A Unified Attention-Based Account of Semantic Impairment in Schizophrenia

Abstract

Background and Hypothesis

People with schizophrenia (SCZ) show characteristic impairments in semantic cognition, yet the cognitive mechanisms underlying these patterns remain unclear. Here, we propose a unified mechanistic account in which semantic retrieval is constrained by limitations in internal attention. Specifically, we approximate semantic space as a network of interconnected concepts and posit two attentional control parameters: representational precision (resolution; the ability to distinguish nearby concepts) and the size of the attended field (the subset of semantic space prioritized). We hypothesize that attentional constraints in SCZ reduce precision and/or narrow the attended field, yielding complementary, theoretically predicted patterns of semantic search.

Study Design

We first constructed a directed semantic graph using a large sample of individuals performing the animal Category Fluency Test (CFT). We then used this network as a normative reference to analyze fluency-derived paths from SCZ and control participants.

Study Results

Relative to controls, CFT paths in the SCZ group showed increased reliance on abstract, category-level structure, and reduced exploration radius. These patterns align with compensatory strategies predicted by the proposed mechanism: reliance on coarse-grained structure when resolution is limited, and restriction of traversal when the effective attended field is narrow.

Conclusions

The findings support an attention-constrained retrieval account rather than a stochastic retrieval deficit, indicating that group-level semantic retrieval differences in SCZ may arise from reduced representational resolution and/or a narrowed attended field. These results highlight the value of considering internal attention mechanisms when interpreting semantic impairments in SCZ, with potential connections to attentional phenotypes of ADHD/ADD.

Introduction

Schizophrenia (SCZ) involves diverse disturbances in semantic cognition,^1–6 often described clinically in terms of disorganized or atypical thought. For instance, clinical descriptions of SCZ highlight loose associations, concretized interpretations, and overly abstract connections.^7,8 Yet, the cognitive mechanisms that give rise to such patterns remain insufficiently understood.⁹ Furthermore, disturbances in semantic cognition have been reported even in patients not presenting with disorganization or pronounced atypicality in speech,^1,4,10 highlighting the need for mechanistic accounts explaining how people with SCZ access, organize, and traverse semantic space.^11,12 Here, we advance a unified account, wherein an impairment in selective attention^13–16 operating over a semantic network can give rise to such diverse manifestations.

Category fluency tasks (CFT) provide a window into semantic search in general^17,18 and into semantic impairment is SCZ in particular.^4,19,20 Converging evidence from studies of SCZ using CFT and other paradigms points to patterns such as transitions that appear locally loose (eg, Labrador → elephant²¹), over-generalization (eg, viewing a camel as a kind of vehicle²²), and concrete associations.^23–30 However, prior work has not clarified the cognitive processes that would generate these patterns within a shared semantic structure.

We build on a recent model³¹ conceptualizing semantic search as a process that identifies a restricted subset of the semantic space, while accounting for context and task-demands.^32,33 Whereas the idea that successful semantic retrieval requires different control processes (known as “semantic control”) is well-established,³⁴ our model adds an “attentional component,” that is used to modulate two control parameters of the attended field: its resolution (the precision with which nearby concepts are distinguished) and its size. On this view, SCZ-related abnormalities in semantic search can be consistent with constraints of internal attention (the “internal counterpart” of widely reported (external) attentional dysfunctions in SCZ^{17–19,35,36}), that reduce resolution and/or narrow the attended field.

A useful metaphor for explaining the concept of resolution in semantic memory involves overlaying semantic space with a pixelated grid (Figure 1-1), where each pixel is a tile whose contents are indistinguishable at the current resolution. For illustration, a pixel labeled “dog” may, at higher resolution, subdivide into Chihuahua vs. Poodle, whereas at lower resolution multiple dog exemplars collapse to a single undifferentiated item.

Figure 1.

(1) (A) Semantic space with a grid of pixels; within a pixel, items are indistinguishable at the current resolution. (B) The attended item in a pixel may be a category name (eg, dog) or another representative (eg, elephant). (2) schematic effects of (A) low resolution, (B) narrow attended field, and (C) both. (3) Narrow attended field favoring concrete thinking (restricted access to abstract, relational structure)

A second component is the attended field—the portion of semantic space currently prioritized. Attentional limitations can (1) reduce resolution (larger pixels; coarse tiling), (2) narrow the attended field (fewer accessible concepts), or (3) do both; these modulations shape downstream thought patterns (Figure 1-2). Figuratively, limited attentional capacity is equivalent to having a smaller pixel budget: with fewer pixels, one can maintain high resolution by covering a smaller area (narrower field), maintain area by using larger pixels (lower resolution), or compromise on both—covering less area at lower resolution.

Although the model is not intended to account for formal thought disorder categories, illustrative examples clarify how its constraints could generate phenomena resembling classical descriptions. Low resolution can produce the appearance of loose associations when a small local step spans a large fine-scale semantic distance (eg, dog → elephant), and can similarly encourage over-general groupings (eg, antelopes and elephants as “savanna animals”). Conversely, a narrowed attended field may limit access to broader or metaphorical relations, yielding more concrete interpretations (Figure 1-3), for instance, treating “the apple doesn’t fall far from the tree” literally when abstract mappings fall outside the current field.

In sum, we propose that characteristic semantic impairments observed in SCZ are consistent with reductions in semantic resolution and/or a restricted attended field, arising from attentional constraints rather than purely stochastic responding. To evaluate this account, we analyze the ordered trajectories formed by each participant’s animal CFT sequence and test two key predictions: (1) Low resolution: the trajectories of individuals with SCZ will reflect tighter transitions between abstracted categories, appearing as a loosening of (local) association; (2) Reduced attended field: individuals with SCZ will exhibit a contracted traversal diameter.

Methods

To test our predictions, we compared CFT performance in individuals with SCZ to that of non-psychiatric controls. Examining individuals’ CFT trajectories requires a “normative semantic map” as a reference. We constructed such a map (particularly, a weighted directed semantic graph, see below) using data from a large, independent sample of unselected individuals who also performed the CFT, using a previously developed procedure.^37,38

Participants

Data were collected from 3 groups: (1) Thirty-seven individuals with SCZ (M:F = 5.16:1; mean age = 43.1 years, range = 26-56; mean education = 10.08 years; mean duration of illness (DOI) = 20.37 years) were recruited from the Sha’ar Menashe Medical Center. Diagnosis was confirmed using the Structured Clinical Interview for DSM-IV – Axis I Disorders (SCID). The study was approved by the Sha’ar Menashe Medical Center ethics committee and conducted in accordance with the Declaration of Helsinki. Only patients deemed capable of providing consent were included, and all gave written informed consent. (2) The control group comprised 180 individuals recruited from the Hebrew University of Jerusalem (HUJI: M:F = 1:1.02; mean age = 27.1 years, range = 18-66) and compensated with coffee coupons. Table 1 summarizes the demographic and clinical characteristics of both groups. (3) Another group (henceforth: Graph group), comprising 1559 individuals, whose CFT responses were used to build the semantic network, was recruited through the website www.panel4all.co.il (Panel4Aall: M:F = 1:1.96; mean age = 29.07 years; range: 18-40) and received gift certificates from the panel4all organization. This strategy of comparing SCZ patients and control participants on the platform of a graph representing a “normal semantic space” is based on the assumption that the lexicon of SCZ is preserved and only the retrieval process is damaged.³⁹ All experimental procedures were approved by the ethics committee of the Faculty of Social Sciences at the Hebrew University of Jerusalem.

Table 1.

Demographic and Clinical Characteristics of the SCZ and Control Groups

Variable	Control	SCZ	Statistical comparison
Age	27.1 ± 7.8 (N = 180)	42.8 ± 7.7 (N = 37)	Mann–Whitney U = 521.5; P < .001
Education	14.2 ± 2.3 (N = 150)	10.1 ± 1.5 (N = 37)	Mann–Whitney U = 5283; P < .001
Gender	Male: 89 (49.4%) Female: 91 (50.6%)	Male: 31 (83.3%) Female: 6 (16.2%)	Chi-square = 14.6; P = .0001
DOI	-	20.4 ± 7.7 (N = 37)	N/A (no counterpart)
SANS	-	52.3 ± 30.8 (N = 37)	N/A (no counterpart)
SAPS	-	52.5 ± 30.4 (N = 31)	N/A (no counterpart)

SAPS and SANS scores were unavailable for 6 patients; only 2 patients had comorbidities. Education data were available for a subset of 150 control participants.

Procedure

Participants were recorded during a 1-minute CFT within the semantic category animal names. Control participants and SCZ patients were recorded using a recording device (Philips DVT4010), while participants from Panel4Aall (Graph group) were recorded using a dedicated phone application. Subsequently, all audio recordings were transcribed using the PRAAT program⁴⁰ which provided the words along with time signatures indicating the start and end of each word. For each participant, two lists were generated: a list of words and a list of timestamps, with each timestamp indicating the retrieval time of a word. The timestamps begin at 0, representing the start of the task, and end at 60 s.

Basic Principles of Constructing a Semantic Graph

A semantic graph is a knowledge representation involving vertices that represent concepts and edges that represent connections between concepts. In a weighted semantic graph there is a number attached to each edge (the edge’s weight) representing the similarity or, alternatively, semantic distance between the concepts that constitute the edge’s endpoints. In a directed semantic graph, for any pair of concepts [A, B] two different edges are possible A → B (how similar is A to B) and B → A (how similar is B to A), and the two need not have the same weight. We created a weighted, directed graph based on the graph group performance in the CFT task.

We define a path as the CFT performance of a single subject, that is, a subject’s ordered list of concepts (animals, in our CFT). A subpath is a section of a path. The weight of an edge , which connects concept to concept , is calculated as the median to subpath length, measured in number of steps from A to B. The median was calculated for the set of subpaths starting at and terminating at , but only for sets that include at least (a parameter “min subjects per edge”; we selected the parameter 𝑐 = 10) subpaths (c is a parameter we call “min subjects per edge”: we attempt to balance the requirements of sufficient data and stability.) if the set was smaller than , the distance was not defined. The weight of an edge from A to B, then, was: if where is an A to B subpath.

Our weights are context-sensitive, meaning that two words (eg, Cow and Chicken), which are related in a specific context (eg, farm), are “close” (connected by a short edge) even if this context rarely occurs. For this reason, we favored our weights over the widely used and effective word embedding models.^24,25 In particular, because our weights are derived from a real-time stream of associations, they preserve transient, context-dependent proximities within the attended field that may not be reflected in the global structure of semantic space (and are often averaged out in static embedding models).

Categories Graph

To examine our hypothesis that CFT performance in SCZ patients may reflect low resolution (and hence category- rather than concept-guided transitions), we created an additional graph in which the nodes represent a higher level of abstraction, namely categories. To construct the categories graph, we first had to determine the category for each animal. Automated clustering algorithms initially produced nonsensical categories for too many animals, likely due to the limitations of our dataset (On one hand, we aimed to use our own edge weights, which are context-sensitive and reflect a real-life stream of associations. On the other hand, our dataset features a participant-to-words-per-participant ratio that is too small. Consequently, the automatic clustering performed with this data yielded unsatisfactory results.). Consequently, three research assistants reviewed the animal list and suggested appropriate categories for each (eg, a lion might belong to the categories feline, predator, and savannah). We then validated these categorizations in a second study, where participants (n = 20) were asked to indicate their agreement with each categorization option for each animal (eg, for the word “lion,” the three categorization options mentioned above), or add a new categorization. A category was deemed relevant to an animal if at least 80% of respondents marked it as such. This process resulted in a list of animals, each assigned to one or more agreed-upon categories.

Next, we constructed a weighted, directed graph of categories, where the vertices represent semantic categories and the edges represent semantic distances between them. The weight of an edge from category A to category B was derived from the animals’ graph. Since the animals’ graph is sparse, with most edges undefined, the distance from A to B was influenced by both the lengths of the existing edges and their relative number (considering the clusters’ sizes)—fewer edges between clusters indicate a greater distance. An animal present in both A and B contributed an edge of length 0. Therefore, we calculated the distance from A to B (Category Distance; Cat-d) as the mean of the existing edge weights, divided by the edge density, defined as the number of existing edges between A and B relative to the potential number of edges (|A|·|B|). If no edges existed between two categories, the distance between them was undefined. Notably, our results hold across multiple methods of computing categories graph weights (ie, distances between categories, see Supplementary V).

The next step was building for each participant a path of categories: for a participant who uttered dog—cat—lion—tiger—leopard, the path on the categories graph would be pets—felines. Note that due to the presence of words assigned to multiple categories (eg, a lion being both a feline and a predator), the process of building the categories’ path is not straightforward. To this end, we employed an algorithm that evaluated each category by determining how many subsequent animals are in the path it included. The category encompassing the largest number of following animals was selected. In cases where this step produced categories of equal length, we adopted a greedy strategy, choosing at each step the category closest (in the categories graph) to the preceding one.

Measurements and Hypotheses

In this section, we will introduce the measures by which we test our model. Recall that we focus on 2 attentional adjustments: reduced resolution, which may give rise to loose associations or over-abstraction, and a narrowed attended field, which may give rise to more concrete thinking.

A straightforward way to operationalize resolution in semantic navigation is by examining distances between successive items on the animal graph. Lower resolution should yield transitions that span larger distances on this fine-grained graph. At the same time, reduced resolution is expected to increase reliance on coarser-grained structure, such that transitions are guided more by category-level organization than by item-level proximity. Under these conditions, transitions may appear distant on the animal graph while gaining coherence on the categories graph. Accordingly, we expect paths produced under lower resolution to show larger average distances between successive animals on the animal graph, but smaller distances between successive categories on the categories graph. For example, identifying lion may activate the broader category felines, and under more coarse-grained guidance, the participant may then navigate to a related category such as predators. This higher-level guidance may lead to producing piranha next. Although this transition appears distant on the animal graph, it is a short step within the categories graph. Thus, what might superficially be viewed as a “loose” or unexpected transition can be understood as a coherent move within a more abstract, category-oriented space.

The distance between words was calculated as the average distance between successive animals in the participant’s path on the graph. Conversely, the distance between categories was calculated for each participant as the average defined Cat-d between successive categories in the participant’s path on the categories graph. To further examine the level of resolution characterizing each participant’s path, we calculated the ratio between the average distance between words and the average Cat-d for each participant. The larger this ratio, the poorer the resolution and the greater the abstraction. High resolution results in concept-oriented paths (short distances between words), while low resolution results in category-oriented paths (short distances between categories). Therefore, the ratio between the average distance between words and the average Cat-d indicates the participant’s tendency to abstract at the expense of coherence (keeping short distances between words); It operationalizes the extent to which the retrieval strategy is guided by lower levels of abstraction (relations between individual words) versus higher levels of abstraction (relations between categories).

Finally, to test the hypothesis that the attended field is narrower in SCZ, we computed the path diameter—the mean pairwise distance among all words in the produced sequence (ie, not limited to adjacent pairs). This measure reflects the size of the semantic environment characterizing the participant’s path. We anticipate finding a smaller diameter among participants with SCZ compared to the control group.

In summary, these two measures—the ratio of paths and the path’s diameter—suffice to define the SCZ group’s pattern of performance in the CFT compared to the control group, and to test our hypothesis concerning poor resolution and narrowed attended field (see Figure 2).

Figure 2.

(A) A subgraph with an example path highlighted in blue. The edge lengths in the figure correspond to distances on the graph, with black edges representing shorter distances than the others. (B) The same shown on the categories graph

Data Analysis Strategy

Our analytic approach was designed to assess whether semantic search measures differ between individuals with SCZ and healthy controls, while accounting for potential confounding variables. We first examined group effects on our primary variables of interest—ratio, diameter, and related semantic measures—using generalized linear models (GLMs) with appropriate link functions. Next, demographic variables (eg, age, gender, education) and low-level variables (eg, number of words, word frequency) were included as covariates to evaluate whether group effects persisted after controlling for these factors. Model robustness was assessed using bootstrapping.

Results

Overview of Variables

Before presenting the main findings, we summarize the conventional CFT measures We first report group differences in standard CFT outcomes—total valid responses (verbal productivity), within-category run length, number of distinct categories —followed by basic lexical properties (eg, word frequency). We then focus on our key two model-linked variables: ratio, which operationalizes resolution in semantic retrieval (relative inter-word vs. inter-category distance), and diameter, which indexes the size of the attended field.

The most consistently reported difference between individuals with SCZ and healthy controls in the CFT performance is the number of words produced, or verbal productivity.^4,41 Additional group differences have been observed in the number of semantic categories generated and the number of successive words produced within each category.⁴¹ Word frequency has also been employed as a performance metric in CFT studies.⁴²

A Kolmogorov–Smirnov 2 samples test showed that SCZ patients (n = 37; mean = 13.6; SD = 4.08) produced significantly fewer words than controls (n = 180; mean = 26.26; SD = 6.2), D = 0.85, P < .001, with a large effect size (d = 2.4). Similarly, SCZ participants generated significantly fewer categories (mean = 6.51; SD = 2.06) than controls (mean = 10.62; SD = 2.6), D = 0.72, P < .001, also with a large effect size (d = 1.74). The average number of successive words per category was also lower among SCZ patients (mean = 1.88; SD = 0.65) compared to controls (mean = 2.10; SD = 0.51), D = 0.29, P = .008, yielding a small effect size (d = 0.38). There was no significant group difference in the number of non-successive words per category.

In contrast, word frequency was significantly higher in the SCZ group (mean = 641; SD = 166) than in the control group (mean = 512.1; SD = 114.81), D = 0.39, P < .001, with a large effect size (d = 0.9).

Main Findings

Our primary objective was to test our hypotheses regarding poor resolution and limited attended field among SCZ patients. As detailed below, the SCZ group differed significantly from the control group across all 4 measures: the distance between words, the distance between categories, and, most importantly, the ratio of these two distances on one hand, and the diameter on the other hand (see Figure 3). Note that controlling for both potential confounders related to low-level task performance (ie, number of words, number of categories, word frequency, and number of successive words per category) and demographic variables, the key effects remained significant.

Figure 3.

Distribution of Values for Distance Between Clusters, Distance Between Words, and the Ratio of These Distances and Diameter in the SCZ and Control Groups

Distance Between Words

To assess loosening of associations, we tested whether group (SCZ vs. control) predicts the variable distance between words. We fitted a generalized linear model⁴³ (GLM; Gaussian family, identity link. The approximately normal distribution of the distance between words variable justified the use of a GLM with a Gaussian family and identity link function.) with group as the predictor. Given the small SCZ sample and unequal variances, robust standard errors were used. The model indicated that SCZ group membership was associated with significantly larger distances between words (n = 37, M = 5.48, SD = 1.12) compared to controls (n = 180, M = 4.81, SD = 0.92), β = 0.66, P < .001, d = 0.64 (medium effect)). To ensure the effect was not driven by outliers, we performed bootstrapping (resampling with replacement, 2000 iterations, original sample-sizes; Unless noted otherwise, all subsequent bootstrap analyses used the same settings). The 95% CI did not include zero (β was strictly positive in >99% of iterations), and the effect was significant in 95.05% of iterations.

Distance Between Categories

To examine overinclusion or abstract thinking, we tested whether group membership predicts the variable distance between categories. A GLM (Gaussian family, identity link. The continuous nature and approximately normal distribution of the distance between categories variable justified the use of a GLM with a Gaussian family and identity link function.) with robust standard errors showed that SCZ group membership was associated with significantly shorter Cat-d (n = 37, M = 157.30, SD = 58.32) compared to controls (n = 180, M = 233.70, SD = 79.94), β = −102.19, P < .001, d = 1.08 (large effect). Bootstrapping confirmed the robustness of the effect, with β strictly negative and significant in all iterations.

Ratio Between Paths

We then tested whether group membership predicts the variable ratio between paths. A GLM (Gaussian family, log link; Log-transformed ratio approximated normality, which justiffied Gaussian GLM with log link) with robust standard errors revealed that SCZ group membership was associated with significantly higher ratios (n = 37, M = 0.040, SD = 0.017) compared to controls (n = 180, M = 0.023, SD = 0.010), β = 0.54, P < .001, d = 1.21 (large effect). Bootstrapping confirmed a strictly positive, significant β in all iterations.

Path Diameter

We tested whether group membership predicts the variable path diameter, operationalizing the size of the attended field. A GLM (Gaussian family, identity link) with robust standard errors indicated that SCZ group membership was associated with significantly smaller diameters (n = 37, M = 6.70, SD = 0.44) compared to controls (n = 180, M = 7.25, SD = 0.26), β = −0.55, P < .0001, d = 1.50 (large effect). Bootstrapping confirmed a strictly negative, significant β in all iterations.

Finally, the association between ratio and diameter was small in both groups (SCZ: Spearman ρ = −0.17, P = .30; Pearson r = −0.11, P = .52; Control: ρ = −0.27, P < .001; r = −0.25, P < .001), indicating that these measures do not capture the same construct but rather reflect distinct characteristics of the CFT retrieval pattern in SCZ.

Distributional Properties of Key Variables

To complement the GLMs—given unequal variances between groups and occasional departures from normality in residual diagnostics—we conducted two-sample Kolmogorov–Smirnov (KS) tests. KS tests indicated significant SCZ–control differences for all variables (distance between words: D = 0.36, P < .001; Cat-d: D = 0.46, P < .001; ratio: D = 0.53, P < .001; diameter: D = 0.65, P < .001).

Robustness to Sampling Variability and Group-Size Imbalance

Because the control group was much larger than the SCZ group, we repeatedly drew equal-N control subsets and re-estimated the group differences. Specifically, we repeatedly drew 2000 random control subsamples of 37 participants each; Sampling was performed without replacement within each subsample, ensuring that no individual appeared twice in the same draw, and with replacement across iterations, allowing individuals to reappear in different subsamples. The SCZ–control difference remained in the expected direction for path ratio (median d = 1.36; positive in all samples) and diameter (median d = −1.20; negative in all samples), and the two-sample Kolmogorov–Smirnov test was significant (P < .05) in 100% of draws across measures.

We also assessed sensitivity to within-group variance using small-N resampling: we drew 20 SCZ and 20 control participants 2000 times and compared group means. Across draws, SCZ showed higher path ratio (median d = 1.30) and smaller diameter (median d = –1.55), with the direction preserved in 100% of draws.

Control for Potential Confounds

We evaluated whether lower-level production and lexical features confounded the association between group (SCZ vs. control) and the key outcomes (ratio, diameter). Candidate confounds were: total number of words, number of categories, mean successive words per category and mean word frequency (see Supplementary I  Table 2).

We first screened for multicollinearity. Variance-inflation diagnostics indicated elevated collinearity for number of words (Number of words showed high collinearity, VIF = 5.76, and was excluded; all remaining VIFs were < 2.1). Since the number of words had the highest VIF score, we excluded it from the model, which reduced the VIF values as follows: group = 1.6, number of categories = 2.1, mean successive words per category = 1.74, frequency = 1.65.), so it was therefore excluded from the multivariable model and examined in a separate sensitivity analyses.

Ratio

We fit a GLM with Gaussian family and log link to predict ratio from group, number of categories, mean successive words per category, and mean word frequency. Frequency was not significant and was dropped. The reduced model showed significantly larger ratio in SCZ than control (β = 0.21, SE = 0.08, P = .01, Full GLM output appears in Supplementary III  Table 4). In a non-parametric check (2000 bootstrap resamples), the group coefficient for ratio was positive in >99% of iterations. Adding Group × confounder interactions did not improve fit (all ns). As a sensitivity test, we reintroduced number of words and tested Group × number of words; the interaction was non-significant and the group effect remained unchanged (β = 0.21, SE = 0.099, P = .034). See appendix III-b for a detailed description of analysis with number of words as a confounder.

Diameter

Given approximate normality across groups, we fit a Gaussian GLM with an identity link. Number of categories and mean successive words per category were excluded as non-significant; The reduced model, retaining only significant predictors, showed significantly smaller diameter in SCZ than control (β = −0.001, SE < 0.001, P < .001, Full GLM output appears in Supplementary III  Table 7). In 2000 bootstrap resamples, the group coefficient was negative and significant in 100% of iterations. Interaction terms did not improve fit (likelihood-ratio tests ns; only the main effect of frequency remained significant). Adding number of words neither altered the conclusions nor reached significance. See Supplementary III-b for a detailed description of analysis with number of words as a confounder.

In sum, across both outcomes—ratio and diameter—the Group effect (SCZ vs. control) remained robust after adjusting for production and lexical confounders; no Group × confounder interactions improved fit, sensitivity analyses (including reintroducing number of words) left conclusions unchanged, and bootstrap checks confirmed the direction and significance of the Group coefficient.

Sociodemographic Checks

Group Differences

Age and gender differed significantly between the SCZ and control groups. In the subset with education data (years of schooling; n = 150 controls), education also differed between groups.

Associations with Outcomes

Spearman correlations showed no association between age and either outcome in either group (SCZ: ρ = 0.13, P = .44 for ratio; ρ = −0.11, P = .50 for diameter; Controls: ρ = 0.03, P = .70 for ratio; ρ = −0.01, P = .87 for diameter). In the education subset, education was likewise uncorrelated with the outcomes (SCZ: ρ = 0.21, P = .20 for ratio; ρ = −0.01, P = .93 for diameter; Controls: ρ = 0.00, P = .92 for ratio; ρ = −0.01, P = .86 for diameter). As for gender, Kolmogorov–Smirnov tests indicated no differences between males and females within groups (SCZ: D(36) = 0.28, P = .73; ratio means M = 0.040, F = 0.041; D(36) = 0.25, P = .80; diameter means M = 6.70, F = 6.73. Controls: D(179) = 0.13, P = .34; ratio means M = 0.022, F = 0.024; D(179) = 0.17, P = .11; diameter means M = 7.29, F = 7.21).

Covariate-Adjusted Models in the Education Subset

To assess whether group differences might be driven by demographic differences, we repeated the analyses in the subset with complete demographic data (SCZ and the 150 controls). We first attempted propensity-score matching, but standardized mean differences (SMDs) remained high (age = 0.72; gender = 0.32; education = 1.67); Even when relaxing the balance criterion from the conventional SMD < 0.10 to a more lenient 0.30 threshold, the matched samples remained very small. We therefore fit covariate-adjusted GLMs.

Ratio

A Gaussian GLM with a log link included group, age, gender, and education as predictors. Group was a significant predictor (β = 0.60, SE = 0.12, P < .001), indicating higher ratio in SCZ; age, gender, and education were not significant. Multicollinearity was low (VIFs <2.3). In bootstrap resampling, the group coefficient was positive in >99% of iterations (and significant in >99%). See Supplementary III for details on model specification and fit diagnostics.

Diameter

We fit a Gaussian GLM with an identity link including group, age, gender, and education. Group was a significant predictor (β = −0.579, SE = 0.093, P < .001), indicating smaller diameters in SCZ; age, gender, and education were not significant. In bootstrap resampling, the group coefficient was negative in 100% of iterations (and significant in >99%), consistent with smaller diameters in SCZ. See Supplementary III for details on model specification and fit diagnostics.

Finally, we conducted the analysis for the whole group of 180 participants, without education, to assess whether group differences might be driven by age and gender in this group.

Ratio

A Gaussian GLM with a log link included group, age and gender as predictors. Model fit was good (deviance explained = 26%). Group was a significant predictor (β = 0.50, SE = 0.091, P < .001), indicating higher ratio in SCZ; age and gender were not significant. Multicollinearity was low (VIFs <1.63). In bootstrap resampling, the group coefficient was positive in 100% of iterations (and significant in >99%). See Supplementary III for details on model specification and fit diagnostics.

Diameter

We fit a Gaussian GLM with an identity link including group, age and gender. Group was a significant predictor (β = −0.546, SE = 0.088, P < .001), indicating smaller diameters in SCZ; age and gender were not significant. In bootstrap resampling, the group coefficient was negative in 100% of iterations (and significant in 100%), consistent with smaller diameters in SCZ. See Supplementary III for details on model specification and fit diagnostics.

Please refer to Supplementary IV for an analysis that controls for cluster centrality, where we examine whether the short distances on the categories graph in the SCZ group result from the specific categories they visit rather than the relationships between successive clusters. In other words, independent of the frequency variable (which was controlled for), SCZ individuals tend to visit categories with a short mean distance to others. We show that even when controlling for the main distance of categories, the paths traveled by the SCZ group remain significantly shorter than those of the control group.

Discussion

Semantic impairments in SCZ encompass diverse alterations in how semantic information is accessed and organized, yet most prior studies using the CFT^35,44–46 have focused primarily on incoherent or disorganized output. The present work offers a complementary perspective by examining whether semantic impairments in SCZ can be understood through constraints on internal attention operating over the semantic network. Rather than targeting formal symptom categories, we asked whether reduced representational resolution and/or a narrowed attended field could provide a unified mechanistic account of characteristic retrieval patterns. We hypothesized that such constraints would manifest as two compensatory adjustments—a shift toward more abstract, category-level guidance and a contraction of the effective traversal radius. Using a path-based analysis of CFT sequences, we derived indices mapping onto these two parameters and found patterns consistent with both predicted adjustments.

A particularly notable finding, which provides specific support for our account, concerns the reduced distances on the categories graph. Indeed, whereas increased semantic distances between subsequent responses are commonly reported in the CFT, and could afford different explanations, a concurrent reduction in distances at a more abstract level suggests a specific underlying organizing principle - a transition to abstraction. We interpret this shift as a compensatory response to attentional constraints.

Our findings also complement the optimal foraging model of semantic search, that focuses on predicting a strategic choice to switch between semantic clusters based on estimated “depletion of concepts” in the current cluster.⁴⁷ Using this model to explain CFT performance of individuals with SCZ, Lundin et al.⁴⁷ showed that individuals with SCZ took longer to transition between words within clusters, while their time to switch between clusters was similar to that of controls. Our account complements these findings of a disrupted balance between local and global transitions, by formalizing foraging within clusters (high-resolution search) and between clusters (low-resolution search) as two ends of a spectrum of abstraction, with individuals with SCZ occupying the more abstract end.

Several potential caveats of this research should be noted: First, the dataset was not designed to link graph-derived measures to formal thought disorder ratings or to specific clinical symptom clusters; accordingly, detailed assessments such as TLC scores, PANSS subscales, neurocognitive batteries, or CPZ-equivalent medication doses were not available. Furthermore, our demonstration that reduced resolution or a narrowed attended field can qualitatively resemble classical symptom descriptions does not imply that these variables serve as clinical markers. Rather, they reflect computational constraints that may contribute to heterogeneous behavioral profiles observed in SCZ. Second, all SCZ patients in our study were undergoing treatment with antipsychotic medications, which complicates the attribution of performance impairments to SCZ itself versus the effects of the medication.

Third, in our account, attention functions as a theoretical mechanism—not a directly measured variable—that organizes our predictions about semantic retrieval. Nevertheless, because we did not administer a direct attention assessment, our inferences are limited to patterns consistent with an attention-constrained account; future work should include convergent measures of attention and semantic control.

Fourth, our results reflect performance in a short, 60-s CFT. This standard method has been repeatedly used for graph/network analyses in aging, neurodegeneration, and SCZ.^45,48–51 Yet, it should be noted that a 1-min sample necessarily limits the number of transitions and may under-sample rarer items and late-phase switches, potentially attenuating estimates of longer-range effects.

Fifth, we did not explicitly code idiosyncratic transitions such as rhyme-/sound-based sequences (eg, cat → bat) or episodic/literary links (eg, bear → snake from The Jungle Book). Within our semantic model, such transitions would tend to appear as mismatches to category structure (eg, longer or atypical edges) and thus likely dilute effects by adding noise. Nevertheless, if a subset of participants consistently relies on these guides, estimates of aperture/diameter and distance-based ratios could be biased.

Finally, we did not quantify or model hub-like words that can function as cross-category bridges (eg, orca, which affords multiple classification routes—mammal/whale/marine predator—versus more canonical items such as goldfish). Conceptually, we use “hub-like” to denote an item’s position in the semantic network (node-level topology), not a local pairwise transition cue. If reliance on such hubs differs between groups, this omission could bias parameter estimates (eg, switch probability, diameter/resolution) or add noise that attenuates effects. Future work should operationalize hubness (eg, multi-category membership; graph degree/betweenness; polysemy/co-classification) and re-estimate models after including hub status as an item-level covariate or after excluding/down-weighting high-hub tokens to test robustness.

Our research generates novel predictions for future studies. Future work could examine how the proposed attentional constraints relate to individual differences in clinical presentation, including formal thought disorder ratings, cognitive functioning, illness course, or medication status. Such studies would enable direct testing of whether graph-derived features correspond to specific symptom dimensions or reflect broader cognitive phenotypes. More broadly, integrating computational, clinical, and neurocognitive measures may clarify how internal attentional limitations interact with other mechanisms contributing to semantic disturbances in SCZ.

Future research could also investigate whether experimentally induced attentional load or dual-task demands in healthy individuals lead to a thinking pattern similar to that observed in SCZ. Healthy individuals performing the CFT task under a divided attention manipulation should exhibit increased overinclusion and concrete thinking in their responses. Similarly, we predict that individuals with attention deficit disorders (ADD) will show a retrieval pattern more similar to that of the SCZ group and less like that of healthy controls; However, the precise degree and nature of this effect may vary depending on the specific attention disorders classified under the ADD umbrella and its severity. If confirmed, these findings would support the view that attentional constraints contribute to semantic-retrieval differences in SCZ. Recognizing SCZ as potentially involving attentional constraints on retrieval could encourage researchers and clinicians to explore treatments and diagnostic methods typically used for ADD in SCZ.

Another promising avenue for future work concerns contextual specificity—how contingent an association is on a narrow context or sense—and its relation to concrete or inflexible thinking. We did not directly quantify contextual specificity in the present study. Instead, we contrasted associations by semantic proximity (shorter vs. longer inter-item distances). In our data, shorter-distance (“close”) associations tended to index intact semantic functioning. Importantly, closeness should not be conflated with impoverished output that arises when the attended field is restricted (small diameter): a narrowed search window limits available candidates, yielding repetitive or overly constrained responses that may appear specific but in fact reflect reduced flexibility. Future studies could explicitly code contextual specificity at the item level (eg, tagging responses as context-bound vs. general, or deriving corpus-based specificity indices) and test whether specificity adds explanatory power for inflexible retrieval beyond the aperture/diameter and resolution parameters in our model.

In summary, our research aimed to deepen the understanding of the mechanisms underlying thought disorders in SCZ. Our results align with the hypothesis that the pattern observed in SCZ CFT performance is consistent with reduced resolution and/or a narrowed attended field. Insights into these mechanisms could advance the development of more precise diagnostic, preventive, and therapeutic approaches. Together, these path-based signatures at the item and category scales are consistent with an attention-constrained retrieval account and motivate future work pairing our trajectory measures with independent assessments of attention and semantic control.

Funding

This work was funded by The Israel Science Foundation grants (grant 1471/20 to AM; grant 1633/24 to IF).

Conflicts of Interest

None declared.