Alpha-Band Construction Decoding Should Scale With Observer-Field Coherence Across Human Listeners

Authors: Sterling Geisel, QBist Lab; Dr. Hideo Tanaka

Abstract

Ramezani, Kinfe, Maier, Schilling, and Krauss report that human EEG signals distinguish argument structure constructions during auditory sentence comprehension. The strongest effects arise near sentence-final object positions and appear most clearly in alpha-band features. The study also claims convergence with recurrent and transformer language models, where constructional representations emerge late, after sufficient relational information has accumulated. This working paper applies the Postulate of Observer As Field. The claim is narrow. If the observer is a spatially and temporally extended field of expectation, then construction decoding should depend not only on stimulus class but also on measurable coherence in the listener’s pre-stimulus neural state. Pudding Theory predicts that higher pre-stimulus phase coherence will increase object-position construction separability, especially for contrasts already near the neural decision boundary. This is not a semantic claim about language. It is a field claim about expectation, integration, and bias in representational geometry.

Source Synopsis

The source paper studies whether human neural activity differentiates Argument Structure Constructions during real-time sentence comprehension. Ten or eleven usable native English participants, depending on the reported analysis subset, listened to 200 synthetic English sentences. The sentences were balanced across four construction types: transitive, ditransitive, caused-motion, and resultative. EEG was recorded with 64 active electrodes at high sampling rate, then filtered, downsampled, cleaned with ICA, segmented into epochs, and analyzed across syntactic roles.

The paper is explicitly motivated by Construction Grammar. In that framework, linguistic constructions are form-meaning pairings, not mere epiphenomena of lexical items or phrase-structure rules. The authors use this framework to ask whether constructions have measurable neural signatures. They also connect the experiment to earlier model work. Recurrent and transformer language models had shown internally differentiated representations of the same construction types. The current experiment tests whether human EEG shows a similar time course and similarity structure.

The main result is temporal. Construction-specific differences are weak or absent at the subject position. They remain limited at the verb position. They become strongest at the object position, where enough argument structure information is available to infer event structure. The source paper states that construction-specific neural signatures emerge “primarily at sentence-final positions, where argument structure information becomes fully disambiguated.” The most informative frequency range is reported as the alpha band, with beta and delta also contributing.

Decoding analyses use pairwise SVM classification with leave-one-out cross-validation, permutation tests, and a verb-controlled procedure to reduce lexical leakage. Ditransitive versus resultative is the strongest contrast, with reported mean observed accuracy near 0.705 against a permutation baseline near 0.456. Caused-motion versus resultative and ditransitive versus transitive remain near chance. The paper interprets this graded pattern as evidence that human neural representations and artificial model representations converge on similar constructional geometry.

Postulate Lens

This paper applies Observer As Field.

The source study is not about a passive detector receiving sentence tokens. It is about a listener whose neural state accumulates relational constraints until a constructional interpretation becomes available. Observer As Field fits because the relevant unit is not a point observer at the moment of classification. It is an extended expectation field, expressed through distributed EEG coherence, that conditions how incoming linguistic evidence is integrated. The source result already depends on accumulation, not on isolated stimulus onset. Subject, verb, and object positions differ because the field of expectation has different available constraints at each time.

The Postulate therefore does not replace Construction Grammar. It sharpens a missing physical variable. In the source model, constructional separability is a property of stimulus category, syntactic position, and learned representation. In Pudding Theory, separability also depends on the observer-field state immediately before and during integration. The listener’s field phase, indexed operationally by pre-stimulus and pre-object alpha coherence, should bias the geometry of EEG features. The prediction is not that the listener chooses the construction. The prediction is that coherent expectation changes the gain of construction-level differentiation once the sentence supplies enough relational information.

This is a conservative use of the Postulate. It does not require distant action. It does not require anomalous semantics. It requires only that expectation is a measurable bias in a distributed observer field.

Pudding Theory Prediction

Pudding Theory predicts that the source paper’s strongest decoding results should split by listener-field coherence. The observable should be computed before the critical object-aligned interval, not after it, to avoid circularity. For each trial, estimate alpha-band phase coherence across a left fronto-temporal and temporal-parietal electrode set during the pre-object window. Then classify trials into high-coherence and low-coherence bins within each participant. The same construction classifier used by Ramezani et al. should then be trained and tested separately inside those bins, with sentence identity, verb identity, and construction balance controlled.

The prediction is specific. Object-aligned construction decoding should be higher in high-coherence trials than in low-coherence trials. The effect should be largest for contrasts that are discriminable but not trivial, such as caused-motion versus ditransitive, transitive versus resultative, and ditransitive versus resultative. For contrasts near chance because their event structures overlap, especially caused-motion versus resultative, coherence should not create a large separation from nothing. It should increase separability only where the sentence supplies distinguishable constructional evidence.

The expected sign is positive. Let delta A be the within-participant difference between high-coherence and low-coherence object-position decoding accuracy, after subtracting each bin’s permutation baseline. Pudding Theory predicts delta A greater than zero across participants for the reliably decodable construction pairs. A strict version predicts a rank-order relation: participants with larger alpha-coherence modulation across trials should also show larger construction decoding gains.

This prediction differs from the source consensus in emphasis. Standard analysis treats alpha activity as a correlate of linguistic integration. Pudding Theory treats coherent alpha organization as part of the observer-field boundary condition that biases representational differentiation. The same stimulus can therefore yield different construction geometry depending on the listener’s pre-object field state. The sentence supplies the evidence. The observer field changes the precision with which that evidence becomes separable.

Falsifiable Observable

The distinguishing observable is the within-participant high-minus-low pre-object alpha phase coherence effect on object-position pairwise construction decoding accuracy, after permutation-baseline correction. If pre-object alpha-coherence decoding gain were measured to be less than or equal to 0.00 accuracy units across the reliably decodable construction pairs, this Postulate would be falsified. The measurement must use held-out trials, verb-controlled folds, and coherence bins defined before the object response window.

Editorial Dialogue

Tanaka: The danger is clear. You are rebranding attention. Alpha coherence already belongs to mainstream cognitive neuroscience. It tracks inhibition, expectancy, and integration. Nothing in the source paper requires a new field ontology.

Sterling: The criticism is valid only if the claim stops at correlation. It does not. The Postulate adds a directional constraint. Pre-object coherence must predict later construction separability after lexical and construction balance are controlled. That is stronger than saying alpha participates in comprehension.

Tanaka: But the experiment has few participants. A binning analysis could be unstable. Individual variability could produce a false gradient.

Sterling: Then the Postulate faces the data. The correct test is hierarchical. Estimate the coherence effect within participant, subtract permutation baselines, and test the group-level slope. If the sign fails, the application fails.

Tanaka: The artificial model convergence still does not imply observer fields. Models have no biological expectation field in your sense.

Sterling: Correct. The model comparison supplies representational geometry, not the field mechanism. The human EEG experiment supplies the place to test the field mechanism. Pudding Theory predicts that biological construction geometry is conditioned by measurable observer coherence. Artificial systems provide a contrast case. Their internal states may converge on similar representational regions, but the present Postulate concerns the human listener’s field during sentence integration.

Discussion

The source paper is a useful test case because its claim is already graded. Some construction pairs separate. Others overlap. That structure prevents an easy all-purpose prediction. Pudding Theory cannot simply predict “more decoding.” It must predict when observer-field coherence changes separability and when semantic overlap limits the effect.

Several limitations matter. The sample is small. The stimuli are synthetic. Sentence length varies. The source analysis relies on extracted features rather than direct waveform comparison. Verb leakage is addressed, but broader lexical and prosodic controls remain incomplete. These constraints do not invalidate the source result, but they limit the precision of any field-based reanalysis.

The proposed test would change the conclusion only if coherence has no prospective effect. If alpha coherence predicts decoding only after the object interval begins, the result belongs to ordinary evoked processing. If pre-object coherence predicts object-position separability, Pudding Theory gains a narrow foothold: expectation is not only a cognitive label but a measurable field condition that shapes representational geometry during linguistic integration.

References

Ramezani, Pegah, Thomas M. Kinfe, Andreas Maier, Achim Schilling, and Patrick Krauss. “Convergent Representations of Linguistic Constructions in Human and Artificial Neural Systems.” arXiv:2603.29617. DOI: doi:10.48550/arxiv.2603.29617.

Geisel, Sterling. “Pudding Theory: A Topological Theory of Information Fields.” QBist Lab Working Paper, September 10, 2025.

Bastiaansen, Marcel, and Peter Hagoort. “Oscillatory neuronal dynamics during language comprehension.” Progress in Brain Research 159 (2006): 179-196.

Fedorenko, Evelina, Michael K. Behr, and Nancy Kanwisher. “Functional specificity for high-level linguistic processing in the human brain.” Proceedings of the National Academy of Sciences 108.39 (2011): 16428-16433.

Fiorentino, Robert, Lauren Covey, and Alison Gabriele. “Individual differences in the processing of referential dependencies: Evidence from event-related potentials.” Neuroscience Letters 673 (2018): 79-84. DOI: https://doi.org/doi:10.1016/j.neulet.2018.02.014.

Goldberg, Adele E. “Constructions: A new theoretical approach to language.” Trends in Cognitive Sciences 7.5 (2003): 219-224.

Tanner, Darren, and Janet G. Van Hell. “ERPs reveal individual differences in morphosyntactic processing.” Neuropsychologia 56 (2014): 289-301.

Solé, Ricard, et al. “Cognition spaces: natural, artificial, and hybrid.” arXiv preprint arXiv:2601.12837 (2026).