Sixteen native American English speaking New York University undergraduate students (4 males; 2 left-handed; age range 18–22) participated in this study. They were given course credit for participation. The institutional review board of New York University approved the study protocol, and written informed consent was obtained from all participants before the study.

The stimuli and procedures were identical to experiment 6 (sentence condition and shuffled sequence condition) of a recent MEG study (Ding et al., 2016). In the sentence condition, each English sentence contains 4 monosyllabic words (Figure 1A). The first two words form a noun phrase (adjective/pronoun + noun) and the last two words form a verb phrase (verb + noun). English words were synthesized independently using the MacinTalk Synthesizer (male voice Alex, in Mac OS X 10.7.5). Each monosyllabic word was adjusted to 320 ms in duration. In each trial, 12 sentences were presented without any acoustic gap between sentences (continuous, isochronous presentation). Therefore, each trial is 15.36 s in duration. Thirty trials were played, eight of which contained outliers. An outlier trial was the same as a normal trial except that 3 consecutive words from a roved position were replaced with 3 random words. The behavioral response was correct in 72% (SD = 9%) trial for the sentence condition.

In the control condition, each syllable in the sentence condition was cut into 5 slices (72-ms in duration with a 10-ms overlap between neighboring slices, which is smoothed by a 10-ms linear ramp) and slices at the same position within a sentence were shuffled across sentences. The control stimulus is not intelligible speech but has similar acoustic properties as the 4-syllable sentences (Figures 1B,C). Detailed procedures of how the control stimuli were made are described in Ding et al. (2016). The control stimulus has the same duration as the sentence stimulus and 30 trials were presented. Eight trials contained outliers consisting of 4 randomly chosen English words embedded in the stimulus. The behavioral response was correct in 81% (SD = 14%) of the trials for the control condition. The behavioral score of one subject (33%) clearly differed from other subjects. If this subject was excluded, the mean correct rate was 84% (SD = 6%).

The spectrogram of an example segment of a stimulus is shown Figure 1B for both conditions. The spectrogram is calculated using an auditory model (Yang et al., 1992). The power spectrum of the temporal envelope of the stimuli is shown in Figure 1C. The temporal envelope is the average of the spectrogram over frequencies, and the power spectrum is calculated using the Discrete Fourier Transform (DFT) without any smoothing window. The power spectrum in Figure 1C is averaged over all 15.36-s duration trials. The stimulus envelope shows a spectral peak only at the syllabic/word rate.

The experiment was conducted in a quiet room. The sentence condition and the control condition were presented in separate sessions and the order of these two sessions was counterbalanced over subjects. The participants were instructed to distinguish normal trials from outlier trials by pressing a response keys at the end of each trial.

EEG was continuously recorded with a 128-channel EEG system (EGI, Inc., Eugene, OR), digitized at a sampling rate of 1,000 Hz (bandpass filter = 0.01–400 Hz) and referenced to the vertex (Cz). The impedance of electrodes was kept below 40 kO (Ferree et al., 2001). EOG artifacts were removed from the EEG recordings using ICA (Delorme and Makeig, 2004). Specifically, the 128-channel EEG signals were dimension reduced to 80 components using PCA and then the 80 principal components were further decomposed using ICA. An independent component was removed if in its topography the mean power over the most frontal 14 channels was more than 10 times stronger than the mean power over all other channels.

The EEG signal was lowpass filtered to 25 Hz, since the signals of interests are in the low-frequency region, at 1/1.28, 2/1.28, and 4/1.28 Hz. Data were re-referenced offline to a common average reference. The response to each trial was epoched. The recorded data from the first sentence of each trial was removed to avoid the transient EEG response to sound onset.

The EEG response in each trial was converted into the frequency domain using the DFT. After the first sentence was removed, each trial was 14.08 s in duration (9 sentences × 1.28 s/sentence) and therefore the frequency resolution of the DFT of the entire trial is 0.071 Hz, i.e., 1/14.08 Hz. If the DFT of the response in trial k is denoted as X_k(f), the evoked power spectrum is shown in equation (1), where K is the total number of trials. X_k(f) is complex-valued Fourier coefficient and is a function of frequency f. The evoked power reflects the power of EEG responses that are synchronized to the speech input. It is the same as the power spectrum of the EEG response waveform averaged over trials.

The inter-trial phase coherence is defined in equation (2), where θ_k is the phase angle of each complex-valued Fourier coefficient, i.e., θ_k = <X_k(f).

The induced power, i.e., the power of EEG responses not synchronized to the speech input, is also calculated as the following formula, where <X(f)> denotes the mean over trials.

The statistical significance of neural entrainment at a target frequency was tested for evoked power and inter-trial phase coherence, respectively. In the power test, to remove the 1/f trend of the response power spectrum, the response power at each frequency was normalized by the neighboring 14 frequency bins (7 bins on each side, which is equivalent to 0.5 Hz). The normalized power (equation 3), which can be viewed as a signal-to-noise measure, is:

where ω denotes frequencies around the target frequency f.

In the phase coherence test, phase coherence values are not normalized by the neighboring frequency bins, since the inter-trial phase coherence spectrum has no 1/f fall-off.

For the power test, the null hypothesis is that the power at the target frequency is not significantly larger than the power in neighboring frequencies. Under the null hypothesis, the normalized power E_n(f) is subject to an F_{(32, 448)} distribution for each channel. When the response power is averaged over channels, since the EEG response is correlated over channels, we conservatively assumes that the normalized power calculated based on the power averaged over channels is also subject to an F_{(32, 448)} distribution. For the phase coherence test, the null hypothesis is that the response phase is not synchronized to the stimulus and the null distribution of θ_k is a uniform distribution. Therefore, we employed the F-test and the Rayleigh test, respectively, to evaluate the statistical significance of the evoked power and phase coherence at each target frequency.

For the response averaged over channels, the null distribution of the evoked power or phase coherence cannot be easily described by a parametric distribution due to the correlation between channels. Therefore, the null distribution of normalized power (or inter-trial phase coherence) is estimated based on the response at non-target frequencies, i.e., the responses at frequencies that are not harmonically related to the sentential rate. The chance-level normalized power (or phase coherence) is pooled over frequencies. When the significance test is applied to individual subjects, the chance-level power (or phase coherence) is pooled over subjects. The statistical significance of the response at a target frequency is the probability that the target-frequency response differs from a chance-level response.

A linear classification analysis is employed to test if the topographic patterns at two frequencies or in two conditions are significantly different. In this analysis, the topographic plots averaged over half of the trials (e.g., the first or last 15 trials) are used to train a classifier. Each subject is viewed as a sample. The classifier's performance is evaluated based on the data averaged over the other half of the trials. A binomial test (N = 16, probability: 0.5) is used to test if the classifier can discriminate the two classes of topographic plots with higher than chance performance.

Whether sentences are represented by hierarchically embedded syntactic structures or linear Markov models during language comprehension is a central question in cognitive science (Chomsky, 1957; Townsend and Bever, 2001; Jackendoff, 2002; Phillips, 2003; Frank et al., 2012; Everaert et al., 2015). Recently, a number of studies have suggested that neural processing of languages cannot be fully explained by a linear Markov model and is consistent with hierarchical syntactic models. For example, using fMRI, Brennan et al. tested how well BOLD signals in each cortical area can be explained by a hierarchical models or Markov models (Brennan et al., 2016). They found that while hierarchical models predicted BOLD signals in the anterior and posterior temporal lobes, a Markov model predicted activity in a broader cortical network including the frontal lobe. Neural entrainment to linguistic structures also provides a useful tool to address how syntactic structures are represented in the brain. For example, previous MEG data show that neural activity can entrain to linguistic structures even without any statistical cues for structural boundaries, demonstrating that statistical cues are not the only cues for syntactic analysis (Ding et al., 2016).

When listening to speech, cortical activity is entrained to the temporal envelope of speech (Luo and Poeppel, 2007; Kerlin et al., 2010; Lalor and Foxe, 2010), which carries the acoustic rhythm of speech. It has also been shown that cortical activity carries phonetic information (Di Liberto et al., 2015). Furthermore, as shown in this study using EEG and previous studies using MEG and ECoG (Ding et al., 2016), cortical activity is also entrained to the rhythms of higher level linguistic structures such as phrases and sentences, in the absence of acoustic cues. Therefore, during speech listening, cortical activity on different time scales is concurrently synchronized to linguistic structures of time scales. Therefore, a hierarchy of linguistic structures are converted into neural dynamics on different time scales, providing a plausible neural basis for the mental representation of hierarchical linguistic structures and the interactions between linguistic levels during speech processing (Townsend and Bever, 2001; Poeppel et al., 2008; Christiansen and Chater, 2016).

Although a syllabic-rate response is observed in both the sentence condition and the control condition, it is weaker in the control condition. This effect may reflect better neural tracking of intelligible monosyllabic words. Alternatively, it is possible that the subjects paid more attention in the sentence condition, which enhances the syllabic/word rate response. The sentence condition may require a higher level of attention, since the behavioral task is more challenging in the sentence condition (~70% correct rate in the sentence condition vs. ~80% correct rate in the control condition).

To dissociate syntactic from prosodic processing, the current study removes prosodic cues in speech. Natural speech, however, contains rich prosodic information which facilitates syntactic analysis. Therefore, neural tracking of phrasal and sentential structures is likely to be more prominent in natural speech. Furthermore, prosodic cues can also directly generate event-related response tracking the structural boundaries, such as the CPS (Steinhauer et al., 1999).

In this study, 30 trials of sentence sequences (~15 s each) are presented and the neural tracking of higher-level linguistic structures, i.e., phrases or sentences, can be detected in more than half of the subjects using EEG. Therefore, the EEG-based paradigm shown here provides a plausible way to measure the neural encoding of higher-level linguistic structures. The sentential and phrasal responses do not reach significance in some subjects, possibly limited by the low SNR of EEG recordings. The subjects in the current study are all young adult native speakers without any language disorders, and therefore individual differences in language ability should be small, given such elementary processing demands. Nevertheless, the performance of detecting random words embedded in a sequence of sentences shows considerable individual differences, which is likely to be driven by cognitive factors, e.g., attention, rather than language ability. Indeed, a recent study shows that neural tracking of phrasal and sentential structure is diminished during sleep (Makov et al., 2017). Future experiments are needed to elucidate the influence of tasks and cognitive states on the neural tracking of phrasal and sentential structures.

It is challenging to detect of low frequency neural activity, since background neural activity generally has a 1/f spectrum. Compared with the 4-Hz syllabic-rate response and the 1-Hz sentential-rate response, however, the 2-Hz phrasal-rate response is especially difficult to detect in individual subjects. A possible reason is the following: The syllables have very clear acoustic boundaries and therefore can drive strong auditory responses. Each sentence is a syntactically and semantically coherent unit and, in this experiment, different sentences are syntactically and semantically disconnected. Therefore, the sentences also have relatively clear perceptual boundaries. The phrases within a sentence, however, are related both syntactically and semantically, which makes the boundaries between them less obvious than those between syllables and sentences.

Finally, since EEG is a commonly available non-invasive neural recording technique, the current paradigm has the potential of being developed into a tool to assess higher-level linguistic processing in populations less able to engage in typical laboratory research, including children and clinical patients. To apply the current paradigm to a special population, however, possibly requires adapting the sentence materials based on the vocabulary familiar to the target population and elucidating how attention and other cognitive factors may influence the neural tracking of higher-level linguistic structures.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.