Co
m
pu
t
er
Science
a
nd
I
nfo
r
m
a
t
io
n T
ec
hn
o
lo
g
ies
Vo
l.
1
,
No
.
2
,
J
u
ly
2
0
2
0
,
p
p
.
84
~
92
I
SS
N:
2722
-
3
2
2
1
,
DOI
: 1
0
.
1
1
5
9
1
/
csit
.
v
1
i
2
.p
84
-
92
84
J
o
ur
na
l
ho
m
ep
a
g
e
:
h
ttp
:
//ia
e
s
p
r
ime.
co
m/in
d
ex
.
p
h
p
/csi
t
Ins
ult
de
tect
io
n
u
sing
a
pa
rtit
io
na
l
CNN
-
LST
M
mo
del
M
o
ha
m
ed
M
a
her
B
en
I
s
m
a
il
Co
m
p
u
ter
S
c
ien
c
e
De
p
a
rtme
n
t,
C
o
ll
e
g
e
o
f
Co
m
p
u
ter an
d
In
fo
rm
a
ti
o
n
S
c
ien
c
e
s
,
Kin
g
S
a
u
d
Un
i
v
e
rsit
y
,
Riy
a
d
h
,
Kin
g
d
o
m
o
f
S
a
u
d
i
Ara
b
i
a
Art
icle
I
nfo
AB
S
T
RAC
T
A
r
ticle
his
to
r
y:
R
ec
eiv
ed
Sep
2
3
,
2
0
1
9
R
ev
is
ed
Ap
r
2
0
,
2
0
2
0
Acc
ep
ted
Ma
y
9
,
2
0
2
0
Re
c
e
n
tl
y
,
d
e
e
p
lea
rn
i
n
g
h
a
s
b
e
e
n
c
o
u
p
le
d
with
n
o
ti
c
e
a
b
le
a
d
v
a
n
c
e
s
in
Na
tu
ra
l
Lan
g
u
a
g
e
P
r
o
c
e
ss
in
g
re
late
d
re
se
a
rc
h
.
In
t
h
is
w
o
rk
,
we
p
r
o
p
o
se
a
g
e
n
e
ra
l
fra
m
e
wo
rk
to
d
e
tec
t
v
e
rb
a
l
o
f
fe
n
se
in
so
c
ial
n
e
two
rk
s
c
o
m
m
e
n
ts.
We
in
tro
d
u
c
e
a
p
a
rti
ti
o
n
a
l
CNN
-
LS
TM
a
rc
h
it
e
c
tu
re
i
n
o
r
d
e
r
to
a
u
t
o
m
a
ti
c
a
ll
y
re
c
o
g
n
ize
v
e
r
b
a
l
o
ffe
n
se
p
a
t
tern
s
in
so
c
ial
n
e
two
r
k
c
o
m
m
e
n
ts.
S
p
e
c
ifi
c
a
ll
y
,
we
u
se
a
p
a
rti
ti
o
n
a
l
CNN
a
lo
n
g
w
it
h
a
LS
T
M
m
o
d
e
l
t
o
m
a
p
th
e
so
c
i
a
l
n
e
two
rk
c
o
m
m
e
n
ts i
n
to
two
p
re
d
e
fin
e
d
c
las
se
s.
In
p
a
rti
c
u
lar,
ra
th
e
r
th
a
n
c
o
n
sid
e
rin
g
a
wh
o
le
d
o
c
u
m
e
n
t
/co
m
m
e
n
ts
a
s
i
n
p
u
t
a
s
p
e
rfo
rm
e
d
u
sin
g
ty
p
ica
l
CNN
,
we
p
a
rti
ti
o
n
t
h
e
c
o
m
m
e
n
ts
in
to
p
a
rts
in
o
r
d
e
r
to
c
a
p
t
u
re
a
n
d
we
i
g
h
t
t
h
e
lo
c
a
ll
y
re
lev
a
n
t
in
f
o
rm
a
ti
o
n
in
e
a
c
h
p
a
rti
ti
o
n
.
Th
e
re
su
lt
in
g
lo
c
a
l
i
n
fo
rm
a
ti
o
n
is
t
h
e
n
se
q
u
e
n
ti
a
ll
y
e
x
p
l
o
it
e
d
a
c
ro
ss
p
a
rti
ti
o
n
s
u
sin
g
L
S
TM
f
o
r
v
e
rb
a
l
o
ffe
n
se
d
e
tec
ti
o
n
.
Th
e
c
o
m
b
in
a
ti
o
n
o
f
th
e
p
a
rti
ti
o
n
a
l
CNN
a
n
d
LS
TM
y
iel
d
s
th
e
in
teg
ra
ti
o
n
o
f
t
h
e
l
o
c
a
l
with
in
c
o
m
m
e
n
ts
in
fo
rm
a
ti
o
n
a
n
d
th
e
lo
n
g
d
istan
c
e
c
o
rre
latio
n
a
c
ro
ss
c
o
m
m
e
n
ts.
Th
e
p
ro
p
o
se
d
a
p
p
ro
a
c
h
wa
s
a
ss
e
ss
e
d
u
sin
g
re
a
l
d
a
tas
e
t,
a
n
d
t
h
e
o
b
tain
e
d
re
su
lt
s
p
ro
v
e
d
th
a
t
o
u
r
s
o
lu
ti
o
n
o
u
tp
e
rfo
r
m
s e
x
isti
n
g
re
lev
a
n
t
so
l
u
ti
o
n
s.
K
ey
w
o
r
d
s
:
Dee
p
lear
n
in
g
,
I
n
s
u
lt d
etec
tio
n
,
So
cial
n
et
wo
r
k
s
,
Su
p
er
v
is
ed
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
:
Mo
h
am
ed
Ma
h
e
r
B
en
I
s
m
ail
,
C
o
m
p
u
ter
Scien
ce
Dep
ar
tm
e
n
t,
C
o
lleg
e
o
f
C
o
m
p
u
ter
an
d
I
n
f
o
r
m
atio
n
Scien
ce
s
,
Kin
g
Sau
d
Un
iv
e
r
s
ity
,
R
iy
ad
h
,
KSA
.
E
m
ail:
ts
@
ee
.
u
ad
.
ac
.
id
1.
I
NT
RO
D
UCT
I
O
N
T
h
e
g
r
o
wth
o
f
th
e
wo
r
ld
p
o
p
u
latio
n
as
well
as
th
e
tech
n
o
lo
g
ical
ad
v
an
ce
s
h
av
e
led
a
n
ew
er
a
o
f
co
m
m
u
n
icatio
n
an
d
s
o
cializa
tio
n
th
r
o
u
g
h
v
ir
tu
al
p
latf
o
r
m
s
s
u
ch
as
Yo
u
T
u
b
e,
I
n
s
tag
r
am
,
T
witter
,
an
d
L
in
k
ed
I
n
.
No
wad
ay
s
,
b
illi
o
n
s
o
f
p
eo
p
le
all
ar
o
u
n
d
th
e
wo
r
l
d
jo
in
ed
s
o
cial
n
etwo
r
k
s
wh
ich
r
eq
u
ir
e
a
b
asic
k
n
o
wled
g
e
o
f
t
h
e
co
m
p
u
ter
f
u
n
d
am
e
n
tals
.
B
esid
es,
th
e
o
u
ts
p
r
ea
d
u
s
e
o
f
th
e
s
m
ar
t
d
ev
ices
alo
n
g
with
th
e
ex
ce
s
s
iv
e
u
s
e
o
f
s
o
cial
n
et
wo
r
k
s
h
as
g
r
an
ted
th
em
t
h
e
ab
ilit
y
to
f
o
r
m
v
ar
io
u
s
v
ir
tu
al
s
o
cieties
wh
er
e
p
eo
p
l
e
ca
n
co
n
tin
u
o
u
s
ly
e
x
ch
an
g
e
id
ea
s
,
in
ter
ests
an
d
co
n
ce
r
n
s
.
T
h
is
r
esu
lted
in
a
n
ew
life
s
ty
le
wh
er
e
th
ey
r
eg
u
lar
l
y
f
o
llo
w,
s
h
ar
e
an
d
g
et
u
p
d
ates
o
n
ev
en
ts
th
at
ar
e
h
eld
in
th
eir
ac
tu
al
s
o
c
iety
.
I
n
f
ac
t,
p
eo
p
le
ar
e
u
s
in
g
s
o
cial
n
etwo
r
k
s
f
o
r
v
ar
io
u
s
p
u
r
p
o
s
es
r
eg
ar
d
less
o
f
th
eir
eth
n
ic
ity
,
n
atio
n
ality
,
ed
u
ca
tio
n
a
n
d
b
ac
k
g
r
o
u
n
d
.
I
n
p
ar
ticu
lar
,
s
o
cial
n
etwo
r
k
s
en
ab
le
th
e
u
s
er
s
to
in
ter
ac
t
with
th
eir
p
ee
r
s
.
Su
ch
in
v
o
lv
em
e
n
t
o
f
th
e
wo
r
ld
wid
e
p
o
p
u
latio
n
in
d
ig
ital
s
o
ciety
h
as
y
ield
ed
v
ar
io
u
s
c
h
allen
g
e
s
an
d
s
id
e
e
f
f
ec
ts
.
Fo
r
in
s
tan
ce
,
s
ec
u
r
ity
,
s
p
am
d
etec
tio
n
an
d
p
r
iv
ac
y
p
r
o
tectio
n
h
as
em
er
g
ed
as
cr
itical
ch
allen
g
es
f
ac
in
g
s
o
cial
n
etwo
r
k
p
r
o
f
ess
io
n
als
an
d
co
m
p
an
ies.
Go
v
er
n
m
en
ts
,
s
u
c
h
as
in
Sau
d
i
Ar
ab
ia,
h
av
e
e
s
tab
lis
h
ed
th
e
C
o
m
m
u
n
icatio
n
s
an
d
I
n
f
o
r
m
atio
n
T
ec
h
n
o
lo
g
y
C
o
m
m
is
s
io
n
to
o
v
er
co
m
e
th
ese
ch
allen
g
es
an
d
t
o
co
p
e
with
th
e
r
ad
ical
ch
a
n
g
es
t
h
at
r
ap
id
ly
h
ap
p
en
in
th
e
d
ig
ital
wo
r
l
d
[
1
]
.
th
ey
h
a
v
e
also
r
eg
u
lated
An
ti
-
C
y
b
er
C
r
im
e
L
aw
to
b
e
im
p
lem
en
ted
th
r
o
u
g
h
g
o
v
er
n
m
en
t
d
ep
ar
tm
en
t
s
u
ch
as
Min
is
tr
y
o
f
I
n
ter
io
r
[
2
]
a
n
d
P
u
b
lic
P
r
o
s
ec
u
tio
n
[
3
]
to
av
o
i
d
an
y
u
n
eth
ical
m
is
u
s
e
o
f
th
e
s
o
cial
n
etwo
r
k
s
an
d
to
p
r
ev
en
t
an
y
v
io
latio
n
s
th
at
m
ay
o
cc
u
r
with
in
th
e
cy
b
er
s
p
ac
e.
T
h
is
p
r
o
v
es
th
at
s
o
m
e
Evaluation Warning : The document was created with Spire.PDF for Python.
C
o
m
p
u
t.
Sci.
I
n
f
.
T
ec
h
n
o
l.
I
n
s
u
lt d
etec
tio
n
u
s
in
g
a
p
a
r
titi
o
n
a
l CN
N
-
LS
TM
mo
d
el
… (
M
o
h
a
med
Ma
h
er B
en
I
s
ma
il
)
85
o
f
th
e
ch
allen
g
es
f
ac
ed
b
y
th
e
d
ig
ital
co
m
m
u
n
ities
ar
e
cr
itical
an
d
r
eq
u
ir
e
r
ea
l
ef
f
o
r
ts
t
o
lim
it
th
eir
im
p
ac
t
o
n
p
eo
p
le
d
aily
life
.
Desp
ite
th
es
e
ef
f
o
r
ts
,
s
o
m
e
s
o
cial
m
ed
ia
u
s
er
s
b
r
ea
k
th
e
co
m
m
u
n
icatio
n
eth
ics
co
d
e
wh
ile
m
ess
ag
in
g
,
d
is
cu
s
s
in
g
o
r
c
o
m
m
en
tin
g
o
n
s
o
cial
m
ed
ia.
T
h
is
b
eh
av
io
r
ca
n
b
e
attr
ib
u
ted
to
m
an
y
f
ac
t
o
r
s
s
u
ch
as
th
eir
p
s
y
c
h
o
lo
g
ical
co
n
d
itio
n
,
l
o
w
ed
u
ca
tio
n
lev
el
o
r
liv
in
g
en
v
ir
o
n
m
en
t.
I
n
p
a
r
ticu
lar
,
tex
tu
al
in
s
u
lt
is
a
ty
p
ical
illu
s
tr
atio
n
o
f
th
is
p
r
o
b
lem
.
A
ty
p
ical
tex
tu
al
in
s
u
lt
co
n
s
is
ts
in
th
e
u
s
e
o
f
v
o
ca
b
u
lar
y
wh
ich
h
ar
m
s
th
e
u
s
er
b
ein
g
co
m
m
u
n
icatin
g
with
.
Su
ch
o
f
f
en
s
e
is
o
f
ten
h
ar
d
to
s
en
s
e
b
ec
au
s
e
its
p
a
tter
n
s
ex
h
ib
it
h
ig
h
v
ar
iab
ilit
y
.
T
y
p
ically
,
it
ca
n
b
e
d
ir
ec
t
in
s
u
lt,
in
tim
id
atio
n
,
s
h
o
u
t
o
r
t
h
r
ea
t.
Ho
wev
e
r
,
wh
atev
er
th
e
f
o
r
m
it
tak
es,
it
r
em
ain
s
u
n
ac
ce
p
tab
le
f
o
r
th
e
m
ajo
r
ity
o
f
u
s
er
s
.
Mo
r
eo
v
er
,
co
n
s
er
v
ativ
e
s
o
cieties
ar
e
m
o
r
e
s
en
s
itiv
e
to
s
u
ch
p
h
en
o
m
en
a.
T
h
u
s
,
ag
g
r
ess
iv
e
b
eh
av
io
r
th
r
o
u
g
h
th
r
ea
ts
b
y
i
m
p
ly
in
g
ab
u
s
e
s
u
ch
as
“Do
n
’
t
y
o
u
d
ar
e
d
o
th
at
o
r
I
’
ll
p
u
n
ch
y
o
u
r
lig
h
ts
o
u
t!
”
is
also
n
o
t
ac
ce
p
ted
.
Similar
ly
,
f
o
wl
n
am
e
u
s
ag
e
s
u
ch
as
“Yo
u
’
r
e
a
s
tu
p
id
g
o
o
d
f
o
r
n
o
th
in
g
!
”
is
n
o
t
to
ler
ated
.
De
s
p
ite
au
th
o
r
ity
ef
f
o
r
ts
to
s
u
ite
u
s
er
s
wh
o
o
f
f
e
n
d
o
th
e
r
s
in
s
o
cial
m
ed
ia
th
r
o
u
g
h
ap
p
r
o
p
r
iate
leg
is
latio
n
,
in
cr
ea
s
in
g
am
o
u
n
t
o
f
in
s
u
lts
ar
e
r
eg
u
lar
ly
r
ep
o
r
te
d
o
n
s
o
cial
n
etwo
r
k
s
.
T
h
u
s
,
v
er
b
a
l
o
f
f
en
s
es
h
av
e
b
ec
o
m
e
th
e
is
s
u
e
th
at
m
o
s
t
o
f
th
e
u
s
er
s
f
ac
e
wh
en
co
n
n
ec
tin
g
to
v
ir
tu
al
s
o
cieties.
Sad
ly
,
u
s
er
s
m
u
s
t
h
an
d
le
m
a
n
u
ally
s
u
ch
c
o
n
ce
r
n
.
Fo
r
ex
a
m
p
le,
th
e
ad
m
in
is
tr
ato
r
s
o
f
Face
b
o
o
k
p
a
g
es
s
h
o
u
ld
s
cr
ee
n
all
co
m
m
en
ts
o
n
e
v
er
y
s
in
g
le
p
o
s
t
an
d
d
is
ca
r
d
in
s
u
lts
.
T
h
is
m
an
u
al
s
o
lu
tio
n
is
s
u
b
jectiv
e
a
n
d
lab
o
r
d
em
an
d
in
g
esp
ec
ially
wh
en
th
e
n
u
m
b
er
o
f
co
m
m
e
n
ts
to
h
an
d
le
is
co
n
s
i
d
er
ab
ly
lar
g
e.
Mo
r
e
o
v
er
,
g
iv
e
n
th
e
co
n
tin
u
o
u
s
ly
g
p
o
win
g
n
u
m
b
e
r
o
f
u
s
er
s
,
b
lo
ck
in
g
th
e
u
s
er
alo
n
g
with
r
ep
o
r
tin
g
th
em
to
th
e
m
o
d
er
ato
r
s
h
as
also
b
ec
o
m
e
an
o
b
s
o
lete
alter
n
ativ
e.
T
h
e
r
ef
o
r
e
,
s
o
lu
tio
n
s
a
b
le
to
au
to
m
atica
lly
d
etec
t
v
er
b
al
in
s
u
lt e
m
e
r
g
ed
as a
n
u
r
g
en
t
n
ee
d
.
On
e
o
f
th
e
ea
r
lies
t
ef
f
o
r
ts
to
s
o
lv
e
th
is
p
r
o
b
le
m
in
an
u
n
s
u
p
er
v
is
ed
m
an
n
e
r
was
to
s
p
ec
if
y
a
lis
t
o
f
p
r
o
h
ib
ited
wo
r
d
s
s
o
th
at
if
an
y
wo
r
d
s
o
f
th
e
lis
t
ap
p
ea
r
ed
in
th
e
u
s
er
m
ess
ag
e,
th
e
m
ess
ag
e
o
r
co
m
m
en
t
will
b
e
r
ejec
ted
.
T
y
p
ically
,
s
u
ch
s
o
lu
tio
n
r
el
y
o
n
a
s
tatic
d
ictio
n
ar
y
alo
n
g
with
s
o
m
e
s
o
cio
-
lin
g
u
is
tic
p
atter
n
s
an
d
s
em
an
tic
r
u
les
[
4
-
6
]
.
Ho
wev
er
its
m
ain
d
r
awb
ac
k
co
n
s
is
ts
in
its
in
ab
ilit
y
to
d
ec
id
e
in
tellig
en
tly
if
th
e
te
x
t
is
an
in
s
u
lt
o
r
n
o
t
.
Fo
r
ex
am
p
le,
if
we
co
n
s
id
e
r
th
e
f
o
llo
win
g
two
c
o
m
m
e
n
ts
:
“T
h
is
id
ea
s
tu
p
id
”
an
d
“Yo
u
ar
e
s
tu
p
i
d
”.
T
h
e
s
ec
o
n
d
o
n
e
is
an
in
s
u
lt
wh
ile
th
e
f
ir
s
t
co
m
m
e
n
t
is
n
o
t.
A
ty
p
ical
p
r
o
h
i
b
ited
lis
t
b
ased
m
eth
o
d
ca
n
n
o
t
d
is
cr
im
in
ate
b
etwe
en
th
em
,
a
n
d
wo
u
ld
eith
er
r
ejec
t
o
r
to
ler
ate
b
o
th
co
m
m
en
ts
.
An
o
th
e
r
a
lter
n
ativ
e
to
d
etec
t
v
er
b
al
a
b
u
s
e
co
n
s
is
ts
in
th
e
f
o
r
m
u
latio
n
o
f
th
e
p
r
o
b
lem
a
s
tex
t
m
in
in
g
an
d
s
u
p
e
r
v
is
ed
lear
n
in
g
p
r
o
b
lem
(
class
if
icatio
n
)
.
I
n
f
ac
t,
th
es
class
if
ier
s
ar
e
in
ten
d
ed
to
d
eter
m
in
e
wh
eth
er
a
co
m
m
en
t
is
an
in
s
u
lt
o
r
n
o
t.
C
o
m
m
o
n
ly
,
s
o
m
e
tr
ain
i
n
g
c
o
m
m
en
ts
ar
e
f
i
r
s
t u
s
ed
to
lear
n
th
e
m
ap
p
i
n
g
b
etwe
en
th
e
an
n
o
tated
co
m
m
e
n
ts
an
d
th
e
two
p
r
ed
ef
in
e
d
class
es.
T
h
en
,
th
e
r
esu
ltin
g
m
ap
p
i
n
g
m
o
d
el
is
u
s
ed
to
au
to
m
atica
lly
p
r
ed
ict
th
e
class
v
alu
e
of
th
e
u
n
la
b
eled
co
m
m
e
n
ts
.
Desp
ite
r
esear
ch
er
’
s
ef
f
o
r
t
to
s
o
lv
e
v
ar
io
u
s
r
ea
l
wo
r
ld
ap
p
licatio
n
s
u
s
in
g
s
u
p
er
v
is
ed
lear
n
in
g
alg
o
r
ith
m
s
[
7
-
11
]
,
a
lim
ited
n
u
m
b
er
o
f
s
o
lu
tio
n
s
ab
le
t
o
d
etec
t
in
s
u
lt
s
in
s
o
cial
n
etwo
r
k
co
m
m
en
ts
in
an
u
n
s
u
p
er
v
is
e
d
m
an
n
er
h
as
b
ee
n
o
u
tlin
e
d
s
o
f
ar
.
L
ately
,
d
ee
p
lear
n
in
g
h
av
e
p
r
o
v
ed
to
b
e
p
r
o
m
is
in
g
ly
ac
cu
r
ate
in
p
r
ed
ic
tin
g
class
es
in
v
ar
io
u
s
ap
p
licatio
n
s
.
I
n
f
ac
t,
v
ar
io
u
s
d
ee
p
lea
r
n
in
g
m
o
d
els
h
av
e
b
ee
n
in
tr
o
d
u
ce
d
a
n
d
d
e
p
lo
y
ed
to
o
v
er
co
m
e
tex
t
class
if
icatio
n
ch
allen
g
es.
I
n
p
ar
ticu
lar
,
th
e
R
ec
u
r
r
en
t
Neu
r
al
Netwo
r
k
(
R
NN)
was
d
esig
n
e
d
to
ca
p
tu
r
e
s
em
an
tic
in
f
o
r
m
ati
o
n
s
eq
u
e
n
tially
th
r
o
u
g
h
f
ix
ed
len
g
th
h
id
d
en
lay
er
v
ec
to
r
s
wh
ich
p
r
o
ce
s
s
co
n
s
ec
u
tiv
e
tim
e
-
s
tep
wo
r
d
s
[
12
]
.
H
o
wev
er
,
s
u
ch
m
o
d
el
m
a
y
ex
h
i
b
it b
ias to
war
d
s
later
wo
r
d
s
wh
en
en
co
d
in
g
th
e
o
v
er
all
s
en
ten
ce
/co
m
m
en
t
s
em
an
tics
.
T
h
is
R
N
N
d
r
awb
ac
k
ca
n
in
ter
p
r
eted
as
a
r
esu
lt
o
f
an
ex
p
lo
d
in
g
g
r
ad
ien
t
wh
i
ch
y
ield
s
lar
g
e
u
p
d
ates
o
f
th
e
m
o
d
el
weig
h
ts
.
T
o
ad
d
r
ess
th
is
is
s
u
e,
th
e
L
o
n
g
Sh
o
r
t
-
T
er
m
Me
m
o
r
y
(
L
STM
)
n
etwo
r
k
was
i
n
tr
o
d
u
ce
d
in
[
13
]
t
o
b
etter
ca
p
t
u
r
e
t
h
e
s
h
o
r
t
a
n
d
lo
n
g
tim
e
d
ep
en
d
e
n
cies.
Mo
r
eo
v
er
,
it
w
as
in
ten
d
ed
to
a
d
d
r
ess
th
e
g
r
a
d
ien
t
ex
p
lo
s
io
n
an
d
g
r
ad
ie
n
t
d
if
f
u
s
io
n
p
r
o
b
le
m
s
in
h
er
ited
f
r
o
m
ty
p
ical
R
NNs
[
13
].
I
n
th
is
p
ap
er
we
p
r
o
p
o
s
e
a
p
ar
titi
o
n
al
C
NN
-
L
STM
ar
ch
i
tectu
r
e
to
b
u
ild
a
s
u
p
er
v
is
ed
l
ea
r
n
in
g
m
o
d
el
ab
l
e
to
d
etec
t
if
a
g
iv
en
co
m
m
en
t/s
en
ten
ce
r
ep
r
esen
ts
a
v
er
b
al
o
f
f
en
s
e.
T
h
e
p
r
o
p
o
s
ed
lo
ca
l
C
NN
p
r
o
ce
s
s
e
s
th
e
u
s
er
co
m
m
en
t
as
a
s
u
b
s
eq
u
en
ce
r
at
h
er
th
an
h
a
n
d
lin
g
th
e
wh
o
le
co
m
m
en
t/s
en
ten
ce
as
d
o
n
e
u
s
in
g
th
e
ty
p
ical
C
NN
m
o
d
els.
I
n
p
ar
ticu
lar
,
it
p
ar
titi
o
n
s
th
e
in
p
u
t
co
m
m
en
ts
in
to
s
eq
u
en
ce
s
in
a
wa
y
th
at
th
e
r
elev
a
n
t
in
f
o
r
m
atio
n
in
ea
ch
p
a
r
titi
o
n
is
ca
p
tu
r
ed
an
d
w
eig
h
ted
b
ased
o
n
its
r
elev
an
ce
to
t
h
e
o
f
f
en
s
e.
T
h
e
ca
p
tu
r
ed
lo
ca
l
i
n
f
o
r
m
atio
n
is
th
en
s
eq
u
en
tially
e
x
p
lo
ited
u
s
in
g
L
STM
an
d
c
o
u
p
le
d
wi
th
th
e
g
lo
b
al
d
e
p
en
d
e
n
cy
ex
tr
ac
ted
u
s
in
g
t
h
e
ty
p
ical
C
NN
in
o
r
d
er
b
etter
m
o
d
el
v
e
r
b
al
o
f
f
en
s
e
s
em
an
tic.
2.
RE
L
AT
E
D
WO
RK
S
I
n
s
u
lt
d
etec
tio
n
in
s
o
cial
n
etwo
r
k
co
m
m
en
ts
is
in
ten
d
ed
t
o
r
ejec
t
co
m
m
e
n
ts
co
n
v
e
y
in
g
in
s
u
ltin
g
m
ess
ag
es
in
an
au
to
m
atic
m
a
n
n
er
.
I
t
was
i
n
tr
o
d
u
ce
d
as
an
alter
n
ativ
e
to
s
u
p
p
o
r
t
a
n
d
/o
r
s
u
b
s
titu
te
th
e
m
an
u
al
ef
f
o
r
t
o
f
th
e
v
ir
tu
al
c
o
m
m
u
n
ity
ad
m
in
is
tr
ato
r
s
.
T
y
p
ically
,
s
u
p
er
v
is
ed
m
ac
h
in
e
lear
n
in
g
te
ch
n
iq
u
es
h
av
e
b
ee
n
ad
o
p
ted
b
y
th
e
r
ec
en
t
v
er
b
al
o
f
f
en
s
e
d
etec
tio
n
a
p
p
r
o
ac
h
es.
I
n
th
e
f
o
llo
win
g
s
u
b
-
s
ec
tio
n
s
,
we
o
u
tlin
e
th
e
s
tate
o
f
th
e
-
ar
t
tex
t
class
if
icatio
n
a
p
p
r
o
ac
h
es
b
ased
o
n
s
u
p
er
v
is
ed
lear
n
in
g
tec
h
n
iq
u
es
as
well
as
th
e
r
elev
a
n
t
d
ee
p
lear
n
in
g
tech
n
iq
u
es,
r
esp
ec
tiv
ely
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
7
2
2
-
3
2
2
1
C
o
m
p
u
t.
Sci.
I
n
f
.
T
ec
h
n
o
l.
,
V
o
l.
1
,
No
.
2
,
J
u
ly
20
2
0
:
84
–
92
86
2
.
1
.
Ver
ba
l
o
f
f
ens
e
det
ec
t
io
n us
i
ng
s
up
er
v
is
e
d lea
rning
t
ec
hn
iqu
es
R
ec
en
tly
,
m
an
y
r
esear
ch
e
r
s
h
a
v
e
co
n
tr
ib
u
ted
to
in
t
r
o
d
u
ce
v
a
r
io
u
s
s
o
lu
tio
n
s
to
ad
d
r
ess
th
e
p
r
o
b
lem
o
f
au
to
m
atic
v
er
b
al
o
f
f
en
s
e
d
etec
tio
n
f
o
r
s
o
cial
n
etwo
r
k
co
m
m
en
ts
.
T
h
e
au
th
o
r
s
in
in
[
12
]
p
r
esen
ted
a
s
o
lu
tio
n
th
at
ad
o
p
ts
a
s
tatic
s
o
cio
-
li
n
g
u
is
tic
b
ased
d
ictio
n
ar
y
t
o
d
e
tect
th
e
co
m
m
en
ts
in
clu
d
in
g
wo
r
d
s
f
r
o
m
th
e
d
ictio
n
ar
y
.
O
n
e
s
h
o
u
ld
n
o
t
e
th
at
th
e
r
ep
o
r
te
d
r
esu
lts
s
h
o
wed
lo
w
co
v
er
ag
e
a
n
d
h
i
g
h
f
a
ls
e
p
o
s
itiv
e
r
ates.
I
n
[
5
]
,
th
e
au
t
h
o
r
s
o
u
tlin
e
d
a
d
is
cr
im
in
atio
n
ap
p
r
o
ac
h
b
etwe
en
r
eg
u
lar
an
d
in
s
u
lt
s
tatem
en
ts
b
ased
o
n
s
en
ten
ce
s
p
ar
s
in
g
an
d
s
em
an
tic
r
u
les
u
s
ag
e.
T
h
e
s
o
lu
tio
n
in
tr
o
d
u
ce
d
in
[
6
]
to
r
ejec
t
in
s
u
ltin
g
co
m
m
en
ts
is
b
ased
o
n
th
e
b
ag
-
o
f
wo
r
d
s
f
ea
t
u
r
es
alo
n
g
with
a
d
ictio
n
a
r
y
th
at
in
cl
u
d
es
th
e
a
b
u
s
in
g
la
n
g
u
a
g
e.
I
n
[
14
]
,
a
lin
g
u
is
tic
an
aly
s
is
b
ased
in
s
u
lt
d
etec
ti
o
n
s
o
lu
tio
n
f
o
r
T
h
ai
tex
t
u
al
co
n
v
er
s
atio
n
s
was
p
r
o
p
o
s
ed
.
T
h
e
au
th
o
r
s
in
[
15
]
p
r
o
p
o
s
ed
an
o
n
lin
e
d
etec
tio
n
s
y
s
tem
th
at
d
etec
ts
h
ar
ass
m
en
t.
T
h
e
m
ain
g
o
al
was
to
d
eter
m
in
e
wh
eth
er
a
co
m
m
en
t
r
e
p
r
esen
ts
an
h
ar
ass
m
en
t
o
r
n
o
t.
No
te
th
at
th
ey
f
o
r
m
u
lated
th
e
h
ar
ass
m
en
t
d
etec
tio
n
as
a
s
en
tim
en
t
an
aly
s
is
p
r
o
b
lem
.
I
n
[
16
]
,
t
h
e
o
u
tlin
ed
s
y
s
tem
aim
s
to
ca
teg
o
r
ize
th
e
u
s
er
co
m
m
en
ts
as b
u
lly
in
g
o
r
n
o
t u
s
in
g
a
Mu
lti
-
C
r
iter
ia
E
v
alu
atio
n
Sy
s
t
em
(
MCES)
wh
ich
r
ev
o
lv
es
ar
o
u
n
d
th
e
co
n
ce
p
t
o
f
weig
h
tin
g
wo
r
d
s
b
ased
o
n
a
s
co
r
e
o
r
a
n
u
m
e
r
ical
v
alu
e.
I
n
[
1
7
]
,
th
e
r
esear
ch
er
s
in
t
r
o
d
u
ce
d
a
s
o
lu
tio
n
th
at
r
elies
o
n
th
e
lin
g
u
is
tic
r
eg
u
lar
ities
ca
p
tu
r
ed
i
n
p
r
o
f
an
e
lan
g
u
a
g
e
u
s
in
g
s
tatis
tical
to
p
ic
m
o
d
elin
g
.
A
s
to
ch
asti
c
g
r
ad
ie
n
t
d
esce
n
t
class
if
ier
was
u
s
ed
in
[
18
]
to
d
etec
t
in
s
u
lts
in
u
s
er
g
en
er
ated
Ar
a
b
ic
n
ewsp
ap
e
r
co
m
m
en
tar
y
.
T
h
e
s
o
lu
tio
n
was
ab
le
to
d
etec
t
m
o
d
er
n
s
tan
d
ar
d
Ar
ab
ic
an
d
c
o
llo
q
u
ial
E
g
y
p
tia
n
Ar
ab
ic.
I
n
[
1
9
]
,
th
e
au
th
o
r
s
p
r
o
p
o
s
ed
a
s
y
s
tem
th
at
r
elies
o
n
m
u
lti
-
lev
el
class
if
icatio
n
to
d
etec
t
f
lam
e
in
an
au
t
o
m
atic
m
an
n
er
.
T
h
is
r
esea
r
ch
a
p
p
lie
d
m
ac
h
i
n
e
lear
n
in
g
tech
n
iq
u
es
f
o
r
a
u
to
m
atic
o
f
f
e
n
s
iv
e
lan
g
u
ag
e
d
etec
tio
n
.
th
e
au
th
o
r
s
u
s
ed
s
u
p
er
v
is
ed
lear
n
i
n
g
m
eth
o
d
s
,
n
am
ely
th
e
Naiv
e
B
ay
es
an
d
th
e
Su
p
p
o
r
t
Vec
to
r
Ma
ch
i
n
e
(
SVM)
t
o
ass
ig
n
co
m
m
en
ts
to
o
n
th
e
“sex
u
al”
o
r
“r
ac
is
t”
ca
teg
o
r
y
.
As
o
n
e
ca
n
n
o
tice
th
e
s
tate
-
of
-
th
e
-
ar
t
i
n
s
u
lt
d
ete
ctio
n
ap
p
r
o
ac
h
es
ab
o
v
e
ty
p
ic
ally
u
s
e
s
u
p
er
v
is
ed
lear
n
in
g
alg
o
r
ith
m
s
to
au
to
m
atica
lly
m
ap
th
e
s
o
cial
m
ed
ia
co
m
m
en
ts
to
th
e
p
r
ed
ef
i
n
ed
class
es.
Sin
ce
s
u
ch
v
er
b
al
o
f
f
e
n
s
e
d
etec
tio
n
s
o
lu
tio
n
s
ar
e
r
elativ
ely
r
ar
e,
we
a
d
d
itio
n
ally
co
v
er
r
elev
a
n
t
tex
t
class
if
icatio
n
ap
p
r
o
ac
h
es.
2
.
2
.
T
y
pica
l t
ex
t
cla
s
s
if
ica
t
io
n
T
y
p
ical
tex
t
class
if
icatio
n
s
y
s
tem
s
r
ely
o
n
tex
t
r
ep
r
esen
tatio
n
an
d
f
ea
tu
r
e
s
elec
tio
n
f
o
r
a
b
etter
d
is
cr
im
in
atio
n
b
etwe
en
th
e
p
r
ed
ef
in
ed
tex
t
ca
teg
o
r
ies.
B
es
id
es,
th
e
f
ea
tu
r
e
s
elec
tio
n
/r
ed
u
ctio
n
ca
n
also
b
e
co
n
d
u
cte
d
to
r
e
d
u
ce
th
e
f
ea
tu
r
e
s
p
ac
e
d
im
en
s
io
n
ality
.
I
n
p
a
r
ticu
lar
,
th
e
L
aten
t
Dir
ich
let
Allo
ca
tio
n
[
20
]
h
as
b
ee
n
ex
p
l
o
ited
to
d
eter
m
i
n
e
th
e
co
r
p
u
s
to
p
ics,
an
d
d
ef
in
e
th
e
f
ea
tu
r
e
s
p
ac
e
ac
co
r
d
in
g
ly
.
Ho
wev
er
,
th
is
ap
p
r
o
ac
h
is
co
n
s
tr
ain
ed
b
y
th
e
lar
g
e
s
iz
e
o
f
th
e
r
esu
ltin
g
v
o
ca
b
u
lar
y
co
m
p
ar
ed
to
t
h
e
s
tan
d
ar
d
B
ag
Of
W
o
r
d
s
(
B
OW
)
r
ep
r
esen
tatio
n
.
I
n
f
ac
t,
d
esp
ite
th
e
p
r
o
m
is
in
g
p
er
f
o
r
m
an
ce
ac
h
iev
ed
in
tex
t
m
in
in
g
ap
p
licatio
n
s
u
s
in
g
wo
r
d
em
b
ed
d
in
g
,
th
e
tr
a
d
itio
n
B
ag
o
f
W
o
r
d
s
(
B
o
W
)
m
o
d
el
is
s
till
ad
o
p
ted
in
v
ar
io
u
s
ap
p
licatio
n
s
an
d
p
r
o
v
e
d
to
p
er
f
o
r
m
r
elativ
ely
well.
T
h
e
B
o
W
m
o
d
el
en
co
d
es
o
n
ly
th
e
k
ey
wo
r
d
s
o
cc
u
r
r
en
ce
f
r
eq
u
en
cy
in
a
g
iv
en
s
et
o
f
d
o
cu
m
e
n
ts
.
I
n
p
ar
ticu
lar
,
T
F
-
I
DF
r
ep
r
esen
tatio
n
p
r
o
v
ed
t
o
b
e
s
u
cc
ess
f
u
l
in
ca
p
tu
r
i
n
g
t
h
e
p
atter
n
s
a
m
o
n
g
th
e
tex
t
s
em
an
tic
ca
teg
o
r
ies.
N
o
te
th
at
n
o
in
f
o
r
m
atio
n
o
n
th
e
s
tr
u
ctu
r
e
o
f
wo
r
d
s
in
a
g
iv
en
d
o
cu
m
en
t
is
en
clo
s
ed
in
s
u
ch
r
ep
r
esen
tatio
n
[
21
]
.
I
n
o
th
er
wo
r
d
s
,
s
p
ar
s
e
r
e
p
r
esen
ta
tio
n
s
r
em
ain
s
ch
allen
g
in
g
f
r
o
m
th
e
c
o
m
p
u
tatio
n
al
an
d
lear
n
in
g
p
o
in
t
o
f
v
iews.
A
s
im
p
le
alter
n
ativ
e
to
lim
i
t
t
h
e
ef
f
ec
t
o
f
th
e
d
ata
s
p
ar
s
ity
co
n
s
is
ts
in
d
i
s
ca
r
d
in
g
th
e
k
ey
wo
r
d
s
with
s
p
ar
s
ity
h
ig
h
er
th
a
n
9
9
%
wh
ich
r
ed
u
ce
s
s
im
u
ltan
eo
u
s
ly
th
e
d
ata
d
i
m
en
s
io
n
ality
.
Oth
er
r
esear
ch
er
s
u
s
ed
g
r
ap
h
r
e
p
r
ese
n
tatio
n
f
o
r
t
ex
t
d
ata
a
n
d
co
u
p
led
it
with
a
p
p
r
o
p
r
iate
d
is
tan
ce
/s
im
ilar
ity
m
ea
s
u
r
es
[
22
]
in
o
r
d
er
to
u
s
e
g
r
a
p
h
m
in
i
n
g
alg
o
r
ith
m
s
.
Sp
ec
if
ically
,
th
e
latter
alg
o
r
ith
m
s
wer
e
in
te
n
d
ed
to
m
in
e
f
r
eq
u
e
n
t
s
u
b
-
g
r
ap
h
s
in
th
e
d
o
cu
m
en
t
co
llectio
n
to
co
n
s
tr
u
ct
th
e
f
ea
tu
r
e
s
p
ac
e
[
21
]
.
H
o
wev
er
,
s
u
ch
r
e
p
r
esen
tatio
n
u
s
u
ally
ex
h
ib
its
h
ig
h
c
o
m
p
u
tatio
n
al
an
d
s
p
ac
e
co
s
ts
.
On
th
e
o
th
e
r
h
an
d
,
h
ier
a
r
ch
ical
class
if
icatio
n
h
as
b
ee
n
also
ad
o
p
ted
f
o
r
tex
t c
lass
if
icatio
n
[
20
]
,
[
23
]
.
I
n
[
7
]
,
a
r
e
v
iew
o
n
th
e
u
s
e
o
f
s
u
p
er
v
is
ed
lear
n
in
g
f
o
r
o
p
in
i
o
n
m
in
i
ng
d
u
r
in
g
th
e
last
d
ec
a
d
e
was
d
o
n
e.
T
h
e
r
esear
ch
er
s
in
[
24
]
in
tr
o
d
u
ce
d
a
n
em
o
tio
n
d
etec
t
io
n
s
y
s
tem
t
h
at
is
in
ten
d
ed
to
r
ec
o
g
n
ize
n
asti
n
ess
an
d
s
ar
ca
s
m
in
o
n
lin
e
c
o
n
v
e
r
s
atio
n
.
B
esid
es,
th
e
au
th
o
r
s
i
n
v
esti
g
ated
th
e
u
s
e
o
f
d
if
f
e
r
en
t
f
ea
tu
r
e
s
ets
alo
n
g
with
tw
o
s
u
p
er
v
is
ed
lear
n
in
g
alg
o
r
ith
m
s
to
im
p
r
o
v
e
th
e
o
v
er
all
class
if
icatio
n
p
er
f
o
r
m
an
ce
.
T
h
e
wo
r
k
i
n
[
25
]
in
tr
o
d
u
ce
d
th
e
k
ey
s
to
n
es
o
f
an
ir
o
n
y
d
etec
tio
n
ap
p
r
o
a
ch
wh
ich
ta
k
es
in
to
co
n
s
id
er
atio
n
th
e
c
u
s
to
m
er
f
ee
d
b
ac
k
in
t
h
e
lear
n
in
g
o
f
th
e
class
if
icatio
n
m
o
d
el.
I
n
[
9
-
11
]
,
t
h
e
au
th
o
r
s
o
u
tlin
ed
th
e
s
tate
-
of
-
th
e
-
a
r
t
s
o
lu
tio
n
s
p
r
o
p
o
s
ed
to
r
ec
o
g
n
ize
r
eg
u
la
r
em
ails
an
d
d
etec
t
ju
n
k
o
n
e
s
[
8
]
.
Desp
ite
s
u
ch
co
n
s
id
er
ab
le
ef
f
o
r
ts
to
o
v
er
co
m
e
r
ea
l
ap
p
licatio
n
s
c
h
allen
g
es,
it
ca
n
b
e
ad
m
itted
th
at
th
er
e
is
n
o
u
n
iv
er
s
al
s
o
lu
tio
n
f
o
r
all
class
if
icatio
n
ch
allen
g
es.
I
n
o
th
e
r
wo
r
d
s
,
it
m
ak
es
n
o
s
en
s
to
claim
th
at
a
cla
s
s
if
icatio
n
tech
n
iq
u
e
o
v
er
tak
es
th
e
o
t
h
er
s
in
all
a
p
p
licatio
n
s
[
4
]
.
T
h
er
ef
o
r
e,
d
ee
p
lear
n
i
n
g
b
ased
class
if
icatio
n
em
er
g
ed
as
a
p
r
o
m
is
in
g
alter
n
ativ
e
to
ad
d
r
e
s
s
th
e
tex
t
class
if
icat
io
n
p
r
o
b
lem
s
.
2
.
3
.
T
ex
t
cla
s
s
if
ica
t
io
n
ba
s
ed
o
n CN
N
a
nd
L
ST
M
Giv
en
th
eir
ab
ilit
y
to
lear
n
th
e
s
tatis
t
ical
p
r
o
p
er
ties
o
f
th
e
i
m
ag
es,
C
NN
h
av
e
b
ee
n
wid
ely
u
s
ed
in
im
ag
e
ca
teg
o
r
izatio
n
a
p
p
licatio
n
s
[
26
]
.
Sp
ec
if
ically
,
C
NNs’
co
n
v
o
lu
tio
n
o
p
e
r
ato
r
ca
p
tu
r
e
s
th
e
lo
wly
v
a
r
ian
t
d
ep
en
d
e
n
cy
b
etwe
en
n
eig
h
b
o
r
in
g
p
i
x
els
in
t
h
e
im
ag
e
r
e
g
io
n
s
.
Su
ch
s
tatis
tical
im
ag
e
ch
ar
ac
t
er
is
tics
ca
n
b
e
also
Evaluation Warning : The document was created with Spire.PDF for Python.
C
o
m
p
u
t.
Sci.
I
n
f
.
T
ec
h
n
o
l.
I
n
s
u
lt d
etec
tio
n
u
s
in
g
a
p
a
r
titi
o
n
a
l CN
N
-
LS
TM
mo
d
el
… (
M
o
h
a
med
Ma
h
er B
en
I
s
ma
il
)
87
f
o
u
n
d
in
tex
tu
al
co
m
m
en
ts
s
in
ce
n
eig
h
b
o
r
in
g
wo
r
d
s
in
a
g
iv
en
co
m
m
en
t
e
x
h
i
b
it
s
o
m
e
d
ep
en
d
en
cy
.
T
h
er
ef
o
r
e,
th
e
k
ey
wo
r
d
s
in
clu
d
e
d
in
a
co
m
m
en
t
s
h
o
u
l
d
b
e
en
co
d
ed
in
o
r
d
er
t
o
b
e
eq
u
iv
ale
n
t
to
th
e
i
m
ag
e
p
ix
els
an
d
f
e
d
to
th
e
C
NN
[
27
]
.
T
y
p
ical
tex
t
r
ep
r
esen
tatio
n
tech
n
iq
u
es
ar
e
u
s
ed
to
in
d
ex
th
e
co
llectio
n
o
f
k
ey
wo
r
d
s
th
at
ar
e
u
s
ed
in
th
e
tex
tu
al
co
m
m
e
n
ts
.
T
h
en
,
th
e
r
esu
ltin
g
m
atr
ix
is
tr
an
s
f
o
r
m
ed
in
t
o
a
lo
wer
d
im
en
s
io
n
al
r
ep
r
esen
tatio
n
af
ter
g
o
in
g
th
r
o
u
g
h
th
e
em
b
e
d
d
in
g
lay
er
[
28
]
.
Su
ch
k
ey
wo
r
d
r
e
p
r
esen
tatio
n
ca
n
b
e
o
b
tai
n
ed
b
y
d
ep
lo
y
in
g
a
d
is
tr
ib
u
tio
n
o
v
e
r
th
e
k
e
y
wo
r
d
wh
ich
r
esu
lts
in
a
f
ix
ed
len
g
t
h
d
en
s
e
v
ec
to
r
.
T
h
is
‘
r
an
d
o
m
ize
d
’
ap
p
r
o
ac
h
is
tu
n
e
d
th
r
o
u
g
h
th
e
C
NN
tr
ain
in
g
p
h
a
s
e.
On
e
s
h
o
u
ld
n
o
te
th
at,
d
e
n
s
e
k
ey
wo
r
d
s
v
ec
to
r
s
o
f
f
ix
ed
le
n
g
th
o
b
tain
ed
u
s
in
g
k
ey
wo
r
d
em
b
e
d
d
in
g
m
eth
o
d
s
lik
e
Glo
Ve
[
2
2
]
an
d
wo
r
d
2
v
ec
[
29
]
ca
n
also
b
e
a
d
o
p
ted
.
T
y
p
ically
,
k
ey
wo
r
d
em
b
ed
d
in
g
r
eq
u
ir
es
a
tr
ain
in
g
p
h
ase
u
s
in
g
lar
g
e
c
o
llectio
n
s
.
Fo
r
in
s
tan
ce
,
th
e
tr
ain
in
g
o
f
th
e
wo
r
d
2
v
ec
m
o
d
e
l
r
elies
o
n
a
co
llec
tio
n
o
f
1
0
0
b
illi
o
n
wo
r
d
s
wh
ich
y
ield
ed
a
3
m
illi
o
n
k
e
y
wo
r
d
v
o
ca
b
u
lar
y
.
Var
io
u
s
s
em
an
tic
co
m
p
o
s
itio
n
ap
p
r
o
ac
h
es
h
av
e
b
ee
n
in
tr
o
d
u
ce
d
to
b
ette
r
r
ep
r
esen
t
th
e
d
o
cu
m
en
ts
/co
m
m
en
ts
in
tex
t
class
if
icatio
n
ap
p
licatio
n
s
.
I
n
p
ar
ticu
lar
,
d
ee
p
lear
n
in
g
p
ar
a
d
ig
m
s
,
s
u
ch
as
R
NN,
C
NN
an
d
L
STM
,
h
av
e
b
ee
n
ad
o
p
ted
t
o
d
esig
n
r
o
b
u
s
t
n
eu
r
al
n
etwo
r
k
s
.
I
n
[
3
0
]
,
a
ty
p
ical
C
NN
n
etw
o
r
k
wh
ich
c
o
m
p
r
is
es
o
n
e
co
n
v
o
lu
tio
n
lay
er
in
clu
d
in
g
f
ilter
s
o
f
v
a
r
io
u
s
wid
th
.
I
n
ad
d
itio
n
,
a
m
ax
p
o
o
lin
g
an
d
f
u
lly
co
n
n
ec
ted
lay
er
s
ar
e
ass
o
ciate
d
f
o
r
s
en
tim
en
t
class
if
icatio
n
.
Oth
er
r
esear
ch
er
s
ass
o
ciate
d
t
h
e
au
to
en
co
d
er
with
R
NN
to
lear
n
a
m
ea
n
in
g
f
u
l
r
ep
r
esen
tatio
n
in
th
e
c
o
n
tex
t
o
f
s
tatis
tical
m
ac
h
in
e
tr
an
s
latio
n
[
3
1
]
.
T
h
e
au
th
o
r
s
in
[
3
2
]
u
s
e
d
m
atr
ices
to
h
a
n
d
le
th
e
n
o
d
es
o
f
t
h
e
tr
ee
s
tr
u
ctu
r
e
o
f
th
eir
R
N
N.
T
h
is
y
ield
ed
b
etter
r
ep
r
esen
tatio
n
o
f
th
e
s
en
tim
en
t
ex
p
r
ess
ed
in
th
e
co
n
s
id
er
ed
s
en
ten
ce
s
.
L
ately
,
as
o
u
tlin
ed
in
[
3
3
]
,
ce
ll
b
lo
ck
s
o
f
L
STM
m
o
d
el
wer
e
in
teg
r
ated
in
R
NN
n
etwo
r
k
t
o
r
e
p
r
esen
t
th
e
n
o
n
leaf
n
o
d
es
o
f
th
e
n
etwo
r
k
tr
ee
s
tr
u
ctu
r
e.
T
h
e
r
esu
ltin
g
m
o
d
el
was
in
ten
d
e
d
to
b
etter
c
ap
tu
r
e
th
e
s
em
an
tic
m
ea
n
in
g
o
f
th
e
tex
t
s
en
te
n
ce
s
.
I
n
[
34
]
,
th
e
au
th
o
r
s
p
r
o
p
o
s
e
d
a
B
o
W
b
ased
C
NN
th
at
r
elies
o
n
a
a
co
n
v
o
lu
tio
n
lay
er
a
n
d
f
ee
d
it
th
e
b
ag
-
of
-
wo
r
d
f
ea
tu
r
e
s
.
I
n
ad
d
itio
n
,
t
h
ey
in
tr
o
d
u
ce
d
a
Seq
u
e
n
tial
C
NN
th
at
is
in
ten
d
ed
to
en
co
d
e
th
e
k
ey
wo
r
d
s
s
eq
u
en
tial
i
n
f
o
r
m
atio
n
th
r
o
u
g
h
th
e
co
n
ca
ten
atio
n
o
f
a
s
in
g
le
v
ec
to
r
o
f
m
u
ltip
le
k
ey
wo
r
d
s
.
T
h
e
r
esear
ch
er
s
in
[
35
]
o
u
tlin
ed
a
d
o
cu
m
en
t
r
ep
r
esen
tatio
n
ap
p
r
o
ac
h
b
ase
d
o
n
n
eu
r
al
n
etwo
r
k
s
th
at
ca
n
lear
n
th
e
r
elatio
n
s
h
ip
s
b
etwe
en
s
en
ten
ce
s
.
Sp
ec
if
ically
,
th
eir
ap
p
r
o
ac
h
co
u
p
les
C
NN
an
d
L
STM
with
w
o
r
d
em
b
ed
d
in
g
to
r
ep
r
esen
t
th
e
s
en
ten
ce
s
.
B
esid
es,
th
ey
ad
o
p
ted
th
e
Gate
d
R
ec
u
r
r
en
t
Un
it
(
GR
U)
,
wh
ich
is
an
ex
ten
s
io
n
o
f
L
STM
,
to
ca
p
tu
r
e
th
e
s
en
ten
ce
’
s
s
em
an
tics
f
o
r
a
m
o
r
e
ac
c
u
r
ate
d
o
c
u
m
en
t
ca
te
g
o
r
izatio
n
.
An
o
th
er
d
ee
p
m
em
o
r
y
n
etwo
r
k
was
u
s
ed
in
[
36
]
to
m
o
d
el
th
e
u
s
er
m
eta
-
d
ata.
Sp
ec
i
f
ically
,
a
L
STM
was
u
s
ed
f
o
r
th
e
d
o
c
u
m
en
t
r
ep
r
esen
tatio
n
,
wh
ile
th
e
d
ee
p
m
em
o
r
y
n
etwo
r
k
was
d
ep
lo
y
ed
to
au
t
o
m
atica
lly
r
ate
n
ew
d
o
cu
m
e
n
ts
.
I
n
[
37
]
,
th
e
a
u
t
h
o
r
s
in
tr
o
d
u
ce
d
an
atten
tio
n
-
b
ased
L
STM
n
etwo
r
k
f
o
r
a
d
o
cu
m
en
t
lev
el
b
ased
s
en
tim
en
t
p
r
ed
ictio
n
.
No
te
th
a
t
r
esu
ltin
g
s
o
lu
tio
n
s
u
p
p
o
r
ts
th
e
E
n
g
lis
h
an
d
C
h
in
ese
lan
g
u
ag
es.
I
n
[
38
]
,
t
h
e
r
es
ea
r
ch
er
s
d
ep
icted
v
ar
io
u
s
v
ar
i
atio
n
s
o
f
th
e
C
NN
b
ased
s
en
tim
en
t
class
if
icat
io
n
ap
p
r
o
ac
h
.
Par
ticu
lar
ly
,
th
ey
in
v
esti
g
ated
th
e
C
NN
-
s
ta
tic
wh
e
r
e
th
ey
p
r
etr
ain
an
d
f
ix
th
e
wo
r
d
e
m
b
ed
d
in
g
a
p
r
io
r
i,
th
e
C
NN
-
r
an
d
wh
er
e
t
h
ey
r
an
d
o
m
l
y
in
itial
ize
th
e
wo
r
d
em
b
ed
d
in
g
,
a
n
d
th
e
C
NN
-
m
u
ltich
an
n
el
wh
er
e
th
e
y
u
s
ed
s
ev
er
al
wo
r
d
em
b
e
d
d
i
n
g
s
ets.
T
h
e
a
u
th
o
r
s
i
n
[
39
]
d
esig
n
ed
a
r
eg
io
n
al
C
NN
-
L
STM
ar
ch
itectu
r
e
th
at
is
in
ten
d
e
d
to
m
ap
th
e
lear
n
ed
tex
t
f
ea
tu
r
es
in
t
o
a
s
et
o
f
p
r
ed
ef
in
e
d
r
ati
n
g
s
ca
teg
o
r
ies.
Similar
ly
in
[
40
]
,
a
C
NN
an
d
L
STM
b
ased
d
ee
p
n
eu
r
al
n
etwo
r
k
was
co
n
s
tr
u
cted
an
d
ass
o
ciate
d
with
lin
g
u
is
tic
em
b
ed
d
in
g
an
d
wo
r
d
2
v
ec
to
class
if
y
s
en
ten
ce
s
as
“f
ee
lin
g
”
o
r
“f
ac
tu
al”.
I
n
[
41
]
,
th
e
r
esea
r
ch
er
s
o
u
tlin
ed
a
n
eu
r
al
n
etwo
r
k
ar
ch
itectu
r
e
b
ased
o
n
two
C
N
Ns
wh
er
e
two
h
id
d
en
lay
er
s
u
s
ed
f
o
r
th
e
f
ea
tu
r
e
r
ep
r
esen
tatio
n
an
d
f
ed
with
b
o
th
th
e
an
n
o
tated
an
d
u
n
an
n
o
tated
in
s
tan
ce
s
.
T
h
e
r
esu
ltin
g
m
o
d
el
was
i
n
ten
d
ed
to
g
en
er
alize
t
h
e
s
en
ten
ce
e
m
b
ed
d
in
g
f
o
r
an
ac
c
u
r
ate
s
en
tim
e
n
t
class
if
icatio
n
.
I
n
o
r
d
er
to
r
ec
o
g
n
ize
th
e
s
en
ten
ce
s
en
tim
en
t
ac
cu
r
ately
,
th
e
au
th
o
r
s
in
[
42
]
p
r
esen
ted
a
m
o
d
el
t
h
at
ex
p
lo
its
th
e
lin
g
u
is
tic
r
eso
u
r
ce
s
an
d
tak
es
in
to
co
n
s
id
er
atio
n
in
f
o
r
m
atio
n
s
u
ch
as
t
h
e
n
eg
atio
n
wo
r
d
s
,
s
en
tim
en
t
lex
i
co
n
,
an
d
in
t
en
s
ity
wo
r
d
s
i
n
to
th
e
L
STM
n
etwo
r
k
.
3.
P
ARTI
T
I
O
N
AL
CNN
-
L
S
T
M
M
O
D
E
L
T
h
e
p
r
o
p
o
s
ed
lo
ca
l
C
NN
-
L
STM
ar
ch
itectu
r
e
is
d
e
p
icted
i
n
Fig
u
r
e
1
.
No
te
t
h
at
to
class
if
y
tex
tu
al
co
m
m
en
ts
u
s
in
g
c
o
n
v
o
lu
tio
n
s
,
we
co
n
v
er
ted
th
e
tex
t
in
s
ta
n
ce
s
in
to
im
ag
es.
T
h
er
ef
o
r
e,
th
e
wo
r
d
2
v
ec
th
at
co
n
s
is
ts
in
a
two
-
lay
er
n
eu
r
al
n
et
was
f
ir
s
t
u
s
ed
to
p
r
o
ce
s
s
th
e
co
m
m
en
t
co
llectio
n
.
Mo
r
e
s
p
ec
if
ically
,
th
e
co
m
m
en
ts
wer
e
co
n
v
er
ted
in
to
s
eq
u
en
ce
s
o
f
k
ey
wo
r
d
s
v
ec
to
r
s
o
f
len
g
th
d
u
s
in
g
wo
r
d
em
b
ed
d
in
g
[
4
1
]
.
T
h
e
r
esu
ltin
g
n
u
m
er
ical
v
ec
to
r
s
ar
e
th
en
f
ed
in
to
th
e
th
e
d
ee
p
n
eu
r
al
n
etwo
r
k
.
I
n
p
ar
ticu
lar
,
th
e
p
r
o
p
o
s
ed
lo
ca
l
C
NN
m
o
d
el
s
p
lit
s
a
co
m
m
en
t
in
to
M
p
ar
titi
o
n
s
{
1
,
2
,
…
,
}
.
R
elev
an
t
f
ea
tu
r
es
ar
e
th
en
ex
tr
ac
ted
f
r
o
m
t
h
ese
p
ar
titi
o
n
s
.
Sp
ec
if
ically
,
th
e
c
o
n
v
o
lu
ti
o
n
al
an
d
m
a
x
p
o
o
lin
g
lay
er
s
p
r
o
ce
s
s
s
eq
u
en
tially
th
e
in
p
u
t
v
ec
to
r
s
in
o
r
d
er
t
o
lear
n
t
h
e
r
elev
a
n
t
f
ea
tu
r
es.
Fin
ally
,
th
e
L
STM
is
u
s
ed
to
in
co
r
p
o
r
ate
s
eq
u
en
tially
th
e
o
b
tain
e
d
lo
ca
l
f
ea
tu
r
es
ac
r
o
s
s
th
e
p
a
r
titi
o
n
s
t
o
f
o
r
m
th
e
o
v
er
all
co
m
m
e
n
t
v
ec
to
r
to
b
e
au
to
m
atica
lly
ca
te
g
o
r
ized
as
in
s
u
lt
o
r
n
o
t.
T
h
e
co
n
v
o
lu
tio
n
al
lay
er
is
in
itially
in
ten
d
ed
to
f
o
r
m
th
e
lo
ca
l
n
g
r
am
f
ea
tu
r
es
f
o
r
ea
c
h
p
ar
titi
o
n
.
L
et
th
e
p
ar
titi
o
n
m
atr
ix
b
e
∈
×
wh
er
e
M
is
th
e
s
eq
u
en
ce
v
o
ca
b
u
la
r
y
s
ize,
an
d
d
is
th
e
k
e
y
wo
r
d
v
ec
to
r
s
d
im
en
s
io
n
ality
.
As
illu
s
tr
ated
in
Fig
u
r
e
1
,
th
e
k
e
y
wo
r
d
v
ec
to
r
s
in
th
e
p
ar
titi
o
n
s
=
{
1
,
2
,
…
,
}
,
=
{
1
,
2
,
…
,
}
an
d
=
{
1
,
2
,
…
,
}
ar
e
ag
g
r
e
g
ate
d
t
o
g
et
th
e
p
ar
titi
o
n
m
atr
ices
,
an
d
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
7
2
2
-
3
2
2
1
C
o
m
p
u
t.
Sci.
I
n
f
.
T
ec
h
n
o
l.
,
V
o
l.
1
,
No
.
2
,
J
u
ly
20
2
0
:
84
–
92
88
As
o
n
e
ca
n
n
o
tice,
C
co
n
v
o
lu
t
io
n
al
f
ilter
s
ar
e
u
s
ed
f
o
r
ea
ch
p
ar
titi
o
n
to
ex
tr
ac
t
th
e
l
o
ca
l
n
-
g
r
am
f
ea
tu
r
es.
I
n
a
s
eq
u
en
ce
o
f
K
k
ey
wo
r
d
s
:
+
−
1
,
th
e
d
ep
lo
y
m
e
n
t o
f
a
f
ilter
H
t,
1
≤
t
≤
T
r
esu
lt
s
in
a
f
ea
tu
r
e
m
a
p
:
=
f
(
:
+
−
1
+
)
(
1
)
W
h
er
e
th
e
o
p
er
ato
r
o
r
ep
r
esen
ts
a
co
n
v
o
l
u
tio
n
,
b
a
n
d
∈
×
ar
e
th
e
b
ias
an
d
th
e
weig
h
t
m
atr
ices
r
esp
ec
tiv
ely
.
On
t
h
e
o
t
h
er
h
an
d
,
is
th
e
d
im
en
s
io
n
o
f
th
e
k
ey
wo
r
d
v
ec
t
o
r
,
ω
is
th
e
f
ilter
len
g
th
an
d
f
d
en
o
tes
th
e
R
eL
U
f
u
n
ctio
n
.
T
h
e
f
ea
tu
r
e
m
ap
s
=
1
,
2
,
…
,
−
+
1
o
f
th
e
f
ilter
H
t
ar
e
o
b
tain
ed
af
ter
a
f
ilter
s
ca
n
s
p
r
o
g
r
ess
iv
ely
f
r
o
m
1
:
−
1
to
+
−
1
:
.
No
te
th
at
th
e
co
m
m
e
n
t
p
ar
titi
o
n
s
e
x
h
ib
it
v
ar
iab
le
tex
t
len
g
th
s
w
h
ich
y
ield
s
v
ar
iab
le
d
im
en
s
io
n
s
f
o
r
.
Nex
t
to
th
e
in
p
u
t
lay
er
o
f
len
g
th
N
,
th
e
o
u
tp
u
t
o
f
th
e
co
n
v
o
lu
tio
n
al
lay
e
r
is
s
u
b
s
am
p
led
in
th
e
Ma
x
-
p
o
o
li
n
g
lay
er
.
I
n
p
ar
ticu
lar
,
p
o
o
lin
g
is
p
er
f
o
r
m
ed
t
h
r
o
u
g
h
t
h
e
a
p
p
li
ca
tio
n
o
f
a
m
a
x
f
u
n
ctio
n
to
th
e
o
u
tp
u
t
o
f
ea
ch
f
ilter
.
T
h
is
o
p
er
atio
n
is
in
ten
d
ed
to
r
ed
u
ce
th
e
co
m
p
u
tatio
n
al
co
s
t
o
f
th
e
u
p
p
er
lay
er
s
an
d
d
is
ca
r
d
th
e
n
o
n
-
m
a
x
im
al
v
alu
es.
I
n
ad
d
itio
n
,
it
p
r
o
ce
s
s
es
th
e
d
if
f
er
en
t
p
ar
titi
o
n
s
an
d
ca
p
tu
r
es
th
e
lo
ca
l
d
ep
en
d
e
n
cy
to
d
ete
r
m
in
e
th
e
m
o
s
t
s
alien
t
in
f
o
r
m
atio
n
.
T
h
e
r
esu
ltin
g
p
ar
titi
o
n
v
ec
to
r
s
ar
e
th
en
p
r
o
v
id
ed
to
a
s
eq
u
e
n
tial
lay
er
.
Fo
r
th
is
s
eq
u
en
tial
lay
er
,
th
e
in
ter
-
p
ar
t
itio
n
lo
n
g
-
d
is
tan
ce
d
e
p
en
d
e
n
c
y
is
ca
p
tu
r
e
d
b
y
a
s
eq
u
en
tial
in
teg
r
atio
n
o
f
th
e
p
ar
titi
o
n
v
ec
to
r
s
i
n
t
o
th
e
co
m
m
en
ts
v
ec
to
r
s
.
No
te
t
h
at
th
e
L
STM
is
in
tr
o
d
u
c
ed
in
th
is
lay
er
in
o
r
d
er
to
a
d
d
r
ess
th
e
ty
p
ical
R
NN
g
r
ad
ien
t
v
an
is
h
in
g
o
r
e
x
p
lo
d
i
n
g
p
r
o
b
lem
.
O
n
ce
all
p
ar
titi
o
n
s
ar
e
s
eq
u
en
tially
tr
av
er
s
ed
b
y
th
e
L
STM
m
em
o
r
y
ce
ll,
t
h
e
last
s
eq
u
en
tial
lay
er
h
i
d
d
en
s
ta
te
ca
n
b
e
p
er
ce
iv
e
d
as
th
e
co
m
m
en
t
r
ep
r
esen
tatio
n
f
o
r
in
s
u
lt
d
etec
tio
n
.
Fin
ally
,
a
ty
p
ic
al
So
f
tm
ax
class
if
ier
is
ad
o
p
ted
f
o
r
th
e
last
lay
er
.
T
h
e
m
in
im
izatio
n
o
f
th
e
m
ea
n
s
q
u
ar
ed
er
r
o
r
b
etwe
en
th
e
g
r
o
u
n
d
tr
u
th
class
v
alu
es
an
d
th
e
p
r
ed
icted
is
u
s
ed
to
tr
ain
th
e
lo
ca
l
C
NN
-
L
STM
.
L
et
=
1
,
2
,
…
,
b
e
a
tr
ain
in
g
s
et
o
f
tex
t
m
a
tr
ix
,
an
d
=
1
,
2
,
…
,
b
e
th
e
co
r
r
esp
o
n
d
in
g
class
v
alu
es.
On
th
e
o
th
er
h
a
n
d
,
we
d
ef
in
e
th
e
lo
s
s
f
u
n
ctio
n
as:
(
,
)
=
1
2
∑
‖
ℎ
(
)
−
‖
2
=
1
(
2
)
B
esid
es,
th
e
b
ac
k
p
r
o
p
ag
atio
n
alg
o
r
ith
m
in
[
4
3
]
b
ased
o
n
t
h
e
s
to
ch
asti
c
g
r
ad
ien
t
d
escen
t
(
SGD)
i
s
u
s
ed
in
th
e
tr
ain
i
n
g
p
h
ase
in
o
r
d
er
to
o
p
tim
ize
th
e
n
etwo
r
k
p
ar
am
eter
s
.
Fig
u
r
e
1
.
Ar
c
h
itectu
r
e
o
f
th
e
p
r
o
p
o
s
ed
m
o
d
el
4.
E
XP
E
R
I
M
E
N
TS
W
e
co
n
d
u
cted
a
r
an
g
e
o
f
e
x
p
er
im
en
ts
to
ev
alu
ate
th
e
p
er
f
o
r
m
a
n
ce
o
f
th
e
p
r
o
p
o
s
ed
ap
p
r
o
ac
h
.
Par
ticu
lar
ly
,
we
u
s
ed
KAGG
L
E
d
ataset
[
4
4
]
wh
ich
r
ep
r
esen
ts
a
co
llectio
n
o
f
c
o
m
m
en
ts
f
r
o
m
v
ar
io
u
s
s
o
cial
m
ed
ia.
T
h
e
6
1
8
3
co
m
m
en
ts
wh
ich
co
m
p
o
s
e
th
is
d
ataset
b
elo
n
g
t
o
th
e
“in
s
u
lt”
an
d
“in
s
u
lt
-
f
r
ee
”
ca
te
g
o
r
ies.
First,
th
ese
co
m
m
en
ts
wer
e
p
r
e
-
p
r
o
ce
s
s
ed
in
o
r
d
er
to
d
is
ca
r
d
s
o
m
e
e
n
co
d
in
g
p
a
r
ts
th
at
m
ay
af
f
ec
t
t
h
e
r
esu
lts
.
Evaluation Warning : The document was created with Spire.PDF for Python.
C
o
m
p
u
t.
Sci.
I
n
f
.
T
ec
h
n
o
l.
I
n
s
u
lt d
etec
tio
n
u
s
in
g
a
p
a
r
titi
o
n
a
l CN
N
-
LS
TM
mo
d
el
… (
M
o
h
a
med
Ma
h
er B
en
I
s
ma
il
)
89
Sp
ec
if
ically
,
th
e
c
o
m
m
en
ts
we
r
e
to
k
e
n
ized
a
n
d
c
o
n
v
e
r
ted
to
lo
wer
ca
s
e.
I
n
a
d
d
itio
n
,
all
p
u
n
ctu
atio
n
c
h
a
r
ac
ter
s
wer
e
er
ased
.
T
h
is
r
esu
lts
in
a
v
o
ca
b
u
lar
y
o
f
1
5
3
2
2
k
e
y
wo
r
d
s
.
T
o
im
p
lem
en
t
th
e
p
r
o
p
o
s
ed
a
p
p
r
o
ac
h
,
a
n
etwo
r
k
with
1
-
D
co
n
v
o
lu
tio
n
al
f
ilter
s
o
f
v
ar
y
in
g
wid
th
s
wer
e
tr
ain
ed
.
No
te
th
at
ea
ch
f
ilter
wid
th
co
r
r
esp
o
n
d
s
to
th
e
n
u
m
b
er
o
f
k
ey
wo
r
d
s
th
e
f
ilter
ca
n
p
r
o
ce
s
s
wh
ich
co
r
r
esp
o
n
d
s
to
th
e
n
-
g
r
am
len
g
th
.
I
n
o
u
r
ex
p
er
im
e
n
ts
,
we
u
s
ed
th
e
p
r
e
-
tr
ain
ed
w
o
r
d
e
m
b
ed
d
in
g
m
o
d
el
(
Fas
tTe
x
t)
[
45
]
.
Fas
tTe
x
t
is
an
E
n
g
lis
h
1
6
B
illi
o
n
T
o
k
en
W
o
r
d
E
m
b
ed
d
in
g
s
u
p
p
o
r
t p
ac
k
ag
e.
T
h
is
m
o
d
el
was a
d
o
p
te
d
to
in
i
tialize
th
e
weig
h
ts
o
f
th
e
em
b
e
d
d
in
g
lay
e
r
.
T
h
is
is
in
ten
d
ed
t
o
to
b
u
ild
3
0
0
-
d
im
en
s
io
n
wo
r
d
v
ec
to
r
s
f
o
r
all
c
o
m
m
en
ts
.
T
h
e
h
y
p
er
-
p
ar
am
et
er
s
o
f
th
e
p
r
o
p
o
s
ed
ar
ch
itectu
r
e
wer
e
o
p
tim
ized
b
ased
o
n
th
e
p
er
f
o
r
m
an
ce
o
f
th
e
tr
ain
in
g
an
d
v
alid
atio
n
p
h
ases
u
s
in
g
th
e
s
ea
r
ch
f
u
n
ctio
n
in
t
r
o
d
u
ce
d
in
[
46
]
.
T
h
is
tu
n
in
g
s
tr
ateg
y
aim
s
to
i
n
v
esti
g
ate
all
ca
n
d
id
ate
p
ar
a
m
eter
co
m
b
in
atio
n
s
,
ass
es
s
th
e
co
r
r
esp
o
n
d
in
g
m
o
d
els
an
d
d
eter
m
in
e
th
e
o
p
tim
al
s
ettin
g
s
.
Fo
r
th
e
co
n
s
id
er
ed
d
ataset,
th
e
o
p
tim
al
p
ar
am
eter
s
o
f
t
h
e
p
r
o
p
o
s
ed
n
e
two
r
k
ar
e
s
h
o
w
n
in
th
e
T
a
b
le
b
elo
w:
T
ab
le
1
.
T
h
e
h
y
p
er
-
p
ar
am
eter
s
o
f
th
e
p
r
o
p
o
s
ed
n
etwo
r
k
ar
c
h
itectu
r
e
#
f
i
l
t
e
r
s
F
i
l
t
e
r
P
o
o
l
D
r
o
p
o
u
t
LSTM
l
a
y
e
r
LSTM
h
i
d
d
e
n
Tr
a
i
n
i
n
g
b
a
t
c
h
(
m)
l
e
n
g
t
h
(
l
)
l
e
n
g
t
h
(
n
)
r
a
t
e
(
p
)
c
o
u
n
t
(
c
)
l
a
y
e
r
(
d
)
si
z
e
(
b
)
/
Ep
o
c
h
s(
s)
64
3
2
0
.
1
2
2
0
0
1
0
0
/
1
0
I
n
o
r
d
er
to
ass
ess
th
e
p
er
f
o
r
m
an
ce
o
f
th
e
p
r
o
p
o
s
ed
ap
p
r
o
ac
h
,
we
u
s
ed
th
e
f
o
llo
win
g
s
tan
d
ar
d
p
er
f
o
r
m
an
ce
m
ea
s
u
r
es in
all
o
u
r
ex
p
e
r
im
en
ts
.
Nam
ely
,
th
e
a
cc
u
r
ac
y
was o
b
tain
e
d
u
s
in
g
:
A
cc
u
r
a
cy
=(
#
C
o
r
r
ec
tPred
icti
o
n
s
)
/(
To
ta
lN
u
mb
erOfPred
ictio
n
s
)
(
3
)
As
o
n
e
ca
n
s
ee
in
Fig
u
r
e
2
,
th
e
v
alid
atio
n
ac
cu
r
ac
y
attain
e
d
b
y
th
e
p
r
o
p
o
s
ed
ap
p
r
o
ac
h
is
8
0
.
8
9
%
with
a
lear
n
in
g
r
ate
o
f
0
.
0
1
.
On
th
e
o
th
er
h
a
n
d
,
th
e
tr
ain
in
g
ac
cu
r
a
cy
r
ea
ch
es 1
0
0
%.
Fig
u
r
e
2
.
T
r
ain
in
g
p
r
o
g
r
ess
: (
a)
ac
cu
r
ac
y
v
s
iter
atio
n
.
(
b
)
lo
s
s
v
s
iter
atio
n
Similar
ly
,
th
e
R
ec
all
an
d
Pre
cisi
o
n
m
etr
ics we
r
e
ca
lcu
lated
u
s
in
g
:
R
ec
a
ll=(
#
C
o
r
r
ec
tlyDete
cted
(
I
n
s
u
lt )
)
/(
To
t
a
lN
u
mb
erOfIn
s
u
lt )
(
4
)
P
r
ec
is
io
n
=(
#
C
o
r
r
ec
tlyDe
tect
ed
(
I
n
s
u
lt
-
fr
ee
)
)
/(
To
ta
lN
u
mb
erOfIn
s
u
lt
-
fr
ee
)
(
5
)
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
7
2
2
-
3
2
2
1
C
o
m
p
u
t.
Sci.
I
n
f
.
T
ec
h
n
o
l.
,
V
o
l.
1
,
No
.
2
,
J
u
ly
20
2
0
:
84
–
92
90
I
n
ad
d
itio
n
,
th
e
F
-
m
ea
s
u
r
e
(
F1
s
co
r
e)
was c
o
n
s
id
er
ed
a
n
d
c
o
m
p
u
ted
u
s
in
g
:
F
1
=2
× (
P
r
ec
is
io
n
× R
ec
a
ll )
/(
P
r
ec
is
io
n
+R
ec
a
ll )
(
6
)
T
ab
le
I
I
r
ep
o
r
ts
th
e
p
er
f
o
r
m
an
ce
m
ea
s
u
r
e
attain
m
en
t
ac
h
iev
e
d
u
s
in
g
t
h
e
p
r
o
p
o
s
ed
ap
p
r
o
ac
h
as
well
as
r
elev
an
t
s
tate
o
f
th
e
ar
t
m
eth
o
d
s
.
As
it
ca
n
b
e
s
ee
n
,
th
e
p
r
o
p
o
s
ed
m
eth
o
d
o
v
e
r
tak
es
th
e
o
th
er
ap
p
r
o
ac
h
es
in
ter
m
s
o
f
Sp
ec
if
icity
,
Acc
u
r
ac
y
an
d
Pre
cisi
o
n
.
I
n
p
ar
ticu
la
r
,
th
e
p
r
o
p
o
s
ed
m
eth
o
d
b
ased
o
n
C
NN
an
d
L
STM
d
etec
ted
ab
o
u
t
3
7
%
m
o
r
e
in
s
u
lt
co
m
m
en
ts
th
an
ty
p
ical
C
NN
-
b
ased
cl
ass
if
icatio
n
.
No
te
t
h
at
th
e
C
NN
-
b
ased
r
esu
lts
wer
e
o
b
tain
ed
af
ter
co
n
v
er
tin
g
th
e
co
m
m
e
n
t
co
llectio
n
in
to
im
ag
es.
B
esid
es,
th
e
in
s
tan
ce
s
wer
e
p
ad
d
ed
in
o
r
d
er
to
h
av
e
a
co
n
s
tan
t
len
g
th
.
Fu
r
th
e
r
m
o
r
e
,
th
e
d
o
c
u
m
en
ts
wer
e
co
n
v
er
ted
in
to
s
eq
u
en
ce
s
o
f
k
ey
wo
r
d
v
ec
t
o
r
s
u
s
in
g
th
e
wo
r
2
v
ec
wo
r
d
e
m
b
ed
d
in
g
[
29
].
Par
ticu
lar
ly
,
t
h
e
im
p
lem
en
ted
n
et
wo
r
k
r
elies
o
n
1
-
D
co
n
v
o
l
u
tio
n
al
f
ilter
s
o
f
v
ar
y
in
g
wid
th
s
.
I
n
o
th
e
r
wo
r
d
s
,
th
e
wid
th
o
f
ea
c
h
f
ilter
f
its
th
e
n
-
g
r
am
len
g
th
.
I
n
f
ac
t,
th
e
d
if
f
e
r
en
t
b
r
an
c
h
es
o
f
co
n
v
o
lu
tio
n
al
lay
e
r
s
o
f
th
e
n
etwo
r
k
h
a
n
d
le
th
e
m
u
ltip
le
n
-
g
r
am
len
g
th
s
.
T
h
e
C
NN
n
etwo
r
k
ar
c
h
itectu
r
e
ca
n
b
e
s
u
m
m
ar
ized
as f
o
llo
ws:
−
B
lo
ck
s
o
f
lay
er
s
wh
ich
co
n
s
is
t
o
f
a
co
n
v
o
lu
tio
n
al
lay
er
,
a
b
atch
n
o
r
m
aliza
tio
n
lay
er
,
a
R
eL
U
lay
er
,
a
d
r
o
p
o
u
t la
y
er
,
an
d
a
m
ax
p
o
o
li
n
g
lay
er
we
r
e
d
esig
n
ed
to
h
a
n
d
le
th
e
n
-
g
r
am
len
g
th
s
2
,
3
,
4
,
an
d
5
.
−
2
0
0
co
n
v
o
l
u
tio
n
al
f
ilter
s
alo
n
g
with
p
o
o
li
n
g
r
e
g
io
n
s
wer
e
u
s
ed
f
o
r
ea
ch
b
lo
ck
.
−
T
h
e
in
p
u
t la
y
er
was c
o
n
n
ec
ted
to
ea
ch
b
l
o
ck
.
−
T
h
e
o
u
tp
u
ts
o
f
th
e
b
lo
ck
s
wer
e
ag
g
r
eg
ate
d
u
s
in
g
a
d
ep
th
co
n
ca
ten
atio
n
lay
e
r
.
−
A
f
u
lly
co
n
n
ec
ted
lay
er
,
a
s
o
f
tm
ax
lay
er
,
an
d
a
class
if
icat
io
n
lay
er
wer
e
in
clu
d
ed
f
o
r
th
e
class
if
icatio
n
ta
s
k
.
T
ab
le
1
.
Per
f
o
r
m
an
ce
m
ea
s
u
r
e
s
o
b
tain
ed
u
s
in
g
t
h
e
m
eth
o
d
in
[
4
7
]
,
ty
p
ical
SVM
class
if
icati
o
n
,
a
C
NN
-
b
ased
m
eth
o
d
an
d
th
e
p
r
o
p
o
s
ed
m
eth
o
d
,
r
esp
ec
tiv
ely
M
e
t
h
o
d
A
c
c
u
r
a
c
y
R
e
c
a
l
l
P
r
e
c
i
s
i
o
n
F1
-
mea
s
u
r
e
M
e
t
h
o
d
i
n
[
47
]
0
.
5
9
8
0
.
5
9
7
0
.
6
8
5
0
.
6
3
8
S
V
M
-
b
a
s
e
d
me
t
h
o
d
0
.
6
0
6
0
.
2
2
3
0
.
7
4
1
0
.
3
4
3
C
N
N
-
b
a
s
e
d
me
t
h
o
d
0
.
7
2
8
0
.
6
8
9
0
.
7
4
2
0
.
7
1
5
P
r
o
p
o
se
d
M
e
t
h
o
d
0
.
8
3
4
0
.
9
4
4
0
.
7
9
3
0
.
8
6
2
Fu
r
th
er
m
o
r
e
,
we
co
n
d
u
cted
a
s
tatis
t
ical
Stu
d
en
t
t
-
test
[
48
]
u
s
in
g
a
c
o
n
f
id
e
n
ce
le
v
el
o
f
9
5
%.
T
h
is
test
was
in
ten
d
ed
to
d
ec
id
e
if
th
e
m
ea
n
s
o
f
two
d
ec
is
io
n
s
ets
o
b
tain
ed
u
s
in
g
two
d
if
f
er
en
t
m
o
d
els
ar
e
r
eliab
ly
d
if
f
er
en
t.
T
h
u
s
,
if
th
e
d
if
f
er
en
c
e
b
etwe
en
t
h
e
m
ea
n
o
f
th
e
p
er
f
o
r
m
an
ce
m
ea
s
u
r
es
is
s
tatis
tical
ly
s
ig
n
if
ic
an
t,
th
en
th
e
n
u
ll
h
y
p
o
th
esis
th
at
ass
u
m
es
th
at
th
e
two
s
am
p
les
f
o
llo
w
s
im
ilar
d
is
tr
ib
u
tio
n
s
is
r
ejec
te
d
.
Sp
ec
if
ically
,
f
o
r
th
e
p
-
v
alu
es
[
4
9
]
b
elo
w
0
.
0
5
,
th
e
class
if
icatio
n
r
esu
lts
wer
e
s
tatis
tically
s
ig
n
if
ican
t.
T
h
er
ef
o
r
e,
th
e
n
u
ll
h
y
p
o
th
eses
wer
e
r
ejec
te
d
b
y
t
h
e
t
-
test
as sh
o
wn
in
T
ab
le
3
.
T
ab
le
2
.
T
-
test
r
esu
lts
b
ased
o
n
th
e
p
er
f
o
r
m
a
n
ce
m
ea
s
u
r
es o
f
th
e
d
if
f
e
r
en
t a
p
p
r
o
ac
h
es
Pro
p
o
s
ed
Me
th
o
d
Vs
Me
th
o
d
in
[
5
9
]
Pro
p
o
s
ed
Me
th
o
d
Vs
SVM
-
b
ased
m
eth
o
d
Pro
p
o
s
ed
Me
th
o
d
Vs
C
NN
-
b
ased
m
eth
o
d
Acc
u
r
ac
y
1
1
1
R
ec
all
1
1
1
Pre
cisi
o
n
1
1
1
Fu
r
th
er
in
v
esti
g
atio
n
s
h
o
wed
o
u
r
a
p
p
r
o
ac
h
ca
te
g
o
r
izes
less
ac
cu
r
ately
n
o
n
-
o
f
f
e
n
s
iv
e
co
m
m
e
n
ts
wh
ich
y
ield
s
lo
wer
s
en
s
itiv
ity
.
Des
p
ite
th
is
co
n
tr
ast
b
etwe
en
th
e
s
p
ec
if
icity
an
d
th
e
s
en
s
itiv
ity
attain
m
en
t,
th
ese
r
esu
lts
ca
n
b
e
co
n
s
id
er
ed
p
r
o
m
is
in
g
.
I
n
f
ac
t,
f
o
r
s
u
ch
in
s
u
lt a
u
to
m
atic
d
etec
tio
n
p
r
o
b
lem
,
o
n
e
ca
n
ass
u
m
e
th
at
th
e
T
r
u
e
Po
s
itiv
e
p
r
ed
ictio
n
s
ar
e
n
o
t
as
im
p
o
r
tan
t
a
s
th
e
T
r
u
e
Neg
ativ
e
in
s
ta
n
ce
s
.
Sp
ec
if
ically
,
th
e
m
is
class
if
icat
io
n
o
f
an
in
s
u
ltin
g
co
m
m
e
n
t
is
n
o
t
co
n
s
id
e
r
ed
as
cr
itical
as
th
e
m
is
cla
s
s
if
i
ca
tio
n
o
f
a
r
eg
u
lar
o
n
e.
I
n
a
d
d
itio
n
,
th
e
ac
c
u
r
ac
y
ca
n
n
o
t
b
e
a
r
eliab
le
p
e
r
f
o
r
m
an
ce
m
ea
s
u
r
e
f
o
r
t
h
is
ap
p
licatio
n
b
ec
au
s
e
th
e
test
in
g
d
ata
in
clu
d
es 7
2
0
v
e
r
b
ally
o
f
f
en
s
iv
e
co
m
m
en
ts
o
n
ly
o
u
t o
f
t
h
e
2
6
7
4
co
m
m
en
ts
,
Evaluation Warning : The document was created with Spire.PDF for Python.
C
o
m
p
u
t.
Sci.
I
n
f
.
T
ec
h
n
o
l.
I
n
s
u
lt d
etec
tio
n
u
s
in
g
a
p
a
r
titi
o
n
a
l CN
N
-
LS
TM
mo
d
el
… (
M
o
h
a
med
Ma
h
er B
en
I
s
ma
il
)
91
5.
CO
NCLU
SI
O
N
I
n
th
is
p
ap
er
we
h
av
e
p
r
o
p
o
s
ed
a
n
o
v
el
ap
p
r
o
ac
h
o
f
au
to
m
atic
in
s
u
lt
d
etec
tio
n
in
s
o
ci
al
n
etwo
r
k
co
m
m
en
ts
.
Sp
ec
if
ically
,
we
p
r
o
p
o
s
ed
a
p
ar
titi
o
n
al
C
NN
-
L
STM
m
o
d
el
in
ten
d
e
d
to
au
t
o
m
atica
lly
r
ec
o
g
n
ize
v
er
b
al
o
f
f
en
s
e
in
s
o
cial
n
et
wo
r
k
co
m
m
en
ts
.
I
n
p
ar
ticu
l
ar
,
we
d
esig
n
e
d
a
p
ar
titi
o
n
a
l
C
NN
an
d
L
STM
ar
ch
itectu
r
e
to
m
ap
s
o
cial
n
etwo
r
k
co
m
m
e
n
ts
in
to
“in
s
u
lt”
o
r
“r
eg
u
lar
”
ca
teg
o
r
ies.
I
n
f
ac
t,
in
s
tead
o
f
co
n
s
id
er
in
g
a
wh
o
le
d
o
cu
m
e
n
t
/co
m
m
en
ts
as in
p
u
t a
s
f
o
r
ty
p
i
ca
l CNN,
we
p
ar
titi
o
n
th
e
co
m
m
en
ts
in
to
p
ar
ts
in
o
r
d
er
to
ca
p
tu
r
e
an
d
weig
h
t
th
e
lo
ca
lly
r
elev
an
t
in
f
o
r
m
atio
n
in
ea
ch
p
ar
titi
o
n
.
T
h
e
o
b
tain
e
d
lo
ca
l
in
f
o
r
m
atio
n
is
th
en
s
eq
u
en
tially
ex
p
lo
ited
ac
r
o
s
s
p
ar
titi
o
n
s
u
s
in
g
L
STM
f
o
r
v
er
b
al
o
f
f
e
n
s
e
d
etec
tio
n
.
T
h
e
ass
o
ciatio
n
o
f
s
u
ch
p
ar
titi
o
n
al
C
NN
an
d
L
STM
allo
ws
th
e
in
teg
r
atio
n
o
f
th
e
lo
ca
l
with
in
co
m
m
en
ts
in
f
o
r
m
atio
n
an
d
th
e
lo
n
g
d
is
tan
ce
co
r
r
elatio
n
ac
r
o
s
s
co
m
m
en
ts
.
T
h
e
o
b
tain
e
d
e
x
p
er
i
m
en
tal
r
esu
lts
p
r
o
v
ed
th
at
th
e
p
r
o
p
o
s
ed
a
p
p
r
o
ac
h
o
v
er
tak
es e
x
is
tin
g
r
ele
v
an
t a
p
p
r
o
ac
h
es.
ACK
NO
WL
E
DG
M
E
N
T
T
h
is
p
r
o
ject
was
s
u
p
p
o
r
te
d
b
y
th
e
R
esear
ch
Gr
o
u
p
s
Pro
g
r
am
(
R
esear
ch
Gr
o
u
p
n
u
m
b
er
RG
-
1439
-
0
3
3
)
,
Dea
n
s
h
ip
o
f
Scien
tific
R
esear
ch
,
Kin
g
Sau
d
Un
iv
er
s
ity
,
R
iy
ad
h
,
Sau
d
i A
r
a
b
ia.
RE
F
E
R
E
NC
E
S
[1
]
“
Co
m
m
u
n
ica
ti
o
n
s
a
n
d
In
fo
rm
a
ti
o
n
Tec
h
n
o
l
o
g
y
Co
m
m
issio
n
.”
[On
l
in
e
].
Av
a
i
lab
le:
h
tt
p
:
//
ww
w.citc.g
o
v
.
sa
/ar/P
a
g
e
s/
d
e
fa
u
lt
.
a
sp
x
[Ac
c
e
ss
e
d
1
No
v
e
m
b
e
r
2
0
1
7
]
.
[2
]
Xia
n
g
,
G
u
a
n
g
&
F
a
n
,
Bi
n
&
Wan
g
,
L
in
g
&
Ho
n
g
,
Ja
so
n
&
Ro
se
,
Ca
ro
ly
n
.
“
De
tec
ti
n
g
o
ffe
n
si
v
e
t
we
e
ts
v
ia
to
p
ica
l
fe
a
tu
re
d
isc
o
v
e
ry
o
v
e
r
a
larg
e
sc
a
le
twit
ter
c
o
r
p
u
s
.”
P
ro
c
e
e
d
i
n
g
s
o
f
t
h
e
2
1
st
ACM
Co
n
fer
e
n
c
e
o
n
In
fo
rm
a
t
io
n
a
n
d
Kn
o
wled
g
e
M
a
n
a
g
e
me
n
t
,
S
h
e
ra
t
o
n
,
M
a
u
i
Ha
wa
ii
,
p
p
.
1
9
8
0
-
1
9
8
4
,
2
0
1
2
.
[3
]
L.
K.
Ha
n
se
n
a
n
d
P
.
“
S
a
lam
o
n
.
Ne
u
ra
l
n
e
two
rk
e
n
se
m
b
les
.
”
IEE
E
T
ra
n
sa
c
ti
o
n
s
o
n
P
a
tt
e
rn
An
a
ly
sis
a
n
d
M
a
c
h
i
n
e
In
telli
g
e
n
c
e
,
v
o
l.
1
2
,
n
o
.
10
,
p
p
.
99
3
–
1
0
0
1
,
1
9
9
0
.
[4
]
El
len
S
p
e
rt
u
s.
“
S
m
o
k
e
y
:
Au
t
o
m
a
ti
c
re
c
o
g
n
it
i
o
n
o
f
h
o
stil
e
m
e
ss
a
g
e
s.
”
In
Pro
c
e
e
d
in
g
s
o
f
t
h
e
Nin
t
h
Co
n
fer
e
n
c
e
o
n
In
n
o
v
a
ti
v
e
Ap
p
li
c
a
ti
o
n
s o
f
Arti
fi
c
i
a
l
In
telli
g
e
n
c
e
,
p
p
.
1
0
5
8
–
1
0
6
5
,
1
9
9
7
.
[5
]
Altaf M
a
h
m
u
d
,
Ka
z
i
Zu
b
a
ir
Ah
m
e
d
,
a
n
d
M
u
m
it
Kh
a
n
.
“
De
tec
ti
n
g
flam
e
s a
n
d
in
su
lt
s in
tex
t.
”
In
Pr
o
c
e
e
d
in
g
s o
f
th
e
S
ixth
I
n
ter
n
a
ti
o
n
a
l
Co
n
fer
e
n
c
e
o
n
Na
tu
r
a
l
L
a
n
g
u
a
g
e
Pro
c
e
ss
in
g
,
2
0
0
8
.
[6
]
Am
ir
H.
Ra
z
a
v
i,
Dia
n
a
I
n
k
p
e
n
,
S
a
sh
a
Uritsk
y
,
a
n
d
S
tan
M
a
twi
n
.
“
Offe
n
siv
e
la
n
g
u
a
g
e
d
e
tec
ti
o
n
u
sin
g
m
u
lt
i
-
le
v
e
l
c
las
sifica
ti
o
n
.
”
In
Pr
o
c
e
e
d
in
g
s o
f
th
e
2
3
rd
C
a
n
a
d
i
a
n
Co
n
fer
e
n
c
e
o
n
Arti
fi
c
i
a
l
I
n
telli
g
e
n
c
e
,
p
p
.
16
–
27
,
2
0
1
0
.
[7
]
“
M
in
istr
y
o
f
I
n
terio
r
i
n
S
a
u
d
i
Ar
a
b
ia
.”
[On
li
n
e
].
A
v
a
il
a
b
le:
h
tt
p
s:/
/
ww
w.m
o
i.
g
o
v
.
sa
/
[Ac
c
e
ss
e
d
1
No
v
e
m
b
e
r
2
0
1
7
].
[8
]
“
P
u
b
l
ic P
ro
se
c
u
ti
o
n
in
S
a
u
d
i
Ara
b
ia
.”
[On
li
n
e
].
Av
a
il
a
b
le:
h
t
tp
s://
ww
w.b
ip
.
g
o
v
.
sa
/
[Ac
c
e
ss
e
d
1
No
v
e
m
b
e
r
2
0
1
7
]
.
[9
]
D.
Lew
is,
K.
Kn
o
wle
s.
“
Th
re
a
d
i
n
g
e
lec
tro
n
ic
m
a
il
:
A
p
re
li
m
i
n
a
ry
s
tu
d
y
.
”
I
n
fo
rm
a
ti
o
n
Pro
c
e
ss
in
g
a
n
d
M
a
n
a
g
e
me
n
t
,
v
o
l.
3
3
,
n
o
.
2
,
p
p
.
2
0
9
–
2
1
7
,
1
9
9
7
.
[1
0
]
W.
Co
h
e
n
.
“
Lea
rn
i
n
g
ru
les
t
h
a
t
c
l
a
ss
ify
e
-
m
a
il
.
”
AA
AI
Co
n
fer
e
n
c
e
,
1
9
9
6
.
[1
1
]
V.
R.
d
e
Ca
rv
a
lh
o
,
W.
C
o
h
e
n
.
“
On
th
e
c
o
ll
e
c
ti
v
e
c
las
sifica
ti
o
n
o
f
e
m
a
il
sp
e
e
c
h
a
c
ts
.”
ACM
S
IGIR
Co
n
fer
e
n
c
e
,
2
0
0
5
.
[1
2
]
K.
Wo
o
d
s,
J.
Ke
g
e
lme
y
e
r,
W.
P
.
,
a
n
d
K.
Bo
w
y
e
r.
“
Co
m
b
in
a
ti
o
n
o
f
m
u
lt
ip
le
c
las
sifiers
u
sin
g
lo
c
a
l
a
c
c
u
ra
c
y
e
stim
a
tes
.
”
IEE
E
T
ra
n
sa
c
ti
o
n
s
o
n
Pa
tt
e
rn
An
a
lys
is a
n
d
M
a
c
h
in
e
I
n
telli
g
e
n
c
e
,
v
o
l.
1
9
,
n
o
.
4
,
p
p
.
4
0
5
–
4
1
0
,
1
9
9
7
.
[1
3
]
Y.
Bi,
D.
Be
ll
,
H.
Wan
g
,
G
.
Gu
o
,
a
n
d
J.
G
u
a
n
.
“
Co
m
b
in
i
n
g
m
u
lt
ip
le
c
las
sifiers
u
si
n
g
d
e
m
p
ste
r
’s
ru
le
f
o
r
tex
t
c
a
teg
o
riza
ti
o
n
.
”
Ap
p
li
e
d
Arti
fi
c
ia
l
In
telli
g
e
n
c
e
,
v
o
l.
2
1
,
n
o
.
3
,
p
p
.
2
1
1
–
2
3
9
,
2
0
0
7
.
[1
4
]
J.
L.
El
m
a
n
.
“
F
in
d
i
n
g
str
u
c
tu
re
i
n
ti
m
e
.
”
Co
g
n
it
ive
S
c
ien
c
e
,
v
o
l
.
1
4
,
n
o
.
2
,
p
p
.
1
7
9
–
2
1
1
,
1
9
9
0
.
[1
5
]
S
.
Ho
c
h
re
it
e
r
a
n
d
J.
S
c
h
m
i
d
h
u
b
e
r
.
“
Lo
n
g
s
h
o
rt
-
term
m
e
m
o
ry
.
”
Ne
u
ra
l
Co
m
p
u
t
a
ti
o
n
,
v
o
l
.
9
,
n
o
.
8
,
p
p
.
1
7
3
5
–
1
7
8
0
,
1
9
9
7
.
[1
6
]
M.
Da
d
v
a
r,
D.
Tri
e
sc
h
n
i
g
g
,
F
.
Jo
n
g
,
“
Ex
p
e
rts
a
n
d
M
a
c
h
i
n
e
s
a
g
a
in
st
B
u
ll
ies
:
A
H
y
b
ri
d
A
p
p
r
o
a
c
h
to
De
tec
t
Cy
b
e
rb
u
ll
ies
.”
A
d
v
a
n
c
e
s in
Arti
fi
c
ia
l
In
tell
ig
e
n
c
e
,
p
p
2
7
5
-
2
8
1
,
2
0
1
4
.
[1
7
]
R.
Ju
st
o
,
T.
C
o
rc
o
ra
n
,
S
.
Lu
k
in
,
M
.
Wal
k
e
r,
M
.
I
n
e
s
T
o
rre
s
.
“
E
x
t
ra
c
ti
n
g
Re
lev
a
n
t
K
n
o
wle
d
g
e
f
o
r
t
h
e
De
tec
ti
o
n
o
f
S
a
rc
a
sm
a
n
d
Na
stin
e
ss
in
th
e
S
o
c
ial
Web
.”
Kn
o
wled
g
e
-
Ba
se
d
S
y
st
e
ms
,
2
0
1
4
.
[1
8
]
G
u
a
n
g
Xia
n
g
Bin
F
a
n
Li
n
g
Wan
g
Ja
so
n
I.
Ho
n
g
Ca
ro
l
y
n
P
.
Ro
se
.
,
“
De
tec
ti
n
g
Offe
n
siv
e
Twe
e
ts
v
ia T
o
p
ica
l
F
e
a
tu
re
Disc
o
v
e
ry
o
v
e
r
a
Larg
e
S
c
a
le
Twit
ter
Co
rp
u
s
.”
P
ro
c
e
e
d
i
n
g
o
f
th
e
2
1
st
ACM
in
ter
n
a
ti
o
n
a
l
c
o
n
fer
e
n
c
e
o
n
I
n
fo
rm
a
ti
o
n
a
n
d
k
n
o
w
led
g
e
m
a
n
a
g
e
me
n
t
(CIK
M
'1
2
)
,
p
p
.
1
9
8
0
1
9
8
4
,
2
0
1
2
.
[1
9
]
Ra
z
a
v
i,
Am
ir,
In
k
p
e
n
,
Dia
n
a
,
Ur
it
sk
y
,
S
a
sh
a
&
M
a
twin
,
S
ta
n
.
“
Offe
n
siv
e
Lan
g
u
a
g
e
De
tec
ti
o
n
Us
in
g
M
u
lt
i
-
lev
e
l
Clas
sifica
ti
o
n
.
”
p
p
.
1
6
-
2
7
,
2
0
1
0
.
[2
0
]
D.
Yin
,
Z.
Qu
e
,
L.
H
o
n
g
.
“
De
tec
ti
o
n
o
f
Ha
ra
ss
m
e
n
t
o
n
Web
2
.
0
.”
CAW
2
.
0
,
S
p
a
i
n
,
2
0
0
9
.
[2
1
]
Wei
Wan
g
,
Die
p
Bich
D
o
,
a
n
d
X
u
e
m
in
Li
n
.
“
Term
g
ra
p
h
m
o
d
e
l
fo
r
tex
t
c
las
sifica
ti
o
n
.
”
I
n
I
n
ter
n
a
t
io
n
a
l
C
o
n
fer
e
n
c
e
o
n
A
d
v
a
n
c
e
d
Da
t
a
M
in
i
n
g
a
n
d
A
p
p
li
c
a
ti
o
n
s
.
S
p
ri
n
g
e
r,
p
p
.
1
9
–
3
0
,
2
0
0
5
.
[2
2
]
Da
v
id
M
.
Blei,
An
d
re
w
Y.
N
g
,
a
n
d
M
ich
a
e
l
I.
Jo
rd
a
n
.
“
Late
n
t
Dir
ich
let
All
o
c
a
ti
o
n
.
”
J
o
u
rn
a
l
o
f
M
a
c
h
in
e
L
e
a
rn
i
n
g
Res
e
a
rc
h
,
p
p
.
9
9
3
–
1
0
2
2
,
2
0
0
3
.
[2
3
]
An
a
n
d
Ra
jara
m
a
n
a
n
d
Je
ffre
y
Da
v
id
Ul
lma
n
.
“
M
i
n
in
g
o
f
M
a
ss
iv
e
Da
tas
e
ts.
”
Ca
m
b
rid
g
e
Un
iv
e
rsity
P
re
ss
,
2
0
1
1
.
[2
4
]
F
ra
n
ç
o
is
R
o
u
ss
e
a
u
,
Emm
a
n
o
u
il
Kia
g
ias
,
a
n
d
M
ich
a
li
s
Va
z
irg
ian
n
is.
“
Te
x
t
c
a
teg
o
riza
ti
o
n
a
s
a
g
ra
p
h
c
las
sifica
ti
o
n
p
ro
b
lem
.
”
In
ACL
,
v
o
l.
1
5
,
p
p
.
1
7
0
2
-
1
7
1
2
,
2
0
1
5
.
[2
5
]
A.
Re
y
e
s,
P
.
Ro
ss
o
.
“
M
a
k
in
g
o
b
jec
ti
v
e
d
e
c
is
io
n
s
fro
m
su
b
jec
ti
v
e
d
a
ta:
De
tec
ti
n
g
ir
o
n
y
in
c
u
st
o
m
e
r
re
v
iew
s
.”
De
c
isio
n
S
u
p
p
o
rt
S
y
ste
ms
,
v
o
l.
5
3
,
n
o
.
4
,
p
p
.
7
5
4
-
7
6
0
,
2
0
1
2
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
7
2
2
-
3
2
2
1
C
o
m
p
u
t.
Sci.
I
n
f
.
T
ec
h
n
o
l.
,
V
o
l.
1
,
No
.
2
,
J
u
ly
20
2
0
:
84
–
92
92
[2
6
]
Y.
Kim
. “
Co
n
v
o
l
u
ti
o
n
a
l
n
e
u
ra
l
n
e
two
rk
s f
o
r
se
n
ten
c
e
c
las
sifica
ti
o
n
.”
a
rXiv
p
re
p
ri
n
t
a
rXiv
:
1
4
0
8
.
5
8
8
2
,
2
0
1
4
.
[2
7
]
Ro
n
a
n
Co
ll
o
b
e
rt,
Ja
so
n
Wes
to
n
,
Léo
n
Bo
tt
o
u
,
M
ic
h
a
e
l
Ka
rlen
,
Ko
ra
y
Ka
v
u
k
c
u
o
g
l
u
,
a
n
d
P
a
v
e
l
Ku
k
sa
.
“
Na
tu
ra
l
Lan
g
u
a
g
e
P
ro
c
e
ss
in
g
(Alm
o
st)
fr
o
m
S
c
ra
tch
.
”
J
.
M
a
c
h
.
L
e
a
rn
.
Res
.
v
o
l.
1
2
,
p
p
.
2
4
9
3
–
2
5
3
7
,
2
0
1
1
.
[2
8
]
Ya
rin
G
a
l
a
n
d
Zo
u
b
in
G
h
a
h
ra
m
a
n
i.
“
A
Th
e
o
re
ti
c
a
ll
y
G
ro
u
n
d
e
d
Ap
p
l
ica
ti
o
n
o
f
Dro
p
o
u
t
i
n
Re
c
u
rre
n
t
Ne
u
ra
l
Ne
two
rk
s.
”
In
A
d
v
a
n
c
e
s
i
n
Ne
u
r
a
l
I
n
fo
rm
a
ti
o
n
Pro
c
e
ss
in
g
S
y
ste
ms
2
9
:
A
n
n
u
a
l
Co
n
fer
e
n
c
e
o
n
Ne
u
ra
l
In
f
o
rm
a
ti
o
n
Pro
c
e
ss
in
g
S
y
ste
ms
,
Ba
rc
e
lo
n
a
,
S
p
a
in
,
p
p
.
1
0
1
9
–
1
0
2
7
,
2
0
1
6
.
[2
9
]
M
ik
o
lo
v
,
To
m
a
s
&
S
u
tsk
e
v
e
r,
Il
y
a
&
Ch
e
n
,
Ka
i
&
Co
rra
d
o
,
G
.
s
&
De
a
n
,
Je
ffre
y
.
“
Distrib
u
ted
Re
p
re
se
n
tatio
n
s
o
f
Wo
rd
s
a
n
d
P
h
ra
se
s
a
n
d
th
e
ir
C
o
m
p
o
siti
o
n
a
li
t
y
.
”
Ad
v
a
n
c
e
s
in
Ne
u
ra
l
In
fo
rm
a
ti
o
n
Pro
c
e
ss
in
g
S
y
ste
ms
.
v
o
l.
26
,
p
p
.
3
1
1
1
–
3
1
1
9
,
2
0
1
3
.
[3
0
]
Je
ffre
y
P
e
n
n
in
g
to
n
,
Rich
a
rd
S
o
c
h
e
r
,
a
n
d
Ch
risto
p
h
e
r
D.
M
a
n
n
in
g
.
“
G
lo
v
e
:
G
lo
b
a
l
v
e
c
to
rs fo
r
wo
rd
r
e
p
re
se
n
tatio
n
.
”
In
EM
N
L
P
,
2
0
1
4
.
[3
1
]
K
Ch
o
,
B
v
a
n
M
e
rrien
b
o
e
r,
C
G
u
lce
h
re
,
F
B
o
u
g
a
re
s,
H
S
c
h
we
n
k
,
“
Lea
rn
in
g
P
h
ra
se
Re
p
re
se
n
tati
o
n
s
u
si
n
g
RNN
En
c
o
d
e
r
-
De
c
o
d
e
r
fo
r
S
tatisti
c
a
l
M
a
c
h
in
e
Tra
n
sla
ti
o
n
”
,
C
o
n
fer
e
n
c
e
o
n
Emp
irica
l
M
e
th
o
d
s
in
Na
tu
ra
l
L
a
n
g
u
a
g
e
Pro
c
e
ss
in
g
(E
M
NL
P)
,
Qa
tar,
2
0
1
4
.
[3
2
]
R.
S
o
c
h
e
r,
B.
Hu
v
a
l,
C.
D.
M
a
n
n
in
g
,
a
n
d
A.
Y.N
g
,
“
S
e
m
a
n
ti
c
c
o
m
p
o
siti
o
n
a
li
ty
t
h
r
o
u
g
h
re
c
u
rsi
v
e
m
a
tri
x
-
v
e
c
to
r
sp
a
c
e
s”
,
P
ro
c
e
e
d
in
g
s
o
f
th
e
2
0
1
2
J
o
i
n
t
C
o
n
fer
e
n
c
e
o
n
Emp
iric
a
l
M
e
t
h
o
d
s
in
Na
t
u
ra
l
L
a
n
g
u
a
g
e
Pro
c
e
ss
in
g
a
n
d
Co
mp
u
t
a
ti
o
n
a
l
N
a
tu
r
a
l
L
a
n
g
u
a
g
e
L
e
a
rn
i
n
g
.
As
so
c
iatio
n
fo
r
Co
m
p
u
tatio
n
a
l
Li
n
g
u
isti
c
s,
p
p
.
1
2
0
1
1
2
1
1
,
2
0
1
2
.
[3
3
]
Tai,
Ka
i
S
h
e
n
g
,
Rich
a
rd
S
o
c
h
e
r
a
n
d
C
h
risto
p
h
e
r
D.
M
a
n
n
in
g
.
“
I
m
p
ro
v
e
d
S
e
m
a
n
ti
c
Re
p
re
se
n
tati
o
n
s
F
ro
m
Tree
-
S
tru
c
t
u
re
d
Lo
n
g
S
h
o
rt
-
Term
M
e
m
o
ry
N
e
two
rk
s
.
”
ArX
iv
a
b
s/
1
5
0
3
.
0
0
0
7
5
,
2
0
1
5
.
[3
4
]
M
ik
a
e
l
He
n
a
ff,
Jo
a
n
Bru
n
a
,
a
n
d
Ya
n
n
LeCu
n
.
“
De
e
p
Co
n
v
o
l
u
ti
o
n
a
l
Ne
two
rk
s
o
n
G
ra
p
h
-
S
tr
u
c
tu
re
d
Da
ta”
,
Co
RR,
2
0
1
5
.
[3
5
]
T.
Y.
Li
u
,
Y.
Ya
n
g
,
H.
Wan
,
H.
Zen
g
,
Z.
C
h
e
n
,
a
n
d
W.
Y.
M
a
.
“
S
u
p
p
o
rt
v
e
c
to
r
m
a
c
h
in
e
s
c
las
sifica
ti
o
n
wit
h
a
v
e
r
y
larg
e
-
sc
a
le t
a
x
o
n
o
m
y
.
”
ACM
S
IG
KDD
Exp
lo
ra
t
io
n
s Ne
wsle
tt
e
r
,
v
o
l.
7
,
n
o
.
1,
p
p
.
36
–
4
3
,
2
0
0
5
.
[3
6
]
G
u
i
-
Ro
n
g
X
u
e
,
Di
k
a
n
Xi
n
g
,
Qia
n
g
Ya
n
g
,
a
n
d
Yo
n
g
Yu
.
“
De
e
p
c
las
sifica
ti
o
n
in
larg
e
-
sc
a
le
tex
t
h
iera
rc
h
ies
.
”
In
S
IGIR
.
p
p
.
6
1
9
–
6
2
6
,
20
08
.
[3
7
]
Jo
h
n
s
o
n
R,
Zh
a
n
g
T.
“
Eff
e
c
ti
v
e
u
se
o
f
wo
rd
o
r
d
e
r
fo
r
tex
t
c
a
teg
o
riza
ti
o
n
wit
h
c
o
n
v
o
lu
ti
o
n
a
l
n
e
u
ra
l
n
e
two
rk
s.
”
I
n
Pro
c
e
e
d
in
g
s
o
f
th
e
C
o
n
fer
e
n
c
e
o
f
th
e
N
o
rth
Ame
ric
a
n
C
h
a
p
ter
o
f
th
e
Asso
c
i
a
ti
o
n
f
o
r
C
o
mp
u
ta
ti
o
n
a
l
L
in
g
u
isti
c
s:
Hu
ma
n
L
a
n
g
u
a
g
e
T
e
c
h
n
o
lo
g
ies
(
NAA
CL
-
HLT
2
0
1
5
)
,
2
0
1
5
.
[3
8
]
Tan
g
D,
Qin
B
,
Li
u
T.
“
Do
c
u
m
e
n
t
m
o
d
e
ll
i
n
g
wi
th
g
a
ted
re
c
u
rre
n
t
n
e
u
ra
l
n
e
two
r
k
fo
r
se
n
t
ime
n
t
c
las
sifica
ti
o
n
.
”
I
n
Pro
c
e
e
d
in
g
s
o
f
t
h
e
Co
n
fer
e
n
c
e
o
n
Emp
irica
l
M
e
th
o
d
s
in
N
a
tu
r
a
l
L
a
n
g
u
a
g
e
Pro
c
e
ss
in
g
(
EM
NL
P
2
0
1
5
)
,
2
0
1
5
.
[3
9
]
Do
u
ZY.
“
Ca
p
t
u
ri
n
g
u
se
r
a
n
d
p
ro
d
u
c
t
In
f
o
rm
a
ti
o
n
fo
r
d
o
c
u
m
e
n
t
le
v
e
l
se
n
ti
m
e
n
t
a
n
a
l
y
sis
wit
h
d
e
e
p
m
e
m
o
ry
n
e
two
rk
.
”
In
Pr
o
c
e
e
d
in
g
s
o
f
t
h
e
C
o
n
fer
e
n
c
e
o
n
Emp
iric
a
l
M
e
th
o
d
s
o
n
Na
t
u
ra
l
L
a
n
g
u
a
g
e
Pro
c
e
ss
in
g
(EM
NL
P
2
0
1
7
)
,
2
0
1
7
.
[4
0
]
Zh
o
u
X,
Wa
n
X,
Xia
o
J.
“
Atten
ti
o
n
-
b
a
se
d
LS
TM
n
e
two
rk
fo
r
c
r
o
ss
-
li
n
g
u
a
l
se
n
ti
m
e
n
t
c
las
sifica
ti
o
n
.
”
I
n
Pro
c
e
e
d
in
g
s
o
f
t
h
e
Co
n
fer
e
n
c
e
o
n
Em
p
irica
l
M
e
th
o
d
s i
n
Na
t
u
ra
l
L
a
n
g
u
a
g
e
Pr
o
c
e
ss
in
g
(E
M
NL
P
2
0
1
6
)
,
2
0
1
6
.
[4
1
]
Kim
Y.
“
Co
n
v
o
lu
ti
o
n
a
l
n
e
u
ra
l
n
e
two
rk
s
fo
r
se
n
ten
c
e
c
las
sifica
ti
o
n
.
”
In
Pro
c
e
e
d
in
g
s
o
f
t
h
e
An
n
u
a
l
M
e
e
ti
n
g
o
f
t
h
e
Asso
c
ia
ti
o
n
f
o
r Co
m
p
u
t
a
ti
o
n
a
l
L
i
n
g
u
isti
c
s (A
CL
2
0
1
4
)
,
2
0
1
4
.
[4
2
]
Wan
g
J,
Yu
L
-
C,
Lai
R.
K.,
a
n
d
Z
h
a
n
g
X.
“
Dim
e
n
sio
n
a
l
se
n
ti
m
e
n
t
a
n
a
ly
sis u
si
n
g
a
re
g
i
o
n
a
l
CNN
-
L
S
TM
m
o
d
e
l.
”
I
n
Pro
c
e
e
d
in
g
s
o
f
t
h
e
A
n
n
u
a
l
M
e
e
ti
n
g
o
f
t
h
e
Asso
c
ia
ti
o
n
f
o
r Co
m
p
u
t
a
ti
o
n
a
l
L
i
n
g
u
isti
c
s (A
CL
2
0
1
6
)
,
2
0
1
6
.
[4
3
]
G
u
g
g
il
la
C,
M
il
ler
T,
G
u
re
v
y
c
h
I.
“
CNN
-
a
n
d
LS
TM
-
b
a
se
d
c
lai
m
c
las
sifica
ti
o
n
in
o
n
li
n
e
u
se
r
c
o
m
m
e
n
ts.
”
In
Pro
c
e
e
d
in
g
s
o
f
t
h
e
In
ter
n
a
ti
o
n
a
l
Co
n
fer
e
n
c
e
o
n
Co
m
p
u
t
a
ti
o
n
a
l
L
i
n
g
u
isti
c
s (COL
ING 2
0
1
6
)
,
2
0
1
6
.
[4
4
]
Yu
J,
Jia
n
g
J.
“
Lea
rn
i
n
g
se
n
ten
c
e
e
m
b
e
d
d
in
g
s
wit
h
a
u
x
i
li
a
ry
tas
k
s
fo
r
c
r
o
ss
-
d
o
m
a
in
se
n
ti
m
e
n
t
c
las
sifica
ti
o
n
.
”
I
n
Pro
c
e
e
d
in
g
s
o
f
t
h
e
Co
n
fer
e
n
c
e
o
n
Emp
irica
l
M
e
th
o
d
s
in
N
a
tu
r
a
l
L
a
n
g
u
a
g
e
Pro
c
e
ss
in
g
(
EM
NL
P
2
0
1
6
)
,
2
0
1
6
.
[4
5
]
Be
n
Ism
a
il
,
M
o
h
a
m
e
d
M
a
h
e
r
&
Bc
h
ir,
Ou
iem
.
“
In
s
u
lt
De
tec
ti
o
n
i
n
S
o
c
ial
Ne
two
r
k
Co
m
m
e
n
ts
Us
i
n
g
P
o
ss
ib
i
li
stic
Ba
se
d
F
u
sio
n
Ap
p
ro
a
c
h
.
”
S
p
ri
n
g
e
r In
ter
n
a
ti
o
n
a
l
P
u
b
li
sh
i
n
g
,
p
p
.
1
5
-
25
,
2
0
1
4
.
[4
6
]
Ya
n
n
LeCu
n
,
Leo
n
Bo
tt
o
u
,
G
e
n
e
v
iev
e
B.
Orr
a
n
d
Kla
u
s
-
Ro
b
e
rt
M
u
l
ler
.
“
Eff
icie
n
t
b
a
c
k
p
ro
p
.
Ne
u
ra
l
n
e
two
rk
s:
Tri
c
k
s
o
f
th
e
trad
e
.”
S
p
rin
g
e
r B
e
rlin
He
i
d
e
lb
e
rg
,
p
p
.
9
-
4
8
,
2
0
1
2
.
[4
7
]
Jo
h
n
Du
c
h
i,
El
a
d
Ha
z
a
n
,
a
n
d
Yo
ra
m
S
in
g
e
r.
“
Ad
a
p
ti
v
e
su
b
g
ra
d
ien
t
m
e
th
o
d
s
fo
r
o
n
li
n
e
lea
rn
i
n
g
a
n
d
st
o
c
h
a
stic
o
p
ti
m
iza
ti
o
n
.
”
T
h
e
J
o
u
rn
a
l
o
f
M
a
c
h
in
e
L
e
a
r
n
in
g
Res
e
a
rc
h
,
v
o
l.
12
,
p
p
.
2
1
2
1
–
2
1
5
9
,
2
0
1
1
.
[4
8
]
Jo
u
li
n
,
Arm
a
n
d
,
Ed
o
u
a
r
d
G
ra
v
e
,
P
io
tr
Bo
ja
n
o
ws
k
i
,
M
a
tt
h
ij
s
Do
u
z
e
,
He
rv
é
Jé
g
o
u
a
n
d
To
m
a
s
M
i
k
o
lo
v
.
“
F
a
stTex
t.
z
ip
:
Co
m
p
re
ss
in
g
tex
t
c
las
sifica
ti
o
n
m
o
d
e
ls.”
ArX
iv
a
b
s/1
6
1
2
.
0
3
6
5
1
,
2
0
1
6
.
[4
9
]
F
a
b
ian
P
e
d
re
g
o
sa
,
G
a
e
l
Va
ro
q
u
a
u
x
,
Ale
x
a
n
d
re
G
ra
m
fo
rt,
Vin
c
e
n
t
M
ich
e
l,
Be
rtran
d
Th
iri
o
n
,
Oli
v
ier
G
rise
l,
M
a
th
ieu
Blo
n
d
e
l,
P
e
ter
P
re
tt
e
n
h
o
fe
r,
R
o
n
Weiss
,
Vin
c
e
n
t
Du
b
o
u
rg
,
Ja
k
e
Va
n
d
e
rp
las
,
Ale
x
a
n
d
re
P
a
ss
o
s,
Da
v
id
Co
u
rn
a
p
e
a
u
,
M
a
tt
h
ieu
Br
u
c
h
e
r,
M
a
tt
h
ie
u
P
e
rro
t,
a
n
d
Ed
o
u
a
r
d
Du
c
h
e
sn
a
y
.
“
S
c
ik
i
t
-
lea
rn
:
M
a
c
h
i
n
e
lea
rn
in
g
i
n
p
y
t
h
o
n
.
”
J
o
u
rn
a
l
o
f
M
a
c
h
i
n
e
L
e
a
r
n
in
g
Res
e
a
rc
h
,
v
o
l.
1
2
,
n
o
.
10
,
p
p
.
2
8
2
5
–
2
8
3
0
,
2
0
1
2
.
Evaluation Warning : The document was created with Spire.PDF for Python.