Experimental methods to study
early language acquisition.
Anne Christophe
Laboratoire de Sciences Cognitives et Psycholinguistique,
EHESS/CNRS/DEC-ENS, Paris
and Maternité Port-Royal, AP-HP, Université Paris Descartes
11th March 2008, Lisbon.
Outline:
• Series of ERP results on syllable perception in very
young infants: work from Ghislaine Dehaene-Lambertz.
=> discussion of the advantages of the ERP technique
for studying phonetic perception in very young infants
• Illustration of several behavioral techniques for studying
early language acquisition (6-24 months infants)
=> discussion of the advantages and drawbacks of each
of these techniques.
Test ing phonet ic per cept ion in inf ant s (and adult s):
Exper iment al Par adigm: habit uat ion-deshabit uat ion
3000 ms
600 ms
time
Ghislaine
DehaeneLambertz
S1
Ba
Da
Da
Ba
S2
Ba
Da
Da
Ba
S3
Ba
Da
Da
Ba
S4
Ba
Ba Response to a change
Da
Da
Nor malizat ion acr oss dif f er ent speaker s
+
Standard
Deviant
z score
/pa/ vs /ta/
same speaker
standard
deviant
Front
R
/pa/ vs /ta/
different speakers
16 sleeping neonates
-2.5 µv
.001 .05
+2.5 µv
.05 .001
p
Dehaene-Lambertz, G., & Peña, M. (2001). Electrophysiological evidence for automatic phonetic
processing in neonates. NeuroReport, 12, 3155-3158
Methodological comparison:
ERP/ Behavioral technique
• ERP result with the
habituationdishabituation
paradigm:
newborn infants can
discriminate between
pa and ta, even with
varied speakers.
• The same result at the
same age could have
been obtained with
non-nutritive sucking,
same habituationdishabituation
paradigm
ERP advantages: 1. response latency, 400ms
2. less infants needed (16 vs 40 infants),
3. passive technique, sleeping infants OK.
Dif f er ent mismat ch r esponses depending on whet her a
phonet ic or an acoust ic char act er ist ic of t he st imulus
changes
z score
Standard Deviant
+
Syllables
/ba/ vs /ga/
-150
100 200 300 400 500
+
Tones
change of
timbre
-150
100 200 300 400 500
- 10
16 awake three-month-old infants
0
.001.01.05
+ 10
µν
p
.05 .01 .001
Front
R
Methodological comparison:
ERP/ Behavioral technique
• ERP result: 3-montholds are able to
discriminate both
syllables and tones
that differ in timber.
• We could test the same
thing with sucking, same
habituationdishabituation paradigm,
inter-subject experimental
design
ERP advantage:
Discrimination responses are different for syllables
(phonetic distinction) and tones (acoustic distinction): this
tells us that different processes are involved.
(no way we could find that out with a behavioral method).
Studies of single subjects
with neonatal cortical lesions
L
R
L. G., left infarct at birth
tested at 2 weeks of age
Dehaene-Lambertz, G., Pena, M., Christophe,
A., Charolais, A., Landrieu, P. (2004). Phoneme
discrimination in a neonate with a left sylvian
infarct. Brain & Language, 88, 26-38.
L
R
S.D., right infarct at birth
tested at 3 months of age
Normal neonates
Front
-2
R
0
+2
µν
L.G.
Same latency for ERP
for L.G. and controls.
600 ms
Global field power
0
600
1200
1800
2400 ms
Discr iminat ion Response in single subj ect s
Timbre discrimination
standard
deviant
Phoneme discrimination
difference
standard
Control infants
S.D., R lesion
L.G., L lesion
deviant
difference
Methodological conclusion:
ERP and habituation-dishabituation.
• The habituation-dishabituation technique is not based on
intrinsic properties of the electrical responses (e.g. N400 =
semantic processing, P600 = response to syntactic violation).
Rather, it is based on a functional description of the task.
• As a result, it allows us to compare subjects of different ages
(where cortical organisation changes, thus modifying the shape
and latency of electrical components).
• Moreover, one can use it to study brain-damaged individuals:
again, the presence of a brain lesion can drastically change the
shape and topography of the electrical responses, which is not
a problem for this type of analysis.
ERP and single-subject testing:
• With ERPs and a passive task that demands no
attention (works with sleeping newborns), one can test
brain-damaged subjects, even though they would be
completely unable to take part in a behavioral task.
• By repeatedly testing a single pathological individual,
one can compute the same statistics than by testing a
group of control, non-pathological subjects.
At present, I know of no behavioral study that yields
individual results (even with test-retest, which has been
attempted by some colleagues). In addition, they all require
some amount of attention...
Word detection with infants:
a variant of conditioned head-turning.
• Session 1: infants are trained to turn their head for
‘paper’ (background: beacon, target paper)
• Session 2: after warm-up trials, infants are tested on
whole sentences:
– containing paper (12):
[The church] [with the most paper spires] [is heavenly].
– containing both syllables of paper (12):
[The man] [with the least pay] [perspires constantly].
– distractor sentences (24).
Word detection: American infants
10-month-olds
13-month-olds
100%
90%
"...paper..."
80%
"...pay#per..."
70%
60%
50%
40%
30%
20%
80%
70%
60%
50%
40%
30%
20%
10%
10%
0%
0%
Target "paper"
Target "pay"
Target "paper"
-6
target = "paper":
difference 34%, t(23)=7,1, p<10
target = "pay":
difference -7%, t(15)=-1.5, p>0.1
Interaction:
t(38)=5.8, p<10
-6
"...paper..."
"...pay#per..."
90%
Percent head-turns
Percent head-turns
100%
Target "pay"
-11
target = "paper":
difference 56%, t(23)=13.5, p<10
target = "pay":
difference -33%, t(15)=-5.5, p<10
Interaction:
t(38)=12.5, p<10
-14
[The church] [with the most paper spires] [is heavenly].
[The man] [with the least pay] [perspires constantly].
Gout, Christophe & Morgan (2004) Journal of Memory and Language
-4
1
Individual results from word-detection
experiments:
a prime
sig. better
than chance
0.5=chance
-10-month-olds: 21/ 24
infants in expected
direction;
17/24 perform
significantly above chance.
0.5
- 13-month-olds: 23/ 24
infants in expected
direction
20/24 perform
significantly above chance;
0
10-month-olds 13-month-olds
American infants
Gout, Christophe & Morgan (2004) Journal of Memory and Language
Prosodic influence of phi boundary
p
paper
a
115.3
pay] [per…
p
112.2
103.5
p
ay
128.5
197.9
0.0
S1-onset
100.0
er
200.0
S1-rhyme
122.3
] [ p
136.5
300.0
400.0
er
86.3
500.0
S2-onset
Word+phrase -final lengthening: 86 ms or 76%, t(23)=8.6, p<0.001
Word+phrase -initial lengthening: 33 ms or 32%, t(23)=7.5, p<0.001
Vowel onset to vowel onset: 132 ms or 61%, t(23)=8.7, p<0.001
600.0
700.0
S2-rhyme
800.0
Methodological comparison:
ERP/ Behavioral technique
• Could we run the same experiment •
with ERP? Problem: high
sensitivity to the acoustic properties
of the stimuli: paper ≠ pay] [per
therefore potentials are bound to be
delayed in one condition relative to
the other. How to interpret it?
The word detection task
allows us to obtain an explicit
response from infants, for
each sentence presented:
- possible to run an item
analysis
- possible to analyse results
from a single subject.
It is almost impossible to imagine an experimental design
where the acoustic properties of stimuli are controlled across
conditions.
Sanders, L. D., & Neville, H. J. (2000). Lexical, syntactic, and stress-pattern cues for
speech segmentation. Journal of Speech, Language, and Hearing Research, 43, 13011321.
Another technique to test word segmentation:
the Head-Turn Preference Procedure
Jusczyk & Aslin, 1995.
• Familiarization phase: infants listen to two words e.g.
‘bike’ and ‘cup’
• Test phase: infants listen to 4 types of passages, 2
containing the ‘ familiar ’ words (‘ bike ’ and ‘ cup ’)
and 2 containing new words (‘ feet ’ and ‘ dog ’)
• Results: infants typically listen longer to passages
containing familiar words.
Word segmentation experiments with French infants
using the Head-Turn Preference Procedure:
12
*
*
listening time (seconds)
10
8
*
6
4
2
0
a mé r ic a ins
bisyllables,
monosyllables,
7.5 months
7.5 months
exp.2
exp.1
bisyllables,
9 months
exp. 3
bisyllables,
9 months
exp. 4
bisyllables,
monosyllables,
10,5 months
10 months
exp. 6
exp. 5
familiar
Jusczyk &
Aslin, 1995
non-familiar
Gout, A. (2001) PhD thesis.
bisyllables,
11 months
exp. 7
bisyllables,
11 months
exp. 8
monosyllables,
13 months
exp. 9
Word detection: French 16-month-olds
100
90
’balcon’-sentences
percent headturns
80
’bal#con’-sentences
70
60
50
40
30
20
10
0
’balcon’ target
’bal’ target
[La rangée de balcons] [fait face au cloître] [du monastère]
[La grande salle de bal] [confère un air solennel] [au château].
Millotte, S. (2005) PhD thesis
Comparison between French and American
infants in the word-detection task:
90%
"paper-sentences"
"pay#per-sentences"
distractor sentences
80%
Percent head-turns
70%
60%
50%
40%
30%
20%
10%
0%
10-month-olds
(24 infants)
12.5-month-olds
(24 infants)
American infants
12.5-month-olds
(15 infants)
16-month-olds
(11 infants)
French infants
16-month-olds
(20 infants)
Comparison between word detection and
HPP, on paper-sentences:
Gout, Christophe & Morgan (2004) Journal
of Memory and Language
10-month-olds
13-month-olds
100%
100%
"...paper..."
"...pay#per..."
80%
70%
60%
50%
40%
30%
20%
10%
"...paper..."
"...pay#per.
90%
Percent head-turns
Percent head-turns
90%
80%
70%
60%
50%
40%
30%
20%
10%
0%
0%
Target "paper"
Target "pay"
Familiarized with ’paper’ and ’beacon’
Test: ’paper’ and ’beacon’-sentences (familiar)
vs ’pay#per’ and ’bee#con’-sentences (new)
Target "paper"
Target "pay"
Methodological advantages of the word detection task:
• An active task, in which the infant has to pay attention
(whereas you can pass HPP even if half-asleep).
• quantitative results:
=> adequate for comparison between infant groups:
different ages, different maternal languages (better than
“all-or-none” as in HPP)
• statistically robust results;
• possibility to measure latencies;
• possibility to do item analyses;
• potentially individual results:
American 10-month-olds: 17/24 perform significantly above chance.
American 13-month-olds: 20/24 perform significantly above chance;
• Drawbacks: difficult to use, high rejection rate: definitely
not good for pathological subjects.
Aside: why do French infants segment
words later than American infants?
Potential differences:
• Language processing itself: it may be harder to segment
French sentences. For instance SW rhythm in American
may help them focus on units smaller than whole
phonological phrases, and pay attention to words earlier.
• Cultural differences: for instance, Americans use very
exaggerated ‘motherese’ with slow speech rate and lots
of pitch excursions. (Quebec/France? States/Britain?).
• Methodological (not too likely, given replication with
different techniques, and different labs, also results from
Thierry Nazzi, in Paris).
The pointing task:
(Can infants exploit the syntactic category of
a new word to constrain its meaning?)
Video
Audio
Noun group
(control)
Familiarization Apple
turning
Regarde, elle
dase ! (look, it’s
dazzing)
Regarde la
dase !
Test
Montre-moi celle
qui dase ! (show
me the one that
dazzes !)
Montre-moi la
dase !
(stupid
question)
Two apples,
one turns,
one does
something
else
Response: pointing (infants are trained to point beforehand on
known words, both objects and actions)
Results: 23-month-old French infants
5
Mean number of pointing responses
Familiar Action
Novel Action
4
3
2
1
0
Experimental Group
Show me the one
that dazzes!
Control Group
Show me the
dazz!
Savita Bernal (2006). PhD thesis; in collaboration with Jeff Lidz.
Bernal, Lidz, Millotte & Christophe (2007) Syntax constrains the acquisition of verb
meaning. Language Learning and Development
Control group:
Video
Familiarization
Test
Apple turning
Verb group
Regarde, elle
dase !
(look, it’s
dazzing !)
Two apples,
Montre-moi
one turns,
celle qui
one does
dase ! (show
something else me the one
that dazzes !)
Noun group
(control)
Regarde la
dase !
(look at the
dazz !)
Montre-moi la
dase !
(show me the
dazz !)
(stupid question)
Results: 23-month-old French infants
(16 in each group)
5
Individual results: Fam ilar - New
4
5
4
3
3
2
Familiar - New
Mean number of pointing responses
Familiar Action
Novel Action
2
1
1
0
Verb Group
-1
Noun Group
-2
-3
-4
0
-5
Experimental Group
Show me the one
that dazzes!
Control Group
Show me the
dazz!
-6
Groupe
Interaction: F(1,30)=11, p<0.005
Bernal, Lidz, Millotte & Christophe (2007) Syntax constrains the acquisition of verb
meaning. Language Learning and Development.
Same experiment in preferential looking:
Proportion looking time towards familiar action
Control group (noun)
Verb group
Before target word
Savita Bernal (2006) PhD thesis
After target word
Methodological advantages of the pointing task:
• An active task, in which the question asked 'show me the
one that dazzes', allows the experimenter to test the
meaning assigned by the child to the new word.
Because the task is explicit, the response is non-ambiguous
(as opposed to preferential looking)
• quantitative results:
=> adequate for comparison between infant groups (better
than “all-or-none” as in preferential looking)
• statistically robust results (much better than preferential
looking);
• possibility to measure latencies;
• Drawbacks: high rejection rate (some infants never point);
does not work with younger infants. Definitely not good
for pathological subjects.
Can 2-year-olds compute syntactic structure
beyond simple transition probabilities?
Syntactic category
Verb
ƒ
ƒ
ƒ
Correct
Alors elle la m an ge
(Then she eats it)
Incorrect
**La fille prend la m an ge
(The girl takes the eat)
Noun
La poule prend la frais e
(The chicken takes the
straw berry )
**Alors il la frais e
(Then he straw berry it)
No particular task: passive listening
Use of known words only (from CDI questionnaire)
To keep infants’ attention focussed, the speaker is playing with
toy objects (e.g. strawberry) while she tells a short story; only her
face is visible when she utters the test sentences.
Example of Script
Sur ma table, je vois une
girafe (N) qui va à
l’école. Elle regarde
(V) la poule
On my table, I see a
giraffe (N) who goes
to school. She looks
(V) at the hen.
1. Donc la poule la
regarde aussi.
1. So the hen looks at
her too.
(Correct)
(Correct)
2. Pourtant, elle la girafe
très vite!
2. However, she giraffes
it really fast!
(Incorrect)
(Incorrect)
Bernal, S., DehaeneLambertz, G., & Christophe,
A. (submitted). Two-yearolds compute syntactic
structure on-line. PNAS.
Incorrect-Correct
Correct
Savita Bernal (2006). PhD
thesis.
Incorrect
2-year-olds detect
the incorrect
sentences
Incorrect
Correct
**La fille prend la m an ge
(The girl takes the eat)
Alors elle la m an ge
(Then she eats it)
La poule prend la frais e
**Alors il la frais e
(Then he straw berry it) (The hen takes the straw berry )
Distinct neural networks for nouns and verbs
original data
modelized data
Nouns
L
L
-4
R
R
0
L
4
40
80
v
P
-4
R
0
4
Verbs
L
L
R
v
P
pA.m
R
L
R
Bernal, Dehaene-Lambertz & Christophe (submitted). Two-year-olds compute
syntactic structure on-line PNAS.
Methodological comparison:
ERP/ Behavioral technique
• ERPs show that infants
react to the
agrammaticality, even
though the acoustic
signal is similar between
condition (e.g. 'la
mange'), and the
transition probabilities
between pairs of words
are all high.
• Could we test the same thing
with a behavioral technique?
Not easily done… Problem
found in all studies of
syntactic processing: it is
difficult to get young
children to perform a
grammaticality judgement
task (after age 3 years:
'truth-value-judgment task')
Additional result in ERPs: different neural networks
involved for nouns and verbs.
Ccl: Methodological comparison:
ERP/ Behavioral technique
ERPs:
Passive method =>
also works on sleeping or
non-cooperative infants,
also if one cannot find a
way to ask the
experimental question.
On-line method
(information on
processing time)
Behavioral method:
Active task =>
explicit response asked
from subject, facilitates
the interpretation (the
most explicit the
response, the easier the
interpretation).
Most methods give little
information as to
processing time.
Methodological comparison:
ERP/ Behavioral technique
ERPs:
Possible to observe
differences in response
topography or latency,
that can be interpreted.
High sensitivity to surface
properties of the stimuli
=> necessary to control
perfectly stimuli across
conditions (not always
feasible).
Behavioral technique:
’All-or-none response’
(most often); e.g.
discrimination Yes/No.
(Note that some
techniques give
quantitative data).
Different experimental
conditions can be easily
compared; generalization
to new conditions can be
tested.
Take-home messages:
• Never assume that results obtained in another
language should replicate 'by default'. Failure to
replicate may either be due to a methodological
difficulty, or to a genuine difference between
languages.
• Select the methodology depending on the
question asked, not the reverse. Be prepared to
design your own methodology if none of the
available ones are satisfactory.
Use of phonological phrases in
on-line syntactic processing
• Locally ambiguous sentences
• Verb sentence:
[le petit chien] [mord la laisse] [qui le retient]…
(the little dog bites the leash that restrains it)
• Adjective sentence :
[le petit chien mort] [sera enterré demain]…
(the little dead dog will be buried tomorrow…)
Æ
Up to the ambiguous words: same phonemic content
Æ only difference: syntactic structure, and therefore prosodic structure
Are prosodic cues exploited on-line? Yes
Task: abstract word detection; e.g. ’mordre’ (to bite)
respond to verb sentences, refrain from responding to adjective
sentences
Results: fast responses only (given at the end of the ambiguous word)
Informative prosody
Neutral prosody
Nombre moyen de réponses adjectif et verbe données
aux phrases ambiguës (prosodie neutre)
Nombre moyen de réponses adjectif et verbe données
aux phrases ambiguës (prosodie informative)
10
10
8
8
6,0
6
4,8
4,4
4
2,8
6
4
2
2
0
0
Phrases Adjectif
Réponses Adjectif
Phrases Verbe
Réponses Verbe
3,6
3,5
Phrases Adjectif
Réponses Adjectif
Millotte, René, Wales & Christophe (in revision) JEP: LMC
2,8
3,4
Phrases Verbe
Réponses Verbe
Comparaison ERP/comportement:
• Pourrait-on faire la même
•
expérience en ERP? Pas facile..
Pb d'acoustique, mord ≠ mort
(à cause de position dans la
structure prosodique)
Mais aussi: quelle différence
de potentiel attendrait-on pour
un adjectif vs un verbe???
La tâche de détection de
mot permet de savoir en
temps réel s'ils ont accédé
au nom ou au verbe.
Même si on imaginait un moyen de contrôler les propriétés
acoustiques des stimuli, il reste le problème du dessin
expérimental. Réfléchi à : phrases 'chimérique' (début de l'une
et fin de l'autre), on attend une erreur…