Content area
Previous studies have shown that the left temporoparietal junction (TPJ) plays a critical role in word reading. Nevertheless, there is still controversy surrounding the phonological and semantic functions of the left TPJ. The parietal unified connectivity-biased computation (PUCC) model posits that the function of the left TPJ depends on both the neurocomputation of this local area and its long-range connectivity. To clarify the specific roles of different TPJ subregions in phonological and semantic processing of Chinese characters, the present study used connectivity-based clustering to identify seven subdivisions within the left TPJ, and conducted comprehensive analyses including functional and structural connectivity, univariate and multivariate analyses (i.e., representational similarity analysis, RSA) on multimodal imaging data (task-state fMRI, resting-state fMRI, and diffusion-weighted imaging [DWI]). Functional and structural connectivity analyses revealed that the left anterior TPJ had stronger connections with the phonological network, while the left posterior TPJ had stronger connections with the semantic network. RSA revealed that the left anterior and posterior TPJ represented phonological and semantic information of Chinese characters, respectively. More importantly, the phonological and semantic representations of the left TPJ were respectively correlated with its functional connectivity to the phonological and semantic networks. Altogether, our results provide a more elaborate perspective on the functional dissociation of the left anterior and posterior TPJ in phonological and semantic processing of Chinese characters, and support the PUCC model.
1 Introduction
Reading relies on orthographic, phonological, and semantic processing of the written words ( Harm and Seidenberg, 2004; Seidenberg, 2005). At the neural level, the reading network primarily comprises three areas: the left ventral occipitotemporal cortex (vOTC), left temporoparietal junction (TPJ), and left inferior frontal gyrus (IFG) ( Caffarra et al., 2021; Dehaene et al., 2015; Taylor et al., 2013; Turker et al., 2023a). The left vOTC, which contains the visual word form area (VWFA), is critical for orthographic processing ( Cohen et al., 2002; Dehaene and Cohen, 2011), while the left TPJ and IFG contribute to high-level phonological and semantic processes ( Cao et al., 2017; Humphreys et al., 2021; Mei et al., 2015a; Oberhuber et al., 2016; Rapp et al., 2016; Taylor et al., 2013; Yi et al., 2019).
The left TPJ, located at the intersection of the left temporal and parietal lobes, is approximately aligned with Brodmann areas 40 and 39, and encompasses the left angular gyrus (AG, von Economo area PG) and supramarginal gyrus (SMG, von Economo area PF) ( Igelström et al., 2015; Schurz et al., 2017). In neuroimaging studies, various atlases have been used to define this region, such as the Talairach and Tournoux atlas ( Lancaster et al., 2000), the Eickhoff-Zilles Anatomy Toolbox ( Eickhoff et al., 2005), the Automated Anatomical Labeling (AAL) atlas ( Tzourio-Mazoyer et al., 2002), and the Harvard-Oxford Atlas ( Desikan et al., 2006). In terms of function, the left TPJ has been found to be involved in a variety of high-level cognitive processes, including language, memory, attention, spatial cognition, and self-awareness ( Seghier et al., 2013; Wang et al., 2017). A large number of studies have revealed that patients with lesions in the left TPJ exhibit impairments in word reading (e.g., Ding et al., 2020; Roux et al., 2014; Sakurai et al., 2010), and individuals with dyslexia exhibit hypoactivity in this region during reading ( Aguilar et al., 2018; Phinney et al., 2007; Richlan et al., 2011; Zhang et al., 2022), indicating that the left TPJ plays a critical role in word reading.
Despite the consensus on the importance of the left TPJ in reading, there is still controversy about the specific functions of the left TPJ in word reading. Previous research has suggested that this area contributes to multiple linguistic processes, including orthographic processing ( Phinney et al., 2007), phonological awareness ( Gabrieli, 2009), grapheme-to-phoneme mapping ( Boros et al., 2016; Braun et al., 2015; Glezer et al., 2016; Price, 2012), and semantic retrieval ( Binder et al., 2009; Davey et al., 2015). Subsequent research examined the precise functions of the left TPJ's subregions and found that the anterior and posterior subregions (i.e., the left SMG and AG) played distinct roles. For example, Numssen et al. (2021) utilized data-driven clustering to divide the inferior parietal lobes (IPL, comprising the SMG and AG) into anterior and posterior subregions in each hemisphere. They found robust activations in the right SMG during attentional reorienting tasks (invalid vs. valid conditions), in the left AG during lexical decision-making tasks (words vs. pseudowords), and in the right AG during perspective-taking tasks (false belief vs. true belief). As for the roles of TPJ subregions in word reading, brain imaging, non-invasive brain stimulation studies, and meta-analyses have revealed that the left posterior TPJ is primarily involved in semantic processing ( Binder et al., 2009; Devereux et al., 2013; Graves et al., 2010; Hartwigsen et al., 2016; Noonan et al., 2013; Seghier et al., 2010; Sliwinska et al., 2015; Turker et al., 2021, 2023a), while the anterior TPJ is primarily engaged in phonological processing ( Hartwigsen et al., 2010, 2016; Sliwinska et al., 2015; Stoeckel et al., 2009; Turker et al., 2021). In addition, there is also evidence that the left posterior TPJ contributes to the process of orthography-to-phonology mapping during word reading (e.g., Tainturier and Rapp, 2003; Tsapkini and Hillis, 2015), and recent studies using multivariate analysis techniques have further found that the anterior TPJ encodes both phonological and semantic information during word reading (e.g., Gao et al., 2022; Graves et al., 2023; Montefinese et al., 2021). Recent studies have also found that the anterior TPJ shows higher activation for pseudowords than words, while the posterior TPJ shows higher activation for words than pseudowords ( Numssen et al., 2021; Turker et al., 2023b, 2023c).
However, the specific role of the left TPJ in Chinese reading is less clear. Researchers have considered that the left TPJ plays different roles between alphabetic and logographic word reading ( Siok et al., 2008; Tan et al., 2005). The left TPJ is believed to subserve the transformation of graphemes to phonemes in alphabetic word reading ( Boros et al., 2016; Glezer et al., 2016; Price, 2012), but it is less involved in the phonological processing of Chinese words ( Bolger et al., 2005; Cao et al., 2009; Oberhuber et al., 2016; Siok et al., 2008; Tan et al., 2003, 2005). Some studies even failed to detect activation of the left TPJ in Chinese reading tasks ( Cao et al., 2009; Liu et al., 2022; Siok et al., 2003; Tan et al., 2003).
These inconsistent results could be due to the following reasons. First, the left TPJ is a structurally and functionally heterogeneous region with multiple cytoarchitectonic subdivisions ( Achal et al., 2016; Caspers et al., 2006, 2008; Humphreys et al., 2022; Igelström et al., 2015; Nelson et al., 2010; Schurz et al., 2017). Most studies investigating the role of the left TPJ in word reading have divided the left TPJ into subregions (e.g., the left SMG and AG) based on atlas-registered approximate segmentation for automated labeling. However, the boundaries of the designations do not completely align among atlases ( Devlin and Poldrack, 2007), and furthermore, dividing the left TPJ into only two or three subregions might not have been sufficient due to the large surface area of TPJ. Therefore, the present study performed a fine-grained data-driven functional partitioning to explore the functions of different subregions of the left TPJ in Chinese word reading.
Second, previous studies on Chinese reading using univariate activation analysis (e.g., Bolger et al., 2005; Liu et al., 2022; Tan et al., 2003; Zhang et al., 2022; Zhu et al., 2014) were unable to examine the nature of neural representations. The multivariate approach (e.g., representational similarity analysis, RSA) can be used to detect fine-grained information of neural patterns due to its high sensitivity to stimulus features ( Kriegeskorte et al., 2008). However, previous studies adopting multivariate approaches (e.g., RSA) mostly used a relatively small set of materials (34 to 80 words) ( Fischer-Baum et al., 2017; Li et al., 2022; X. Liu et al., 2022; Qu et al., 2022; Xu et al., 2018; Zhao et al., 2017). To overcome this limitation, the present study used a relatively large set of materials (i.e., 198 Chinese characters, with each character being one word), and combined univariate and multivariate analyses to decode the left TPJ's function in Chinese reading.
Third, the parietal unified connectivity-biased computation theory (PUCC) posits that the function of the left TPJ depends not only on the general neural computations of this region, but also on its connectivity with other regions ( Humphreys et al., 2015, 2017; 2020, 2021). The left TPJ is a complex region through which various white matter fibres pass, such as arcuate fasciculus, superior longitudinal fasciculus, and vertical occipital fasciculus ( Martino et al., 2013). These fibres are associated with phonological processing, auditory comprehension, and articulatory processing ( Kellmeyer et al., 2013; Nakajima et al., 2020; Saur et al., 2010; Schotten et al., 2014; Vanderauwera et al., 2015). In addition to these structural connections, the left TPJ is also functionally connected with the frontal-temporal reading-related network ( Achal et al., 2016; Humphreys et al., 2022; Yeo et al., 2011). Nevertheless, no research has clarified the phonological and semantic roles of the left TPJ from the perspective of its connections with reading networks. Besides, there is increasing evidence that multimodal brain imaging approaches can provide a more comprehensive understanding of brain functions compared to unimodal approaches due to the complementary nature of information hidden in various modalities (for a review, see Calhoun and Sui, 2016). The present study functionally defined phonological- and semantic-specific networks, and examined the left TPJ's functional and structural connectivity patterns with the two networks. We combined the characteristics of connectivity and local neural patterns of the left TPJ to comprehensively explore its functions in Chinese character reading.
In sum, to systematically investigate the functions of different TPJ subregions in phonological and semantic processing of Chinese characters, the present study conducted resting-state functional connectivity (rs-FC) analysis, structural connectivity (probabilistic fibre tracking) analysis, activation analysis, and representational similarity analysis (RSA) on multimodal imaging data (task-state fMRI, resting-state fMRI, and diffusion-weighted imaging [DWI] data). Specifically, we first used the phonological and semantic judgement tasks to identify the phonological- and semantic-specific networks, assessed functional and structural connectivities between the left TPJ subregions and the two networks, and used connectivity-based clustering to identify seven subregions within the left TPJ. Then, we examined the activation patterns of the left TPJ subregions in phonological and semantic tasks. Finally, we used a large set of stimuli (i.e., 198 words) and RSA to decode the phonological and semantic information underlying the neural representation of the left TPJ subregions.
2 Materials and methods
2.1 Participants
Two groups of native Chinese speakers participated in this study (
Fig. 1
All participants were right-handed according to the Edinburgh Inventory ( Snyder et al., 1993), and their eyesight was normal or corrected to normal. None of them had a previous history of head injuries, neurological disorders, or psychiatric illnesses. Written consent was obtained from them before the experiment. The Institutional Review Board (IRB) of the School of Psychology at South China Normal University approved the study.
2.2 Materials and fMRI tasks
The first group of 77 participants engaged in two runs of the phonological and semantic judgement tasks (
Fig. 2
The materials and procedures of the phonological and semantic judgement tasks were taken from our previous study ( Li et al., 2024). They consisted of 72 pairs of Chinese characters: half for phonological judgement (i.e., rhyme judgement) and half for semantic judgement. A block design was used, with each run consisting of 3 blocks of phonological judgement task, 3 blocks of semantic judgement task, and 3 blocks of perceptual judgement ( Fig. 2A). The perceptual judgement task was used to address other research questions unrelated to the current study, and thus excluded from the analysis in this paper. The blocks were organized according to the Latin square design, with 6 trials per block. A 0.5 s fixation was presented at the beginning of each trial, followed by two Chinese characters presented sequentially. The initial character was displayed for 0.5 s, followed by a 0.5 s blank. The second character was displayed for a maximum of 2.5 s and disappeared after a button press. Participants were asked to press one of the two buttons to indicate whether the second character rhymed with the first (phonological judgement task) or was semantically related to it (semantic judgement task). Each run lasted for 376 s.
The second group of 32 individuals completed six runs of the familiarity judgement task (
Fig. 2B). This task contained a large number of stimuli (i.e., 198 Chinese characters), which provided sufficient power to decode the phonological and semantic information underlying the neural representation of the left TPJ. These words included adjectives (e.g., “冷”/cold/), verbs (e.g., “写”/write/), and nouns (e.g., “纸”/paper/). They consisted of 4–19 strokes (mean = 10.09, SD = 2.81), and had medium to high usage frequency (mean = 217.05 per million, SD = 907.72). We used an event-related design for the familiarity judgement task to determine the BOLD signal corresponding to each trial. Each run consisted of 66 trials and lasted for 396 s. A 1 s fixation marked the start of each trial, after which a Chinese character was shown for 0.5 s. Participants judged how often they encountered the corresponding words in their daily lives and rated each word from 1 (“rarely or never”) to 3 (“often”) by pressing one of the three buttons. Each trial was followed by a blank with variable duration between 2.5 and 11.5 s (mean = 4.5 s). To improve the design efficiency, trial sequences were optimized with Optseq2 (
2.3 MRI data acquisition
All the MRI images were acquired utilizing a 3.0 T Siemens Prisma MRI scanner at the MRI Center of South China Normal University. Scanning parameters are as follows. The resting-state and task-state functional MR images were acquired with a single-shot T2*-weighted gradient-echo EPI sequence (58 axial slices, TR/TE/θ = 2000ms/30ms/90°, FOV = 224 × 224 mm, matrix size = 112 × 112, slice thickness = 2 mm, voxel size = 2 × 2 × 3 mm). Resting-state data were acquired with participants' eyes being open. All participants confirmed that they remained awake and alert during the scanning session. The duration of the resting scan was 8 min.
DWI data were acquired using a single-shot T2-weighted gradient-echo EPI sequence with 64 contiguous axial slices along 64 non-collinear directions, in addition to 10 vol without diffusion encoding ( b 0) as anatomical reference for head-motion and eddy-current corrections (TR/TE = 8300ms/84 ms, FOV = 256 × 256 mm, matrix size = 128 × 128, slice thickness = 2 mm, voxel size = 2 × 2 × 2 mm, b = 700 s/mm 2). The duration of the DTI scan was 10 min. Seven participants' DWI data were removed from the analysis due to inconsistent parameters, resulting in 105 participants in subsequent DTI analysis.
In addition, a high-resolution anatomical image was acquired for each participant with a T1-weighted, three-dimensional, gradient-echo pulse-sequence (176 sagittal slices, TR/TE/θ = 2530ms/1.94ms/7°, FOV = 256 × 256 mm, matrix size = 256 × 256, slice thickness = 1 mm).
2.4 Data-driven parcellation of the left temporoparietal junction
We used connectivity-based clustering (
Bray et al., 2013;
Mars et al., 2011) to identify subdivisions within the left TPJ with relatively homogeneous intrinsic functional connectivity patterns. The resting-state images were preprocessed using DPARSF Version 5.3 (
We first created a left TPJ mask consisting of the left anterior supramarginal gyrus (SMG), left posterior SMG, and left angular gyrus (AG) as defined in the Harvard-Oxford Atlas (Maximal Probability Threshold: 25 %) within FSL. We selected the Harvard-Oxford Atlas because it is widely used and readily available within FSL. A correlation matrix was then generated for each participant in which each row corresponded to a voxel in the TPJ mask, each column corresponded to a voxel in the rest of the brain, and each matrix entry contained the Pearson correlation between the resting-state time courses for the two corresponding voxels. The correlation coefficients were transformed into Fisher's Z-scores and then averaged across 112 participants. The averaged correlation matrix was entered into k-means clustering. We employed the elbow criterion to determine the optimal number of clusters for segregation, and three clusters were identified as the optimal choice. However, as mentioned in the Introduction, using only three clusters would result in a loss of finer-grained information. Therefore, we divided the left TPJ into seven clusters because it was observed that 7 of the 17 networks (which were widely used in the rs-fMRI literature,
Yeo et al., 2011) have connections to the left TPJ. Nevertheless, all the subsequent data analyses were also conducted on the TPJ subregions for the 3-cluster solution. This allows us to compare the results between the coarser and the finer-grained clustering solutions and provides a more comprehensive view of the function of the left TPJ. The k-means algorithm started with random assignment and was run 20 times to find the assignment that minimized the total distance to the center. As shown in
Fig. 3
Each cluster served as a seed to identify the intrinsic connectivity network. As shown in the periphery of Fig. 3, TP1 was connected to the dorsal attention and control networks, TP2 to the dorsal attention and ventral attention networks, TP3 to the control network, TP4 to the ventral attention and control networks, TP5 to the temporal parietal and default networks, TP6 to the control and default networks, and TP7 to the default network. The proportions of overlapped regions between each TPJ cluster's network and Yeo's 17 networks are shown in Supplementary Table 1.
We computed the connectivity pattern similarities across the 7 clusters using the following procedure. The correlation values between voxels in each TPJ cluster and voxels in the rest of the brain were averaged across the TPJ cluster to obtain its connectivity pattern. Then, all pairs of the 7 connectivity patterns were subjected to Pearson correlation analysis, resulting in a 7 × 7 matrix of connectivity pattern similarity (bottom left panel of Fig. 3). Heat map revealed that the connectivity patterns of TP1, TP2, TP3, and TP4 were positively correlated, while the connectivity patterns of TP5, TP6, and TP7 were positively correlated.
2.5 Activation analysis
Activation analysis was conducted on the phonological and semantic judgement tasks to identify the networks specific to phonological and semantic processing, respectively. The image preprocessing was performed using FEAT Version 6.00 in FSL (FMRIB's Software Library,
The general linear model was used to model the data at the first level of analysis for every participant and every run. The regressors utilized in the general linear model were created by convolving the event durations and onsets with a double-gamma hemodynamic response function. For each participant and each run, the contrast of phonological judgement minus semantic judgement was computed to obtain the phonological network, and the reverse contrast was calculated to obtain the semantic network. Subsequently, a second-level analysis was conducted for each participant to merge the imaging data from two runs with a fixed-effects model. Group-level analysis was conducted utilizing FLAME (FSL's local analysis of mixed effects) Stage 1 only ( Beckmann et al., 2003; Woolrich et al., 2004). We report only group images that passed a height threshold of Z > 3.1 and a cluster significance level of p < 0.05, with GRF correction.
To compare the neural activations between phonological and semantic tasks in 7 subregions of the left TPJ, we performed a region of interest (ROI) analysis. Specifically, the beta values of each condition from the first-level model were extracted and then averaged across all voxels in each ROI for each participant. The percent signal changes were estimated using the following formula [contrast image / (mean of run)] × ppheight × 100 %, where ppheight represents the peak height of the hemodynamic response versus the baseline level of activity.
2.6 Resting-state functional connectivity
To investigate the role of the left TPJ in phonological and semantic processing in terms of functional connectivity, we computed the resting-state functional connectivity (rsFC) between each of the 7 TPJ subregions and the phonological and semantic networks. To prevent the impact of self-correlation, the phonological and semantic masks were constructed by subtracting the left TPJ mask from the activation maps of phonological judgement > semantic judgement and the reverse contrast. In addition, as language has long been known to be lateralized to the left hemisphere ( Geschwind and Levitsky, 1968; Olulade et al., 2020; Vigneau et al., 2006), we only included areas in the left hemisphere as language networks (see Supplementary Fig. 1A-B). We conducted Pearson correlation on the mean time course of each TPJ subregion and those of phonological and semantic networks, respectively. The correlation coefficients were all converted to Fisher's Z-scores.
Due to the large difference in volume size between the phonological network (3968 voxels) and the semantic network (12,782 voxels), we then created a voxel-matched semantic network using the activation map of semantic judgement > phonological judgement with a threshold of Z > 4.6 (see Supplementary Fig. 1C). The rsFC was recalculated between the TPJ subregions and the voxel-matched phonological and semantic networks.
2.7 Structural connectivity
In addition to rsFC, we also assessed the structural connectivity between each TPJ subregion and the phonological and semantic networks (Supplementary Fig. 1). The preprocessing procedure included motion and eddy-current corrections, and registration of the DWI to the
b
0 image using a 12-parameter affine transformation in the FSL FDT toolbox. Using the FSL Bedpostx tool, we performed Bayesian estimation of diffusion parameters to model crossing fibres and to build distributions of diffusion parameters within each voxel with the “ball and stick” multi-compartment decomposition model (
Behrens et al., 2003). The interpolated streamline algorithm was used to estimate fibre tracts between ROIs (
Behren et al., 2003). We conducted the probabilistic tractography by using the FMRIB's diffusion toolbox (FDT v5.0,
2.8 Representational similarity analysis
Finally, we performed representational similarity analysis (RSA) on the data of the familiarity judgement task to decode the phonological and semantic information underlying the neural representation of the left TPJ. First, the neural representational dissimilarity matrix (RDM) was created for the 198 Chinese characters. The data were preprocessed using the same approach utilized in the activation analysis, with one difference that spatial smoothing was not applied. The onset and the duration (i.e., 0.5 s) of each trial were modeled as one predictor. Following previous studies ( Mumford et al., 2012; Xue et al., 2010), the parameters for the model were calculated using least squares estimation and ridge regression. For each participant, we generated activation maps reflecting the neural activation patterns for each trial. To obtain a stable neural RDM, the activation maps were averaged across all participants. From these activation maps, the neural RDMs of each TPJ subregion were computed as 1 minus the Pearson correlation for all pairs of characters (198 × (198–1)/2 = 19,503 pairs).
We created three model-based predicted RDMs (i.e., phonological, semantic, and visual RDMs). The phonology of each character comprised 59 binary units, with 23 onsets, 34 vowels, and an additional unit representing null onsets in Chinese pinyin system. The phonological RDM was calculated as 1 minus the Pearson correlation for all pairs of characters. As for the semantic model, a recent study on English words has compared different semantic models (i.e., based on corpus, semantic categories, and semantic dimensions), and found that the model based on semantic dimensions consistently outperformed other models in predicting the neural pattern similarity of the semantic network ( Fernandino et al., 2022). Therefore, we selected a semantic model from a large semantic database of Chinese words—the Six Semantic Dimension Database (SSDD, Wang et al., 2023). The six semantic dimensions include vision, motor, socialness, emotion, time, and space. Each dimension encodes the relative importance of an experiential attribute according to Word2Vec ratings. The semantic RDM was computed as the sum of squared differences in the six dimensions between all pairs of characters. We then calculated the Spearman's rank correlation between the phonological RDM and semantic RDM and found that the two RDMs were not correlated (rho = 0.004, p = 0.611). The visual representation of each character was constructed as binary silhouette images (604 × 604 pixels), in which all background pixels had the value 0 and all character pixels had the value 1. Each binary silhouette image was then converted to a pixel vector. The visual RDM was calculated as 1 minus the Pearson correlation for all pairs of characters. The dissimilarity coefficients were all converted to Fisher's Z-scores.
Finally, we calculated Spearman's partial correlation between the neural RDM and one predicted RDM, with the other two predicted RDMs as covariates, and transformed the correlation coefficients into Fisher's Z-scores. Non-parametric permutation test was conducted to determine the significance level ( Eklund et al., 2016). Specifically, the correlation between the neural RDM and predicted RDM was regarded as the true value. We randomly permuted the three predicted RDMs and calculated Spearman's partial correlation between the neural RDM and one predicted RDM, using the other two predicted RDMs as covariates, to derive a null correlation value for each permutation. All correlation coefficients were converted into Fisher's Z-scores. We executed this process 5000 times and compared the observed Z-value with the null distribution derived from 5000 permutations. The significance of the true Z-value was determined using the equation: p = ((the number of permutated Z-values > the true Z-value) + 1)/5001.
In addition to the ROI-based RSA, we also performed searchlight-based RSA to locate areas that represented phonological and semantic information within the left TPJ. The neural RDMs were constructed on cubic regions (5 × 5 × 5 voxels) centered on each voxel of the left TPJ one by one. Spearman's rank correlation was calculated between neural RDMs and predicted RDMs. As the occipital lobe is well-known for its function in visual processing ( Baker et al., 2018), we further defined the bilateral occipital pole as a control area based on the Harvard-Oxford atlas (Maximal Probability Threshold: 25 %) within FSL, and decoded the visual information within the bilateral occipital pole.
2.9 Correlations between information representations and TPJ's functional connectivity with language networks
To investigate whether the phonological and semantic representations in the voxels of the left TPJ were correlated with its functional connectivity to the phonological and semantic networks during word reading, we did a task-state functional connectivity analysis with the phonological and semantic networks as two seed regions. Following previous studies ( Rissman et al., 2004; White et al., 2023), we computed functional connectivity by correlating the trial-to-trial fluctuations between each voxel of the left TPJ and the two seed regions. First, for each region, we subtracted single-trial responses from the group-averaged activation maps. Second, we computed the Pearson correlations between responses in each voxel of the left TPJ and the mean responses in seed regions (i.e., phonological and semantic networks). Finally, Pearson correlations were computed to further explore the associations between the phonological/semantic representations in each voxel of the left TPJ and their functional connectivities with the phonological/semantic networks.
3 Results
3.1 Behavioral performance
For the first group of participants (who performed the phonological judgement and semantic judgement tasks), the data of three participants were excluded due to poor task performance in the phonological judgement task (accuracies were lower than 0.65), resulting in 74 participants in the data analyses. As shown in Fig. 1C, the mean accuracies of the phonological judgement task and semantic judgement task were 0.95 (SD = 0.05) and 0.95 (SD = 0.06), respectively. Paired-sample t-test revealed that there was no significant difference between the accuracies of the two tasks ( t = 0.05, p = 0.957). The reaction time for the phonological judgement task (mean = 898.78 ms, SD = 204.11) was higher than that for the semantic judgement task (mean = 765.59 ms, SD = 126.39) ( t = 8.28, p < 0.001). For the second group of participants (who performed the familiarity judgement task), the mean reaction time was 1082.45 ms (SD = 228.49).
3.2 The phonological task recruited anterior and superior TPJ, whereas the semantic task recruited posterior TPJ
As shown in
Fig. 4
We then performed ROI analysis by extracting the percent signal changes in each TPJ subregion (
3.3 The left anterior and posterior TPJ had stronger functional and structural connections with phonological and semantic networks, respectively
To clarify the functions of different TPJ subregions from the perspective of connectivity, we assessed the rsFC and structural connectivity between each TPJ subregion and the phonological and semantic networks (based on the analyses in the previous section). As shown in
Fig. 5
Structural connectivity analysis showed that ( Fig. 5B) TP1 ( t = 14.916, p < 0.001), TP2 ( t = 9.304, p < 0.001), TP3 ( t = 22.764, p < 0.001), and TP6 ( t = 7.758, p < 0.001) had stronger structural connectivity with the phonological network than with the semantic network after Bonferroni correction, while TP5 ( t = 33.490, p < 0.001) and TP7 ( t = 11.483, p < 0.001) had stronger structural connectivity with the semantic network than with the phonological network after Bonferroni correction.
The results of the voxel-matched ROI analysis were essentially in line with the above findings (see Supplementary Fig. 2). Taken together, the left anterior and posterior TPJ had stronger functional and structural connections with the phonological and semantic networks, respectively.
3.4 The left anterior and posterior TPJ represented phonological and semantic information of Chinese characters, respectively
Finally, to decode phonological and semantic information within the neural patterns of the left TPJ, we performed RSA within each TPJ subregion using the data from the familiarity judgement task. As shown in
Fig. 6
The above results were further confirmed by searchlight-based RSA (see Supplementary Fig. 3). Specifically, the superior and anterior TPJ could encode the phonological information of Chinese characters, whereas the inferior and posterior TPJ could encode the semantic information of Chinese characters. The visual information was encoded in the bilateral occipital pole, but not in the left TPJ. Taken together, these results indicate that the left anterior and posterior TPJ represented phonological and semantic information of Chinese characters, respectively.
3.5 Information representations in the left TPJ were positively correlated with this region's functional connectivity to the language network
We further performed correlational analysis to test whether phonological and semantic representations in the left TPJ were correlated with this region's functional connectivity to the phonological and semantic networks. As shown in
Fig. 7
3.6 Results of the 3-cluster solution
Using the same methods as in the 7-cluster solution, we re-ran activation analysis, rsFC analysis, structural connectivity analysis, and RSA for each TPJ subregion based on the 3-cluster solution. The results are shown in Supplementary Fig. 4. Specifically, the left TPJ was divided into anterior, middle, and posterior subregions. Activation analysis revealed that the anterior ( t = 7.075, p < 0.001) and middle ( t = 2.596, p = 0.034) TPJ exhibited higher activations during the phonological task than during the semantic task, while the posterior TPJ ( t = 5.744, p < 0.001) showed greater activations during the semantic task than during the phonological task after Bonferroni correction. Resting-state functional connectivity and structural connectivity analyses revealed that the anterior and middle TPJ had stronger rsFC (anterior TPJ: t = 29.57, p < 0.001; middle TPJ: t = 4.986, p < 0.001) and structural connectivity (anterior TPJ: t = 13.353, p < 0.001; middle TPJ: t = 20.48, p < 0.001) with the phonological network than with the semantic network, while the posterior TPJ had stronger rsFC ( t = 23.638, p < 0.001) and structural connectivity ( t = 25.625, p < 0.001) with the semantic network than with the phonological network after Bonferroni correction. RSA revealed a relatively weak phonological representation in the anterior TPJ ( r = 0.012, p = 0.041, uncorrected), but a robust semantic representation in the posterior TPJ ( r = 0.073, p < 0.001, after Bonferroni correction). None of the three subregions represented visual information. In summary, the results based on the 3-cluster solution are essentially consistent with those based on the 7-cluster solution, indicating that the left anterior TPJ is more involved in phonological processing, while the posterior TPJ is more engaged in semantic processing.
4 Discussion
The present study utilized multimodal neuroimaging techniques to systematically investigate the phonological and semantic functions of the left TPJ in Chinese character reading. The left TPJ was divided into seven clusters using connectivity-based clustering. We used data from the phonological and semantic judgement tasks to identify the networks related to phonological and semantic processing, and computed the functional and structural connectivities between each subregion of the left TPJ and the two networks. The results of functional and structural connectivity analyses revealed that the left anterior TPJ had stronger connections with the phonological network, while the left posterior TPJ had stronger connections with the semantic networks. Activation analysis revealed that the phonological task recruited anterior and superior TPJ, whereas the semantic task engaged posterior TPJ. Finally, by using a relatively large set of materials (i.e., 198 Chinese characters) and RSA, we further decoded the phonological and semantic information underlying the neural representations in each subregion of the left TPJ. Results showed that the left anterior and posterior TPJ represented phonological and semantic information of Chinese characters, respectively. More importantly, phonological representation in the left TPJ was correlated with its functional connectivity to the phonological network, whereas semantic representation in the left TPJ was correlated with its functional connectivity to the semantic network. Altogether, our results revealed the functional divisions of the left anterior and posterior TPJ for phonological and semantic processing in Chinese character reading.
The findings from our research provide three key insights into the functions of the left TPJ in word reading. First, we identified seven functionally heterogeneous subregions of the left TPJ by using a connectivity-based clustering and examined the functional characteristics of the seven subregions based on their functional networks. Our results are aligned with previous studies ( Achal et al., 2016; Bray et al., 2013; Mars et al., 2011; Yeo et al., 2011), but more importantly, they suggest that these subregions belong to different functional networks and subserve different functions. Besides, functional connectivity pattern similarity analysis showed that the anterior and posterior subregions of the left TPJ exhibited discrepancies in connectivity patterns. These findings indicate that the left TPJ is a heterogeneous region with multiple functional organizations, and its functions are supported by functionally distinct subregions.
Second, by combining the univariate activation analysis and multivariate RSA, our study clearly dissociated the functions of different subregions in the left TPJ during Chinese character reading. Multivariate RSA revealed that the anterior TPJ (i.e., TP1, TP2, and TP3) could encode phonological information, and the posterior TPJ (i.e., TP5, TP6, and TP7) could encode semantic information of Chinese characters. Our results extended previous findings demonstrating that the left anterior and posterior TPJ were primarily involved in phonological and semantic processing, respectively ( Farahibozorg et al., 2022; Humphreys et al., 2021; Oberhuber et al., 2016; Yi et al., 2019).
Although several previous studies have shown that the left anterior TPJ is less engaged in phonological processing during Chinese reading than English reading ( Cao et al., 2009; Liu et al., 2022; Siok et al., 2003; Tan et al., 2003), by using a relatively large set of materials with a high ratio of signal to noise, our study found that the left anterior TPJ could encode phonological features of Chinese characters, indicating that the left anterior TPJ may carry out a cross-lingual function in phonological processing during word reading. This conclusion is also supported by the finding of Feng et al. (2020) that Chinese and French readers exhibited universal neural activations in the left TPJ. Future research on the phonological processing of Chinese characters should pay attention to the left anterior TPJ.
The left posterior TPJ (specifically the AG) has been identified as a critical node in the semantic network. Studies have shown that the left AG exhibits higher activation for words compared to pseudowords ( Numssen et al., 2021; Turker et al., 2023b, 2023c). Other studies have provided causal evidence for AG's role in semantic processing (e.g., Hartwigsen et al., 2015). In our study, the left AG showed greater activation during the semantic task than during the phonological task. Moreover, previous studies also suggest that the AG is not just a localized processor, but rather a hub that coordinates semantic information flow across multiple brain regions. It has been reported that offline repetitive transcranial magnetic stimulation (rTMS) over the AG alone does not significantly slow down semantic decision, but combining it with online rTMS over the anterior inferior frontal gyrus (aIFG) does lengthen the decision time significantly ( Hartwigsen et al., 2016). Subsequent studies have found that a disruption of the AG leads to widespread suppression of activity not only in the AG itself but also in other key semantic regions, including the aIFG and the posterior middle temporal gyrus ( Hartwigsen et al., 2017). Consistent with these results, we found that the posterior TPJ (i.e., TP5, TP6, and TP7) in the left AG not only represented semantic information of Chinese characters, but also had stronger functional connections with the semantic network.
It is worth noting that the function of TP4 (positioned at the center of the left TPJ) was not clearly demonstrated in the present study. TP4 seemed not to be involved in either phonological or semantic processing. As the functional network of TP4 overlapped with the ventral attention and control networks, we speculate that TP4 may have a more general function such as serving as a bridge for the coordination between phonological and semantic processing during word reading. Future studies should further explore the precise function of TP4 in word reading. It should also be noted that TP7 (which belongs to the left AG) could also encode phonological information of Chinese characters. This might be due to the fact that the phonological access of Chinese characters relies on whole-word mapping and recruits ventral regions including the left AG, fusiform gyrus, and middle temporal gyrus ( Li et al., 2022; Mei et al., 2015b; Taylor et al., 2013).
The third contribution of our study is to provide direct evidence for the PUCC model from the perspective of word reading. This model posits that the function of the left TPJ depends on both the generalized neurocomputation of this local area and its long-range connectivity ( Humphreys et al., 2015, 2017; 2020, 2021). Here, we functionally defined the phonological and semantic networks, and assessed the long-range connectivities between the left TPJ subregions and the two functional networks. We found that the left anterior and posterior TPJ had stronger connections with the phonological and semantic networks, respectively. This was consistent with the results of local functional decoding (i.e., activation analysis and RSA), indicating that the phonological and semantic functions of the left TPJ were also supported by its connections with the frontotemporal reading networks. More importantly, we found that in the left TPJ, the enhancement of the phonological/semantic representation was accompanied by strengthened functional connectivity with the phonological/semantic networks during word reading. This means that the strength of information representations in the left TPJ might reflect top-down modulation from the frontotemporal reading networks. Although previous studies have extensively investigated the phonological and semantic functions of the left TPJ, to our knowledge, this is the first study to confirm the phonological and semantic functions of the anterior and posterior TPJ from the perspective of connectivity and local neural representation. Future research should combine multiple modalities and analytical methods to comprehensively investigate the functions of a brain region with intricate functional organizations.
Despite the consistent results summarized above, it is worth mentioning that our study found some distinct patterns of structural and functional connectivity in the left TPJ. For instance, TP6 exhibited stronger rsFC but weaker structural connectivity with the semantic network than with the phonological network. This discrepancy might be attributed to the fact that structural and functional connectivity represent distinct aspects of brain connectivity ( Fotiadis et al., 2024; Huang et al., 2016; Liu et al., 2022). Specifically, structural connectivity refers to the actual physical connections between brain regions, which are typically connected via anatomical nerve fibres such as white matter. In the present study, TP6 is physically closer to the parietal cortex of the phonological network, which may account for its stronger structural connectivity with the phonological network. In contrast, functional connectivity refers to the synchrony of activity between brain regions and reflects the dynamic interactions of brain regions during specific tasks or in the resting state. Therefore, although structural connectivity serves as the anatomical foundation for functional connectivity ( Guo et al., 2022), it may not entirely coincide with functional connectivity ( Baum et al., 2020; Gu et al., 2021; Vázquez-Rodríguez et al., 2019; Zamani Esfahlani et al., 2022). Indeed, a recent study found that multimodal areas such as the parietal and temporal cortex exhibited a low level of coupling in structural-functional connectivity ( Popp et al., 2024). This low level of coupling might be due to the brain's adaptation to new experiences as well as the integration of information from different senses, which can shape the functional connectivity between distant brain regions ( Buckner and Krienen, 2013; Vázquez-Rodríguez et al., 2019). Given the left-hemispheric dominance for language processing, the functional connectivity of the left TPJ with the language network might be dynamically shaped by increasing language experience, leading to patterns that diverge from its underlying structural connectivity. Future studies should employ longitudinal designs and compare the left and right TPJ to further elucidate how language experience and evolution impact the connectivity of the left TPJ.
Finally, several limitations of this study need to be discussed. First, in the present study, we defined the left TPJ using the Harvard-Oxford Atlas. Future research should consider the adoption of alternative atlases to further refine our understanding of the left TPJ. Second, the present study utilized k-means clustering to segment the left TPJ following several previous studies (e.g., Mars et al., 2012; Numssen et al., 2021). However, various other cortical parcellation methods exist in the field (see review Eickhoff et al., 2018), including spectral clustering (e.g., Wang et al., 2017, 2020) and hierarchical clustering (e.g., Blumensath et al., 2013). Future research could explore a wider range of methods to provide a more comprehensive understanding of the functional subdivisions within the left TPJ. Finally, our study focused on group-averaged results and did not examine individual differences for two reasons. In our study, the resting-state fMRI scans were only 8 min long, which may not be sufficient to capture stable individual connectivity patterns. In addition, most participants in our study exhibited ceiling effects in their behavioral performance (mean accuracy = 95 % in both phonological and semantic judgement tasks), which did not allow us to link individual differences in behavioral data to brain connectivity. Future research could employ longer scan durations and more challenging behavioral tasks to explore the relationship between language skills and brain structure more comprehensively.
5 Conclusion
To summarize, the present study utilized multimodal neuroimaging techniques to comprehensively investigate the roles of different subregions of the left TPJ in word reading. Functional and structural connectivity analyses revealed that the left anterior and posterior TPJ had stronger connections with the phonological and semantic networks, respectively. Univariate activation and RSA revealed that the left anterior and posterior TPJ were responsible for the phonological and semantic processing of Chinese characters, respectively. Furthermore, the phonological and semantic representations in the left TPJ were positively correlated with its functional connectivity to the phonological and semantic networks, respectively. Our results deepen our understanding of the functional dissociation of the left anterior and posterior TPJ in phonological and semantic processing, and provide converging evidence for the parietal unified connectivity-biased computation model.
Data and code availability statement
The raw MRI data and codes are available upon reasonable request to the corresponding author, LM, given appropriate ethical, data protection, and data-sharing agreements.
CRediT authorship contribution statement
Aqian Li: Writing – review & editing, Writing – original draft, Visualization, Investigation, Formal analysis, Data curation. Chuansheng Chen: Writing – review & editing. Yuan Feng: Writing – review & editing, Data curation. Rui Hu: Writing – review & editing. Xiaoxue Feng: Writing – review & editing. Jingyu Yang: Writing – review & editing. Xingying Lin: Writing – review & editing. Leilei Mei: Writing – review & editing, Supervision, Methodology, Funding acquisition, Conceptualization.
Declaration of competing interest
The authors declare no competing interests.
Acknowledgments
This study was supported by grants from the
Supplementary materials
Supplementary material associated with this article can be found, in the online version, at
Appendix Supplementary materials
Table 1
| Cluster | Center of gravity | Voxels | Proportions | Color in Fig. 3 | ||
| SMG_ant | SMG_pos | AG | ||||
| TP1 | −50 −32 40 | 336 | 100 % | 0 % | 0 % | Yellow |
| TP2 | −61 −31 30 | 459 | 97 % | 3 % | 0 % | Cyan |
| TP3 | −48 −43 46 | 334 | 27 % | 70 % | 3 % | Violet |
| TP4 | −57 −42 37 | 293 | 26 % | 74 % | 0 % | Orange |
| TP5 | −56 −48 17 | 469 | 0 % | 74 % | 26 % | Red |
| TP6 | −49 −54 42 | 288 | 0 % | 18 % | 82 % | Blue |
| TP7 | −50 −56 24 | 415 | 0 % | 0 % | 100 % | Green |
© 2025 The Author(s)