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

The human capacity for bilingualism, the ability to fluently process more than one language, presents a fascinating challenge to neuroscience. Understanding the neural mechanisms underlying this ability is critical for understanding the brain's plasticity and cognitive architecture. Two prominent theoretical viewpoints attempt to explain how the brain handles multiple languages. The first proposes that language processing is mediated by a single, innate language module that is adapted to each language through parameter adjustments. This perspective is aligned with the ideas of Noam Chomsky's universal grammar, suggesting an underlying, shared neural mechanism for all languages. Evidence supporting this view often cites studies showing overlapping brain activations in areas such as the left inferior frontal gyrus (IFG) for both first (L1) and second (L2) languages. However, a contrasting perspective emerges from numerous studies revealing the existence of separate brain mechanisms associated with each language. These distinctions might arise from various factors, including language type, acquisition order, and age of acquisition. To reconcile these disparate observations, Perfetti and colleagues proposed the assimilation-accommodation hypothesis. This hypothesis posits that the brain utilizes existing neural networks supporting L1 for processing L2 (assimilation) while simultaneously recruiting new neural networks to accommodate unique L2 linguistic features (accommodation). Previous research has shown that the extent of assimilation and accommodation is influenced by factors such as the linguistic distance between languages, age of L2 acquisition, and L2 proficiency. However, the fundamental organizational principles governing these neural changes remain largely unclear. Our prior research using multivariate analysis demonstrated that bilinguals process L1 and L2 through interleaved but functionally independent neural populations within cortical regions. This suggests that distinct neural computations are crucial for language-appropriate use of each language, even when underlying language preprocessing logic might remain largely consistent. Word recognition, for example, involves processing and integrating orthographic, phonological, and semantic representations within brain language networks. This study builds upon this foundation. It investigates the precise organizational rules governing L1 and L2 within the entire language network to reveal how these networks accomplish the same linguistic functions (orthographic, phonological, and semantic) across languages and anatomically identifiable regions.

The literature on bilingual language processing is extensive, with studies using various neuroimaging techniques, including fMRI, EEG, and MEG. Early studies primarily focused on identifying brain regions activated during language tasks, often revealing overlapping activations for L1 and L2, particularly in areas traditionally associated with language processing. However, more recent studies, particularly those employing advanced multivariate analysis techniques, have begun to reveal more nuanced patterns. For instance, studies using Representational Similarity Analysis (RSA) have shown that even when there is overlapping activation, the patterns of activation within those regions can differ significantly for L1 and L2. These studies support the idea that while some aspects of language processing might be shared across languages, there are also distinct neural representations specific to each language. Furthermore, the influence of various factors on the neural organization of bilingual language processing has been explored. Age of acquisition is a consistently highlighted factor; earlier acquisition is often associated with greater neural overlap between L1 and L2, while later acquisition is associated with greater separation. The type of languages involved also matters, with larger linguistic distances between L1 and L2 often leading to more distinct neural representations. The level of proficiency in L2 is also a relevant factor, with higher proficiency frequently being associated with more assimilation of L2 into the L1 neural network. This review of literature underscores the need for more refined analyses that move beyond simple comparisons of brain activation regions and delve into the complexities of distributed neural activity patterns, which are crucial for unraveling the intricate mechanisms of bilingual language processing.

This study employed a carefully designed fMRI experiment involving 51 Chinese-English bilingual participants (25 males; mean age 23.4 years). Participants were native Chinese speakers who had begun learning English as their L2 between 3 and 15 years of age. They were carefully screened to exclude individuals with neurological or psychiatric disorders, and their handedness was assessed using the Edinburgh Handedness Inventory. The fMRI experiment used an event-related design with three conditions: Chinese word reading, English word reading, and Chinese pinyin reading. Each condition consisted of 40 trials, carefully matched for factors like picture size, number of strokes, and word frequency to minimize confounds. Stimuli were presented visually for 1000ms, followed by a 4-6s inter-stimulus interval. Participants identified the language type after semantic access by pressing one of three buttons. Accuracy and reaction time data were recorded. Brain imaging data were acquired using a 3T Siemens Prisma scanner, with both functional (EPI) and anatomical (T1-weighted) images obtained. Extensive preprocessing of fMRI data was performed using SPM12 software, including slice timing correction, spatial realignment, coregistration, normalization to MNI space, and spatial smoothing. Statistical analysis of fMRI data utilized a two-stage general linear model (GLM). A first-level GLM performed a fixed-effects analysis for each participant and condition, with the convolution of stimulus onset time and canonical hemodynamic response function serving as independent variables. The second-level analysis involved one-sample t-tests to identify group-level activation patterns. The core methodological innovation lies in the use of Representational Similarity Analysis (RSA). RSA was employed to compare brain activity patterns with behavioral measures. Three behavioral representational dissimilarity matrices (RDMs) were constructed: logographeme RDMs (based on word units), phonological RDMs (based on phonetic features), and semantic RDMs (based on word2vec embeddings). Neural RDMs were constructed from voxel-wise beta-value maps. Searchlight RSA, with a 6mm radius, was used to map the relationships between behavioral and neural RDMs, identifying voxels sensitive to each language-processing component. Cognitive loads for each linguistic component were calculated by summing the correlation values of significant voxels. Support vector regression (SVR) with leave-one-subject-out cross-validation was employed to classify language types using brain representation maps, exploring the separation of neural responses for different linguistic components. A recursive feature elimination scheme helped rank brain regions by their contribution to language type prediction. Analyses also included the calculation of overlapping brain areas for different components across languages to investigate assimilation. Finally, partial correlation analyses examined the relationship between cognitive loads and reaction times, and validation analyses were conducted by using different participant subsamples and controlling for age of acquisition (AOA).

The study's primary findings center around the co-representation and distinct spatial distribution of three linguistic components (logographeme, phonology, semantics) in the brains of Chinese-English bilinguals. The key observations include: 1. **Co-representation and Spatial Separation:** The three linguistic components were co-represented within most language-related cortical regions but with separate neural populations for each language. This means the same brain regions were involved in processing the same components, but the patterns of activation were distinct for each language. The high accuracy (100% in some cases) of language classification based on these distinct brain activation patterns further supports this observation. 2. **Accommodation and Assimilation:** The distinct spatial distribution patterns, predominantly found in the left hemisphere's middle frontal gyrus, inferior parietal gyrus, angular gyrus, supramarginal gyrus, and fusiform gyrus, highlight accommodation for language-specific features. Conversely, co-representation in regions like the opercular and triangular parts of the inferior frontal gyrus, temporal pole, superior and middle temporal gyrus, precentral gyrus, and supplementary motor areas, suggests the assimilation of cognitive processing across languages. 3. **Functional Localization:** RSA and subsequent analyses functionally localized voxels sensitive to each language processing component. This enabled the study to identify regions uniquely or specifically involved in particular linguistic processes. 4. **Cognitive Load and Reaction Time:** Significant correlations were found between reaction time and cognitive loads for specific components and languages, suggesting a direct link between neural processing effort and behavioral performance. For instance, reaction time for Chinese word reading correlated with logo-grapheme brain loads, and Chinese pinyin reading reaction time correlated with semantic brain loads. 5. **Age of Acquisition Effects:** While both early and late AOA groups demonstrated intersecting neural populations supporting linguistic components, late AOA participants showed more extensive brain region activation for L2 processing, consistent with the idea of greater accommodation for L2 in later learners. 6. **Validation:** Robustness of the findings was verified through validation analyses using random subsamples and excluding non-right-handed participants. The overall results remained consistent across analyses, supporting the reproducibility and reliability of the key findings.

This study significantly advances our understanding of bilingual language processing by demonstrating both the separation and sharing of neural resources for different languages and their components. The findings strongly support and enrich the assimilation-accommodation hypothesis, revealing the complexity of neural adaptations to long-term, multilingual experience. The high accuracy of language classification based on the patterns of neural activity highlights the functional independence of language systems. Moreover, the identification of shared neural resources in core language regions suggests a degree of assimilation of the L2 into the L1 network. The observed differences in brain activation patterns for different linguistic components across languages reflect the accommodation of language-specific features. The study's findings have implications for clinical applications, such as intraoperative localization during brain surgeries. The significant contribution of the left middle frontal areas in logo-grapheme representation could guide neurosurgical procedures. The detailed investigation of linguistic components provides crucial information for models of bilingual language processing. The correlation between cognitive loads and reaction time demonstrates a clear link between neural processing demands and behavioral performance. Future research should build upon these findings by exploring the role of other factors, such as language proficiency and individual differences.

This research provides compelling evidence for the coexistence of both distinct and shared neural representations in bilingual language processing. The findings support and refine the assimilation-accommodation hypothesis, showing how the brain adapts to the demands of processing multiple languages. The precise identification of neural mechanisms underlying bilingualism is crucial for understanding brain plasticity and has practical implications for language acquisition, education, and clinical neurology. Future research could focus on exploring the influence of additional factors, such as language proficiency and individual differences in learning styles, on neural adaptation in bilinguals. Further investigation using more sophisticated analyses could further refine our understanding of the dynamic interplay between assimilation and accommodation in bilingual language processing.

The study's limitations include the relatively homogeneous participant sample (primarily right-handed, Chinese-English bilinguals), which might limit the generalizability of findings to other bilingual populations. While the searchlight RSA provided a fine-grained analysis, it could not completely define the nature of the neural manipulation occurring within the identified regions. Future studies could incorporate additional methodologies to more precisely clarify these mechanisms. Finally, while the study controlled for several potential confounds, other factors such as language proficiency might have subtle influences on the results that could be investigated further.