Alpha oscillations in left perisylvian cortices support semantic processing and predict performance

Abstract

Semantic processing is the ability to discern and maintain conceptual relationships among words and objects. While the neural circuits serving semantic representation and controlled retrieval are well established, the neuronal dynamics underlying these processes are poorly understood. Herein, we examined 25 healthy young adults who completed a semantic relation word-matching task during magnetoencephalography (MEG). MEG data were examined in the time–frequency domain and significant oscillatory responses were imaged using a beamformer. Whole-brain statistical analyses were conducted to compare semantic-related to length-related neural oscillatory responses. Time series were extracted to visualize the dynamics and were linked to task performance using structural equation modeling. The results indicated that participants had significantly longer reaction times in semantic compared to length trials. Robust MEG responses in the theta (3–6 Hz), alpha (10–16 Hz), and gamma (64–76 Hz and 64–94 Hz) bands were observed in parieto-occipital and frontal cortices. Whole-brain analyses revealed stronger alpha oscillations in a left-lateralized network during semantically related relative to length trials. Importantly, stronger alpha oscillations in the left superior temporal gyrus during semantic trials predicted faster responses. These data reinforce existing literature and add novel temporal evidence supporting the executive role of the semantic control network in behavior.

1. Introduction

Semantic processing is the ability to discern conceptual relationships and understand information relating to language including words, objects, and concepts presented both verbally and nonverbally. This ability allows us to comprehend a variety of complex stimuli and interact with the world around us. Semantic processing is not only critical for language function, but also the ability to recognize objects and their conceptual relationships during activities of daily living (Kindell et al. 2014). When putting on shoes, for instance, you must first identify the object as a shoe and then recognize its relationship with a foot in order to know what to do with the shoe. Consequently, impairments in semantic processing, in dementias and aphasias for example, can greatly impact quality of life.

Semantic knowledge can be probed by requiring participants to passively view words or sentences with different levels of semantic association or congruency (McDonald et al. 2010; Mollo et al. 2018; Mamashli et al. 2019) or by having participants make an active semantic judgment pertaining to the presented stimuli. Common tasks for semantic judgment include responding if the presented stimulus is a word or nonword (Almeida and Poeppel 2013; Zhu et al. 2015; Della Rosa et al. 2018), judging if 2 words are semantically related (Teige et al. 2018), matching a picture with its correct descriptor (Thompson-Schill et al. 1997; Davey et al. 2015), matching 2 semantically related words (Thompson-Schill et al. 1997; Wagner et al. 2001; Badre et al. 2005; Hoffman et al. 2010), or judging if a sentence is semantically accurate (Kielar et al. 2015, 2016, 2018).

The dominant model in this area suggests that semantic processing is composed of at least 2 separate processes, conceptual representation and executive control, which regulates the activation and selection of supporting semantic areas in a stimulus-appropriate manner (Jefferies and Lambon Ralph 2006; Jackson et al. 2015; Lambon Ralph et al. 2017). The networks and brain regions involved in each of these processes are well established in the literature. Activation of the bilateral anterior temporal lobes (ATLs) is primarily responsible for the conceptual representation portion of semantic knowledge (Jefferies and Lambon Ralph 2006; Jackson et al. 2015; Lambon Ralph et al. 2017; Mollo et al. 2017; Rice et al. 2018), although other regions are believed to be extensively involved. Specifically, a hub-and-spoke model has been proposed for neural activation serving conceptual representation (Lambon Ralph et al. 2017), with the ATL acting as the multimodal “hub” that integrates information from multiple modality-specific “spokes” (e.g. primary and secondary sensorimotor cortices for action information, fusiform gyrus for color knowledge, posterior inferolateral temporal regions for motion information, etc.; Visser et al. 2010; Lambon Ralph et al. 2017; Chen et al. 2020). In theory, when presented with a stimulus, numerous modality-specific representations (“spokes”) will be activated such as its shape, name, motion, etc. These areas will then communicate with the ATL as a “hub” to form multimodal associations and representations (Lambon Ralph et al. 2017; Chen et al. 2020).

Other areas including the left posterior temporal cortex, left inferior parietal cortices, and left lateral prefrontal cortex have been implicated in the executive control aspects of semantic processing (Jefferies and Lambon Ralph 2006; Badre and Wagner 2007; Noonan et al. 2009; Lambon Ralph et al. 2017; Methqal et al. 2017; Palacio and Cardenas 2019). This semantic control network has a less clear role in semantic processing. Some theories have proposed that these areas may contribute to controlled retrieval of relevant information (Wagner et al. 2001), while others have proposed that the network is more involved in selection of appropriate representations in order to resolve competition (Thompson-Schill et al. 1997). Both theories have been empirically supported, which may indicate that the semantic control network plays a role in both controlled retrieval and selection (Badre et al. 2005). Further, this overall 2-process framework has been widely supported by neuropsychological and neuroimaging research on patients with semantic dementia and lesions secondary to strokes (Jefferies and Lambon Ralph 2006; Lambon Ralph et al. 2006; Noonan et al. 2009; Rogers et al. 2015; Thompson et al. 2017, 2018).

Most studies of semantic processing to date have used functional magnetic resonance imaging (fMRI) and focused on identifying the anatomical correlates of these semantic networks. Studies that have utilized more temporally precise methods, such as electroencephalography (EEG), have shown robust alpha oscillations (i.e. event-related desynchronizations [ERD], or reductions in power relative to baseline) during semantic tasks, although these studies have generally not attempted to localize such responses beyond noting their left hemispheric laterality (Klimesch et al. 1994, 1997; Röhm et al. 2001; Maguire et al. 2010). Such left-lateralized alpha oscillations have also been shown in verbal memory tasks and are thought to represent active processing of the presented stimuli (van Dijk et al. 2008; Proskovec et al. 2016, 2019; Wiesman et al. 2016; Embury et al. 2019; Koshy et al. 2020). Previous magnetoencephalography (MEG) studies of semantic processing have shown more varied responses in alpha and beta bands but have largely focused on early responses to semantic stimuli (Cornelissen et al. 2009; McDonald et al. 2010; Teige et al. 2018). Thus, very few studies have examined the later, sustained neural responses involved in semantic decisions.

In this study, we utilized MEG to quantify the oscillatory dynamics serving semantic processing to better understand its spatial, temporal, and spectral complexities. Twenty-nine healthy young adults performed a word-matching task during MEG recording. In one condition, participants were instructed to match target words based on their semantic-relatedness to the probes, and in the other condition, participants matched based on the length of the word. We applied advanced oscillatory analysis methods to identify differences in neural activity during the semantic and nonsemantic conditions and extracted voxel time series data from peaks exhibiting condition-wise differences at the whole-brain level to visualize the dynamics. We also examined associations between neural oscillatory activity and task behavior using structural equation modeling. We hypothesized that spectrally specific oscillations in a left-lateralized network would be stronger during semantic versus nonsemantic trials of the language processing task and that the spatiotemporal dynamics would predict behavioral performance on the task.

2. Methods

2.1. Participants

We studied 29 healthy young adults; 4 were excluded due to artifactual MEG data and/or technical problems during data acquisition. The remaining 25 participants included 13 males (mean age: 23.76 years, range: 19–29 years, 23 right-handed). Exclusionary criteria included any medical illness affecting the CNS, any neurological or psychiatric disorder, history of head trauma, current substance abuse, and the MEG Laboratory’s standard exclusion criteria (e.g. ferromagnetic implants). The University of Nebraska Medical Center’s Institutional Review Board (IRB) approved this study protocol. Written informed consent was obtained from each participant after a full description of the study.

2.2. Experimental paradigm

Participants were seated in a nonmagnetic chair within a magnetically shielded room. During MEG recording, participants performed a semantic processing task (Fig. 1). Participants were instructed to maintain fixation on a centrally located crosshair throughout the task. Each trial began with presentation of a crosshair for 2,000 ms (±200 ms). Next, 3 words were presented for 2,500 ms, 1 at the top of the screen (target) and 2 at the bottom (probe) to the left and the right of the target word, and participants were instructed to respond to 1 of 2 task conditions. In the semantic condition (“semantic”), participants were instructed to identify which probe word was related semantically to the target word with their right index or middle finger (left and right presented probes, respectively). For the control condition (“length”), participants were instructed to identify the probe word that matched the length of the target word and respond with their right index or middle finger for left and right presented probe words, respectively. Both conditions used the same body of words and all words were balanced for frequency across probes and targets, although the target/probe sets were different between conditions to ensure conditional differences were not biased by memory effects (i.e. the participants remembering the pairs from the first block). The blocked task conditions were counterbalanced across participants, such that half of participants completed the semantic and then the control condition, while the others completed the blocks in reverse order. All participants completed a brief practice session of the task prior to the MEG recording to ensure adherence to the task instructions. Each trial lasted 4,500 ms (±200 ms) and there were a total of 230 trials (i.e. 115 length and 115 semantic trials), resulting in a total run time of approximately 18.5 min.

Fig. 1.

Semantic processing task and behavioral performance. A fixation cross was presented for 2,000 ms (±200 ms). Next, 3 words were presented for 2,500 ms, 1 at the top of the screen and 2 at the bottom. In the length condition, participants responded as to which of the bottom words had the same number of letters as the top target word. Participants responded with their right index finger if the word on the left matched or with their right middle finger if the word on the right matched. In the semantic condition, participants responded as to which of the bottom words’ meaning was most closely associated with the top target word using the same fingers. RTs were significantly different between length and semantic conditions such that participants responded faster in the length condition. Accuracy did not differ between conditions. ^*P < 0.01. Error bars reflect one standard error of the mean.

2.3. MEG data acquisition

All recordings were performed in a magnetically shielded room with active shielding engaged to compensate for environmental noise. Neuromagnetic responses were sampled continuously at 1 kHz with an acquisition bandwidth of 0.1–330 Hz using an MEGIN/Elekta MEG system (Elekta, Helsinki, Finland) with 306 magnetic sensors, including 204 planar gradiometers and 102 magnetometers. Participants were monitored throughout data acquisition using a real-time audio–video feed from inside the magnetically shielded room. Data from each participant were individually corrected for head motion and subjected to noise reduction using the signal space separation method with a temporal extension (Taulu et al. 2005; Taulu and Simola 2006).

2.4. Structural MRI processing and MEG coregistration

Prior to MEG measurement, 4 coils were attached to the subject’s head and localized, together with the 3 fiducial points and scalp surface, with a 3D digitizer (Fastrak 3SF0002; Polhemus Navigator Sciences, Colchester, VT). Once the subject was positioned for MEG recording, an electric current with a unique frequency label (e.g. 322 Hz) was fed to each of the coils. This induced a measurable magnetic field and allowed each coil to be localized in reference to the sensors throughout the recording session. Since coil locations were also known in head coordinates, all MEG measurements could be transformed into a common coordinate system. With this coordinate system, each participant’s MEG data were co-registered with their structural MRI data prior to source space analyses using BESA MRI (Version 2.0; BESA GmbH, Gräfelfing, Germany). All structural MRI data were aligned parallel to the anterior and posterior commissures and transformed into standardized space. Following source analysis (i.e. beamforming), each participant’s 4.0 × 4.0 × 4.0 mm MEG functional images were also transformed into standardized space using the transform that was previously applied to the structural MRI volume and spatially resampled.

2.5. MEG time–frequency transformation and statistics

Cardiac and ocular artifacts (e.g. blinks, eye movement) were removed from the data using signal-space projection, which was accounted for during source reconstruction (Uusitalo and Ilmoniemi 1997). The continuous magnetic time series was divided into epochs of 4,000 ms duration, with the stimulus presentation defined as 0 ms and the baseline being defined as −600 to −100 ms prior to the onset of the stimuli. Epochs containing artifacts were rejected based on individual amplitude and gradient thresholds surpassing 2.5 median absolute deviations, supplemented with visual inspection. The average amplitude threshold was 1055.15 (SD = 349.32) fT and the average gradient threshold was 202.00 (SD = 144.91) fT/s across all participants. Only trials where participants responded correctly were used for analysis. On average, 96 (SD = 7.4) semantic and 98 (SD = 5.4) length trials remained after artifact rejection and were used in subsequent analyses. Importantly, the number of trials did not significantly differ between the length and semantic conditions (P = 0.34).

Artifact-free epochs were transformed into the time–frequency domain using complex demodulation and the resulting spectral power estimations per sensor were averaged over trials to generate time–frequency plots of mean spectral density (Kovach and Gander 2016). These sensor-level data were normalized per time–frequency bin using the respective bin’s baseline power, which was calculated as the mean power during the −600 to −100 ms time period.

The specific time–frequency windows used for imaging were determined by statistical analysis of the sensor-level spectrograms across all participants and all trials, restricted to the entire array of gradiometers. To reduce the risk of false-positive results while maintaining reasonable sensitivity, a 2-stage procedure was followed to control for Type I error. In the first stage, paired-samples t-tests against baseline were conducted on each data point and the output spectrogram of t-values was thresholded at P < 0.05 to define time–frequency bins containing potentially significant oscillatory deviations relative to baseline across all participants and conditions. In Stage 2, time–frequency bins that survived the threshold were clustered with temporally and/or spectrally neighboring bins that were also above the (P < 0.05) threshold, and a cluster value was derived by summing all the t-values of all data points in the cluster. Nonparametric permutation testing was then used to derive a distribution of cluster-values, and the significance level of the observed clusters (from Stage 1) was tested directly using this distribution (Ernst 2004; Maris and Oostenveld 2007). For each comparison, 1,000 permutations were computed to build a distribution of cluster values. Based on these analyses, only the time–frequency windows that contained significant oscillatory events across all trials were subjected to imaging, and these source images were then used to test our hypothesized effects.

2.6. MEG source imaging and statistics

Cortical responses were imaged through the dynamic imaging of coherent sources beamformer (Gross et al. 2001), which applies spatial filters to time–frequency sensor data to calculate voxel-wise source power for the entire brain volume. The regularization parameter was set to 0.0001%, which enables spatially selective image reconstructions and allows for estimation of numerous sources in the brain. The single images are derived from the cross spectral densities of all combinations of MEG gradiometers averaged over the time–frequency range of interest, and the solution of the forward problem for each location on a grid specified by input voxel space. Such images are typically referred to as pseudo-t maps, with units (pseudo-t) that reflect noise-normalized power differences (i.e. active vs. passive) per voxel. Following convention, we computed noise-normalized, source power per voxel in each participant using active (i.e. task) and passive (i.e. baseline) periods of equal duration and bandwidth (Hillebrand et al. 2005). MEG preprocessing and imaging used the Brain Electrical Source Analysis (BESA V7.0) software.

Normalized differential source power was computed for the statistically determined time–frequency bands over the entire brain volume per participant at 4.0 × 4.0 × 4.0 mm resolution. The resulting 3D maps of brain activity were averaged across all participants and both conditions to assess the anatomical basis of the significant oscillatory responses identified through the sensor-level analysis.

2.7. Statistical and virtual sensor analyses

To identify the neuroanatomical basis of semantic relative to length processing, we performed whole-brain paired samples t-tests (i.e. semantic vs. length) on each of the beamformer maps which corresponded to the oscillatory responses of interest, as determined by the sensor-level analyses. To correct for multiple comparisons, the resulting images underwent nonparametric permutation testing using a cluster-based permutation method similar to that performed on the sensor-level spectrograms, with 10,000 permutations per comparison (threshold of P < 0.005; 2-tailed).

To visualize the dynamics of oscillatory power at regions exhibiting significant condition-wise effects, voxel time series (i.e. “virtual sensors”) were extracted from each participant’s data using the peak voxel in each condition-wise difference cluster. To compute the virtual sensors, we applied the sensor weighting matrix derived through the forward computation to the preprocessed signal vector, which yielded a time series for the specific coordinate in source space. Using these data, we computed the relative (i.e. baseline-corrected) response time series for each participant per task condition.

Finally, we used a multiple regression approach (i.e. structural equation modeling) to interrogate the predictive capacity of neural oscillations for processing semantically related words on task performance. Specifically, the relative power time series extracted from peak condition-wise difference voxels were used to compute a difference score of neural oscillatory activity (semantic – length). This difference score was computed for each identified peak and for task performance (i.e. reaction time [RT]). The neural difference scores were modeled as simultaneous predictors of behavioral difference scores in reaction time. All model parameters, including correlations between each of the neural difference scores, were freely estimated. Modeling was conducted in Mplus (Version 8.1).

3. Results

3.1. Behavioral analysis

Participants performed well on the task responding accurately to 99.16% (SD = 0.92%) of the length trails and 99.30% (SD = 0.83%) of the semantic trials. There was not a significant difference in accuracy between conditions (t[24] = 0.68, P = 0.505). Interestingly, there was a significant difference in RTs between conditions (t[24] = 2.95, P = 0.007), such that participants responded faster to length trials (mean RT = 1351.79 ms; SD = 291.60 ms) relative to semantic trials (mean RT = 1493.88 ms; SD = 287.92 ms; Fig. 1). From this, a difference score (i.e. semantic – length) was computed per participant to index the behavioral slowing demonstrated when processing semantically related words compared to their nonsemantically related matches (M_difference = 142.09 ms, SD_difference = 240.53 ms). Importantly, this difference score was used for subsequent analyses interrogating brain–behavior relationships (below).

3.2. Sensor-level analysis

Statistical analysis of the time–frequency spectrograms across all trials, irrespective of condition, revealed significant clusters of theta (3–6 Hz), alpha (10–16 Hz), and gamma (64–76 Hz and 64–94 Hz) oscillatory activity in gradiometers near occipital, parietal, and frontal cortices across all participants and conditions (Fig. 2; P < 0.05, corrected). Importantly, we focused subsequent source analyses on the timeframe between stimulus onset and average RT, as we were interested in the semantic processing of the stimuli and not movement-related activity. Specifically, a strong increase in theta activity (3–6 Hz) was detected shortly after stimulus onset and this lasted for ~350 ms (i.e. 0–350 ms) in gradiometers near the visual cortex (Fig. 2). In addition, a more temporally sustained theta (3–6 Hz) response was observed in frontal sensors and occurred from 350 to 850 ms following stimulus onset. In contrast, strong decreases in alpha power (i.e. alpha ERD; 10–16 Hz) were detected in the 200–1,200 ms time window in gradiometers near the visual cortices. Importantly, this response was divided into early (200–700 ms), middle (450–950 ms), and later (700–1,200 ms) bins to enable each of these time windows of neural processing to be resolved and probed for differences between semantic and length conditions. Further, we used a sliding window approach to image the evolution of alpha responses to maximize the amount of data included in each image and thereby improve the signal-to-noise ratio (SNR) and spatial precision (Fig. 2). Finally, we observed a transient, broadband increase in gamma activity (64–94 Hz) shortly after stimulus onset (i.e. 100–200 ms) in posterior parieto-occipital gradiometers, as well as a more sustained increase in gamma activity (64–76 Hz) in parieto-occipital sensors from ~200 to 700 ms. All of these time–frequency bins were significant at P < 0.05, corrected.

Fig. 2.

Sensor-level analysis of oscillatory strength. MEG time–frequency spectrograms with time (ms) on the x-axis and frequency (Hz) on the y-axis. The onset of the stimuli occurred at 0 ms. Power is shown in percentage change relative to the baseline period (−600 to −100 ms), with color scale bars beneath each panel. Data have been averaged across all trials, including both conditions and all participants. We found robust transient increases in the theta (3–6 Hz; bottom panel) and gamma (64–94 Hz; top panel) range at the onset of the stimuli in spectrograms near the occipital cortices. In addition, sustained increases in theta after initial target onset (i.e. 350–850 ms; middle panel) were observed in a cluster of spectrograms near the left prefrontal cortex and in occipital cortices in the gamma range (64–76 Hz, 200–700 ms; top panel). Finally, strong, temporally sustained alpha oscillations (i.e. decreases in power or ERD) were seen shortly after stimulus onset and were divided into early (200–700 ms), mid (450–950 ms), and late (700–1,200 ms) responses. White boxes denote the time–frequency windows used for source imaging analyses. 2D power topographies across the whole array averaged over the time window enclosed in the white box are shown to the right of each spectrogram. All time–frequency windows were significant at P < 0.05, corrected. θ = theta. α = alpha. γ = gamma.

3.3. Beamformer analysis

To identify the brain regions generating the significant sensor-level oscillations, these time–frequency windows were imaged using a beamformer. The resulting maps were grand averaged across all participants and conditions (Fig. 3). Increases were observed in theta power (3–6 Hz) from 0 to 350 ms in bilateral medial occipital cortices and from 350 to 850 ms in the left inferior frontal cortex. In contrast, strong decreases in alpha power (i.e. alpha ERD; 10–16 Hz) were observed in lateral occipital cortices bilaterally during early, middle, and late processing windows. Finally, both transient and sustained (64–94 Hz, 100–200 ms; 64–76 Hz, 200—700 ms) increases in gamma power were observed in bilateral medial occipital cortices.

Fig. 3.

Source-level oscillatory activity during target processing. The beamformer images (pseudo-t) were averaged across all participants and conditions for each time–frequency window determined in our sensor-level analyses. Early, transient increases in the theta band were seen in bilateral primary visual cortices, while sustained increases were seen in the left inferior frontal cortex and bilateral primary visual. In addition, increases in gamma activity were observed during early and late target processing in bilateral visual cortices. In contrast, decreases in alpha power (i.e. alpha ERD) were observed in lateral occipital cortices bilaterally throughout early, mid, and late temporal windows. θ = theta. α = alpha. γ = gamma.

3.4. Conditional differences on alpha oscillatory responses

Whole-brain paired-sample t-tests per oscillatory response were conducted to identify the neural regions involved in semantic processing compared to control (i.e. length) processing. These tests revealed significant differences in the late alpha time window prior to response onset (i.e. 700–1,200 ms), such that stronger decreases in alpha power from baseline (i.e. greater alpha ERD) were observed in a left-lateralized network during semantic processing relative to the control condition (i.e. length judgment). Specifically, stronger alpha oscillations (i.e. greater decreases from baseline) in the 700–1,200 ms time window were observed in the left pars orbitalis, left inferior frontal gyrus (IFG), and left posterior superior temporal gyrus (STG; P < 0.005, corrected; Fig. 4). Peak voxel time series were extracted from each of these regions to visualize the evolution of left-lateralized alpha dynamics (Fig. 4). Since participants responded with their right hand and the time course of these alpha responses likely overlapped with motor preparation, we performed additional control analyses to ensure left-lateralized motor responses did not contribute to the purportedly semantic-related alpha responses. These analyses are described in the Supplemental Materials and showed that movement-related alpha and beta responses were centered on the left precentral gyrus hand region and were much more dorsal and medial (see Supplementary Fig. S1) than the semantic-related alpha responses identified in our primary analyses. Finally, there were no other conditional differences observed in any theta, gamma, or additional alpha (i.e. early and middle) maps that survived our rigorous multiple comparisons correction approach.

Fig. 4.

Conditional effects in late alpha activity. (Left): Whole-brain paired-sample t-tests on the late alpha window (i.e. 700–1,200 ms) revealed significantly stronger oscillations (i.e. greater decrease from baseline) during semantic relative to length processing in left-lateralized language regions including the left pars orbitalis, left inferior frontal gyrus, and left superior temporal gyrus (P < 0.005, corrected). (Middle): Baseline-corrected virtual sensors were extracted from each region exhibiting the significant conditional effect to visualize the temporal dynamics. Time series denote time (in ms) on the x-axis and relative power (%) on the y-axis. Progressively stronger decreases in alpha power were seen in these regions following stimulus onset and peaking about 1,150–1,200 ms. (Right): Average relative power by condition during the late window (700–1,200 ms). Error bars reflect one standard error of the mean.

Finally, we evaluated whether alpha activity in this left-lateralized network predicted changes in task performance. As described in Section 2.7, we computed neural difference scores (semantic – length) from the relative (i.e. baseline-corrected) time series extracted from each of the 3 regions exhibiting significant condition-wise effects (i.e. left pars orbitalis, IFG, and STG). The 3 regions were modeled as predictors of behavioral difference scores in RT (semantic – length). Collectively, these alpha difference scores accounted for 44% of the variance in RT differences (P = 0.003). The left STG was the strongest predictor and was the only one to reach statistical significance (β = 0.504, P = 0.038; Fig. 5). Essentially, this suggests that stronger alpha oscillations (i.e. greater decreases from baseline) in the left STG during the processing of semantic relative to length trials predict faster RTs on semantic trials. The association between the activity in the left pars orbitalis and task behavior trended in the same direction, though it was not statistically significant in the current sample (β = 0.300, P = 0.137). Importantly, these results suggest that when alpha oscillations in left-lateralized language networks are similar during the processing of both semantically and nonsemantically matched words, behavioral decrements are present for trials where a conceptual match is required (i.e. slower RT for semantic trials).

Fig. 5.

Left-lateralized alpha activity modulates behavioral performance. (Left): Conceptual model probed using structural equation modeling. (Top left): We hypothesized that stronger alpha oscillations (i.e. greater ERD) in left-lateralized language areas during the processing of semantically related targets would lead to equalized behavior between task conditions. In other words, the additional behavioral “cost” of semantic over length processing would be reduced in the context of stronger alpha oscillations. (Bottom left): Likewise, we hypothesized that more equalized neural responses in the alpha range during semantic and length trials would lead to a greater behavioral cost in RT on the task. (Right): Structural equation model of task-related alpha activity in left-lateralized language regions on the RT difference (semantic – length). Solid lines with single-headed arrow denote significant predictive paths, while dashed lines denote nonsignificant paths. Solid curved double-headed arrows indicate significant correlations between variables. All data are represented as standardized coefficients. Regression analyses revealed that left-lateralized alpha oscillations accounted for 44% of the variance in RT differences between task conditions. Specifically, stronger oscillations in the left superior temporal gyrus, but not other left-lateralized nodes led to significantly smaller RT differences on the task. α = alpha.

4. Discussion

In this study, we investigated the neural oscillatory responses involved in a semantic processing task. Whole-brain pairwise analyses revealed significant differences in oscillatory activity between semantic and length processing trials in the left superior temporal and prefrontal cortices. Specifically, a left-lateralized network involving the left pars orbitalis, left IFG, and left STG showed stronger alpha oscillations (i.e. greater alpha ERD) during the 700–1,200 ms time range. Follow-up analyses using the time series from the peak voxel in each region showed that the stronger alpha oscillations needed for semantic processing predicted the conditional differences in RTs and that this effect was driven by activity in the left STG. Essentially, stronger alpha oscillations in the left STG during the semantic condition were associated with faster responses and more equalized behavior across conditions. Below, we discuss the implications of these findings for understanding the cognitive processes involved in semantic processing.

Previous investigations of alpha oscillatory activity consider decreases in power (i.e. desynchronizations) in a given region to indicate activation of that region, as would be shown in fMRI (Cabeza and Nyberg 2000; Medendorp et al. 2006; Klimesch et al. 2007; Jensen and Mazaheri 2010; Klimesch 2012). For example, previous studies utilizing verbal memory tasks have implicated left-lateralized alpha oscillations as a representation of active processing of verbal stimuli (van Dijk et al. 2008; Proskovec et al. 2016; Wiesman et al. 2016; Embury et al. 2019; Proskovec et al. 2019; Koshy et al. 2020). Here, we found sustained alpha oscillatory activity (i.e. ERD) in the left-lateralized semantic control network, including regions of the left IFG and left STG, suggesting that semantic decisions require recruitment of executive control areas and that activity in this area is sustained through almost the entire process (i.e. up to the motor response). To our knowledge, we are the first to show such continued active processing of semantic stimuli in subsequent time windows in the left-lateralized semantic control network. In contrast, transient and sustained increases and theta and gamma activity were not modulated by task condition. Previous studies have implicated early occipital theta and gamma activity in visual stimulus registration and early visual processing (Wiesman et al. 2017, 2018; Wiesman and Wilson 2019) and the current results suggest that these early oscillations as well as the later sustained activity are likely more involved in pre-semantic visual processing.

Our most important finding was likely the relationship between the strength of alpha oscillations in the left STG and task performance. Essentially, we found that stronger alpha oscillations (i.e. greater decrease from baseline) in the STG during semantically related trials compared to length trials are associated with smaller behavioral “costs” for semantic processing (i.e. smaller RT differences between the 2 conditions). The left posterior STG, which includes Wernicke’s area, has not generally been associated with semantic control and processing. In fact, this region has traditionally been associated with speech comprehension and production of intelligible speech (Binder et al. 2009; Binder 2017), though some studies have proposed a role in semantic cognition. For example, at least one study has shown the STG to be activated more during a verbal word-matching task when compared to nonverbal semantic tasks (Marconi et al. 2013). Likewise, using a semantic decision task and independent component analysis (ICA), Kim et al. (2011) found a left-lateralized network including the left STG among other more established control regions (e.g. left inferior frontal gyrus). Other supportive evidence can be gleaned from lesion studies where patients with semantic aphasia exhibit impaired performance on semantic tasks that are more executively demanding (Noonan et al. 2009; Rogers et al. 2015).

The left IFG, including the pars orbitalis and pars triangularis, is widely known as a critical node in the semantic control network (Noppeney et al. 2004; Bedny et al. 2008; Binder et al. 2009; Noonan et al. 2009; Hoffman et al. 2010, 2015; Kim et al. 2011; Marconi et al. 2013; Thompson et al. 2017; Della Rosa et al. 2018; Teige et al. 2018). One theory proposes that the left IFG is responsible for controlled retrieval of relevant semantic representations, while another theory suggests that this region may be important for post-retrieval selection of semantic representations or resolution of competing alternatives (Thompson-Schill et al. 1997; Wagner et al. 2001). In support of these theories, studies have shown greater activation during tasks requiring greater executive control (Della Rosa et al. 2018; Teige et al. 2018), as well as tasks requiring competition resolution (Bedny et al. 2008; Hoffman et al. 2015), which lend support to the controlled retrieval and selection theories, respectively. Furthermore, previous work has proposed that the left IFG serves both of these processes through anatomically dissociable regions, with the left pars orbitalis serving controlled retrieval and the left pars triangularis being involved in post-retrieval selection (Badre et al. 2005; Badre and Wagner 2007). Transcranial magnetic stimulation (TMS) studies that have stimulated the left pars triangularis have shown that “virtual lesions” in this area disrupt the ability to make executively demanding judgments but have a limited effect on decisions involving strong associations that are thought to utilize more automatic processes (Hoffman et al. 2010; Whitney et al. 2011, 2012). These studies, however, were not able to distinguish selection from controlled retrieval. The left IFG response observed in the current study appeared to be centered on the left pars opercularis, though it extended into the pars triangularis and thus likely reflected activation of both regions.

Teige et al. (2018) suggested that controlled retrieval would likely be associated with early activation in the left IFG, while later activation would be seen if this area is responsible for selection or competition resolution. Given the supposed anatomic dissociation of these two processes, it would follow that the left orbitalis would show early activation followed by late activation of the left pars triangularis. However, our data show simultaneous activation of the left pars orbitalis and the more posterior regions of the IFG, contradicting these theories. Interestingly, Teige et al. (2018) proposed that sustained activation in the semantic control network, as we observed, could represent an attempt to continually shape selection within the ATL to eventually arrive at the most accurate semantic representation.

Before closing, it is important to note limitations of this study. First, the nonsemantic control task involved matching words based on number of letters. While participants were not asked to make a judgment regarding the meaning of the words, the presentation of the words may have resulted in automatic activation of semantic representations in ATLs and other regions. Thus, we were limited in our ability to restrict activation of semantic representation networks during the length condition. To prevent this, future studies could use strings of letters or other nonsemantic stimuli as a control condition. Of course, such stimuli would come with their own limitations and would not be as matched as those in the current study. Secondly, a common pitfall of neuroimaging studies is the susceptibility to overfit models and this study is no exception. Future work would benefit from evaluating the models identified in the current study in independent samples. Additionally, we implemented cluster-based permutation testing to control for Type I error in determining conditional differences in neural activity. While this is considered a very rigorous approach, it could be overconservative and thus it is possible that other brain regions were activated during semantic processing and were not detected.

In conclusion, the current study investigated oscillatory responses during a semantic processing paradigm. We observed stronger alpha oscillatory activity (i.e. greater alpha ERD) in the semantic condition relative to the nonsemantic length condition, with peaks in the left orbitalis, left IFG, and left STG. Further, these differences in activity, particularly in the left STG, explained differences in RTs between conditions. Our findings provide new evidence for the importance of left-lateralized alpha activity in language regions during semantic tasks and suggest that the strength of alpha oscillations in the left STG directly impacts the behavioral cost of semantic processing. These data reinforce existing literature and add novel insight on the temporal dynamics supporting the executive role of the semantic control network in behavior.

Funding

This research was supported by grants R01-MH116782 (TWW), R01-DA047828 (TWW), R01-MH118013 (TWW), and P20-GM144641 (TWW) from the National Institutes of Health of the United States of America.

Conflict of Interest statement. None declared.