The neural basis of semantic cognition in Mandarin Chinese: A combined fMRI and TMS study

Abstract While converging sources of evidence point to the possibility of a large‐scale distributed network for semantic cognition, a consensus regarding the underlying subregions and their specific function in this network has not been reached. In the current study, we combined functional magnetic resonance imaging (fMRI) and transcranial magnetic stimulation (TMS) methodology to investigate the neural basis of semantic cognition in Mandarin Chinese. In the fMRI experiment, strong activations were observed in left inferior frontal gyrus (IFG) and left middle temporal gyrus (MTG) for semantic judgment task. Moreover, functional connectivity was found from seed region left IFG to left MTG. Meanwhile, negative correlation between performance and extracted parameter estimates from left IFG to left MTG was detected in semantic task. Subsequent TMS stimulation over left IFG resulted in performance deficits in semantic judgment task, in contrast to other three sites: left MTG, right intraparietal sulcus (IPS) and a control site. We concluded that the neural basis of semantic processing for Mandarin Chinese closely resembled that for alphabetic languages such as English, supporting a language‐universal view on semantic cognition.

the neural basis of semantic cognition (Bolger, Perfetti, & Schneider, 2005;Tan, Laird, Li, & Fox, 2005;Wu, Ho, & Chen, 2012), in particular, it is not clear whether there is a language-specific or a language-general neural circuit of semantic processing across different languages. For example, in a meta-analysis study, Tan et al. (2005) found that the left temporoparietal cortex has a unique contribution for the conversions from the grapheme to phoneme in alphabetic languages. Whereas, left middle frontal gyrus (MFG) plays a special role in Chinese language processing. In another meta-analysis, a universal network across the writing systems was proposed by Bolger et al. (2005). What is more, some differential activations were also identified for Chinese processing, which was highly consistent with the meta-analysis found in Tan et al.'s (2005).
However, using functional magnetic resonance imaging (fMRI) in a semantic task with words written in cursive font, Nakamura et al. (2012) demonstrated that two universal neural circuits, one for reading by eye and one for reading by hand, had similar activation and showed identical patterns of activation and repetition in Chinese and French language groups. Their findings suggested that languages with distinctive features of orthographies only regulated a fixed set of invariant macroscopic brain circuits, supporting a cross-cultural commonality for semantic processing. More recently, another functional MRI study was conducted by Rueckl et al. (2015) to examine reading and speech perception in four highly different languages: Spanish, English, Hebrew, and Chinese. Using three complementary analytic approaches, they found small clusters of activation for Chinese in which speech-print correlation was greater compared with other alphabetic languages. However, those clusters were not located in regions such as left middle frontal which had been claimed to be unique for Chinese language by former studies (Bolger et al., 2005;Tan et al., 2005). Taken together, there is still lack of controversy about the neural basis of semantic processing for Mandarin Chinese.
Most of recent studies on Chinese semantic processing have been performed by fMRI (Dong, Nakamura, Okada, Hanakawa, & Fukuyama, 2005;Rueckl et al., 2015;Tan et al., 2000;Yu, Mo, Li, & Mo, 2015;Zhang et al., 2017;Zhao et al., 2014), but these techniques may not be the best way to study systematically the specific role of each particular language area. Evidently, fMRI has been used to measure the whole-brain activity to detect the cortical areas involved in certain brain undertakings. However, it cannot detect specific cortical area which plays a causal role in certain functions. Instead, by inducing focal and transient disruption of neural processing, the TMS technique is an effective tool to complement the results of fMRI (Pascual-Leone, Tormos, & Keenan, 1998;Pascual-Leone, Walsh, & Rothwell, 2000;Walsh & Rushworth, 1999). Thus, using a combined fMRI-TMS methodology, the present study aimed to investigate the neural basis of semantic cognition in Mandarin Chinese and the specific function of the implicated brain regions.
First, in the fMRI experiment, we sought to investigate the neural basis of semantic processing for Mandarin Chinese. In particular, we wondered whether there exists a similar large-scale distributed network of semantic cognition to those of alphabetic languages or whether there exists some unique cortical regions merely for Chinese semantic processing. Then, to clarify the different functional roles of these regions, we used online TMS to generate focal "virtual lesions" in healthy volunteers to reveal whether these regions have causal relationships to the processing of semantic cognition in Chinese.

| Participants
Twenty healthy individuals participated in the fMRI study (10 males, mean age, 21.4 ± 1.96, range from 20 to 24 years) and another 6 subjects, along with the 20 participants in fMRI study, were selected for the subsequent TMS experiments. Two subjects quitted the TMS experiment later for personal reasons. Thus, altogether 24 subjects participated in the TMS study (16 males and 8 females, mean age, 21.63 ± 2.02, from 20 to 25 years). With normal or corrected-to-normal vision, all subjects neither had any personal or family history of neurological or psychiatric illness nor took medication in the course of the experiment. Before the formal experiment, they were all informed of the potential side effects of TMS and they offered written informed consent.
All the subjects passed safety screening for TMS (Wassermann, 1998) and MRI. This study was approved by the local committee for the Protection of Human Subjects for the University of Electronic Science and Technology of China. The methods in our study were conducted in line with the approved guidelines and all experiments were in accordance with the declaration of Helsinki.

| Experimental session design
The experiments were carried out in 6 days (see Figure 1). On Day 1, participants received behavioral training for the semantic and numerical judgment tasks. Then, on Day 2 for the fMRI experiment, participants performed the same tasks and the functional neuroimaging results would be the basis for the location of the regions of interest (ROIs) to be stimulated in the subsequent TMS experiment. Finally, on the following Days 3-6, we applied online TMS at four cortical regions when participants performed the same tasks as those in the fMRI experiment. Three of the cortical regions were ROIs we got in the former fMRI experiment: left IFG, left MTG, and right IPS. The final region was Vertex as a control site, which was defined as position Cz according to the international 10-20-system for EEG (Gold, Chang, Wang, Zhu, & Juan, 2014;Jung, Bungert, Bowtell, & Jackson, 2016;Sandrini, Umiltà, & Rusconi, 2011;Smittenaar, Fitzgerald, Romei, Wright, & Dolan, 2013) at Brodmann area 6. The order of TMS stimulation for the four cortical regions in Days 3-6 were counterbalanced across all participants and separated by at least a week.

| Task
The presentation of task stimuli was performed by a PC with E-Prime 2.0 software (Psychology software; Psychology Software Tools). All participants performed a semantic judgment task and a numerical judgment task (see Figure 1). The semantic judgment task was adapted from the previous fMRI and TMS studies (Binney, Embleton, Jefferies, Parker, & Lambon Ralph, 2010;Cao, Bin, Li, & Yan, 2014;Pobric, Jefferies, & Lambon Ralph, 2007, 2010a, 2010b, 2010cVisser, Jefferies, Embleton, & Lambon Ralph, 2012). In this task, subjects were required to select a target word which is most related in meaning to a probe word from three words. Accordingly, each trial contained four written words: a probe word (e.g., 河流, river in English), the target word (e.g., 湖泊, lake in English), and two unrelated words (e.g., 医生, 广场, doctor and square in English). The four words used in each trial were matched for image ability (p = .42), word length (p = .65), and word frequency (p = .98). The numerical judgment task was also adapted from previous studies (Binney et al., 2010;Cao et al., 2014;Pobric et al., 2007;Pobric, Jefferies, & Lambon Ralph, 2010a, 2010b, 2010cVisser et al., 2012). Similar to the semantic judgment task, in each trial, a probe number (e.g., 374) was shown on upper half of the screen and three choice numbers (e.g., 319, 368, and 334) were presented underneath. Subjects were asked to choose which of the three choice numbers was closest in value to the probe number.

| fMRI procedure
Each participant performed two runs with mini-block design in the fMRI experiment. Within each run, there were 15 semantic task blocks and 15 numerical task blocks. Those task blocks were ordered in random and interleaved by fixed blocks lasting for 8 s. Within each task block, a trial began with 1 s fixation followed by a word or a number stimuli presented for 3 s. Each task block consisted of four trials.
Each run took 12 min. Subjects were required to select the target stimuli by pressing 1 of 3 designated buttons on a magnetic resonance compatible response box. Because of their small number of characters, Arabic numbers were presented in a larger font than semantic words (Arabic 1.6 × 1 visual angle; average of semantic words, 7.8× 0.8 ).

| fMRI acquisition
MRI scanning was done using a 3.0 T GE Sigma scanner at the High-Field Magnetic Resonance Brain Imaging Key Laboratory at the University of Electronic Science and Technology of China, Chengdu, China. We acquired functional MRI images using a gradient echo planar F I G U R E 1 Experimental tasks, session design and TMS stimulation sites. (a) Experimental tasks and trial sequence. Each trial began with a fixation cross with 1,000 ms. Thereafter, subjects were required to choose either the target word that was semantically related with the cue word or the target number that was numerically close to the cue number. 43 slices. We also acquired a high-resolution, whole-brain structural T1-weighted image with a magnetization-prepared gradient echo sequence. The scanning parameters were the following: TR = 1900 ms; uk/spm). For each participant, the first five volumes of each functional time series were discarded for signal stabilization. Images were corrected for head movement between scans by an affine registration. The remaining functional images were preprocessed including slice timing correction, three-dimensional motion correction, co-registration to individual T1 structural image, and normalization by DARTEL (Ashburner, 2007) to the Montreal Neurological Institute (ICBM152 brain template)

mm 3 ) and spatial smoothing with an 8 mm
Gaussian kernel (full-width at half maximum). Low-frequency signal drifts were removed with a temporal high-pass filter (default cutoff of 128 s).
No subject had more than 2 mm of translation or 2 rotation.
At the single subject level, we operated statistical analysis on the basis of the context-based general linear model (GLM) of the experiment. The individual trial onset sequence for semantic judgment task and numerical judgment task were obtained by convolving a canonical hemodynamic response function to form regressors of the design matrix (Gazzaley et al., 2007;Passaro et al., 2013;Robitaille et al., 2010;Yang, Fan, Wang, & Li, 2017).
At the group level, we first carried out paired t test to evaluate different activations between the semantic judgment task and the control judgment task (semantic-number) and vice versa (numbersemantic) and between each task and the fixation blocks (semanticrest and control-rest). Then, the following whole brain multi-subject analysis was conducted by using a random effects model with a onesample t test on the summary statistic. The statistical contrast maps were thresholded at p < .001 (corrected for false discovery rate [FDR]) to control for multiple comparison (Benjamini & Hochberg, 1995;Genovese, Lazar, & Nichols, 2002).

| Defining ROIs for TMS targeting
We selected the ROIs according to contrast maps. All ROIs were identified on the peak activation of each cluster with 6-mm radius sphere.
Bold signal changes were analyzed according to the mean signal intensity of each ROI with the fixation epochs as a baseline.
Two ROI regions were localized at the left hemisphere of the frontal and temporal cortices, which were usually considered to be important in semantic judgment task. The average MNI coordinates across our sub-

| TMS stimulation
In the TMS experiment, online triple-pulse TMS (tpTMS) was applied at the stimuli onset to transiently disrupt the processing of semantic cognition. According to the ROIs in the fMRI experiment, online tpTMS stimulation was administered with a Magstim super rapid magnetic stimulator and a figure-of-eight coil (diameter 70 mm) (Magstim Company Limited, Whiteland, United Kingdom). TMS session was conducted in accordance with the published safety guidelines (Rossi et al., 2009;Wassermann et al., 1996).
The coil was placed tangentially to the skull with the coil handle oriented perpendicular to the target cortex, guided by the online BrainSight frameless stereotaxy system (BrainSight Frameless, Rogue Research, Montreal, Canada). Triple-pulse TMS was stimulated with at a pulse interval of 25 ms (40 Hz) at 100% resting motor threshold (RMT) as that used in our previous TMS studies on language comprehension (Zhang et al., 2018). tpTMS has proved to be more suitable for induction of measurable behavior effect size than single pulse TMS in higher cognitive functions and has thus been widely used in language cognition (Sack, 2005;Schuhmann, Schiller, Goebel, & Sack, 2009, 2012Zhang et al., 2018). The RMT was identified immediately before the delivery of tpTMS. It was set as the stimulation intensity which had 50% chance to produce motor-evoked potentials larger than 50 μV peak-to-peak in the contralateral first dorsal interosseous muscle, following stimulation over the hand area of the participant's right motor cortex Yan, Wei, Zhang, Jin, & Li, 2016;Zhang et al., 2018). EMG activity was recorded with 9-mmdiameter Ag-AgCl surface cup electrodes and displayed on a conventional electromyography, which was also used to trigger the stimulator (Magstim Company Limited). Mean stimulation intensities for left IFG, left MTG, right IPS, and Vertex were 32.2 ± 5.0%, 32.3 ± 5.5%, 32.5 ± 5.5%, and 32.1 ± 5.0% of total stimulator output, respectively.

| Statistical data analysis
Statistical data analysis was done using SPSS Statistics Release 20 (IBM, Somers, NY) GLM. Reaction time (RT) and accuracy rate (ACC) was measured. We used repeated measures ANOVA (Bonferroni corrected) to compare RT and ACC, with stimulation sites (IFG, MTG, IPS vs. Vertex) and tasks (semantic judgment task and numerical judgment task) as within-subject factors. Post hoc t test, using Bonferroni corrected for multi comparisons, was conducted to compare the RT and ACC across different TMS sites within each task.
The adjusted p value is .0083 for significant threshold.

| Data availability
The authors confirm that all data underlying the findings are fully available without restriction. All relevant data are within the Supporting Information files.

| RESULTS
We combined fMRI and TMS methodology to investigate the neural basis of semantic cognition in Mandarin Chinese. First, in the fMRI experiment, participants performed semantic and numerical judgment task in Mandarin Chinese. According to each subject's fMRI data, we identified three sites for TMS stimulation: left IFG, left MTG, right IPS, and a control site Vertex. Then, we delivered online triple-pulse TMS (tpTMS) stimulation to each of these regions on separate days.

| Whole-brain analysis
The data were initially processed as a whole brain cluster analysis, comparing semantic judgment tasks with numerical judgment tasks (semantics > numbers). The statistical image was evaluated at p < .001 (FDR-corrected). The clusters that exceeded 30 voxels were shown in Table 1 and were also shown on the color scale in Figure 2. Activation was achieved in the ventrolateral frontal cortex of the left hemisphere, including the pars triangularis (BA45), the pars orbitalis (BA47), and also parts of the pars opercularis (BA44). We also obtained a cluster in the left MTG of the left hemisphere. In addition, a large cluster extending to the left and right occipital lobe (BA17/18) was observed, which may reflect the greater visual processing required by orthographic over digit stimuli or even semantic feedback to early visual areas (Hon et al., 2009 Contrasting the numerical task against the semantic task (numerical > ), we found activation in the bilateral inferior parietal lobe, the right precuneus, the right frontal cortex (middle/superior gyrus), and the right superior occipital gyrus (corrected to p < .001 using FDR).

| Functional connectivity analysis
To explore whether any observed core regions worked in concert with other regions as a network for cognitive processing, we used general- According to the whole brain analysis, four ROIs (IFG, MTG, IPS, and Vertex) were taken as seed regions with a 6 mm radius sphere centered at the peak activation coordinates. To test the network of cognitive processing, two PPI contrasts were set: (semantic task > numerical task) and (numerical task > semantic task).

| TMS results
The TMS-induced effects in RT and ACC were shown in Figure 5 for semantic and numerical judgment task, respectively. and Vertex). Mean RT and ACC for the semantic and numerical judgment task at four TMS stimulation sites are listed in Table 2.

| TMS-induced changes in RT
Overall, the two-factorial ANOVA revealed a main effect of task

| TMS-induced changes in ACC
A two-factorial ANOVA with task (semantic task vs. numerical task) and TMS sites (IFG, MTG, IPS, and Vertex) as the two within-subject factors revealed a significant main effect of task [F (1, 23) = 108.611, p < .05], indicating that ACC was significantly higher when subjects made numerical judgments compared to semantic judgments. Exploratory analyses also demonstrate a significant main effect of TMS sites

| DISCUSSION
The present study aimed to investigate the neural basis of semantic cognition in Mandarin Chinese and the specific functions of its underlying subregions using a combined fMRI-TMS methodology. First, in the fMRI experiment, strong activation in left IFG and left MTG were detected in semantic judgment task, coupled with significant functional connectivity between these regions. Meanwhile, functional connectivity analysis between left IFG and MTG showed strong correlation with performance in the same task. Interestingly, no significant activation was found in left MFG and superior parietal gyri, which were claimed to be unique for Chinese language processing. Moreover, TMS stimulation over left IFG largely disrupted performance in semantic task, compared to the other three sites: left MTG, right IPS, and Vertex. However, no such deficits were found at left MTG in the same task.
More recently, studies applying TMS on healthy participants were conducted to investigate the causal role of left IFG in semantic cognition (Whitney et al., 2011). In a recent TMS study, using material in English, Whitney et al. (2011) investigated the specific roles of left IFG and MTG in semantic processing in two different manipulations of semantic cognition (associative strength and feature selection). Their data analysis revealed that left IFG along with MTG support both the controlled semantic retrieval underpinned by a bottom-up automatic spreading activation mechanism and the feature selection of semantic knowledge which involved a more top-down executive mechanism.
With similar semantic retrieval task, the work presented here confirmed not only the recruitment but also the causal role of left IFG for  (Corbett et al., 2009;Noonan et al., 2010), as well as functional neuroimaging studies of healthy individuals (Davey et al., 2016;Jefferies & Lambon Ralph, 2006;Schwartz et al., 2009) have also noted that instead of a passive store for semantic attributes, left MTG may be involved in the strategic retrieval of semantic information.
First, in line with the above fMRI studies on semantic control at left MTG, we identified greater activation for Chinese semantic judgment task at this region along with inferior regions of the left frontal context. In addition, PPI analysis revealed strong functional connectivity between left IFG and MTG, coupling with correlation with performance in semantic task, further supporting the view that left MTG works in concert with left prefrontal cortex in order to allow strategic access of semantic information (Friederici, 2009). Therefore, the present fMRI findings in Mandarin Chinese are consistent with the recruitment of MTG, alongside IFG, for semantic processing (Badre et al., 2005;Davey et al., 2015Davey et al., , 2016Noonan et al., 2013;Poldrack et al., 1999;Snijders et al., 2009).
However, in the online TMS results study, TMS-induced changes for semantic selection task at left MTG were not observed as expected. One possible interpretation for the discrepancy between fMRI and TMS results may lie in the different contribution of those sites in semantic control network. Davey et al., (2016) investigated the different functional roles of IFG, MTG, and IPS in semantic cognition with a comparison with the meta-analysis of Noonan et al. (2013). Their data provided converging evidence for the three components of semantic cognition: (a) default mode network (DMN) preferentially for the automatic spreading activation with semantic representation, which is supported by ATL and other regions (Jackson, Hoffman, Pobric, & Lambon Ralph, 2016;Lau, Gramfort, Hämäläinen, & Kuperberg, 2013;Power & Petersen, 2013;Wirth et al., 2011); (b) multiple-demand executive network (MDN) (Duncan, 2010) for topdown allocation of attention and usually supported by frontoparietal control systems (Power & Petersen, 2013) (Gardner et al., 2012;Jefferies, Baker, Doran, & Ralph, 2007).
Likewise, it is reasonable that TMS-induced effects over left MTG were not observed due to the less contribution of MTG (compared with that of left IFG) in the process of semantic retrieval. Thus, our analysis is consistent with the studies on alphabetic languages that left MTG, coupling with left IFG, are essential components of a distributed cortical network underpinning executive semantic processing.

| CONCLUSION
Our study investigated the neural basis of semantic cognition in Mandarin Chinese and the specific functions of its underlying subregions using a combined fMRI-TMS methodology. We have demonstrated that left IFG specifically plays a causal role in semantic cognition, along with subordinate recruitment of other regions such as MTG. These findings are convergent with others studies on alphabetic languages.
Therefore, we propose that the neural basis of semantic processing for Mandarin Chinese closely resembled that for alphabetic languages such as English, supporting a language-universal view on semantic cognition.

DATA AVAILABILITY STATEMENT
The authors confirm that all data underlying the findings are fully available without restriction. All relevant data are within the Supporting Information files.