The Anterior Temporal Semantic Hub in the Left and Right Hemispheres: A Unitary System or Two Separate Systems for Meaning?

Abstract

Research indicates that the anterior temporal lobes (ATLs) in the left and right hemispheres constitute a supramodal hub for verbal and nonverbal semantic information. However, it remains unclear whether (1) left and right ATLs are respectively specialized for verbal and nonverbal semantic processing (functional specialization models) or (2) bilateral ATLs equally contribute to multimodal semantic processing (unitary “hub-and-spoke” models). The present study examined the question using repetitive transcranial magnetic stimulation (rTMS) with an implicit semantic priming paradigm minimizing top-down influences from cognitive control mechanisms. After receiving rTMS over the left ATL, right ATL or vertex, healthy human adults of either sex made real-versus-unreal judgments about visual words or objects each preceded by semantically related or unrelated primes. While a similar amount of semantic priming survived vertex stimulation for both words and objects, we found that left ATL stimulation, but not right ATL stimulation, eliminated semantic priming for words, suggesting that the verbal semantics is represented in the left ATL. In contrast, rTMS over the ATL eliminated semantic priming for objects irrespective of the side of stimulation. This latter finding suggests a twofold mechanism whereby semantic processing of visual objects depends on semantic computations in the right ATL as well as automatic activation of semantic word representations in the left ATL generating “verbal label” feedback. The observed functional asymmetry of the semantic ATL hub overall supports functional specialization models enriched by verbal label feedback accounts, unveiling that verbal semantic knowledge plays a guiding role in the semantic analysis of visual objects.

Significance Statement

Research indicates that bilateral anterior temporal lobes (ATLs) represent a supramodal hub for semantic information. However, it remains open whether left and right ATLs equally contribute to verbal and nonverbal semantic processing. Using transcranial magnetic stimulation with a semantic priming paradigm, we found that left, but not right, ATL stimulation eliminated priming for words, suggesting that verbal semantics is represented in left ATL. In contrast, ATL stimulation eliminated priming for objects regardless of the stimulation side, suggesting that semantic processing of nonverbal stimuli requires both semantic computations in right ATL and automatic activation of word representations in left ATL. Our results demonstrate the functional asymmetry of the semantic ATL hub, unveiling a guiding role of verbal knowledge in semantic processing.

Introduction

A central issue in cognitive neuroscience is understanding how object meanings are stored as coherent concepts in the brain. Early neurocognitive models assumed that the brain's semantic system comprises multiple modality-specific subsystems (Damasio, 1989; Damasio, 1996), each of which integrates various features of stimuli within a sensorimotor modality (Martin, 2007). In contrast, the more recent “hub-and-spoke model” proposes that the anterior temporal lobes (ATLs) in the left and right hemispheres play a pivotal role in generating higher-order, abstract knowledge of objects (Patterson et al., 2007; Ralph et al., 2017). Currently, this latter theoretical framework may represent the most comprehensive neurocognitive model of semantic memory, accounting for various behavioral manifestations in semantic dementia (SD), a neurodegenerative condition with multimodal semantic deficits (Hodges and Patterson, 2007). It remains unclear, however, whether the left and right ATLs play distinct roles in semantic processing.

Namely, the hemispheric specialization model assumes that left and right ATLs are respectively specialized for verbal and nonverbal semantics (Gainotti, 2014; Mesulam et al., 2014), consistent with the fact that focal lesions in left and right ATLs affect semantic processing in verbal (Snowden et al., 2004; Mesulam et al., 2013) and nonverbal (Mion et al., 2010) domains, respectively. In contrast, the unitary hub-and-spoke model posits that both ATLs equally contribute to multimodal semantic processing (Ralph et al., 2017), which has also gained support from neuropsychological and neuroimaging data (Patterson et al., 2007; Rice et al., 2015).

While these existing data seem inconsistent about the functional properties of the ATLs, transcranial magnetic stimulation (TMS) in healthy adults provides a powerful alternative test for addressing the question. TMS can transiently disrupt the function of target brain sites with a spatial resolution of <10 mm, allowing for a more rigorous testing of the causal link between structure and behavior than lesion models with heterogeneous damage patterns and neuroimaging data derived from correlational inferences (Pascual-Leone et al., 2000). However, previous brain stimulation studies also seem rather inconsistent, with some aligning with the hemispheric specialization models (Bonni et al., 2015; 2023) and others supporting the unitary models (Pobric et al., 2010).

Importantly, the apparent inconsistency might arise from explicit “semantic association” tasks used in those studies, which require multiple late-stage cognitive components beyond automatic semantic processing, including executive control, working memory, and response selection (Lambon Ralph, 2014). Consequently, any potential differences in behavioral measures between left and right ATLs might be confounded by more global and nonspecific effects arising from those strategic components. Such late-stage modulatory influences should be controlled for when testing the functional asymmetry in the ATLs, which constitute the computational core hub for automatic/implicit semantic processing (Teige et al., 2018).

We addressed the question using repetitive TMS with an implicit semantic priming paradigm, which allowed us to separate automatic aspects of semantic processing from late-stage strategic influences. Participants made real-versus-nonreal judgments about words or objects after receiving TMS applied over the left or right ATL (Fig. 1). According to the unitary hub-and-spoke model, unilateral ATL stimulation, whether left or right, should not disrupt semantic priming for either words or objects, because both ATLs equally contribute to semantic processing for words and objects, such that functional disruption of one ATL should be compensated by its homolog in the other hemisphere (Ralph et al., 2017). Such compensatory inter-hemispheric cooperation is well known to operate in other brain areas, including the posterior temporal cortex for speech perception (Hickok et al., 2008) and middle fusiform cortex for face recognition (Bartolomeo and Thiebaut de Schotten, 2016). In contrast, hemispheric specialization models would predict that left ATL stimulation disrupts semantic priming for words but not for objects, and vice versa for right ATL stimulation. Yet another possibility is that verbal representations play a dominant and guiding role in semantic analysis (Lupyan, 2012). Hence, left ATL stimulation should affect semantic priming not only for words but also for objects, whereas right ATL stimulation should affect only objects.

Figure 1.

Behavioral task and cortical target regions. A, Sequence of visual stimuli used for the lexical decision (LD) and object decision (OD) tasks. On each trial, participants viewed a prime–target pair which could be either real or nonreal in each task and made real versus nonreal judgment about targets. Real targets were preceded by primes which could be either semantically related or unrelated to the targets. B, Stimulation sites. Repetitive TMS was delivered to the left anterior temporal lobe (left ATL: x = −52, y = 2, z = −28), right homologous (right ATL: x = 53, y = 4, z = −32) and vertex in a counterbalanced order across participants (see Materials and Methods).

Materials and Methods

Participants

Twenty-five healthy right-handed volunteers (13 females; age range, 18–45 years) participated in the experiment. All of them were native Japanese speakers with normal or corrected-to-normal vision. None of them had known neurological or psychiatric disorders. All provided written informed consent prior to the experiment. The sample size was determined using G*Power (Faul et al., 2009) for repeated-measures ANOVA with eight measurements (2 sites × 2 tasks × 2 priming conditions), with a statistical power (1 − β) of 0.8 and the significance level (α) at 0.05. Since TMS typically produces small-to-moderate effects on cognitive functions in healthy participants (Klaus and Schutter, 2018), we expected an effect-size size (η_p²) of 0.042 (corresponding to ∼0.42 in Hedge's g) according to several meta-analysis studies (Klaus and Schutter, 2018; Patel et al., 2020; Ntemou et al., 2023; Wang et al., 2024; Zhang et al., 2024). This analysis revealed that at least 22 participants were required to adequately power the study. Since two participants were excluded from analyses because of technical errors during data acquisition, we additionally performed post hoc power analysis (alpha level = 0.05) and confirmed that statistical power (1 − β) exceeded 0.95 for each of the significant differences reported (see Results). The protocol of this study was approved by the ethical review committee of the National Rehabilitation Center for Persons with Disabilities.

Behavioral tasks

We initially selected 76 Japanese nouns denoting object names and corresponding 76 line-drawings from Cycowicz et al. (1997). Using these items as primes, we constructed 76 semantically related prime–target pairs (e.g., orange-lemon) and 76 unrelated pairs (e.g., orange-eye) for each of the lexical decision (LD) and object decision (OD) tasks (Fig. 1). Lexical frequency of targets was indexed by Log10 of the number of search results in the Google engine (Blair et al., 2002; Moreno-Martinez et al., 2014) and matched between the related and unrelated pairs [mean (SE) = 7.68 (0.09) vs 7.73 (0.08); t (150) = −0.41, p = 0.68]. To compare the semantic relatedness between the related and unrelated pairs, we calculated the semantic similarity between primes and targets using Sematch (Zhu and Iglesias, 2017) and confirmed that the similarity score was significantly higher for the related than for the unrelated pairs [mean (SE) = 0.82 (0.01) vs 0.37 (0.01); t₍₁₅₀₎ = 28.71, p = 7.97 × 10⁻⁶³]. Written word stimuli for LD were presented in phonographic kana script, with their word length being matched between the related and unrelated pairs (t₍₁₅₀₎ = −0.70, p = 0.48). The perceptual familiarity and visual complexity of target objects were also matched between the related and unrelated pairs (t₍₁₅₀₎ = 0.39, p = 0.70 for familiarity and t₍₁₅₀₎ = 0.81, p = 0.42 for complexity). For the LD task, we further created 76 nonwords by shuffling characters pseudorandomly among real target words. For the OD task, 76 nonobjects were also created by decomposing and rearranging parts of the real target object pictures (see https://osf.io/t7mnw/ for all stimulus materials).

Stimulus presentation and response collection were controlled using the E-prime 3.0 software (Psychology Software Tools). The sequence of events used for the rTMS experiment is illustrated in Figure 1. Each trial comprised central fixation (1,000 ms), a prime stimulus (200 ms), blank screen (100 ms), and a target stimulus (500 ms) followed by a blank response screen (1,500 ms). This prime–target stimulus onset asynchrony is sufficiently short to tap the implicit semantic processing of visual stimuli (∼ 400 ms; Neely, 1977; Ortells et al., 2003). In half of all trials, target stimuli were pseudorandomly replaced with nonwords in the LD task or with nonobject in the OD task. In the LD task, participants decided as accurately and as quickly as possible whether or not targets represented real words, whereas they decided whether or not targets were real objects in the OD task. In both tasks, participants responded by button press with their right index and middle fingers. Each task consisted of 76 trials and lasted ∼4 min, yielding 19 trials for each of the four priming conditions (i.e., related vs unrelated × real vs nonreal).

Participants were seated at a viewing distance of 57 cm with a head and chin rest to reduce head movements. The rTMS experiment consisted of three blocks in each of which participants received rTMS over one of the three cortical targets (left ATL, right ATL, and vertex, see below) for ∼8 min, followed immediately by the LD and OD tasks (for ∼4 min each). The order of stimulation sites and tasks was randomly counterbalanced across participants such that tasks and stimulation sites are not correlated with each other. To eliminate possible carryover effects of rTMS, participants took a >30 min break between blocks. Before the start of the experiment, they performed a short practice block for each task.

TMS procedure

For each participant, structural T1-weighted MRI scan was obtained prior to the rTMS experiment. To determine cortical stimulation sites, we used the MarsBaR toolbox (https://marsbar-toolbox.github.io/) to construct two 5-mm-radius spherical regions of interest (ROIs) in the left ATL and right ATL in the standardized brain space defined by the Montreal Neurological Institute. The center coordinates of the ROIs were located at x = −52, y = 2, z = −28 for left ATL and x = 53, y = 4, z = −32 for right ATL, respectively (Fig. 1B), according to the spatial coordinates data provided by previous brain stimulation studies (Lambon Ralph et al., 2009; Teichmann et al., 2016a). These two ROIs were then mapped onto each participant's native brain space using the spatial normalization procedure implemented in Statistical Parametric Mapping 12 (https://www.fil.ion.ucl.ac.uk). To control for global nonspecific effects of rTMS on behavioral outcomes, we additionally included the vertex as a control site located at Cz according to the international 10–20 system (Steinmetz et al., 1989).

Single-pulse monophasic TMS was generated using a 70 mm figure-of-eight coil with MagStim rapid (Magstim Company). The coil positions were guided by the Brainsight neuronavigation system (Rogue Research). The intensity of TMS was set to 100% of the resting motor threshold in the first dorsal interosseous muscles, which corresponded to ∼60% of the maximum stimulator output [mean (SD) = 59.26 (6.17) %]. Prior to the experiment, the motor threshold was defined as the lowest stimulus intensity required to elicit motor-evoked potentials in at least five consecutive trials. Such low-frequency rTMS is known to transiently suppress variously different types of neurocognitive processing (see Silvanto and Cattaneo, 2017 for review). A total of 492 pulses of 1 Hz rTMS were applied to either of the three sites in each block, so that the total number of TMS pulses, including those required for determining the motor threshold, did not exceed 1,500 pulses for each participant in accordance with the clinical practice guideline by the Japanese Society of Clinical Neurophysiology.

Statistical analysis

Behavioral data were analyzed using R (http://www.r-project.org). Prior to analysis, reaction time (RT) data were screened for errors and outliers (±3 SD from the mean) for each experiment. Since participants performed the two tasks with high accuracy but responded more slowly in LD than in OD (see Results), we then compared speed–accuracy relations between the two tasks. We modeled trial-level accuracy using a mixed-effects logistic regression (Davidson and Martin, 2013; Haeuser and Kray, 2023), predicting accuracy from RT, task, and their interaction, with random intercepts and RT slopes for participants. This analysis revealed a significant interaction between RT and task (b = 8.99 × 10⁻⁴, SE = 4.19 × 10⁻⁴, z = 2.15, p = 0.032), suggesting a non-negligible difference in speed–accuracy trade-off between LD and OD.

We therefore used composite RT-accuracy scores (Townsend and Ashby, 1983) as a primary performance measure to compare the behavioral effects of TMS between the LD and OD tasks (Fig. 2). This behavioral index, typically referred to as “adjusted RT,” has been widely used for controlling for speed–accuracy trade-offs in behavioral science (Watson et al., 2010; Han and Marois, 2013; Ritchie and de Beeck, 2019; Lee and Han, 2020). Adjusted RTs are also useful and commonly employed in TMS studies (Chambers et al., 2004; Cattaneo et al., 2009; Jackson et al., 2015; Uno et al., 2025), because magnetic stimulation per se is known to exert nonspecific facilitatory effects on behavior via increased intersensory coupling or motor cortical excitability (Terao et al., 1997; Duecker et al., 2013; Luber and Lisanby, 2014). More specifically, these adjusted RTs were calculated for each participant by dividing RT by accuracy for each priming condition for each task for each site. The adjusted RTs are therefore equal to plain RTs in the absence of errors and increase in proportion with the number of errors such that higher scores overall represent poorer performance. Since fast decisions tend to be less accurate (Heitz, 2014), this composite index allowed to control for the speed–accuracy trade-offs in behavioral performance across different tasks by weighing fast performance against poor accuracy (Townsend and Ashby, 1983).

Figure 2.

Behavioral performance during the LD and OD tasks. A, Mean accuracy and reaction times are plotted with respect to prime–target semantic relatedness for each target site. Error bars represent the standard error of the mean. Participants overall responded faster but made slightly more errors in OD than in LD. B, Adjusted RTs were then calculated to control for speed–accuracy trade-offs by dividing RT by accuracy for each condition for each task for each site (see Materials and Methods). C, Net effects of ATL stimulation on semantic priming were obtained by subtracting the adjusted RTs for the vertex from those for the left and right ATLs. The asterisk indicates a significant site × task interaction at p < 0.005. Semantic priming for LD was eliminated by the left ATL stimulation but survived the right ATL stimulation. In contrast, semantic priming for OD was eliminated by ATL stimulation irrespective of the side of stimulation.

Statistical analysis was performed using the adjusted RTs for correct responses to real words and objects. Since the resulting behavioral measures did not meet the assumptions of normality (p < 0.001, Shapiro–Wilk test) and homogeneity of variance (p < 0.001, Levene test), these values were examined using permutation ANOVA implemented in R (https://cran.r-project.org). To confirm baseline effects of semantic priming under the present experimental procedure, we first examined adjusted RTs for the vertex stimulation using 2 × 2 repeated-measures ANOVA with 5,000 permutations which included task (LD vs OD) and prime–target relation (related vs unrelated) as within-participant factors (Fig. 2A). Next, to examine the region-specific net effects of TMS on semantic priming, adjusted RTs for the vertex were subtracted from those for the left and right ATLs for each task for each priming condition (Fig. 2B) and then submitted to a 2 × 2 × 2 permutation ANOVA which included site (left and right ATLs), task (LD and OD), and prime–target relation (related and unrelated) as within-participant factors.

E-field modelling

The spatial distribution of the TMS-induced electric field (E-field) was estimated using the SimNIBS software package (https://simnibs.github.io/simnibs/). For each participant, the high-resolution T1-weighted MRI volume was used to construct a tetrahedral surface mesh head model which included five types of head tissue (scalp, skull, cerebrospinal fluid, gray and white brain matter) with standard isotropic conductivity values (Wagner et al., 2004; Thielscher et al., 2011). For each of the left and right ATL target sites, the position and orientation of the TMS coil were exported from the Brainsight neuronavigation system and mapped into the SimNIBS coordinates to determine the magnetic vector potential at each mesh node. The strength of the E-field was then calculated using the finite element method for each target site for each participant. In group analysis, the individual-level E-field maps were normalized and rendered onto the surface-based MNI brain space using FNIRT (https://fsl.fmrib.ox.ac.uk/) and averaged across participants for each target site (Fig. 3).

Figure 3.

Computational modeling of the rTMS-induced electric field (E-field). For each cortical target site, the strength of the E-field was estimated from the position and orientation of the TMS coil using the finite element method for each participant and mapped onto the surface-based MNI brain space across all participants. The resulting cortical E-field estimates show that TMS pulses were successfully delivered to the three target sites during the rTMS experiment.

Results

Mean accuracy and RT data are illustrated for each task for each target site in Figure 2A. Overall, participants performed the two tasks with high accuracy but responded more slowly in LD than in OD (92.92% and 608 ms for LD, 92.57% and 572 ms for OD). To control for the between-task difference in speed–accuracy trade-off and nonspecific facilitatory effects of TMS on behavior (Terao et al., 1997; Duecker et al., 2013; Luber and Lisanby, 2014), we calculated adjusted RTs (Fig. 2B) to compare the behavioral effects of rTMS on semantic priming across tasks and sites (see Materials and Methods). We first examined the adjusted RTs for the vertex stimulation to determine the baseline effects of semantic priming with 2 × 2 repeated-measures permutation ANOVA with task (LD and OD) and prime–target relation (related and unrelated) as within-participant factors. The main effects of task and relation were both significant (F_(1,22) = 5.252, p = 0.031, η_p² = 0.193 and F_(1,22) = 64.911, p = 2.000 × 10⁻⁴, η_p² = 0.747) but did not interact with each other (F_(1,22) = 2.840, p = 0.109, η_p² = 0.114), confirming that the LD and OD tasks produced significant and similar amounts of semantic priming during vertex stimulation.

We next examined net effects of rTMS on semantic priming by subtracting the adjusted RTs for the vertex from those for the left and right ATLs for each task for each condition (Fig. 2C). These difference scores were then submitted to 2 × 2 × 2 permutation ANOVA including site (left and right ATLs), task (LD and OD), and prime–target relation (related and unrelated). The main effects of task, site, and relation were all nonsignificant (F_(1,22) = 0.048, p = 0.829, η_p² = 0.002 for task, F_(1,22) = 0.403, p = 0.535, η_p² = 0.018 for site, F_(1,22) = 1.050, p = 0.313, η_p² = 0.046 for relation). There was a significant interaction between site and relation (F_(1,22) = 7.361, p = 0.011, η_p² = 0.251), suggesting that the impact of rTMS on semantic priming differed between the left and right ATLs. The effect of task interacted neither with that of site nor with that of relation (F_(1,22) = 1.614, p = 0.209, η_p² = 0.068 and F_(1,22) = 0.087, p = 0.769, η_p² = 0.004). Importantly, there was a significant three-way interaction between the effects of task, site and relation (F_(1,22) = 4.608, p = 0.036, η_p² = 0.173). To describe this three-way interaction, we further examined the site × priming interaction separately for each of the LD and OD tasks.

For the LD task, the direction of semantic priming (i.e., related vs unrelated) was seemingly reversed when rTMS was delivered to the left ATL as compared with the right ATL (Fig. 3C). This was indeed validated by significant cross-over interaction between site and relation (F_(1,22) = 10.910, p = 0.003, η_p² = 0.332), with the two main effects being both nonsignificant (F_(1,22) = 1.525, p = 0.238, η_p² = 0.065 for site and F_(1,22) = 0.220, p = 0.640, η_p² = 0.010 for relation). By restricting the analyses to each stimulation site, we further confirmed that semantic priming survived after right ATL stimulation (F_(1,22) = 4.687, p = 0.037, η_p² = 0.176) but was eliminated by the left ATL stimulation (F_(1,22) = 1.882, p = 0.185, η_p² = 0.079).

For OD, the main effects of site and relation were neither significant (F_(1,22) = 0.034, p = 0.849, η_p² = 0.002 and F_(1,22) = 0.820, p = 0.373, η_p² = 0.036, respectively) nor interacted with each other (F_(1,22) = 0.021, p = 0.890, η_p² = 9.466 × 10⁻⁴). Unlike in LD, these findings therefore confirmed that semantic priming for OD was uniformly eliminated by rTMS, irrespective of whether magnetic stimulation was applied to the left ATL (F_(1,22) = 0.400, p = 0.536, η_p² = 0.018) or to the right ATL (F_(1,22) = 0.947, p = 0.3337, η_p² = 0.041).

Discussion

Using an implicit semantic priming paradigm with rTMS, we examined whether the left and right ATLs play different causal roles in the automatic semantic processing of verbal and nonverbal information. Overall, we obtained a significant three-way site × task × relation interaction followed by site × relation interaction for LD. These findings strongly suggest that left and right ATL stimulations have distinct effects on verbal and nonverbal semantic priming and are therefore at odds with the unitary hub-and-spoke model assuming that the left and right ATLs are equally involved in multimodal semantic processing. Our results seem also partially inconsistent with the hemispheric specialization model assuming that the left and right ATLs specifically contribute to verbal and nonverbal semantic processing, respectively. This latter model is well consistent with the finding that left but not right ATL stimulation disrupted semantic priming in LD but seems inconsistent with the finding that both left and right ATL stimulations impaired priming in OD.

Specifically, the behavioral effects of rTMS observed in LD indicate that the left ATL plays a specific and causal role in verbal semantic processing, which supports the hemispheric specialization model assuming that the left ATL is specifically dedicated to verbal semantics. This finding also concurs with neuropsychological data showing that left ATL damage affects verbal semantics but spares nonverbal semantics (Mesulam et al., 2013) as well as with neuroimaging data showing that left ATL lesions cause verbal semantic impairments (Mion et al., 2010; Gefen et al., 2013).

The behavioral effects observed in OD concur with the previous work by Pobric et al. (2010), which showed that rTMS over the ATLs impaired semantic processing of visual objects irrespective of the stimulation side. Their observations, along with ours, do not fit the unitary models of a bilaterally distributed semantic hub with left/right compensatory mechanisms. Namely, such models predict no observable change in behavior because functional disruption of one ATL should be compensated by its homologous area in the other hemisphere (Lambon Ralph et al., 2010; Ralph et al., 2017). In other words, the raw fact that unilateral ATL stimulation, either left or right, impaired semantic processing in LD or OD is inconsistent with the unitary hub-and-spoke model, which posits a redundantly organized bilaterally distributed semantic hub (Patterson et al., 2007; Lambon Ralph et al., 2010; Ralph et al., 2017).

On the other hand, since semantic priming in OD was disrupted by ATL stimulation irrespective of the stimulation side, it may be argued that semantic processing of objects relies on broader neural systems including both the left and right ATLs, seemingly inconsistent with the notion that the right ATL is specifically responsible for nonverbal semantic processing. However, these effects of TMS observed in OD align well with the hemispheric specialization model if the left and right ATL are functionally connected with each other during semantic processing. Specifically, the apparent inconsistency can be reconciled if semantic processing of nonverbal objects partially relies on the multimodal integration of visuo-perceptual and linguistic information via the functional coordination between bilateral ATLs. Namely, a growing body of behavioral and neuroimaging studies shows that perceptual and cognitive processing of objects is influenced by their verbal labels (McKelvie, 1976; Lupyan, 2008a,b; Boutonnet and Lupyan, 2015; Fotiadis and Protopapas, 2023). Inspired by the time-honored notion of linguistic relativity (Wolff and Holmes, 2011), the “label feedback hypothesis” proposes that visual objects rapidly and automatically activate their verbal representations which in turn produces facilitatory feedback activation in neural systems for nonverbal processing, including those involved in lower-level sensory processing (Lupyan, 2012). Our results from the OD task are in line with this view while suggesting that object processing is disrupted either by right ATL stimulation altering directly semantic processing of nonverbal information or by left ATL stimulation inhibiting semantic word representations, which are indispensable for the generation of verbal feedback activation. Such automatic feedback activation from linguistic representations is known to strongly influence behavior in healthy adults (Lupyan, 2012) and also concurs with neuropsychological data showing that left ATL atrophy in SD correlates with both verbal and nonverbal semantic impairments (Adlam et al., 2006).

Relatedly, it is important to note that our results overall suggest that the left ATL plays a general and dominant role in semantic processing. That is, we observed that semantic processing is generally impaired by left ATL stimulation, since left ATL stimulation, but not right ATL stimulation, eliminated semantic priming for both words and objects. Hence, such general reduction in semantic processing efficiency may be interpreted as suggesting that the left ATL acts as a “promotor” boosting semantic processing in both verbal and nonverbal modalities via verbal labels.

Taken together, our rTMS results from the LD and OD tasks are consistent with the hemispheric specialization model enriched by verbal label feedback accounts. Yet, we do not assume general claims of the linguistic relativity theory positing that language shapes semantic merging, reasoning, or thinking. Rather, our findings reveal that language-related semantic codes are required to reinforce efficient processing of multimodal/nonverbal semantic representations through bottom-up feedback processes. Furthermore, our TMS results from an implicit priming paradigm demonstrate that the ATL hub holds a causal role in automatic semantic processing, asymmetrically distributed across the ATLs, rather than in controlled stages of semantics (Lau et al., 2013; Teige et al., 2018). Such regulatory downstream processes are shown to be implemented by a “semantic control network” connecting the ATL hub with inferior prefrontal and posterior middle temporal regions (Hagoort, 2005; Thompson-Schill et al., 2005; Binder et al., 2009).

More generally, the functional left-versus-right ATL asymmetry seems in good accordance with the Hebbian framework, whereby the left and right ATLs each have strong connectivity with left language areas and right visuo-perceptual areas (Lambon Ralph et al., 2001; Gainotti, 2007). Such hemispheric asymmetry has been demonstrated by a growing number of neuroimaging studies with healthy and brain-damaged people unveiling that the left and right ATLs have different patterns of functional/structural connectivity with brain areas involved in domain-specific representations (Hurley et al., 2015; Papinutto et al., 2016). For example, fiber tract connections between the right ATL and right posterior temporal cortex involved in visual object representations correlate with nonverbal semantics whereas those between the left ATL and left posterior temporal cortex involved in lexical representations correlate with verbal semantics (Sundqvist et al., 2020). Different functional connectivity between the ATLs and other brain areas is also shown to reflect the individual variability in behavioral performance in post-stroke aphasia during verbal and nonverbal processing (Warren et al., 2009). In sum, our results combined with prior findings show that semantic specialization of the left and right ATL is consistent with their asymmetric and distinct connectivity with remote modality-specific areas.

There are a few limitations of the study that should be considered. First, our behavioral paradigms involving verbal feedback mechanisms did not unveil double dissociation between verbal and nonverbal dimensions following left versus right ATL stimulation, which would have provided further support for hemispheric specialization models. To substantiate the specific role of the right ATL in nonverbal semantics, future research needs to uncover such double dissociations by exploring split-brain patients or by controlling for the effects of verbal feedback. This latter could be achieved by using nonverbal perceptual tasks (e.g., face familiarity judgment), which probes nonverbal semantic representations while minimally involving verbal representations. Second, although our priming paradigm shed light on automatic semantic activation while minimizing possible influences from postsemantic executive control, such top-down modulatory effects might not be entirely eliminated from behavioral measures because conscious perception of visual stimuli may still induce strategic analysis of semantic information to some extent (Dehaene et al., 2006; Kouider and Dehaene, 2007). This remaining issue may be resolved by using experimental paradigms which can more rigorously restrict strategic influences from executive control (e.g., shorter prime duration or visual masking). Third, it is possible that magnetic stimulation applied to the ATLs not only yields not only local inhibitory effects but also changes the activation level of cortico-subcortical networks involved in meaning. Since the difficulty in disentangling these local and remote effects is an inherent methodological limitation of TMS, we cannot fully exclude that the observed effects of ATL stimulation are influenced by the potential modulation of other task-relevant neural systems. Fourth, our results with a limited sample size of 23 participants require confirmatory investigations with larger populations, possibly in cross-linguistic or cross-cultural designs.

In conclusion, our results demonstrate that the left ATL plays a specific and causal role in automatic semantic processing of verbal information, whereas both left and right ATLs are involved in the processing of nonverbal objects. These findings enrich hemispheric specialization models by integrating verbal label feedback mechanisms and suggest that efficient semantic processing of objects relies on verbal representations. Our results also have clinical/therapeutic implications since noninvasive brain stimulation has generated promising results in patients with semantic deficits (Teichmann et al., 2016b). By unveiling functional specializations of the ATLs, they may provide guidance for selecting optimal stimulation targets in brain-damaged patients suffering from verbal or nonverbal semantic disabilities.

Data Availability

Stimulus materials used in the present study and behavioral data analyzed are available at the OSF (https://osf.io/t7mnw/). Other behavioral and MRI datasets are archived in the National Rehabilitation Center for Persons with Disabilities and available from the corresponding author on reasonable request. All codes used for data analysis are available at the OSF (https://osf.io/t7mnw/).