The functional organization of language in the brain has captivated scientists and the public alike for nearly two centuries. For most of the field’s existence, classical models of language representation have largely focused on simplified notions of neurological language representation, mostly limited to the familiar “Broca and Wernicke” areas. However, the advent of modern neuroimaging and direct mapping techniques has revealed the degree to which the classical models are underspecified and inadequate in explaining the complex neurological substrate for the cognitive phenomenon that is human language (1). It is now widely recognized that human language relies on the orchestrated cooperation of various intricate bilateral, cortical and subcortical networks.

For the neurosurgeon tasked with resection of a tumor in a brain area implicated in language, understanding the behavior of language networks is paramount for the preservation and maximal retention of cognitive ability. The extent to which specific regions of neurological tissue are involved in language networks is tested intraoperatively during awake craniotomies using direct electrocortical stimulation (DES). The widespread use of DES in awake neurosurgeries to preserve patients’ cognitive abilities such as language and even musical performance has not only drastically improved the patient quality of life following surgery but also the scientific understanding of language cortical representation. As the most direct mechanism of assessing neurological activity, DES is widely recognized as the gold-standard in measuring brain function.

With a growing majority of the human population being bilingual or multilingual, understanding the basis for the coexistence of two or more language systems in the brain is of increasing relevance. In the context of surgery, the multilingual brain has unique functional characteristics that must be accounted for in order to achieve maximal retention of linguistic abilities for each of a patient’s languages.

Prevailing theories for human language organization share an emphasis on dynamically modulated networks that integrate bilateral, cortical and subcortical networks. One salient example is Hickok and Poeppel’s (2) model for the functional organization of speech processing in the human brain which proposed a dual-stream model for language consisting of two broad distinct but interacting pathways. This model, analogous to the ventral-dorsal stream model for visual processing (“what” and “how” pathways originating in the occipital lobe, respectively), highlights a corresponding functional basis for auditory speech processing (3). While the original proposal of this model focused largely on the cortical foundations of the respective streams, data from DES studies on cortical and subcortical regions helped attribute the role of specific white matter tracts to this theoretical framework (4).

The dual-stream model has evolved and adapted over the last two decades with contributions from functional imaging, tractography, and DES data (3). The first of these parallel interconnected pathways is a ventral stream for lexical-semantic processing of speech signals that involves the superior and middle temporal lobes bilaterally, connecting to the ventrolateral prefrontal cortex via the extreme capsule, uncinate fasciculus (UF), and inferior fronto-occipital fasciculus (IFOF; 1, 4–6). The other is a dorsal stream that involves the posterior dorsal-most aspects of the temporal lobe and parietal operculum, and the posterior frontal lobe via the superior longitudinal fasciculus (SLF)/arcuate fasciculus (AF) system that includes the white matter subcomponents AF, SLF I, II, III and SLF-tp (6). This pathway is mostly involved in the translation of speech signals to articulatory representations supporting speech production and phonological working memory (1, 7). Additionally, the dorsal stream likely plays a role in sensorimotor integration with the ventral premotor area and inferior frontal gyrus (IFG) pars opercularis through the mapping of phonological information onto articulatory motor representations (3, 6).

Similar to the Hickok-Poeppel model, the Rauschecker-Scott model has built upon this foundation with evidence from non-human primates. These authors argue that the dual-stream model is hierarchically and topographically organized in a way that is continuous with our evolutionary development of auditory processing mechanisms (8). Additionally, they argue that the sylvian parietotemporal region (involved in sensorimotor integration characteristic of dorsal stream) is also responsible for auditory spatial processing. With evidence from anatomic and physiologic studies of primate auditory cortex along with human DTI data, their model proposes a loop originating from the primary auditory cortex with one branch extending postero-dorsally towards premotor areas, and another branch extending antero-ventrally towards the inferior frontal gyrus. Thus, Rauschecker & Scott (8) extend the dual-stream scheme to close the loop between speech perception and production.

Overall, this dual-stream framework recognizes the bilateral nature of language while still maintaining a level of left-hemisphere dominance. While this framework developed with the invaluable contribution of modern neuroimaging data, it is worth noting that this schema for language processing in the human brain has its origins in classical models. Chang et al. (3) astutely point out in their review of modern neurolinguistic theory:

Thus, while the dual-stream model represents a major innovative advancement beyond the Wernicke’s conceptual framework, formalizing a bilateral evidence-based organization of language networks, it also provides a bridge between classical and modern perspectives. Contemporary frameworks developed in the last decade with evidence from DES studies and connectomics, however, have extended this framework towards a “hodotopic” understanding of language networks (9). This view (from the Greek hodos, “path” and topos, “place”) is a paradigmatic shift towards a highly distributed and dynamic understanding of the central nervous system which holds that complex cognitive phenomena such as language emerge from dynamic, plastic, and highly interconnected cortical-subcortical pathways rather than discrete cortical loci (9, 10) In multilinguals, hodotopy accounts for the idiosyncratic and partially overlapping neural representations of different languages, as well as differing reorganization patterns shaped by individual language history and the specific tracts affected by pathology (11). This perspective provides a step forward in explaining the interaction of various languages in a multilingual subject, but it is not wholly sufficient per se. When it comes to understanding the maintenance of multiple languages in the brain, many critical factors must be considered, namely age of acquisition (AoA), proficiency level, and cognitive control mechanisms, all of which impact the inter-subject variability of language networks.

With regards to a model for multilingualism, the literature suggests shows that all languages within a subject are largely underpinned by shared networks, but with critical nuances. For example, on the basis of phonology, multilinguals may have additional processing demands as a product of competing representations extending from articulatory planning to post-articulatory monitoring. On grammar, the sharing of biological substrates across languages exists with variations in activation patterns that depend on aforementioned factors such proficiency, language distance, and AoA (12). Low proficiency level and/or exposure in multilinguals may impact the biology of lexico-semantic processing by requiring greater recruitment of the prefrontal cortex (12).

Multilingual language control relies on executive attention networks in addition to language networks, implying that multilingualism has neurological effects beyond language processing with potentially beneficial effects at the neural and cognitive level (12). For example, lifelong bilingualism has been shown to be associated with greater white matter integrity, enhanced cognitive reserve in later life, and structural differences in regions involved in language control and executive function (13). Furthermore, studies on multilingualism indicate that multiple languages in the brain requires the co-activation of dynamic neural mechanisms for language switching and inhibition. One such cognitive linguistics experiment by Starreveld et al. (14) demonstrated the co-activation effect of the non-target language during word production in English-Dutch bilinguals. This co-activation and constant need for monitoring of language control in multilingual individuals is believed to contribute to the enhancement of neuroplasticity (15).

Despite the general consensus in the literature regarding overlapping biological substrates for distinct languages, this theoretical schema is not always observed in clinical practice, such as in neurosurgical contexts where language mapping is required. A recent compelling systematic review by Połczyńska' & Bookheimer (16) on awake brain mapping studies in bilinguals analyzed 28 studies with 207 patients and found evidence for generally separate cortical areas for different languages in both anterior and posterior sites. In cases in which there was overlap between L1 (first language) and L2 (second language), the relationship was explained by either early L2 AoA and small linguistic distance, a quantification of the phylogenetic relationship and mutual intelligibility between language families. This review suggests that AoA, proficiency, and exposure are associated with increased neuroanatomical overlap. In other words, these data suggest that frequent everyday use of both languages can lead to an increased sharing of neural substrates for different languages. Połczyńska & Bookheimer (16)’s review demonstrates that the heterogeneity of a multilingual’s linguistic profile has demonstrable effects on a patient’s language mapping pattern, which has critical implications for surgical planning that will be further discussed in Section 6 of this review. In sum, these studies highlight the high degree of individual variability in language neuroanatomical overlap, influenced by diverse factors such as cognitive control, linguistic similarity, a patient’s unique language usage behaviors (16–18).

Instances of pathology can be particularly insightful in understanding the neurobiology of language in multilingual patients as they reveal how language networks adapt, reorganize, and fail under unique pathological conditions. This understanding is vital for clinical care and research given that patients with glioma have considered language to be the most important function to preserve, even over motor ability, memory, and problem solving, (19). Important pathological factors influencing language outcomes include the type and location of the lesion, as well as, in the case of brain tumors, the degree of white matter involvement and the rate of tumor growth.

Brain pathology may impact monolinguals differently than bilingual and multilinguals. ReFaey et al. (27) investigated bilingualism as a prognostic factor in a retrospective review of 56 patients (14 bilingual) undergoing left-sided awake craniotomy. Bilingual patients demonstrated higher tolerance to direct electrical stimulation (DES) currents and fewer intraoperative seizures (although not statistically significant). These findings suggest that bilingualism may enhance patients’ ability to engage distributed and redundant neural representations that buffer against surgical or pathological disruption. Furthermore, functional imaging studies in bilinguals show that L1 and L2 have both shared and distinct subcortical connectivity patterns, and tumor-induced damage may result in differential deficits and recovery trajectories for each language (25). A further discussion on the impact of brain pathology in multilinguals, and specifically how tumors may affect cortical and subcortical reorganization in this population, is provided in Section 5.

Awake brain mapping through direct electrocortical stimulation (DES) is unapparelled in many respects for allowing resection of tumors in formerly inaccessible areas (such as the insula), its protection against permanent post-operative language impairment, and its direct insight into language representation (19). DES provides several advantages over noninvasive imaging. For example, it is causal – directly disrupting neural activity with millisecond-level temporal precision – rather than correlational, such as fMRI which relies on hemodynamic responses that may be subject to neurovascular uncoupling in the context of an insult (15). Intraoperative DES studies have shed light on the variability of language representation across subjects. One of the early landmark DES studies investigating essential language sites across 117 patients demonstrated a definitive mosaic of cortical representation and the need for revision of the classical models (28).

More recently, intraoperative DES studies have revealed more specific characteristics of language networks, de-emphasizing strict interpretations of the classical model in favor of bilateral probabilistic maps for anatomic epicenters for language functions, such as phonologic and semantic hubs, subserved by parallel networks, which is more practical for consideration in neurosurgical settings (9, 29, 30). Mugler et al. (31) found that articulatory gestures and phonemes are differentially represented in the precentral and inferior frontal gyri, highlighting the role of the primary and premotor cortices. This provides evidence for the importance of the sensorimotor system in speech production and the embodied nature of cognitive phenomena such as language (32). Hsieh et al. (33) analyzed cortical sites involved in speech arrest and language errors, finding that these regions were more strongly associated with inter-community connectivity (module connectors), which suggests that cortical sites critical for language function serve as key connectors between distinct language subnetworks, facilitating communication and integration across the broader language network.

The accuracy of language mapping has improved over recent decades, especially when combined with preoperative techniques such as resting-state fMRI (rs-fMRI), MRI-based tractography, and navigated transcranial magnetic stimulation (nTMS), which boasts a high sensitivity and good correlation with intraoperative DES (34, 35). In the context of multilingualism, the use of electrocorticography (ECoG) in multilingual epilepsy patients in conjunction with DES was able to identify language-specific sites that DES could not in 75% of patients, providing evidence that ECoG can complement DES in discriminating the cortical representations of separate languages (36).

It is worthy to note the limitations of DES, as it can only be accomplished during neurosurgery and can therefore only be used to investigate neurological function in the setting of pathology, such as a tumor. This makes it challenging to generalize the findings to healthy function and normative models. Prior to noninvasive neuroimaging techniques such as fMRI and EEG, brain function could only be investigated in the context of pathology – e.g. a lesion in a particular region identified postmortem may have been linked to a particular cognitive deficit while the patient was alive. This gave rise to the misleading impression that broad functions such as language, emotions, and memory could be entirely localized. Now, it is understood that cognitive phenomena can rarely be linked to just one site. This mode of reasoning, known as the lesion-deficit tradition, dominated the scientific understanding of the brain until the late 20^th century and is important to keep in mind when interpreting the conclusions drawn from DES studies on patients who obligatorily will have a neuropathology (15).

Many aspects of the neuroscience of multilingualism remain debated, and the reorganization of multiple language networks in the setting of pathology has even more research potential. While it is known that tumor growth can induce language reorganization in monolinguals, there is not as much known for bilinguals or multilinguals. In a study of five bilingual tumor patients who underwent awake craniotomy, Quiñones et al. (25) found that brain tumors lead to reorganization of language networks to the right hemisphere and ipsilesional left hemisphere areas, and that L1 and L2 followed distinct reshaping patterns following surgery. The authors conclude that neuroplasticity impacts the compensatory involvement of executive control regions, “supporting the allocation of cognitive resources as a consequence of increased attentional demands” (25). Thus, bilingual brains likely follow different reshaping patterns after tumor resection.

A systematic review of 7 studies with 25 multilingual patients with left frontal lobe tumors (mostly gliomas) who underwent language mapping indicated heterogeneity in the level of overlap of cortical sites subserved by L1, L2, and L3, finding that L3 tends to be more unpredictable (18). In general, L1 and L2 shared many sites near the pars triangularis and opercularis (18). The review agreed with the findings from Połczyńska & Bookheimer (16) in that languages learned earlier tend to have a higher degree of shared cortical sites than those learned later, and that languages with a later AoA generally exhibit activation in a greater number of sites, especially distal ones (37). These data support the notion that younger AoA (and likely higher proficiency and exposure level) are correlated with more cortical integration of different language networks.

However, these findings are not always the case. For example, one study of 13 multilingual individuals with lesions in the left, dominant hemisphere found the opposite trend where younger AoA was associated with greater, more distinct cortical representation than later acquired languages (38). Moreover, this study found that late-acquired languages largely overlap with early-acquired language sites, with the highest overlap (71%) occurring between early- and late-learned languages (38). Notably, the patients in this cohort were a mix of fast-growing tumors (n=5) and slow-growing tumors (n=7). As discussed in Section 3, tumors may trigger reorganization to the cortex adjacent to the lesion or within the same hemisphere in a network of areas that are not language typical (17). Importantly, the temporal pattern of lesional growth is known to cause differences in language function redistribution, with slow-growing lesions, such as low-grade gliomas, allowing for more effective neuroplastic reorganization that may lead to better outcomes (22). This difference in behavior of slow and fast-growing lesions may explain the results found by Fernández-Coello (38). The heterogeneity in the literature highlights the complex, multifaceted contribution of variables related to tumor pathology as well as patient-specific language profile in characterizing their overall effect on patient outcomes. The final section of this review proposes a strategy to standardize data collection and while acknowledging logistical challenges in patient care to address the heterogeneity in the field.