Schizophrenia Research 259 (2023) 97–103

Contents lists available at ScienceDirect

Schizophrenia Research
journal homepage: www.elsevier.com/locate/schres

Language network self-inhibition and semantic similarity in first-episode

schizophrenia: A computational-linguistic and effective
connectivity approach
María Francisca Alonso-Sánchez a, b, *, Roberto Limongi b, Joseph Gati b, c, d, e,
Lena Palaniyappan b, c, d, e
a
CIDCL, Fonoaudiología, Facultad de Medicina, Universidad de Valparaíso, Chile
b
Robarts Research Institute, Western University, London, Ontario, Canada
c
Centre for Youth Mental Health Service Innovation, Research and Training, Douglas Mental Health University Institute, Montreal, Quebec, Canada
d
Department of Psychiatry, McGill University, Montreal, Quebec, Canada
e
Department of Medical Biophysics, Schulich School of Medicine and Dentistry, Western University, London, Ontario, Canada

A R T I C L E I N F O A B S T R A C T

Keywords: Introduction: A central feature of schizophrenia is the disorganization and impoverishment of language. Recently,
Psychosis we observed higher semantic similarity in first-episode-schizophrenia (FES) patients. In this study, we investigate
Lexical access if this aberrant similarity relates to the ‘causal’ connectivity between two key nodes of the word production
fMRI
system: inferior frontal gyrus (IFG) and the semantic-hub at the ventral anterior temporal lobe (vATL).
Spectral dynamic causal modelling
Broca's area
Methods: Resting-state fMRI scans were collected from 60 participants (30 untreated FES and 30 healthy con­
Disorganization trols). The semantic distance was measured with the CoVec semantic tool based on GloVe. A spectral dynamic
Formal thought disorder causal model with Parametrical Empirical Bayes was constructed modelling the intrinsic self-inhibitory and
extrinsic-excitatory connections within the brain regions. We estimated the parameters of a fully connected
model with the semantic distance as a covariate.
Results: FES patients chose words with higher semantic similarity when describing the pictures compared to the
HC group. Among patients, an increased semantic similarity was related with an increase in intrinsic connections
within both the vATL and IFG, suggesting that reduced ‘synaptic gain’ in these regions likely contribute to
aberrant sampling of the semantic space during discourse in schizophrenia.
Conclusions: Lexical impoverishment relates to increased self-inhibition in both the IFG and vATL. The associated
reduction in synaptic gain may relate to reduced precision of locally generated neural activity, forcing the choice
of words that are already ‘activated’ in a lexical network. One approach to improve word sampling may be via
promoting synaptic gain via supra-physiological stimulation within the Broca's-vATL network; this proposal
needs verification.

1. Introduction semantic system while preserving the thematic progression of discourse

aligned to the communicational goal (Levelt, 2001). The rapid and
When we speak, we choose words relevant to the specific context by precise production of diverse words is critical to fluent language;
drawing on our knowledge of what those words represent. Semantic restricting the use of words to a limited repertoire of concepts may
memory is seen as the storage of words (“bird”) and many related con­ interfere with the efficiency of communication.
cepts they represent (they fly, chirp, have feathers or even that some of Semantic impairment is a prominent linguistic feature of schizo­
them only swim, but they are still in the same category). Word selection phrenia (Brown and Kuperberg, 2015). Previous studies have demon­
process involves overcoming competition (i.e., resolution of conflicts) strated a failure to inhibit the ‘retrieval’ or ‘activation’ i.e., the use of
but this occurs swiftly without affecting the efficiency of a discourse. A conceptually related words that are not appropriate for the context in
top-down control is said to drive the lexical unit selection within the patients with schizophrenia (Almeida and Radanovic, 2021). This

* Corresponding author at: Robarts Research Institute, 1151 Richmond St., UWO, London, Ontario N6A 5B7, Canada.
E-mail address: mariafrancisca.alonso@uv.cl (M.F. Alonso-Sánchez).

https://doi.org/10.1016/j.schres.2022.04.007
Received 24 March 2022; Received in revised form 28 April 2022; Accepted 29 April 2022
Available online 11 May 2022
0920-9964/© 2022 Elsevier B.V. All rights reserved.
M.F. Alonso-Sánchez et al. Schizophrenia Research 259 (2023) 97–103

increased semantic priming is seen especially in those with clinically performance in patients with schizophrenia, specifically related to the
detectable thought disorder (Pomarol-Clotet et al., 2008), with a higher cognitive control of lexical/semantic selection. Surprisingly, few pub­
degree of similarity between words contributing to redundant discourse lished studies have examined the aspects of the causal connectivity of
with reduced information content (Alonso-Sánchez et al., 2022). Se­ brain regions in relation to aberrant lexical selection in schizophrenia.
mantic deficits have been generally described as part of disorganization The term effective connectivity refers to the causal influence that one
symptoms (Elvevåg et al., 2007; Hoffman et al., 2020) and poverty of neural system exerts over other neural systems, while discounting other
content (e.g., lower semantic density (Rezaii et al., 2019)). These fea­ influences. This directed causal relation is widely different from the
tures often appear before the onset of psychosis (Mota et al., 2017; structural connectivity that uses white matter tracts (physical architec­
Spencer et al., 2021) and relate to social and occupational outcomes in ture), or from functional connectivity that estimates only statistical
early stages of the illness (Palaniyappan et al., 2019). dependencies (correlation) between brain regions. Dynamic causal
A wide variety of computational tools (natural language processing, modelling (DCM) is a generative model framework for effective con­
a.k.a. NLP) are now accessible to detect speech features that are not nectivity hypothesis testing between neural populations (Zeidman et al.,
readily perceptible to the human eye/ear. These tools can readily cap­ 2019a; Zeidman et al., 2019b). DCM uses a forward model to estimate
ture surface features such as word repetitions (type-token ratio), and the influence of one brain region over another by modelling the neural
also unravel the content of the texts and the relationship between the fluctuation through Bayesian inversion of the brain blood oxygen level
words that compose them (Levy et al., 2015; Mikolov et al., 2013). One response (Friston, 2011; Friston et al., 2003). For resting-state fMRI
of these approaches, termed distributional semantics, clusters words instead of estimating the neuronal states time-varying fluctuation, we
into neighbourhoods based on the frequency of their co-occurrence in estimate the parameters of the cross-correlation functions (spectral
everyday language (e.g. ‘snack’ and ‘plate’) while distancing widely DCM) by modelling endogenous neural fluctuations.
those that are highly improbable to appear together (‘snack’ and Notwithstanding the considerable amount of literature that relates
‘planet’). Specifically, the GloVe semantic similarity tool benefits from a semantic memory with the anterior temporal lobe and cognitive control
high-dimensional word vectorization framework and is based on a log- with the inferior frontal gyrus (Binder et al., 2009; Chiou and Lambon
bilinear regression model that uses both local and global statistics of a Ralph, 2019; Jackson, 2021; Jackson et al., 2016; Rice et al., 2015;
corpus, unlike other methods (Word2Vec uses only local while Latent Visser et al., 2010), to our knowledge, there have been no attempts to
Semantic Analysis uses only global statistics) (Pennington et al., 2014). examine the effective connectivity between frontal and temporal regions
The local statistic computation captures the relation between words, as in patients with schizophrenia and subtle semantic disturbances. Several
shown in the analogy task (king – man + woman = queen) while the confounds need to be addressed when undertaking such investigations.
global statistics profit from all the document and not only by the sur­ First, as long-term exposure to antipsychotic medications can affect
rounding words. This approach computes the ratio of co-occurrence connectivity parameters and the inferred causal influences (Schmidt
probabilities between adjacent words within a frame. Thus, we can es­ et al., 2016), we recruited a sample of untreated, but first-episode pa­
timate the cosine similarity within words to unravel the semantic di­ tients. Second, subjective measures of speech are replete with problems
versity in a discourse sample. In our study, we infer the lexical unit of reliability; employing automated measures with corpus-based quan­
selection from the putative semantic library using the GloVe cosine tification of statistical co-occurrence of semantic distance is critical for a
similarity from a sample of a descriptive discourse (Kanerva, 2009; reliable approach. We employ an NLP measure of lexical similarity and
Pennington et al., 2014). explore the effective connectivity between two core areas of the se­
The anatomical aspect of brain network recruitment in lexical/se­ mantic network: IFG and vATL in untreated first episode of schizo­
mantic tasks is still a matter of intense debate, but a growing body of phrenia patients (mean 6 days of lifetime daily dose exposure). We
evidence supports the view that the ventral Anterior Temporal Lobe hypothesize that lexical inefficiency, indexed by higher semantic simi­
(vATL) is critical for semantic cognition, constituting a putative Se­ larity between words, will relate to the effective connectivity of IFG and
mantic Hub (Visser et al., 2010). The hub is part of a larger network that vATL in patients with first episode of schizophrenia.
comprises the uncinate fasciculus, frontal-occipital fasciculus and
ventral occipitotemporal areas (Chiou et al., 2018; Jackson et al., 2019; 2. Materials and methods
Krieger-Redwood and Jefferies, 2014). The inferior frontal gyrus (IFG) is
related to the regulation of automatic spreading activation/inhibition of 2.1. Participants
candidates/competitors within the lexical network (Chiou et al., 2018;
Python et al., 2018). The top-down control mediated by IFG contributes Resting-state fMRI scans were collected from 60 participants at the
to both lexical retrieval (the activation of several possible relevant re­ Prevention and Early Intervention Program for Psychosis in London,
sponses) and selection (the narrowing of choices to the most appropriate Ontario, Canada. The sample included 30 subjects with untreated First
one) (Conner et al., 2019; Snyder et al., 2011; Wagner et al., 2001; Episode of schizophrenia (FES group) and 30 healthy controls (HC
Whitney et al., 2009). group). 68% of this sample has been included in our prior work that
In patients with schizophrenia, the frontotemporal network was re­ focused on longitudinal speech data (Alonso-Sánchez et al., 2022).
ported as underactive during semantic processing and the authors sug­ Severity of symptoms was confirmed with the Positive and Negative
gested that these results might be explained by an inhibition impairment Syndrome Scale (PANSS) (Kay et al., 1987). Language disorganization,
(Arcuri et al., 2012). Horn reported a deviant semantic network with poverty of speech, and perseveration of ideas were assessed with the
increased severity of formal thought disorder in which the difference Thought and Language Index (TLI) scale. The Impoverishment in
was due to a decreased contribution of the left pars orbitalis, and the Thinking is the sum of poverty of speech, weakening of goal and
authors highlighted the possible influence of semantic retrieval demands perseveration of ideas; while disorganization of thinking is the scores
in the context of selection conflict (Horn and Federspiel, 2012). sum of looseness, distractibility, peculiar use of words, sentences, and
Adopting a similar interpretation, when patients with schizophrenia logic. The FEP group was drug-naïve for antipsychotics at the time of
showed greater activation in the left inferior frontal gyrus (BA45) while assessment (total antipsychotic use less than 14 days). The inclusion
performing a semantic task, the authors argued that this greater acti­ criteria for the HC group were not having (a) personal or family mental
vation was due to a difficulty in semantic representation access (Chen illness or (b) neurological diseases. All participants provided written
et al., 2013). Surbeck reported structural aberrations in white matter informed consent before assessment and ethics approval was granted by
tracts underlying semantic processing in the ventral language stream of the Human Research Ethics Board at Western University, London,
the brain (Surbeck et al., 2020). Taken together, these studies support a Ontario.
role for the inferior frontal gyrus in the disruption of semantic

98
M.F. Alonso-Sánchez et al. Schizophrenia Research 259 (2023) 97–103

2.2. Assessments field, realignment, co-registration and segmentation. A DARTEL tem­

plate was built with all T1 images of the subjects, and the fMRI images
All participants were assessed with the Social and Occupational were normalized to the Montreal Neurological Institute (MNI) space. We
Functioning Assessment Scale (SOFAS) to quantify the functioning level fitted a general linear model with six head movement parameters and
in social and occupational domains (Rybarczyk, 2011) and with the the white matter and cerebrospinal fluid time series and a cosine basis
Socioeconomic Status (SES) to measure the parental level of occupation set with frequencies ranging from 0.0078 to 0.1 Hz as regressors. All
and employment from 1 (Managerial and professional occupations) to 5 images were high pass filtered to discard slow frequency drifts (0.0078
(routine occupations) (Boyd, 2008). In addition, we used The Edinburgh Hz). The volumes of interest were defined a priori as Brodmann area 45
Inventory (Oldfield, 1971) to define the handedness of the participants, (BA45) in the inferior frontal gyrus (MNI: − 44 28 22) due to its role in
the scale is scored from − 12 (totally left-handed) to +12 (totally right- lexical selection (Heim et al., 2008) and the sematic hub located in the
handed). ventral anterior temporal lobe (MNI: − 40 − 15 − 33) (Chiou and Lambon
Ralph, 2019) (shown in Fig. 1).
2.3. Semantic measures
2.5. Effective connectivity
To elicit the language sample, the participants were instructed to
describe 3 pictures of the Thematic Apperception Test images (Murray, To compare the effective connectivity between groups we used a
1943) during 1 min per picture and were prompted if they stopped spectral dynamic causal modelling (DCM) approach with Parametrical
speaking before one-minute time window elapsed. The task was recor­ Empirical Bayes (PEB) (Friston, 2011; Friston et al., 2014). PEB is a
ded and later transcribed by research assistants. We analyzed the se­ hierarchical Bayesian model that uses both non-linear and linear ana­
mantic distance between words with the Covington Vector semantic tool lyses to measure the group level effects and subject-by-subject vari­
(CoVec version 1.0.5912) (Covington, 2016). CoVec is a natural lan­ ability on parameter estimates. We use bilinear modulatory effects with
guage processing tool based on data from GloVe Project, with 840 billion two states per region, excitatory (inter-region) and inhibitory (within-
words in English with 300-element vectors (Pennington et al., 2014). region), without stochastic effects (Friston et al., 2014).
CoVec measures the likelihood of co-occurrence of words through vector We evaluate the evidence in support of the hypothesis that both
cosine similarity. The text preprocessing removes punctuation, marks group membership and semantic similarity better explained the effective
‘stopwords’ (e.g. “a”), and finally, ignores words which are not found in connectivity within the two-node network than group alone. We also
the GloVe dataset (displays a warning of all the missing words). CoVec assessed the evidence for an interaction between group and semantic
reports the average of similarity of successive words with an n-word similarity in explaining the 2-node network's effective connectivity. To
frame segment, using all the positions of the frame. In this analysis we this end, we compared the following as regressors predicting connec­
used a full frame including all the text per each picture. tivity: Model 1: Group; Model 2: Group and Semantic similarity; Model
3: Group, Semantic Similarity, and group × semantic similarity inter­
2.4. fMRI acquisition and analysis action. The semantic regressor was mean-centered and the interaction
between variables was summarized as the product of the linguistic
The acquisition and processing are described in detail in our prior measure and group. For the comparison we estimated the free energy
work reporting baseline data from this sample (Dey et al., 2021; Limongi (accuracy minus complexity) of the three models considering group and
et al., 2020). In brief, all data were acquired at the Centre for Functional semantic co-occurrence and interaction as covariates. We select the
and Metabolic Mapping (CFMM) at the University of Western Ontario on model with higher free energy and analyze the parameter of each vol­
a Siemens 7 T Plus (Erlangen, Germany) using a multi-band EPI acqui­ ume of interest with the intrinsic and extrinsic connectivity strength and
sition sequence (Feinberg et al., 2010; Moeller et al., 2010) with the the corresponding posterior probabilities.
following imaging parameters: TE/TR = 20 ms/1000 ms, flip angle =
30◦ , 63 slices with multi-band factor = 3, iPat = 3 and an isotropic 2.6. Statistical analysis
resolution of 2 mm. A total of 360 functional volumes were collected for
a total acquisition time of 6 min, 26 s. T1-weighted MP2RAGE Demographic and clinical data were analyzed using descriptive and
anatomical volume (TE/TR = 2.83/6000 ms) at 750 μm isotropic reso­ Bayesian statistics. We use Bayesian t-test for continuous variables, and a
lution was also acquired for anatomical registration. fMRI data was pre- Bayesian independent multinomial test for categorical variables. Log
processed with SPM12 (available free at http://www.fil.ion.ucl.ac.uk/ odd ratios were computed with the 95% credible interval for categorical
spm/software/spm12/) on Matlab. The pre-processing included variables. Bayes factors were considered substantial over 3, strong over
manual reorientation on the anterior and posterior commissure, voxel 10 and very strong over 30 for all the hypotheses tested according to
displacement maps calculation for inhomogeneities in the magnetic Jeffreys schema (Jarosz and Wiley, 2014). All the statistical analyses

Fig. 1. Anatomical map of voxel positioning for dynamic causal modelling. Brodmann area 45, ventral anterior temporal lobe. Model for DCM. Plot made with
BrainNet Viewer (Xia et al., 2013).

99
M.F. Alonso-Sánchez et al. Schizophrenia Research 259 (2023) 97–103

used JASP version 0.14.0.1 (JASP Team, 2020) and the figures plots
were made on Python in Jupyter Notebook 6.1.5 (Kluyver et al., 2016).

3. Results

3.1. Demographic

Clinical and demographic data are shown in Table 1. The groups did
not differ regarding age, gender or educational level. There was no
group difference in the use of English as the first language (83% FES and
90% HC had English as the first language, BF10 = 0.281, 95% CI: − 0.93,
1.96), in the self-reported immigrant status as first generation (20% FES
and 26% HC were first generation immigrants, BF10 = 0.317, 95% CI:
− 1.50, 0.81) and in the ethnicity of the participants (67% FES and 73%
HC were white, BF10 = 0.334, 95% CI: − 1.37, 0.77). All the participants
had English as their transactional language. In the FES group the 50%
was antipsychotics naïve, while the other 50% was expose to a mean of Fig. 2. 95% highest credible interval of semantic similarity per group. HC:
healthy controls; FES: first-episode schizophrenia.
6 days of a lifetime daily dose (SD: 3.2, range: 2–12).

the pictures offered, patients generate content that is likely to be as

3.2. Semantic similarity specific and as diverse as healthy subjects, but still choose words that are
semantically close to each other.
The groups did not differ in the number of words spoken during the
descriptive discourse (HC: 70.5 ± 12.8, FES: 72.7 ± 33.9, BF = 0.266,
3.3. Model selection
effect size δ 95% CI = − 0.50, 0.42), but as hypothesized, the FES group
had higher semantic similarity than the HC group with very strong ev­
When comparing the models regarding the regressors included on
idence favoring the between-group difference hypothesis (HC: 0.330 ±
them, the free energy of the second model (group and semantic simi­
0.02, FES: 0.364 ± 0.03, BF = 1623, effect size δ 95% CI = − 1.71,
larity) exceeded the other models' performance (Fig. 3). The third model
− 0.59) (Fig. 2).
(Group, semantic similarity, and interaction) was an overfit for the data.
We also examined whether the higher semantic similarity in FEP
comes from generating content that is non-specific to the picture (i.e.,
3.4. Effective connectivity
specificity) and/or from content that is ‘stuck’ in describing limited
features of the picture at hand (i.e. insufficient diversity). These results,
The parameter estimates from the winning model indicated that the
presented in the Supplementary materials, indicate that when describing
intrinsic inhibitory connections of BA45 and vATL increased as a func­
tion of semantic similarity in the FES group (FES coded as 1; HC coded as
Table 1
− 1). The parameters (Fig. 3) indicate that the intrinsic inhibitory
Mean and standard deviations are shown for continuous variables, with per­
connection of BA45 and vATL increased linearly with semantic simi­
centages for categorical variables. BF10: Bayes Factor. PANSS: Positive and
larity. The mean estimated connection strength within BA45 [intrinsic]
Negative syndrome scale. SOFAS: Social and Occupational Functioning Assess­
ment Scale. TLI – Thought and Language Index. is 9.125 (posterior probability, PP = 1), from BA45 to vATL -3.464 (PP:
0.90), from vATL to BA45 -3.060 (PP: 0.87) and, within vATL [intrinsic]
HC (n = FES (n = BF10 Effect size
is 11.034 (PP: 1.0).
30) 30) δ 95% CI
Mean ± Mean ±
SD SD 4. Discussion
Age 21.4 ± 21.4 ± 0.262 − 0.46,
3.5 2.2 0.46 In this spectral DCM study leveraging an NLP measure of semantic
Gender 73% 73% 0.277 − 1.11, similarity, we demonstrate 3 major findings. First, patients with FES
male male 1.13 chose words with higher degree of semantic similarity (i.e., lower se­
Educational level (<12/≥12 30%/ 46%/ 0.713 − 1.73,
years) 70% 54% 0.35
mantic distance from each other), despite speaking comparable number
Edinburgh Handedness 10.4 ± 9.8 ± 4.2 0.308 − 0.32, of words when describing pictures. Second, at resting state, we recov­
Inventory 3.1 0.60 ered a fully connected network between the temporal semantic hub
PANSS-8 Positive – 12.3 ± – – (vATL) and inferior frontal Broca's area (BA45); the connectivity of this
3.5
network was better explained by considering inter-individual differ­
PANSS-8 Negative – 7.6 ± 4.3 – –
PANSS-8 total – 26.4 ± – – ences in semantic similarity along with diagnosis, compared to diag­
7.4 nostic status alone. Third, we show that in FES, higher intrinsic
SOFAS 81.9 ± 39.0 ± >10,000 inhibitory connectivity within vATL and BA45 at rest relates to an
5.0 14.7 aberrant increase in semantic similarity of word choices when partici­
Parental SES (<3/≥3) 44%/ 30%/ 1.00 − 0.24,
56% 70% 1.95
pants describe a picture. These findings provide a neural model linking
Total TLI 0.29 ± 1.46 ± 736.2 − 1.64, speech production and semantic storage regions in the lexical impair­
0.3 1.3 − 0.53 ment seen in first episode schizophrenia.
TLI Impoverishment in 0.13 ± 0.47 ± 6.8 − 1.16, Higher inhibitory connections reflect a loss of synaptic gain within
thinking 0.2 0.6 − 0.14
cortical microcircuits. In an excitation/inhibition framework, an in­
TLI Disorganization in 0.15 ± 0.99 ± 97.6 − 1.45,
thinking 0.2 1.1 − 0.37 crease in inhibitory tone may be seen as a reflection of a primary py­
Antipsychotic exposure days at – 3.07 ± – – ramidal dysfunction in schizophrenia (as proposed by Adams et al.,
time of scan 10.9 2022) or an increase interneuron-related GABAergic activity (Pinotsis
Duration of untreated – 9.5 ± – – et al., 2012). Based on the findings from N-methyl-D-aspartate receptor
psychosis 13.2
antagonist (ketamine), autoantibody models, postmortem studies and

100
M.F. Alonso-Sánchez et al. Schizophrenia Research 259 (2023) 97–103

Fig. 3. Top left: Model selection. G: Model group; G + S: Model group and semantic similarity; G + S + G * S: Model group, semantic similarity and Interaction
between group and semantic similarity. Top right: Second level effect of Semantic similarity on intrinsic inhibitory and extrinsic excitatory connections. Posterior
parameter estimates with bars represent the estimated effect size and the lines represent 90% credible interval. Bottom: Brain network with the connectivity strength.

knockout mouse models, Jeevakumar and Kroener (2016) and Naka­ studies have examined the association between the anterior temporal
zawa et al. (2017) proposed that NMDA hypofunction initially occurs in lobe and schizophrenia features of speech. As NLP measures become
cortical GABAergic neurons during early postnatal development, more commonly available, they have the potential to replace subjective
reducing intrinsic excitability and GABA release, leading to a disinhi­ rating scales as instruments of psychopathological quantification. In this
bition of pyramidal neurons and subsequent homeostatic down regula­ context, eliciting the neural basis of key NLP measures in schizophrenia
tion of pyramidal excitability. is of critical importance. While a plethora of derived measures are
Reduced synaptic gain (reflected in increased inhibitory tone), may available, we chose a lexical measure based on world choices during a
relate to reduced precision of locally generated neural activity, as well as discourse, given the wealth of information available on semantic ab­
a reduction in the generation of spontaneous neural oscillations at rest. normalities in schizophrenia (Pomarol-Clotet et al., 2008), This
One physiological consequence of higher intrinsic inhibitory tone is a approach provides a precedence for examining the neural basis of syn­
reduction in the neuronal disinhibition or ‘release’ required for excit­ tactic abnormalities (e.g. connective use (Mackinley et al., 2021)) and
atory activity. In terms of neuronal recruitment in service of a cognitive linguistic style (e.g. analytic thinking (Silva et al., 2021)) that are shown
task, this may make a brain region more dependent on extrinsically to be relevant in schizophrenia. However, such studies do not address
driven stimulatory processes. If this increased regional inhibitory tone the inhibitory or excitatory connectivity of the vATL so it is challenging
continues during speech production, for our 2-node language network to compare our results with previous reports.
model, this may mean that aberrant word sampling cannot be overcome Our study has several strengths as well as some limitations. To our
in the presence of a physiologically normal level of top-down stimula­ knowledge, this is the first study exploring the effective connectivity of
tion (from IFG) arriving at the vATL. IFG and vATL in patients with first episode of schizophrenia in relation
The strength of the connection between the vATL and IFG was not to lexical efficiency. We recruited an extremely early-stage sample who
affected by diagnostic status or semantic similarity in our sample. It is had a mean of 3 days of lifetime antipsychotic exposure and were still in
possible that such an effect on extrinsic connections may be present the untreated state when speech and fMRI data were obtained. We also
when speaking, but not at the resting state when fMRI scans were ob­ used unambiguous picture stimuli which are shown to be more reliable
tained. In fact, dynamic causal influence between left IFG and temporal in NLP studies of schizophrenia (Morgan et al., 2021). Given the head
cortex has been linked to the ‘priming’ activity in a previous study motion challenges when people are speaking in the scanner, we used
(Matsumoto and Kakigi, 2013). Experimental verification of this possi­ resting state fMRI data to make inferences about speech outside the
bility requires a speech production task in the scanner wherein demands scanner. This has conceptual limitations and may explain why Broca-
placed on the control system from IFG to vATL (thus, extrinsic signals to vATL connectivity was not related to the semantic similarity measure
and from the semantic hub), can be studied directly (Jackson, 2021). as expected. The relationship between task and resting-state effective
Our results complement previous observations of aberrant Broca's connectivity (as inferred using DCM) is less well understood. Jung and
area connectivity in first-episode schizophrenia (Du et al., 2021; Li et al., colleagues demonstrated a strong correlation in the effective connec­
2017). Interestingly, while previous studies have focused on the IFG, few tivity among regions engaged in a n-back task between the rest and the

101
M.F. Alonso-Sánchez et al. Schizophrenia Research 259 (2023) 97–103

task states and concluded that the task-dependent connectivity endows Appendix A. Supplementary data
resting- state connectivity with a context sensitivity that predicts task
performance (Jung et al., 2018). Applying these results to our findings, Supplementary data to this article can be found online at https://doi.
we can expect that in the context of online speech production, the BA45- org/10.1016/j.schres.2022.04.007.
vATL connectivity pattern may exhibit ‘context sensitive’ changes in line
with their resting-state connectivity patterns. More specifically, we can References
expect the intrinsic inhibitory tone of BA45 and vATL to be modulated
by speech production when in-scanner speech is generated. We caution Adams, R.A., Pinotsis, D., Tsirlis, K., Unruh, L., Mahajan, A., Horas, A.M., Convertino, L.,
Summerfelt, A., Sampath, H., Du, X.M., Kochunov, P., Ji, J.L., Repovs, G., Murray, J.
the readers that this is a speculative interpretation; a conclusive answer D., Friston, K.J., Hong, L.E., Anticevic, A., 2022. Computational modeling of
requires empirical data to test within-subject rest-task concordance. electroencephalography and functional magnetic resonance imaging paradigms
Finally, we had access to limited longitudinal imaging data from this indicates a consistent loss of pyramidal cell synaptic gain in schizophrenia. Biol.
Psychiatry 91 (2), 202–215. https://doi.org/10.1016/j.biopsych.2021.07.024.
sample, precluding an examination of changes in brain connectivity over Almeida, V.N., Radanovic, M., 2021. Semantic priming and neurobiology in
time to complement our recent report of worsening semantic similarity schizophrenia: a theoretical review. Neuropsychologia, 108058. https://doi.org/
in FES (Alonso-Sánchez et al., 2022). 10.1016/j.neuropsychologia.2021.108058.
Alonso-Sánchez, M.F., Ford, S.D., MacKinley, M., Silva, A.M., Limongi, R.,
Palaniyappan, L., 2022. Progressive changes in descriptive discourse in first episode
5. Conclusion schizophrenia: a longitudinal computational semantics study. Schizophrenia 8 (1),
1–9. https://doi.org/10.1038/s41537-022-00246-8.
Arcuri, S.M., Broome, M.R., Giampietro, V., Jr, E.A., Kircher, T.T.J., Williams, S.C.R.,
In conclusion, we report an aberrant inhibitory connection strength
Andrew, C.M., Brammer, M., Morris, R.G., Mcguire, P.K., 2012. Faulty Suppression
in the IFG and the vATL in relation to inefficiency of semantic selection of Irrelevant Material in Patients With Thought Disorder Linked to Attenuated
in schizophrenia. Our results suggest that NLP-derived semantic anom­ Frontotemporal Activation, 2012. https://doi.org/10.1155/2012/176290.
alies in schizophrenia have a likely neurophysiological basis, and lexical Binder, J.R., Desai, R.H., Graves, W.W., Conant, L.L., 2009. Where is the semantic
system? A critical review and meta-analysis of 120 functional neuroimaging studies.
selection may be linked to reduced regional synaptic gain within the left Cereb. Cortex 19 (12), 2767–2796. https://doi.org/10.1093/cercor/bhp055.
IFG and the vATL. Boyd, M., 2008. A socioeconomic scale for Canada: measuring occupational status from
the census. Can. Rev. Sociol. 45 (1), 51–91. https://doi.org/10.1111/j.1755-
618X.2008.00003.x.
CRediT authorship contribution statement Brown, M., Kuperberg, G.R., 2015. In: Framework of Language Processing : Linking
Language Perception , Interpretation, and Production Abnormalities in
MFA and LP conceptualized the project, JG supervised the fMRI Schizophrenia, 9, pp. 1–23. https://doi.org/10.3389/fnhum.2015.00643
(November).
project, MF analyzed the data, MF and LP wrote the manuscript, RL and Chen, P.J., Fan, L.Y., Hwang, T.J., Hwu, H.G., Liu, C.M., Chou, T.L., 2013. The deficits on
all authors critically reviewed and approved the final version of the a cortical-subcortical loop of meaning processing in schizophrenia. NeuroReport 24
manuscript. (3), 147–151. https://doi.org/10.1097/WNR.0b013e32835df562.
Chiou, R., Humphreys, G.F., Jung, J.Y., Lambon Ralph, M.A., 2018. Controlled semantic
cognition relies upon dynamic and flexible interactions between the executive
Declaration of competing interest ‘semantic control’ and hub-and-spoke ‘semantic representation’ systems. Cortex 103,
100–116. https://doi.org/10.1016/j.cortex.2018.02.018.
Chiou, R., Lambon Ralph, M.A., 2019. Unveiling the dynamic interplay between the hub-
LP reports personal fees from Otsuka Canada, SPMM Course Limited,
and spoke-components of the brain’s semantic system and its impact on human
UK, Canadian Psychiatric Association; book royalties from Oxford Uni­ behaviour. NeuroImage 199 (April), 114–126. https://doi.org/10.1016/j.
versity Press; investigator-initiated educational grants from Janssen neuroimage.2019.05.059.
Conner, C.R., Kadipasaoglu, C.M., Shouval, H.Z., Hickok, G., Tandon, N., 2019. Network
Canada, Sunovion and Otsuka Canada outside the submitted work. All
dynamics of Broca’s area during word selection. PLoS ONE 14 (12), 1–30. https://
other authors report no relevant conflicts. doi.org/10.1371/journal.pone.0225756.
Covington, M.A., 2016. Covington Vector Semantics Tools. Covington Innovations.
Acknowledgements Dey, A., Dempster, K., MacKinley, M., Jeon, P., Das, T., Khan, A., Gati, J.,
Palaniyappan, L., 2021. Conceptual disorganization and redistribution of resting-
state cortical hubs in untreated first-episode psychosis: a 7T study. NPJ Schizophr. 7
We appreciate all the participants and their families for the time and (1), 1–9. https://doi.org/10.1038/s41537-020-00130-3.
effort to contribute to this study. We thank Michael Covington (Cov­ Du, J., Palaniyappan, L., Liu, Z., Cheng, W., Gong, W., Zhu, M., Wang, J., Zhang, J.,
Feng, J., 2021. The genetic determinants of language network dysconnectivity in
ington Innovations) for providing us his CoVec NLP tool. We thank all drug-naïve early stage schizophrenia. NPJ Schizophr. 7 (1) https://doi.org/10.1038/
research team members of the NIMI lab and all the staff members of the s41537-021-00141-8.
PEPP London team, particularly Drs. Kara Dempster (currently at Dal­ Elvevåg, B., Foltz, P.W., Weinberger, D.R., Goldberg, T.E., 2007. Quantifying
incoherence in speech: an automated methodology and novel application to
housie University), Julie Richard, Priya Subramanian and Hooman schizophrenia. Schizophr. Res. 93 (1–3), 304–316. https://doi.org/10.1016/j.
Ganjavi and (late) Dr. Raj Harricharan for their assistance in patient schres.2007.03.001.
recruitment and supporting clinical care. We also appreciate the assis­ Feinberg, D.A., Moeller, S., Smith, S.M., Auerbach, E., Ramanna, S., Glasser, M.F.,
Miller, K.L., Ugurbil, K., Yacoub, E., 2010. Multiplexed echo planar imaging for sub-
tance offered by Peter Jeon and Michael Mackinley for data acquisition, second whole brain fmri and fast diffusion imaging. PLoS ONE 5 (12). https://doi.
Dr. Ali Khan for the fMRI protocol, Hannah Ke, Jenny Chan and Sabrina org/10.1371/journal.pone.0015710.
Ford for data curation. Friston, K.J., 2011. Functional and effective connectivity: a review. Brain Connect. 1 (1),
13–36. https://doi.org/10.1089/brain.2011.0008.
Friston, K.J., Harrison, L., Penny, W., 2003. Dynamic causal modelling. NeuroImage 19
Funding (4), 1273–1302. https://doi.org/10.1016/S1053-8119(03)00202-7.
Friston, K.J., Kahan, J., Biswal, B., Razi, A., 2014. NeuroImage a DCM for resting state
fMRI. NeuroImage 94, 396–407. https://doi.org/10.1016/j.
This study was funded by The Canadian Institutes of Health Research
neuroimage.2013.12.009.
(CIHR) Foundation Grant (375104/2017) and the Canada First Research Heim, S., Eickhoff, S.B., Amunts, K., 2008. Specialisation in Broca’s region for semantic,
Excellence Fund (CFREF) to BrainsCAN. This work was also supported phonological, and syntactic fluency? NeuroImage 40 (3), 1362–1368. https://doi.
by the National Agency for Research and Development (ANID), Schol­ org/10.1016/j.neuroimage.2008.01.009.
Hoffman, P., Cogdell-Brooke, L., Thompson, H.E., 2020. Going off the rails: impaired
arship Program, Becas Chile 2019, Postdoctoral Fellow 74200048 coherence in the speech of patients with semantic control deficits. Neuropsychologia
(MFA). We also acknowledge support from the Bucke Family Fund, The 146 (June), 107516. https://doi.org/10.1016/j.neuropsychologia.2020.107516.
Chrysalis Foundation and The Arcangelo Rea Family Foundation (Lon­ Horn, H., Federspiel, A., 2012. In: Semantic Network Disconnection in Formal Thought
Disorder, pp. 14–23. https://doi.org/10.1159/000337133.
don, Ontario). LP acknowledges personal chair support from the Tanna Jackson, R.L., 2021. The neural correlates of semantic control revisited. NeuroImage
Schulich Endowment (Schulich School of Medicine, Western University) 224, 117444. https://doi.org/10.1016/j.neuroimage.2020.117444 (October 2020).
and Monique H. Bourgeois Endowment (Douglas Research Centre,
McGill University).

102
M.F. Alonso-Sánchez et al. Schizophrenia Research 259 (2023) 97–103

Jackson, R.L., Cloutman, L.L., Lambon, M.A., 2019. ScienceDirect exploring distinct Oldfield, R.C., 1971. The assessment and analysis of handedness: the Edinburgh
default mode and semantic networks using a systematic ICA approach. Cortex 113, inventory. Neuropsychologia 9 (1), 97–113. https://doi.org/10.1016/0028-3932
279–297. https://doi.org/10.1016/j.cortex.2018.12.019. (71)90067-4.
Jackson, R.L., Hoffman, P., Pobric, X.G., Ralph, X.M.A.L., 2016. In: The Semantic Palaniyappan, L., Mota, N.B., Oowise, S., Balain, V., Copelli, M., Ribeiro, S., Liddle, P.F.,
Network at Work and Rest : Differential Connectivity of Anterior Temporal Lobe 2019. Speech structure links the neural and socio-behavioural correlates of psychotic
Subregions, 36(5), pp. 1490–1501. https://doi.org/10.1523/JNEUROSCI.2999- disorders. Prog. Neuro-Psychopharmacol. Biol. Psychiatry 88, 112–120. https://doi.
15.2016. org/10.1016/j.pnpbp.2018.07.007 (April 2018).
Jarosz, A., Wiley, J., 2014. What are the odds? A practical guide to computing and Pennington, J., Socher, R., Manning, C.D., 2014. GloVe: global vectors for word
reporting Bayes factors. J. Probl. Solving 7, 2–9. https://doi.org/10.7771/1932- representation. In: Conference: Proceedings of the 2014 Conference on Empirical
6246.1167. Methods in Natural Language Processing (EMNLP), pp. 1532–1543. https://doi.org/
JASP Team, 2020. JASP (version 0.14.1). University of Amsterdam. https://jasp-stats. 10.3115/v1/D14-1162.
org/. Pinotsis, D.A., Moran, R.J., Friston, K.J., 2012. Dynamic causal modeling with neural
Jeevakumar, V., Kroener, S., 2016. Ketamine administration during the second postnatal fields. Neuroimage 59 (2), 1261–1274. https://doi.org/10.1016/j.
week alters synaptic properties of fast-spiking interneurons in the medial prefrontal neuroimage.2011.08.020.
cortex of adult mice. Cereb. Cortex 26 (3), 1117–1129. https://doi.org/10.1093/ Pomarol-Clotet, E., Oh, T.M.S.S., Laws, K.R., McKenna, P.J., 2008. Semantic priming in
cercor/bhu293. schizophrenia: systematic review and meta-analysis. Br. J. Psychiatry 192 (2),
Jung, K., Friston, K.J., Pae, C., Choi, H.H., Tak, S., Choi, Y.K., Park, B., Park, C.A., 92–97. https://doi.org/10.1192/bjp.bp.106.032102.
Cheong, C., Park, H.J., 2018. Effective connectivity during working memory and Python, G., Glize, B., Laganaro, M., 2018. The involvement of left inferior frontal and
resting states: a DCM study. NeuroImage 169, 485–495. https://doi.org/10.1016/j. middle temporal cortices in word production unveiled by greater facilitation effects
neuroimage.2017.12.067 (December 2017). following brain damage. Neuropsychologia 121 (October), 122–134. https://doi.
Kanerva, P., 2009. Hyperdimensional computing: an introduction to computing in org/10.1016/j.neuropsychologia.2018.10.026.
distributed representation with high-dimensional random vectors. Cogn. Comput. 1 Rezaii, N., Walker, E., Wolff, P., 2019. OPEN A machine learning approach to predicting
(2), 139–159. https://doi.org/10.1007/s12559-009-9009-8. psychosis using semantic density and latent content analysis. NPJ Schizophr.
Kay, S.R., Fiszbein, A., Opler, L.A., 1987. The positive and negative syndrome scale https://doi.org/10.1038/s41537-019-0077-9. December 2018.
(PANSS) for schizophrenia. Schizophr. Bull. 13 (2), 261–276. https://doi.org/ Rice, G.E., Hoffman, P., Lambon Ralph, M.A., 2015. Graded specialization within and
10.1093/schbul/13.2.261. between the anterior temporal lobes. Ann. N. Y. Acad. Sci. 1359 (1), 84–97. https://
Kluyver, T., Ragan-Kelley, B., Pérez, F., Granger, B., Bussonnier, M., Frederic, J., doi.org/10.1111/nyas.12951.
Kelley, K., Hamrick, J., Grout, J., Corlay, S., Ivanov, P., Avila, D., Abdalla, S., Rybarczyk, B., 2011. Encyclopedia of clinical neuropsychology. In: Kreutzer, J.,
Willing, C., Team, J.development, 2016. Jupyter notebooks – a publishing format for DeLuca, J., Caplan, B. (Eds.), Encyclopedia of Clinical Neuropsychology. Springer.
reproducible computational workflows. In: Loizides, F., Scmidt, B. (Eds.), Positioning https://doi.org/10.1007/978-0-387-79948-3.
and Power in Academic Publishing: Players, Agents and Agendas. IOS Press, Schmidt, A., Palaniyappan, L., Smieskova, R., Simon, A., Riecher-Rössler, A., Lang, U.E.,
pp. 87–90. https://doi.org/10.3233/978-1-61499-649-1-87. Fusar-Poli, P., McGuire, P., Borgwardt, S.J., 2016. Dysfunctional insular connectivity
Krieger-Redwood, K., Jefferies, E., 2014. TMS interferes with lexical-semantic retrieval in during reward prediction in patients with first-episode psychosis. J. Psychiatry
left inferior frontal gyrus and posterior middle temporal gyrus: evidence from Neurosci. 41 (6), 367–376. https://doi.org/10.1503/jpn.150234.
cyclical picture naming. Neuropsychologia 64, 24–32. https://doi.org/10.1016/j. Silva, A., Limongi, R., MacKinley, M., Palaniyappan, L., 2021. Small words that matter:
neuropsychologia.2014.09.014. linguistic style and conceptual disorganization in untreated first-episode
Levelt, W.J.M., 2001. Spoken word production: a theory of lexical access. Proc. Natl. schizophrenia. Schizophr. Bull. Open 2 (1), 1–10. https://doi.org/10.1093/
Acad. Sci. U. S. A. 98 (23), 13464–13471. https://doi.org/10.1073/ schizbullopen/sgab010.
pnas.231459498. Snyder, H.R., Banich, M.T., Munakata, Y., 2011. Choosing our words: retrieval and
Levy, O., Goldberg, Y., Dagan, I., 2015. Improving distributional similarity with lessons selection processes recruit shared neural substrates in left ventrolateral prefrontal
learned from word embeddings. Trans. Assoc. Comput. Linguistics 3, 211–225. cortex. J. Cogn. Neurosci. 23 (11), 3470–3482. https://doi.org/10.1162/jocn_a_
https://doi.org/10.1162/tacl_a_00134. 00023.
Li, Y., Li, P., Yang, Q.X., Eslinger, P.J., Sica, C.T., Karunanayaka, P., 2017. Lexical- Spencer, T.J., Thompson, B., Oliver, D., Diederen, K., Demjaha, A., Weinstein, S.,
semantic search under different covert verbal fluency tasks: an fMRI study. Front. Morgan, S.E., Day, F., Valmaggia, L., Rutigliano, G., de Micheli, A., Mota, N.B.,
Behav. Neurosci. 11 (August), 1–15. https://doi.org/10.3389/fnbeh.2017.00131. Fusar-Poli, P., McGuire, P., 2021. Lower speech connectedness linked to incidence of
Limongi, R., Jeon, P., Mackinley, M., Das, T., Dempster, K., Théberge, J., Bartha, R., psychosis in people at clinical high risk. Schizophr. Res. 228, 493–501. https://doi.
Wong, D., Palaniyappan, L., 2020. Glutamate and dysconnection in the salience org/10.1016/j.schres.2020.09.002.
network: neurochemical, effective connectivity, and computational evidence in Surbeck, W., Hänggi, J., Scholtes, F., Viher, P.V., Schmidt, A., Stegmayer, K.,
schizophrenia. Biol. Psychiatry 88 (3), 273–281. https://doi.org/10.1016/j. Studerus, E., Lang, U.E., Riecher-Rössler, A., Strik, W., Seifritz, E., Borgwardt, S.,
biopsych.2020.01.021. Quednow, B.B., Walther, S., 2020. Anatomical integrity within the inferior fronto-
Mackinley, M., Chan, J., Ke, H., Dempster, K., Palaniyappan, L., 2021. Linguistic occipital fasciculus and semantic processing deficits in schizophrenia spectrum
determinants of formal thought disorder in first episode psychosis. Early Interv. disorders. Schizophr. Res. 218, 267–275. https://doi.org/10.1016/j.
Psychiatry 15 (2), 344–351. https://doi.org/10.1111/eip.12948. schres.2019.12.025.
Matsumoto, A., Kakigi, R., 2013. Subliminal semantic priming changes the dynamic Visser, M., Jefferies, E., Lambon Ralph, M.A., 2010. Semantic processing in the anterior
causal influence between the left frontal and temporal cortex. J. Cogn. Neurosci. 26, temporal lobes: a meta-analysis of the functional neuroimaging literature. J. Cogn.
165–174. https://doi.org/10.1162/jocn_a_00472. Neurosci. 22 (6), 1083–1094. https://doi.org/10.1162/jocn.2009.21309.
Mikolov, T., Sutskever, I., Chen, K., Corrado, G., Dean, J., 2013. Distributed Wagner, A., Paré-Blagoev, J., Clark, J., Poldrack, R., 2001. Recovering meaning: left
representations of words and phrases and their compositionality. Adv. Neural Inf. prefrontal cortex guides controlled semantic retrieval. Neuron 31, 329–338. https://
Proces. Syst. 1389–1399. https://doi.org/10.18653/v1/d16-1146. doi.org/10.1016/S0026-0495(96)90074-8.
Moeller, S., Yacoub, E., Olman, C.A., Auerbach, E., Strupp, J., Harel, N., Uğurbil, K., Whitney, C., Grossman, M., Kircher, T.T.J., 2009. The influence of multiple primes on
2010. Multiband multislice GE-EPI at 7 Tesla, with 16-fold acceleration using partial bottom-up and top-down regulation during meaning retrieval: evidence for 2 distinct
parallel imaging with application to high spatial and temporal whole-brain FMRI. neural networks. Cereb. Cortex 19 (11), 2548–2560. https://doi.org/10.1093/
Magn. Reson. Med. 63 (5), 1144–1153. https://doi.org/10.1002/mrm.22361. cercor/bhp007.
Morgan, S.E., Diederen, K., Vértes, P.E., Ip, S.H.Y., Wang, B., Thompson, B., Demjaha, A., Xia, M., Wang, J., He, Y., 2013. BrainNet viewer: a network visualization tool for human
Micheli, A.de, 2021. Assessing psychosis risk using quantitative markers of brain connectomics. PLoS ONE 8 (7). https://doi.org/10.1371/journal.
disorganised speech. NPJ Schizophr. https://doi.org/10.1101/ pone.0068910.
2021.01.04.20248717. December 2020. Zeidman, P., Jafarian, A., Corbin, N., Seghier, M.L., Razi, A., Price, C.J., Friston, K.J.,
Mota, N.B., Copelli, M., Ribeiro, S., 2017. Thought disorder measured as random speech 2019. A guide to group effective connectivity analysis, part 1: first level analysis with
structure classifies negative symptoms and schizophrenia diagnosis 6 months in DCM for fMRI. NeuroImage 200 (May), 174–190. https://doi.org/10.1016/j.
advance. NPJ Schizophr. 3 (1), 1–10. https://doi.org/10.1038/s41537-017-0019-3. neuroimage.2019.06.031.
Murray, H., 1943. Thematic Apperception Test. Harvard University Press. Zeidman, P., Jafarian, A., Seghier, M.L., Litvak, V., Cagnan, H., Price, C.J., Friston, K.J.,
Nakazawa, K., Jeevakumar, V., Nakao, K., 2017. Spatial and temporal boundaries of 2019. A guide to group effective connectivity analysis, part 2: second level analysis
NMDA receptor hypofunction leading to schizophrenia. NPJ Schizophr. 3 (1), 0–1. with PEB. NeuroImage 200 (May), 12–25. https://doi.org/10.1016/j.
https://doi.org/10.1038/s41537-016-0003-3. neuroimage.2019.06.032.

103