hierarchical functional connectivity between the core language system and the working memory system

www.sciencedirect.com

c o r t e x x x x ( 2 0 1 3 ) 1e8

Available online at

Journal homepage: www.elsevier.com/locate/cortex

Research report

Hierarchical functional connectivity between the corelanguage system and the working memory system

Michiru Makuuchi and Angela D. Friederici*

Department of Neuropsychology, Max Planck Institute for Human Cognitive and Brain Sciences, Leipzig, Germany

a r t i c l e i n f o

Article history:

Received 30 July 2012

Reviewed 19 October 2012

Revised 12 November 2012

Accepted 13 January 2013

Action editor Stefano Cappa

Published online xxx

Keywords:

Complex syntax

Working memory

Dynamic causal modeling

Language network

fMRI

* Corresponding author. Max Planck Institute1a, 04103 Leipzig, Germany.

E-mail address: [email protected] (A.D

Please cite this article in press as: Makuusystem and the working memory system

0010-9452/$ e see front matter ª 2013 Elsevhttp://dx.doi.org/10.1016/j.cortex.2013.01.007

a b s t r a c t

Language processing inevitably involves working memory (WM) operations, especially for

sentences with complex syntactic structures. Evidence has been provided for a neuro-

anatomical segregation between core syntactic processes and WM, but the dynamic rela-

tion between these systems still has to be explored. In the present functional magnetic

resonance imaging (fMRI) study, we investigated the network dynamics of regions involved

in WM operations which support sentence processing during reading, comparing a set of

dynamic causal models (DCM) with different assumptions about the underlying connec-

tional architecture. The DCMs incorporated the core language processing regions (pars

opercularis and middle temporal gyrus), WM related regions (inferior frontal sulcus and

intraparietal sulcus), and visual word form area (fusiform gyrus). The results indicate

a processing hierarchy from the visual to WM to core language systems, and moreover,

a clear increase of connectivity between WM regions and language regions as the pro-

cessing load increases for syntactically complex sentences.

ª 2013 Elsevier Ltd. All rights reserved.

1. Introduction finding was the existence of direct and indirect (relayed by the

The arcuate fasciculus (AF), a white matter fiber bundle con-

necting Broca’s area to Wernicke’s area, has been considered

to play a central role in language processing for more than

a century (Dejerine, 1901; Geschwind, 1965a, 1965b; Wernicke,

1874). Using diffusion tensor imaging (DTI) to enable the vis-

ualization of white matter fibers, a comparative study among

humans, chimpanzees, and macaques revealed that humans

possess the most developed AF (Rilling et al., 2008), which

appears to be consistent with the fact that only the humans

enjoy fully developed language.

After applying DTI to humans, the classical view of the AF

and its role in language processingwas revised. The first novel

for Human Cognitive and

. Friederici).

chi M, Friederici AD, , Hi, Cortex (2013), http://dx

ier Ltd. All rights reserved

inferior parietal lobe) dorsal connections between Broca’s and

Wernicke’s areas (Catani et al., 2005). The two dorsal con-

nections which also differ in their termination regions in the

prefrontal cortex appear to serve different functions. The

connection that terminates in the premotor cortex supports

sound-to-motor mapping (Hickok and Poeppel, 2007; Saur

et al., 2008) and the connection that terminates in the poste-

rior portion of Broca’s area subserves the processing of syn-

tactically complex sentences (Friederici, 2009; Brauer et al.,

2011; Wilson et al., 2011). The second eminent finding was

the reappraisal of a ventral pathway connecting Broca’s and

Wernicke’s areas via the extreme capsule fibers system (ECFS;

Makris et al., 2009; Petrides and Pandya, 2009) and its possible

Brain Sciences, Department of Neuropsychology, Stephanstraße

erarchical functional connectivity between the core language.doi.org/10.1016/j.cortex.2013.01.007

.

mailto:[email protected]

www.sciencedirect.com/science/journal/00109452

www.elsevier.com/locate/cortex

http://dx.doi.org/10.1016/j.cortex.2013.01.007



c o r t e x x x x ( 2 0 1 3 ) 1e82

role as part of the semantic processing system (Hickok and

Poeppel, 2007; Rolheiser et al., 2011; Saur et al., 2008; Weiller

et al., 2009).

More recently, studies revealed that both the dorsal and

the ventral pathways are involved in syntactic processing

(Friederici et al., 2006; Tyler et al., 2011). The ventral pathway

appears to support local combinational processes while the

dorsal pathway subserves the processing of complex syntax in

particular (Friederici et al., 2006) as indicated by devel-

opmental (Brauer et al., 2011) and patient studies (Wilson

et al., 2011). These latter studies showed that the processing

of syntactically complex sentences is deficient when the

dorsal pathway connecting to the posterior portion of Broca’s

area is not fully matured or lesioned by degenerative

processes.

The full merit of the anatomical connections will become

obvious when the effective connectivity among the language-

relevant brain regions during processing is identified. Effective

connectivity analysis of functional magnetic resonance im-

aging (fMRI) data, especially dynamic causal model (DCM)

(Friston et al., 2003; Stephan and Friston, 2010) can estimate

inter-regional relationships in a neural network, such as the

direction of the causal influence and its change due to

experimental interventions.

Here we investigated the dynamics of the neural network

supporting the processing of sentences with varying syntactic

complexity. Syntactic complexity was chosen as a focus of

interest sincemany prior studies on language processing have

reported activation increase in Broca’s area, in particular the

pars opercularis (PO) and middle and superior temporal re-

gions for syntactically complex object-first compared to

subject-first sentences (e.g., Stromswold et al., 1996;

Bornkessel et al., 2005; Santi and Grodzinsky, 2007; Caplan

et al., 2008; Friederici et al., 2009; Newman et al., 2010). In

object-first sentences, the object noun is dislocated from its

original position leaving a trace behind, thus creating a dis-

tance between the new and the original positionswhich needs

to be dealt with during processing. The processing of syntac-

tically complex sentences has long been considered to be an

interplay between working memory (WM) and syntactic pro-

cessing (Caplan and Waters, 1999; Fiebach et al., 2001; Cooke

et al., 2002; Novais-Santos et al., 2007; Santi and Grodzinsky,

2007; Makuuchi et al., 2009).

In this context, twoWM components have been discussed:

syntactic WM with its main focus located in the inferior

frontal gyrus (IFG) (Caplan andWaters, 1999) and phonological

WMwith its main focus located in the parietal cortex (Novais-

Santos et al., 2007). The syntactic WM is taken to be syntax-

specific (Caplan and Waters, 1999) whereas the phonological

WM is viewed to be involved in memory for word lists (Owen

et al., 2005) as well as for memory-demanding sentence

structures (Meyer et al., 2012). A recent fMRI experiment has

located the syntactic WM, i.e., the region directly related to

syntactic complexity in the inferior frontal sulcus (IFS)

(Makuuchi et al., 2009). This study had used embedded sen-

tence structures requiring the storage of multiple subject-

noun phrases of the different embedded sentences. Another

fMRI experiment investigated the storage of one subject or

object-noun phrase across a number intervening words in

a sentence and found the temporo-parietal cortex to activate

Please cite this article in press as: Makuuchi M, Friederici AD, , Hisystem and the working memory system, Cortex (2013), http://d

as a function of storage distance (Meyer et al., 2012). Relating

the localization to findings in the literature (Owen et al., 2005),

this storage component was identified as phonological WM.

In the present study, DCMwas applied to fMRI data froman

experiment (Makuuchi et al., 2013) varying syntactic com-

plexity (subject-first canonical structures vs object-first non-

canonical structures) and the concomitant WM demands.

The WM demands are resulted from the non-canonical

structures in which the object-noun phrase is encountered

early in the sentence but has to be kept in memory before the

subject-noun phrase and the verb are perceived. This may

recruit the phonological WM as the object-noun phrase has to

be stored in memory, but also the syntactic WM as structure-

building is required while holding the object-noun phrase. In

the original fMRI study object-first sentences which resulted

from different syntactic operation compared to subject-first

sentences led to increased activations in a number of brain

region: the PO and the IFS as well as the middle part of the

middle temporal gyrus (mMTG), and the intraparietal sulcus

(IPS) of the left hemisphere. Based on the functional data,

we included these regions in the DCM analysis. Activations

were also observed in the two right hemisphere regions, PO

and IPS, which are homolog areas to activated regions in the

left hemispheric network. Such right hemispheric homolog

activations are often reported in language studies without

discussing their specific function. They may reflect inter-

hemispheric coactivation mediated by the corpus callosum.

To keep the number of possible models at a reasonable level,

we decided to take only the left hemispheric network into

account. We further added the fusiform gyrus (FG) since it is

assumed to represent the visual word form area (Cohen et al.,

2000, 2002) and to act as the initial cerebral gate of the visual

linguistic information.With these regions we constructed and

estimated fifteen alternative DCMs.

2. Methods

The present study used data collected in a previous fMRI

study. The details of the experimental stimuli and procedure

of this study are found in the paper published elsewhere

(Makuuchi et al., 2013). For a comprehensive description of

stimuli and procedure see also below.

2.1. Participants

Twenty-two young, right-handed, healthy participants were

examined (eleven females). Handedness (mean 93.9, range

80e100) was assessed with the Edinburgh Inventory (Oldfield,

1971). Themean agewas 25.0 years old (range 20e33 years). All

participants were native German speakers. Reading span

was measured by a German version of the Daneman and

Carpenter (1980) reading span test (mean 4.0, range 3e5.5).

No participant had a history of neurological disorders. One

participant who had low performance was excluded from the

group analysis as in the previous report (Makuuchi et al.,

2013), and also excluded from the DCM analysis. The exper-

imental procedures were approved by the Research Ethics

Committees of the University of Leipzig. Written informed

consent was given by all participants.

erarchical functional connectivity between the core languagex.doi.org/10.1016/j.cortex.2013.01.007



c o r t e x x x x ( 2 0 1 3 ) 1e8 3

2.2. Stimuli and procedure

The stimuli were German sentences. Basic subject-first sen-

tences were transformed into object-first sentences created by

dislocating object-noun phrases from their original position

(called “gap”) to a new position (called “filler”), making a sen-

tence syntactically more complex. Dislocation was achieved by

linguistic operations called Movement and Scrambling and

varied the filler-gap distance. The basic subject-first sentences

were eleven words long and comprised three noun phrases,

a verb, and a temporal adverb at the end or the beginning of the

sentence: for example,DieserMann, glaube ich, zeigte demKind den

Onkel gestern Abend/This man, I think, showed the uncle to the

child yesterday night or Gestern Abend, glaube ich, zeigte dieser

Mann dem Kind den Onkel/Yesterday evening, I think, this man

showed the uncle to the child. Object-first sentences were con-

structed fromthese subject-first sentences bymoving either the

direct or the indirect object in front of the subject. This design

resulted in six variants of a basic sentence, which are all eleven

words long and expressed a similar meaning. Sentences were

presented visually, word by word, with a duration of 500 msec

and an inter-word-interval of 100 msec. Mean sentence onset

asynchronywas 11.2 sec. Forty distinct sentences per condition

were presented, resulting in a total of 240 trials. In 20% of the

trials, participants were required to perform a probe sentence

verification task. The probes assessed the thematic role

assignment of the different noun phrases (NPs) (whowas doing

what towhom).Half of theprobes restatedpart of the content of

the sentences previously perceived, while the other half did not

match the content of the sentence previously perceived. Par-

ticipants were requested to judge whether the probe expressed

the same content or not by pressing response buttons using the

index and middle finger of the right hand. The 20% trails with

probes were not included in the analysis of the hemodynamic

response.

2.3. Image acquisition

Functional MRI data were acquired with a whole-body 3 T

Magnetom TRIO (Siemens Medical Solution, Erlangen, Ger-

many) with a gradient-echo echo planar imaging (EPI)

sequence. The brain was covered with 2.5 mm thick 24 axial

images with .5 mm gaps [repetition time (TR) ¼ 1.6 sec, echo

time (TE) ¼ 30 msec, Flip angle ¼ 90�, field of view

(FOV) ¼ 19.2 � 19.2 cm2, 64 � 64 matrix]. The resulting

voxel size was 3 � 3 � 3 mm3. The slices were aligned to the

anterior commissure-posterior commissure (ACePC) plane

and placed to cover thewhole of Broca’s andWernicke’s areas.

The same slices were scanned with a T1-weighted modified

driven equilibrium fourier transform (MDEFT) sequence

(TR ¼ 1300 msec, TE ¼ 7.4 msec, 256 � 256 matrix) for the

spatial coregistration of EPI images to high-resolution ana-

tomical images. The participants had one session of fMRI

scanning with 1687 volumes per session in about 45 min.

Structural high-resolution images of the participants were

also collected on a different day with a three-dimensional

MDEFT sequence [TR ¼ 1300 msec, TE ¼ 3.93 msec, inversion

time (TI) ¼ 650 msec, Flip angle ¼ 10�, FOV ¼ 25.6 � 24 cm2,

256 � 240 matrix, sagittal 128 slices, 1 mm thick, 2 number of

excitation (NEX)].

Please cite this article in press as: Makuuchi M, Friederici AD, , Hisystem and the working memory system, Cortex (2013), http://dx

2.4. Image analysis

The first five volumes of the fMRI session were discarded to

eliminate magnetic saturation effects, and a total of 1682

volumes were used. The data analysis was carried out using

SPM8 (available at http://www.fil.ion.ucl.ac.uk/spm/) on Linux

PC workstations. Structural images were co-registered to

individuals’ functional images and normalized using the

Diffeomorphic Anatomical Registration using Exponentiated

Lie algebra (DARTEL) procedure (Ashburner, 2007), in which

individual structural images are segmented into gray and

white matter, and mean images of all individual’s images

serve as templates. The DARTEL normalization proceeds in six

steps with increasing spatial resolution, with the final step

being the linear transformation into the Montreal Neuro-

logical Institute (MNI) space. For functional data preprocess-

ing, EPI images were realigned to the first image and resliced

with correction for geometrical warping using a deformation

field map scan. Subsequently, the difference in the slice

acquisition time was corrected and all volumes were resliced

again. The first-level statistics were computed with the

unnormalized and unsmoothed images, and the resulting

statistical images were normalized using the DARTEL pa-

rameters with voxel resampling at 3 � 3 � 3 mm3 and

smoothing of 8 mm full width at half maximum (FWHM) and

fed into the second-level analysis of variance (ANOVA). We

also normalized individual structural and functional data

using the individual structural images normalized by DARTEL

as the target images.

2.5. Main effects

The main effect of DISTANCE (as t contrast) involves the left

IFG extending from PO [peak z ¼ 5.47, (�51 15 18)] to the IFS

[z¼ 5.50, (�36 6 33)], themMTG [z¼ 4.30 (�54�36�6)], and the

IPS [z ¼ 5.21, (�33 �51 36)]. The main effect of TYPE of dis-

location (as t contrast) was found only in occipital regions

bilaterally. For details see Makuuchi et al. (2013).

2.6. Dynamic causal modeling

To understand the causal relations between the activated re-

gions as revealed by the main effect of filler-gap DISTANCE in

the ANOVA, we performed a DCM analysis. Five volumes-of-

interest (VOI) in the left hemisphere were selected to con-

struct a systematic set of alternative DCMs. These VOIs are the

PO, IFS, mMTG, IPS, and FG (with FG acting as the cerebral gate

for visually presented words) (Fig. 1).

As a first step, we reanalyzed the functional data with new

design matrices that explicitly encoded the input and the two

modulatory effects, to follow the framework of the DCM

analysis. In this analysis, the experimental events were

regrouped into the input (all sentences that were not followed

by the probe sentence were encoded with ‘1’ and the others

with ‘0’). The modulatory effect of DISTANCE was encoded

categorically with ‘1’ for all sentences containing a dis-

location, i.e., object-first sentences and with ‘0’ for all sen-

tences with no dislocation, i.e., subject-first sentences. The

modulatory effect of TYPE of dislocation was encoded with ‘1’

for Scrambling andwith ‘0’ for Movement. The sentenceswith


http://www.fil.ion.ucl.ac.uk/spm/



Fig. 1 e Dynamic causal models. Top panel: The five VOIs used in the DCM analysis. The VOIs are shown at the mean

coordinates of the individual maxima on a template brain in theMNI space. Bottom panel: The fifteen DCMs. The PO, mMTG,

IFS, and IPS are schematically presented and labeled in model 1. These regions are assumed to be interconnected

bidirectionally, and the connections between the FG and other regions are allowed to have fifteen different patterns. The

Bayesian model comparison revealed that the best explanation (in terms of balance between fit and complexity) among the

fifteen alternative models was provided by model 13 (highlighted in yellow). In model 13, the FG is connected to the PO, IFS,

and IPS, but not to the mMTG.

c o r t e x x x x ( 2 0 1 3 ) 1e84

probes were modeled as a distinct condition to avoid com-

plication in estimation of blood oxygen level dependent

(BOLD) signal for sentences segregated from the probe, and six

motion parameters were also modeled as covariates of no

interest in the design matrices. The differences between the

present and the previous (Makuuchi et al., 2013) statistical

analyses are summarized as follows. The main six conditions

(S0, S1, S2, M1, M0, and M2) are modeled as a combination of

modulatory effects of DISTANCE and TYPE, namely [0,0], [1,0],

[1,0], [0,1], [1,1], and [1,1] [(DISTANCE, TYPE)]. Note that DIS-

TANCE 1 and 2 were collapsed. In terms of design matrix, the

previous model had six columns for conditions to represent

the six conditions, but the present model had two columns

which represent DISTANCE and TYPE, and one column for

input.

To find activation foci near to the group maxima found in

the ANOVA, we calculated the “effect of interest” F-contrast

(SPM{F}). This contrast is an F-test against the null hypothesis

that none of the input and the two modulatory effects causes

activation. Each subject’s local maxima were picked up in the

left hemisphere fromhis/her SPM{F}s as the closestmaxima to

the group maxima within the anatomical regions (Fig. 1). The

spherical VOIs of a 6 mm radius were created with the indi-

vidual local maxima as the centers. SPM{F}s were thresholded

with p < .05 without correction and the VOI time series data


were extracted as eigenvariates (the first principal compo-

nent) of the supra-threshold voxels’ time series data in the

VOI. The time series were adjusted for the effect of interest;

i.e., the contributions from the sentence with probe condition

and the motion parameters were subtracted from the time

series data. The group maxima coordinates, the mean of local

maxima across the participants, and its standard deviation

(SD) were as follows: the PO group maxima [�51 15 18], mean

[�53 12 20], SD [3.1 4.9 5.5]; IFS groupmaxima [�36 6 34], mean

[�39 7 31], SD [6.1 5.6 4.5]; mMTG group maxima [�54 �36 �6],

mean [�56�37�2], SD [4.2 6.9 7.2]; IPS groupmaxima [�33�51

36], mean [�33 �52 40], SD [4.8 4.6 5.3]; FG group maxima [�44

�58 �15], mean [�43 �59 �12], SD [ 5 4.9 4.7].

For the left frontal cluster, we defined two distinct VOIs,

namely, the PO and IFS, since a previous study (Makuuchi et al.,

2009) indicated that the PO and the IFS are functionally dis-

sociable into reflecting syntactic complexity and syntax related

WM, respectively. ThemMTG is included in themodel because

it was revealed by the ANOVA, and because the involvement of

theMTGhas been reported for a number of studies on sentence

processing (Tyler and Marslen-Wilson, 2008; Caplan et al.,

2008). The IPS was taken into the DCM, since it was revealed

by the ANOVA and is reported to be activated not only during

verbal WM tasks (e.g., Jonides et al., 1998; Owen et al., 2005;

Smith and Jonides, 1998), but also for sentence processingwhen




c o r t e x x x x ( 2 0 1 3 ) 1e8 5

WM demands are high (e.g., Novais-Santos et al., 2007; Meyer

et al., 2012), suggesting to reflect phonological WM as the

common feature. The FGVOIwas selected to tap into the visual

word form area (Cohen et al., 2000, 2002) on the basis of the

activation coordinates reported in ameta-analysis (x¼�44� 4,

y ¼ �58 � 5, z ¼ �15 � 6) (Jobard et al., 2003). Although the

ANOVA did not reveal FG activation for the main effects or

interaction, this region is activated by the current sentence

reading task, as the SPM{F} for the effect of interest in the in-

dividual analysis with the new model for DCM detected con-

sistent activation in this region.

A total of fifteenDCMswere constructed for five regions (PO,

IFS, mMTG, IPS, and FG), and then estimated and compared

(Fig. 1). The four regions (PO, IFS, mMTG, IPS) were supposed to

fully interconnect based on several previous anatomical stud-

ies (Catani et al., 2005; Frey et al., 2008; Petrides and Pandya,

2009; Seltzer and Pandya, 1978, 1984; Yeterian et al., 2011).

Since the connection between the FG and other regions is un-

clear, we made fifteen variations of the connection pattern. All

connections were modeled as bidirectional (i.e., IFSePOmeans

IFS / PO and PO / IFS), and all connections including recur-

sive connections (e.g., PO/ PO) were allowed to bemodulated

by the factors filler-gap DISTANCE and TYPE of dislocation.

Models 1e4 were “one-connection from the FG”models, where

one of the four regions (PO, IFS, mMTG, IPS) was connected

with the FG. Similarly, models 5e10 were the “two-connection

from the FG” models, models 11e15 were the “three-con-

nection from the FG” models, and the last model was the “full

connection” model, in which all the four regions were inter-

connected with the FG. Sets of fifteen DCMs for twenty-one

subjects were submitted to a random effects Bayesian model

comparison procedure (Stephan et al., 2009). The posterior

means of estimated connection strengths from the winning

model were subsequently evaluated by one-sample t-tests,

correcting for multiple tests by controlling the false discovery

rate (FDR) at q < .05 (Benjamini and Hochberg, 1995).

Table 1 e (a) Endogeneous connectivity; (b) Modulation by fact

To\from PO IFS

(a)

PO .117 (.349) p ¼ .14 .12

IFS .091 (.353) p ¼ .25 .01

mMTG .169 (.198) p ¼ .00088 .111 (.272) p ¼ .077

IPS .245 (.296) p ¼ .0012 .101 (.193) p ¼ .026 .11

FG �.017 (.329) p ¼ .82 .154 (.375) p ¼ .074

FDR q < .05, p ¼ .0256.

(b)

PO .068 (.107) p ¼ .0089 .070 (.112) p ¼ .0098 �.01

IFS .063 (.118) p ¼ .024 .117 (.087) p ¼ .000005 .04

mMTG .009 (.115) p ¼ .74 .057 (.099) p ¼ .016 .06

IPS .023 (.103) p ¼ .32 .071 (.091) p ¼ .0020 .02

FG .000 (.091) p ¼ .99 .039 (.091) p ¼ .060

FDR q < .05, p ¼ .0167.

(c)

PO .069 (.110) p ¼ .0095 .032 (.096) p ¼ .14 .00

IFS .024 (.101) p ¼ .28 .082 (.099) p ¼ .0012 .04

mMTG .021 (.089) p ¼ .28 .001 (.083) p ¼ .96 .04

IPS .008 (.060) p ¼ .55 .039 (.075) p ¼ .027 �.00

FG �.021 (.097) p ¼ .33 �.019 (.094) p ¼ .36

FDR q < .05, p ¼ .0022.


3. Results

The Bayesian model comparison revealed that the best model

(in terms of providing the best trade-off between accuracy and

model fit and generalizability) was model 13, in which the FG

is connected to the PO, IFS, and IPS, but not to mMTG (Fig. 1).

The t-tests on the connection strength revealed the results

summarized in Table 1 and Fig. 2. The exceedance probability

of the winning model was .217, while the exceedance proba-

bilities for the remaining models were less than .0596. One-

sample t-tests were performed on the posterior means of the

estimated connection strengths of model 13. The significant

(FDR correction at q < .05) endogenous connections and

modulatory effects are shown in Table 1 and Fig. 2.

4. Discussion

Against a background of rapidly developing knowledge of the

neuroanatomical pathways for language, we obtained esti-

mates of effective connectivity among language regions from

functional imaging data using dynamic causal modeling

(DCM). Based on data from a study on the processing of syn-

tactically complex sentences by reading, we grouped the re-

gions into the core language system (PO and mMTG), the WM

system [with a syntactic WM (IFS) and a phonological WM

(IPS)], and the visual system (FG). The DCM results indicate

a clear hierarchical relationship between the three systems

regarding the direction of the causal influence from the visual

systemto theWMsystemto the core language system (Fig. 2A).

Although our connectivity estimates refer to neuronal in-

fluences that are expressed at fairly low frequencies, and

should be taken with caution, the results suggest that during

sentence reading information is first conveyed from visual

system (FG) to theWM systems, i.e., the syntactic (IFS) and the

or DISTANCE; (c) Modulation by factor TYPE.

mSTS IPS FG

9 (.255) p ¼ .031 .243 (.302) p ¼ .0015 .105 (.314) p ¼ .14

5 (.322) p ¼ .84 .078 (.188) p ¼ .073 .315 (.419) p ¼ .0026

.174 (.225) p ¼ .0020 NaN

4 (.229) p ¼ .034 .235 (.284) p ¼ .0011

NaN .055 (.300) p ¼ .41

9 (.103) p ¼ .40 .013 (.107) p ¼ .59 .021 (.135) p ¼ .49

2 (.101) p ¼ .073 .072 (.107) p ¼ .0056 .041 (.150) p ¼ .23

7 (.118) p ¼ .017 .061 (.114) p ¼ .023 NaN

8 (.095) p ¼ .20 .089 (.117) p ¼ .0024 .036 (.167) p ¼ .34

NaN .036 (.076) p ¼ .045 �.051 (.127) p ¼ .079

6 (.093) p ¼ .77 �.006 (.118) p ¼ .81 �.041 (.142) p ¼ .20

7 (.097) p ¼ .040 .026 (.108) p ¼ .29 �.076 (.161) p ¼ .043

1 (.088) p ¼ .046 .000 (.084) p ¼ .99 NaN

0 (.088) p ¼ .98 .045 (.094) p ¼ .041 �.028 (.139) p ¼ .37

NaN .006 (.094) p ¼ .76 �.074 (.165) p ¼ .054




Fig. 2 e Significant endogenous connections and significant modulatory effects on connections by the increased WM load.

(A) Hierarchical diagram for the functionally segregated regions (WM, core language, and visual systems) and their

connections. Statistically significant endogenous connections (Table 1a, FDR correction q < .05) and significant modulation

(Table 1b) of connections are shown by arrows. The factor TYPE showed a significant modulatory effect on the self-

connection of the IFS only (Table 1c). Black arrows indicate endogenous connections. All self-connections (e.g., FG / FG)

were significant and are not indicated on the figure. Red arrows indicate significantly increased connection strengths by the

factor DISTANCE. The modulation of all self-connections was significant (not indicated on the figure), except in the FG. The

vertical gray arrow to the FG represents the input to the visual system. (B) Significant endogenous connections plotted on

a schematic brain. (C) Significantly modulated connections by the factor DISTANCE plotted on a schematic brain.

c o r t e x x x x ( 2 0 1 3 ) 1e86

phonological (IPS) WMs, before effectively connecting to the

language system. During the processing of sentences with low

WM demand, i.e., subject-first (canonical word order) sen-

tences (eleven words long) without dislocation, predominant

causal influences exist from the FG via the IPS to the core

language system, i.e., PO and mMTG (Fig. 2B). In contrast,

processing sentences with highWM demands, i.e., object-first

(non-canonical word order) sentences (eleven words long)

with larger filler-gap distance due to dislocation introduces an

additional bidirectional influence between the two WM foci

(i.e., IFS and IPS) and, moreover, leads to an enhancement of

the causal influence from the IFS to the core language regions

(Fig. 2C). The shift of the information mediation from the IPS

to IFS for object-first sentences suggests a qualitative shift

from phonological WM to syntactic WM (Friederici, 2012b).

The IFG, including the IFS, is reported to generally activate

more intensively during WM for sequences that contain an

underlying structure, compared to unstructured sequences

across domains (Bor et al., 2003, 2004; Jubault et al., 2007). In

language, the IFG is activated as a function of WM involved in

sentence processing (Santi and Grodzinsky, 2007, 2010), and

the IFS has been revealed to support syntactic WM, in par-

ticular (Amici et al., 2007; Makuuchi et al., 2009). In the present

study the activation in the IFS can be associatedwith syntactic

WM necessary to process the filler-gap distance of the dis-

located object-noun phrase while building up the hierarchical

structure underlying the syntactically complex, object-first

sentence. In contrast, the IPS which appears to be associated

with temporal storage and retrieval of linguistic information

in phonological form in verbal WM tasks (Paulesu et al., 1993;

Jonides et al., 1998; Smith and Jonides, 1998; Majerus et al.,

2010; Champod and Petrides, 2010) may reflect neural activ-

ities to hold words as phonological WM (Meyer et al., 2012).

With respect to the functionsof thewhitematter pathways,

the present results suggest a strong involvement of the dorsal

pathway terminating in the PO, in the processing of complex

sentences (Friederici, 2009; Anwander et al., 2007; Brauer et al.,

2011). Within the core language regions, a significant intrinsic


connectionwas found unidirectionally from the frontal cortex

to the temporal cortex, i.e., PO to mMTG. This connection is

either direct (PO / mMTG) or mediated by the parietal lobe

(PO / IPS / mMTG), in line with the finding of direct and

indirect connections between language regions (Catani et al.,

2005). With regard to effective connectivity, the present data

are compatible with a recent DCM study on the auditory sen-

tence processing showing functional connectivity between the

frontal and the posterior temporal cortices (den Ouden et al.,

2012). Functionally, these fronto-to-temporal connectivities

are interpreted as topedown processes in the service of argu-

ment assignment and sentence interpretation (Friederici,

2012a; Yvert et al., 2012).

Another interesting DCM finding is the lack of significant

effective connectivity between the visual system (FG) and the

core language system (PO andmMTG). This result implies that

visual linguistic information may be transmitted indirectly

to the core language system, with the WM system serving

as a buffer or a mediator of interaction between the sensory

and the core language systems. More generally, WM system

might emerge as an interface between the systems that pro-

foundly differ in the content, manner, or speed of information

processing.

In conclusion, the present study suggests that the brain

mechanisms of sentence processing during reading can be

viewed as consisting of different sub-systems: namely, the vi-

sual system, the WM system, and the core language system

which interact in a hierarchically structured manner. Beyond

the ‘static’ activation map, the present DCM analysis provides

information about the direction and intensity of the causal in-

fluences between regions, with a clear shift in connection

strengths asmemory demands in sentence processing increase.

Acknowledgments

We thank Klaas Enno Stephan for his most helpful comments

on the manuscript. We are grateful to Annet Wiedemann,




c o r t e x x x x ( 2 0 1 3 ) 1e8 7

Wipper Simone andAnke Kummer forMRI data acquisition, to

Kerstin Flake and Andrea Gast-Sandmann for the graphics,

and to Helga Smallwood for English editing.

r e f e r e n c e s

Amici S, Brambati SM, Wilkins DP, Ogar J, Dronkers NL, Miller BL,et al. Anatomical correlates of sentence comprehension andverbal working memory in neurodegenerative disease. Journalof Neuroscience, 27(23): 6282e6290, 2007.

Anwander A, Tittgemeyer M, von Cramon DY, Friederici AD, andKnoesche TR. Connectivity-based parcellation of Broca’s area.Cerebral Cortex, 17(4): 816e825, 2007.

Ashburner J. A fast diffeomorphic image registration algorithm.NeuroImage, 38(1): 95e113, 2007.

Benjamini Y and Hochberg Y. Controlling the false discovery rate:A practical and powerful approach to multiple testing. Journalof the Royal Statistical Society Series B, 57(1): 289e300, 1995.

Bor D, Cumming N, Scott CE, and Owen AM. Prefrontal corticalinvolvement in verbal encoding strategies. European Journal ofNeuroscience, 19(12): 3365e3370, 2004.

Bor D, Duncan J, Wiseman RJ, and Owen AM. Encoding strategiesdissociate prefrontal activity from working memory demand.Neuron, 37(2): 361e367, 2003.

Bornkessel I, Zyssett S, Friederici AD, von Cramon DY, andSchlesewsky M. Who did what to whom? The neural basis ofargument hierarchies during language comprehension.NeuroImage, 26(1): 221e233, 2005.

Brauer J, Anwander A, and Friederici AD. Neuroanatomicalprerequisites for language functions in the maturing brain.Cerebral Cortex, 21(2): 459e466, 2011.

Caplan D and Waters GS. Verbal working memory and sentencecomprehension. Behavioural Brain Sciences, 22(1): 77e94, 1999.

Caplan D, Chen E, and Waters G. Task-dependent and task-independent neurovascular responses to syntactic processing.Cortex, 44(3): 257e275, 2008.

Catani M, Jones DK, and Ffytche DH. Perisylvian language networksof the human brain. Annals of Neurology, 57(1): 8e16, 2005.

Champod AS and Petrides M. Dissociation within thefrontoparietal network in verbal working memory: Aparametric functional magnetic resonance imaging study.Journal of Neuroscience, 30: 3849e3856, 2010.

Cohen L, Dehaene S, Naccache L, Lehericy S, Dehaene-Lambertz G, Henaff MA, et al. The visual word form area:Spatial and temporal characterization of an initial stage ofreading in normal subjects and posterior split-brain patients.Brain, 123: 291e307, 2000.

Cohen L, Lehericy S, Chochon F, Lemer C, Rivaud S, andDehaene S. Language-specific tuning of visual cortex?Functional properties of the visual word form area. Brain, 125:1054e1069, 2002.

Cooke A, Zurif EB, DeVita C, Alsop D, Koenig P, Detre J, et al. Neuralbasis for sentencecomprehension:Grammatical andshort-termmemory components. Human Brain Mapping, 15(2): 80e94, 2002.

Daneman M and Carpenter PA. Individual differences in workingmemory and reading. Journal of Verbal Learning and VerbalBehavior, 19(4): 450e466, 1980.

Dejerine JJ. Anatomie des centres nerveux. Paris: Rueff, 1901.den Ouden DB, Saur D, Mader W, Schelter B, Lukic S, Wali E, et al.

Network modulation during complex syntactic processing.NeuroImage, 59(1): 815e823, 2012.

Fiebach CJ, Schlesewsky M, and Friederici AD. Syntactic workingmemory and the establishment of filler-gap dependencies:Insights from ERPs and fMRI. Journal of Psycholinguistic Research,30(3): 321e338, 2001.


Frey S, Campbell JS, Pike GB, and Petrides M. Dissociating thehuman language pathways with high angular resolutiondiffusion fiber tractography. Journal of Neuroscience, 28(45):11435e11444, 2008.

Friederici AD. Pathways to language: Fiber tracts in the humanbrain. Trends in Cognitive Sciences, 13(4): 175e181, 2009.

Friederici AD. Language development and the ontogeny of thedorsal pathway. Frontiers in Evolutionary Neuroscience 4: 3.http://dx.doi.org/10.3389/fnevo.2012.00003, 2012a.

Friederici AD. The cortical language circuit: From auditoryperception to sentence comprehension. Trends in CognitiveSciences, 16(5): 262e268, 2012b.

Friederici AD, Bahlmann J, Heim S, Schubotz RI, and Anwander A.The brain differentiates human and non-human grammars:Functional localization and structural connectivity. Proceedingsof the National Academy of Sciences of the USA, 103(7): 2458e2463,2006.

Friederici AD, Makuuchi M, and Bahlmann J. The role of theposterior superior temporal cortex in sentencecomprehension. NeuroReport, 20(6): 563e568, 2009.

Friston KJ, Harrison L, and Penny W. Dynamic causal modelling.NeuroImage, 19(4): 1273e1302, 2003.

Geschwind N. Disconnexion syndromes in animals and man. I.Brain, 88: 237e294, 1965a.

Geschwind N. Disconnexion syndromes in animals and man. II.Brain, 88: 585e644, 1965b.

Hickok G and Poeppel D. The cortical organization of speechprocessing. Nature Reviews Neuroscience, 8(5): 393e402, 2007.

Jobard G, Crivello F, and Tzourio-Mazoyer N. Evaluation of thedual route theory of reading: A metanalysis of 35neuroimaging studies. NeuroImage, 20(2): 693e712, 2003.

Jonides J, Schumacher EH, Smith EE, Koeppe RA, Awh E, Reuter-Lorenz PA, et al. The role of parietal cortex in verbal workingmemory. Journal of Neuroscience, 18(13): 5026e5034, 1998.

Jubault T, Ody C, and Koechlin E. Serial organization of humanbehavior in the inferior parietal cortex. Journal of Neuroscience,27(41): 11028e11036, 2007.

Majerus S, D’Argembeau A, Martinez Perez T, Belayachi S, Van derLinden M, Collette F, et al. The commonality of neuralnetworks for verbal and visual short-term memory. Journal ofCognitive Neuroscience, 22: 2570e2593, 2010.

Makris N, Papadimitriou GM, Kaiser JR, Sorg S, Kennedy DN, andPandya DN. Delineation of the middle longitudinal fascicle inhumans: A quantitative, in vivo, DT-MRI study. Cerebral Cortex,19(4): 777e785, 2009.

Makuuchi M, Bahlmann J, Anwander A, and Friederici AD.Segregating the core computational faculty of humanlanguage from working memory. Proceedings of the NationalAcademy of Sciences of the USA, 106(20): 8362e8367, 2009.

Makuuchi M, Grodzinsky Y, Amunts K, Santi A, and Friederici A.Processing non-canonical sentences in Broca’s region:Reflections of movement distance and type. Cereb Cortex, 23(3):694e702, 2013.

Meyer L, Obleser J, Anwander A, and Friederici AD. Linkingordering in Broca’s area to storage in left temporo-parietalregions: The case of sentence processing. NeuroImage, 62(3):1987e1998, 2012.

Newman SD, Ikuta T, and Burns T. The effect of semanticrelatedness on syntactic analysis: An fMRI study. Brain andLanguage, 113(2): 51e58, 2010.

Novais-Santos S, Gee J, Shah M, Troiani V, Work M, andGrossman M. Resolving sentence ambiguity with planning andworking memory resources: Evidence from fMRI. NeuroImage,37(1): 361e378, 2007.

Oldfield RC. The assessment and analysis of handedness: TheEdinburgh inventory. Neuropsychologia, 9(1): 97e113, 1971.

Owen AM, McMillan KM, Laird AR, and Bullmore E. N-backworking memory paradigm: A meta-analysis of normative


http://dx.doi.org/10.3389/fnevo.2012.00003



c o r t e x x x x ( 2 0 1 3 ) 1e88

functional neuroimaging studies. Human Brain Mapping, 25(1):46e59, 2005.

Paulesu E, Frith CD, and Frackowiak RS. The neural correlates ofthe verbal component of working memory. Nature, 362(6418):342e345, 1993.

Petrides M and Pandya DN. Distinct parietal and temporalpathways to the homologues of Broca’s area in the monkey.PLoS Biology, 7(8): e1000170, 2009.

Rilling JK, Glasser MF, Preuss TM, Ma X, Zhao T, Hu X, et al. Theevolution of the arcuate fasciculus revealed with comparativeDTI. Nature Neuroscience, 11(4): 426e428, 2008.

Rolheiser T, Stamatakis EA, and Tyler LK. Dynamic processing inthe human language system: Synergy between the arcuatefascicle and extreme capsule. Journal of Neuroscience, 31(47):16949e16957, 2011.

Santi A and Grodzinsky Y. Working memory and syntax interactin Broca’s area. NeuroImage, 37(1): 8e17, 2007.

Santi A and Grodzinsky Y. fMRI adaptation dissociates syntacticcomplexity dimensions. NeuroImage, 51(4): 1285e1293, 2010.

Saur D, Kreher BW, Schnell S, Kummerer D, Kellmeyer P, Vry MS,et al. Ventral and dorsal pathways for language. Proceedings ofthe National Academy of Sciences of the USA, 105(46):18035e18040, 2008.

Seltzer B and Pandya DN. Afferent cortical connections andarchitectonics of the superior temporal sulcus andsurrounding cortex in the rhesus monkey. Brain Research,149(1): 1e24, 1978.

Seltzer B and Pandya DN. Further observations on parieto-temporal connections in the rhesus monkey. ExperimentalBrain Research, 55(2): 301e312, 1984.

Smith EE and Jonides J. Neuroimaging analyses of humanworking memory. Proceedings of the National Academy of Sciencesof the USA, 95(20): 12061e12068, 1998.


Stephan KE and Friston KJ. Analyzing effective connectivity withfunctional magnetic resonance imaging. WileyInterdisciplinary Reviews in Cognitive Science, 1(3): 446e459,2010.

Stephan KE, Penny WD, Daunizeau J, Moran RJ, and Friston KJ.Bayesian model selection for group studies. NeuroImage, 46(4):1004e1017, 2009.

Stromswold K, Caplan D, Alpert N, and Rauch S. Localization ofsyntactic comprehension by positron emission tomography.Brain and Language, 52(3): 452e473, 1996.

Tyler LK and Marslen-Wilson W. Fronto-temporal brain systemssupporting spoken language comprehension. PhilosophicalTransaction of the Royal Society of London B Biological Sciences,363(1493): 1037e1054, 2008.

Tyler LK, Marslen-Wilson WD, Randall B, Wright P, Devereux BJ,Zhuang J, et al. Left inferior frontal cortex and syntax:Function, structure and behaviour in patients with lefthemisphere damage. Brain, 134(2): 415e431, 2011.

Weiller C, Musso M, Rijntjes M, and Saur D. Please don’tunderestimate the ventral pathway in language. Trends inCognitive Sciences, 13(9): 369e370, 2009.

Wernicke C. Der Aphasische Symptomencomplex. Breslau: Cohenand Weigert, 1874.

Wilson SM, Galantucci S, Tartaglia MC, Rising K, Patterson DK,Henry ML, et al. Syntactic processing depends on dorsallanguage tracts. Neuron, 72(2): 397e440, 2011.

Yeterian EH, Pandya DN, Tomaiuolo F, and Petrides M. Thecortical connectivity of the prefrontal cortex in the monkeybrain. Cortex, 48(1): 58e81, 2011.

Yvert G, Perrone-Bertolotti M, Baciu M, and David O. Dynamiccausal modeling of spatiotemporal integration of phonologicaland semantic processes: An electroencephalographic study.Journal of Neuroscience, 32(12): 4297e4306, 2012.




hierarchical functional connectivity between the core language system and the working memory system

Documents