I
nte
rna
t
io
na
l J
o
urna
l o
f
E
lect
rica
l a
nd
Co
m
pu
t
er
E
ng
ineering
(
I
J
E
CE
)
Vo
l.
1
6
,
No
.
4
,
A
u
g
u
s
t
20
2
6
,
p
p
.
2
0
6
1
~
2
0
7
3
I
SS
N:
2088
-
8
7
0
8
,
DOI
: 1
0
.
1
1
5
9
1
/ijece.
v
1
6
i
4
.
pp
2
0
6
1
-
2
0
7
3
2061
J
o
ur
na
l ho
m
ep
a
g
e
:
h
ttp
:
//ij
ec
e.
ia
esco
r
e.
co
m
A
hy
brid
co
nten
t
a
nd chara
cter
fea
ture
m
ethod
for
S
M
S
spa
m
detec
tion
G
er
t
rude
Sela
s
e
G
o
s
u
1
,
E
d
w
a
rd
Appa
u Nk
et
ia
h
2
,
L
i Wa
ng
1
,
J
o
s
hu
a
F
enuk
u
3
,
Xia
o
y
a
Xu
1
1
C
o
l
l
e
g
e
o
f
C
o
m
p
u
t
e
r
S
c
i
e
n
c
e
a
n
d
Te
c
h
n
o
l
o
g
y
,
Ta
i
y
u
a
n
U
n
i
v
e
r
s
i
t
y
o
f
T
e
c
h
n
o
l
o
g
y
,
Ta
i
y
u
a
n
,
C
h
i
n
a
2
C
o
l
l
e
g
e
o
f
M
a
t
h
e
ma
t
i
c
s,
T
a
i
y
u
a
n
U
n
i
v
e
r
si
t
y
o
f
T
e
c
h
n
o
l
o
g
y
,
Ta
i
y
u
a
n
,
C
h
i
n
a
3
C
o
l
l
e
g
e
o
f
M
e
c
h
a
n
i
c
a
l
En
g
i
n
e
e
r
i
n
g
,
Ta
i
y
u
a
n
U
n
i
v
e
r
si
t
y
o
f
Te
c
h
n
o
l
o
g
y
,
T
a
i
y
u
a
n
,
C
h
i
n
a
Art
icle
I
nfo
AB
S
T
RAC
T
A
r
ticle
his
to
r
y:
R
ec
eiv
ed
Feb
1
0
,
2
0
2
6
R
ev
is
ed
J
u
n
6
,
2
0
2
6
Acc
ep
ted
J
u
l 2
2
,
2
0
2
6
S
h
o
rt
m
e
ss
a
g
e
se
rv
ice
(S
M
S
)
sp
a
m
re
m
a
in
s
a
sig
n
ifi
c
a
n
t
c
h
a
ll
e
n
g
e
d
u
e
t
o
it
s
imp
a
c
t
o
n
u
se
r
se
c
u
rit
y
a
n
d
c
o
m
m
u
n
ica
ti
o
n
e
fficie
n
c
y
.
T
h
is
stu
d
y
p
ro
p
o
se
s
a
h
y
b
rid
sp
a
m
d
e
tec
ti
o
n
m
o
d
e
l,
c
o
n
v
o
lu
ti
o
n
a
l
n
e
u
ra
l
n
e
tw
o
rk
wi
th
c
o
n
ten
t
a
n
d
c
h
a
ra
c
ter
-
b
a
se
d
fe
a
tu
re
s
(
CNN
-
CCB
)
,
wh
ich
i
n
teg
ra
tes
wo
rd
-
lev
e
l
fe
a
tu
re
s
a
n
d
c
h
a
ra
c
ter
-
lev
e
l
fe
a
tu
re
s,
with
h
a
n
d
c
ra
fted
c
o
n
te
n
t
-
b
a
se
d
a
n
d
c
h
a
ra
c
ter
-
b
a
se
d
fe
a
tu
re
s
in
a
u
n
ifi
e
d
d
e
e
p
lea
rn
i
n
g
fra
m
e
wo
rk
.
Un
li
k
e
c
o
n
v
e
n
ti
o
n
a
l
CNN
a
n
d
h
y
b
rid
m
o
d
e
ls
t
h
a
t
re
ly
p
r
ima
ril
y
o
n
lea
rn
e
d
re
p
re
se
n
tatio
n
s,
t
h
e
p
ro
p
o
se
d
a
p
p
ro
a
c
h
in
c
o
rp
o
ra
tes
stru
c
tu
ra
l
f
e
a
tu
re
s
to
e
n
h
a
n
c
e
d
e
tec
ti
o
n
o
f
s
h
o
rt
a
n
d
n
o
isy
tex
t
p
a
tt
e
rn
s.
F
irstl
y
,
t
h
e
tex
t
d
a
ta
we
re
to
k
e
n
ize
d
a
n
d
p
ro
c
e
ss
e
d
th
r
o
u
g
h
d
u
a
l
c
o
n
v
o
l
u
ti
o
n
a
l
b
ra
n
c
h
e
s,
wh
il
e
h
a
n
d
c
ra
fted
fe
a
tu
re
s
a
re
fu
se
d
to
imp
ro
v
e
c
las
sifi
c
a
ti
o
n
.
S
e
c
o
n
d
ly
,
c
las
s
we
ig
h
ti
n
g
is
a
p
p
li
e
d
t
o
a
d
d
re
ss
d
a
ta
imb
a
lan
c
e
wh
il
e
m
a
in
tain
i
n
g
p
re
d
ictiv
e
re
li
a
b
il
it
y
.
F
in
a
ll
y
,
re
g
u
lariz
a
ti
o
n
tec
h
n
i
q
u
e
s
a
re
e
m
p
l
o
y
e
d
t
o
p
re
v
e
n
t
o
v
e
rfit
ti
n
g
.
Th
e
e
x
p
e
rime
n
tal
re
su
lt
s
sh
o
w
th
a
t
CNN
-
CCB
a
c
h
iev
e
s
h
ig
h
p
e
rfo
rm
a
n
c
e
,
wit
h
a
n
a
c
c
u
ra
c
y
o
f
0
.
9
9
7
,
p
re
c
isio
n
o
f
0
.
9
8
8
,
re
c
a
ll
o
f
0
.
9
9
8
,
a
n
d
F
1
-
sc
o
re
o
f
0
.
9
9
3
,
o
u
tp
e
rfo
r
m
in
g
b
a
se
li
n
e
m
o
d
e
ls
su
c
h
a
s
lo
n
g
sh
o
rt
-
term
m
e
m
o
ry
(LS
T
M
)
d
a
n
g
a
ted
re
c
u
rre
n
t
u
n
it
(G
RU)
.
T
h
e
m
o
d
e
l
d
e
m
o
n
stra
tes
c
o
n
siste
n
t
p
e
rf
o
rm
a
n
c
e
a
c
ro
ss
th
re
e
d
a
tas
e
ts,
in
d
i
c
a
ti
n
g
t
h
e
m
o
d
e
l’s
ro
b
u
stn
e
ss
a
n
d
g
e
n
e
ra
l
iza
b
il
it
y
.
Th
e
se
fin
d
in
g
s
su
g
g
e
s
t
th
a
t
t
h
e
p
ro
p
o
se
d
h
y
b
ri
d
fra
m
e
wo
rk
is
e
ffe
c
ti
v
e
fo
r
S
M
S
sp
a
m
d
e
tec
ti
o
n
a
n
d
h
a
s
p
o
ten
ti
a
l
a
p
p
li
c
a
ti
o
n
s
i
n
m
o
b
il
e
se
c
u
rit
y
a
n
d
re
a
l
-
ti
m
e
c
o
m
m
u
n
ica
ti
o
n
sy
ste
m
s.
K
ey
w
o
r
d
s
:
C
h
ar
ac
ter
f
ea
tu
r
es
C
o
n
ten
t f
ea
tu
r
es
Hy
b
r
id
s
p
am
d
etec
tio
n
m
o
d
el
Neu
r
al
n
etwo
r
k
s
SMS sp
am
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
:
L
i Wan
g
C
o
lleg
e
o
f
C
o
m
p
u
ter
Scien
ce
an
d
T
ec
h
n
o
lo
g
y
,
T
aiy
u
an
Un
i
v
er
s
ity
o
f
T
ec
h
n
o
lo
g
y
W
an
b
ailin
Dis
tr
ict,
T
aiy
u
an
,
Sh
an
x
i,
0
3
0
0
2
4
,
C
h
in
a
E
m
ail: w
an
g
li@
ty
u
t.e
d
u
.
cn
1.
I
NT
RO
D
UCT
I
O
N
Sh
o
r
t
m
ess
ag
e
s
er
v
ice
(
SM
S)
h
as
b
ec
o
m
e
in
cr
ea
s
in
g
ly
f
r
eq
u
e
n
t
an
d
r
ea
d
ily
ac
ce
s
s
ib
le
o
n
co
m
m
u
n
icatio
n
d
ev
ices.
T
h
is
h
as
b
ee
n
an
in
d
is
p
u
tab
le
m
e
d
iu
m
o
f
co
m
m
u
n
icatio
n
in
o
u
r
d
aily
liv
es.
“T
h
e
tech
n
o
lo
g
y
o
f
SMS
is
o
n
e
s
er
v
ice
th
at
is
m
o
s
tly
s
tab
le
an
d
wid
ely
u
s
ed
m
o
b
ile
co
m
m
u
n
i
ca
tio
n
m
eth
o
d
s
”
[
1
]
,
th
er
ef
o
r
e,
it
is
p
r
im
ar
ily
u
s
ed
b
y
in
d
iv
id
u
als
t
o
co
n
n
ec
t
to
o
n
e
an
o
th
er
.
Mo
r
e
in
d
iv
i
d
u
als wo
r
ld
wid
e
m
a
k
e
u
s
e
o
f
m
o
b
ile
p
h
o
n
es;
h
e
n
ce
,
t
h
ey
co
n
tin
u
o
u
s
ly
im
m
e
r
s
e
th
em
s
elv
es
in
s
en
d
in
g
tex
t
m
ess
ag
es.
I
n
th
e
y
ea
r
2
0
1
7
,
“th
e
wo
r
ld
s
en
t
8
.
3
tr
illi
o
n
te
x
t
m
ess
ag
es:
th
e
n
u
m
b
er
o
f
tex
t
m
ess
ag
es
s
en
t
m
o
n
th
ly
was
6
9
0
b
illi
o
n
”
[
2
]
,
m
ak
in
g
SMS
an
im
p
o
r
tan
t
co
m
m
u
n
icatio
n
to
o
l.
C
u
r
r
en
tly
,
ce
ll
p
h
o
n
e
u
s
er
s
ac
r
o
s
s
th
e
g
lo
b
e
h
av
e
i
n
cr
ea
s
ed
to
ab
o
u
t
9
0
.
9
3
%
o
f
th
e
w
o
r
ld
’
s
p
o
p
u
latio
n
[
3
]
,
a
n
d
c
o
m
m
u
n
icatio
n
in
to
d
ay
’
s
s
o
ci
ety
is
s
ig
n
if
ican
tly
d
o
m
in
ated
b
y
te
x
t
m
ess
ag
in
g
.
I
t
is
th
e
f
astes
t
way
to
s
en
d
an
d
r
ec
eiv
e
i
n
f
o
r
m
atio
n
.
T
h
e
u
s
e
o
f
SMS
h
as
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
0
8
8
-
8
7
0
8
I
n
t J E
lec
&
C
o
m
p
E
n
g
,
Vo
l.
1
6
,
No
.
4
,
Au
g
u
s
t
20
2
6
:
2
0
6
1
-
2073
2062
in
cr
ea
s
ed
o
v
er
th
e
y
ea
r
s
,
s
u
ch
th
at
“8
0
%
o
f
p
eo
p
le
d
o
n
o
t
r
ea
d
e
-
m
ails
,
wh
ile
9
8
%
o
f
ce
ll
p
h
o
n
e
u
s
er
s
r
ea
d
th
eir
SMS d
aily
”
[
4
]
.
Sp
am
tex
t
is
a
m
ess
ag
e
t
h
at
a
p
p
ea
r
s
to
b
e
r
ea
l
b
u
t
h
as
im
m
o
r
al
o
r
e
v
il
in
ten
tio
n
s
,
wh
ich
i
s
r
ec
eiv
ed
with
o
u
t
ask
in
g
o
r
s
o
licitin
g
f
o
r
it.
Sp
am
m
ess
ag
es
ar
e
u
n
wan
ted
b
u
t
ar
e
u
n
a
v
o
id
a
b
le
b
ec
au
s
e
o
f
m
o
d
e
r
n
co
m
m
u
n
icatio
n
tech
n
o
l
o
g
y
[
5
]
.
Sp
am
g
e
n
er
ates
p
r
o
b
le
m
s
in
p
eo
p
le’
s
liv
es,
ca
u
s
in
g
co
n
g
esti
o
n
o
n
co
m
m
u
n
icatio
n
s
er
v
ice,
an
d
c
au
s
in
g
p
u
b
lic
s
ec
u
r
ity
an
d
s
o
c
ial
s
tab
ilit
y
p
r
o
b
lem
s
,
wh
ich
n
eg
ativ
ely
in
f
lu
en
ce
th
e
cr
ed
ib
ilit
y
o
f
co
m
m
u
n
icati
o
n
s
y
s
tem
[
6
]
.
T
h
e
i
n
v
ad
e
r
s
teals secr
et
in
f
o
r
m
atio
n
b
y
s
en
d
in
g
s
p
am
m
ess
ag
es
to
v
ictim
s
.
I
n
n
o
ce
n
t
p
eo
p
le
wh
o
g
et
th
ese
m
ess
ag
es
f
r
o
m
h
ac
k
e
r
s
p
an
ic
f
o
llo
win
g
i
n
s
tr
u
ctio
n
s
f
r
o
m
th
e
h
ac
k
er
s
an
d
g
iv
i
n
g
o
u
t
th
eir
s
en
s
itiv
e
in
f
o
r
m
atio
n
,
lead
in
g
to
lo
s
s
o
f
ass
et
s
an
d
r
eso
u
r
ce
s
[
7
]
.
Peo
p
le
wh
o
s
en
d
s
p
am
m
ess
ag
es,
p
e
r
f
o
r
m
ce
r
tain
ac
tiv
ities
to
en
tice
u
s
er
s
in
to
d
o
i
n
g
w
h
at
th
ey
wan
t,
a
n
d
is
a
d
o
m
in
atin
g
p
r
o
b
lem
i
n
th
e
wo
r
ld
to
d
a
y
w
h
ich
h
as witn
ess
ed
an
in
cr
ea
s
e
in
r
ec
en
t tim
es
[
8
]
.
T
h
e
s
p
r
ea
d
o
f
s
p
am
m
ess
ag
es
ca
n
n
o
t
b
e
f
u
lly
co
n
tr
o
lled
,
h
en
ce
is
d
if
f
ic
u
lt
to
s
to
p
co
n
s
u
m
er
s
f
r
o
m
r
ec
eiv
in
g
th
em
[
9
]
.
“Sp
a
m
tex
t
m
o
v
ed
u
p
b
y
5
8
%
o
f
te
x
t
m
e
s
s
ag
es
in
2
0
2
2
,
1
6
%
o
f
p
eo
p
le
h
av
e
p
atr
o
n
ized
a
s
ca
m
m
er
’
s
tex
t
to
p
u
r
c
h
ase
a
s
er
v
ice,
an
d
an
esti
m
ated
1
m
illi
o
n
s
p
am
tex
ts
ar
e
s
en
t
ev
er
y
m
in
u
te
g
lo
b
ally
”
[
5
]
.
T
h
er
ef
o
r
e,
th
e
c
o
n
tin
u
o
u
s
im
p
lem
en
tatio
n
o
f
s
p
am
d
etec
tio
n
an
d
f
ilter
in
g
m
eth
o
d
s
till
p
er
s
is
ts
.
Un
lik
e
em
ail
o
r
s
o
cial
m
ed
ia
tex
ts
,
SMS
m
e
s
s
ag
es
ar
e
o
f
te
n
s
h
o
r
ter
,
n
o
is
ier
,
lin
g
u
is
ticall
y
d
iv
er
s
e,
an
d
m
ay
in
v
o
lv
e
f
r
eq
u
e
n
t
u
s
e
o
f
ab
b
r
e
v
iatio
n
s
,
co
llo
q
u
ialis
m
s
,
an
d
m
u
ltil
in
g
u
al
co
d
e
-
s
witch
in
g
.
SMS
s
p
am
m
ess
ag
es
in
f
lu
en
ce
u
s
er
s
s
ig
n
if
ican
tly
b
ec
au
s
e
ea
c
h
m
ess
ag
e
r
ec
eiv
ed
o
n
p
h
o
n
e
is
v
iewe
d
[
1
0
]
.
C
o
n
s
eq
u
en
tly
,
co
n
s
id
er
a
b
le
a
tten
tio
n
to
s
p
am
p
r
o
b
lem
s
is
ch
an
n
eled
to
war
d
s
d
ev
el
o
p
in
g
ap
p
r
o
ac
h
es
to
en
ab
le
u
s
er
s
escap
e
th
is
p
r
o
b
l
em
.
T
h
o
u
g
h
th
er
e
ar
e
s
o
m
e
p
r
io
r
s
tu
d
ies
o
n
s
p
am
d
etec
tio
n
,
th
ey
ar
e
m
ain
ly
f
o
cu
s
ed
o
n
s
in
g
le
m
o
d
els
[
1
1
]
–
[
1
3
]
,
m
ak
in
g
th
em
less
ef
f
icien
t
with
r
o
b
u
s
t
tex
t
d
at
a.
Oth
er
s
tu
d
ies
in
tr
o
d
u
ce
d
h
y
b
r
id
m
o
d
els
with
s
eq
u
en
tial
lear
n
in
g
[
1
]
,
[
1
4
]
,
b
u
t
m
an
y
o
f
th
ese
s
tu
d
ies
wer
e
f
o
cu
s
ed
o
n
o
n
l
y
wo
r
d
-
lev
el
f
ea
tu
r
es
o
r
ch
ar
ac
ter
-
lev
el
f
ea
tu
r
es,
with
th
e
u
s
e
o
f
wo
r
d
v
ec
to
r
,
h
ash
i
n
g
v
ec
t
o
r
izer
,
c
o
u
n
t
v
ec
to
r
ize
r
,
ter
m
f
r
eq
u
en
cy
-
in
v
er
s
e
d
o
cu
m
e
n
t
f
r
eq
u
e
n
cy
(
T
F
-
I
DF)
,
b
ag
-
of
-
wo
r
d
s
,
[
1
5
]
,
[
1
6
]
,
as
well
as
I
n
f
o
Gain
an
d
C
h
i
-
s
q
u
ar
e
[
1
7
]
.
T
h
is
p
o
s
itio
n
ed
o
u
r
wo
r
k
as
r
esp
o
n
s
e
b
y
p
r
o
p
o
s
in
g
a
n
o
v
el
ap
p
r
o
ac
h
th
at
in
teg
r
ates
o
u
r
h
an
d
cr
af
ted
co
n
ten
t
-
b
ased
an
d
c
h
ar
ac
ter
-
b
ased
f
ea
tu
r
es
to
e
n
h
an
ce
d
etec
tio
n
o
f
s
p
am
.
Un
lik
e
co
n
v
en
tio
n
al
d
ee
p
lear
n
in
g
ap
p
r
o
ac
h
es
th
at
r
ely
s
o
lely
o
n
wo
r
d
-
le
v
el
em
b
ed
d
i
n
g
s
,
o
u
r
m
eth
o
d
in
teg
r
ates
d
if
f
er
en
t
co
m
p
lem
en
tar
y
f
ea
tu
r
e
ty
p
es.
Mo
r
e
s
o
,
e
x
is
tin
g
SMS
s
p
am
d
etec
tio
n
m
o
d
els
wer
e
m
ain
ly
p
r
o
p
o
s
ed
b
ased
o
n
s
m
all
an
d
h
ig
h
ly
im
b
alan
ce
d
d
atasets
,
with
m
o
s
t
o
f
th
em
f
o
cu
s
in
g
o
n
e
-
m
ails
tex
t
d
ata
[
1
8
]
–
[
2
0
]
,
lim
itin
g
th
e
g
en
er
aliza
b
ilit
y
a
n
d
r
o
b
u
s
tn
ess
o
f
t
h
ese
m
o
d
els.
Hen
ce
,
we
cr
ea
ted
a
n
ew
d
ata
(
o
r
ig
in
al
d
ata)
f
r
o
m
d
if
f
er
e
n
t
SMS d
ata
s
am
p
les.
Ou
r
s
tu
d
y
p
r
o
p
o
s
ed
a
co
n
v
o
lu
tio
n
al
n
e
u
r
al
n
etwo
r
k
wit
h
co
n
te
n
t
an
d
ch
ar
ac
ter
-
b
ase
d
f
ea
tu
r
es
(
C
NN
-
C
C
B
)
m
o
d
el,
wh
ich
i
n
tr
o
d
u
ce
s
a
h
y
b
r
id
f
ea
tu
r
e
l
ea
r
n
in
g
f
r
am
ewo
r
k
.
T
h
e
f
r
a
m
ewo
r
k
in
te
g
r
ates
s
em
an
tic,
s
ty
lis
tic,
an
d
s
tr
u
ctu
r
al
f
ea
tu
r
es
f
o
r
im
p
r
o
v
e
d
SMS
s
p
am
d
etec
tio
n
.
T
h
e
m
ain
co
n
tr
ib
u
tio
n
s
o
f
th
is
s
tu
d
y
ar
e:
−
I
n
co
r
p
o
r
atio
n
o
f
h
an
d
cr
af
ted
f
ea
tu
r
es
in
to
d
ee
p
lear
n
i
n
g
:
we
in
teg
r
ated
co
n
ten
t
-
b
ased
an
d
ch
ar
ac
ter
-
b
ased
s
tatis
t
ical
f
ea
tu
r
es to
co
m
p
lem
en
t o
u
r
lea
r
n
ed
r
ep
r
esen
tatio
n
s
.
−
Du
al
-
r
ep
r
esen
tatio
n
lear
n
in
g
o
f
C
NN
a
r
ch
itectu
r
e:
we
d
esig
n
ed
a
p
ar
allel
C
NN
ar
c
h
itectu
r
e
th
at
ca
p
tu
r
es
b
o
th
wo
r
d
-
lev
el
a
n
d
ch
a
r
ac
ter
-
lev
el
p
atter
n
s
.
−
A
h
y
b
r
id
f
ea
tu
r
e
in
teg
r
atio
n
f
r
am
ewo
r
k
:
we
u
s
ed
a
u
n
if
ied
m
o
d
el
th
at
jo
in
tly
in
te
g
r
ates
wo
r
d
-
lev
el
an
d
ch
ar
ac
ter
-
lev
el
p
atter
n
s
,
with
co
n
ten
t
-
b
ased
a
n
d
c
h
ar
ac
ter
-
b
ased
s
tatis
t
ical
f
ea
tu
r
es.
−
C
o
m
p
r
eh
en
s
iv
e
em
p
ir
ical
v
ali
d
atio
n
:
we
co
n
d
u
cte
d
an
ex
te
n
s
iv
e
ev
alu
atio
n
u
s
in
g
s
tatis
ti
ca
l
s
ig
n
if
ican
ce
test
in
g
an
d
ab
latio
n
s
tu
d
y
to
q
u
an
tify
th
e
c
o
n
tr
ib
u
tio
n
o
f
ea
c
h
f
ea
tu
r
e
t
y
p
e.
Un
lik
e
tr
ad
itio
n
al
C
NN
-
b
ase
d
s
p
am
d
etec
tio
n
m
o
d
els
th
at
r
ely
s
o
lely
o
n
wo
r
d
-
lev
el
em
b
ed
d
in
g
s
,
an
d
s
o
m
e
ex
is
tin
g
h
y
b
r
id
m
o
d
els
th
at
ty
p
ically
co
m
b
in
e
o
n
ly
wo
r
d
an
d
ch
ar
ac
te
r
-
lev
el
r
ep
r
esen
tatio
n
s
,
o
u
r
p
r
o
p
o
s
ed
C
NN
-
C
C
B
m
o
d
el
in
tr
o
d
u
ce
s
a
m
o
r
e
co
m
p
r
e
h
en
s
iv
e
f
ea
tu
r
e
in
teg
r
atio
n
s
tr
ateg
y
.
I
t
jo
in
tly
in
co
r
p
o
r
ates
h
an
d
c
r
af
ted
co
n
ten
t
-
b
ased
an
d
ch
ar
ac
ter
-
b
as
ed
f
ea
tu
r
es
with
in
a
u
n
if
ied
f
r
am
ewo
r
k
,
with
co
n
s
id
er
atio
n
o
f
wo
r
d
-
lev
el
a
n
d
ch
ar
ac
te
r
-
lev
el
p
atter
n
s
as
well.
Mo
r
eo
v
er
,
wh
ile
p
r
i
o
r
ap
p
r
o
ac
h
es
f
o
cu
s
p
r
im
ar
ily
o
n
lear
n
ed
r
ep
r
esen
tatio
n
s
,
th
e
in
clu
s
io
n
o
f
s
tr
u
ctu
r
al
f
ea
tu
r
es
(
e.
g
.
,
h
y
p
er
lin
k
p
r
esen
ce
an
d
m
ess
ag
e
len
g
th
)
en
h
a
n
ce
s
th
e
m
o
d
el’
s
a
b
ilit
y
to
ca
p
tu
r
e
p
a
tter
n
s
th
at
ar
e
d
if
f
icu
lt
f
o
r
o
th
er
m
o
d
els
to
lear
n
.
T
h
is
co
m
b
in
atio
n
p
r
o
v
id
es
a
m
o
r
e
r
o
b
u
s
t
an
d
in
ter
p
r
etab
le
ap
p
r
o
ac
h
to
s
p
am
d
etec
tio
n
,
p
ar
ticu
lar
ly
f
o
r
n
o
is
y
tex
t.
T
h
e
k
e
y
n
o
v
elty
o
f
o
u
r
p
r
o
p
o
s
ed
m
o
d
el
in
clu
d
es;
−
Mu
lti
-
lev
el
f
ea
tu
r
e
f
u
s
io
n
:
T
h
e
m
o
d
el
co
m
b
in
es
lear
n
e
d
r
ep
r
esen
tatio
n
s
th
r
o
u
g
h
C
NN
with
clea
r
h
an
d
cr
a
f
ted
f
ea
tu
r
es,
en
ab
lin
g
it to
ca
p
tu
r
e
b
o
th
im
p
licit a
n
d
ex
p
licit c
h
ar
ac
ter
is
tics
o
f
s
p
a
m
m
ess
ag
es.
−
Par
allel
C
NN
a
r
ch
itectu
r
e:
W
e
in
tr
o
d
u
ce
d
th
e
u
s
e
o
f
d
u
al
co
n
v
o
lu
tio
n
al
b
r
a
n
ch
es,
wh
ich
allo
w
s
s
im
u
ltan
eo
u
s
lear
n
in
g
o
f
s
em
a
n
tic
an
d
s
ty
lis
tic
p
atter
n
s
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J E
lec
&
C
o
m
p
E
n
g
I
SS
N:
2088
-
8
7
0
8
A
h
yb
r
id
co
n
ten
t
a
n
d
ch
a
r
a
ct
er fea
tu
r
e
meth
o
d
fo
r
S
MS
s
p
a
m
d
etec
tio
n
(
Gertr
u
d
e
S
ela
s
e
Go
s
u
)
2063
−
E
n
h
an
ce
d
r
o
b
u
s
tn
ess
f
o
r
n
o
is
y
tex
t
:
B
y
in
co
r
p
o
r
atin
g
ch
ar
ac
ter
-
lev
el
in
f
o
r
m
atio
n
,
o
u
r
m
o
d
el
ef
f
ec
tiv
el
y
h
an
d
les m
is
s
p
ellin
g
s
,
am
b
ig
u
o
u
s
wo
r
d
s
,
an
d
in
f
o
r
m
al
lan
g
u
ag
es c
o
m
m
o
n
l
y
f
o
u
n
d
in
SMS sp
am
.
T
h
is
s
tu
d
y
ad
d
r
ess
es
th
e
lim
itatio
n
s
o
f
e
x
is
tin
g
SMS
s
p
am
d
etec
tio
n
m
eth
o
d
s
th
at
r
ely
p
r
im
ar
ily
o
n
wo
r
d
-
lev
el
f
ea
tu
r
e
r
e
p
r
esen
tatio
n
o
r
o
n
ly
f
o
c
u
s
o
n
c
h
ar
ac
te
r
-
lev
el
p
atter
n
s
,
wh
ic
h
o
f
ten
f
ail
to
ca
p
tu
r
e
th
e
d
iv
er
s
e
an
d
n
o
is
y
p
atter
n
s
p
r
e
s
en
t
in
s
h
o
r
t
tex
t
m
ess
ag
es.
T
o
o
v
e
r
co
m
e
th
is
g
ap
,
we
p
r
o
p
o
s
e
a
h
y
b
r
id
d
ee
p
lear
n
in
g
m
o
d
el
C
NN
-
C
C
B
th
at
in
teg
r
ates
a
h
an
d
cr
af
ted
co
n
ten
t
-
b
ased
a
n
d
ch
ar
ac
ter
-
b
ased
s
tatis
tical
f
ea
tu
r
es
with
in
a
u
n
if
ied
f
r
am
ewo
r
k
.
T
h
e
r
esu
lts
d
em
o
n
s
tr
ate
th
a
t
o
u
r
p
r
o
p
o
s
ed
m
o
d
el
co
n
s
is
ten
tly
o
u
tp
er
f
o
r
m
s
b
aselin
e
m
o
d
els,
ac
h
iev
in
g
h
i
g
h
p
er
f
o
r
m
an
ce
ac
r
o
s
s
m
u
ltip
le
ev
alu
atio
n
m
etr
ics.
T
h
e
a
b
latio
n
s
tu
d
y
f
u
r
th
e
r
co
n
f
ir
m
s
th
at
ea
c
h
f
ea
tu
r
e
ty
p
e
co
n
tr
ib
u
tes
p
o
s
itiv
ely
to
th
e
o
v
er
all
p
er
f
o
r
m
a
n
ce
,
wh
ile
t
h
e
s
tatis
tical
an
aly
s
i
s
v
alid
ates
th
e
r
eliab
ilit
y
o
f
t
h
e
m
o
d
el’
s
im
p
r
o
v
em
e
n
t.
T
h
ese
f
in
d
in
g
s
im
p
ly
th
at
co
m
b
in
in
g
m
u
lti
-
lev
el
f
ea
tu
r
es
s
ig
n
if
ican
tly
en
h
a
n
ce
s
th
e
r
o
b
u
s
tn
ess
an
d
ef
f
ec
tiv
en
ess
o
f
s
p
am
d
etec
tio
n
s
y
s
tem
s
.
T
h
e
p
r
o
p
o
s
ed
ap
p
r
o
ac
h
is
s
u
itab
le
f
o
r
s
h
o
r
t
an
d
n
o
is
y
tex
t
d
o
m
ain
s
,
s
u
ch
as
SMS
an
d
o
th
er
s
o
cial
m
ed
i
a
ch
ats,
wh
er
e
b
o
th
s
em
an
tic
an
d
s
ty
lis
tic
p
atter
n
s
ar
e
d
o
m
in
a
n
t.
2.
M
E
T
H
O
D
T
h
e
p
r
o
p
o
s
ed
m
o
d
el
f
o
cu
s
es
o
n
f
ilter
in
g
tex
t
m
ess
ag
es
b
y
m
er
g
in
g
two
d
if
f
er
en
t
f
ea
tu
r
e
ex
tr
ac
tio
n
tech
n
iq
u
es.
T
h
e
f
r
a
m
ewo
r
k
in
teg
r
ates
b
o
th
s
em
an
tic
co
n
ten
t
f
ea
tu
r
es
(
s
y
n
tax
,
s
em
an
tics
,
len
g
th
o
f
tex
t,
an
d
p
r
esen
ce
o
f
h
y
p
er
lin
k
s
)
an
d
ch
ar
ac
ter
-
b
ased
f
ea
tu
r
es
(
n
u
m
b
er
o
f
ch
a
r
ac
ter
s
,
r
ep
etitio
n
s
,
ca
p
italizatio
n
p
atter
n
s
,
an
d
s
p
ec
ial
s
y
m
b
o
ls
)
.
Fig
u
r
e
1
also
s
h
o
ws th
e
im
p
l
em
en
tatio
n
o
f
th
e
p
r
o
p
o
s
ed
m
o
d
el.
Fig
u
r
e
1
.
Pro
p
o
s
ed
C
NN
-
C
C
B
m
o
d
el
2
.
1
.
E
x
t
r
a
ct
io
n
o
f
co
nte
nt
a
nd
cha
ra
ct
er
f
ea
t
ures
Fo
r
co
n
te
n
t
-
b
ased
f
ea
tu
r
es,
we
d
ef
in
ed
“
ex
tr
ac
t
_
co
n
te
n
t_
f
ea
tu
r
es(te
x
t)
”
.
T
h
is
e
x
tr
ac
t
s
s
y
n
tax
,
s
em
an
tic,
tex
t
len
g
th
,
an
d
p
r
e
s
en
ce
o
f
h
y
p
er
lin
k
s
u
s
in
g
Sp
aCy
in
n
atu
r
al
lan
g
u
ag
e
p
r
o
c
ess
in
g
(
NL
P).
T
h
is
f
u
n
ctio
n
en
ab
les
th
e
m
o
d
el
to
tr
an
s
f
o
r
m
u
n
s
tr
u
ct
u
r
ed
te
x
tu
al
d
ata
in
to
s
tr
u
ct
u
r
ed
n
u
m
e
r
ical
r
ep
r
esen
tatio
n
s
,
wh
ich
ca
n
th
en
b
e
ef
f
ec
tiv
el
y
u
s
ed
f
o
r
d
o
wn
s
tr
ea
m
task
s
s
u
ch
as
class
if
icatio
n
,
clu
s
ter
in
g
,
o
r
s
im
ilar
ity
an
aly
s
is
.
T
h
er
ef
o
r
e
,
th
e
co
n
te
n
t
-
b
ased
f
ea
tu
r
es f
o
r
ea
c
h
tex
t
(
)
ar
e:
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
0
8
8
-
8
7
0
8
I
n
t J E
lec
&
C
o
m
p
E
n
g
,
Vo
l.
1
6
,
No
.
4
,
Au
g
u
s
t
20
2
6
:
2
0
6
1
-
2073
2064
−
Sy
n
tax
co
u
n
t:
is
th
e
n
u
m
b
er
o
f
s
y
n
tactic
elem
en
ts
ex
t
r
ac
ted
u
s
in
g
Sp
aCy
.
T
h
er
ef
o
r
e,
it
is
t
h
e
to
tal
n
u
m
b
er
o
f
to
k
e
n
s
in
th
e
Sp
aCy
d
o
cu
m
en
t w
h
ich
ar
e
n
o
t u
n
iq
u
e
s
y
n
tax
ty
p
es.
(
)
=
|
|
=
∑
1
|
|
=
1
.
−
Sem
an
tic
co
u
n
t
:
is
th
e
n
u
m
b
er
o
f
m
ea
n
in
g
f
u
l
to
k
en
s
/en
titi
es
(
e.
g
.
,
n
o
u
n
s
,
v
er
b
s
,
n
a
m
ed
en
titi
es).
I
t
is
th
e
to
tal
n
u
m
b
er
o
f
to
k
e
n
s
in
th
e
d
o
cu
m
e
n
t,
wh
ich
ar
e
e
n
tity
lab
els
co
u
n
ted
p
e
r
to
k
en
,
r
eg
a
r
d
less
o
f
b
ein
g
em
p
ty
o
r
n
o
t.
(
)
=
∑
1
(
_
(
)
≠
0
|
|
=
1
.
−
T
ex
t le
n
g
th
:
is
th
e
to
tal
n
u
m
b
e
r
o
f
wo
r
d
s
/to
k
en
s
in
th
e
m
ess
ag
e,
in
clu
d
i
n
g
s
p
ac
es.
(
ℎ
)
=
|
.
(
)
|
.
−
Hy
p
er
lin
k
p
r
esen
ce
:
is
th
e
b
in
ar
y
f
ea
tu
r
e
in
d
icatin
g
p
r
esen
c
e
o
f
s
u
b
s
tr
in
g
“h
ttp
”
o
r
“ww
w”,
h
av
in
g
1
if
UR
L
ex
is
ts
,
el
s
e
0
.
(
)
=
{
1
,
“http”
or
“w
w
w”
,
0
,
Ot
he
r
wi
s
e
.
All th
ese
f
o
r
m
s
th
e
co
n
te
n
t f
e
atu
r
e
v
ec
to
r
as sh
o
wn
in
(
1
)
:
(
)
=
[
(
)
,
(
)
,
(
ℎ
)
,
(
)
]
.
(
1
)
Fo
r
c
h
ar
ac
ter
-
b
ased
f
ea
tu
r
es,
we
d
ef
in
ed
“
ex
tr
ac
t_
ch
a
r
ac
ter
_
f
ea
tu
r
es(te
x
t)
”
wh
ich
e
x
tr
ac
ts
th
e
n
u
m
b
er
o
f
ch
a
r
ac
ter
s
,
co
u
n
t
o
f
r
ep
ea
ted
ch
a
r
ac
ter
s
,
n
u
m
b
er
o
f
ca
p
italized
wo
r
d
s
,
an
d
f
r
e
q
u
en
cy
o
f
s
p
ec
if
ic
s
p
ec
ial
ch
ar
ac
ter
s
.
T
h
ese
ch
a
r
ac
ter
-
b
ased
m
etr
ics
h
elp
ca
p
tu
r
e
tech
n
ical
p
atter
n
s
in
th
e
tex
t,
wh
ich
ca
n
b
e
u
s
ef
u
l
f
o
r
d
is
tin
g
u
is
h
in
g
b
e
twee
n
d
if
f
er
e
n
t
wr
itin
g
b
eh
av
io
r
s
,
d
etec
tin
g
an
o
m
alies,
o
r
i
m
p
r
o
v
in
g
th
e
p
er
f
o
r
m
an
ce
o
f
tex
t c
lass
if
icatio
n
m
o
d
els.
Fo
r
ea
ch
tex
t
(
)
,
th
e
ch
ar
ac
ter
-
b
ased
f
ea
tu
r
es we
r
e
d
ef
in
ed
as:
−
Nu
m
b
er
o
f
ch
ar
ac
te
r
s
: th
is
is
t
h
e
to
tal
n
u
m
b
er
o
f
ch
ar
ac
te
r
s
in
th
e
tex
t.
(
ℎ
)
=
(
)
=
|
|
.
−
R
ep
ea
ted
ch
ar
ac
ter
s
:
is
th
e
c
o
u
n
t
o
f
co
n
s
ec
u
tiv
e
r
e
p
ea
ted
c
h
ar
ac
ter
s
(
e.
g
.
,
“
o
o
o
”,
“!
!
!
”)
.
C
h
ar
ac
ter
s
th
at
ap
p
ea
r
m
o
r
e
th
a
n
o
n
ce
in
th
e
t
ex
t,
th
at
is
all
o
cc
u
r
r
e
n
ce
s
.
(
)
=
|
{
∈
|
.
(
)
>
1
}
|
.
−
C
ap
italized
wo
r
d
s
:
is
th
e
n
u
m
b
er
o
f
wo
r
d
s
f
u
lly
i
n
u
p
p
er
ca
s
e
(
s
p
lit b
y
wh
ites
p
ac
e)
.
(
)
=
|
{
∈
.
(
)
|
=
(
)
}
|
.
−
Sp
ec
ial
ch
ar
ac
ter
f
r
e
q
u
en
c
y
:
is
th
e
co
u
n
t
o
f
s
y
m
b
o
ls
o
r
ch
ar
a
cter
s
b
elo
n
g
in
g
to
th
e
s
et
(
@
,
#
,
$
,
%,
&
,
*
)
.
(
)
=
∑
1
(
∈
{
@
,
#
,
$
,
%
,
&
,
∗
}
)
.
∈
All th
ese
f
o
r
m
s
th
e
ch
ar
ac
ter
f
ea
tu
r
e
v
ec
to
r
as sh
o
wn
in
(
2
)
:
ℎ
(
)
=
[
(
ℎ
)
,
(
)
,
(
)
,
(
)
]
.
(
2
)
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J E
lec
&
C
o
m
p
E
n
g
I
SS
N:
2088
-
8
7
0
8
A
h
yb
r
id
co
n
ten
t
a
n
d
ch
a
r
a
ct
er fea
tu
r
e
meth
o
d
fo
r
S
MS
s
p
a
m
d
etec
tio
n
(
Gertr
u
d
e
S
ela
s
e
Go
s
u
)
2065
2
.
2
.
Wo
rd
t
o
k
eniza
t
i
o
n a
nd
pa
dd
ing
Usi
n
g
to
k
en
izer
,
ea
ch
te
x
t
is
co
n
v
er
ted
to
a
s
eq
u
en
ce
o
f
in
teg
er
s
r
ep
r
esen
tin
g
wo
r
d
s
.
E
a
ch
wo
r
d
is
to
k
en
ized
as
s
h
o
wn
in
(
3
)
.
W
e
th
en
ap
p
lied
p
ad
d
in
g
to
ac
h
iev
e
a
f
ix
ed
len
g
th
m
a
x
_
len
s
eq
u
en
ce
f
o
r
n
eu
r
a
l
n
etwo
r
k
in
p
u
t
in
(
4
)
.
T
h
is
en
s
u
r
es
th
at
all
in
p
u
t
s
eq
u
e
n
ce
s
h
av
e
u
n
if
o
r
m
d
im
e
n
s
io
n
s
,
al
lo
win
g
th
e
n
eu
r
al
n
etwo
r
k
to
p
r
o
ce
s
s
b
atch
es e
f
f
icien
tly
with
o
u
t b
ein
g
af
f
ec
ted
b
y
v
a
r
y
in
g
s
en
ten
ce
len
g
t
h
s
.
T
h
e
to
k
en
ize
w
o
r
d
s
eq
u
en
ce
i
s
:
_
(
)
=
[
1
,
2
,
…
,
]
.
(
3
)
T
h
e
p
ad
d
ed
wo
r
d
s
eq
u
e
n
ce
is
:
(
)
=
[
1
,
2
,
…
,
,
0
,
0
,
…
,
0
]
,
(
4
)
wh
er
e
is
th
e
in
d
ex
o
f
ea
c
h
wo
r
d
in
tex
t.
2
.
3
.
Cha
ra
c
t
er
t
o
k
eniza
t
io
n
a
nd
pa
dd
ing
Fo
r
ch
ar
ac
te
r
to
k
en
izatio
n
,
ea
ch
ch
a
r
ac
ter
is
m
a
p
p
ed
to
its
u
n
iq
u
e
o
r
d
in
al
v
alu
e
u
n
d
e
r
th
e
Un
ico
d
e
en
co
d
in
g
s
ch
em
e
in
(
5
)
.
T
h
is
is
p
ad
d
ed
to
ac
h
iev
e
a
f
ix
ed
m
ax
_
ch
ar
len
g
th
i
n
(
6
)
.
T
h
is
en
s
u
r
es
th
at
all
tex
t
in
p
u
ts
ar
e
r
ep
r
esen
ted
in
a
co
n
s
is
ten
t
n
u
m
er
ical
f
o
r
m
at,
allo
win
g
th
em
to
b
e
ef
f
icien
tly
p
r
o
ce
s
s
ed
b
y
m
ac
h
in
e
lear
n
in
g
m
o
d
els r
eg
ar
d
less
o
f
th
eir
o
r
ig
in
al
le
n
g
th
.
T
h
e
to
k
en
ize
c
h
ar
ac
ter
s
eq
u
en
ce
is
:
ℎ
_
(
)
=
[
(
1
)
,
(
2
)
,
…
,
(
)
]
.
(
5
)
T
h
e
p
ad
d
ed
ch
a
r
ac
ter
s
eq
u
e
n
c
e
is
:
(
)
=
[
(
1
)
,
(
2
)
,
…
,
(
)
,
0
,
0
,
…
,
0
]
,
(
6
)
wh
er
e
(
)
is
th
e
Un
ico
d
e
v
alu
e
o
f
ea
ch
ch
ar
ac
ter
.
2
.
4
.
CNN
f
e
a
t
ure
lea
rning
W
e
b
u
ilt
a
Ker
as
f
u
n
ctio
n
al
ap
p
licatio
n
p
r
o
g
r
am
in
g
in
te
r
f
ac
e
(
API
)
m
o
d
el
with
two
p
ar
allel
b
r
an
ch
es,
with
ea
ch
b
r
an
ch
c
o
n
s
is
tin
g
o
f
:
E
m
b
ed
d
i
n
g
→
C
o
n
v
1
D
→
Glo
b
al
Ma
x
Po
o
lin
g
.
T
h
is
ar
ch
itectu
r
e
allo
ws
th
e
m
o
d
el
t
o
lear
n
co
m
p
lem
en
tar
y
f
ea
tu
r
es
f
r
o
m
th
e
s
am
e
in
p
u
t
b
y
p
r
o
ce
s
s
in
g
it
th
r
o
u
g
h
two
p
ar
allel
co
n
v
o
l
u
tio
n
al
p
at
h
way
s
.
T
h
e
u
s
e
o
f
em
b
ed
d
in
g
lay
er
s
co
n
v
er
ts
d
is
cr
ete
to
k
e
n
s
in
to
d
en
s
e
v
ec
to
r
s
,
wh
ile
C
o
n
v
1
D
an
d
g
lo
b
al
m
ax
p
o
o
li
n
g
ef
f
icien
tly
ca
p
tu
r
e
an
d
r
ef
i
n
e
th
e
m
o
s
t salien
t lo
ca
l p
atter
n
s
in
ea
ch
b
r
an
ch
.
−
C
NN
wo
r
d
-
lev
el
B
r
an
ch
with
(
)
as in
p
u
t f
r
o
m
th
e
wo
r
d
s
eq
u
e
n
ce
.
E
m
b
ed
d
in
g
:
(
)
→
(
)
∈
ℝ
_
×
_
(
7
)
C
o
n
v
o
lu
tio
n
a
n
d
ac
tiv
atio
n
:
(
)
=
(
1
(
(
)
)
)
.
(
8
)
Glo
b
al
Ma
x
Po
o
lin
g
:
(
)
=
ma
x
(
(
)
)
∈
ℝ
_
.
(
9
)
−
C
NN
ch
ar
ac
ter
-
lev
el
B
r
an
ch
with
(
)
as in
p
u
t f
o
r
ch
ar
ac
ter
s
eq
u
en
ce
.
E
m
b
ed
d
in
g
:
(
)
→
(
)
∈
ℝ
_
ℎ
×
_
.
(
1
0
)
C
o
n
v
o
lu
tio
n
a
n
d
ac
tiv
atio
n
:
(
)
=
(
1
(
(
)
)
)
.
(
1
1
)
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
0
8
8
-
8
7
0
8
I
n
t J E
lec
&
C
o
m
p
E
n
g
,
Vo
l.
1
6
,
No
.
4
,
Au
g
u
s
t
20
2
6
:
2
0
6
1
-
2073
2066
Glo
b
al
Ma
x
Po
o
lin
g
:
(
)
=
ma
x
(
(
)
)
∈
ℝ
_
,
(
1
2
)
wh
er
e
ℝ
d
en
o
tes s
et
o
f
r
ea
l n
u
m
b
er
s
,
an
d
R
eL
U
is
th
e
ac
tiv
atio
n
f
u
n
ctio
n
u
s
ed
.
2
.
5
.
F
e
a
t
ure
f
us
io
n a
nd
cla
s
s
if
ica
t
io
n
T
h
e
two
f
ea
tu
r
e
v
ec
to
r
s
co
m
p
u
ted
ea
r
lier
,
(
)
an
d
ℎ
(
)
an
d
th
e
two
p
ar
allel
b
r
an
ch
v
ec
to
r
s
,
(
)
an
d
(
)
,
ar
e
co
m
b
in
ed
in
t
o
a
s
in
g
le
v
ec
to
r
as
s
h
o
wn
in
(
1
3
)
.
T
h
is
is
f
o
llo
wed
b
y
th
e
d
en
s
e
lay
er
,
th
en
th
e
o
u
t
p
u
t
lay
e
r
,
as
s
h
o
wn
in
(
1
4
)
an
d
(
1
5
)
r
esp
ec
ti
v
ely
,
wh
e
r
e
we
a
p
p
lied
weig
h
t
m
atr
ices.
T
h
ese
o
p
er
atio
n
s
f
u
s
e
th
e
co
m
p
lem
en
tar
y
f
ea
tu
r
es
f
r
o
m
b
o
th
b
r
a
n
ch
es
in
to
a
u
n
if
ied
r
e
p
r
esen
t
atio
n
,
en
ab
lin
g
th
e
n
etwo
r
k
to
lea
r
n
h
i
g
h
er
-
le
v
el
p
atter
n
s
b
ef
o
r
e
p
r
o
d
u
cin
g
th
e
f
in
al
p
r
ed
ictio
n
.
C
o
n
ca
ten
atio
n
:
(
)
=
[
(
)
,
(
)
,
(
)
,
ℎ
(
)
]
.
(
1
3
)
Den
s
e
lay
er
:
(
)
→
(
1
(
)
+
1
)
.
(
1
4
)
Ou
tp
u
t la
y
er
:
̂
(
)
=
(
2
(
)
+
2
)
,
(
1
5
)
wh
er
e
1
is
th
e
weig
h
t
m
at
r
ix
f
o
r
t
h
e
d
en
s
e
lay
e
r
,
m
ap
p
in
g
th
e
f
u
s
ed
f
ea
tu
r
e
(
)
to
th
e
h
i
d
d
en
d
en
s
e
r
ep
r
esen
tatio
n
(
)
.
2
is
th
e
weig
h
t
m
atr
ic
f
o
r
th
e
o
u
tp
u
t
lay
e
r
,
m
ap
p
in
g
th
e
f
u
s
ed
f
ea
tu
r
e
(
)
to
th
e
f
in
al
p
r
ed
ictio
n
̂
(
)
.
(
)
is
th
e
s
ig
m
o
id
f
u
n
ctio
n
s
in
ce
th
is
is
a
b
in
ar
y
class
if
icat
io
n
p
r
o
b
lem
.
1
is
t
h
e
b
ias
v
ec
to
r
f
o
r
d
e
n
s
e
lay
er
,
an
d
2
is
th
e
b
ias v
ec
to
r
f
o
r
th
e
o
u
tp
u
t l
ay
er
.
3.
E
XP
E
R
I
M
E
N
T
A
L
SE
T
T
I
N
G
S
T
h
e
d
esig
n
ch
o
ice
o
f
o
u
r
m
o
d
el
ar
ch
itectu
r
e
was
b
ec
au
s
e
cr
ea
tin
g
p
ar
allel
b
r
an
ch
es
ca
p
tu
r
es
b
o
th
s
em
an
tic
an
d
s
ty
lis
tic
p
atter
n
s
o
f
tex
t.
T
h
e
ex
p
er
im
e
n
t
was
i
m
p
lem
en
ted
in
Py
th
o
n
u
s
in
g
T
en
s
o
r
Flo
w/Ker
as.
Sp
aCy
(
en
_
co
r
e
_
web
_
s
m
)
wa
s
u
s
ed
f
o
r
lin
g
u
is
tic
f
ea
tu
r
e
e
x
tr
ac
tio
n
.
W
e
u
s
ed
a
k
e
r
n
el
s
ize
o
f
5
d
u
e
to
its
ef
f
ec
tiv
en
ess
to
c
ap
tu
r
e
s
h
o
r
t
tex
t
p
atter
n
s
,
an
d
th
e
g
lo
b
al
m
ax
p
o
o
lin
g
r
etain
s
s
tr
o
n
g
est
f
e
atu
r
es
an
d
r
e
d
u
ce
s
o
v
er
f
itti
n
g
.
Als
o
,
to
r
ed
u
ce
o
v
er
f
itti
n
g
,
we
in
clu
d
ed
s
tr
o
n
g
r
eg
u
lar
izatio
n
with
a
d
r
o
p
o
u
t
o
f
0
.
5
.
Fig
u
r
e
2
illu
s
tr
ates th
e
co
m
p
lete
ex
p
er
i
m
en
tal
wo
r
k
f
l
o
w
f
o
r
r
ep
r
o
d
u
c
ib
ilit
y
.
−
Data
s
p
litt
in
g
:
T
h
e
d
ataset
wa
s
d
iv
id
ed
in
to
tr
ain
in
g
an
d
tes
tin
g
s
ets
u
s
in
g
a
r
atio
o
f
8
0
:2
0
,
with
a
f
ix
ed
r
an
d
o
m
s
tate
o
f
4
2
,
to
en
s
u
r
e
r
ep
r
o
d
u
cib
ilit
y
.
−
Pre
p
r
o
ce
s
s
in
g
s
tep
s
:
T
h
e
tex
t
d
ata
wer
e
p
r
o
ce
s
s
ed
u
s
in
g
th
e
Sp
aCy
lib
r
ar
y
(
e
n
_
co
r
e_
web
_
s
m
)
f
o
r
lin
g
u
is
tic
f
ea
tu
r
e
e
x
tr
ac
tio
n
.
C
o
n
ten
t
-
b
ased
f
ea
tu
r
es
wer
e
co
m
p
u
ted
as
well
as
C
h
ar
ac
ter
-
b
ased
f
ea
t
u
r
es.
T
h
e
wo
r
d
s
eq
u
en
ce
s
wer
e
to
k
en
ized
u
s
in
g
Ker
as
T
o
k
en
izer
with
a
v
o
ca
b
u
lar
y
s
ize
o
f
1
0
,
0
0
0
an
d
p
ad
d
e
d
to
a
f
ix
ed
len
g
th
o
f
2
0
0
.
W
h
ile
th
e
ch
ar
ac
ter
s
eq
u
en
ce
s
wer
e
en
co
d
e
d
u
s
in
g
Un
ico
d
e
(
o
r
d
)
an
d
p
ad
d
e
d
to
a
m
ax
im
u
m
le
n
g
th
o
f
2
0
0
.
−
Hy
p
er
p
ar
a
m
eter
s
:
Fo
r
t
h
e
h
y
p
er
p
ar
am
eter
s
,
we
u
s
ed
an
e
m
b
ed
d
in
g
d
im
en
s
io
n
o
f
1
2
8
,
an
d
n
u
m
b
er
o
f
co
n
v
o
l
u
tio
n
f
ilter
s
u
s
ed
was
6
4
.
W
e
u
s
ed
a
d
r
o
p
o
u
t
r
ate
o
f
0
.
5
a
n
d
a
b
atch
s
ize
o
f
6
4
.
T
h
e
n
u
m
b
e
r
o
f
ep
o
ch
s
was
1
0
0
,
an
d
R
eL
U
was
u
s
ed
as
th
e
ac
tiv
atio
n
f
u
n
ctio
n
f
o
r
th
e
h
id
d
en
lay
er
w
h
ile
Sig
m
o
id
was
u
s
ed
f
o
r
t
h
e
o
u
t
p
u
t la
y
er
.
−
Op
tim
izatio
n
:
Fo
r
o
p
tim
izatio
n
,
we
u
s
ed
Ad
am
with
a
lear
n
in
g
r
ate
o
f
0
.
0
0
1
an
d
a
B
in
a
r
y
cr
o
s
s
-
en
tr
o
p
y
as
o
u
r
lo
s
s
f
u
n
ctio
n
.
T
o
ad
d
r
e
s
s
th
e
i
s
s
u
e
o
f
cla
s
s
im
b
alan
ce
,
class
weig
h
ts
wer
e
u
s
ed
f
o
r
b
o
th
d
en
s
e
an
d
o
u
tp
u
t la
y
e
r
.
3
.
1
.
E
x
perim
ent
a
l
f
ra
m
ewo
rk
Fig
u
r
e
2
s
h
o
ws
th
e
ex
p
e
r
im
en
tal
f
r
am
ewo
r
k
o
f
th
e
s
tu
d
y
.
T
h
e
f
lo
w
o
f
th
is
ex
p
er
im
en
t
i
n
c
lu
d
es
d
ata
p
r
e
-
p
r
o
ce
s
s
in
g
at
th
e
in
itial
s
t
ag
e,
as
well
as
d
ata
p
r
o
ce
s
s
in
g
,
wh
ich
in
clu
d
es
to
k
e
n
izatio
n
an
d
p
ad
d
in
g
.
T
h
e
s
eq
u
en
ce
s
wer
e
th
e
n
m
o
v
ed
in
d
ep
en
d
en
tly
to
th
e
two
p
a
r
allel
b
r
an
ch
es.
T
h
ese
ar
e
f
u
s
ed
in
with
f
ea
tu
r
es
ex
tr
ac
ted
,
th
en
th
e
m
o
d
el
tr
ain
in
g
tak
es
p
lace
,
wh
ich
is
f
o
llo
wed
b
y
t
h
e
ev
alu
atio
n
.
Fin
ally
,
tex
ts
ar
e
p
r
ed
icted
u
s
in
g
th
e
m
o
d
el
to
d
is
tin
g
u
is
h
if
th
ey
ar
e
s
p
am
o
r
n
o
t.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J E
lec
&
C
o
m
p
E
n
g
I
SS
N:
2088
-
8
7
0
8
A
h
yb
r
id
co
n
ten
t
a
n
d
ch
a
r
a
ct
er fea
tu
r
e
meth
o
d
fo
r
S
MS
s
p
a
m
d
etec
tio
n
(
Gertr
u
d
e
S
ela
s
e
Go
s
u
)
2067
Fig
u
r
e
2
.
E
x
p
er
im
e
n
tal
f
r
am
e
wo
r
k
3
.
2
.
Da
t
a
s
et
s
d
escript
io
n
T
h
e
o
r
i
g
in
al
d
ataset
u
s
ed
i
n
th
is
s
tu
d
y
was
co
n
s
tr
u
cte
d
b
y
ag
g
r
eg
atin
g
m
u
ltip
le
SM
S
d
ataset
s
o
b
tain
ed
f
r
o
m
Kag
g
le,
wh
ich
i
s
a
m
ac
h
in
e
lear
n
in
g
r
ep
o
s
ito
r
y
.
All d
atasets
wer
e
m
er
g
ed
in
to
a
u
n
if
ied
c
o
r
p
u
s
m
ak
in
g
it
th
e
lar
g
est
SMS
d
ataset
as
co
m
p
a
r
ed
to
t
h
e
o
n
es
u
s
ed
in
p
r
ev
i
o
u
s
s
tu
d
i
es.
T
h
e
d
ata
was
p
r
ep
r
o
ce
s
s
ed
b
y
r
em
o
v
in
g
m
is
s
in
g
o
r
n
u
ll
en
tr
ies,
an
d
it
was
lab
elled
as
1
=
s
p
am
an
d
0
=n
o
n
-
s
p
am
.
T
h
e
co
n
s
tr
u
cted
d
ataset
co
n
tain
s
1
3
,
9
4
6
SMS m
ess
ag
es,
with
3
,
4
3
5
(
2
5
%)
s
p
am
m
ess
ag
es a
n
d
1
0
,
5
1
1
(
7
5
%)
n
o
n
-
s
p
am
m
ess
ag
es.
Ho
wev
er
,
o
u
r
p
r
o
p
o
s
ed
m
o
d
el
was
ev
alu
ated
o
n
two
o
th
er
d
a
tasets
to
test
its
g
en
er
aliza
b
ilit
y
:
“
Star
ter
d
ataset
”
an
d
“
Sp
am
m
ails
”
.
T
h
e
Star
ter
d
ataset
h
as
5
,
5
7
2
S
MS
m
ess
ag
es,
with
4
,
8
5
2
n
o
n
-
s
p
am
m
ess
ag
es
an
d
7
4
7
s
p
am
m
ess
ag
es.
W
h
ile
th
e
Sp
am
m
ails
co
n
tain
5
,
1
7
1
em
ail
m
ess
ag
es,
with
3
,
6
7
2
n
o
n
-
s
p
am
an
d
1
,
4
9
9
s
p
am
m
ails
.
T
h
ese
also
h
av
e
th
e
s
am
e
lab
ellin
g
as
th
e
o
r
ig
in
al
d
ata
co
n
s
tr
u
cted
.
4.
RE
SU
L
T
S AN
D
D
I
SCU
SS
I
O
NS
T
h
e
p
r
o
p
o
s
ed
C
NN
-
C
C
B
m
o
d
el
o
u
tp
e
r
f
o
r
m
s
o
th
er
m
o
d
els
ac
r
o
s
s
all
ev
alu
atio
n
m
etr
ics,
ac
h
iev
in
g
h
ig
h
er
ac
cu
r
ac
y
,
p
r
ec
is
io
n
,
r
e
ca
ll,
an
d
F1
-
s
co
r
e.
T
h
is
d
em
o
n
s
tr
ates
th
e
ef
f
ec
tiv
en
ess
o
f
i
n
teg
r
atin
g
co
n
ten
t
-
b
ased
an
d
ch
ar
ac
ter
-
b
ased
f
e
atu
r
es
f
o
r
s
p
am
d
etec
tio
n
.
C
o
m
p
ar
ed
with
o
th
e
r
m
ac
h
in
e
l
ea
r
n
in
g
tech
n
iq
u
es,
T
ab
le
1
s
h
o
ws
th
at
th
e
p
r
o
p
o
s
ed
C
NN
-
C
C
B
(
0
.
9
9
7
)
o
u
tp
e
r
f
o
r
m
s
k
-
n
ea
r
est
n
eig
h
b
o
r
(
KNN)
(
0
.
8
9
4
)
,
lo
g
is
tic
r
eg
r
ess
io
n
(
L
R
)
(
0
.
8
7
2
)
,
s
u
p
p
o
r
t
v
ec
to
r
m
ac
h
in
e
(
SVM)
(
0
.
8
5
8
)
,
d
ec
is
io
n
tr
ee
(
DT
)
(
0
.
8
4
2
)
,
a
n
d
n
aïv
e
B
ay
es
(
NB
)
(
0
.
7
5
3
)
in
ac
c
u
r
ac
y
.
T
h
e
p
r
o
p
o
s
ed
C
NN
-
C
C
B
m
o
d
el
also
o
u
tp
er
f
o
r
m
ed
t
h
ese
m
ac
h
i
n
e
lear
n
in
g
m
o
d
els
in
p
r
ec
is
io
n
,
F
1
-
s
co
r
e,
an
d
r
ec
all,
wh
ich
en
s
u
r
es th
e
co
r
r
ec
tn
ess
o
f
th
e
m
o
d
el’
s
p
r
ed
ictio
n
.
T
ab
le
1
.
Me
tr
ics
co
m
p
ar
is
o
n
o
f
p
r
o
p
o
s
ed
m
o
d
el
with
m
ac
h
i
n
e
lear
n
in
g
alg
o
r
ith
m
s
A
l
g
o
r
i
t
h
m
A
c
c
u
r
a
c
y
P
r
e
c
i
s
i
o
n
F1
-
S
c
o
r
e
R
e
c
a
l
l
Lo
g
i
s
t
i
c
s
r
e
g
r
e
ss
i
o
n
0
.
8
7
2
0
.
7
7
6
0
.
7
2
1
0
.
6
7
3
N
a
ï
v
e
B
a
y
e
s
c
l
a
ssi
f
i
e
r
0
.
7
5
3
0
.
4
9
6
0
.
5
4
7
0
.
6
0
9
D
e
c
i
s
i
o
n
t
r
e
e
c
l
a
ss
i
f
i
e
r
0
.
8
4
2
0
.
7
6
0
0
.
5
4
6
0
.
4
2
6
K
-
n
e
a
r
e
s
t
n
e
i
g
h
b
o
r
(
K
N
N
)
0
.
8
9
4
0
.
9
1
1
0
.
7
4
5
0
.
6
3
0
S
u
p
p
o
r
t
v
e
c
t
o
r
ma
c
h
i
n
e
(
S
V
M
)
0
.
8
5
8
0
.
7
8
8
0
.
6
6
5
0
.
5
7
5
P
r
o
p
o
se
d
m
o
d
e
l
(
C
N
N
-
C
C
B
)
0
.
9
9
7
0
.
9
8
8
0
.
9
9
3
0
.
9
9
8
Un
lik
e
th
e
b
aselin
e
m
o
d
els
th
at
r
ely
p
r
im
ar
ily
o
n
wo
r
d
-
le
v
el
r
ep
r
esen
tatio
n
s
,
o
u
r
p
r
o
p
o
s
ed
C
NN
-
C
C
B
m
o
d
el
in
teg
r
ated
b
o
th
co
n
ten
t
an
d
ch
ar
ac
te
r
f
ea
tu
r
e
s
,
th
er
eb
y
m
a
k
in
g
th
e
p
r
o
p
o
s
ed
m
o
d
el
c
o
n
s
id
e
r
p
atter
n
s
th
at
ar
e
co
m
m
o
n
in
s
p
am
m
ess
ag
es
b
u
t
wer
e
o
f
ten
o
v
er
lo
o
k
ed
b
y
wo
r
d
-
lev
el
m
o
d
els.
T
h
is
ex
p
lain
s
th
e
h
ig
h
p
er
f
o
r
m
a
n
ce
o
f
th
e
p
r
o
p
o
s
ed
C
NN
-
C
C
B
m
o
d
el
ag
ain
s
t
L
STM
an
d
GR
U.
Fro
m
T
ab
le
2
,
with
o
u
t
th
e
in
clu
s
io
n
o
f
d
r
o
p
o
u
t
r
ate
to
th
e
b
aselin
e
m
o
d
els,
o
u
r
p
r
o
p
o
s
ed
m
o
d
el
p
er
f
o
r
m
e
d
co
m
p
etitiv
ely
with
th
e
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
0
8
8
-
8
7
0
8
I
n
t J E
lec
&
C
o
m
p
E
n
g
,
Vo
l.
1
6
,
No
.
4
,
Au
g
u
s
t
20
2
6
:
2
0
6
1
-
2073
2068
b
aselin
e
m
o
d
els.
H
o
wev
er
,
w
ith
eq
u
al
d
r
o
p
o
u
t
r
ates
(
0
.
5
)
,
o
u
r
p
r
o
p
o
s
ed
m
o
d
el
o
u
tp
er
f
o
r
m
ed
th
e
b
aselin
e
m
o
d
els
ac
r
o
s
s
m
etr
ics.
Als
o
,
with
o
u
t
d
r
o
p
o
u
t
r
ates
f
o
r
all
th
r
ee
m
o
d
els,
o
u
r
p
r
o
p
o
s
ed
m
o
d
el
s
ig
n
if
ican
tly
o
u
tp
er
f
o
r
m
ed
th
e
b
aselin
e
m
o
d
els.
T
o
s
u
p
p
o
r
t
o
u
r
ar
g
u
m
en
t
th
at
o
u
r
p
r
o
p
o
s
ed
m
o
d
el
is
g
en
er
ally
m
o
r
e
ef
f
ec
tiv
e
f
o
r
s
p
am
d
etec
tio
n
,
we
co
n
d
u
cted
a
s
tatis
tical
s
i
g
n
if
ican
ce
test
as
s
h
o
wn
in
T
ab
le
3
.
T
h
is
is
to
em
p
ir
ically
af
f
ir
m
th
e
p
e
r
f
o
r
m
an
ce
o
f
o
u
r
p
r
o
p
o
s
ed
m
o
d
el
a
g
ain
s
t th
e
b
aselin
e
m
o
d
els.
T
ab
le
2
.
C
o
m
p
a
r
is
o
n
o
f
p
r
o
p
o
s
ed
m
o
d
el
with
b
aselin
e
m
o
d
e
ls
A
l
g
o
r
i
t
h
m
A
c
c
u
r
a
c
y
P
r
e
c
i
s
i
o
n
F1
-
S
c
o
r
e
R
e
c
a
l
l
Lo
ss
W
i
t
h
o
u
t
d
r
o
p
o
u
t
f
o
r
a
l
l
mo
d
e
l
s
LSTM
0
.
9
9
6
0
.
9
9
5
0
.
9
9
3
0
.
9
9
1
0
.
0
2
5
G
R
U
0
.
9
9
5
0
.
9
9
5
0
.
9
9
2
0
.
9
8
9
0
.
0
2
6
P
r
o
p
o
se
d
m
o
d
e
l
(
C
N
N
-
C
C
B
)
0
.
9
9
7
0
.
9
9
6
0
.
9
9
6
0
.
9
9
6
0
.
0
1
4
W
i
t
h
o
u
t
d
r
o
p
o
u
t
f
o
r
b
a
s
e
l
i
n
e
m
o
d
e
l
s
LSTM
0
.
9
9
6
0
.
9
9
5
0
.
9
9
3
0
.
9
9
1
0
.
0
2
5
G
R
U
0
.
9
9
5
0
.
9
9
5
0
.
9
9
2
0
.
9
8
9
0
.
0
2
6
P
r
o
p
o
se
d
m
o
d
e
l
(
C
N
N
-
C
C
B
)
0
.
9
9
7
0
.
9
8
8
0
.
9
9
3
0
.
9
9
8
0
.
0
4
8
W
i
t
h
d
r
o
p
o
u
t
r
a
t
e
f
o
r
a
l
l
mo
d
e
l
s
LSTM
0
.
9
8
8
0
.
9
9
2
0
.
9
7
5
0
.
9
5
9
0
.
1
3
0
G
R
U
0
.
9
8
5
0
.
9
9
2
0
.
9
6
9
0
.
9
4
7
0
.
0
6
2
P
r
o
p
o
se
d
m
o
d
e
l
(
C
N
N
-
C
C
B
)
0
.
9
9
7
0
.
9
8
8
0
.
9
9
3
0
.
9
9
8
0
.
0
4
8
T
ab
le
3
.
Statis
tical
s
ig
n
if
ican
ce
o
f
m
o
d
el
p
er
f
o
r
m
a
n
ce
M
o
d
e
l
C
o
m
p
a
r
i
so
n
T
-
v
a
l
u
e
P
-
v
a
l
u
e
P
a
i
r
e
d
t
-
t
e
s
t
C
N
N
-
C
C
B
v
s L
S
T
M
2
4
.
5
2
8
9
0
.
0
0
0
C
N
N
-
C
C
B
v
s GRU
3
9
.
5
2
8
5
0
.
0
0
0
SMS
d
ata
ar
e
ty
p
ically
s
h
o
r
t,
in
f
o
r
m
al,
an
d
n
o
is
y
,
th
e
r
ef
o
r
e
wo
r
d
-
lev
el
m
o
d
els
s
u
ch
as
L
STM
an
d
GR
U
m
ay
s
tr
u
g
g
le
with
lim
it
ed
co
n
tex
tu
al
in
f
o
r
m
atio
n
.
I
n
co
n
tr
ast,
th
e
ch
ar
ac
ter
-
lev
el
b
r
an
ch
in
o
u
r
m
o
d
el
en
ab
les
it
to
d
etec
t
s
u
b
-
wo
r
d
p
atter
n
s
an
d
ir
r
eg
u
lar
tex
t
s
tr
u
ctu
r
es,
m
ak
in
g
it
m
o
r
e
r
o
b
u
s
t
to
ab
b
r
ev
iatio
n
s
,
m
is
s
p
ellin
g
s
,
an
d
co
m
p
licated
wr
itin
g
s
ty
les u
s
ed
in
s
p
am
.
T
h
e
s
tatis
tical
te
s
ts
in
T
ab
le
3
s
h
o
w
th
at
th
e
p
er
f
o
r
m
an
ce
o
f
o
u
r
p
r
o
p
o
s
ed
C
NN
-
C
C
B
m
o
d
el
o
v
er
th
e
b
aselin
e
m
o
d
els
is
s
tatis
tically
s
ig
n
if
ican
t
(
p
-
v
alu
e
<
0
.
0
5
)
,
b
ec
au
s
e
th
e
co
m
p
ar
is
o
n
o
f
o
u
r
p
r
o
p
o
s
ed
m
o
d
el
t
o
b
ase
m
o
d
els
p
r
o
d
u
ce
d
v
er
y
lo
w
p
-
v
alu
es
a
n
d
h
ig
h
T
-
v
alu
es.
T
h
is
also
c
o
n
f
ir
m
s
th
at
th
e
o
b
s
er
v
ed
p
er
f
o
r
m
an
ce
ac
r
o
s
s
th
e
ev
alu
atio
n
m
etr
ices
is
n
o
t
d
u
e
t
o
r
an
d
o
m
v
ar
iatio
n
b
u
t
ar
e
attr
ib
u
tab
le
to
th
e
in
teg
r
atio
n
o
f
th
e
c
o
n
ten
t a
n
d
ch
ar
ac
ter
f
ea
tu
r
es.
T
ab
le
4
an
d
Fig
u
r
e
3
s
h
o
w
th
e
ab
ilit
y
o
f
th
e
p
r
o
p
o
s
ed
s
p
a
m
d
etec
tio
n
m
o
d
el
(
C
NN
-
C
C
B
)
f
o
r
SMS,
to
p
r
ed
ict
m
ess
ag
es
as
s
p
am
o
r
n
o
n
-
s
p
am
.
Usi
n
g
a
test
d
ata
s
am
p
le
o
f
2
,
7
9
0
,
with
2
,
1
0
7
n
o
n
-
s
p
am
m
ess
ag
es
an
d
6
8
3
an
d
s
p
am
m
ess
ag
es,
th
e
co
n
f
u
s
io
n
m
atr
ix
p
r
o
v
e
s
th
at
th
e
p
r
o
p
o
s
ed
m
o
d
el
h
as
h
ig
h
p
r
e
d
ictiv
e
ca
p
ac
ity
an
d
a
v
er
y
lo
w
ch
an
ce
o
f
m
is
class
if
icatio
n
.
T
h
is
i
n
d
icate
s
th
e
r
o
b
u
s
tn
ess
o
f
o
u
r
m
o
d
el
to
h
an
d
le
r
ea
l
-
life
s
ce
n
ar
io
s
.
T
ab
le
4
.
C
o
n
f
u
s
io
n
m
atr
ix
f
o
r
C
NN
-
C
C
B
o
n
test
d
ata
P
r
e
d
i
c
t
e
d
t
r
u
e
P
r
e
d
i
c
t
e
d
f
a
l
s
e
A
c
t
u
a
l
t
r
u
e
(
6
8
3
)
6
8
2
(
TP)
1
(
F
N
)
A
c
t
u
a
l
f
a
l
s
e
(
2
,
1
0
7
)
8
(
F
P
)
2
,
0
9
9
(
TN
)
Fig
u
r
e
3
.
C
o
n
f
u
s
io
n
m
atr
i
x
o
f
p
r
o
p
o
s
ed
m
o
d
el
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J E
lec
&
C
o
m
p
E
n
g
I
SS
N:
2088
-
8
7
0
8
A
h
yb
r
id
co
n
ten
t
a
n
d
ch
a
r
a
ct
er fea
tu
r
e
meth
o
d
fo
r
S
MS
s
p
a
m
d
etec
tio
n
(
Gertr
u
d
e
S
ela
s
e
Go
s
u
)
2069
T
h
e
v
alu
es o
b
tain
ed
f
r
o
m
th
e
co
n
f
u
s
io
n
m
atr
ix
m
ea
n
s
th
at;
−
T
h
e
m
o
d
el
p
r
ed
icted
6
8
2
m
ess
ag
es
to
b
e
s
p
am
wh
ich
ar
e
in
d
ee
d
s
p
am
,
s
ig
n
if
y
in
g
a
tr
u
e
c
lass
if
icatio
n
o
f
th
e
m
o
d
el.
−
T
h
e
m
o
d
el
p
r
ed
icte
d
8
m
ess
ag
es
to
b
e
s
p
am
wh
ich
ar
e
r
ath
er
n
o
n
-
s
p
am
m
ess
ag
e,
im
p
ly
in
g
v
er
y
lo
w
m
is
class
if
icatio
n
.
−
T
h
e
m
o
d
el
p
r
e
d
icted
1
m
e
s
s
ag
e
to
b
e
n
o
n
-
s
p
am
w
h
ich
is
r
ath
er
s
p
am
,
in
d
icati
n
g
a
v
er
y
l
o
w
m
is
class
if
icatio
n
as we
ll.
−
T
h
e
m
o
d
el
c
o
r
r
ec
tly
p
r
ed
icted
2
,
0
9
9
m
ess
ag
es a
s
n
o
n
-
s
p
am
wh
ich
ar
e
in
d
ee
d
n
o
n
-
s
p
am
m
ess
ag
es.
T
h
e
p
r
o
p
o
s
ed
C
NN
-
C
C
B
m
o
d
el
was
ev
alu
ated
ac
r
o
s
s
th
r
ee
d
atasets
:
s
tar
ter
,
m
ails
,
an
d
th
e
co
n
s
tr
u
cted
o
r
ig
in
al
d
ataset.
T
h
e
m
o
d
el
co
n
s
is
ten
tly
ac
h
iev
es
h
ig
h
p
er
f
o
r
m
an
ce
ac
r
o
s
s
th
ese
d
atasets
,
in
d
icatin
g
its
g
en
er
aliza
b
ilit
y
an
d
v
iab
ilit
y
.
T
h
e
h
ig
h
p
er
f
o
r
m
an
ce
ac
r
o
s
s
m
u
ltip
le
d
atasets
,
s
u
g
g
est
th
at
th
e
lear
n
ed
f
ea
tu
r
es a
r
e
n
o
t
d
ataset
-
s
p
ec
if
ic
b
u
t c
ap
tu
r
e
g
en
e
r
al
p
atter
n
s
o
f
s
p
am
m
ess
ag
es wit
h
in
o
th
er
d
o
m
ai
n
s
.
Fro
m
T
ab
le
5
,
all
t
h
e
d
atas
ets
o
b
tain
ed
h
ig
h
m
ea
s
u
r
em
en
t
m
etr
ic
ab
o
v
e
9
0
%,
in
d
i
ca
tin
g
th
e
ef
f
ec
tiv
en
ess
o
f
th
e
p
r
o
p
o
s
ed
m
o
d
el.
Sp
ec
if
ically
,
th
e
o
r
ig
i
n
al
d
ata
o
b
tain
ed
an
ac
cu
r
ac
y
o
f
0
.
9
9
7
,
p
r
ec
is
io
n
o
f
0
.
9
8
8
,
F
1
-
s
co
r
e
o
f
0
.
9
9
3
,
an
d
r
ec
all
o
f
0
.
9
9
8
.
Star
ter
d
ataset
o
b
tain
ed
an
ac
cu
r
ac
y
o
f
0
.
9
8
6
,
p
r
ec
is
io
n
o
f
0
.
9
4
7
,
F
1
-
s
co
r
e
o
f
0
.
9
5
0
,
a
n
d
a
r
ec
all
o
f
0
.
9
5
3
.
W
h
ile
m
ails
d
ataset
o
b
tain
ed
a
n
ac
cu
r
ac
y
0
.
9
6
5
,
p
r
e
cisi
o
n
o
f
0
.
9
0
8
,
F
1
-
s
co
r
e
0
.
9
4
0
,
an
d
r
ec
all
o
f
0
.
9
7
6
.
T
h
o
u
g
h
s
tar
ter
d
a
taset
o
u
tp
er
f
o
r
m
s
m
ails
,
m
ails
o
b
tain
ed
a
h
ig
h
er
r
ec
all
v
alu
e.
T
ab
le
5
.
Mo
d
el
an
aly
s
is
o
v
er
d
if
f
er
en
t
d
atasets
A
l
g
o
r
i
t
h
m
A
c
c
u
r
a
c
y
P
r
e
c
i
s
i
o
n
F1
-
S
c
o
r
e
R
e
c
a
l
l
Lo
ss
S
t
a
r
t
e
r
D
a
t
a
se
t
0
.
9
8
6
0
.
9
4
7
0
.
9
5
0
0
.
9
5
3
0
.
1
0
9
M
a
i
l
s D
a
t
a
se
t
0
.
9
6
5
0
.
9
0
8
0
.
9
4
0
0
.
9
7
6
0
.
1
1
8
O
r
i
g
i
n
a
l
D
a
t
a
0
.
9
9
7
0
.
9
8
8
0
.
9
9
3
0
.
9
9
8
0
.
0
4
8
T
h
e
ab
latio
n
s
tu
d
y
in
Fig
u
r
e
4
h
ig
h
lig
h
ts
th
e
c
o
n
tr
ib
u
tio
n
o
f
ea
ch
c
o
m
p
o
n
en
t
i
n
th
e
p
r
o
p
o
s
ed
C
NN
-
C
C
B
m
o
d
el.
T
h
e
f
u
ll
m
o
d
el
a
ch
iev
es
th
e
h
ig
h
est
p
e
r
f
o
r
m
an
ce
ac
r
o
s
s
all
m
etr
ics
(
ac
cu
r
ac
y
,
p
r
ec
is
io
n
,
r
ec
all,
an
d
F1
-
s
co
r
e
)
,
c
o
n
f
ir
m
i
n
g
t
h
e
ef
f
ec
tiv
en
ess
o
f
c
o
m
b
in
i
n
g
m
u
ltip
le
f
ea
tu
r
e
ty
p
es
with
class
weig
h
tin
g
.
Firstl
y
,
r
em
o
v
in
g
co
n
ten
t
f
ea
tu
r
es
r
esu
lts
in
a
s
lig
h
t
d
ec
lin
e
in
th
e
p
er
f
o
r
m
an
ce
m
etr
ics.
T
h
is
in
d
icate
s
th
at
co
n
ten
t
f
ea
tu
r
es
p
r
o
v
id
e
u
s
ef
u
l
co
n
tex
tu
al
s
ig
n
als
th
at
im
p
r
o
v
e
t
h
e
c
lass
if
icatio
n
.
W
ith
o
u
t
th
e
ch
ar
ac
ter
f
ea
tu
r
es
lead
s
to
a
m
o
r
e
n
o
ticea
b
le
d
r
o
p
in
t
h
e
ev
alu
atio
n
m
etr
ics.
T
h
is
s
h
o
ws
th
at
ch
ar
ac
ter
f
ea
tu
r
e
p
att
er
n
s
ar
e
p
ar
ticu
lar
ly
im
p
o
r
tan
t f
o
r
co
r
r
ec
tly
id
en
tif
y
in
g
s
p
am
m
ess
ag
es.
T
h
e
m
o
d
el
p
r
o
d
u
ce
d
th
e
lo
west p
er
f
o
r
m
an
ce
with
o
u
t b
o
th
co
n
ten
t
a
n
d
ch
ar
ac
ter
f
ea
tu
r
es,
wh
ich
d
em
o
n
s
tr
ates
th
at
r
el
y
in
g
s
o
lely
o
n
wo
r
d
e
m
b
ed
d
in
g
s
is
in
s
u
f
f
icien
t
f
o
r
ca
p
tu
r
in
g
th
e
d
i
v
er
s
e
p
atter
n
s
p
r
esen
t
in
SMS
s
p
am
.
I
n
ad
d
itio
n
,
in
clu
d
in
g
class
weig
h
tin
g
h
elp
s
th
e
m
o
d
el
t
o
b
etter
im
p
r
o
v
e
d
etec
tio
n
s
en
s
itiv
ity
,
b
ec
au
s
e
ex
clu
d
in
g
class
weig
h
tin
g
r
e
d
u
ce
s
o
v
e
r
all
p
er
f
o
r
m
an
ce
as we
ll.
Fig
u
r
e
4
.
Ab
latio
n
s
tu
d
y
with
co
n
ten
t f
ea
tu
r
es,
ch
ar
ac
ter
f
ea
tu
r
es,
an
d
class
weig
h
tin
g
Fro
m
T
ab
le
6
,
t
h
e
m
ea
n
an
d
s
tan
d
ar
d
d
e
v
iatio
n
q
u
an
tif
y
th
e
s
tab
ilit
y
a
n
d
c
o
n
s
is
ten
cy
o
f
th
e
p
r
o
p
o
s
ed
m
o
d
el’
s
p
er
f
o
r
m
an
c
e
ac
r
o
s
s
m
u
ltip
le
r
u
n
s
.
T
h
e
m
ea
n
r
e
p
r
esen
ts
th
e
a
v
er
ag
e
p
er
f
o
r
m
an
ce
ac
r
o
s
s
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
0
8
8
-
8
7
0
8
I
n
t J E
lec
&
C
o
m
p
E
n
g
,
Vo
l.
1
6
,
No
.
4
,
Au
g
u
s
t
20
2
6
:
2
0
6
1
-
2073
2070
m
u
ltip
le
r
u
n
s
,
in
d
icatin
g
th
e
e
f
f
ec
tiv
en
ess
o
f
th
e
m
o
d
el.
W
h
ile
th
e
s
tan
d
a
r
d
d
ev
iatio
n
m
ea
s
u
r
es
th
e
m
o
d
el’
s
co
n
s
is
ten
cy
ac
r
o
s
s
r
u
n
s
with
d
if
f
er
en
t r
an
d
o
m
d
ata
s
p
lits
.
T
h
e
lo
w
s
tan
d
ar
d
d
e
v
iatio
n
v
al
u
e
s
co
n
f
ir
m
s
th
at
th
e
m
o
d
el
is
s
tab
le
an
d
h
as
h
i
g
h
lev
el
o
f
g
e
n
er
aliza
b
ilit
y
.
T
h
is
lo
w
v
ar
ian
ce
in
d
icate
s
th
at
th
e
p
r
o
p
o
s
ed
m
o
d
el
p
r
o
d
u
ce
s
co
n
s
is
ten
t
r
esu
lts
u
n
d
er
d
if
f
er
e
n
t
tr
ain
in
g
co
n
d
itio
n
s
,
an
d
th
is
co
n
s
is
ten
cy
s
tr
en
g
th
en
s
th
e
claim
o
f
g
en
er
aliza
tio
n
.
T
ab
le
6
.
Var
iab
ilit
y
a
n
aly
s
is
o
f
p
r
o
p
o
s
ed
m
o
d
el
M
e
a
n
S
t
a
n
d
a
r
d
d
e
v
i
a
t
i
o
n
A
c
c
u
r
a
c
y
0
.
9
9
7
0
.
0
0
0
1
P
r
e
c
i
s
i
o
n
0
.
9
8
8
0
.
0
0
0
3
R
e
c
a
l
l
0
.
9
9
7
0
.
0
0
0
2
F1
-
S
c
o
r
e
0
.
9
9
2
0
.
0
0
0
2
T
ab
le
7
s
h
o
ws
th
at
th
e
p
r
o
p
o
s
ed
m
o
d
el
(
C
NN
-
C
C
B
)
f
o
r
SMS
s
p
am
d
etec
tio
n
o
u
tp
er
f
o
r
m
s
o
th
er
m
o
d
els
p
r
o
p
o
s
ed
b
y
e
x
is
tin
g
s
tu
d
ies.
So
m
e
s
tu
d
ies
also
h
a
v
e
ex
p
l
o
r
ed
t
h
e
u
s
e
o
f
C
NN
f
o
r
s
p
am
d
etec
tio
n
[
2
1
]
,
[
2
2
]
,
b
u
t
th
eir
m
o
d
els’
ef
f
ec
tiv
en
ess
wer
e
lo
w
co
m
p
a
r
e
d
to
o
u
r
p
r
o
p
o
s
ed
m
o
d
el.
T
h
is
m
ay
b
e
as
a
r
esu
lt
o
f
th
ese
s
tu
d
ies
f
o
cu
s
in
g
o
n
wo
r
d
-
lev
el
to
k
e
n
izer
s
alo
n
e,
wh
ich
af
f
ir
m
s
th
e
u
s
ef
u
ln
ess
o
f
o
u
r
ar
ch
itectu
r
al
d
esig
n
ch
o
ice.
I
n
th
e
q
u
est
to
im
p
r
o
v
e
s
p
a
m
class
if
icatio
n
,
s
o
m
e
s
tu
d
ies
p
r
o
p
o
s
ed
h
y
b
r
id
m
o
d
els
b
y
am
alg
am
atin
g
C
NN
an
d
L
ST
M
[
1
]
,
[
1
4
]
.
Ho
wev
er
,
th
ey
wer
e
f
o
c
u
s
ed
o
n
s
m
all
d
atase
t,
p
r
o
d
u
cin
g
r
elativ
e
lo
w
p
er
f
o
r
m
an
ce
,
as
co
m
p
a
r
e
d
to
o
u
r
p
r
o
p
o
s
ed
m
o
d
el.
Similar
ca
s
es
o
cc
u
r
r
ed
with
m
ac
h
i
n
e
lear
n
in
g
m
o
d
els,
s
u
ch
as
th
e
in
teg
r
atio
n
o
f
Naïv
e
B
ay
es
an
d
SVM
[
2
3
]
.
L
STM
,
b
ein
g
a
g
o
o
d
tex
t
class
if
ier
was
ex
p
lo
r
ed
in
d
ep
en
d
en
tly
b
y
s
o
m
e
r
esear
ch
er
s
an
d
s
h
o
wn
r
esu
lts
o
f
8
8
.
3
3
%
an
d
9
7
.
5
8
%
[
9
]
,
[
2
2
]
,
m
ak
in
g
o
u
r
C
NN
-
C
C
B
m
o
r
e
ef
f
icien
t
f
o
r
s
p
am
d
etec
tio
n
.
I
n
th
e
u
s
e
o
f
b
id
ir
ec
tio
n
al
lo
n
g
s
h
o
r
t
-
ter
m
m
em
o
r
y
(
B
iLST
M
)
wh
ich
in
clu
d
es
th
e
f
o
r
war
d
a
n
d
b
ac
k
war
d
lear
n
i
n
g
tec
h
n
iq
u
e
,
th
o
u
g
h
a
p
p
ea
r
s
to
attain
r
elativ
ely
g
o
o
d
p
er
f
o
r
m
an
ce
o
f
9
8
.
4
6
%
an
d
9
8
.
0
5
%,
s
till
u
n
d
er
p
er
f
o
r
m
ed
in
co
m
p
ar
i
s
o
n
to
o
u
r
p
r
o
p
o
s
ed
C
NN
-
C
C
B
m
o
d
el.
laten
t
s
em
an
tic
in
d
ex
in
g
(
L
SI)
tech
n
iq
u
e
was
em
p
lo
y
ed
with
d
e
ep
lear
n
in
g
m
eth
o
d
f
o
r
ef
f
ec
t
iv
e
s
p
am
d
etec
tio
n
,
an
d
a
tr
a
n
s
f
o
r
m
e
r
-
b
ased
m
o
d
el:
B
id
ir
ec
tio
n
al
en
co
d
e
r
r
e
p
r
esen
tatio
n
s
f
r
o
m
tr
an
s
f
o
r
m
e
r
s
(
B
E
R
T
)
was
also
u
s
ed
[
2
4
]
,
[
2
5
]
.
An
o
t
h
er
s
tu
d
y
h
ig
h
lig
h
ted
th
e
ab
ilit
y
o
f
B
E
R
T
m
o
d
el
to
e
x
tr
ac
t
f
ea
t
u
r
es
b
ef
o
r
e
p
ass
in
g
o
n
to
m
ac
h
in
e
lear
n
in
g
alg
o
r
ith
m
[
2
6
]
,
b
u
t o
u
r
p
r
o
p
o
s
ed
m
o
d
el
o
u
tp
er
f
o
r
m
s
B
E
R
T
+
ML
a
lg
o
r
it
h
m
.
T
h
o
u
g
h
s
o
m
e
o
f
th
ese
p
r
ev
io
u
s
s
tu
d
ies
u
s
ed
s
tate
-
of
-
th
e
-
ar
t
tex
t
class
if
ies,
th
ey
wer
e
f
o
cu
s
ed
o
n
th
e
u
s
e
o
f
wo
r
d
-
lev
el
f
ea
tu
r
es,
w
h
ich
m
ay
n
o
t
b
e
s
u
f
f
icien
t
f
o
r
SMS
s
p
am
d
etec
tio
n
in
r
ea
l
t
im
e,
b
ec
au
s
e
SMS
m
ess
ag
es
co
u
ld
s
o
m
etim
es
b
e
s
p
o
n
tan
e
o
u
s
.
Ad
d
itio
n
ally
,
u
n
lik
e
th
ese
p
r
ev
io
u
s
s
tu
d
ies
wh
ich
wer
e
f
o
cu
s
ed
o
n
s
m
all
d
ata
s
ize
th
at
a
r
e
v
e
r
y
im
b
ala
n
ce
d
,
o
u
r
s
tu
d
y
wo
r
k
ed
with
d
ata
th
at
was
cr
ea
te
d
f
r
o
m
o
th
er
s
m
all
SMS sp
am
d
ata,
m
ak
in
g
it lar
g
er
an
d
m
in
im
izin
g
p
r
o
b
lem
s
ass
o
ciate
d
with
im
b
alan
ce
d
at
a.
T
ab
le
7
.
Me
tr
ics
co
m
p
ar
is
o
n
o
f
p
r
o
p
o
s
ed
m
o
d
el
with
m
ac
h
i
n
e
lear
n
in
g
alg
o
r
ith
m
s
A
l
g
o
r
i
t
h
m
A
c
c
u
r
a
c
y
(
%
)
P
r
e
c
i
s
i
o
n
(
%
)
G
h
a
n
e
m
a
n
d
Er
b
a
y
(
2
0
2
0
)
C
N
N
9
8
.
0
0
-
W
a
n
i
e
t
a
l
.
(
2
0
2
4
)
LSTM
w
i
t
h
W
o
r
d
2
V
e
c
9
7
.
5
8
9
7
.
5
W
a
n
i
e
t
a
l
.
(
2
0
2
4
)
C
N
N
w
i
t
h
W
o
r
d
2
V
e
c
9
7
.
6
1
9
7
.
0
R
a
j
a
se
k
h
a
r
e
t
a
l
.
(
2
0
2
3
)
LSTM
8
8
.
3
3
-
G
h
o
u
r
a
b
i
e
t
a
l
.
(
2
0
2
0
)
C
N
N
-
LSTM
9
8
.
3
7
9
5
.
3
9
R
a
i
h
e
n
e
t
a
l
.
(
2
0
2
4
)
B
ER
T
9
8
.
8
1
9
7
.
2
5
W
a
n
i
e
t
a
l
.
(
2
0
2
4
)
B
i
LST
M
9
8
.
4
6
9
8
.
0
5
H
o
ssa
i
n
e
t
a
l
.
(
2
0
2
2
)
C
N
N
-
LSTM
9
8
.
4
0
9
8
.
0
0
Al
-
O
t
a
i
b
i
e
t
a
l
.
(
2
0
2
5
)
DL
-
LSI
8
9
.
0
0
8
9
.
0
0
V
a
su
d
e
v
a
n
e
t
a
l
.
(
2
0
2
4
)
N
a
ï
v
e
B
a
y
e
s
+
S
V
M
9
7
.
4
9
-
G
u
o
e
t
a
l
.
(
2
0
2
2
)
B
ER
T+
M
L
A
l
g
o
r
i
t
h
m
-
9
7
.
8
6
P
r
o
p
o
se
d
m
o
d
e
l
(
C
N
N
-
C
C
B
)
9
9
.
7
9
8
.
8
5.
CO
NCLU
SI
O
N
W
e
p
r
o
p
o
s
ed
an
SMS
s
p
am
d
etec
tio
n
m
o
d
el
(
C
NN
-
C
C
B
)
,
tak
in
g
in
to
co
n
s
id
er
atio
n
co
n
ten
t
an
d
ch
ar
ac
ter
-
b
ased
f
ea
tu
r
es
in
v
o
l
v
ed
in
tex
ts
.
T
h
e
co
n
ca
ten
atio
n
o
f
th
ese
f
ea
tu
r
es
en
ab
les
th
e
m
o
d
el
to
lear
n
h
ig
h
er
-
le
v
el
in
ter
ac
tio
n
s
b
etw
ee
n
v
ar
iety
o
f
tech
n
ical
f
ea
tu
r
es,
wh
ich
en
h
a
n
ce
d
class
if
icatio
n
p
er
f
o
r
m
a
n
ce
ac
r
o
s
s
all
ev
alu
atio
n
m
etr
ics.
T
h
e
u
s
e
o
f
p
ar
allel
co
n
v
o
lu
tio
n
al
lay
er
s
is
to
allo
w
th
e
m
o
d
el
ca
p
tu
r
e
lo
ca
l
n
-
g
r
am
p
atter
n
s
as
well
as
ch
ar
ac
ter
p
atter
n
s
ef
f
icien
tly
.
Als
o
,
class
weig
h
ts
wer
e
em
p
lo
y
ed
to
h
elp
s
o
lv
e
th
e
p
r
o
b
lem
o
f
im
b
alan
ce
d
d
ata
wh
ich
s
o
m
etim
es lea
d
to
p
o
o
r
p
r
ed
ictio
n
.
Evaluation Warning : The document was created with Spire.PDF for Python.