HAL Id: hal-02272288
https://hal.inria.fr/hal-02272288
Submitted on 27 Aug 2019
HAL is a multi-disciplinary open access archive for the deposit and dissemination of sci- entific research documents, whether they are pub- lished or not. The documents may come from teaching and research institutions in France or abroad, or from public or private research centers.
L’archive ouverte pluridisciplinaire HAL, est destinée au dépôt et à la diffusion de documents scientifiques de niveau recherche, publiés ou non, émanant des établissements d’enseignement et de recherche français ou étrangers, des laboratoires publics ou privés.
Syntactic Parsing versus MWEs: What can fMRI signal tell us
Murielle Fabre, Yoann Dupont, Éric Villemonte de la Clergerie
To cite this version:
Murielle Fabre, Yoann Dupont, Éric Villemonte de la Clergerie. Syntactic Parsing versus MWEs:
What can fMRI signal tell us. PARSEME-FR 2019 consortium meeting, Jun 2019, Blois, France.
�hal-02272288�
Syntactic Parsing versus MWEs
what can fMRI signal tell us
Murielle Fabre, Yoann Dupont, Eric de la Clergerie
Parsème - Blois Murielle Fabre
Project & Approach
Cereb
Left hemisphere Sentence Network
Parsème - Blois Murielle Fabre
Bring together
computational linguistics and cognitive neuro-imaging
to shed light on sentence comprehension and its neural bases
Project & Approach
Cereb
Left hemisphere Sentence Network
Bring together
computational linguistics and cognitive neuro-imaging to shed light on sentence comprehension and its neural bases
Parsème - Blois Murielle Fabre
Does MWE processing pattern together
with sentence- structure building
effects in the
brain?
Naturalistic Corpus : The Little Prince in 3 languages
,
Il y a six ans déjà que mon ami s'en est allé avec son mouton.
Si j'essaie ici de le décrire, c'est afin de ne pas l'oublier.
C'est triste d'oublier un ami.
French
Chinese English Six years ago that my friend left
wit his sheep. If I try to describe him, it’s ignorer not to forget him. It is sad to forget a friend.
Parsème - Blois Murielle Fabre
Naturalistic Corpus : The Little Prince in 3 languages
"Everyday listening" conditions in English
,
Il y a six ans déjà que mon ami s'en est allé avec son mouton.
Si j'essaie ici de le décrire, c'est afin de ne pas l'oublier.
C'est triste d'oublier un ami.
French
Chinese English Six years ago that my friend left
wit his sheep. If I try to describe him, it’s ignorer not to forget him. It is sad to forget a friend.
Parsème - Blois Murielle Fabre
Participants (51)
American native speakers (32 women, 18-37 years old ) Task: Listen to the audiobook The Little Prince (1 h 38 min) 9 runs
+ Comprehension questions after each run
Recoding : 3T MRI scanner 32-channel head coil at the Cornell MRI Facility. Muti-echo sequence.
Preprocessing : FSL, AFNI + the signal-to-noise ratio, using multi-echo independent
components analysis (ME-ICA) (Kundu et al., 2013)
Analysis : General Linear Model (GLM - SPM12)
Control Regressors : pitch (f0), acoustic volume
(RMS), Word rate, Word frequency.
The Team for this study
Murielle Fabre
Christophe Pallier Hazem
Al-saied
ATILF UMR 7118 (CNRS/Université de Lorraine)
Mathieu Constant Eric de la
Clergerie Yoann
Dupont
Parsème - Blois Murielle Fabre
Road map for today
Identifying MWEs
Stability of PMIs Parsing vs. MWEs
Next steps fMRI results
A B C D E
Parsème - Blois Murielle Fabre
Parsing versus MWEs
Parsème - Blois Murielle Fabre
A
Parsing vs. MWEs Identifying Stability of PMIs fMRI results Next steps
Scientific Questions and Hypotheses
Caractéristiques de l’apprentissage
de la lecture chinoise:
- Conscience phonologique :
Phonologie convergente Sémantique divergente - Conscience graphique :
Identification des unités graphiques et
de la structure interne du caractère
- Conscience morphologique
Multi-word expressions
Parsème - Blois Murielle Fabre
Frequently co-occurring word sequences, known as Multiword Expressions (MWEs) are likely to be processed differently by the language
network.
Parsing vs. MWEs Identifying Stability of PMIs fMRI results Next steps
Scientific Questions and Hypotheses
Caractéristiques de l’apprentissage
de la lecture chinoise:
- Conscience phonologique :
Phonologie convergente Sémantique divergente - Conscience graphique :
Identification des unités graphiques et
de la structure interne du caractère
- Conscience morphologique
Multi-word expressions
Parsème - Blois Murielle Fabre
Frequently co-occurring word sequences, known as Multiword Expressions (MWEs) are likely to be processed differently by the language
network.
MWEs are a perfect testing ground to understand how expressions like break
the ice, boa constrictor, see to it, in spite of are processed in the brain.
Structure building MWE Processing
Parsing vs. MWEs Identifying Stability of PMIs fMRI results Next steps
One has to look after lamps
Sound (Hz)
Scientific Questions and Hypotheses
Caractéristiques de l’apprentissage
de la lecture chinoise:
- Conscience phonologique :
Phonologie convergente Sémantique divergente - Conscience graphique :
Identification des unités graphiques et
de la structure interne du caractère
- Conscience morphologique
Multi-word expressions
Parsème - Blois Murielle Fabre
Frequently co-occurring word sequences, known as Multiword Expressions (MWEs) are likely to be
processed differently by the language network.
MWEs are a perfect testing ground to understand how expressions like break the ice, boa constrictor, see to it, in spite of are processed in the brain.
Structure building MWE Processing
Parsing vs. MWEs Identifying Stability of PMIs fMRI results Next steps
one
has_to
look_after
lamps
has to
after look
One has to look after lamps
Sound (Hz)
Scientific Questions and Hypotheses
Caractéristiques de l’apprentissage
de la lecture chinoise:
- Conscience phonologique :
Phonologie convergente Sémantique divergente - Conscience graphique :
Identification des unités graphiques et
de la structure interne du caractère
- Conscience morphologique
Multi-word expressions
Parsème - Blois Murielle Fabre
Frequently co-occurring word sequences, known as Multiword Expressions (MWEs) are likely to be processed differently by the language
network.
MWEs are a perfect testing ground to understand how expressions like break
the ice, boa constrictor, see to it, in spite of are processed in the brain.
Structure building MWE Processing
Parsing vs. MWEs Identifying Stability of PMIs fMRI results Next steps
one
has_to
look_after
lamps
has to
after look
One has to look after lamps
Sound (Hz)
Scientific Questions and Hypotheses
Caractéristiques de l’apprentissage
de la lecture chinoise:
- Conscience phonologique :
Phonologie convergente Sémantique divergente - Conscience graphique :
Identification des unités graphiques et
de la structure interne du caractère
- Conscience morphologique
Multi-word expressions
Parsème - Blois Murielle Fabre
Frequently co-occurring word sequences, known as Multiword Expressions (MWEs) are likely to be processed differently by the language
network.
MWEs are a perfect testing ground to understand how expressions like break
the ice, boa constrictor, see to it, in
spite of are processed in the brain. a computational graded quantification identifying expressions
likely to be processed as units, rather than built-up compositionally
a computational measure tracking tree-building work
needed in composed syntactic phrases
Structure building MWE Processing
Parsing vs. MWEs Identifying Stability of PMIs fMRI results Next steps
Scientific Questions and Hypotheses
Caractéristiques de l’apprentissage
de la lecture chinoise:
- Conscience phonologique :
Phonologie convergente Sémantique divergente - Conscience graphique :
Identification des unités graphiques et
de la structure interne du caractère
- Conscience morphologique
one
has_to
look_after
lamps
has to
after look
MWE Processing Structure building
One has to look after lamps
One has to look after lamps
Sound (Hz)
8 6 4 2 computational metrics 0
Frequently co-occurring word sequences, known as Multiword Expressions (MWEs) are likely to be processed differently by the language
network.
MWEs are a perfect testing ground to understand how expressions like break
the ice, boa constrictor, see to it, in spite of are processed in the brain.
Parsème - Blois Murielle Fabre
PMI a computational graded quantification identifying expressions likely to be processed as units, rather than built-up compositionally, BU tracks tree-building work needed in composed syntactic phrases.
Multi-word expressions
Parsing vs. MWEs Identifying Stability of PMIs fMRI results Next steps
Scientific Questions and Hypotheses
Caractéristiques de l’apprentissage
de la lecture chinoise:
- Conscience phonologique :
Phonologie convergente Sémantique divergente - Conscience graphique :
Identification des unités graphiques et
de la structure interne du caractère
- Conscience morphologique
one
has_to
look_after
lamps
has to
after look
MWE Processing Structure building
One has to look after lamps
One has to look after lamps
Sound (Hz)
8 6 4 2 computational metrics 0
time (s)
0 1 1 2 2
2s 6s 10s 14s 18s 22s 26s 30s
2 1.5
1 0.5
0
estimated BOLD signal 2 6 10 14 18 22 26 30
Frequently co-occurring word sequences, known as Multiword Expressions (MWEs) are likely to be processed differently by the language
network.
MWEs are a perfect testing ground to understand how expressions like break
the ice, boa constrictor, see to it, in spite of are processed in the brain.
Parsème - Blois Murielle Fabre
PMI a computational graded quantification identifying expressions likely to be processed as units, rather than built-up compositionally, BU tracks tree-building work needed in composed syntactic phrases.
Multi-word expressions
Parsing vs. MWEs Identifying Stability of PMIs fMRI results Next steps
Bottom-Up Parser actions count
Parsème - Blois Murielle Fabre
Investigating syntax in the brain through parsers
Parser actions Hierarchical
representation
Parsing vs. MWEs Identifying Stability of PMIs fMRI results Next steps
Parser actions Bottom-Up
Word-by-word measure of syntactic-structure building:
- Can instantiate
constituent-structure building the phrase/sentence.
as it builds and collects sub-parses towards the end of the phrase or sentence.
- The rules of a grammar are
applied at each incoming word
Parsème - Blois Murielle Fabre
Investigating syntax in the brain through parsers
Parser actions Hierarchical
representation
Parsing vs. MWEs Identifying Stability of PMIs fMRI results Next steps
S
PP
NP
SBAR
S
VP
VP
NP
NN drawing JJ
first PRP$
my VB
make TO
to NP
NN
pencil VBN
coloured DT
a IN
with S
VP
VBD
managed NP
PRP
I PP
NP
NN
turn IN
in CC
and S
VP
PP
NP
PP
NP
NN
jungle DT
the IN
of NP
NNS
adventures DT
the IN
about NP
NN
lot DT
a VBD
thought NP
PRP
I IN
So
1 2 1 1 2 1 1 2 1 1 7 1 1 3 2 3 1 1 1 2 1 1 1 1 10
Bottom-up parser action count
Parser actions count
Murielle Fabre
Sentence hierarchical representation + computational complexity metrics
Number of REDUCE actions taken since last word
Parsing vs. MWEs Identifying Stability of PMIs fMRI results Next steps
Association measure : Point-wise Mutual Information
PMI : A computational measure to link the degree of cohesiveness
of MWEs
Parsème - Blois Murielle Fabre
Computational graded quantification identifying expressions likely to be processed as units, rather than built-
up compositionally
Parsing vs. MWEs Identifying Stability of PMIs fMRI results Next steps
Association measure : Point-wise Mutual Information
PMI : A computational measure to link the degree of cohesiveness
of MWEs
where
and
American english corpus : Coca 560 millions
Parsème - Blois Murielle Fabre
Computational graded quantification identifying expressions likely to be processed as units, rather than built-
up compositionally
Parsing vs. MWEs Identifying Stability of PMIs fMRI results Next steps
Association measure : Point-wise Mutual Information
PMI : A computational measure to link the degree of cohesiveness of MWEs
where
and
American english corpus : Coca 560 millions
Parsème - Blois Murielle Fabre
Parsing vs. MWEs Identifying Stability of PMIs fMRI results Next steps
S
PP
NP
SBAR
S
VP
VP
NP
NN drawing JJ
first PRP$
my VB
make TO
to NP
NN
pencil VBN
coloured DT
a IN
with S
VP
VBD
managed NP
PRP
I PP
NP
NN
turn IN
in CC
and S
VP
PP
NP
PP
NP
NN
jungle DT
the IN
of NP
NNS
adventures DT
the IN
about NP
NN
lot DT
a VBD
thought NP
PRP
I IN
So
1 2 1 1 2 1 1 2 1 1 7 1 1 3 2 3 1 1 1 2 1 1 1 1 10
Bottom-up parser action count
Parser actions count
Murielle Fabre
Sentence hierarchical representation + computational complexity metrics
0 0 0 0 2.62 5
0 0 0 0 0 0 0 0 4.083 0 0 0 0 0 0 0 0 0 0 0
MWE cohesion strength -> PMI
Parsing vs. MWEs Identifying Stability of PMIs fMRI results Next steps
Identifying MWEs : English
Parsème - Blois Murielle Fabre
B
Question Identifying Expriment fMRI Results Next steps
Parsing vs. MWEs Identifying Stability of PMIs fMRI results Next steps
Computational toolkit ot identify MWEs
MWEs were identified using a statistical tagger (Al Saied et al. 2017). trained on Children’s Book Test dataset.
Hazem Al-saied Mathieu Constant
Question Identifying Expriment fMRI Results Next steps Parsing vs. MWEs Identifying Stability of PMIs fMRI results Next steps
Parsème - Blois Murielle Fabre
Computational toolkit ot identify MWEs
MWEs were identified using a statistical tagger (Al Saied et al. 2017). trained on Children’s Book Test dataset.
Hazem Al-saied Mathieu Constant
Question Identifying Expriment fMRI Results Next steps Parsing vs. MWEs Identifying Stability of PMIs fMRI results Next steps
Parsème - Blois Murielle Fabre
Computational toolkit ot identify MWEs
MWEs were identified using a statistical tagger (Al Saied et al. 2017). trained on Children’s Book Test dataset.
Hazem Al-saied Mathieu Constant
Question Identifying Expriment fMRI Results Next steps Parsing vs. MWEs Identifying Stability of PMIs fMRI results Next steps
Parsème - Blois Murielle Fabre
Computational toolkit ot identify MWEs
MWEs were identified using a statistical tagger (Al Saied et al. 2017). trained on Children’s Book Test dataset.
Hazem Al-saied Mathieu Constant
Question Identifying Expriment fMRI Results Next steps Parsing vs. MWEs Identifying Stability of PMIs fMRI results Next steps
Parsème - Blois Murielle Fabre
Computational toolkit to identify English MWEs
MWEs were identified using a statistical tagger (Al Saied et al. 2017). Trained on Children’s Book Test dataset.
Hazem Al-saied Mathieu Constant
Parsème - Blois Murielle Fabre
Question Identifying Expriment fMRI Results Next steps
Parsing vs. MWEs Identifying Stability of PMIs fMRI results Next steps
MWE English
Parsème - Blois Murielle Fabre
Question Identifying Expriment fMRI Results Next steps
Parsing vs. MWEs Identifying Stability of PMIs fMRI results Next steps
MWE English
Parsème - Blois Murielle Fabre
Question Identifying Expriment fMRI Results Next steps
Parsing vs. MWEs Identifying Stability of PMIs fMRI results Next steps
MWE English
Matthieu Constant and Isabelle Tellier. 2012. Evaluating the impact of external lexical resources into a crf-based multiword segmenter and part-of-speech tagger. In 8th International Conference on Language Resources and Evaluation (LREC’12), pages
646–650. Parsème - Blois Murielle Fabre
Question Identifying Expriment fMRI Results Next steps
Parsing vs. MWEs Identifying Stability of PMIs fMRI results Next steps
MWE English
Matthieu Constant and Isabelle Tellier. 2012. Evaluating the impact of external lexical resources into a crf-based multiword segmenter and part-of-speech tagger. In 8th International Conference on Language Resources and Evaluation (LREC’12), pages
646–650. Parsème - Blois Murielle Fabre
Question Identifying Expriment fMRI Results Next steps
Parsing vs. MWEs Identifying Stability of PMIs fMRI results Next steps
Identifying MWEs : French
Parsème - Blois Murielle Fabre
B
Question Identifying Expriment fMRI Results Next steps
Parsing vs. MWEs Identifying Stability of PMIs fMRI results Next steps
Automatized identification of French MWEs
Parsème - Blois Murielle Fabre
Yoann Dupont
Eric de la Clergerie Extracting MWE
FRMG patterns matching Wiktionary
entries
Euro-Parliament Corpus - 41,5 millions
both written and oral style
Verifying patterns
in a second corpus Wikisource Corpus - 64 millions
Narrative written style
Filtering results
> to the average
PMI
> to the average occurence
Calculating PMIs scores
&
French euro-parliament + French wikisource
+ French wikipedia (179 millions) Average occurence
in
Question Identifying Expriment fMRI Results Next steps
Parsing vs. MWEs Identifying Stability of PMIs fMRI results Next steps
Automatized identification of French MWEs
Parsème - Blois Murielle Fabre
Yoann Dupont
Eric de la Clergerie Extracting MWE
FRMG patterns matching Wiktionary
entries
Euro-Parliament Corpus - 41,5 millions
both written and oral style
Verifying patterns
in a second corpus Wikisource Corpus - 64 millions
Narrative written style
Filtering results
> to the average
PMI
> to the average occurence
Calculating PMIs scores
&
French euro-parliament + French wikisource
+ French wikipedia (179 millions) Average occurence
in
Hand-picked selection
1300 Le Petit
Prince Audio
a) point de sa chute b) point de chute
Question Identifying Expriment fMRI Results Next steps
Parsing vs. MWEs Identifying Stability of PMIs fMRI results Next steps
MWE Patterns : French
Parsème - Blois Murielle Fabre
-> Ordre des mots des adjectifs antéposé post-posés :
• donner une fausse idée
• pas m' étonner beaucoup
• ne éprouver plus le besoin
-> Avantage d’avoir une représentation en dépendances :
MWE longue
• entourer le cou de son bras
MWE avec inclusion
• entrer à son tour dans la danse
Figures de style imagées non-incluses
• s’enroula autour de sa cheville, comme un bracelet d’or
• entasser l’humanité sur un îlot
Question Identifying Expriment fMRI Results Next steps
Parsing vs. MWEs Identifying Stability of PMIs fMRI results Next steps
MWE patterns English vs. French
Parsème - Blois Murielle Fabre
000 00 000 600 24 71
Table 3:
English French
Verbale 631 Nominales 380 Adverbiales 32 Prépositionnelles 305
Question Identifying Expriment fMRI Results Next steps
Parsing vs. MWEs Identifying Stability of PMIs fMRI results Next steps
Identification of French MWEs : patterns
Parsème - Blois Murielle Fabre
Nominal Adverbial
Verbal Prepositional
oeil clos
passant ordinaire pas de course couleur de miel drôle de bête éclat de rire
économie de temps
geste de lassitude mèche de cheveu messe de minuit peine de mort
mouvement de regret poupée de chiffon source de malentendu
clr habiller à le européen clr voir important comme écraser son nez contre le vitre ébaucher un sourire
être bien obliger
lever le oeil vers le ciel
ne clr avancer pas à grand-chose ne manger pas de pain
apaiser le soif aimer les chiffres
avoir mouiller le tempe boire le dernier goutte ce ne être pas mon faute cld habiller le coeur
clr enfoncer dans une rêverie parler toujours le premier au fond de son coeur
à tout hasard ni faim ni soif sur terre
comme un fontaine dans le désert contre tout espérance
en larme
faute de patience tout doucement bien loin
peut-être bien que tout à fait
un peu juste oui ou non à peine plus
Question Identifying Expriment fMRI Results Next steps
Parsing vs. MWEs Identifying Stability of PMIs fMRI results Next steps
Stability of PMI scores
Parsème - Blois Murielle Fabre
C
Question Expriment fMRI Results Next steps
Parsing vs. MWEs Expriment fMRI Stability of PMIs fMRI results Next steps
Parsing vs. MWEs Identifying Stability of PMIs Stability of PMI fMRI results Next steps
Preliminary steps to calculate PMIs - French
1. Building of a corpus of a comparable size to the COCA coprus.
2. Building a corpus of children books comparable to the CBT 3. Dependency parsing to capture more open MWEs
Parsème - Blois Murielle Fabre
Question Expriment fMRI Results Next steps
Parsing vs. MWEs Expriment fMRI Stability of PMIs fMRI results Next steps
Parsing vs. MWEs Identifying Stability of PMIs Stability of PMI fMRI results Next steps
French Children books Corpus Children's books Test Corpus
Size
CBT English
Register Littérature jeunesse et
des classiques
contenants des dialogues livres et histoires
pour la jeunesse
CBT Français
108 books 6 millions
Wikisource - Gutenberg 6 millions
Children's book dataset taken as reference : fb.ai/babi / http://arxiv.org/abs/1511.02301
Go through Project Gutenberg to find children's books, then stripping out the Project Gutenberg headers (which is sadly nontrivial). They have a lot of public domain works already transcribed in .txt form.
Parsème - Blois Murielle Fabre
Question Expriment fMRI Results Next steps
Parsing vs. MWEs Expriment fMRI Stability of PMIs fMRI results Next steps
Parsing vs. MWEs Identifying Stability of PMIs Stability of PMI fMRI results Next steps
French coca-style Corpus
COCA corpus ( Davies, 2008 )
Corpus of Contemporary American English
COCA Américain
spoken, fiction,
popular magazines, newspapers, and academic texts
COCA-fr Français
560 millions
20 million words each year 1990-2017
500 millions
Parsème - Blois Murielle Fabre
spoken, fiction,
popular magazines, newspapers, and academic texts Size
Register
Leverage on already published open access corpora + stripping out Whikisource headers (which is sadly nontrivial) and constitute an agglomerated corpus of public domain works transcribed in .txt form.
Question Expriment fMRI Results Next steps
Parsing vs. MWEs Expriment fMRI Stability of PMIs fMRI results Next steps
Parsing vs. MWEs Identifying Stability of PMIs Stability of PMI fMRI results Next steps
Stability of PMI scores across corpora
Parsème - Blois Murielle Fabre
41,5 m 6 m 200 m 6 m 180 m
Question Expriment fMRI Results Next steps
Parsing vs. MWEs Expriment fMRI Stability of PMIs fMRI results Next steps
Parsing vs. MWEs Identifying Stability of PMIs Stability of PMI fMRI results Next steps
Stability of PMI measures
Parsème - Blois Murielle Fabre
Stability of PMI measures
Parsème - Blois Murielle Fabre
Stability of PMI measures
Parsème - Blois Murielle Fabre
fMRI results - English
Parsème - Blois Murielle Fabre
D
Question Expriment fMRI Results Next steps
Parsing vs. MWEs Expriment fMRI Stability of PMIs fMRI results Next steps
Parsing vs. MWEs Identifying Stability of PMIs fMRI results Next steps
S
PP
NP
SBAR
S
VP
VP
NP
NN drawing JJ
first PRP$
my VB
make TO
to NP
NN
pencil VBN
coloured DT
a IN
with S
VP
VBD
managed NP
PRP
I PP
NP
NN
turn IN
in CC
and S
VP
PP
NP
PP
NP
NN
jungle DT
the IN
of NP
NNS
adventures DT
the IN
about NP
NN
lot DT
a VBD
thought NP
PRP
I IN
So
1 2 1 1 2 1 1 2 1 1 7 1 1 3 2 3 1 1 1 2 1 1 1 1 10
Bottom-up parser action count
Parser actions count
Murielle Fabre
Sentence hierarchical representation + computational complexity metrics
0 0 0 0 2.62 5
0 0 0 0 0 0 0 0 4.083 0 0 0 0 0 0 0 0 0 0 0
MWE cohesion strength -> PMI
PMI a computational graded quantification identifying expressions likely to be processed as units, rather than built-up compositionally, BU tracks tree-building work needed in composed syntactic phrases.
Question Expriment fMRI Results Next steps
Parsing vs. MWEs Expriment fMRI Stability of PMIs fMRI results Next steps
Parsing vs. MWEs Identifying Stability of PMIs fMRI results Next steps
Bottom-up parser action count
Cerebellum Crus I
p < 0.05 FWE
Bottom- up parser action count - English
Bilateral network involving IFG and ATL
Parsème - Blois Murielle Fabre
Analysis of MWEs and parser action counts in naturalistic spoken story comprehension supports a dissociation between Temporal and Parietal brain structures and anterior Frontal regions such as IFG and ATL, as respectively sub-serving the retrieval of
memorized expressions and structure-building processes.
Question Expriment fMRI Results Next steps
Parsing vs. MWEs Expriment fMRI Stability of PMIs fMRI results Next steps
Parsing vs. MWEs Identifying Stability of PMIs fMRI results Next steps
Bottom-up parser action count
Cerebellum Crus I
p < 0.05 FWE
Bottom- up parser action count - English
Bilateral network involving IFG and ATL
Parsème - Blois Murielle Fabre
Analysis of MWEs and parser action counts in naturalistic spoken story comprehension supports a dissociation between Temporal and Parietal brain
structures and anterior Frontal regions such as IFG and ATL, as respectively sub-serving the retrieval of
memorized expressions and structure-building processes.
Question Expriment fMRI Results Next steps
Parsing vs. MWEs Expriment fMRI Stability of PMIs fMRI results Next steps
Parsing vs. MWEs Identifying Stability of PMIs fMRI results Next steps
Increasing MWE cohesion strength (PMI) Decreasing MWE cohesion strength (PMI)
Positive and negative correlation with PMI - English
Parsème - Blois Murielle Fabre
-> PMI as a proxy of lexical cohesiveness
The results show an overlap between the significant effect for decreasing MWE cohesiveness and Bottom-up parser action count in left IFG and posterior temporal lobe. Highly cohesive MWEs implicate the Precuneus and the SMA, suggesting that only truly lexicalized linguistic expressions rely on these areas rather than traditional frontal and temporal nodes of the language network.
Question Expriment fMRI Results Next steps
Parsing vs. MWEs Expriment fMRI Stability of PMIs fMRI results Next steps
Parsing vs. MWEs Identifying Stability of PMIs fMRI results Next steps
IFG orb
MFG/
p < 0.05 FWE SFG
L R
IFG Tri
pMTG
IFG orb
Increasing MWE cohesion strength (PMI) Decreasing MWE cohesion strength (PMI)
SMA Precuneus
Positive and negative correlation with PMI - English
Parsème - Blois Murielle Fabre
-> PMI as a proxy of lexical cohesiveness
The results show an overlap between the significant effect for decreasing MWE cohesiveness and Bottom-up parser action count in left IFG and posterior temporal lobe. Highly cohesive MWEs implicate the Precuneus and the SMA, suggesting that only truly lexicalized linguistic expressions rely on these areas rather than traditional frontal and temporal nodes of the language network.
Question Expriment fMRI Results Next steps
Parsing vs. MWEs Expriment fMRI Stability of PMIs fMRI results Next steps
Parsing vs. MWEs Identifying Stability of PMIs fMRI results Next steps
IFG orb
MFG/
p < 0.05 FWE SFG
L R
IFG Tri
pMTG
IFG orb
Increasing MWE cohesion strength (PMI) Decreasing MWE cohesion strength (PMI)
SMA Precuneus
Positive and negative correlation with PMI - English
Question Teams Expriment fMRI Results Next steps
Parsème - Blois Murielle Fabre
-> PMI as a proxy of lexical cohesiveness
The results show an overlap between the significant effect for decreasing MWE cohesiveness and Bottom-up parser action count in left IFG and posterior temporal lobe. Highly cohesive MWEs implicate the Precuneus and the SMA, suggesting that only truly lexicalized linguistic expressions rely on these areas rather than traditional frontal and temporal nodes of the language network.
1 - PMI : word-by-word computational measure of lexical cohesiveness —> cognitively plausible computational measure of the balance between compositionality versus cohesiveness in MWEs
Our study is showing that this association measure in MWEs produces a neuro-cognitive effect during naturalistic story listening.
2 - Effect of less cohesive MWEs and phase-structure building effect (BU) We observe an overlap between the significant effect for decreasing MWE cohesiveness and Bottom-up parser action count in left IFG and posterior temporal lobe.
fMRI Results Summary
The results show an overlap between the significant effect for decreasing MWE cohesiveness and Bottom-up parser action count in left IFG and posterior temporal lobe. Highly cohesive MWEs implicate the Precuneus and the SMA, suggesting that only truly lexicalized linguistic expressions rely on these areas rather than traditional frontal and temporal nodes of the language network.
Parsème - Blois Murielle Fabre
Expérience Enseignement Recherche Next steps Profil
Question Expriment fMRI Results Next steps
Parsing vs. MWEs Expriment fMRI Stability of PMIs fMRI results Next steps
Parsing vs. MWEs Identifying Stability of PMIs fMRI results Next steps
1 - PMI : word-by-word computational measure of lexical cohesiveness in French
—> Confirm that PMI is a cognitively plausible computational measure of the balance between compositionality and cohesiveness in MWEs
- Compare French and English on the different degrees of compositionality measured with PMI scores - Compare French and English in terms of morphology of the identified MWEs : Nominal versus Verbal
patterns
2 - Effect of less cohesive MWEs and phase-structure building effect (BU) in French
- Replicate the overlap between the significant effect for decreasing MWE cohesiveness and Bottom-up parser action count in left IFG and posterior temporal lobe.
- Leverage on the open MWEs identified thanks to dependency parsing, and observe if different processes/activation patterns are elicited by open versus contiguous MWEs
Next steps in French
Parsème - Blois Murielle Fabre
Question Expriment fMRI Results Next steps
Parsing vs. MWEs Expriment fMRI Stability of PMIs fMRI results Next steps
Parsing vs. MWEs Identifying Stability of PMIs fMRI results Next steps
Thanks for your attention
Parsème - Blois Murielle Fabre