Neural harmonics of syntactic structure

Can neural rhythms reflect purely internal syntactic processes in multi-word constructions? To test this controversial conjecture - relevant to language in particular and cognition more broadly - we recorded electroencephalographic and behavioural data as participants listened to isochronously presented sentences of varying in syntactic complexity. Each trial comprised ten concatenated sentences and was either fully grammatical (regular) or rendered ungrammatical via randomly distributed word order violations. We found that attending the regular repetition of abstract syntactic categories (phrases and sentences) generates neural rhythms whose harmonics are mathematically independent of word rate. This permits to clearly separate endogenous syntactic rhythms from exogenous speech rhythms. We demonstrate that endogenous but not exogenous rhythms predict participants’ grammaticality judgements, and allow for the neural decoding of regular vs. irregular trials. Neural harmonic series constitute a new form of behaviourally relevant evidence for syntactic competence.

Information for a complete list of sentences (N = 300). Using a signal detection approach [32], we estimated task sensitivity and response  all ps ≤ 1.080*10 -31 , FDR-corrected (threshold p = 0.001, Figure S2a,b,c). No other 1 8 7 significant rhythmic components were found, suggesting that acoustic input could not 1 8 8 drive higher-order, syntactic rhythms (see Supporting Information). For neural data, we calculated inter-trial phase coherence estimates as the ratio of tagging effects were investigated using all trials, regardless of grammaticality. Figure   1 9 5 4, left column, illustrates phase coherence peaks for each sentence structure.

5 8
We evaluated the amount of evidence provided by phase coherence data using a 2 5 9 Bayesian rmANOVA. The Bayes factor indicates that the data -across sentence  We have determined that the brain encodes the duration of regularly repeated period. However, the emergence of significant peaks of neural activity does not imply 2 7 5 that they encode perceived grammaticality. To test this, we re-calculated SNR 2 7 6 estimates of inter-trial phase coherence separately for hits (correctly identified 2 7 7 irregular trials) and correct rejections (correctly identified regular trials). Omnibus 2 7 8 rmANOVAs with factors frequency peak, sentence structure and grammaticality, 2 7 9 separately for each sentence context, highlighted a significant grammaticality effect: 0.24, SD = 0.13 (see Figure 5a). For the fast five words context, a significant Using a Bayesian rmANOVA with factors sentence context and grammaticality indicates decisive evidence in favour of the alternative hypothesis. As for the context 2 9 8 factor, decisive evidence was obtained for the difference between fast five-and four- is modulated by trial grammaticality. Harmonics encode grammaticality judgements 3 1 6 When participants decide whether a trial contained ungrammatical sentences or not 3 1 7 (irregular trial), their judgment must be driven by implicit syntactic competence of the 3 1 8 German language. However, violations of expected word order also represent a form exogenous index did not: F (1,29) = 0.70, p = 0.407 (see Figure 6a). This suggests that participants' ability to decide whether a trial was regular or irregular: endogenous, adjR 2 = 0.28 (see Figure 6b). To decide whether the results for exogenous rhythms 3 3 8 in irregular trials reflect sensory surprise or grammatical violation detection, we 3 3 9 resorted to a neural decoding approach. We calculated standardized phase feature dimension to classify hits and correct rejection trials from neural rhythms.

4 3
The performance of the classifier -scored by computing the nonparametric measure Crucially, however, it was at chance for entrainment to word rate: .50, CI: 0.47-0.53 3 4 8 (see Figure 6c). Hence, we conclude that only top-down rhythms, constituting a 3 4 9 harmonic series of syntactic node repetition period, encode information about a trial The strength of harmonic series predicts the degree to which native German 3 9 9 speakers are sensitive to word order violations, i.e., the grammaticality manipulation Germanic language family, English, sensibly, assigns word order more weight for 4 0 3 comprehension than German does [38]. We therefore predict that the sensitivity to By correctly characterizing the correspondence between linguistic functions and 4 1 2 brain data, we contribute to linguistic theory by proving no evidence for the existence structure of NPs and VPs (see also [39][40]). In this respect, a frequency tagging 4 1 7 approach to abstract multi-word categories has the potential to contribute directly to The delta band (0.5-4 Hz) and infra-slow (< 0.5 Hz) rhythms we measured lie within violations as a function of neural response strength.

4
In conclusion, we have shown that internally constructed neural rhythms and their Stimuli. To reduce implicit word-and phrase-level prosodic cues, word stimuli were Praat, v. 6.0.16, www.praat.org). If an original stimulus was longer than the relative 4 7 0 SOA, its duration was compressed to fit the stimulation window while preserving 4 7 1 pitch.

7 2
Individual trials were composed of ten sentences, identical from the viewpoint of Information). Within a trial, no sentence was ever repeated. There were two types of 4 7 7 trial: regular trials, which contained only grammatically correct sentences; and 4 7 8 irregular trials, which contained also two ungrammatical sentences, obtained by periods between blocks were self-determined by each participant.

9 6
Trial presentation was blocked by sentence structure. Stimuli within a trial were NP is a phrase with a noun as head (e.g., "The girls"), while a VP has a verb as head 5 0 4 (e.g., "play rugby"). NPs usually perform the grammatical functions of verb subject or sentences. See Figure 2 for an exemplary analysis of a sentence used in each 5 0 9 condition. Stimulus sequences were created using custom scripts written in Matlab minus zscore(False Alarms)) and response bias (criterion = -5 1 7 0.5*(zscore(Hits)+zscore(False Alarms))) for each participant and condition [32].

1 8
We used an actiCAP 64-channel, active electrode set (10-10 system, Brain Vision 5 1 9 Recorder, Brain Products, brainproducts.com) to record electroencephalographic pass), and ICs marked as artefactual were rejected before trial epoching. Each ten- the average of two samples before and two samples after it.

4
To verify the presence of significant neural peak and the effect of syntactic approach to trial classification was run on the standardized (z-scored) phase 5 6 0 coherence scores for each participant and each harmonic series peak, with the 5 6 1 addition of the word rate peak. Three independent, numerically balanced subsets 5 6 2 ("folds") were created for each peak from the scores of all participants ("samples"), tested by predicting whether a trial label was irregular or regular. Accuracy was 5 6 7 quantified using the area under the curve of the receiver-operating characteristic This analysis was repeated three times, with each trial subset serving as testing set Anonymised pre-processed datasets for this experiment, as well as the relevant 5 7 6 analysis scripts, are in the process of being made available in OSF 5 7 7 (https://osf.io/rqfws/, currently private). Acknowledgements 5 8 0 The authors would like to thank Cornelius Abel, Jana Gessert, Freya Materne,