bioRxiv preprint doi: https://doi.org/10.1101/2022.10.15.512349; this version posted October 17, 2022.

The copyright holder for this

preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.

A common neural code for meaning in discourse production and

comprehension

Tanvi PATEL
Matías MORALES
Martin J. PICKERING
Paul HOFFMAN*

School of Philosophy, Psychology & Language Sciences, University of Edinburgh, UK

* Correspondence to:
Dr. Paul Hoffman
School of Philosophy, Psychology & Language Sciences,
University of Edinburgh,
7 George Square, Edinburgh, EH8 9JZ, UK
Tel: +44 (0) 131 650 4654
Email: [email protected]

Acknowledgements
PH was supported by a BBSRC grant (BB/T004444/1). Imaging was carried out at the
Edinburgh Imaging Facility (www.ed.ac.uk/edinburgh-imaging), University of Edinburgh, which
is part of the SINAPSE collaboration (www.sinapse.ac.uk). For the purpose of open access,
the authors have applied a Creative Commons Attribution (CC BY) licence to any Author
Accepted Manuscript version arising from this submission.

1
bioRxiv preprint doi: https://doi.org/10.1101/2022.10.15.512349; this version posted October 17, 2022. The copyright holder for this
preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.

Abstract

How does the brain code the meanings conveyed by language? Neuroimaging studies have
investigated this by linking neural activity patterns during discourse comprehension to
semantic models of language content. Here, we applied this approach to the production of
discourse for the first time. Participants underwent fMRI while producing and listening to
discourse on a range of topics. We quantified the semantic similarity of different speech
passages and identified where similarity in neural activity was predicted by semantic similarity.
A widely distributed and bilateral network, similar to that found for comprehension, showed
such effects when participants produced their own discourse. Critically, cross-task neural
similarities between comprehension and production were also predicted by similarities in
semantic content. These results indicate that discourse semantics engages a common neural
code during both comprehension and production. Furthermore, common coding in right-
hemisphere regions challenges the idea that language production processes are strongly left-
lateralised.

2
bioRxiv preprint doi: https://doi.org/10.1101/2022.10.15.512349; this version posted October 17, 2022. The copyright holder for this
preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.

Introduction

Both the production and comprehension of language are rooted in our conceptual
knowledge of the world. At the neural level, this semantic processing is carried out through
the activation of a widely distributed network of systems that store and retrieve knowledge
[1-3]. Traditionally, work on semantic representation has linked broad semantic categories
(e.g., animals vs. tools) to regional engagement of semantic processing areas [4, 5]. But more
recently, the development of sophisticated multivariate pattern analysis (MVPA) techniques
has allowed researchers to examine how conceptual knowledge is encoded across distributed
patterns of neural activation [6, 7]. These methods have moved the field from a coarse
category-based approach to semantic representation into a fine-grained multidimensional
understanding of the neural coding of meaning. They have also expanded the space of
investigation from individual words and objects into the rich array of concepts and situations
that occur in natural discourse. In this study we use Representational Similarity Analysis
(RSA), an MVPA technique that tests how similarity in the properties of stimuli can predict
similarity in the neural activation patterns they elicit [7, 8]. Previous studies have used RSA
at the single-word level to investigate how semantic content is coded in the brain [9-12]. The
present study applies this approach to the discourse level and asks a critical but under-
investigated question: is the coding of semantic content similar when people produce their
own narratives, compared to when they listen to another person’s speech? In other words,
does meaning in language comprehension and production rely on a common neural code?
MVPA has been used to investigate questions of semantic representation at different
levels of stimulus complexity. Studies using stimuli at the single word [e.g., 10, 11, 13] and
sentence level [e.g., 14, 15, 16] have demonstrated that patterns of neural activation can be
predicted from vector-based semantic models. These include models that code conceptual
relationships based on associated sensory-motor experiences [17, 18] and models of natural
language processing, which derive semantic representations from lexical statistics in large text
corpora [19-21]. Beyond the word and sentence levels, studies have used voxel-wise [22, 23]
and multivariate approaches to examine language comprehension at the discourse level. Such

3
bioRxiv preprint doi: https://doi.org/10.1101/2022.10.15.512349; this version posted October 17, 2022. The copyright holder for this
preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.

studies use more naturalistic stimuli, such as written stories [24, 25] spoken narratives [22,
26, 27] or movies [28]. These studies have revealed a widely distributed set of frontal,
temporal, and parietal areas in which activation patterns can be predicted by the semantic
content of language.
The critical question in our study is the degree to which these neural codes are similar
across different language tasks – specifically, comprehending versus producing speech. MVPA
studies have shown that the neural representations of concepts are robust to changes in the
stimuli used to evoke them. In other words, concepts elicit similar patterns of activation in
the brain’s semantic network, irrespective of the stimuli used to evoke them. This cross-
stimulus generalisation has been observed for object names vs. object pictures [11, 29]; for
written vs. auditory words [30]; for stories vs. movie stimuli [28]; and for event descriptions
that vary in lexical and syntactic content [31]. This is true even across languages – neural
representations of concepts show cross-linguistic invariance at the individual word [32] and
story-level [25]. These findings are consistent with contemporary accounts of semantic
cognition, which hold that higher-order conceptual knowledge is supported by “hub” regions
that permit generalisation of knowledge across contexts, exemplars and tasks [1, 2, 33, 34].
In the present study, we test whether the same kind of generalisation occurs when
comparing activation evoked during different language tasks. Theories of language production
and comprehension suggest that these are interwoven processes that recruit overlapping
neural networks [35-38]. At the level of discourse, both comprehension and production are
thought to involve the construction of situation models which encode the entities and events
described [35, 39]. In addition, one particular view postulates that comprehension uses a
prediction-by-production mechanism, whereby the listener covertly imitates the linguistic
form of the speaker’s utterances in order to derive their intentions, and then runs this
through their production system to predict the speaker’s upcoming utterances [38, 39].
However, there is currently limited evidence that the neural representations engaged when
people produce discourse are similar to those elicited when they listen to it. This is in part
because, while MVPA studies of discourse comprehension are now commonplace, analogous
studies of discourse production are rare.

4
bioRxiv preprint doi: https://doi.org/10.1101/2022.10.15.512349; this version posted October 17, 2022. The copyright holder for this
preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.

Univariate neuroimaging studies suggest some differences in the neural networks

activated during comprehension and production. While production and comprehension of
speech activate common language areas in the left hemisphere, narrative comprehension
tends to activate right frontal and temporal regions more than production [40-42]. These
findings align with theoretical views that while speech comprehension engages a bilateral
network, language production is more left-lateralised [43-47]. More recently, a series of
studies have investigated correlations in neural activity between a speaker producing a story
and a listener comprehending it [42, 48-50]. These indicate a more bilateral picture, with
significant production-comprehension temporal coupling in left and right prefrontal, temporal
and inferior parietal cortices. This inter-subject coupling suggests a large degree of shared
neural processing across comprehension and production. However, these studies have
investigated correlations between different individuals and not within the same individual
performing different language tasks. In addition, inter-subject correlations in neural activity
could be due to shared processes at any level of the language system and do not indicate
where commonalities are due specifically to its semantic content.
Here, we used RSA to investigate the degree to which neural patterns during speech
production and comprehension align with a single model of semantic content. In the scanner,
participants listened to pre-recorded speech samples, and produced extended passages of
speech on a variety of topics. We used a distributional semantic model [latent semantic
analysis; 21] to quantify the degree to which different speech chunks contained similar
semantic content. We then used RSA to assess how well semantic similarity could predict
similarity in the neural responses evoked during different periods of heard or spoken language.
Critically, as well as performing these analyses separately for comprehension and production
data, we conducted a cross-task analysis that looked at similarities between comprehension
and production periods (within participants). In this way, we were able to identify brain
regions which share a common neural code for semantic content during speaking and listening.
We used a searchlight approach to seek semantic-coding regions across the whole
brain. In addition, we investigated effects in a set of regions of interest known to be involved
in semantic processing. These comprised parts of the anterior temporal lobes, which act as a
hub for representation of semantic knowledge [2, 33], inferior prefrontal and posterior

5
bioRxiv preprint doi: https://doi.org/10.1101/2022.10.15.512349; this version posted October 17, 2022. The copyright holder for this
preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.

temporal regions that regulate access to this knowledge [51, 52] and the angular gyrus, a key
node in the default mode network which is involved in constructing mental models of events
and ongoing experiences [1, 53]. By investigating effects in the left and right hemisphere
homologues of these regions, we were also able to probe the extent of lateralization during
production and comprehension of speech.

Results

We performed a series of analyses on fMRI data in which participants produced and listened
to speech samples on a range of topics (see Figure 1). After reporting basic characteristics of
the language samples, we report the results of univariate analyses which investigated the
degree to which brain regions were engaged during discourse comprehension and production,
compared with a baseline of automatic speech. Following this, we use RSA to identify regions
in which similarities in activation patterns are predicted by similarities in the semantic content
of language. As well as performing these analyses separately on listening and speaking data,
we perform a cross-task analysis that tests for semantic neural coding that generalises across
language tasks. Finally, we directly compare neural similarity patterns across semantic-related
brain regions to investigate the degree to which different regions are similarly influenced by
discourse content.

Figure 1. (A) Structure of trials in the experiment. On each discourse trial, participants were
presented with a topic prompt for 6s and were then required to either speak about this topic for
50s or listen to a recording of another person speaking about it for 50s. In the baseline condition,
participants either recited or listened to a well-known British nursery rhyme. (B) Example semantic
and quantity dissimilarity matrices. Topic labels indicate different discourse topics, each of which
was divided into five 10s segments. The semantic DSM codes how similar speech segments are in
lexical-semantic content. The quantity DSM (used as a control) codes how similar segments are
terms of the number of words they contain. (C) Procedure for main RSA analyses. The analysis tested
the relationship between a neural DSM, generated by comparing local activation patterns for
different speech segments, and the semantic DSM, while controlling for the effects of the quantity
DSM. (D) Procedure for comparison of neural DSMs. This analysis test the relationship between
two neural DSMs obtained from different brain regions. Speak1 = speech production run 1; Speak2
= speech production run 2; Listen1 = comprehension run 1; Listen2 = comprehension run 2. DSM
= dis-similiarity matrix.

6
bioRxiv preprint doi: https://doi.org/10.1101/2022.10.15.512349; this version posted October 17, 2022. The copyright holder for this
preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.

7
bioRxiv preprint doi: https://doi.org/10.1101/2022.10.15.512349; this version posted October 17, 2022. The copyright holder for this
preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.

Characteristics of speech: In the discourse production task, participants spoke about a series
of topics for 50s at a time (see Method for details). They produced a mean of 129 words per
topic (SD = 37, range = 67-320). In the discourse comprehension task, participants listened
to recordings of another person speaking about a different set of topics. These recordings
contained a mean of 141 words per topic (SD = 20, range = 102-170). Example speech samples
from comprehension and production tasks are provided in Supplementary Materials .After
scanning, participants received 12 comprehension questions relating to the speech presented
in the listening task. They answered 10/12 questions correctly on average. They also provided
audibility rating with a mean rating of 5.5/7, suggesting that they were able to understand the
discourse samples presented in the scanner.

Univariate analyses: Whole-brain activation maps are shown in Figure 2. These show effects
for each discourse task relative to a “low-level” speech baseline (reciting/listening to a familiar
nursery rhyme). The baseline conditions involved language processing and so had similar
perceptual and motor demands to the discourse tasks, but they did not require participants
to process novel, meaningful verbal information. In line with previous studies, the results
indicate close correspondence in the areas recruited for production and comprehension of
novel discourse, particularly in the left hemisphere. Both tasks activated similar left-lateralised
networks associated with semantic processing, including lateral and medial prefrontal cortex,
lateral temporal cortex, the ventral anterior temporal lobe and the angular gyrus. Listening
produced significantly greater activation than speaking in right prefrontal regions, left superior
temporal sulcus and bilateral post-central gyrus. Speech production was associated with
greater activity in the cerebellum, medial prefrontal cortex and the occipital lobe. The overall
picture, however, was that participants engaged broadly similar networks whether they were
speaking or listening.
We also assessed activation in five regions of interest (ROIs) that are frequently
implicated in semantic processing, and their right-hemisphere homologues (see Figure 3A).
Results are shown in Figure 3B and reveal a left-lateralised pattern for both tasks. 2 x 5 x 2
ANOVA showed a main effect of hemisphere (F(1,24) = 64.4, p < .001; left > right) as well as
ROI (F(4,96) = 20.7, p < .001), and task (F(1,24) = 5.7, p = .026; listening > speaking). All of

8
bioRxiv preprint doi: https://doi.org/10.1101/2022.10.15.512349; this version posted October 17, 2022. The copyright holder for this
preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.

the interactions between these factors were also significant (F > 6.9, p < .001). FDR-corrected
post-hoc tests indicated that activation was left-lateralised for both tasks in most regions (see
Figure 3B). In the IFG, pMTG and AG, this effect was larger in the speaking task, with right-
hemisphere regions deactivating during production, relative to the baseline speech condition.
Similar results were obtained when discourse was contrasted against rest rather than baseline
speech conditions (see Supplementary Figure 1). Thus, univariate analyses suggest that the
left hemisphere is dominant in semantic aspects of discourse processing, particularly when
participants produced, rather than heard, discourse.

Figure 2: Univariate activation contrasts. Areas activated by each discourse task compared with its
baseline, and for the direct contrast of speaking vs. listening. Results are shown at cluster-corrected
p < 0.05. Peak activation co-ordinates are reported in Supplementary Table 1.

9
bioRxiv preprint doi: https://doi.org/10.1101/2022.10.15.512349; this version posted October 17, 2022. The copyright holder for this
preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.

Figure 3: Region of interest analyses. (A) ROI locations shown on the left hemisphere. Regions were
selected based on known involvement in semantic processing and defined using an anatomical atlas
(see Method for details). (B) Univariate activation values for discourse vs. baseline, where pale bars
represent left-hemisphere activation and bright bars represent the right hemisphere. Asterisks below
the x-axis indicate significant hemispheric differences. (C) RSA results, showing the Fisher-
transformed correlation between the semantic DSM and neural DSM in each ROI. Asterisks above
the x-axis indicate correlations significantly greater than zero, while those below the x-axis indicate
significant hemispheric differences. * = p < 0.05; ** = p < 0.01; *** = p < 0.001, all FDR-corrected.
Error bars show 1 SEM. IFG = inferior frontal gyrus; lATL = lateral anterior temporal lobe; vATL =
ventral anterior temporal lobe; pMTG = posterior middle temporal gyrus; AG = angular gyrus.

RSA searchlights: These analyses investigated whether similarity in the activation produced
during different passages of speech can be explained by similarity in the semantic content of
those passages. The RSA method assesses this by first computing a neural dis-similarity matrix
(DSM) that measures the dis-similarity (1 – Pearson correlation) between activation patterns
elicited at different points during task performance. To generate neural DSMs, we divided
each 50s discourse period into 5 segments of 10s duration. We calculated the pairwise dis-

10
bioRxiv preprint doi: https://doi.org/10.1101/2022.10.15.512349; this version posted October 17, 2022. The copyright holder for this
preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.

similarities between the activation patterns evoked during these segments of speech (see
Methods for details). We then tried to predict the values in the neural DSM using a semantic
DSM, which coded dis-similarity between the content of segment using an established vector-
based model of semantics [21]. We also controlled for similarity in the quantity of speech
contained in each segment (see Figure 1C and Methods). This process was repeated across a
series of “searchlights” to build a whole-brain map of semantic-neural correlations. As shown
in Figure 4, we performed three separate analyses: comparing listening segments with other
listening segments, comparing speaking segments with other speaking segments and, in a
cross-task analysis, comparing speaking segments with listening segments.
Results of the three searchlight analyses are shown in Figure 4. When participants
listened to discourse, semantic similarity predicted neural similarity in regions along the length
of the superior temporal sulci bilaterally. In the anterior temporal lobes, effects extended
ventrally into the middle and inferior temporal gyri, particularly in the right hemisphere.
Significant correlations were also observed in a large area of posterior medial cortex
encompassing posterior cingulate, cuneus and precuneus. These results converge with those
of previous studies in indicating that the bilateral lateral temporal lobe regions, as well as
other default mode network regions (posterior cingulate, medial prefrontal cortex), track
semantic content during spoken narrative comprehension [26, 27].
The second analysis extended the use of semantic RSA to the domain of discourse
production for the first time. Here, activity in a more extensive set of regions was found to
correlate with semantic similarity (Figure 4B). The strongest correlations were found in left
posterior pMTG and inferior parietal cortex, with significant associations also found in IFG,
the default mode regions of posterior cingulate and ventromedial prefrontal cortex, as well
as parts of the intraparietal sulci and motor cortices. Thus, when individuals produce
discourse on similar topics, this is reflected in similar activation patterns within a widely
distributed brain network. As in the listening analysis, the distribution of significant regions
was largely bilateral, in contrast to the highly left-lateralised pattern observed in the univariate
analysis.

11
Figure 4: Searchlight analyses. The top panel in the figure indicates which parts of the DSMs were used in each analysis. DSMs consist of pairwise
comparisons between different 10-second segments of discourse processing. In the Listening and Speaking analyses, we used pairs taken from the same
language task but taken from different scanning runs (since segments in the same run may be affected by temporal autocorrelation in the BOLD signal;
see Methods). In the cross-task analysis, pairs consisted of one Speaking and one Listening segment (which were always from different scanning runs).
Brain maps show regions where the correlation between neural and semantic DSMs was significantly greater than zero (at cluster-corrected p < 0.05).
Colour scales show the group-average Fisher-transformed correlation coefficient between semantic and neural DSMs; note that the scale is different in
the cross-task analysis. Peak effect co-ordinates are reported in Supplementary Table 2.

12
Figure 5: Relationships between neural DSMs. This analysis tested the correlations between neural DSMs obtained in different regions. Top panel shows
correlations between neural DSMs for each analysis. Strong correlations indicate that neural DSMs were more similar to one another. Bottom panel shows
hierarchical cluster analysis performed using the correlation data. The plots cluster regions according to how similar their neural DSMs are.

13
bioRxiv preprint doi: https://doi.org/10.1101/2022.10.15.512349; this version posted October 17, 2022. The copyright holder for this
preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.

Finally, the cross-task analysis tested for predictive effects of semantic similarity when
comparing activation patterns across comprehension and production tasks. Significant effects
here indicate that similarities in activation across language tasks can be predicted by the
underlying semantic content of the discourse being processed. We interpret this as evidence
for common coding of language content across speaking and listening tasks. The cross-task
analysis revealed semantic-neural correlations in an extensive and bilateral set of regions
(Figure 4C). These include regions classically associated with semantic and language
processing, such as the anterior and posterior temporal lobes and lateral prefrontal cortices,
but also nodes of the default mode network (inferior parietal lobes, posterior cingulate and
medial prefrontal cortex) and left hippocampus and parahippocampal gyrus. There are two
key implications of these results. First, they indicate that an array of brain regions encode the
content of one’s own speech in a similar way to the content of speech produced by others,
suggesting that the semantics of comprehension and production share a common neural code.
Second, they indicate that activation in right-hemisphere regions is as sensitive as left-
hemisphere regions to the meaning of discourse. Thus, while univariate analyses suggest that
the right hemisphere is less selectively engaged during discourse processing, its activation is
nevertheless predicted by the semantic content, suggesting a potential functional role for
these regions.
The cross-task analysis identified many of the same regions as the separate task-
specific analyses. When comparing these results, however, it is important to note that the
magnitudes of the neural-semantic correlations in the cross-task analysis are lower than in
the separate listening and speaking analyses. The cross-task analysis had greater statistical
power because a greater number of speech segment pairs were available when comparing
across different tasks (see Figure 4C). This allowed weaker semantic-neural correlations to
reach statistical significance in this analysis.
In the control analyses, neural DSMs were predicted used the quantity DSMs (which class
speech segments as dis-similar if they contain very different number of words) rather than
semantic DSMs. In the listening task, effects were found bilaterally in ventral premotor and
superior temporal regions, centred on Heschl’s gyrus (see Supplementary Figure 2). The
cross-task analyses revealed similar small clusters in the left and right superior temporal lobes,

14
bioRxiv preprint doi: https://doi.org/10.1101/2022.10.15.512349; this version posted October 17, 2022. The copyright holder for this
preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.

again centred on primary auditory cortex. Analysis of the speaking task revealed a more
distributed set of regions, including bilateral motor and premotor cortices, some medial parts
of the default mode network and left inferior parietal cortex, though with weaker correlations
than those seen in the main analyses. None of these analyses revealed the patterns of
temporal, inferior prefrontal and parietal effects seen in the main analysis. Thus, the control
analyses suggest that effects in Figure 4 are specific to predictors that capture high-level
conceptual content of language and not its lower-level properties.

RSA effects in regions of interest: These analyses investigated effects in our set of targeted
ROIs that are known to be key nodes in the brain’s semantic network [2, 5]. Correlations
between semantic and neural DSMs for each ROI are shown in Figure 3C. These are
consistent with the widespread and bilateral effects seen in the searchlight analyses. For the
listening task, semantic similarity predicted neural similarity in all regions except IFG and right
AG (at FDR-corrected p < 0.05). For the speaking task, all regions showed significant effects
with the exception of vATL and right AG. In the cross-task analysis, significant correlations
were observed in all regions except vATL. As with the searchlight results, in almost all cases
semantic DSMs predicted neural similarity to a similar degree in the left and right hemispheres.
Direct comparisons of the effects in left and right-hemisphere homologues revealed only two
significant differences. In the production task, the correlation was significantly stronger in left
vATL compared with right vATL. In the cross-task analysis, a significant difference was found
in pMTG, but with right pMTG showing the stronger effect. Thus, while our univariate ROI
analysis showed clearly that semantic regions in the left-hemisphere showed a stronger BOLD
response during discourse processing, here we found that activation in both hemispheres
tracked the semantic content of discourse to a similar degree.
Results of control ROI analyses, using the quantity DSMs as the predictor, are
provided in Supplementary Figure 3. A much more limited set of correlations was observed
here. In the listening task, quantity DSMs predicted neural patterns in some regions,
predominantly in the temporal lobes (left and right lATL, left vATL and right AG). In the
speaking task, only one region (left AG) showed a correlation above zero, and only right IFG
showed a cross-task effect. These results suggest that neural patterns in semantic brain

15
bioRxiv preprint doi: https://doi.org/10.1101/2022.10.15.512349; this version posted October 17, 2022. The copyright holder for this
preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.

regions were primarily sensitive to the content of discourse and not to the number of words
processed in each segment.

Comparison of similarity structure in different brain regions: The final analysis explored the
degree to which different parts of the semantic network represent the content of discourse
similarly. Rather than comparing each region’s neural DSM to a semantic DSM, here we
directly compared the neural DSMs of different ROIs with each other (see Figure 1D). The
correlations between ROIs for each analysis are shown in the top panel of Figure 5. All of the
correlations were positive. This indicates consistency in neural responses across the brain:
pairs of speech segments that elicited similar activation patterns in one ROI tended to elicit
similar patterns in all of the other ROIs. In general, the strongest correlations between ROIs
were observed in the listening task, with less convergence in the speaking task and the cross-
task analysis. Strong correlations also tended to be observed between left and right
homologue regions (visible as a diagonal line in the lower-right portion of each correlation
plot). Correlations between left and right AG were particularly strong. The strong cross-
hemispheric coupling in activation is also evident in the hierarchical cluster plots that group
ROIs by similarity in their DSMs. In every case, left and right homologues were most similar
to one another. This suggests a degree of functional association between left and right-
hemisphere regions, despite the fact that the left hemisphere regions consistently showed
stronger BOLD responses in the univariate analysis. Otherwise, the three cluster analyses
showed broadly similar relationships between regions, with the pMTG and AG most strongly
related and also clustered with the IFG, while ATL regions showed more distinct neural
patterns.

Discussion

Recent years have seen an explosion in investigations of how language content is

represented in the brain, driven by advances in decoding and pattern analysis methods allied
with more naturalistic fMRI paradigms [22, 24-27]. Crucially, however, prior studies have

16
bioRxiv preprint doi: https://doi.org/10.1101/2022.10.15.512349; this version posted October 17, 2022. The copyright holder for this
preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.

focused almost exclusively on how language is encoded during comprehension tasks, rather
than when people produce their own discourse. Here, for the first time, we used a model of
semantic discourse content to predict fMRI activation patterns during the production of
language, comparing this to comprehension-based patterns in the same individuals. We
imaged participants while they produced and listened to passages of discourse in response to
topic prompts and we used RSA to relate neural similarities between different passages to
similarity in their semantic content, using a vector-based model of distributional semantics
[21]. When participants produced discourse with similar content, they showed similar
patterns of neural activity in a wide range of frontal, temporal and parietal sites, which were
more extensive than those observed during comprehension. Moreover, semantic similarity
also predicted neural similarities when comparing between the two language tasks, showing
that concepts activated during production and comprehension share similar neural
representations. Finally, the regions that showed semantic coding of discourse content in both
production and comprehension were strikingly bilateral, suggesting a greater role for the right
hemisphere in discourse processing than simple activation analyses would suggest.
The major contribution of the present work is to demonstrate similar neural coding
of semantic content during production and comprehension of high-level discourse. Previous
studies have shown inter-subject correlations when comparing the activation of a speaker and
a comprehender [42, 48, 49], and have demonstrated that this coupling is greatest when both
participants understand narratives in the same way [54]. The present work makes two
important new contributions. First, we show that similar neural patterns for production and
comprehension occur within the same individuals as they switch between language tasks.
Second, we show that comprehension-production similarities can be predicted by a
distributional model of semantics, with semantically similar spoken/heard discourse passages
eliciting more similar activation patterns. Alignment between comprehension and production
is predicted by a range of theories which propose that higher levels of language processing
are largely shared between production and comprehension [35-37]. Potential benefits of this
shared cognitive machinery include the ability of the production system to make forward
predictions that aid comprehension [38, 55] and the ability for interlocutors to align their
mental frameworks during conversation [39].

17
bioRxiv preprint doi: https://doi.org/10.1101/2022.10.15.512349; this version posted October 17, 2022. The copyright holder for this
preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.

Semantic effects that crossed between language tasks were observed in a wide range
of brain regions. Prominent among these were key nodes of the default-mode network:
inferior parietal cortices (AG), posterior cingulate and medial prefrontal regions.
Neuroimaging evidence implicates these default-mode network regions in the construction
of mental models that represent the details of events (situation models) in a range of contexts,
including movie-watching [28], story-listening and reading [42, 54, 56], memory retrieval [57]
and in social interactions [53]. Consistent with these findings, our previous work showed
increased activation in this network when participants produced or heard discourse that was
low in coherence, and thus was harder to construct a situation model for [58]. The present
results provide more direct evidence for the role of default-mode regions in representing
discourse content, by showing that they represent the semantics of discourse consistently
across language processing modes.
We also found effects in ATL regions. Though parts of the ATL are frequently
identified as belonging to the default-mode network, theories of semantic cognition typically
ascribe a different role to this region compared to other default-mode network regions like
AG [59-62]. While AG is thought to be involved in combining concepts to represent events
and situations [1, 62, 63], theories of ATL function focus on its role in coding conceptual
structure at the single-word and concept level [2, 33]. Our data support this idea that ATL is
functionally distinct from AG, since cluster analyses indicated a strong separation between
ATL similarity patterns and those of the other semantic regions we studied (including AG but
also pMTG and IFG, which we will come to shortly). Nevertheless, there was some evidence
that the ATLs were sensitive to the semantic content of speech. Lateral ATLs encoded
semantic similarity during both conditions and in the cross-task analysis. The ventral part of
the ATL was sensitive to semantic similarity only during speech comprehension. The lack of
vATL effects during speech production may be due to this region’s high susceptibility to fMRI
signal dropout [64].
ROI analyses also revealed effects of semantic similarity in the posterior temporal
lobes (pMTG) and prefrontal cortices (IFG). pMTG has been implicated in the semantics of
events and actions [65-67], processes which are likely to be critical when understanding or
generating narrative discourse. In addition, both pMTG and IFG form part of the semantic

18
bioRxiv preprint doi: https://doi.org/10.1101/2022.10.15.512349; this version posted October 17, 2022. The copyright holder for this
preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.

control network: a set of regions which show increased engagement when semantic
processing requires high levels of cognitive control [68, 69]. Functions of this network include
supporting retrieval of context-relevant semantic information as well as selection processes
that arbitrate between multiple competing concepts to ensure contextually relevant
information is attended to [2]. Semantic RSA effects in these regions might therefore reflect
systematic variations in the cognitive control demands of different discourse topics.
At the whole-brain level, the most striking finding was the widespread right-
hemisphere coding of discourse content. This occurred even though right-hemisphere regions
were not activated to the same extent as left-hemisphere regions and even, in the case of
production, often deactivated relative to baseline conditions or to rest. The cross-ROI
analysis further demonstrated that semantic content evoked similar activation patterns across
left and right homologues of semantic processing regions. Thus, despite showing different
levels of BOLD response, left and right-hemisphere regions appear to track the semantics of
discourse in a similar way.
These results are consistent with the view that the right hemisphere makes specific
contributions to understanding natural language, such as understanding metaphors and jokes
[70, 71], making inferences [72] and comprehending narrative structures [73]. One particular
framework suggests that semantic processing occurs bilaterally, with each hemisphere
undertaking its own type of neurocomputation but working interactively with the other [74].
The left hemisphere shows fine coding, rapidly activating a network of strongly linked semantic
associates; while the right hemisphere encodes more coarse associations, and as such is
sensitive to more distant semantic relations and broader conceptual meanings. This coarse
coding is thought to be more important for capturing the nuanced meanings contained in
more complex language acts, such as extended narratives. Indeed, our findings are consistent
with previous studies that have implicated a bilateral network in comprehension of extended,
naturalistic speech [e.g., 22]. Importantly, however, we have shown that this right-hemisphere
contribution is also present in discourse production. While few neuroimaging studies have
investigated language production, lesion studies indicate that patients with right-hemisphere
damage produce disorganised and poorly structured narratives, suggesting a right-hemisphere
role in discourse planning processes [75-77]. Studies of ATL function also point to a bilateral

19
bioRxiv preprint doi: https://doi.org/10.1101/2022.10.15.512349; this version posted October 17, 2022. The copyright holder for this
preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.

pattern of engagement in lexical-semantic processing [78, 79]. Our results are broadly in line
with these findings in suggesting that right-hemisphere homologues of semantic regions code
the semantic properties of discourse during production as well as comprehension.
We end by considering ways in which the present approach could be extended in
future work. First, our semantic model uses information about patterns of word usage in
natural language to determine similarity in meaning. Though such distributional approaches
can very accurately mimic human judgements of semantic relatedness [80], they have been
criticized on the grounds that they do not capture perceptual properties of objects such as
their size, shape or colour [81-83]. Future studies may benefit from the use of more complex
semantic models that make use of both lexical co-occurrence and experiential attributes [84-
87]. Applying such an approach to discourse (cf. single-word comprehension) will present
significant challenges, as discourse contains descriptions of complex situations whose
perceptual characteristics cannot easily be predicted from their lexical constituents. Second,
conceptual representations are dynamic and are shaped by the individual’s personal context,
in terms of their past experiences, individual processing preferences and abilities [88]. We
controlled for these individual differences by directly comparing comprehension and
production in the same individuals (as opposed to previous studies which compared speaking
and listening in different people). Nevertheless, the degree to which people’s discourse
representations vary in their content and neural instantiation remains an important question
for future work.
To conclude, we have used a distributional model of semantics to predict neural
similarity patterns during discourse processing. We have shown that a broad set of brain
regions code language content during speech production in a similar way to when hearing
someone else’s speech, suggesting common coding across different language processing
modes. We believe that this work can stimulate further insights into how the semantic
systems of the brain drive the generation, as well as understanding, of language.

20
bioRxiv preprint doi: https://doi.org/10.1101/2022.10.15.512349; this version posted October 17, 2022. The copyright holder for this
preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.

Method

Participants: 25 young adults participated in the study in exchange for payment. Their mean
age was 24 (SD = 4.4, range = 18-35), all were native English speakers and they were all
classified as right-handed using the Edinburgh Handedness Inventory [89]. The study was
approved by the Psychology Research Ethics Committee of the University of Edinburgh and
all participants gave informed consent. Unrelated analyses of this dataset, investigating effects
of discourse coherence and other psycholinguistic properties, have been reported previously
[58, 90]. In contrast, the present study used multivariate analyses to investigate the neural
representation of discourse content.

Materials: In the speaking task, 12 prompts were used to elicit discourse from participants.
Each prompt probed common semantic knowledge on a specific topic (e.g., Describe how you
would make a cup of tea or coffee; see Supplementary Materials for a complete list of prompts
in both tasks). This discourse speaking condition was contrasted with a baseline of automatic
speech that involved repeated recitation of the English nursery rhyme, Humpty Dumpty [41,
91, 92]. For the listening task, we generated audio samples of discourse using transcripts of
speech provided by participants in a previous behavioural study, in response to 12 topic
prompts [93]. Listening topics were different from those used in the speaking task. For each
topic, we selected two different responses provided by different participants in the Hoffman
et al. (2018) study. One sample was highly consistent with the topic and the other deviated
somewhat from the specified topic (as indicated by global coherence measured using the
methods of Hoffman et al [91, 93]). We did this to increase the variability in the semantic
content of the stimuli. We divided the samples into two counterbalanced sets, so that each
participant would hear one of the two sets. To generate the audio recordings that would be
presented in the scanner, the 24 transcripts were read aloud by the same male native English
speaker and edited so that their duration was 50s each (a sample transcript is provided in
Supplementary Materials). A 10s recording of the same speaker reciting the Humpty Dumpty
rhyme was also made to provide a baseline listening condition.

21
bioRxiv preprint doi: https://doi.org/10.1101/2022.10.15.512349; this version posted October 17, 2022. The copyright holder for this
preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.

Design and procedure: Each participant completed two speaking and two listening runs,
presented in an alternating sequence. The order of runs was counterbalanced over
participants. Each run lasted approximately 8 minutes and included six discourse trials and
five baseline trials, with the order of discourse trials in each run fully randomised for each
participant.
Speaking trials began with the presentation of a written topic prompt on screen for
6s (see Figure 1A). Participants were asked to prepare to speak during this period and to
start speaking when a green circle replaced the prompt in the centre of screen. They were
instructed to speak about the topic for 50s, after which a red cross would replace the green
circle. At this point participants were instructed to wait for the next prompt to appear on
screen. The procedure for listening trials was the same, except participants were asked to
listen to the speaker attentively for 50s while the green circle was on screen. For the baseline
automatic speech conditions, participants were instructed to recite or to listen to the
Humpty Dumpty rhyme for a 10s period. When speaking, they were asked to start reciting
again from the beginning if they reached the end of the nursery rhyme before the 10s had
elapsed. The baseline conditions therefore involved grammatically well-formed continuous
speech, but without the requirement to generate or understand novel discourse. All trials
were presented with an inter-stimulus interval jittered between 3s and 7s (M = 5s).
Before scanning, participants were presented with training trials to familiarise them
with the tasks. To ensure attention during listening runs, they were told that they would
receive a memory test on the material after scanning. In this test, participants answered 12
multiple choice questions, one for each topic presented during listening. They were also asked
to rate how well they could hear the speech samples in the scanner from 1 (inaudible) to 7
(perfectly audible).

Processing of speech samples: Spoken responses were digitally recorded with an MRI-
compatible microphone and processed with noise cancellation software [94] to reduce noise
from the scanner. They were then manually transcribed. For analysis, each 50s response was
divided into 5 segments of 10s duration.

22
bioRxiv preprint doi: https://doi.org/10.1101/2022.10.15.512349; this version posted October 17, 2022. The copyright holder for this
preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.

For each participant, we constructed two dissimilarity matrices (DSMs) based on the
properties of the speech produced or heard in each 10s segment. Our main DSM indexed
the semantic relatedness of speech segments using Latent Semantic Analysis (LSA) [21]. Like
other distributional models of semantic representation [19, 20], LSA represents words as
embeddings in a high-dimensional semantic space. The proximity of two words in this space
indicates the degree to which they are used in similar contexts in natural language, which is
taken as a proxy for similarity in meaning. We used LSA vectors generated from the British
National Corpus, which we have used previously for analyses of discourse production [91,
93]. As in previous studies, a vector representation for each speech segment was generated
by averaging the LSA vectors of all of the words contained in the segment [excluding function
words and weighting vectors by their log frequency in the segment and their entropy in the
original corpus; 93]. Once a vector had been calculated for each speech segment, a semantic
DSM was calculated using a cosine similarity metric. An example of a semantic DSM for a
single participant is shown in Figure 1B. Each participant’s semantic DSM was unique as each
participant produced different information in response to the topic prompts.
We computed a second DSM that captured variation in the quantity of speech
produced or heard in different segments. Here, dissimilarity between two segments was
defined as the difference in the number of words contained in the two segments; thus,
segments containing a similar quantity of speech were represented as similar to one another.
A quantity DSM for a single participant is shown in Figure 1B. We included the quantity DSM
as a covariate in our semantic analyses, to ensure that our effects were being driven by the
semantic content of the discourse and were not confounded by variation in the amount of
speech participants were processing.

Image acquisition and processing: Participants were scanned on a 3T Siemens Prisma scanner
using a 32-channel head coil. fMRI data were acquired at three echo times (13ms, 31ms, and
48ms) using a whole-brain multi-echo acquisition protocol. Data from these three echo series
were weighted and combined, and the resulting time-series were denoised using independent
components analysis (ICA). This approach improves the signal quality in regions that typically
suffer from susceptibility artefacts (e.g., the ventral anterior temporal lobes) and is helpful in

23
bioRxiv preprint doi: https://doi.org/10.1101/2022.10.15.512349; this version posted October 17, 2022. The copyright holder for this
preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.

reducing motion-related artefacts [95]. The TR was 1.7 s and images consisted of 46 slices
with an 80 x 80 matrix and isotropic voxel size of 3mm. Multiband acceleration with a factor
of 2 was used and the flip angle was 73°. Four runs of 281 volumes (477.7s) were acquired. A
high-resolution T1-weighted structural image was also acquired for each participant using an
MP-RAGE sequence with 1mm isotropic voxels, TR = 2.5 s, TE = 4.6 ms.
Images were pre-processed and analysed using SPM12 and the TE-Dependent Analysis
Toolbox 0.0.7 (Tedana) [96]. Estimates of head motion were obtained using the first BOLD
echo series. Slice-timing correction was carried out and images were then realigned using the
previously obtained motion estimates. Tedana was used to combine the three echo series
into a single-time series and to divide the data into components classified as either BOLD-
signal or noise-related based on their patterns of signal decay over increasing TEs [95].
Components classified as noise were discarded. After that, images were unwarped with a B0
fieldmap to correct for irregularities in the scanner's magnetic field. Finally, functional images
were spatially normalised to MNI space using SPM’s DARTEL tool [97].
For univariate analysis, images were smoothed with a kernel of 8mm FWHM. Data
were treated with a high-pass filter with a cut-off of 128s and the four experimental runs
were analysed using a single general linear model. Four types of speech block were modelled:
discourse speaking (50s), baseline speaking (10s), discourse listening (50s), and baseline
listening (10s). Prompt presentation periods were modelled as blocks of 6s, with separate
regressors for discourse prompts and Humpty Dumpty prompts. Covariates consisted of six
motion parameters and their first-order derivatives.
For multivariate analysis, images were smoothed with a kernel of 4mm FWHM, as a
small amount of smoothing has been shown to improve the sensitivity of multivariate analyses
[98, 99]. Each run was modelled in a separate general linear model. Regressors for prompts
and baseline conditions were the same as in the univariate analysis. Discourse blocks were
divided into segments of 10s duration (5 per prompt) and each segment was modelled using
a separate regressor. Thus, we obtained a separate beta map for each 10s segment of speech
heard/produced by each participant. Beta maps were converted into t-maps for RSA analysis.
Motion covariates were included as in the univariate analysis.

24
bioRxiv preprint doi: https://doi.org/10.1101/2022.10.15.512349; this version posted October 17, 2022. The copyright holder for this
preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.

Regions of interest: In addition to whole-brain analyses, we defined five anatomical regions in

the left and right hemispheres based on their involvement in semantic cognition. These are
shown in Figure 3A. Four of the five ROIs were defined using probability distribution maps
from the Harvard-Oxford brain atlas [100], including all voxels with a >30% probability of
falling within the following regions:
1. Inferior frontal gyrus (IFG): the pars orbitalis and pars triangularis regions of inferior
frontal gyrus, with voxels more medial than x = ±30 removed to exclude medial
orbitofrontal cortex
2. Lateral anterior temporal lobe (lATL): the anterior division of the superior and middle
temporal gyri
3. Ventral anterior temporal lobe (vATL): the anterior division of the inferior temporal and
fusiform gyri
4. Posterior middle temporal gyrus (pMTG): the temporo-occipital part of the middle
temporal gyrus
The final ROI covered the angular gyrus (AG) and included voxels with a >30% probability of
falling within this region in the LPBA40 atlas [101]. A different atlas was used in this case
because the AG region defined in the Harvard-Oxford atlas is small and does not include
parts of the inferior parietal cortex typically implicated in semantic processing.

Univariate analyses: We obtained whole-brain activation maps for speech production

(discourse minus baseline) and speech comprehension (discourse minus baseline). We
computed a subtraction of these two maps to identify differences between speaking and
listening at the discourse level. For these analyses we employed a voxel-height threshold of p
< 0.005 for one-sample t-tests with correction for multiple comparisons at the cluster level,
using SPM’s random field theory. The direct comparison of speaking and listening was
restricted to voxels that showed an effect of discourse > baseline in at least one of the two
tasks (at a liberal threshold of p < 0.05 uncorrected). In addition to whole-brain analysis,
contrast estimates for discourse minus baseline were extracted for each ROI in each task and
were entered into a 2 x 2 x 5 (task x hemisphere x ROI) ANOVA.

25
bioRxiv preprint doi: https://doi.org/10.1101/2022.10.15.512349; this version posted October 17, 2022. The copyright holder for this
preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.

Representational similarity analysis: RSA was used to investigate the degree to which neural
patterns were predicted by the semantics of the discourse that participants heard and
produced. Analyses were performed using CoSMoMVPA [102]. Searchlight analyses were run
using a spherical searchlight of radius 12mm (4 voxels) and ROI analyses were performed in
the 10 semantic regions described previously. Because the ROIs varied substantially in size, a
voxel selection criterion was applied to ensure that all ROI analyses used the same number
of voxels. For each participant, voxels in each ROI were ordered by their effect size in the
contrast of discourse over baseline and the 100 most active voxels in each ROI were selected
for the RSAs. This ensured that all ROI analyses were equally powered and were based on
the voxels that were most engaged by the task.
Separate RSAs were performed for the listening and speaking tasks, as well as a cross-
task analysis that tested pattern similarity across comprehension and production. For
listening, a neural DSM was generated by calculating 1 minus Pearson correlations between
activation patterns for pairs of speech segments presented during the listening task. To avoid
any confounding effects of temporal auto-correlation in the BOLD signal, only pairs of
segments from different scanning runs were compared [103] (see Figure 4A). This also meant
that segments from within the same speech topic were never compared. Activation patterns
were mean-centred within each run, ensuring that each voxel had a mean activation of zero
[as recommended by 104]. For the speaking task, the same process was followed but pairs
were taken from the speaking runs (see Figure 4B). For the cross-task analysis, each pair
consisted of one speaking segment and one listening segment (see Figure 4C). Thus, the cross-
task analysis tested the degree to which similarity in neural patterns across language tasks
could be explained by their semantic content.
The association between neural DSMs and their corresponding semantic DSMs was
measured using partial Spearman correlations, which were Fisher-transformed prior to
statistical inference. Analyses included the quantity DSMs as a covariate, ensuring that effects
were not dependent on the number of words in each segment (see Figure 1C). To test the
degree to which our results were specific to semantic processing, control analyses were
conducted using quantity DSMs as the predictor and semantic DSMs as the covariate. The
results of these are reported in Supplementary Materials.

26
bioRxiv preprint doi: https://doi.org/10.1101/2022.10.15.512349; this version posted October 17, 2022. The copyright holder for this
preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.

To determine whether semantic DSMs significantly predicted neural dissimilarity

patterns, permutation tests were performed using the two-stage method introduced by
Stelzer et al. [105]. For each participant, we calculated Spearman partial correlations using a
semantic DSM in which the order of segments had been randomly permuted within each run.
This process was repeated 100 times to provide a distribution of results for each participant
under the null hypothesis. Following this, a Monte Carlo approach was taken to generate a
null distribution at the group level. We randomly selected one correlation map from each
participant’s null distribution and averaged these to give a group mean. This process was
repeated 10,000 times to generate a distribution of the expected group correlations under
the null hypothesis. In ROI analyses, the position of the observed result in this null distribution
was used to determine a p-value (e.g., if the observed accuracy was greater than 99% of values
in the null distribution, this would represent a p-value of 0.01). For searchlight analyses,
observed and null maps were entered into CoSMoMVPA’s Monte Carlo cluster statistics
function, which returned a statistical map corrected for multiple comparisons using threshold-
free cluster enhancement [106]. These maps were thresholded at corrected p < 0.05.

Comparison of similarity structure in different brain regions: Finally, we investigated the

relationship between neural DSMs in different ROIs (see Figure 1D). The purpose of this
analysis was to explore the degree to which different parts of the semantic network
represented discourse content in a similar way. For each participant, Spearman correlations
were calculated between pairs of neural DSMs, for all possible pairs of ROIs. This resulted in
a correlation matrix that coded the degree to which pairs of ROIs shared similar neural DSMs.
These were then averaged over participants to give a single correlation matrix for the whole
group. This process was performed within the listening and speaking tasks, as well as across
tasks. To visualise the relationships between ROIs, group correlation matrices were
converted to distance matrices and hierarchical cluster analyses performed using R, with
Ward’s distance metric.

27
bioRxiv preprint doi: https://doi.org/10.1101/2022.10.15.512349; this version posted October 17, 2022. The copyright holder for this
preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.

Data and Code Availability: Data and code associated with this study are available at:
https://osf.io/s3z8y/. Unthresholded activation and correlation maps are archived at:
https://neurovault.org/collections/12792/.

References
1. Binder, J.R. and R.H. Desai, The neurobiology of semantic memory. Trends in Cognitive
Sciences, 2011. 15(11): p. 527-536.
2. Lambon Ralph, M.A., E. Jefferies, K. Patterson, and T.T. Rogers, The neural and
computational bases of semantic cognition. Nature Reviews Neuroscience, 2017. 18: p.
42-55.
3. Pulvermüller, F., How neurons make meaning: brain mechanisms for embodied and
abstract-symbolic semantics. Trends in cognitive sciences, 2013. 17(9): p. 458-470.
4. Martin, A., The representation of object concepts in the brain. Annu Rev Psychol, 2007.
58: p. 25-45.
5. Binder, J.R., R.H. Desai, W.W. Graves, and L.L. Conant, Where is the semantic
system? A critical review and meta-analysis of 120 functional neuroimaging studies.
Cerebral Cortex, 2009. 19: p. 2767-2796.
6. Norman, K.A., S.M. Polyn, G.J. Detre, and J.V. Haxby, Beyond mind-reading: multi-voxel
pattern analysis of fMRI data. Trends in cognitive sciences, 2006. 10(9): p. 424-430.
7. Kriegeskorte, N., M. Mur, and P.A. Bandettini, Representational similarity analysis-
connecting the branches of systems neuroscience. Frontiers in systems neuroscience,
2008: p. 4.
8. Haxby, J.V., A.C. Connolly, and J.S. Guntupalli, Decoding neural representational spaces
using multivariate pattern analysis. Annual review of neuroscience, 2014. 37(1): p.
435-456.
9. Wang, X., et al., Organizational principles of abstract words in the human brain.
Cerebral Cortex, 2018. 28(12): p. 4305-4318.

28
bioRxiv preprint doi: https://doi.org/10.1101/2022.10.15.512349; this version posted October 17, 2022. The copyright holder for this
preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.

10. Carota, F., N. Kriegeskorte, H. Nili, and F. Pulvermüller, Representational similarity

mapping of distributional semantics in left inferior frontal, middle temporal, and motor
cortex. Cerebral Cortex, 2017. 27(1): p. 294-309.
11. Devereux, B.J., A. Clarke, A. Marouchos, and L.K. Tyler, Representational Similarity
Analysis Reveals Commonalities and Differences in the Semantic Processing of Words and
Objects. Journal of Neuroscience, 2013. 33(48): p. 18906-18916.
12. Fischer-Baum, S., D. Bruggemann, I.F. Gallego, D.S. Li, and E.R. Tamez, Decoding levels
of representation in reading: A representational similarity approach. Cortex, 2017. 90: p.
88-102.
13. Mitchell, T.M., S.V. Shinkareva, A. Carlson, K.M. Chang, V.L. Malave, R.A. Mason, and
M.A. Just, Predicting human brain activity associated with the meanings of nouns.
Science, 2008. 320(5880): p. 1191-5.
14. Anderson, A.J., J.R. Binder, L. Fernandino, C.J. Humphries, L.L. Conant, M. Aguilar, X.
Wang, D. Doko, and R.D. Raizada, Predicting neural activity patterns associated with
sentences using a neurobiologically motivated model of semantic representation. Cerebral
Cortex, 2017. 27(9): p. 4379-4395.
15. Pereira, F., B. Lou, B. Pritchett, S. Ritter, S.J. Gershman, N. Kanwisher, M. Botvinick,
and E. Fedorenko, Toward a universal decoder of linguistic meaning from brain
activation. Nature communications, 2018. 9(1): p. 1-13.
16. Wang, J., V.L. Cherkassky, and M.A. Just, Predicting the brain activation pattern
associated with the propositional content of a sentence: modeling neural representations
of events and states. Human brain mapping, 2017. 38(10): p. 4865-4881.
17. Fernandino, L., J.R. Binder, R.H. Desai, S.L. Pendl, C.J. Humphries, W.L. Gross, L.L.
Conant, and M.S. Seidenberg, Concept representation reflects multimodal abstraction: A
framework for embodied semantics. Cerebral cortex, 2016. 26(5): p. 2018-2034.
18. Binder, J.R., L.L. Conant, C.J. Humphries, L. Fernandino, S.B. Simons, M. Aguilar, and
R.H. Desai, Toward a brain-based componential semantic representation. Cognitive
neuropsychology, 2016. 33(3-4): p. 130-174.
19. Mikolov, T., K. Chen, G. Corrado, and J. Dean, Efficient estimation of word
representations in vector space. arXiv preprint arXiv:1301.3781, 2013.

29
bioRxiv preprint doi: https://doi.org/10.1101/2022.10.15.512349; this version posted October 17, 2022. The copyright holder for this
preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.

20. Pennington, J., R. Socher, and C.D. Manning. Glove: Global vectors for word
representation. in Proceedings of the 2014 conference on empirical methods in natural
language processing (EMNLP). 2014.
21. Landauer, T.K. and S.T. Dumais, A solution to Plato's problem: The latent semantic
analysis theory of acquisition, induction and representation of knowledge. Psychological
Review, 1997. 104: p. 211-240.
22. Huth, A.G., W.A. De Heer, T.L. Griffiths, F.E. Theunissen, and J.L. Gallant, Natural
speech reveals the semantic maps that tile human cerebral cortex. Nature, 2016.
532(7600): p. 453-458.
23. de Heer, W.A., A.G. Huth, T.L. Griffiths, J.L. Gallant, and F.E. Theunissen, The
hierarchical cortical organization of human speech processing. Journal of Neuroscience,
2017. 37(27): p. 6539-6557.
24. Wehbe, L., B. Murphy, P. Talukdar, A. Fyshe, A. Ramdas, and T. Mitchell,
Simultaneously uncovering the patterns of brain regions involved in different story reading
subprocesses. PloS one, 2014. 9(11): p. e112575.
25. Dehghani, M., et al., Decoding the neural representation of story meanings across
languages. Human brain mapping, 2017. 38(12): p. 6096-6106.
26. Schrimpf, M., I.A. Blank, G. Tuckute, C. Kauf, E.A. Hosseini, N. Kanwisher, J.B.
Tenenbaum, and E. Fedorenko, The neural architecture of language: Integrative
modeling converges on predictive processing. Proceedings of the National Academy of
Sciences, 2021. 118(45): p. e2105646118.
27. Zhang, Y., K. Han, R. Worth, and Z. Liu, Connecting concepts in the brain by mapping
cortical representations of semantic relations. Nature communications, 2020. 11(1): p.
1-13.
28. Baldassano, C., U. Hasson, and K.A. Norman, Representation of real-world event
schemas during narrative perception. Journal of Neuroscience, 2018. 38(45): p. 9689-
9699.
29. Fairhall, S.L. and A. Caramazza, Brain Regions That Represent Amodal Conceptual
Knowledge. Journal of Neuroscience, 2013. 33(25): p. 10552-10558.

30
bioRxiv preprint doi: https://doi.org/10.1101/2022.10.15.512349; this version posted October 17, 2022. The copyright holder for this
preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.

30. Liuzzi, A.G., R. Bruffaerts, R. Peeters, K. Adamczuk, E. Keuleers, S. De Deyne, G.

Storms, P. Dupont, and R. Vandenberghe, Cross-modal representation of spoken and
written word meaning in left pars triangularis. NeuroImage, 2017. 150: p. 292-307.
31. Asyraff, A., R. Lemarchand, A. Tamm, and P. Hoffman, Stimulus-independent neural
coding of event semantics: Evidence from cross-sentence fMRI decoding. NeuroImage,
2021. 236: p. 118073.
32. Correia, J., E. Formisano, G. Valente, L. Hausfeld, B. Jansma, and M. Bonte, Brain-
based translation: fMRI decoding of spoken words in bilinguals reveals language-
independent semantic representations in anterior temporal lobe. Journal of
Neuroscience, 2014. 34(1): p. 332-338.
33. Patterson, K., P.J. Nestor, and T.T. Rogers, Where do you know what you know? The
representation of semantic knowledge in the human brain. Nature Reviews
Neuroscience, 2007. 8(12): p. 976-987.
34. Margulies, D.S., et al., Situating the default-mode network along a principal gradient of
macroscale cortical organization. Proceedings of the National Academy of Sciences,
2016. 113(44): p. 12574-12579.
35. Kintsch, W. and T.A. Vandijk, Toward a Model of Text Comprehension and Production.
Psychological Review, 1978. 85(5): p. 363-394.
36. Levelt, W.J., A. Roelofs, and A.S. Meyer, A theory of lexical access in speech production.
Behavioral and brain sciences, 1999. 22(1): p. 1-38.
37. Hagoort, P., MUC (Memory, Unification, Control) and beyond. Front Psychol, 2013. 4: p.
416.
38. Pickering, M.J. and S. Garrod, An integrated theory of language production and
comprehension. Behav Brain Sci, 2013. 36(4): p. 329-47.
39. Pickering, M.J. and S. Garrod, Toward a mechanistic psychology of dialogue. Behavioral
and brain sciences, 2004. 27(2): p. 169-190.
40. Awad, M., J.E. Warren, S.K. Scott, F.E. Turkheimer, and R.J. Wise, A common system
for the comprehension and production of narrative speech. J Neurosci, 2007. 27(43): p.
11455-64.

31
bioRxiv preprint doi: https://doi.org/10.1101/2022.10.15.512349; this version posted October 17, 2022. The copyright holder for this
preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.

41. AbdulSabur, N.Y., Y. Xu, S. Liu, H.M. Chow, M. Baxter, J. Carson, and A.R. Braun,
Neural correlates and network connectivity underlying narrative production and
comprehension: A combined fMRI and PET study. Cortex, 2014. 57: p. 107-127.
42. Silbert, L.J., C.J. Honey, E. Simony, D. Poeppel, and U. Hasson, Coupled neural systems
underlie the production and comprehension of naturalistic narrative speech. Proceedings
of the National Academy of Sciences of the United States of America, 2014.
111(43): p. E4687-E4696.
43. Hickok, G. and D. Poeppel, The cortical organization of speech processing. Nature
Reviews Neuroscience, 2007. 8(5): p. 393-402.
44. Poeppel, D., K. Emmorey, G. Hickok, and L. Pylkkänen, Towards a new neurobiology of
language. Journal of Neuroscience, 2012. 32(41): p. 14125-14131.
45. Lambon Ralph, M.A., J.L. McClelland, K. Patterson, C.J. Galton, and J.R. Hodges, No
right to speak? The relationship between object naming and semantic impairment:
Neuropsychological abstract evidence and a computational model. Journal of Cognitive
Neuroscience, 2001. 13(3): p. 341-356.
46. Schapiro, A.C., J.L. McClelland, S.R. Welbourne, T.T. Rogers, and M.A. Lambon
Ralph, Why bilateral damage is worse than unilateral damage to the brain. J Cogn
Neurosci, 2013. 25(12): p. 2107-23.
47. Federmeier, K.D., Thinking ahead: The role and roots of prediction in language
comprehension. Psychophysiology, 2007. 44(4): p. 491-505.
48. Stephens, G.J., L.J. Silbert, and U. Hasson, Speaker-listener neural coupling underlies
successful communication. Proc Natl Acad Sci U S A, 2010. 107(32): p. 14425-30.
49. Liu, Y., E.A. Piazza, E. Simony, P.A. Shewokis, B. Onaral, U. Hasson, and H. Ayaz,
Measuring speaker–listener neural coupling with functional near infrared spectroscopy.
Scientific reports, 2017. 7(1): p. 1-13.
50. Jiang, J., B. Dai, D. Peng, C. Zhu, L. Liu, and C. Lu, Neural synchronization during face-
to-face communication. Journal of Neuroscience, 2012. 32(45): p. 16064-16069.
51. Badre, D. and A.D. Wagner, Left ventrolateral prefrontal cortex and the cognitive control
of memory. Neuropsychologia, 2007. 45(13): p. 2883-2901.

32
bioRxiv preprint doi: https://doi.org/10.1101/2022.10.15.512349; this version posted October 17, 2022. The copyright holder for this
preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.

52. Jefferies, E., The neural basis of semantic cognition: Converging evidence from
neuropsychology, neuroimaging and TMS. Cortex, 2013. 49: p. 611-625.
53. Yeshurun, Y., M. Nguyen, and U. Hasson, The default mode network: where the
idiosyncratic self meets the shared social world. Nature Reviews Neuroscience, 2021.
22(3): p. 181-192.
54. Heidlmayr, K., K. Weber, A. Takashima, and P. Hagoort, No title, no theme: The
joined neural space between speakers and listeners during production and comprehension
of multi-sentence discourse. cortex, 2020. 130: p. 111-126.
55. Dell, G.S. and F. Chang, The P-chain: Relating sentence production and its disorders to
comprehension and acquisition. Philosophical Transactions of the Royal Society B:
Biological Sciences, 2014. 369(1634): p. 20120394.
56. Ferstl, E.C., J. Neumann, C. Bogler, and D.Y. von Cramon, The extended language
network: a meta-analysis of neuroimaging studies on text comprehension. Hum Brain
Mapp, 2008. 29(5): p. 581-93.
57. Ranganath, C. and M. Ritchey, Two cortical systems for memory-guided behaviour.
Nature reviews neuroscience, 2012. 13(10): p. 713-726.
58. Morales, M., T. Patel, A. Tamm, M. Pickering, and P. Hoffman, Similar neural networks
respond to coherence during comprehension and production of discourse. Cerebral
Cortex, in press.
59. Humphreys, G.F., P. Hoffman, M. Visser, R.J. Binney, and M.A. Lambon Ralph,
Establishing task- and modality-dependent dissociations between the semantic and default
mode networks. Proceedings of the National Academy of Sciences of the United
States of America, 2015. 112(25): p. 7857-7862.
60. Farahibozorg, S.-R., R.N. Henson, A.M. Woollams, and O. Hauk, Distinct roles for the
Anterior Temporal Lobe and Angular Gyrus in the spatio-temporal cortical semantic
network. Biorxiv, 2019: p. 544114.
61. Humphreys, G.F., M.A. Lambon Ralph, and J.S. Simons, A unifying account of angular
gyrus contributions to episodic and semantic cognition. Trends in Neurosciences, 2021.
44(6): p. 452-463.

33
bioRxiv preprint doi: https://doi.org/10.1101/2022.10.15.512349; this version posted October 17, 2022. The copyright holder for this
preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.

62. Mirman, D., J.-F. Landrigan, and A.E. Britt, Taxonomic and thematic semantic systems.
Psychological bulletin, 2017. 143(5): p. 499.
63. Price, A.R., M.F. Bonner, J.E. Peelle, and M. Grossman, Converging evidence for the
neuroanatomic basis of combinatorial semantics in the angular gyrus. The Journal of
Neuroscience, 2015. 35(7): p. 3276-3284.
64. Ojemann, J.G., E. Akbudak, A.Z. Snyder, R.C. McKinstry, M.E. Raichle, and T.E.
Conturo, Anatomic localization and quantitative analysis of gradient refocused echo-
planar fMRI susceptibility artifacts. Neuroimage, 1997. 6(3): p. 156-167.
65. Davey, J., et al., Exploring the role of the posterior middle temporal gyrus in semantic
cognition: Integration of anterior temporal lobe with executive processes. NeuroImage,
2016. 137: p. 165-177.
66. Liljeström, M., A. Tarkiainen, T. Parviainen, J. Kujala, J. Numminen, J. Hiltunen, M.
Laine, and R. Salmelin, Perceiving and naming actions and objects. Neuroimage, 2008.
41(3): p. 1132-1141.
67. Watson, C.E., E.R. Cardillo, G.R. Ianni, and A. Chatterjee, Action concepts in the brain:
an activation likelihood estimation meta-analysis. J Cogn Neurosci, 2013. 25(8): p.
1191-205.
68. Noonan, K.A., E. Jefferies, M. Visser, and M.A. Lambon Ralph, Going beyond Inferior
Prefrontal Involvement in Semantic Control: Evidence for the Additional Contribution of
Dorsal Angular Gyrus and Posterior Middle Temporal Cortex. Journal of Cognitive
Neuroscience, 2013. 25(11): p. 1824-1850.
69. Jackson, R.L., The neural correlates of semantic control revisited. NeuroImage, 2021.
224: p. 117444.
70. Rapp, A.M., D.E. Mutschler, and M. Erb, Where in the brain is nonliteral language? A
coordinate-based meta-analysis of functional magnetic resonance imaging studies.
Neuroimage, 2012. 63(1): p. 600-610.
71. Marinkovic, K., S. Baldwin, M.G. Courtney, T. Witzel, A.M. Dale, and E. Halgren,
Right hemisphere has the last laugh: neural dynamics of joke appreciation. Cognitive,
Affective, & Behavioral Neuroscience, 2011. 11(1): p. 113-130.

34
bioRxiv preprint doi: https://doi.org/10.1101/2022.10.15.512349; this version posted October 17, 2022. The copyright holder for this
preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.

72. Mason, R.A. and M.A. Just, How the brain processes causal inferences in text: A
theoretical account of generation and integration component processes utilizing both
cerebral hemispheres. Psychological Science, 2004. 15(1): p. 1-7.
73. Knutson, K.M., J.N. Wood, and J. Grafman, Brain activation in processing temporal
sequence: an fMRI study. Neuroimage, 2004. 23(4): p. 1299-1307.
74. Jung-Beeman, M., Bilateral brain processes for comprehending natural language. Trends
in cognitive sciences, 2005. 9(11): p. 512-518.
75. Marini, A., S. Carlomagno, C. Caltagirone, and U. Nocentini, The role played by the
right hemisphere in the organization of complex textual structures. Brain and language,
2005. 93(1): p. 46-54.
76. Davis, G.A., T.M. O'Neil-Pirozzi, and M. Coon, Referential cohesion and logical
coherence of narration after right hemisphere stroke. Brain and Language, 1997. 56(2):
p. 183-210.
77. Bartels-Tobin, L.R. and J.J. Hinckley, Cognition and discourse production in right
hemisphere disorder. Journal of Neurolinguistics, 2005. 18(6): p. 461-477.
78. Rice, G.E., P. Hoffman, and M.A. Lambon Ralph, Graded specialization within and
between the anterior temporal lobes. Annals of the New York Academy of Sciences,
2015. 1359(1): p. 84-97.
79. Rice, G.E., M.A. Lambon Ralph, and P. Hoffman, The roles of left versus right anterior
temporal lobes in conceptual knowledge: an ALE meta-analysis of 97 functional
neuroimaging studies. Cerebral Cortex, 2015. 25(11): p. 4374-4391.
80. Pereira, F., S. Gershman, S. Ritter, and M. Botvinick, A comparative evaluation of off-
the-shelf distributed semantic representations for modelling behavioural data. Cognitive
neuropsychology, 2016. 33(3-4): p. 175-190.
81. Bruffaerts, R., S. De Deyne, K. Meersmans, A.G. Liuzzi, G. Storms, and R.
Vandenberghe, Redefining the resolution of semantic knowledge in the brain: advances
made by the introduction of models of semantics in neuroimaging. Neuroscience &
Biobehavioral Reviews, 2019. 103: p. 3-13.

35
bioRxiv preprint doi: https://doi.org/10.1101/2022.10.15.512349; this version posted October 17, 2022. The copyright holder for this
preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.

82. Glenberg, A.M. and D.A. Robertson, Symbol grounding and meaning: A comparison of
high-dimensional and embodied theories of meaning. Journal of Memory and Language,
2000. 43(3): p. 379-401.
83. Glenberg, A.M. and S. Mehta, Constraint on covariation: it’s not meaning. Italian Journal
of Linguistics. Alessandro Lenci (guest editor): From Context to Meaning:
Distributional Models of the Lexicon in Linguistics and Cognitive Science, 2008.
20(1): p. 241-264.
84. Andrews, M., G. Vigliocco, and D. Vinson, Integrating Experiential and Distributional
Data to Learn Semantic Representations. Psychological Review, 2009. 116(3): p. 463-
498.
85. Hoffman, P., J.L. McClelland, and M.A. Lambon Ralph, Concepts, control and context: A
connectionist account of normal and disordered semantic cognition. Psychological
Review, 2018. 125: p. 293-328.
86. Davis, C.P. and E. Yee, Building semantic memory from embodied and distributional
language experience. Wiley Interdisciplinary Reviews: Cognitive Science, 2021. 12(5):
p. e1555.
87. Fernandino, L., J.-Q. Tong, L.L. Conant, C.J. Humphries, and J.R. Binder, Decoding the
information structure underlying the neural representation of concepts. Proceedings of
the National Academy of Sciences, 2022. 119(6): p. e2108091119.
88. Yee, E. and S.L. Thompson-Schill, Putting concepts into context. Psychonomic bulletin
& review, 2016. 23(4): p. 1015-1027.
89. Oldfield, R.C., The assessment and analysis of handedness: the Edinburgh inventory.
Neuropsychologia, 1971. 9(1): p. 97-113.
90. Wu, W., M. Morales, T. Patel, M.J. Pickering, and P. Hoffman, Modulation of brain
activity by psycholinguistic information during naturalistic speech comprehension and
production. Cortex, 2022. 155: p. 287-306.
91. Hoffman, P., Reductions in prefrontal activation predict off-topic utterances during speech
production. Nature communications, 2019. 10(1): p. 515.
92. Blank, S.C., S.K. Scott, K. Murphy, E. Warburton, and R.J. Wise, Speech production:
Wernicke, Broca and beyond. Brain, 2002. 125(8): p. 1829-1838.

36
bioRxiv preprint doi: https://doi.org/10.1101/2022.10.15.512349; this version posted October 17, 2022. The copyright holder for this
preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.

93. Hoffman, P., E. Loginova, and A. Russell, Poor coherence in older people's speech is
explained by impaired semantic and executive processes. eLife, 2018. 7: p. e38907.
94. Cusack, R., N. Cumming, D. Bor, D. Norris, and J. Lyzenga, Automated post‐hoc noise
cancellation tool for audio recordings acquired in an MRI scanner. Human brain mapping,
2005. 24(4): p. 299-304.
95. Kundu, P., V. Voon, P. Balchandani, M.V. Lombardo, B.A. Poser, and P.A. Bandettini,
Multi-echo fMRI: A review of applications in fMRI denoising and analysis of BOLD signals.
Neuroimage, 2017. 154: p. 59-80.
96. DuPre, E., T. Salo, R. Markello, P. Kundu, K. Whitaker, and D. Handwerker, ME-
ICA/tedana: 0.0.7. https://doi.org/10.5281/zenodo.3786890, 2019.
97. Ashburner, J., A fast diffeomorphic image registration algorithm. Neuroimage, 2007.
38(1): p. 95-113.
98. Hendriks, M.H., N. Daniels, F. Pegado, and H.P. Op de Beeck, The effect of spatial
smoothing on representational similarity in a simple motor paradigm. Frontiers in
neurology, 2017. 8: p. 222.
99. Gardumi, A., D. Ivanov, L. Hausfeld, G. Valente, E. Formisano, and K. Uludağ, The
effect of spatial resolution on decoding accuracy in fMRI multivariate pattern analysis.
Neuroimage, 2016. 132: p. 32-42.
100. Makris, N., J.M. Goldstein, D. Kennedy, S.M. Hodge, V.S. Caviness, S.V. Faraone, M.T.
Tsuang, and L.J. Seidman, Decreased volume of left and total anterior insular lobule in
schizophrenia. Schizophrenia research, 2006. 83(2): p. 155-171.
101. Shattuck, D.W., M. Mirza, V. Adisetiyo, C. Hojatkashani, G. Salamon, K.L. Narr, R.A.
Poldrack, R.M. Bilder, and A.W. Toga, Construction of a 3D probabilistic atlas of human
cortical structures. Neuroimage, 2008. 39(3): p. 1064-1080.
102. Oosterhof, N.N., A.C. Connolly, and J.V. Haxby, CoSMoMVPA: multi-modal
multivariate pattern analysis of neuroimaging data in Matlab/GNU Octave. Frontiers in
neuroinformatics, 2016. 10: p. 27.
103. Mumford, J.A., T. Davis, and R.A. Poldrack, The impact of study design on pattern
estimation for single-trial multivariate pattern analysis. Neuroimage, 2014. 103: p. 130-
138.

37
bioRxiv preprint doi: https://doi.org/10.1101/2022.10.15.512349; this version posted October 17, 2022. The copyright holder for this
preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.

104. Diedrichsen, J. and N. Kriegeskorte, Representational models: A common framework for

understanding encoding, pattern-component, and representational-similarity analysis.
PLoS computational biology, 2017. 13(4): p. e1005508.
105. Stelzer, J., Y. Chen, and R. Turner, Statistical inference and multiple testing correction
in classification-based multi-voxel pattern analysis (MVPA): random permutations and
cluster size control. Neuroimage, 2013. 65: p. 69-82.
106. Smith, S.M. and T.E. Nichols, Threshold-free cluster enhancement: addressing problems
of smoothing, threshold dependence and localisation in cluster inference. Neuroimage,
2009. 44(1): p. 83-98.

38