Multiscale neural features of tonal bilingualism: linking regional differences in brain network degree centrality to neurotransmitter-gene signatures in Bai-Mandarin bilinguals

Abstract

Understanding neural mechanisms of tonal language processing is crucial for revealing language-specific brain adaptations, particularly in tonal bilinguals. While hemispheric differences have been identified, network-level mechanisms and their underlying multiscale biological bases remain unclear. Using resting-state fMRI data from Bai-Mandarin bilinguals (BMB) and Mandarin monolinguals (MM), this study investigated brain network topology through degree centrality (DC) analysis, followed by neurotransmitter mapping and transcriptomic analyses. BMB exhibited significantly lower DC in the left middle frontal gyrus (MFG), left inferior parietal lobule (IPL), and left middle temporal gyrus (MTG), but higher DC in bilateral medial prefrontal cortex (mPFC). These differences predominantly manifested across higher-order cognitive networks, including the frontoparietal network (FPN), dorsal attention network (DAN), and default mode network (DMN). Neurotransmitter mapping explained 37% of the group-level variance in DC, with significant contributions from serotonin transporter (5-HTT), dopamine receptors (D₁, D₂), and γ-aminobutyric acid (GABA) systems. Transcriptomic analysis revealed 1,801 genes associated with DC differences (30.07% variance explained), enriched in protein localization, transport, and cellular morphogenesis, with differential expression evident in microglia, excitatory and inhibitory neurons. These findings reveal that tonal bilingualism shapes brain network architecture through coordinated multiscale mechanisms, providing novel insights into the neurobiological basis of tonal language processing.

Supplementary Information

The online version contains supplementary material available at 10.1038/s41598-026-38523-6.

Introduction

Neural signatures of tonal languages

Languages across cultures exhibit a wide range of phonological systems, shaped by diverse communicative functions and historical trajectories^1,2. A fundamental typological distinction exists between tonal languages, such as Mandarin and Cantonese, which use pitch variations to distinguish lexical meaning, and non-tonal languages, such as English and German, where pitch serves primarily intonational or pragmatic purposes rather than lexical differentiation^3–5. An intriguing question concerns how the brain represents and manages tonal features, particularly in individuals exposed to multiple languages with differing pitch systems.

First, compared to non-tonal language users, tonal language users (e.g., Mandarin Chinese) exhibit higher gray and white matter densities in the right anterior temporal lobe and left insula⁶. They also show specific activations in the bilateral temporo-parietal (particularly the middle temporal gyrus, MTG) and subcortical regions during pitch processing, revealing the neural adaptations of tonal language users for semantic processing³. Such comparisons often involve the combined effects of multiple linguistic dimensions, including differences in writing systems and phonological features⁷.

More specifically, within tonal language systems, studies on tonal bilinguals also reveal neural differences compared to Mandarin monolinguals (MM). For instance, in the domain of tonal language perception, Hakka-Mandarin bilinguals exhibit significantly shorter mismatch negativity (MMN) latencies under Mandarin syllable conditions, indicating a more efficient auditory processing mechanism⁸. Additionally, Cantonese-Mandarin bilinguals demonstrate greater gray matter volume in the posterior cerebellum and enhanced functional connectivity within networks associated with language control and phonological processing^9,10. These findings reveal the impact of diverse tonal language experiences on brain structure and function.

Bai-Mandarin bilingualism as a research model

Previous studies have extensively investigated the neural correlates of balanced bilingualism in non-tonal language contexts. For instance, simultaneous bilinguals of English–French show enhanced inter-hemispheric connectivity and greater whole-brain efficiency compared to monolinguals¹¹. Similarly, Spanish–Catalan bilinguals demonstrate increased subcortical gray-matter volume within the language-control and speech-monitoring network (e.g., the putamen, caudate nucleus, and thalamus) in immersive bilingual environments^12,13. These findings suggest that balanced use of two languages from birth shapes brain structure and function in characteristic ways. In contrast, the neural effects of balanced bilingualism in tonal language contexts remain less explored, despite the unique perceptual and cognitive demands involved in lexical tone processing. With respect to simultaneous bilingualism in tonal languages, the bilingualism of the Bai people is particularly noteworthy. Primarily located in Yunnan Province, southwest China, they attain high proficiency in Mandarin through the national education system while continuing to use Bai in daily life. Crucially, Bai-Mandarin simultaneous bilinguals receive equal exposure to both languages from birth, achieving balanced proficiency and usage frequency. This unique linguistic context effectively controls for variables such as the age of second language acquisition, providing valuable conditions for studying differences between tonal languages.

Bai and Mandarin Chinese both belong to the Sino-Tibetan language family and are tonal languages that share similarities while exhibiting distinct differences. Both languages involve complex pitch perception and tonal decoding during phonological processing. However, Bai’s tonal system is more intricate, comprising six to eight tones, compared to Mandarin’s four lexical tones¹⁴. Furthermore, Bai and Mandarin differ in syllable structure and lexical characteristics. Bai demonstrates greater complexity in syllable structure, featuring unique phonemes and combinations. For instance, the labiodental sound [v] can function as both an initial consonant and a final vowel, serving as an independent phoneme, whereas Mandarin follows more regular phonological patterns¹⁴. Lexically, Bai retains a larger inventory of core words associated with the Tibeto-Burman branch and significantly adapts the phonology of Mandarin loanwords, illustrating remarkable linguistic adaptability^14,15. These shared yet distinct characteristics offer a distinctive experimental window for investigating the neurophysiological mechanisms of Bai-Mandarin bilinguals (BMB). Such research helps elucidate how speakers with different linguistic experiences of tonal languages generate differentiated neural representations based on similar linguistic foundations.

Previous study using resting-state functional magnetic resonance imaging (rs-fMRI) has shown that BMB exhibit significant right-hemisphere dominance compared to MM ¹⁶. Specifically, they demonstrate enhanced functional connectivity between the right inferior frontal gyrus (IFG) and other brain regions, along with a significant increase in gray matter volume in the triangular part of the right IFG. This structural and functional covariance suggests that the right IFG, as an integrative hub for tonal language processing, plays a crucial role in bilingual language control and the processing of acoustic features. These findings challenge the functional hypothesis (domain-specific model), which posits left-hemisphere dominance when pitch variation conveys semantic information^5,17. However, tonal language processing involves multiple cognitive functions, including lexical tone perception and phonological processing, which rely on the involvement of multiple brain regions, such as the MTG and the inferior parietal lobule (IPL)^18,19. Yet, it remains unclear how the interactions among these regions are dynamically recruited to support tonal language processing.

Current knowledge gaps and the present study

Previous research on tonal languages has revealed that they exhibit distinctive brain region activity influenced by linguistic experience and bilingual exposure, highlighting the complex interplay of tone perception at cultural, cognitive, and neural levels^9,20. While these studies have provided valuable insights into the neurophysiological mechanisms of tonal language users, several research gaps remain. Rather than focusing on hemispheric lateralization debates, adopting a holistic perspective to investigate the connectivity of brain networks can provide more systematic insights into the mechanisms of tonal language processing^10,17. More recently, individualized functional connectomics has enabled the identification of the language network even in resting-state and non-related task data²¹. To this end, we first applied degree centrality (DC) analysis to rs-fMRI data, a topological metric in network neuroscience that identifies hub regions critical for information integration in the brain^22,23, to detect key regions that differentiate BMB and MM.

In the present study, we selected DC as the primary network metric and adopted a whole-brain, data-driven analytical framework. DC provides a voxel-wise, threshold-free index of the overall connectedness of each region, offering a robust characterization of global functional integration without relying on predefined seeds or regions of interest (ROIs). Compared with higher-order graph metrics that require thresholding and are sensitive to sparsity selection, DC demonstrates superior test–retest reliability and cross-sample stability, making it suitable for detecting experience-related variations in large-scale network organization^22–24. At the same time, a whole-brain exploratory approach was deemed necessary because existing findings regarding regions involved in bilingual experience or tone processing remain heterogeneous and insufficient to define a single a priori ROI. Given that bilingual experience likely influences distributed large-scale networks rather than isolated regions, restricting analyses to a small ROI set could overlook broader system-level modulation. The adoption of voxel-wise DC thus enables an unbiased assessment of potential network-level alterations, aligning methodologically with the exploratory scope of the present study^25–27.

Furthermore, beyond adaptive adjustments at the brain level, language is intricately linked to genetics²⁸, with dialectal distinctions influencing genetic population structures²⁹. Neurogenetic studies have identified dopamine-related genes (e.g., DRD1, DRD2) and the ASPM gene as being associated with grammar learning and lexical tone perception^30,31. Nonetheless, the precise mechanisms through which neurotransmitter systems and genetic factors contribute to tonal language processing remain poorly understood. By integrating network topological metrics from DC with analyses of neurotransmitter distribution and gene expression patterns, researchers can better understand how structural and functional differences in the brain emerge across multiple biological scales. This cross-scale research approach, combining genetics, neurochemistry, brain imaging, and network analysis, can provide systematic perspectives into between-group differences in tonal language.

Therefore, building on network topological differences identified through DC analysis, we further examined how neurotransmitter systems and gene expression patterns interact to shape these functional differences. Specifically, we investigated (1) the neurochemical basis of DC differences through neurotransmitter receptor density analysis and (2) the gene expression patterns associated with DC differences through transcriptome-wide association analysis, with particular attention to functional enrichment of implicated genes.

Results

Comparison of brain network topological features between BMB and MM

To investigate the differences in brain network topology between BMB and MM groups, DC was calculated at the voxel level for each participant. No significant between-group differences were observed for demographic variables. There were no significant group differences in sex distribution (χ² = 0.64, P = 0.42) or age (two-sample t-test: t = − 0.77, P = 0.44). A two-sample t-test was performed to compare DC values between the two groups, controlling for age and sex (see Figure S1 in the appendix). The results revealed that BMB exhibited significantly lower DC in brain regions including the left middle frontal gyrus (MFG), left IPL, and left MTG, whereas they showed significantly higher DC in the bilateral medial prefrontal cortex (mPFC) (AlphaSim correction, P < 0.005; Fig. 1A). These regions with significant group differences were subsequently mapped onto seven resting-state networks. The affected regions were mainly distributed across higher-order cognitive networks (Fig. 1B), including the frontoparietal network (FPN), dorsal attention network (DAN), and default mode network (DMN).

Fig. 1.

Regional DC differences, network distribution, and functional associations between BMB and MM groups. (A) Statistical maps showing DC differences between BMB and MM groups, thresholded at q < 0.005. Positive values indicate regions where DC is higher in BMB than MM, whereas negative values indicate regions where DC is lower in BMB than in MM. Notably, significant differences were observed in regions including the mPFC, MFG, IPL, and MTG. (B) Distribution of the differential regions across the seven canonical functional brain networks, highlighting that the frontoparietal network (FPN) and default mode network (DMN) contained the largest proportions of altered regions. (C) Functional decoding analysis of the regions showing significant DC differences suggests that alterations in the mPFC and other cortical areas are predominantly associated with social cognition and emotional processing.

To further characterize the functional attributes of these DC-difference regions, a function-based decoding approach was employed. Functional annotations were performed for the significantly different regions, including the mPFC, MFG, IPL, and MTG. The decoding results revealed that these regions were broadly involved in multiple cognitive domains (Fig. 1C). Specifically, the mPFC was closely related to social cognition, autobiographical memory, and emotional processing; the MFG was associated with working memory, cognitive control, and language processing, highlighting its critical role in executive control and high-level cognitive regulation; the IPL was implicated in working memory and numerical cognition; and the MTG was particularly sensitive to language, declarative memory and visual semantics.

Neurotransmitter correlates of group differences in brain function

To further decode the neurochemical basis underlying the observed neuroimaging differences between BMB and MM groups, a multiple linear regression model was employed to assess the contribution of major neurotransmitter systems to group differences in DC (Fig. 2A). The model explained 37% of the group-level variance in DC (P_spin = 0.020, adjusted R² = 0.37; Fig. 2A), indicating that spatial patterns of neurotransmitter distribution significantly contributed to the brain network differences between the two groups.

Fig. 2.

Neurochemical correlates of DC differences between BMB and MM groups. (A) Multiple linear regression was used to predict DC differences between BMB and MM groups based on 19 neurotransmitter systems. The scatter plot shows the relationship between predicted and observed DC differences, indicating good model performance. The intensity of blue reflects the density of points. Darker shades indicate regions with a higher concentration of points, while lighter shades indicate sparser point distribution. The color bar indicates the number of observations per bin. A null model preserving spatial autocorrelation was employed to validate the robustness of the predictions. Brain maps illustrating group-level DC differences across the cortex. Voxel-wise t-values derived from the group comparison were averaged within each region of the 100-region Schaefer parcellation. The color scale represents the resulting regional mean t-values, with vmin and vmax indicating the lower and upper bounds of the displayed t-value range, respectively. (B) Contributions of individual neurotransmitters to the regression model. Darker colors indicate neurotransmitters whose contributions were statistically significant after 10,000 bootstrap iterations.

Further analysis identified several neurotransmitter receptors or transporters as significant contributors to the model (Fig. 2B). Specifically, the serotonin transporter (5-HTT, z = 3.37, P_FDR < 0.001), dopamine D₁ receptor (D₁, z = 3.25, P_FDR < 0.001), dopamine D₂ receptor (D₂, z = 3.02, P_FDR = 0.003), M₁ muscarinic acetylcholine receptor (M₁, z = − 3.05, P_FDR = 0.002), µ-opioid receptor (MOR, z = − 3.36, P_FDR < 0.001), γ-aminobutyric acid (GABA, z = − 3.86, P_FDR < 0.001), and dopamine transporter (DAT, z = − 4.04, P_FDR < 0.001) all showed significant explanatory power for the observed group differences in DC.

Transcriptomic association analysis of functional brain differences between BMB and MM

To explore the molecular basis underlying functional brain differences between BMB and MM groups, we conducted a transcriptomic-level association analysis to identify gene expression patterns related to differences in DC between the two groups. The first PLS component (PLS1) explained 30.07% of the variance in DC differences. The spatial distribution of PLS1 scores across brain regions showed a significant positive correlation with the t-statistic map of DC differences (r = 0.55, P_spin < 0.001; Fig. 3A). This positive association indicates that regions exhibiting larger DC differences between the two groups also show higher PLS1 scores, suggesting that the multivariate gene-expression pattern captured by PLS1 is spatially aligned with the functional topology alterations. In other words, genes with stronger PLS1 weights tend to be more influential in regions where DC differences are more pronounced, highlighting a coordinated gene–phenotype coupling across the cortex.

Fig. 3.

Transcriptomic basis of DC differences between BMB and MM groups: Gene expression patterns, functional enrichment, and cell-type contributions. PLS regression (PLSR) was applied to whole-brain expression data of 15,633 genes and the t-statistics map of DC differences between groups. (A) The first PLS component (PLS1) explained 30.07% of the variance in DC differences, and its spatial distribution was significantly positively correlated with the t-statistics map (r = 0.55, Pspin < 0.001), indicating that regions with larger DC alterations exhibit higher PLS1 scores. The color bar indicates the number of observations per bin. (B) Gene enrichment analysis of genes with significant contributions to PLS1 was performed with FDR correction (q < 0.05), highlighting overrepresented biological functions. (C) The mean expression levels of significant genes across regions were negatively correlated with DC differences (r = − 0.37, Pspin < 0.010), suggesting that absolute expression levels do not directly parallel the spatial pattern of DC alterations. (D) Regional mean expression of significant genes was examined across the seven canonical functional brain networks, showing higher expression in the frontoparietal network (FPN) and lower expression in the limbic network (LIM). (E) Cell-type enrichment analysis revealed specific contributions of astrocytes, excitatory and inhibitory neurons, oligodendrocytes, microglia, and endothelial cells to the observed transcriptomic–DC associations. Astro: Astrocytes; Endo: Endothelial cells; Micro: Microglia; Neuro-Ex: Excitatory neurons; Neuro-In: Inhibitory neurons; OPC: Oligodendrocyte precursor cells; Oligo: Oligodendrocytes. Note. All transcriptomic analyses were conducted in the left hemisphere. * P < 0.05, ** P < 0.01, *** P < 0.001.

Data Availability .

The dataset used in the current study is available from the corresponding author on reasonable request.

After applying FDR correction to the normalized weights of PLS1, we identified 744 positively weighted genes (PLS1 + gene set) and 1,057 negatively weighted genes (PLS1 − gene set) with statistical significance (P_FDR < 0.05). Functional enrichment analysis showed that the top 20 significantly enriched terms were primarily associated with protein localization and transport, including activities such as the targeting of proteins to organelles or membranes, intracellular protein transport, and import into cells (Fig. 3B). In addition, they were involved in various aspects of cellular morphogenesis, particularly in the formation and regulation of plasma membrane-bounded cell projections. These biological processes also encompassed key aspects of neural development and neurite formation, such as brain and head development, neuron projection morphogenesis, and morphogenetic processes involved in neuronal differentiation. Furthermore, the genes were implicated in the organization and modulation of membrane structures, as well as in cellular responses to hormonal stimuli and post-translational modifications, indicating a complex molecular interplay underlying the structural and functional architecture of the brain.

We further visualized the whole-brain average expression pattern of genes associated with DC differences and found a significant negative correlation with the t-map of DC differences (r = − 0.37, P_spin = 0.010; Fig. 3C). This negative association indicates that, although these genes contribute to the multivariate expression pattern captured by PLS1, their absolute mean expression levels tend to be higher in regions showing smaller DC alterations. Conversely, this also suggests that genes with stronger PLS1 weights may possibly be influential in regions exhibiting more pronounced DC differences through relative under-expression in those regions. Regional expression characteristics showed higher expression levels in the FPN and lower levels in the limbic network (LIM), suggesting their potential roles in regulating local functional network topology (Fig. 3D).

Considering that differences in genetic background, lifestyle, and environmental exposure between BMB and MM groups may influence brain microstructure, we incorporated the cortical single-nucleus transcriptomic sequencing data (SNDROP-seq) published by Seidlitz et al. ³² to identify cell-type-specific gene categories differentially expressed across eight transcriptionally defined cell types, including synapse-related genes, excitatory neurons (N-EX), inhibitory neurons (N-IN), astrocytes, microglia, endothelial cells, oligodendrocytes, and oligodendrocyte precursor cells (OPCs). Subsequent cell-type enrichment analysis revealed that gene sets related to microglia, N-EX, N-IN, oligodendrocytes, and synapses were significantly enriched among genes associated with imaging phenotype differences, further supporting their critical roles in the group-level functional brain organization (Fig. 3E).

Discussion

Summary of key findings

Our investigation examined differences in brain network topology between BMB and MM, and their associations with neurochemical and genetic profiles. At the network level, BMB exhibited lower DC in the left MFG, IPL, and MTG, but higher DC in the bilateral mPFC, compared with MM, with differences primarily distributed across higher-order cognitive networks, including FPN, DAN, and DMN. Neurochemical analyses demonstrated that multiple neurotransmitter systems collectively explained 37% of the group-level variance in DC, with significant contributions from serotonin, dopamine, acetylcholine, GABA, and opioid systems. Transcriptomic analysis identified 1,801 genes (744 positively weighted and 1,057 negatively weighted) that were significantly associated with DC differences and enriched for processes related to protein localization, cellular morphogenesis, and neural development. Cell-type enrichment analysis revealed significant involvement of microglia, excitatory and inhibitory neurons, oligodendrocytes, and synapse-related genes. The multilevel examination provides new insights into the neurobiological basis of bilinguals in tonal languages.

Neurobiological mechanisms underlying group differences

Interpretation of network topology differences

In BMB compared to MM, the observed lower DC in multiple key language nodes, including the left MFG, IPL, and MTG, as well as the higher DC in the bilateral mPFC, is thought to reflect adaptive changes associated with tonal bilingual experience. Previous research on BMB emphasized right hemispheric specialization, particularly the integrative role of right IFG in tonal language processing¹⁶. Our DC analysis extends previous findings from the perspective of a more distributed bilateral pattern. This difference is likely due to our focus on whole-brain network topology rather than specific regional connectivity, suggesting that tonal bilingualism involves both localized hemispheric specialization and broader network-level integration¹⁷.

Specifically, the left IPL and MTG may reflect specialized patterns for processing multiple tonal systems. The IPL plays a key role in L2 learning success³³ and is anatomically connected to language areas via the superior longitudinal fasciculus³⁴. The MTG, which is involved in categorical phonemic tone processing and likely stores lexical tone knowledge while serving as a lexical interface between phonetic and semantic representations^35–37, may reflect specialized neural tuning for processing multiple tonal inventories. In addition, the role of the left MFG in addressing phonology has been widely evidenced in Chinese reading³⁸. The reduced DC observed in the frontal cortex of BMB compared to MM may reflect enhanced neural efficiency^39,40, suggesting that bilingual speakers require less reliance on top-down mechanisms compared to monolinguals⁴¹. Thus, the reduced centrality in these regions may indicate the pruning of less efficient connections, resulting in a more simplified and efficient network architecture⁴².

In contrast, we observed higher DC in bilateral mPFC in BMB compared to MM, possibly reflecting neural plasticity shaped by both cultural and linguistic factors. The activation of the mPFC is influenced by cultural or linguistic factors during theory of mind tasks⁴³. For self-referential processing, mPFC typically shows greater activity when judging self-related traits compared to others⁴⁴. Compared to MM, BMB individuals may engage in prolonged self-monitoring to integrate dual linguistic-cultural identities into their self-concept, a process that entails self-referential evaluation of which language is more appropriate in a given context. Accordingly, the enhanced centrality observed in the mPFC suggests that it functions as a critical hub for information integration and distribution⁴², potentially reflecting culturally embedded cognitive patterns that integrate social and linguistic information differently than in MM individuals.

In addition, the observed differences in higher-order cognitive networks (FPN, DAN, and DMN) between the two groups highlight how tonal bilingualism shapes large-scale brain organization. We identified a mixed pattern: BMB showed increased DC in the bilateral mPFC of the DMN but decreased DC in regions associated with the FPN and DAN (such as left MFG and IPL). This pattern points to network-specific adaptations rather than uniform alterations. Bilingual language processing requires frequent reversal of functions dominated by different hemispheres across multiple networks including the DMN, DAN, and FPN⁴⁵, indicating that the altered topology we observed may support the dynamic hemispheric switching demands involved in managing multiple tonal systems.

Specifically, the DAN represents task-positive regions activated during goal-directed tasks⁴⁶ and interacts with the prefrontal cortex at the junction of the dorsal and ventral attention systems⁴⁷. In contrast, the DMN is traditionally characterized as a task-negative network that supports internal mentation and introspective processes^48–50, exhibiting gradual up- and down-regulation depending on cognitive demands⁵¹. Our findings of decreased DC in the DAN alongside increased DC in the DMN align with prior reports that simultaneous bilinguals show stronger negative correlations between the DMN and task-positive attention networks than sequential bilinguals⁵². This may indicate enhanced cognitive control abilities⁵² in the BMB group, particularly given their social-cognitive demands and communicative requirements when shifting between two social group identities.

Furthermore, the FPN is implicated in the initiation and modulation of cognitive control across diverse tasks⁵³. Research has shown that bilinguals exhibit stronger functional connectivity than monolinguals in the cingulo-opercular network but not in the FPN⁵⁴, which is consistent with our observation of lower DC in FPN regions (such as left MFG and IPL). These findings suggest that bilingual network connectivity reflects reorganization rather than simple enhancement, with connectivity patterns modulated by bilingual experience⁵⁵. Particularly for tonal language users who rely on categorical perception of pitch patterns, parallels may also be drawn with absolute pitch musicians. These individuals exhibit reduced global neural connectivity alongside local hyperconnectivity in temporal-parietal regions associated with auditory and language processing⁵⁶, again indicating targeted alterations in network architecture. Therefore, the distinctive organization of brain network topology in BMB may reflect a functional modification consistent with the adaptive control hypothesis⁵⁷, enabling more flexible switching between different tonal systems while concentrating hub connectivity and reducing unnecessary connections.

Neurotransmitter systems and bilingual language processing

The significant contribution of serotonin and dopamine systems to group differences between BMB and MM in brain network topology provides crucial insights into the neurochemical basis of tonal bilingualism. Among these systems, 5-HTT showed the most robust positive association with DC differences, indicating that regions with higher 5-HTT density tend to show more positive DC-difference values (BMB–MM). Notably, Selinger et al. reported that individuals with genetically determined low serotonin transporter expression display enhanced subcortical auditory speech encoding with higher signal-to-noise ratios and stronger pitch strength representation⁵⁸. This apparent paradox may reflect layer- or region-specific roles of serotonergic modulation. Reduced 5-HTT expression in subcortical auditory pathways appears to support precise and robust signal extraction, whereas higher 5-HTT density in cortical hub regions may facilitate the network-level integration necessary for managing dual tonal systems. This pattern suggests a functionally heterogeneous distribution of 5-HTT across the auditory hierarchy, with region-specific specialization for either precise encoding or flexible integration in tonal bilinguals.

The positive associations with both D₁ and D₂ receptors indicate that regions with higher dopamine receptor density exhibit more positive DC-difference values. This underscores dopamine’s critical role in adaptive learning. Genetic variants linked to dopamine function have been shown to predict both cognitive flexibility and second language learning^59–61. D₁ receptors shape prefrontal synaptic transmission by modulating recurrent excitation within local circuits, a key mechanism for working memory^62,63, while D₂ receptor binding and receptor density associate with category fluency and implicit sequence learning^64,65. These dopaminergic contributions may underlie the enhanced DC observed in the mPFC of BMB individuals and may reflect their increased flexibility in accessing lexical items across dual linguistic categories.

The negative associations of M₁, MOR, and GABA with DC differences suggest that brain regions with higher densities of these neurotransmitter systems tend to show more negative DC-difference values. M₁ muscarinic receptors, which are involved in mnemonic, attentional, and cognitive processes⁶⁶, may be related to the reduced DC observed in DAN regions. A previous study has shown that hippocampal M₁ binding predicts limbic-temporal hyperactivation underlying learning and may play a role in functional responses related to learning and memory⁶⁷. Similarly, the MOR, which plays a key role in reward, motivation, and emotional responses⁶⁸, may differentially modulate language learning motivation and reward processing in the context of tonal bilingualism.

The strong negative contribution of GABA aligns with its role as the primary inhibitory neurotransmitter, with GABA levels negatively correlating with coordinated activity within resting motor networks⁶⁹. The differential GABA distribution between BMB and MM groups may reflect distinct inhibitory control mechanisms required for managing multiple tonal systems. The most robust negative association with DC differences was observed for DAT, which dynamically regulates dopamine signaling to modulate movement, motivation, and learning behavior⁷⁰. This suggests that regions with lower DAT density exhibit more positive DC-difference values, potentially reflecting group variations in dopamine-mediated cognitive and behavioral processes such as reward processing and learning strategies.

These neurotransmitter systems likely support tonal language processing through dynamic interactions across multiple timescales and spatial scales. Neurotransmitter receptor densities follow the organizational principles of brain connectomes⁷¹, with language-related areas sharing similar receptor density profiles that differ from non-language regions⁷². The interplay between excitatory and inhibitory systems may fine-tune the neural dynamics required for language switching and processing the complex pitch patterns characteristic of tonal languages.

Transcriptomic signatures of neural differences

The identification of 744 positively weighted and 1,057 negatively weighted genes through PLS analysis reveals the molecular basis of differences in brain function between BMB and MM groups. PLS1 accounted for 30.07% of the variance in DC differences, with its spatial pattern positively correlating with the t-statistic map. Notably, the mean expression levels of PLS1-significant genes were negatively correlated with DC differences across regions. These findings suggest that genes with higher PLS1 weights tend to exert greater influence in regions with larger DC differences between BMB and MM, while exhibiting relatively lower expression levels in those same regions.

Functional enrichment analyses revealed coherent biological processes underlying neural specializations. Enrichment in protein localization, cellular morphogenesis, and plasma membrane projections indicates alterations in neuronal architecture establishment and maintenance. The involvement of brain development, neuron projection morphogenesis, and neuronal differentiation genes points to divergent developmental trajectories between BMB and MM groups. Different patterns of neuronal activity induce distinct transcriptional profiles with varying temporal dynamics⁷³, while variants like CNTNAP2 (rs7794745) shape the neuronal architecture of the language faculty through genetically determined effects⁷⁴.

The negative correlation between average gene expression and DC differences, combined with network-specific expression patterns (higher in FPN, lower in limbic networks), suggests region-specific molecular mechanisms regulating local functional topology. Neurodevelopmental genes continue to be expressed in adult cortex, maintaining regionally specialized circuitry for language networks⁷⁵. Research on specific language-related genes has revealed remarkable specificity. For instance, variations in dopamine-related genes (e.g., DRD2/ANKK1) may indirectly influence language acquisition by modulating the cortico-striatal pathway involved in procedural learning⁷⁶. Moreover, ASPM variants show specific associations with lexical tone perception related to processing pitch patterns within syllable-level timeframes rather than general musical or auditory abilities³¹. Cross-linguistic population-scale studies support a weak negative effect of ASPM-D on tone presence, suggesting that observed linguistic diversity may be partly driven by genetic diversity⁷⁷.

The cell-type enrichment findings provide a neurobiological framework for understanding how genetic differences manifest at the cellular level. The significant enrichment in microglia-related genes is particularly noteworthy given microglia’s role as highly dynamic cells that survey the brain and make essential contributions to the central nervous system development^78–80. The enrichment in genes related to both excitatory and inhibitory neurons suggests alterations in the excitation-inhibition balance crucial for network dynamics. Excitatory neurons show layer-specific gene expression differences between the frontal and temporal language cortex, which are associated with white matter connectivity and language-related conditions⁸¹. The presence of inhibitory neurons is crucial for the emergence and consolidation of modular structures in neural networks, with their numbers directly related to memory capacity⁸². Oligodendrocytes, which facilitate fast nerve conduction through myelination⁸³, show enrichment patterns that may reflect differences in information transmission efficiency, while synapse-related gene enrichment suggests underlying distinctions in synaptic organization and plasticity. These cell-type-specific patterns indicate that brain differences between BMB and MM groups arise from coordinated changes across multiple cellular populations.

Integration of multi-level findings

The convergence of findings across network topology, neurochemistry, and genetics necessitates a comprehensive theoretical framework. We propose the Multilevel Neurolinguistic Adaptation (MNA) framework, which conceptualizes brain differences between BMB and MM as emerging from dynamic interactions across genetic, molecular, cellular, and network levels, all shaped by cultural-linguistic experience and environmental factors. This framework builds upon existing developmental systems perspectives that emphasize how phenotypes emerge rather than being predetermined⁸⁴ and incorporates principles from developmental cultural neuroscience that integrate culture, development, and cognitive neuroscience⁸⁵.

At the genetic level, allelic distributions between BMB and MM create differential neural responsiveness, which, together with language experience, determine bilingual language control⁸⁶. This bottom-up genetic influence operates through multiple interconnected pathways that cascade from molecular to network levels. Genetic variants between BMB and MM groups affect neurotransmitter expression, creating distinct neurochemical landscapes where differences in serotonin, dopamine, and GABA systems act as molecular mediators translating genetic variation into functional consequences. These neurochemical differences influence the development of structural connectivity, as structural networks partially mediate genetic influences on cognition⁸⁷, while simultaneously driving cell-type-specific expression patterns that create local differences in neural computation. Through developmental trajectories, these multilevel genetic influences establish persistent neural organization differences that ultimately converge at the network level, producing the characteristic pattern of reduced centrality in classical language regions but enhanced centrality in social-cognitive hubs like the mPFC.

Environmental factors interact with this biological substrate through bidirectional mechanisms across the lifespan. Cultural evolution constitutes a primary factor shaping linguistic structure⁸⁸, while epigenetic mechanisms enable cultural experiences to mold the brain through socialization and adaptation⁸⁹. Therefore, different forms of bilingualism and cultural contexts engage brain networks in distinct ways⁵⁷. For instance, functional brain connectivity is shaped by both past and current linguistic experiences⁹⁰. In addition, the demands of processing multiple tonal systems enhance pitch acuity⁹¹, and environmental stressors create epigenetic “developmental switches” that program long-term neural responses⁹². Notably, development can strongly influence activity-dependent gene expression⁷³. Consequently, these top-down environmental influences create a dynamic interplay between genetic predispositions and experiential factors, ultimately establishing different neural architectures between BMB and MM groups.

Conclusion

In conclusion, our study provides a multiscale analysis of the neural mechanisms underlying brain topology configuration in tonal bilingualism. The decreased degree centrality in left-lateralized language regions (MFG, IPL, MTG) and increased DC in bilateral mPFC among Bai-Mandarin bilinguals reflect a distinctive network architecture of tonal bilingualism, primarily distributed across higher-order cognitive networks including the FPN, DAN, and DMN. Our neurotransmitter mapping revealed notable contributions from monoaminergic (serotonin, dopamine) and inhibitory (GABA) systems, explaining 37% of the observed group differences and highlighting the neurochemical underpinnings of tonal bilingual brain patterns. The transcriptomic analysis further illuminated molecular mechanisms, identifying 1,801 genes associated with DC differences that are implicated in protein transport, cellular morphogenesis, and neural development pathways, with differential expression across microglia, excitatory and inhibitory neurons.

This integration of network, neurochemical, and molecular perspectives demonstrates that tonal bilingualism shapes brain connectivity through coordinated biological mechanisms spanning multiple scales. Future research should examine how these neurobiological signatures relate to specific linguistic features of tonal languages and cognitive characteristics in bilinguals, potentially informing educational strategies and interventions for language learning. Our findings provide novel insights into the neurobiological basis of tonal language processing and establish a framework for investigating language-specific brain adaptations across multiple scales.

Limitations

Several limitations constrain the interpretation of our findings. First, the multifaceted differences between BMB and MM groups, such as the number of languages acquired and tonal complexity, make it challenging to attribute the observed multilevel neural differences specifically to tonal processing. These confounding factors create interpretive ambiguity regarding the specific contributions of tonal versus non-tonal bilingual experience to neural organization. Second, the absence of comprehensive socioeconomic status (SES) data represents a notable confound, as family SES profoundly influences language development and may interact with bilingual effects^93,94.

Additionally, our methodological approach presents several constraints that limit causal interpretation. The cross-sectional design and reliance on static brain network topology measures cannot capture the dynamic nature of language experience, which is inherently a long-term process rather than a static binary variable⁹⁵. This approach may miss crucial developmental trajectories and individual variations in the neural profiles of tonal bilinguals. Future studies should integrate longitudinal, multimodal approaches combining behavioral experiments, task-based fMRI, EEG/ERP, and intervention studies to establish causal relationships.

Finally, in this study, functional data were normalized using an EPI template rather than the participants’ T1-weighted structural images. This approach was chosen to avoid potential errors arising from cross-modal EPI–T1 registration and to simplify the processing pipeline, thereby ensuring consistency across multiple participants. However, we note that this method may slightly reduce spatial precision in small subcortical regions or peripheral structures.

Methodology

Participants

A total of 58 healthy adult participants were enrolled in this study (aged 20–36 years, see Table 1), including 28 MM individuals (12 males and 16 females; mean age = 26.07 years, SD = 2.03) and 30 BMB individuals (16 males and 14 females; mean age = 25.33 years, SD = 4.57). Age differences were examined using independent samples t tests, whereas sex differences were analyzed using chi square tests. During participant recruitment, brief background interviews were conducted to verify language exposure patterns. Participants in the BMB group had been exposed to both Bai and Mandarin from birth through dual cultural immersion (Bai at home and Mandarin in school and workplace settings), and they reported balanced daily use and equivalent proficiency in both languages. Participants in the MM group had lived exclusively within a Mandarin-speaking cultural environment since birth. To control for potential confounds, we excluded individuals with: (1) significant exposure to languages other than Bai and Mandarin (e.g., advanced English proficiency beyond basic classroom instruction, or fluency in other Chinese dialects); (2) a history of language or speech disorders; and (3) neurological disorders or MRI contraindications. All participants were right-handed college students and were scanned using identical MRI systems and acquisition parameters. Written informed consent was obtained from all participants prior to the study. The study procedures were conducted in accordance with the latest revision of the Declaration of Helsinki and received full approval from the local ethics committee at the Kunming Medical University.

Table 1.

Demographic and linguistic characteristics of participants.

Characteristic	BMB (N = 30, 14 female)	MM (N = 28, 16 female)
Quantitative measures	Mean ± SD	Mean ± SD
Age (years)	25.33 ± 4.57	26.07 ± 2.03
Qualitative measures
Age of acquisition	Both Bai and Mandarin from birth	Mandarin from birth
Daily language use	Balanced (~ 50% each)	Mandarin dominant
Language proficiency	Equivalent in both	Native Mandarin only

Resting-state fMRI data acquisition

Resting-state functional MRI (rs-fMRI) scans were acquired using a 3.0 T Siemens MAGNETOM Allegra syngo scanner. During the scanning session, participants were instructed to relax, keep their eyes closed, and remain awake. The fMRI acquisition parameters were as follows: repetition time (TR) = 2000 ms, echo time (TE) = 22 ms, flip angle = 90°, matrix size = 64 × 64, voxel size = 3.4 × 3.4 × 4.6 mm³, and a total of 240 volumes were collected.

Resting-state fMRI data preprocessing

Preprocessing of the rs-fMRI data was performed using the DPABI toolbox (v7.0)⁹⁶ and included the following steps. To achieve magnetic equilibrium, the first 10 volumes were discarded. The remaining images were realigned to the first volume to correct for head motion. Functional images were realigned to the first volume of each run to preserve fine spatial details, as realignment to the mean image, while more robust to noise, may blur subtle anatomical features. All fMRI images were normalized to the EPI template (MNI152 EPI template with a resolution of 3 × 3 × 3 mm³) and resampled to a voxel size of 3 × 3 × 3 mm³. Spatial smoothing was performed using a Gaussian kernel with a full-width at half-maximum (FWHM) of 6 mm. Linear trends were removed to minimize signal drifts. Several nuisance covariates were regressed out, including the Friston-24 head motion parameters, mean white matter (WM) signal, and cerebrospinal fluid (CSF) signal. A band-pass filter (0.01–0.1 Hz) was applied to reduce low-frequency drift and high-frequency noise. To control for motion-related artifacts, participants with head motion exceeding one voxel in any direction or with a mean framewise displacement (FD) > 0.2 mm were excluded. No participants were excluded based on these criteria.

Graph-theoretical network analysis

We employed the DC metric to assess the topological organization of the brain’s functional connectivity network, aiming to uncover potential neural functional differences between BMB and MM groups. DC reflects the extent of connectivity of a given brain region within the entire functional network; higher DC values indicate a more prominent hub role of the region. DC maps were first computed at the whole-brain voxel level for each individual participant using the DPABI toolbox (v7.0)⁹⁶. For each voxel, the preprocessed BOLD time series was correlated (Pearson’s r) with the time series of every other voxel in the brain to generate a voxel-by-voxel correlation matrix. Correlation coefficients were transformed to Fisher’s z values prior to subsequent processing. DC was operationalized as the sum (count) of suprathreshold positive connections for each voxel, where suprathreshold was defined as either a correlation magnitude cutoff (r > 0.2). Resulting DC maps were normalized (z-scored) across the brain to allow group-level comparisons. Subsequently, group-level comparisons were conducted using two-sample t-tests. Spatial statistical correction was performed using the AlphaSim method, with a significance threshold of q < 0.005.

In addition to DC, several global and higher-order network metrics were computed, including global efficiency and the clustering coefficient, based on the 100-region Schaefer parcellation. However, only DC exhibited significant group differences in our analyses. To maintain clarity and focus on the primary findings, only DC results are reported in the main text, whereas other metrics, which did not show significant effects, are not detailed.

Brain parcellation

The voxel-level statistical difference map of the whole brain was parcellated into 100 cortical regions of interest (ROIs). These 100 ROIs were derived from the local–global functional parcellation (scale-100 version) proposed by Schaefer et al.⁹⁷. To interpret the findings within the framework of canonical resting-state networks, we adopted an established cortical parcellation based on the seven-network scheme initially proposed by Yeo et al.⁹⁸. These networks include: Visual (VIS), Somatomotor (SOM), Salience/Ventral Attention (SAL), Dorsal Attention (DAN), Limbic (LIM), Frontoparietal (FPN), and Default Mode Network (DMN). Specifically, functional images were first normalized to the 2-mm isotropic MNI152 template using nonlinear registration⁹⁹, and DC was subsequently computed directly within this standardized space. To derive region- and network-level indices, two widely used parcellation schemes—the Schaefer-100 functional atlas and the Yeo seven-network template, both resampled to 2-mm resolution—were applied. Within each parcel or intrinsic network, mean DC values were calculated by averaging across all voxels contained in that spatial unit, yielding parcel-level or network-level measures of DC alterations. This voxel-to-region aggregation approach has been widely adopted in large-scale connectome studies^100,101, demonstrating that region-averaged measures can robustly capture the spatial distribution of voxel-wise effects while reducing noise, although fine-grained spatial variations and parcel-boundary effects may be attenuated.

Term-based meta-analysis

To decode the observed functional brain differences between BMB and MM groups, we employed the Neurosynth database (https://neurosynth.org/), an online platform for large-scale meta-analyses of fMRI studies. A set of 24 topic terms was selected, covering a comprehensive range of behavioral and cognitive domains previously examined in the literature^102,103(See Table S1 in the Appendix). A binary mask was created by assigning a value of 1 to regions with significant differences and 0 elsewhere. This mask was used as input to the meta-analysis. z-scores for each topic term were then weighted by this binary mask and subsequently re-ranked and visualized. A significance threshold of z > 3.1 was applied.

Neuroimaging–neurotransmitter association analysis

To further explore the neurobiological underpinnings of functional brain differences between BMB and MM groups, we utilized neurotransmitter receptor density maps derived from a cohort of over 1,200 healthy individuals (18–94 years), as reported by Hansen et al. ⁷¹. Although population-specific receptor atlas for Chinese cohorts are not currently available, the neurotransmitter maps represents one of the most comprehensive and methodologically robust in vivo receptor datasets to date and therefore provides a suitable reference for exploratory cross-modal mapping. This atlas has been widely applied across numerous neuroimaging studies^101,104,105. These maps, originally provided at the voxel level, encompassed nineteen neurotransmitter systems. For receptors and transporters with multiple tracer images, weighted averages were computed based on the number of participants contributing to each image. This yielded 19 neurotransmitter receptor and transporter maps, including: 5-HT_1A, 5-HT_1B, 5-HT_2A, 5-HT₄, 5-HT₆, 5-HTT, α₄β₂, CB₁, D₁, D₂, DAT, GABA_a/BZ, H₃, M₁, mGluR₅, MOR, NET, VAChT, and NMDA. To obtain region-level measures, the voxel-wise maps were subsequently parcellated using the Schaefer-100 atlas. Within each parcel, mean values were computed by averaging across all constituent voxels, providing parcel-level estimates of neurotransmitter receptor density. No additional weighting or smoothing was applied, ensuring that the voxel-level quantitative information was preserved as accurately as possible.

We applied multiple linear regression to quantify the contribution of these 19 neurotransmitter systems to the observed DC differences between BMB and MM groups. Given the relatively low dimensionality of the neurotransmitter data and minimal multicollinearity among receptor types, linear regression provides a straightforward and interpretable framework that allows direct estimation of each receptor system’s contribution to the imaging phenotype. In this model, the regional neurotransmitter receptor density values were treated as the predictors (a matrix comprising 100 cortical ROIs across 19 neurotransmitter receptor density maps), whereas the imaging-derived functional maps served as the response variables (i.e., group-level degree centrality (DC) difference t-map summarized within the 100-region Schaefer parcellation). The latter were represented as vectorized spatial distributions of the imaging phenotypes across the 100 ROIs. A spin-permutation null model was used to assess the statistical significance of the regression model^106,107, and FDR correction (q < 0.05) was applied to control for multiple comparisons.

Specifically, we projected the group-level functional difference map onto the fsLR32k cortical surface space to generate a surface-based parcellation¹⁰⁸. The spatial coordinates of each parcel were defined using the centroid vertex of the closest matching vertex on the average spherical surface. These parcel coordinates were then randomly rotated, and the original parcels were reassigned values based on their nearest rotated neighbors (repeated 10,000 times). For each permutation, a new regression model was fitted using the redistributed surface values.

To evaluate the contribution of each neurotransmitter to the model, z-scores were calculated by dividing the original regression coefficient by the standard deviation of the coefficient distribution derived from the 10,000 bootstrap resamples^104,109.

Neuroimaging–transcriptomics association analysis

Gene expression data at the donor and probe levels, along with corresponding spatial coordinates, were obtained from the Allen Human Brain Atlas (AHBA) via the Allen Institute website (https://humanbrain-map.org), where postmortem tissue samples were collected with informed consent from the donors’ next of kin¹¹⁰. The dataset consists of transcriptomic profiles from six postmortem adult human brains. Microarray data were preprocessed using the abagen toolbox^111,112, and the Schaefer-100 parcellation was applied to generate a 100 × 15,633 matrix of gene expression data. Due to the availability of gene expression data from only two donors for the right hemisphere, analyses were restricted to the left hemisphere, resulting in a final matrix of 50 × 15,633 gene expression values.

We used partial least squares (PLS) regression to examine the relationship between group differences in brain function and gene expression patterns, thereby identifying potential molecular underpinnings. Transcriptomic data are high-dimensional and highly collinear, with thousands of genes exhibiting strong spatial autocorrelation. PLS regression addresses these challenges by extracting latent components that maximize the covariance between gene expression patterns and imaging-derived features, improving robustness, reducing dimensionality, and yielding stable and biologically interpretable gene–imaging associations. To generate the component vectors of the functional map used in the PLS regression, voxel-wise DC difference values were mapped onto the brain parcellation template (Schaefer-100) and averaged within each parcel to obtain a 100 × 1 regional vector for each participant, where N represents the number of parcels, thereby yielding the final functional map. The resulting regional vectors were then standardized (z-scored) to ensure compatibility with the high-dimensional gene expression matrix. These 100 × 1 vectors served as the response variables in the PLS model, with the 100 × 15,633 gene expression matrix used as predictors.

To preserve spatial autocorrelation, we tested the null hypothesis that the covariance between PLS1 and genome-wide expression could arise by chance using spin permutations (10,000 rotations of the response variable). Bootstrapping was used to estimate gene weights in the PLS model, from which z-scores were calculated and used to rank genes based on their contribution to PLS1. To identify enriched Gene Ontology (GO) biological processes and KEGG pathways, genes with absolute Z-scores |Z| > 2.76 (FDR-corrected q < 0.05) were selected for enrichment analysis using Metascape (https://metascape.org/gp/index.html#/main/step1), an automated meta-analysis platform that integrates over 40 independent knowledgebases for gene list annotation and analysis¹¹³. All results were corrected for multiple comparisons using FDR (q < 0.05).

Next, we computed the mean expression level for each of the 100 brain ROIs by averaging the expression values of all selected genes within that region. We then performed a Pearson correlation analysis between this regional mean expression vector and the corresponding DC difference vector, thereby assessing the extent to which regional transcriptional intensity is associated with alterations in graph-theoretical topology. To evaluate statistical significance, a permutation test with 10,000 iterations was performed, preserving the spatial autocorrelation of the data. This regional average expression vector was subsequently projected onto the intrinsic brain 7-network architecture to visualize the distribution of gene expression at the network level.

Cell-type enrichment analyses

Cell-type-specific gene lists were compiled from Seidlitz et al. ³², who integrated significantly differentially expressed genes from five independent single-cell RNA-seq studies. Cell-type-specific gene lists were compiled for major cortical cell classes, including excitatory neurons (Neuro-Ex), inhibitory neurons (Neuro-In), astrocytes (Astro), oligodendrocytes (Olig), oligodendrocyte precursor cells (OPC), microglia (Mirco), endothelial cells (Endo), and synapses. For each cell type, the full set of annotated genes was included, rather than only genes identified as significant in the current PLS analysis. Regional gene expression values were extracted from the 100 × 15,633 Schaefer-100 region-by-gene matrix. Significant genes identified by the PLS model were assigned corresponding weights. Enrichment analyses were performed using the aggregate fold change method¹¹⁴, in which the observed gene weights for each list—derived from a predefined set of cell-type-specific genes—were compared against the mean weights obtained from 1,000 random permutations. To control for multiple comparisons, both Z-scores and permutation p-values were adjusted using FDR correction to ensure statistical validity. The significance threshold was set at q < 0.05.

Statistical analyses

Statistical significance was defined using a two-tailed α = 0.05 unless otherwise specified. All statistical tests were conducted using two-sided p-values. To ensure uniform reporting of statistical results across the manuscript, p-values were standardized according to widely accepted conventions. Specifically, when p > 0.001, values were reported to three decimal places (e.g., p = 0.002); when p < 0.001, values were uniformly presented as p < 0.001. Significance thresholds (e.g., α levels) were reported separately as p < 0.05 to avoid confusion with statistical outcomes. For analyses involving multiple comparisons—including ROI-level, network-level, neurotransmitter-based, and gene-level tests—the false discovery rate (FDR) was controlled using the Benjamini–Hochberg procedure, with a significance threshold set at q < 0.05. Only results surviving FDR correction are reported as statistically significant in the main text.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

Lu Zhang and Haoyu Xu conceptualized the study. Lu Zhang and Yang Yang developed the methodology. Yin Mo was responsible for data collection. Lu Zhang, Haoyu Xu, and Yang Yang drafted the manuscript. Lu Zhang, Haoyu Xu, Yang Yang, Yin Mo, and Li Gao reviewed the manuscript. Lu Zhang acquired the funding.

Funding

This research was funded by “Research on the Key Integration Pathways and Mechanisms of Intelligent Technology for Chinese as a Second Language Discourse Comprehension”, 2024 China National Social Science Fund Youth Project (Project No: 24CYY085).

Data availability

The dataset used in the current study is available from the corresponding author on reasonable request.

Declarations

Competing interests

The authors declare no competing interests.