Engineering Intelligent NLP Applications Using Deep Learning – Part 2
UROP Poster pdf
1. The
EEG
Signature
of
Same
Words
Spoken
by
Different
Speakers
S.
Elahi,
Professor
J.
Brennan
Department
of
LinguisBcs
at
the
University
of
Michigan
Results
• A:
n1
–
early
stages
of
acousBc
processing
à
possible
acousBc
priming
for
same
talk
• B:
p2
–
later
stages
of
acousBc
processing
à
lack
of
acousBc
priming
for
diff
talker
• C:
n400
–
semanBc
processing
à
lack
of
semanBc
priming
for
non
repeated
words
n400
–
semanBc
processing
à
reduced
for
both
repeBBon
condiBons
• D:
possible
p300
à
unexpected
sBmuli
processing
–
with
diff
talker
Conclusion
PaMerns
we
see
at
this
point
and
Bme:
• repeBBon
effects
for
the
same
talker
appear
earlier,
~100
milliseconds
• repeBBon
effects
for
different
talkers
do
not
appear
later,
>300
milliseconds
S4muli
Condi4ons
and
Hypothesized
Priming
Effect
R:
repeat;
NR:
non-‐repeat;
+:
feature
is
repeated
in
sBmuli
pair;
–:
feature
is
not
repeated
in
sBmuli
pair
Data
CollecBon
Materials
Experiment
ParBcipants
Our
study
obtained
data
from
X
parBcipants:
each
subject
was
right-‐handed,
learned
English
as
his
or
her
first
language,
and
possessed
no
severe
neurological
impairments
or
disorders.
Sessions
lasted
for
an
hour
and
a
half
and
all
took
place
in
the
ComputaBonal
NeurolinguisBcs
Lab
in
Lorch
Hall.
Methods
Introduc4on
• Hearing
a
word
affects
an
individual’s
understanding
of
the
word
that
follows.
For
example,
if
one
was
to
hear
“dog”,
a
repeated
uMerance
“dog”
would
be
processed
quicker
than
another
word.
Priming
effects
occur
at
each
linguisBc
level
of
speech
percepBon-‐acousBcally,
phoneBcally,
phonologically,
lexically,
and
conceptually.
• Furthermore,
we
understand
a
given
word
with
apparent
effortlessness
even
when
spoken
by
different
speakers.
Though
the
semanBcs
of
a
given
word
remains
the
same,
the
phoneBc
and
acousBc
aspect
of
the
word
varies
with
speaker
–
yet
we
are
sBll
able
to
understand
that
“dog”
means
dog.
• At
issue
is
how
we
recognize
words
given
acousBc
and
phoneBc
variaBon.
The
prototype
theory
states
that
we
possess
the
same
mental
representaBon
for
a
given
object
and
compare
any
new
sBmuli
of
that
object
to
a
single
“prototype”
we
have
constructed
in
our
head.
AlternaBvely,
the
exemplar
theory
proposes
that
we
categorize
objects
by
comparison
of
any
new
sBmuli
with
instances
of
that
word
we
have
stored
in
memory.
• Recent
research
indicates
that
more
rapid
processing
occurs
when
the
same
word
is
repeated
by
the
same
speaker
as
opposed
to
when
it
is
repeated
by
different
speakers.
We
use
electroencephalography
(EEG)
to
record
charges
in
scalp
electrical
potenBals
caused
by
neural
acBvity
while
the
parBcipants
perform
an
experiment
regarding
this
effect.
We
compare
the
Bme-‐course
of
Event-‐Related
Poten4als
(ERP)
between
condiBons
to
test
for
what
stage
of
processing
is
being
primed.
Word
Non-‐Word
Same
Speaker
Different
Speaker
Same
Speaker
Different
Speaker
R
NR
R
NR
R
NR
R
NR
Acous4cs
+
(+)
–
–
+
(+)
–
–
Phone4cs
+
–
(–)
–
+
–
(–)
–
Phonology
+
–
+
–
+
–
+
–
Seman4cs
+
–
+
–
–
–
–
–
We
used
electroencephalography
(EEG)
to
record
electrical
acBvity
in
the
brain.
We
situated
each
subject
in
a
nylon
cap
containing
61
acBvely
amplified
electrodes
evenly
distributed
across
the
scalp.
These
electrodes
measured
electrical
potenBal,
which
reflected
the
current
generated
by
the
summary
of
corBcal
neurons
in
the
brain.
We
injected
electrolyte
gel
to
minimize
impedance
while
it
is
being
monitored.
Low
impedances
means
the
electrical
connecBon
between
the
skin
and
electrode
are
stabilized.
Hearing
threshold
was
idenBfied
for
each
subject
using
a
1000
Hz
tone.
The
volume
was
set
to
be
45
dB
above
this
threshold.
The
task
of
the
parBcipants
was
to
iden4fy
if
the
s4mulus
was
a
word
or
a
non-‐word
as
quickly
and
accurately
as
possible.
This
indicaBon
was
made
by
pressing
either
the
lei
or
right
buMon
on
a
gamepad
console.
Subjects
were
presented
with
a
total
of
700
words,
separated
into
secBons
containing
50
words
each.
Each
word
was
separated
from
each
other
by
an
inter-‐sBmulus
interval
that
ranged
from
400
-‐600
milliseconds.
The
image
to
the
lei
depicts
the
brain
signals
we
monitor
during
the
experiment.
Abstract
Speech
percepBon
research
shows
that
individuals
with
normal
language
processing
rapidly
understand
the
meaning
of
a
spoken
word
despite
the
fact
the
arBculaBon
and
acousBcs
of
these
words
differ
-‐
someBmes
significantly
-‐
across
speakers
of
the
same
dialect.
While
the
semanBcs
of
the
word
spoken
by
two
separate
speakers
would
remain
the
same,
the
phonology,
phoneBcs,
and
acousBcs
of
the
word
are
different
among
the
two.
In
order
to
contribute
to
the
mapping
of
this
capacity
for
"speech
normalizaBon",
we
test
the
Bming
of
repeBBon
effects
within
and
across
speakers
using
electroencephalography.
This
allows
us
to
idenBfy
the
Bming
of
cogniBve
processes
that
allow
the
brain
to
recognize
that
the
same
word
uMered
by
two
different
speakers
are,
in
fact,
the
same
word.
We
used
electroencephalography
(EEG)
to
record
charges
in
electrical
potenBals
that
are
caused
by
neural
acBvity
while
the
human
subjects
listened
to:
words
or
non-‐words,
either
repeated
or
no
repeated,
and
if
repeated,
by
the
same
speaker
or
by
different
speakers.
The
results
from
this
experiment
will
indicate
how
quickly
we
recognize
two
words
(that
are
the
same)
are
the
same
Experimental
Design
No
rep
grassx
dogA
catA
boMlex
Diff
talk
grassx
dogA
dogB
boMlex
Same
talk
grassx
dogA
dogA
boMlex
Hypothesis
and
Objec4ves
Our
objecBve
is
to
discover
at
what
point
in
4me
is
the
difference
in
acous4cs
and
phone4cs
of
repeated
words,
spoken
by
different
speakers,
factored
out
seman4c
aUribute
of
the
word
manifests
within
our
minds.
We
do
this
by
looking
for
what
brain
signals
were
sensiBve
when
words
were
repeated
(either
by
the
same
speaker
or
by
different
speakers)
compared
to
when
they
were
not
repeated.
We
predict
that
early
components
will
be
reduced
for
within-‐speaker
priming
due
to
repeBBon
effects
on
acousBc
and/or
phoneBc
processing.
In
contrast,
we
predict
that
between-‐speaker
priming
will
only
show
aMenuaBon
on
later
components,
reflecBng
lexical
and/or
conceptual
priming.