networks. Here, we present interpretable and cognitively grounded models of disambiguation and memory retrieval and evaluate their neural alignment and spatio-temporal correlates using human magnetoencephalography responses to naturalistic narrative speech. Multivariate temporal response function modeling demonstrates firstly that these human-bias-informed models fare equally well in accounting for observed human language processing data as their Transformer counterparts. This same modeling framework then suggests that surprisal and NAE temporally dissociate in the cortical language network -- surprisal being predictive of bilateral superior temporal gyrus and supramarginal gyrus activation [~]300-500 ms, and NAE being predictive of activity in the same regions, but later [~]750-850 ms. By demonstrating that interpretable neurocomputational models can achieve meaningful brain alignment while maintaining explanatory transparency, this work offers a methodological blueprint for bridging the gap between algorithmic theory and neural implementation.

Iconicity refers to systematic links between word form and meaning. Although evidence for iconicity in natural language continues to grow, its neural basis remains unclear. Using functional magnetic resonance imaging (fMRI) and multivariate pattern analysis (MVPA), we examined iconic shape associations of auditory real words and pseudowords. The pseudowords were matched to the real words in phonemic and phonotactic properties, while differing primarily in the absence of learned semantic representations. Participants listened to each item and judged whether it sounded rounded or pointed. Searchlight MVPA revealed significant decoding for both stimulus types. For real words, iconic shape associations were decoded above chance in regions associated with visual and haptic shape processing (left lateral occipital complex and left anterior intraparietal sulcus), visual imagery (bilateral precuneus), phonological processing (bilateral supramarginal gyri), and semantic processing (left middle frontal and right superior frontal gyri). For pseudowords, significant decoding was found in regions associated with multisensory feature organization (right posterior intraparietal sulcus) and language processing (left angular and inferior frontal gyri). Together, these findings provide evidence for neural mechanisms mediating iconic associations, with language-related areas involved for both real words and pseudowords, and visual processing for real words.

BackgroundAlzheimers disease (AD) causes progressive decline in language and cognition. Automated speech analysis has emerged as a promising screening tool, yet clinical data scarcity limits progress. To address this, we generated a large-scale simulated speech dataset to model linguistic and acoustic deterioration across cognitive stages, Control, Mild Cognitive Impairment (MCI), and AD. MethodsUsing Monte Carlo simulations, we emulated the Pitt DementiaBank "Cookie Theft" narratives. Acoustic features (speech rate, pause duration, jitter, shimmer) and linguistic features (type-token ratio, unique-word count, filler usage) were synthetically sampled from real-world DementiaBank distributions. We trained an XGBoost classifier to distinguish diagnostic groups, and applied SHAP (Shapley Additive exPlanations) to assess feature importance. ResultsThe model achieved high discriminative performance (AUC {approx} 0.94; accuracy {approx} 85%). Compared to controls, simulated MCI and AD groups showed progressive declines in fluency and lexical diversity, and increases in disfluencies and voice instability. SHAP analysis revealed that key predictors included reduced type-token ratio, higher pause and filler rates, and elevated jitter/shimmer. Classification was most accurate for Control vs. AD; MCI misclassifications highlighted intermediate profiles. InterpretationOur framework, FMN (Forget Me Not), captures clinically relevant speech changes using simulated data, offering an explainable and scalable approach for cognitive screening. While not a substitute for real datasets, FMN validates a pipeline that mirrors known AD markers and can guide future real-world deployments. External validation remains a key next step for translational impact.

The internal representations of large language models (LLMs) correlate, or "align", with human neural activity during language comprehension. One view holds that this alignment reflects shared sensitivity to statistical patterns in LLMs and humans, while others hold that it reflects, at least in part, the emergence of shared linguistic representations in these systems. Here, we investigate whether hierarchical linguistic composition, a property believed to be fundamental to human language, modulates LLM-brain alignment. To this end, we manipulated syntax, compositional semantics, and associative semantics in English sentences that were presented to both an LLM and human participants during an electroencephalography (EEG) experiment. We matched linguistically manipulated stimuli in predictability, which allows us to tease apart alignment induced by linguistic structure from statistical factors. By comparing LLM-EEG alignment scores that were derived using a linear encoding model across predictability-matched conditions, we evaluate how linguistic manipulations modulate the alignment between human EEG reading data and contextual embeddings extracted word-by-word from the hidden layers of GPT2-XL. Three key patterns emerge: (1) increased alignment for word sequences with syntactic structure, (2) decreased alignment for sentences with compositional semantics, and (3) associative semantics does not modulate alignment. These observed linguistic modulations of LLM-EEG alignment take place above and beyond predictability. Our results indicate that associative semantics is encoded similarly by LLMs and the brain, as are at least some aspects of syntactic structure, while compositional semantics is more uniquely encoded in the human brain.

Computational models are a linchpin in our understanding of the neurocognitive basis of reading. These models can simulate idealized profiles of alexia syndromes, but in reality, individuals with alexia present with a wide range of mixed deficits rather than idealized syndromes. To provide a complete cognitive theory of reading, computational models must be able to account for this individual variation. However, this has never been demonstrated. We test oral reading and non-reading phonological and semantic processing in 83 left-hemisphere stroke survivors. We show that individual alexia profiles can be simulated by applying graded phonology and semantic lesions to an artificial neural network model of reading, creating "matched models" that represent individual stroke survivors. The severity of damage to the semantic and phonological layers of the matched models was highly correlated with directly-measured semantic and phonological processing deficits. However, we also identify systematic ways in which the models fail to simulate the reading performance of their matched stroke survivors. Our results support theories of alexia that rely on process-based deficits, demonstrate the feasibility of large-scale individualized modelling of alexia, and suggest ways to further improve the correspondence of models and human reading behavior.