Intersecting distributed networks support convergent linguistic functioning across different languages in bilinguals

The study addresses a central question in bilingual neurocognition: to what extent first (L1) and second (L2) languages share common neural systems versus relying on distinct, specialized systems. Building on the assimilation-accommodation hypothesis, which proposes that L2 processing both reuses L1 networks (assimilation) and recruits additional language-specific networks (accommodation), the authors note that prior work has not resolved whether these principles operate at finer-grained anatomical and functional levels. They focus on the three core linguistic components involved in reading—orthographic (logo-grapheme), phonological, and semantic processing—and ask whether bilingual brains use interleaved but functionally independent neural populations within regions to implement the same components across languages, and whether the same regions subserve these components across languages. The study hypothesizes (1) the presence of interleaved neural populations for the same components across languages (accommodation) and (2) convergence of the same components within the same regions across languages (assimilation).

Two contrasting views frame bilingual language processing: a Chomskyan parameter-setting view implying a universal language mechanism with shared regions for L1 and L2 (e.g., left inferior frontal gyrus), and evidence for separate mechanisms influenced by language type, acquisition order, and age of acquisition (AoA). The assimilation-accommodation hypothesis reconciles these by positing shared L1 networks for L2 (assimilation) along with recruitment of new networks for L2-specific features (accommodation), modulated by language distance, AoA, and proficiency. Prior multivariate analyses suggest distinct, interleaved patterns for L1 versus L2 within regions, implying separate neural populations. Reading research indicates that orthographic, phonological, and semantic representations are processed and integrated across language networks. However, principles governing how bilingual brains organize these components across the whole network remain unclear, motivating the present fine-grained representational similarity analysis (RSA) approach.

Participants: Fifty-one Chinese-English bilingual adults (25 males; mean age 23.4 years), native Chinese speakers without neurological/psychiatric history, normal/corrected vision, learned English (L2) between ages 3–15 (23 early AoA 3–8 years; 28 late AoA 9–15 years). All were educated in Chinese pinyin and passed College English Test Band 4. Handedness: 41 right-handed, 10 balanced. Ethics approval and informed consent obtained. Task and stimuli: Event-related fMRI with three reading conditions (Chinese words, Chinese pinyin, English words), 40 trials each (120 total). Chinese words were two-character words; Chinese pinyin corresponded to the same two-character words; English words had no semantic equivalents in other conditions. Stimuli (white on black) were presented for 1000 ms with 4–6 s inter-stimulus fixation. Visual angle controls and matching for picture size, pixel percentage, strokes, and word frequency were applied. Participants made a 3-button response identifying the language after semantic access (button-language mapping randomized), emphasizing comprehension rather than speed to avoid responses based solely on visual features. A post-scan re-recognition checklist was administered (not analyzed). Due to a machine fault, behavioral responses were recorded for 44 participants. MRI acquisition: 3T Siemens Prisma. Functional EPI: TE 33 ms, flip angle 52°, matrix 110×96, FOV 220×196 mm, slice thickness 2 mm, 72 axial slices, TR 720 ms. Structural T1: TE 2.56 ms, flip angle 8°, matrix 320×320, FOV 256×256 mm, slice thickness 0.8 mm, 208 sagittal slices, TR 3000 ms. Preprocessing (SPM12): Slice-timing correction, unwrapped realignment for motion/field distortion, coregistration (T1 to mean EPI), segmentation and normalization to MNI space, application of deformation fields to EPIs, and 6-mm FWHM Gaussian smoothing. Univariate activation analysis: First-level GLM per subject with condition regressors (onsets convolved with HRF), 6 motion parameters as regressors, high-pass filtering. Contrasts of each condition vs fixation. Second-level one-sample t-tests with FDR p<0.05, cluster size >10. Representational Similarity Analysis (RSA): Behavioral RDMs (40×40 per condition) constructed for (a) logo-grapheme: 1 minus overlapping ratio of basic orthographic units (logo-graphemes for Chinese; letters/tones for pinyin; letters for English), (b) phonology: 1 minus ratio of shared phonetic units (initials/finals/tones for Chinese and pinyin; vowels/consonants for English, position-agnostic), and (c) semantics: 1 minus cosine similarity of 300-dim word2vec embeddings (skip-gram; window 5, negative samples 5, lr 0.025, subsampling 1e-4) trained on Wikipedia Chinese or English corpora. Neural RDMs: First-level GLM with 120 trial-specific regressors (40 per condition) plus motion parameters. For each voxel, 6-mm searchlight sphere used to compute neural dissimilarities (1−Spearman’s rho) between trial patterns, yielding neural RDMs per condition. Analyses constrained by a language-related mask (AAL-based bilateral regions: middle and inferior frontal gyri, precentral, supplementary motor, inferior parietal, supramarginal, angular, superior and middle temporal gyri and poles, fusiform, superior/middle/inferior occipital; intersected with gray matter probability >0.2). Voxels with low BOLD variance (<1/8 mean) excluded. For each linguistic component, partial Spearman correlations between neural RDM and the corresponding behavioral RDM were computed while controlling the other two behavioral RDMs, per language. Fisher r-to-z transformed correlation maps were obtained. Group-level significant voxels were identified with one-tailed one-sample t-tests (H0: z>0), p<0.05 uncorrected, cluster size >10. Cognitive loads: Defined as the sum of z-values in voxels showing significant positive correlations for a given component and language (validated also by number of significant voxels). Support Vector Regression (SVR): For each subject and language, feature vectors comprised brain representations for three components across 30 ROIs (bilateral versions of the above regions; 90 features per sample). Labels encoded language category (Chinese word=1; English word=2; Chinese pinyin=3). Leave-one-subject-out cross-validation with fold-specific group-level masks avoided information leakage. Recursive feature elimination (RFE) assessed feature importance by rMSE increase when a feature was omitted; an elbow criterion determined the optimal feature set. Statistics: Two-way repeated-measures ANOVA tested differences across languages and components for reaction time, accuracy, and brain loads (overall, early AoA, late AoA), with Bonferroni-corrected post hocs. Spearman correlations assessed RDM interrelations. Linear mixed-effects modeling produced adjusted reaction times; partial correlations examined relations between cognitive loads and adjusted reaction times with Bonferroni correction. Validation included subsampling (n=31–51, 1000 iterations), right-handed-only analyses, and mapping to HCPex template.

Behavioral: Accuracy was high and did not differ across conditions: Chinese words 98.4% (87.5–100%, SD 2.4%), English words 98.1% (80–100%, SD 4.0%), Chinese pinyin 97.16% (80–100%, SD 5.0%); ANOVA F(2,86)=1.20, p=0.31. Reaction times differed significantly: Chinese 968 ms (574–1770, SD 293), English 1182 ms (682–1933, SD 359), Pinyin 1497 ms (718–2744, SD 518), ANOVA F(2,86)=19.3, p<0.01; post hoc ps<0.01 (Bonferroni), increasing from Chinese < English < Pinyin. Univariate activation: Activation maps were highly similar across languages (Chinese vs Pinyin r=0.90; Chinese vs English r=0.93; Pinyin vs English r=0.97; all p<0.001). Commonly activated regions included middle temporal gyrus, fusiform, middle occipital gyrus, middle frontal gyrus, superior temporal gyrus, and inferior frontal gyrus. RSA and accommodation: Across all cortical language-mask regions, each linguistic component (logo-grapheme, phonology, semantics) showed anatomically interdigitated but separate neural populations for different languages, indicating accommodation. Cognitive load patterns differed by language: for logo-grapheme, minimal loads during Chinese word reading; phonology loads were highest for English words; semantic loads were highest for Chinese pinyin. Classification: Using representation maps as features, SVR classified language types with high accuracy. With all components: Chinese vs English 100%, Chinese vs Pinyin 100%, English vs Pinyin 96.1%. Single components also discriminated languages: logo-grapheme (100%, 100%, 82.4%), phonology (100%, 100%, 96.1%), semantics (100%, 100%, 100%). RFE indicated the left middle frontal cortex (logo-grapheme representation) as the top contributor, with additional contributions from supplementary motor, temporal, occipital, precentral, and inferior frontal regions. Assimilation and overlaps: Despite voxel-wise separations, regional-level convergence (assimilation) was evident. Overlapping component-specific representations across language pairs were found, e.g., logo-grapheme: right fusiform, right superior temporal cortex, left middle frontal cortex, right inferior frontal cortex (Chinese–English); right STG, right IFG, right inferior parietal, right precentral (Chinese–Pinyin); left middle temporal, left inferior occipital, right STG (Pinyin–English). Phonology: right and left middle/superior temporal and occipital areas and right dorsal IFG (Chinese–English); right fusiform, left superior occipital, right supramarginal, right precentral (Pinyin–English). Semantics: left precentral (Chinese–English); left ventral IFG, left middle temporal, left angular, right middle temporal (Chinese–Pinyin); left ventral IFG and left supramarginal (Pinyin–English). Triangular left IFG for semantics appeared consistent across languages but did not meet the cluster >10 threshold. Regional component profiles: Many regions were consistently involved in multiple components across languages (assimilation), notably the opercular and triangular parts of the IFG, temporal pole, superior and middle temporal gyri, precentral gyrus, and supplementary motor areas. Some regions showed language-dependent component profiles (accommodation), e.g., left middle frontal gyrus, left angular gyrus, left inferior parietal gyrus, left supramarginal gyrus, and fusiform. Cognitive load patterns by language: Chinese words showed roughly equivalent loads across logo-grapheme, phonology, and semantics; English words showed higher phonological loads; Chinese pinyin showed higher semantic loads. Brain–behavior relations: Significant partial correlations linked higher logo-grapheme loads with longer reaction times for Chinese words (r=0.454, p<0.05 Bonferroni-corrected), and higher semantic loads with longer reaction times for Chinese pinyin (r=0.528, p<0.01 corrected). Marginal associations: English phonological loads with shorter reaction times (r=−0.309, p=0.047 uncorrected), and pinyin logo-grapheme loads with longer reaction times (r=0.303, p=0.051 uncorrected). Total loads were not significantly correlated with reaction time. AoA validation: Both early (3–8 years) and late (9–15 years) AoA subgroups exhibited intersecting, separated neural populations for components across languages; late AoA engaged more extensive L2-related regions than early AoA (permutation p=0.060). The interaction between language type and component (equal loads for Chinese; reversed-V for English with high phonology; V-shape for pinyin with high semantics) held in both subgroups. Robustness: Patterns were stable across subsamples, right-handed-only analyses, alternative load definitions (voxel counts), and when mapped to the HCPex template.

Findings demonstrate that bilingual language processing is supported by intersecting, distributed neural populations that are functionally independent across languages at a voxel-wise level (accommodation), while converging at the regional level to perform the same linguistic components (assimilation). This dual organization reconciles debates about shared versus separate bilingual networks. The ability to classify language type perfectly or near-perfectly from component-specific representations indicates that each writing system evokes distinct population codes even for the same linguistic component, aligning with clinical observations of selective language impairments and recoveries in bilinguals. The left middle frontal cortex’s strong contribution to logo-grapheme-based classification underscores its importance in orthographic processing and may inform clinical mapping. The observation that most language-related regions contribute to multiple components supports recent network-level hypotheses (expansion–fractionation–specialization) in association cortex. Component-specific overlap patterns across languages, and language-dependent component load profiles (e.g., higher phonology for English, higher semantics for pinyin), reflect differences in orthographic transparency, phonological assembly demands, and semantic mapping (e.g., homophony in Chinese). Significant correlations between cognitive loads and reaction time support the behavioral relevance of these neural measures. Age of acquisition modulates the extent of L2 network engagement (greater in late AoA), but the overarching intersecting-population organization persists. Together, the results strengthen and refine the assimilation-accommodation hypothesis by specifying fine-grained anatomical and functional principles across the bilingual reading network.

The study provides convergent evidence that bilingual reading engages anatomically interdigitated but functionally distinct neural populations for different languages, while maintaining convergent regional processing for core linguistic components (orthographic, phonological, semantic). These findings validate and extend the assimilation-accommodation hypothesis at a fine-grained spatial scale, demonstrate language-type-specific neural coding that can decode language categories with high accuracy, and link neural cognitive loads to behavior. Future research should combine precision RSA with experimental manipulations to pinpoint specific processes, incorporate individual-difference factors such as proficiency and usage, and advance neural decoding of language comprehension across modalities and writing systems.

First, searchlight RSA identifies linguistic feature-related processes but does not specify the exact cognitive manipulations; combining RSA with more targeted experimental designs or methods is needed. Second, additional factors such as language proficiency and usage, known to affect bilingual brain responses, were not explicitly modeled and should be considered in future studies.