I
nte
rna
t
io
na
l J
o
urna
l o
f
E
lect
rica
l a
nd
Co
m
pu
t
er
E
ng
ineering
(
I
J
E
CE
)
Vo
l.
15
,
No
.
2
,
A
p
r
il
20
25
,
p
p
.
2
0
7
0
~
2
0
8
0
I
SS
N:
2088
-
8
7
0
8
,
DOI
: 1
0
.
1
1
5
9
1
/ijece.
v
15
i
2
.
pp
2
0
7
0
-
2
0
8
0
2070
J
o
ur
na
l ho
m
ep
a
g
e
:
h
ttp
:
//ij
ec
e.
ia
esco
r
e.
co
m
Ens
emble o
f
winn
ing
t
icke
ts
:
p
runi
ng
bidirec
tiona
l e
nco
der
from the
t
ra
nsfo
r
mers
a
tt
en
tion he
a
ds for enha
nce
d
mo
del
eff
icie
ncy
Ny
a
la
la
ni Sm
a
rt
s
,
Ra
j
a
la
k
s
hm
i Selv
a
ra
j
,
Venu
M
a
dh
a
v
K
utha
di
D
e
p
a
r
t
me
n
t
o
f
C
o
mp
u
t
i
n
g
a
n
d
I
n
f
o
r
m
a
t
i
c
s,
S
c
h
o
o
l
o
f
P
u
r
e
a
n
d
A
p
p
l
i
e
d
o
f
S
c
i
e
n
c
e
s,
B
o
t
sw
a
n
a
I
n
t
e
r
n
a
t
i
o
n
a
l
U
n
i
v
e
r
si
t
y
o
f
S
c
i
e
n
c
e
a
n
d
Te
c
h
n
o
l
o
g
y
,
P
a
l
a
p
y
e
,
B
o
t
sw
a
n
a
Art
icle
I
nfo
AB
S
T
RAC
T
A
r
ticle
his
to
r
y:
R
ec
eiv
ed
J
u
l 8
,
2
0
2
4
R
ev
is
ed
Oct
1
5
,
2
0
2
4
Acc
ep
ted
Oct
2
3
,
2
0
2
4
Th
e
a
d
v
a
n
c
e
d
m
o
d
e
ls
o
f
d
e
e
p
n
e
u
ra
l
n
e
two
rk
s
li
k
e
b
i
d
irec
ti
o
n
a
l
e
n
c
o
d
e
r
fro
m
th
e
tra
n
sfo
rm
e
rs
(BERT
)
a
n
d
o
th
e
rs,
p
o
se
s
c
h
a
ll
e
n
g
e
s
in
term
s
o
f
c
o
m
p
u
tati
o
n
a
l
re
so
u
rc
e
s
a
n
d
m
o
d
e
l
siz
e
.
In
o
rd
e
r
to
tac
k
le
t
h
e
se
issu
e
s,
tec
h
n
iq
u
e
s
o
f
m
o
d
e
l
p
r
u
n
in
g
h
a
v
e
su
rfa
c
e
d
a
s
th
e
m
o
st
u
se
fu
l
m
e
th
o
d
s
in
a
d
d
re
ss
in
g
t
h
e
issu
e
s
o
f
m
o
d
e
l
c
o
m
p
lex
it
y
.
Th
is
re
se
a
rc
h
p
a
p
e
r
e
x
p
l
o
re
s
t
h
e
c
o
n
c
e
p
t
o
f
p
ru
n
in
g
BERT
a
tt
e
n
t
io
n
h
e
a
d
s
a
c
ro
ss
th
e
e
n
se
m
b
le
o
f
win
n
i
n
g
ti
c
k
e
ts
in
o
r
d
e
r
to
e
n
h
a
n
c
e
th
e
e
fficie
n
c
y
o
f
th
e
m
o
d
e
l
with
o
u
t
s
a
c
rifi
c
in
g
p
e
rfo
rm
a
n
c
e
.
E
x
p
e
rime
n
tal
e
v
a
lu
a
ti
o
n
s
s
h
o
w
h
o
w
e
ffe
c
ti
v
e
t
h
e
a
p
p
ro
a
c
h
is,
in
a
c
h
iev
i
n
g
sig
n
ifi
c
a
n
t
m
o
d
e
l
c
o
m
p
re
ss
io
n
w
h
il
e
sti
ll
m
a
in
tain
in
g
c
o
m
p
e
ti
ti
v
e
p
e
rf
o
rm
a
n
c
e
a
c
ro
ss
d
iffere
n
t
n
a
t
u
ra
l
lan
g
u
a
g
e
p
ro
c
e
s
sin
g
tas
k
s.
Th
e
k
e
y
fin
d
i
n
g
s
o
f
t
h
is
st
u
d
y
i
n
c
lu
d
e
m
o
d
e
l
siz
e
th
a
t
h
a
s
b
e
e
n
r
e
d
u
c
e
d
b
y
3
6
%
,
wi
th
o
u
r
e
n
se
m
b
le
m
o
d
e
l
re
a
c
h
in
g
g
re
a
ter p
e
rfo
rm
a
n
c
e
a
s
c
o
m
p
a
re
d
to
th
e
b
a
se
li
n
e
BERT
m
o
d
e
l
o
n
b
o
t
h
S
tan
f
o
rd
S
e
n
ti
m
e
n
t
Tree
b
a
n
k
v2
(S
S
T
-
2
)
a
n
d
C
o
rp
u
s
o
f
Li
n
g
u
ist
ic
Ac
c
e
p
t
a
b
il
it
y
(Co
LA)
d
a
tas
e
ts.
Th
e
re
su
lt
s
fu
rt
h
e
r
sh
o
w
a
F
1
-
sc
o
re
o
f
9
4
%
a
n
d
9
6
%
,
re
sp
e
c
ti
v
e
ly
,
a
n
d
a
c
c
u
ra
c
y
sc
o
re
s
o
f
9
5
%
a
n
d
9
6
%
o
n
t
h
e
two
d
a
t
a
se
ts.
Th
e
fin
d
in
g
s
o
f
t
h
is
re
se
a
rc
h
p
a
p
e
r
c
o
n
tri
b
u
te
to
th
e
o
n
g
o
i
n
g
e
ffo
rts
i
n
e
n
h
a
n
c
in
g
t
h
e
e
fficie
n
c
y
o
f
larg
e
-
sc
a
le
lan
g
u
a
g
e
m
o
d
e
ls.
K
ey
w
o
r
d
s
:
Atten
tio
n
h
ea
d
s
B
id
ir
ec
tio
n
al
en
co
d
er
f
r
o
m
th
e
tr
an
s
f
o
r
m
er
s
L
o
tter
y
tick
ets h
y
p
o
th
esis
Natu
r
al
lan
g
u
ag
e
p
r
o
ce
s
s
in
g
Pru
n
in
g
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
:
Ny
alala
n
i Sm
ar
ts
Dep
ar
tm
en
t o
f
C
o
m
p
u
tin
g
an
d
I
n
f
o
r
m
atics,
Facu
lty
o
f
Scien
ce
s
,
B
o
ts
wan
a
I
n
ter
n
atio
n
al
U
n
iv
er
s
ity
o
f
Scien
ce
an
d
T
ec
h
n
o
lo
g
y
Plo
t 1
0
0
7
1
,
Palap
y
e,
B
o
ts
wan
a
E
m
ail:
Ny
alala
n
i.sm
ar
ts
@
s
tu
d
en
tm
ail.
b
iu
s
t.a
c.
b
w
1.
I
NT
RO
D
UCT
I
O
N
Dee
p
n
eu
r
al
n
etwo
r
k
s
lik
e
b
id
ir
ec
tio
n
al
en
co
d
er
f
r
o
m
t
h
e
tr
an
s
f
o
r
m
er
s
(
B
E
R
T
)
[
1
]
h
a
v
e
to
tally
ch
an
g
ed
n
atu
r
al
lan
g
u
a
g
e
p
r
o
ce
s
s
in
g
(
NL
P)
task
s
.
Pre
v
io
u
s
r
esear
ch
in
th
e
liter
atu
r
e
h
av
e
d
em
o
n
s
tr
ated
th
at
B
E
R
T
i
s
a
p
o
wer
f
u
l
m
o
d
el
f
o
r
NL
P
task
s
.
Ho
wev
er
,
th
e
alar
m
in
g
r
ate
at
wh
ich
th
e
m
o
d
els
ar
e
g
r
o
win
g
i
n
ter
m
s
o
f
m
o
d
el
co
m
p
lex
ity
with
b
illi
o
n
s
o
f
p
ar
am
eter
s
[
2
]
,
[
3
]
h
as
lead
to
in
cr
ea
s
ed
d
em
an
d
f
o
r
co
m
p
u
tatio
n
al
r
eso
u
r
ce
s
a
n
d
th
e
lar
g
e
m
o
d
el
s
izes
n
ee
d
ed
f
o
r
tr
ain
i
n
g
a
n
d
d
ep
lo
y
m
en
t
[
4
]
,
[
5
]
.
T
h
e
co
m
p
licatio
n
s
o
f
th
e
m
o
d
els
m
ak
e
th
e
p
r
ac
ticality
an
d
s
ca
lab
ilit
y
o
f
th
ese
m
o
d
els
in
r
ea
l
-
wo
r
ld
ap
p
licatio
n
s
to
b
e
d
if
f
icu
lt.
T
h
u
s
,
t
h
er
e
is
a
p
r
ess
in
g
n
ee
d
f
o
r
m
o
r
e
e
f
f
icie
n
t a
n
d
ef
f
ec
tiv
e
ap
p
r
o
ac
h
es to
NL
P
[
6
]
.
Pre
v
io
u
s
s
tu
d
ies
lik
e
th
o
s
e
b
y
[
7
]
–
[
9
]
h
av
e
s
h
o
wn
B
E
R
T
c
ap
ab
ilit
ies
in
r
eg
ar
d
to
NL
P
t
ask
s
,
with
th
e
an
o
th
e
r
s
tu
d
y
b
y
Ma
r
u
f
et
a
l.
[
1
0
]
d
em
o
n
s
tr
atin
g
its
s
ig
n
if
ican
t
im
p
r
o
v
em
en
ts
in
m
ac
h
in
e
tr
a
n
s
latio
n
m
o
d
els
f
o
r
d
o
c
u
m
en
t
-
lev
el
tr
a
n
s
latio
n
,
th
e
s
tu
d
y
[
1
1
]
illu
s
tr
atin
g
its
ca
p
ab
ilit
ies
in
co
n
tex
tu
al
u
n
d
e
r
s
tan
d
in
g
an
d
th
e
s
tu
d
y
b
y
Z
a
k
r
ao
u
i
et
a
l.
[
1
2
]
in
n
eu
r
al
m
ac
h
in
e
tr
a
n
s
latio
n
(
NM
T
)
.
I
n
f
o
r
m
atio
n
f
r
o
m
th
e
liter
atu
r
e
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J E
lec
&
C
o
m
p
E
n
g
I
SS
N:
2088
-
8
7
0
8
E
n
s
emb
le
o
f wi
n
n
in
g
ticke
ts
:
p
r
u
n
in
g
b
id
ir
ec
tio
n
a
l e
n
co
d
er fr
o
m
th
e
tr
a
n
s
fo
r
mers
…
(
N
ya
la
la
n
i S
ma
r
ts
)
2071
s
h
o
ws
th
at
B
E
R
T
h
as
al
s
o
ex
ce
lled
in
s
en
tim
en
t
an
aly
s
is
task
s
an
d
h
as
p
r
o
v
ed
its
im
p
o
r
tan
ce
in
v
ar
io
u
s
p
r
ac
tical
ap
p
licatio
n
s
.
Stu
d
ies
d
em
o
n
s
tr
ate
th
at
it
h
as
b
ee
n
a
p
p
lied
to
an
aly
ze
c
u
s
to
m
er
r
e
v
iews
an
d
f
ee
d
b
ac
k
o
n
e
-
co
m
m
er
ce
p
latf
o
r
m
s
[
1
3
]
.
I
t
h
as
also
b
ee
n
u
s
ed
to
a
n
al
y
ze
u
s
er
g
e
n
er
ated
c
o
n
ten
t
r
el
ated
to
p
r
o
d
u
ct
a
n
d
s
er
v
ice
ev
alu
atio
n
s
[
1
4
]
an
d
o
th
er
NL
P
task
s
lik
e
q
u
esti
o
n
an
s
wer
in
g
a
n
d
tex
t
class
if
icatio
n
.
R
eg
ar
d
less
o
f
th
ese
ac
h
iev
em
en
ts
,
th
ese
s
tu
d
ies
wh
ich
h
ig
h
lig
h
ted
ab
o
v
e
also
n
o
te
th
e
ch
allen
g
es
p
o
s
ed
b
y
B
E
R
T
's
lar
g
e
s
ize
an
d
h
ig
h
co
m
p
u
tatio
n
al
r
eq
u
ir
em
e
n
ts
.
C
o
n
s
eq
u
en
tly
,
th
ese
in
d
icate
th
at
m
eth
o
d
s
t
h
at
ca
n
r
ed
u
ce
th
e
co
m
p
lex
ity
o
f
m
o
d
els
lik
e
B
E
R
T
with
o
u
t
s
ac
r
if
icin
g
p
er
f
o
r
m
a
n
ce
ar
e
n
ee
d
ed
.
C
u
r
r
en
tly
,
th
e
p
r
u
n
in
g
tech
n
iq
u
es
th
at
ar
e
av
ailab
le
u
s
u
ally
lead
to
a
tr
ad
e
-
o
f
f
b
etwe
en
m
o
d
el
s
ize
an
d
ac
cu
r
ac
y
,
wh
ich
m
ak
es
th
em
less
ef
f
ec
tiv
e.
T
h
is
s
tu
d
y
aim
s
to
ad
d
r
ess
th
ese
ch
allen
g
es b
y
d
ev
elo
p
in
g
a
m
o
r
e
e
f
f
icien
t
p
r
u
n
in
g
m
eth
o
d
.
W
e
i
n
t
r
o
d
u
c
e
a
n
e
w
p
r
u
n
i
n
g
t
e
c
h
n
i
q
u
e
t
h
a
t
w
e
c
al
l
e
n
s
e
m
b
l
e
o
f
w
i
n
n
i
n
g
t
i
c
k
e
t
s
w
h
i
c
h
u
t
i
l
i
z
es
t
h
e
l
o
t
t
e
r
y
t
i
c
k
et
h
y
p
o
t
h
e
s
is
t
o
e
n
h
a
n
c
e
B
E
R
T
’
s
p
e
r
f
o
r
m
a
n
c
e
m
e
t
r
i
c
s
w
h
i
l
e
d
r
as
ti
c
a
ll
y
r
e
d
u
c
i
n
g
i
t
s
c
o
m
p
l
e
x
it
y
.
I
n
t
h
i
s
p
r
o
p
o
s
e
d
a
p
p
r
o
a
c
h
,
w
e
i
m
p
l
e
m
e
n
t
a
n
i
t
e
r
a
t
i
v
e
a
tt
e
n
t
i
o
n
h
e
a
d
p
r
u
n
i
n
g
f
o
r
B
E
R
T
r
e
ta
i
n
i
n
g
t
h
e
m
i
n
i
m
a
l
f
e
a
t
u
r
e
s
o
f
t
h
e
m
o
d
e
l
.
I
n
o
u
r
ap
p
r
o
a
c
h
t
h
e
c
o
m
b
i
n
e
d
p
r
u
n
e
d
s
u
b
n
e
t
w
o
r
k
s
o
r
wi
n
n
i
n
g
t
i
c
k
et
s
a
r
e
c
o
m
b
i
n
e
d
i
n
t
o
a
r
o
b
u
s
t
e
n
s
e
m
b
l
e
m
o
d
el
w
h
i
c
h
g
i
v
e
s
u
p
t
o
3
6
%
r
e
d
u
c
t
i
o
n
i
n
m
o
d
e
l
s
i
z
e
w
i
t
h
o
u
t
d
et
e
r
i
o
r
a
t
i
n
g
t
h
e
p
e
r
f
o
r
m
a
n
c
e
.
Pru
n
in
g
tech
n
i
q
u
es
h
av
e
b
ec
o
m
e
im
p
o
r
tan
t
in
o
v
er
co
m
i
n
g
th
e
ch
allen
g
es
o
f
in
cr
ea
s
in
g
m
o
d
el
co
m
p
lex
ity
i
n
d
ee
p
n
eu
r
al
n
etwo
r
k
s
.
B
y
s
elec
tiv
ely
r
em
o
v
in
g
r
ed
u
n
d
a
n
t
o
r
less
im
p
ac
tf
u
l
co
m
p
o
n
e
n
ts
,
p
r
u
n
in
g
s
tr
ea
m
lin
es
n
etwo
r
k
s
tr
u
ctu
r
es,
im
p
r
o
v
in
g
i
n
f
er
en
c
e
s
p
ee
d
,
r
ed
u
cin
g
m
em
o
r
y
u
s
ag
e,
an
d
en
h
a
n
cin
g
co
m
p
u
tatio
n
al
e
f
f
icien
cy
[
1
5
]
–
[
1
7
]
.
On
e
p
r
o
m
is
in
g
a
p
p
r
o
ac
h
is
th
e
co
n
ce
p
t
o
f
“
win
n
i
n
g
tick
ets,
”
wh
ich
id
en
tifie
s
ef
f
icien
t
s
u
b
n
etwo
r
k
s
with
in
s
p
ar
s
e
d
ee
p
n
eu
r
al
n
etwo
r
k
s
th
at
m
atch
o
r
ex
ce
e
d
th
e
p
er
f
o
r
m
an
ce
o
f
th
e
o
r
ig
in
al
m
o
d
el
[
1
8
]
,
[
1
9
]
.
I
n
th
is
s
tu
d
y
we
m
ak
e
th
e
f
o
llo
win
g
k
ey
co
n
tr
ib
u
tio
n
s
:
i)
Dev
elo
p
m
en
t
o
f
an
en
s
em
b
le
m
eth
o
d
f
o
r
p
r
u
n
in
g
d
ee
p
lear
n
i
n
g
m
o
d
els
;
ii)
I
ter
ativ
e
p
r
u
n
in
g
o
f
B
E
R
T
atten
tio
n
h
ea
d
s
to
r
etain
o
n
ly
th
e
m
o
s
t
cr
itical
co
m
p
o
n
en
ts
;
an
d
iii)
I
n
teg
r
atio
n
o
f
p
r
u
n
ed
s
u
b
n
etwo
r
k
s
in
t
o
a
r
o
b
u
s
t
en
s
em
b
le,
r
esu
ltin
g
in
a
3
6
%
r
ed
u
ctio
n
in
m
o
d
el
s
ize
with
o
u
t
lo
s
s
o
f
p
er
f
o
r
m
an
ce
.
T
h
e
r
em
ain
d
e
r
o
f
th
is
p
a
p
er
is
s
tr
u
ctu
r
ed
as:
s
ec
tio
n
2
d
escr
ib
es
th
e
p
r
o
p
o
s
ed
en
s
em
b
le
o
f
win
n
in
g
tick
ets
m
eth
o
d
in
d
etail.
Sectio
n
3
p
r
esen
ts
e
x
p
er
im
en
tal
r
esu
lts
d
em
o
n
s
tr
atin
g
th
e
ef
f
icac
y
o
f
o
u
r
ap
p
r
o
ac
h
an
d
d
is
cu
s
s
es
th
e
im
p
licatio
n
s
o
f
o
u
r
f
i
n
d
in
g
s
an
d
p
o
ten
tial f
u
tu
r
e
wo
r
k
.
Fin
ally
,
s
ec
tio
n
4
co
n
clu
d
es th
e
p
ap
e
r
.
2.
M
E
T
H
O
D
Ou
r
ap
p
r
o
ac
h
in
clu
d
es
th
e
u
t
ilizatio
n
o
f
a
well
-
estab
lis
h
ed
Stan
f
o
r
d
Sen
tim
en
t
T
r
ee
b
a
n
k
(
SS
T
-
2
)
[
2
0
]
a
n
d
C
o
r
p
u
s
o
f
L
in
g
u
is
tic
Acc
ep
tab
ilit
y
(
C
o
L
A)
[
2
1
]
d
atasets
with
tr
ain
in
g
an
d
d
e
v
elo
p
m
en
t
s
ets,
th
e
ca
r
ef
u
l
in
itializatio
n
o
f
B
E
R
T
-
b
ase
lo
wer
ca
s
e
m
o
d
el
f
r
o
m
t
h
e
t
r
an
s
f
o
r
m
er
s
lib
r
ar
y
[
2
2
]
.
T
h
e
ap
p
r
o
ac
h
also
u
s
es
th
e
in
n
o
v
ativ
e
u
s
e
o
f
en
s
em
b
le
tech
n
iq
u
es,
m
eticu
lo
u
s
ly
ch
o
s
en
tr
ain
in
g
co
n
f
i
g
u
r
at
io
n
s
an
d
ev
alu
ated
u
s
in
g
th
e
m
etr
ics
o
u
tlin
ed
in
T
ab
le
1
.
E
ac
h
s
tep
is
d
esig
n
e
d
to
o
p
tim
ize
th
e
lea
r
n
in
g
p
r
o
ce
s
s
an
d
e
n
h
an
ce
m
o
d
el
ef
f
icien
c
y
,
p
r
o
v
id
i
n
g
a
co
m
p
r
eh
e
n
s
iv
e
f
r
am
ew
o
r
k
.
T
h
e
p
r
o
p
o
s
ed
m
eth
o
d
g
en
er
al
ar
ch
itectu
r
e
is
s
h
o
wn
in
Fig
u
r
e
1
wh
ich
d
em
o
n
s
tr
ates
h
o
w
th
is
s
tu
d
y
o
p
tim
izes
B
E
R
T
m
o
d
el
u
s
in
g
p
r
u
n
e
d
s
u
b
n
etwo
r
k
s
,
ca
lled
wi
n
n
in
g
tick
ets.
Star
tin
g
with
a
f
u
lly
tr
ain
ed
B
E
R
T
m
o
d
el,
iter
ativ
e
p
r
u
n
in
g
r
e
m
o
v
es
u
n
n
ec
ess
ar
y
p
a
r
am
ete
r
s
,
id
en
tify
in
g
ef
f
icien
t
s
u
b
n
etwo
r
k
s
th
at
r
etain
p
er
f
o
r
m
an
ce
.
T
h
ese
s
m
aller
B
E
R
T
m
o
d
els
ca
lled
win
n
in
g
tick
ets
ar
e
co
m
b
in
e
d
in
to
an
en
s
em
b
le
[
2
3
]
,
en
h
an
cin
g
ef
f
icien
c
y
an
d
p
er
f
o
r
m
an
ce
o
v
er
th
e
o
r
ig
in
al
m
o
d
el.
T
h
e
im
p
lem
en
tatio
n
o
f
th
e
en
s
em
b
le
p
r
u
n
in
g
tech
n
iq
u
e
f
o
r
B
E
R
T
atten
tio
n
h
ea
d
s
in
v
o
lv
es selectiv
ely
id
en
tify
in
g
an
d
r
etai
n
in
g
in
f
o
r
m
a
tiv
e
atten
tio
n
h
ea
d
s
b
ased
o
n
th
e
p
r
in
ci
p
les o
f
th
e
l
o
tter
y
tick
et
h
y
p
o
th
esis
.
T
a
b
l
e
1
.
C
o
L
A
,
S
S
T
-
2
t
a
s
k
s
,
d
a
t
a
s
et
s
i
z
es
,
a
n
d
m
e
t
r
i
cs
f
o
r
t
h
is
s
t
u
d
y
Ta
sk
D
a
t
a
s
e
t
Tr
a
i
n
D
e
v
M
e
t
r
i
c
s
Li
n
g
u
i
s
t
i
c
a
c
c
e
p
t
a
b
i
l
i
t
y
C
o
r
p
u
s
o
f
Li
n
g
u
i
st
i
c
A
c
c
e
p
t
a
b
i
l
i
t
y
Ju
d
g
e
me
n
t
s
1
0
k
1k
A
c
c
u
r
a
c
y
F1
-
sc
o
r
e
S
e
n
t
i
me
n
t
a
n
a
l
y
si
s
Th
e
S
t
a
n
f
o
r
d
S
e
n
t
i
m
e
n
t
Tr
e
e
b
a
n
k
6
7
k
8
7
2
A
c
c
u
r
a
c
y
F1
-
sc
o
r
e
Fig
u
r
e
1
.
S
tep
s
o
f
o
p
tim
izin
g
B
E
R
T
m
o
d
els th
r
o
u
g
h
an
en
s
e
m
b
le
o
f
p
r
u
n
ed
s
u
b
n
etwo
r
k
s
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
0
8
8
-
8
7
0
8
I
n
t J E
lec
&
C
o
m
p
E
n
g
,
Vo
l.
15
,
No
.
2
,
Ap
r
il
20
25
:
2
0
7
0
-
2
0
8
0
2072
2
.
1
.
E
ns
em
ble
prun
ing
m
et
ho
ds
R
ec
en
t
s
tu
d
ies
h
av
e
e
x
p
lo
r
e
d
v
ar
i
o
u
s
m
eth
o
d
s
f
o
r
lev
er
ag
in
g
m
u
ltip
le
p
r
u
n
e
d
s
u
b
n
e
two
r
k
s
to
en
h
an
ce
m
o
d
el
p
er
f
o
r
m
an
ce
an
d
r
o
b
u
s
tn
ess
,
o
f
ten
th
r
o
u
g
h
th
e
len
s
o
f
th
e
lo
tter
y
tick
et
h
y
p
o
th
esis
an
d
en
s
em
b
le
tech
n
iq
u
es
[
2
4
]
.
A
n
o
th
er
s
tu
d
y
b
y
C
h
en
et
a
l.
[
2
5
]
in
tr
o
d
u
ce
d
s
tr
u
ctu
r
e
d
win
n
in
g
tick
ets
at
in
itial
B
E
R
T
tr
ain
in
g
an
d
Ko
b
ay
as
h
i
et
a
l.
[
2
6
]
d
is
co
v
er
ed
f
o
u
r
win
n
in
g
tick
ets
d
u
r
in
g
th
e
f
in
e
-
tu
n
in
g
p
r
o
ce
s
s
.
Fu
r
th
er
m
o
r
e
,
Go
n
g
et
a
l.
[
2
7
]
r
eg
u
lar
ized
th
e
p
r
o
ce
s
s
o
f
f
in
e
-
tu
n
in
g
u
s
in
g
a
tech
n
i
q
u
e
ca
ll
ed
L
1
d
is
tan
ce
an
d
ex
p
lo
r
ed
th
e
s
tr
u
ctu
r
e
o
f
th
e
d
is
co
v
er
ed
s
u
b
n
etwo
r
k
s
.
An
o
th
er
ex
am
p
le
o
f
th
e
im
p
l
em
en
tatio
n
o
f
th
is
tech
n
iq
u
e
ca
n
b
e
s
ee
n
in
th
e
s
tu
d
y
[
2
8
]
,
in
th
eir
r
esear
ch
th
ey
ap
p
lied
th
e
l
o
tter
y
tick
et
h
y
p
o
th
esis
to
p
r
u
n
e
atten
tio
n
h
ea
d
s
in
m
ac
h
in
e
tr
an
s
latio
n
.
I
n
th
is
s
tu
d
y
we
u
s
e
th
e
lo
tter
y
tick
et
h
y
p
o
t
h
esis
to
iter
ativ
ely
p
r
u
n
e
th
e
atten
tio
n
h
ea
d
s
,
id
e
n
tify
i
n
g
win
n
in
g
tick
ets
b
y
r
etain
in
g
o
n
ly
th
e
m
o
s
t
cr
itical
atten
tio
n
h
ea
d
s
f
r
o
m
a
B
E
R
T
m
o
d
el.
T
h
is
m
eth
o
d
r
e
f
in
es
th
e
s
u
b
n
etwo
r
k
s
tr
u
ctu
r
e
iter
ativ
ely
an
d
in
teg
r
ates th
ese
p
r
u
n
ed
s
u
b
n
etwo
r
k
s
in
to
a
r
o
b
u
s
t e
n
s
em
b
le.
2
.
2
.
Da
t
a
s
et
a
nd
p
re
pro
ce
s
s
i
ng
W
e
u
tili
ze
d
th
e
SS
T
-
2
an
d
C
o
L
A
d
atasets
f
r
o
m
th
e
g
en
er
al
lan
g
u
ag
e
u
n
d
er
s
tan
d
in
g
ev
alu
atio
n
(
GL
UE
)
b
en
c
h
m
ar
k
f
o
r
o
u
r
e
x
p
er
im
en
ts
,
d
iv
id
in
g
an
d
s
h
u
f
f
lin
g
th
e
d
atasets
u
s
in
g
d
if
f
e
r
en
t
r
an
d
o
m
s
ee
d
s
to
en
s
u
r
e
th
at
th
e
m
o
d
el
s
ee
th
e
d
ata
in
d
if
f
er
en
t
o
r
d
er
,
with
t
h
e
B
E
R
T
m
o
d
el
ch
ec
k
p
o
in
t
“
b
er
t
-
b
ase
-
u
n
ca
s
ed
”
em
p
lo
y
ed
.
T
h
e
SS
T
-
2
d
ataset
is
a
b
in
ar
y
s
en
tim
en
t
class
if
i
ca
tio
n
(
p
o
s
itiv
e/n
eg
ativ
e
s
en
ti
m
en
t)
d
ataset
u
s
ed
f
o
r
s
en
ten
ce
-
lev
el
s
en
tim
en
t
an
aly
s
is
.
T
h
e
d
ataset
d
o
es
n
o
t
n
ee
d
to
b
e
p
r
o
ce
s
s
ed
an
d
is
wid
ely
u
s
ed
in
NL
P
task
s
,
en
s
u
r
in
g
th
at
s
en
ten
ce
s
ar
e
well
-
f
o
r
m
atted
an
d
lab
els
ar
e
co
n
s
is
ten
t.
Fo
r
th
e
SS
T
-
2
d
ataset,
th
e
tr
ain
in
g
s
et
was
s
lig
h
tly
r
ed
u
ce
d
t
o
4
0
,
0
0
0
s
h
u
f
f
led
s
am
p
les,
th
e
v
alid
atio
n
s
et
co
m
p
r
is
ed
8
0
0
s
a
m
p
les,
an
d
t
h
e
test
s
et
in
clu
d
ed
1
,
5
0
0
s
am
p
les
r
ea
s
o
n
b
ein
g
to
allo
w
f
ast
ex
p
er
im
en
tatio
n
,
to
m
i
n
im
ize
th
e
r
eq
u
ir
em
e
n
ts
o
f
co
m
p
u
tatio
n
al
r
eso
u
r
ce
s
,
in
clu
d
in
g
m
em
o
r
y
,
p
r
o
ce
s
s
in
g
p
o
w
er
,
an
d
tim
e.
Fo
r
th
e
C
o
L
A
d
a
taset wh
ich
is
u
s
ed
f
o
r
lin
g
u
is
tic
ac
ce
p
tab
ilit
y
class
if
icatio
n
to
te
s
ts
a
m
o
d
el’
s
u
n
d
er
s
tan
d
in
g
o
f
lin
g
u
is
tic
r
u
les.
T
h
e
C
o
L
A
d
ataset
is
well
-
o
r
g
an
ized
,
wit
h
g
r
a
m
m
atica
lly
ac
ce
p
ta
b
le
a
n
d
u
n
ac
ce
p
tab
le
s
en
ten
ce
s
alr
ea
d
y
lab
ele
d
an
d
h
as
b
ee
n
d
o
n
e
b
y
lin
g
u
is
tic
ex
p
e
r
ts
,
to
en
s
u
r
e
q
u
ality
an
d
r
eli
ab
le
d
ata.
Fo
r
o
u
r
ex
p
e
r
im
en
ts
th
e
tr
ain
in
g
s
et
co
n
tain
ed
8
,
0
0
0
s
h
u
f
f
led
s
am
p
les,
th
e
v
alid
atio
n
s
et
co
n
s
is
ted
o
f
1
,
0
0
0
s
am
p
les,
an
d
th
e
test
s
e
t
h
ad
1
,
0
0
0
s
am
p
les.
T
h
is
s
p
ec
if
ic
s
p
litt
in
g
an
d
s
h
u
f
f
lin
g
en
s
u
r
e
d
th
at
o
u
r
m
o
d
el
tr
ain
in
g
a
n
d
ev
alu
atio
n
wer
e
b
ased
o
n
d
iv
er
s
e
an
d
r
ep
r
esen
tativ
e
d
at
a
s
am
p
les.
2
.
3
.
M
o
del a
nd
t
o
k
enizer
i
nitia
liza
t
io
n
W
e
in
itialized
th
e
B
E
R
T
m
o
d
el
f
o
r
s
eq
u
en
ce
class
if
icatio
n
an
d
th
e
co
r
r
esp
o
n
d
i
n
g
to
k
e
n
izer
u
s
in
g
p
r
e
-
tr
ain
ed
c
h
ec
k
p
o
in
ts
.
T
h
is
s
tep
en
s
u
r
es
th
at
o
u
r
m
o
d
el
b
en
ef
its
f
r
o
m
tr
an
s
f
er
lear
n
in
g
,
lev
er
ag
in
g
th
e
v
ast
k
n
o
wled
g
e
e
n
co
d
e
d
in
th
e
B
E
R
T
ar
ch
itectu
r
e.
T
h
e
B
E
R
T
m
o
d
el
an
d
A
u
to
To
ke
n
iz
er
f
r
o
m
th
e
h
u
g
g
in
g
f
ac
e
tr
an
s
f
o
r
m
er
s
lib
r
ar
y
ass
o
ciate
d
with
b
ert
-
b
a
s
e
-
u
n
ca
s
ed
ca
ll
ed
B
ertT
o
ke
n
iz
er
,
wer
e
ca
r
ef
u
lly
ch
o
s
en
to
alig
n
with
o
u
r
ex
p
er
im
en
tal
r
e
q
u
ir
e
m
en
ts
,
en
ab
lin
g
e
f
f
icien
t a
n
d
ac
cu
r
ate
s
eq
u
en
ce
class
if
icatio
n
task
s
.
2
.
4
.
E
ns
em
ble o
f
winn
ing
t
ick
et
s
An
em
p
ty
en
s
em
b
le
s
et
was
in
itialized
to
s
to
r
e
th
e
win
n
in
g
tick
ets.
T
o
f
ac
ilit
ate
th
e
p
r
u
n
in
g
o
f
atten
tio
n
h
ea
d
s
,
a
m
ask
m
at
r
ix
with
all
v
alu
es
s
et
to
1
wer
e
cr
ea
ted
.
T
h
is
m
ask
m
a
tr
ix
is
cr
u
cial
f
o
r
id
en
tify
in
g
an
d
r
etain
in
g
th
e
m
o
s
t
cr
itical
co
m
p
o
n
en
ts
o
f
B
E
R
T
ar
ch
itectu
r
e.
A
s
u
b
-
n
e
two
r
k
(
tick
et)
was
r
an
d
o
m
l
y
in
itialized
b
ased
o
n
th
e
B
E
R
T
ar
ch
itectu
r
e,
u
s
in
g
th
e
s
p
ec
if
ied
ch
ec
k
p
o
i
n
t.
T
h
e
o
p
tim
izer
an
d
lo
s
s
f
u
n
ctio
n
f
o
r
th
e
s
u
b
-
n
etwo
r
k
wer
e
s
et
to
Ad
am
an
d
cr
o
s
s
-
en
tr
o
p
y
lo
s
s
,
r
esp
ec
tiv
ely
.
T
h
ese
co
n
f
ig
u
r
atio
n
s
allo
wed
f
o
r
ef
f
e
ctiv
e
o
p
tim
iza
tio
n
an
d
tr
ain
in
g
o
f
t
h
e
s
u
b
-
n
etwo
r
k
,
en
s
u
r
in
g
its
p
er
f
o
r
m
a
n
ce
alig
n
s
with
th
e
o
v
er
all
m
o
d
el
g
o
als.
2
.
5
.
T
ra
ini
ng
c
o
nfig
ura
t
io
n
T
h
e
tr
ain
in
g
p
ar
am
ete
r
s
wer
e
m
eticu
lo
u
s
ly
ch
o
s
en
to
o
p
t
im
ize
th
e
lear
n
in
g
p
r
o
ce
s
s
.
W
e
s
e
t
th
e
n
u
m
b
er
o
f
tr
ain
in
g
iter
atio
n
s
to
f
iv
e,
with
a
lear
n
in
g
r
ate
o
f
2
e
-
5
an
d
a
b
atch
s
ize
o
f
1
6
.
A
co
n
v
e
r
g
en
ce
th
r
esh
o
ld
o
f
0
.
0
0
1
was
estab
li
s
h
ed
to
en
s
u
r
e
s
tab
le
tr
ain
in
g
.
T
h
e
tr
ain
in
g
ar
g
u
m
en
ts
in
clu
d
ed
p
ar
am
eter
s
f
o
r
s
av
in
g
th
e
m
o
d
el,
lo
ad
i
n
g
th
e
b
est m
o
d
el
at
th
e
en
d
o
f
tr
ain
i
n
g
,
s
ettin
g
th
e
n
u
m
b
e
r
o
f
t
r
ain
in
g
ep
o
c
h
s
to
th
r
ee
,
an
d
d
e
f
in
in
g
th
e
e
v
alu
atio
n
a
n
d
s
av
in
g
s
tr
ateg
ies.
T
h
e
m
et
r
ices
f
o
r
s
elec
tin
g
th
e
b
est
m
o
d
el
was
s
et
to
th
e
F1
-
s
co
r
e,
ac
cu
r
ac
y
an
d
th
e
t
o
tal
n
u
m
b
e
r
o
f
s
av
ed
m
o
d
el
s
was
lim
ited
to
1
0
.
T
h
ese
co
n
f
ig
u
r
atio
n
s
wer
e
d
esig
n
ed
to
en
s
u
r
e
th
at
o
u
r
m
o
d
el
tr
ain
i
n
g
was
b
o
th
ef
f
icien
t
an
d
r
o
b
u
s
t,
p
r
e
v
en
tin
g
o
v
er
f
itti
n
g
wh
ile
m
ax
im
izin
g
p
e
r
f
o
r
m
an
ce
.
2
.
6
.
Select
io
n o
f
info
rma
t
iv
e
a
t
t
ent
io
n hea
ds
us
ing
t
he
lo
t
t
er
y
t
ic
k
et
hy
po
t
hes
is
A
s
alien
cy
s
co
r
e
is
a
m
etr
ic
u
s
ed
to
m
ea
s
u
r
e
th
e
im
p
o
r
ta
n
ce
o
r
r
elev
an
ce
o
f
in
d
iv
id
u
a
l
elem
en
ts
with
in
a
m
o
d
el,
s
u
ch
as
n
eu
r
o
n
s
,
f
ea
tu
r
es,
o
r
atten
tio
n
h
ea
d
s
,
b
ased
o
n
th
eir
co
n
tr
ib
u
tio
n
to
th
e
m
o
d
el'
s
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J E
lec
&
C
o
m
p
E
n
g
I
SS
N:
2088
-
8
7
0
8
E
n
s
emb
le
o
f wi
n
n
in
g
ticke
ts
:
p
r
u
n
in
g
b
id
ir
ec
tio
n
a
l e
n
co
d
er fr
o
m
th
e
tr
a
n
s
fo
r
mers
…
(
N
ya
la
la
n
i S
ma
r
ts
)
2073
p
er
f
o
r
m
an
ce
.
Salien
cy
s
co
r
es
ar
e
ty
p
ically
d
er
iv
e
d
f
r
o
m
g
r
ad
ien
ts
o
r
o
th
er
s
en
s
itiv
ity
m
ea
s
u
r
es
co
m
p
u
te
d
with
r
esp
ec
t
to
th
e
in
p
u
t
f
ea
tu
r
es
o
r
p
ar
am
eter
s
o
f
th
e
m
o
d
e
l
[
2
9
]
.
I
n
th
e
ca
s
e
o
f
atten
tio
n
-
b
ased
m
o
d
els
lik
e
B
E
R
T
,
s
alien
cy
s
co
r
es
wer
e
co
m
p
u
ted
f
o
r
th
e
atten
tio
n
h
ea
d
s
,
wh
ich
r
ep
r
esen
t
th
e
r
elev
an
ce
o
f
d
if
f
er
e
n
t
p
ar
ts
o
f
th
e
in
p
u
t to
t
h
e
m
o
d
el's o
u
tp
u
t
[
3
0
]
.
L
et
d
en
o
te
th
e
th
atten
tio
n
h
ea
d
.
T
h
e
s
alien
cy
s
co
r
e
f
o
r
at
ten
tio
n
h
ea
d
d
en
o
ted
as
we
r
e
co
m
p
u
ted
b
ased
o
n
its
im
p
ac
t
o
n
th
e
m
o
d
el’
s
l
o
s
s
o
r
o
u
tp
u
t
.
W
e
m
ea
s
u
r
ed
th
e
c
h
an
g
e
in
th
e
m
o
d
el’
s
lo
s
s
wh
en
is
p
er
tu
r
b
ed
o
r
m
ask
ed
:
=
(
(
,
(
)
)
)
−
(
(
)
)
(
1
)
Her
e,
m
ask
r
ep
r
esen
ts
a
m
o
d
i
f
icatio
n
o
f
t
h
e
m
o
d
el
wh
er
e
th
e
th
atten
tio
n
h
ea
d
is
m
ask
ed
o
r
ze
r
o
e
d
o
u
t.
2
.
7
.
P
r
a
ct
ica
l
c
a
lcula
t
io
ns
a
nd
a
lg
o
rit
hm
s
f
o
r
identif
y
ing
winn
ing
t
ick
et
s
I
n
th
is
s
u
b
s
ec
tio
n
,
we
s
p
ec
if
y
th
e
s
tep
s
in
d
etail
an
d
th
e
al
g
o
r
ith
m
th
at
was
u
s
ed
to
ca
r
r
y
o
u
t
th
e
en
s
em
b
le
o
f
win
n
in
g
tick
ets
,
wh
ich
p
ar
tic
u
lar
ly
h
elp
s
in
r
e
d
u
cin
g
th
e
n
u
m
b
er
o
f
atten
tio
n
h
ea
d
s
o
f
B
E
R
T
m
o
r
e
ef
f
icien
tly
with
o
u
t
co
m
p
r
o
m
is
in
g
o
n
th
e
p
e
r
f
o
r
m
an
c
e
o
f
th
e
m
o
d
el.
T
h
e
m
ain
g
o
al
is
to
tr
ain
th
e
s
u
b
n
ets
o
f
th
e
B
E
R
T
m
o
d
el
a
n
d
p
r
u
n
e
th
em
,
s
ea
r
c
h
in
g
f
o
r
p
o
s
s
ib
le
s
u
b
n
ets
wh
er
e
th
ei
r
p
er
f
o
r
m
an
ce
ca
n
b
e
eq
u
al
o
r
h
ig
h
e
r
th
an
th
e
p
e
r
f
o
r
m
an
ce
o
f
th
e
o
r
ig
in
al
m
o
d
e
l.
Her
e
we
d
escr
ib
e
m
eth
o
d
s
wh
ich
m
ak
e
u
s
e
o
f
in
d
ep
en
d
en
t p
r
u
n
ed
m
o
d
els
t
o
en
h
a
n
ce
th
e
f
in
al
m
o
d
el
p
er
f
o
r
m
an
ce
b
y
u
s
in
g
m
o
d
el
en
s
em
b
le
tech
n
iq
u
e.
T
h
e
A
lg
o
r
ith
m
1
g
i
v
en
b
elo
w
p
r
esen
ts
a
s
tep
-
by
-
s
tep
s
tr
ateg
y
f
o
r
estab
lis
h
in
g
an
d
c
r
ea
tin
g
t
h
e
s
e
win
n
in
g
tick
ets
lev
er
ag
in
g
b
o
th
tr
ain
in
g
an
d
v
alid
atio
n
d
atasets
to
en
s
u
r
e
o
p
tim
al
p
er
f
o
r
m
a
n
ce
.
T
h
e
p
r
o
ce
s
s
lead
s
u
p
to
ag
g
r
eg
atin
g
p
r
ed
ictio
n
s
f
r
o
m
m
u
ltip
le
p
r
u
n
ed
s
u
b
n
etwo
r
k
s
,
u
s
in
g
s
im
p
le
av
er
a
g
in
g
to
g
en
er
ate
r
o
b
u
s
t
an
d
r
eliab
le
f
in
al
p
r
ed
ictio
n
s
.
T
h
is
m
eth
o
d
aim
s
f
o
r
b
etter
p
er
f
o
r
m
a
n
ce
in
ter
m
s
o
f
co
m
p
u
tatio
n
al
tim
e
b
y
p
r
o
v
id
i
n
g
n
o
r
m
aliza
tio
n
a
n
d
e
n
h
an
cin
g
th
e
ef
f
ec
tiv
en
ess
o
f
t
h
e
lar
g
e
m
o
d
els,
s
u
ch
as B
E
R
T
.
Alg
o
r
ith
m
1
.
E
n
s
em
b
le
o
f
win
n
in
g
tick
ets ap
p
r
o
ac
h
f
o
r
B
E
R
T
I
n
p
u
t
:
−
A
p
r
e
-
tr
ain
ed
B
E
R
T
m
o
d
el:
−
Nu
m
b
er
o
f
en
s
em
b
les to
b
e
cr
ea
ted
:
−
Valid
atio
n
d
ataset:
−
T
est d
ataset:
Ou
tp
u
t
:
−
An
en
s
em
b
le
o
f
win
n
in
g
tic
k
e
ts
:
Step
s
1.
L
o
ad
th
e
p
r
e
-
tr
ain
e
d
B
E
R
T
m
o
d
el:
Step
s
2.
I
n
itialize
an
em
p
ty
e
n
s
em
b
le
s
et
to
s
to
r
e
th
e
win
n
in
g
tick
ets:
=
{
}
Step
s
3.
Set th
e
n
u
m
b
er
o
f
e
n
s
em
b
les to
b
e
cr
ea
ted
:
Step
s
4.
Fo
r
ea
ch
en
s
em
b
le
iter
atio
n
f
r
o
m
1
to
:
a.
I
n
itialize
a
m
ask
m
atr
ix
f
o
r
th
e
atten
tio
n
h
ea
d
s
with
all
v
al
u
es set to
1
:
=
1
×
b.
R
an
d
o
m
ly
in
itialize
a
s
u
b
-
n
etwo
r
k
(
tick
et)
b
ased
o
n
th
e
B
E
R
T
ar
ch
itectu
r
e:
=
_
(
)
c.
T
r
ain
th
e
s
u
b
-
n
etwo
r
k
f
o
r
a
f
i
x
ed
n
u
m
b
er
o
f
iter
atio
n
s
o
r
u
n
til co
n
v
er
g
e
n
ce
o
n
a
tr
ain
in
g
d
ataset:
=
(
,
)
d.
E
v
alu
ate
th
e
s
u
b
-
n
etwo
r
k
'
s
p
er
f
o
r
m
a
n
ce
o
n
th
e
v
alid
atio
n
d
a
taset:
=
(
,
)
e.
I
f
th
e
s
u
b
-
n
etwo
r
k
m
ee
ts
th
e
d
esire
d
p
er
f
o
r
m
an
ce
c
r
iter
ia:
−
Ap
p
ly
th
e
m
ask
m
atr
ix
to
th
e
s
u
b
-
n
etwo
r
k
,
ef
f
ec
tiv
ely
p
r
u
n
in
g
t
h
e
s
elec
ted
atten
tio
n
h
ea
d
s
:
=
(
,
)
−
Ad
d
th
e
p
r
u
n
e
d
s
u
b
-
n
etwo
r
k
(
win
n
in
g
tick
et)
to
th
e
en
s
em
b
l
e
s
et:
=
∪
Step
s
5.
Fo
r
ea
ch
s
am
p
le
in
th
e
test
d
ataset:
a.
Feed
th
e
s
am
p
le
th
r
o
u
g
h
ea
ch
win
n
in
g
tick
et
in
t
h
e
en
s
em
b
le
s
et:
=
(
)
b.
Ag
g
r
eg
ate
th
e
p
r
ed
ictio
n
s
f
r
o
m
ea
ch
win
n
in
g
tick
et
to
o
b
tai
n
th
e
f
in
al
p
r
ed
ictio
n
:
(
)
=
(
1
(
)
,
2
(
)
,
…
,
(
)
)
=
1
∑
=
1
(
)
Step
s
6.
E
v
alu
ate
th
e
p
e
r
f
o
r
m
an
ce
o
f
t
h
e
en
s
em
b
le
o
n
th
e
test
d
ataset
u
s
in
g
ap
p
r
o
p
r
iate
m
et
r
ics:
=
(
,
)
Step
s
7.
R
ep
ea
t step
s
4
to
6
f
o
r
th
e
d
es
ir
ed
n
u
m
b
er
o
f
iter
atio
n
s
o
r
en
s
em
b
les.
Step
s
8.
Ag
g
r
eg
ate
th
e
p
er
f
o
r
m
an
ce
m
etr
ics ac
r
o
s
s
all
en
s
em
b
les to
ass
es
s
th
e
o
v
er
all
p
er
f
o
r
m
a
n
ce
o
f
th
e
en
s
em
b
le
m
eth
o
d
:
=
_
(
)
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
0
8
8
-
8
7
0
8
I
n
t J E
lec
&
C
o
m
p
E
n
g
,
Vo
l.
15
,
No
.
2
,
Ap
r
il
20
25
:
2
0
7
0
-
2
0
8
0
2074
a.
No
tatio
n
−
E
n
s
em
b
le
o
f
win
n
in
g
tick
ets
:
c
o
m
b
in
in
g
m
u
ltip
le
p
r
u
n
e
d
m
o
d
els
wh
er
e
=
1
,
2
,
…
,
.
(
)
=
(
1
(
)
,
2
(
)
,
…
,
(
)
)
(
2
)
−
Ag
g
r
eg
atio
n
f
u
n
ctio
n
:
a
g
g
r
eg
a
tes th
e
p
r
ed
ictio
n
s
o
f
th
e
p
r
u
n
ed
m
o
d
els u
s
in
g
s
im
p
le
a
v
er
a
g
in
g
.
(
)
=
(
1
(
)
,
2
(
)
,
…
,
(
)
)
=
1
∑
=
1
(
)
(
3
)
b.
Pro
ce
s
s
o
f
f
o
r
m
in
g
an
e
n
s
em
b
le
−
Fo
r
m
th
e
en
s
em
b
le
s
et
:
=
=
⋃
=
1
(
(
,
1
)
∣
(
,
)
)
(
4
)
w
h
er
e:
=
(
(
1
)
,
)
(
5
)
−
Ma
k
e
p
r
ed
ictio
n
s
with
ag
g
r
e
g
atio
n
:
Ob
tain
in
d
iv
id
u
al
p
r
ed
ictio
n
s
f
r
o
m
ea
c
h
p
r
u
n
ed
s
u
b
-
n
etwo
r
k
:
(
)
=
1
,
2
,
…
,
(
6
)
Ag
g
r
eg
ate
th
ese
p
r
ed
ictio
n
s
u
s
in
g
th
e
ag
g
r
eg
atio
n
f
u
n
ctio
n
:
(
)
=
(
1
(
)
,
2
(
)
,
…
,
(
)
)
=
1
∑
=
1
(
)
(
7
)
T
h
e
s
t
e
p
s
o
u
t
li
n
e
d
i
n
t
h
e
a
l
g
o
r
i
t
h
m
a
b
o
v
e
d
e
m
o
n
s
t
r
at
e
t
h
e
p
r
o
c
e
s
s
o
f
c
r
e
a
ti
n
g
a
n
e
n
s
em
b
l
e
f
r
o
m
p
r
u
n
e
d
a
t
t
e
n
t
i
o
n
h
e
a
d
s
o
f
a
B
E
R
T
m
o
d
e
l
a
n
d
u
s
i
n
g
t
h
is
e
n
s
e
m
b
l
e
f
o
r
a
g
g
r
e
g
a
t
e
d
p
r
e
d
i
c
t
i
o
n
s
.
T
h
e
c
h
o
s
e
n
m
e
t
h
o
d
s
e
n
s
u
r
e
t
h
e
r
e
l
i
a
b
i
l
it
y
a
n
d
r
e
p
r
o
d
u
c
i
b
i
l
i
t
y
o
f
r
e
s
u
lt
s
.
T
h
e
S
S
T
-
2
a
n
d
C
o
L
A
d
a
ta
s
e
t
s
f
r
o
m
t
h
e
G
L
UE
b
e
n
c
h
m
a
r
k
p
r
o
v
i
d
e
s
a
r
o
b
u
s
t
f
o
u
n
d
a
t
i
o
n
f
o
r
o
u
r
s
t
u
d
y
,
w
h
i
l
e
t
h
e
B
E
R
T
m
o
d
e
l
p
r
o
v
i
d
e
s
a
s
t
a
n
d
a
r
d
i
n
s
e
q
u
e
n
c
e
c
l
a
s
s
i
f
ic
a
t
i
o
n
t
as
k
s
,
b
e
n
e
f
it
s
f
r
o
m
p
r
e
-
t
r
a
i
n
e
d
c
h
e
c
k
p
o
i
n
t
s
t
o
e
n
h
a
n
c
e
p
e
r
f
o
r
m
a
n
c
e
.
T
h
e
e
n
s
e
m
b
l
e
o
f
w
i
n
n
i
n
g
t
i
c
k
e
ts
a
p
p
r
o
a
c
h
a
i
m
s
t
o
i
m
p
r
o
v
e
m
o
d
e
l
e
f
f
i
ci
e
n
c
y
b
y
f
o
c
u
s
i
n
g
o
n
t
h
e
m
o
s
t
c
r
u
c
i
al
c
o
m
p
o
n
e
n
t
s
o
f
t
h
e
B
E
R
T
a
r
c
h
i
t
e
c
t
u
r
e
.
T
h
e
t
r
a
i
n
i
n
g
c
o
n
f
i
g
u
r
a
t
i
o
n
s
a
r
e
d
e
s
i
g
n
e
d
t
o
o
p
t
i
m
i
z
e
t
h
e
l
e
a
r
n
i
n
g
p
r
o
c
ess
w
h
i
l
e
p
r
e
v
e
n
t
i
n
g
o
v
e
r
f
i
t
t
i
n
g
,
e
n
s
u
r
i
n
g
t
h
e
r
o
b
u
s
tn
e
s
s
o
f
o
u
r
f
i
n
d
i
n
g
s
.
F
i
g
u
r
e
2
s
h
o
w
s
t
h
e
s
i
m
u
l
at
i
o
n
o
f
t
h
e
e
n
s
e
m
b
l
e
m
o
d
e
l
d
u
r
i
n
g
t
h
e
t
r
a
i
n
i
n
g
p
r
o
c
e
s
s
.
O
u
r
i
m
p
l
em
e
n
t
a
t
i
o
n
c
o
d
e
s
c
a
n
b
e
f
o
u
n
d
a
t
G
i
tH
u
b
:
h
t
t
p
s
:
//
g
i
t
h
u
b
.
c
o
m
/n
k
s
m
a
r
t
s
/
E
n
s
em
b
l
e
-
of
-
w
i
n
n
i
n
g
-
t
ic
k
e
ts
/
b
l
o
b
/m
a
i
n
/N
e
w
_
p
r
u
n
i
n
g
_
B
E
R
T
_
a
tt
e
n
t
i
o
n
_
h
e
a
d
s
_
F
1
_
s
c
o
r
e
.
i
p
y
n
b
.
Fig
u
r
e
2
.
Simu
latio
n
o
f
e
n
s
em
b
le
m
o
d
el
d
u
r
in
g
th
e
tr
ain
i
n
g
p
r
o
ce
s
s
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J E
lec
&
C
o
m
p
E
n
g
I
SS
N:
2088
-
8
7
0
8
E
n
s
emb
le
o
f wi
n
n
in
g
ticke
ts
:
p
r
u
n
in
g
b
id
ir
ec
tio
n
a
l e
n
co
d
er fr
o
m
th
e
tr
a
n
s
fo
r
mers
…
(
N
ya
la
la
n
i S
ma
r
ts
)
2075
3.
RE
SU
L
T
S AN
D
D
I
SCU
SS
I
O
N
T
h
is
s
ec
tio
n
p
r
esen
ts
a
d
etailed
o
v
er
v
iew
o
f
o
u
r
e
x
p
er
im
e
n
ts
,
r
esu
lts
,
an
d
d
is
cu
s
s
io
n
s
,
f
o
cu
s
in
g
o
n
th
e
p
er
f
o
r
m
a
n
ce
a
n
d
ef
f
icien
c
y
o
f
th
e
B
E
R
T
m
o
d
el
as
co
m
p
ar
ed
to
o
u
r
en
s
em
b
le
m
o
d
el
o
n
th
e
C
o
L
A
an
d
SST
-
2
d
atasets
.
T
h
e
an
aly
s
is
in
clu
d
es
a
p
e
r
f
o
r
m
an
ce
c
o
m
p
a
r
is
o
n
o
f
th
e
p
r
u
n
e
d
en
s
em
b
le
m
o
d
els
ag
ain
s
t
th
e
b
aselin
e
m
o
d
els,
o
f
f
er
in
g
in
s
ig
h
ts
in
to
th
e
ef
f
icien
cy
-
p
er
f
o
r
m
an
ce
tr
ad
e
-
o
f
f
.
Fu
r
th
e
r
m
o
r
e
,
we
in
v
esti
g
ate
th
e
ef
f
ec
ts
o
f
p
r
u
n
i
n
g
o
n
ac
c
u
r
ac
y
,
in
f
e
r
en
ce
s
p
ee
d
,
an
d
co
m
p
u
tatio
n
al
ex
p
e
n
s
es,
p
o
in
tin
g
o
u
t
th
e
b
e
n
ef
its
an
d
s
h
o
r
tco
m
in
g
s
o
f
u
s
in
g
p
r
u
n
ed
m
o
d
els.
T
h
e
r
esu
lts
em
p
h
asize
o
n
th
e
a
d
v
an
ta
g
e
o
f
th
e
e
n
s
em
b
le
m
eth
o
d
to
b
r
in
g
o
u
t a
n
im
p
r
o
v
ed
p
er
f
o
r
m
an
ce
in
ef
f
icien
cy
with
o
u
t sacr
if
icin
g
th
e
task
p
e
r
f
o
r
m
an
ce
.
3
.
1
.
E
x
perim
ent
s
s
et
up
W
e
s
tar
t
with
th
e
B
E
R
T
b
ase
-
u
n
ca
s
ed
m
o
d
el,
f
ea
tu
r
in
g
1
2
lay
er
s
,
1
2
atten
tio
n
h
ea
d
s
,
1
1
0
m
illi
o
n
p
ar
am
eter
s
,
a
n
d
7
6
8
h
i
d
d
en
u
n
its
p
er
lay
er
.
Usi
n
g
d
if
f
er
e
n
t
r
an
d
o
m
s
ee
d
s
,
we
ev
alu
ate
with
F1
-
s
co
r
e
an
d
ac
cu
r
ac
y
.
T
r
ain
in
g
p
a
r
am
ete
r
s
in
clu
d
e
a
b
atch
s
ize
o
f
1
6
,
lear
n
in
g
r
ate
o
f
2
e
-
5
,
5
iter
atio
n
s
,
an
d
a
co
n
v
er
g
en
ce
th
r
esh
o
ld
o
f
0
.
0
0
1
,
we
u
s
e
Go
o
g
le
C
o
llab
a
c
lo
u
d
s
er
v
ice
o
n
a
s
in
g
le
NVI
DI
A
T
esla
T
4
GPU.
W
e
u
s
e
th
e
Ad
am
o
p
tim
izer
a
n
d
cr
o
s
s
-
en
tr
o
p
y
l
o
s
s
f
u
n
ctio
n
to
u
p
d
ate
p
ar
am
eter
s
an
d
m
in
im
ize
lo
s
s
.
3
.
2
.
P
er
f
o
r
m
a
nce
c
o
m
pa
riso
n
T
h
e
ef
f
ec
tiv
en
ess
o
f
o
u
r
e
n
s
em
b
le
m
o
d
el
is
e
v
id
en
t
i
n
it
s
p
er
f
o
r
m
an
ce
o
n
th
e
C
o
L
A
an
d
SS
T
-
2
d
atasets
,
as
s
h
o
wn
in
T
ab
le
2
.
W
e
co
m
p
a
r
e
o
u
r
en
s
em
b
le
m
o
d
el
to
B
E
R
T
b
aselin
e
m
o
d
el,
E
ar
ly
B
E
R
T
,
an
d
d
o
m
in
an
t
win
n
in
g
tick
et
u
s
in
g
th
e
d
ef
a
u
lt
tr
ain
in
g
s
ettin
g
s
f
o
r
p
r
e
-
tr
ain
in
g
an
d
f
in
e
-
tu
n
in
g
.
Ou
r
e
n
s
em
b
le
ac
h
iev
es
s
u
p
er
io
r
ac
c
u
r
ac
y
a
n
d
F1
-
s
co
r
es
o
f
9
5
%
an
d
9
4
%
o
n
C
o
L
A,
an
d
9
6
%
an
d
9
6
%
o
n
SS
T
-
2
.
T
h
ese
r
esu
lts
in
d
icate
th
e
ef
f
icien
cy
g
ain
s
o
f
o
u
r
ap
p
r
o
ac
h
,
b
ala
n
cin
g
th
e
h
ig
h
er
p
e
r
f
o
r
m
an
ce
o
f
th
e
m
o
d
el
wh
ile
m
an
ag
in
g
its
co
m
p
lex
ity
.
Ou
r
f
in
d
i
n
g
s
alig
n
with
p
r
ev
io
u
s
s
tu
d
ies
wh
ich
d
em
o
n
s
tr
ate
th
at
en
s
em
b
le
m
eth
o
d
s
h
av
e
s
h
o
wn
p
r
o
m
is
e
in
en
h
an
cin
g
m
o
d
el
p
er
f
o
r
m
a
n
ce
ac
r
o
s
s
v
ar
io
u
s
task
s
b
y
lev
er
ag
in
g
th
e
s
tr
en
g
th
s
o
f
m
u
ltip
le
m
o
d
els
[
2
5
]
,
p
ar
ticu
lar
l
y
with
[
2
6
]
,
wh
ich
also
id
en
tifie
d
f
o
u
r
s
u
b
n
etwo
r
k
s
d
u
r
i
n
g
th
e
f
in
e
-
t
u
n
in
g
p
r
o
ce
s
s
.
T
h
e
co
n
s
is
ten
t
o
u
tp
er
f
o
r
m
an
ce
o
f
o
u
r
en
s
em
b
le
m
o
d
el
o
v
er
o
th
e
r
b
aselin
es
in
d
icate
s
h
o
w
im
p
o
r
tan
t
an
en
s
em
b
le
tech
n
iq
u
e
is
in
n
at
u
r
al
lan
g
u
a
g
e
p
r
o
ce
s
s
in
g
.
T
h
e
s
u
b
s
tan
tial
r
ed
u
ctio
n
in
p
a
r
am
eter
s
f
r
o
m
1
1
0
m
illi
o
n
in
t
h
e
B
E
R
T
b
asel
in
e
to
7
0
m
illi
o
n
af
ter
p
r
u
n
i
n
g
s
h
o
ws th
at
p
r
u
n
i
n
g
n
o
t o
n
ly
e
n
h
an
ce
s
ef
f
icien
c
y
b
u
t a
ls
o
im
p
r
o
v
es
g
en
er
aliza
tio
n
ac
r
o
s
s
d
atasets
.
Ou
r
en
s
em
b
le
m
o
d
el
o
f
f
e
r
s
a
co
m
p
ellin
g
k
ey
f
o
r
task
s
r
e
q
u
i
r
in
g
h
i
g
h
ac
cu
r
ac
y
an
d
ef
f
icien
c
y
.
T
h
e
r
esu
lts
p
r
esen
ted
s
u
p
p
o
r
t
th
e
h
y
p
o
th
esis
th
at
co
m
b
in
in
g
p
r
u
n
in
g
tec
h
n
i
q
u
e
with
en
s
em
b
le
m
eth
o
d
s
ca
n
g
r
ea
tly
b
o
o
s
t d
ee
p
lear
n
in
g
m
o
d
el
p
er
f
o
r
m
an
ce
.
T
ab
le
2
.
Mo
d
el
p
er
f
o
r
m
a
n
ce
c
o
m
p
ar
is
o
n
with
o
th
er
e
n
s
em
b
l
e
tech
n
iq
u
es
M
o
d
e
l
D
a
t
a
s
e
t
M
e
t
r
i
c
s
R
e
f
e
r
e
n
c
e
B
ER
T
C
o
LA
6
0
.
5
[
1
]
Ea
r
l
y
B
E
R
T
C
o
LA
52
[
2
5
]
D
o
mi
n
a
n
t
w
i
n
n
i
n
g
t
i
c
k
e
t
C
o
LA
68
[
2
7
]
O
u
r
e
n
sem
b
l
e
m
o
d
e
l
C
o
LA
9
5
.
0
-
B
ER
T
SST
-
2
9
2
.
7
[
1
]
Ea
r
l
y
B
E
R
T
SST
-
2
9
0
.
7
1
[
2
5
]
D
o
mi
n
a
n
t
w
i
n
n
i
n
g
t
i
c
k
e
t
SST
-
2
9
6
.
4
[
2
7
]
O
u
r
e
n
sem
b
l
e
m
o
d
e
l
SST
-
2
9
6
.
0
-
3
.
3
.
E
x
perim
ent
a
l
f
ind
ing
s
o
n m
o
del
perf
o
rma
nce
t
re
nd
s
T
h
e
p
er
f
o
r
m
an
ce
o
f
o
u
r
en
s
em
b
le
m
o
d
el,
as
s
h
o
wn
in
Fig
u
r
es
3
,
h
ig
h
lig
h
ts
k
ey
f
in
d
in
g
s
in
its
ap
p
licatio
n
to
th
e
C
o
L
A
a
n
d
SS
T
-
2
d
atasets
.
Sp
ec
if
ically
,
o
n
t
h
e
C
o
L
A
d
ataset
Fig
u
r
e
3
(
a)
,
th
e
F1
-
s
co
r
e
ex
h
ib
its
an
in
itial
s
tead
y
in
c
r
ea
s
e,
p
ea
k
in
g
at
1
7
0
0
a
n
d
2
0
0
0
tr
ain
in
g
s
tep
s
b
ef
o
r
e
g
r
ad
u
ally
d
ec
lin
in
g
,
wh
er
ea
s
ac
cu
r
ac
y
im
p
r
o
v
es
s
ig
n
if
ican
tly
af
ter
8
0
0
s
tep
s
,
r
ea
ch
in
g
its
p
ea
k
at
1
9
0
0
s
tep
s
with
m
in
o
r
f
lu
ctu
atio
n
s
th
er
ea
f
ter
.
I
n
co
n
tr
ast,
th
e
SS
T
-
2
d
ataset
in
Fi
g
u
r
e
3
(
b
)
r
ev
ea
ls
a
m
o
r
e
c
o
n
s
is
ten
t
p
er
f
o
r
m
a
n
ce
,
with
th
e
F1
-
s
co
r
e
an
d
ac
cu
r
a
cy
b
o
t
h
s
h
o
win
g
co
n
ti
n
u
o
u
s
i
m
p
r
o
v
e
m
en
ts
,
p
ea
k
in
g
at
ar
o
u
n
d
2
4
0
0
an
d
2
3
0
0
s
tep
s
,
r
esp
ec
tiv
ely
.
T
h
e
m
o
d
el's
s
u
p
er
io
r
p
er
f
o
r
m
an
ce
o
n
th
e
SS
T
-
2
d
ataset,
with
an
F1
-
s
co
r
e
o
f
9
6
%,
u
n
d
er
s
co
r
es
its
ef
f
ec
tiv
en
ess
in
s
en
tim
en
t
an
aly
s
is
task
s
,
es
p
ec
ially
wh
en
co
m
p
a
r
ed
to
its
r
esu
lts
o
n
th
e
m
o
r
e
ch
allen
g
in
g
,
im
b
alan
ce
d
C
o
L
A
d
ataset,
n
ev
e
r
th
eless
o
u
r
en
s
em
b
le
m
o
d
el
h
as
s
h
o
wn
a
r
em
ar
k
ab
le
p
er
f
o
r
m
an
ce
with
a
s
co
r
e
o
f
9
5
% c
o
m
p
ar
e
d
to
o
t
h
er
m
o
d
els wh
ich
h
av
e
f
ar
less
p
er
f
o
r
m
an
ce
.
W
h
en
co
m
p
a
r
ed
to
p
r
ev
i
o
u
s
s
tu
d
ies,
s
u
ch
as
th
o
s
e
h
ig
h
lig
h
tin
g
th
e
ca
p
ab
ilit
ies
o
f
B
E
R
T
in
n
atu
r
al
lan
g
u
ag
e
u
n
d
er
s
tan
d
in
g
[
1
3
]
,
o
u
r
en
s
em
b
le
m
o
d
el'
s
p
er
f
o
r
m
an
ce
alig
n
s
well,
p
a
r
ticu
lar
ly
in
s
en
tim
en
t
an
aly
s
is
.
Ho
wev
er
,
th
e
m
o
d
el'
s
p
er
f
o
r
m
an
ce
o
n
th
e
C
o
L
A
d
ataset
s
u
g
g
ests
th
at
wh
ile
it
is
h
ig
h
ly
ef
f
ec
tiv
e
,
it
m
ay
r
eq
u
i
r
e
ad
d
itio
n
al
o
p
tim
izatio
n
s
to
f
u
lly
ad
d
r
ess
th
e
ch
allen
g
es
p
o
s
ed
b
y
im
b
alan
ce
d
d
atasets
.
T
h
e
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
0
8
8
-
8
7
0
8
I
n
t J E
lec
&
C
o
m
p
E
n
g
,
Vo
l.
15
,
No
.
2
,
Ap
r
il
20
25
:
2
0
7
0
-
2
0
8
0
2076
s
tu
d
y
'
s
s
tr
en
g
th
lies
in
d
e
m
o
n
s
tr
atin
g
th
e
m
o
d
el'
s
ad
ap
tab
i
lity
ac
r
o
s
s
d
if
f
e
r
en
t
task
s
,
alt
h
o
u
g
h
th
e
n
ee
d
f
o
r
task
-
s
p
ec
if
ic
ad
ju
s
tm
en
ts
,
s
u
ch
as
ea
r
ly
s
to
p
p
in
g
d
u
r
in
g
tr
ain
in
g
o
n
C
o
L
A,
h
i
g
h
lig
h
ts
a
lim
itatio
n
in
its
g
en
er
aliza
b
ilit
y
with
o
u
t
f
i
n
e
-
tu
n
in
g
.
Un
ex
p
ec
ted
ly
,
th
e
co
n
s
is
ten
t
im
p
r
o
v
em
en
t
i
n
SS
T
-
2
p
er
f
o
r
m
an
ce
s
u
g
g
ests
th
at
th
e
m
o
d
el
is
p
ar
ticu
lar
ly
well
-
s
u
ited
f
o
r
s
en
tim
en
t
an
aly
s
is
,
p
o
s
s
ib
ly
d
u
e
to
th
e
d
ataset'
s
b
alan
ce
d
n
atu
r
e
co
m
p
ar
e
d
to
C
o
L
A.
T
h
e
s
tu
d
y
d
em
o
n
s
tr
ates
th
e
ef
f
ec
tiv
en
ess
o
f
o
u
r
en
s
em
b
le
m
o
d
el
in
en
h
an
cin
g
p
e
r
f
o
r
m
an
ce
ac
r
o
s
s
s
en
tim
en
t
an
aly
s
is
an
d
lin
g
u
is
tic
ac
ce
p
tab
ilit
y
in
n
atu
r
al
lan
g
u
ag
e
u
n
d
er
s
tan
d
i
n
g
task
s
,
with
p
ar
ticu
lar
ly
im
p
r
ess
iv
e
r
esu
lts
in
s
en
tim
en
t
a
n
aly
s
is
as
ev
id
en
ce
d
b
y
t
h
e
p
er
f
o
r
m
an
ce
m
etr
ics.
T
h
e
f
in
d
in
g
s
h
ig
h
lig
h
t
th
e
im
p
o
r
tan
c
e
o
f
task
-
s
p
ec
if
ic
o
p
tim
izatio
n
s
an
d
en
s
em
b
lin
g
tech
n
iq
u
es in
b
o
o
s
tin
g
th
e
p
er
f
o
r
m
an
ce
o
f
th
e
m
o
d
el.
(
a)
(
b
)
Fig
u
r
e
3
.
Me
tr
ics o
v
e
r
s
tep
s
: (
a)
C
o
L
A
d
ataset
-
F1
-
s
co
r
e
an
d
a
cc
u
r
ac
y
s
co
r
es
a
n
d
(
b
)
SS
T
-
2
d
ataset
-
F1
an
d
a
cc
u
r
ac
y
s
co
r
es
3
.
4
.
M
o
del
e
f
f
iciency
W
e
ass
e
s
s
m
o
d
el
ef
f
icien
cy
u
s
in
g
two
p
r
im
ar
y
m
etr
ics:
tr
ain
in
g
lo
s
s
an
d
v
alid
atio
n
lo
s
s
.
T
h
e
tr
ain
in
g
lo
s
s
d
ec
r
ea
s
es
s
tead
il
y
ac
r
o
s
s
ep
o
ch
s
,
wh
ich
in
d
icat
e
ef
f
ec
tiv
e
lear
n
in
g
a
n
d
ad
a
p
tatio
n
to
th
e
tr
ain
in
g
d
ata'
s
co
m
p
lex
ities
.
I
n
th
is
s
t
u
d
y
,
b
o
th
r
u
n
s
(
R
u
n
1
a
n
d
R
u
n
2
)
r
ep
r
esen
ted
in
Fig
u
r
e
4
s
h
o
ws
a
s
ig
n
if
ican
t
d
ec
r
ea
s
e
in
tr
ain
in
g
lo
s
s
o
v
er
ep
o
ch
s
,
d
e
m
o
n
s
tr
atin
g
e
f
f
ec
t
iv
e
lear
n
in
g
.
R
u
n
1
s
tar
ts
with
a
s
lig
h
tly
lo
wer
in
itial
tr
ain
in
g
lo
s
s
(
0
.
4
7
9
4
)
c
o
m
p
ar
ed
t
o
R
u
n
2
(
0
.
4
8
7
1
)
,
w
ith
b
o
th
co
n
v
er
g
i
n
g
to
s
im
ilar
v
alu
es
b
y
th
e
th
ir
d
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J E
lec
&
C
o
m
p
E
n
g
I
SS
N:
2088
-
8
7
0
8
E
n
s
emb
le
o
f wi
n
n
in
g
ticke
ts
:
p
r
u
n
in
g
b
id
ir
ec
tio
n
a
l e
n
co
d
er fr
o
m
th
e
tr
a
n
s
fo
r
mers
…
(
N
ya
la
la
n
i S
ma
r
ts
)
2077
ep
o
ch
(
0
.
1
6
7
9
f
o
r
R
u
n
1
an
d
0
.
1
7
8
3
f
o
r
R
u
n
2
)
.
T
h
e
co
n
s
is
ten
t
r
ed
u
ctio
n
o
f
tr
ain
in
g
lo
s
s
s
h
o
ws
th
at
th
e
m
o
d
el
is
ef
f
ec
tiv
ely
f
itti
n
g
th
e
tr
ain
in
g
d
ata.
W
h
en
co
m
p
ar
in
g
v
alid
atio
n
l
o
s
s
es,
R
u
n
1
s
h
o
ws
a
s
tead
y
d
ec
r
ea
s
e
f
r
o
m
0
.
4
3
9
9
6
1
to
0
.
2
2
9
7
2
2
,
s
u
g
g
esti
n
g
g
o
o
d
g
en
e
r
aliza
tio
n
.
I
n
co
n
tr
ast,
R
u
n
2
s
h
o
ws
an
in
itial
d
ec
r
ea
s
e
f
r
o
m
0
.
4
6
2
1
9
8
to
0
.
3
0
1
4
5
9
,
f
o
llo
wed
b
y
a
s
lig
h
t
in
cr
ea
s
e
to
0
.
2
9
9
9
9
7
.
T
h
is
d
if
f
er
e
n
ce
s
h
in
ts
th
at
th
e
m
o
d
el
co
u
ld
b
e
p
o
te
n
tially
o
v
er
f
itti
n
g
o
r
th
er
e’
s
v
a
r
iab
il
ity
in
m
o
d
el
p
er
f
o
r
m
a
n
ce
.
T
h
e
lo
wer
v
alid
atio
n
lo
s
s
in
R
u
n
1
s
u
g
g
ests
it
g
en
er
alize
s
s
lig
h
tly
b
etter
,
s
h
o
win
g
th
e
im
p
o
r
tan
ce
o
f
s
tab
le
v
alid
atio
n
p
er
f
o
r
m
a
n
ce
f
o
r
m
o
d
el
r
o
b
u
s
tn
ess
.
T
h
e
s
tu
d
y
co
n
f
i
r
m
s
th
at
th
e
m
o
d
el
lear
n
s
a
n
d
g
en
er
alize
s
we
ll,
with
d
ec
r
ea
s
in
g
t
r
ain
in
g
an
d
v
alid
atio
n
lo
s
s
es.
Mo
n
ito
r
in
g
th
ese
m
etr
ics
o
f
f
er
s
in
s
ig
h
ts
in
to
th
e
m
o
d
el'
s
g
en
er
aliza
tio
n
ab
ilit
y
t
o
u
n
s
ee
n
d
ata
an
d
g
u
id
es
ad
ju
s
tm
en
ts
s
u
ch
as
r
eg
u
lar
izatio
n
tech
n
iq
u
es
o
r
m
o
d
if
i
ca
tio
n
s
to
th
e
m
o
d
el
a
r
ch
it
ec
tu
r
e
to
im
p
r
o
v
e
p
er
f
o
r
m
an
ce
an
d
p
r
ev
en
t
o
v
er
f
itti
n
g
.
Fig
u
r
e
4
.
T
r
ain
in
g
a
n
d
v
alid
atio
n
l
o
s
s
v
s
e
p
o
ch
s
Fig
u
r
e
5
s
h
o
ws
th
e
d
ev
elo
p
m
en
t
o
f
en
s
em
b
le
iter
atio
n
s
an
d
ev
alu
atio
n
m
etr
ics
ac
r
o
s
s
ep
o
ch
s
,
p
r
o
v
id
i
n
g
a
v
is
u
al
r
e
p
r
esen
tat
io
n
o
f
:
(
a)
th
e
n
u
m
b
e
r
o
f
e
n
s
em
b
le
iter
atio
n
s
ex
ec
u
ted
p
er
ep
o
ch
an
d
(
b
)
th
e
co
r
r
esp
o
n
d
in
g
e
v
alu
atio
n
m
e
tr
ics
ac
h
iev
ed
at
ea
ch
ep
o
c
h
.
Fig
u
r
e
5
(
a)
s
h
o
ws
th
e
f
lu
c
tu
atin
g
n
u
m
b
er
o
f
en
s
em
b
le
iter
atio
n
s
p
er
e
p
o
c
h
,
d
ep
icted
b
y
b
lu
e
b
a
r
s
.
Fig
u
r
e
5
(
b
)
d
is
p
lay
s
th
e
u
p
war
d
tr
en
d
in
e
v
alu
atio
n
m
etr
ics
ac
r
o
s
s
ep
o
c
h
s
,
r
ep
r
e
s
en
ted
b
y
g
r
e
en
b
ar
s
.
F
ig
u
r
es
5
(
a)
an
d
5
(
b
)
s
h
o
ws
in
s
ig
h
tf
u
l
o
b
s
er
v
atio
n
s
r
eg
ar
d
in
g
th
e
e
f
f
icien
cy
a
n
d
p
er
f
o
r
m
an
ce
o
f
o
u
r
m
o
d
el
u
tili
zin
g
en
s
em
b
le
p
r
u
n
i
n
g
tech
n
iq
u
es
ac
r
o
s
s
1
3
ep
o
ch
s
.
R
em
ar
k
ab
l
y
,
tr
ain
in
g
an
d
f
in
e
-
tu
n
in
g
th
e
m
o
d
e
l
p
er
3
ep
o
ch
s
to
o
k
ab
o
u
t
4
5
m
in
u
tes
to
1
h
o
u
r
,
alth
o
u
g
h
m
o
r
e
t
h
an
5
h
o
u
r
s
was n
ee
d
ed
to
f
in
e
-
tu
n
e
u
s
in
g
1
3
ep
o
c
h
s
.
A
co
m
p
ar
is
o
n
ca
n
b
e
d
r
awn
b
etwe
en
th
e
s
tu
d
y
o
f
[
2
4
]
u
tili
zin
g
u
p
to
2
0
ep
o
ch
s
to
f
in
e
-
tu
n
e
th
eir
m
o
d
el,
wh
er
ea
s
in
o
u
r
s
tu
d
y
,
we
u
s
ed
o
n
ly
3
a
n
d
1
3
ep
o
ch
s
an
d
s
till
o
u
r
en
s
em
b
le
m
o
d
el
ac
h
iev
ed
o
u
ts
tan
d
in
g
p
er
f
o
r
m
an
ce
.
T
h
e
r
ef
o
r
e,
t
h
e
in
cr
ea
s
e
in
e
n
s
em
b
le
iter
atio
n
s
o
v
er
e
p
o
ch
s
(
r
a
n
g
in
g
f
r
o
m
2
to
9
)
alig
n
s
with
f
in
d
in
g
s
b
y
Z
h
en
g
et
a
l.
[
2
4
]
,
wh
o
n
o
ted
t
h
at
en
s
em
b
le
m
eth
o
d
s
ca
n
im
p
r
o
v
e
th
e
p
er
f
o
r
m
an
ce
o
f
th
e
m
o
d
el
b
y
lev
er
a
g
in
g
m
u
ltip
le
lear
n
in
g
alg
o
r
it
h
m
s
to
y
ield
im
p
r
o
v
e
d
p
r
ed
ictio
n
ac
cu
r
ac
y
.
Fu
r
th
er
m
o
r
e,
in
th
is
s
tu
d
y
,
we
f
o
cu
s
ed
o
n
o
p
tim
izin
g
th
e
p
r
u
n
in
g
p
r
o
ce
s
s
with
in
an
e
n
s
em
b
le
m
o
d
el
b
y
tar
g
etin
g
atten
tio
n
h
ea
d
s
,
em
p
lo
y
i
n
g
a
f
ix
e
d
lear
n
in
g
r
ate
o
f
2
e
-
5
,
an
d
s
ettin
g
th
e
p
r
u
n
in
g
th
r
esh
o
ld
to
0
.
0
0
1
.
Ou
r
ap
p
r
o
ac
h
allo
wed
f
o
r
th
e
s
y
s
tem
atic
r
ed
u
ctio
n
o
f
th
e
m
o
d
el'
s
co
m
p
lex
ity
,
p
ar
ticu
lar
ly
b
y
p
r
u
n
i
n
g
en
tire
b
lo
ck
s
o
r
s
tr
u
ctu
r
es o
f
atten
tio
n
h
ea
d
s
,
r
esu
ltin
g
in
a
s
tr
ea
m
lin
ed
m
o
d
e
l a
r
ch
itectu
r
e.
C
o
m
p
ar
ativ
ely
,
p
r
ev
io
u
s
s
tu
d
ies
lik
e
th
at
o
f
[
2
4
]
–
[
2
7
]
u
s
ed
iter
ativ
e
m
ag
n
it
u
d
e
p
r
u
n
i
n
g
,
wh
ic
h
f
o
cu
s
es
o
n
in
d
iv
id
u
al
weig
h
ts
.
T
h
is
tech
n
iq
u
e
allo
ws
f
o
r
m
o
r
e
f
i
n
e
g
r
ain
p
r
u
n
in
g
b
u
t
ca
n
en
d
u
p
with
a
less
r
ed
u
ctio
n
in
m
o
d
el
s
ize
as
co
m
p
ar
ed
to
o
u
r
a
p
p
r
o
ac
h
.
W
h
ile
o
u
r
m
eth
o
d
is
f
o
cu
s
e
d
in
p
r
u
n
in
g
lar
g
e
r
s
tr
u
ctu
r
es
with
in
th
e
m
o
d
el,
lik
ely
lead
in
g
to
m
o
r
e
r
e
d
u
cti
o
n
s
in
m
o
d
el
co
m
p
le
x
ity
,
it
m
ay
also
in
tr
o
d
u
ce
m
o
r
e
v
ar
iab
ilit
y
in
m
o
d
el
p
e
r
f
o
r
m
a
n
ce
.
Pre
v
i
o
u
s
s
tu
d
ies
f
o
cu
s
ed
o
n
u
s
in
g
weig
h
t
p
r
u
n
in
g
with
a
s
m
aller
th
r
esh
o
ld
wh
ich
p
r
o
v
id
es
f
in
e
r
co
n
tr
o
l
o
v
er
m
o
d
el
p
r
u
n
in
g
an
d
m
ig
h
t
b
etter
p
r
eser
v
e
m
o
d
el
ac
cu
r
ac
y
.
T
h
e
d
if
f
er
en
t
f
o
cu
s
a
r
ea
s
o
f
p
r
u
n
i
n
g
wh
ich
ar
e
weig
h
ts
v
s
.
atte
n
tio
n
h
ea
d
s
s
h
o
w
th
e
tr
ad
e
-
o
f
f
s
b
etwe
en
p
r
ec
is
io
n
an
d
s
im
p
lify
in
g
th
e
s
tr
u
ctu
r
al
o
f
th
e
m
o
d
el.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
0
8
8
-
8
7
0
8
I
n
t J E
lec
&
C
o
m
p
E
n
g
,
Vo
l.
15
,
No
.
2
,
Ap
r
il
20
25
:
2
0
7
0
-
2
0
8
0
2078
T
h
e
p
r
im
ar
y
p
u
r
p
o
s
e
o
f
th
is
s
tu
d
y
was
to
ex
p
lo
r
e
th
e
ef
f
ec
t
iv
en
ess
o
f
atten
tio
n
h
ea
d
p
r
u
n
in
g
with
in
en
s
em
b
le
m
o
d
els
an
d
its
im
p
ac
t
o
n
m
o
d
el
e
f
f
icien
cy
.
O
u
r
f
i
n
d
in
g
s
d
r
aw
atten
tio
n
to
th
e
p
o
ten
tial
f
o
r
a
s
ig
n
if
ican
t
m
o
d
el
r
ed
u
ctio
n
t
h
r
o
u
g
h
atten
tio
n
h
ea
d
p
r
u
n
i
n
g
wh
ile
m
ain
tain
in
g
p
er
f
o
r
m
an
ce
,
p
ar
ticu
lar
l
y
with
in
th
e
co
n
tex
t
o
f
an
en
s
em
b
le
.
T
h
is
s
tu
d
y
h
ig
h
lig
h
ts
th
e
im
p
o
r
tan
ce
o
f
ex
p
lo
r
in
g
alter
n
ativ
e
p
r
u
n
in
g
s
tr
ateg
ies,
s
u
ch
as
b
lo
ck
-
lev
el
p
r
u
n
in
g
,
an
d
in
v
ites
f
u
tu
r
e
r
es
ea
r
ch
to
d
elv
e
d
ee
p
er
in
to
th
e
ef
f
ec
ts
o
f
d
if
f
er
e
n
t
th
r
esh
o
ld
lev
els an
d
lear
n
in
g
r
ates o
n
m
o
d
el
p
er
f
o
r
m
an
ce
.
(
a)
(
b
)
Fig
u
r
e
5
.
E
n
s
em
b
le
iter
atio
n
s
an
d
ev
alu
atio
n
m
etr
ics:
(
a)
e
n
s
em
b
le
iter
atio
n
s
p
er
e
p
o
ch
an
d
(
b
)
ev
al
u
atio
n
m
etr
ics p
er
ep
o
ch
4.
CO
NCLU
SI
O
N
T
h
is
p
ap
er
in
tr
o
d
u
ce
s
a
n
o
v
e
l
ap
p
r
o
ac
h
to
en
h
an
ce
B
E
R
T
m
o
d
el
ef
f
icien
cy
b
y
p
r
u
n
in
g
atten
tio
n
h
ea
d
s
u
s
in
g
an
e
n
s
em
b
le
o
f
wi
n
n
in
g
tick
ets
.
E
x
p
er
im
en
tal
r
e
s
u
lts
d
em
o
n
s
tr
ate
s
ig
n
if
ican
t r
ed
u
ctio
n
s
in
m
o
d
el
co
m
p
lex
ity
wh
ile
m
ain
tain
in
g
o
r
im
p
r
o
v
in
g
p
er
f
o
r
m
a
n
ce
ac
r
o
s
s
v
ar
io
u
s
NL
P
ta
s
k
s
.
Ou
r
ex
p
er
im
en
ts
o
n
C
o
L
A
an
d
SS
T
-
2
d
atasets
s
h
o
w
th
e
p
r
u
n
ed
e
n
s
em
b
le
m
o
d
el
r
e
d
u
ce
s
p
ar
a
m
eter
s
f
r
o
m
1
1
0
m
illi
o
n
to
7
0
m
illi
o
n
,
ac
h
ie
v
in
g
s
u
p
er
io
r
ac
cu
r
ac
y
a
n
d
F1
-
s
co
r
es
co
m
p
ar
ed
to
th
e
o
r
ig
in
al
B
E
R
T
.
Sp
ec
if
ically
,
o
u
r
en
s
em
b
le
m
o
d
el
ac
h
iev
ed
9
5
%
ac
cu
r
ac
y
an
d
9
4
%
F1
-
s
co
r
e
o
n
SS
T
-
2
,
an
d
9
6
%
f
o
r
b
o
t
h
m
etr
ics
o
n
C
o
L
A.
T
h
is
u
n
d
er
s
co
r
es
th
e
ef
f
icac
y
o
f
p
r
u
n
in
g
i
n
f
o
r
m
ativ
e
at
ten
tio
n
h
ea
d
s
to
o
p
tim
ize
m
o
d
el
p
e
r
f
o
r
m
an
ce
ef
f
icien
tly
.
Ou
r
an
aly
s
is
als
o
em
p
h
asizes
s
elec
tin
g
ap
p
r
o
p
r
iate
ev
alu
atio
n
m
etr
ics
tailo
r
ed
to
d
ataset
ch
ar
ac
ter
is
tics
;
F1
-
s
co
r
e
f
o
r
C
o
L
A's
im
b
alan
ce
d
d
ata
an
d
ac
cu
r
ac
y
f
o
r
s
en
tim
en
t
an
al
y
s
is
in
SS
T
-
2
.
I
n
co
n
clu
s
io
n
,
th
e
e
n
s
em
b
le
o
f
win
n
in
g
tick
ets
o
f
f
e
r
s
a
p
r
o
m
is
in
g
av
en
u
e
to
o
p
tim
ize
l
ar
g
e
-
s
ca
le
lan
g
u
a
g
e
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J E
lec
&
C
o
m
p
E
n
g
I
SS
N:
2088
-
8
7
0
8
E
n
s
emb
le
o
f wi
n
n
in
g
ticke
ts
:
p
r
u
n
in
g
b
id
ir
ec
tio
n
a
l e
n
co
d
er fr
o
m
th
e
tr
a
n
s
fo
r
mers
…
(
N
ya
la
la
n
i S
ma
r
ts
)
2079
m
o
d
els,
b
alan
cin
g
e
f
f
icien
cy
an
d
p
er
f
o
r
m
an
ce
,
with
p
o
te
n
tial
ap
p
licatio
n
s
ac
r
o
s
s
d
iv
er
s
e
NL
P
task
s
an
d
b
ey
o
n
d
.
Fu
tu
r
e
r
esear
ch
co
u
l
d
ex
p
lo
r
e
ex
ten
d
i
n
g
th
ese
tec
h
n
iq
u
es
to
o
t
h
er
NL
P
task
s
o
r
in
v
esti
g
atin
g
th
e
ef
f
ec
ts
o
f
d
if
f
er
en
t
p
r
u
n
i
n
g
s
tr
ateg
ies
o
n
en
s
em
b
le
m
o
d
e
ls
.
Fu
r
th
er
in
v
esti
g
atio
n
in
to
th
e
v
ar
iab
ilit
y
in
p
er
f
o
r
m
an
ce
ac
r
o
s
s
d
if
f
er
en
t
p
r
u
n
in
g
tech
n
iq
u
es
co
u
ld
o
f
f
er
v
alu
a
b
le
in
s
ig
h
ts
f
o
r
o
p
tim
izin
g
m
o
d
e
l
ef
f
icien
cy
an
d
r
o
b
u
s
tn
ess
in
v
ar
io
u
s
d
ee
p
lear
n
in
g
a
p
p
licati
o
n
s
.
ACK
NO
WL
E
DG
E
M
E
NT
S
T
h
e
au
th
o
r
s
ar
e
g
r
atef
u
l
to
B
o
ts
wan
a
I
n
ter
n
atio
n
al
Un
iv
er
s
ity
o
f
Scien
ce
an
d
T
ec
h
n
o
lo
g
y
f
o
r
its
s
u
p
p
o
r
t.
RE
F
E
R
E
NC
E
S
[
1
]
J.
D
e
v
l
i
n
,
M
.
-
W
.
C
h
a
n
g
,
K
.
L
e
e
,
a
n
d
K
.
T
o
u
t
a
n
o
v
a
,
“
B
ER
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
m
e
r
s
f
o
r
l
a
n
g
u
a
g
e
u
n
d
e
r
s
t
a
n
d
i
n
g
,
”
i
n
P
ro
c
e
e
d
i
n
g
s
o
f
t
h
e
2
0
1
9
C
o
n
f
e
re
n
c
e
o
f
t
h
e
N
o
rt
h
,
S
t
r
o
u
d
s
b
u
r
g
:
A
ss
o
c
i
a
t
i
o
n
f
o
r
C
o
m
p
u
t
a
t
i
o
n
a
l
Li
n
g
u
i
st
i
c
s,
2
0
1
9
,
p
p
.
4
1
7
1
–
4
1
8
6
,
d
o
i
:
1
0
.
1
8
6
5
3
/
v
1
/
N
1
9
-
1
4
2
3
.
[
2
]
A
.
Ta
n
g
,
P
.
Q
u
a
n
,
L
.
N
i
u
,
a
n
d
Y
.
S
h
i
,
“
A
su
r
v
e
y
f
o
r
s
p
a
r
s
e
r
e
g
u
l
a
r
i
z
a
t
i
o
n
b
a
s
e
d
c
o
m
p
r
e
ssi
o
n
m
e
t
h
o
d
s
,
”
A
n
n
a
l
s
o
f
D
a
t
a
S
c
i
e
n
c
e
,
v
o
l
.
9
,
n
o
.
4
,
p
p
.
6
9
5
–
7
2
2
,
A
u
g
.
2
0
2
2
,
d
o
i
:
1
0
.
1
0
0
7
/
s4
0
7
4
5
-
0
2
2
-
0
0
3
8
9
-
6.
[
3
]
C
.
Y
u
,
T
.
C
h
e
n
,
a
n
d
Z.
G
a
n
,
“
B
o
o
st
t
r
a
n
sf
o
r
mer
-
b
a
se
d
l
a
n
g
u
a
g
e
mo
d
e
l
s
w
i
t
h
G
P
U
-
f
r
i
e
n
d
l
y
s
p
a
r
si
t
y
a
n
d
q
u
a
n
t
i
z
a
t
i
o
n
,
”
i
n
Fi
n
d
i
n
g
s
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:
AC
L
2
0
2
3
,
S
t
r
o
u
d
s
b
u
r
g
,
P
A
,
U
S
A
:
A
sso
c
i
a
t
i
o
n
f
o
r
C
o
m
p
u
t
a
t
i
o
n
a
l
Li
n
g
u
i
s
t
i
c
s,
2
0
2
3
,
p
p
.
2
1
8
–
2
3
5
,
d
o
i
:
1
0
.
1
8
6
5
3
/
v
1
/
2
0
2
3
.
f
i
n
d
i
n
g
s
-
a
c
l
.
1
5
.
[
4
]
C.
-
C
.
J
.
K
u
o
a
n
d
A
.
M
.
M
a
d
n
i
,
“
G
r
e
e
n
l
e
a
r
n
i
n
g
:
i
n
t
r
o
d
u
c
t
i
o
n
,
e
x
a
m
p
l
e
s
a
n
d
o
u
t
l
o
o
k
,
”
J
o
u
rn
a
l
o
f
Vi
s
u
a
l
C
o
m
m
u
n
i
c
a
t
i
o
n
a
n
d
I
m
a
g
e
Re
p
r
e
se
n
t
a
t
i
o
n
,
v
o
l
.
9
0
,
p
p
.
1
0
3
6
8
5
,
F
e
b
.
2
0
2
3
,
d
o
i
:
1
0
.
1
0
1
6
/
j
.
j
v
c
i
r
.
2
0
2
2
.
1
0
3
6
8
5
.
[
5
]
R
.
V
e
r
d
e
c
c
h
i
a
,
J.
S
a
l
l
o
u
,
a
n
d
L.
C
r
u
z
,
“
A
sy
s
t
e
m
a
t
i
c
r
e
v
i
e
w
o
f
g
r
e
e
n
A
I
,
”
WIREs
D
a
t
a
Mi
n
i
n
g
a
n
d
K
n
o
w
l
e
d
g
e
D
i
sc
o
v
e
ry
,
v
o
l
.
1
3
,
n
o
.
4
,
Ju
l
.
2
0
2
3
,
d
o
i
:
1
0
.
1
0
0
2
/
w
i
d
m.
1
5
0
7
.
[
6
]
J.
D
.
N
u
n
e
s,
M
.
C
a
r
v
a
l
h
o
,
D
.
C
a
r
n
e
i
r
o
,
a
n
d
J.
S
.
C
a
r
d
o
s
o
,
“
S
p
i
k
i
n
g
n
e
u
r
a
l
n
e
t
w
o
r
k
s:
a
s
u
r
v
e
y
,
”
I
EEE
Ac
c
e
ss
,
v
o
l
.
1
0
,
p
p
.
6
0
7
3
8
–
6
0
7
6
4
,
2
0
2
2
,
d
o
i
:
1
0
.
1
1
0
9
/
A
C
C
ESS
.
2
0
2
2
.
3
1
7
9
9
6
8
.
[
7
]
S
.
S
h
r
e
y
a
s
h
r
e
e
,
P
.
S
u
n
a
g
a
r
,
S
.
R
a
j
a
r
a
j
e
s
w
a
r
i
,
a
n
d
A
.
K
a
n
a
v
a
l
l
i
,
“
A
l
i
t
e
r
a
t
u
r
e
r
e
v
i
e
w
o
n
b
i
d
i
r
e
c
t
i
o
n
a
l
e
n
c
o
d
e
r
r
e
p
r
e
s
e
n
t
a
t
i
o
n
s
f
r
o
m
t
r
a
n
s
f
o
r
m
e
r
s
,
”
i
n
I
n
v
e
n
t
i
v
e
C
o
m
p
u
t
a
t
i
o
n
a
n
d
I
n
f
o
r
m
a
t
i
o
n
T
e
c
h
n
o
l
o
g
i
e
s
:
P
r
o
c
e
e
d
i
n
g
s
o
f
I
C
I
C
I
T
2
0
2
1
,
S
p
r
i
n
g
e
r
,
2
0
2
2
,
p
p
.
3
0
5
–
3
2
0
.
[
8
]
W
.
K
h
a
n
,
A
.
D
a
u
d
,
K
.
K
h
a
n
,
S
.
M
u
h
a
mm
a
d
,
a
n
d
R
.
H
a
q
,
“
E
x
p
l
o
r
i
n
g
t
h
e
f
r
o
n
t
i
e
r
s
o
f
d
e
e
p
l
e
a
r
n
i
n
g
a
n
d
n
a
t
u
r
a
l
l
a
n
g
u
a
g
e
p
r
o
c
e
ss
i
n
g
:
a
c
o
mp
r
e
h
e
n
s
i
v
e
o
v
e
r
v
i
e
w
o
f
k
e
y
c
h
a
l
l
e
n
g
e
s
a
n
d
e
mer
g
i
n
g
t
r
e
n
d
s,
”
N
a
t
u
ra
l
L
a
n
g
u
a
g
e
Pr
o
c
e
ssi
n
g
J
o
u
rn
a
l
,
v
o
l
.
4
,
p
p
.
1
0
0
0
2
6
,
S
e
p
.
2
0
2
3
,
d
o
i
:
1
0
.
1
0
1
6
/
j
.
n
l
p
.
2
0
2
3
.
1
0
0
0
2
6
.
[
9
]
D
.
K
h
u
r
a
n
a
,
A
.
K
o
l
i
,
K
.
K
h
a
t
t
e
r
,
a
n
d
S
.
S
i
n
g
h
,
“
N
a
t
u
r
a
l
l
a
n
g
u
a
g
e
p
r
o
c
e
ssi
n
g
:
s
t
a
t
e
o
f
t
h
e
a
r
t
,
c
u
r
r
e
n
t
t
r
e
n
d
s
a
n
d
c
h
a
l
l
e
n
g
e
s
,
”
Mu
l
t
i
m
e
d
i
a
T
o
o
l
s
a
n
d
Ap
p
l
i
c
a
t
i
o
n
s
,
v
o
l
.
8
2
,
n
o
.
3
,
p
p
.
3
7
1
3
–
3
7
4
4
,
Ja
n
.
2
0
2
3
,
d
o
i
:
1
0
.
1
0
0
7
/
s1
1
0
4
2
-
0
2
2
-
1
3
4
2
8
-
4.
[
1
0
]
S
.
M
a
r
u
f
,
F
.
S
a
l
e
h
,
a
n
d
G
.
H
a
f
f
a
r
i
,
“
A
su
r
v
e
y
o
n
d
o
c
u
me
n
t
-
l
e
v
e
l
n
e
u
r
a
l
mac
h
i
n
e
t
r
a
n
sl
a
t
i
o
n
,
”
A
C
M
C
o
m
p
u
t
i
n
g
S
u
r
v
e
y
s
,
v
o
l
.
5
4
,
n
o
.
2
,
p
p
.
1
–
3
6
,
M
a
r
.
2
0
2
2
,
d
o
i
:
1
0
.
1
1
4
5
/
3
4
4
1
6
9
1
.
[
1
1
]
A
.
B
a
n
d
i
,
P
.
V
.
S
.
R
.
A
d
a
p
a
,
a
n
d
Y
.
E.
V
.
P
.
K
.
K
u
c
h
i
,
“
T
h
e
p
o
w
e
r
o
f
g
e
n
e
r
a
t
i
v
e
A
I
:
a
r
e
v
i
e
w
o
f
r
e
q
u
i
r
e
me
n
t
s
,
mo
d
e
l
s,
i
n
p
u
t
–
o
u
t
p
u
t
f
o
r
m
a
t
s
,
e
v
a
l
u
a
t
i
o
n
me
t
r
i
c
s
,
a
n
d
c
h
a
l
l
e
n
g
e
s,”
F
u
t
u
re
I
n
t
e
r
n
e
t
,
v
o
l
.
1
5
,
n
o
.
8
,
p
p
.
2
6
0
,
Ju
l
.
2
0
2
3
,
d
o
i
:
1
0
.
3
3
9
0
/
f
i
1
5
0
8
0
2
6
0
.
[
1
2
]
J.
Z
a
k
r
a
o
u
i
,
M
.
S
a
l
e
h
,
S
.
A
l
-
M
a
a
d
e
e
d
,
a
n
d
J.
M
.
A
l
j
a
’
a
m,
“
A
r
a
b
i
c
ma
c
h
i
n
e
t
r
a
n
s
l
a
t
i
o
n
:
a
s
u
r
v
e
y
w
i
t
h
c
h
a
l
l
e
n
g
e
s
a
n
d
f
u
t
u
r
e
d
i
r
e
c
t
i
o
n
s,
”
I
E
EE
A
c
c
e
ss
,
v
o
l
.
9
,
p
p
.
1
6
1
4
4
5
–
1
6
1
4
6
8
,
2
0
2
1
,
d
o
i
:
1
0
.
1
1
0
9
/
A
C
C
ESS
.
2
0
2
1
.
3
1
3
2
4
8
8
.
[
1
3
]
M
.
S
.
S
a
y
e
e
d
,
V
.
M
o
h
a
n
,
a
n
d
K
.
S
.
M
u
t
h
u
,
“
B
ER
T:
a
r
e
v
i
e
w
o
f
a
p
p
l
i
c
a
t
i
o
n
s
i
n
se
n
t
i
me
n
t
a
n
a
l
y
s
i
s,
”
H
i
g
h
T
e
c
h
a
n
d
I
n
n
o
v
a
t
i
o
n
J
o
u
rn
a
l
,
v
o
l
.
4
,
n
o
.
2
,
p
p
.
4
5
3
–
4
6
2
,
Ju
n
.
2
0
2
3
,
d
o
i
:
1
0
.
2
8
9
9
1
/
H
I
J
-
2
0
2
3
-
04
-
02
-
0
1
5
.
[
1
4
]
K
.
A
.
A
l
s
h
a
i
k
h
,
O
.
A
.
A
l
m
a
t
r
a
f
i
,
a
n
d
Y
.
B
.
A
b
u
s
h
a
r
k
,
“
B
E
R
T
-
b
a
s
e
d
m
o
d
e
l
f
o
r
a
s
p
e
c
t
-
b
a
s
e
d
s
e
n
t
i
m
e
n
t
a
n
a
l
y
s
i
s
f
o
r
a
n
a
l
y
z
i
n
g
A
r
a
b
i
c
o
p
e
n
-
e
n
d
e
d
s
u
r
v
e
y
r
e
s
p
o
n
s
e
s
:
a
c
a
s
e
s
t
u
d
y
,
”
I
E
E
E
A
c
c
e
s
s
,
v
o
l
.
1
2
,
p
p
.
2
2
8
8
–
2
3
0
2
,
2
0
2
4
,
d
o
i
:
1
0
.
1
1
0
9
/
A
C
C
E
S
S
.
2
0
2
3
.
3
3
4
8
3
4
2
.
[
1
5
]
T.
L
i
a
n
g
,
J.
G
l
o
ssn
e
r
,
L.
W
a
n
g
,
S
.
S
h
i
,
a
n
d
X
.
Zh
a
n
g
,
“
P
r
u
n
i
n
g
a
n
d
q
u
a
n
t
i
z
a
t
i
o
n
f
o
r
d
e
e
p
n
e
u
r
a
l
n
e
t
w
o
r
k
a
c
c
e
l
e
r
a
t
i
o
n
:
a
su
r
v
e
y
,
”
N
e
u
ro
c
o
m
p
u
t
i
n
g
,
v
o
l
.
4
6
1
,
p
p
.
3
7
0
–
4
0
3
,
O
c
t
.
2
0
2
1
,
d
o
i
:
1
0
.
1
0
1
6
/
j
.
n
e
u
c
o
m.
2
0
2
1
.
0
7
.
0
4
5
.
[
1
6
]
B
.
L
i
,
Y
.
M
i
a
o
,
Y
.
W
a
n
g
,
Y
.
S
u
n
,
a
n
d
W
.
W
a
n
g
,
“
I
mp
r
o
v
i
n
g
t
h
e
e
f
f
i
c
i
e
n
c
y
a
n
d
e
f
f
e
c
t
i
v
e
n
e
ss
f
o
r
B
E
R
T
-
b
a
se
d
e
n
t
i
t
y
r
e
s
o
l
u
t
i
o
n
,
”
i
n
Pro
c
e
e
d
i
n
g
s
o
f
t
h
e
AA
AI
C
o
n
f
e
re
n
c
e
o
n
Ar
t
i
f
i
c
i
a
l
I
n
t
e
l
l
i
g
e
n
c
e
,
2
0
2
1
,
p
p
.
1
3
2
2
6
–
1
3
2
3
3
,
d
o
i
:
1
0
.
1
6
0
9
/
a
a
a
i
.
v
3
5
i
1
5
.
1
7
5
6
2
.
[
1
7
]
V
.
K
.
V
e
r
ma,
N
.
M
e
h
t
a
,
S
.
S
i
,
R
.
H
e
n
a
o
,
a
n
d
L
.
C
a
r
i
n
,
“
P
u
s
h
i
n
g
t
h
e
e
f
f
i
c
i
e
n
c
y
l
i
mi
t
u
si
n
g
s
t
r
u
c
t
u
r
e
d
s
p
a
r
se
c
o
n
v
o
l
u
t
i
o
n
s,”
i
n
2
0
2
3
I
E
EE/
C
VF
W
i
n
t
e
r
C
o
n
f
e
re
n
c
e
o
n
A
p
p
l
i
c
a
t
i
o
n
s
o
f
C
o
m
p
u
t
e
r
V
i
si
o
n
(
WA
C
V)
,
I
EEE,
Ja
n
.
2
0
2
3
,
p
p
.
6
4
9
2
–
6
5
0
2
,
d
o
i
:
1
0
.
1
1
0
9
/
W
A
C
V
5
6
6
8
8
.
2
0
2
3
.
0
0
6
4
4
.
[
1
8
]
Y
.
Y
a
n
g
,
C
.
Z
h
a
n
g
,
B
.
W
a
n
g
,
a
n
d
D
.
S
o
n
g
,
“
D
o
g
e
t
i
c
k
e
t
s
:
u
n
c
o
v
e
r
i
n
g
d
o
ma
i
n
-
g
e
n
e
r
a
l
l
a
n
g
u
a
g
e
m
o
d
e
l
s
b
y
p
l
a
y
i
n
g
l
o
t
t
e
r
y
t
i
c
k
e
t
s,
”
i
n
C
C
F
I
n
t
e
rn
a
t
i
o
n
a
l
C
o
n
f
e
re
n
c
e
o
n
N
a
t
u
r
a
l
L
a
n
g
u
a
g
e
Pr
o
c
e
ssi
n
g
a
n
d
C
h
i
n
e
se
C
o
m
p
u
t
i
n
g
,
2
0
2
2
,
p
p
.
1
4
4
–
1
5
6
,
d
o
i
:
1
0
.
1
0
0
7
/
9
7
8
-
3
-
031
-
1
7
1
2
0
-
8
_
1
2
.
[
1
9
]
Z.
G
a
n
e
t
a
l
.
,
“
P
l
a
y
i
n
g
l
o
t
t
e
r
y
t
i
c
k
e
t
s
w
i
t
h
v
i
s
i
o
n
a
n
d
l
a
n
g
u
a
g
e
,
”
i
n
P
ro
c
e
e
d
i
n
g
s o
f
t
h
e
AAAI
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
,
2
0
2
2
,
p
p
.
6
5
2
–
6
6
0
,
d
o
i
:
1
0
.
1
6
0
9
/
a
a
a
i
.
v
3
6
i
1
.
1
9
9
4
5
.
[
2
0
]
M
.
W
a
n
k
h
a
d
e
,
A
.
C
.
S
.
R
a
o
,
a
n
d
C
.
K
u
l
k
a
r
n
i
,
“
A
su
r
v
e
y
o
n
se
n
t
i
me
n
t
a
n
a
l
y
si
s
me
t
h
o
d
s,
a
p
p
l
i
c
a
t
i
o
n
s,
a
n
d
c
h
a
l
l
e
n
g
e
s,”
Ar
t
i
f
i
c
i
a
l
I
n
t
e
l
l
i
g
e
n
c
e
Re
v
i
e
w
,
v
o
l
.
5
5
,
n
o
.
7
,
p
p
.
5
7
3
1
–
5
7
8
0
,
O
c
t
.
2
0
2
2
,
d
o
i
:
1
0
.
1
0
0
7
/
s1
0
4
6
2
-
0
2
2
-
1
0
1
4
4
-
1.
[
2
1
]
A
.
W
a
r
st
a
d
t
,
A
.
S
i
n
g
h
,
a
n
d
S
.
R
.
B
o
w
m
a
n
,
“
N
e
u
r
a
l
n
e
t
w
o
r
k
a
c
c
e
p
t
a
b
i
l
i
t
y
j
u
d
g
m
e
n
t
s,
”
T
r
a
n
s
a
c
t
i
o
n
s
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
s
t
i
c
s
,
v
o
l
.
7
,
p
p
.
6
2
5
–
6
4
1
,
N
o
v
.
2
0
1
9
,
d
o
i
:
1
0
.
1
1
6
2
/
t
a
c
l
_
a
_
0
0
2
9
0
.
[
2
2
]
T.
W
o
l
f
e
t
a
l
.
,
“
Tr
a
n
sf
o
r
m
e
r
s:
s
t
a
t
e
-
of
-
t
h
e
-
a
r
t
n
a
t
u
r
a
l
l
a
n
g
u
a
g
e
p
r
o
c
e
ssi
n
g
,
”
i
n
Pr
o
c
e
e
d
i
n
g
s
o
f
t
h
e
2
0
2
0
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
ssi
n
g
:
S
y
s
t
e
m
D
e
m
o
n
s
t
ra
t
i
o
n
s
,
S
t
r
o
u
d
s
b
u
r
g
,
P
A
,
U
S
A
:
A
ss
o
c
i
a
t
i
o
n
f
o
r
C
o
m
p
u
t
a
t
i
o
n
a
l
Li
n
g
u
i
s
t
i
c
s,
2
0
2
0
,
p
p
.
3
8
–
45
,
d
o
i
:
1
0
.
1
8
6
5
3
/
v
1
/
2
0
2
0
.
e
m
n
l
p
-
d
e
m
o
s.
6
.
[
2
3
]
A
.
M
o
h
a
m
m
e
d
a
n
d
R
.
K
o
r
a
,
“
A
c
o
m
p
r
e
h
e
n
s
i
v
e
r
e
v
i
e
w
o
n
e
n
s
e
m
b
l
e
d
e
e
p
l
e
a
r
n
i
n
g
:
O
p
p
o
r
t
u
n
i
t
i
e
s
a
n
d
c
h
a
l
l
e
n
g
e
s
,
”
J
o
u
r
n
a
l
o
f
K
i
n
g
S
a
u
d
U
n
i
v
e
r
s
i
t
y
-
C
o
m
p
u
t
e
r
a
n
d
I
n
f
o
r
m
a
t
i
o
n
S
c
i
e
n
c
e
s
,
v
o
l
.
3
5
,
n
o
.
2
,
p
p
.
7
5
7
–
7
7
4
,
F
e
b
.
2
0
2
3
,
d
o
i
:
1
0
.
1
0
1
6
/
j
.
j
k
s
u
c
i
.
2
0
2
3
.
0
1
.
0
1
4
.
[
2
4
]
R
.
Zh
e
n
g
e
t
a
l
.
,
“
R
o
b
u
s
t
l
o
t
t
e
r
y
t
i
c
k
e
t
s
f
o
r
p
r
e
-
t
r
a
i
n
e
d
l
a
n
g
u
a
g
e
m
o
d
e
l
s,
”
i
n
Pr
o
c
e
e
d
i
n
g
s
o
f
t
h
e
6
0
t
h
A
n
n
u
a
l
M
e
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
(
V
o
l
u
m
e
1
:
L
o
n
g
Pa
p
e
rs)
,
A
sso
c
i
a
t
i
o
n
f
o
r
C
o
mp
u
t
a
t
i
o
n
a
l
L
i
n
g
u
i
st
i
c
s,
2
0
2
2
,
p
p
.
2
2
1
1
–
2
2
2
4
,
d
o
i
:
1
0
.
1
8
6
5
3
/
v
1
/
2
0
2
2
.
a
c
l
-
l
o
n
g
.
1
5
7
.
[
2
5
]
X
.
C
h
e
n
,
Y
.
C
h
e
n
g
,
S
.
W
a
n
g
,
Z.
G
a
n
,
Z.
W
a
n
g
,
a
n
d
J
.
L
i
u
,
“
E
a
r
l
y
B
E
R
T
:
e
f
f
i
c
i
e
n
t
B
E
R
T
t
r
a
i
n
i
n
g
v
i
a
e
a
r
l
y
-
b
i
r
d
l
o
t
t
e
r
y
t
i
c
k
e
t
s
,
”
i
n
Pro
c
e
e
d
i
n
g
s
o
f
t
h
e
5
9
t
h
A
n
n
u
a
l
M
e
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
s
t
i
c
s
a
n
d
t
h
e
1
1
t
h
I
n
t
e
rn
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
r
a
l
L
a
n
g
u
a
g
e
Pr
o
c
e
ssi
n
g
(
V
o
l
u
m
e
1
:
L
o
n
g
P
a
p
e
rs)
,
S
t
r
o
u
d
s
b
u
r
g
,
P
A
,
U
S
A
:
A
sso
c
i
a
t
i
o
n
f
o
r
C
o
m
p
u
t
a
t
i
o
n
a
l
Li
n
g
u
i
s
t
i
c
s,
2
0
2
1
,
p
p
.
2
1
9
5
–
2
2
0
7
,
d
o
i
:
1
0
.
1
8
6
5
3
/
v
1
/
2
0
2
1
.
a
c
l
-
l
o
n
g
.
1
7
1
.
Evaluation Warning : The document was created with Spire.PDF for Python.