Verb and sentence processing with TMS: A systematic review and meta-analysis

Transcranial magnetic stimulation (TMS) has provided relevant evidence regarding the neural correlates of language. The aim of the present study is to summarize and assess previous ﬁndings regarding linguistic levels (i.e


Introduction
Several claims have been put forward in the lesion and neuroimaging literature regarding linguistic processes (i.e., semantics and morpho-syntax) and brain structures that are needed to produce and comprehend verbs and sentences (e.g. Rofes & Mahon, 2021;Vigliocco, Vinson, Druks, Barber, & Cappa, 2011). In the current study, we aim to summarize and assess some of these claims with a focus on transcranial magnetic stimulation (TMS) studies. We focus in particular on topics and stimuli used, target stimulated regions, and reported associations between cortical regions and linguistic functions. We focus our attention on TMS because it is a noninvasive neuroscientific technique that provides causal evidence and has high spatial accuracy (Genon, Reid, Langner, Amunts, & Eickhoff, 2018;Siebner, Hartwigsen, Kassuba, & Rothwell, 2009;. Regarding previous reviews of TMS research in language processing, the following topics have been covered: the first 15 years of TMS language research (Devlin & Watkins, 2007), the role of the right hemisphere in language (Hartwigsen & Siebner, 2012), clinical applications in people with brain tumors (Jeltema et al., 2021;Lefaucheur & Picht, 2016), effects of TMS on reading processes (Arrington, Ossowski, Baig, Persichetti, & Morris, 2022), and the use of TMS in language rehabilitation (Galletta, Rao, & Barrett, 2011;Pisano & Marangolo, 2020;Torres, Drebing, & Hamilton, 2013). However, to the best of our knowledge, a systematic review and meta-analysis of TMS studies of verbs and sentences in healthy speakers does not exist. Since TMS and verb tasks are increasingly being used during preoperative language mapping of individuals with brain tumors, this work is highly relevant from a clinical point of view (Hauck et al., 2015;Ntemou, et al., 2021a;Ohlerth et al., 2021Ohlerth et al., , 2022. From a theoretical perspective, when TMS is performed with healthy individuals, it offers causal evidence about language processing in the brain unencumbered by neuroplastic changes that may occur in neuropathologies. It is hence a valuable method to assess previous claims about language in the brain. In the current study, we will first contextualize our aims by introducing some of the claims currently discussed in the lesion (e.g., post-stroke aphasia, neurodegeneration) and neuroimaging literature. We then summarize those claims that connect cortical structure and linguistic function by addressing first morphosyntactic and then semantic processes.
To produce sentences, we need to engage three main cognitive levels (Garrett, 1980;Levelt, Roelofs, & Meyer, 1999;Rofes et al., 2018): (1) the message/conceptual level: which consists of the encoding of conceptual information and communicative goals; (2) the functional level: the first phase of this level consists of the retrieval of semantic representations and the second of the assignment of grammatical features and thematic roles to each verbal argument; and (3) the positional level: which consists of assembly of syntactic features and framebuilding for morphemes and lexical representations. Localizing such linguistic processes in the human brain has gathered attention since the late 20th century (Geschwind, 1972) and has led to specific linguistic processes being attributed to certain cortical regions. For example, frontal cortical regions have been connected to morphosyntactic processes, whereas temporoparietal regions have mainly been connected to semantic processes.
Current lesion and neuroimaging studies therefore indicate that morphosyntactic processes are connected to frontal Reaction times regions, whereas semantic processes are connected to temporoparietal regions. Neither lesion nor neuroimaging studies, however, can offer evidence that is both spatially constrained and causal. This is why results from studies employing different methods, such as TMS, can help us to lend additional validity to these claims. TMS is a neuroscientific technique that (a) unlike neuroimaging methods, it non-invasively provides causal evidence for language processing in the brain and (b) unlike studies of lesion distribution, offers high spatial accuracy (Genon et al., 2018;Siebner et al., 2009). Despite the large number of TMS contributions regarding semantic and morpho-syntactic aspects of verb and sentence processing, we are unaware of a systematic review and meta-analysis on this specific topic. To fill this gap in the literature, we are conducting a qualitative review and meta-analysis in order to synthesize findings from TMS studies that employed verb and sentence experimental stimuli. Our aim is to, first, systematically gather and summarize causal evidence for different processing levels and, second, quantify functional specializations of cortical areas via the meta-analysis of effect sizes.

The present study
We summarize TMS studies and assess previous claims regarding linguistic processes (semantic/morpho-syntactic) and brain structures utilized in verb and sentence processing. Hence, we ask the following research questions: Systematic review.
1. Which topics and stimuli contrasts have been investigated in TMS research? 2. Which cortical areas have been targeted? 3. What are the reported associations between cortical area and linguistic function?
Meta-analysis (difference in effect sizes of TMS-induced effects).
1. Is there an effect size difference between experimental and control conditions? 2. Is there a difference between anterior and posterior stimulated regions between effect sizes of semantic contrasts? 3. Is there a difference between anterior and posterior stimulated regions between effect sizes of morphosyntactic contrasts?
Specific to the systematic review, we expect that areas that have received the biggest number of stimulations comprise the left perisylvian cortical regions with a specific focus on pIFG, motor cortex, and posterior temporal regions. Following neuroimaging and lesion work, we expect to find more reported associations between areas of the left frontal lobe and morpho-syntactic processes, whereas more associations with semantic processes are expected for left temporal areas. Regarding the meta-analysis, we expect a larger effect size for experimental stimuli (i.e., the stimuli in question) compared to control stimuli. In terms of effect size differences, if indeed anterior regions are mainly associated with morphosyntactic processes, we expect to see larger effect sizes for morphosyntactic contrasts when frontal regions are stimulated. If posterior regions are primarily connected to semantic processes, we expect to find larger effects sizes for semantic as opposed to morphosyntactic contrasts when temporoparietal regions are stimulated.

Systematic review
A systematic review was conducted according to the PRISMA-P guidelines (Moher et al., 2015) using the following databases: PubMed, Embase, PsycInfo, and Web of Science (Core collection). The queries were conducted in March 2021. Our search strings included the search terms ("Transcranial Magnetic Stimulation", "repetitive Transcranial Magnetic Stimulation", "navigated Transcranial Magnetic Stimulation ", "TMS", "nTMS", "rTMS") AND ("verb*", "sentence*" Papers identified through the initial search were entered into Covidence (Covidence Systematic Review software, Veritas Health Innovation, Melbourne, Australia. Available at www.covidence.org). Two independent reviewers then screened the titles and abstracts for eligibility (i.e., EN & CS). Reviewers were blinded to each other's decisions and when disagreements occurred the abstracts in question were discussed until a consensus was reached. No disagreements persisted following re-examination of the abstracts. Following duplicate removal and identification of relevant abstracts, full texts of the remaining records were reviewed by the first author to decide final inclusion for data extraction.
The two reviewers identified the following information for each article included for full text screening: title, authors, year of publication, general framework/topic, aim(s), participant characteristics (sample size, mean age/range), experimental task used, stimuli characteristics (e.g., action verbs vs abstract verbs), outcome measures, offline/online stimulation, use of navigation, stimulation protocol (e.g., repetitive TMS, 1 Hz, 15min), targeted brain areas, summary of results, and cortical areas that responded to stimulation (i.e., cortical areas that, when stimulated, showed an effect for the examined contrast).

Meta-analysis
A meta-analysis using a random effects model was conducted with the eligible papers included in the systematic review. To calculate effect sizes, we extracted the means, standard deviations, and sample sizes for all experimental and control conditions. In case these data were only given in figures, the application "Web Digitizer" (https://automeris.io/ WebPlotDigitizer/) was used to extract relevant values. Data points were classified according to condition (experimental vs control), stimulated brain region (e.g., posterior inferior frontal gyrus, supramarginal gyrus, etc.), stimulated brain region coarse (anterior vs posterior), stimulus type (e.g., whether the stimulus contrast is targeting semantic or morphosyntactic processing) and measure (e.g., RTs, accuracy).
Effect sizes and their 95% confidence intervals (CIs) were calculated as standardized mean differences (Hedge's g) for all reported contrasts (stimulationdsham/baseline condition; Hedges & Olkin, 1984). Random effects models were fitted to the outcome effect sizes. Study heterogeneity was assessed based on Cochran's Q, t 2 , and І 2 statistics. To test whether condition (i.e., experimental vs control), stimulus type (i.e., semantic vs morphosyntactic), and location of stimulation (i.e., anterior vs posterior) had an effect on the standardized mean differences, we performed meta-analyses as well as ANOVAs. For the effect of Condition, we meta-analyzed effect sizes for Experimental compared to Control conditions. To examine the interplay between stimulus type and location of stimulation, we divided studies into stimuli targeting Semantic or Morphosyntactic processes, and stimulated locations into Anterior and Posterior regions. We then performed ANOVAs with interaction between Condition and location of stimulation for each linguistic process (i.e., Semantic and Morphosyntactic). Semantic stimuli contrasts include: 1) highly expected sentence endings vs unexpected sentence endings, 2) semantically anomalous vs semantically typical sentences, 3) literal vs idiomatic sentence meaning, and 4) action vs abstract verb processing. Morphosyntactic stimuli contrasts were: 1) irregular vs regular past tense generation, 2) syntactically complex vs syntactically simple sentences, and 3) syntactically anomalous vs syntactically typical sentences. Anterior stimulations consisted of regions of the left frontal lobe and the anterior temporal pole, whereas the posterior group included stimulations that targeted middle and posterior temporal as well as parietal areas. Meta-analyses and regression models were performed on R (R Core Team, 2020), using the metafor (Viechtbauer, 2010) and meta packages (Schwarzer, Carpenter, & Rü cker, 2015). Visualizations were created using ggplot2 (Wickham, 2016).

Systematic review: search results
A search among 4 databases yielded 2301 records. Three additional studies were identified via citation search (i.e., we checked the reference list of 22 review papers). Of the total 2304 records, 1109 records were removed as duplicates and 1071 were excluded during abstract screening because (1) they were not original experimental papers, (2) the participants were part of a clinical population, (3) the neurostimulation method of choice was not TMS, and (4) abstracts were published as part of conference proceedings. Full texts were identified for the remaining 121 records, out of which 79 were excluded during full-text screening (for exclusion reasons, see Fig. 1). The final 45 eligible records were included in the systematic review. Fig. 1 presents the flowchart of the literature search and screening process. Reasons for exclusion from the meta-analysis can be found in Supplementary Material 1.

Topics and stimuli contrasts
The included studies (n ¼ 45) fall within four topics. For the investigation of each topic, different stimuli have been employed and contrasted. From more to less frequent, the investigated topics were: 1) the embodiment of verb processing, 2) morphosyntactic effects on sentence processing, 3) semantic effects on sentence processing (other than embodied cognition), 4) morphological and word class differences of verbs compared to nouns. Almost half of the papers (46.5%; n ¼ 20) focused on embodied cognition. The majority of studies on embodied cognition compared action verbs with semantically abstract verbs (e.g., to write vs to think), sometimes, in different contexts (e.g., negation vs affirmation), tenses (e.g., I sewed vs I will sew), or person inflections (e.g., I sew vs he sews (Candidi, Leone-Fernandez, Barber, Carreiras, & Aglioti, 2010;Liuzza, Candidi, & Aglioti, 2011;Papeo, Corradi-Dell'Acqua, & Rumiati, 2011).

Targeted cortical regions
Fifty-eight percent of studies (n ¼ 26) targeted either a single cortical region or one target cortical region and one control region (e.g., the homologous area of the right hemisphere). Thirty-three percent (n ¼ 15) of studies targeted 2 areas of interest in order to form dissociations between cortical region and linguistic function, and nine percent (n ¼ 4) of studies examined 3 or more cortical areas. Regarding the location, out of the total 56 stimulated regions, 70% of stimulations targeted the left frontal lobe (n ¼ 43), 16% (n ¼ 9) the left temporal lobe and 8% the left parietal lobe (n ¼ 4). Only two studies (4%) that employed the presurgical language mapping protocol stimulated the entire cortex (bilaterally) and 1 study examined the left and right cerebellum (2%). Within the left frontal lobe, the areas that received the most stimulations were the hand area of the motor cortex as well as the inferior frontal gyrus (IFG), whereas within the left temporoparietal cortices more than half of the posterior stimulations targeted the left superior and middle temporal gyri (STG, MTG). See Fig. 2 for a heatmap of target cortical regions in the left hemisphere.
Regarding the neural correlates of verbs as a word class, 62.5% (n ¼ 5) of studies associated the left anterior (n ¼ 4) and posterior (n ¼ 1) MFG as necessary for processing verbs compared to nouns. Studies that employed the preoperative language mapping protocol (37.5%; n ¼ 2) and stimulated the entire cortex, did not report significant associations between the left MFG and verb processing relative to noun processing. Fig. 3 presents a summary of left hemispheric regions and their reported associated functions.
Right hemisphere regions have primarily been stimulated as control areas for different verb and sentence contrasts. Twenty-eight percent (n ¼ 12) of the included papers stimulated homologous regions of the target left hemisphere areas as control sites. An exception to this comprised early investigations of idiom and metaphorical language processing (e.g., Papagno et al., 2002). Based on evidence from aphasia, these studies investigated right temporal regions as the locus of non-literal language processing, without however confirming predictions of lesion literature. Unlike studies on idiom comprehension, Manenti, Cappa, Rossini, and Miniussi (2008) reported lowered performance during comprehension of syntactically reversible transitive sentences after right DLPFC stimulation. However, besides the outcomes of Manenti et al. (2008) no other effects of linguistic contrasts were observed during right hemisphere stimulations.

Meta-analysis
Overall, 72 effect sizes were calculated from 22 studies. Data of 43 studies that were included in the systematic review could not be included in the meta-analysis due to missing data or the use of non-parametric statistical tests. Reasons for exclusion from the meta-analysis are listed in Supplementary Material. Regarding the effect of TMS during verb/sentence processing, we found a significant total effect size of .35 (95% CI: .28e.43, z ¼ 7.94; p < .0001). The test for heterogeneity was not significant (Q ¼ 63.97, p ¼ .71). Fig. 4 demonstrates the effect sizes alongside the 95% confidence intervals according to study.

Condition: experimental vs control
Total standardized mean difference for the control conditions was .26 (95% CI: .17e.35) and for the experimental condition .45 (95% CI: .33e.57). The analysis for group differences was significant, indicating that experimental conditions of TMS studies had an overall larger effect size (Q ¼ 6.71, p ¼ .009; see

Stimulus type: semantics
A main effect of Condition was found (F(1, 43) ¼ 5.801, p ¼ .02). The effect of Brain region coarse (anterior vs posterior) was not significant and the same holds for the interaction between Condition and Brain region coarse (F(1, 43) ¼ .14, p ¼ .7). See Fig. 6a.

Discussion
The present review and meta-analysis summarized and assessed the effects of TMS on the processing of verb and sentence stimuli in neurotypical speakers. We reviewed the examined topics and stimuli, target stimulated regions, and reported associations between cortical region and linguistic function. To examine whether different verb stimuli resulted in different effect sizes, we meta-analysed the effect sizes of the included studies and compared them according to condition (experimental vs control), stimulus type, and location of stimulation (anterior vs posterior).

Stimuli contrasts and target regions
Several stimuli targeting semantic or morphosyntactic contrasts have been employed within TMS studies. However, in the current literature, the majority of studies choose to stimulate one target cortical region that is predicted to relate to one linguistic function. To do that, the target region is stimulated during experimental and control conditions and then the induced behavioural effect is contrasted to a control region (e.g., Cappa et al., 2002;Finocchiaro et al., 2008;Glenberg et al., 2008;Holland & Lambon Ralph, 2010;Innocenti et al., 2014;Kuipers et al., 2013;Papeo et al., 2011;Shapiro et al., 2001). The control region is primarily selected due to having little to no presumed language function (e.g., homologous area of the target in the right hemisphere, areas of the occipital lobe) or because it comprises an anatomical landmark that causes TMS to diffuse (e.g., vertex). This methodology of stimulating one cortical region and connecting it to one linguistic function might answer to the question "Does area X respond more to stimulus Y?", but it does not address the specificity of the effect in "area X". For example, the comparison between object relative and subject relative clauses while stimulating only the left IFG, allows researchers to report whether the area is causally involved in performance of object relative clauses. However, it cannot determine whether the left posterior IFG is the only language related area that responds to this contrast. Hence, in this scenario we can conclude that the left posterior IFG is crucial for object rather than subject relative clauses, but we cannot demonstrate whether this is the only crucial cortical region for this contrast.
Determining the specificity of an effect is essential to disentangle contributions of different cortical regions to language processing. Currently, language is not considered to be merely a cognitive function located in Broca's and Wernicke's areas, but a dynamic model consisting of different cortical regions as well as subcortical white matter streams (Catani et al., 2012;Duffau, Moritz-Gasser, & Mandonnet, 2014;Forkel & Catani, 2019;Friederici & Gierhan, 2013;Indefrey, 2011;Indefrey & Levelt, 2000;Tuncer et al., 2021). Hence, the mapping of different linguistic components to specific cortical regions may be established not only by demonstrating the region that is causally involved in a linguistic function, but also by elucidating language relevant regions that are not involved. A suggestion to cope with this issue is to stimulate multiple language relevant areas while using the same stimuli contrast. By doing so, we may establish neuropsychological double dissociations and demonstrate the specificity of causal involvement in a linguistic function (Davies, 2010;Teuber, 1955).
The underlying connectivity between cortical areas can provide a solution to these methodological issues. As previously discussed, the language network does not consist only of cortical regions but part of it are also white matter fiber bundles (Catani & Mesulam, 2008;Fekonja et al., 2019;Forkel & Catani, 2019;Tuncer et al., 2021). The connectivity of white matter pathways can be used to guide the placement of stimulated regions and lead the dissociations. Initial evidence within the preoperative language mapping setting show that tractography-based target placement leads to a larger effect size compared to literature-based target placement methods without taking tractography into account (Reisch et al., under review (Ntemou, Picht, et al., 2021). Following a tractographybased method could result in providing potential dissociations not only between two or more language-relevant regions, but also regions that are interconnected.

Associations between brain structures and linguistic functions: anterior regions
We demonstrate that TMS studies that employed verb stimuli have connected specific cortical regions with certain linguistic structures. Starting with areas of the left frontal lobe, the processing of verb specific morphology has been linked to the MFG (Cappa et al., 2002;Cappelletti et al., 2008;Finocchiaro et al., 2008;Lo Gerfo et al., 2008;Shapiro et al., 2001). However, studies that stimulated the entire cortical surface during verb and noun production tasks have not reported the MFG to be significantly more involved in verb processing (Hauck et al., 2015;Ohlerth et al., 2021). On the contrary, these studies suggest that the highest numbers of errors during verb production compared to object naming were concentrated in posterior cortical regions (e.g., posterior STG/MTG, anterior SMG; Hauck et al., 2015;Ohlerth et al., 2021).
Also, when stimulated, the left IFG affects processing of complex sentences as well as judgments on the grammaticality of sentences (Kuhnke et al., 2017;Lauro, Reis, Cohen, Cecchetto, & Papagno, 2010;Manenti et al., 2008;Meyer et al., 2018). In that regard, contributions of the left IFG to the comprehension of complex sentences have been dissociated from the non-linear syntactic functions carried out by the posterior STG (Kuhnke et al., 2017). Unlike the adjacent MFG, the left IFG seems not to make a distinction between word classes (Cappelletti et al., 2008), and within the TMS literature has been specifically associated with functions of argument reordering, syntactic ambiguity resolution, and decision for the syntactic well-formedness of sentences ( Hand and leg areas of the M1 are cortical regions primarily associated with semantic functions (Pulvermü ller & Fadiga, 2010). TMS and action verbs have proven useful in the investigations of embodied theories of language processing. It has been shown that while listening to action verbs, TMS elicits larger MEPs compared to listening to abstract verbs (e.g., Buccino et al., 2005;Gianelli & Dalla Volta, 2014;Glenberg et al., 2008;Innocenti et al., 2014). Importantly, perturbation of motor areas has also been shown to have behavioral effects, as inhibition of hand and leg regions specifically modulates performance on action verbs compared to abstract verbs (Pulvermü ller et al., 2005;Reilly, Howerton, & Desai, 2019;Repetto, Colombo, Cipresso, & Riva, 2013;Vukovic, Feurra, Shpektor, Myachykov, & Shtyrov, 2017). Interestingly, the effect of action verbs is found not only while stimulating M1 areas but also when stimulating SMA areas (Tremblay, Sato, & Small, 2012;Willems, Labruna, D'Esposito, Ivry, & Casasanto, 2011). Even though the embodied effect of the left M1 has been replicated, it seems to be influenced by linguistic factors such as context (i.e., action verbs within affirmative vs negated contexts; Liuzza et al., 2011;Papeo, Hochmann, & Battelli, 2016), person inflection (I run vs he runs; Papeo et al., 2011), tense inflection (e.g., I will run vs I ran; Candidi et al., 2010), and previous perturbation of posterior temporal regions (Papeo et al., 2015). Hence, it is worth to employ the connectivity or motor and premotor regions (e.g., frontal aslant tract) to choose potential dissociating areas during action verb processing. Also, similarly to Papeo et al. (2015), the specificity of the action verb effect of the M1 should be contrasted with other semantic related areas.

Associations between brain structures and linguistic functions: posterior regions
Posterior cortical regions have received less attention than frontal areas within the TMS literature of verb processing. This is not surprising given that the importance of temporoparietal regions in grammatical processes has been recently emphasized (Fridriksson et al., 2018;Yagata et al., 2017). Adding to neuroimaging and lesion findings, TMS studies support morphosyntactic functions of the parietal lobe during sentence comprehension. In a series of publications, the intraparietal sulcus (IPS) has been connected to thematic role reanalysis/reassignment during processing of passive transitive reversible sentences (Finocchiaro et al., 2015(Finocchiaro et al., , 2021Vercesi et al., 2020). However, this finding has not been yet contrasted with the involvement of other regions in thematic role assignment during sentence comprehension (e.g., posterior IFG, posterior STG/MTG (Ben-Shachar, Hendler, Kahn, Ben-Bashat, & Grodzinsky, 2003;Malyutina & den Ouden, 2017). As a result, it remains unclear whether the effect of IPS stimulation during sentence comprehension of passive reversible sentences is indeed specific to this region. In addition, stimulation of adjacent areas of the IPS (i.e., SMG and AnG) has been reported to impair comprehension of long sentences, evidence that connects the inferior parietal lobe with verbal working memory (Lauro et al., 2010). These findings are important for three reasons: 1) it has been reported that passive reversible sentences carry a higher cognitive load than active reversible sentences (Caplan, Waters, DeDe,  1  2  3  4  5  6  7  8  9  10  11  12  13  14  15  16  17  18  19  20  21  22  23  24  25  26  27  28  29  30  31  32  33  34  35  36  37  38  39  40  41  42  43  44  45  46  47  48  49  50  51  52  53  54  55  56  57  58  59  60  61  62  63  64 Michaud, & Reddy, 2007), 2) performance on passive reversibles predicts performance on working memory tasks (Sung, Eom, & Lee, 2018), and 3) according to Walsh and Pascual-Leone (2003, p. 57) TMS affects approximately a 4 by 3 cm cortical area. Hence, poor performance on passive reversible sentences during IPS stimulation might also be due to a working memory effect driven by unavoidable costimulation of the inferior parietal cortex.
Finally, posterior temporal regions of the left hemisphere have primarily been studied with TMS in relation to semantic functions. The pSTG and pMTG have been associated with integration of sentential meaning as the input becomes available to the listener for both literal as well as metaphorical sentences (Fogliata et al., 2007;Franzmeier et al., 2012;Oliveri et al., 2004;Papagno et al., 2002). Interestingly, inhibition of posterior temporal regions prior to the recording of MEPs affects the embodied effect of action verbs (Papeo et al., 2015). This finding led researchers to claim that the pSTG/pMTG functions as a repository of verb meaning. This argument is also in line with contemporary models of single word retrieval that place lemma and phonological code selection in left posterior temporal regions (Indefrey, 2011).

4.2.
Meta-analysis: frontal regions for syntax, temporal regions for semantics?
Similarly to previous meta-analyses of neurostimulation effects during language tasks (Gatti, Rinaldi, Cristea, & Vecchi, 2021;Klaus & Schutter, 2018), we also report a small overall effect size (.35, CI: .26e.43) of TMS studies on verb and sentence processing. Given that we only included behavioral effects in the meta-analysis (e.g., RT/accuracy modulations), our results are similar to average TMS effects on cognitive tasks (i.e., range .24e.40 SMD; Beynel et al., 2019;Gatti et al., 2021;Klaus & Schutter, 2018). These small effect sizes, albeit demonstrating that TMS has the ability to modulate performance on verb tasks, may be due to the varying TMS protocols used to investigate verb processing. As Klaus and Schutter (2018) also point out, language investigations with TMS follow no clear guidelines. In our meta-analysis, stimulation intensities varied from 90% to 120% of the RMT and duration ranged from one or two hundred milliseconds to minutes for both online as well as offline protocols. Also, the determination of TMS intensities for language research based on the RMT of each participant has been proven a safe and established practice (Krieg et al., 2017). However, it remains unclear whether the excitability of the motor cortex is the most suitable measure to determine TMS intensities for the modulation of linguistic behavior (Klaus & Schutter, 2018). Early TMS studies on language production used the induction of speech arrests as a measure to determine individual intensities for language mapping (Jennum, Friberg, Fuglsang-Frederiksen, & Dam, 1994;Pascual-Leone, Gates, & Dhuna, 1991). Albeit this method being abandoned, it should be noted that similarly to the mapping of motor and visual functions that require motor and vision-specific thresholds to be determined (e.g., Kammer, Beck, Erb, & Grodd, 2001;Picht et al., 2009), language assessment with TMS might benefit from the determination of language-specific stimulation thresholds.
Even though lesion and neuroimaging literature have suggested that morphosyntactic processes of verbs require further frontal rather than temporal regions, this distinction does not emerge from the present meta-analysis. Indeed, we found a lack of interaction between Condition and Location of stimulation for both morphosyntactic as well as semantic contrasts. As fMRI work has shown, besides frontal regions, also posterior temporal regions respond to morphosyntactic contrasts (Grodzinsky & Friederici, 2006;Matchin, Liao, Gaston, & Lau, 2019). According to  model of the cortical organization of syntax, posterior temporal regions are engaged during hierarchical syntactic processes. This implies that damage to these regions can have detrimental effects to sentence production and comprehension, even with simple sentence structures . The lack of significantly larger effect sizes when stimulating anterior compared to posterior cortical regions with TMS, supports the argument that morphosyntactic functions are not localized only in one frontal area but are rather distributed across left hemisphere perisylvian regions. To disentangle whether anterior and posterior regions subserve distinct aspects of syntactic processing, however, we need to further probe these areas by using the same stimuli (e.g., see Acheson & Hagoort, 2013; Kuhnke et al., 2017;Lauro et al., 2010).
Similarly to the analysis of morphosyntactic contrasts, we did not observe an interaction between Condition and Location of stimulation for studies that examined semantic contrasts. Traditionally, left temporoparietal areas are considered the core semantic network. Temporal and inferior parietal areas have been reported to store semantic representations and damage to these regions may lead to semantic deficits (Binder & Desai, 2011;Dronkers, Wilkins, Van Valin, Redfern, & Jaeger, 2004;Hillis & Caramazza, 1991;Lambon Ralph, Ehsan, Baker, & Rogers, 2012). The semantic system, however, is not only a storage of units of meaning. Control processes that guide word selection and interpretation are also involved (Binder, Desai, Graves, & Conant, 2009;Lambon-Ralph, Jefferies, Patterson, & Rogers, 2017). These distinct processes that constitute the semantic system are distributed across frontal regions rather than being localized in specific temporo-parietal areas (for a meta-analysis of semantic related fMRI activations see: Binder et al., 2009). Such evidence is not only reported in the TMS literature, but also in fMRI studies and in intraoperative studies with direct-electrical stimulation (Gobbo et al., 2021;Manenti et al., 2008;Ri es, Dronkers, & Knight, 2016).
Besides, specific action verb features are also related to areas of the motor cortex (Buccino et al., 2005;Pulvermü ller et al., 2005). Even though this semantic effect is quite established (see also section 3.2.3; e.g., Buccino et al., 2005;Pulvermü ller et al., 2005;Vukovic et al., 2017), what is less known is that adjacent regions of the motor cortex also respond to these semantic contrasts (e.g., SMA; Tremblay et al., 2012;Willems et al., 2011). Hence, with both control processes and specific action related features being processed by frontal regions, the lack of interaction may be explained by the existence of a distributed semantic network across the perisylvian cortex (Binder & Desai, 2011;Binder et al., 2009;Lambon-Ralph et al., 2017).

Limitations and future directions
In order to acquire the biggest sample size and because TMS has been shown to have both inhibitory as well as excitatory effects on cortical activity , we included eligible studies without analyzing possible effects of the specific TMS protocols or the direction of the induced effect. While this could be problematic because it has not been yet addressed whether both virtual lesion and excitatory protocols allow for comparable causal inferences, the majority of studies we included used inhibitory stimulation (61%). In any case, future work should disentangle potential differences between inhibitory and excitatory protocols. In addition, due to specific cortical regions being tested only with stimuli that targeted either morphosyntax or semantics, we were not able to perform meta-analyses for delineated cortical regions. Even though, stimuli contrasts clearly targeted either morphosyntactic or semantic functions, the two processing stages can hardly be teased apart, as even uninflected verbs (i.e., to play, to run) have been argued to engage morphosyntactic functions (den Ouden et al., 2009;Meltzer-Asscher, Mack, Barbieri, & Thompson, 2015;Thompson et al., 2007).
Regarding future research, morphosyntactic variables of verbs and sentences should be investigated further by including lexical variables such as frequency, different argument structure complexities, instrumentality, name relatedness to the noun, etc. To investigate the specificity of an effect and establish dissociations, more than one cortical region needs to be targeted. Tractography of white matter anatomy may be a promising method to target connected cortical regions (Reisch et al., under review;Ntemou, Ohlerth, et al., 2021). In addition, more TMS studies are needed to shed light on the role of the left IPS in sentence processing and dissociate its contribution with working memory functions of inferior parietal regions.

Conclusion
The present review and meta-analysis set out to fill a gap in the literature by synthesizing findings from TMS studies that employed verb and sentence experimental stimuli. The systematic review demonstrates that TMS studies with verbs and sentences provide crucial insights into the neural correlates of language processing. Findings concerning frontal stimulations support the claims put forward in the lesion and neuroimaging literature that inferior and middle frontal areas participate in the morphosyntactic processing of verbs, whereas probing the motor cortex has demonstrated behavioral effects on action verb processing. Further, TMS studies support the claims that regions of the temporoparietal cortex are connected not only to semantic access and meaning integration, but also to thematic role assignment. The metaanalysis showed that, although TMS has a small overall effect size on verb processing, it is a method capable of establishing double dissociations via the stimulation of multiple areas while using the same linguistic stimuli. However, it did not provide conclusive evidence that frontal regions are exclusively engaged for morphosyntax and temporoparietal regions for semantics.