Convergent representations of linguistic constructions in human and artificial neural systems

Understanding how the brain processes linguistic constructions is a central challenge in cognitive neuroscience and linguistics. Recent computational studies have shown that artificial neural language models spontaneously develop internally differentiated representations of Argument Structure Constructions (ASCs), generating specific predictions about when and how construction-level information should emerge during language processing. The present study tests these predictions in human neural activity using electroencephalography (EEG). Data were recorded from 12 native English speakers, with 11 participants retained for all reported analyses after preprocessing. Construction-related activity was analyzed using time-frequency analysis, statistical feature extraction, and machine learning classification. The results reveal construction-specific neural signatures that emerge primarily at sentence-final positions, where argument structure information becomes fully disambiguated, and are most prominent in the alpha frequency band. Pairwise classification further demonstrated reliable differentiation between specific construction pairs, particularly ditransitive and resultative constructions, while other pairs exhibited overlapping neural signatures. Importantly, the temporal emergence and similarity structure of these neural effects closely mirror patterns previously observed in recurrent and transformer-based language models, in which constructional representations arise during integrative stages of processing. Together, these findings provide converging evidence that linguistic constructions are neurally encoded as distinct form-meaning mappings, supporting core assumptions of Construction Grammar and suggesting that both biological and artificial learning systems converge on similar representational solutions during predictive language processing. More broadly, this convergence is consistent with the idea that learning systems may discover stable regions within an underlying representational landscape - recently referred to as a Platonic representational space - that constrains which linguistic abstractions emerge as cognitively and computationally efficient forms.

Understanding how the brain processes and represents language is a fundamental challenge in cognitive neuroscience (Pulvermüller, 2002). This paper adopts a usage-based constructionist approach, which views language as a system of form-meaning pairs (constructions) that link patterns to specific communicative functions (Goldberg, 2009, 2003). Argument Structure Constructions (ASCs), such as transitive, ditransitive, caused-motion, and resultative constructions, are significant for language comprehension and production (Goldberg, 1995, 2006, 2019). These constructions are key to syntactic theory and essential for constructing meaning in sentences. Exploring the neural and computational mechanisms underlying the processing of these constructions can yield significant insights into language and cognition (Pulvermüller, 2012, 2021; Henningsen and Pulvermüller, 2022; Pulvermüller, 2023).

In neurolinguistics, growing evidence suggests that both syntactic and semantic combinatorial processes recruit a left-lateralized fronto-temporal network, including inferior frontal and superior temporal regions (Fedorenko et al., 2011). Event-related potential (ERP) and time-frequency studies further indicate that different aspects of syntactic integration are reflected in neural oscillations, particularly in the alpha and beta bands (Tanner and Van Hell, 2014). However, despite extensive work on morphosyntax, agreement, and phrase structure, surprisingly little is known about the neural representation of constructions as theorized in Construction Grammar. Unlike traditional syntactic theories, CxG posits that constructions themselves—not only lexemes—are psychologically real units of processing. Direct neural evidence for this claim remains scarce.

Recent advances in computational linguistics have begun to shed light on this issue. Neural network language models, including both recurrent architectures and transformer-based models, have been shown to encode abstract syntactic and constructional information in their internal representations (Surendra et al., 2023; Krauss et al., 2025), and to be able to predict language processing in the brain (Schrimpf et al., 2020). Recent work has demonstrated that neural language models can encode structured information related to argument structure constructions (ASCs), including the emergence of clustered representations under controlled conditions (Ramezani et al., 2025a,b). However, a growing body of research suggests that such representations are often partial and exhibit important limitations. In particular, large language models tend to capture syntactic regularities more robustly than the full semantic properties associated with constructions, showing reduced generalization in tasks requiring deeper event-structural understanding (Scivetti et al., 2025; Bonial and Tayyar Madabushi, 2025; Weissweiler et al., 2022).

Accordingly, constructional representations in artificial systems are best characterized as structured but incomplete, reflecting emergent abstractions shaped by statistical learning rather than fully developed form-meaning pairings in the sense of Construction Grammar. This more nuanced view provides an important context for the present study: rather than assuming full equivalence between artificial and human representations, we test whether human neural activity exhibits convergent patterns that may reflect similar underlying computational pressures, potentially extending beyond the limitations observed in current language models.

What remains unclear is whether human neural activity mirrors these computational representations. If both artificial and biological neural systems converge on similarly differentiated representations of ASCs, this would provide strong evidence for the cognitive reality of constructions and for shared algorithmic principles underlying language processing in brains and machines.

The present study addresses this gap by investigating whether human EEG signals differentiate between four ASCs—transitive, ditransitive, caused-motion, and resultative—during naturalistic auditory sentence comprehension. Using time-frequency analysis, statistical feature extraction, and machine learning classification, we test whether construction-specific neural signatures can be detected and whether their discriminability parallels the representational structure observed in language models. By directly comparing human electrophysiology with computational models, this work aims to bridge neurolinguistics and artificial intelligence, providing converging evidence for the neural reality of constructions and offering a novel perspective on how form–meaning pairings are processed in the brain.

The present study adopts an emerging methodological perspective in which computational models are used as tools for generating testable hypotheses about human cognition. In this framework, artificial neural networks are not treated as direct models of the brain, but as formalized systems that implement specific learning and processing principles. These principles can be analyzed to derive predictions about when and how particular representations should emerge during human language processing. Recent work has begun to formalize this approach, using language models to generate experimentally testable hypotheses about linguistic generalization and structure (Misra and Kim, 2024).

From this perspective, the goal is not to demonstrate that human neural processing mirrors artificial systems, but to test whether similar representational patterns arise under shared computational constraints. Convergence between model-derived predictions and neural data is therefore interpreted as evidence that both biological and artificial systems may arrive at comparable representational solutions when solving the same computational problem—in this case, the integration of argument structure during language comprehension.

Recent computational work has demonstrated that artificial neural language models develop internally differentiated representations of Argument Structure Constructions (ASCs) during predictive learning, even in the absence of explicit syntactic supervision (Ramezani et al., 2025a,b). These findings generate specific, testable predictions about when and how construction-level information should emerge during incremental language processing. In particular, they suggest that constructional distinctions arise at later stages of processing, once sufficient argument structure information has been accumulated, and that these representations exhibit graded similarity relations reflecting shared event-structural properties. The present study tests these predictions in human EEG data by examining whether construction-specific neural signatures emerge during sentence comprehension and whether their temporal and spectral characteristics align with patterns observed in computational models.

Participants

We recorded the neural responses of 12 participants (eight females, mean age 43 years, range 26–62 years) while they listened to an audiobook. One participant was excluded from further analysis due to insufficient data quality after preprocessing (e.g., excessive noise and artifact contamination), resulting in a final sample of 11 participants included in all reported analyses. All participants were healthy, right-handed individuals, English native speakers with normal hearing, and reported no history of neurological disorders or substance abuse. The study protocol was approved by the Ethics Committee of the University Hospital Erlangen (No: 22-361-2, PK). All participants provided written informed consent before their inclusion in the study, and all experiments were performed in accordance with relevant guidelines and regulations.

Stimulation protocol and EEG recording

Participants listened to the audio format of sentences generated by GPT-4. This dataset was designed to include sentences that exemplify four distinct ASCs: transitive, ditransitive, caused-motion, and resultative constructions as shown in Table 1. Each ASC category comprised 50 sentences, yielding a total of 200 sentences.

The sentence materials used in the present EEG experiment were drawn from the stimulus set employed in our previous computational studies investigating constructional representations in recurrent and transformer-based language models (Ramezani et al., 2025a,b). Because presenting the complete set of 2,000 sentences would have resulted in an impractically long recording session, we selected a random subset of 200 sentences, comprising 50 sentences from each of the four Argument Structure Constructions (transitive, ditransitive, caused-motion, and resultative). The same subset was presented to all participants, allowing for a direct comparison between neural representations in humans and the corresponding model representations.

This design was chosen to ensure direct comparability between artificial and biological systems. By presenting identical linguistic input across computational and neural experiments, the study enables a controlled test of whether similar representational distinctions emerge across systems when processing the same constructional variations. This approach strengthens the interpretation that any observed convergence reflects shared representational principles rather than differences in stimulus composition.

Our materials were restricted to sub-constructions of ASCs so that the sentences varied in construction type without introducing significant structural differences that could obscure compositional effects. The generated text was converted to audio using the Google text-to-speech model.1 The total number of sentences was 200, divided into three parts. In the middle of each part, participants could rest as much as they needed.

A challenge in natural-language EEG experiments is that words within the same syntactic category vary in duration, complicating comparisons of signals of different lengths. As shown in Figure 1, our stimuli exhibit such length variation. For ERP analyses, we aligned epochs to the maximum duration, whereas for statistical analyses, we used extracted features, which are independent of signal length.

EEG was recorded at 2,500 Hz using a Brain Products ActiCAP system with 64 active electrodes arranged according to the international 10–20 system. Electrode impedances were kept below 25 kΩ, and an experimenter monitored all recordings.

Data preprocessing

A systematic approach was used to identify and correct faulty channels. Channels showing abnormal characteristics—such as zero variability or unusually high noise—were flagged in the metadata. These channels were then reconstructed using spatial interpolation. In spherical spline interpolation, the scalp is modeled as a smooth surface, and a curvature-minimizing function is fit through the good electrodes; the missing electrode's value is estimated from this function, ensuring reconstruction without artificial peaks.

Next, an finite impulse response (FIR) band-pass filter between 1 and 45 Hz was applied using MNE. FIR filters convolve the signal with fixed coefficients so that each output sample is a weighted sum of past inputs, and their linear-phase properties preserve temporal structure. The 1 Hz cutoff removes slow drifts and movement-related artifacts. In comparison, the 45 Hz cutoff attenuates high-frequency noise (e.g., muscle and environmental interference) while retaining theta, alpha, beta, and lower-gamma activity. Filtering before downsampling also prevents aliasing.

The data were downsampled to 250 Hz to reduce computational load. By the Nyquist theorem, sampling must exceed twice the highest frequency present; with a 45 Hz cutoff, anything above 90 Hz is sufficient, so 250 Hz provides a wide margin. Downsampling was performed only after low-pass filtering to ensure no energy above the new 125 Hz Nyquist limit remained.

Independent Component Analysis (ICA; fastICA, 30 components, MNE) was applied to remove ocular artifacts. Following the approach used in Koelbl et al. (2023), components strongly associated with eye movements—most clearly visible in the Fp1 channel—or exceeding an artifact-correlation threshold of 0.2, along with components typically dominated by artifacts (e.g., ICs 0 and 1), were removed. This procedure effectively eliminated blink- and eye-movement-related activity.

After artifact correction, the continuous data were segmented into epochs time-locked to stimulus onset. Baseline correction was applied by subtracting the mean amplitude of the pre-stimulus window from each epoch to compensate for slow drifts. Trials with excessive amplitude fluctuations were automatically rejected.

Data analyses

After preprocessing, we assessed signal quality and verified trigger accuracy by inspecting ERPs time-locked to sentence onset. The averaged ERP across all participants and all four construction types is shown in Figure 2, revealing a clear P200 component. The prominent P200 component observed in the grand-average ERP (Figure 2) reflects early auditory and perceptual processing stages and serves as an indicator of reliable, time-locked neural responses to the auditory stimuli. Consistent with previous research, the P200 has been associated with early perceptual and phonological processing rather than higher-level linguistic integration. In the context of the present study, this component is therefore not interpreted as reflecting construction-level processing. Instead, it provides evidence for the overall quality and temporal alignment of the EEG data. The analyses relevant to constructional differentiation focus on later processing stages, particularly those associated with the accumulation of argument structure information across the sentence.

Statistical significance of decoding performance was assessed using permutation testing (1,000 label shuffles per participant and construction pair).

For control analyses, cross-validation was additionally performed using verb identity as a grouping factor to prevent lexical overlap between training and test sets.

Time-frequency analysis

Because sentence lengths vary, as shown in Figure 3. Epoch duration was standardized using the third-quartile length (covering 74% of sentences) to avoid overlap with subsequent stimuli.

Time-frequency representations were computed separately for each construction type. Because the sentences contained different syntactic items, a joint analysis was not possible; however, comparing the four plots allows us to assess whether corresponding sentence positions exhibit shared or construction-specific activity patterns.

In Figure 4, it is shown that the strongest construction-specific power occurs at lower frequencies, roughly between 2 and 5 Hz. To isolate construction-specific oscillatory activity, we subtracted a grand-average time-frequency spectrum computed across all construction conditions. Specifically, for each time-frequency point, the average power across the four construction types was calculated and subsequently subtracted from the corresponding values of each individual condition. This procedure removes components that are shared across constructions, such as general auditory processing and task-related baseline activity, thereby emphasizing relative differences between conditions. The resulting time-frequency representations therefore reflect residual, construction-sensitive deviations from the overall average response, allowing a more direct visualization of construction-specific spectral patterns.

Statistical analysis

A statistical analysis was conducted to identify activity patterns associated with each construction by first extracting relevant features from the EEG data. To enable comparison across categories, we focused on epochs aligned to typical syntactic roles—subject, verb, and object. Because sentence durations varied, direct waveform comparison was not possible; instead, we computed multiple statistical features from the broadband signal and from specific frequency bands (delta, theta, alpha, beta, and gamma).

Pairwise t -tests were then performed for each construction pair, with p -values corrected using the Benjamini–Hochberg procedure. Features with p < 0.05 were examined to determine which channels, frequency bands, or feature types distinguished the constructions. All statistical tests were conducted separately for each participant.

In Figure 5, the number of features showing significant differences (p < 0.05) across syntactic roles and construction pairs is summarized.

As expected, no significant differences emerged for the subject position, likely because participants could not distinguish constructions based solely on hearing the subject. A few important differences appeared for verbs, although verbs are generally not construction-specific. In contrast, the object position—occurring near the end of the sentence, when listeners have sufficient contextual information—showed substantially more significant differences than both subject and verb.

Pairwise comparisons revealed that not all construction pairs were distinguishable. The most significant differences appeared between transitive and resultative constructions, followed by caused-motion and ditransitive constructions. Other pairs showed only minor effects, and caused-motion vs. resultative yielded no significant differences.

Statistical analysis of frequency bands

As shown in Figure 6, although frequency-band contributions varied across participants, the alpha band exhibited the highest number of discriminative features, followed by beta and delta, while gamma showed the fewest effects. Importantly, this result reflects feature-based sensitivity rather than a direct comparison of spectral power across conditions. The counts of significant features therefore provide an indirect measure of how strongly different frequency bands contribute to construction discriminability, rather than indicating exclusive or dominant involvement in construction processing.

Complementary information about spectral dynamics is provided by the time-frequency representations (Figure 4), which illustrate how oscillatory activity evolves across time and frequency for each construction type. Together, these analyses suggest that lower-frequency bands, particularly in the alpha range, are sensitive to constructional variation, while cautioning against overinterpreting these findings as definitive evidence for band-specific processing. Alternative analyses based on average spectra would also be possible, but they were not the focus of the present feature-based approach.

Statistical analysis of feature type

The extracted features included mean amplitude, peak amplitude, standard deviation, root-mean-square, kurtosis, and zero-crossing rate. As shown in Figure 7, kurtosis and peak amplitude yielded the largest numbers of discriminative features, while mean amplitude and standard deviation showed more moderate effects and zero-crossing rate contributed only minimally.

Assessing construction separability via SVM classification

For classification, we chose a simple model—not to maximize accuracy, but to test whether construction differences were detectable at all—and used an SVM. We employed a pairwise classification approach, training an SVM with an RBF kernel and evaluating performance using Leave-One-Out (LOO) cross-validation. Accuracy, F1-score, and recall were computed for each construction pair to assess how well the corresponding brain responses could be distinguished.

Token classification

We implemented a classification task to assess whether the model could distinguish between constructions. Because signal lengths varied, we focused on two previously effective features—signal kurtosis and peak amplitude—which were extracted from subject, verb, and object epochs.

Pairwise decoding results

When classifying all four constructions simultaneously, the model achieved an average accuracy of 30%, only slightly above the 25% chance level. Consistent with our statistical results, not all classes were separable. We therefore switched to pairwise classification to examine each construction pair individually and obtain a clearer picture of their discriminability.

The relatively low performance in the four-class classification setting, despite stronger results in pairwise comparisons, likely reflects the combined influence of limited signal-to-noise ratio in speech-evoked EEG data and overlapping neural representations across construction types. While specific construction pairs exhibit sufficient differences to support reliable binary classification, the simultaneous separation of all four constructions presents a more demanding task, particularly when some constructions share structural and semantic properties. This pattern suggests that constructional representations are not uniformly distinct but exhibit graded similarity, which reduces global separability while still allowing robust discrimination between selected pairs.

In the two-class setting, the chance level is 50%, so accuracies at or below this threshold indicate no meaningful separation between classes. As shown in Figure 8, caused-motion vs. resultative and ditransitive vs. transitive showed the lowest classification performance, whereas ditransitive vs. resultative achieved the highest performance.

It is noteworthy that precision and recall values are nearly identical across classification conditions. This reflects the balanced and symmetric nature of the pairwise classification setup, in which both classes are equally represented and the decision boundary is optimized under comparable conditions for both classes. Under such circumstances, precision and recall are expected to converge, particularly when classification performance is relatively stable across class labels.

Group-level inference

Classification was performed separately for each participant, and decoding performance was evaluated for each pairwise construction contrast based on object-aligned epochs. Because multi-class classification remained close to chance level, subsequent analyses focused on pairwise contrasts, which provide a more sensitive assessment of construction discriminability.

Decoding performance was summarized at the group level by averaging accuracy across participants. To ensure that group averages reflected reliable above-chance decoding at the participant level, accuracy was expressed relative to chance (accuracy minus chance level), and subject-level inference was performed across participants using Wilcoxon signed-rank tests against zero. This approach allows assessing whether decoding performance was consistently above chance across participants rather than being driven by a subset of individuals. The resulting statistics for each construction pair are reported in Table 2.

Permutation-based validation

To further assess whether decoding performance exceeded chance levels, statistical significance was additionally evaluated using permutation testing. For each participant and construction pair, class labels were randomly shuffled 1,000 times, and decoding accuracy was recomputed to generate a contrast-specific null distribution. Observed decoding accuracies were then compared against these null distributions to estimate empirical permutation-based significance. An example of such a null distribution for Subject 06 in the caused-motion versus transitive contrast is shown in Figure 9.

Across construction pairs, observed accuracies consistently exceeded the corresponding permutation-based null accuracies for contrasts that showed strong decoding performance, whereas contrasts with low decoding accuracy overlapped with the null distribution. These results confirm that the observed discriminability cannot be explained by chance-level fluctuations in the data. Summary statistics across construction pairs are reported in Table 3.

Control analysis: verb leakage

To exclude the possibility that decoding performance was driven by lexical overlap rather than construction-level information, we performed an additional control analysis addressing potential verb leakage between training and test data. Inspection of the stimulus set revealed that several verbs occurred across multiple sentences and constructions, raising the possibility that classifiers could exploit verb-specific cues rather than constructional structure.

To control for this effect, decoding was repeated using a verb-controlled cross-validation scheme in which all sentences containing the same verb were assigned exclusively to either the training or the test set within each fold. This procedure prevents the classifier from relying on verb-specific information and requires generalization across verbs. Decoding performance under this constraint showed a slight reduction in overall accuracy but preserved the relative pattern of discriminability across construction pairs, as shown in Figure 10. In particular, contrasts that exhibited strong separability in the standard analysis remained above chance level.

These results indicate that construction-specific decoding cannot be explained by lexical overlap between constructions and instead reflects construction-level neural differentiation.

Model-derived predictions and neural confirmation

The present study was designed to test whether representational principles previously identified in computational language models generalize to human neural processing during sentence comprehension. In two preceding studies (Ramezani et al., 2025a,b), we showed that both a recurrent neural language model and a transformer-based language model develop internally differentiated representations of Argument Structure Constructions (ASCs) despite substantial architectural differences and minimal explicit supervision. Importantly, these computational results do not merely demonstrate that artificial systems can represent constructions; they generate concrete predictions about when and how construction-level information should emerge during incremental language processing. The EEG results presented here allow these predictions to be evaluated at the level of human neural dynamics.

A first prediction concerns the temporal locus of constructional differentiation. In both the recurrent and transformer-based models, construction-specific representations emerged only after sufficient argument structure information had accumulated to determine the underlying event structure of the sentence. Early elements such as subjects or determiners carried little construction-specific information, whereas later elements—particularly those associated with the object or completion of argument structure—showed the strongest differentiation. The EEG findings closely follow this predicted pattern. Construction-specific neural differences were largely absent at early sentence positions and became most pronounced at the object position, where listeners can reliably infer the constructional meaning of the sentence. This convergence suggests that construction processing in the human brain reflects an integrative stage of interpretation rather than early syntactic categorization, supporting usage-based accounts in which constructions emerge from accumulated relational information during comprehension.

A second prediction concerns the internal similarity structure among constructions. Computational analyses revealed that ASC representations are not uniformly separated but instead organized according to graded similarities reflecting shared structural and semantic properties. Some construction pairs formed clearly distinct clusters, whereas others remained partially overlapping. The EEG results exhibit the same qualitative structure: ditransitive and resultative constructions showed robust separability, whereas caused-motion and resultative constructions exhibited substantial overlap. This correspondence indicates that the convergence between artificial and biological systems extends beyond the detection of constructional differences per se and includes the relative geometry of these representations. Such alignment is consistent with the interpretation that both systems converge on similar representational solutions because they are constrained by common demands of event-structure integration during predictive language processing.

A third point of convergence concerns the level of processing at which constructional distinctions become neurally expressed. In computational models, construction-specific information emerges in intermediate and late representational stages associated with contextual integration and abstraction. In the EEG data, construction differences were most prominent in lower-frequency oscillatory activity, particularly in the alpha band, which has been associated with integrative and unification processes during language comprehension (Bastiaansen and Hagoort, 2006; Meyer, 2018). Alpha-band activity has been linked to the coordination of distributed neural processes involved in combining lexical and syntactic information into coherent representations, consistent with its proposed role in higher-level language integration. The parallel emergence of construction sensitivity at integrative processing stages across systems suggests that ASC representations reflect the construction of higher-level relational meaning rather than the processing of isolated lexical or syntactic features.

Taken together, the results provide converging evidence that construction-level representations follow similar computational principles across artificial neural networks and the human brain. Across recurrent models, transformer architectures, and human electrophysiology, construction distinctions emerge late in processing, depend on the availability of argument structure information, and exhibit comparable similarity relations between construction types. Rather than reflecting architecture-specific mechanisms, this convergence supports the view that constructions constitute stable representational solutions that arise naturally in systems optimized for predictive language processing. In this sense, the present findings strengthen the interpretation that constructional knowledge represents a psychologically real level of linguistic organization, grounded in general principles of learning, prediction, and event-structure integration.

Convergence between human EEG and neural language models

The present study demonstrates that human EEG signals differentiate between four Argument Structure Constructions (ASCs)—transitive, ditransitive, caused-motion, and resultative—during naturalistic sentence comprehension. Across complementary analytical approaches, including time-frequency analysis, statistical feature extraction, and machine-learning classification, the results consistently indicate that constructional differences are reflected in neural dynamics. These findings support the interpretation that the human brain represents ASCs as distinguishable form–meaning mappings rather than as purely lexical or syntactic combinations. Importantly, the observed neural patterns do not arise in isolation but align closely with representational structures previously identified in computational models of language processing, including both recurrent neural networks and transformer-based architectures. Taken together, this convergence strengthens the constructionist assumption that ASCs constitute psychologically real units of linguistic knowledge and suggests that similar abstraction principles may operate across biological and artificial neural systems.

A central contribution of the present work lies in the correspondence between neural signatures observed in EEG and the internal representational organization reported in computational models. In both the LSTM and BERT studies, sentence representations formed distinct clusters according to construction type, even under tightly controlled syntactic conditions and minimal semantic variability. The EEG findings exhibit a comparable structure of discriminability, indicating that construction-level distinctions emerge at analogous stages of processing across systems.

First, constructional differentiation was strongest at sentence-final positions, particularly at the object role. This mirrors the computational findings, where construction-specific information was most pronounced in representations associated with later integrative stages of processing. In recurrent models, hidden-state representations corresponding to the object position showed maximal separability between constructions, while in the transformer model object-related embeddings provided the most informative signal for construction identification. The EEG results show the same temporal profile: construction-specific neural differences emerge most clearly when sufficient contextual information has accumulated to determine the underlying event structure. The stability of this effect across permutation testing and verb-controlled decoding further indicates that the observed differentiation reflects construction-level processing rather than lexical confounds or statistical fluctuations.

Second, the similarity relations between constructions follow comparable patterns in EEG and computational models. Not all construction pairs were equally distinguishable; instead, separability reflected structural and semantic relatedness between constructions. Ditransitive and resultative constructions showed the strongest differentiation, whereas caused-motion and resultative constructions exhibited substantial overlap. The presence of this shared similarity structure suggests that convergence occurs not only at the level of detecting constructional differences but also in the organization of representational space itself. Such correspondence is consistent with the interpretation that both biological and artificial systems are constrained by similar demands of event-structure representation during predictive language comprehension.

Third, the spectral characteristics of the EEG effects provide an additional point of alignment with computational findings. Construction-specific differences were most prominent in the alpha frequency band, with additional contributions from beta and delta activity. Alpha oscillations have been consistently linked to integrative and unification processes during language comprehension, reflecting the coordination of distributed neural activity involved in combining lexical, syntactic, and semantic information into coherent meaning representations (Bastiaansen and Hagoort, 2006; Meyer, 2018). This interpretation aligns with the present findings, in which alpha-band activity shows sensitivity to constructional variation at stages of processing associated with higher-level integration. This interpretation aligns with the modeling results, in which construction-level distinctions emerged in representational stages associated with contextual integration rather than early lexical encoding. The sensitivity of alpha-band activity to constructional variation may therefore reflect a shared computational principle across systems: the integration of role-based information into higher-level relational meaning.

Taken together, these converging observations suggest that artificial neural networks—despite their architectural differences from biological systems—arrive at internal encodings that resemble those observed in human neural processing. Rather than implying direct mechanistic equivalence, this convergence indicates that both classes of systems may discover similar representational solutions when optimized for predictive language processing. In this sense, the parallel findings support the view that contemporary language models capture aspects of the computational structure underlying human sentence comprehension and provide a useful framework for generating testable hypotheses about neural language processing.

Construction-specific neural signatures and their linguistic interpretation

While the previous section focused on the convergence between neural and computational representations, the EEG results also allow a more direct interpretation of how constructions themselves are neurally encoded during sentence comprehension. The fact that subject and verb positions exhibited almost no construction-specific differences, whereas objects yielded robust effects, indicates that construction processing is incremental and context-dependent. Subjects do not yet constrain construction type, and verbs—although semantically informative—do not uniquely identify ASC membership across all items. Only when the object (and the associated argument structure) becomes available can listeners infer the construction's conceptual schema.

The pairwise EEG distinctions further support this view. Transitive and resultative constructions differed strongly, consistent with their divergent event structures. Caused-motion and ditransitive constructions also showed separability, albeit to a lesser degree. However, caused-motion vs. resultative did not differ at all. This is linguistically meaningful: many resultative verbs encode changes of location (e.g., pushed the box open), making them semantically similar to caused-motion constructions. These fine-grained semantic overlaps likely contribute to their neural similarity. This interpretation is further supported by the verb-controlled decoding analysis, in which construction discriminability remained largely preserved when lexical overlap between training and test data was eliminated, indicating that the observed differences cannot be explained by verb-specific cues.

An important alternative interpretation concerns the role of predictive and integrative processes in shaping the observed object-position effects. Within the Bayesian brain and predictive processing framework, language comprehension is viewed as a continuous process of updating probabilistic beliefs based on the interaction between prior expectations and incoming sensory evidence (Rao and Ballard, 1999; Körding and Wolpert, 2004; Friston, 2010; Hohwy, 2013). Verbs provide strong prior information about likely argument structure configurations (e.g., the verb give increasing the probability of a ditransitive construction), but these predictions are typically probabilistic and remain underspecified until additional linguistic input is encountered. The object position represents a stage at which sufficient evidence has accumulated to substantially reduce uncertainty about the intended construction, enabling a more complete interpretation of the underlying event structure.

From this perspective, the observed construction-specific effects may reflect the emergence of stable posterior representations resulting from the integration of prior expectations with incoming linguistic evidence, rather than the activation of constructional representations in isolation. Importantly, this interpretation is fully compatible with the constructionist framework adopted here. In Construction Grammar, constructions can be understood as emergent form–meaning pairings that develop through repeated experience with recurrent event structures (Goldberg, 1995, 2003, 2006, 2009, 2019). Likewise, predictive processing does not posit fixed symbolic representations, but interprets linguistic knowledge as hierarchical probabilistic models that are continuously updated by prediction errors. The present findings are therefore consistent with the view that constructional representations become most strongly expressed at stages of processing where sufficient relational information has accumulated to support coherent event interpretation and where prediction uncertainty has been substantially reduced (Clark, 2013; Nieuwland et al., 2018).

Importantly, these effects emerged primarily at low frequencies (2–5 Hz), which are associated with chunking, phrase-level integration, and hierarchical parsing. This suggests that ASC distinctions arise at the level of conceptual event-structure integration rather than at the phonological or lexical level. Consistent with this interpretation, permutation-based validation confirmed that reliable decoding emerged only for construction pairs showing clear conceptual differences, further suggesting that ASC distinctions reflect higher-level integration processes rather than chance-level fluctuations or surface-level lexical variation.

Frequency band contributions and time windows

From a neurocognitive perspective, the spectral profile of the observed effects provides additional insight into the processing stage at which constructional distinctions arise. Consistent with previous ERP and time-frequency studies on combinatorial syntax and semantics (Bastiaansen and Hagoort, 2006), the alpha band was the most sensitive spectral range for detecting ASC differences. Alpha oscillations have been linked to hierarchical processing, anticipatory unification, and controlled retrieval of structured linguistic information. The additional contributions from beta and delta bands align with their known roles in syntactic prediction and temporal integration.

The timing of these effects - emerging near the object position - further supports the interpretation that ASCs are processed as event-level templates rather than as lexical fragments. The EEG therefore appears to reflect not merely the recognition of individual words but the incremental construction of a relational meaning pattern. This interpretation is further supported by the verb-controlled decoding analysis, which indicates that construction-specific effects persist when lexical overlap between constructions is removed.

Individual differences and cognitive variability

The analyses revealed noteworthy individual variation in the degree and location of construction-specific neural effects. Such variability is consistent with previous ERP research demonstrating differences in morphosyntactic processing strategies among native speakers (Tanner and Van Hell, 2014). Variation in linguistic performance, cognitive control, or attentional strategies can also modulate construction processing (Fiorentino et al., 2018). This raises a broader theoretical question discussed in Construction Grammar: Do all native speakers share the same inventory of constructions, or are there idiosyncratic, experience-based variants?

Our data support the latter possibility. While all participants showed construction sensitivity at the group level, as confirmed by subject-level inference across decoding analyses, the precise channels, frequency ranges, and feature types that distinguished constructions differed across individuals. This suggests multiple neurocognitive pathways for successful construction comprehension, aligning with usage-based theories that emphasize learning histories and experience-shaped representations. In this sense, the EEG results suggest that constructional representations emerge at the level of event-structure integration, providing a linguistic interpretation of the representational convergence discussed above.

Limitations

As with many EEG studies employing controlled linguistic materials, several methodological limitations should be acknowledged when interpreting the present findings.

First, although the stimulus set was carefully designed, we were unable to control sentence length and lexical content with full precision. This variability introduces noise into the EEG signal and complicates the alignment of neural responses across stimuli.

Second, the dataset was small both in terms of the number of participants and the number of trials per construction type. Such constraints can increase the influence of individual differences and may reduce statistical power, potentially obscuring more subtle construction-specific effects. To mitigate these limitations, decoding results were validated using permutation testing, subject-level statistical inference, and verb-controlled cross-validation, which together reduce the likelihood that the reported effects reflect chance fluctuations or lexical confounds.

An additional limitation concerns the use of synthetic speech stimuli generated via text-to-speech systems. While this approach provides a high degree of experimental control over linguistic content, it lacks the rich prosodic cues present in natural speech, such as intonation, rhythm, and stress patterns, which can facilitate the disambiguation of argument structure during comprehension. The absence of such cues may have influenced both the magnitude and temporal dynamics of the observed neural responses, potentially attenuating construction-specific effects. At the same time, the use of controlled synthetic stimuli was essential for isolating constructional differences while minimizing confounding variability across conditions.

Methodological considerations: sub-constructions and experimental constraints

The present experimental design deliberately focused on sub-types of ASCs rather than highly diverse surface forms in order to maximize experimental control. This design had clear advantages: it allowed aligning syntactic roles across sentences and isolating constructional effects from extraneous linguistic variability. However, the similarity among sub-constructions—particularly between caused-motion and change-of-state/resultative verbs—also makes strong neural dissociations more difficult to detect.

The fact that we nonetheless observed clear construction-specific signatures, supported by permutation-based validation and verb-controlled decoding analyses, indicates that ASCs have a robust neural footprint even under conservative stimulus control. Future studies may test broader ASC families, examine cross-linguistic generalization, or manipulate semantic dimensions within constructions (e.g., location-changing vs. property-changing resultatives) to further refine the representational landscape.

Our results provide converging evidence that EEG signals carry construction-specific information during real-time sentence comprehension and that the structure of these neural representations closely mirrors those observed in computational language models. This neural–computational correspondence strengthens the case for the cognitive reality of constructions and provides evidence that both brains and artificial neural networks rely on comparable learning principles—incremental prediction, role-based integration, and event-structure abstraction—when processing linguistic input.

Future studies may extend the present work by increasing sample size and broadening the set of constructions examined, thereby supporting more robust generalization and finer distinctions among ASC types. Examining L2 learners represents another promising direction, with the potential to illuminate how construction-specific neural signatures vary across proficiency and learning histories. Additionally, cross-linguistic investigations employing parallel constructions across different native languages could help determine whether ASCs rest on universal cognitive principles or emerge from language-specific experience.

Although our study focused on neural signals, the results have broader implications for behavioral theories of sentence processing. Behavioral studies often show graded effects of argument structure predictability, thematic role assignment, and event-structure inference. The present EEG data suggest that these behavioral patterns may emerge from underlying construction-sensitive neural dynamics that accumulate evidence across the sentence until a constructional interpretation becomes available.

The convergence of three lines of evidence—behavioral patterns reported in psycholinguistics, neural signatures shown here, and emergent representations in recurrent and transformer-based language models—is consistent with the view that ASCs constitute a psychologically and computationally real level of linguistic organization, guiding human sentence comprehension and emerging naturally in predictive learning systems.

More broadly, the observed convergence between neural and computational representations can be interpreted within the conceptual framework of a structured representational landscape, in which learning systems may converge toward stable and computationally efficient solutions. Such ideas have recently been discussed in terms of “Platonic” regularities in representational space (Huh et al., 2024; Metzner, 2025a,b; Levin, 2025; Solé et al., 2026). However, this interpretation should be approached with caution. The present data do not provide direct evidence for the existence of such a space, but are merely consistent with this perspective.

Importantly, alternative explanations must also be considered. In particular, artificial language models are trained on large corpora of human-generated language, which may already reflect the statistical structure of human cognition. As a result, similarities between artificial and biological systems may arise not only from shared computational principles, but also from shared input distributions. The observed convergence should therefore be interpreted as suggestive rather than definitive, motivating future work to disentangle these contributing factors.