I
nd
o
ne
s
ia
n J
o
urna
l o
f
E
lect
rica
l En
g
ineering
a
nd
Co
m
pu
t
er
Science
Vo
l.
3
6
,
No
.
2
,
N
o
v
em
b
e
r
20
24
,
p
p
.
1
2
3
4
~
1
241
I
SS
N:
2
502
-
4
7
52
,
DOI
:
1
0
.
1
1
5
9
1
/ijeecs.v
3
6
.
i
2
.
p
p
1
2
3
4
-
1
2
4
1
1234
J
o
ur
na
l ho
m
ep
a
g
e
:
h
ttp
:
//ij
ee
cs
.
ia
esco
r
e.
co
m
Exploring
t
h
e pot
ential o
f
Distil
BE
RT
a
rchite
ctur
e f
o
r
a
utom
a
tic
ess
a
y
sco
ring
t
a
sk
So
um
ia
I
k
is
s
1
,
Na
j
im
a
Da
o
ud
i
2
,
M
a
na
r
Abo
urez
q
2
,
M
o
s
t
a
f
a
B
ella
f
k
ih
1
1
N
a
t
i
o
n
a
l
I
n
st
i
t
u
t
e
o
f
P
o
s
t
s a
n
d
T
e
l
e
c
o
mm
u
n
i
c
a
t
i
o
n
s,
R
a
b
a
t
,
M
o
r
o
c
c
o
2
S
c
h
o
o
l
o
f
I
n
f
o
r
ma
t
i
o
n
S
c
i
e
n
c
e
s,
R
a
b
a
t
,
M
o
r
o
c
c
o
Art
icle
I
nfo
AB
S
T
RAC
T
A
r
ticle
his
to
r
y:
R
ec
eiv
ed
J
an
17
,
2
0
2
4
R
ev
is
ed
Au
g
3
,
2
0
2
4
Acc
ep
ted
A
ug
11
,
2
0
2
4
Au
to
m
a
ti
c
a
ss
e
ss
m
e
n
t
o
f
writi
n
g
e
ss
a
y
s,
o
r
th
e
p
ro
c
e
ss
o
f
u
si
n
g
c
o
m
p
u
ters
to
e
v
a
lu
a
te
a
n
d
a
ss
ig
n
g
ra
d
e
s
to
writt
e
n
te
x
t,
is
v
e
ry
n
e
e
d
e
d
i
n
t
h
e
e
d
u
c
a
ti
o
n
sy
ste
m
a
s
a
n
a
lt
e
rn
a
ti
v
e
to
re
d
u
c
e
h
u
m
a
n
b
u
r
d
e
n
a
n
d
ti
m
e
c
o
n
su
m
p
ti
o
n
,
e
sp
e
c
ially
fo
r
larg
e
-
sc
a
le
tes
ts.
T
h
is
tas
k
h
a
s
re
c
e
iv
e
d
m
o
re
a
tt
e
n
ti
o
n
i
n
th
e
las
t
fe
w
y
e
a
rs,
b
e
in
g
o
n
e
o
f
th
e
m
a
jo
r
u
se
s
fo
r
n
a
tu
ra
l
lan
g
u
a
g
e
p
ro
c
e
ss
in
g
(NLP
)
.
Tra
d
it
i
o
n
a
l
a
u
t
o
m
a
ti
c
sc
o
rin
g
sy
ste
m
s
t
y
p
ica
ll
y
re
l
y
o
n
h
a
n
d
c
ra
fted
fe
a
tu
re
s,
wh
e
re
a
s
re
c
e
n
t
st
u
d
ies
h
a
v
e
u
se
d
d
e
e
p
n
e
u
ra
l
n
e
two
r
k
s.
S
i
n
c
e
th
e
a
d
v
e
n
t
o
f
tran
sfo
rm
e
rs,
p
re
-
train
e
d
lan
g
u
a
g
e
m
o
d
e
ls
h
a
v
e
p
e
rfo
rm
e
d
we
ll
in
m
a
n
y
d
o
w
n
stre
a
m
tas
k
s.
We
u
ti
li
z
e
th
e
Ka
g
g
le
b
e
n
c
h
m
a
rk
i
n
g
a
u
to
m
a
ted
stu
d
e
n
t
a
ss
e
ss
m
e
n
t
p
rize
d
a
tas
e
t
to
fin
e
-
t
u
n
e
t
h
e
p
re
-
train
e
d
Disti
lBE
RT
in
th
re
e
d
iffere
n
t
sc
e
n
a
rio
s,
a
n
d
we
c
o
m
p
a
re
re
su
lt
s
with
th
e
e
x
ist
in
g
n
e
u
ra
l
n
e
two
rk
-
b
a
se
d
a
p
p
r
o
a
c
h
e
s
to
a
c
h
iev
e
imp
ro
v
e
d
p
e
rfo
rm
a
n
c
e
in
th
e
a
u
to
m
a
ti
c
e
ss
a
y
sc
o
ri
n
g
tas
k
.
W
e
u
ti
l
ize
q
u
a
d
ra
ti
c
we
i
g
h
ted
Ka
p
p
a
(QWK)
a
s
th
e
m
a
in
m
e
tri
c
to
e
v
a
l
u
a
te
th
e
p
e
rfo
rm
a
n
c
e
o
f
o
u
r
p
r
o
p
o
se
d
m
e
th
o
d
.
Re
su
lt
s
sh
o
w t
h
a
t
f
in
e
-
tu
n
i
n
g
Dis
ti
lBE
RT
g
iv
e
s g
o
o
d
re
su
lt
s,
e
sp
e
c
ially
wit
h
th
e
sc
e
n
a
rio
o
f
trai
n
in
g
a
ll
p
a
ra
m
e
ters
,
wh
ich
a
c
h
iev
e
0
.
9
0
o
f
QWK
a
n
d
o
u
t
p
e
rfo
rm
n
e
u
ra
l
n
e
two
r
k
m
o
d
e
l
s.
K
ey
w
o
r
d
s
:
Au
to
m
atic
ess
ay
s
co
r
in
g
Dis
til
B
E
R
T
Fin
e
-
tu
n
in
g
Natu
r
al
lan
g
u
ag
e
p
r
o
ce
s
s
in
g
T
r
an
s
f
o
r
m
e
r
s
T
h
is i
s
a
n
o
p
e
n
a
c
c
e
ss
a
rticle
u
n
d
e
r th
e
CC B
Y
-
SA
li
c
e
n
se
.
C
o
r
r
e
s
p
o
nd
ing
A
uth
o
r
:
So
u
m
ia
I
k
is
s
Natio
n
al
I
n
s
titu
te
o
f
Po
s
ts
an
d
T
elec
o
m
m
u
n
icatio
n
s
Ma
d
in
at
E
l I
r
f
a
n
e
,
R
ab
at,
Mo
r
o
cc
o
E
m
ail: so
u
m
ia.
ik
is
s
9
7
@
g
m
ail.
co
m
1.
I
NT
RO
D
UCT
I
O
N
I
n
h
ig
h
er
e
d
u
ca
tio
n
,
e
-
lear
n
in
g
s
y
s
tem
s
ar
e
b
ec
o
m
in
g
m
o
r
e
an
d
m
o
r
e
co
m
m
o
n
,
as
u
n
iv
e
r
s
ities
an
d
in
s
titu
tio
n
s
en
lar
g
e
p
r
o
v
is
io
n
s
an
d
m
o
r
e
s
tu
d
en
ts
g
et
en
r
o
lled
.
Ass
es
s
m
en
t
an
d
ev
alu
atio
n
o
f
s
tu
d
en
ts
’
k
n
o
wled
g
e
a
n
d
ca
p
ab
ilit
ies
is
an
in
teg
r
al
p
ar
t
o
f
ed
u
ca
tio
n
al
s
y
s
tem
s
;
th
u
s
,
th
e
ef
f
e
ctiv
en
ess
o
f
an
e
-
ass
es
s
m
en
t
s
y
s
tem
is
cr
u
cial
a
n
d
q
u
ite
ap
p
ea
lin
g
to
ac
a
d
em
i
c
in
s
titu
tio
n
s
to
r
ed
u
ce
tim
e
c
o
n
s
u
m
p
tio
n
a
n
d
th
e
wo
r
k
lo
ad
o
f
h
u
m
an
r
a
ter
s
.
I
n
th
e
ed
u
ca
tio
n
al
en
v
ir
o
n
m
en
t,
ass
es
s
m
en
t
is
g
en
er
ally
d
o
n
e
t
h
r
o
u
g
h
e
x
am
s
an
d
test
s
;
h
en
ce
,
th
e
p
u
r
p
o
s
e
o
f
s
u
ch
an
e
-
ass
ess
m
en
t
s
y
s
tem
is
to
au
to
m
atica
lly
e
v
alu
ate
as
s
ig
n
m
en
ts
an
d
test
s
ac
co
m
p
lis
h
ed
b
y
th
e
lear
n
e
r
s
an
d
d
eter
m
in
e
an
ap
p
r
o
p
r
iate
s
co
r
e
.
T
h
e
ass
ig
n
m
en
ts
ca
n
b
e
g
iv
en
in
d
if
f
er
e
n
t
ty
p
es;
th
e
m
o
s
t
co
m
m
o
n
ar
e
eith
er
s
elec
tio
n
an
d
f
ill
-
in
-
th
e
-
b
lan
k
task
s
,
wh
ich
ar
e
r
elativ
ely
ea
s
y
to
im
p
lem
en
t
as
th
e
an
s
wer
s
ar
e
u
s
u
ally
p
r
ed
eter
m
in
e
d
,
o
r
th
e
y
ca
n
b
e
in
th
e
f
o
r
m
o
f
a
s
h
o
r
t
an
s
wer
o
r
ess
ay
,
wh
i
ch
ex
am
i
n
e
h
i
g
h
er
-
o
r
d
e
r
ab
ilit
ies
s
u
ch
as
lo
g
ical
th
i
n
k
in
g
an
d
cr
itical
r
ea
s
o
n
in
g
.
I
n
ess
ay
wr
itin
g
test
s
,
a
lear
n
er
wr
ites
a
p
iece
o
f
tex
t
in
r
esp
o
n
s
e
to
a
g
iv
en
to
p
ic,
ca
lled
a
p
r
o
m
p
t,
an
d
h
u
m
an
r
ater
s
ev
alu
ate
an
d
g
r
ad
e
t
h
o
s
e
ess
ay
s
.
T
h
e
p
r
o
c
ess
is
ex
p
en
s
iv
e
an
d
tim
e
-
co
n
s
u
m
in
g
a
n
d
is
n
o
t
alwa
y
s
c
o
n
s
is
ten
t
am
o
n
g
an
d
with
in
r
ater
s
.
Au
to
m
atic
ass
es
s
m
en
t
o
f
wr
itin
g
ess
ay
s
aim
s
t
o
en
ab
le
co
m
p
u
ter
s
to
an
aly
ze
an
d
ass
ig
n
g
r
ad
es
to
lear
n
er
s
’
s
ess
ay
s
au
t
o
m
atica
lly
with
o
u
t
h
u
m
a
n
in
te
r
f
er
en
ce
[
1
]
.
I
t
is
a
h
ig
h
ly
ch
allen
g
i
n
g
task
th
at
r
e
q
u
ir
es
ex
p
lo
itin
g
m
u
ltip
le
asp
ec
ts
th
at
in
f
lu
en
ce
th
e
q
u
ality
o
f
ess
ay
s
,
s
u
ch
as c
o
n
ten
t,
r
ele
v
an
ce
,
a
n
d
co
h
er
en
ce
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
d
o
n
esian
J
E
lec
E
n
g
&
C
o
m
p
Sci
I
SS
N:
2
5
0
2
-
4
7
52
E
xp
lo
r
in
g
th
e
p
o
ten
tia
l o
f D
is
tilBE
R
T a
r
ch
itectu
r
e
fo
r
a
u
to
ma
tic
ess
a
y
s
co
r
in
g
ta
s
k
(
S
o
u
mi
a
I
kiss
)
1235
T
h
e
ess
ay
as
s
ess
m
en
t
ta
s
k
h
a
s
b
ee
n
th
e
s
u
b
ject
o
f
ex
ten
s
iv
e
in
v
esti
g
atio
n
.
An
d
is
o
n
e
o
f
th
e
m
o
s
t
ch
allen
g
in
g
task
s
in
n
at
u
r
al
la
n
g
u
ag
e
p
r
o
ce
s
s
in
g
(
NL
P).
T
h
e
q
u
ality
o
f
s
u
c
h
a
m
o
d
el
is
g
en
er
ally
r
elate
d
to
its
ab
ilit
y
to
e
x
tr
ac
t
u
s
ef
u
l
f
ea
tu
r
es
th
at
r
ep
r
esen
t
th
e
m
ax
im
u
m
p
o
s
s
ib
le
ch
a
r
ac
ter
is
tics
o
f
t
h
e
ess
ay
,
wh
ich
lead
s
to
ac
cu
r
ately
ev
alu
atin
g
an
d
p
r
ed
ictin
g
th
e
co
r
r
esp
o
n
d
in
g
s
co
r
e.
Su
ch
f
ea
tu
r
es
m
ig
h
t
b
e
ex
tr
ac
ted
eith
er
m
an
u
ally
(
h
an
d
cr
af
ted
f
ea
tu
r
e
s
)
o
r
au
to
m
atica
lly
u
s
in
g
n
eu
r
al
n
etwo
r
k
-
b
ased
tec
h
n
iq
u
es
[
2
]
.
T
h
e
ea
r
ly
wo
r
k
s
m
ain
ly
u
s
e
f
ea
tu
r
es
d
esig
n
ed
m
an
u
ally
b
y
h
u
m
a
n
e
x
p
er
ts
b
ased
o
n
p
r
io
r
k
n
o
wled
g
e
of
li
n
g
u
is
tics
,
s
u
ch
as
g
r
am
m
ar
,
te
rm
f
r
eq
u
e
n
cy
,
an
d
s
y
n
tax
[
3
]
,
[
4
]
.
Pro
ject
ess
ay
g
r
ad
e
(
PEG)
[
5
]
is
am
o
n
g
th
e
f
ir
s
t
co
m
p
u
ter
p
r
o
g
r
a
m
s
f
o
r
g
r
a
d
in
g
.
I
t
u
s
ed
lin
ea
r
r
eg
r
ess
io
n
to
p
r
ed
ict
th
e
s
co
r
e
o
f
a
g
iv
e
n
tex
t b
ased
o
n
th
e
s
ty
le
o
f
ess
ay
s
an
d
lin
g
u
is
tic
f
ea
tu
r
es
o
n
ly
.
Ov
er
4
0
y
ea
r
s
f
o
llo
win
g
p
ag
e
’
s
in
itial
wo
r
k
,
m
an
y
au
to
-
s
co
r
in
g
ess
ay
s
o
f
twar
e
p
r
o
g
r
a
m
s
h
av
e
a
s
tr
ik
in
g
r
es
em
b
lan
ce
to
PEG.
I
n
1
9
9
9
,
th
e
in
tellig
en
t
ess
ay
ass
e
s
s
o
r
(
I
E
A)
u
s
ed
laten
t
s
em
an
tic
an
aly
s
is
(
L
SA)
f
o
r
s
em
an
tic
s
im
ilar
ity
.
Ma
n
y
o
t
h
e
r
co
m
m
e
r
cial
ass
ess
m
en
t
s
y
s
t
em
s
h
av
e
c
o
m
e
o
u
t,
n
am
ely
th
e
elec
tr
o
n
ic
ess
ay
r
ater
(E
-
R
ater
)
a
n
d
C
-
R
ater
[
1
]
,
wh
ich
h
a
v
e
b
e
e
n
u
s
ed
f
o
r
T
OE
FL
an
d
GR
E
ex
am
in
atio
n
s
.
R
ec
en
tly
,
au
to
m
atic
ess
ay
-
s
co
r
in
g
r
esear
ch
h
as
b
ee
n
b
ased
o
n
n
e
u
r
al
n
et
wo
r
k
s
an
d
u
s
es
en
d
-
to
-
en
d
m
o
d
els
[
6
]
.
T
h
ese
m
o
d
els
ar
e
m
ain
ly
b
ased
o
n
wo
r
d
em
b
ed
d
in
g
tech
n
iq
u
es
to
lea
r
n
h
ig
h
-
d
im
en
s
io
n
al
f
ea
tu
r
es
f
r
o
m
d
ata
an
d
ca
p
tu
r
e
th
e
m
o
s
t
p
o
s
s
ib
le
in
f
o
r
m
atio
n
.
I
n
2
0
1
2
,
Kag
g
le
ca
m
e
u
p
with
ASAP.
W
i
s
h
is
a
lar
g
e
-
s
ca
le
d
ataset
th
at
co
m
es
with
th
e
in
cr
ea
s
e
o
f
d
ee
p
lear
n
in
g
tech
n
iq
u
es
th
at
n
ee
d
h
u
g
e
d
ata
to
ac
h
iev
e
g
o
o
d
p
er
f
o
r
m
a
n
ce
.
Sin
ce
th
e
n
,
th
er
e
h
as
b
ee
n
a
r
e
n
ewe
d
in
ter
est
in
au
to
m
atic
ass
ess
m
en
t,
an
d
th
e
m
o
s
t
r
ec
en
t
wo
r
k
s
u
s
e
th
e
ASAP
d
ataset
to
m
ea
s
u
r
e
th
eir
p
er
f
o
r
m
an
ce
.
Mo
r
e
r
ec
en
t
r
esear
ch
h
as
ap
p
lied
n
eu
r
al
n
etwo
r
k
s
f
o
r
au
to
m
atic
s
co
r
in
g
[
7
]
-
[
1
3
]
.
T
a
g
h
ip
o
u
r
an
d
Ng
[
1
0
]
u
s
e
b
o
th
r
e
cu
r
r
e
n
t
n
e
u
r
a
l
n
etwo
r
k
s
(
R
NN)
an
d
co
n
v
o
lu
tio
n
al
n
eu
r
al
n
etwo
r
k
s
(
C
NN)
,
th
ey
u
s
e
C
NN
to
ex
tr
ac
t
lo
ca
l
in
f
o
r
m
atio
n
f
r
o
m
tex
t
an
d
lo
n
g
s
h
o
r
t
-
ter
m
m
em
o
r
y
(
L
STM
)
t
o
g
en
e
r
ate
tem
p
o
r
al
co
n
te
x
t
a
n
d
lo
n
g
h
is
to
r
y
.
L
ater
o
n
,
u
s
ed
C
NN
with
L
STM
an
d
p
er
f
o
r
m
e
d
th
e
o
u
tp
u
t sco
r
e
with
th
e
atten
tio
n
m
ec
h
an
is
m
[
1
1
]
.
T
h
e
tr
ad
itio
n
al
s
y
s
tem
s
th
at
a
d
ap
ted
h
an
d
c
r
af
ted
f
ea
tu
r
es
a
ch
iev
ed
g
o
o
d
r
esu
lts
an
d
r
e
v
o
lu
tio
n
ized
th
e
f
ield
o
f
au
t
o
m
atic
s
co
r
in
g
.
Ho
wev
er
,
th
ey
t
y
p
ically
u
s
e
co
m
p
licated
,
h
a
n
d
cr
af
te
d
f
ea
tu
r
es
f
r
o
m
a
s
m
all
am
o
u
n
t o
f
d
ata
[
1
4
]
-
[
1
6
]
.
T
h
e
s
e
tech
n
iq
u
es o
f
f
ea
tu
r
e
ex
tr
ac
tio
n
ca
n
b
e
m
o
d
if
ie
d
,
m
ad
e
s
im
p
le
to
ex
p
lain
an
d
ad
ap
t
n
ew
f
ea
t
u
r
es,
y
et
t
h
ey
ca
n
b
e
u
n
p
er
f
o
r
m
e
d
to
u
n
d
er
s
tan
d
s
o
m
e
d
ee
p
in
f
o
r
m
atio
n
in
th
e
tex
t.
Neu
r
al
n
etwo
r
k
m
o
d
els
ar
e
m
ain
ly
b
ased
o
n
wo
r
d
em
b
ed
d
i
n
g
tec
h
n
iq
u
es
to
lear
n
h
ig
h
-
d
im
en
s
io
n
al
f
ea
tu
r
es
f
r
o
m
d
ata
an
d
ca
p
tu
r
e
th
e
m
o
s
t
p
o
s
s
ib
le
in
f
o
r
m
atio
n
.
Nev
e
r
th
el
ess
,
th
ey
n
ee
d
an
en
o
r
m
o
u
s
a
m
o
u
n
t
o
f
an
n
o
tated
d
ata
to
r
etr
iev
e
p
r
o
f
o
u
n
d
ly
s
em
an
tic
f
ea
tu
r
es
f
r
o
m
th
e
tex
t
d
u
r
in
g
tr
ai
n
in
g
,
as
lo
w
d
ata
ca
n
r
esu
lt
in
p
o
o
r
p
er
f
o
r
m
an
ce
.
Vaswan
i
et
a
l.
[
1
7
]
ca
m
e
o
u
t
wit
h
tr
a
n
s
f
o
r
m
e
r
s
,
wh
ich
ar
e
n
e
u
r
al
n
etwo
r
k
s
th
at
ar
e
b
u
ilt
o
n
th
e
atten
tio
n
m
ec
h
an
is
m
[
1
8
]
.
T
h
e
tr
an
s
f
o
r
m
er
a
r
ch
itectu
r
e
a
llo
ws
p
ar
allelis
m
in
d
ata
p
r
o
ce
s
s
in
g
an
d
av
o
id
s
v
an
is
h
in
g
g
r
a
d
ien
ts
.
Nu
m
er
o
u
s
p
r
e
-
tr
ain
ed
lan
g
u
ag
e
m
o
d
els
h
av
e
b
ee
n
d
ev
elo
p
e
d
u
s
in
g
th
is
ar
ch
itectu
r
e,
an
d
th
ey
h
av
e
s
h
o
wn
p
r
o
m
is
in
g
o
u
tco
m
es in
a
v
ar
iety
o
f
d
o
wn
s
tr
ea
m
NL
P a
p
p
licatio
n
s
b
y
f
in
e
-
tu
n
in
g
p
r
e
-
tr
ain
e
d
m
o
d
els
s
u
ch
as
b
id
ir
ec
tio
n
al
en
co
d
er
r
ep
r
esen
tatio
n
s
f
r
o
m
tr
an
s
f
o
r
m
er
s
(
B
E
R
T
)
.
Few
wo
r
k
s
o
n
au
to
m
atic
s
co
r
in
g
h
av
e
u
s
ed
th
is
m
o
d
el;
[
1
9
]
-
[
2
1
]
h
av
e
f
i
n
e
-
tu
n
e
d
th
e
B
E
R
T
m
o
d
el
f
o
r
th
e
task
.
Nev
er
th
eless
,
it h
as n
o
t
b
ee
n
ef
f
ec
tiv
ely
ap
p
lied
t
o
s
u
r
p
ass
o
th
er
d
ee
p
lea
r
n
in
g
m
o
d
els
lik
e
L
STM
in
th
e
au
to
m
ated
ess
ay
s
co
r
in
g
d
o
m
ain
.
T
h
e
B
E
R
T
m
o
d
el
s
u
f
f
er
s
f
r
o
m
f
i
x
ed
in
p
u
t
len
g
th
an
d
s
ize
lim
itatio
n
s
,
wo
r
d
p
iece
em
b
e
d
d
in
g
p
r
o
b
lem
s
,
a
n
d
c
o
m
p
u
tatio
n
al
co
m
p
lex
ities
.
I
n
th
is
wo
r
k
,
w
e
in
v
esti
g
ate
th
e
p
o
ten
tial
o
f
f
in
e
-
tu
n
in
g
a
p
r
e
-
tr
ain
ed
m
o
d
el,
n
am
ely
Dis
til
B
E
R
T
[
2
2
]
,
a
m
o
r
e
co
m
p
ac
t,
f
aster
tr
an
s
f
o
r
m
er
-
b
ased
ar
c
h
itectu
r
e
cr
ea
ted
b
y
ap
p
ly
in
g
k
n
o
wl
ed
g
e
d
is
till
atio
n
[
2
3
]
t
o
th
e
B
E
R
T
ar
ch
itectu
r
e,
wh
ich
o
f
f
er
s
a
m
o
r
e
s
ca
lab
le
an
d
ec
o
n
o
m
ical
alter
n
ativ
e
to
B
E
R
T
.
T
h
e
r
em
ain
in
g
s
ec
tio
n
s
o
f
t
h
is
p
ap
er
ar
e
o
r
g
a
n
ized
ac
co
r
d
i
n
g
to
th
e
f
o
llo
win
g
s
tr
u
ctu
r
e:
I
n
th
e
n
e
x
t
s
ec
tio
n
2
,
ex
p
lain
o
u
r
m
et
h
o
d
an
d
r
esear
c
h
d
esig
n
.
I
n
th
is
p
ar
t,
we
d
etail
s
tep
-
by
-
s
tep
o
u
r
p
r
o
p
o
s
ed
ar
ch
itectu
r
e
an
d
d
escr
ib
e
o
u
r
n
u
m
er
ical
s
tu
d
y
,
in
clu
d
i
n
g
d
ata
ac
q
u
is
itio
n
,
ev
alu
a
tio
n
m
etr
ics,
an
d
im
p
lem
en
tatio
n
.
T
h
e
ex
p
er
im
en
tal
f
in
d
in
g
s
an
d
an
aly
s
is
ar
e
s
h
o
wn
in
s
ec
tio
n
5
.
I
n
co
n
clu
s
io
n
,
s
ec
tio
n
6
g
iv
es a
s
u
m
m
ar
y
a
n
d
p
r
o
v
i
d
es su
g
g
esti
o
n
s
f
o
r
f
u
r
th
e
r
r
esear
ch
an
d
p
er
s
p
ec
tiv
es.
2.
M
E
T
H
O
D
I
n
th
is
s
ec
tio
n
,
we
f
ir
s
t
d
e
s
cr
ib
e
th
e
ar
ch
itectu
r
e
an
d
d
esig
n
o
f
th
e
p
r
o
p
o
s
ed
m
o
d
el.
T
h
en
,
we
p
r
o
v
id
e
d
etails o
n
o
u
r
n
u
m
er
ical
s
tu
d
y
,
in
clu
d
in
g
d
ata
ac
q
u
is
iti
o
n
,
ex
p
er
im
e
n
tal
s
ettin
g
s
,
an
d
m
etr
ics u
s
ed
f
o
r
p
e
r
f
o
r
m
an
ce
ev
al
u
atio
n
.
2
.
1
.
M
o
del descript
io
n
T
h
e
f
o
llo
win
g
s
u
b
-
s
ec
tio
n
d
escr
ib
es
th
e
p
r
o
p
o
s
ed
ap
p
r
o
ac
h
,
s
tar
tin
g
with
an
o
v
er
v
i
ew
o
f
th
e
p
er
tain
ed
Dis
til
B
E
R
T
m
o
d
el
an
d
th
en
g
iv
in
g
th
e
d
etailed
a
r
ch
itectu
r
e
o
f
o
u
r
m
eth
o
d
t
o
a
d
ap
t
Dis
til
B
E
R
T
to
o
u
r
task
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
5
0
2
-
4
7
52
In
d
o
n
esian
J
E
lec
E
n
g
&
C
o
m
p
Sci
,
Vo
l.
3
6
,
No
.
2
,
No
v
em
b
er
20
24
:
1
2
3
4
-
1
2
4
1
1236
2
.
1
.
1
.
Dis
t
il
B
E
RT
m
o
del
Dis
til
B
E
R
T
i
s
a
p
r
e
-
tr
ain
e
d
lan
g
u
ag
e
m
o
d
el
c
r
ea
ted
b
y
ap
p
ly
in
g
th
e
m
ec
h
an
is
m
o
f
k
n
o
wled
g
e
d
is
till
atio
n
[
2
2
]
to
th
e
la
r
g
e
m
o
d
el
o
f
B
E
R
T
[
2
4
]
.
T
h
e
g
en
er
al
ar
ch
itectu
r
e
is
s
im
ilar
to
B
E
R
T
,
wh
er
e
th
e
b
ase
u
n
it
is
th
e
tr
a
n
s
f
o
r
m
er
,
esp
e
cially
th
e
en
c
o
d
er
b
lo
ck
,
wh
er
e
th
e
tr
a
n
s
f
o
r
m
e
r
is
an
en
co
d
er
a
n
d
d
ec
o
d
e
r
.
Fig
u
r
e
1
d
ep
icts
th
e
g
en
er
al
ar
ch
ite
ctu
r
e
o
f
th
e
B
E
R
T
en
c
o
d
er
,
wh
er
e
t
h
e
m
ain
b
l
o
ck
u
s
es
a
s
elf
-
atten
tio
n
m
ec
h
an
is
m
to
g
en
e
r
ate
a
co
n
t
ex
tu
al
em
b
ed
d
in
g
r
e
p
r
esen
tatio
n
.
As
with
th
e
o
r
ig
in
al
B
E
R
T
,
Dis
til
B
E
R
T
wa
s
also
tr
ain
ed
o
n
d
ata
co
llected
f
r
o
m
th
e
E
n
g
lis
h
W
ik
ip
ed
ia
an
d
th
e
T
o
r
o
n
to
B
o
o
k
C
o
r
p
u
s
.
T
h
e
m
ain
d
if
f
er
en
ce
b
etwe
en
th
em
lies
in
th
e
to
k
e
n
-
ty
p
e
e
m
b
ed
d
in
g
s
an
d
p
o
o
le
r
s
th
at
ar
e
o
m
itted
f
r
o
m
Dis
tilB
E
R
T
.
I
n
ad
d
itio
n
,
th
e
d
is
till
ed
m
o
d
el
u
s
es
h
u
g
e
b
atch
es
with
th
e
h
elp
o
f
g
r
ad
i
en
t
ac
cu
m
u
latio
n
u
s
in
g
d
y
n
a
m
ic
m
ask
in
g
in
p
lace
o
f
th
e
m
ask
in
g
u
s
e
d
in
th
e
o
r
ig
in
al
B
E
R
T
an
d
with
o
u
t
th
e
n
ex
t
s
en
ten
ce
p
r
e
d
ictio
n
(
NS
P)
o
b
jectiv
e
d
u
r
i
n
g
tr
ain
in
g
.
T
h
e
n
u
m
b
er
o
f
tr
a
n
s
f
o
r
m
er
lay
er
s
(
e
n
co
d
er
s
)
in
t
h
e
B
E
R
T
b
ase
(
1
2
lay
er
s
)
h
as
b
ee
n
r
ed
u
ce
d
to
s
ix
.
T
h
u
s
,
th
e
d
is
till
ed
v
er
s
io
n
is
ap
p
r
o
x
im
ately
s
m
aller
with
6
6
m
illi
o
n
p
ar
am
ete
r
s
,
wh
er
e
th
e
n
u
m
b
er
is
r
ed
u
ce
d
to
n
ea
r
ly
h
alf
(
4
0
%)
an
d
f
aster
with
6
0
% wh
ile
p
r
eser
v
in
g
m
o
r
e
th
an
9
5
% o
f
B
E
R
T
p
er
f
o
r
m
an
ce
.
T
h
is
m
ak
es
Dis
ti
B
E
R
T
an
id
ea
l c
h
o
ice
to
p
er
f
o
r
m
o
u
r
s
tu
d
y
.
Fig
u
r
e
1.
T
h
e
g
e
n
er
al
ar
ch
itec
tu
r
e
o
f
B
E
R
T
2
.
1
.
2
.
P
ro
po
s
ed
a
rc
hite
ct
ure
I
n
th
is
s
u
b
s
ec
tio
n
,
we
i
n
tr
o
d
u
ce
th
e
o
v
e
r
all
ar
ch
itectu
r
e
o
f
th
e
p
r
o
p
o
s
ed
m
o
d
el.
Fig
u
r
e
2
p
r
esen
ts
a
s
ch
em
atic
r
ep
r
esen
tatio
n
o
f
th
e
en
tire
ar
c
h
itectu
r
e.
T
h
e
m
o
d
el
p
er
f
o
r
m
s
f
in
e
-
t
u
n
in
g
o
n
th
e
Dis
tilB
E
R
T
m
o
d
el
to
tack
le
th
e
p
r
o
b
lem
o
f
ess
ay
s
co
r
in
g
.
T
h
e
m
o
d
el
r
ec
ei
v
es
as
an
in
p
u
t
X,
w
h
ich
r
ep
r
esen
ts
an
ess
ay
(
a
s
eq
u
en
ce
o
f
wo
r
d
s
)
.
T
h
e
in
p
u
t
s
eq
u
en
ce
s
ar
e
to
k
e
n
ized
u
s
in
g
Dis
til
B
E
R
T
to
k
en
ize
an
d
co
n
v
er
ted
to
a
s
et
o
f
em
b
e
d
d
in
g
v
ec
t
o
r
s
.
T
h
e
n
ap
p
ly
th
e
tr
an
s
f
o
r
m
er
e
n
co
d
e
r
an
d
u
s
e
a
s
elf
-
atten
tio
n
m
ec
h
an
is
m
to
l
ea
r
n
th
e
co
n
tex
tu
alize
d
em
b
ed
d
i
n
g
s
.
L
ater
,
th
e
co
n
te
x
tu
alize
d
em
b
e
d
d
in
g
s
ar
e
c
o
n
ca
ten
ated
i
n
to
a
s
in
g
le
v
ec
to
r
an
d
u
s
ed
as
in
p
u
t
t
o
a
r
e
g
r
ess
io
n
l
ay
er
ad
d
ed
o
n
to
p
to
f
in
e
-
tu
n
e
th
e
p
r
e
-
tr
ain
ed
Dis
tilB
E
R
T
o
n
th
e
ess
ay
s
co
r
in
g
task
an
d
p
r
ed
ict
a
s
co
r
e
f
o
r
ea
ch
ess
ay
.
I
n
wh
at
f
o
llo
ws,
we
p
r
o
ce
ed
with
th
e
m
ai
n
co
m
p
o
n
en
ts
o
f
th
e
m
o
d
el
in
d
etail.
Fig
u
r
e
2
.
T
h
e
p
r
o
p
o
s
ed
ess
ay
s
co
r
in
g
m
o
d
el
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
d
o
n
esian
J
E
lec
E
n
g
&
C
o
m
p
Sci
I
SS
N:
2
5
0
2
-
4
7
52
E
xp
lo
r
in
g
th
e
p
o
ten
tia
l o
f D
is
tilBE
R
T a
r
ch
itectu
r
e
fo
r
a
u
to
ma
tic
ess
a
y
s
co
r
in
g
ta
s
k
(
S
o
u
mi
a
I
kiss
)
1237
a)
I
n
p
u
t
e
m
b
ed
d
in
g
s
:
th
e
to
k
en
iz
er
in
v
o
l
v
es
b
r
ea
k
in
g
d
o
wn
th
e
in
p
u
t
s
eq
u
e
n
ce
in
to
u
n
i
q
u
e
to
k
en
s
to
en
a
b
le
th
e
m
o
d
el
to
u
n
d
e
r
s
tan
d
th
e
m
ea
n
in
g
o
f
th
e
tex
t.
W
e
u
s
e
th
e
Dis
til
B
E
R
T
to
k
en
izer
to
s
p
lit
th
e
in
p
u
t
ess
ay
in
to
to
k
en
s
an
d
ad
d
t
h
e
s
p
ec
ial
to
k
en
s
[
C
L
S]
at
th
e
f
ir
s
t
p
o
s
itio
n
an
d
[
SEP]
at
th
e
en
d
o
f
th
e
s
eq
u
en
ce
.
T
h
e
t
o
k
en
s
ar
e
u
s
ed
to
g
en
e
r
ate
v
ec
to
r
s
o
f
w
o
r
d
,
s
eg
m
en
t,
an
d
p
o
s
itio
n
al
em
b
e
d
d
in
g
s
.
T
h
e
latter
ar
e
t
h
en
s
u
m
m
e
d
u
p
to
o
n
e
em
b
ed
d
in
g
v
ec
t
o
r
f
o
r
ea
ch
to
k
en
t
o
b
e
p
ass
ed
to
t
h
e
tr
an
s
f
o
r
m
er
en
co
d
er
.
b)
T
r
an
s
f
o
r
m
e
r
en
co
d
er
:
th
e
co
n
tex
tu
al
in
f
o
r
m
atio
n
is
ex
tr
a
cted
v
ia
p
r
e
-
tr
ain
i
n
g
u
s
in
g
t
h
e
p
r
e
-
tr
ain
e
d
Dis
til
B
E
R
T
en
co
d
er
to
c
o
n
v
e
r
t
in
p
u
t
to
k
en
s
i
n
to
c
o
n
tex
tu
al
em
b
ed
d
in
g
s
.
T
h
e
en
c
o
d
er
co
m
p
r
is
es
a
s
tack
o
f
s
ix
la
y
er
s
,
ea
ch
o
f
wh
ich
co
n
tain
s
a
f
ee
d
-
f
o
r
war
d
n
eu
r
al
n
etwo
r
k
an
d
a
m
u
lti
-
h
ea
d
s
elf
-
a
tten
tio
n
m
ec
h
an
is
m
th
at
aim
s
to
r
ec
o
g
n
ize
c
o
n
tex
tu
al
co
n
n
ec
tio
n
s
b
etwe
en
to
k
en
s
.
Fig
u
r
e
3
d
escr
ib
es
a
s
in
g
l
e
lay
er
o
f
t
h
e
en
co
d
er
co
m
p
o
n
e
n
t.
Fig
u
r
e
3
.
A
s
in
g
le
lay
er
o
f
th
e
en
co
d
er
c
o
m
p
o
n
en
t
c)
Fin
e
-
tu
n
in
g
o
n
ess
ay
s
co
r
in
g
:
to
h
an
d
le
o
u
r
task
o
f
ess
ay
s
co
r
in
g
,
we
f
in
e
-
tu
n
e
d
th
e
last
lay
er
o
f
th
e
en
co
d
er
b
lo
ck
,
wh
ic
h
o
u
tp
u
ts
7
6
8
d
im
e
n
s
io
n
al
h
id
d
e
n
u
n
it
s
.
T
h
e
class
if
icatio
n
lay
er
is
r
ep
lace
d
with
a
d
en
s
e
lay
er
,
wh
ich
c
o
n
s
is
ts
o
f
a
f
u
lly
co
n
n
ec
ted
n
etwo
r
k
an
d
is
th
en
f
ed
to
t
h
e
r
eg
r
ess
io
n
lay
er
to
p
r
e
d
ict
ess
ay
s
co
r
es.
2
.
2
.
E
x
perim
ent
s
I
n
th
is
s
u
b
-
s
ec
tio
n
,
we
in
tr
o
d
u
ce
a
d
escr
ip
tio
n
o
f
th
e
e
x
p
e
r
im
en
t’
s
p
r
o
ce
d
u
r
e,
in
clu
d
in
g
th
e
u
s
ed
d
ataset,
p
er
f
o
r
m
an
ce
m
etr
ics,
an
d
ex
p
er
im
en
tal
s
etu
p
.
2
.
2
.
1
.
Da
t
a
a
cquis
it
io
n
T
o
im
p
lem
en
t
t
h
e
ess
ay
s
co
r
i
n
g
task
an
d
v
alid
ate
th
e
p
r
o
p
o
s
ed
m
o
d
el,
th
e
ex
p
e
r
im
en
ts
wer
e
ca
r
r
ied
o
u
t
u
s
in
g
th
e
au
t
o
m
ated
s
tu
d
e
n
t
ass
es
s
m
en
t
p
r
ize
(
ASAP
)
d
ataset
(
h
ttp
s
://ww
w.
k
ag
g
le.
co
m
/c/asap
-
ae
s
/d
ata
)
,
a
f
r
eq
u
e
n
tly
u
s
ed
c
o
r
p
u
s
as
a
r
ef
er
en
ce
d
ataset
f
o
r
task
s
in
v
o
lv
in
g
a
u
to
m
atic
ess
ay
g
r
ad
i
n
g
.
I
t
is
a
lar
g
e
-
s
ca
le
d
ataset
in
tr
o
d
u
ce
d
a
n
d
s
u
p
p
o
r
ted
b
y
th
e
Hew
lett,
W
illi
am
,
an
d
Flo
r
a
Hew
lett
Fo
u
n
d
atio
n
as
p
ar
t
o
f
a
Kag
g
le
co
m
p
etitio
n
in
2
0
1
2
.
I
t
is
co
m
p
o
s
ed
o
f
1
2
,
9
7
6
lab
eled
ess
ay
s
with
eig
h
t
p
r
o
m
p
ts
.
T
ab
le
1
s
h
o
ws
m
o
r
e
d
etails
ab
o
u
t th
e
ASAP d
ataset.
T
ab
le
1
.
Statis
tic
d
etails o
f
Kag
g
le
’
s
ASAP d
ataset
P
e
o
m
t
s se
t
N
u
mb
e
r
o
f
e
ss
a
y
s
Th
e
me
a
n
l
e
n
g
t
h
o
f
e
ssa
y
s
S
c
o
r
e
r
a
n
g
e
1
1
,
7
8
3
3
5
0
1
-
12
2
1
,
8
0
0
3
5
0
1
-
6
3
1
,
7
2
6
1
5
0
0
-
3
4
1
,
7
7
2
1
5
0
0
-
3
5
1
,
8
0
5
1
5
0
0
-
4
6
1
,
8
0
0
1
5
0
0
-
4
7
1
,
5
6
9
2
5
0
0
-
30
8
7
2
3
6
5
0
0
-
30
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
5
0
2
-
4
7
52
In
d
o
n
esian
J
E
lec
E
n
g
&
C
o
m
p
Sci
,
Vo
l.
3
6
,
No
.
2
,
No
v
em
b
er
20
24
:
1
2
3
4
-
1
2
4
1
1238
2
.
2
.
2
.
E
v
a
lua
t
io
n m
e
t
ric
T
o
m
ea
s
u
r
e
th
e
p
er
f
o
r
m
an
ce
,
we
u
s
e
q
u
ad
r
atic
weig
h
ted
Kap
p
a
(
QW
K
)
,
wh
ich
is
th
e
s
tan
d
ar
d
ev
alu
atio
n
m
etr
ic
f
o
r
t
h
e
ASAP
co
m
p
etitio
n
,
o
f
f
icially
d
esig
n
ated
b
y
Kag
g
le.
T
h
e
m
etr
i
c
is
h
ig
h
ly
s
en
s
itiv
e
to
in
co
r
r
ec
t
p
r
ed
ictio
n
s
,
wh
ic
h
ca
n
h
elp
to
e
v
alu
ate
th
e
c
o
n
s
is
ten
cy
o
f
s
u
ch
m
o
d
el
p
r
ed
ictio
n
s
.
I
t
m
ea
s
u
r
es
th
e
ag
r
ee
m
e
n
t
b
etwe
en
two
r
ater
s
,
s
p
ec
if
ically
b
etwe
en
th
e
h
u
m
an
r
ater
s
’
s
co
r
e
an
d
th
e
s
y
s
tem
’
s
p
r
ed
icted
s
co
r
e.
T
y
p
ically
,
th
e
QW
K
s
co
r
e
f
alls
b
etwe
en
0
an
d
1
,
wh
ich
in
d
icate
s
n
o
co
n
s
is
ten
cy
b
etwe
en
r
ater
s
an
d
co
m
p
lete
co
n
s
is
ten
cy
,
r
esp
ec
ti
v
ely
.
QW
K
is
ca
lcu
lated
u
s
in
g
in
th
e
(
1
)
:
=
1
−
∑
,
,
,
∑
,
,
,
(
1
)
w
h
er
e
th
e
weig
h
ts
,
o
b
s
er
v
ed
s
co
r
es,
an
d
p
r
ed
icted
s
co
r
es
ar
e
r
e
p
r
esen
ted
b
y
th
e
m
atr
ices
w,
O,
an
d
E
,
r
esp
ec
tiv
ely
.
T
h
e
n
u
m
b
e
r
o
f
e
s
s
ay
s
th
at
ea
r
n
a
s
co
r
e
o
f
i
f
r
o
m
th
e
f
ir
s
t
r
ater
an
d
a
s
co
r
e
o
f
j
f
r
o
m
th
e
s
ec
o
n
d
r
ater
is
r
ep
r
esen
ted
b
y
th
e
v
al
u
e
o
f
,
.
T
h
e
weig
h
ts
ar
e
:
,
=
(
−
)
2
(
−
1
)
2
(
2
)
w
h
er
e
N
r
ep
r
esen
ts
th
e
to
tal
p
o
ten
tial
s
co
r
e.
T
h
e
o
u
ter
p
r
o
d
u
ct
o
f
two
s
co
r
e
h
is
to
g
r
am
v
ec
t
o
r
s
is
u
s
ed
to
co
m
p
u
te
m
atr
ix
E
.
Ma
tr
i
x
O
an
d
m
atr
i
x
E
ar
e
th
en
n
o
r
m
al
ized
s
o
th
at
th
eir
s
u
m
s
ar
e
eq
u
al.
An
d
af
ter
th
at,
th
e
QW
K
s
co
r
e
is
ca
lcu
lated
.
2
.
2
.
3
.
I
m
plem
ent
a
t
i
o
n set
t
ing
s
W
e
u
s
ed
th
e
GPU
h
ar
d
war
e
ac
ce
ler
ato
r
p
latf
o
r
m
f
r
o
m
Go
o
g
le
C
o
lab
an
d
th
e
T
en
s
o
r
f
l
o
w
en
v
ir
o
n
m
en
t
to
b
u
ild
o
u
r
m
o
d
el.
Fo
r
o
u
r
ex
p
er
im
en
ts
,
we
in
s
tan
tiated
th
e
tr
an
s
f
o
r
m
er
m
o
d
el,
n
am
ely
“D
is
tib
er
t
-
b
ase
-
u
n
ca
s
ed
”
f
r
o
m
th
e
h
u
g
g
in
g
f
ac
e
lib
r
ar
y
,
a
n
d
m
o
d
i
f
ied
th
e
ar
ch
itectu
r
e
f
o
r
th
e
ess
ay
s
co
r
in
g
task
in
o
r
d
er
to
f
in
e
-
t
u
n
e
th
e
weig
h
ts
u
s
in
g
t
h
e
lab
eled
A
SAP
d
ataset.
As
in
p
u
t
t
o
th
e
m
o
d
el,
we
p
e
r
f
o
r
m
to
k
en
izatio
n
f
o
r
ea
ch
ess
ay
f
r
o
m
th
e
ASAP
d
ataset
u
s
in
g
th
e
Dis
til
B
E
R
T
o
k
en
izer
class
,
wh
ich
co
n
v
e
r
ts
ea
ch
s
en
ten
ce
in
to
a
s
eq
u
en
ce
o
f
t
o
k
en
s
an
d
th
en
a
s
eq
u
e
n
ce
o
f
n
u
m
b
er
s
.
[
C
L
S]
an
d
[
SEP]
to
k
en
s
w
er
e
in
s
er
ted
b
ef
o
r
e
an
d
af
ter
ea
ch
s
eq
u
en
ce
,
r
esp
ec
tiv
ely
.
Ou
r
to
k
e
n
izatio
n
f
u
n
ctio
n
en
d
s
u
p
r
etu
r
n
in
g
two
ar
r
ay
s
:
a
n
ar
r
ay
o
f
I
Ds
th
at
r
ep
r
esen
t
th
e
s
eq
u
en
ce
o
f
wo
r
d
s
in
a
n
ess
ay
en
co
d
ed
as
s
eq
u
en
ce
s
o
f
n
u
m
b
er
s
,
an
d
a
n
ar
r
ay
o
f
atten
tio
n
-
m
ask
A
wis
h
is
a
b
in
ar
y
s
eq
u
en
ce
th
at
h
elp
s
th
e
m
o
d
el
eith
er
p
ay
atten
tio
n
to
th
e
co
r
r
esp
o
n
d
in
g
I
D
o
r
n
o
t.
T
h
e
two
ar
r
ay
s
ar
e
th
en
f
ed
to
o
u
r
m
o
d
el
as
in
p
u
ts
.
T
o
b
u
ild
o
u
r
m
o
d
el
ar
c
h
itectu
r
e,
we
b
eg
in
b
y
im
p
o
r
tin
g
“
D
is
tilb
er
t
-
b
ase
-
u
n
ca
s
ed
”
f
r
o
m
th
e
tr
an
s
f
o
r
m
e
r
p
ac
k
ag
e
an
d
m
ak
in
g
s
o
m
e
ch
an
g
es to
ar
g
u
m
en
ts
in
th
e
“
Dis
tilB
er
t
C
o
n
f
ig
u
r
atio
n
”
class
b
y
i
n
cr
ea
s
in
g
d
r
o
p
o
u
t
an
d
atten
tio
n
-
d
r
o
p
o
u
t
f
r
o
m
0
.
1
to
0
.
2
.
t
h
en
we
in
itialize
th
e
b
ase
Dis
t
ilB
E
R
T
u
s
in
g
th
e
“
T
FDis
til
B
er
tMo
d
el
”
class
w
ith
n
o
ad
d
ed
class
if
icatio
n
h
ea
d
o
n
to
p
.
W
e
te
m
p
o
r
ar
ily
f
r
ee
ze
Dis
tilb
er
t
’
s
p
r
e
-
tr
ai
n
ed
weig
h
ts
to
p
r
ev
en
t
th
em
f
r
o
m
u
p
d
atin
g
d
u
r
in
g
th
e
tr
ain
in
g
o
f
o
u
r
ad
d
e
d
lay
er
s
.
W
e
tr
y
s
ev
er
al
f
in
e
-
tu
n
in
g
s
ce
n
ar
io
s
,
as
s
h
o
wn
in
T
ab
le
2
.
First,
we
k
ee
p
th
e
weig
h
ts
o
f
th
e
p
r
e
-
tr
ain
ed
Dis
tilB
er
t
f
r
o
ze
n
,
an
d
we
tr
ai
n
o
n
ly
th
e
a
d
d
ed
o
u
tp
u
t
lay
er
s
.
W
e
tr
ain
with
o
n
e
d
en
s
e
lay
er
f
o
r
s
ix
ep
o
ch
s
in
th
e
f
ir
s
t
s
ce
n
ar
io
,
th
en
with
th
r
ee
lay
er
s
o
f
1
0
0
,
3
2
,
an
d
1
f
o
r
f
i
v
e
ep
o
c
h
s
in
th
e
s
ec
o
n
d
.
An
o
th
er
ca
s
e
o
f
f
in
e
-
tu
n
in
g
th
at
we
e
x
p
er
im
en
ted
with
was
to
tr
ain
all
lay
er
s
.
T
o
d
o
th
at,
we
m
ad
e
th
e
p
a
r
am
eter
s
tr
ain
ab
le
an
d
tr
ai
n
ed
with
a
s
m
all
lear
n
in
g
r
ate
o
f
1
e
-
5
f
o
r
2
ep
o
ch
s
.
As
we
b
u
ilt
o
u
r
m
o
d
el
ar
ch
itectu
r
e,
f
o
r
tr
ain
in
g
p
u
r
p
o
s
es,
we
co
n
s
id
e
r
ed
6
0
%
o
f
th
e
d
ata
f
o
r
tr
ain
in
g
,
2
0
%
f
o
r
v
alid
atio
n
,
an
d
2
0
%
f
o
r
test
in
g
to
m
ea
s
u
r
e
t
h
e
p
er
f
o
r
m
an
ce
a
f
ter
tr
ain
in
g
.
W
e
u
s
e
Ad
am
o
p
tim
izer
,
r
ec
tifie
d
lin
ea
r
u
n
it
(
Re
L
U)
f
o
r
th
e
ac
tiv
atio
n
f
u
n
cti
o
n
,
a
n
d
3
2
b
at
ch
s
izes.
T
ab
le
2.
E
x
p
er
im
en
t scen
a
r
io
s
D
i
st
i
l
B
e
r
t
e
x
p
e
r
i
me
n
t
s
P
a
r
a
me
t
e
r
s
F
i
n
e
t
u
n
i
n
g
1
F
r
e
e
z
e
b
a
se
=
t
r
u
e
,
e
p
o
c
h
s
=
6
,
l
e
a
r
n
i
n
g
r
a
t
e
=
5
e
-
5
,
a
d
d
e
d
l
a
y
e
r
s
=
d
e
n
se
1
F
i
n
e
t
u
n
i
n
g
2
F
r
e
e
z
e
b
a
se
=
t
r
u
e
,
e
p
o
c
h
s
=
5
,
l
e
a
r
n
i
n
g
r
a
t
e
=
5
e
-
5
,
a
d
d
e
d
l
a
y
e
r
s
=
d
e
n
se
1
0
0
a
n
d
d
e
n
se
3
2
a
n
d
d
e
n
se
1
F
i
n
e
t
u
n
i
n
g
3
F
r
e
e
z
e
b
a
se
=
f
a
l
se
,
e
p
o
c
h
s
=
2
,
l
e
a
r
n
i
n
g
r
a
t
e
=
1
e
-
5
,
a
d
d
e
d
l
a
y
e
r
s
=
d
e
n
se
1
3.
RE
SU
L
T
S AN
D
D
I
SCU
SS
I
O
N
I
n
th
is
s
ec
tio
n
,
we
ev
alu
ate
th
e
p
er
f
o
r
m
an
ce
o
f
ea
ch
o
f
th
e
s
ce
n
ar
io
s
we
tr
ied
in
o
u
r
e
x
p
er
im
en
ts
.
Fig
u
r
e
4
s
h
o
ws
th
e
r
esu
lts
o
f
QW
K
f
o
r
th
e
th
r
ee
f
in
e
-
tu
n
in
g
m
o
d
els.
I
n
t
h
is
p
ar
t
,
we
ev
alu
ate
th
e
p
er
f
o
r
m
an
ce
o
f
ea
c
h
o
f
th
e
s
ce
n
ar
io
s
we
tr
ied
in
o
u
r
ex
p
e
r
im
en
ts
.
Fig
u
r
e
4
s
h
o
ws
th
e
r
es
u
lts
o
f
QW
K
f
o
r
th
e
th
r
ee
f
in
e
-
tu
n
in
g
m
o
d
els.
I
t
is
clea
r
th
at
f
in
e
-
tu
n
in
g
3
,
wh
ich
co
n
s
is
ts
o
f
u
p
d
atin
g
all
lay
e
r
s
o
f
th
e
p
r
e
-
tr
ain
ed
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
d
o
n
esian
J
E
lec
E
n
g
&
C
o
m
p
Sci
I
SS
N:
2
5
0
2
-
4
7
52
E
xp
lo
r
in
g
th
e
p
o
ten
tia
l o
f D
is
tilBE
R
T a
r
ch
itectu
r
e
fo
r
a
u
to
ma
tic
ess
a
y
s
co
r
in
g
ta
s
k
(
S
o
u
mi
a
I
kiss
)
1239
D
is
tib
er
t,
g
iv
es
b
etter
r
esu
lt
s
,
esp
ec
ially
in
ter
m
s
o
f
Q
W
K,
wh
ich
ac
h
iev
es
9
0
%.
Fin
e
-
tu
n
in
g
b
y
tr
ain
in
g
all
th
e
p
ar
am
ete
r
s
o
f
a
p
r
etr
ain
ed
l
ar
g
e
la
n
g
u
a
g
e
m
o
d
el
t
o
p
e
r
f
o
r
m
b
etter
o
n
n
ew
task
s
th
e
m
o
d
el
wasn
’
t
tr
ai
n
ed
o
n
.
Fo
r
f
in
etu
n
in
g
1
a
n
d
2
,
wh
ich
co
n
s
is
t
o
f
f
r
ee
zin
g
th
e
p
r
e
-
tr
ain
ed
lay
er
s
an
d
o
n
l
y
tr
ain
in
g
th
e
n
ewly
ad
d
ed
lay
er
s
.
W
e
ca
n
s
ee
th
at
f
in
etu
n
in
g
2
with
two
lay
e
r
s
o
n
to
p
with
1
0
0
an
d
3
2
,
r
esp
ec
tiv
el
y
,
an
d
a
r
eg
r
ess
io
n
lay
er
ac
h
ie
v
es
8
4
%.
QW
K
,
wh
ich
is
h
i
g
h
er
th
an
7
9
%
ac
h
ie
v
ed
b
y
f
in
etu
n
i
n
g
1
,
wh
ich
is
s
im
ilar
to
f
in
etu
n
in
g
1
,
b
u
t
th
e
d
i
f
f
er
en
ce
h
er
e
is
to
f
r
ee
ze
th
e
p
a
r
am
eter
s
o
f
Di
s
tilb
er
t
an
d
tr
ain
o
n
ly
th
e
o
u
t
p
u
t
r
eg
r
ess
io
n
lay
e
r
.
T
h
ese
r
esu
lts
s
h
o
w
th
at
f
in
e
-
tu
n
in
g
with
m
o
r
e
la
y
er
s
g
iv
es
b
etter
p
er
f
o
r
m
an
ce
;
h
o
wev
er
,
it
is
wo
r
th
m
en
tio
n
in
g
th
at
tr
ain
in
g
with
o
u
t f
r
ee
zin
g
in
cr
ea
s
es tim
e
co
n
s
u
m
p
tio
n
.
Fig
u
r
e
4
.
Per
f
o
r
m
an
c
e
in
ter
m
s
o
f
QW
K
an
d
ac
cu
r
ac
y
C
o
m
p
ar
in
g
o
u
r
f
in
d
in
g
s
with
b
aselin
e
m
o
d
els,
t
h
e
r
esu
lts
ar
e
lis
ted
in
T
ab
le
3
.
Fro
m
th
e
tab
le
,
we
ca
n
s
ee
th
at
E
ASE
(
SVR
)
[
4
]
,
wh
ich
a
p
p
lies
r
eg
r
ess
io
n
tech
n
iq
u
es
to
a
s
et
o
f
h
a
n
d
-
e
n
g
in
ee
r
ed
f
ea
tu
r
es,
h
as
lo
w
p
er
f
o
r
m
an
ce
in
co
m
p
ar
is
o
n
to
n
eu
r
al
n
etwo
r
k
-
b
a
s
ed
m
eth
o
d
s
.
T
h
is
ex
p
lain
s
t
h
e
p
o
ten
tial
o
f
d
ee
p
lear
n
in
g
to
ex
tr
ac
t
u
s
ef
u
l
f
ea
tu
r
es
th
at
h
elp
ass
ig
n
a
s
co
r
e
th
at
is
clea
r
ly
co
n
s
is
ten
t
with
a
h
u
m
an
o
n
e.
C
o
m
b
in
in
g
W
o
r
d
2
V
ec
with
n
eu
r
al
n
etwo
r
k
s
im
p
r
o
v
es
p
er
f
o
r
m
an
ce
o
v
e
r
m
eth
o
d
s
th
at
e
m
p
lo
y
d
ir
ec
tly
C
NN
an
d
/o
r
L
STM
[
1
0
]
,
[
2
5
]
.
R
2
B
E
R
T
[
2
1
]
p
r
o
p
o
s
es
a
f
in
e
-
t
u
n
in
g
o
f
th
e
p
r
e
-
tr
ain
e
d
lan
g
u
ag
e
m
o
d
el
B
E
R
T
.
I
t
co
m
b
i
n
es
r
eg
r
ess
io
n
alo
n
g
with
r
an
k
in
g
u
s
in
g
m
u
ltip
le
lo
s
s
es
o
f
th
e
s
am
e
task
.
T
h
e
m
o
d
el
g
ets
g
o
o
d
r
esu
lts
b
u
t
f
ails
to
o
v
er
co
m
e
a
ll
n
eu
r
al
n
etwo
r
k
-
b
ased
m
o
d
els.
T
h
e
o
b
s
er
v
ed
QW
K
g
o
tten
b
y
Dis
ti
B
E
R
T
ca
n
o
u
tp
er
f
o
r
m
o
th
er
b
aselin
e
m
o
d
els,
esp
ec
ially
f
o
r
f
in
etu
n
in
g
3
,
in
d
icatin
g
th
e
ef
f
ec
tiv
e
n
ess
o
f
p
r
e
-
t
r
ain
ed
m
o
d
els
f
o
r
au
to
m
atic
ess
ay
s
co
r
in
g
task
s
r
eg
ar
d
in
g
th
eir
c
ap
ac
ity
to
ca
p
tu
r
e
d
ee
p
l
y
s
em
an
tic
in
f
o
r
m
atio
n
f
r
o
m
t
h
e
tex
t u
s
in
g
th
e
atten
tio
n
m
ec
h
an
is
m
.
Ov
er
all,
r
esu
lts
s
h
o
wed
th
at
th
e
p
r
e
-
t
r
ain
ed
d
is
till
atio
n
o
f
B
E
R
T
is
f
ea
s
ib
le
f
o
r
th
e
au
to
m
atic
ass
ess
m
en
t ta
s
k
an
d
ca
n
co
r
r
ec
tl
y
ass
ess
es
s
ay
s
.
T
ab
le
3
.
P
er
f
o
r
m
an
ce
r
esu
lts
ac
h
iev
ed
b
y
d
if
f
er
en
t m
o
d
els b
ased
o
n
QW
K
M
o
d
e
l
s
Q
W
K
EA
S
E
(
S
V
R
)
[
4
]
0
.
6
9
LSTM
_
C
N
N
[
1
0
]
0
.
7
6
R
2
B
E
R
T
[
2
1
]
0
.
7
9
S
K
I
P
F
LO
W
[
2
5
]
0
.
7
6
F
i
n
e
t
u
n
i
n
g
1
0
.
7
9
F
i
n
e
t
u
n
i
n
g
2
0
.
8
4
F
i
n
e
t
u
n
i
n
g
3
0
.
9
0
4.
CO
NCLU
SI
O
N
AND
F
U
T
U
RE
WO
RK
E
v
alu
atin
g
ess
ay
s
au
to
m
atica
lly
h
elp
s
to
o
v
er
co
m
e
th
e
b
i
as
o
f
h
u
m
an
e
v
alu
atio
n
an
d
m
ak
es
th
e
p
r
o
ce
s
s
m
o
r
e
co
n
s
is
ten
t.
T
h
u
s
,
a
m
o
d
el
t
h
at
o
v
er
co
m
es
th
e
s
h
o
r
tco
m
in
g
s
an
d
im
p
r
o
v
es
th
e
g
r
ad
i
n
g
p
er
f
o
r
m
an
ce
is
cr
itical.
Ab
o
v
e
,
we
ad
d
r
ess
th
e
p
o
ten
tial
o
f
a
p
p
ly
in
g
t
h
e
t
r
an
s
f
o
r
m
er
ar
ch
it
ec
tu
r
e
to
th
e
ess
ay
-
s
co
r
in
g
task
.
Sp
ec
if
ically
,
we
f
in
e
-
tu
n
e
d
Dis
til
B
E
R
T
,
wh
ich
is
a
co
m
p
r
ess
ed
an
d
f
aster
v
a
r
ian
t
o
f
th
e
B
E
R
T
m
o
d
el
f
o
r
ess
ay
s
co
r
in
g
task
s
.
W
e
tr
ied
d
if
f
er
en
t
im
p
lem
e
n
tatio
n
s
ce
n
ar
io
s
o
f
f
i
n
etu
n
in
g
u
s
in
g
th
e
Kag
g
le
ASAP
d
ata
s
et,
an
d
r
esu
lts
s
h
o
w
th
at
tr
ain
in
g
Dis
ti
l
B
E
R
T
b
y
u
p
d
atin
g
all
p
ar
am
eter
s
g
iv
es
b
etter
r
esu
lts
th
an
f
in
etu
n
in
g
with
a
f
r
ee
zin
g
b
a
s
e
b
u
t
c
o
m
es
with
a
n
i
n
cr
ea
s
ed
co
s
t
in
ter
m
s
o
f
tim
e.
T
o
s
h
o
w
th
e
b
en
ef
it
o
f
Dis
til
B
E
R
T
o
v
er
th
e
p
r
ev
io
u
s
wo
r
k
s
,
we
co
m
p
ar
e
it with
d
if
f
er
en
t b
aselin
e
m
o
d
els f
r
o
m
th
e
s
tate
o
f
th
e
ar
t in
ter
m
s
o
f
th
e
QW
K
m
etr
ic,
a
n
d
o
u
r
m
eth
o
d
s
h
o
ws
g
o
o
d
r
esu
lts
o
v
er
th
em
,
esp
ec
ially
in
c
o
m
p
ar
is
o
n
t
o
n
e
u
r
al
n
etwo
r
k
-
b
ased
m
eth
o
d
s
.
Ou
r
m
eth
o
d
wo
r
k
s
well
f
o
r
p
r
o
m
p
t
-
s
p
ec
if
ic
ess
ay
s
co
r
in
g
task
s
,
as
lab
eled
d
ata
is
79
84
90
0
50
100
fin
e
tu
n
i
n
g
1
fin
e
tu
n
i
n
g
2
fin
e
tu
n
i
n
g
3
V
A
L
UE
MO
DE
L
Q
W
K
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
5
0
2
-
4
7
52
In
d
o
n
esian
J
E
lec
E
n
g
&
C
o
m
p
Sci
,
Vo
l.
3
6
,
No
.
2
,
No
v
em
b
er
20
24
:
1
2
3
4
-
1
2
4
1
1240
av
ailab
le
wh
ile
tr
ain
in
g
.
Ou
r
tar
g
et
f
o
r
th
e
n
ex
t
wo
r
k
is
to
g
en
er
alize
th
e
m
o
d
el
f
o
r
u
n
s
e
en
d
ata
as
a
cr
o
s
s
-
p
r
o
m
p
t
task
,
wh
ich
is
a
r
ea
l
-
wo
r
ld
ca
s
e
wh
er
e
lab
eled
d
at
a
ar
e
n
o
t
alwa
y
s
av
ailab
le
a
n
d
r
eq
u
i
r
e
h
u
g
e
e
f
f
o
r
t
an
d
tim
e
to
p
r
ep
ar
e
.
RE
F
E
R
E
NC
E
S
[
1
]
S
.
V
a
l
e
n
t
i
,
F
.
N
e
r
i
,
a
n
d
A
.
C
u
c
c
h
i
a
r
e
l
l
i
,
“
A
n
o
v
e
r
v
i
e
w
o
f
c
u
r
r
e
n
t
r
e
se
a
r
c
h
o
n
a
u
t
o
ma
t
e
d
e
ss
a
y
g
r
a
d
i
n
g
,
”
J
o
u
rn
a
l
o
f
I
n
f
o
rm
a
t
i
o
n
T
e
c
h
n
o
l
o
g
y
Ed
u
c
a
t
i
o
n
:
Re
s
e
a
r
c
h
,
v
o
l
.
2
,
p
p
.
3
1
9
–
3
3
0
,
2
0
0
3
,
d
o
i
:
1
0
.
2
8
9
4
5
/
3
3
1
.
[
2
]
M
.
A
.
H
u
s
sei
n
,
H
.
H
a
ssa
n
,
a
n
d
M
.
N
a
ssef,
“
A
u
t
o
m
a
t
e
d
l
a
n
g
u
a
g
e
e
ss
a
y
sco
r
i
n
g
sy
s
t
e
ms
:
a
l
i
t
e
r
a
t
u
r
e
r
e
v
i
e
w
,
”
P
e
e
rJ
C
o
m
p
u
t
e
r
S
c
i
e
n
c
e
,
v
o
l
.
2
0
1
9
,
n
o
.
8
,
p
.
e
2
0
8
,
A
u
g
.
2
0
1
9
,
d
o
i
:
1
0
.
7
7
1
7
/
p
e
e
r
j
-
c
s.
2
0
8
.
[
3
]
A
.
K
.
.
M
a
y
a
,
J.
N
a
z
u
r
a
,
a
n
d
B
.
L.
M
u
r
a
l
i
d
h
a
r
a
,
“
R
e
c
e
n
t
t
r
e
n
d
s
i
n
a
n
sw
e
r
scri
p
t
e
v
a
l
u
a
t
i
o
n
–
a
l
i
t
e
r
a
t
u
r
e
s
u
r
v
e
y
,
”
i
n
Pro
c
e
e
d
i
n
g
s
o
f
t
h
e
3
r
d
I
n
t
e
r
n
a
t
i
o
n
a
l
C
o
n
f
e
re
n
c
e
o
n
I
n
t
e
g
r
a
t
e
d
I
n
t
e
l
l
i
g
e
n
t
C
o
m
p
u
t
i
n
g
C
o
m
m
u
n
i
c
a
t
i
o
n
&
S
e
c
u
r
i
t
y
(
I
C
I
I
C
2
0
2
1
)
,
2
0
2
1
,
v
o
l
.
4
,
d
o
i
:
1
0
.
2
9
9
1
/
a
h
i
s.
k
.
2
1
0
9
1
3
.
0
1
4
.
[
4
]
P
.
P
h
a
n
d
i
,
K
.
M
.
A
.
C
h
a
i
,
a
n
d
H
.
T.
N
g
,
“
F
l
e
x
i
b
l
e
d
o
m
a
i
n
a
d
a
p
t
a
t
i
o
n
f
o
r
a
u
t
o
ma
t
e
d
e
ssa
y
s
c
o
r
i
n
g
u
s
i
n
g
c
o
r
r
e
l
a
t
e
d
l
i
n
e
a
r
r
e
g
r
e
ss
i
o
n
,
”
i
n
C
o
n
f
e
r
e
n
c
e
Pr
o
c
e
e
d
i
n
g
s
-
E
MN
L
P
2
0
1
5
:
C
o
n
f
e
r
e
n
c
e
o
n
Em
p
i
r
i
c
a
l
Me
t
h
o
d
s
i
n
N
a
t
u
ra
l
L
a
n
g
u
a
g
e
Pro
c
e
ss
i
n
g
,
2
0
1
5
,
p
p
.
4
3
1
–
4
3
9
,
d
o
i
:
1
0
.
1
8
6
5
3
/
v
1
/
d
1
5
-
1
0
4
9
.
[
5
]
E.
B
.
P
a
g
e
,
“
C
o
m
p
u
t
e
r
g
r
a
d
i
n
g
o
f
s
t
u
d
e
n
t
p
r
o
se,
u
si
n
g
m
o
d
e
r
n
c
o
n
c
e
p
t
s
a
n
d
so
f
t
w
a
r
e
,
”
J
o
u
rn
a
l
o
f
Ex
p
e
ri
m
e
n
t
a
l
E
d
u
c
a
t
i
o
n
,
v
o
l
.
6
2
,
n
o
.
2
,
p
p
.
1
2
7
–
1
4
2
,
J
a
n
.
1
9
9
4
,
d
o
i
:
1
0
.
1
0
8
0
/
0
0
2
2
0
9
7
3
.
1
9
9
4
.
9
9
4
3
8
3
5
.
[
6
]
M
.
U
t
o
,
“
A
r
e
v
i
e
w
o
f
d
e
e
p
-
n
e
u
r
a
l
a
u
t
o
ma
t
e
d
e
ssa
y
s
c
o
r
i
n
g
m
o
d
e
l
s,”
Be
h
a
v
i
o
rm
e
t
ri
k
a
,
v
o
l
.
4
8
,
n
o
.
2
,
p
p
.
4
5
9
–
4
8
4
,
J
u
l
.
2
0
2
1
,
d
o
i
:
1
0
.
1
0
0
7
/
s
4
1
2
3
7
-
021
-
0
0
1
4
2
-
y.
[
7
]
W
.
S
o
n
g
,
D
.
W
a
n
g
,
R
.
F
u
,
L.
Li
u
,
T.
Li
u
,
a
n
d
G
.
H
u
,
“
D
i
sc
o
u
r
se
mo
d
e
i
d
e
n
t
i
f
i
c
a
t
i
o
n
i
n
e
ssa
y
s
,
”
i
n
AC
L
2
0
1
7
-
5
5
t
h
An
n
u
a
l
Me
e
t
i
n
g
o
f
t
h
e
Ass
o
c
i
a
t
i
o
n
f
o
r
C
o
m
p
u
t
a
t
i
o
n
a
l
L
i
n
g
u
i
st
i
c
s,
P
ro
c
e
e
d
i
n
g
s
o
f
t
h
e
C
o
n
f
e
r
e
n
c
e
(
L
o
n
g
P
a
p
e
rs
)
,
2
0
1
7
,
v
o
l
.
1
,
p
p
.
1
1
2
–
1
2
2
,
d
o
i
:
1
0
.
1
8
6
5
3
/
v
1
/
P
1
7
-
1
0
1
1
.
[
8
]
F
.
D
o
n
g
a
n
d
Y
.
Z
h
a
n
g
,
“
A
u
t
o
m
a
t
i
c
f
e
a
t
u
r
e
s
f
o
r
e
ssa
y
s
c
o
r
i
n
g
-
a
n
e
m
p
i
r
i
c
a
l
st
u
d
y
,
”
i
n
EM
N
L
P
2
0
1
6
-
C
o
n
f
e
re
n
c
e
o
n
Em
p
i
r
i
c
a
l
Me
t
h
o
d
s
i
n
N
a
t
u
ra
l
L
a
n
g
u
a
g
e
Pr
o
c
e
s
si
n
g
,
Pr
o
c
e
e
d
i
n
g
s
,
2
0
1
6
,
p
p
.
1
0
7
2
–
1
0
7
7
,
d
o
i
:
1
0
.
1
8
6
5
3
/
v
1
/
d
1
6
-
1
1
1
5
.
[
9
]
D
.
A
l
i
k
a
n
i
o
t
i
s
,
H
.
Y
a
n
n
a
k
o
u
d
a
k
i
s
,
a
n
d
M
.
R
e
i
,
“
A
u
t
o
m
a
t
i
c
t
e
x
t
s
c
o
r
i
n
g
u
s
i
n
g
n
e
u
r
a
l
n
e
t
w
o
r
k
s
,
”
i
n
Pro
c
e
e
d
i
n
g
s
o
f
t
h
e
5
4
t
h
An
n
u
a
l
Me
e
t
i
n
g
o
f
t
h
e
Asso
c
i
a
t
i
o
n
f
o
r
C
o
m
p
u
t
a
t
i
o
n
a
l
L
i
n
g
u
i
s
t
i
c
s
(
Vo
l
u
m
e
1
:
L
o
n
g
Pa
p
e
rs)
,
2
0
1
6
,
p
p
.
7
1
5
–
7
2
5
,
d
o
i
:
1
0
.
1
8
6
5
3
/
v
1
/
P
1
6
-
1
0
6
8
.
[
1
0
]
K
.
T
a
g
h
i
p
o
u
r
a
n
d
H
.
T.
N
g
,
“
A
n
e
u
r
a
l
a
p
p
r
o
a
c
h
t
o
a
u
t
o
m
a
t
e
d
e
ss
a
y
sc
o
r
i
n
g
,
”
i
n
E
MN
L
P
2
0
1
6
-
C
o
n
f
e
r
e
n
c
e
o
n
Em
p
i
r
i
c
a
l
Me
t
h
o
d
s
i
n
N
a
t
u
ra
l
L
a
n
g
u
a
g
e
Pr
o
c
e
s
si
n
g
,
Pr
o
c
e
e
d
i
n
g
s
,
2
0
1
6
,
p
p
.
1
8
8
2
–
1
8
9
1
,
d
o
i
:
1
0
.
1
8
6
5
3
/
v
1
/
d
1
6
-
1
1
9
3
.
[
1
1
]
F
.
D
o
n
g
,
Y
.
Zh
a
n
g
,
a
n
d
J.
Y
a
n
g
,
“
A
t
t
e
n
t
i
o
n
-
b
a
s
e
d
r
e
c
u
r
r
e
n
t
c
o
n
v
o
l
u
t
i
o
n
a
l
n
e
u
r
a
l
n
e
t
w
o
r
k
f
o
r
a
u
t
o
m
a
t
i
c
e
ss
a
y
sc
o
r
i
n
g
,
”
i
n
C
o
N
L
L
2
0
1
7
-
2
1
st
C
o
n
f
e
r
e
n
c
e
o
n
C
o
m
p
u
t
a
t
i
o
n
a
l
N
a
t
u
r
a
l
L
a
n
g
u
a
g
e
L
e
a
r
n
i
n
g
,
Pr
o
c
e
e
d
i
n
g
s
,
2
0
1
7
,
p
p
.
1
5
3
–
1
6
2
,
d
o
i
:
1
0
.
1
8
6
5
3
/
v
1
/
k
1
7
-
1
0
1
7
.
[1
2]
Y
.
W
a
n
g
,
Z.
W
e
i
,
Y
.
Z
h
o
u
,
a
n
d
X
.
H
u
a
n
g
,
“
A
u
t
o
ma
t
i
c
e
ss
a
y
s
c
o
r
i
n
g
i
n
c
o
r
p
o
r
a
t
i
n
g
r
a
t
i
n
g
s
c
h
e
ma
v
i
a
r
e
i
n
f
o
r
c
e
men
t
l
e
a
r
n
i
n
g
,
”
i
n
Pro
c
e
e
d
i
n
g
s
o
f
t
h
e
2
0
1
8
C
o
n
f
e
re
n
c
e
o
n
Em
p
i
r
i
c
a
l
Me
t
h
o
d
s
i
n
N
a
t
u
ra
l
L
a
n
g
u
a
g
e
Pr
o
c
e
ss
i
n
g
,
EM
N
L
P
2
0
1
8
,
2
0
1
8
,
p
p
.
7
9
1
–
7
9
7
,
d
o
i
:
1
0
.
1
8
6
5
3
/
v
1
/
d
1
8
-
1
0
9
0
.
[
1
3
]
R
.
K
u
mar
,
S
.
M
a
t
h
i
a
s,
S
.
S
a
h
a
,
a
n
d
P
.
B
h
a
t
t
a
c
h
a
r
y
y
a
,
“
M
a
n
y
h
a
n
d
s
mak
e
l
i
g
h
t
w
o
r
k
:
u
s
i
n
g
e
ssa
y
t
r
a
i
t
s
t
o
a
u
t
o
mat
i
c
a
l
l
y
sco
r
e
e
ss
a
y
s,”
i
n
N
A
AC
L
2
0
2
2
-
2
0
2
2
C
o
n
f
e
re
n
c
e
o
f
t
h
e
N
o
rt
h
Am
e
r
i
c
a
n
C
h
a
p
t
e
r
o
f
t
h
e
Ass
o
c
i
a
t
i
o
n
f
o
r
C
o
m
p
u
t
a
t
i
o
n
a
l
L
i
n
g
u
i
st
i
c
s:
H
u
m
a
n
L
a
n
g
u
a
g
e
T
e
c
h
n
o
l
o
g
i
e
s,
Pr
o
c
e
e
d
i
n
g
s
o
f
t
h
e
C
o
n
f
e
r
e
n
c
e
,
2
0
2
2
,
p
p
.
1
4
8
5
–
1
4
9
5
,
d
o
i
:
1
0
.
1
8
6
5
3
/
v
1
/
2
0
2
2
.
n
a
a
c
l
-
mai
n
.
1
0
6
.
[
1
4
]
I
.
P
e
r
si
n
g
a
n
d
V
.
N
g
,
“
M
o
d
e
l
i
n
g
a
r
g
u
m
e
n
t
s
t
r
e
n
g
t
h
i
n
st
u
d
e
n
t
e
ssa
y
s,
”
i
n
AC
L
-
I
J
C
N
L
P
2
0
1
5
-
5
3
rd
An
n
u
a
l
Me
e
t
i
n
g
o
f
t
h
e
As
so
c
i
a
t
i
o
n
f
o
r
C
o
m
p
u
t
a
t
i
o
n
a
l
L
i
n
g
u
i
st
i
c
s
a
n
d
t
h
e
7
t
h
I
n
t
e
r
n
a
t
i
o
n
a
l
J
o
i
n
t
C
o
n
f
e
re
n
c
e
o
n
N
a
t
u
ra
l
L
a
n
g
u
a
g
e
P
ro
c
e
ssi
n
g
o
f
t
h
e
Asi
a
n
Fe
d
e
r
a
t
i
o
n
o
f
N
a
t
u
r
a
l
L
a
n
g
u
a
g
e
Pr
o
c
e
ssi
n
g
,
Pro
c
e
e
d
i
n
g
s
o
f
t
h
e
C
o
n
f
e
re
n
c
e
,
2
0
1
5
,
v
o
l
.
1
,
p
p
.
5
4
3
–
5
5
2
,
d
o
i
:
1
0
.
3
1
1
5
/
v
1
/
p
1
5
-
1
0
5
3
.
[1
5]
H
.
V
.
N
g
u
y
e
n
a
n
d
D
.
J
.
L
i
t
m
a
n
,
“
A
r
g
u
m
e
n
t
m
i
n
i
n
g
f
o
r
i
m
p
r
o
v
i
n
g
t
h
e
a
u
t
o
mat
e
d
sc
o
r
i
n
g
o
f
p
e
r
su
a
s
i
v
e
e
s
say
s
,
”
3
2
n
d
A
AAI
C
o
n
f
e
re
n
c
e
o
n
Art
i
f
i
c
i
a
l
I
n
t
e
l
l
i
g
e
n
c
e
,
AAAI
2
0
1
8
,
v
o
l
.
3
2
,
n
o
.
1
,
p
p
.
5
8
9
2
–
5
8
9
9
,
A
p
r
.
2
0
1
8
,
d
o
i
:
1
0
.
1
6
0
9
/
a
a
a
i
.
v
3
2
i
1
.
1
2
0
4
6
.
[
1
6
]
A
.
L
o
u
i
s
a
n
d
A
.
N
e
n
k
o
v
a
,
“
A
u
t
o
mat
i
c
a
l
l
y
a
sses
si
n
g
ma
c
h
i
n
e
su
mm
a
r
y
c
o
n
t
e
n
t
w
i
t
h
o
u
t
a
g
o
l
d
s
t
a
n
d
a
r
d
,
”
C
o
m
p
u
t
a
t
i
o
n
a
l
L
i
n
g
u
i
st
i
c
s
,
v
o
l
.
3
9
,
n
o
.
2
,
p
p
.
2
6
7
–
3
0
0
,
J
u
n
.
2
0
1
3
,
d
o
i
:
1
0
.
1
1
6
2
/
C
O
LI
_
a
_
0
0
1
2
3
.
[
1
7
]
A
.
V
a
sw
a
n
i
e
t
a
l
.
,
“
A
t
t
e
n
t
i
o
n
i
s
a
l
l
y
o
u
n
e
e
d
,
”
A
d
v
a
n
c
e
s
i
n
N
e
u
ra
l
I
n
f
o
rm
a
t
i
o
n
Pr
o
c
e
ssi
n
g
S
y
s
t
e
m
s
,
v
o
l
.
2
0
1
7
-
D
e
c
e
m
,
n
o
.
N
i
p
s,
p
p
.
5
9
9
9
–
6
0
0
9
,
2
0
1
7
.
[
1
8
]
D
.
B
a
h
d
a
n
a
u
,
K
.
H
.
C
h
o
,
a
n
d
Y
.
B
e
n
g
i
o
,
“
N
e
u
r
a
l
m
a
c
h
i
n
e
t
r
a
n
s
l
a
t
i
o
n
b
y
j
o
i
n
t
l
y
l
e
a
r
n
i
n
g
t
o
a
l
i
g
n
a
n
d
t
r
a
n
s
l
a
t
e
,
”
3
r
d
I
n
t
e
rn
a
t
i
o
n
a
l
C
o
n
f
e
re
n
c
e
o
n
L
e
a
rn
i
n
g
Re
p
res
e
n
t
a
t
i
o
n
s
,
I
C
L
R
2
0
1
5
-
C
o
n
f
e
re
n
c
e
T
r
a
c
k
P
ro
c
e
e
d
i
n
g
s
,
2
0
1
5
.
[
1
9
]
P
.
U
.
R
o
d
r
i
g
u
e
z
,
A
.
J
a
f
a
r
i
,
a
n
d
C
.
M
.
O
r
mero
d
,
“
La
n
g
u
a
g
e
m
o
d
e
l
s
a
n
d
a
u
t
o
m
a
t
e
d
e
ssa
y
sc
o
r
i
n
g
,
”
a
r
Xi
v
,
2
0
1
9
,
[
O
n
l
i
n
e
]
.
A
v
a
i
l
a
b
l
e
:
h
t
t
p
:
/
/
a
r
x
i
v
.
o
r
g
/
a
b
s
/
1
9
0
9
.
0
9
4
8
2
.
[
2
0
]
E.
M
a
y
f
i
e
l
d
a
n
d
A
.
W
.
B
l
a
c
k
,
“
S
h
o
u
l
d
y
o
u
f
i
n
e
-
t
u
n
e
B
E
R
T
f
o
r
a
u
t
o
ma
t
e
d
e
ssa
y
sc
o
r
i
n
g
?
,
”
i
n
Pr
o
c
e
e
d
i
n
g
s o
f
t
h
e
A
n
n
u
a
l
Me
e
t
i
n
g
o
f
t
h
e
Asso
c
i
a
t
i
o
n
f
o
r
C
o
m
p
u
t
a
t
i
o
n
a
l
L
i
n
g
u
i
st
i
c
s
,
2
0
2
0
,
p
p
.
1
5
1
–
1
6
2
,
d
o
i
:
1
0
.
1
8
6
5
3
/
v
1
/
2
0
2
0
.
b
e
a
-
1
.
1
5
.
[
2
1
]
R
.
Y
a
n
g
,
J.
C
a
o
,
Z
.
W
e
n
,
Y
.
W
u
,
a
n
d
X
.
H
e
,
“
E
n
h
a
n
c
i
n
g
a
u
t
o
ma
t
e
d
e
ssay
s
c
o
r
i
n
g
p
e
r
f
o
r
m
a
n
c
e
v
i
a
f
i
n
e
-
t
u
n
i
n
g
p
r
e
-
t
r
a
i
n
e
d
l
a
n
g
u
a
g
e
mo
d
e
l
s
w
i
t
h
c
o
m
b
i
n
a
t
i
o
n
o
f
r
e
g
r
e
ssi
o
n
a
n
d
r
a
n
k
i
n
g
,
”
i
n
Fi
n
d
i
n
g
s
o
f
t
h
e
As
so
c
i
a
t
i
o
n
f
o
r
C
o
m
p
u
t
a
t
i
o
n
a
l
L
i
n
g
u
i
st
i
c
s
Fi
n
d
i
n
g
s
o
f
A
C
L
:
EM
N
L
P
2
0
2
0
,
2
0
2
0
,
p
p
.
1
5
6
0
–
1
5
6
9
,
d
o
i
:
1
0
.
1
8
6
5
3
/
v
1
/
2
0
2
0
.
f
i
n
d
i
n
gs
-
e
m
n
l
p
.
1
4
1
.
[
2
2
]
V
.
S
a
n
h
,
L.
D
e
b
u
t
,
J
.
C
h
a
u
m
o
n
d
,
a
n
d
T.
W
o
l
f
,
“
D
i
st
i
l
B
E
R
T,
a
d
i
s
t
i
l
l
e
d
v
e
r
si
o
n
o
f
B
ER
T:
sm
a
l
l
e
r
,
f
a
s
t
e
r
,
c
h
e
a
p
e
r
a
n
d
l
i
g
h
t
e
r
,
”
a
rXi
v
,
2
0
1
9
,
[
O
n
l
i
n
e
]
.
A
v
a
i
l
a
b
l
e
:
h
t
t
p
:
/
/
a
r
x
i
v
.
o
r
g
/
a
b
s
/
1
9
1
0
.
0
1
1
0
8
.
[
2
3
]
J.
G
o
u
,
B
.
Y
u
,
S
.
J
.
M
a
y
b
a
n
k
,
a
n
d
D
.
Ta
o
,
“
K
n
o
w
l
e
d
g
e
d
i
st
i
l
l
a
t
i
o
n
:
a
su
r
v
e
y
,
”
I
n
t
e
rn
a
t
i
o
n
a
l
J
o
u
r
n
a
l
o
f
C
o
m
p
u
t
e
r
V
i
si
o
n
,
v
o
l
.
1
2
9
,
n
o
.
6
,
p
p
.
1
7
8
9
–
1
8
1
9
,
J
u
n
.
2
0
2
1
,
d
o
i
:
1
0
.
1
0
0
7
/
s
1
1
2
6
3
-
0
2
1
-
0
1
4
5
3
-
z.
[
2
4
]
J.
D
e
v
l
i
n
,
M
.
-
W
.
C
h
a
n
g
,
K
.
Le
e
,
K
.
T.
G
o
o
g
l
e
,
a
n
d
A
.
I
.
La
n
g
u
a
g
e
,
“
B
E
R
T:
p
r
e
-
t
r
a
i
n
i
n
g
o
f
d
e
e
p
b
i
d
i
r
e
c
t
i
o
n
a
l
t
r
a
n
sf
o
r
mers
f
o
r
l
a
n
g
u
a
g
e
u
n
d
e
r
s
t
a
n
d
i
n
g
,
”
N
a
a
c
l
-
H
l
t
2
0
1
9
,
n
o
.
M
l
m,
p
p
.
4
1
7
1
–
4
1
8
6
,
2
0
1
8
,
[
O
n
l
i
n
e
]
.
A
v
a
i
l
a
b
l
e
:
h
t
t
p
s:
/
/
a
c
l
a
n
t
h
o
l
o
g
y
.
o
r
g
/
N
1
9
-
1
4
2
3
.
p
d
f
.
[
2
5
]
Y
.
Ta
y
,
M
.
P
h
a
n
,
L.
A
.
T
u
a
n
,
a
n
d
S
.
C
.
H
u
i
,
“
S
k
i
p
F
l
o
w
:
i
n
c
o
r
p
o
r
a
t
i
n
g
n
e
u
r
a
l
c
o
h
e
r
e
n
c
e
f
e
a
t
u
r
e
s
f
o
r
e
n
d
-
to
-
e
n
d
a
u
t
o
mat
i
c
t
e
x
t
sco
r
i
n
g
,
”
Pr
o
c
e
e
d
i
n
g
s
o
f
t
h
e
AA
A
I
C
o
n
f
e
r
e
n
c
e
o
n
Art
i
f
i
c
i
a
l
I
n
t
e
l
l
i
g
e
n
c
e
,
v
o
l
.
3
2
,
n
o
.
1
,
p
p
.
5
9
4
8
–
5
9
5
5
,
A
p
r
.
2
0
1
8
,
d
o
i
:
1
0
.
1
6
0
9
/
a
a
a
i
.
v
3
2
i
1
.
1
2
0
4
5
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
d
o
n
esian
J
E
lec
E
n
g
&
C
o
m
p
Sci
I
SS
N:
2
5
0
2
-
4
7
52
E
xp
lo
r
in
g
th
e
p
o
ten
tia
l o
f D
is
tilBE
R
T a
r
ch
itectu
r
e
fo
r
a
u
to
ma
tic
ess
a
y
s
co
r
in
g
ta
s
k
(
S
o
u
mi
a
I
kiss
)
1241
B
I
O
G
RAP
H
I
E
S O
F
AUTH
O
RS
S
o
u
m
ia
Ik
iss
is
a
P
h
.
D
.
s
tu
d
e
n
t
a
t
th
e
Na
ti
o
n
a
l
I
n
stit
u
te
o
f
P
o
sts
a
n
d
Tele
c
o
m
m
u
n
ica
ti
o
n
s,
Ra
b
a
t,
M
o
r
o
c
c
o
sin
c
e
2
0
2
1
.
G
ra
d
u
a
ted
fr
o
m
th
e
Un
i
v
e
rsity
o
f
S
u
lt
a
n
e
M
o
u
la
y
S
li
m
a
n
e
,
Be
n
i
M
e
ll
a
l,
M
o
ro
c
c
o
in
2
0
2
0
wit
h
a
m
a
ste
r
’
s
d
e
g
re
e
in
‘Bu
sin
e
ss
In
telli
g
e
n
c
e
’.
He
r
re
se
a
rc
h
a
re
a
s
a
re
a
rti
ficia
l
in
tell
ig
e
n
c
e
,
d
e
e
p
lea
rn
in
g
,
a
n
d
n
a
tu
ra
l
lan
g
u
a
g
e
p
ro
c
e
ss
in
g
a
n
d
th
e
ir
a
p
p
li
c
a
ti
o
n
to
re
a
l
-
wo
rl
d
p
ro
b
lem
s.
S
h
e
is
c
u
rre
n
tl
y
wo
r
k
in
g
o
n
th
e
a
p
p
li
c
a
ti
o
n
o
f
a
rti
f
icia
l
in
telli
g
e
n
c
e
in
t
h
e
e
d
u
c
a
ti
o
n
i
n
d
u
stry
,
m
a
in
ly
o
n
t
h
e
e
v
a
l
u
a
ti
o
n
p
a
rt
.
S
h
e
c
a
n
b
e
c
o
n
tac
ted
a
t
e
m
a
il
:
so
u
m
ia.i
k
iss9
7
@g
m
a
il
.
c
o
m
.
Na
jim
a
Da
o
u
d
i
is
a
fu
ll
p
ro
fe
ss
o
r
a
t
th
e
S
c
h
o
o
l
o
f
In
f
o
rm
a
ti
o
n
S
c
ien
c
e
s,
Ra
b
a
t
,
M
o
ro
c
c
o
.
S
h
e
h
a
s
a
n
e
n
g
i
n
e
e
rin
g
d
e
g
re
e
fro
m
t
h
e
Na
ti
o
n
a
l
I
n
stit
u
t
e
o
f
S
tatisti
c
s
a
n
d
A
p
p
li
e
d
Eco
n
o
m
ics
(INSE
A),
Ra
b
a
t,
M
o
ro
c
c
o
,
a
n
d
a
P
h
.
D.
i
n
Co
m
p
u
t
e
r
S
c
ien
c
e
fro
m
ENS
IAS,
Ra
b
a
t,
M
o
r
o
c
c
o
.
He
r
re
se
a
r
c
h
in
tere
st
s
in
c
lu
d
e
a
rti
ficia
l
in
telli
g
e
n
c
e
,
e
-
lea
rn
in
g
,
re
c
o
m
m
e
n
d
a
ti
o
n
sy
ste
m
s
,
a
n
d
n
a
tu
ra
l
lan
g
u
a
g
e
p
ro
c
e
ss
in
g
.
S
h
e
c
a
n
b
e
c
o
n
tac
ted
a
t
e
m
a
il
:
n
d
a
o
u
d
i@e
si.ac
.
m
a
.
Ma
n
a
r
A
b
o
u
r
e
z
q
is
a
p
ro
fe
ss
o
r
a
t
th
e
S
c
h
o
o
l
o
f
In
fo
rm
a
ti
o
n
S
c
ien
c
e
s,
M
o
ro
c
c
o
.
S
h
e
is
a
n
E
n
g
i
n
e
e
r
o
f
th
e
Na
ti
o
n
a
l
S
c
h
o
o
l
f
o
r
C
o
m
p
u
ter
S
c
ien
c
e
a
n
d
S
y
ste
m
s
An
a
ly
sis
a
n
d
h
o
l
d
s
a
P
h
.
D
.
i
n
Co
m
p
u
ter
S
c
ien
c
e
fro
m
M
o
h
a
m
m
e
d
V
U
n
iv
e
rs
it
y
in
Ra
b
a
t,
M
o
ro
c
c
o
.
He
r
re
se
a
rc
h
in
tere
sts
in
c
lu
d
e
a
rti
ficia
l
in
tel
l
ig
e
n
c
e
,
Ara
b
ic
n
a
t
u
ra
l
lan
g
u
a
g
e
p
r
o
c
e
ss
in
g
,
a
n
d
c
lo
u
d
c
o
m
p
u
ti
n
g
.
S
h
e
c
a
n
b
e
c
o
n
tac
ted
a
t
e
m
a
il
:
m
a
n
a
r.
a
b
o
u
re
z
q
@g
m
a
il
.
c
o
m
.
Mo
sta
f
a
Be
ll
a
fk
ih
is
a
p
ro
fe
ss
o
r
a
t
Th
e
Na
ti
o
n
a
l
In
stit
u
te
o
f
P
o
sts
a
n
d
Tele
c
o
m
m
u
n
ica
ti
o
n
s
(INP
T)
i
n
R
a
b
a
t,
M
o
r
o
c
c
o
.
re
c
e
iv
e
d
a
P
h
.
D
.
t
h
e
sis
in
C
o
m
p
u
ter
S
c
ien
c
e
fro
m
th
e
Un
i
v
e
rsity
o
f
P
a
ris
6
,
F
ra
n
c
e
,
in
Ju
n
e
1
9
9
4
a
n
d
a
d
o
c
to
ra
te
S
c
ien
c
e
in
Co
m
p
u
te
r
S
c
ien
c
e
(o
p
ti
o
n
n
e
two
r
k
s)
fro
m
th
e
Un
i
v
e
rsity
o
f
M
o
h
a
m
m
e
d
V
in
Ra
b
a
t,
M
o
r
o
c
c
o
.
His
re
se
a
rc
h
in
tere
sts
in
c
lu
d
e
n
e
two
rk
m
a
n
a
g
e
m
e
n
t,
k
n
o
wle
d
g
e
m
a
n
a
g
e
m
e
n
t,
AI,
d
a
ta
m
in
i
n
g
,
a
n
d
d
a
ta
b
a
se
s.
He
c
a
n
b
e
c
o
n
tac
te
d
a
t:
m
b
e
ll
a
fk
ih
@
y
a
h
o
o
.
c
o
m
.
Evaluation Warning : The document was created with Spire.PDF for Python.