Functional connectivity during orthographic, phonological, and semantic processing of Chinese characters identifies distinct visuospatial and phonosemantic networks

Abstract While neuroimaging studies have identified brain regions associated with single word reading, its three constituents, namely, orthography, phonology, and meaning, and the functional connectivity of their networks remain underexplored. This study examined the neurocognitive underpinnings of these neural activations and functional connectivity of the identified brain regions using a within‐subject design. Thirty‐one native Mandarin speakers performed orthographic, phonological, and semantic judgment tasks during functional magnetic resonance imaging. The results indicated that the three processes shared a core network consisting of a large region in the left prefrontal cortex, fusiform gyrus, and medial superior frontal gyrus but not the superior temporal gyrus. Orthographic processing more strongly recruited the left dorsolateral prefrontal cortex, left superior parietal lobule and bilateral fusiform gyri; semantic processing more strongly recruited the left inferior frontal gyrus and left middle temporal gyrus, whereas phonological processing more strongly activated the dorsal part of the precentral gyrus. Functional connectivity analysis identified a posterior visuospatial network and a frontal phonosemantic network interfaced by the left middle frontal gyrus. We conclude that reading Chinese recruits cognitive resources that correspond to basic task demands with unique features best explained in connection with the individual reading subprocesses.


| INTRODUCTION
Neuroimaging studies in the past three decades have well identified a left-lateralized brain network for reading, comprising the inferior frontal (or Broca's area) and occipitotemporal and temporoparietal regions (which coincide with Wernicke's area) (see Price, 2012 for a review).
This neural network for reading overlaps largely with the spoken language network and is argued to be culturally universal (e.g., Feng et al., 2020;Nakamura et al., 2012;Rueckl et al., 2015), although some studies have shown that the neural mechanisms underlying reading and reading impairments may be shaped by language-specific factors (Bolger et al., 2005;Paulesu et al., 2000;Siok et al., 2008;Siok et al., 2004;Siok et al., 2009;Tan et al., 2005;Tan et al., 2003;). For example, Chinese readers seldom use the left temporoparietal region, particularly the left superior temporal gyrus (STG), but more robustly recruit the left middle frontal gyrus (LMFG), left superior parietal lobule (SPL), bilateral fusiform gyri (FFG), and bilateral lingual gyri and cuneus to process orthophonological information Wu et al., 2012). These divergent findings may arise due to variability in the tasks used that tap into different aspects of reading, such as visuo-orthographic, phonological, and semantic processing. Indeed, studies reporting universal reading networks used implicit reading tasks that measured automatic reading processing (Krafnick et al., 2016) or simple decision tasks that involved language-general semantic processing (Dehaene & Cohen, 2011;Hu et al., 2010;Nakamura et al., 2012;Rueckl et al., 2015), whereas studies claiming cultural-specific reading networks examined phonological activation during visual word identification (Siok et al., 2004;Siok et al., 2008;Tan et al., 2005). Since reading is a complex phenomenon involving multiple cognitive and linguistic processes, these inconsistent findings highlight the need to delimit the neural networks underlying various aspects of reading processing. The purpose of this study was to examine the neurocognitive bases of the three constituents of reading (i.e., orthography, phonology, and meaning) in Chinese.
Spoken language uses sound to convey meaning. Thus, language acquisition involves the mapping of sound to meaning, rendering the brain networks for phonological and semantic processing in young children highly overlapped and connected (e.g., Mathur et al., 2020;Weiss et al., 2018). It has long been reported that languages used across the world are commonly supported by a left-lateralized language network comprising Broca's and Wernicke's areas. When children start learning to read, they have to associate the newly acquired written symbols with the already existing phonological and semantic constituents , resulting in the wiring and rewiring of the visual cortex with the spoken language network (Dehaene, 2009;Orton, 1925;Schlaggar & McCandliss, 2007;Skeide et al., 2017). Here, the learning process and the resultant neural mechanisms for reading may be dependent on the visual configuration of the written symbols used and the manner of print-to-sound mapping.
In alphabetic scripts, a grapheme (the smallest written unit, either a letter or letter cluster) represents a phoneme (the smallest contrastive speech unit) in a quasi-systematic manner, with word meanings derived through the mediation of letter-sound correspondence. Thus, processing alphabetic scripts requires the dorsal circuit, including the left temporoparietal regions, for rule-based phonological processing and the ventral circuit, including the occipitotemporal regions, for memory-based word recognition (Pugh et al., 2001). The Chinese writing system does not represent individual phonemes of a word. Each Chinese character encodes a morphosyllabic unit, and its pronunciation is usually underdetermined from the stroke patterns or character components. Accordingly, character pronunciation must be memorized by rote and cannot be deduced by recourse to a prelexical, rule-based phonological computation process that is available in alphabetic systems. Thus, not surprisingly, the left temporoparietal region, particularly the STG, is seldom observed to be involved in Chinese orthography-to-phonology processing . This type of memory-based phonological process, known as addressed phonology, is also available in alphabetic systems for reading words with high frequency or irregular spelling patterns (Coltheart, 2006), and reading written scripts in both systems activates the ventral reading circuit. However, contrary to alphabetic systems, phonology does not mediate meaning access in Chinese character reading, as homophones abound , although phonological activation precedes semantics (Tan et al., 1995;Tan et al., 1996) and is obligatory during meaning access (Spinks et al., 2000). This one-to-many relationship between phonology and meaning may weaken the connection between them, whereas the mostly one-to-one mapping between orthography and phonology may make these two processes more closely associated . Based on the above analysis, the reading network for Chinese reading may involve two possible patterns of relationships among orthography, phonology, and meaning: phonology and meaning are more integrated because this relationship is shaped by spoken language experience, or phonology and orthography are more associated because form-form relations are more reliable than form-meaning relations . A second purpose of this study was to examine these two possibilities. We examined the relationships among the three networks by depicting the relative strength in activation and connectivity among specialized regions of interest (ROIs).
One of the often reported peculiarities of neural activation during Chinese reading is activation in the left dorsolateral frontal regions.
We addressed two questions. The first question concerns the cortical networks underlying individual types of processing in Chinese reading: Are there specialized and distinct regions for orthographic, phonological, and semantic processing? Studies comparing these processes suggested that the left inferior frontal gyrus (IFG), FFG, and SPL are the loci for task-specific activations (Booth et al., 2006;Guo & Burgund, 2010;Kuo et al., 2004;Liu et al., 2006;Liu et al., 2009). However, these studies either compared only two of the processes (Booth et al., 2006;Cao et al., 2009;Kuo et al., 2004), focused on specific ROIs (Guo & Burgund, 2010) or did not fully report on all possible comparisons . To fully understand the functional specialization of the three networks, a more powerful yet stringent direct comparison among the three processes using whole-brain analysis is required. In this study, we utilized a withinsubject design for the between-task comparisons. As no studies to date have attempted a connectivity-based comparison of the three processes, we also probed into functional connectivity among the orthographic, phonological, and semantic processing networks in Chinese speakers. Next, we asked how the three networks are interrelated among one another. In alphabetic languages, as the pronunciation of a word can be derived from the spelling through grapheme-to-phoneme conversion (GPC), orthographic and phonological processing cannot be fully disentangled. This suggests that these two networks should be similar. Nevertheless, previous studies have demonstrated segregation of the two networks. Phonological processing of alphabetic languages shows activation in the left supramarginal gyrus (SMG), STG, and IFG (e.g., Booth et al., 2002;Paulesu et al., 1996;Pugh et al., 2001), while orthographic processing (usually in terms of spelling judgment) elicits activation in the bilateral FFG and SPL (Booth et al., 2002;Cohen et al., 2000;Petersen et al., 1990;Tagamets et al., 2000). Similarly, phonological and semantic networks have been found to be distinct in the left IFG (Poldrack et al., 1999), middle temporal cortex, SMG, and precentral gyrus (Price et al., 1997). The above descriptions provide two possible types of relationships between the neural networks of the subcomponents of Chinese reading, and this study determined whether Chinese character reading follows one of those patterns.
Component, homophone, and synonym judgment tasks were used to probe into orthographic, phonological, and semantic processing, respectively. The phonological and semantic tasks were paired with a visual-complexity-matched font-size judgment baseline task to account for activations arising from orthographic analysis. The orthographic task was paired with a line pattern judgment baseline task to extract the activation for finer visuo-orthographic analysis of the internal structure of Chinese characters. The activation and connectivity profiles of the three processes were analyzed by whole-brain, ROI, and ROI-to-ROI connectivity approaches. It was predicted that regions specialized for a particular process will exhibit stronger activation levels and interconnections among themselves, hence forming a functional unit. Commonly activated regions that are not specialized for any particular task, probably the LMFG, may act as an integration hub that connects separate processes in Chinese reading.

| Participants
Here, 31 native Mandarin speakers were recruited at Peking University (21 males and 10 females; mean age = 20.9 years, SD = 1.68 years). All participants were undergraduate or graduate students, had normal or corrected-to-normal vision and were free of any neurological or psychiatric disorders. All participants except one were right-handed as judged by the Edinburgh handedness inventory (Oldfield, 1971). The one exception obtained a score of +40, which is considered marginal between ambidextrous and right-handed. The activation maps of this participant showed a typical left-lateralized pattern and were included in the group analysis. Written informed consent was obtained from each participant, and the study was approved by the Human Research Ethics Committee at the University of Hong Kong. All scanning protocols were approved by the Peking University Institutional Review Boards.

| Materials
A component judgment task, a homophone decision task, and a synonym judgment task were used to probe the orthographic, phonological, and semantic processing of Chinese character reading, respectively, in different runs ( Figure 1a). For each task, 28 pairs of characters with shared components, homophones, or synonyms were displayed, and the participants were asked to judge if the two characters shared the same component, pronunciation or meaning in the orthographic, phonological, and semantic tasks, respectively. An extra 28 pairs of characters were included in each task as fillers to balance the number of correct yes and no responses. The frequency and stroke number of the characters were matched across the three tasks (Table 1). In each trial, a pair of characters were simultaneously presented above and below a fixation cross for 2 s, followed by a 500-ms blank. Trials were organized into blocks of 14 and were presented in the same pseudorandomized order for each participant. In the orthographic task, component judgment blocks were alternated with line pattern judgment blocks (Figure 1b) in which the participants judged whether two line patterns were the same. In the phonological and semantic tasks, experimental blocks alternated with font-size judgment blocks in which the participants judged if the two characters had the same size. A task cue of 2 s appeared at the beginning of each block ( Figure 1c). The three tasks were performed during three separate scans, which lasted for 310 s each.

| Stimuli presentation and behavioral data acquisition
Stimuli presentation and response data-logging were interfaced by Eprime 2.0. The participants lay in a supine position while seeing stimuli back-projected onto a screen through a mirror mounted in a 64-channel head coil. The participants indicated a "yes" response by pressing a button with their right index finger and a "no" response with their right middle finger. The participants were familiarized with all the tasks using materials that did not overlap with the experimental stimuli before the scans. The order of performing the three tasks was counterbalanced among the participants.

| Image acquisition
Functional magnetic resonance imaging (fMRI) data were acquired with a 3 T Siemens MAGNETOM Prisma scanner in the Centre for MRI Research at Peking University. High-resolution (0.5 Â 0.5 Â 1 mm 3 ) T1-weighted anatomical brain images were acquired using a three-dimensional magnetization-prepared rapid acquisition images with framewise displacement greater than 0.9 mm or with global BOLD signal changes greater than 5 SDs were identified for scrubbing (Nieto-Castanon, 2020). The functional images were then segmented and normalized to the ICBM standard template in MNI space at a resolution of 2 Â 2 Â 2 mm, and smoothed with an isotropic 8-mm full-width at half-maximum Gaussian kernel. Participants with head motion exceeding 1 mm within a single run were excluded.
In total, twenty-two scans from six participants were identified as motion-or signal-change outliers; the maximum number of scans discarded within the same run was 4. The maximum head motion of all participants was less than 1 mm; thus, no participants were excluded due to excessive head motion.

| Common and task-specific activations
Preprocessed images of individual participants were submitted to a first-level analysis using a general linear model. The experimental task blocks and the baseline blocks were modeled by two separate regressors that were convolved with a canonical hemodynamic response function. The six head motion parameters estimated during the motion correction step, independent regressors for the scrubbed scans and the global BOLD signal change were included to regress out motion artifacts and to remove the effects of motion and signal change spikes, respectively. The time series data were high-pass filtered at 128 s and modeled by FAST (Corbin et al., 2018). The group activation t-maps for each task were then obtained by performing second-level random-effect analysis on the contrast images between  the task and baseline blocks for each participant. To determine the regions common to the three processes, a conjunction analysis at the second level was conducted. The effects of the three tasks were tested against the conjunction null (Nichols et al., 2005). To discover the regions specific to individual tasks, the t-contrast maps obtained from the three tasks were entered into a one-way within-subject ANOVA with unequal variance. An uncorrected p value of .001 and a familywise error rate (FWE)-corrected p value of .05 were applied to the voxelwise and cluster-level statistical significance for all contrasts.
The ROI clusters were sorted by hierarchical clustering of their functional similarity. We also tested for connectivity differences among the three tasks at the cluster and ROI-to-ROI connection levels.

| Behavioral results
The accuracy rates and reaction times for the three tasks are reported in Table 2. A task-by-block-type (3 Â 2) factorial ANOVA was performed on the accuracy rates and reaction time in R. The main effects and interaction effect with the accuracy data were not significant

| fMRI contrasts
Due to the imbalance of sex, we first conducted an independentsample t test between males and females to check if there were sex effects. No clusters survived the task contrast comparison (voxellevel: p = .001, uncorrected; cluster-level: p = .05, FWE corrected).
fMRI data for all participants were pooled together for subsequent analyses.
Activation patterns and peak coordinates in individual tasks are reported in Figure 2a and the Supplementary materials (Tables S1-S3). The results of the conjunction analysis and withinsubject F contrasts are reported in Figure 2b and Table 3. Here, we focus on the commonalities and differences among the three tasks.
Consistent with previous meta-analyses, the conjunction analysis revealed that orthographic, phonological, and semantic processing of

| Activation profile in the left frontal region, bilateral FFG, and left S/MTG and SPL
In Figure 3, the contrast estimates from each task were plotted along the five preselected axes (three in the left frontal cortex and two in the bilateral fusiform regions). The corresponding statistics are reported in Table 4. Along the superior axis in the frontal area (sMFG, superior portion of the MFG), contrast estimates from all three tasks showed an inverted-U shape. At the most posterior and superior coordinate (À48, 0, 50; BA 6), activation during phonological processing was the strongest. Activation in all tasks peaked between (À48, 12, 41; BA 8) and (À48, 24, 32; BA 9) and then decreased. Activation levels during orthographic processing were consistently higher than those during phonological and semantic processing at the second anterior coordinates (i.e., À48, 36, 23; BA 9), while activation levels during phonological processing closely paralleled levels during semantic processing. Along the middle axis (MFG), activation during all tasks also showed an inverted-U shape. Orthographic processing was the strongest in the most posterior coordinate, although it was not statistically significantly stronger than the other two. Orthographic processing was surpassed by phonological and semantic processing as the coordinates moved rostrally and ventrally. Semantic processing exhibited a trend towards stronger activation compared to phonological processing and orthographic processing at (À48, 36, 2; BA 45) and (À48, 24 12; BA 45) respectively, though the omnibus-F tests were not significant. Along the most ventral axis (IFG), activation during phonological and semantic processing showed an inverted-U shape, while that of orthographic processing showed a gradually increasing trend when moving in rostral and ventral directions, despite its relatively low activation level. Activations were significantly stronger during both phonological and semantic processing than orthographic processing at (À46, 16, 0; junction of BA 44, 45 and 47). At (À46, 26, À7; BA 47), the pairwise t test showed that activation was stronger during semantic processing than both orthographic and phonological processing, though the F test was not statistically significant.
T A B L E 2 Behavioral data from the three fMRI tasks. SD, standard deviation   F I G U R E 3 Plots of contrast estimates from the three functional magnetic resonance imaging (fMRI) tasks at selected regions of interest (ROIs). Asterisks indicate a significant difference in activation levels at the ROI. †: p < .1; *: p < .05; **: p < .005; ***: p < .0005, Bonferroni corrected In the bilateral FFG, activation during orthographic processing showed an apparent increasing trend along the anterior-posterior axis. Its activation was significantly stronger than phonological and semantic processing at all coordinates. No comparisons between phonological and semantic processing reached statistical significance.
Again, there was a close parallel between the patterns of phonological and semantic processing.
Consistent with previous findings, significantly stronger activation during semantic processing was found in the left MTG (BA 21). In the left STG (BA 22), phonological and semantic activation were significantly stronger than orthographic activation in the pairwise comparison. Finally, activation was significantly stronger during orthographic processing than both phonological and semantic processing in the left SPL (BA 7).

| Connectivity analysis
ROI clusters, connectivity matrices among the eight ROIs during the three reading processes and the differences in connectivity patterns are presented in Figure 4, and the statistics can be found in the Supplementary materials (Table S4) (iii) BA 7 and 37. All cluster-level connections were significant during the homophone and synonym judgments.
As the clustering patterns were different across the three tasks, we first manually reset the clusters in the phonological and semantic conditions so that they followed the clustering pattern in the orthographic condition. The justification is that both BA 9 and 46, as well as BA 7 and 37, were always clustered together across tasks, while the grouping of BA 44, 47, 21, and 6 varied. With this grouping, significant connectivity differences (after FDR correction) were found

| DISCUSSION
To the best of our knowledge, this is the first study that examined the brain activation and connectivity for Chinese reading during orthographic, phonological, and semantic processing using a withinsubject design. There were two main findings from our results. First, the three processes were supported by a left-lateralized task-common network, which included the left dorsolateral prefrontal cortex, FFG, and bilateral medial frontal gyri. Second, orthographic processing recruited a specialized network that is distinct from phonological and semantic processing, while the phonological and semantic networks were highly similar. Differences across tasks were mainly exhibited in the form of differences in activation strength and connectivity pat- Previous studies discovered a developmental increase in activation in the left SPL, precuneus, MFG, bilateral MOG, and FFG during Chinese reading (Cao et al., 2010;Cao et al., 2015). Both the involvement of the right MOG and FFG have been suggested to be responsible for the visual analysis of the more complex and pictorial nature of Chinese characters (Cao et al., 2015;Guo & Burgund, 2010). It should be noted that activation of the FFG remained significant during phonological and semantic processing after the subtraction of activation during font-size judgment. This implies that the mid-fusiform area is involved in the finer visuo-orthographic analysis of characters and not the recognition of the general outline of characters, and may be responsible for the integration of visual, phonological and semantic information (Qin et al., 2021).
The left SPL and LMFG have been hypothesized to be a frontalparietal network that contributes to spatial working memory F I G U R E 4 Results of the connectivity analysis. (a) Connectomes and connectivity matrices associated with the three reading tasks.
Reported connections survived at a false discovery rate (FDR)corrected p value of .05. (b). Significant connectivity differences within/between clusters. The color bars show the relative effect sizes of the indicated connectivity (solid black curves) (Curtis, 2006;Ikkai & Curtis, 2011). Together with the FFG, these three regions exhibited significant connectivity during the orthography task. It can be inferred that orthographic processing in Chinese requires the recognition of the overall word form and the retention of the fine visuospatial details of stroke patterns. The concurrent activation of the FFG, and LMFG in the three tasks, together with the stronger connectivity between BA 7 and 37 during orthographic processing, suggests that orthographic processing is a nontrivial component in Chinese reading and requires cognitive support from an independent visuospatial network.

| Phonological processing
Significantly stronger activation during phonological processing was found in the left precentral gyrus (BA 6) and IFG (BA 44). Our results were consistent with most of the previous studies and meta-analyses (Chen et al., 2002;Kuo et al., 2004;Siok et al., 2003;Siok et al., 2020;Tan et al., 2005;Wu et al., 2012). BA 44, being part of Broca's area, is a well-documented region for speech sound processing, such as speech rehearsal (Price, 2012) and syllable counting (Poldrack et al., 1999). Our results reflected task requirements that the participants explicitly retrieve the pronunciation of the characters and the syllabic nature of Chinese characters.
Unlike some previous studies, we failed to find reliable evidence for the involvement of the STG/S. This region is in the proximity of the auditory cortex and has been identified with auditory processing, such as the analysis of speech and nonspeech sounds (Binder et al., 1996) and rapid frequency transitions (Poldrack et al., 2001).
Reliance on the IFG and not the STS/G for phonological processing can be ascribed to the principles of phonological decoding: GPC does not exist with Chinese characters, so pronunciations must be retrieved through addressed phonology. In fact, Chen et al. (2002) found both IFG and STG activation in the pinyin > character contrast but only IFG activation in the character > pinyin contrast. Cao et al. (2010) and Cao et al. (2015) also found a developmental decrease in the activation of the left and right STG, respectively, which was attributed to reduced phonological representation among adults. With the absence of GPC and weak phonological representation of Chinese characters, it is not surprising that the STG/S does not constitute a part of the phonological processing circuit.

| Semantic processing
Semantic processing was found to uniquely utilize the pars triangularis and pars orbitalis in the left IFG (BA 45/47) and MTG (BA 21).
Together with phonological processing, the activation level in the pars opercularis (BA 44) was also significantly stronger than that during orthographic processing. Consistent with both alphabetic and Chinese studies, these regions were commonly reported to be at the center of semantic retrieval, representation and comprehension (Bookheimer, 2002;Booth et al., 2006;Chee et al., 2001;Poldrack et al., 1999). With the exception of the task-common regions of the LMFG and SPL, we did not find other peculiarities of the Chinese semantic network.

| Functional segregation of the left lateral frontal cortex and the role of the LMFG
The ROI analysis revealed that the LMFG (BA 9/46) is specialized for orthographic processing in Chinese. The connectivity analysis further showed that BA 9 and 46 formed a functional unit that was consistent across tasks. In contrast, although BA 6, 44 and 47 are specialized for phonological and semantic processing, they were not consistently grouped together. We can, therefore, separate the left lateral frontal cortex into an orthographic network (LMFG) and a phonosemantic network (BA 6/44/47). During orthographic processing, we did not find connections between the visuospatial cluster (BA 7/37) and the phonosemantic cluster (BA 6/21/44/47). Instead, relatively strong connections were found between BA7 and BA9, BA9 and BA46, and BA46 and BA44/47 ( Figure 4a and Table S4). The strong connectivity between the SPL and LMFG is consistent with the proposal that the LMFG is a projection from the SPL that is responsible for spatial working memory (Kravitz et al., 2011). Together with its domain-general functions, such as working memory, executive control and attention, we hypothesize that the LMFG performs a major function in visuospatial analysis and acts as a platform for integrating visual and language functions, that is, intergrating the information from the visuospatial network and phonosemantic network. This hypothesis would explain the activation of the LMFG in various Chinese reading tasks, such as phonological (e.g., Kuo et al., 2004;Siok et al., 2003;Tan, Feng, et al., 2001) and semantic (e.g., Booth et al., 2006;Chee et al., 2000;Ding et al., 2003;Siok et al., 2004;Tan, Liu, et al., 2001;Wu et al., 2012) judgments. In the orthographic task, the LMFG itself performed its visuospatial analysis function, which led to a higher level of activation. Our results suggest that the LMFG is both an orthographic processing center and an interface between the visuospatial information and the phonosemantic networks.

| Implications for the neural networks associated with Chinese reading
The whole-brain activation patterns, ROI activation levels and ROI-to-ROI connectivity patterns suggest that orthographic processing vastly differs between phonological and semantic processing in Chinese. A dedicated network, which includes the left FFG, SPL, and MFG, is specialized for visuo-orthographic processing of Chinese characters. In contrast, there was a high similarity between the phonological and semantic networks. In the connectivity analysis, there was no obvious clustering of BA 6, 21, 44, and 47 and the within-cluster connections of these four regions were higher during phonological and semantic processing, implying that they were all involved in phonological and semantic processing. Additionally, in the ROI analysis, with the exception of the left precentral gyrus, stronger phonological processing than orthographic processing always implied stronger semantic processing. The activation pattern during phonological processing resembled that during semantic processing, although semantic processing elicited stronger activation of the left pars triangularis, pars orbitalis, and MTG. The close parallel between the phonology and semantic networks, with further specialization for semantic processing, suggests that phonological processing is always present when meaning is accessed, but the reverse is not true (i.e., semantic processing is not always performed when sounds are accessed). This pattern echoes the findings of behavioral studies that phonological activation is obligatory in Chinese reading (Spinks et al., 2000) while semantic processing is not.
In Chinese, most characters map to only one pronunciation but have many meanings, and many characters share the same pronunciation. Thus, even though phonology is activated early during character identification (Tan et al., 1995), it does not mediate access to meaning in Chinese. Therefore, character meaning is difficult to determine without context. This asymmetry in mapping is termed the determinacy principle ). During the phonological task, the participants were not forced to select a particular meaning while it was required in the synonym judgment task. As a result, we failed to identify semantics-specific regions during the phonological task, but extra resources from the semantic processing centers were recruited during the semantic task on top of the activation of the automatic phonological network. This explains why the semantic network contains all the major regions recruited during phonological processing.
In general, our findings reveal that orthographic processing in Chinese utilizes the bilateral FFG, left MFG and left SPL, whereas phonological and semantic processing recruits the left IFG (Broca's area) but not the left superior temporal region (Wernicke's area). These peculiarities should be interpreted with reference to the cognitive demand that the Chinese writing system imposes. In alphabetic languages, an indirect route by applying GPC is possible. The explicit decoding of graphemes into speech sound is closer to the auditory processing of language. Hence, the speech processing center (Wernicke's area) is involved in reading alphabetic languages. In Chinese, due to the absence of GPC, the phonology of a character is retrieved after visual form analysis. This phonological retrieval is believed to be done by the frontal regions. We do not claim that auditory processing does not play any role in Chinese. In fact, Broca's area was strongly activated during the phonological task, and a trend of stronger activation was exhibited in the STG during the phonological and semantic task than during the purely visual orthographic task (Table 4 and Figure 3).
However, the mapping principle in Chinese weakens the role of auditory processing and rendered activation in the STG nonsignificant. In fact, a meta-analysis of fMRI studies on dyslexic readers reported that only dyslexic readers of shallow orthographies demonstrate underactivation in the temporoparietal region (including STG/S), while dyslexic readers of deep orthographies do not (Martin et al., 2016). This lends support to the hypothesis that reading, or word decoding, engages the STG/S to different extents depending on the print-to-sound mapping principle.
Similarly, the stronger activation in the LMFG (BA 9/46) and SPL and their connectivity during the orthographic task should be interpreted with reference to the high visuospatial cognitive demand during Chinese reading rather than as a fundamentally different network from the occipitotemporal network for word form processing. Chinese characters are made up of basic stroke patterns that are assembled to form components, which are in turn organized into a square shape to form characters. Unlike alphabetic languages, in which a small set of letters (e.g., only 26 in English) are arranged linearly into words, there are approximately 560 components in Chinese (State Language Commission, 1998), and these components are packed into a twodimensional space. The distinction among components can be minute, such as 已 versus 己 (the 乚 stroke starts above the left end of the lower horizontal stroke in the first character, but it just touches the left end of the lower horizontal stroke in the second character) and 未 versus 末 (the upper horizontal stroke is shorter in the first character but longer in the second character). Moreover, different characters can be constructed from the same components but with different spatial arrangements, such as 棘 versus 棗 (the two 朿 's are arranged horizontally or vertically). These features tax resources for both visual perceptual skills (comparing the relative position and length of strokes) and spatial perceptual skills (comparing the relative position of components). Therefore, the right FFG is recruited for extra effort required for the visual analysis. In addition to recruiting the "what" ventral visual pathway from the visual areas (I/MOG) to the FFG for component identification, Chinese character recognition requires the "where" pathway (Mishkin et al., 1983;Ungerleider & Haxby, 1994) for the spatial analysis of components. Hence, the FFG and SPL formed a core functional unit supporting the visuospatial analysis of the characters. The brain network involved in processing a writing script is dependent on its linguistics features.
Our results highlight the importance of investigating the constituent components of reading separately. As reading is an umbrella term that encompasses more specific cognitive processes, it would be misleading to use a single reading task to represent the reading network of a language as a whole.
As a concluding remark, although our results appear to suggest that the network of Chinese reading differs from that of reading alphabetic languages, it should be stressed that this study did not include direct cross-language comparisons, so it would be premature to draw conclusions. To determine whether Chinese reading engages a universal brain network, future studies should compare different orthographies. However, to avoid conflating the constituent reading processes, the orthographic, phonological, and semantic processing components should be investigated simultaneously to fully elucidate the reading profile of the studied languages.

| CONCLUSION
We conclude that Chinese reading utilizes two major networks: the visuospatial network, which includes the left FFG, SPL, and MFG, to handle visuo-orthographic information and the phonosemantic network, which includes the left precentral gyrus, IFG, and MTG, to retrieve and process phonological and semantic information. We propose that the neural mechanisms underlying Chinese reading correspond to the cognitive demands imposed by its peculiar visual and print-to-sound mapping features. With its capacity in visuospatial working memory, the LMFG acts as an interface between the posterior visuospatial network and the frontal phonosemantic network, supporting the processing of visually demanding Chinese scripts. Our study demonstrates the importance of considering the constituents of a writing system when examining its reading networks.
ACKNOWLEDGMENT This research was supported by an internal research grant (#202011159012) from the University of Hong Kong.

DATA AVAILABILITY STATEMENT
Data are available upon request from the corresponding author.