Brain & Language 164 (2017) 5362

Contents lists available at ScienceDirect

Brain & Language

journal homepage: www.elsevier.com/locate/b&l

Cortico-striatal language pathways dynamically adjust for syntactic

complexity: A computational study
Krisztina Szaliszny a,f,,1, David Silverstein b,1, Marc Teichmann c,d, Hugues Duffau e, Anja Smits g,h
a
Department of Neuroscience, Psychiatry, University Hospital, Uppsala University, 751 85 Uppsala, Sweden
b
Department of Computational Science and Technology, KTH Royal Institute of Technology; Stockholm Brain Institute, Karolinska Institutet, Stockholm, Sweden
c
Department of Neurology, Institut de la mmoire et de la maladie dAlzheimer, Centre de Rfrence Dmences Rares, Hopital de la Piti-Salpetrire, AP-HP, Paris, France
d
Institut du Cerveau et de la Moelle Epinire (ICM), ICM-INSERM 1127, FrontLab, Paris, France
e
Department of Neurosurgery, Gui de Chauliac Hospital, University of Montpellier, Institute for Neurosciences of Montpellier, Montpellier, France
f
Computational Neuroscience Group, Wigner Research Institute, Hungarian Academy of Sciences, P.O. Box 49, Budapest, Hungary
g
Department of Neuroscience, Neurology, Uppsala University, Uppsala, Sweden
h
Department of Clinical Neurosciences and Rehabilitation, Sahlgrenska Academy, Gothenburg, Institute of Neurosciences and Physiology, University of Gothenburg,
Gothenburg, Sweden

a r t i c l e i n f o a b s t r a c t

Article history: A growing body of literature supports a key role of fronto-striatal circuits in language perception. It is
Received 17 January 2016 now known that the striatum plays a role in engaging attentional resources and linguistic rule computa-
Revised 5 August 2016 tion while also serving phonological short-term memory capabilities. The ventral semantic and the dorsal
Accepted 14 August 2016
phonological stream dichotomy assumed for spoken language processing also seems to play a role in
Available online 25 October 2016
cortico-striatal perception. Based on recent studies that correlate deep Broca-striatal pathways with com-
plex syntax performance, we used a previously developed computational model of frontal-striatal syntax
Keywords:
circuits and hypothesized that different parallel language pathways may contribute to canonical and non-
Sentence comprehension
Striatum
canonical sentence comprehension separately. We modified and further analyzed a thematic role assign-
Cortico-striatal language network ment task and corresponding reservoir computing model of language circuits, as previously developed by
Reservoir computing Dominey and coworkers. We examined the models performance under various parameter regimes, by
influencing how fast the presented language input decays and altering the temporal dynamics of acti-
vated word representations. This enabled us to quantify canonical and non-canonical sentence compre-
hension abilities. The modeling results suggest that separate cortico-cortical and cortico-striatal circuits
may be recruited differently for processing syntactically more difficult and less complicated sentences.
Alternatively, a single circuit would need to dynamically and adaptively adjust to syntactic complexity.
2016 Published by Elsevier Inc.

1. Introduction
One of the most complex functions of the human brain is its
capacity to communicate by language (Medaglia, Lynall, &
Bassett, 2015). Exploring the neural basis of syntactic processing,
i.e. the unique ability to combinatorially assemble words to form
meaningful sentences, is one of the key challenges in cognitive
neuroscience. It is generally accepted that syntactical operations
are processed by a network of different cortical regions. The two
main cortical regions with a central role in syntax are the inferior-
posterior frontal cortex including Brocas area and the superior
Corresponding author at: Department of Neuroscience, Psychiatry, University
Hospital, Uppsala University, 751 85 Uppsala, Sweden.
E-mail address: krisztina.szalisznyo@neuro.uu.se (K. Szaliszny).
1
These authors contributed equally to this work.
temporal cortex (Friederici & Kotz, 2003; Grodzinsky & Santi,
2008; Pallier, Devauchelle, & Dehaene, 2011). However, subcortical
grey matter structures also affect syntax and there is accumulating
evidence that cortico-striatal pathways play a major role in syntac-
tic language processing.
The striatum as the major input nucleus to the basal ganglia
may impact language processing in several different ways
(Dominey & Inui, 2009; Dominey, Inui, & Hoen, 2009). First, its role
in non-language-specific functions such as attentional resources
may have a general impact on language performances. Second,
the striatum presumably subserves phonological short-term mem-
ory capacities which are necessary for accurate word perception
in phrasal contexts. This latter role may contribute to morpholog-
ical and syntactic rule application through maintaining intermedi-
ate language chunks such as morphemes and grammatical groups.
Evidence exists that the basal ganglia chunks the representations

http://dx.doi.org/10.1016/j.bandl.2016.08.005
0093-934X/ 2016 Published by Elsevier Inc.
54 K. Szaliszny et al. / Brain & Language 164 (2017) 5362
of motor and cognitive action sequences (Destrebecqz et al., 2005;
Kotz, Schwartze, & Schmidt-Kassow, 2009). Nigro-striatal circuits
signal the initiation or termination of each action sequence, which
occurs during sequence learning. Altering the function of the stri-
atal circuit disrupts the development of start/stop activity and
selectively impairs sequence learning (Badgaiyan, Fischman, &
Alpert, 2007; Jin & Costa, 2010; Yin, 2010). It has been suggested
that the striatum may also play a role in linguistic rule computa-
tion applying to particular kinds of morpho-syntactic rules
(Teichmann, Darcy, Bachoud-Levi, & Dupoux, 2009).
Human cortico-striatal projection tracing with diffusion tensor
imaging (Lehericy, Ducros, Krainik, et al., 2004) and anterograde
tracing (Wiesendanger, Clarke, Kraftsik, & Tardif, 2004) showed
that cortico-striatal connections in humans are organized in multi-
ple overlapping circuits (Alexander, DeLong, & Strick, 1986). It was
proposed that cortico-striatal projection neurons in Brodmann area
47 (BA 47) project to the head of the caudate, in a cortico-striato-
nigro-thalamo-cortical circuit. The BA47 cortical region has a role
in working memory for semantic features, thematic structure
(Friederici, 2002), the unification of individual semantic features
into an overall representation at the multi-word level (Turken &
Dronkers, 2011). It seems that there exists a grammatical struc-
ture circuit in the cortico-striatal system with distinct and segre-
gated paths, similar to the proposal of Ullman (2001, 2006).
Language disturbances related to cortico-striatal dysfunction are
commonly encountered in a variety of neurological disorders.
Patients with subcortical cerebrovascular damage, Parkinsons dis-
ease and Huntingtons disease have difficulties suppressing seman-
tically appropriate alternatives when trying to inflect novel verbs
(Nemeth et al., 2012). This pathology may result from impaired
function when the striatum is serving a non-language specific role
in late inhibitory processing (Chan, Ryan, & Bever, 2011).
A recent study demonstrated that in patients with primary pro-
gressive aphasia, syntactic processing depends primarily on dorsal
cortical and subcortical language tracts (Wilson et al., 2011). There
was a strong correlation between reduced fractional anisotropy in
the superior longitudinal fasciculus (including its arcuate compo-
nent) and deficits in syntactic comprehension. It was suggested
(Teichmann et al., 2009) that the ventral portions of the striatum
may subserve linguistic rule computations, whereas more dorsally
situated portions may underpin lexical operations (Teichmann &
Gaura, 2008). Such sub-portions of the striatum are supposed to
receive input from cortical areas which have been implicated in
rule application and lexical processing (Brocas area and posterior
temporal cortices, respectively). So there is some evidence for a
striatal division of labor, similar to what is seen in the cortical
dual-stream model (Ueno, Saito, Rogers, & Lambon, 2011). The dor-
sal striatums role seems to be tied to motor responses (actor) ver-
sus the more cognitive-learning role (critic) of the ventral striatum.
In the actor-critic architecture, the critic monitors and identifies
patterns, such as motor contingencies, while the actor alters
behaviors so as to improve the synchronization between environ-
mental circumstances and behavioral responses (Reiss et al., 2005).
Recently, Dominey and colleagues presented a cortico-striatal
reservoir computing model that reflects the recurrent fronto-
cortical networks and where the output is represented by plastic
cortico-striatal neurons. In this reservoir model, BA47 receives
recurrent on-line input on word categories during sentence pro-
cessing with plastic connections between striatum and cortex
(Hinaut & Dominey, 2013). The reservoir network consists of fixed
connections encoding the spatiotemporal structure of an input
sequence, while the connections to a readout layer are trained to
produce a desired output in response to input sequences. It was
shown that the recurrent reservoir successfully decodes grammat-
ical structure from sentences in real-time. In the present study, we
have adapted this model to further study cortico-striatal function
in sentence processing (Hinaut & Dominey, 2013).
Sentence comprehension is based on the integration of associa-
tive and symbolic information (Townsend & Bever, 2001). More
recently, it has been suggested that thematic role assignment for
canonical and non-canonical sentence comprehension may require
distinct spatio-temporal brain activity (del Rio et al., 2011). By using
voxel-based lesion symptom mapping (VLSM) and fiber tracking in
patients with fronto-striatal lesions, it was shown that perception of
canonical sentences involved the BA44/45 areas and extended cau-
dally to the underlying white matter, with no involvement of the
deep white matter regions or the caudate (Teichmann et al.,
2015). Non-canonical sentence perception, on the other hand, was
associated with a region extending in BA44/BA45/BA47, throughout
the intervening white matter, to the caudate head of the left stria-
tum (Teichmann et al., 2015). The cortico-cortical pathways showed
little overlap with the non-canonical syntax related VLSM clusters.
These findings extend current syntax network models and correlate
anatomically identified cortico-striatal pathways with complex
syntax performance (Lehericy, Ducros, Krainik, et al., 2004;
Lehericy, Ducros, Vd Moortele, et al., 2004; Teichmann et al., 2015).
In the present study, we elaborated on this concept by using a
computational model of fronto-striatal syntax circuits (Hinaut &
Dominey, 2013). We hypothesized that comprehension abilities
may differ between syntactically more simple and more complicated
sentences and that the processing of sentences with different syntac-
tic complexity may occur through contributions of separate cortico-
striatal circuits. We examined the thematic role assignment ability
of a reservoir network, where those model parameters were varied
which influence the property of the network to retain information
about the input for a certain time. Our study supports the observa-
tion that canonical and non-canonical sentence types requires differ-
ent memory capacity for optimal processing. We propose that this
might be via spatially distinct pathways (del Rio et al., 2011) or that
the same network is capable of dynamically adjusting to the syntac-
tic complexity. This second working hypothesis would be consistent
with a more continuous, gradient-wise processing of syntactic com-
plexity, with varying lengths of syntactic dependence.
2. Methods
2.1. Design
To illustrate the aim and the clinical implications of our compu-
tational study, we first present a selection of patient data from our
recent publication (Teichmann et al., 2015). The rationale of that
study was to explore whether the syntax network, in addition to
the known cortico-cortical pathways, also involves a fronto-
striatal pathway that accounts for the impact of the striatum on
syntactic processing. The clinical test results of phrasal syntax task
support the existence of a deep Broca-caudate pathway dedicated
to particular aspects of syntax. We describe this here in detail
(Teichmann et al., 2015).
2.2. Clinical case illustrations
Assessment of phrasal syntactic capacities was performed as
part of our previous study (Teichmann et al., 2015). In summary,
12 patients were enrolled with lesions in the left frontal lobe
and/or left striatum (Fig. 1 and Fig. 2) due to a WHO grade II glioma
(Teichmann et al., 2015). Tumor resection was performed at least
nine months prior to testing. Patients were right-handed, native
French speakers and without previous history of neurological or
psychiatric disease. In addition, 15 healthy controls matched with
 K. Szaliszny et al. / Brain & Language 164 (2017) 5362
patients by age, gender, handedness and number of years of educa-
tion were tested. Syntactic capacity of sentence comprehension
was evaluated using sentence-picture pairs consisting of non-
canonical (passives, object-relatives) or canonical sentences
(actives, subject-relatives). Sentences were either plausable (The
girl waters the flower) or non-plausable (The flower waters the
girl). Canonical and non-canonical sentences were paired with a
specific picture depicting the plausable and non-plausable versions
of each sentence. Participants were asked to respond with YES or
NO, depending on whether the auditorily presented sentence and
simultaneously visually presented picture was correctly matched.
A total of 32 sentence-picture pairs were presented, which was
obtained by crossing three parameters (canonicity, structure and
plausibility) with the same number of frequency-matched content
words. VSLM was performed followed by fiber tracking of the
cortico-cortical fiber systems (arcuate, uncinate fasciculus, inferior
fronto-occipital fasciculus and aslant) and of the putative
Broca-striatum tract. Significant VSLM clusters were overlapped
with probability maps of the fiber tracts, and the lesion load of
each tract was determined for patients and controls. In the final
step, the relationship between test results for syntactic scores
and lesion load of the different tracts on MRI was quantified.
2.3. Model description
We have used a modeling framework of reservoir computing
model developed previously by Hinaut and Dominey (2013).
Fig. 1. Axial MRI slices with lesion segmentation images of 6 (16 of 12) right-handed patients with lesions in the left frontal lobe and/or left striatum who showed reduced
non-canonical phrasal syntax performance (Table 1). Slices are shown in accordance with neurological convention (left is left).
55
Syntactic comprehension was tested after presenting training sen-
tences (corpus with different grammatical constructions) to the
model. The model generates the coded meaning, the thematic role
(agent, object, recipient) in the sentence for each semantic word.
The reservoir neural network, which represents the BA 44/45/47
cortical area, is composed of leaky integrator neurons (Eq. (1)).
Variable xt represents the reservoir state; ut is the input; dt is
the time step; s is the time constant (leaking rate); and f N is
the hyperbolic tangent (tanh) activation function, identically as
described previously (Hinaut & Dominey, 2013). W in is the connec-
tion weight matrix from inputs to the reservoir. W res is the recur-
rent connections within the cortical reservoir network. W out is
the output weight matrix. The activation time (AT) is the time dur-
ing which each word is presented. The time step dt equals the
reciprocal of AT (i.e. 1=AT) (Hinaut & Dominey, 2013).
 
dt dt
xt 1 1  xt f W in ut W res xt 1
s s
yt W out 1; xt 2
yt is the output (striatal) activity (Eq. (2)). In order to train the
connection weights W out between the reservoir and the readout
(striatum), ridge regression was used. The readout activity repre-
sents the coded meaning of the input sentence. A winner-take-all
algorithm determines which of the 3 possible coded meanings
(agent, object, recipient) is activated. The network performance
was quantified by the meaning error, as the percentage of all
56 K. Szaliszny et al. / Brain & Language 164 (2017) 5362

Fig. 2. Axial MRI slices with lesion segmentation images of 6 (712) right-handed patients with lesions in the left frontal lobe and/or left striatum who showed reduced
non-canonical phrasal syntax performance (Table 1). Slices are shown in accordance with neurological convention (left is left).

neurons being incorrect. The scale factor for the input weight
matrix W in (input scaling) was varied and the effect of parameter
change was examined (Hinaut & Dominey, 2013). The density of
the input connections (input connection probability) was also var-
ied and analyzed. Internal weights W res were chosen from a normal
distribution with mean 0 and standard deviation 1, with a connec-
tivity of 10%. This connectivity value was also altered. The spectral
radius is the largest absolute eigenvalue of the generated matrix
Wres. The spectral radius was rescaled to 1 in most of the simula-
tions, but was also altered and analyzed. The time constant (s)
was also altered and the effect was analyzed. A corpus containing
45 sentence constructions was first used, containing both canonical
and non-canonical sentences, created by Hinaut and Dominey
(2013). The input signal of the incoming words is generated as an
activated square wave during AT. In this study the syntactic com-
plexity (canonicity) was defined by subject-first and subject-
embedded (canonical) or object-first and object-embedded (non-
canonical) clauses (Hinaut & Dominey, 2013; Meltzer, McArdle,
Schafer, & Braun, 2010). Other measures and aspects of syntactic
complexity (e.g. passive, active) were not examined in detail
(Rogalsky, Matchin, & Hickok, 2008).
Subject embedded sentence example:
The grandmother who saw the postman is smiling.
Object embedded sentence example:
The grandmother whom the postman saw is smiling.
In sentences containing an object-embedded relative clause, the
object is placed before the subject of the sentence (Meltzer et al.,
2010). We divided the corpus into subject-first (canonical) and
object-first (non-canonical parts) (one containing 21 sentences,
the other containing 24 sentences). Ridge regression was used to
train the readout weights to minimize the mean squared error
between the target output and the projection of the reservoir state
through the trained readout weights. For cross-validation, one con-
struction was left out for testing and the model was trained on the
remaining sentences (Hinaut & Dominey, 2013). We varied those
parameters that influence the memory capacity of the network.
According to the literature, a longer spectral radius enables a
longer short-term memory of the input (Lukosevicius, 2012). The
input scaling also influences the relative effect of the current input
on reservoir state as opposed to the history (Lukosevicius, 2012).
The leaking rate of the reservoir units can be matched to the speed
of the dynamics of the input, with a longer leaking rate better
matched to slower changing inputs (Lukosevicius, 2012).
In the final step, the original corpus of 45 sentences was
extended to 462 sentences (Hinaut & Dominey, 2013). As previ-
ously demonstrated with this extended corpus, the network can
generalize on the grammatical structure of the training set without
overfitting the input data Hinaut and Dominey (2013). We defined
subgroups in the extended sentence corpus, containing 3, 4 or 5
nouns. It should be noted, that one noun can play more than one
role in a sentence, and it was the number of assigned roles, and
 K. Szaliszny et al. / Brain & Language 164 (2017) 5362 57

not the number of nouns per se which were quantified. Model per- Table 2
formance was analyzed for these subgroups when varying the time Test results of phrasal syntax for healthy controls (n = 15) (originally described by
Teichmann et al. (2015)). Values are presented as percentage of correct answers.
constant of the reservoir neural elements. We then analyzed the
effect of input scaling while keeping the number of nouns constant CONTROLS Canonical Non-canonical
(same length with 4 nouns), when training on either a subject control1 100 100
first (canonical) or an object first (non-canonical) extended control2 93.75 93.75
corpus. control3 100 100
control4 100 93.75
control5 100 100
3. Results control6 100 100
control7 100 93.75
control8 93.75 87.5
3.1. Clinical case illustrations control9 100 100
control10 100 100
The results of the phrasal syntax tests for 12 patients (Fig. 1 and control11 100 93.75
control12 93.75 93.75
Fig. 2) and 15 healthy controls are presented in Table 1 (patients)
control13 100 93.75
and Table 2 (controls). As shown, non-canonical sentence perfor- control14 100 100
mance differed significantly between patients (77.6%, standard control15 100 100
error 3.1) and healthy controls (96.7%, standard error 1.1). Within mean value 98.8 96.7
the group of patients, there was a statistically significant correla-
tion between non-canonical syntactic capacity and lesion load
values of the Broca-caudate tracts (Teichmann et al., 2015). In con-
trast, no significant correlations were found between performance
of non-canonical sentences and lesion load of the cortico-cortical
tracts (arcuate, uncinate fasciculus, inferior fronto-occipital fasci-
culus and aslant). Short-term memory capacities were only slightly
lower for patients compared to healthy controls and did not
account for the observed differences in non-canonical syntax
capacities between the two groups.

3.2. Computational modeling

3.2.1. Varying the time constant of the reservoir neural elements

In the first set of simulations we analyzed the effects of chang-
ing the time constant (leaking rate) of the cortical reservoir neural
elements on the network performance. We assumed a uniform Fig. 3. Effect of the time constant of the neural elements (leaking rate) on the
network performance, while distributing the corpus to canonical and non-canonical
leaking rate in our simulations for all the reservoir units. For this
subsets.
experiment, we modified the corpus and distributed it into syntac-
tically more simple (canonical) and syntactically more complicated
(non-canonical) sentences. We kept all parameters the same and necessary for equivalent processing of syntactically more compli-
varied only the time constant (both input matrix density and inter- cated non-canonical sentences.
nal matrix density were 0.1, number of neurons 300, spectral
radius 1, activation time AT = 20). On canonical sentence process-
3.2.2. Sensitivity analysis of the spectral radius
ing we found that the longer the time constant of the neural ele-
We then chose different time constants (leaking rates) for the
ments, the lower the meaning error and the better the resulting
canonical and non-canonical networks, obtained according to the
performance. In the case of the non-canonical corpus, the meaning
results from the first set of simulations (Fig. 3), in order to have
error was always higher than in the case of canonical sentences for
approximately similar network performances (Fig. 4), with a
the same time constant value (Fig. 3). Our simulation results
predict that longer time constants of the neural elements are

Table 1
Test results of phrasal syntax for patients (n = 12) (originally described by Teichmann
et al. (2015)). Values are presented as percentage of correct answers.

Patient Canonical Non-canonical

patient1 87.5 68.75
patient2 87.5 56.25
patient3 100 100
patient4 87.5 87.5
patient5 93.75 93.75
patient6 100 68.75
patient7 93.75 62.5
patient8 87.5 68.75
patient9 100 100
patient10 93.75 68.75
patient11 100 93.75 Fig. 4. Effect of spectral radius on the network performance, while distributing the
patient12 93.75 62.5 corpus to canonical and non-canonical subsets. We used a time constants of 15 in
mean value 93.8 77.6 the case of canonical and 30 in the case of non-canonical corpus (see Fig. 3). The
spectral radius between 0.5 and 10 was examined.
58 K. Szaliszny et al. / Brain & Language 164 (2017) 5362
Fig. 5. Effect of spectral radius on the network performance, while distributing the
corpus to canonical and non-canonical subsets. We used a time constant of 15 in the
case of canonical and 30 in the case of non-canonical corpus (see Fig. 3). The
spectral radius interval between 0.5 and 2 was examined.
spectral radius of 1. The spectral radius influences how fast the
input dies out in a reservoir over time (Jaeger, Lukosevicius,
Popovici, & Siewert, 2007). With higher spectral radius values, a
longer input history is processed by the reservoir elements. In
the case of lower values, the performance of the network depends
more on the recent input history. The effect of spectral radius
change was analyzed for both canonical and non-canonical perfor-
mance. We found that when using a higher spectral radius, the net-
work trained with non-canonical sentences had lower meaning
error values (Fig. 4). However, with a decreasing spectral radius,
the performance of the network trained with canonical sentences
improved, with lower meaning error values (Fig. 5). Both input
matrix density and internal matrix density were 0.1, input scaling
was chosen as 0.75, and number of neurons 300, activation time
AT = 20.
3.2.3. Changing the input matrix density and the number of elements
In the next set of simulations we altered the sparseness of the
input connection weight matrix and the number of neural ele-
ments in the reservoir. The input weight matrix density corre-
sponds to the cortico-cortical connections from other dorsal
cortical language tracts (middle temporal gyrus (MTG), MTG via
the inferior fronto-occipital fasciculus (IFOF), superior temporal
sulcus (STS/BA39), anterior superior temporal gyrus STG/BA22
Fig. 6. Effect of input matrix density and the number neurons in the reservoir on meaning error in case of a network trained on canonical corpus (A) and on non-canonical
corpus (B).
(Turken & Dronkers, 2011)), Wilson et al. (2011) to the BA
44/45/47. The neural correlates correspond to both the input
sparseness which depends on input from other dorsal language
tract cortical regions as well as white matter properties (Wilson
et al., 2011). According to the literature, the cortical part of syntac-
tic processing depends mostly on the dorsal language tracts
(Wilson et al., 2011). On simulations, the input weight matrix den-
sity was a default value 0.1 (10%), equal to the value used in the
original modeling study (Hinaut & Dominey, 2013). Here we exam-
ined the network performance while varying the input sparseness
as well as the number of units in the reservoir and analyzed a net-
work trained on both a canonical (Fig. 6A) and a non-canonical cor-
pus (Fig. 6B). We found that the network performance, quantified
by measuring meaning error, increases (with decreasing error) as
the input weight matrix becomes more dense (at the parameter
regime with higher number of neurons) and improves with the
number of neurons in the reservoir. With increased connectivity
between the input cortical layer and the reservoir, the performance
increases. When reducing the input matrix density and increasing
the number of neurons in the reservoir, functional performance
can be maintained, due to a conserved number of input connec-
tions. In larger networks, the susceptibility from input connectivity
loss is lower, which was the case both for canonical and non-
canonical networks (Fig. 6A and B). Increasing the number of neu-
rons improved the network performance when trained individually
both with canonical and non-canonical sentences and the two
dimensional parameter space of input matrix density and number
of elements did not differ qualitatively in these cases. Input scaling
0.75, spectral radius 1.0, tau of 6, activation time AT = 1 were used
for this set of simulations.
3.2.4. Altering the internal matrix density and the number of elements
In the 4th set of simulations we altered the sparseness of the
internal connection weight matrix as well as the number of neural
elements of the reservoir, in both the case of the canonical and
non-canonical trained networks (Fig. 7A and B). The connection
weight matrix corresponds to the cortico-cortical connections
within the BA 44/45/47. The weight matrix sparseness had a
default value of 0.1 in the rest of the simulations, (10%), equal to
the value used in the original modeling study (Hinaut &
Dominey, 2013). Here we examined the network performance
while varying the internal weight matrix sparseness together with
the number of units in the reservoir. We found that the internal
 K. Szaliszny et al. / Brain & Language 164 (2017) 5362 59

Fig. 7. Effect of internal matrix sparseness and the number neural elements in the reservoir on meaning error in the case of a network trained on a canonical corpus (A) and
on a non-canonical corpus (B).

weight matrix sparseness has no significant effect on the network
performance. This was the case for both canonical and non-
canonical sentences (Fig. 7A and B). It was pointed out earlier by
other groups that sparser connectivity pushes the networks toward
more stable regimes and has the tendency to stabilize the reser-
voirs (Boedecker, Obst, Lizier, Mayer, & Asada, 2012). We did not
observe similar effects; the network performance quantified by
meaning error was not significantly dependent on the sparseness
of reservoir connectivity. Increasing the number of neurons
improved the performance both in the case of canonical and
non-canonical networks and the two dimensional parameter space
did not differ significantly on the meaning error in these two cases.
Input scaling 0.75, spectral radius 1, input matrix density 0.1, tau 6
and activation time AT = 1 were used.
3.2.5. Sensitivity analysis of input scaling and input matrix density
We have again used 2 networks with different corpora, one with
canonical and the other with non-canonical sentences. We chan-
ged the input matrix density and varied the input scaling, which
represents the value of the elements in the input matrix. In both
canonical and non-canonical cases, generally the meaning error
increased with a more dense input matrix and with higher input
scaling (Fig. 8A and B). We propose that with an increasing input
matrix density, a decreasing value for input scaling can partially
compensate (Fig. 8A and B). According to the literature, input scal-
ing influences the relative effect of the current input (Lukosevicius,
2012). When input scaling is low, it decreases the influence of the
input signal and increases the influence of the recurrently con-
nected reservoir neurons, which may improve memory. However,
when higher, it can lead to a loss of memory patterns
(Lukosevicius, 2012), thus higher values for input scaling can
decrease the temporal memory of the reservoir and the inner
states will be completely driven by the actual signal (Petrenas,
Marozas, Lukosevicius, & Sakalauskas, 2012, chap. 2). We observed
that with low input scaling values, and low input matrix density
values the model with a canonical training corpus performed
slightly better in absolute terms than with the non-canonical case
(Fig. 8A versus B). In the non-canonical case we observed a shift
towards higher input scaling values for optimal performance
(Fig. 8B) (Fig. 8A versus B). A tau of 15, number of neurons 300,
spectral radius of 1, activation time AT = 1 were used for this set
of simulations.

Fig. 8. Effect of the input scaling and input matrix density on the performance of the network, quantified by meaning error in a network which was trained on a canonical (A)
and a non-canonical corpus (B).
60 K. Szaliszny et al. / Brain & Language 164 (2017) 5362

3.2.6. Sensitivity analysis of the time constant while also altering the
length of the input sentences
While varying the time constant of the reservoir neural ele-
ments, we simultaneously varied the length of the input sentences,
by choosing sentences with a specific number of nouns. For this set
of simulations we used the extended corpus. We found that in the
case of shorter sentences the performance of the network is better,
with all time constant values. The simulation results predict that
longer time constants of the neural elements are necessary for
equivalent processing of longer sentences (more nouns) than
shorter ones, but only in a quite narrow parameter range, when
the time constant was very low. It should be noted that the number
of sentences in the training corpora with different numbers of
nouns were not the same, which might have influenced these
results. We then analyzed the effect of input scaling while keeping
the number of nouns constant (4 nouns), training on subject-first
(canonical) and object-first (non-canonical) extended corpus. We
found that the performance improved on both object-first and
subject-first trained networks, as the input scaling decreased up
to a point, before stimulation became perhaps too small to drive
the network. Both input matrix density and internal matrix density
were 0.1, number of neurons 300, spectral radius 1.

Fig. 9. Schematic figure of the cortico-striatal pathways which are significant to canonical and non-canonical sentence perception. The yellow part of the figure illustrates
cortical parts of the language pathways, while the blue half represents striatal parts. Red lines represent cortico-cortical pathways while blue lines illustrate cortico-striatal
connections. The dashed (red and blue) line indicates a suggested recruitment of cortico-striatal pathways with increased syntactic load (Teichmann et al., 2015). Clinical
evidence suggests that dependent on syntactic complexity load cortico-cortical and cortico-striatal pathways are recruited for canonical and non-canonical sentence
conprehension (Teichmann et al., 2015) (Fig. A). Alternatively, the same cortico-striatal network is capable of dynamically adjusting to syntactic complexity and this is only
partially explained by changing working memory load (Fig. B).
 K. Szaliszny et al. / Brain & Language 164 (2017) 5362
4. Discussion
There is accumulating evidence for a genuine role of the striatum
rule-application (De-Diego-Balaguer et al., 2008; Teichmann et al.,
2015). Such a role in syntactic processing requires the online detec-
tion of distant dependencies and the creation of syntactic categories
by means of rule learning (Futrell, Mahowald, & Gibson, 2015;
Sambin et al., 2012). Here we used a computational model to com-
plement recent clinical data showing that Broca-striatal pathways
subtend complex syntax with non-canonical sentence comprehen-
sion (Teichmann et al., 2015). Our simulation results are consistent
with this concept and propose that the neural substrates for the per-
ception of sentences with various syntactic complexities are differ-
ent, utilizing varying memory capacity of the network (Fig. 9). An
alternative possibility is that the Broca-striatal circuit is capable of
dynamic changes, dependent on the syntactic input load (Jaeger
et al., 2007; Makuuchi, Grodzinsky, Amunts, Santi, & Friederici,
2013).
Likewise, two associated and functionally coupled pathways
involving the cortico-striatal system were found for acoustic signal
perception (Geiser, Notter, & Gabrieli, 2012). Participants in this
study listening to either periodic or non-periodic sequences of
tones, performed significantly better on intensity discrimination
during periodic than non-periodic tone sequences (Geiser et al.,
2012). It was shown that activation in the putamen was stronger
for periodic than for non-periodic sequences, whereas for non-
periodic sequences the activation in the primary and secondary
auditory cortices was stronger. The authors concluded that there
is a cortico-striatal mechanism that enhances auditory perception
through temporal context processing (Geiser et al., 2012). It is pos-
sible that the putamen serves as a neural substrate for participat-
ing in sentence processing (Geiser et al., 2012). Interestingly,
Brocas aphasics could not comprehend sentences that required
syntactic analysis but had good single word comprehension
(Caramazza & Zurif, 1976). These patients were able to understand
complex sentences when semantic or pragmatic cues were avail-
able. However, their comprehension of complex sentences
dropped to chance when forced to rely solely on grammatical
information. It has also been demonstrated that patients with
Parkinsons disease have increased difficulty processing
non-canonical sentences, which has been attributed to impaired
thematic role assignment in case of syntactically complex
sentences (Arnott et al., 2011; Macdonald & Monchi, 2011;
Tremblay et al., 2010).
It is intriguing that syntactically more complicated non-
canonical sentences can have equivalent processing with simpler
canoninal sentences when longer time constants were used. Non-
canonical sentences were on average 13 percent longer in the sim-
pler (shorter) corpus than the canonical sentences. So far there is
only indirect support for a spatiotemporal organization of
cortico-striatal language networks. For example, magnetoen-
cephalography was used to track cortical activity during sentence
comprehension in Spanish speakers reading subject-first and
non-canonical ambiguous object-first cleft sentences (del Rio
et al., 2011). The authors found a time-modulation of a medial
frontal network that adapted performance to task demands. It
has also been shown that during early phases of sequential learn-
ing, changes occur significantly earlier in the striatum (caudate)
than in the cortex (Pasupathy & Miller, 2005).
The complex interplay between syntactic complexity and work-
ing memory and the neurophysiological implementation of this
interaction have been extensively studied (Friederici, 2011;
Makuuchi et al., 2013). We examined the thematic role assignment
ability of a reservoir network, and varied those model parameters
that influence the property of the network to retain input informa-
61
tion for a certain period of time. To mimic the memory contribu-
tion to the sentence comprehension ability, we tested the effect
of increasing sentence length by varying the number of nouns in
the input sentences (Teichmann et al., 2015). We found that the
network trained on a syntactically more complex corpus per-
formed better in tasks requiring longer memory, if parameters
had been adjusted for this.
The group of patients with fronto-striatal damage reported in
this study was assessed for memory deficits to exclude interaction
between performance of sentences and short-term memory
(Teichmann et al., 2015). Digit span assessment showed only
slightly lower short-term memory measurements between
patients and controls that also did not show the poor performance
with non-canonical sentence comprehension (Teichmann et al.,
2015). The modeling results presented here are consistent with
the clinical data, and show that the differences in thematic role
assignment performance between syntactically less and more
complex sentences are only partially explained by differences in
working memory (Friederici, 2011; Makuuchi et al., 2013). It is
known that frontal-basal ganglia circuits play a key role in proce-
dural memory. Procedural memory systems are thought to be
essential for cognitive and motor skills/habits via learning rules
and sequences (Ullman & Pierpont, 2005). We observed that the
time constant (tau), input scaling and spectral radius all had an
impact on the memory of the network and had higher impact on
non-canonical syntactically more complicated sentences. It is pos-
sible, that although short-term memory measurements do not
reflect the quantified function loss with cortico-striatal damage
(Teichmann et al., 2015), procedural memory function may more
closely mirror the observed syntax perception decline, in corre-
spondence with the Procedural Deficit Hypothesis (Ullman, 2004;
Ullman & Pierpont, 2005).
In conclusion, we used a previously developed cortico-striatal
computational model (Hinaut & Dominey, 2013) and found distinc-
tive comprehension abilities while separating canonical and more
complicated non-canonical cortico-striatal processing, by applying
different parameters for the two different tasks. Higher perfor-
mance on non-canonical sentence comprehension was observed
with increased network memory function. Still, memory which
sustains too long might cause interference between continuous
sentences. When lowering the input matrix density and increasing
the number of neurons, network function can be maintained. These
results support the possible restoring effect of compensatory plas-
ticity, where the loss of input connections can be compensated for
by increased number of neural elements or increased weight of the
remaining connections (Szalisznyo, Silverstein, Duffau, & Smits,
2013). The model indirectly predicts that longer memory capabili-
ties of the reservoir network is beneficial for processing syntacti-
cally more complex non-canonical sentences, suggesting a
complex recruitment of cortico-striatal language networks during
thematic role assignment.
Acknowledgements
We wish to state that we have used and modified the open
access model code developed and published by Hinaut and
Dominey (2013). K.Sz. was supported by an ALF (unrestricted grant
from Uppsala County Council) grant from Uppsala University, Insti-
tute of Neuroscience. Some neural simulations were run at
UppMax computing center at Uppsala University.
References
Alexander, G., DeLong, M., & Strick, P. (1986). Parallel organization of functionally
segregated circuits linking basal ganglia and cortex. Annual Review of
Neuroscience, 9, 357381.
62 K. Szaliszny et al. / Brain & Language 164 (2017) 5362
Arnott, W., Copland, D., Chenery, H., Murdoch, B., Silburn, P., & Angwin, A. (2011).
The influence of dopamine on automatic and controlled semantic activation in
Parkinsons disease. Parkinsons Disease, 157072. http://dx.doi.org/10.4061/
2011/157072.
Badgaiyan, R., Fischman, A., & Alpert, N. (2007). Striatal dopamine release in
sequential learning. Neuroimage, 38(3), 549556.
Boedecker, J., Obst, O., Lizier, J., Mayer, N., & Asada, M. (2012). Information
processing in echo state networks at the edge of chaos. Theory in Biosciences, 131
(3), 205213.
Caramazza, A., & Zurif, E. (1976). Dissociation of algorithmic and heuristic processes
in language comprehension: Evidence from aphasia. Brain and Language, 3,
572582.
Chan, S., Ryan, L., & Bever, T. (2011). Role of the striatum in language: Syntactic and
conceptual sequencing. Brain and Language, 125(3), 283294.
De-Diego-Balaguer, R., Couette, M., Dolbeau, G., Durr, A., Youssov, K., & Bachoud-
Levi, A. (2008). Striatal degeneration impairs language learning: Evidence from
Huntingtons disease. Brain, 131(Pt 11), 28702881.
del Rio, D., Maestu, F., Lopez-Higes, R., Moratti, S., Gutirrez, R., Maestu, C., & del
Pozo, F. (2011). Conflict and cognitive control during sentence comprehension:
Recruitment of a frontal network during the processing of Spanish object-first
sentences. Neuropsychologia, 49(3), 382391.
Destrebecqz, A., Peigneux, P., Laureys, S., Degueldre, C., DelFiore, G., Aerts, J., ...
Maquet, P. (2005). The neural correlates of implicit and explicit sequence
learning: Interacting networks revealed by the process dissociation procedure.
Learning and Memory, 12(5), 480490.
Dominey, P., & Inui, T. (2009). Cortico-striatal function in sentence
comprehension: Insights from neurophysiology and modeling. Cortex, 45(8),
10121018.
Dominey, P., Inui, T., & Hoen, M. (2009). Neural network processing of natural
language: II. Towards a unified model of cortico-striatal function in learning
sentence comprehension and non-linguistic sequencing. Brain and Language,
109(23), 8092.
Friederici, A. (2002). Towards a neural basis of auditory sentence processing. Trends
in Cognitive Science, 6(2), 7884.
Friederici, A. (2011). The brain basis of language processing: From structure to
function. Physiological Reviews, 91(4), 13571392.
Friederici, A., & Kotz, S. (2003). The brain basis of syntactic processes: Functional
imaging and lesion studies. Neuroimage, 20, 1:581:17.
Futrell, R., Mahowald, K., & Gibson, E. (2015). Large-scale evidence of dependency
length minimization in 37 languages. Proceedings of the National Academy of
Sciences of the United States of America, 112(33), 1033610341.
Geiser, E., Notter, M., & Gabrieli, J. (2012). A corticostriatal neural system enhances
auditory perception through temporal context processing. The Journal of
Neuroscience, 32(18), 61776182.
Grodzinsky, Y., & Santi, A. (2008). The battle for Brocas region. Trends in Cognitive
Sciences, 12(12), 474480.
Hinaut, X., & Dominey, P. (2013). Real-time parallel processing of grammatical
structure in the fronto-striatal system: A recurrent network simulation study
using reservoir computing. PLoS One, 8(2), e52946.
Jaeger, H., Lukosevicius, M., Popovici, D., & Siewert, U. (2007). Optimization and
applications of echo state networks with leaky-integrator neurons. Neural
Networks, 20(3), 335352.
Jin, X., & Costa, R. (2010). Start/stop signals emerge in nigrostriatal circuits during
sequence learning. Nature, 466(7305), 457462.
Kotz, S., Schwartze, M., & Schmidt-Kassow, M. (2009). Non-motor basal ganglia
functions: A review and proposal for a model of sensory predictability in
auditory language perception. Cortex, 45(8), 982990.
Lehericy, S., Ducros, M., Krainik, A., Francois, C., Van de Moortele, P., Ugurbil, K., &
Kim, D. (2004). 3-d diffusion tensor axonal tracking shows distinct SMA and
pre-SMA projections to the human striatum. Cerebral Cortex, 14(12),
13021309.
Lehericy, S., Ducros, M., Vd Moortele, P., Francois, C., Thivard, L., Poupon, C., ... Kim,
D. (2004). Diffusion tensor fiber tracking shows distinct corticostriatal circuits
in humans. Annals of Neurology, 55(4), 522529.
Lukosevicius, M. (2012). A practical guide to applying echo state networks. In: The
series Lecture Notes in Computer Science (Vol. 7700, pp. 659686).
Macdonald, P., & Monchi, O. (2011). Differential effects of dopaminergic therapies
on dorsal and ventral striatum in Parkinsons disease: Implications for cognitive
function. Parkinsons Disease, 6, 572743.
Makuuchi, M., Grodzinsky, Y., Amunts, K., Santi, A., & Friederici, A. (2013).
Processing noncanonical sentences in Brocas region: Reflections of movement
distance and type. Cerebral Cortex, 23(3), 694702.
Medaglia, J., Lynall, M., & Bassett, D. (2015). Cognitive network neuroscience. Journal
of Cognitive Neuroscience, 27(8), 14711491.
Meltzer, J., McArdle, J., Schafer, R., & Braun, A. (2010). Neural aspects of sentence
comprehension: Syntactic complexity, reversibility, and reanalysis. Cerebral
Cortex, 20(8), 18531864.
Nemeth, D., Dye, C., Sefcsik, T., Janacsek, K., Turi, Z., Londe, Z., ... Ullman, M. (2012).
Language deficits in pre-symptomatic Huntingtons disease: Evidence from
Hungarian. Brain and Language, 121(3), 248253.
Pallier, C., Devauchelle, A., & Dehaene, S. (2011). Cortical representation of the
constituent structure of sentences. Proceedings of the National Academy of
Sciences of the United States of America, 108(6), 25222527.
Pasupathy, A., & Miller, E. (2005). Different time courses of learning-related activity
in the prefrontal cortex and striatum. Nature, 433(7028), 873876.
Petrenas, A., Marozas, V., Lukosevicius, A., & Sakalauskas, A. (2012). Extraction of f-
waves in electrocardiograms using echo state network based nonlinear
adaptive filters. Digital Image and Signal Processing for Measurement Systems, 59.
Reiss, J., Campbell, D., Leslie, W., Paulus, M., Stroman, P., Polimeni, J., ... Sareen, J.
(2005). The role of the striatum in implicit learning: A functional magnetic
resonance imaging study. Neuroreport, 16(12), 12911295.
Rogalsky, C., Matchin, W., & Hickok, G. (2008). Brocas area, sentence
comprehension, and working memory: An fMRI study. Frontiers in Human
Neuroscience, 6(2), 14.
Sambin, S., Teichmann, M., deDiego Balaguer, R., Giavazzi, M., Sportiche, D.,
Schlenker, P., & Bachoud-Levi, A. (2012). The role of the striatum in sentence
processing: Disentangling syntax from working memory in Huntingtons
disease. Neuropsychologia, 50(11), 26252635.
Szalisznyo, K., Silverstein, D., Duffau, H., & Smits, A. (2013). Pathological neural
attractor dynamics in slowly growing gliomas supports an optimal time frame
for white matter plasticity. PLoS One, 26(8(7)).
Teichmann, M., Darcy, I., Bachoud-Levi, A., & Dupoux, E. (2009). The role of the
striatum in phonological processing: Evidence from early stages of Huntingtons
disease. Cortex, 45(7), 839849.
Teichmann, M., & Gaura, V. E. A. (2008). Language processing within the striatum:
Evidence from a pet correlation study in Huntingtons disease. Brain, 131(4),
10461056.
Teichmann, M., Rosso, C., Martini, J., Bloch, I., Brugieres, P., Duffau, H., ... Bachoud-
Levi, A. (2015). A cortical-subcortical syntax pathway linking Brocas area and
the striatum. Human Brain Mapping, 36(6), 22702283.
Townsend, D., & Bever, T. (2001). Sentence comprehension: The integration of habits
and rules. Cambridge, MA: MIT Press.
Tremblay, P., Bedard, M., Langlois, D., Blanchet, P., Lemay, M., & Parent, M. (2010).
Movement chunking during sequence learning is a dopamine-dependant
process: A study conducted in Parkinsons disease. Experimental Brain
Research, 205(3), 375385.
Turken, A., & Dronkers, N. (2011). The neural architecture of the language
comprehension network: Converging evidence from lesion and connectivity
analyses. Frontiers in Systems Neuroscience, 10, 10.3389.
Ueno, T., Saito, S., Rogers, T., & Lambon, M. (2011). Lichtheim 2: Synthesizing
aphasia and the neural basis of language in a neurocomputational model of the
dual dorsal-ventral language pathways. Neuron, 72(2), 385396.
Ullman, M. (2001). A neurocognitive perspective on language: The declarative/
procedural model. Nature Reviews Neuroscience, 2(10), 717726.
Ullman, M. (2004). Contributions of memory circuits to language: The declarative/
procedural model. Cognition, 92(12), 231270.
Ullman, M., & Pierpont, E. I. (2005). Specific language impairment is not specific to
language: The procedural deficit hypothesis. Cortex, 41(3), 399433.
Ullman, M. (2006). Is Brocas area part of a basal ganglia thalamocortical circuit?
Cortex, 42(4), 480485.
Wiesendanger, E., Clarke, S., Kraftsik, R., & Tardif, E. (2004). Topography of cortico-
striatal connections in man: Anatomical evidence for parallel organization.
European Journal of Neuroscience, 20(7), 19151922.
Wilson, S., Galantucci, S., Tartaglia, M., Rising, K., Patterson, D., Henry, M., ... Gorno-
Tempini, M. (2011). Syntactic processing depends on dorsal language tracts.
Neuron, 72(2), 397403.
Yin, H. (2010). The sensorimotor striatum is necessary for serial order learning.
Journal of Neuroscience, 3;30(44), 1471914723.