Language disturbances are a fundamental characteristic of schizophrenia (SZ) (1, 2). These abnormalities have been described as the expression of the formal thought disorder (FTD), whether as part of the positive (speech disorganization) or negative (speech impoverishment) symptoms (3). Researchers have examined the neurobiology of language disturbances in SZ through the framework of FTD (4–10). FTD was conceptualized since Kraepelin’s early work on incoherence descriptions (1) and loosening of associations by Bleuler (4). Since then, it has been associated with various aspects of language and communication. The assessment of FTD involves the use of speech-based scales, such as Andreasen’s (5) Thought, Language, and Communication (TLC) or Liddle’s (6) Thought and Language Index (TLI), which subjectively capture discourse features. It is important to note that these scales include items that cluster into subscales related to impoverished thought/language (such as poverty of speech and weakening of goal) and disorganized thought/language (including looseness, peculiar words, sentences, and logic). Thus, within this framework, language and thought are essentially considered equivalent.

Neuroimaging studies have provided evidence that FTD in SZ is associated with alterations in the brain’s language network (7–10). The relationship between FTD and the brain language network suggests that the impaired organization of thoughts and speech observed in individuals with SZ may stem from underlying neural dysfunctions in language-related brain regions. Specifically, abnormalities have been observed in areas such as the prefrontal cortex (PFC), inferior frontal gyrus (IFG), temporal lobe, and superior temporal gyrus (STG), which are known to play critical roles in language comprehension and production (7, 8). Disruptions in the connectivity and activation patterns within these regions have been associated with the presence and severity of FTD. These dysfunctions can impede the integration and coordination of different linguistic processes, leading to the characteristic disorganized and incoherent speech patterns observed in FTD. In fact, according to Wensing´s (7) review, three clusters show convergent aberrant activation linked to FTD: the left STG was associated with speech and auditory processing, and two clusters in the left posterior middle temporal gyrus (MTG) were involved in semantic processing.

However, it is crucial to acknowledge that different levels of language (i.e. morphology, syntax, lexical) entail distinct cognitive challenges and may involve different brain regions. In the same vein, it is also essential to consider that these language levels do not function autonomously but rather interact with each other and that linguistic meaning is generated as a result of this interconnection. Therefore, it is important not to confuse micro processes, such as the processing of lexical units (lexical meaning), and macro processes, such as the interpretation of a text based on the context in which it occurs (semantic meaning).

We sacrifice specificity when investigating brain activation patterns using macro linguistic markers or complex symptoms such as FTD. For instance, when we analyze the weakening of goal, we are actually analyzing the product of numerous linguistic and cognitive micro-processes working in tandem. Examining language disorders and their connection to the brain in detail, focusing on the signs of linguistic micro processes, can provide us with more precise insights. For instance, in the domain of semantics, it is important to differentiate between lexical access and sentence comprehension despite their shared involvement in meaning. By delving into these granular aspects, we can gain a deeper understanding of the complexities of language atypicalities and their neural mechanisms.

In the context of exploring language networks in individuals with SZ, it is especially important to focus on signs rather than symptoms. While symptoms provide valuable insights into the patient’s personal experience, signs provide specific measures of the language disorder. Previous reviews have focused on subjective observation-based instruments to measure symptoms of thought/language disturbances in SZ. While these rating scales have proven clinically useful, they may not fully capture subtle deviations across all dimensions of language. Alternatively, tools from cognitive neuroscience, neuropsychology, and computational sciences offer a better opportunity to explore the multidimensional nature of language. Experimental paradigms such as lexical decision, semantic integration, and semantic association provide insights into cognitive subprocesses involved in language processing. Additionally, computational linguistic tools enable quantitative analyses such as part-of-speech tagging, syntactic parsing, semantic similarity, and speech graphs, among other approaches, to provide objective patterns and specific measures of language impairments. By leveraging language models, linguistic features extracted from oral or written text can be employed to explore the relationship between language impairments and the underlying neural networks in SZ.

Understanding the relationship between language and language networks in patients with SZ is essential for improving diagnostic tools and developing targeted interventions. By combining cognitive neuroscience, linguistic and computational linguistic techniques with neuroimaging data, we can enhance our understanding of the language network functioning in persons with SZ, improve diagnostic approaches, and advance the development of targeted interventions of neuromodulation for language impairments in this population. In this context, the aim of this review is to analyze the brain language networks and related areas in individuals with Psychosis spectrum in relation to linguistic signs.

As mentioned above, we analyzed the results of 39 studies based on the type of patients, type of language task and brain measure. In the following sections, we discuss the findings of this review based on the production and comprehension of language.

In this review, we have compiled studies focusing on examining language networks in individuals with SZ using a sign-based approach rather than focusing on symptoms of thought disorders. We observed recurring and consistent patterns in some studies, while differences emerged in others. Overall, the findings related to the production of language suggested different brain hemodynamic responses in the prefrontal and temporal cortex and reduced functional connectivity between the medial prefrontal cortex and subcortical regions during verbal fluency tasks; different patterns of activation about the amount of speech and syntactic complexity and; related to semantic tasks there was a covariation between linguistic features and the brain activity/connectivity of the temporal and frontal cortex. Regarding comprehension of language, the findings showed less left lateralization in the group of patients, different activation in the IFG, AG, FG, precuneus and STG (some studies observed reduction while others enhanced) in lexical tasks; in the semantic tasks, the studies observed different patterns of activation in IFG, medial PFC, ACC, precuneus, STG and MTG; and in the memory encoding tasks the studies showed reduced gray matter in the group of patients and decrease activity in the DMN. We will now delve into the methodological distinctions that may contribute to these studies’ discrepancies.

Concerning Verbal fluency, several studies have examined the relationship between performance and brain measures. Regarding structural brain measures, in UHR individuals, two studies (39, 42) observed that the verbal fluency deficit was related to reduced grey matter density and lower left lateralization of pars triangularis. The reduced grey matter density was also observed in FES patients but in basal ganglia and thalamus regions (44). Notably, Habets (44) also found reduced grey matter in frontal, temporal and insular areas, but they were not related to the verbal fluency task. In Psychosis patients, two studies found tract differences in the left superior longitudinal fasciculus, inferior longitudinal fasciculus and inferior frontal-occipital fasciculus (28, 35). These results suggest that the grey-and-white matter of the language network predicts, to some extent, better verbal fluency performance. Concerning the functional measures, all the studies (29, 31, 32, 38, 52) except for one (55) observed reduced brain activity in relation to verbal fluency tasks. This difference in results is attributable to the fact that Ragland (55) included switching conditions to increase the cognitive demand of the task. The findings indicate that in over-learned conditions, the neural processes involved in lexical retrieval and word generation remain relatively unaffected in individuals with schizophrenia and that poorly modulated neural response during standard fluency tasks is secondary to well-documented deficits in cognitive control. The areas involved were IFG, STG, bilateral PFC, brainstem, insula, and the left hippocampus. This is intuitively plausible, considering that deficiencies in strategic memory typically correlate with frontal lobe dysfunction or subcortical structures associated with circuits involving the frontal lobe. Furthermore, even though the verbal fluency task is straightforward to apply, the analysis of the studies we reviewed lacks adequacy in explaining the intricate retrieval system. Indeed, incorporating new approaches, such as computational linguistics, is essential to gain a detailed understanding of this type of data (67).

In the domain of semantics, specifically concerning the productive dimension, three articles employed NLP techniques to analyze language samples. Two articles correlated these linguistic metrics with resting-state brain activity, while the third article integrated an additional task during the scanning process. It is not surprising that the results among these studies were inconsistent, given the diverse nature of the experiments. Tagamets (57) utilized NLP metrics to establish correlations with brain activity; however, participants were engaged in a task during the scanning process where they had to monitor whether the stimulus matched the previous one in terms of spelling and sound. Their findings revealed a cluster of activity in the bilateral SFG, right supramarginal gyrus, right FG, right STP, and right culmen/vermis. In the other two studies, participants were in a resting state during the scanning process. They used different types of speech samples (autobiographic life changes versus picture descriptions), yet the results were consistent. Alonso’s study (22) suggests that lexical impoverishment observed in individuals with SZ may be attributed to increased self-inhibition in the IFG and vATL, possibly stemming from reduced synaptic gain and diminished precision of locally generated neural activity (22). Similarly, In Haas’s (25) study, semantic coherence measured through LSA was related to functional connectivity between the temporal and prefrontal semantic hubs. We hypothesize that the differences in the cognitive demand during scanner, such as homophones and homographs matching (25) versus resting state (22), may play a crucial role in accounting for these disparities. While none of the tasks is ideally suited for addressing the relation between lexical selection and brain connectivity, the evidence indicates a striking similarity in the connectivity of the semantic network between the active semantic task and the resting-state data (68). In other words, although it may be less goal-directed and stimulus-driven, the process of conceptualization is almost unavoidable during the resting state.

Regarding syntax, we only found three studies, and two were from the same author. Kircher’s (50, 51) studies showed that the HC group showed correlations between the number of complex sentences produced and activation in the posterior portion of the right MTG and the left SFG. In contrast, these correlations were absent in the patient group. Liang observed that the mean length of T-units (complexity) in the speech and the lexical cohesion were related to the cortical thickness (59). Since the nature of the brain measure is quite different, is not possible to compare the results. Additionally, there are variations in the method of measuring syntactic complexity. Kirchner differentiated between simple and complex sentences based on clauses, while Liang utilized T units of the sentences.

Regarding comprehensive lexical tasks, most studies focused on conventional language regions, notably the IFG and the STG. The only two studies addressed lateralization of language-related brain activation and observed consistently less lateralization in the group of FES (40, 43). Regarding the comparison of brain activity, Han (48) observed enhanced IFG activity with a lexical decision task. At the same time, Chen (47) found less activity in the same brain region in a lexical association with Chinese characters task. In this same line, Kuperberg’s (54) study localized the N400 priming effect spatially in the left temporal fusiform cortex. These results may suggest an elevated and prolonged automatic spreading activation during the semantic processing in the patient group. Some studies found increased connectivity or activity, while others found the opposite. The variances observed could potentially originate from the specific task employed in each study or the clinical characteristics of the sample. For example, Woodward (33) found decreased connectivity, while Lavigne (36) found higher connectivity. However, the first compared SZ with HC, whereas the second compared SZ patients with and without hallucinations. Overall, using a priming task, several studies observed heightened hemodynamic responses among the IFG, FG and STG.

In the semantic meaning paradigms, the results were not consistent in the studies. Using passive listening, Royer’s study (34) identified differences in the leftward functional laterality indices between patients and controls. Kubicki’s (53) study revealed decreased activation in the left IFG and increased activation in the left STG during a semantic judgment between abstract and concrete words. While Horn’s (49) study observed reduced left IFG activation in the patient’s group during passive word reading. The difference in the STG activation between these results could be explained by the nature of the task and the brain recruitment. Indeed, the STG is one of the key candidates for abstract concept processing (69), while semantic processing during passive word reading is more related to MTG (70). The consistent pattern of reduced activity in the IFG is worth noting, with evidence pointing towards the IFG’s role in controlled semantic tasks.

The remaining studies categorized as semantic comprehension paradigms involved higher cognitive demands, resulting in more extensive brain engagement. In Sabb’s (46) experiment, participants were tasked with monitoring whether the auditory stimuli belonged to the same topic and whether they were congruent, thus encompassing a broad range of semantic processing. They observed an increased activation in bilateral medial PFC, left IFG, left MTG, and ACC in the CHR group. Moreover, the CHR psychosis converters exhibited heightened neural activity in left IFG, caudate and STG regions that may serve as a potential marker for identifying those who are at a higher risk of developing psychosis.

Adamczyk (24) utilized a cognitively demanding task focused on metaphor comprehension and found decreased activation in the left caudate, interhemispheric dorsal ACC, left IFG, mFG, and SFG, left insula, Heschl gyrus, STG, precuneus, and inferior parietal lobule including the angular and supramarginal gyri. Zhang (30) used a self-reflection task and observed reduced functional connectivity between the ventral-anterior insula and the precuneus/posterior CC when engaging in close other-reflection tasks. Here emerges a pattern of heightened engagement of cognitive domains associated with complex tasks involving the brain’s activation of the precuneus and insula.

Dickey (41) explored prosodic integration in schizotypal personality disorder individuals and observed differences in the right/left STG coupling compared to HC. Wroblewski (27) studied gestural processing and observed significantly weaker coupling from the MTG to the STS in SSD individuals. Finally, McIntosh (45) employed a task with a more executive profile, specifically the Hayling task, in which they observed a different activity pattern in the dorsal PFC and insula (45).

Some differences in the results may be due to the broad sense in which the concept of “semantics” is used to describe properties of the discourse of SZ patients. In fact, some studies focus on the overall meaning of discourse (22, 24, 25, 27, 30, 34, 45, 46, 49, 57, 66), while others focus on lexical meanings (33, 36, 40, 43, 47, 48, 53, 56, 58). This difference is relevant because language phenomena at both levels of analysis are different. In Linguistics terms, discourse’s overall meaning, or semantics, emerges from the interplay among various language levels, including phonological, morphological, lexical, syntactic, and interclausal components. These levels interact with each other and elements specific to the cultural and situational context in which the discourse takes place (71). The text, as a unit of meaning, conveys semantically coherent and lexically and syntactically cohesive ideas aimed towards a specific goal, influenced by contextual factors such as the speakers’ task (conversation, discussion, presentation) and the discourse modality [narration, explanation, argumentation, instruction (72)]. Within this framework, studies explore preservation, semantic distance between text elements (22, 25), discourse generation based on text comprehension (24, 46), semantic judgment (30), and semantic coherence (57). On the other hand, lexical meanings, or lexical semantics, pertain to the meanings of individual lexical units within a language and the relationships between them within specific lexical-semantic fields. Psycholinguistic terms also involves speakers’ ability to construct and retrieve lexical units from their mental lexicon. Within this context, studies investigate lexical decision tasks, associations between lexical units (33), meaning generation (36), judgment of lexical units (40, 47, 53, 56, 58), and generation of new lexical units (43).

Similarly, in neurolinguistic terms, activation patterns vary based on cognitive demands. For instance, tasks associated with lexical selection heavily engage the inferior frontal gyrus (IFG), while those requiring greater executive functions recruit more prefrontal areas. Similarly, more intricate processes like metaphor processing involve multiple brain regions. Notably, prosodic processing exhibits a more substantial reliance on the right hemisphere. The significance of the superior temporal gyrus (STG) in semantic processing is also prominently observed. In summary, the results indicate that the language network, particularly the semantic network (sometimes more narrow and at others more broad), manifests distinct activation patterns in patients.

Finally, when examining memory encoding paradigms, it was found that individuals with SSD had lower gray matter volume in the CC, juxtapositional lobule, right STG, and precuneus in one study (46) and lower cerebellar gray matter density in another (44) in relation to a verbal learning test.

The variations in the results can be attributed to the group comparisons. In the first study, SZ individuals were compared with a control group, whereas in the second study, SZ individuals were compared with non-psychotic first-degree relatives. Moreover, this semantic processing was functionally related to decreased linguistic processing network activity and reduced DMN suppression (23). This is a plausible assertion, given that both the DMN and the language network are key candidates for neural markers of schizophrenia.

There are some limitations in the results of this work. In our review of the language network, we made an effort to explore its widely defined components based on existing literature. Several studies pre-defined specific brain areas, resulting in variations in the comparison of brain regions across patients and HC. For instance, certain areas of the language network, such as the IFG, consistently exhibit differences, however, it is one of the few areas that is considered in all analyses. Therefore, not all brain areas were consistently examined in all studies included in our analysis.

Furthermore, another constraint lies in the fact that only a limited number of studies have employed machine learning models for analyzing language features. Computational language models have comparative advantages over alternative other approaches to describe linguistic data. A specific illustration of their utility can be observed in the language model employed to extract all the metadata from this very review. Nevertheless, the studies of this review predominantly explore language variables using experimental paradigms designed to isolate one of the variables by controlling the others. However, extracting variables from language phenomena in this manner is highly intricate and results in a significant loss of information. Language is inherently contextual, and a substantial portion of the meaning is intricated in the different layers, levels of language with multiple micro processes in this case.

In that scenario, a large and structured set of texts -oral or written corpus- produced in ecological conditions allows researchers to test existing neuro and psycholinguistics theories and models, assessing their validity with data based on contextual language, and eventually, develop new theoretical frameworks to explain them. More specifically, the metrics of the language structure not only provide information about its use, also allow relating these occurrences with neurobiological evidence. Based on the limitations identified in this literature review, we propose certain guidelines for future research in this field. Examining larger networks involved in language processing -or networks that work together (eg. DMN and Salience Network)- can provide a more comprehensive understanding of the neural mechanisms underlying linguistic/semantic processing in individuals with SSD.

Moreover, extensive evidence exists pointing towards discernible patterns in the speech and communication of individuals with schizophrenia (SZ). Similarly, brain patterns have been identified in association with SZ. However, to date, no specific biomarkers have been identified for SZ, either in linguistics or neuroimaging. It is worth noting that language models have demonstrated the capability to predict brain measurements (73). Thus, there is a pressing need to integrate and combine these models with brain metrics, not solely for classification purposes, but also to gain a qualitative understanding of the language processes involved in SZ. This understanding can then serve as a foundation for identifying potential targets for neuromodulation interventions. In summary, future research should integrate language models and neuroimaging while considering broader brain networks in analyses. By following these guidelines, we can advance our knowledge of language processing in individuals with severe mental illness and enhance the validity and generalizability of findings in this field.

Beyond the limitations of the review’s results, there are also inherent limitations within the review itself. Firstly, its generalizability is constrained as the findings primarily pertain to SZ, potentially limiting applicability to other psychiatric or neurological conditions. We acknowledge the importance of transdiagnosis research, however most of the results were comparing SZ patients with controls. Additionally, an over-reliance on neuroimaging might overshadow other crucial aspects of language processing in SZ. The inclusion of additional techniques, such as EEG, was impractical due to scope limitations and this would suffice to constitute a new comprehensive review. Furthermore, there is a possibility of bias in interpretation stemming from the backgrounds and prior studies of the review’s authors, potentially influencing study selection and interpretation. Lastly, given the rapidly evolving nature of the fields of neuroimaging and computational linguistics, certain sections of the review may require updating in the coming years.

The original contributions presented in the study are included in the article/ Supplementary Material . Further inquiries can be directed to the corresponding author.

MA-S, PA-F, and MO contributed to conception and design of the study. JP performed the search. MA-S, LZ-R, and PA-F analyzed the results. MA-S, PA-F, MO, LZ-R, and AC wrote the manuscript. All authors contributed to manuscript revision, read, and approved the submitted version.