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
9
,
No
.
2
,
A
u
g
u
s
t
2
0
2
5
,
p
p
.
1
249
~
1
2
6
0
I
SS
N:
2
5
0
2
-
4
7
5
2
,
DOI
: 1
0
.
1
1
5
9
1
/ijeecs.v
3
9
.i
2
.
pp
1
2
4
9
-
1
2
6
0
1249
J
o
ur
na
l ho
m
ep
a
g
e
:
h
ttp
:
//ij
ee
cs.ia
esco
r
e.
co
m
Deep learni
ng
-
ba
sed multi
-
tie
r sen
sitiv
ity a
na
ly
sis
net
wo
rk
for
do
cument se
nsitiv
ity cla
ss
ificatio
n
Sa
diy
a
Ans
a
ri,
Sh
a
m
ee
m
Akt
her
D
e
p
a
r
t
me
n
t
o
f
F
a
c
u
l
t
y
o
f
E
n
g
i
n
e
e
r
i
n
g
a
n
d
Te
c
h
n
o
l
o
g
y
,
K
B
N
U
n
i
v
e
r
s
i
t
y
,
K
a
l
a
b
u
r
a
g
i
,
I
n
d
i
a
Art
icle
I
nfo
AB
S
T
RAC
T
A
r
ticle
his
to
r
y:
R
ec
eiv
ed
Au
g
21
,
2
0
2
4
R
ev
is
ed
May
16
,
2
0
2
5
Acc
ep
ted
J
u
l
2
,
2
0
2
5
In
t
h
e
d
ig
i
tal
a
g
e
,
th
e
e
x
p
o
n
e
n
ti
a
l
g
ro
wt
h
o
f
d
a
ta
n
e
c
e
ss
it
a
tes
ro
b
u
st
a
n
d
e
fficie
n
t
sy
ste
m
s
fo
r
d
o
c
u
m
e
n
t
c
las
sifica
ti
o
n
to
m
a
in
tain
d
a
ta
se
c
u
rit
y
a
n
d
c
o
m
p
li
a
n
c
e
.
Tex
t
c
las
sifica
ti
o
n
p
lay
s
a
c
ru
c
ial
r
o
le
i
n
id
e
n
t
ify
i
n
g
se
n
siti
v
e
in
fo
rm
a
ti
o
n
b
y
a
u
t
o
m
a
ti
c
a
ll
y
c
a
teg
o
rizi
n
g
d
o
c
u
m
e
n
ts
b
a
se
d
o
n
t
h
e
ir
c
o
n
ten
t.
Us
in
g
a
d
v
a
n
c
e
d
m
a
c
h
in
e
lea
rn
in
g
a
n
d
d
e
e
p
lea
rn
in
g
m
o
d
e
ls,
it
a
n
a
ly
z
e
s
tex
t
to
d
e
tec
t
k
e
y
w
o
rd
s,
p
a
tt
e
r
n
s,
a
n
d
c
o
n
te
x
tu
a
l
c
u
e
s
t
h
a
t
i
n
d
ica
te
t
h
e
p
re
se
n
c
e
o
f
se
n
siti
v
e
d
a
ta.
Th
is
p
a
p
e
r
p
r
e
se
n
ts
a
n
o
v
e
l
fra
m
e
wo
r
k
,
t
h
e
m
u
lt
i
-
ti
e
r
se
n
siti
v
it
y
a
n
a
ly
sis
n
e
two
r
k
(
M
TS
AN
),
d
e
si
g
n
e
d
to
a
c
c
u
ra
tely
c
las
sify
d
o
c
u
m
e
n
ts
in
t
o
p
u
b
li
c
,
p
ri
v
a
te,
a
n
d
c
o
n
fid
e
n
ti
a
l
c
a
teg
o
ries
.
T
h
e
p
ro
p
o
se
d
sy
ste
m
in
teg
ra
tes
se
v
e
ra
l
a
d
v
a
n
c
e
d
c
o
m
p
o
n
e
n
ts,
i
n
c
lu
d
in
g
th
e
m
u
lt
i
-
ti
e
r
se
n
siti
v
it
y
e
n
c
o
d
in
g
n
e
two
r
k
(M
TS
EN).
M
TS
AN
lev
e
ra
g
e
s
a
c
o
m
b
in
a
ti
o
n
o
f
c
o
n
v
o
l
u
ti
o
n
a
l
n
e
two
rk
s
a
n
d
g
ra
p
h
c
o
n
v
o
l
u
ti
o
n
a
l
n
e
two
r
k
s
(
G
CN
s)
to
c
a
p
tu
re
b
o
th
l
o
c
a
l
a
n
d
g
l
o
b
a
l
c
o
n
tex
t
u
a
l
in
f
o
rm
a
ti
o
n
.
T
h
e
d
u
a
l
-
sc
o
p
e
g
ra
p
h
c
o
n
v
o
lu
ti
o
n
b
lo
c
k
(DSG
CB)
is
in
tr
o
d
u
c
e
d
t
o
a
d
d
re
ss
b
o
t
h
g
lo
b
a
l
d
e
p
e
n
d
e
n
c
ies
a
n
d
lo
c
a
l
d
y
n
a
m
ics
,
e
m
p
lo
y
i
n
g
a
n
o
v
e
l
fu
si
o
n
m
e
c
h
a
n
ism
t
o
m
e
rg
e
g
lo
b
a
l
a
n
d
lo
c
a
l
fe
a
tu
re
s
e
ffe
c
ti
v
e
ly
.
Ad
d
it
i
o
n
a
ll
y
,
t
h
e
c
ro
ss
-
ti
e
r
in
fo
rm
a
ti
o
n
f
u
sio
n
b
lo
c
k
(
CTI
F
B)
fa
c
il
it
a
tes
th
e
se
a
m
les
s
in
te
g
ra
ti
o
n
o
f
m
u
lt
i
-
lev
e
l
fe
a
tu
re
s,
f
u
rth
e
r
re
fi
n
in
g
th
e
c
las
sifica
ti
o
n
p
ro
c
e
ss
.
T
h
e
re
su
lt
s
d
e
m
o
n
stra
te
t
h
a
t
t
h
e
p
ro
p
o
se
d
M
TS
AN
m
o
d
e
l
o
u
tp
e
rf
o
rm
s
trad
it
io
n
a
l
m
a
c
h
in
e
lea
rn
in
g
a
p
p
ro
a
c
h
e
s
a
n
d
c
o
n
tem
p
o
ra
ry
d
e
e
p
lea
rn
in
g
m
o
d
e
ls
su
c
h
a
s
b
id
irec
ti
o
n
a
l
e
n
c
o
d
e
r
re
p
r
e
se
n
tatio
n
s
fr
o
m
tra
n
sfo
rm
e
rs
(
BERT
)
,
a
c
h
iev
in
g
su
p
e
rio
r
a
c
c
u
ra
c
y
a
n
d
F
1
sc
o
re
s
i
n
c
las
sify
i
n
g
se
n
siti
v
e
in
fo
rm
a
ti
o
n
.
K
ey
w
o
r
d
s
:
Do
cu
m
en
t
s
en
s
itiv
ity
c
lass
if
icatio
n
Mu
lti
-
tier
s
en
s
itiv
ity
an
aly
s
is
n
etwo
r
k
E
n
cr
y
p
tio
n
Du
al
-
s
co
p
e
g
r
a
p
h
co
n
v
o
l
u
tio
n
b
lo
ck
Dee
p
lear
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
:
Sad
iy
a
An
s
ar
i
Dep
ar
tm
en
t o
f
Facu
lty
o
f
E
n
g
i
n
ee
r
in
g
a
n
d
T
ec
h
n
o
lo
g
y
,
KB
N
Un
iv
er
s
ity
Kala
b
u
r
ag
i,
I
n
d
ia
E
m
ail: sad
iy
aa
n
s
ar
i_
k
b
n
u
@
r
ed
if
f
m
ail.
co
m
1.
I
NT
RO
D
UCT
I
O
N
I
n
to
d
ay
'
s
d
ig
ital
er
a,
th
e
ex
p
o
n
en
tial
g
r
o
wth
o
f
d
ata
h
as
b
ec
o
m
e
b
o
th
a
b
o
o
n
an
d
a
c
h
a
llen
g
e
f
o
r
o
r
g
an
izatio
n
s
.
As
in
f
o
r
m
atio
n
p
r
o
life
r
ates,
s
o
d
o
es
th
e
n
ee
d
f
o
r
ef
f
ec
tiv
e
d
ata
m
an
ag
em
en
t
an
d
class
if
icatio
n
.
T
h
e
class
if
icatio
n
o
f
d
o
cu
m
e
n
ts
in
to
ca
teg
o
r
ies
s
u
ch
as
p
u
b
lic,
p
r
iv
ate,
an
d
co
n
f
id
en
ti
al
is
es
s
en
tial
f
o
r
m
ain
tain
in
g
d
ata
s
ec
u
r
ity
,
p
r
i
v
ac
y
,
an
d
c
o
m
p
lian
ce
with
le
g
al
an
d
r
eg
u
lato
r
y
f
r
am
ewo
r
k
s
.
T
h
is
n
ec
ess
ity
i
s
p
ar
ticu
lar
ly
p
r
o
n
o
u
n
ce
d
in
s
ec
to
r
s
lik
e
f
in
an
ce
,
h
ea
lth
ca
r
e,
a
n
d
tech
n
o
lo
g
y
,
wh
e
r
e
s
en
s
it
iv
e
in
f
o
r
m
atio
n
m
u
s
t
b
e
m
eticu
lo
u
s
ly
p
r
o
tecte
d
to
p
r
ev
e
n
t
u
n
au
t
h
o
r
ized
ac
c
ess
an
d
b
r
ea
ch
es.
E
n
s
u
r
in
g
th
e
s
ec
u
r
ity
an
d
co
n
f
id
en
tiality
o
f
d
ata
to
s
af
eg
u
ar
d
s
en
s
itiv
e
in
f
o
r
m
atio
n
f
r
o
m
u
n
au
th
o
r
ized
ac
ce
s
s
[
1
]
.
Fo
r
d
ata
t
o
b
e
au
th
en
tic,
it
n
ee
d
s
to
co
m
e
f
r
o
m
a
tr
u
s
two
r
th
y
s
o
u
r
ce
an
d
s
tay
u
n
ch
a
n
g
ed
.
E
n
cr
y
p
tio
n
an
d
s
ig
n
atu
r
e
s
y
s
tem
s
ar
e
ess
en
tial
f
o
r
m
ain
tain
in
g
co
n
f
id
en
tiality
a
n
d
v
er
if
y
i
n
g
au
t
h
en
ticity
.
T
h
er
e
a
r
e
t
h
r
ee
ty
p
es
o
f
d
ata
s
en
s
itiv
ity
p
u
b
lic,
p
r
iv
ate,
an
d
co
n
f
id
e
n
tial.
Pu
b
lic
d
ata
r
e
f
er
s
to
in
f
o
r
m
atio
n
th
at
is
o
p
en
ly
ac
ce
s
s
ib
le
an
d
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
5
0
2
-
4
7
5
2
I
n
d
o
n
esian
J
E
lec
E
n
g
&
C
o
m
p
Sci
,
Vo
l.
3
9
,
No
.
2
,
Au
g
u
s
t
20
25
:
1
2
4
9
-
1
260
1250
p
o
s
es
m
in
im
al
r
is
k
if
d
is
clo
s
e
d
.
I
t
in
clu
d
es
c
o
n
ten
t
s
u
ch
as p
r
ess
r
elea
s
es,
p
u
b
licly
av
aila
b
le
f
in
an
cial
r
ep
o
r
ts
,
an
d
m
ar
k
etin
g
m
ater
ials
.
T
h
e
s
e
d
o
cu
m
en
ts
ar
e
in
ten
d
e
d
f
o
r
wid
e
d
is
tr
ib
u
tio
n
an
d
d
o
n
o
t
r
eq
u
ir
e
p
r
o
tectio
n
fro
m
u
n
au
th
o
r
ized
ac
ce
s
s
.
I
n
co
n
tr
ast,
p
r
iv
ate
d
ata
in
cl
u
d
es
in
f
o
r
m
atio
n
t
h
at
is
n
o
t
m
ea
n
t
f
o
r
p
u
b
lic
d
is
clo
s
u
r
e
b
u
t
is
n
o
t
n
ec
ess
ar
ily
h
ig
h
ly
s
en
s
itiv
e
[
2
]
.
T
h
is
ca
teg
o
r
y
m
ay
en
co
m
p
ass
in
ter
n
a
l
co
m
m
u
n
icatio
n
s
,
p
er
s
o
n
al
o
p
in
io
n
s
in
cu
s
to
m
er
r
ev
iews,
an
d
in
ter
n
al
m
e
m
o
s
.
W
h
ile
p
r
iv
ate
d
ata
s
h
o
u
ld
b
e
p
r
o
tecte
d
to
m
ain
tain
p
r
iv
ac
y
,
its
u
n
a
u
th
o
r
ized
d
is
clo
s
u
r
e
ty
p
ically
p
o
s
es
less
r
is
k
th
an
co
n
f
id
en
tial
in
f
o
r
m
atio
n
.
C
o
n
f
id
en
tial
d
ata
,
o
n
t
h
e
o
th
er
h
an
d
,
in
clu
d
es
h
ig
h
ly
s
en
s
itiv
e
in
f
o
r
m
atio
n
th
at,
if
ex
p
o
s
ed
,
co
u
ld
r
esu
lt
in
s
ig
n
if
ican
t
leg
al,
f
in
an
cial,
o
r
r
ep
u
tatio
n
al
d
a
m
ag
e.
T
h
is
ca
t
eg
o
r
y
c
o
v
er
s
a
b
r
o
ad
r
a
n
g
e
o
f
d
o
cu
m
e
n
ts
,
s
u
ch
as
m
ed
ical
r
ec
o
r
d
s
p
r
o
tecte
d
u
n
d
er
laws
lik
e
HI
PAA
(
h
ea
lth
in
s
u
r
an
ce
p
o
r
tab
ilit
y
a
n
d
ac
co
u
n
tab
ilit
y
ac
t
)
,
f
in
an
cial
d
ata,
p
r
o
p
r
ietar
y
b
u
s
i
n
ess
in
f
o
r
m
atio
n
,
an
d
in
t
er
n
al
co
r
p
o
r
ate
co
m
m
u
n
icati
o
n
s
[
3
]
,
[
4
]
.
T
h
e
p
r
o
tectio
n
o
f
co
n
f
id
en
tial
d
ata
is
p
ar
am
o
u
n
t,
as
b
r
ea
c
h
es
ca
n
lead
t
o
s
ev
er
e
co
n
s
eq
u
en
ce
s
,
in
clu
d
in
g
id
e
n
tity
th
ef
t,
f
in
an
cial
lo
s
s
,
an
d
lo
s
s
o
f
in
tellectu
al
p
r
o
p
er
ty
.
T
ex
t
class
if
icati
o
n
,
a
s
u
b
f
ield
o
f
n
atu
r
al
lan
g
u
ag
e
p
r
o
ce
s
s
in
g
(
NL
P),
p
lay
s
a
cr
u
cial
r
o
le
in
au
to
m
atin
g
th
e
class
if
icatio
n
o
f
d
o
cu
m
en
ts
i
n
to
t
h
ese
ca
teg
o
r
ies.
T
ex
t
class
if
icatio
n
in
v
o
lv
es
ass
ig
n
in
g
p
r
ed
ef
in
e
d
lab
els
to
tex
t
d
o
cu
m
en
ts
b
ased
o
n
th
eir
co
n
ten
t.
T
r
ad
itio
n
al
m
e
th
o
d
s
r
elied
h
ea
v
ily
o
n
m
an
u
al
r
ev
iew
an
d
r
u
le
-
b
ased
s
y
s
tem
s
,
wh
ich
ar
e
n
o
t
s
ca
lab
le
g
iv
e
n
th
e
v
ast
am
o
u
n
ts
o
f
d
ata
g
e
n
er
ated
d
aily
.
T
h
ese
m
eth
o
d
s
o
f
ten
lack
th
e
f
lex
ib
i
lity
an
d
ac
cu
r
ac
y
n
ee
d
e
d
to
h
an
d
le
co
m
p
lex
an
d
n
u
a
n
ce
d
l
an
g
u
ag
e
in
d
i
v
er
s
e
d
o
cu
m
e
n
ts
.
T
h
e
ad
v
en
t
o
f
d
ee
p
lear
n
in
g
h
as
r
ev
o
lu
t
io
n
ized
tex
t
class
if
icatio
n
,
o
f
f
er
in
g
s
ig
n
if
ican
t
im
p
r
o
v
em
e
n
ts
in
ac
cu
r
ac
y
a
n
d
ef
f
icien
cy
.
Dee
p
lear
n
in
g
m
o
d
els,
s
u
ch
as
co
n
v
o
lu
tio
n
al
n
eu
r
al
n
etwo
r
k
s
(
C
NNs),
r
ec
u
r
r
en
t
n
eu
r
al
n
et
wo
r
k
s
(
R
NNs),
an
d
tr
an
s
f
o
r
m
er
-
b
ased
ar
c
h
itectu
r
es
lik
e
B
E
R
T
(
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
)
,
h
av
e
d
em
o
n
s
tr
ated
ex
ce
p
tio
n
al
p
er
f
o
r
m
a
n
c
e
in
u
n
d
er
s
tan
d
in
g
an
d
p
r
o
ce
s
s
in
g
n
atu
r
al
lan
g
u
a
g
e
[
5
]
.
T
h
ese
m
o
d
els
lev
er
ag
e
lar
g
e
d
atasets
an
d
p
o
wer
f
u
l
co
m
p
u
tin
g
r
eso
u
r
ce
s
to
lear
n
c
o
m
p
lex
p
atter
n
s
an
d
r
elatio
n
s
h
ip
s
with
in
tex
t,
e
n
ab
lin
g
th
em
to
class
if
y
d
o
c
u
m
en
ts
with
h
ig
h
p
r
ec
is
io
n
.
C
NNs,
ty
p
ically
u
s
ed
in
im
ag
e
p
r
o
ce
s
s
in
g
,
h
av
e
b
ee
n
s
u
cc
ess
f
u
lly
a
d
ap
ted
f
o
r
tex
t
class
if
icatio
n
b
y
tr
ea
tin
g
tex
t
as
a
s
eq
u
e
n
ce
o
f
wo
r
d
s
o
r
c
h
ar
ac
ter
s
,
id
en
tif
y
in
g
lo
ca
l
p
atter
n
s
,
an
d
co
m
b
i
n
in
g
th
em
to
f
o
r
m
a
co
m
p
r
e
h
en
s
iv
e
u
n
d
er
s
tan
d
i
n
g
.
R
NNs,
an
d
th
eir
v
a
r
ian
t
l
o
n
g
s
h
o
r
t
-
ter
m
m
em
o
r
y
n
etwo
r
k
s
(
L
STM
s
)
,
ar
e
well
-
s
u
ited
f
o
r
s
eq
u
en
tial
d
a
ta,
ca
p
tu
r
in
g
th
e
c
o
n
tex
t
an
d
d
ep
en
d
en
cies
with
in
tex
t.
T
r
an
s
f
o
r
m
e
r
-
b
ase
d
m
o
d
els,
s
u
ch
as
B
E
R
T
,
ex
ce
l
in
h
an
d
lin
g
l
o
n
g
-
r
an
g
e
d
e
p
en
d
en
cies
an
d
c
o
n
tex
t
b
y
u
s
in
g
s
elf
-
atten
tio
n
m
ec
h
an
is
m
s
,
m
ak
in
g
th
em
p
ar
ticu
lar
ly
ef
f
ec
tiv
e
f
o
r
task
s
in
v
o
lv
in
g
n
u
an
ce
d
lan
g
u
ag
e
an
d
co
n
te
x
tu
al
u
n
d
er
s
tan
d
i
n
g
[
6
]
.
T
h
e
im
p
le
m
en
tatio
n
o
f
a
r
o
b
u
s
t
d
o
cu
m
en
t
class
if
icatio
n
s
y
s
tem
n
o
t
o
n
ly
en
h
a
n
ce
s
d
ata
s
ec
u
r
ity
b
u
t
also
s
tr
ea
m
lin
es
o
p
er
atio
n
s
b
y
e
n
s
u
r
in
g
th
at
s
en
s
itiv
e
in
f
o
r
m
atio
n
is
ap
p
r
o
p
r
iately
h
an
d
led
a
n
d
ac
ce
s
s
ed
.
I
t
en
ab
les
o
r
g
an
izati
o
n
s
to
co
m
p
ly
with
d
ata
p
r
o
te
ctio
n
r
eg
u
latio
n
s
,
s
af
eg
u
a
r
d
i
n
tellect
u
al
p
r
o
p
er
t
y
,
an
d
m
ain
tain
c
u
s
to
m
er
tr
u
s
t.
As th
e
lan
d
s
ca
p
e
o
f
d
ata
c
o
n
ti
n
u
es to
ev
o
lv
e,
th
e
d
e
v
elo
p
m
e
n
t a
n
d
r
e
f
in
em
en
t
o
f
th
ese
class
if
icatio
n
s
y
s
tem
s
r
em
ain
a
cr
itical
ar
ea
o
f
r
esear
c
h
an
d
i
n
n
o
v
atio
n
[
7
]
.
I
n
to
d
a
y
'
s
d
ig
ital
er
a,
th
e
r
ap
id
g
r
o
wth
o
f
d
ata
n
ec
ess
itates
ef
f
ec
tiv
e
class
if
icatio
n
s
y
s
tem
s
,
p
ar
ticu
lar
ly
f
o
r
s
en
s
itiv
e
in
f
o
r
m
atio
n
in
s
ec
to
r
s
lik
e
f
in
a
n
ce
,
h
ea
lth
ca
r
e,
an
d
tech
n
o
l
o
g
y
.
E
n
s
u
r
in
g
d
ata
s
ec
u
r
ity
an
d
p
r
i
v
ac
y
is
cr
u
cia
l
to
p
r
ev
en
t
u
n
au
th
o
r
ized
ac
c
ess
an
d
b
r
ea
ch
es,
wh
ile
also
m
ee
tin
g
r
eg
u
lat
o
r
y
co
m
p
lian
ce
r
eq
u
ir
em
en
ts
s
u
ch
as
HI
PA
A
an
d
GDP
R
(
g
en
er
al
d
ata
p
r
o
tectio
n
r
eg
u
latio
n
)
[
8
]
,
[
9
]
.
T
r
ad
itio
n
al
m
an
u
al
m
eth
o
d
s
ar
e
in
s
u
f
f
icien
t
to
h
an
d
le
th
e
s
h
ee
r
v
o
lu
m
e
o
f
d
ata,
p
r
o
m
p
tin
g
t
h
e
n
ee
d
f
o
r
a
u
to
m
ated
,
ef
f
icien
t,
an
d
ac
c
u
r
ate
class
if
icatio
n
m
ec
h
a
n
is
m
s
.
R
ec
en
t
ad
v
an
ce
m
en
ts
in
d
ee
p
lear
n
in
g
a
n
d
NL
P,
in
clu
d
in
g
C
NNs,
R
NNs,
an
d
t
r
an
s
f
o
r
m
e
r
-
b
ased
m
o
d
els
lik
e
B
E
R
T
,
o
f
f
er
p
r
o
m
is
in
g
s
o
lu
tio
n
s
[
1
0
]
.
T
h
ese
tech
n
o
lo
g
ies
ca
n
en
h
an
ce
t
h
e
ac
cu
r
ac
y
a
n
d
r
o
b
u
s
tn
ess
o
f
s
en
s
itiv
e
in
f
o
r
m
atio
n
d
etec
tio
n
,
m
itig
atin
g
r
is
k
s
ass
o
ciate
d
with
m
is
class
if
icatio
n
an
d
en
s
u
r
in
g
p
r
o
p
e
r
d
ata
h
an
d
lin
g
.
T
h
e
m
o
tiv
atio
n
f
o
r
th
is
r
esear
ch
is
to
lev
er
ag
e
th
ese
tech
n
o
lo
g
ical
ad
v
an
ce
m
en
ts
t
o
d
ev
elo
p
a
s
tate
-
of
-
t
h
e
-
ar
t
s
y
s
tem
f
o
r
ca
teg
o
r
izin
g
d
o
cu
m
en
ts
in
to
p
u
b
lic,
p
r
iv
ate,
an
d
co
n
f
id
en
tial
lev
el
s
,
th
er
eb
y
im
p
r
o
v
i
n
g
o
p
er
atio
n
al
ef
f
icien
cy
a
n
d
c
o
m
p
lian
ce
.
−
Dev
elo
p
m
en
t
o
f
MT
SAN
f
r
am
ewo
r
k
:
in
tr
o
d
u
ce
d
th
e
m
u
lti
-
tier
s
en
s
itiv
ity
an
aly
s
is
n
etwo
r
k
(
MT
SAN)
f
o
r
p
r
ec
is
e
class
if
icatio
n
o
f
d
o
cu
m
en
ts
in
to
p
u
b
lic,
p
r
iv
ate,
an
d
co
n
f
id
e
n
tial
ca
teg
o
r
ies.
−
I
n
n
o
v
ativ
e
m
u
lti
-
tier
f
ea
tu
r
e
en
co
d
i
n
g
:
im
p
lem
e
n
ted
t
h
e
m
u
lti
-
tier
s
en
s
itiv
ity
en
co
d
in
g
n
etwo
r
k
(
MT
SEN)
to
ca
p
tu
r
e
m
u
lti
-
tier
d
o
cu
m
en
t f
ea
tu
r
es,
p
r
o
v
id
in
g
a
n
u
a
n
ce
d
u
n
d
er
s
tan
d
in
g
o
f
co
n
ten
t.
−
E
n
h
an
ce
d
co
n
tex
t
u
al
u
n
d
er
s
tan
d
in
g
with
DSGC
B
:
d
ev
elo
p
ed
th
e
d
u
al
-
s
co
p
e
g
r
ap
h
co
n
v
o
lu
tio
n
b
lo
c
k
(
DSGC
B
)
to
ef
f
ec
tiv
ely
in
teg
r
ate
g
lo
b
al
an
d
lo
ca
l c
o
n
tex
tu
al
in
f
o
r
m
atio
n
.
−
Ad
v
an
ce
d
f
ea
tu
r
e
f
u
s
io
n
wit
h
C
T
I
FB
:
u
til
ized
th
e
cr
o
s
s
-
tier
in
f
o
r
m
atio
n
f
u
s
io
n
b
l
o
c
k
(
C
T
I
FB
)
t
o
s
ea
m
less
ly
m
er
g
e
m
u
lti
-
lev
el
f
ea
tu
r
es,
en
h
a
n
cin
g
class
if
icat
io
n
ac
cu
r
ac
y
2.
RE
L
AT
E
D
WO
RK
T
h
e
p
r
im
ar
y
o
b
jectiv
e
o
f
th
i
s
s
tu
d
y
is
to
d
e
v
elo
p
a
r
o
b
u
s
t
an
d
e
f
f
icien
t
s
y
s
tem
f
o
r
c
lass
if
y
in
g
d
o
cu
m
e
n
ts
in
to
p
u
b
lic,
p
r
iv
at
e,
an
d
co
n
f
i
d
en
tial
ca
teg
o
r
ies,
p
ar
ticu
lar
ly
in
th
e
co
n
tex
t
o
f
h
an
d
lin
g
lo
n
g
a
n
d
co
m
p
lex
d
o
c
u
m
en
ts
.
T
h
is
s
y
s
t
em
aim
s
to
le
v
er
ag
e
ad
v
a
n
ce
d
m
ac
h
in
e
lear
n
in
g
an
d
d
ee
p
le
ar
n
in
g
tech
n
iq
u
es
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:
2502
-
4
7
5
2
Dee
p
lea
r
n
in
g
-
b
a
s
ed
mu
lti
-
ti
er sen
s
itivi
ty
a
n
a
lysi
s
n
etw
o
r
k
fo
r
…
(
S
a
d
iy
a
A
n
s
a
r
i
)
1251
to
ac
cu
r
ately
id
e
n
tify
a
n
d
d
if
f
er
en
tiate
s
en
s
itiv
e
in
f
o
r
m
atio
n
,
th
e
r
eb
y
im
p
r
o
v
in
g
d
o
cu
m
e
n
t
m
an
a
g
em
en
t
an
d
en
s
u
r
in
g
d
ata
s
ec
u
r
ity
.
Pu
ja
r
et
a
l.
[
1
0
]
u
tili
ze
d
B
E
R
T
to
s
eg
m
en
t
len
g
th
y
te
x
ts
,
g
en
er
atin
g
in
itial
r
ep
r
esen
tatio
n
s
f
o
r
ea
ch
s
eg
m
en
t.
T
h
e
in
te
r
ac
tio
n
s
b
etw
ee
n
s
eg
m
en
ts
wer
e
s
u
b
s
eq
u
e
n
tly
m
o
d
ele
d
u
s
in
g
eith
er
a
r
ec
u
r
r
en
t
lay
e
r
o
r
a
t
r
an
s
f
o
r
m
er
.
B
u
ild
in
g
o
n
th
is
,
[
1
1
]
in
tr
o
d
u
ce
d
E
R
NI
E
-
DOC
,
wh
ich
in
clu
d
es
a
r
etr
o
s
p
ec
tiv
e
f
ee
d
m
ec
h
a
n
is
m
th
at
allo
ws
f
o
r
th
e
i
n
teg
r
atio
n
o
f
s
em
an
tic
in
f
o
r
m
ati
o
n
f
r
o
m
t
h
e
en
tire
d
o
cu
m
e
n
t.
T
o
b
etter
r
ep
r
esen
t
th
e
s
tr
u
ctu
r
al
f
ea
tu
r
es
o
f
lo
n
g
d
o
cu
m
en
ts
,
[
1
2
]
p
r
o
p
o
s
ed
t
h
e
h
ier
ar
ch
ical
g
r
a
p
h
co
n
v
o
l
u
tio
n
al
n
etwo
r
k
s
(
GC
Ns)
,
co
n
s
tr
u
ctin
g
b
o
th
s
ec
tio
n
an
d
w
o
r
d
g
r
ap
h
s
to
ex
p
licitly
m
o
d
el
m
ac
r
o
a
n
d
m
icr
o
s
tr
u
ctu
r
al
in
f
o
r
m
atio
n
.
Ad
d
itio
n
ally
,
s
en
ten
ce
s
tr
u
ctu
r
es we
r
e
in
co
r
p
o
r
ate
d
in
to
wo
r
d
g
r
ap
h
m
o
d
elin
g
to
en
h
an
ce
f
ea
tu
r
e
lear
n
in
g
ca
p
a
b
ilit
ies.
T
o
ad
d
r
ess
th
e
co
m
p
u
tatio
n
al
co
m
p
lex
ity
in
h
er
e
n
t
i
n
th
e
s
elf
-
atten
tio
n
m
ec
h
an
is
m
o
f
tr
an
s
f
o
r
m
er
s
,
[
1
3
]
p
r
o
p
o
s
ed
v
a
r
io
u
s
s
p
ar
s
e
a
tten
tio
n
m
ec
h
a
n
is
m
s
,
aim
in
g
t
o
r
estrict
th
e
r
a
n
g
e
o
f
to
k
e
n
in
ter
ac
tio
n
s
an
d
r
ed
u
ce
co
m
p
u
tatio
n
al
o
v
er
h
ea
d
.
T
h
r
ee
m
ain
ty
p
es
o
f
s
p
ar
s
ity
p
atter
n
s
h
av
e
b
ee
n
id
en
tifie
d
:
1
)
f
i
x
ed
p
atter
n
,
wh
ich
in
clu
d
es
win
d
o
w
-
b
ased
,
g
l
o
b
al,
a
n
d
r
an
d
o
m
atten
tio
n
tech
n
iq
u
es.
Fo
r
in
s
tan
ce
,
[
1
4
]
p
r
esen
te
d
a
h
y
b
r
id
ap
p
r
o
ac
h
th
at
co
m
b
in
es
w
in
d
o
wed
lo
ca
l
-
co
n
tex
t
atten
tio
n
with
task
-
d
r
iv
e
n
g
lo
b
al
atten
tio
n
,
e
f
f
ec
tiv
ely
r
e
d
u
cin
g
c
o
m
p
u
tatio
n
al
co
m
p
le
x
ity
f
r
o
m
q
u
a
d
r
atic
to
lin
ea
r
.
T
h
is
ap
p
r
o
ac
h
was
f
u
r
th
er
r
ef
in
e
d
b
y
[
1
5
]
,
w
h
o
i
n
co
r
p
o
r
ated
r
an
d
o
m
atten
tio
n
,
m
ain
tain
in
g
lin
ea
r
co
m
p
lex
it
y
wh
ile
im
p
r
o
v
in
g
m
o
d
el
p
er
f
o
r
m
an
ce
.
2
)
L
ea
r
n
ab
le
p
atter
n
,
ex
em
p
lifie
d
b
y
[
1
6
]
,
wh
ich
d
y
n
am
ically
d
eter
m
in
es
th
e
ass
o
ciate
d
r
eg
io
n
s
f
o
r
ea
ch
to
k
e
n
,
th
er
e
b
y
en
h
a
n
cin
g
th
e
c
ap
tu
r
e
o
f
s
em
an
tic
co
r
r
elatio
n
s
.
3
)
L
o
w
-
r
an
k
p
atter
n
s
,
as
ex
p
lo
r
ed
b
y
[
1
7
]
,
wh
er
e
s
elf
-
atten
tio
n
m
atr
ices
ar
e
p
r
o
j
ec
ted
in
to
lo
wer
-
d
im
en
s
io
n
al
s
p
ac
es
to
r
ed
u
ce
co
m
p
lex
ity
,
lev
er
a
g
in
g
th
e
o
b
s
er
v
atio
n
th
at
th
ese
m
atr
ices
o
f
ten
ex
h
ib
it
lo
w
-
r
an
k
p
r
o
p
e
r
ties
.
Kitaev
et
a
l.
[
1
8
]
i
n
tr
o
d
u
ce
d
th
e
Dee
p
Do
c
class
if
ier
,
u
tili
zin
g
a
d
ee
p
C
N
N
with
th
e
Alex
Net
ar
ch
itectu
r
e.
T
h
is
m
o
d
el
was
p
r
etr
ain
ed
o
n
th
e
I
m
ag
eNe
t d
a
t
aset,
s
ig
n
if
ican
tly
im
p
r
o
v
in
g
p
r
ev
io
u
s
ap
p
r
o
ac
h
es b
y
co
n
v
e
r
tin
g
co
n
v
o
lu
tio
n
al
lay
er
s
in
to
f
lex
ib
le
f
ea
tu
r
e
e
x
tr
ac
to
r
s
.
Fu
r
th
er
m
o
r
e
,
W
an
g
et
a
l.
[
1
9
]
co
m
b
in
e
d
tex
tu
al
d
ata
f
r
o
m
co
m
m
er
cial
OC
R
s
y
s
tem
s
with
r
aw
im
ag
e
d
ata.
T
h
is
d
ata
was
th
e
n
p
r
o
ce
s
s
ed
b
y
a
N
L
P
m
o
d
el,
wh
ich
tr
an
s
lated
t
h
e
tex
t
in
t
o
th
e
f
ea
tu
r
e
s
p
ac
e.
An
ex
tr
em
e
lear
n
in
g
m
ac
h
in
e
was
em
p
lo
y
ed
to
m
an
ip
u
la
te
f
r
o
ze
n
c
o
n
v
o
lu
tio
n
al
lay
e
r
s
tr
ain
ed
with
th
e
Alex
Net
m
o
d
el,
r
esu
ltin
g
in
im
p
r
o
v
e
d
o
u
tp
u
t
with
o
u
t
c
o
m
p
r
o
m
is
in
g
ac
c
u
r
ac
y
.
I
n
th
e
r
e
alm
o
f
h
ier
a
r
ch
ical
tex
t
class
if
icatio
n
,
Ma
r
ten
s
an
d
Pro
v
o
s
t
[
2
0
]
in
tr
o
d
u
ce
d
th
e
Seq
2
L
ab
el
f
r
am
ew
o
r
k
,
w
h
ich
em
p
lo
y
s
a
r
an
d
o
m
g
en
er
ativ
e
a
p
p
r
o
ac
h
t
o
lear
n
i
n
g
lab
el
h
ier
ar
ch
ies
b
y
s
h
u
f
f
lin
g
lab
el
s
eq
u
e
n
ce
s
d
u
r
in
g
tr
ain
i
n
g
.
Ver
m
eir
e
et
a
l.
[
2
1
]
p
r
o
p
o
s
ed
th
e
v
ar
iatio
n
al
c
o
n
tin
u
o
u
s
lab
el
d
is
tr
ib
u
tio
n
l
ea
r
n
in
g
(
VC
L
DL
)
f
r
am
ewo
r
k
,
wh
ich
tr
ea
ts
lab
el
d
is
tr
ib
u
tio
n
as
a
co
n
tin
u
o
u
s
d
en
s
ity
f
u
n
ctio
n
in
laten
t
s
p
ac
e
.
T
h
is
m
eth
o
d
estab
lis
h
es
a
r
elatio
n
s
h
ip
b
etwe
en
th
e
f
ea
tu
r
e
an
d
lab
el
s
p
ac
es,
u
n
co
v
e
r
in
g
in
f
o
r
m
atio
n
h
id
d
en
in
o
b
s
er
v
a
b
le
lo
g
ical
lab
els.
Ad
d
itio
n
ally
,
L
an
g
et
a
l.
[
2
2
]
in
tr
o
d
u
ce
d
th
e
p
r
o
m
p
t
-
b
ased
lab
el
-
awa
r
e
f
r
am
ewo
r
k
f
o
r
m
u
lti
-
lab
el
tex
t
clas
s
if
icatio
n
(
PLAM
L
)
.
T
h
is
f
r
am
ewo
r
k
e
n
h
an
ce
s
p
r
o
m
p
t
-
b
ased
lear
n
in
g
with
th
r
ee
k
e
y
tech
n
iq
u
es:
a
to
k
en
weig
h
tin
g
alg
o
r
ith
m
co
n
s
id
er
in
g
la
b
el
c
o
r
r
elatio
n
s
,
a
tem
p
late
f
o
r
a
u
g
m
en
tin
g
tr
ain
in
g
s
am
p
les
t
o
m
ak
e
th
e
p
r
o
ce
s
s
lab
el
-
awa
r
e,
an
d
a
d
y
n
a
m
ic
th
r
esh
o
ld
m
ec
h
a
n
is
m
to
r
ef
in
e
t
h
e
p
r
ed
ictio
n
co
n
d
itio
n
o
f
ea
c
h
lab
el.
Z
h
ao
et
a
l.
[
2
3
]
f
u
r
th
er
r
e
f
in
ed
th
e
m
u
lti
-
lab
el
class
if
icat
io
n
ap
p
r
o
ac
h
b
y
p
r
o
p
o
s
in
g
PLAM
L
,
wh
ich
s
p
ec
if
ically
ad
d
r
ess
es
th
e
ch
allen
g
es
ass
o
ciate
d
with
m
u
lti
-
lab
el
class
if
icatio
n
.
T
h
ese
ad
v
an
ce
m
en
ts
co
llectiv
ely
co
n
tr
ib
u
te
to
th
e
d
e
v
elo
p
m
e
n
t
o
f
m
o
r
e
ef
f
icien
t a
n
d
ac
cu
r
ate
d
o
cu
m
e
n
t c
lass
if
icatio
n
s
y
s
te
m
s
.
Desp
ite
s
ig
n
if
ican
t
a
d
v
a
n
ce
m
en
ts
in
d
o
cu
m
e
n
t
class
if
icatio
n
,
ex
is
tin
g
m
eth
o
d
s
,
s
u
c
h
as
th
o
s
e
lev
er
ag
in
g
B
E
R
T
,
E
R
NI
E
-
D
OC
,
an
d
h
ier
ar
ch
ical
GC
N
s
,
o
f
ten
f
ac
e
c
h
allen
g
es
in
ac
c
u
r
ately
ca
teg
o
r
izin
g
s
en
s
itiv
e
in
f
o
r
m
atio
n
,
esp
ec
ially
in
lo
n
g
an
d
co
m
p
lex
d
o
c
u
m
en
ts
.
T
h
e
co
m
p
u
tatio
n
al
o
v
er
h
ea
d
ass
o
ciate
d
with
s
elf
-
atten
tio
n
m
ec
h
a
n
is
m
s
an
d
th
e
lim
ited
ab
ilit
y
t
o
ca
p
tu
r
e
n
u
a
n
ce
d
c
o
n
tex
tu
al
r
elatio
n
s
h
ip
s
f
u
r
th
e
r
ex
ac
er
b
ate
th
ese
ch
allen
g
es.
Ad
d
itio
n
ally
,
wh
ile
f
r
am
ewo
r
k
s
lik
e
PLAM
L
an
d
VC
L
DL
h
av
e
im
p
r
o
v
ed
m
u
lti
-
lab
el
class
if
icatio
n
,
th
ey
m
a
y
s
till
f
all
s
h
o
r
t
in
s
ce
n
ar
io
s
r
eq
u
ir
in
g
p
r
ec
is
e
d
if
f
er
en
ti
atio
n
o
f
d
o
cu
m
e
n
t
s
en
s
itiv
ity
lev
els
[
2
4
]
,
[
2
5
]
.
T
h
is
r
esear
ch
aim
s
to
a
d
d
r
ess
th
ese
g
ap
s
b
y
d
e
v
elo
p
in
g
a
r
o
b
u
s
t
class
if
icatio
n
s
y
s
tem
th
at
lev
er
ag
es
ad
v
an
ce
d
d
ee
p
-
lear
n
i
n
g
tech
n
iq
u
e
s
to
en
h
an
ce
t
h
e
ac
cu
r
ac
y
an
d
ef
f
icien
cy
o
f
id
en
tify
in
g
a
n
d
ca
teg
o
r
izin
g
s
en
s
itiv
e
in
f
o
r
m
atio
n
.
3.
P
RO
P
O
SE
D
M
E
T
H
O
D
T
h
is
s
ec
tio
n
o
f
th
e
p
r
o
p
o
s
ed
s
tu
d
y
f
o
cu
s
es
o
n
th
e
p
r
o
p
o
s
e
d
m
o
d
el
MT
SAN
th
at
aim
s
to
class
if
y
s
en
s
itiv
e
d
ata.
T
h
is
is
ca
teg
o
r
ized
in
to
v
a
r
io
u
s
s
tag
es,
n
am
el
y
p
r
iv
ate,
p
u
b
lic,
a
n
d
c
o
n
f
id
e
n
tial
s
tag
es
h
av
in
g
an
in
cr
ea
s
ed
ac
c
u
r
ac
y
.
T
h
e
p
r
o
p
o
s
ed
m
o
d
el
h
as
th
r
ee
m
aj
o
r
p
h
ases
,
th
e
MT
SEN,
th
e
p
r
o
p
o
s
ed
i
n
ter
ac
tiv
e
n
etwo
r
k
,
a
n
d
th
e
i
n
teg
r
ated
s
e
n
s
itiv
ity
f
ea
tu
r
e
f
u
s
io
n
(
I
SF
F).
Fig
u
r
e
1
s
h
o
ws th
e
MT
SAN
m
o
d
el.
T
h
e
MT
SEN
co
n
s
u
m
es
th
e
in
p
u
t
d
o
c
u
m
en
t
an
d
s
tu
d
ies
th
e
m
u
lti
-
lev
el
ex
p
r
ess
io
n
s
f
o
r
s
en
ten
ce
s
,
wo
r
d
s
as
well
as
s
ec
tio
n
s
.
F
u
r
th
er
,
t
h
e
p
r
o
p
o
s
ed
“
MT
SAN”
is
im
p
lem
en
ted
f
o
r
d
ata
i
n
ter
ac
tio
n
b
etwe
en
v
ar
io
u
s
lev
els
o
f
th
e
c
o
n
v
o
l
u
tio
n
al
n
etwo
r
k
.
Her
e,
th
e
c
o
n
v
o
lu
ti
o
n
al
n
etwo
r
k
is
d
esig
n
ed
p
a
r
ticu
lar
ly
to
h
an
d
le
len
g
th
y
d
o
c
u
m
en
ts
,
w
h
ile
th
e
co
n
v
o
l
u
tio
n
al
n
etwo
r
k
f
o
r
wo
r
d
s
as
well
as
s
en
ten
ce
s
is
in
v
ested
in
g
r
asp
in
g
th
e
m
icr
o
s
tr
u
ctu
r
e
o
f
th
e
d
o
cu
m
en
t.
T
h
e
f
in
e
-
tu
n
ed
n
o
d
es
th
at
ar
e
r
e
d
u
n
d
an
t
ar
e
o
m
itted
an
d
th
e
in
te
r
ac
tio
n
o
f
th
e
m
u
lti
-
lev
el
d
ata
is
im
p
r
o
v
is
ed
,
s
ec
tio
n
-
aid
ed
s
eg
m
en
ts
f
o
r
p
o
o
lin
g
as
well
as
C
T
I
FB
ar
e
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
5
0
2
-
4
7
5
2
I
n
d
o
n
esian
J
E
lec
E
n
g
&
C
o
m
p
Sci
,
Vo
l.
3
9
,
No
.
2
,
Au
g
u
s
t
20
25
:
1
2
4
9
-
1
260
1252
im
p
lem
en
ted
p
r
i
o
r
as
well
a
s
af
ter
ev
er
y
lay
er
o
f
th
e
w
o
r
d
as
well
as
s
en
ten
ce
co
n
v
o
lu
tio
n
s
.
T
h
ese
ar
e
u
tili
ze
d
to
co
m
b
in
e
th
e
t
h
r
ee
-
l
ev
el
g
r
ap
h
s
in
to
o
n
e
in
teg
r
ated
u
n
it.
W
e
also
in
tr
o
d
u
ce
a
g
l
o
b
al
lo
ca
l
g
r
ap
h
ical
co
n
v
o
l
u
tio
n
al
s
eg
m
en
t
to
d
y
n
am
ically
g
r
asp
th
e
lo
ca
l
as
we
ll
as
th
e
g
lo
b
al
f
ea
tu
r
es
in
s
id
e
th
e
s
en
ten
ce
n
o
d
es
wh
ich
th
er
ef
o
r
e
in
cr
ea
s
es
th
e
ca
p
ab
ilit
ies
o
f
f
ea
tu
r
e
r
ep
r
esen
tatio
n
.
L
astl
y
,
th
e
I
S
FF
is
im
p
lem
en
ted
f
o
r
in
teg
r
atio
n
o
f
th
e
p
r
io
r
m
u
lti
-
l
ev
el
f
ea
tu
r
es f
o
r
th
e
last
o
f
len
g
th
y
d
o
cu
m
e
n
t c
ateg
o
r
izatio
n
.
Fig
u
r
e
1
.
Mu
lti
-
tier
s
en
s
itiv
ity
an
aly
s
is
n
etwo
r
k
3
.
1
.
M
ulti
-
t
ier
s
ens
it
iv
it
y
enco
din
g
net
wo
rk
A
d
o
cu
m
en
t
is
n
o
r
m
ally
m
ad
e
u
p
o
f
s
ec
tio
n
s
th
at
co
u
ld
b
e
f
u
r
th
er
b
r
o
k
en
d
o
wn
in
t
o
s
en
ten
ce
s
an
d
wo
r
d
s
.
T
h
e
d
o
c
u
m
en
t
is
s
p
lit
in
to
s
ec
tio
n
s
f
o
r
a
f
ix
ed
d
im
e
n
s
io
n
to
g
r
asp
th
e
lay
e
r
ed
d
ata
r
ep
r
esen
tatio
n
s
o
m
itti
n
g
r
e
d
u
n
d
an
cy
as
is
e
x
p
r
ess
ed
as
{
1
,
2
,
…
.
,
}
.
W
er
e,
=
{
0
,
1
,
…
,
}
is
th
e
−
ℎ
s
ec
tio
n
h
av
in
g
+
1
to
k
en
s
.
T
o
m
ain
tain
a
s
im
ilar
len
g
th
f
o
r
ev
er
y
s
e
ctio
n
,
p
a
d
d
in
g
is
u
tili
ze
d
if
n
ee
d
ed
.
E
v
er
y
s
ec
tio
n
is
s
to
r
ed
in
th
e
p
r
io
r
tr
ain
ed
en
co
d
er
ℎ
(
∙
,
)
,
th
e
p
ar
am
eter
s
h
er
e
ar
e
d
en
o
ted
as
.
Her
e,
in
th
e
last
pha
s
e,
th
e
to
k
e
n
is
co
n
s
id
er
ed
a
s
ec
tio
n
attr
ib
u
te.
W
h
er
e
as,
th
e
o
th
er
to
k
en
s
ar
e
co
n
s
id
er
ed
f
o
r
wo
r
d
attr
ib
u
tes.
Fu
r
th
er
,
r
esu
ltin
g
in
s
ec
tio
n
-
lev
el
attr
ib
u
tes
f
o
r
len
g
th
y
d
o
cu
m
e
n
ts
=
[
1
,
2
,
…
,
]
×
an
d
f
o
r
w
o
r
d
lev
el,
it
is
g
iv
e
n
as
=
[
11
,
…
.
.
,
1
,
…
,
1
,
…
.
.
,
]
×
.
T
h
e
attr
ib
u
te
d
im
en
s
io
n
is
d
e
n
o
ted
as
.
T
h
e
s
en
ten
ce
ex
p
r
ess
io
n
s
ar
e
attain
ed
b
y
co
m
b
in
in
g
th
e
wo
r
d
attr
ib
u
tes
f
o
r
ev
er
y
s
en
ten
ce
v
ia
p
o
o
lin
g
.
Sen
ten
ce
m
ask
in
g
o
p
er
atio
n
is
u
s
ed
.
L
et
u
s
ass
u
m
e
th
er
e
ar
e
s
en
ten
ce
s
in
th
e
d
o
c
u
m
en
t,
th
er
e
ar
e
,
=
1
,
2
,
…
,
wo
r
d
s
in
th
e
−
ℎ
s
en
ten
ce
.
T
h
is
is
ex
p
r
ess
ed
as
=
[
1
,
…
,
1
,
2
,
…
.
,
2
,
3
,
…
.
,
3
,
…
.
,
,
…
.
,
]
.
Her
e,
th
e
n
u
m
b
er
s
th
at
ar
e
id
en
tical
s
h
o
wca
s
e
th
at
th
e
wo
r
d
s
at
t
h
o
s
e
l
o
c
a
t
i
o
n
s
a
r
e
f
r
o
m
t
h
e
s
a
m
e
s
en
t
e
n
c
e
,
t
h
e
v
a
l
u
e
o
f
t
h
e
n
u
m
b
e
r
s
e
x
p
r
e
s
s
e
s
t
h
e
p
o
s
it
i
o
n
a
l
d
at
a
o
f
t
h
e
s
e
n
t
e
n
c
e
s
.
A
n
o
t
h
e
r
p
r
o
j
e
c
t
i
o
n
l
a
y
e
r
is
a
d
d
e
d
t
o
o
b
t
a
i
n
t
h
e
s
e
n
t
e
n
c
e
a
tt
r
i
b
u
t
e
s
w
h
ic
h
a
r
e
e
x
p
r
e
s
s
e
d
a
s
g
i
v
e
n
i
n
(
1
)
.
=
(
[
=
=
1
]
)
,
wh
er
e
=
1
,
…
,
=
+
,
=
1
,
…
,
(
1
)
I
n
th
is
eq
u
atio
n
,
we
o
b
s
er
v
e
t
h
e
m
ax
p
o
o
lin
g
is
im
p
lem
en
t
ed
co
lu
m
n
-
wis
e.
T
h
e
tr
ain
a
b
l
e
attr
ib
u
tes
ar
e
r
ep
r
esen
ted
as
×
.
I
n
ad
d
itio
n
to
s
en
ten
ce
m
ask
in
g
,
t
h
er
e
ar
e
two
ad
d
ed
t
r
an
s
f
er
m
ask
in
g
u
tili
ze
d
in
th
is
n
etwo
r
k
m
o
d
e
l
th
at
ar
e
d
en
o
ted
as
−
×
an
d
−
×
,
wh
ich
is
u
s
ed
to
estab
lis
h
th
e
r
elatio
n
s
h
ip
b
etwe
en
wo
r
d
s
,
s
en
ten
ce
s
as
well
a
s
s
ec
tio
n
s
.
T
h
e
d
ef
in
itio
n
g
iv
en
f
o
r
(
2
)
is
s
u
ch
th
at
if
a
wo
r
d
o
r
a
s
en
ten
ce
b
elo
n
g
s
to
a
p
ar
ticu
lar
s
ec
tio
n
th
en
th
e
m
ask
in
g
s
co
r
e
is
s
et
to
1
o
th
er
wis
e
it is
s
et
to
0
.
[
−
]
=
{
1
,
ℎ
0
,
[
−
]
=
{
1
,
ℎ
0
,
(
2
)
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:
2502
-
4
7
5
2
Dee
p
lea
r
n
in
g
-
b
a
s
ed
mu
lti
-
ti
er sen
s
itivi
ty
a
n
a
lysi
s
n
etw
o
r
k
fo
r
…
(
S
a
d
iy
a
A
n
s
a
r
i
)
1253
3
.
2
.
M
ulti
-
t
ier
s
ens
it
iv
it
y
a
na
ly
s
is
net
wo
rk
T
h
e
p
r
o
p
o
s
ed
m
o
d
el
co
n
s
is
ts
o
f
s
ev
er
al
g
r
ap
h
ical
co
n
v
o
l
u
tio
n
al
n
etwo
r
k
s
f
o
r
wo
r
d
s
,
an
d
s
en
ten
ce
s
as
well
as
s
ec
tio
n
s
to
g
r
asp
i
n
tr
a
an
d
in
ter
-
lev
el
co
r
r
elatio
n
s
f
o
r
len
g
th
y
d
o
cu
m
e
n
ts
.
C
o
n
s
id
er
in
g
th
e
Sectio
n
g
r
ap
h
,
th
e
s
ec
tio
n
s
ca
le
attr
ib
u
tes
ar
e
g
iv
en
as
×
.
Self
-
atten
tio
n
m
ec
h
a
n
is
m
is
im
p
lem
en
te
d
to
b
u
ild
n
o
d
es f
o
r
th
e
c
o
m
p
let
ely
lin
k
ed
g
r
ap
h
as g
iv
en
in
(
3
)
.
=
(
,
,
)
,
=
(
(
(
)
×
(
)
)
(
)
−
1
2
)
(
3
)
Her
e,
th
e
ed
g
e
s
et
is
d
en
o
ted
as
,
an
d
th
e
a
d
jace
n
t
m
atr
ix
is
g
iv
en
as
f
o
r
th
e
g
r
ap
h
.
N
o
r
m
aliza
tio
n
wh
ich
is
r
o
w
-
wis
e
i
s
im
p
lem
en
ted
an
d
th
e
weig
h
ts
ar
e
g
iv
en
as
×
,
×
.
T
h
e
d
im
en
s
io
n
o
f
th
e
attr
ib
u
t
e
is
g
iv
en
as
.
T
h
e
o
u
tp
u
t
f
o
r
th
e
g
r
ap
h
ical
co
n
v
o
lu
tio
n
al
n
e
two
r
k
f
o
r
th
e
−
ℎ
lay
er
is
as
g
iv
en
b
el
o
w,
wh
er
e
th
e
d
iag
o
n
a
l
n
o
d
e
m
atr
i
x
is
g
iv
e
n
as
̃
,
=
∑
(
,
)
.
×
with
as th
e
ac
tiv
atio
n
f
u
n
ctio
n
.
+
1
=
(
1
√
̃
√
̃
)
(
4
)
W
o
r
d
,
as
well
as
co
n
tex
tu
al
-
s
en
s
itiv
ity
g
r
ap
h
s
,
h
av
e
n
o
d
es
th
at
b
eg
in
with
a
h
u
g
e
q
u
a
n
tity
th
at
h
as
v
is
ib
le
co
m
p
lex
ity
wh
ile
co
n
s
id
er
in
g
th
e
co
m
p
u
tatio
n
s
in
v
o
lv
e
d
.
T
h
er
ef
o
r
e,
we
in
tr
o
d
u
ce
s
ec
tio
n
-
aid
e
d
p
o
o
lin
g
s
eg
m
en
ts
f
o
r
s
en
ten
ce
s
as
well
as
s
ec
tio
n
-
aid
ed
p
o
o
lin
g
s
eg
m
en
ts
f
o
r
wo
r
d
s
th
at
allo
w
p
o
o
lin
g
o
p
er
atio
n
s
iter
ativ
el
y
.
T
h
is
h
elp
s
to
o
m
it
n
o
d
es
th
at
ar
e
r
ed
u
n
d
a
n
t
as
well
as
d
ec
r
ea
s
e
s
th
e
co
m
p
u
tatio
n
a
l
co
m
p
lex
ity
an
d
r
e
p
r
esen
ts
g
r
a
p
h
s
f
o
r
wo
r
d
s
a
n
d
s
en
ten
ce
s
a
t
v
ar
io
u
s
lev
els.
T
h
is
im
p
r
o
v
e
s
th
e
r
e
p
r
esen
tatio
n
th
at
is
p
r
o
d
u
ce
d
f
o
r
t
h
e
co
n
cl
u
d
in
g
attr
ib
u
tes.
W
e
ex
p
r
ess
th
e
p
r
o
jectio
n
v
ec
to
r
as
=
(
)
,
ap
p
li
ed
o
n
th
e
attr
ib
u
tes
o
f
th
e
−
ℎ
lay
er
o
f
t
h
e
m
ac
r
o
-
s
en
s
itiv
ity
g
r
ap
h
,
wh
er
e
.
T
h
e
in
d
ex
f
o
r
t
h
e
o
v
er
h
ea
d
n
o
d
es
is
r
etr
iev
ed
h
av
in
g
s
ca
lar
p
r
o
jectio
n
s
co
r
es
f
o
r
s
en
ten
ce
s
an
d
s
im
ilar
ly
f
o
r
wo
r
d
s
.
T
h
is
is
ex
p
r
ess
ed
as
=
‖
‖
,
=
(
,
)
an
d
=
‖
‖
,
=
(
,
)
.
Her
e,
th
e
r
an
k
in
g
o
f
n
o
d
es
is
ex
p
r
ess
ed
as
(
,
)
an
d
(
,
)
th
at
r
esu
lts
in
th
e
lar
g
est
v
alu
e
in
an
d
lar
g
est
v
alu
e
in
.
T
h
e
d
ata
p
r
o
p
ag
atio
n
f
o
r
s
ec
tio
n
-
aid
ed
p
o
o
lin
g
s
eg
m
en
ts
f
o
r
s
en
ten
ce
s
as we
ll a
s
s
ec
tio
n
-
aid
ed
p
o
o
lin
g
s
eg
m
e
n
ts
f
o
r
wo
r
d
s
is
f
o
r
m
u
lated
as g
iv
e
n
in
(
5
)
.
,
̃
=
(
,
:
)
−
+
1
=
−
(
,
:
)
̃
=
(
(
)
)
=
̃
⊗
(
̃
1
)
(
5
)
,
̃
=
(
,
:
)
−
+
1
=
−
(
,
:
)
−
+
1
=
−
(
,
)
−
+
1
=
−
(
,
)
̃
=
(
(
)
)
=
̃
⊗
(
̃
1
)
(
6
)
C
o
n
s
id
er
in
g
th
e
(
4
)
an
d
(
5
)
,
th
e
s
u
b
-
m
at
r
ices
ar
e
(
,
:
)
,
(
,
:
)
,
−
(
,
:
)
,
−
(
,
:
)
,
−
(
,
)
an
d
−
(
,
)
b
y
c
h
o
o
s
in
g
a
r
o
w
o
r
co
lu
m
n
ac
c
o
r
d
in
g
to
an
d
.
E
lem
en
t
-
b
ased
m
u
ltip
licatio
n
is
r
ep
r
esen
ted
as
⊗
.
C
o
n
s
id
er
in
g
th
e
co
n
tex
tu
al
-
s
en
s
itiv
ity
g
r
ap
h
,
we
p
r
o
p
o
s
ed
a
DSGC
B
s
eg
m
en
t
t
o
r
eso
lv
e
th
e
is
s
u
e
o
f
u
n
wan
te
d
d
ata
g
ath
er
in
g
b
y
ca
p
t
u
r
in
g
s
em
an
tic
r
elatio
n
s
f
o
r
v
ar
io
u
s
g
lo
b
al
ar
ea
s
.
T
h
is
r
eso
lu
tio
n
in
clu
d
es
two
c
o
n
v
o
lu
tio
n
al
g
r
ap
h
s
,
o
n
e
b
ei
n
g
g
l
o
b
al
an
d
t
h
e
o
t
h
er
o
n
e
b
ein
g
lo
ca
l.
T
h
e
g
lo
b
al
g
r
ap
h
is
r
es
p
o
n
s
ib
le
f
o
r
tak
in
g
c
ar
e
o
f
d
e
p
en
d
e
n
cies o
n
lo
n
g
-
te
r
m
s
en
ten
ce
n
o
d
es.
Ass
u
m
e
is
th
e
p
o
o
led
s
en
ten
c
e
attr
ib
u
te,
th
e
atten
tio
n
co
e
f
f
icien
ts
ar
e
ev
alu
ated
with
d
o
t
p
r
o
d
u
ctio
n
f
o
r
v
ar
io
u
s
n
o
d
es.
C
o
n
s
id
er
as
th
e
n
u
m
b
er
o
f
n
o
d
es
in
th
e
g
l
o
b
al
g
r
ap
h
.
T
h
e
atten
tio
n
m
atr
i
x
is
d
en
o
ted
as
=
(
1
√
.
̃
)
,
h
er
e
=
.
T
h
er
ef
o
r
e,
we
f
o
r
m
u
late
as p
er
g
iv
e
n
(
7
)
.
̃
=
,
(
,
)
×
(
7
)
Her
e,
,
an
d
,
×
ar
e
ex
p
r
ess
ed
as
p
r
o
jectio
n
m
atr
ices.
T
h
e
o
u
tp
u
t
f
o
r
th
e
−
ℎ
h
ea
d
is
ev
alu
ated
as
=
(
,
)
,
wh
er
e
th
e
m
at
r
ix
th
at
is
tr
ain
ab
le
is
g
iv
en
as
,
×
.
T
h
e
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
5
0
2
-
4
7
5
2
I
n
d
o
n
esian
J
E
lec
E
n
g
&
C
o
m
p
Sci
,
Vo
l.
3
9
,
No
.
2
,
Au
g
u
s
t
20
25
:
1
2
4
9
-
1
260
1254
co
n
clu
d
in
g
in
teg
r
ated
r
esu
lt
is
g
iv
en
as
f
o
llo
ws,
with
0
×
as
th
e
attr
ib
u
te
m
atr
ix
as
g
iv
en
in
(
8
)
.
=
[
{
}
=
1
]
0
×
(
8
)
W
h
ile
lo
o
k
in
g
in
to
th
e
lo
ca
l
co
n
v
o
lu
tio
n
al
g
r
ap
h
s
eg
m
e
n
t,
is
d
ev
elo
p
ed
to
ca
ter
to
th
e
lo
ca
l
d
ep
en
d
e
n
cies
c
o
n
s
id
er
i
n
g
t
h
e
d
y
n
am
ic
lo
ca
l
d
ata.
An
atten
tio
n
-
m
ask
in
g
win
d
o
w
is
also
d
e
v
elo
p
ed
×
,
h
av
in
g
a
win
d
o
w
d
im
en
s
io
n
o
f
,
f
o
r
s
h
ap
in
g
th
e
r
elatio
n
b
etwe
en
th
e
n
o
d
es
as
well
as
th
eir
lo
ca
l
lo
ca
tio
n
s
.
A
th
r
esh
o
ld
im
p
lem
en
ted
o
n
th
e
atten
tio
n
h
ea
d
c
o
ef
f
icien
ts
f
o
r
th
e
g
lo
b
a
l
g
r
ap
h
s
eg
m
en
t to
c
h
o
o
s
e
th
e
r
ig
h
t sem
an
tic
n
eig
h
b
o
r
g
lo
b
all
y
,
wh
ich
is
ex
p
r
ess
ed
as g
iv
en
in
(
9
)
.
[
]
,
=
{
[
]
,
,
[
]
,
ℎ
0
,
,
ℎ
(
9
)
Her
e,
th
e
−
ℎ
an
d
−
ℎ
r
o
w
as
well
a
s
th
e
co
lu
m
n
o
f
th
e
atten
tio
n
co
ef
f
icien
t
is
ex
p
r
ess
ed
as
[
]
,
.
L
astl
y
,
th
e
lay
e
r
-
b
ased
d
ata
p
r
o
p
a
g
atio
n
f
o
r
th
e
lo
ca
l
c
o
n
v
o
lu
tio
n
al
g
r
a
p
h
s
eg
m
en
t,
s
i
m
ilar
to
t
h
e
g
lo
b
al
co
n
v
o
l
u
tio
n
al
g
r
ap
h
s
eg
m
en
t
is
g
iv
en
as
g
iv
e
n
in
(
1
0
)
.
W
h
er
e
th
e
tr
ain
a
b
le
p
a
r
am
eter
s
a
r
e
g
iv
en
as
0
∗
an
d
,
∗
.
On
th
e
in
teg
r
atio
n
o
f
t
h
e
g
lo
b
al
as
well
as
th
e
lo
ca
l
co
n
v
o
lu
tio
n
al
g
r
ap
h
s
eg
m
en
ts
,
we
u
s
e
an
attr
ib
u
te
f
u
s
io
n
g
ate
th
at
is
u
s
ed
in
o
b
tain
in
g
t
h
e
c
o
n
clu
d
i
n
g
attr
i
b
u
te
r
e
p
r
esen
tatio
n
f
o
r
s
en
te
n
ce
s
.
T
h
er
e
f
o
r
e,
we
o
b
tain
th
e
(
1
1
)
.
=
(
,
∗
,
∗
)
,
=
1
,
2
,
…
=
[
{
}
=
1
]
0
∗
×
(
1
0
)
=
(
1
[
;
]
+
1
)
,
=
1
,
2
,
…
⃗
⃗
=
⨁
+
(
1
−
)
⨁
(
1
1
)
Her
e,
th
e
o
p
er
atio
n
f
o
r
co
n
ca
ten
atio
n
elem
e
n
t
-
wis
e
is
ex
p
r
ess
ed
as
⨁
,
1
2
×
an
d
1
ar
e
t
h
e
lear
n
in
g
v
ar
ia
b
les.
T
h
e
wo
r
d
s
o
f
th
e
s
am
e
s
ec
tio
n
o
r
s
en
ten
ce
n
o
r
m
ally
h
a
v
e
m
o
r
e
ess
en
tial
d
ata.
T
h
e
lay
er
ed
s
tr
u
ctu
r
e
o
f
d
ata
is
im
p
lem
e
n
ted
f
o
r
in
t
r
a
-
s
ec
tio
n
lear
n
i
n
g
g
r
a
p
h
s
to
en
h
an
ce
th
e
attr
ib
u
te
in
ter
ac
tio
n
.
i
s
u
s
ed
to
r
ep
r
esen
t
th
e
a
d
jace
n
t
m
atr
ix
th
at
r
esu
lts
f
r
o
m
th
e
o
p
er
atio
n
o
f
s
elf
-
atten
tio
n
o
n
th
e
p
o
o
lin
g
wo
r
d
attr
ib
u
te
,
is
in
itially
d
ec
o
u
p
led
as
an
in
tr
a
-
s
ec
tio
n
m
atr
ix
th
at
is
also
ad
jace
n
t
an
d
e
x
p
r
ess
ed
as
in
ter
s
ec
tio
n
m
atr
ix
f
o
r
th
e
len
g
th
o
f
t
h
e
s
ec
tio
n
.
T
h
e
m
ask
i
n
g
o
f
s
en
ten
ce
s
as
well
a
s
s
ec
tio
n
s
ar
e
co
m
b
in
e
d
with
to
f
o
r
m
u
l
ate
th
e
co
n
v
o
l
u
tio
n
al
g
r
a
p
h
f
o
r
in
tr
a
-
s
ec
tio
n
as g
iv
en
in
(
1
2
)
.
=
(
−
+
1
+
−
+
1
)
̂
=
(
(
)
⨁
)
=
(
̂
)
(
1
2
)
W
h
er
e
an
d
ar
e
u
s
ed
to
d
en
o
te
h
y
p
er
p
ar
am
eter
s
.
Fo
r
in
ter
-
s
ec
tio
n
co
n
v
o
lu
tio
n
al
g
r
ap
h
a
s
m
en
tio
n
ed
in
(
1
3
)
.
A
Gate
t
ec
h
n
iq
u
e
is
ap
p
l
ied
b
etwe
en
an
d
th
at
r
esu
lts
in
as g
iv
en
in
(
1
4
)
.
̂
=
(
)
=
(
̂
)
+
(
1
3
)
=
(
2
+
3
)
+
2
′
=
⨁
+
(
1
−
)
⨁
(
1
4
)
Her
e,
th
e
lear
n
in
g
v
ar
iab
les
f
o
r
th
e
g
ate
ar
e
ex
p
r
ess
ed
as
2
,
3
,
an
d
2
.
T
h
e
attr
ib
u
te
in
ter
ac
tio
n
b
etwe
en
th
e
m
ac
r
o
-
s
en
s
itiv
ity
g
r
ap
h
,
co
n
te
x
tu
al
-
s
en
s
itiv
ity
g
r
ap
h
as
well
as
k
ey
wo
r
d
-
s
en
s
itiv
ity
g
r
ap
h
s
h
av
e
to
b
e
in
te
g
r
ate
d
wh
ich
is
p
e
r
f
o
r
m
ed
u
s
in
g
a
C
T
I
FB
.
I
f
a
w
o
r
d
o
r
s
en
ten
ce
n
o
d
e
is
f
r
o
m
a
s
ec
tio
n
th
en
as
g
iv
en
in
(
1
5
)
.
Her
e,
th
e
lear
n
in
g
v
ar
iab
les
o
f
tr
an
s
f
er
f
u
s
io
n
ar
e
d
en
o
ted
as
an
d
.
T
h
e
d
etailed
wo
r
k
in
g
o
f
th
e
p
r
o
p
o
s
ed
m
o
d
el
is
ex
p
lain
ed
in
t
h
e
g
i
v
en
A
lg
o
r
ith
m
1
.
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:
2502
-
4
7
5
2
Dee
p
lea
r
n
in
g
-
b
a
s
ed
mu
lti
-
ti
er sen
s
itivi
ty
a
n
a
lysi
s
n
etw
o
r
k
fo
r
…
(
S
a
d
iy
a
A
n
s
a
r
i
)
1255
=
−
+
1
+
1
+
1
=
(
,
⃗
⃗
)
=
−
+
1
+
1
+
1
=
(
,
′
)
(
1
5
)
Alg
o
r
ith
m
1
.
MT
SAN
Input
Dataset
=
{
(
,
)
}
1
, language model
ℎ
(
∙
,
)
that is prior
-
trained and
,
,
,
,
,
,
,
=
1
,
…
,
Output
Trained model
that categorizes
classes
Step 1
For every iteration
=
1
,
2
,
…
,
do
Step 2
Sample of a batch
from
Step 3
Multi
-
tier sensitivity encoding network (MTSEN) part:
Step 4
Representation through words and sections
Step 5
Sentence representation using (1)
Step 6
S
RMG using (2) and (3)
Step 7
Implement
Multi
-
Tier Sensitivity Analysis Network (MTSAN)
Step 8
For every layer
=
1
,
2
,
…
.
,
do
Step 9
Global convolutional network graph based on sections using (3) and (4)
Step 10
section aided pooling seg
ments for sentences using (5) and (6)
Step 11
Global c
onvolu
tiona
l grap
h
for con
textua
l
-
s
e
ns
i
t
i
v
it
y
g
r
a
p
h
s
e
g
m
e
nt
u
s
in
g
(
7
)
–
(11)
Step 12
Global convolutional graph for keyword
-
sensitivity graph segment using (12)
-
(14)
Step 13
Interactions using transfer
fusion with (15)
Step 14
End for
Step 15
Feature fusion part
Step 16
Feature fusion calculation using equation 16
Step 17
The
Multi
-
Tier Sensitivity Analysis Network (MTSAN)
is evaluated using (17)
Step 18
End For
Step 19
Return
3
.
3
.
I
nte
g
ra
t
ed
s
ens
it
iv
it
y
f
ea
t
ure
f
us
io
n
A
p
o
o
lin
g
o
p
er
atio
n
b
ased
o
n
th
e
co
lu
m
n
is
ap
p
lied
to
g
at
h
e
r
m
o
r
e
d
ata
at
d
if
f
er
e
n
t
lay
er
s
wh
ich
ar
e
ex
p
r
ess
ed
as
(
∙
)
,
th
e
f
in
al
o
u
tp
u
t
is
th
e
r
esu
lt
o
f
th
e
ev
alu
atio
n
as
g
iv
en
in
(
1
6
)
.
T
h
e
co
u
n
t
o
f
lay
er
s
in
th
e
p
r
o
p
o
s
ed
m
o
d
el
is
in
d
icate
d
b
y
.
T
h
en
,
an
o
th
er
m
a
x
p
o
o
li
n
g
o
p
e
r
atio
n
is
p
er
f
o
r
m
e
d
to
in
teg
r
ate
th
e
p
r
i
o
r
f
ea
tu
r
es o
f
th
e
lay
er
s
wh
ich
ar
e
,
,
an
d
wh
ich
is
f
o
r
m
u
lated
as
g
iv
en
b
el
o
w
in
(
1
7
)
.
=
(
)
=
[
(
1
)
,
…
.
,
(
)
)
]
=
[
(
1
)
,
…
.
,
(
)
)
]
(
1
6
)
=
(
[
,
,
]
)
(
1
7
)
4.
P
E
RF
O
RM
A
NCE
E
VA
L
U
AT
I
O
N
T
h
e
p
ap
er
in
tr
o
d
u
ce
s
th
e
MT
SAN,
a
n
o
v
el
f
r
am
ewo
r
k
f
o
r
class
if
y
in
g
d
o
cu
m
e
n
ts
in
to
p
u
b
lic,
p
r
iv
ate,
an
d
co
n
f
id
en
tial
ca
te
g
o
r
ie
s
.
L
ev
er
a
g
in
g
a
d
v
an
ce
d
d
ee
p
lear
n
i
n
g
tech
n
iq
u
es,
in
cl
u
d
in
g
c
o
n
v
o
lu
tio
n
al
n
etwo
r
k
s
an
d
GC
N
s
,
MT
SA
N
ca
p
tu
r
es
b
o
th
lo
ca
l
an
d
g
lo
b
al
co
n
tex
tu
al
in
f
o
r
m
atio
n
.
I
t
in
co
r
p
o
r
ates
s
ev
er
al
k
ey
co
m
p
o
n
en
ts
,
s
u
ch
as
th
e
MT
SEN,
DSG
C
B
,
an
d
C
T
I
FB
,
to
en
h
an
ce
f
ea
tu
r
e
r
ep
r
esen
tatio
n
an
d
class
if
icatio
n
ac
cu
r
ac
y
.
T
h
e
p
r
o
p
o
s
ed
s
y
s
tem
was
ev
alu
ate
d
o
n
d
atasets
lik
e
th
e
2
0
n
ewsg
r
o
u
p
s
,
e
n
r
o
n
em
ail
,
an
d
MI
MI
C
-
III
clin
ical
d
ata
b
ase
,
d
em
o
n
s
tr
atin
g
s
u
p
er
i
o
r
p
er
f
o
r
m
a
n
ce
o
v
e
r
tr
ad
itio
n
al
m
o
d
els
lik
e
s
u
p
p
o
r
t
v
ec
to
r
m
ac
h
in
e
(
SVM
)
an
d
c
o
n
tem
p
o
r
ar
y
d
ee
p
lear
n
in
g
m
o
d
els
lik
e
B
E
R
T
.
T
h
is
s
tu
d
y
ad
d
r
ess
es
ex
is
tin
g
g
ap
s
in
s
en
s
itiv
e
in
f
o
r
m
atio
n
class
if
ica
tio
n
,
p
ar
ticu
lar
ly
f
o
r
co
m
p
lex
a
n
d
le
n
g
th
y
d
o
c
u
m
e
n
ts
,
b
y
p
r
o
v
i
d
in
g
a
m
o
r
e
r
o
b
u
s
t a
n
d
ef
f
icien
t c
lass
if
icatio
n
s
y
s
tem
.
4
.
1
.
Da
t
a
s
et
d
et
a
ils
a
nd
co
mp
a
riso
n m
ec
ha
nis
m
4
.
1
.
1
.
20
newsg
ro
up
s
da
t
a
s
et
T
h
is
d
ataset
co
n
s
is
ts
o
f
ar
o
u
n
d
2
0
,
0
0
0
n
ewsg
r
o
u
p
d
o
cu
m
en
ts
,
p
a
r
titi
o
n
ed
ac
r
o
s
s
2
0
d
if
f
er
e
n
t
n
ewsg
r
o
u
p
s
,
c
o
v
er
i
n
g
a
wid
e
r
an
g
e
o
f
t
o
p
ics.
T
h
e
v
ast
m
aj
o
r
ity
o
f
th
e
co
n
te
n
t
in
th
e
2
0
New
s
g
r
o
u
p
s
d
ataset
ca
n
b
e
class
if
ied
as
p
u
b
lic
.
T
h
ese
ar
e
d
is
cu
s
s
io
n
s
f
r
o
m
v
ar
i
o
u
s
n
ewsg
r
o
u
p
s
,
co
v
e
r
in
g
a
wid
e
ar
r
ay
o
f
to
p
ics,
in
ten
d
ed
f
o
r
p
u
b
lic
v
iewin
g
an
d
s
h
ar
in
g
.
T
h
er
e
is
m
in
im
al
r
is
k
ass
o
ciate
d
with
th
is
d
ata
b
ein
g
p
u
b
licly
ac
ce
s
s
ib
le,
as th
e
co
n
ten
t is m
ea
n
t to
b
e
o
p
en
t
o
all
u
s
er
s
o
f
t
h
e
n
ewsg
r
o
u
p
s
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
5
0
2
-
4
7
5
2
I
n
d
o
n
esian
J
E
lec
E
n
g
&
C
o
m
p
Sci
,
Vo
l.
3
9
,
No
.
2
,
Au
g
u
s
t
20
25
:
1
2
4
9
-
1
260
1256
4
.
1
.
2
.
E
nro
n
em
a
il da
t
a
s
e
t
Ma
n
y
em
ails
in
t
h
is
d
ataset
ca
n
b
e
co
n
s
id
er
e
d
p
r
iv
ate
.
As
th
ey
in
v
o
lv
e
in
ter
n
al
co
r
p
o
r
ate
co
m
m
u
n
icatio
n
s
th
at
ar
e
n
o
t
m
ea
n
t
f
o
r
p
u
b
lic
d
is
clo
s
u
r
e.
T
h
ey
d
o
n
o
t
n
ec
ess
ar
ily
c
o
n
t
ain
h
ig
h
l
y
s
en
s
itiv
e
in
f
o
r
m
atio
n
.
4
.
1
.
3
.
M
I
M
I
C
-
I
I
I
clinica
l da
t
a
ba
s
e
T
h
e
clin
ical
n
o
tes
an
d
m
ed
ica
l
d
ata
in
th
e
MI
MI
C
-
III
d
atab
ase
ar
e
h
ig
h
ly
co
n
f
id
en
tial.
T
h
is
d
ataset
in
clu
d
es
s
en
s
itiv
e
p
atien
t
in
f
o
r
m
atio
n
,
m
ed
ical
h
is
to
r
ies,
d
i
ag
n
o
s
es,
an
d
tr
ea
tm
en
t
p
la
n
s
.
T
h
e
HI
PAA
in
th
e
Un
ited
States
,
f
o
r
in
s
tan
ce
,
m
a
n
d
ates stric
t c
o
n
f
id
en
tiality
f
o
r
s
u
ch
d
ata
to
p
r
o
tect
p
atien
t p
r
iv
ac
y
.
4
.
1
.
4
.
Su
pp
o
rt
v
ec
t
o
r
m
a
chi
ne
SVMs
ar
e
a
p
o
p
u
lar
m
ac
h
i
n
e
-
lear
n
in
g
tech
n
iq
u
e
u
s
ed
f
o
r
cl
ass
if
icatio
n
task
s
.
T
h
ey
w
o
r
k
b
y
f
i
n
d
in
g
th
e
o
p
tim
al
h
y
p
e
r
p
lan
e
th
at
s
ep
ar
ates
d
if
f
er
e
n
t
class
es
in
a
h
ig
h
-
d
im
en
s
io
n
al
s
p
ac
e
.
I
n
t
h
e
co
n
tex
t
o
f
d
o
cu
m
e
n
t
class
if
icatio
n
,
SV
Ms
ar
e
ef
f
ec
tiv
e
d
u
e
to
th
eir
ab
ilit
y
to
h
an
d
le
s
p
ar
s
e
d
ata
a
n
d
h
ig
h
-
d
im
en
s
io
n
al
f
ea
tu
r
e
s
p
ac
es,
ty
p
ical
in
te
x
t
-
b
ased
ap
p
licatio
n
s
.
4
.
1
.
5
.
B
idi
re
ct
io
na
l
enco
der
re
presenta
t
io
ns
f
ro
m
t
ra
ns
f
o
rm
er
s
B
E
R
T
,
a
s
tate
-
of
-
th
e
-
ar
t
d
ee
p
lear
n
in
g
m
o
d
el
f
o
r
NL
P,
lev
er
ag
es
th
e
tr
an
s
f
o
r
m
e
r
ar
ch
itectu
r
e.
B
E
R
T
is
p
r
etr
ain
ed
o
n
v
ast
a
m
o
u
n
ts
o
f
te
x
t
d
ata
an
d
f
in
e
-
tu
n
ed
f
o
r
s
p
ec
if
ic
task
s
,
m
ak
i
n
g
it
ex
ce
p
tio
n
ally
g
o
o
d
at
u
n
d
er
s
tan
d
in
g
co
n
tex
t
an
d
s
em
an
tics
.
Un
lik
e
tr
a
d
itio
n
al
m
o
d
els,
B
E
R
T
r
ea
d
s
te
x
t
b
i
-
d
ir
ec
tio
n
ally
,
co
n
s
id
er
in
g
th
e
c
o
n
tex
t
f
r
o
m
b
o
th
p
r
ec
e
d
in
g
an
d
f
o
llo
wi
n
g
wo
r
d
s
in
a
s
en
ten
ce
,
th
u
s
ca
p
tu
r
in
g
r
ich
er
in
f
o
r
m
atio
n
4
.
2
.
Resul
t
s
Fig
u
r
e
2
co
m
p
a
r
es
th
e
o
v
er
all
ac
cu
r
ac
y
o
f
th
r
ee
d
if
f
e
r
en
t
m
o
d
els
;
SVM,
B
E
R
T
,
an
d
th
e
MT
SAN
in
class
if
y
in
g
d
o
cu
m
e
n
ts
in
to
p
u
b
li
c,
p
r
iv
ate,
a
n
d
co
n
f
id
en
tial
ca
teg
o
r
ies.
T
h
e
SVM
m
o
d
el
a
ch
iev
es a
n
ac
cu
r
ac
y
o
f
ar
o
u
n
d
8
4
%,
i
n
d
icatin
g
its
b
asic
ca
p
ab
ilit
y
to
p
e
r
f
o
r
m
th
e
class
if
icatio
n
task
,
th
o
u
g
h
it
s
tr
u
g
g
les
with
m
o
r
e
co
m
p
lex
d
is
tin
ctio
n
s
.
T
h
e
B
E
R
T
m
o
d
el,
lev
e
r
ag
in
g
d
ee
p
l
ea
r
n
in
g
tech
n
iq
u
es
an
d
tr
an
s
f
o
r
m
er
a
r
ch
itectu
r
e,
s
ig
n
if
ican
tly
im
p
r
o
v
es
ac
cu
r
a
cy
to
ap
p
r
o
x
im
ately
9
1
%,
d
e
m
o
n
s
tr
atin
g
a
b
etter
u
n
d
er
s
ta
n
d
in
g
o
f
n
u
a
n
ce
d
lan
g
u
ag
e
an
d
co
n
tex
t.
T
h
e
M
T
SAN
o
u
tp
er
f
o
r
m
s
b
o
th
SVM
an
d
B
E
R
T
,
ac
h
iev
in
g
a
n
ac
c
u
r
ac
y
o
f
a
b
o
u
t
9
6
%.
T
h
is
s
u
g
g
ests
th
at
th
e
PS
m
o
d
el
in
co
r
p
o
r
ates
ad
v
an
ce
d
f
ea
tu
r
es
o
r
o
p
tim
izatio
n
s
,
m
a
k
in
g
it
p
ar
ticu
lar
l
y
ef
f
ec
tiv
e
at
ac
cu
r
ately
ca
teg
o
r
izin
g
d
o
c
u
m
en
ts
ac
r
o
s
s
d
if
f
e
r
en
t
s
en
s
itiv
ity
lev
els.
T
h
e
s
u
p
er
io
r
p
e
r
f
o
r
m
an
ce
o
f
th
e
PS
m
o
d
el
h
ig
h
lig
h
ts
its
p
o
ten
tial
as
a
r
eliab
le
to
o
l
f
o
r
s
en
s
itiv
e
d
ata
class
if
icat
io
n
,
alig
n
in
g
well
to
en
s
u
r
e
p
r
ec
is
e
id
en
tific
atio
n
a
n
d
h
an
d
lin
g
o
f
s
en
s
itiv
e
in
f
o
r
m
atio
n
.
Fig
u
r
e
2
.
C
o
m
p
a
r
es th
e
o
v
er
al
l a
cc
u
r
ac
y
o
f
th
r
ee
d
if
f
er
e
n
t m
o
d
els
I
llu
s
tr
ates
th
e
F1
-
s
co
r
e
f
o
r
p
u
b
lic
d
ata
class
if
icatio
n
ac
h
iev
e
d
b
y
th
r
ee
m
o
d
els:
SVM,
B
E
R
T
,
an
d
th
e
MT
SAN.
T
h
e
SVM
m
o
d
e
l
s
co
r
es
ar
o
u
n
d
8
4
.
5
%,
i
n
d
icat
in
g
a
r
elativ
el
y
b
alan
ce
d
p
r
ec
i
s
io
n
an
d
r
ec
all
b
u
t
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:
2502
-
4
7
5
2
Dee
p
lea
r
n
in
g
-
b
a
s
ed
mu
lti
-
ti
er sen
s
itivi
ty
a
n
a
lysi
s
n
etw
o
r
k
fo
r
…
(
S
a
d
iy
a
A
n
s
a
r
i
)
1257
with
lim
itatio
n
s
in
h
a
n
d
lin
g
t
h
e
co
m
p
lex
ity
o
f
p
u
b
lic
d
ata
co
n
ten
t.
T
h
e
B
E
R
T
m
o
d
el
im
p
r
o
v
es
u
p
o
n
t
h
is
,
ac
h
iev
in
g
ap
p
r
o
x
i
m
ately
9
1
.
5
%,
b
en
ef
itin
g
f
r
o
m
its
ab
ilit
y
t
o
u
n
d
er
s
tan
d
co
n
tex
t
an
d
s
em
an
tics
th
r
o
u
g
h
d
ee
p
lear
n
in
g
.
T
h
e
MT
SAN
ac
h
iev
es
th
e
h
ig
h
est
F1
-
s
co
r
e
at
ar
o
u
n
d
9
6
.
5
%,
s
h
o
wca
s
in
g
its
s
u
p
er
io
r
ca
p
a
b
ilit
y
in
ac
cu
r
ately
class
if
y
in
g
p
u
b
lic
d
o
c
u
m
en
ts
.
T
h
is
h
ig
h
p
er
f
o
r
m
an
ce
s
u
g
g
ests
th
at
th
e
PS
m
o
d
el
ef
f
ec
ti
v
ely
b
alan
ce
s
p
r
ec
is
io
n
an
d
r
ec
all,
m
in
im
izin
g
b
o
th
f
alse
p
o
s
itiv
es
an
d
f
alse
n
eg
ativ
es.
T
h
e
r
esu
lts
h
ig
h
lig
h
t
th
e
PS
m
o
d
el'
s
ad
v
an
ce
d
p
r
o
ce
s
s
in
g
ca
p
a
b
ilit
ies,
m
ak
in
g
it
th
e
m
o
s
t
r
eliab
le
ch
o
ice
f
o
r
ca
teg
o
r
izin
g
p
u
b
li
c
d
ata
with
in
th
is
s
et
o
f
m
o
d
els.
Fig
u
r
e
3
co
m
p
ar
es th
e
F1
-
s
co
r
e
o
f
p
u
b
lic
d
ata
f
o
r
th
r
ee
d
if
f
e
r
en
t
m
o
d
els.
Fig
u
r
e
3
.
C
o
m
p
a
r
es th
e
F1
-
s
c
o
r
e
o
f
p
u
b
lic
d
ata
f
o
r
th
r
ee
d
if
f
er
en
t m
o
d
els
T
h
e
b
ar
c
h
ar
t
d
is
p
lay
s
th
e
F1
-
s
co
r
e
f
o
r
class
if
y
in
g
p
r
iv
ate
d
ata
ac
r
o
s
s
th
r
ee
m
o
d
els:
SV
M,
B
E
R
T
,
an
d
th
e
MT
SAN.
T
h
e
SV
M
m
o
d
el
s
h
o
ws
an
F1
-
s
co
r
e
o
f
ap
p
r
o
x
im
ately
8
2
.
5
%,
r
ef
lectin
g
m
o
d
e
r
ate
p
er
f
o
r
m
an
ce
in
b
alan
cin
g
p
r
e
cisi
o
n
an
d
r
ec
all
f
o
r
p
r
iv
ate
d
ata,
b
u
t
with
s
o
m
e
s
h
o
r
tco
m
in
g
s
lik
ely
d
u
e
to
its
less
co
m
p
lex
h
an
d
lin
g
o
f
n
u
an
ce
s
in
th
e
d
ata.
B
E
R
T
,
lev
er
ag
in
g
a
d
v
an
ce
d
NL
P
ca
p
ab
ilit
ies,
ac
h
iev
es
an
im
p
r
o
v
e
d
F1
-
s
co
r
e
o
f
ar
o
u
n
d
8
9
.
5
%,
in
d
icatin
g
a
m
o
r
e
ac
c
u
r
ate
class
if
icatio
n
o
f
p
r
iv
ate
d
o
cu
m
e
n
ts
,
th
an
k
s
to
its
ab
ilit
y
to
u
n
d
er
s
tan
d
th
e
co
n
t
ex
t
an
d
in
tr
icate
d
etails.
T
h
e
MT
SAN
lead
s
wi
th
th
e
h
ig
h
est
F1
-
s
co
r
e
at
ap
p
r
o
x
im
ately
9
4
.
5
%,
d
em
o
n
s
tr
atin
g
ex
ce
p
tio
n
al
p
r
ec
is
io
n
a
n
d
r
ec
all
i
n
id
e
n
tify
in
g
p
r
i
v
ate
in
f
o
r
m
atio
n
.
T
h
is
r
esu
lt
in
d
icate
s
th
at
th
e
PS
m
o
d
el
is
p
ar
ticu
lar
ly
ef
f
e
ctiv
e
in
m
an
ag
in
g
th
e
co
m
p
lex
ities
o
f
p
r
iv
ate
d
ata,
en
s
u
r
in
g
a
h
ig
h
lev
el
o
f
ac
cu
r
ac
y
in
class
if
icatio
n
,
th
u
s
o
f
f
er
in
g
th
e
m
o
s
t
r
eliab
le
p
er
f
o
r
m
an
ce
am
o
n
g
th
e
ev
alu
ated
m
o
d
els.
Fig
u
r
e
4
co
m
p
ar
es th
e
F1
-
s
co
r
e
o
f
p
r
iv
ate
d
ata
f
o
r
th
r
ee
d
i
f
f
er
en
t
m
o
d
el
s
.
Fig
u
r
e
4
.
C
o
m
p
a
r
es th
e
F1
-
s
c
o
r
e
o
f
p
r
iv
ate
d
ata
f
o
r
th
r
ee
d
if
f
er
en
t m
o
d
els
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
5
0
2
-
4
7
5
2
I
n
d
o
n
esian
J
E
lec
E
n
g
&
C
o
m
p
Sci
,
Vo
l.
3
9
,
No
.
2
,
Au
g
u
s
t
20
25
:
1
2
4
9
-
1
260
1258
Fig
u
r
e
5
illu
s
tr
ates
th
e
F1
-
s
co
r
e
f
o
r
th
e
class
if
icatio
n
o
f
co
n
f
id
en
tial
d
ata
ac
r
o
s
s
th
r
ee
m
o
d
els:
SVM,
B
E
R
T
,
an
d
th
e
MT
SA
N.
T
h
e
SVM
m
o
d
el
h
as
th
e
lo
we
s
t
F1
-
s
co
r
e
at
ap
p
r
o
x
im
ately
7
9
%,
s
u
g
g
e
s
tin
g
a
s
tr
u
g
g
le
with
ac
cu
r
ately
id
e
n
tify
in
g
a
n
d
d
is
tin
g
u
is
h
in
g
co
n
f
id
en
tial
in
f
o
r
m
ati
o
n
,
p
o
s
s
ib
ly
d
u
e
to
its
lim
itatio
n
s
in
h
an
d
lin
g
th
e
s
u
b
tleties
o
f
s
en
s
itiv
e
co
n
ten
t.
T
h
e
B
E
R
T
m
o
d
el
im
p
r
o
v
es
o
n
th
is
with
an
F1
-
s
co
r
e
o
f
a
r
o
u
n
d
8
8
%,
b
e
n
ef
itin
g
f
r
o
m
its
d
ee
p
lear
n
in
g
ar
c
h
itectu
r
e
th
at
b
etter
ca
p
tu
r
es
co
n
tex
t
an
d
s
em
an
tic
n
u
an
ce
s
,
th
u
s
en
h
an
cin
g
its
p
er
f
o
r
m
a
n
ce
in
class
if
y
in
g
co
n
f
id
en
tial
d
o
cu
m
e
n
ts
.
T
h
e
MT
SAN
ac
h
iev
es
th
e
h
ig
h
est
F1
-
s
co
r
e
at
a
b
o
u
t
9
3
.
5
%,
in
d
icatin
g
s
u
p
er
i
o
r
p
r
ec
is
io
n
an
d
r
ec
all.
T
h
is
s
u
g
g
es
ts
th
at
th
e
PS
m
o
d
el
is
h
ig
h
ly
ef
f
ec
tiv
e
at
m
an
ag
i
n
g
th
e
co
m
p
lex
ities
in
h
er
en
t
in
co
n
f
id
en
tial
d
ata,
en
s
u
r
in
g
a
m
o
r
e
ac
cu
r
ate
an
d
r
eliab
le
class
if
icatio
n
.
T
h
e
h
ig
h
p
er
f
o
r
m
an
ce
o
f
th
e
PS
m
o
d
el
h
ig
h
lig
h
ts
its
ca
p
ab
ilit
y
to
m
ain
tain
a
b
alan
ce
b
etwe
en
co
r
r
ec
tl
y
id
en
tify
i
n
g
co
n
f
id
en
tial
in
f
o
r
m
atio
n
an
d
m
in
im
izin
g
f
alse
p
o
s
itiv
es
an
d
n
eg
ativ
es,
m
ak
in
g
it th
e
m
o
s
t e
f
f
ec
tiv
e
m
o
d
el
am
o
n
g
th
o
s
e
ev
alu
ated
f
o
r
th
is
ta
s
k
.
Fig
u
r
e
5
.
F1
-
s
co
r
e
f
o
r
th
e
clas
s
if
icatio
n
o
f
co
n
f
id
en
tial d
ata
4
.
3
.
Co
m
pa
ra
t
iv
e
a
na
ly
s
is
T
h
e
co
m
p
ar
ativ
e
a
n
aly
s
is
d
e
m
o
n
s
tr
ates
th
at
th
e
p
r
o
p
o
s
ed
MT
SAN
s
ig
n
if
ican
tly
o
u
tp
e
r
f
o
r
m
s
b
o
t
h
tr
ad
itio
n
al
m
ac
h
in
e
lear
n
in
g
m
o
d
els,
s
u
ch
as
SVM,
an
d
ad
v
an
ce
d
d
ee
p
lear
n
in
g
m
o
d
els
l
ik
e
B
E
R
T
.
MT
SA
N
ac
h
iev
ed
a
h
i
g
h
er
ac
c
u
r
ac
y
r
ate
o
f
ap
p
r
o
x
im
ately
9
6
%,
c
o
m
p
ar
ed
to
8
4
%
f
o
r
SVM
an
d
9
1
%
f
o
r
B
E
R
T
,
s
h
o
wca
s
in
g
its
s
u
p
er
io
r
ca
p
a
b
ilit
y
in
co
r
r
ec
tly
ca
teg
o
r
izin
g
s
en
s
itiv
e
in
f
o
r
m
atio
n
i
n
to
p
u
b
lic,
p
r
iv
ate,
a
n
d
co
n
f
id
en
tial
ca
teg
o
r
ies.
T
h
e
F
1
-
s
co
r
es
also
f
av
o
r
ed
MT
SA
N,
p
ar
ticu
lar
ly
in
h
an
d
lin
g
c
o
m
p
lex
d
is
tin
ctio
n
s
in
co
n
f
i
d
en
tial
d
ata,
wh
er
e
it
r
ea
ch
ed
9
3
.
5
%,
co
m
p
ar
e
d
to
B
E
R
T
'
s
8
8
%.
Key
im
p
r
o
v
em
en
ts
o
v
er
B
E
R
T
in
clu
d
e
th
e
in
tr
o
d
u
ctio
n
o
f
th
e
MT
SE
N,
wh
ich
p
r
o
ce
s
s
es
m
u
lti
-
lev
el
d
o
cu
m
e
n
t
f
ea
t
u
r
es
s
u
ch
as
s
ec
tio
n
s
,
s
en
ten
ce
s
,
an
d
wo
r
d
s
,
o
f
f
er
in
g
a
m
o
r
e
n
u
an
ce
d
u
n
d
e
r
s
tan
d
in
g
o
f
d
o
c
u
m
en
t
s
tr
u
ctu
r
e.
T
h
e
DSGC
B
en
h
an
ce
s
s
em
an
tic
r
elatio
n
s
h
ip
d
etec
tio
n
b
y
ca
p
t
u
r
in
g
b
o
th
g
lo
b
al
d
ep
e
n
d
en
ci
es
an
d
lo
ca
l
d
y
n
am
ics,
s
u
r
p
ass
in
g
B
E
R
T
'
s
to
k
en
-
lev
el
atten
tio
n
m
ec
h
an
is
m
.
Ad
d
itio
n
ally
,
th
e
C
T
I
FB
in
MT
SAN
ef
f
ec
tiv
ely
in
teg
r
ates
m
u
lti
-
lev
el
f
ea
tu
r
es,
a
ca
p
ab
ilit
y
n
o
t
p
r
esen
t
in
B
E
R
T
'
s
ar
ch
itectu
r
e,
th
u
s
lev
er
ag
in
g
b
o
th
m
ac
r
o
an
d
m
icr
o
-
s
tr
u
ctu
r
al
in
f
o
r
m
atio
n
f
o
r
m
o
r
e
ac
cu
r
ate
class
if
icatio
n
.
Fu
r
th
er
m
o
r
e,
MT
SAN
ad
d
r
ess
es
th
e
ch
allen
g
e
o
f
h
an
d
lin
g
lo
n
g
a
n
d
co
m
p
lex
d
o
cu
m
e
n
ts
m
o
r
e
ef
f
i
cien
tly
th
an
B
E
R
T
,
wh
ich
ca
n
s
tr
u
g
g
le
with
len
g
th
y
in
p
u
ts
d
u
e
to
its
q
u
ad
r
atic
co
m
p
lex
ity
.
T
h
e
s
ec
tio
n
-
ai
d
ed
p
o
o
lin
g
an
d
f
ea
tu
r
e
f
u
s
io
n
m
ec
h
an
is
m
s
in
MT
SAN
f
ac
ili
ta
te
th
e
m
an
ag
em
en
t
o
f
d
o
c
u
m
en
ts
with
m
ix
e
d
co
n
ten
t
ty
p
es
an
d
v
ar
y
in
g
s
en
s
i
tiv
ity
lev
els.
Ov
er
all,
th
ese
e
n
h
an
ce
m
e
n
ts
m
ak
e
MT
SAN
a
m
o
r
e
r
eliab
le
an
d
ef
f
ec
tiv
e
t
o
o
l
f
o
r
s
en
s
itiv
e
in
f
o
r
m
atio
n
class
if
icatio
n
,
p
r
o
v
id
in
g
a
co
m
p
r
eh
e
n
s
iv
e
an
d
ef
f
icien
t s
o
lu
tio
n
f
o
r
m
a
n
ag
in
g
s
en
s
itiv
e
d
ata.
5.
CO
NCLU
SI
O
N
T
h
e
p
r
o
p
o
s
ed
MT
SAN
p
r
esen
ts
a
r
o
b
u
s
t
an
d
ef
f
ec
tiv
e
s
o
lu
tio
n
f
o
r
class
if
y
in
g
d
o
cu
m
en
ts
in
to
p
u
b
lic,
p
r
i
v
ate,
an
d
c
o
n
f
i
d
en
tial
ca
teg
o
r
ies
.
B
y
lev
er
ag
in
g
ad
v
an
ce
d
d
ee
p
lea
r
n
in
g
te
ch
n
iq
u
es,
in
clu
d
in
g
co
n
v
o
l
u
tio
n
al
n
etwo
r
k
s
an
d
GC
N
s
,
MT
SAN
s
u
cc
es
s
f
u
lly
ca
p
tu
r
es
b
o
th
lo
ca
l
an
d
g
lo
b
al
co
n
tex
tu
al
Evaluation Warning : The document was created with Spire.PDF for Python.