The rhythms of language: an overview of linguistic processes and neural oscillations

Abstract

For the last decades neuroscientists have grown interest in the analysis of the rhythmic activity of the brain synchronized at temporal and spatial level. These neural oscillations, grouped by their frequency, have been proposed to govern all cognitive processes. In the field of the neurobiology of language, considerable research has linked speech processing and language comprehension to neural oscillations. On one hand, neural rhythmic activity is thought to synchronize to relevant spectral information of speech on three-time scales – which physically reflect phoneme, syllable and phrase processing. On the other hand, syntactic and semantic processing is subserved by faster oscillatory patterns not necessarily related to the acoustic properties of speech. For each linguistic process, this article summarizes the neural oscillations involved. Further evidence comes from studies on language-related pathologies.

1. Introduction

Neuroscientists using magnetoencephalography (MEG) and electroencephalography (EEG) have greatly relied on event-related potentials (ERPs) and event-related magnetic fields (ERFs) to investigate the major components involved in linguistic processes – a positive or negative deflection of the signal in respect of a baseline – such as P100, N100, P200, P300, N400, P600 (Swaab et al. 2012). Thereby, considerable research has tested the involvement of distinct brain regions and the concomitant electrical/magnetic activity for various linguistic processes, highlighting the spatiotemporal distribution of neural activation. This approach is based on the idea that each stimulus (visual, auditory and so on) a subject is exposed may elicit a time-locked neural response. However, this response cannot be seen in the raw EEG signal, due to the overlapping of ongoing background activity. To determine these systematic responses, a certain experimental paradigm is repeated a number of times: each time a stimulus is presented to the subject, a marker is set into EEG/ERF signal to obtain time-locked epochs to the experimental event of interest: that is, the components previous cited. By averaging all the epochs, only the systematic response should remain (Sauseng and Klimesch 2008).

Recent advancement in neurolinguistic research have seen a shift in paradigm: a central question has become not only which brain region is responsible for which function but also how brain regions interact with each other. In fact, it is necessary to explore “not only what is connected, but how and in what directions regions of the brain are connected” (Kopell et al. 2014, 1319) by adding a functional perspective to understand how the brain’s regions are involved in producing and processing brain signals (Murphy 2015).

Although ERPs/ERFs have proved to be extremely useful, this approach overshadows that the EEG activity of the human brain is not flat and that functions, especially complex ones, involve different areas. In fact, all the electrical activity recorded at the scalp is characterized by rhythms, which are driven by fluctuations in excitability of large-sized populations of neurons, with specific spatiotemporal patterns that differ in amplitude, timing, and frequency (Cohen 2017). Depending on their frequency, such rhythms are grouped in delta (δ: ~0.5–4 Hz), theta (θ: ~4–8 Hz), alpha (α: ~8–12 Hz), beta (β: ~12–30 Hz) and gamma (γ: ~30–120 Hz). Neural oscillations have found wide use in clinical applications, providing useful information about levels of consciousness, psychological states, or presence of neurological disorders.

Recently, researchers have come to realize that these oscillatory rhythms also subserve a wide variety of cognitive processes: it has been argued that the synchronization and desynchronization of these oscillations in distinct clusters can shape input gain and assist information transfer (Akam and Kullmann 2010; Muller et al. 2018). In fact, strong evidence suggests that the reorganization of ongoing oscillatory patterns might explain some of the features of ERPs/ERFs, due to phase reset (i.e., the reshaping of the signal) once a stimulus is presented to the subject (Başar et al. 2001; Başar 2011). Therefore, event-related oscillations, further than to have the time-locked EEG information, permits the retrieval of non-phase locked EEG information related to the cognitive activity induced by the stimulus

As for language, neural oscillations have been linked to a number of linguistic operations. This article provides an overview of neural oscillations subserving linguistic operations. Following Meyer (2018), a dichotomy between lower-level functions of speech processing and higher-level functions of language comprehension will be assumed: on one hand, linguistically meaningful units must be segmented from speech, based on temporal and spectral cues recognized by the auditory system; on the other hand, two streams of language comprehension are assumed to occur to decode the meaning of words (semantic stream) and the relations between words (syntactic stream).

2. The delta-(theta-gamma) neural code for speech processing

2.1. Speech processing on three timescales

By paraphrase of Meyer (2018), the segmentation and identification of discrete phonological units have been found to occur in a particular range of operational frequencies. Phonological units decreasing in granularity hierarchically build speech: the combination of phonemes result into syllables; the combination of syllables result into intonation phrases. Each phonological unit has an acoustics-temporal counterpart (Gussenhoven and Jacobs 2017).

In the last decade, researchers have found that neural oscillations might subserve a set of neural operations that allows the segmentation and identification of discrete phonological units. In fact, during speech processing, three frequency bands, gamma, theta and delta bands, seem to synchronize respectively with the pace of phonemes, syllables and intonational phrases, by tracking linguistically meaningful acoustic properties of speech on three different time scales (Bourguignon et al. 2020; Giraud and Poeppel 2012; Molinaro and Lizarazu 2018). The synchronization of neural oscillations to speech is thought to occur thanks to the socalled neural entrainment which relies on phase synchronization and amplitude synchronization (Obleser and Kayser 2019): on one hand, bottom-up modulations of neural oscillations are stimulus-dependent, relying on acoustic properties of speech; on the other hand, neural oscillations have been found to internally organize, building hierarchical structures, where lower-frequency bands top-down modulate higher-frequency bands, regardless of stimulus properties (Fontolan et al. 2014).

2.2. Stimulus-bound processing

Starting with bottom-up modulations, the phonemic time scale falls within the gamma band frequencies (30-120 Hz). As discussed by Meyer (2018), low and high gamma band oscillations may subserve acoustic and categorical processing, respectively: low gamma-band phase synchronization seems to be related to acoustic processing (Gross et al. 2013), while phonemic-categorical perception is subserved by amplitude synchronization of high gamma bands (Lehongre et al. 2011), which reflect the spiking activity of neurons in the auditory cortex sensitive to phonemes (Mesgarani et al. 2014). In addition, it has been argued that low gamma band synchronization occurs more strongly with the acoustic amplitude envelope compared to phonemic-categorical information (Di Liberto, O’Sullivan, and Lalor 2015).

Going up in granularity, theta bands oscillations (4-8 Hz) capture the pace of syllables, thus subserving syllabic processing. In fact, recent works have suggested that theta oscillations phase-synchronize to the onset of syllables, allowing the segmentation of syllables (Luo and Poeppel 2007; Howard and Poeppel 2012; Peelle, Gross, and Davis 2013; Doelling et al. 2014; references from Meyer 2018). Moreover, further evidence suggests a relationship between the amplitude modulations of speech and the phase of neural oscillations (Gross et al. 2013; Vander Ghinst et al. 2016; Molinaro, Monsalve, and Lizarazu 2016).

Lastly, delta bands (0.5-4 Hz) have shown increased phase coherence to the fundamental frequency envelope of speech: delta oscillations have been proposed to aid the segmentation of intonational phrases (Giraud and Poeppel 2012), due to the amplitude extrema of the pitch contour marking the boundaries of intonational phrases. Interestingly, delta bands also capture the pace of syntactic phrases (Ding et al. 2016; Molinaro and Lizarazu 2018), which do not have a direct physical counterpart, in the case prosodic cues were explicitly removed (Ding et al. 2017). However, the role of delta oscillations in speech processing is still under debate (see for example Boucher, Gilbert, and Jemel 2019).

2.3. Top-down modulations

While strong evidence suggests bottom-up modulations of neural oscillations, it has been argued that these oscillations hierarchically self-organize, regardless of acoustic properties of speech: particularly, the phase of lower-frequency bands top-down modulates the amplitude of higher-frequency bands (Giraud and Poeppel 2012; Fontolan et al. 2014), opening a new window on brain dynamics of speech processing. In fact, theta-gamma cross frequency coupling in the left hemisphere have been proposed to subserve the concatenation of phonemes into syllables (Canolty et al. 2006), although this hypothesis contrasts with a number of studies that show a theta-gamma coupling in the right auditory cortex (Luo and Poeppel 2007; Abrams et al. 2008; Hämäläinen et al. 2012; Gross et al. 2013; Howard and Poeppel 2012; Peelle, Gross, and Davis 2013; references from Meyer 2018). Interestingly, it has been shown that phase-amplitude coupling between theta and gamma oscillations adapts to speech rate (Lizarazu, Lallier, and Molinaro 2019). The combination of syllables into intonational phrases is then subserved by delta-theta cross-frequency coupling (Giraud and Poeppel 2012; Ding et al. 2016). It is worth noticing that the neural underpinnings of prosody are still not clear and need future research (for a discussion see Myers, Lense, and Gordon 2019), given that, at least in some cases, prosody conveys crucial information on the syntactic structure which suggests a tight relation with top-down information.

3. Language comprehension along two streams

3.1. Syntactic processing

Once all phonological units are segmented from speech and the auditory system is temporarily aligned, the brain must decode the relations between words which are recursively combined into syntactic phrases (Chomsky 1957). Recent evidence suggests that the grouping of words into phrases might be subserved by delta band cycles through phase resetting. In fact, Ding and colleagues (2016) have found an increase in delta-power associated with internal syntactic phrase generation. These findings were later confirmed by Bonhage et al. (2017): subjects involved in this study showed an increase in delta band power while exposed to a list of words that could be combined into syntactic phrases, while a decrease in delta band power was found for a list of words that could not be grouped into phrases. As mentioned above, delta band oscillations also play a role in the segmentation of intonational phrases. Interestingly, these results may not contradict each other: in fact, Ghitza (2017) argued that the relation between delta bands and intonational phrases would reflect a stimulus-bound bottom-up segmentation, while delta oscillations subserving syntactic chunking would reflect top-down generation based on a priori syntactic knowledge. However, the role of delta bands in syntactic phrase generation is still under debate. While a sentence unfolds word by word, phrases have to be stored in verbal working memory and retrieved later on to assess their dependencies with other phrases and generate syntactic hierarchies. A number of studies have linked the storage of phrases in verbal working memory with an increase in alpha band activity (Haarmann and Cameron 2005; Weiss et al. 2005; Meyer, Obleser, and Friederici 2013; Bonhage et al. 2017; references from Meyer 2018). Particularly, alpha band power increases with storage demands, localized in the inferior parietal cortex.

Interesting findings come from violation studies that examined time-locked neural oscillations related to syntactic anomalies. Many of these studies have found a consistent pattern in response to syntactic violations, such as gender and number agreement violation (Bastiaansen, van Berkum, and Hagoort 2002; Davidson and Indefrey 2007; Schneider et al. 2016), mostly confirming the aforementioned findings.

A groundbreaking result concerns the possible role of gamma band in structure-building operations, what generative linguists call Merge: (Nelson and colleagues (2017) found a specific gamma band pattern that they claim as evidence for Merge, the binding. Particularly, gamma power increases every time a new word is added to an unfolding sentence, while it sharply decreases when words can be compressed into a syntactic node. In addition, a recent study has shown a difference in high gamma response for the syntactic disambiguation of homophones phrases (Artoni et al. 2020). However, these results may be in contrast with aforementioned studies linking delta activity to syntactic processing.

3.2. Semantic processing

Along the syntactic parsing, language comprehension also implies a semantic processing which has been linked primarily with beta and gamma oscillations. Evidence of beta-bands involvement in semantic processing have come from a number of violation studies, focused on semantic anomalies (Kielar et al. 2014; 2015; Wang et al. 2012a; Luo et al. 2010). Particularly, these studies have found a decrease both in alpha and beta oscillations related to semantic anomalies. Willems, Oostenveld, and Hagoort (2008) have linked both alpha and beta decrease to audio-visual semantic anomalies. However, they found that alpha activity decreases where both a visual and linguistic context mismatch occurs.

Interestingly, other research on semantic anomalies has also found an increase in theta power (Hagoort et al. 2004; Hald, Bastiaansen, and Hagoort 2006; Davidson and Indefrey 2007; M. Bastiaansen and Hagoort 2015; Wang, Zhu, and Bastiaansen 2012; references from Prystauka and Lewis 2019). Bastiaansen, Mazaheri, and Jensen (2012) proposed that theta power increase due to semantic anomalies might reflect the integration of the anomalous word into the sentence. Another interesting proposal comes from Prystauka and Lewis (2019): given that theta increase has also been found in syntactic violation studies and has been proposed to aid lexical-semantic retrieval (Bastiaansen, Mazaheri, and Jensen 2012; Marko, Cimrová, and Riečanský 2019), an increase in theta power may reflect a general error detection mechanism. This idea is also supported by other studies on incorrect solutions to mathematical equations (Tzur and Berger 2007) or motor error in reaching a task (Arrighi et al. 2016).

However, violation studies do not give the full picture: semantic processing also relies on predictions of upcoming words. Each word is stored in the long-term memory with a certain probability of occurring in a given context, prior and after other words (Hagoort et al. 2004; Kutas and Federmeier 2010). Top-down predictions, independent of stimuli, have also been linked with beta bands power which increases when expectations of upcoming words are confirmed and decreases when such predictions do not match the sequence of incoming words (Lewis and Bastiaansen 2015; Lewis et al. 2016). For example, Wang et al. (2012) performed a cloze test, finding a beta power decrease in sentence ending that did not matched expectations. These findings were further confirmed by Lewis et al. (2017) that compared short stories of sentences, observing an increase in beta power in semantically coherent stories and a decrease in beta power for semantically incongruent stories. An interesting proposal about the role of beta bands and prediction comes from Lewis et al. (2016), yielding that they might subserve predictions across different linguistic levels, from the auditory domain to the syntactic level (Kim and Chung 2008; Sabine Weiss and Mueller 2012; Arnal, Wyart, and Giraud 2011; Arnal and Giraud 2012). However, Meyer (2018) argues that beta bands only subserve lexical-semantic predictions for two reasons: beta bands power increase during contextual prediction of upcoming words correlates with the amplitude of N400, indicating the lexical-semantic predictability of a word (Kutas and Federmeier 2010; Wang et al. 2012; Hale 2016; Lewis et al. 2016) but not its syntactic category (Levy 2008; Frank et al. 2015); beta-bands have been shown to be modulated by syntactic factors only in syntactic violation studies (e.g.: the syntactic category of the upcoming words do not match expectations), possibly yielding that semantic processing does not occur when syntactic parsing is not accomplished (Steinhauer and Drury 2012).

Alongside top-down predictions based on the occurrence frequency of each word in context stored in long-term memory, lexical-semantic representations of incoming words must be checked. When lexical-semantic predictions are fulfilled, gamma power has been found to increase (Wang, Zhu, and Bastiaansen 2012; Molinaro, Barraza, and Carreiras 2013; Monsalve, Pérez, and Molinaro 2014). Conversely, gamma power decreases when the incoming word does not match expectations (Hald, Bastiaansen, and Hagoort 2006; Penolazzi, Angrilli, and Job 2009; Rommers, Dijkstra, and Bastiaansen 2012; references from Meyer 2018).

The interplay between beta and gamma bands has been included in the predictive coding framework (e.g. Friston 2005). In fact, data collected on beta and gamma bands are compatible with the predictive coding framework (Lewis and Bastiaansen 2015; for a discussion see Meyer 2018; Prystauka and Lewis 2019).

4. Language-related disorders and neural oscillations

Further evidence of the implications of neural oscillations into linguistic operations comes from research on language-related pathologies and disorders. In fact, a number of studies has confirmed the aforementioned findings on linguistic operations and neural oscillations.

Current models of aphasia classifications still rely on the Wernicke-Lichtheim model which links damages in a brain area with a specific function. This model has the advantage of being simple: for example, a damage in the motor area of language, Broca’s area, will be linked to a nonfluent aphasic syndrome, while damage in the sensory area of language, Wernicke’s area, will be linked to fluent aphasia syndrome (Lichteim 1885; Wernicke 1974).

A number of studies have observed an alteration of neural oscillations both at resting state and while performing a linguistic task. Spironelli and Angrilli (2009), for example, demonstrated that an increase in delta amplitude in the perilesional area is a marker of brain damage in chronic non-fluent aphasic patients. In fact, delta band might be an index of neural inhibition. Other research has shown that focal lesions in the left hemispheric language regions may lead to a change in brain physiology. For example, Meinzer et al. (2004) found an increase in spontaneous delta activity in the perilesional area in a group of stroke patients suffering from different aphasia types, while also reporting a decrease in spontaneous perilesional delta activity after an intense speech and language therapy. Dubovik et al. (2012) also found a shift from fast to slow spontaneous neural oscillations, particularly in delta and theta frequency range. Interestingly, Nicolo et al. (2015) reported that more coherent beta oscillations in lesioned Broca’s area in early post stroke recovery patients predicts future language improvement during recovery.

Functional restoration of the brain in post-stroke patients seems to be related to an increase in spontaneous alpha-band synchronization (Westlake et al. 2012; Dubovik et al. 2012). Moreover, Kielar et al. (2016) investigated the functional reorganization of language networks: particularly, in a group of subjects suffering from different types of aphasia, they found a decrease in alpha and beta power in the left hemisphere, where the lesion occurred, in response to semantic anomalies during sentence comprehension.

A previous study also reported a possible role of beta activity as an index of the reorganization of language networks in aphasic patients: Spironelli, Manfredi, and Angrilli (2013) reported that non-fluent aphasic subjects, after linguistic recovery, showed a reduced beta activation in the core damaged area during a phonological and semantic task, while also showing an increased delta activity compared to healthy control subjects. They also found an increased high beta-activity in the left anterior sites during the phonological and orthographic task. The authors of the study have interpreted these findings as an index of the reorganization of language in recovered non-fluent aphasic subjects at the left prefrontal sites.

In addition, many studies on dyslexic subjects have confirmed the role of neural oscillation into linguistic operations. For example, Lehongre et al. (2011) linked a decreased entrainment of the lower gamma band to phonological deficits, while Leong and Goswami (2014) suggested that rhythmic entrainment at the syllabic time-scale is disrupted in dyslexic subjects. These results were recently confirmed by Lizarazu et al. (2021), proving an impairment of cortical entrainment in the delta and theta range to speech in dyslexic subjects.

7. Conclusions

Although this article is far from offering an exhaustive overview, it is clear that neural oscillations provide a new window on brain dynamics related to linguistic operations. The number of studies following the oscillation-based frame-work has been growing in the last decade, yielding an increasing interest in brain’s oscillatory nature. On one hand, speech processing seems to be subserved by delta, theta and gamma bands, respectively at phrase, syllable and phoneme timescales. On the other hand, language comprehension is subserved by a variety of frequency bands involved in syntactic and semantic processing, including more general cognitive functions such as the implication of short- and long-term memory.

Bottom-up and top-down modulations of neural oscillations may provide a neural code for linguistic operations: the cyclicity of oscillatory rhythms’ synchronization and desynchronization may represent a neural coding (and decoding) that matches linguistic computations, shortening the gap between broader neuroscientific investigations and more fine-grained linguistic investigations (Granularity Mismatch Problem, (Embick and Poeppel 2015). Accordingly, formal proposals of hierarchical organization of neural oscillations have emerged (Murphy 2019; Grimaldi 2019).