I
nte
rna
t
io
na
l J
o
urna
l o
f
E
lect
rica
l a
nd
Co
m
p
ute
r
E
ng
in
ee
ring
(
I
J
E
CE
)
Vo
l.
10
,
No
.
4
,
A
u
g
u
s
t
2
0
2
0
,
p
p
.
3
7
3
4
~
3
7
4
2
I
SS
N:
2
0
8
8
-
8708
,
DOI
: 1
0
.
1
1
5
9
1
/
i
j
ec
e
.
v
1
0
i
4
.
p
p
3
7
3
4
-
3
7
4
2
3734
J
o
ur
na
l ho
m
ep
a
g
e
:
h
ttp
:
//ij
ec
e.
ia
esco
r
e.
co
m/in
d
ex
.
p
h
p
/I
JE
C
E
Tex
t
cla
ss
ificatio
n bas
ed on g
a
ted
recurren
t
un
it
co
m
bin
es w
ith
supp
o
rt
v
ector
ma
chine
M
uh
a
mm
a
d
Z
u
lqa
rna
in,
Ro
za
ida
G
ha
za
li,
Ya
na
M
a
z
w
in M
o
hm
a
d H
a
s
s
i
m
,
M
uh
a
m
m
a
d Re
ha
n
F
a
c
u
lt
y
o
f
Co
m
p
u
ter S
c
ien
c
e
a
n
d
In
f
o
rm
a
ti
o
n
T
e
c
h
n
o
l
o
g
y
,
Un
iv
e
rs
it
i
T
u
n
Hu
ss
e
in
On
n
M
a
la
y
sia
,
M
a
la
y
sia
Art
icle
I
nfo
AB
ST
RAC
T
A
r
ticle
his
to
r
y:
R
ec
eiv
ed
J
an
22
,
2
0
1
9
R
ev
i
s
ed
J
an
18
,
2
0
2
0
A
cc
ep
ted
Feb
1
,
2
0
2
0
A
s
th
e
a
m
o
u
n
t
o
f
u
n
str
u
c
tu
re
d
t
e
x
t
d
a
ta
th
a
t
h
u
m
a
n
it
y
p
ro
d
u
c
e
l
a
rg
e
l
y
a
n
d
a
lo
t
o
f
te
x
ts
a
r
e
g
ro
w
s
o
n
th
e
In
t
e
rn
e
t,
so
th
e
o
n
e
o
f
th
e
in
telli
g
e
n
t
tec
h
n
iq
u
e
is
re
q
u
ire
p
r
o
c
e
ss
in
g
it
a
n
d
e
x
trac
ti
n
g
d
if
f
e
r
e
n
t
ty
p
e
s
o
f
k
n
o
w
led
g
e
f
ro
m
it
.
G
a
t
e
d
re
c
u
rre
n
t
u
n
it
(G
RU)
a
n
d
su
p
p
o
rt
v
e
c
to
r
m
a
c
h
in
e
(S
V
M
)
h
a
v
e
b
e
e
n
su
c
c
e
ss
f
u
ll
y
u
se
d
to
Na
tu
ra
l
Lan
g
u
a
g
e
P
ro
c
e
ss
in
g
(NL
P
)
s
y
st
e
m
s
w
it
h
c
o
m
p
a
ra
ti
v
e
,
re
m
a
r
k
a
b
le
re
su
lt
s.
G
RU
n
e
t
w
o
rk
s
p
e
r
f
o
r
m
w
e
ll
in
se
q
u
e
n
ti
a
l
lea
rn
in
g
tas
k
s
a
n
d
o
v
e
rc
o
m
e
t
h
e
issu
e
s
o
f
“
v
a
n
ish
in
g
a
n
d
e
x
p
lo
si
o
n
o
f
g
ra
d
ien
ts
in
sta
n
d
a
rd
re
c
u
rre
n
t
n
e
u
ra
l
n
e
tw
o
rk
s
(RNN
s)
w
h
e
n
c
a
p
tu
re
in
g
lo
n
g
-
term
d
e
p
e
n
d
e
n
c
ies
.
In
th
is
p
a
p
e
r,
w
e
p
ro
p
o
se
d
a
tex
t
c
la
ss
if
ic
a
ti
o
n
m
o
d
e
l
b
a
se
d
o
n
i
m
p
ro
v
e
d
a
p
p
ro
a
c
h
e
s
to
th
is
n
o
rm
b
y
p
re
se
n
ti
n
g
a
li
n
e
a
r
su
p
p
o
rt
v
e
c
to
r
m
a
c
h
in
e
(S
V
M
)
a
s
th
e
re
p
lac
e
m
e
n
t
o
f
S
o
f
t
m
a
x
in
th
e
f
in
a
l
o
u
t
p
u
t
lay
e
r
o
f
a
G
R
U
m
o
d
e
l.
F
u
rth
e
rm
o
re
,
th
e
c
ro
ss
-
e
n
tro
p
y
f
u
n
c
ti
o
n
sh
a
ll
b
e
re
p
lac
e
d
w
it
h
a
m
a
r
g
i
n
-
b
a
se
d
f
u
n
c
ti
o
n
.
Em
p
iri
c
a
l
re
su
lt
s
p
re
se
n
t
th
a
t
th
e
p
ro
p
o
se
d
G
RU
-
S
V
M
m
o
d
e
l
a
c
h
iev
e
d
c
o
m
p
a
ra
ti
v
e
l
y
b
e
tt
e
r
re
su
lt
s
th
a
n
th
e
b
a
se
li
n
e
a
p
p
ro
a
c
h
e
s BL
S
T
M
-
C,
DA
BN.
K
ey
w
o
r
d
s
:
Dee
p
lear
n
in
g
GR
U
Su
p
p
o
r
t v
ec
to
r
m
ac
h
i
n
e
T
ex
t c
lass
if
icatio
n
Co
p
y
rig
h
t
©
2
0
2
0
In
st
it
u
te o
f
A
d
v
a
n
c
e
d
E
n
g
i
n
e
e
rin
g
a
n
d
S
c
ien
c
e
.
Al
l
rig
h
ts re
se
rv
e
d
.
C
o
r
r
e
s
p
o
nd
ing
A
uth
o
r
:
R
o
za
id
a
Gh
az
ali,
Facu
lt
y
o
f
C
o
m
p
u
ter
Scien
ce
an
d
I
n
f
o
r
m
atio
n
T
ec
h
n
o
lo
g
y
,
Un
i
v
er
s
iti T
u
n
H
u
s
s
ei
n
On
n
Ma
la
y
s
ia,
”
8
6
4
0
0
,
P
ar
it R
aj
a,
B
atu
P
ah
at,
J
o
h
o
r
,
Ma
lay
s
ia.
E
m
ail:
r
o
za
id
a@
u
th
m
.
ed
u
.
m
y
1.
I
NT
RO
D
UCT
I
O
N
W
ith
th
e
q
u
ick
l
y
a
n
d
co
n
ti
n
u
o
u
s
p
r
o
d
u
ctio
n
o
f
i
n
f
o
r
m
atio
n
s
tec
h
n
o
lo
g
ies
i
n
d
ig
it
al
f
o
r
m
at,
th
e
r
ap
id
l
y
in
cr
ea
s
in
g
i
n
th
e
n
u
m
b
er
o
f
elec
tr
o
n
ic
te
x
t
m
a
s
s
a
g
e,
o
n
th
e
w
o
r
ld
w
id
e
w
e
b
(
www
)
,
s
u
c
h
as
i
m
p
o
r
tan
t
n
e
w
s
s
i
tes,
f
o
r
u
m
s
,
b
lo
g
s
an
d
s
o
o
n
tex
t
class
if
ica
t
io
n
h
as
b
ec
o
m
e
v
er
y
s
er
io
u
s
p
r
o
b
lem
f
o
r
th
e
b
ig
o
r
g
an
izatio
n
a
n
d
co
m
p
an
ies
t
o
m
a
n
ag
e
an
d
o
r
g
a
n
ize
th
e
t
ex
tu
a
l
d
ata.
T
ex
t
cla
s
s
i
f
icat
io
n
i
s
a
tech
n
iq
u
e
o
f
class
i
f
y
in
g
u
n
o
r
g
a
n
ized
te
x
t
a
u
to
m
at
icall
y
ac
co
r
d
in
g
to
th
e
p
r
ed
ef
in
e
d
class
e
s
o
r
ca
teg
o
r
ies
b
ased
o
n
th
e
te
x
t
co
n
ten
t
s
th
r
o
u
g
h
a
g
iv
e
n
clas
s
if
icatio
n
s
y
s
te
m
.
No
w
ad
a
y
s
,
T
ex
t
class
if
ica
tio
n
b
ec
o
m
es
a
n
i
m
p
o
r
tan
t
ta
s
k
f
o
r
au
to
m
at
icall
y
c
lass
if
y
i
n
g
t
h
e
d
o
cu
m
en
t
s
to
th
eir
r
esp
ec
t
iv
e
ca
teg
o
r
ies.
F
u
r
t
h
er
m
o
r
e,
tex
t
clas
s
i
f
icatio
n
t
ec
h
n
iq
u
e
h
a
v
e
to
u
s
e
in
m
a
n
y
ap
p
licatio
n
s
o
f
N
L
P
,
s
u
c
h
as
s
p
a
m
f
ilter
i
n
g
[
1
]
,
u
s
er
i
n
t
en
tio
n
an
al
y
s
is
[
2
]
,
tex
t
clas
s
i
f
icatio
n
[3
,
4]
,
p
e
r
s
o
n
alize
d
n
e
w
s
r
ec
o
m
m
e
n
d
atio
n
[
5
]
,
em
ail
ca
te
g
o
r
izati
o
n
[
6
]
,
s
en
ti
m
e
n
t
an
al
y
s
is
[
7
]
,
an
d
s
e
n
ta
n
ce
s
class
i
f
icatio
n
[
8
]
,
in
w
h
ic
h
r
eq
u
ir
e
to
all
o
ca
te
p
r
e
-
d
ef
i
n
ed
ca
teg
o
r
ies
to
a
s
eq
u
e
n
tial
o
f
te
x
t.
A
m
o
s
t
o
f
g
e
n
er
al
a
n
d
b
asic
tec
h
n
iq
u
e
o
f
r
ep
r
ese
n
ti
n
g
tex
t
s
i
s
b
a
g
-
of
-
w
o
r
d
s
.
W
h
ile,
in
b
ag
-
of
-
w
o
r
d
s
ap
p
r
o
ac
h
h
as
an
is
s
u
e
to
lo
s
e
th
e
w
o
r
d
s
s
eq
u
en
ce
a
n
d
ig
n
o
r
e
th
e
s
e
m
an
tic
o
f
w
o
r
d
s
[
5
]
.
T
h
er
ef
o
r
e,
in
t
h
is
co
n
te
x
t,
h
o
w
to
o
r
g
a
n
ize
an
d
u
s
e
th
e
s
e
lar
g
e
a
m
o
u
n
ts
o
f
te
x
t
in
f
o
r
m
at
io
n
b
ec
o
m
e
s
p
ar
ticu
lar
l
y
i
m
p
o
r
tan
t.
F
u
r
th
er
m
o
r
e,
at
p
r
esen
t
m
o
s
t o
f
t
h
e
co
m
m
o
n
an
d
i
m
p
o
r
ta
n
t c
las
s
i
f
i
er
h
as b
ee
n
ad
o
p
ted
in
m
ac
h
in
e
lear
n
i
n
g
-
b
ased
al
g
o
r
ith
m
s
to
in
it
iated
in
tr
ad
itio
n
al
tex
t
clas
s
i
f
icatio
n
ar
e:
n
aiv
e
B
a
y
e
s
,
s
u
p
p
o
r
t
v
ec
to
r
m
ac
h
in
e,
n
eu
r
al
n
et
w
o
r
k
,
k
-
n
ea
r
est
n
ei
g
h
b
o
r
s
,
f
ast
te
x
t,
an
d
d
ec
is
io
n
tr
ee
s
alg
o
r
ith
m
[
9
]
.
C
o
m
p
ar
ativ
e
l
y
,
SVM
h
as
g
r
ea
ter
i
m
p
le
m
en
ta
tio
n
r
eq
u
ir
e
m
en
ts
to
f
u
lf
ill
t
h
e
t
h
eo
r
y
,
b
u
t
it
h
as
b
ee
n
u
s
ed
to
p
r
o
d
u
ce
b
etter
r
esu
lts
in
m
an
y
f
ie
ld
s
[
1
0
,
1
1
]
.
Ho
w
e
v
er
,
th
eir
p
er
f
o
r
m
a
n
ce
d
ep
en
d
s
g
e
n
e
r
all
y
o
n
t
h
e
q
u
al
it
y
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:
2
0
8
8
-
8708
Text
cla
s
s
ific
a
tio
n
b
a
s
ed
o
n
g
a
ted
r
ec
u
r
r
en
t u
n
it
co
mb
in
es
w
ith
… (
Mu
h
a
mma
d
Zu
lq
a
r
n
a
i
n
)
3735
o
f
h
a
n
d
-
cr
a
f
ted
f
ea
tu
r
e
s
[
1
2
]
.
B
in
[
1
3
]
p
r
o
p
o
s
ed
th
e
latest
t
y
p
e
o
f
R
NN
as
k
n
o
w
n
GR
U,
w
h
ic
h
u
til
ize
s
th
e
ad
v
an
ta
g
e
s
o
f
GR
U
an
d
C
NN
j
o
in
tl
y
,
to
r
ec
o
g
n
ize
m
ed
icall
y
r
elatio
n
s
i
n
clin
ical
r
ec
o
r
d
s
,
w
it
h
o
n
l
y
w
o
r
d
s
e
m
b
ed
d
in
g
f
ea
tu
r
es.
Ov
er
t
h
e
y
ea
r
s
,
v
ar
io
u
s
d
ee
p
l
ea
r
n
in
g
ap
p
r
o
ac
h
es
h
av
e
b
ee
n
p
r
esen
ted
to
ad
d
r
ess
t
h
is
is
s
u
e
an
d
h
a
v
e
ex
ten
d
ed
to
i
m
p
le
m
e
n
t
o
f
R
NNs
to
r
ea
l
-
w
o
r
ld
is
s
u
es.
H
o
w
e
v
er
,
b
ased
o
n
th
e
liter
atu
r
e
r
ev
ie
w
s
an
d
o
u
r
r
esear
ch
ex
p
er
i
m
e
n
t,
w
e
f
o
u
n
d
th
a
t
s
o
m
e
d
ef
icien
c
ies
o
f
s
tan
d
ar
d
R
NN
ar
e
t
h
e
g
r
ad
ien
t
v
a
n
is
h
i
n
g
an
d
ex
p
lo
d
in
g
is
s
u
es.
I
t
m
ak
e
s
th
e
tr
ain
i
n
g
o
f
R
N
N
d
if
f
icu
lt
,
in
t
w
o
w
a
y
s
:
(
i)
it
ca
n
n
o
t
p
r
o
ce
s
s
v
er
y
lo
n
g
s
eq
u
en
ce
s
if
u
s
i
n
g
h
y
p
er
b
o
lic
tan
h
ac
t
iv
atio
n
f
u
n
ctio
n
;
(
ii)
it
is
v
er
y
u
n
s
tab
le
if
u
s
i
n
g
th
e
r
ec
tif
ied
li
n
ea
r
u
n
it
(
R
eL
U)
as
a
n
ac
tiv
atio
n
f
u
n
ctio
n
.
I
n
th
i
s
r
esear
ch
,
f
o
r
m
a
n
ag
e
to
o
v
er
c
o
m
e
th
e
s
e
i
s
s
u
es
w
e
h
av
e
to
u
s
e
th
e
late
s
t
t
y
p
e
o
f
r
ec
u
r
r
en
t
n
e
t
w
o
r
k
s
s
u
c
h
a
s
G
R
U
n
et
w
o
r
k
,
in
itia
ll
y
in
tr
o
d
u
ce
d
b
y
K
y
u
n
g
h
y
u
n
C
h
o
2
0
1
4
.
GR
U
m
o
d
el
in
d
icate
s
a
s
tr
o
n
g
p
er
f
o
r
m
a
n
ce
o
n
r
ea
l
-
w
o
r
l
d
s
eq
u
en
t
ial
tas
k
s
.
I
n
f
ac
t,
t
h
e
i
n
f
lu
e
n
ce
o
f
th
is
r
esear
ch
w
e
h
a
v
e
to
co
m
b
i
n
e
t
w
o
m
o
s
t
o
f
t
h
e
s
tr
o
n
g
s
ta
te
-
of
-
th
e
-
ar
t
ap
p
r
o
ac
h
es,
g
ated
r
e
cu
r
r
en
t
u
n
it
(
G
R
U)
an
d
s
u
p
p
o
r
t
v
ec
to
r
m
ac
h
in
e
(
SVM)
,
b
o
th
o
f
w
h
ic
h
ca
n
b
e
tr
ac
ed
b
ac
k
[
1
4
]
.
GR
U
h
as
th
e
ca
p
ab
ilit
y
o
f
a
d
y
n
a
m
ical
s
y
s
te
m
w
h
er
e
t
h
e
n
et
w
o
r
k
o
u
tp
u
t
w
ill
b
e
d
ec
i
d
ed
n
o
t
in
d
iv
id
u
all
y
b
y
th
e
c
u
r
r
en
t
i
n
p
u
t
b
u
t
al
s
o
b
y
d
ec
id
e
cu
r
r
en
t
n
et
w
o
r
k
s
t
ate.
T
h
e
o
u
tp
u
t
is
b
ase
o
n
th
e
p
r
ev
io
u
s
co
m
p
u
tatio
n
s
.
I
t
is
ch
ar
ac
ter
ized
b
y
th
e
g
o
o
d
m
o
d
el
o
f
s
eq
u
e
n
tial
d
ata
an
d
co
m
p
letel
y
u
tili
za
tio
n
s
o
f
s
eq
u
e
n
ce
in
f
o
r
m
at
io
n
.
Fu
r
t
h
er
m
o
r
e,
in
r
ec
e
n
t
y
ea
r
g
ated
r
ec
u
r
r
en
t
u
n
it
(
G
R
U)
h
a
v
e
u
s
ed
b
ec
au
s
e
o
f
t
h
e
s
h
o
r
t
-
co
m
in
g
s
o
f
s
tan
d
ar
d
R
NN
s
ar
e
ad
d
r
ess
in
g
w
it
h
lo
n
g
te
x
t.
T
h
e
ad
v
an
t
ag
e
o
f
t
h
e
GR
U
m
o
d
el
is
th
a
t
th
e
y
d
eter
m
i
n
e
i
n
th
e
p
r
o
ce
s
s
o
f
s
y
n
t
h
es
is
to
co
n
tr
o
l
h
o
w
m
u
ch
i
n
f
o
r
m
atio
n
s
h
o
u
ld
b
e
r
ec
eiv
ed
i
n
t
h
e
cu
r
r
en
t
s
y
n
th
e
s
is
s
tep
s
,
h
o
w
m
u
c
h
is
f
o
r
g
o
tte
n
,
an
d
h
o
w
m
u
c
h
i
n
f
o
r
m
atio
n
is
p
ass
ed
b
ac
k
.
T
h
r
o
u
g
h
t
h
ese
g
ate
co
n
tr
o
ls
,
GR
U
h
as
a
p
r
o
f
icien
t
lear
n
i
n
g
ab
ili
t
y
f
o
r
lo
n
g
tex
t.
Gi
v
e
n
th
e
e
x
is
ti
n
g
liter
atu
r
e,
th
ese
m
o
d
els
h
a
v
e
b
ee
n
i
m
p
le
m
en
t
ed
all
m
e
m
o
r
y
f
r
ee
.
C
o
n
s
eq
u
en
t
l
y
,
t
h
e
R
N
N
n
e
t
w
o
r
k
s
w
it
h
m
e
m
o
r
y
o
p
er
atio
n
ar
e
o
b
v
io
u
s
l
y
m
o
r
e
ap
p
r
o
p
r
iate
f
o
r
th
is
t
y
p
e
o
f
tas
k
.
I
n
t
h
is
r
es
ea
r
ch
p
ap
er
,
w
e
m
ai
n
l
y
p
er
f
o
r
m
a
th
r
ee
f
o
ld
co
n
tr
ib
u
tio
n
.
Firstl
y
,
w
e
co
n
ce
n
tr
ate
o
n
g
at
ed
r
ec
u
r
r
en
t
u
n
it
(
GR
U)
n
et
wo
r
k
,
a
ty
p
e
o
f
g
ated
R
N
N
s
,
wh
ich
g
r
ea
tl
y
m
i
n
i
m
izes
th
e
v
an
i
s
h
in
g
g
r
a
d
ien
t
a
n
d
ex
p
lo
d
in
g
is
s
u
e
s
o
f
R
NN
s
o
v
er
g
ati
n
g
m
ec
h
an
is
m
t
h
at
co
n
s
tr
u
cts
th
e
s
i
m
p
le
ar
ch
itect
u
r
e
w
h
il
e
p
r
eser
v
in
g
t
h
e
i
m
p
ac
t
o
f
L
ST
M
(
is
a
a
n
o
t
h
er
t
y
p
e
o
f
R
NN
).
Seco
n
d
l
y
,
w
e
p
r
o
p
o
s
ed
to
r
ep
lace
th
e
h
y
p
er
b
o
lic
tan
g
e
n
t
ac
ti
v
atio
n
f
u
n
ctio
n
(
ta
n
h
)
w
i
th
lea
k
y
r
ec
tif
ied
li
n
ea
r
u
n
i
t
(
L
R
el
u
)
ac
ti
v
atio
n
i
n
th
e
s
tat
e
to
th
e
m
o
d
if
ied
eq
u
atio
n
.
L
R
elu
u
n
its
h
av
e
b
ee
n
d
e
m
o
n
s
tr
ated
to
b
e
b
etter
p
er
f
o
r
m
a
n
ce
t
h
a
n
s
i
g
m
o
id
n
o
n
-
li
n
ea
r
itie
s
f
o
r
f
ee
d
-
f
o
r
w
ar
d
DNNs
.
T
h
ir
d
l
y
,
w
e
p
r
esen
t
a
r
ev
is
i
n
g
to
t
h
i
s
s
tan
d
ar
d
b
y
i
m
p
le
m
e
n
ti
n
g
lin
ea
r
s
u
p
p
o
r
t
v
ec
to
r
m
ac
h
i
n
e
(
SVM)
as
th
e
r
ep
lace
m
e
n
t
o
f
tr
ad
itio
n
al
So
f
t
m
a
x
i
n
th
e
f
in
al
o
u
tp
u
t la
y
er
o
f
a
GR
U
m
o
d
el.
2.
SUPP
O
RT
V
E
CT
O
R
M
ACH
I
NE
T
h
e
s
u
p
p
o
r
t
v
ec
to
r
m
ac
h
i
n
e
w
a
s
i
n
tr
o
d
u
ce
d
b
y
Vap
n
ek
i
n
1
9
9
0
s
an
d
it
s
p
o
p
u
lar
it
y
h
a
s
in
cr
ea
s
ed
q
u
ick
l
y
.
I
t
i
s
es
s
en
t
iall
y
d
u
e
to
its
ac
h
iev
ed
e
x
ce
lle
n
t
“
s
ta
te
-
of
-
t
h
e
-
ar
t
p
er
f
o
r
m
a
n
ce
s
i
n
v
ar
io
u
s
r
ea
l
-
w
o
r
ld
a
p
p
licatio
n
s
an
d
g
en
er
alize
d
b
etter
o
n
in
v
is
ib
le
d
ata
s
et
s
.
Fu
r
th
er
m
o
r
e
,
th
e
alter
n
ate
s
i
g
n
i
f
ica
n
t
co
m
p
o
n
en
t
is
th
at,
d
is
s
i
m
il
ar
n
e
u
r
al
n
et
w
o
r
k
s
,
SVM
g
en
er
ate
co
n
s
is
te
n
t
r
esu
lt
s
.
SVM
w
as
o
r
ig
i
n
all
y
d
es
ig
n
ed
f
o
r
p
r
o
b
lem
s
o
f
t
w
o
clas
s
es
a
n
d
co
n
ta
in
s
b
o
th
p
o
s
iti
v
e
an
d
n
e
g
ati
v
e
o
b
jects.
T
h
e
b
asic
id
ea
o
f
SVMs
is
as
s
i
g
n
th
e
in
p
u
t
v
ec
to
r
to
an
e
n
tit
y
s
p
ac
e
w
i
th
a
co
m
p
lex
d
i
m
en
s
io
n
a
l
f
ea
tu
r
e
an
d
f
i
n
d
t
h
e
b
est
p
o
s
s
ib
le
lin
ea
r
h
y
p
er
p
lan
e
t
h
a
t
s
p
lits
i
n
to
t
w
o
tar
g
et
c
lass
e
s
at
th
e
m
a
x
i
m
all
y
t
h
e
m
ar
g
in
[
1
5
]
.
Fu
r
th
er
m
o
r
e,
m
a
n
y
s
u
p
p
o
r
t v
ec
to
r
s
ar
e
u
s
ed
to
esti
m
ate
t
h
e
g
en
er
aliza
tio
n
p
e
r
f
o
r
m
an
ce
,
th
e
n
u
m
b
er
d
ec
r
e
ases
a
s
t
h
e
n
u
m
b
er
o
f
s
u
p
p
o
r
t
v
ec
to
r
s
in
cr
ea
s
e
s
.
T
h
e
f
ea
tu
r
es
s
p
ac
e
m
ap
p
in
g
f
r
o
m
th
e
in
p
u
t
s
p
ac
e
is
ac
h
iev
ed
t
h
r
o
u
g
h
t
h
e
k
er
n
el
tr
ick
s
,
w
h
ich
allo
w
s
m
ap
p
in
g
to
a
h
ig
h
er
s
p
ac
e
d
i
m
en
s
io
n
a
lit
y
w
i
t
h
o
u
t
th
e
e
s
s
e
n
tial to
e
x
p
licitl
y
f
i
n
d
th
is
s
p
ac
e.
T
h
e
k
er
n
el
ca
n
b
e
d
eter
m
in
ed
a
s
an
y
f
u
n
c
tio
n
alit
y
t
h
at
s
atis
f
ie
s
Me
r
ce
r
’
s
t
h
eo
r
e
m
[
1
6
]
.
T
h
e
r
ad
ials
b
ase
k
er
n
el
an
d
li
n
e
ar
k
er
n
e
l
ar
e
b
et
w
ee
n
t
h
e
m
o
s
t e
x
te
n
s
iv
el
y
ap
p
lied
k
er
n
els
w
i
th
SVM.
Ho
w
e
v
er
,
d
ev
elo
p
m
en
t
o
f
th
e
latest
k
er
n
e
ls
th
a
t
s
h
o
w
s
th
e
s
i
m
ilar
ities
w
i
th
v
ar
io
u
s
ap
p
li
ca
tio
n
s
is
an
ac
tiv
e
ex
p
lo
r
atio
n
f
ield
an
d
s
ev
er
al
k
er
n
e
ls
h
a
v
e
b
ee
n
d
ev
elo
p
ed
in
r
ec
en
t
y
ea
r
s
to
h
an
d
le
p
ar
ticu
lar
ap
p
licatio
n
.
SVM
w
as
in
itiall
y
d
esi
g
n
ed
as
b
in
ar
y
clas
s
i
f
ier
s
s
o
s
ev
er
al
m
eth
o
d
s
h
a
v
e
u
s
e
d
to
in
cr
ea
s
e
SVM
s
to
th
e
m
u
lt
ip
le
class
e
s
is
s
u
e
[
1
7
]
.
T
h
e
p
r
in
cip
le
m
et
h
o
d
s
”
ar
e
“
o
n
e
a
g
ain
s
t
all”
(
O
AA
)
an
d
“
o
n
e
ag
ai
n
s
t
o
n
e”
(
OAO)
w
h
er
e
in
t
h
e
1
s
t
tech
n
iq
u
e
n
SVM
class
if
ier
s
w
il
l
b
e
co
n
s
tr
u
ct,
o
n
e
f
o
r
ea
ch
class
,
a
n
d
(
−
)
b
in
ar
y
“
SVM
class
i
f
ier
is
ap
p
lied
;
th
e
m
ai
n
s
tr
ea
m
v
o
ti
n
g
b
e
t
w
ee
n
th
e
s
e
class
if
ier
s
h
a
v
e
u
s
ed
in
p
r
ed
ictio
n
to
n
e
w
p
o
in
ts
ci
te
-
m
u
lticla
s
s
s
u
p
p
o
r
t
v
ec
to
r
m
ac
h
in
e
s
.
T
h
e
b
asic
e
q
u
atio
n
s
o
f
t
h
e
SVM
th
at
ar
e
ap
p
lied
to
esti
m
atio
n
t
h
e
d
ec
is
io
n
f
u
n
ct
i
o
n
f
r
o
m
a
tr
ai
n
in
g
d
ataset
[
1
8
]
ar
e
as f
o
llo
w
s
:
ƒ
(
x
)
=
(
∑
.
(
,
)
+
=
1
)
0
(
1
)
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
0
8
8
-
8708
I
n
t J
E
lec
&
C
o
m
p
E
n
g
,
Vo
l.
10
,
No
.
4
,
A
u
g
u
s
t 2
0
2
0
:
3
7
3
4
-
3742
3736
w
h
er
e
b
is
th
e
ter
m
o
f
b
ias
an
d
is
p
r
esen
ti
n
g
th
e
n
u
m
b
er
o
f
th
e
s
u
p
p
o
r
t
v
ec
to
r
s
,
y
n
∈
{−
1
,
+1
}
is
th
e
cla
s
s
s
ig
n
s
to
w
h
ich
th
e
s
u
p
p
o
r
t
v
ec
to
r
b
elo
n
g
a
n
d
α
is
g
o
t
as
th
e
s
o
lu
t
io
n
o
f
th
e
s
u
b
s
eq
u
en
t
q
u
ad
r
atic
o
p
tim
izatio
n
i
s
s
u
es:
”
m
i
n
1
2
+
∑
=
1
(
2
)
s
.
t
(
(
)
+
)
≥
1
−
≥
0
,
=
1
,
…
i
n
t
h
e
d
ataset,
th
e
n
u
m
b
er
o
f
d
ata
p
o
in
ts
ca
n
n
o
t
in
cr
ea
s
e
th
e
n
u
m
b
er
o
f
s
u
p
p
o
r
t
v
ec
to
r
s
;
p
r
ef
er
ab
ly
it
s
h
o
u
ld
b
e
a
co
m
p
ar
ativ
el
y
s
m
all
p
o
r
ti
o
n
o
f
th
e
d
ataset
[
1
9
]
.
3.
G
RU
-
B
AS
E
D
M
O
DE
L
T
h
e
GR
U
w
as
r
ec
en
tl
y
in
tr
o
d
u
ce
d
b
y
[
2
0
]
.
GR
U
i
s
o
n
e
o
f
t
h
e
n
e
w
er
g
e
n
er
atio
n
s
o
f
g
ated
R
NN
s
t
h
a
t
h
av
e
ap
p
lied
to
s
o
l
v
e
t
h
e
b
asi
c
is
s
u
e
s
o
f
v
an
i
s
h
in
g
g
r
ad
ien
t
an
d
ex
p
lo
d
in
g
in
s
ta
n
d
ar
d
R
N
Ns
w
h
e
n
ca
p
t
u
r
in
g
lo
n
g
-
ter
m
d
ep
en
d
en
cie
s
.
T
h
is
GR
U
ap
p
r
o
ac
h
i
s
b
ased
o
n
d
y
n
a
m
icall
y
a
n
al
y
s
i
s
t
h
at
s
u
f
f
er
s
f
r
o
m
g
r
ad
ien
t
s
ex
p
lo
s
io
n
d
u
e
to
th
eir
n
o
n
-
lin
ea
r
d
y
n
a
m
ic.
I
t
w
as
d
esi
g
n
ed
to
p
r
o
p
er
ly
u
p
d
ate
o
r
r
eset
it
s
m
e
m
o
r
y
co
n
te
n
t
s
an
d
is
a
li
g
h
t
l
y
m
o
r
e
s
i
m
p
l
if
y
v
ar
iatio
n
o
f
L
ST
M
[
2
1
]
.
I
t
is
co
m
b
in
a
tio
n
o
f
t
h
e
i
n
p
u
t
a
n
d
f
o
r
g
et
g
ates
i
n
to
a
s
in
g
le
“u
p
d
ate
g
ate”
a
n
d
h
a
s
an
ad
d
itio
n
al
“
r
eset
g
ate”
.
T
h
e
GR
U
m
o
d
el
i
s
s
i
m
p
le
a
n
d
h
as f
e
w
er
p
ar
a
m
e
ter
s
as c
o
m
p
ar
e
to
tr
ad
itio
n
al
L
ST
M
m
o
d
els an
d
ar
e
g
ai
n
p
r
o
g
r
ess
i
v
el
y
p
o
p
u
lar
it
y
.
Fig
u
r
e
1
illu
s
tr
ati
n
g
,
th
e
m
u
l
tip
le
n
eu
r
o
n
s
ar
e
co
m
p
o
s
ed
in
a
s
in
g
le
i
n
p
u
t
la
y
er
,
t
h
e
n
u
m
b
er
s
o
f
a
n
eu
r
o
n
ar
e
d
ec
id
ed
b
y
t
h
e
s
c
o
p
e
o
f
th
e
f
ea
t
u
r
es
s
p
ac
e.
E
q
u
iv
ale
n
tl
y
,
th
e
o
u
tp
u
t
la
y
er
n
e
u
r
o
n
s
ar
e
r
ese
m
b
les
to
th
e
o
u
tp
u
t
s
p
ac
e.
Ho
w
e
v
er
,
d
is
ti
n
ct
t
h
e
L
ST
M,
th
e
G
R
U
co
m
p
letel
y
e
x
p
o
s
ed
it
’
s
m
e
m
o
r
y
co
n
ten
t
ea
c
h
ti
m
e
s
tep
an
d
b
alan
ce
s
a
m
o
n
g
th
e
n
e
w
m
e
m
o
r
y
co
n
ten
t
s
an
d
th
e
p
r
ev
io
u
s
m
e
m
o
r
y
co
n
te
n
ts
s
e
v
er
el
y
u
s
e
o
f
th
e
leak
y
co
m
b
i
n
atio
n
,
alth
o
u
g
h
w
i
th
it
s
ad
j
u
s
tab
le
ti
m
e
co
n
tin
u
o
u
s
l
y
co
n
tr
o
l b
y
u
p
d
ate
g
a
te.
Fig
u
r
e
1
.
Stru
ct
u
r
e
o
f
th
e
G
R
U
-
b
ased
m
o
d
el
T
h
e
ac
tiv
atio
n
ℎ
o
f
th
e
GR
U
at
ti
m
e
t
i
s
a
li
n
ea
r
i
n
ter
p
o
latin
g
a
m
o
n
g
t
h
e
p
r
ev
io
u
s
ac
ti
v
atio
n
s
ℎ
−
1
an
d
th
e
ca
n
d
id
ate
ac
ti
v
atio
n
ĥ
:
ℎ
=
(
1
−
ƶ
)
ℎ
−
1
+
ƶ
ĥ
(
3
)
T
h
e
u
p
d
ate
g
ate
ƶ
d
eter
m
i
n
e
h
o
w
m
u
ch
t
h
e
u
n
it
u
p
d
ates its
ac
tiv
atio
n
:
ƶ
=
s
ig
m
(
ƶ
+
ƶ
ℎ
−
1
)
ͥ
(
4
)
T
h
e
ca
n
d
id
ate
ac
tiv
atio
n
ĥ
is
ca
l
cu
lated
s
i
m
ilar
to
th
e
u
p
d
ate
g
ate:
ĥ
=
tan
h
(
+
(
∗
ℎ
−
1
))
ͥ
(
5
)
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:
2
0
8
8
-
8708
Text
cla
s
s
ific
a
tio
n
b
a
s
ed
o
n
g
a
ted
r
ec
u
r
r
en
t u
n
it
co
mb
in
es
w
ith
… (
Mu
h
a
mma
d
Zu
lq
a
r
n
a
i
n
)
3737
w
h
er
e
is
th
e
in
p
u
t
v
ec
to
r
,
ℎ
is
th
e
h
id
d
en
s
tate
v
ec
to
r
,
ar
e
r
eset
g
ates
an
d
∗
d
en
o
ted
b
y
ele
m
e
n
t
-
w
is
e
p
r
o
d
u
ct.
W
h
en
th
e
r
eset
g
a
te
is
o
f
f
(
=
=
0
)
it
p
er
m
its
t
h
e
u
n
it
to
f
o
r
g
et
th
e
p
r
ev
io
u
s
in
f
o
r
m
at
io
n
.
T
h
is
is
s
i
m
ilar
to
allo
w
lear
n
i
n
g
o
f
in
p
u
t
s
eq
u
e
n
tial
s
ym
b
o
l.
T
h
e
r
eset
g
ate
is
e
v
alu
ated
u
s
in
g
th
e
s
u
b
s
eq
u
en
t
eq
u
atio
n
.
=
s
ig
m
(
+
ℎ
−
1
)
ͥ
(
6
)
Up
d
ate
g
ate
ƶ
d
ec
id
es
h
o
w
m
u
c
h
t
h
e
p
r
ev
io
u
s
i
n
f
o
r
m
atio
n
(
f
r
o
m
th
e
p
r
ev
io
u
s
ti
m
e
s
te
p
t
)
s
h
o
u
ld
th
r
o
w
a
w
a
y
an
d
w
h
at
n
e
w
s
tat
e
t
o
ad
d
.
Un
its
o
f
r
eset
g
ate
ar
e
p
r
o
p
er
ly
ac
ti
v
e
o
n
s
h
o
r
t
-
ter
m
d
ep
en
d
en
cie
s
.
Un
it
s
w
it
h
lo
n
g
-
ter
m
d
ep
en
d
e
n
cies
h
av
e
an
ac
ti
v
e
u
p
d
ate
g
a
te.
4.
T
H
E
P
RO
P
O
SE
D
G
R
U
-
SV
M
ARCH
I
T
E
CT
U
RE
In
th
e
ex
it
in
g
s
y
s
te
m
m
o
s
tl
y
r
esear
ch
er
s
h
a
v
e
u
s
ed
m
ac
h
i
n
e
lear
n
in
g
an
d
d
ata
m
i
n
i
n
g
alg
o
r
ith
m
s
to
class
i
f
y
th
e
tex
t.
B
u
t
i
n
o
u
r
p
r
o
p
o
s
ed
m
o
d
el
w
e
co
n
ce
n
tr
ate
o
n
th
e
d
ee
p
lear
n
in
g
al
g
o
r
ith
m
s
to
clas
s
i
f
y
th
e
tex
t.
I
n
s
e
v
er
al
s
t
u
d
ies
a
n
d
r
esear
ch
es
v
er
if
ied
t
h
at
d
ee
p
lear
n
in
g
al
g
o
r
ith
m
s
p
r
o
d
u
c
ed
b
etter
ac
cu
r
ac
y
th
an
m
ac
h
in
e
lear
n
in
g
an
d
tr
ad
itio
n
al
alg
o
r
ith
m
s
.
I
n
th
i
s
p
ap
er
,
w
e
d
is
c
u
s
s
ed
th
e
R
NN
al
g
o
r
ith
m
to
ac
q
u
ir
e
b
etter
ac
cu
r
ac
y
.
T
h
e
ar
ch
itect
u
r
e
o
f
o
u
r
p
r
o
p
o
s
ed
s
y
s
te
m
i
s
s
h
o
w
n
in
F
ig
u
r
e
2
.
Fig
u
r
e
2
.
T
h
e
p
r
o
p
o
s
ed
GR
U
-
SVM
ar
ch
i
tect
u
r
e
m
o
d
el,
w
it
h
ti
m
es
tep
,
GR
U
i
n
p
u
t
u
n
i
t
an
d
SVM
as i
t
s
class
i
f
ier
Si
m
i
lar
l
y
,
to
t
h
e
w
o
r
k
o
f
A
l
alsh
a
k
m
u
b
ar
ak
[
2
2
]
an
d
s
h
o
w
s
t
h
at
th
e
p
r
o
p
o
s
e
to
ap
p
ly
SVM
a
s
th
e
clas
s
i
f
ier
i
n
a
d
ee
p
lear
n
in
g
ar
c
h
itect
u
r
e
esp
ec
iall
y
GR
U.
T
h
en
,
t
h
e
p
ar
a
m
eter
s
ar
e
tr
ain
ed
ac
r
o
s
s
th
e
g
ati
n
g
m
ec
h
a
n
is
m
o
f
G
R
U
ar
ch
i
tectu
r
e
[
2
3
]
.
Ho
w
e
v
er
,
g
ated
r
ec
u
r
r
en
t
u
n
it
GR
U
o
v
er
co
m
es
s
u
c
h
d
ef
icien
c
ies
o
f
ex
is
tin
g
R
NN
s
b
y
ad
d
in
g
t
h
e
R
NN
w
ith
an
u
p
d
ate
an
d
r
eset
g
ates
th
at
tak
e
as
an
in
p
u
t
,
ℎ
−
1
,
an
d
g
en
er
ate
s
ĥ
,
b
y
t
h
e
f
o
llo
w
i
n
g
:
-
U
p
d
ate
g
ate:
=
σ (
.
[
ℎ
−
1
,
]
)
(
7
)
-
R
eset
g
ate:
=
σ (
.
[
ℎ
−
1
,
]
)
(
8
)
-
C
an
d
id
ate
h
id
d
en
s
tate:
ĥ
=
ta
n
h
(
.
[
∗
ℎ
−
1
,
]
)
(
9
)
T
h
e
s
ec
o
n
d
m
o
d
if
icat
io
n
co
n
s
is
t
s
o
f
r
ep
lacin
g
t
h
e
s
tan
d
ar
d
h
y
p
er
b
o
lic
tan
g
en
t
w
i
th
L
e
ak
y
-
R
e
L
U
(
L
R
el
u
)
ac
ti
v
atio
n
f
u
n
ctio
n
.
I
n
p
ar
ticu
lar
,
w
e
m
o
d
i
f
y
th
e
ca
lc
u
latio
n
o
f
ca
n
d
id
ate
s
tate
ĥ
in
(
9
)
as f
o
llo
w
s
:
ĥ
=
LRe
lu
(
.
[
∗
ℎ
−
1
,
]
)
(
1
0
)
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
0
8
8
-
8708
I
n
t J
E
lec
&
C
o
m
p
E
n
g
,
Vo
l.
10
,
No
.
4
,
A
u
g
u
s
t 2
0
2
0
:
3
7
3
4
-
3742
3738
-
Hid
d
en
s
tate:
ℎ
= (
1
−
)
∗
ℎ
−
1
+
∗
ĥ
(
1
1
)
w
h
er
e
W
,
U
ar
e
w
eig
h
t
m
atr
ic
es,
σ
is
a
lo
g
i
s
tic
s
i
g
m
o
id
ac
ti
v
atio
n
f
u
n
ct
io
n
,
an
d
ℎ
−
1
is
th
e
h
id
d
en
s
tate
(
f
r
o
m
p
r
ev
io
u
s
t
i
m
e
s
tep
t
)
.
B
u
t
as
a
f
i
n
al
la
y
er
i
n
tr
o
d
u
ce
d
b
y
SV
M
an
d
it
s
r
ep
lace
m
e
n
t
o
f
tr
ad
i
tio
n
al
s
o
f
t
m
a
x
la
y
er
in
a
G
R
U
”
t
h
e
o
p
ti
m
u
m
s
ep
ar
atin
g
h
y
p
er
-
p
la
n
e
p
ar
a
m
eter
s
ar
e
also
s
o
lv
e
d
th
e
s
u
b
s
eq
u
e
n
t
r
eg
u
lar
izat
io
n
b
y
i
m
p
r
o
v
i
n
g
t
h
e
p
r
in
cip
al
f
u
n
cti
o
n
o
f
SVM
as
s
h
o
w
n
in
eq
u
ati
o
n
12
:
Ɉ
(
ῶ
,
)
=
m
i
n
1
2
ῶ
+
∑
=
1
(
1
2
)
(
ῶ
(
ẍ
)
+
)
≥
1
−
an
d
≥
0
,
=
1
,
…
(
1
3
)
C
o
n
s
id
er
a
tr
ain
in
g
d
ata
s
et
G
=
{[
ẍ
,
]
}
=
1
,
ẍ
∈
is
th
e
i
th
in
p
u
t
d
esi
g
n
a
n
d
is
its
eq
u
iv
ale
n
t
o
b
s
er
v
ed
r
esu
lt
∈
{+
1
,
-
1
}.
I
n
tes
t
clas
s
i
f
icatio
n
m
o
d
el,
d
en
o
ted
b
y
attr
ib
u
te
s
o
f
tex
t
v
e
cto
r
,
is
clas
s
lab
el,
w
h
er
e
c
i
s
a
co
n
s
ta
n
t
r
e
p
r
esen
tin
g
a
tr
ad
eo
f
f
a
m
o
n
g
t
h
e
ed
g
e
a
n
d
t
h
e
ca
lc
u
latio
n
o
f
to
tal
er
r
o
r
s
.
“
ϕ
(
ẍ
)
is
a
n
o
n
-
lin
ea
r
f
u
n
ctio
n
t
h
at
p
l
o
ts
th
e
in
p
u
t
s
p
ac
e
in
to
a
s
u
p
er
io
r
d
im
e
n
s
io
n
al
f
ea
t
u
r
e
s
p
ac
e.
T
h
e
m
ar
g
i
n
b
et
w
ee
n
th
e
t
w
o
p
ar
ts
is
2
/
w
.
T
h
e
L
2
-
SVM
h
a
s
to
ap
p
ie
d
f
o
r
t
h
e
p
r
o
p
o
s
ed
GR
U
-
S
VM
s
tr
u
ct
u
r
e
.
A
s
f
o
r
t
h
e
p
r
ed
ictio
n
,
th
e
d
ec
is
io
n
f
u
n
ctio
n
f
(
x
)
=
s
i
g
n
(
wx
+
b
)
g
e
n
er
ates
a
s
co
r
e
v
ec
to
r
f
o
r
ea
ch
cla
s
s
es.
So
,
to
d
ev
elo
p
th
e
p
r
ed
ict
class
lab
el
y
o
f
a
d
ata
x
,
th
e
ar
r
m
ax
f
u
n
ctio
n
is
e
m
p
lo
y
ed
:
P
r
ed
ict
-
class
=
ar
g
m
ax
(
s
ig
n
(
wx
+
b
)
)
(
1
4
)
T
h
e
p
r
o
p
o
s
ed
GR
U
-
SVM
ap
p
r
o
ac
h
is
s
u
m
m
ar
ized
as f
o
llo
ws:
a.
I
n
p
u
t t
h
e
d
ataset
f
ea
tu
r
es {
x
i
|
x
i
∈
R
m
}
to
th
e
G
R
U
n
et
w
o
r
k
.
b.
I
n
itialize
ar
b
itra
r
y
v
al
u
es
w
it
h
th
e
lear
n
i
n
g
p
ar
a
m
eter
s
w
ei
g
h
ts
a
n
d
b
iases
(
th
e
y
w
i
ll
b
e
ad
j
u
s
ted
th
r
o
u
g
h
tr
ain
i
n
g
)
.
c.
T
h
e
GR
U
n
et
w
o
r
k
ce
ll
s
tate
is
ca
lcu
lated
b
ased
o
n
th
e
i
n
p
u
t
f
ea
t
u
r
es
x
i
,
a
n
d
its
lear
n
in
g
p
ar
a
m
eter
s
v
alu
e
s
.
d.
A
t
t
h
e
last
t
i
m
e
s
tep
,
th
e
p
r
ed
ictio
n
o
f
th
e
m
o
d
el
is
co
m
p
u
ted
u
s
i
n
g
th
e
d
ec
is
i
o
n
f
u
n
ct
io
n
o
f
SVM:
ƒ
(
x
)
=
s
ig
n
(
w
x
+b
)
.
e.
An
o
p
ti
m
izatio
n
m
et
h
o
d
is
e
m
p
lo
y
ed
f
o
r
lo
s
s
m
i
n
i
m
iza
tio
n
(
A
d
a
m
o
p
ti
m
izer
w
as
ap
p
lied
in
t
h
is
r
esear
ch
)
.
Op
ti
m
izatio
n
ad
j
u
s
t
s
th
e
w
ei
g
h
ts
a
n
d
b
iases
b
ased
o
n
th
e
co
m
p
u
ted
lo
s
s
.
f.
T
h
is
p
r
o
ce
s
s
is
r
ep
ea
ted
u
n
til
th
e
r
ec
u
r
r
en
t
n
e
u
r
al
n
et
w
o
r
k
a
ch
iev
e
s
t
h
e
ap
p
r
o
p
r
iate
ac
cu
r
ac
y
”
o
r
ac
h
iev
e
s
th
e
h
i
g
h
est p
o
s
s
ib
le
ac
cu
r
ac
y
.
5.
E
XP
E
R
I
M
E
NT
S
E
T
UP
I
n
th
i
s
s
ec
tio
n
,
w
e
e
x
p
lai
n
th
e
ex
p
er
i
m
e
n
tal
s
et
tin
g
s
an
d
em
p
ir
ical
r
esu
lt
s
o
f
t
h
e
p
r
o
p
o
s
ed
m
o
d
el.
5
.
1
.
Co
llect
i
o
n o
f
da
t
a
s
et
s
I
n
th
i
s
s
ec
tio
n
,
w
e
ad
o
p
ted
t
h
e
o
r
d
in
ar
y
e
x
p
er
i
m
e
n
tal
d
at
asets
t
h
at
co
n
tai
n
C
h
in
e
s
e
c
o
r
p
u
s
an
d
co
llected
b
y
Fu
d
a
n
Un
iv
er
s
it
y
Dr
.
L
i
R
o
n
g
la.
I
n
t
h
e
a
u
to
m
at
icall
y
te
x
t c
la
s
s
i
f
icat
io
n
s
y
s
te
m
,
t
h
e
ex
p
er
i
m
en
ta
l
d
ataset
is
t
y
p
icall
y
d
iv
id
ed
i
n
to
2
p
ar
ts
:
th
e
tr
ain
i
n
g
s
et
a
n
d
test
i
n
g
s
et.
W
e
r
an
d
o
m
l
y
s
el
ec
ted
1
0
ca
teg
o
r
ies
f
r
o
m
th
e
co
r
p
u
s
a
n
d
d
elete
d
s
o
m
e
er
r
o
r
d
o
cu
m
en
t
s
.
Fin
all
y
,
th
e
co
r
p
u
s
co
n
s
is
t
s
of
1
0
ca
teg
o
r
ies
an
d
1
8
8
5
d
o
cu
m
en
t
s
i
n
th
e
t
r
ain
in
g
s
ets
a
n
d
test
i
n
g
s
et
s
c
o
n
tain
9
4
0
d
o
cu
m
e
n
ts
.
A
cc
o
r
d
in
g
to
th
e
te
x
t
ca
teg
o
r
izatio
n
s
y
s
te
m
s
ett
in
g
,
ea
ch
clas
s
s
h
o
u
ld
i
n
cl
u
d
e
a
p
ar
ticu
lar
a
m
o
u
n
t
o
f
tr
ai
n
in
g
tex
t.
E
ac
h
o
n
e
o
f
th
ese
te
x
ts
w
as
co
n
c
lu
d
ed
u
s
i
n
g
t
h
e
class
if
ier
,
”
an
d
t
h
en
t
h
e
class
if
icati
o
n
r
es
u
lts
d
is
t
in
c
tio
n
to
th
e
ac
cu
r
ate
d
eter
m
in
at
io
n
.
T
h
er
ef
o
r
e,
w
e
ca
n
d
eter
m
i
n
e
t
h
e
m
o
d
el
i
m
p
ac
t
w
ith
th
e
SVM
clas
s
i
f
ier
.
Deta
il
d
escr
ip
tio
n
o
f
d
ataset
is
p
r
esen
ted
i
n
T
ab
le
1.
I
n
“
ad
d
itio
n
,
th
er
e
ar
e
th
r
ee
ev
alu
atio
n
i
n
d
ex
m
ea
s
u
r
es t
h
e
e
f
f
icie
n
c
y
o
f
tex
t c
la
s
s
i
f
icat
io
n
:
R
ec
a
ll
=
+
(
1
5
)
P
r
ec
is
io
n
=
+
(
1
6
)
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:
2
0
8
8
-
8708
Text
cla
s
s
ific
a
tio
n
b
a
s
ed
o
n
g
a
ted
r
ec
u
r
r
en
t u
n
it
co
mb
in
es
w
ith
… (
Mu
h
a
mma
d
Zu
lq
a
r
n
a
i
n
)
3739
F1
=
2
∗
∗
+
(
1
7
)
w
h
er
e
r
ep
r
esen
t
th
e
p
o
s
i
tiv
e
in
s
tan
ce
test
d
o
cu
m
e
n
tatio
n
s
th
at
ar
e
u
s
ed
ac
c
u
r
atel
y
clas
s
if
ied
t
h
e
cla
s
s
a
s
b
elo
n
g
t
o
th
e
n
u
m
b
er
o
f
th
i
s
ca
teg
o
r
y
.
ar
e
s
h
o
w
in
g
t
h
e
n
e
g
ati
v
e
in
s
tan
ce
o
f
test
d
o
cu
m
en
tatio
n
s
th
a
t
ar
e
er
r
o
r
class
if
ied
f
o
r
b
elo
n
g
s
to
th
e
n
u
m
b
er
o
f
th
i
s
ca
teg
o
r
y
.
ar
e
th
e
p
o
s
itiv
e
in
s
ta
n
ce
s
o
f
test
d
o
cu
m
e
n
t
atio
n
s
th
at
ar
e
b
e
er
r
o
r
class
if
ied
f
o
r
d
o
es
n
o
t
b
elo
n
g
to
th
e
n
u
m
b
er
o
f
th
is
ca
te
g
o
r
y
.
F
u
r
th
er
m
o
r
e,
h
ig
h
er
p
r
ec
is
io
n
m
ea
n
s
t
h
at
a
m
o
d
el
r
et
u
r
n
d
r
a
m
atica
ll
y
m
o
r
e
r
elate
d
r
esu
lt
s
th
a
n
ir
r
elev
a
n
t,
al
s
o
h
i
g
h
r
e
ca
ll
m
ea
n
s
t
h
at
a
n
al
g
o
r
ith
m
h
a
s
r
etu
r
n
ed
b
est
o
f
th
e
r
elev
a
n
t
r
esu
lt
s
.
So
,
w
e
asp
ec
t
to
o
b
tain
h
ig
h
er
v
alu
e
s
o
f
p
r
ec
is
io
n
an
d
r
ec
all.
T
ab
le
1
.
E
x
p
er
im
e
n
t
al
d
ataset
T
ex
t c
lass
T
r
ain
in
g
s
et
s
T
est s
ets
C
o
mp
u
t
e
r
1
3
3
67
A
r
t
1
6
5
83
En
t
e
r
t
a
i
n
me
n
t
1
3
4
67
Ed
u
c
a
t
i
o
n
1
4
9
74
Ec
o
n
o
my
2
1
4
1
0
5
P
o
l
i
c
y
3
3
2
1
6
4
S
p
o
r
t
3
0
0
1
5
0
M
e
d
i
c
i
n
e
1
3
6
68
B
u
s
i
n
e
ss
1
8
1
90
T
r
a
n
sp
o
r
t
a
t
i
o
n
1
4
1
70
5
.
2
.
H
y
perpa
ra
m
et
er
s
a
nd
s
o
f
t
w
a
re
Data
p
r
ep
r
o
ce
s
s
in
g
a
n
d
m
an
i
p
u
late
h
av
e
p
er
f
o
r
m
ed
i
n
P
y
t
h
o
n
3
.
6
,
b
asis
o
n
th
e
s
k
lear
n
,
n
u
m
p
y
a
n
d
p
an
d
as
p
ac
k
ag
e
s
.
Dee
p
lear
n
in
g
GR
U
n
et
w
o
r
k
a
n
d
tr
ad
itio
n
al
DNNs
h
a
v
e
to
ex
ec
u
ted
w
it
h
T
en
s
o
r
Flo
w
,
an
o
p
en
s
o
u
r
ce
s
o
f
t
w
ar
e
lib
r
a
r
y
f
o
r
n
u
m
er
icall
y
co
m
p
u
tatio
n
s
t
h
r
o
u
g
h
d
ata
f
lo
w
s
g
r
ap
h
.
P
er
f
o
r
m
a
n
ce
o
f
all
m
et
h
o
d
s
w
a
s
b
ased
o
n
p
r
e
-
d
ef
i
n
ed
ass
es
s
m
en
t
m
ea
s
u
r
es.
C
o
m
p
letel
y
s
i
m
u
latio
n
s
w
er
e
i
m
p
le
m
e
n
ted
o
n
I
n
t
el
C
o
r
e
i7
-
3
7
7
0
XP
U
@
3
.
4
0
”
GHz
,
an
d
4
GB
o
f
R
AM
m
ac
h
i
n
e.
T
h
e
d
etail
o
f
h
y
p
er
p
ar
a
m
eter
s
o
f
th
e
p
r
o
p
o
s
ed
m
o
d
el
is
p
r
esen
t
ed
in
T
ab
le
2.
T
ab
le
2
.
Deta
il o
f
h
y
p
er
p
ar
a
m
eter
s
u
s
ed
i
n
p
r
o
p
o
s
ed
GR
U
-
S
VM
m
o
d
el
H
y
p
er
p
ar
am
e
ter
s
GR
U
-
SVM
B
a
t
c
h
S
i
z
e
1
2
8
C
e
l
l
S
i
z
e
1
2
8
D
r
o
p
o
u
t
R
a
t
e
0
.
8
5
Ep
o
c
h
s
15
L
e
a
r
n
i
n
g
R
a
t
e
1e
-
5
S
V
M
_
C
0
.
5
N
o
.
o
f
c
l
a
sse
s
10
”
6.
RE
SU
L
T
S AN
D
D
I
SCU
SS
I
O
N
T
o
ev
alu
ates
a
n
d
co
m
p
ar
ed
th
e
p
er
f
o
r
m
an
ce
o
f
G
R
U
-
S
VM
tex
t
clas
s
i
f
icatio
n
a
lg
o
r
ith
m
w
it
h
b
aselin
e
D
A
B
N
a
n
d
B
L
STM
-
C
clas
s
i
f
icatio
n
a
l
g
o
r
ith
m
s
,
w
e
h
a
v
ex
p
er
i
m
e
n
t
ed
o
n
w
ell
-
k
n
o
w
n
C
h
i
n
ese
co
r
p
u
s
d
atasets
.
w
e
h
av
e
co
n
clu
d
ed
th
at
t
h
e
G
R
U
-
SVM
a
p
p
r
o
ac
h
h
as
b
est
te
x
t
clas
s
i
f
i
ca
tio
n
s
ca
p
ab
ilit
y
in
th
e
ter
m
o
f
th
e
r
ec
all,
p
r
ec
is
io
n
an
d
th
e
F1
co
m
p
ar
e
w
it
h
th
e
e
x
is
tin
g
ap
p
r
o
ac
h
es
o
f
d
ee
p
au
to
en
co
d
er
b
elief
n
et
w
o
r
k
s
(
D
A
B
N)
[
2
4
]
an
d
B
i
d
ir
ec
tio
n
al
L
o
n
g
Sh
o
r
t
T
er
m
Me
m
o
r
y
w
i
th
co
n
v
o
lu
tio
n
al
la
y
er
“
(
B
L
ST
M
-
C
)
[
2
5
]
.
T
h
e
p
r
o
p
o
s
ed
tech
n
iq
u
e
p
e
r
f
o
r
m
ed
b
etter
tex
t
clas
s
if
icatio
n
o
n
t
h
e
d
atasets
s
u
c
h
as
co
m
p
u
ter
,
ar
t,
en
ter
tain
m
en
t,
p
o
liti
cs,
ed
u
ca
tio
n
,
s
p
o
r
ts
,
an
d
tr
an
s
p
o
r
tatio
n
.
W
h
ile
p
r
o
p
o
s
ed
im
p
r
o
v
e
d
GR
U
-
SVM
m
o
d
el
p
er
f
o
r
m
a
n
ce
is
n
o
t
g
o
o
d
to
class
if
y
ac
c
u
r
ate
class
i
n
th
e
k
in
d
o
f
b
u
s
i
n
es
s
,
ec
o
n
o
m
y
,
a
n
d
m
ed
icin
e.
T
h
is
m
a
y
b
e
b
ec
au
s
e
in
eli
m
i
n
ati
n
g
r
elate
d
f
ea
t
u
r
es
o
f
a
test
r
es
u
lt,
a
n
d
lo
s
t
s
ev
er
al
in
f
o
r
m
at
io
n
.
T
h
er
ef
o
r
e,
th
at
th
e
r
ec
all
r
ate
cr
iter
ia
h
a
v
e
af
f
ec
ted
.
C
o
n
s
eq
u
en
t
l
y
,
it
is
also
r
eq
u
ir
e
f
o
r
f
u
r
th
er
i
m
p
r
o
v
e
m
e
n
t.
T
h
er
ef
o
r
e,
th
e
p
r
o
p
o
s
ed
GR
U
-
SVM
m
o
d
el
ca
n
p
r
o
v
id
e
a
p
o
w
er
f
u
l
al
g
o
r
ith
m
i
n
p
er
f
o
r
m
in
g
te
x
t c
la
s
s
i
f
icatio
n
task
s
.
Fi
g
u
r
e
3
to
Fig
u
r
e
5
”
s
h
o
w
s
th
e
p
er
f
o
r
m
a
n
ce
o
f
t
h
e
p
r
o
p
o
s
ed
GR
U
-
SVM
m
o
d
el
w
it
h
th
e
tr
ad
itio
n
a
l
b
aselin
e
ap
p
r
o
ac
h
es.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
0
8
8
-
8708
I
n
t J
E
lec
&
C
o
m
p
E
n
g
,
Vo
l.
10
,
No
.
4
,
A
u
g
u
s
t 2
0
2
0
:
3
7
3
4
-
3742
3740
Fig
u
r
e
3
.
C
o
m
p
ar
is
o
n
p
er
f
o
r
m
an
ce
o
f
t
h
e
p
r
o
p
o
s
ed
m
o
d
el
w
i
th
b
asel
in
e
ap
p
r
o
ac
h
es in
th
e
ter
m
o
f
p
r
ec
is
io
n
Fig
u
r
e
4
.
C
o
m
p
ar
is
o
n
p
er
f
o
r
m
an
ce
o
f
th
e
p
r
o
p
o
s
ed
m
o
d
el
w
i
th
b
asel
in
e
ap
p
r
o
ac
h
es in
t
h
e
ter
m
o
f
r
e
ca
ll
Fig
u
r
e
5
.
C
o
m
p
ar
is
o
n
p
er
f
o
r
m
an
ce
o
f
t
h
e
p
r
o
p
o
s
ed
m
o
d
el
w
ith
b
aseli
n
e
ap
p
r
o
ac
h
es in
t
h
e
ter
m
o
f
F1
7.
CO
NCLU
SI
O
N
I
n
th
i
s
p
ap
er
,
w
e
h
a
v
e
p
r
o
p
o
s
ed
a
g
ated
r
ec
u
r
r
en
t
u
n
it
n
et
w
o
r
k
a
n
d
r
ep
lace
its
tr
ad
itio
n
al
s
o
f
t
m
a
x
o
u
tp
u
t
la
y
er
w
i
th
SVM
f
o
r
te
x
t
class
if
icatio
n
.
W
e
h
a
v
e
p
er
f
o
r
m
ed
an
e
x
p
er
i
m
en
t
o
n
t
h
e
“
well
-
k
n
o
w
n
C
h
in
e
s
e
co
r
p
u
s
d
atasets
to
e
v
al
u
ate
t
h
e
p
er
f
o
r
m
a
n
ce
o
f
t
h
e
G
R
U
-
SV
M.
T
h
e
d
ataset
i
s
p
u
b
licl
y
av
a
ilab
le
an
d
i
n
cl
u
d
ed
1
8
8
5
d
o
cu
m
e
n
t
s
f
o
r
tr
ain
i
n
g
an
d
9
4
0
d
o
cu
m
e
n
ts
f
o
r
test
in
g
.
T
h
e
r
esu
lts
h
av
e
b
ee
n
c
o
m
p
ar
ed
w
it
h
t
w
o
s
tate
-
of
-
th
e
-
ar
t
D
A
B
N
a
n
d
B
L
ST
M
-
C
m
o
d
els.
T
h
r
o
u
g
h
th
e
p
r
ac
ticall
y
d
ata
e
x
p
er
im
en
t,
th
e
p
r
o
p
o
s
ed
GR
U
-
SVM
m
o
d
el
ac
h
iev
ed
th
e
b
est
tex
t
class
if
ica
tio
n
ac
cu
r
ac
y
/p
er
f
o
r
m
a
n
ce
r
ate
o
f
9
4
.
7
5
%,
w
h
ic
h
d
em
o
n
s
tr
ated
th
e
p
o
ten
tial
o
f
i
m
p
le
m
en
tin
g
GR
U
-
SVM
o
n
m
u
l
ti
-
c
lass
cla
s
s
i
f
icatio
n
is
s
u
es
w
it
h
a
s
m
al
l
n
u
m
b
er
o
f
clas
s
es.
F
u
r
th
er
m
o
r
e,
o
u
r
p
r
o
p
o
s
ed
m
et
h
o
d
ac
h
iev
ed
m
u
c
h
b
etter
p
er
f
o
r
m
a
n
ce
in
t
h
e
ter
m
o
f
p
r
ec
is
io
n
,
r
ec
all
an
d
F1
th
a
n
DA
B
N
a
n
d
B
L
ST
M
-
C
,
p
ar
tic
u
lar
l
y
w
h
e
n
th
e
s
ize
o
f
t
h
e
s
to
r
ag
e
li
m
ited
.
ACK
NO
WL
E
D
G
E
M
E
NT
S
T
h
e
au
th
o
r
s
w
o
u
ld
li
k
e
to
t
h
an
k
s
Mi
n
i
s
tr
y
o
f
E
d
u
ca
tio
n
Ma
la
y
s
ia,
Un
iv
er
s
iti
T
u
n
Hu
s
s
ei
n
O
n
n
Ma
la
y
s
ia
a
n
d
R
e
s
ea
r
ch
Ma
n
a
g
e
m
e
n
t
C
en
ter
(
R
MC)
f
o
r
f
u
n
d
in
g
t
h
is
r
esear
c
h
ac
ti
v
it
y
u
n
d
er
th
e
Fu
n
d
u
m
e
n
tal
R
esear
ch
Gr
a
n
t Sc
h
e
m
e
(
FR
G
S),
v
o
te
n
o
.
1
6
4
1
.
RE
F
E
R
E
NC
E
S
[1
]
G
.
Tzo
rtzis
a
n
d
A
.
L
ik
a
s,
“
De
e
p
Be
li
e
f
Ne
t
w
o
rk
s
f
o
r
S
p
a
m
F
il
terin
g
,
”
1
9
th
IEE
E
In
t.
Co
n
f.
T
o
o
ls
wit
h
Art
if
.
I
n
tell
.
2
0
0
7
)
,
p
p
.
3
0
6
–
3
0
9
,
2
0
0
7
.
[2
]
Y.
Nin
g
e
t
a
l.
,
“
M
u
lt
i
-
tas
k
d
e
e
p
lea
rn
in
g
f
o
r
u
se
r
in
ten
ti
o
n
u
n
d
e
rsta
n
d
in
g
in
sp
e
e
c
h
in
tera
c
ti
o
n
sy
ste
m
s,”
3
1
st A
AA
I
Co
n
f.
Arti
f.
In
tell
.
AA
AI
2
0
1
7
,
p
p
.
1
6
1
–
1
6
7
,
2
0
1
7
.
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:
2
0
8
8
-
8708
Text
cla
s
s
ific
a
tio
n
b
a
s
ed
o
n
g
a
ted
r
ec
u
r
r
en
t u
n
it
co
mb
in
es
w
ith
… (
Mu
h
a
mma
d
Zu
lq
a
r
n
a
i
n
)
3741
[3
]
M
.
Z
u
lq
a
rn
a
i
n
,
R
.
G
h
a
z
a
li
,
Y.
M
.
M
.
Ha
ss
i
m
,
a
n
d
M
.
Re
h
a
n
,
“
A
c
o
m
p
a
ra
ti
v
e
re
v
ie
w
o
n
d
e
e
p
lea
r
n
in
g
m
o
d
e
ls
f
o
r
tex
t
c
las
si
f
ica
ti
o
n
,
”
In
d
o
n
e
s.
J
.
El
e
c
tr.
En
g
.
Co
mp
u
t.
S
c
i.
,
v
o
l.
1
9
,
n
o
.
1
,
p
p
.
1
8
5
6
–
1
8
6
6
,
2
0
2
0
.
[4
]
R.
G
h
a
z
a
li
,
N.
A
.
Hu
sa
in
i,
L
.
H.
Ism
a
il
,
T
.
He
ra
w
a
n
,
a
n
d
Y.
M
.
M
.
Ha
ss
im
,
“
T
h
e
p
e
r
f
o
rm
a
n
c
e
o
f
a
Re
c
u
rre
n
t
HO
NN
f
o
r
te
m
p
e
ra
tu
re
ti
m
e
se
rie
s p
re
d
ictio
n
,
”
Pr
o
c
.
In
t.
J
t
.
C
o
n
f.
Ne
u
ra
l
Ne
two
rk
s
,
n
o
.
Ju
ly
,
p
p
.
5
1
8
–
5
2
4
,
2
0
1
4
.
[5
]
K.
J.
Oh
,
W
.
J.
L
e
e
,
C.
G
.
L
i
m
,
a
n
d
H.
J.
Ch
o
i,
“
P
e
rs
o
n
a
li
z
e
d
n
e
ws
re
c
o
m
m
e
n
d
a
ti
o
n
u
sin
g
c
las
sif
ie
d
k
e
y
w
o
rd
s
to
c
a
p
tu
re
u
se
r
p
re
f
e
re
n
c
e
,
”
In
t.
Co
n
f.
Ad
v
.
Co
mm
u
n
.
T
e
c
h
n
o
l.
ICACT
,
p
p
.
1
2
8
3
–
1
2
8
7
,
2
0
1
4
.
[6
]
G
.
M
u
jt
a
b
a
,
L
.
S
h
u
ib
,
R.
G
.
Ra
j,
N.
M
a
jee
d
,
a
n
d
M
.
A
.
A
l
-
Ga
ra
d
i,
“
Ema
il
Clas
sif
ic
a
ti
o
n
Re
se
a
rc
h
T
re
n
d
s:
Re
v
ie
w
a
n
d
Op
e
n
Iss
u
e
s,”
IEE
E
Acc
e
ss
,
v
o
l.
5
,
p
p
.
9
0
4
4
–
9
0
6
4
,
2
0
1
7
.
[7
]
S
.
M
.
S
a
q
i
b
,
F
.
M
.
Ku
n
d
i,
a
n
d
S
.
A
h
m
a
d
,
“
Un
su
p
e
rv
ise
d
L
e
a
rn
in
g
M
e
th
o
d
f
o
r
S
o
r
ti
n
g
P
o
siti
v
e
a
n
d
Ne
g
a
ti
v
e
Re
v
ie
w
s
Us
in
g
L
S
I
(L
a
t
e
n
t
S
e
m
a
n
ti
c
In
d
e
x
in
g
)
w
it
h
A
u
to
m
a
ti
c
G
e
n
e
r
a
ted
Qu
e
ries
,
”
IJ
CS
NS
In
t.
J
.
Co
mp
u
t.
S
c
i.
Ne
tw.
S
e
c
u
r.
,
v
o
l
.
1
8
,
n
o
.
1
,
p
p
.
5
6
–
6
2
,
2
0
1
8
.
[8
]
M
.
Zu
l
q
a
rn
a
in
,
R.
G
h
a
z
a
li
,
S
.
H.
Kh
a
lee
f
a
h
,
a
n
d
A
.
Re
h
a
n
,
“
A
n
I
m
p
ro
v
e
d
th
e
P
e
rf
o
r
m
a
n
c
e
o
f
GR
U
M
o
d
e
l
b
a
se
d
o
n
Ba
tch
No
rm
a
li
z
a
ti
o
n
f
o
r
S
e
n
te
n
c
e
Clas
sif
i
c
a
ti
o
n
,
”
In
t.
J
.
C
o
mp
u
t.
S
c
i.
Ne
tw.
S
e
c
u
r.
,
v
o
l.
1
9
,
n
o
.
9
,
p
p
.
1
7
6
–
1
8
6
,
2
0
1
9
.
[9
]
V
.
Ko
rd
e
a
n
d
C.
N.
M
a
h
e
n
d
e
r,
“
T
e
x
t
Clas
si
f
ic
a
ti
o
n
a
n
d
Clas
sif
ier
s:
A
S
u
rv
e
y
,
”
In
t.
J
.
Arti
f.
In
tell.
Ap
p
l
.
,
v
o
l.
3
,
n
o
.
2
,
p
p
.
8
5
–
9
9
,
2
0
1
2
.
[1
0
]
W
.
S
h
a
rif
,
N.
A
.
S
a
m
su
d
in
,
M
.
M
.
De
ris,
a
n
d
M
.
A
a
m
ir,
“
I
m
p
ro
v
e
d
re
lativ
e
d
isc
rim
in
a
ti
v
e
c
rit
e
rio
n
f
e
a
tu
re
ra
n
k
in
g
tec
h
n
iq
u
e
f
o
r
tex
t
c
las
sif
ic
a
ti
o
n
,
”
In
t.
J
.
Arti
f.
I
n
tell.
,
v
o
l.
1
5
,
n
o
.
2
,
p
p
.
6
1
–
7
8
,
2
0
1
7.
[1
1
]
R.
G
h
a
z
a
li
,
Z.
A
.
Ba
k
a
r,
Y.
M
a
z
w
in
,
a
n
d
M
.
Ha
ss
im
,
“
F
u
n
c
ti
o
n
a
l
L
in
k
Ne
u
ra
l
Ne
tw
o
rk
w
it
h
M
o
d
if
ied
Cu
c
k
o
o
S
e
a
rc
h
T
ra
in
in
g
A
lg
o
rit
h
m
,
”
In
t.
Co
n
f.
In
tell
.
Co
m
p
u
t
.
,
p
p
.
2
8
5
–
2
9
1
,
2
0
1
4
.
[1
2
]
J.
L
iu
,
W
.
-
C.
Ch
a
n
g
,
Y
.
W
u
,
a
n
d
Y.
Ya
n
g
,
“
De
e
p
L
e
a
rn
in
g
f
o
r
Ex
tr
e
m
e
M
u
lt
i
-
lab
e
l
T
e
x
t
Clas
sifica
ti
o
n
,
”
Pr
o
c
.
4
0
t
h
In
t.
ACM
S
IGIR
Co
n
f.
Res
.
De
v
.
In
f.
Retr.
-
S
IGIR
’1
7
,
p
p
.
1
1
5
–
1
2
4
,
2
0
1
7
.
[1
3
]
B.
He
,
Y.
G
u
a
n
,
a
n
d
R.
Da
i,
“
Co
n
v
o
lu
ti
o
n
a
l
G
a
ted
Re
c
u
rre
n
t
Un
it
s
f
o
r
M
e
d
ica
l
Re
latio
n
Clas
si
f
ica
t
io
n
,
”
J
.
M
a
c
h
.
L
e
a
rn
.
Res
.
,
n
o
.
j
u
ly
,
p
p
.
1
1
8
–
1
2
7
,
2
0
1
8
.
[1
4
]
L
.
W
e
i,
B.
Wei,
a
n
d
B.
W
a
n
g
,
“
T
e
x
t
Clas
si
f
ica
ti
o
n
Us
in
g
S
u
p
p
o
r
t
V
e
c
to
r
M
a
c
h
in
e
w
it
h
M
ix
tu
re
o
f
Ke
rn
e
l,
”
A
J
.
S
o
ft
w
.
En
g
.
Ap
p
l.
,
v
o
l.
2
0
1
2
,
n
o
.
De
c
e
m
b
e
r,
p
p
.
5
5
–
5
8
,
2
0
1
2
.
[1
5
]
M
.
G
o
u
d
j
il
,
M
.
K
o
u
d
il
,
M
.
Be
d
d
a
,
a
n
d
N.
G
h
o
g
g
a
li
,
“
A
No
v
e
l
A
c
ti
v
e
Lea
rn
in
g
M
e
th
o
d
Us
i
n
g
S
V
M
f
o
r
T
e
x
t
Clas
sif
ic
a
ti
o
n
,
”
In
t
.
J
.
A
u
to
m
.
Co
mp
u
t.
,
v
o
l.
1
5
,
n
o
.
3
,
p
p
.
2
9
0
–
2
9
8
,
2
0
1
8
.
[1
6
]
M
.
K.
S
.
V
a
rm
a
,
“
P
ix
e
l
-
b
a
se
d
Clas
sif
i
c
a
ti
o
n
Us
in
g
S
u
p
p
o
rt
V
e
c
to
r
M
a
c
h
in
e
Clas
sif
ier,”
2
0
1
6
IEE
E
6
th
I
n
t.
Co
n
f.
Ad
v
.
Co
m
p
u
t
.
,
p
p
.
5
1
–
5
5
,
2
0
1
6
.
[1
7
]
V
.
Ho
o
sh
m
a
n
d
M
o
g
h
a
d
d
a
m
a
n
d
J.
Ha
m
id
z
a
d
e
h
,
“
Ne
w
He
r
m
it
e
o
rth
o
g
o
n
a
l
p
o
ly
n
o
m
ial
k
e
rn
e
l
a
n
d
c
o
m
b
in
e
d
k
e
rn
e
ls i
n
S
u
p
p
o
rt
V
e
c
to
r
M
a
c
h
in
e
c
las
si
f
ier,”
Pa
tt
e
rn
Rec
o
g
n
it
.
,
v
o
l.
6
0
,
p
p
.
9
2
1
–
9
3
5
,
2
0
1
6
.
[1
8
]
G
.
In
d
ra
wa
n
,
I.
K.
P
.
S
u
d
iars
a
,
a
n
d
K.
A
g
u
stin
i,
“
S
m
o
o
th
S
u
p
p
o
r
t
V
e
c
to
r
M
a
c
h
i
n
e
f
o
r
S
u
icid
e
-
Re
l
a
ted
Be
h
a
v
io
u
rs
P
re
d
ictio
n
,
”
In
t.
J
.
E
lec
tr.
Co
mp
u
t.
En
g
.
,
v
o
l.
8
,
n
o
.
5
,
p
p
.
3
3
9
9
–
3
4
0
6
,
2
0
1
8
.
[1
9
]
W
.
S
h
a
rif
,
U.
T
u
n
,
H.
On
n
,
I.
T
ri
,
a
n
d
R.
Ya
n
t
o
,
“
A
n
Op
ti
m
ise
d
S
u
p
p
o
rt
V
e
c
to
r
M
a
c
h
i
n
e
w
it
h
Rin
g
e
d
S
e
a
l
S
e
a
rc
h
A
l
g
o
rit
h
m
f
o
r
Eff
ici
e
n
t
T
e
x
t
Clas
sif
ica
ti
o
n
,
”
J
.
En
g
.
S
c
i
.
T
e
c
h
n
o
l.
,
v
o
l.
1
4
,
n
o
.
3
,
p
p
.
1
6
0
1
–
1
6
1
3
,
2
0
1
9
.
[2
0
]
K.
Ch
o
e
t
a
l.
,
“
L
e
a
rn
in
g
P
h
ra
s
e
Re
p
re
se
n
tatio
n
s
u
si
n
g
RNN
En
c
o
d
e
r
-
De
c
o
d
e
r
f
o
r
S
tatisti
c
a
l
M
a
c
h
in
e
T
r
a
n
sla
ti
o
n
,
”
a
rXiv
,
n
o
.
S
e
p
tem
b
e
r,
p
p
.
1
–
1
5
,
2
0
1
4
.
[2
1
]
M
.
Z
u
lq
a
rn
a
i
n
,
S
.
A
.
Ish
a
k
,
R.
Gh
a
z
a
li
,
a
n
d
N.
M
.
Na
w
i,
“
A
n
I
m
p
ro
v
e
d
De
e
p
L
e
a
rn
in
g
A
p
p
ro
a
c
h
b
a
se
d
o
n
V
a
rian
t
Tw
o
-
S
tate
G
a
ted
Re
c
u
rr
e
n
t
Un
it
a
n
d
W
o
rd
Em
b
e
d
d
in
g
s
f
o
r
S
e
n
ti
m
e
n
t
Clas
sif
i
c
a
ti
o
n
,
”
In
t.
J
.
Ad
v
.
Co
mp
u
t
.
S
c
i.
Ap
p
l
.
,
v
o
l.
1
1
,
n
o
.
1
,
p
p
.
5
9
4
–
6
0
3
,
2
0
2
0
.
[2
2
]
A
.
A
lalsh
e
k
m
u
b
a
ra
k
a
n
d
L
.
S
.
S
m
it
h
,
“
A
No
v
e
l
A
p
p
ro
a
c
h
Co
m
b
in
in
g
Re
c
u
rre
n
t
Ne
u
ra
l
Ne
t
wo
rk
a
n
d
S
u
p
p
o
rt
V
e
c
to
r
M
a
c
h
in
e
s
F
o
r
T
i
m
e
S
e
rie
s
Clas
sif
ic
a
ti
o
n
,
”
2
0
1
3
9
th
In
t
.
C
o
n
f.
In
n
o
v
.
I
n
f.
T
e
c
h
n
o
l.
,
n
o
.
A
u
g
u
st,
p
p
.
4
2
–
4
7
,
2
0
1
3
.
[2
3
]
M
.
Zu
l
q
a
rn
a
in
,
R.
G
h
a
z
a
li
,
M
.
G
.
G
h
o
u
se
,
a
n
d
M
.
F
.
M
u
sh
taq
,
“
Eff
i
c
ien
t
P
ro
c
e
ss
in
g
o
f
G
RU
Ba
se
d
o
n
W
o
rd
Em
b
e
d
d
in
g
f
o
r
T
e
x
t
Clas
sif
ic
a
ti
o
n
,
”
In
t
.
J
.
I
n
fo
rm
a
ti
c
s V
is.
,
v
o
l.
3
,
n
o
.
4
,
p
p
.
3
7
7
–
3
8
3
,
2
0
1
9
.
[2
4
]
Z.
Ya
n
g
a
n
d
X
.
P
a
n
g
,
“
Re
se
a
r
c
h
a
n
d
Im
p
le
m
e
n
tatio
n
o
f
T
e
x
t
Clas
sif
ica
ti
o
n
M
o
d
e
l
Ba
se
d
o
n
C
o
m
b
in
a
ti
o
n
o
f
D
A
E
a
n
d
DBN
,
”
2
0
1
7
1
0
t
h
In
t.
S
y
mp
.
Co
mp
u
t
.
In
tell.
De
s.
,
p
p
.
1
9
0
–
1
9
3
,
2
0
1
7
.
[2
5
]
Y.
L
i,
X
.
W
a
n
g
,
a
n
d
P
.
X
u
,
“
Ch
in
e
se
T
e
x
t
Clas
si
f
ica
ti
o
n
M
o
d
e
l
Ba
se
d
o
n
De
e
p
L
e
a
rn
in
g
,
”
J
.
Ki
n
g
S
a
u
d
Un
iv.
-
Co
mp
u
t
.
In
f.
S
c
i.
,
n
o
.
No
v
e
m
b
e
r,
p
p
.
2
0
–
3
2
,
2
0
1
8
.
B
I
O
G
RAP
H
I
E
S
O
F
AUTH
O
RS
M
u
h
a
m
m
a
d
Zu
l
q
a
r
n
a
in
re
c
e
iv
e
d
h
is
Ba
c
h
e
lo
r
a
n
d
M
a
s
ter
d
e
g
re
e
in
Co
m
p
u
ter
S
c
ien
c
e
a
n
d
In
f
o
rm
a
ti
o
n
T
e
c
h
n
o
l
o
g
y
f
ro
m
T
h
e
Isla
m
ia
Un
iv
e
rsit
y
o
f
Ba
h
a
wa
lp
u
r
(IUB),
P
a
k
istan
.
He
re
c
e
iv
e
d
h
is
M
.
P
h
il
d
e
g
re
e
(M
a
ste
r
o
f
P
h
i
lo
so
p
h
y
)
f
ro
m
Na
ti
o
n
a
l
Co
ll
e
g
e
o
f
Bu
sin
e
ss
A
d
m
in
istratio
n
a
n
d
Eco
n
o
m
ics
,
L
a
h
o
re
,
P
a
k
istan
.
H
e
is
c
u
rre
n
tl
y
p
u
rsu
in
g
P
h
.
D
f
ro
m
Un
iv
e
rsit
y
T
u
n
Hu
ss
e
in
On
n
M
a
la
y
sia
.
His
re
se
a
rc
h
in
tere
st
is
M
a
c
h
in
e
L
e
a
rn
in
g
a
n
d
De
e
p
lea
rn
in
g
f
o
r
n
a
tu
ra
l
la
n
g
u
a
g
e
p
ro
c
e
ss
in
g
a
n
d
it
s a
p
p
li
c
a
ti
o
n
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
0
8
8
-
8708
I
n
t J
E
lec
&
C
o
m
p
E
n
g
,
Vo
l.
10
,
No
.
4
,
A
u
g
u
s
t 2
0
2
0
:
3
7
3
4
-
3742
3742
Ro
z
a
i
d
a
G
h
a
z
a
li
is
c
u
rre
n
tl
y
a
P
r
o
f
e
ss
o
r
a
t
th
e
F
a
c
u
lt
y
o
f
Co
m
p
u
ter
S
c
ien
c
e
a
n
d
I
n
f
o
rm
a
ti
o
n
T
e
c
h
n
o
lo
g
y
,
Un
iv
e
r
siti
T
u
n
Hu
ss
e
in
On
n
M
a
lay
sia
(U
T
HM).
S
h
e
g
ra
d
u
a
ted
w
it
h
a
P
h
.
D.
d
e
g
re
e
in
Hig
h
e
r
Ord
e
r
Ne
u
ra
l
Ne
t
w
o
rk
s
f
ro
m
th
e
S
c
h
o
o
l
o
f
Co
m
p
u
ti
n
g
a
n
d
M
a
th
e
m
a
ti
c
a
l
S
c
ien
c
e
s
a
t
L
iv
e
rp
o
o
l
Jo
h
n
M
o
o
re
s
Un
iv
e
rsity
,
Un
it
e
d
Kin
g
d
o
m
in
2
0
0
7
.
Earl
ier,
in
2
0
0
3
sh
e
c
o
m
p
lete
d
h
e
r
M
.
S
c
.
d
e
g
re
e
in
Co
m
p
u
ter
S
c
ien
c
e
f
ro
m
Un
iv
e
rsiti
Tek
n
o
lo
g
i
M
a
la
y
sia
(U
T
M
).
S
h
e
re
c
e
iv
e
d
h
e
r
B.
S
c
.
(Ho
n
s)
d
e
g
re
e
in
Co
m
p
u
ter S
c
ien
c
e
f
ro
m
Un
iv
e
rsiti
S
a
in
s M
a
la
y
sia
(USM
)
in
1
9
9
7
.
I
n
2
0
0
1
,
Ro
z
a
id
a
jo
i
n
e
d
t
h
e
a
c
a
d
e
m
ic
sta
ff
in
UT
HM.
He
r
re
se
a
rc
h
a
re
a
in
c
lu
d
e
s
n
e
u
ra
l
n
e
tw
o
rk
s,
s
w
a
r
m
in
telli
g
e
n
c
e
,
o
p
ti
m
iza
ti
o
n
,
d
a
ta
m
in
in
g
,
a
n
d
ti
m
e
se
ries
p
re
d
ictio
n
.
S
h
e
h
a
s
su
c
c
e
ss
f
u
ll
y
su
p
e
rv
ise
d
a
n
u
m
b
e
r
o
f
P
h
D
a
n
d
m
a
ste
r
stu
d
e
n
ts
a
n
d
p
u
b
li
sh
e
d
m
o
re
th
a
n
1
0
0
a
rti
c
les
in
v
a
rio
u
s
in
ter
n
a
ti
o
n
a
l
jo
u
r
n
a
ls
a
n
d
c
o
n
f
e
re
n
c
e
p
ro
c
e
e
d
in
g
s.
S
h
e
a
c
ts
a
s
a
re
v
ie
w
e
r
f
o
r
v
a
rio
u
s
jo
u
rn
a
ls
a
n
d
c
o
n
f
e
re
n
c
e
s,
a
n
d
a
s
a
n
e
d
it
o
r
in
a
f
e
w
S
p
rin
g
e
r
c
o
n
f
e
r
e
n
c
e
p
ro
c
e
e
d
in
g
s.
S
h
e
h
a
s
a
lso
se
r
v
e
d
a
s
a
c
o
n
f
e
re
n
c
e
c
h
a
ir,
a
n
d
a
s a
tec
h
n
ica
l
c
o
m
m
it
tee
f
o
r
n
u
m
e
ro
u
s in
tern
a
ti
o
n
a
l
c
o
n
f
e
re
n
c
e
s.
Ya
n
a
M
a
z
w
in
M
o
h
m
a
d
H
a
ss
i
m
is
a
se
n
io
r
lec
tu
re
r
a
t
th
e
F
a
c
u
lt
y
o
f
Co
m
p
u
ter
S
c
ien
c
e
a
n
d
In
f
o
rm
a
ti
o
n
T
e
c
h
n
o
l
o
g
y
,
Un
iv
e
r
siti
T
u
n
Hu
ss
e
in
O
n
n
M
a
la
y
sia
(UT
HM).
S
h
e
g
ra
d
u
a
ted
w
it
h
a
P
h
D
d
e
g
re
e
f
ro
m
Un
iv
e
rsiti
T
u
n
Hu
ss
e
in
On
n
M
a
lay
sia
(U
T
H
M
)
in
2
0
1
6
.
Earlier,
in
2
0
0
6
sh
e
c
o
m
p
let
e
d
h
e
r
M
a
ste
r
'
s
d
e
g
r
e
e
in
Co
m
p
u
ter
S
c
ien
c
e
f
ro
m
Un
iv
e
rsiti
o
f
M
a
la
y
a
(UM)
.
S
h
e
re
c
e
iv
e
d
h
e
r
Ba
c
h
e
lo
r
o
f
I
n
f
o
rm
a
ti
o
n
T
e
c
h
n
o
lo
g
y
(
Ho
n
s)
d
e
g
re
e
m
a
jo
rin
g
in
In
d
u
stria
l
Co
m
p
u
ti
n
g
f
ro
m
Un
iv
e
rsiti
Ke
b
a
n
g
sa
a
n
M
a
la
y
sia
(UK
M
)
in
2
0
0
1
.
In
2
0
0
3
,
Ya
n
a
M
a
z
w
in
jo
in
e
d
th
e
a
c
a
d
e
m
ic
sta
ff
in
U
T
HM.
H
e
r
re
se
a
rc
h
a
re
a
in
c
lu
d
e
s
n
e
u
ra
l
n
e
tw
o
rk
s,
s
w
a
r
m
in
telli
g
e
n
c
e
,
o
p
ti
m
iza
ti
o
n
a
n
d
c
las
sif
ica
ti
o
n
.
M
u
h
a
m
m
a
d
Re
h
a
n
is
a
n
A
ss
istan
t
p
r
o
f
e
ss
o
r
a
t
th
e
De
p
a
rtme
n
t
o
f
Co
m
p
u
ter
S
c
ien
c
e
a
n
d
In
f
o
rm
a
ti
o
n
T
e
c
h
n
o
lo
g
y
,
Ba
q
a
i
M
e
d
ica
l
Un
iv
e
r
sity
(BM
U)
K
a
r
a
c
h
i,
P
a
k
istan
.
He
re
c
e
iv
e
d
h
is
M
.
P
h
i
l.
De
g
re
e
(M
a
ste
r
o
f
P
h
il
o
so
p
h
y
)
f
ro
m
th
e
F
a
c
u
lt
y
o
f
Co
m
p
u
ter
S
c
ien
c
e
a
n
d
In
f
o
rm
a
ti
o
n
T
e
c
h
n
o
lo
g
y
,
Ha
m
d
a
rd
Un
iv
e
rsit
y
c
o
re
a
re
a
is
Co
m
p
u
ter
Ne
t
w
o
rk
s
in
2
0
1
6
.
Ka
ra
c
h
i,
P
a
k
istan
.
He
re
c
e
i
v
e
d
h
is
Ba
c
h
e
lo
r
o
f
C
o
m
p
u
ter
S
c
ien
c
e
d
e
g
r
e
e
m
a
jo
rin
g
in
S
o
f
tw
a
re
En
g
in
e
e
rin
g
f
ro
m
Da
d
a
b
h
o
y
In
stit
u
te
o
f
Hig
h
e
r
Ed
u
c
a
ti
o
n
(DIH
E)
Ka
ra
c
h
i,
P
a
k
istan
in
2
0
0
4
.
He
is
c
u
rre
n
tl
y
p
u
rsu
i
n
g
h
is P
h
.
D.
f
ro
m
Un
iv
e
rsity
T
u
n
Hu
ss
e
in
On
n
M
a
lay
sia
(U
T
HM).
His r
e
se
a
r
c
h
a
re
a
is
D
a
ta
M
in
i
n
g
a
n
d
M
a
c
h
in
e
L
e
a
rn
in
g
f
o
r
Op
ti
m
iza
ti
o
n
a
n
d
it
s a
p
p
li
c
a
ti
o
n
.
Evaluation Warning : The document was created with Spire.PDF for Python.