I
n
t
e
r
n
at
ion
al
Jou
r
n
al
of
E
lec
t
r
ical
an
d
Com
p
u
t
e
r
E
n
gin
e
e
r
in
g
(
I
JE
CE
)
Vol.
15
,
No.
1
,
F
e
br
ua
r
y
20
25
,
pp.
700
~
710
I
S
S
N:
2088
-
8708
,
DO
I
:
10
.
11591/i
jec
e
.
v
15
i
1
.
pp
7
00
-
710
700
Jou
r
n
al
h
omepage
:
ht
tp:
//
ij
e
c
e
.
iaes
c
or
e
.
c
om
D
e
n
ig
r
a
t
io
n
a
n
al
ysi
s of
T
w
itt
e
r
d
a
t
a u
si
n
g c
yc
li
c
l
e
ar
n
in
g
r
at
e
b
ase
d
l
o
n
g sh
o
r
t
-
t
e
r
m
m
e
m
o
r
y
S
u
h
as
B
h
ar
ad
waj
Raj
e
n
d
r
a
1
,
S
a
m
p
at
h
Ku
z
h
alv
aim
oz
h
i
1
,
Ve
d
ava
t
h
i
Nage
n
d
r
a
P
r
as
ad
2
1
D
e
pa
r
tm
e
nt
of
I
nf
or
ma
ti
on S
c
ie
nc
e
a
nd E
ngi
ne
e
r
in
g, T
he
N
a
ti
ona
l
I
ns
ti
tu
te
of
E
ngi
ne
e
r
in
g, M
ys
ur
u,
A
f
f
il
ia
te
d t
o
V
is
ve
s
va
r
a
ya
T
e
c
hnol
ogi
c
a
l
U
ni
ve
r
s
it
y, B
e
la
ga
vi
, I
ndi
a
2
D
e
pa
r
tm
e
nt
of
C
omput
e
r
S
c
ie
nc
e
a
nd
E
ngi
ne
e
r
in
g
, V
id
ya
va
r
d
ha
ka
C
ol
le
ge
of
E
ngi
ne
e
r
in
g, M
ys
ur
u, I
ndi
a
Ar
t
icle
I
n
f
o
AB
S
T
RA
CT
A
r
ti
c
le
h
is
tor
y
:
R
e
c
e
ived
F
e
b
19,
2024
R
e
vis
e
d
J
ul
16,
2024
Ac
c
e
pted
Oc
t
1,
2024
T
ech
n
o
l
o
g
i
ca
l
i
n
n
o
v
a
t
i
o
n
h
a
s
g
i
v
e
n
ri
s
e
t
o
a
n
ew
fo
rm
o
f
b
u
l
l
y
i
n
g
,
o
ft
e
n
l
ead
i
n
g
t
o
s
i
g
n
i
fi
ca
n
t
h
arm
t
o
o
n
e'
s
rep
u
t
a
t
i
o
n
w
i
t
h
i
n
s
o
ci
a
l
ci
rcl
e
s
.
W
h
en
a
s
i
n
g
l
e
p
ers
o
n
b
eco
me
s
t
arg
et
t
o
an
i
m
o
s
i
t
y
an
d
h
aras
s
me
n
t
i
n
a
cy
b
erb
u
l
l
y
i
n
g
i
n
c
i
d
e
n
t
,
i
t
i
s
t
ermed
a
s
d
e
n
i
g
ra
t
i
o
n
.
Man
y
d
i
ffere
n
t
c
y
b
erb
u
l
l
y
i
n
g
d
e
t
ect
i
o
n
t
ech
n
i
q
u
e
s
are
carri
e
d
o
u
t
t
o
co
u
n
t
er
t
h
i
s
,
co
n
cen
t
ra
t
i
n
g
o
n
w
o
r
d
-
b
a
s
ed
d
at
a
an
d
u
s
er
acco
u
n
t
feat
u
res
o
n
l
y
.
T
h
e
mai
n
o
b
j
ect
i
v
e
o
f
t
h
i
s
res
earch
i
s
t
o
en
h
a
n
ce
t
h
e
l
earn
i
n
g
rat
e
o
f
l
o
n
g
s
h
o
r
t
-
t
erm
mem
o
ry
(
L
ST
M)
u
s
i
n
g
c
y
cl
i
c
l
earn
i
n
g
rat
e
(CL
R).
T
h
erefo
re,
i
n
t
h
i
s
re
s
earch
,
c
y
b
erb
u
l
l
y
i
n
g
i
n
s
o
c
i
a
l
med
i
a
i
s
d
e
t
ect
e
d
b
y
d
e
v
el
o
p
i
n
g
a
framew
o
rk
b
a
s
ed
o
n
L
ST
M
-
CL
R
w
h
i
ch
i
s
mo
re
s
t
a
b
l
e
f
o
r
en
h
a
n
ci
n
g
c
l
as
s
i
f
i
cat
i
o
n
accu
rac
y
w
i
t
h
o
u
t
t
h
e
n
ee
d
fo
r
mu
l
t
i
p
l
e
t
ri
a
l
s
a
n
d
m
o
d
i
fi
ca
t
i
o
n
s
.
T
h
e
effec
t
i
v
en
e
s
s
o
f
t
h
e
s
u
g
g
es
t
ed
L
ST
M
-
CL
R
i
s
as
s
es
s
ed
f
o
r
i
d
en
t
i
f
y
i
n
g
c
y
b
er
b
u
l
l
y
i
n
g
u
s
i
n
g
T
w
i
t
t
er
d
a
t
a.
T
h
e
at
t
a
i
n
e
d
res
u
l
t
s
s
h
o
w
t
h
a
t
t
h
e
p
ro
p
o
s
ed
L
ST
M
-
CL
R
o
b
t
ai
n
s
8
2
%
accu
racy
,
8
0
%
p
reci
s
i
o
n
,
8
3
%
recal
l
an
d
8
1
%
F
-
meas
u
re
i
n
t
h
e
cl
as
s
i
fi
c
at
i
o
n
o
f
cy
b
er
b
u
l
l
y
i
n
g
t
w
eet
s
,
w
h
i
ch
i
s
s
u
p
er
i
o
r
w
h
e
n
co
mp
ar
ed
w
i
t
h
t
h
e
ex
i
s
t
i
n
g
mu
l
t
i
l
ay
er
p
erce
p
t
r
o
n
(ML
P)
an
d
b
i
d
i
rect
i
o
n
al
en
c
o
d
e
r
rep
res
e
n
t
a
t
i
o
n
s
fro
m
t
ran
s
fo
rmer
s
(BE
RT
)
m
o
d
e
l
s
.
K
e
y
w
o
r
d
s
:
C
ybe
r
bull
ying
C
yc
li
c
lea
r
ning
r
a
te
De
nigr
a
ti
on
L
ong
s
hor
t
-
ter
m
memor
y
T
witt
e
r
twe
e
ts
Th
i
s
i
s
a
n
o
p
en
a
c
ces
s
a
r
t
i
c
l
e
u
n
d
e
r
t
h
e
CC
B
Y
-
SA
l
i
ce
n
s
e.
C
or
r
e
s
pon
din
g
A
u
th
or
:
S
uha
s
B
ha
r
a
dwa
j
R
a
jendr
a
De
pa
r
tm
e
nt
of
I
nf
or
mat
ion
S
c
ienc
e
a
nd
E
nginee
r
i
ng,
T
he
Na
ti
ona
l
I
ns
ti
tut
e
of
E
nginee
r
ing
,
Af
f
il
iate
d
to
Vis
ve
s
va
r
a
ya
T
e
c
hnologi
c
a
l
Unive
r
s
it
y
B
e
laga
vi,
Ka
r
na
taka
E
mail:
s
uha
s
br
.
r
e
s
e
a
r
c
h@gmail.
c
om
1.
I
NT
RODU
C
T
I
ON
De
nigr
a
ti
on
c
ons
ti
tut
e
s
a
f
or
m
of
c
ybe
r
bull
yin
g
whe
r
e
a
n
indi
vidual
pur
pos
e
f
ull
y
unde
r
mi
ne
s
s
omeone
e
ls
e
's
r
e
putation
or
s
oc
ial
c
onne
c
ti
ons
by
s
pr
e
a
ding
unf
a
vor
a
ble
int
e
r
ne
t
r
umor
s
or
gos
s
ip.
C
ha
r
a
c
ter
a
s
s
a
s
s
ination
of
publi
c
f
igur
e
s
s
uc
h
a
s
poli
ti
c
ians
a
nd
c
e
l
e
br
it
ies
a
r
e
a
c
omm
on
f
or
m
of
c
ybe
r
bull
ying
of
f
e
ns
e
[
1]
,
[
2]
.
De
nigr
a
ti
on
is
a
pr
omi
ne
nt
thr
e
a
t
in
c
ontempor
a
r
y
s
oc
iety
a
nd
ha
s
s
igni
f
ica
nt
ne
ga
ti
ve
c
ons
e
que
n
c
e
s
f
or
victim
s
,
be
ing
quit
e
de
s
tr
uc
ti
ve
be
c
a
us
e
of
the
s
tr
ong
pr
opa
ga
ti
on
a
nd
f
r
e
que
nc
y
[
3]
.
S
ome
r
e
c
e
nt
e
f
f
or
ts
ha
ve
a
ddr
e
s
s
e
d
thi
s
is
s
ue
with
the
us
e
of
s
e
ve
r
a
l
c
las
s
ica
l
mac
hine
lea
r
ning
(
M
L
)
a
nd
de
e
p
lea
r
ning
(
DL
)
methods
to
de
tec
t
de
nigr
a
ti
on.
T
he
powe
r
o
f
s
oc
ial
ne
twor
king
s
it
e
s
c
r
e
a
tes
a
f
e
r
ti
le
gr
ound
f
o
r
s
pr
e
a
ding
de
f
a
mator
y
r
umo
r
s
,
whic
h
is
a
nother
f
o
r
m
of
de
nigr
a
ti
on
bull
ying.
T
he
s
e
tar
ge
ted
malicious
r
e
mar
ks
r
e
f
lec
t
a
mongs
t
a
huge
numbe
r
of
r
e
c
ipi
e
nts
,
a
nd
a
r
e
a
c
ha
ll
e
nge
to
be
r
e
c
ti
f
ie
d
in
be
s
t
c
a
s
e
s
c
e
na
r
io
[
4]
.
T
he
models
that
a
r
e
de
s
igned
with
the
a
im
of
unve
il
ing
a
nd
a
na
lyzing
e
xpr
e
s
s
ions
of
ins
ult
s
of
ten
f
ound
in
s
uc
h
pos
ts
s
e
r
ve
a
s
ins
tr
u
ments
to
he
lp
identif
y
ha
te
s
pe
e
c
h,
ins
ult
a
nd
bul
lyi
ng
[
5]
.
T
his
r
e
qui
r
e
s
a
mec
ha
nis
m
o
f
de
tec
ti
on
f
o
r
mi
ti
g
a
ti
ng
of
the
unde
r
lyi
ng
ha
r
m
f
ul
e
f
f
e
c
ts
[
6
]
.
An
i
ntr
us
ion
de
tec
ti
on
s
ys
tem
(
I
DS)
s
e
r
ve
s
a
s
a
s
of
twa
r
e
a
ppli
c
a
ti
on
f
or
moni
tor
ing
da
ta
tr
a
f
f
ic
f
low
a
c
r
os
s
a
ne
twor
k
to
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
E
lec
&
C
omp
E
ng
I
S
S
N:
2088
-
8708
De
nigr
ati
on
analys
is
of
T
w
it
ter
data
us
ing
c
y
c
li
c
lear
ning
r
ate
…
(
Suhas
B
har
adw
aj
R
ajendr
a
)
701
identif
y
potential
ins
tanc
e
s
of
malicious
be
ha
vior
[
7]
.
R
e
c
e
ntl
y,
e
nha
nc
e
ments
in
c
ybe
r
bull
ying
c
las
s
if
ica
ti
on
ha
ve
r
e
s
ult
e
d
f
r
om
the
a
utom
a
ti
c
identif
ica
ti
on
o
f
de
nigr
a
ti
on
[
8]
,
[
9]
.
C
ybe
r
bull
ying
is
li
nke
d
to
c
a
us
ing
de
tr
im
e
ntal
im
pa
c
ts
on
menta
l
a
nd
phys
ica
l
he
a
lt
h
,
a
c
a
de
mi
c
pe
r
f
or
manc
e
,
de
pr
e
s
s
ion,
a
nd
a
n
incr
e
a
s
e
d
r
is
k
of
s
uicida
l
thought
s
,
a
s
pe
r
va
r
ious
s
tudi
e
s
[
10]
,
[
11]
.
As
a
r
e
s
ult
,
the
quick
identi
f
ica
ti
on
of
de
nig
r
a
ti
on
is
c
r
it
ica
l
to
mi
ni
mi
z
e
it
s
de
tr
im
e
ntal
c
ons
e
que
nc
e
s
on
victim
s
.
F
u
r
ther
mor
e
,
the
r
e
c
ur
r
ing
na
tur
e
o
f
de
n
igr
a
ti
on
make
s
it
c
r
it
ica
l
to
noti
c
e
a
nd
e
li
mi
na
te
it
a
t
the
e
a
r
li
e
s
t
[
12]
.
T
he
goa
l
of
c
ybe
r
a
ggr
e
s
s
ion
is
to
identif
y
the
a
ggr
e
s
s
or
s
,
while
a
ls
o
s
uppor
ti
ng
the
victim
s
.
T
he
pr
e
domi
na
nt
a
ppr
oa
c
h
in
a
ddr
e
s
s
ing
c
ybe
r
bull
ying
ha
s
lar
ge
ly
c
e
nter
e
d
a
r
ound
s
c
r
uti
nizing
a
nd
qua
nti
f
ying
a
c
c
ur
a
c
y
s
c
or
e
s
with
the
a
ppli
c
a
ti
on
o
f
lea
r
ning
methodologi
e
s
[
13]
.
DL
a
ppr
oa
c
he
s
s
uc
h
a
s
c
onv
olut
ional
ne
ur
a
l
ne
twor
k
(
C
NN
)
,
long
s
hor
t
-
ter
m
memor
y
(
L
S
T
M
)
,
a
nd
bidi
r
e
c
ti
ona
l
long
s
hor
t
-
ter
m
m
e
mor
y
(
B
iL
S
T
M
)
ha
ve
gr
own
in
f
a
vor
of
d
e
tec
ti
ng
c
ybe
r
bull
ying
[
14
]
.
De
s
pit
e
the
numer
ous
p
r
e
ve
nti
on
a
nd
int
e
r
ve
nti
on
a
ppr
oa
c
he
s
,
c
ybe
r
bull
ying
a
c
ti
on
ha
s
not
r
e
duc
e
d
in
the
las
t
de
c
a
de
[
15]
.
T
he
c
ur
r
e
nt
a
na
lys
e
s
ha
ve
obs
e
r
ve
d
r
e
pe
a
tedly
de
tec
ti
ng
c
ybe
r
bull
ying
oc
c
ur
r
e
nc
e
s
that
a
r
e
f
ound
to
be
e
f
f
e
c
ti
ve
in
de
t
e
c
ti
ng
c
ybe
r
bull
ying.
How
e
ve
r
,
their
a
c
c
ur
a
c
y
is
r
e
duc
e
d
onc
e
the
da
ta
s
ize
is
e
nlar
ge
d.
T
he
r
e
f
o
r
e
,
lea
r
n
ing
models
may
not
be
pe
r
f
e
c
t
in
de
a
li
ng
with
r
e
gular
langua
ge
unc
e
r
tainti
e
s
typi
c
a
l
f
o
r
c
ybe
r
bu
ll
ying
[
1
6]
.
T
o
ove
r
c
ome
the
li
mi
tations
of
the
e
xis
ti
ng
me
thods
,
a
DL
model
na
med
L
S
T
M
whic
h
is
int
e
gr
a
ted
wit
h
c
yc
li
c
lea
r
ning
r
a
te
(
C
L
R
)
a
ppr
oa
c
h
is
pr
opos
e
d
in
thi
s
r
e
s
e
a
r
c
h
f
or
e
f
f
e
c
ti
ve
de
tec
ti
on
of
de
nigr
a
ti
on.
I
ts
e
f
f
e
c
ti
ve
ne
s
s
is
mea
s
ur
e
d
by
c
ompar
ing
it
with
the
s
tate
-
of
-
the
-
a
r
t
methods
na
mely,
c
onvolut
ional
ne
ur
a
l
ne
t
wor
k
(
C
NN
)
,
r
e
c
ur
r
e
nt
ne
ur
a
l
ne
twor
k
(
R
NN
)
,
a
nd
ga
ted
r
e
c
ur
r
e
nt
unit
(
GR
U)
.
R
a
j
e
t
al.
[
17
]
s
ugge
s
ted
a
c
ybe
r
bull
ying
de
tec
ti
on
s
ys
tem
us
ing
a
de
e
p
lea
r
ning
f
r
a
mew
or
k
by
e
va
luating
r
e
a
l
-
ti
me
twe
e
ts
a
nd
pos
ts
on
s
oc
ial
media
.
S
e
ve
r
a
l
ne
ur
a
l
ne
two
r
ks
we
r
e
e
xa
mi
ne
d,
a
n
d
it
wa
s
lea
r
ne
d
that
the
C
NN
-
B
i
L
S
T
M
a
c
hieve
d
pr
e
f
e
r
a
ble
c
las
s
if
ica
ti
on
a
c
c
ur
a
c
y
in
de
tec
ti
ng
c
ybe
r
bull
ying
texts
a
s
it
lea
r
nt
the
global
f
e
a
tur
e
s
a
nd
long
-
ter
m
de
pe
nd
e
nc
ies
.
How
e
ve
r
,
thi
s
ne
twor
k
r
e
quir
e
d
a
lar
ge
a
mount
of
da
ta
a
nd
ti
me
f
or
tr
a
ini
ng
.
Aldua
il
a
j
a
nd
B
e
lghi
th
[
18]
de
ve
loped
a
c
ybe
r
bull
ying
identif
ica
ti
on
mec
ha
nis
m
with
a
s
uppor
t
ve
c
tor
mac
hine
us
ing
a
r
e
a
l
da
tas
e
t
f
r
om
YouT
ube
a
nd
T
witt
e
r
.
A
na
tu
r
a
l
langua
ge
T
oolKit
wa
s
us
e
d
f
or
p
r
e
-
pr
oc
e
s
s
ing
the
da
ta,
a
nd
the
wor
ds
we
r
e
e
xtr
a
c
ted
ba
s
e
d
on
dif
f
e
r
e
nt
s
c
e
na
r
ios
us
ing
ter
m
f
r
e
que
nc
y
-
inver
s
e
doc
ument
f
r
e
que
nc
y
(
T
F
-
I
DF)
a
nd
ba
g
o
f
wo
r
ds
(
B
oW
)
.
T
he
s
uppor
t
ve
c
tor
mac
hine
(
S
VM
)
model
wa
s
tr
a
ined
us
ing
r
e
a
l
-
ti
me
da
tas
e
t
a
nd
a
c
hieve
d
be
tt
e
r
c
las
s
if
ica
ti
on
r
e
s
ult
s
.
How
e
ve
r
,
tr
a
ini
ng
the
model
with
huge
da
ta
a
c
c
ompl
is
he
d
be
t
ter
r
e
s
ult
s
.
M
ur
s
he
d
e
t
al
.
[
19
]
de
ve
loped
a
hybr
id
DL
model
to
identif
y
c
ybe
r
bull
ying
us
ing
T
witt
e
r
da
ta
.
T
he
p
r
e
s
e
nted
a
ppr
oa
c
h
wa
s
de
ve
loped
by
int
e
gr
a
ti
ng
E
lm
a
n
-
type
R
NN
with
a
n
opt
im
ize
d
Dolphin
e
c
holoca
ti
on
a
lgor
it
hm
.
T
h
is
wa
s
done
to
f
ine
-
tune
the
E
lm
a
n
R
NN
’
s
pa
r
a
mete
r
s
a
nd
les
s
tr
a
ini
ng
ti
me.
T
his
a
ppr
oa
c
h
r
e
s
ult
e
d
in
the
a
c
c
ur
a
te
de
tec
ti
on
of
c
ybe
r
bull
ying
on
s
oc
ial
media
.
How
e
ve
r
,
the
r
e
wa
s
s
ti
ll
a
n
ope
n
r
e
s
e
a
r
c
h
a
r
e
a
f
or
de
tec
ti
ng
c
ybe
r
bu
ll
ying
f
r
om
i
mage
s
,
vid
e
os
,
a
nd
a
udio.
S
he
lk
e
a
nd
Attar
[
20]
pr
opos
e
d
a
hyb
r
id
b
i
dir
e
c
ti
ona
l
L
S
T
M
wi
th
a
mul
ti
laye
r
pe
r
c
e
pt
r
on
(
B
iL
S
T
M
-
M
L
P
)
model
f
or
c
ybe
r
bull
ying
de
tec
ti
on
us
ing
r
e
a
l
-
wor
ld
a
nd
be
nc
hmar
k
da
tas
e
ts
on
T
witt
e
r
.
B
iL
S
T
M
wa
s
us
e
d
f
or
wor
d
e
mbedding
a
nd
wa
s
c
ombi
ne
d
with
mul
ti
laye
r
pe
r
c
e
ptr
on
(
M
L
P
)
by
us
ing
pos
t
-
wis
e
f
e
a
tur
e
s
,
whic
h
im
pr
ove
d
the
a
c
c
ur
a
c
y.
T
h
is
model
mai
nly
f
oc
us
e
d
on
text
,
c
ontent
-
ba
s
e
d,
a
nd
lexic
a
l
c
a
tegor
y
f
e
a
tur
e
s
f
or
c
ybe
r
bu
ll
ying
de
tec
ti
on
.
How
e
ve
r
,
th
e
dr
a
wba
c
k
wa
s
that
it
d
id
not
f
oc
us
on
mu
lt
im
e
d
ia
-
ba
s
e
d
f
e
a
tur
e
s
.
R
a
j
e
t
al.
[
21
]
pr
opos
e
d
a
mec
ha
nis
m
to
de
tec
t
c
y
be
r
bull
ying
by
us
ing
a
hybr
id
bidi
r
e
c
ti
ona
l
ga
ted
r
e
c
ur
r
e
nt
u
nit
(
B
i
-
GR
U)
a
nd
C
NN
-
B
iL
S
T
M
on
r
e
a
l
-
wor
ld
c
ybe
r
bull
ying
.
B
i
-
GR
U
wa
s
us
e
d
f
or
text
r
e
pr
e
s
e
ntation
by
us
ing
global
ve
c
tor
s
(
GloVe
)
,
while
C
NN
-
B
iL
S
T
M
wa
s
e
mpl
oye
d
in
the
c
las
s
if
ica
ti
on
model.
T
he
model
of
f
e
r
e
d
a
r
obus
t
mec
ha
nis
m
with
s
ha
ll
ow
ne
ur
a
l
ne
twor
ks
to
r
e
duc
e
the
r
e
quir
e
ment
of
c
ompl
e
x
ne
ur
a
l
ne
twor
ks
.
B
ut
it
a
ls
o
e
xhibi
ted
a
li
mi
tation
of
be
ing
una
ble
to
a
c
hieve
high
a
c
c
ur
a
c
ies
whe
n
met
with
e
xtens
ive
da
tas
e
ts
in
s
upe
r
vis
e
d
c
las
s
if
ica
ti
on.
B
e
hl
e
t
al
.
[
22]
de
ve
loped
a
M
L
P
with
a
n
o
pti
mi
z
e
r
f
or
the
e
f
f
e
c
ti
ve
c
las
s
if
ica
ti
on
of
T
witt
e
r
twe
e
ts
du
r
ing
the
C
OV
I
D
-
19
c
r
is
is
.
T
he
thr
e
e
c
a
tegor
ies
c
ons
ider
e
d
we
r
e
r
e
s
our
c
e
ne
e
ds
,
r
e
s
our
c
e
a
va
il
a
bil
it
y,
a
nd
other
s
.
B
e
tt
e
r
c
las
s
if
ica
ti
on
r
e
s
ult
s
we
r
e
obta
ined
by
e
mpl
oying
loca
l
int
e
r
p
r
e
table
model
-
a
gnos
ti
c
e
xp
lana
ti
ons
to
e
xa
mi
ne
the
be
ha
vior
of
the
pr
opos
e
d
model.
How
e
ve
r
,
due
to
les
s
tr
a
ini
ng
da
ta,
it
dis
playe
d
po
or
r
obus
tnes
s
with
the
tr
a
ini
ng
pe
r
iod
o
f
the
c
las
s
if
ier
be
ing
higher
.
De
b
a
nd
C
ha
nda
[
23]
c
ompar
e
d
the
e
f
f
i
c
ienc
y
of
the
bidi
r
e
c
ti
ona
l
e
nc
ode
r
r
e
pr
e
s
e
ntatio
ns
f
r
om
tr
a
ns
f
or
mer
s
(
B
E
R
T
)
e
mbedding
model
to
pr
e
di
c
t
the
dis
a
s
ter
f
r
om
T
witt
e
r
da
ta
.
T
he
B
E
R
T
e
mbedding
model
wa
s
c
ont
r
a
s
ted
with
the
tr
a
dit
ional
c
ontext
ua
l
e
mbedding
models
,
whe
r
e
the
f
indi
ngs
s
howe
d
that
the
B
E
R
T
e
mbedding
model
a
c
hieve
d
s
upe
r
ior
r
e
s
ult
s
.
How
e
ve
r
,
tr
a
ns
f
or
mer
-
ba
s
e
d
ne
ur
a
l
ne
twor
k
mo
de
ls
li
ke
B
E
R
T
r
e
quir
e
d
a
s
igni
f
ica
nt
a
mount
of
memor
y
s
tor
a
ge
f
or
tr
a
ini
ng,
whe
r
e
in
the
p
r
e
diction
a
c
c
ur
a
c
y
dim
ini
s
he
d
whe
n
the
length
of
the
twe
e
ts
incr
e
a
s
e
d.
Na
s
uti
on
a
nd
S
e
ti
a
wa
n
[
24]
de
mons
tr
a
ted
a
hybr
id
method
c
a
ll
e
d
C
NN
a
nd
B
iL
S
T
M
f
or
e
nha
nc
ing
c
ybe
r
bull
ying
de
tec
ti
on
on
I
ndone
s
ian
T
witt
e
r
.
T
he
major
objec
ti
ve
wa
s
to
a
s
s
e
s
s
the
pr
e
s
e
ntation
ge
ne
r
a
ted
by
F
a
s
tT
e
xt
-
e
nha
nc
e
d
f
e
a
tur
e
e
xpa
ns
ion,
a
nd
hyb
r
id
C
NN
a
nd
B
iL
S
T
M
.
C
ons
e
que
ntl
y,
the
outcome
s
c
onf
ir
med
the
de
letion
of
twe
e
ts
c
ompr
is
ing
c
ybe
r
bull
y
ing
to
be
mor
e
pr
e
c
is
e
a
nd
on
tar
ge
t,
de
ve
lopi
ng
a
s
e
ns
e
of
s
e
c
ur
it
y
in
c
ons
umer
s
.
Ne
ve
r
thele
s
s
,
in
s
ome
c
a
s
e
s
,
ther
e
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2088
-
8708
I
nt
J
E
lec
&
C
omp
E
ng
,
Vol
.
15
,
No.
1
,
F
e
br
ua
r
y
20
25
:
700
-
7
10
702
wa
s
a
r
is
e
in
the
a
c
c
ur
a
c
y
whe
n
other
methods
s
uc
h
a
s
Allg
r
a
m
a
nd
Unig
r
a
m+B
igr
a
m
we
r
e
de
pl
oye
d,
a
s
oppos
e
d
to
the
pr
opos
e
d
model
.
F
r
om
the
ove
r
a
ll
a
na
lys
is
,
i
t
is
c
lea
r
ly
obs
e
r
ve
d
that
the
e
xis
ti
ng
methods
ha
ve
major
dr
a
wba
c
ks
s
uc
h
a
s
the
r
e
quir
e
ment
of
a
lar
ge
a
mount
of
da
ta
a
nd
ti
me
f
o
r
tr
a
ini
ng,
les
s
e
r
r
obus
tnes
s
of
the
s
ys
t
e
m,
high
c
omput
a
ti
ona
l
c
ompl
e
xit
y,
incons
ider
a
ti
on
of
m
ult
im
e
dia
-
ba
s
e
d
f
e
a
tur
e
s
,
a
nd
low
pr
e
diction
a
c
c
ur
a
c
ies
.
Als
o,
the
lar
ge
memor
y
s
tor
a
ge
de
mands
of
C
NN
or
B
iL
S
T
M
dur
ing
the
tr
a
ini
ng
s
tage
br
oke
the
s
c
a
labili
ty
a
nd
a
ppli
c
a
ti
on
of
thes
e
models
to
lar
ge
da
tas
e
ts
.
F
ur
ther
mo
r
e
,
thos
e
tec
hniques
that
only
uti
li
z
e
d
text
a
na
lys
is
we
r
e
unli
ke
ly
to
de
tec
t
thos
e
im
pli
c
it
c
lue
s
that
we
r
e
pa
r
t
o
f
mul
ti
media
c
ontent
s
uc
h
a
s
pict
ur
e
s
a
nd
videos
.
As
a
r
e
s
ult
,
the
a
f
or
e
mentioned
methods
we
r
e
not
ve
r
y
good
a
t
pinpoi
nti
ng
c
ybe
r
bull
y
ing
of
f
e
nde
r
s
,
ther
e
by
li
mi
ti
ng
the
s
ys
tem’
s
e
f
f
e
c
ti
ve
ne
s
s
.
On
the
other
ha
nd,
the
number
of
input
s
to
da
ta
inc
r
e
a
s
e
d
with
lar
ge
da
ta
s
e
ts
whic
h
we
r
e
r
e
leva
nt
to
s
ome
model
s
,
but
we
r
e
les
s
e
f
f
e
c
ti
ve
.
Although
thes
e
a
ppr
oa
c
he
s
ha
ve
made
be
tt
e
r
r
e
s
ult
s
in
de
tec
ti
ng
c
ybe
r
bull
ying,
a
n
e
f
f
icie
nt
a
nd
s
e
c
ur
e
s
ys
tem
i
s
e
s
s
e
nti
a
l
f
or
f
or
mul
a
ti
ng
e
f
f
e
c
ti
ve
,
s
c
a
lable
,
a
nd
a
ll
-
inclus
ive
s
tr
a
tegie
s
tha
t
s
top
c
ybe
r
bull
ying.
T
o
ove
r
c
ome
the
pr
e
vious
r
e
s
e
a
r
c
h’
s
dr
a
wba
c
ks
,
the
pr
opos
e
d
r
e
s
e
a
r
c
h
is
f
oc
us
e
d
on
de
ve
lopi
ng
a
r
obus
t
a
nd
e
f
f
icie
nt
DL
ba
s
e
d
L
S
T
M
-
C
L
R
de
nigr
a
ti
on
s
ys
tem
to
de
tec
ti
ng
c
ybe
r
bull
ying.
T
he
major
c
ontr
ibut
ions
of
thi
s
r
e
s
e
a
r
c
h
a
r
e
s
pe
c
if
ied
a
s
f
oll
ows
:
i)
I
de
nti
f
y
ing
c
ybe
r
bull
ying
on
s
oc
ial
media
by
de
ve
lopi
ng
a
de
nigr
a
ti
on
de
tec
ti
on
s
ys
tem
t
o
a
ddr
e
s
s
the
pr
e
s
s
ing
c
onc
e
r
ns
on
s
oc
ial
media
platf
or
m
;
ii
)
R
e
duc
ing
the
r
is
k
of
the
pr
opos
e
d
model
be
ing
tr
a
ppe
d
in
the
loca
l
mi
nim
a
by
c
ombi
ning
C
L
R
int
o
L
S
T
M
t
r
a
ini
ng
,
whic
h
f
ur
the
r
f
a
c
il
it
a
tes
the
f
a
s
ter
c
onve
r
ge
nc
e
by
e
nha
nc
ing
the
opti
mi
z
a
ti
on
e
f
f
icie
nc
y
a
nd
c
ybe
r
b
ull
ying
de
tec
ti
on
c
a
pa
bil
it
ies
;
a
nd
i
ii
)
T
he
lea
r
nin
g
r
a
te
of
L
S
T
M
is
inc
r
e
a
s
e
d
with
C
L
R
,
int
ur
n
inc
r
e
a
s
ing
the
model’
s
c
a
pa
c
it
y
a
nd
pr
ovid
ing
a
p
r
omi
s
ing
wa
y
to
tac
kle
onli
ne
c
ybe
r
bull
ying.
T
he
r
e
s
t
of
the
pa
pe
r
is
a
r
r
a
nge
d
a
s
f
oll
ows
:
t
he
pr
opos
e
d
methodology
is
e
xplaine
d
in
s
e
c
ti
on
2
.
T
he
pr
oc
e
s
s
of
L
S
T
M
with
c
yc
li
c
lea
r
ning
r
a
te
is
e
xplaine
d
in
s
e
c
ti
on
3,
whi
le
the
r
e
s
ult
s
a
nd
it
s
c
om
pa
r
is
ons
a
r
e
given
in
s
e
c
ti
on
4
.
At
las
t,
the
c
onc
lus
ion
of
thi
s
r
e
s
e
a
r
c
h
pa
pe
r
is
s
umm
a
r
ize
d
in
s
e
c
ti
on
5
.
2.
P
ROP
OS
E
D
M
E
T
HO
D
A
DL
-
ba
s
e
d
L
S
T
M
c
las
s
if
ier
with
C
L
R
is
pr
op
os
e
d
in
th
is
r
e
s
e
a
r
c
h
to
de
tec
t
the
de
nigr
a
ti
on
o
f
pe
r
s
ons
thr
ough
s
oc
ial
media
.
T
he
r
e
s
e
a
r
c
h
pr
opo
s
e
s
a
de
e
p
lea
r
ning
-
ba
s
e
d
L
S
T
M
c
las
s
if
ier
a
ugme
nted
with
C
L
R
to
e
f
f
e
c
ti
ve
ly
de
tec
t
ins
tanc
e
s
of
pe
r
s
on
de
nigr
a
ti
on
on
s
oc
ial
media
platf
or
ms
.
B
y
leve
r
a
ging
the
L
S
T
M
's
s
e
que
nti
a
l
modeling
c
a
pa
bil
it
ies
,
the
c
las
s
if
ier
a
na
lyze
s
textua
l
da
ta
in
c
ontext
a
nd
c
a
ptur
e
s
nua
nc
e
s
in
langua
ge
us
a
ge
.
I
ntegr
a
ti
ng
C
L
R
e
nha
nc
e
s
the
model's
tr
a
ini
ng
dyna
mi
c
s
,
potentially
im
pr
oving
i
ts
a
bil
it
y
to
ge
ne
r
a
li
z
e
a
c
r
os
s
va
r
ying
leve
ls
of
de
nigr
a
ti
on
e
xpr
e
s
s
ions
.
T
he
methodology
invol
ve
s
c
oll
e
c
ti
ng
T
witt
e
r
da
ta,
pr
e
-
pr
oc
e
s
s
ing
the
da
ta
,
f
e
a
tur
e
e
xtr
a
c
ti
on,
a
nd
c
las
s
if
ica
ti
on
us
ing
L
S
T
M
with
C
L
R
.
T
he
p
r
opos
e
d
f
r
a
mew
or
k’
s
f
low
diagr
a
m
is
r
e
pr
e
s
e
nted
in
F
igur
e
1.
F
igur
e
1.
F
low
diagr
a
m
o
f
the
pr
opos
e
d
de
nigr
a
ti
o
n
de
tec
ti
on
f
r
a
mew
or
k
2.
1.
T
wit
t
e
r
d
at
a
T
he
da
ta
is
c
oll
e
c
ted
f
r
om
the
T
wit
ter
f
o
r
de
tec
ti
ng
de
nigr
a
ti
on
a
nd
non
-
de
nigr
a
ti
on
e
ve
nts
.
T
he
da
tas
e
t
c
ons
i
s
ts
of
2
,
000
c
omm
e
nts
,
1
,
000
c
omm
e
nts
with
r
e
putation
r
umor
s
,
a
nd
1
,
000
c
omm
e
nts
with
non
-
r
e
putation
r
u
mor
s
.
T
he
c
oll
e
c
ted
da
ta
is
given
a
s
i
nput
to
the
pr
e
-
pr
oc
e
s
s
ing
s
tage
whic
h
is
c
lea
r
ly
e
xplaine
d
in
the
f
ol
lowing
s
ub
-
s
e
c
ti
ons
.
2.
2.
Dat
a
p
r
e
-
p
r
oc
e
s
s
in
g
I
t
is
a
n
int
e
gr
a
l
pa
r
t
of
the
na
tu
r
a
l
langua
ge
pr
o
c
e
s
s
ing
(
NL
P
)
to
r
e
buil
d
the
or
ig
inal
da
ta
int
o
a
mea
ningf
ul
f
or
mat
.
Va
r
ious
methods
s
uc
h
a
s
s
temmi
ng
a
nd
lemmatiza
ti
on,
a
nd
text
r
e
moval
a
r
e
c
a
r
r
ied
out
a
s
a
pa
r
t
of
pr
e
-
pr
oc
e
s
s
ing.
E
a
c
h
of
thes
e
tec
hniques
is
dis
c
us
s
e
d
be
low.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
E
lec
&
C
omp
E
ng
I
S
S
N:
2088
-
8708
De
nigr
ati
on
analys
is
of
T
w
it
ter
data
us
ing
c
y
c
li
c
lear
ning
r
ate
…
(
Suhas
B
har
adw
aj
R
ajendr
a
)
703
2.
2.
1.
S
t
e
m
m
in
g
an
d
l
e
m
m
a
t
izat
ion
T
he
twe
e
ts
do
not
c
ontain
a
s
tanda
r
d
f
or
mat
s
in
c
e
a
s
ingl
e
wor
d
with
the
s
a
me
mea
ning
c
a
n
be
e
xpr
e
s
s
e
d
dif
f
e
r
e
ntl
y.
T
his
pr
oblem
is
pr
e
ve
nted
b
y
us
ing
lemmatiza
ti
on
a
nd
s
temmi
ng
,
whe
r
e
the
pr
oc
e
s
s
of
lemmatiza
ti
on
is
to
c
onve
r
t
a
ll
wor
ds
int
o
their
di
c
ti
ona
r
y
-
ba
s
e
d
f
or
m
known
a
s
a
lemma.
S
temmi
n
g
is
the
pr
oc
e
s
s
of
r
e
mov
ing
the
wor
d's
las
t
f
e
w
letter
s
to
obtain
a
mea
ningf
ul
ba
s
e
.
T
his
c
ontr
ibut
e
s
to
e
nha
nc
ing
the
pr
e
diction
a
c
c
ur
a
c
y
of
the
s
ys
tem.
2.
2.
2.
S
t
op
wor
d
/t
e
xt
r
e
m
oval
W
or
ds
s
uc
h
a
s
“
to
,
”
“
me
,
”
“
my
,
”
a
nd
“
our
s
,
”
a
nd
s
o
on,
a
r
e
known
a
s
s
top
wor
ds
whic
h
do
not
pr
ovide
e
nough
s
igni
f
ica
nc
e
to
the
s
e
ntenc
e
s
a
nd
c
a
us
e
nois
e
in
the
da
tas
e
t.
T
o
r
e
move
thes
e
wor
ds
,
a
P
ython
pa
c
ka
ge
li
br
a
r
y
c
a
ll
e
d
s
topwor
d
is
uti
li
z
e
d
[
25]
.
T
he
e
xtr
a
c
ti
on
of
r
e
leva
nt
f
e
a
tur
e
s
is
pe
r
f
or
med
a
f
ter
the
r
e
moval
of
s
topwor
ds
whic
h
is
given
a
s
input
to
th
e
e
xtr
a
c
ti
on
pr
oc
e
s
s
whe
r
e
global
ve
c
tor
s
a
r
e
a
ppli
e
d.
2.
3.
F
e
a
t
u
r
e
e
xt
r
ac
t
ion
Onc
e
pr
e
-
pr
oc
e
s
s
ing
is
pe
r
f
or
med,
a
p
r
e
-
tr
a
ined
wor
d
e
mbedding
model
c
a
ll
e
d
Glove
(
2
b
il
li
on
twe
e
ts
,
27
bil
li
on
tokens
,
a
nd
1.
2
voc
a
bular
ies
)
i
s
de
ployed
f
or
ge
ne
r
a
ti
ng
a
wo
r
d
ve
c
tor
mat
r
ix
with
200
dim
e
ns
ions
.
Globa
l
ve
c
tor
(
Glove
)
is
a
2ve
c
-
ba
s
e
d
wor
d
r
e
pr
e
s
e
ntation
that
a
ids
in
the
e
f
f
icie
nt
lea
r
ning
of
wor
d
e
mbeddings
f
r
om
textua
l
doc
uments
.
T
he
glove
model
is
c
ombi
ne
d
with
T
F
-
I
DF
to
de
ter
mi
ne
the
pr
opos
e
d
a
ppr
oa
c
h’
s
e
f
f
e
c
ti
ve
ne
s
s
.
T
F
-
I
DF
pr
e
s
e
nts
a
r
e
lative
f
r
e
que
nc
y
o
f
the
wo
r
d
that
is
pr
e
s
e
nt
in
the
textua
l
doc
ument
a
nd
the
I
DF
is
us
e
d
in
the
pr
oc
e
s
s
of
s
c
a
li
ng
wi
th
the
tot
a
l
c
ount
of
doc
uments
.
E
a
c
h
input
wor
d
is
r
e
pr
e
s
e
nted
a
s
a
token
a
nd
e
ve
r
y
indi
vidual
wor
d
is
tr
a
ns
f
or
med
to
a
wo
r
d
ve
c
tor
of
d
im
e
ns
ion
.
S
o,
the
dim
e
ns
ionalit
ies
of
e
a
c
h
wor
d
ve
c
tor
a
r
e
r
e
pr
e
s
e
nted
a
s
a
nd
the
input
text
mat
r
ix
c
r
e
a
ted
is
de
noted
a
s
=
{
1
,
2
,
…
,
}
∈
×
.
E
qua
ti
on
(
1)
r
e
pr
e
s
e
nts
the
f
e
a
tur
e
ve
c
tor
f
or
doc
ument
c
onc
a
tena
ti
on
a
nd
wor
d
e
mbeddings
.
T
he
text
r
e
pr
e
s
e
ntation
is
e
nha
nc
e
d
by
c
ombi
ning
the
pr
e
-
tr
a
ined
Glove
a
nd
T
F
-
I
DF
we
ighi
ng
whic
h
is
numer
ica
ll
y
r
e
pr
e
s
e
nted
in
(
2)
:
=
1
⊕
2
⊕
3
…
⊕
−
1
⊕
(
1)
=
,
×
(
2)
w
he
r
e
,
the
wor
d
ve
c
tor
matr
ix
is
s
tate
d
a
s
,
a
nd
the
we
ight
e
d
va
lue
of
the
doc
ument
a
nd
the
ter
m
is
s
tate
d
a
s
a
nd
,
r
e
s
pe
c
ti
ve
ly.
T
he
s
ugge
s
ted
a
ppr
oa
c
h
s
olves
dim
e
ns
ionalit
y
is
s
ue
s
r
e
late
d
to
high
-
dim
e
ns
ional
matr
ix.
Onc
e
the
text
r
e
pr
e
s
e
ntation
f
r
om
the
wo
r
d
r
e
pr
e
s
e
ntation
laye
r
is
r
e
c
e
ived,
the
Ga
us
s
ian
nois
e
a
nd
Ga
us
s
ian
dr
opout
a
r
e
ge
ne
r
a
ted.
T
he
Ga
us
s
ian
nois
e
a
nd
Ga
us
s
ian
dr
opout
pr
oc
e
s
s
e
s
a
r
e
e
mpl
oye
d
to
r
e
gular
ize
the
model
by
r
e
nde
r
ing
it
les
s
s
us
c
e
pti
ble
to
ove
r
f
it
ti
ng
.
T
he
L
S
T
M
c
las
s
if
ier
is
tr
a
ined
with
the
C
L
R
f
or
e
f
f
e
c
ti
ve
c
las
s
if
ica
ti
on,
whic
h
is
de
s
c
r
ibed
in
the
f
oll
owing
s
e
c
ti
on
[
26]
.
3.
CL
ASS
I
F
I
CA
T
I
ON
OF
S
E
NT
I
M
E
NT
S
USI
NG
L
ONG
S
HO
RT
-
T
E
RM
M
E
M
ORY
WI
T
H
CY
CL
I
C
L
E
AR
NI
NG
RA
T
E
T
he
e
xtr
a
c
ted
f
e
a
tur
e
s
a
r
e
c
las
s
if
ied
us
ing
L
S
T
M
ne
twor
k,
whe
r
e
the
input
a
nd
output
f
e
a
tur
e
s
a
r
e
c
onc
a
tena
ted
f
or
the
r
e
gular
iza
ti
on
o
f
e
ve
r
y
lay
e
r
.
Unlike
othe
r
c
las
s
if
ier
s
,
L
S
T
M
ha
s
the
a
dva
ntage
of
ove
r
c
omi
ng
the
ove
r
f
it
ti
ng
p
r
oblem,
a
nd
the
s
e
lec
ted
f
e
a
tur
e
s
a
r
e
f
e
d
to
the
top
laye
r
to
e
nha
nc
e
the
f
e
a
tur
e
s
.
T
he
L
S
T
M
ha
s
mul
t
ipl
ica
ti
ve
c
e
ll
s
f
o
r
med
by
tempor
a
l
a
nd
mul
ti
pl
ica
ti
ve
unit
s
made
up
of
va
r
ious
c
ha
r
a
c
ter
s
that
ha
ndle
da
ta
s
tr
e
a
m
in
the
memor
y
block.
F
igur
e
2
il
lus
tr
a
tes
the
a
r
c
hit
e
c
tur
e
of
L
S
T
M
model.
T
hr
e
e
ga
tes
na
mely,
f
or
ge
t
ga
te
,
input
ga
te
,
a
n
d
output
ga
te
play
a
s
igni
f
ica
nt
r
ole
in
s
tor
ing
t
he
memor
y
c
omponents
a
nd
r
e
gulate
s
the
inf
or
mation
f
low.
T
he
output
ga
te
p
r
ovides
the
f
inal
output
,
the
f
o
r
ge
t
ga
te
s
e
lec
ts
whic
h
da
ta
to
e
r
a
s
e
f
r
om
the
c
e
ll
,
a
nd
the
input
ga
te
s
e
lec
ts
whic
h
da
ta
to
a
dd
to
th
e
c
e
ll
s
tate
.
T
he
s
e
ga
tes
he
lp
in
t
r
a
ns
mi
tt
ing
the
da
ta
a
nd
s
a
ve
the
memor
y
c
omponents
.
T
he
pr
oc
e
s
s
ing
of
n
ode
s
in
L
S
T
M
with
th
r
e
e
ga
tes
is
given
th
r
ough
(
3
)
to
(
8)
:
=
(
.
[
ℎ
−
1
,
]
+
)
(
3)
=
(
.
[
ℎ
−
1
,
]
+
)
(
4)
̃
=
ℎ
(
.
[
ℎ
−
1
,
]
+
)
(
5)
=
∗
−
1
+
∗
̃
(
6)
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2088
-
8708
I
nt
J
E
lec
&
C
omp
E
ng
,
Vol
.
15
,
No.
1
,
F
e
br
ua
r
y
20
25
:
700
-
7
10
704
=
(
0
.
[
ℎ
−
1
,
]
+
0
(
7)
ℎ
=
∗
ta
nh
(
)
(
8)
w
he
r
e
,
de
notes
f
or
ge
t
ga
te,
de
notes
the
s
igm
oid
f
u
nc
ti
on,
ℎ
−
1
de
notes
the
hidden
s
tate
of
the
pr
ior
laye
r
,
is
the
input
of
the
c
ur
r
e
nt
laye
r
,
s
tate
s
the
we
ight
s
,
a
nd
de
notes
the
bias
s
tate
.
is
the
input
ga
te,
is
the
c
e
ll
s
tate
in
the
ne
xt
ye
a
r
,
̃
is
the
int
e
r
media
te
tempor
a
r
y
s
tate
,
−
1
is
c
e
ll
s
tate
pr
e
s
e
nt
in
the
pr
e
c
e
di
ng
laye
r
.
is
the
output
laye
r
a
nd
ℎ
is
the
hidden
s
tate
of
the
ne
xt
laye
r
.
T
he
pur
pos
e
of
the
pr
e
s
e
nt
method
is
to
tr
a
in
L
S
T
M
us
ing
a
n
L
R
that
c
yc
les
thr
ough
e
ve
r
y
ba
tch.
T
o
c
a
lcula
te
the
L
R
of
L
S
T
M
,
the
C
L
R
is
e
mpl
oye
d
t
o
va
r
y
the
L
R
dur
ing
tr
a
ini
ng
in
a
c
yc
li
c
manne
r
,
t
ypica
ll
y
by
os
c
il
lating
it
be
twe
e
n
a
lowe
r
a
nd
uppe
r
bound.
T
his
he
lps
in
a
c
hieving
f
a
s
ter
c
onve
r
ge
nc
e
a
nd
po
tentially
f
indi
ng
be
tt
e
r
mi
nim
a
in
the
los
s
s
c
e
na
r
io.
C
L
R
a
ll
ows
the
L
R
to
va
r
y
dur
ing
t
r
a
ini
ng,
whic
h
he
lps
th
e
model
c
onve
r
ge
f
a
s
ter
.
B
y
us
ing
a
higher
L
R
dur
ing
c
e
r
tain
pha
s
e
s
of
tr
a
ini
ng,
the
model
make
s
lar
ge
r
upda
tes
to
it
s
pa
r
a
mete
r
s
,
potentially
s
pe
e
ding
up
c
onve
r
ge
nc
e
.
Highe
r
L
R
s
dur
ing
c
yc
les
of
tr
a
ini
ng
he
lp
e
xp
lor
e
the
pa
r
a
mete
r
s
pa
c
e
mor
e
br
oa
dly
,
whi
le
lowe
r
L
R
s
a
ll
ow
f
or
mor
e
f
ine
-
gr
a
ined
a
djus
tm
e
nts
a
nd
e
xploi
tation
of
pr
omi
s
ing
r
e
gions
.
T
he
C
L
R
c
ons
is
ts
of
pa
r
a
mete
r
s
s
uc
h
a
s
b
a
tch
s
ize
,
s
tep
s
iz
e
,
ba
tch
or
it
e
r
a
ti
on,
c
yc
le,
a
nd
_
a
nd
_
.
C
ons
ider
ing
a
s
e
poc
h
number
f
r
om
1
to
e
poc
hs
,
e
a
c
h
e
poc
h's
it
e
r
a
ti
on
is
r
e
pr
e
s
e
nted
by
the
s
ymbol
t.
T
he
s
e
que
nc
e
{
,
}
,
1
ge
ne
r
a
ted
a
s
the
output
f
r
om
the
tr
a
ini
ng
is
s
tate
d
a
s
given
be
low
:
{
1
,
1
,
1
,
2
,
…
.
1
,
}
whe
n
=
1
,
1
e
poc
h
{
2
,
1
,
2
,
2
,
…
.
2
,
}
whe
n
=
2
,
2
e
poc
h
{
1
,
1
,
1
,
2
,
…
.
1
,
}
whe
n
=
,
ℎ
e
poc
h
w
he
r
e
,
de
notes
the
mos
t
r
e
c
e
nt
e
poc
h
.
T
he
s
e
que
nc
e
that
is
f
or
med
a
f
te
r
e
ve
r
y
e
poc
h
is
.
whe
r
e
in
ha
s
the
va
lue
.
T
he
c
u
r
r
e
nt
e
poc
h’
s
f
ir
s
t
i
ter
a
ti
on
make
s
us
e
of
e
xpr
e
s
s
ions
given
in
(
9
)
a
nd
(
10)
.
,
=
−
1
,
(
9)
,
+
1
=
,
−
∇
,
(
,
)
(
10)
I
n
(
9)
,
de
notes
the
ne
xt
ne
w
e
poc
h’
s
f
ir
s
t
it
e
r
a
ti
on
with
index
,
a
nd
−
1
s
tate
s
the
pr
ior
e
poc
h.
S
im
il
a
r
ly,
in
(
10)
,
a
nd
∇
,
de
note
the
gr
a
dients
of
the
ne
twor
k
laye
r
with
r
e
s
pe
c
t
to
e
poc
hs
a
nd
it
e
r
a
ti
on.
+
1
is
a
n
i
ter
a
ti
on
o
f
the
ne
xt
ne
w
e
poc
h
with
the
L
R
.
T
he
L
R
f
or
the
s
ubs
e
que
nt
it
e
r
a
ti
on
a
tt
a
ined
ba
s
e
d
on
the
pr
io
r
it
e
r
a
ti
on
is
r
e
p
r
e
s
e
nted
in
(
11
)
.
T
he
a
ve
r
a
ge
s
tocha
s
ti
c
gr
a
dient
(
s
g)
of
a
n
e
poc
h
is
given
by
(
12
)
:
+
1
=
‖
‖
2
|
|
(
11)
,
+
1
=
(
1
−
)
,
+
∇
(
,
)
(
12)
w
he
r
e
,
=
1
,
2
,
3
,
…
.
,
a
nd
r
e
pr
e
s
e
nt
the
pr
e
-
de
f
ined
s
moot
hing
f
a
c
tor
s
that
r
e
gulate
the
de
gr
e
e
of
de
c
a
y.
T
he
dif
f
e
r
e
nc
e
in
the
c
ur
r
e
nt
g
r
a
dients
a
nd
pr
ior
e
poc
hs
is
r
e
pr
e
s
e
nt
e
d
in
(
13)
.
T
o
ge
t
a
ne
w
L
R
f
or
the
c
ur
r
e
nt
e
poc
h,
the
a
ve
r
a
ge
of
a
ll
L
R
s
f
r
om
the
pr
i
or
e
poc
h
is
us
e
d.
T
he
a
ve
r
a
ge
f
indi
ng
f
or
mul
a
is
given
in
(
14)
.
T
he
ne
w
L
R
f
or
the
it
e
r
a
ti
on
of
the
ne
xt
e
poc
h
is
given
in
(
15)
.
=
,
−
−
1
,
(
13)
=
1
∑
=
1
(
14)
=
,
>
,
ℎ
(
15)
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
E
lec
&
C
omp
E
ng
I
S
S
N:
2088
-
8708
De
nigr
ati
on
analys
is
of
T
w
it
ter
data
us
ing
c
y
c
li
c
lear
ning
r
ate
…
(
Suhas
B
har
adw
aj
R
ajendr
a
)
705
F
ur
ther
mor
e
,
a
nd
a
r
e
c
ompar
e
d
to
de
ter
mi
ne
the
va
lue
f
or
the
f
ol
lowing
e
poc
h
in
the
L
R
s
c
he
duler
.
T
he
s
a
me
is
take
n
int
o
a
c
c
ount
if
is
lar
ge
r
than
,
e
ls
e
is
take
n
int
o
a
c
c
ount
f
or
the
f
o
ll
owing
e
poc
h
[
27]
.
T
he
va
lues
obtaine
d
a
c
c
ount
f
or
a
nythi
ng
mi
nim
a
l
.
As
a
r
e
s
ult
,
tr
a
ini
ng
is
pe
r
f
or
med
mo
r
e
quickly
with
the
f
oll
owing
pa
r
a
m
e
ter
s
.
−
B
a
tch
s
ize
:
B
a
tch
s
ize
de
ter
mi
ne
s
the
c
ount
of
tr
a
i
ning
s
a
mpl
e
s
f
or
us
e
in
a
n
it
e
r
a
ti
on
whic
h
is
c
ons
i
de
r
e
d
a
s
128
in
thi
s
r
e
s
e
a
r
c
h.
−
S
tep
s
ize
:
S
tep
s
ize
de
ter
mi
ne
s
the
it
e
r
a
ti
on
c
oun
t
to
c
ompl
e
te
ha
l
f
the
c
yc
le
whic
h
is
c
ons
ider
e
d
a
s
1
he
r
e
.
−
B
a
tch
or
it
e
r
a
ti
on:
I
t
de
ter
mi
ne
s
the
s
e
t
of
s
a
mpl
e
s
with
a
ba
tch
s
ize
of
128
with
100
it
e
r
a
ti
ons
/batc
h
e
s
to
c
ompl
e
te
one
e
poc
h.
−
_
:
T
he
ba
s
e
L
R
or
the
mi
nim
um
L
R
c
ons
ider
e
d
in
t
his
r
e
s
e
a
r
c
h
is
0.
00001.
−
_
:
T
he
maxi
mum
L
R
c
ons
ider
e
d
is
0.
05
.
T
he
pe
r
f
o
r
manc
e
of
the
pr
opos
e
d
L
S
T
M
-
C
L
R
a
c
hieve
s
be
tt
e
r
c
las
s
if
ica
ti
on
a
c
c
ur
a
c
y.
T
he
e
va
luation
of
the
pr
opos
e
d
model
is
pe
r
f
or
med
on
both
tr
a
ini
ng
a
nd
tes
ti
ng
da
ta.
T
he
a
c
c
ur
a
c
y
a
nd
lo
s
s
of
the
pr
opos
e
d
L
S
T
M
-
C
L
R
model
on
tr
a
ined
da
ta
giv
e
s
r
is
e
to
s
upe
r
ior
va
lues
,
a
s
obs
e
r
ve
d
in
F
igur
e
s
3
a
nd
4
r
e
s
pe
c
ti
ve
ly.
T
he
ps
e
udoc
ode
of
pr
opos
e
d
c
las
s
if
ier
is
mentioned
a
s
Algo
r
it
hm
1
.
F
igur
e
2.
Ar
c
hit
e
c
tur
e
of
L
S
T
M
F
igur
e
3.
Ac
c
ur
a
c
y
e
va
luation
f
or
t
r
a
ini
ng
a
nd
tes
ti
ng
da
ta
F
igur
e
4.
L
os
s
e
va
luation
of
t
r
a
ini
ng
a
nd
tes
ti
ng
da
ta
Algor
it
hm
1
.
P
s
e
udoc
ode
f
o
r
pr
opos
e
d
c
las
s
if
ier
Input: Training dataset
Output
Normalize the dataset
Select the training feature size
for n epochs and batch size
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2088
-
8708
I
nt
J
E
lec
&
C
omp
E
ng
,
Vol
.
15
,
No.
1
,
F
e
br
ua
r
y
20
25
:
700
-
7
10
706
Train the
network
(LSTM)
end
for
Run Predictions using LSTM
Calculate the loss function
if cyclic_lr (epoch, current_lr)
base_lr=0.00001 # Initial LR
max_lr=0.05 # Maximum LR
step_size=1 # Number of epochs in half a cycle
cycle=floor (1+epoch/(2*step_size))
x=abs (epoch/step_size
-
2*cycle+1)
new_lr=base_lr+(max_lr
-
base_lr)*max(0, (1
-
x))
Return new_lr
End if
Train model with CLR
for epoch in range(epochs):
current_lr=cyclic_lr (epoch, model.optimizer.lr. numpy ())
model.optimizer.lr. assign(current_lr)
fitting
the model,
the training and testing values
Evaluating the parameter
End for
Return model
At
f
ir
s
t,
the
L
S
T
M
model’
s
a
r
c
hit
e
c
tu
r
e
is
ini
ti
a
li
z
e
d
by
s
pe
c
if
ying
the
input
s
ha
pe
,
hidden
laye
r
s
,
a
nd
output
unit
s
.
T
he
C
L
R
pa
r
a
mete
r
s
a
r
e
then
s
p
e
c
if
ied
a
s
ba
s
e
L
R
(
_
)
,
max
L
R
(
_
)
,
a
nd
s
tep
s
ize
.
T
he
L
R
s
c
he
duler
is
then
ini
ti
a
li
z
e
d
with
C
L
R
:
=
(
_
,
_
,
_
)
.
T
he
L
S
T
M
is
then
c
ompi
led
with
the
C
L
R
s
c
he
duler
.
Dur
ing
e
a
c
h
it
e
r
a
ti
on,
the
C
L
R
upda
tes
the
L
R
,
ther
e
by
he
lpi
ng
the
L
S
T
M
c
onve
r
ge
f
a
s
ter
a
nd
e
xplo
r
e
a
br
oa
de
r
pa
r
a
mete
r
s
pa
c
e
.
Af
ter
tr
a
ini
ng
is
c
ompl
e
ted,
the
L
S
T
M
model
is
now
f
ine
-
tuned
us
ing
C
L
R
.
T
he
model
is
then
e
va
luate
d
on
the
va
li
da
ti
o
n
s
e
t
f
or
pe
r
f
or
manc
e
a
s
s
e
s
s
ment.
At
las
t,
the
tr
a
ined
L
S
T
M
=
C
L
R
model
is
r
e
a
dy
f
or
making
pr
e
dictions
on
ne
w
s
e
que
nc
e
s
.
C
L
R
a
djus
ts
the
L
R
dur
ing
t
r
a
ini
ng
in
a
c
yc
li
c
pa
tt
e
r
n,
e
nha
nc
ing
the
L
S
T
M
c
onve
r
ge
n
c
e
.
T
he
ba
s
e
a
nd
max
L
R
s
a
r
e
s
pe
c
if
ied,
a
ll
owing
the
model
to
e
xplor
e
a
wide
r
r
a
nge
of
pa
r
a
mete
r
s
.
T
he
L
R
s
c
he
duler
is
incor
por
a
ted
int
o
the
L
S
T
M
model’
s
tr
a
ini
ng
pr
oc
e
s
s
.
F
ur
ther
,
the
C
L
R
a
lt
e
r
na
tes
the
L
R
be
twe
e
n
the
s
pe
c
if
ied
r
a
nge
s
dur
ing
the
tr
a
ini
ng
c
yc
les
.
T
he
os
c
il
lating
L
R
he
lps
e
s
c
a
pe
loca
l
mi
nim
a
a
nd
a
c
c
e
ler
a
tes
c
onve
r
ge
nc
e
.
T
he
L
S
T
M
model
is
u
pda
ted
it
e
r
a
ti
ve
ly,
a
djus
ti
ng
we
ight
s
ba
s
e
d
on
the
C
L
R
s
c
he
dule.
T
he
C
L
R
-
e
nh
a
nc
e
d
tr
a
ini
ng
pr
oc
e
s
s
ba
lanc
e
s
e
xplor
a
ti
on
a
nd
e
xploi
tation
pha
s
e
s
.
Af
ter
t
r
a
ini
ng,
the
L
S
T
M
-
C
L
R
model
is
e
xpe
c
ted
to
ha
ve
e
nha
nc
e
d
ge
ne
r
a
li
z
a
ti
on
a
nd
im
pr
ove
d
pe
r
f
o
r
ma
nc
e
.
T
he
e
va
luation
on
a
s
e
pa
r
a
te
va
li
da
ti
on
s
e
t
e
ns
ur
e
s
the
model's
e
f
f
e
c
ti
ve
ne
s
s
.
F
inally,
the
tr
a
ined
L
S
T
M
-
C
L
R
model
is
r
e
a
dy
f
o
r
de
ploym
e
nt
a
nd
f
o
r
making
pr
e
dictions
on
ne
w
s
e
que
nti
a
l
da
ta.
4.
RE
S
UL
T
S
AN
D
DI
S
CU
S
S
I
ON
T
he
pr
opos
e
d
C
L
R
-
L
S
T
M
model
is
s
im
ulate
d
on
Ana
c
onda
Na
vigator
3.
5
.
2.
0
(
64
-
bit
)
,
P
ython
3
.
7,
OS:
W
indows
10
(
64
-
bit
)
,
P
r
oc
e
s
s
or
:
int
e
l
c
or
e
i7,
R
AM
:
16
GB
.
T
he
pe
r
f
or
manc
e
is
e
va
luate
d
on
the
f
oll
owing
metr
ics
,
a
s
given
in
(
16)
to
(
19)
.
T
he
pe
r
f
or
manc
e
is
mea
s
ur
e
d
in
ter
ms
of
t
r
ue
p
os
it
ive
(
T
P
)
,
tr
ue
ne
ga
ti
ve
(
T
N)
,
f
a
ls
e
pos
it
ive
(
F
P
)
,
a
nd
f
a
ls
e
ne
ga
ti
ve
(
F
N)
whic
h
a
r
e
the
pos
it
ive
a
nd
ne
ga
ti
ve
c
las
s
e
s
of
t
r
ue
a
nd
f
a
ls
e
pr
e
dictions
.
=
(
+
)
(
16)
=
(
=
)
(
17)
=
(
18)
−
=
2
∗
∗
+
(
19)
4.
1.
Qu
an
t
it
a
t
ive
an
alys
is
T
he
pe
r
f
o
r
manc
e
of
L
S
T
M
with
a
nd
without
C
L
R
is
a
na
lyze
d
in
thi
s
s
e
c
ti
on
r
e
s
pe
c
ti
ve
to
the
tr
a
dit
ional
c
las
s
if
ier
s
s
uc
h
a
s
C
NN
,
R
NN
,
a
nd
GR
U,
a
s
dis
playe
d
in
T
a
bles
1
a
nd
2.
T
he
s
e
tabula
r
v
a
lues
a
r
e
gr
a
phica
ll
y
r
e
pr
e
s
e
nted
in
F
igu
r
e
s
5
a
nd
6.
F
r
o
m
T
a
ble
s
1
a
nd
2
,
it
is
c
lea
r
that
L
S
T
M
a
c
hiev
e
s
be
tt
e
r
c
las
s
if
ica
ti
on
with
C
L
R
whe
n
c
ompar
e
d
to
pe
r
f
or
mi
ng
without
int
e
gr
a
ti
on
of
C
L
R
.
T
he
L
S
T
M
model
de
mons
tr
a
tes
the
highes
t
a
c
c
ur
a
c
y,
pr
e
c
is
ion,
r
e
c
a
ll
,
a
nd
F
-
mea
s
ur
e
a
mong
the
c
las
s
if
ier
s
,
s
howc
a
s
ing
it
s
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
E
lec
&
C
omp
E
ng
I
S
S
N:
2088
-
8708
De
nigr
ati
on
analys
is
of
T
w
it
ter
data
us
ing
c
y
c
li
c
lear
ning
r
ate
…
(
Suhas
B
har
adw
aj
R
ajendr
a
)
707
e
f
f
ica
c
y
in
ha
ndli
ng
s
e
que
nti
a
l
da
ta
with
long
-
ter
m
de
pe
nde
nc
ies
.
R
NN
a
nd
G
R
U
pe
r
f
or
m
pr
e
f
e
r
a
bl
y
,
but
lag
be
hind
L
S
T
M
due
to
their
li
mi
tations
in
c
a
ptu
r
in
g
int
r
ica
te
tempor
a
l
r
e
lations
hips
.
As
s
e
e
n
in
T
a
b
le
2,
the
GR
U
c
las
s
if
ier
de
mons
tr
a
tes
be
tt
e
r
pe
r
f
or
manc
e
,
while
R
NN
a
nd
C
NN
f
a
il
in
pe
r
f
o
r
manc
e
be
c
a
us
e
of
GR
U
a
nd
L
S
T
M
,
indi
c
a
ti
ng
their
c
ompar
a
ti
ve
li
m
it
a
ti
on
s
in
c
a
ptur
ing
pa
tt
e
r
ns
withi
n
the
da
tas
e
t.
Ove
r
a
ll
,
the
table
s
ugge
s
ts
that
L
S
T
M
is
e
f
f
e
c
ti
ve
f
or
the
de
nigr
a
ti
on
identif
ica
ti
on
tas
k
a
nd
s
howc
a
s
ing
their
r
obus
tnes
s
in
ha
ndli
ng
s
e
que
nti
a
l
da
ta
with
c
ompl
e
x
de
pe
nde
nc
i
e
s
.
T
a
ble
1.
P
e
r
f
o
r
manc
e
a
na
lys
is
of
c
las
s
if
ier
s
withou
t
C
L
R
C
la
s
s
if
ie
r
s
A
c
c
ur
a
c
y (
%
)
P
r
e
c
is
io
n (
%
)
R
e
c
a
ll
(
%
)
F
-
M
e
a
s
ur
e
(
%
)
C
N
N
71.35
72.73
75.65
74.55
R
N
N
73.57
74.23
77.48
75.65
G
R
U
76.23
76.34
78.44
77.67
L
S
T
M
78.45
79.12
80.54
78.63
T
a
ble
2.
P
e
r
f
o
r
manc
e
a
na
lys
is
of
c
las
s
if
ier
s
with
C
L
R
C
la
s
s
if
ie
r
s
A
c
c
ur
a
c
y (
%
)
P
r
e
c
is
io
n (
%
)
R
e
c
a
ll
(
%
)
F
-
M
e
a
s
ur
e
(
%
)
C
N
N
74.27
73.27
77.44
76.29
R
N
N
77.73
75.27
79.42
77.94
G
R
U
79.56
77.41
80.56
79.85
L
S
T
M
82.00
80.00
83.00
81.00
F
igur
e
5.
P
e
r
f
or
manc
e
a
na
lys
is
of
c
las
s
if
ier
s
without
C
L
R
F
igur
e
6.
P
e
r
f
or
manc
e
a
na
lys
is
of
c
las
s
if
ier
s
with
C
L
R
4.
2.
Com
p
ar
a
t
ive
an
alys
is
T
he
L
S
T
M
-
C
L
R
pe
r
f
or
manc
e
is
c
ontr
a
s
ted
with
t
he
tr
a
dit
ional
methods
s
uc
h
a
s
M
L
P
[
22]
,
B
E
R
T
[
23]
a
nd
C
NN
-
B
iL
S
T
M
[
24]
,
a
s
s
hown
in
T
a
bl
e
3.
F
r
om
T
a
ble
3,
it
is
obs
e
r
ve
d
that
the
L
S
T
M
-
C
L
R
a
c
hieve
s
be
tt
e
r
c
las
s
if
ica
ti
on
a
c
c
ur
a
c
y
with
82%
a
s
the
e
xis
ti
ng
methods
ha
ve
li
mi
tat
ions
.
M
L
P
[
22]
ha
s
li
mi
tations
of
les
s
tr
a
ini
ng
da
ta,
poor
r
obus
tnes
s
of
the
s
ys
tem,
a
nd
longer
tr
a
ini
ng
pe
r
iod
o
f
the
c
las
s
if
ier
.
T
he
B
E
R
T
[
23]
ha
s
li
mi
tations
of
dim
ini
s
hing
pr
e
diction
a
c
c
ur
a
c
y
whe
n
the
length
of
the
twe
e
ts
is
i
nc
r
e
a
s
e
d.
T
he
s
e
li
mi
tations
a
r
e
ove
r
c
ome
in
thi
s
r
e
s
e
a
r
c
h
by
int
r
oduc
ing
C
L
R
to
the
L
S
T
M
f
or
f
a
s
t
a
nd
a
c
c
ur
a
te
tr
a
ini
ng.
T
a
ble
3.
C
ompar
a
ti
ve
a
na
lys
is
of
p
r
opos
e
d
L
S
T
M
-
C
L
R
with
e
xis
ti
ng
a
ppr
oa
c
he
s
C
la
s
s
if
ie
r
s
A
c
c
ur
a
c
y (
%
)
P
r
e
c
is
io
n (
%
)
R
e
c
a
ll
(
%
)
F
-
M
e
a
s
ur
e
(
%
)
M
L
P
(
a
c
c
ur
a
c
y on mi
xe
d da
ta
s
e
t
s
)
[
22]
88
-
-
-
B
E
R
T
[
23]
C
N
N
-
B
iL
S
T
M
[
24]
79
80.55
-
-
-
-
74
-
L
S
T
M
-
C
L
R
82
80
83
81
M
L
P
's
s
hor
tcomings
include
r
e
li
a
nc
e
on
li
mi
ted
tr
a
ini
ng
da
ta
a
nd
e
xtende
d
tr
a
ini
ng
pe
r
iods
,
wi
th
a
n
e
mphas
is
on
the
ne
e
d
f
or
mor
e
s
ophis
ti
c
a
ted
models
c
a
pa
ble
of
ha
ndli
ng
the
dyna
mi
c
a
nd
diver
s
e
na
tur
e
of
s
oc
ial
media
c
ontent.
W
hil
e
B
E
R
T
ini
ti
a
l
ly
ke
e
ps
pr
omi
s
e
in
unde
r
s
tanding
c
ontextua
l
ter
ms
in
lang
ua
ge
,
it
s
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2088
-
8708
I
nt
J
E
lec
&
C
omp
E
ng
,
Vol
.
15
,
No.
1
,
F
e
br
ua
r
y
20
25
:
700
-
7
10
708
pe
r
f
or
manc
e
f
a
il
s
with
longer
twe
e
ts
,
making
it
dif
f
icult
f
o
r
a
ppli
c
a
ti
on
in
r
e
a
l
-
wor
ld
s
c
e
na
r
ios
w
he
r
e
text
lengths
va
r
y
wide
ly.
T
he
p
r
opos
e
d
method
L
S
T
M
-
C
L
R
a
ddr
e
s
s
e
s
thes
e
c
ha
ll
e
nge
s
by
f
a
c
il
it
a
ti
ng
f
a
s
ter
a
nd
mor
e
a
c
c
ur
a
te
tr
a
ini
ng
.
C
L
R
opti
mi
z
e
s
the
lea
r
n
ing
r
a
te
du
r
ing
t
r
a
ini
ng,
a
ll
owing
the
model
to
na
vigate
c
ompl
e
x
da
ta
dis
tr
ibut
ions
mor
e
e
f
f
icie
ntl
y
a
nd
a
da
pt
to
va
r
ying
twe
e
t
lengths
.
C
ons
e
que
ntl
y,
L
S
T
M
-
C
L
R
not
only
ove
r
c
omes
the
li
mi
tations
o
f
M
L
P
a
nd
B
E
R
T
,
but
a
ls
o
s
howc
a
s
e
s
the
potential
to
e
nh
a
nc
e
the
e
f
f
ica
c
y
of
s
e
nti
ment
a
na
lys
is
a
nd
de
nigr
a
ti
on
de
te
c
ti
on
in
s
oc
ial
media
c
ontexts
.
4.
3.
Dis
c
u
s
s
ion
T
his
r
e
s
e
a
r
c
h
a
im
s
to
a
ddr
e
s
s
the
is
s
ue
of
de
tec
ti
n
g
pe
r
s
on
de
nigr
a
ti
on
on
s
oc
ial
media
platf
o
r
ms
by
pr
opos
ing
L
S
T
M
-
C
L
R
to
a
c
hieve
be
tt
e
r
c
las
s
if
ica
ti
on
a
c
c
ur
a
c
y,
a
s
oppos
e
d
to
t
r
a
dit
ional
methods
s
uc
h
a
s
M
L
P
[
23]
,
B
E
R
T
[
23]
a
nd
C
NN
-
B
iL
S
T
M
[
25]
.
T
he
main
objec
ti
ve
is
to
e
nha
nc
e
the
L
S
T
M
’
s
lea
r
n
ing
r
a
te
(
L
R
)
with
the
he
lp
of
C
L
R
.
De
tec
ti
ng
a
nd
mi
ti
g
a
ti
ng
pe
r
s
on
de
nigr
a
ti
on
is
c
r
uc
ial
not
only
f
or
p
r
otec
ti
ng
indi
viduals
'
menta
l
we
ll
-
be
ing,
but
a
ls
o
f
or
f
os
ter
ing
a
s
a
f
e
r
a
nd
mor
e
inclus
ive
onli
ne
e
nvir
onment.
T
he
r
e
f
or
e
,
i
n
thi
s
r
e
s
e
a
r
c
h,
c
ybe
r
bull
ying
in
s
oc
ial
media
is
de
tec
ted
by
de
ve
lopi
ng
a
f
r
a
mew
or
k
ba
s
e
d
on
the
L
S
T
M
.
F
r
om
the
r
e
s
ult
a
na
lys
is
,
t
he
L
S
T
M
-
C
L
R
pe
r
f
or
manc
e
is
c
ontr
a
s
ted
with
the
tr
a
dit
ional
met
hods
s
uc
h
a
s
M
L
P
[
23]
,
B
E
R
T
[
23
]
a
nd
C
NN
-
B
iL
S
T
M
[
25]
.
F
r
om
T
a
ble
3
,
it
is
obs
e
r
ve
d
that
the
L
S
T
M
-
C
L
R
a
tt
a
ins
c
omm
e
nda
ble
c
las
s
if
ica
ti
on
a
c
c
ur
a
c
y
of
82%
,
pr
e
c
is
ion
a
s
80%
,
r
e
c
a
ll
a
s
83%
,
a
nd
F
-
mea
s
ur
e
a
s
81%
.
W
hil
e
the
e
xis
ti
ng
methods
M
L
P
[
23]
a
tt
a
ins
88%
of
a
c
c
ur
a
c
y;
the
e
xis
ti
ng
B
E
R
T
[
23
]
a
tt
a
ins
79%
a
nd
74%
of
a
c
c
ur
a
c
y
a
nd
F
-
mea
s
ur
e
,
r
e
s
pe
c
ti
ve
ly.
On
the
othe
r
ha
nd,
the
e
xis
ti
ng
C
NN
-
B
iL
S
T
M
obtains
a
n
a
c
c
ur
a
c
y
of
80.
55%
.
T
he
s
e
r
e
s
ult
s
p
r
ove
im
por
tanc
e
of
de
pl
oying
a
da
va
nc
e
d
tec
hnologi
e
s
to
tac
kle
today's
p
r
oblems
,
a
longs
ide
s
howing
how
de
e
p
-
lea
r
ning
a
ppr
oa
c
he
s
make
a
s
igni
f
ica
nt
c
ontr
ibut
ion
to
the
ongoing
f
igh
t
a
ga
ins
t
c
ybe
r
bull
ying
de
tec
ti
on
.
S
ti
ll
,
a
high
L
R
s
ome
ti
mes
c
a
us
e
s
va
r
ying
los
s
f
unc
ti
on
a
nd
pr
oble
ms
with
c
onve
r
ge
nc
e
a
s
f
indi
ng
the
global
be
s
t
is
c
ha
ll
e
nging.
W
hil
e
a
s
mall
L
R
s
lows
down
the
ne
twor
k’
s
lea
r
ning
s
pe
e
d
a
nd
make
s
it
ha
r
d
to
identif
y
the
global
be
s
t.
T
he
r
e
f
o
r
e
,
the
f
inal
c
onve
r
ge
nc
e
e
f
f
e
c
t
in
the
ne
twor
k
model
is
s
igni
f
ica
ntl
y
a
f
f
e
c
ted
by
L
R
,
ther
e
by
m
a
king
the
s
e
tt
ing
of
L
R
a
major
f
oc
us
of
the
a
pp
li
e
d
DL
model.
Als
o,
the
pr
e
s
e
nt
s
tudy
doe
s
not
dis
ti
nguis
h
a
mong
the
c
ybe
r
bull
ying
c
a
tegor
ies
,
c
ons
e
que
ntl
y
mot
ivating
to
e
xtend
a
nd
s
tudy
i
f
the
p
r
opos
e
d
L
S
T
M
-
C
L
R
c
a
n
e
xe
c
ute
f
ine
-
gr
a
ined
c
ybe
r
bull
ying
c
las
s
if
ica
ti
ons
in
the
f
utur
e
.
5.
CONC
L
USI
ON
T
he
r
e
s
e
a
r
c
h
f
indi
ngs
o
f
f
e
r
a
n
a
dva
nc
e
ment
in
s
e
nti
ment
a
na
lys
is
,
e
nha
nc
ing
de
nigr
a
ti
on
de
tec
ti
on
on
s
oc
ial
media
th
r
ough
L
S
T
M
-
C
L
R
.
I
n
the
r
e
s
e
a
r
c
h
f
ield
,
thi
s
s
igni
f
ies
a
s
hif
t
towa
r
ds
mor
e
s
ophis
ti
c
a
ted
de
e
p
lea
r
ning
tec
hniques
a
nd
pr
omi
s
e
s
a
s
a
f
e
r
onli
ne
e
nvir
onment,
f
os
ter
ing
digi
tal
c
ivi
li
ty
a
nd
c
ounter
ing
c
ybe
r
bull
ying
with
im
pr
ove
d
a
c
c
ur
a
c
y
a
nd
e
f
f
icie
nc
y.
T
he
f
u
r
ther
inves
ti
ga
ti
ons
a
r
e
f
oc
us
e
d
on
the
im
pleme
ntation
of
p
r
opos
e
d
method
with
mul
ti
-
m
oda
l
da
ta.
I
t
is
c
r
uc
ial
to
us
e
the
tec
hnology's
po
tential
f
o
r
the
good
of
the
s
oc
iety
a
s
it
de
ve
lops
f
ur
the
r
,
ther
e
f
or
e
f
os
ter
ing
a
mor
e
s
e
c
ur
e
a
nd
inclus
ive
onli
ne
c
omm
unit
y
f
o
r
a
ll
.
T
hr
ough
the
in
tr
oduc
ti
on
of
a
nove
l
f
r
a
mew
or
k
that
make
s
us
e
of
L
S
T
M
with
C
L
R
,
thi
s
s
tudy
a
dds
to
the
ongoing
e
f
f
or
ts
to
tac
kle
c
ybe
r
bull
ying.
T
he
main
goa
l
is
to
incr
e
a
s
e
L
S
T
M
's
L
R
us
ing
C
L
R
,
whic
h
make
s
it
e
a
s
ier
to
s
pot
a
nd
de
a
l
wi
th
c
a
s
e
s
of
c
ybe
r
bull
ying
on
s
oc
ial
media
.
P
r
omi
s
in
g
r
e
s
ult
s
a
r
e
obtaine
d
whe
n
the
pr
opos
e
d
f
r
a
mew
or
k
is
e
v
a
luate
d
a
ga
ins
t
the
e
xis
ti
ng
a
ppr
oa
c
he
s
na
mely,
M
L
P
a
nd
B
E
R
T
uti
li
z
ing
T
witt
e
r
da
ta
a
s
a
be
nc
hmar
k.
F
r
om
the
r
e
s
ult
s
,
i
t
is
s
e
e
n
that
the
pr
opos
e
d
L
S
T
M
-
C
L
R
a
c
c
ompl
is
he
s
82%
a
c
c
ur
a
c
y,
80%
of
pr
e
c
is
ion,
83
%
of
r
e
c
a
ll
a
nd
81
%
of
F
-
mea
s
ur
e
.
T
his
indi
c
a
tes
how
we
ll
the
pr
opos
e
d
f
r
a
mew
or
k
wor
ks
to
identi
f
y
twe
e
ts
t
ha
t
invol
ve
c
ybe
r
bu
ll
ying.
M
or
e
ove
r
,
e
xplor
ing
te
c
hniques
to
mi
ti
ga
te
bias
e
s
in
t
r
a
ini
ng
da
ta
a
nd
im
pr
ove
the
model's
r
e
s
il
ienc
e
to
a
dve
r
s
a
r
ial
a
tt
a
c
ks
e
nh
a
nc
e
s
it
s
pr
a
c
ti
c
a
l
uti
li
ty
in
r
e
a
l
-
wor
ld
s
e
tt
ings
.
F
utur
e
r
e
s
e
a
r
c
h
c
ould
a
ls
o
f
oc
us
on
de
ve
lopi
ng
e
n
s
e
mbl
e
me
thods
or
hybr
id
a
r
c
hit
e
c
tur
e
s
that
int
e
gr
a
te
mul
ti
ple
modalit
ies
f
or
the
pr
a
c
ti
c
a
l
im
pli
c
a
ti
ons
o
f
de
nigr
a
ti
on
de
t
e
c
ti
on.
RE
F
E
RE
NC
E
S
[
1]
A
.
P
.
R
odr
ig
ue
s
e
t
al
.
,
“
R
e
a
l
-
ti
me
T
w
it
te
r
s
pa
m
de
te
c
ti
on
a
nd
s
e
nt
im
e
nt
a
na
ly
s
is
u
s
in
g
ma
c
hi
ne
le
a
r
ni
ng
a
nd
de
e
p
le
a
r
ni
ng
te
c
hni
que
s
,”
C
om
put
at
io
nal
I
nt
e
ll
ig
e
nc
e
and N
e
ur
o
s
c
ie
nc
e
, vol
. 2022, pp. 1
–
14,
A
pr
. 2022,
doi
:
10.1155/2022/
5211949.
[
2]
N
.
P
a
vi
th
a
e
t
al
.
,
“
M
ovi
e
r
e
c
omm
e
nda
ti
on
a
nd
s
e
nt
im
e
nt
a
n
a
ly
s
is
u
s
in
g
ma
c
hi
ne
le
a
r
ni
ng,”
G
lo
bal
T
r
an
s
it
io
ns
P
r
oc
e
e
di
ngs
,
vol
. 3, no. 1, pp. 279
–
284, J
un. 2022, doi:
10.1016/j
.gl
tp
.2022.03.012.
[
3]
P
.
K
ouka
r
a
s
,
C
.
N
ou
s
i,
a
nd
C
.
T
jo
r
tj
is
,
“
S
to
c
k
ma
r
ke
t
pr
e
di
c
ti
on
us
in
g
mi
c
r
obl
oggi
ng
s
e
nt
im
e
nt
a
na
ly
s
is
a
nd
m
a
c
hi
ne
le
a
r
ni
ng,
”
T
e
le
c
om
, vol
. 3, no. 2, pp. 358
–
378, M
a
y 2022, doi:
10.3390/t
e
l
e
c
om3020019.
[
4]
Al
-
K
how
a
r
iz
mi
,
I
.
P
.
S
a
r
i,
a
nd
H
.
M
a
ul
a
na
,
“
O
pt
im
iz
a
ti
on
of
s
uppor
t
ve
c
to
r
ma
c
hi
n
e
w
it
h
c
ubi
c
ke
r
ne
l
f
unc
ti
on
to
de
te
c
t
c
ybe
r
bul
ly
in
g
in
s
oc
ia
l
n
e
twor
ks
,”
T
e
lk
om
ni
k
a
(
T
e
le
c
om
m
u
ni
c
at
io
n
C
om
put
in
g
E
le
c
tr
oni
c
s
and
C
ont
r
ol
)
,
vol
.
22,
no.
2,
pp. 329
–
339, Apr
. 2024, doi:
10.12928/T
E
L
K
O
M
N
I
K
A
.v22i2.25437.
[
5]
R
. H
. C
ha
nne
gow
da
, P
. K
a
r
th
ik
, R
.
S
r
in
iv
a
s
a
ia
h, a
nd M
. S
hi
va
r
a
j,
“
C
us
to
mi
z
e
d ma
s
k r
e
gi
on ba
s
e
d c
onvolut
io
na
l
ne
ur
a
l
ne
tw
or
ks
f
or
un
-
uni
f
or
me
d
s
ha
pe
te
xt
de
te
c
ti
on
a
nd
te
xt
r
e
c
ogni
ti
on,”
I
nt
e
r
nat
io
nal
J
our
nal
of
E
le
c
tr
ic
al
and
C
om
put
e
r
E
ngi
ne
e
r
in
g
,
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
E
lec
&
C
omp
E
ng
I
S
S
N:
2088
-
8708
De
nigr
ati
on
analys
is
of
T
w
it
ter
data
us
ing
c
y
c
li
c
lear
ning
r
ate
…
(
Suhas
B
har
adw
aj
R
ajendr
a
)
709
vol
. 13, no. 1, pp. 413
–
424, F
e
b. 2023, doi:
10.11591/i
je
c
e
.v13i1.pp413
-
424.
[
6]
N
.
G
a
r
g
a
nd
K
.
S
ha
r
ma
,
“
T
e
xt
pr
e
-
pr
oc
e
s
s
in
g
of
mul
ti
li
ngua
l
f
or
s
e
nt
im
e
nt
a
na
ly
s
is
ba
s
e
d
on
s
oc
ia
l
ne
twor
k
da
ta
,”
I
nt
e
r
nat
io
nal
J
our
nal
of
E
le
c
tr
ic
al
and C
om
put
e
r
E
ngi
ne
e
r
in
g
, vol
. 12, no. 1,
pp. 776
–
784, F
e
b. 2022,
doi
:
10.11591/i
je
c
e
.v12i1.pp776
-
784.
[
7]
N
.
K
e
w
s
uw
un
a
nd
S
.
K
a
jo
r
nka
s
ir
a
t,
“
A
s
e
nt
im
e
nt
a
na
ly
s
is
mo
de
l
of
a
gr
it
e
c
h
s
ta
r
tu
p
on
F
a
c
e
book
c
omm
e
nt
s
u
s
in
g
na
iv
e
B
a
ye
s
c
la
s
s
if
ie
r
,”
I
nt
e
r
nat
io
nal
J
our
nal
of
E
le
c
tr
ic
al
and
C
om
put
e
r
E
ngi
ne
e
r
in
g
,
vol
.
12,
no.
3,
pp.
2829
–
2838,
J
un.
2
022,
doi
:
10.11591/i
je
c
e
.v12i3.pp2829
-
2838.
[
8]
K
.
N
a
it
ha
ni
a
nd
Y
.
P
.
R
a
iwa
ni
,
“
R
e
a
li
z
a
ti
on
of
na
tu
r
a
l
la
ng
ua
ge
pr
oc
e
s
s
in
g
a
nd
ma
c
hi
ne
l
e
a
r
ni
ng
a
ppr
oa
c
he
s
f
or
te
xt
-
b
a
s
e
d
s
e
nt
im
e
nt
a
na
ly
s
i
s
,”
E
x
pe
r
t
Sy
s
te
m
s
, vol
. 40, no. 5, Aug. 202
3,
doi
:
10.1111/e
xs
y.13114.
[
9]
W
.
F
.
S
a
tr
ya
,
R
.
A
pr
il
li
ya
ni
,
a
nd
E
.
H
.
Y
os
s
y,
“
S
e
nt
im
e
nt
a
na
l
ys
is
of
I
ndone
s
ia
n
pol
ic
e
c
hi
e
f
us
in
g
mul
ti
-
le
ve
l
e
ns
e
mbl
e
mod
e
l,
”
P
r
oc
e
di
a C
om
put
e
r
S
c
ie
nc
e
, vol
. 216, pp. 620
–
629, 2022, doi:
10.1016/j
.pr
oc
s
.2022.12.177.
[
10]
I
.
A
lm
a
li
s
,
E
.
K
oul
oumpr
is
,
a
nd
I
.
V
la
ha
v
a
s
,
“
S
e
c
to
r
-
le
ve
l
s
e
nt
im
e
nt
a
na
ly
s
is
w
it
h
de
e
p
le
a
r
ni
ng,”
K
now
l
e
dge
-
B
as
e
d
Sy
s
te
m
s
,
vol
. 258, De
c
. 2022, doi:
10.1016/j
.knos
ys
.2022.109954.
[
11]
U
.
N
a
s
e
e
m,
I
.
R
a
z
z
a
k,
M
.
K
hus
hi
,
P
.
W
.
E
kl
und,
a
nd
J
.
K
im
,
“
C
O
V
I
D
S
e
nt
i:
a
la
r
ge
-
s
c
a
l
e
be
nc
hma
r
k
T
w
it
te
r
da
ta
s
e
t
f
or
C
O
V
I
D
-
19
s
e
nt
im
e
nt
a
na
ly
s
is
,”
I
E
E
E
T
r
ans
ac
ti
ons
on C
om
put
at
io
nal
Soc
ia
l
Sy
s
te
m
s
,
vol
.
8,
no.
4,
pp.
976
–
988,
A
ug.
2021, doi:
10.1109/T
C
S
S
.2021.3051189.
[
12]
N
.
A
s
l
a
m,
F
.
R
u
s
t
a
m
,
E
.
L
e
e
,
P
.
B
.
W
a
s
hi
ng
to
n,
a
n
d
I
.
A
s
hr
a
f
,
“
S
e
nt
i
me
nt
a
n
a
l
y
s
i
s
a
n
d
e
m
ot
io
n
d
e
t
e
c
ti
on
on
c
r
y
pt
oc
ur
r
e
n
c
y
-
r
e
l
a
t
e
d
t
w
e
e
t
s
u
s
i
ng
e
n
s
e
mb
l
e
L
S
T
M
-
G
R
U
mo
de
l,
”
I
E
E
E
A
c
c
e
s
s
,
vo
l.
10
,
p
p.
3
93
13
–
3
93
24
,
20
22
,
doi
:
10
.1
10
9/
A
C
C
E
S
S
.2
02
2.
31
65
62
1.
[
13]
G
.
C
ha
ndr
a
s
e
ka
r
a
n,
N
.
A
nt
oa
ne
la
,
G
.
A
ndr
e
i,
C
.
M
oni
c
a
,
a
nd
J
.
H
e
ma
nt
h,
“
V
is
ua
l
s
e
nt
im
e
nt
a
na
ly
s
is
us
in
g
d
e
e
p
le
a
r
ni
ng
mo
de
ls
w
it
h s
oc
ia
l
me
di
a
da
ta
,”
A
ppl
ie
d Sc
ie
n
c
e
s
, vol
. 12, no. 3,
J
a
n. 2
022, doi:
10.3390/app120310
30.
[
14]
H
.
S
a
le
h,
S
.
M
os
ta
f
a
,
A
.
A
lh
a
r
bi
,
S
.
E
l
-
S
a
ppa
gh,
a
nd
T
.
A
lk
ha
li
f
a
h,
“
H
e
te
r
oge
ne
ous
e
ns
e
mbl
e
de
e
p
le
a
r
ni
ng
mode
l
f
or
e
nha
nc
e
d
A
r
a
bi
c
s
e
nt
im
e
nt
a
na
ly
s
i
s
,”
Se
n
s
or
s
, vol
. 22, no. 10, M
a
y 2022,
doi
:
10.3390/s
22103707.
[
15]
B
.
A
lB
a
da
ni
,
R
.
S
hi
,
a
nd
J
.
D
ong,
“
A
nove
l
ma
c
hi
n
e
le
a
r
ni
ng
a
ppr
oa
c
h
f
or
s
e
nt
im
e
nt
a
na
ly
s
is
on
T
w
it
te
r
in
c
or
por
a
ti
ng
th
e
uni
ve
r
s
a
l
la
ngua
ge
mode
l
f
in
e
-
tu
ni
ng a
nd S
V
M
,”
A
ppl
ie
d Sy
s
te
m
I
nnov
at
io
n
, vol
. 5, no. 1, J
a
n. 2022,
doi
:
10.3390/a
s
i5
010013
.
[
16]
H
.
S
a
le
h,
S
.
M
o
s
ta
f
a
,
L
.
A
.
G
a
br
a
ll
a
,
A
.
O
.
A
s
e
e
r
i,
a
nd
S
.
E
l
-
S
a
ppa
gh,
“
E
nha
nc
e
d
A
r
a
bi
c
s
e
nt
im
e
nt
a
na
ly
s
is
u
s
in
g
a
n
ove
l
s
ta
c
ki
ng e
ns
e
mbl
e
of
hybr
id
a
nd de
e
p l
e
a
r
ni
ng mode
ls
,”
A
ppl
ie
d Sc
ie
nc
e
s
, vol
. 12, no. 18, S
e
p. 2022, doi:
10.3390/app121889
67.
[
17]
M
.
R
a
j,
S
.
S
in
gh,
K
.
S
ol
a
nki
,
a
nd
R
.
S
e
lv
a
na
mbi
,
“
A
n
a
ppl
ic
a
ti
on
to
de
te
c
t
c
ybe
r
bul
ly
in
g
us
in
g
m
a
c
hi
ne
le
a
r
ni
ng
a
nd
d
e
e
p
le
a
r
ni
ng t
e
c
hni
que
s
,”
SN
C
om
put
e
r
S
c
ie
nc
e
, vol
. 3, no. 5,
J
ul
. 2
022, doi:
10.1007/s
42979
-
022
-
01308
-
5.
[
18]
A
.
M
.
A
ld
u
a
il
a
j
a
nd
A
.
B
e
lg
hi
th
,
“
D
e
t
e
c
ti
ng
A
r
a
bi
c
c
ybe
r
b
ul
ly
in
g
twe
e
ts
u
s
in
g
ma
c
hi
n
e
le
a
r
ni
ng,”
M
ac
hi
ne
L
e
ar
ni
ng
and
K
now
le
dge
E
x
tr
ac
ti
on
, vol
. 5, no. 1, pp. 29
–
42, J
a
n. 2023, doi:
10.3390/m
a
ke
5010003.
[
19]
B
.
A
.
H
.
M
ur
s
he
d,
J
.
A
ba
w
a
jy
,
S
.
M
a
ll
a
ppa
,
M
.
A
.
N
.
S
a
if
,
a
nd
H
.
D
.
E
.
A
l
-
A
r
ik
i,
“
D
E
A
-
R
N
N
:
a
hyb
r
id
de
e
p
le
a
r
ni
ng
a
ppr
o
a
c
h
f
or
c
ybe
r
bul
ly
in
g
de
te
c
ti
on
in
T
w
it
te
r
s
oc
ia
l
me
di
a
pl
a
tf
or
m,”
I
E
E
E
A
c
c
e
s
s
,
vol
.
10,
pp.
25857
–
25871,
2022,
doi
:
10.1109/AC
C
E
S
S
.2022.3153675.
[
20]
S
.
S
he
lk
e
a
nd
V
.
A
tt
a
r
,
“
R
umor
de
te
c
ti
on
in
s
oc
ia
l
ne
twor
k
b
a
s
e
d
on
us
e
r
,
c
ont
e
nt
a
nd
le
xi
c
a
l
f
e
a
tu
r
e
s
,”
M
ul
ti
m
e
di
a
T
ool
s
and
A
ppl
ic
at
io
ns
, vol
. 81, no. 12, pp. 17347
–
17368, M
a
r
. 2022, doi:
10.1007/s
11042
-
022
-
12761
-
y.
[
21]
C
.
R
a
j,
A
.
A
ga
r
w
a
l,
G
.
B
ha
r
a
th
y,
B
.
N
a
r
a
ya
n,
a
nd
M
.
P
r
a
s
a
d,
“
C
ybe
r
bul
ly
in
g
de
te
c
ti
on:
h
ybr
id
mode
ls
ba
s
e
d
on
ma
c
hi
ne
le
a
r
ni
ng a
nd na
tu
r
a
l
la
ngua
ge
pr
oc
e
s
s
in
g t
e
c
hni
que
s
,”
E
le
c
tr
oni
c
s
, vol
. 10, no. 22, Nov. 20
21,
doi
:
10.3390/e
le
c
tr
oni
c
s
1022281
0.
[
22]
S
.
B
e
hl
,
A
.
R
a
o,
S
.
A
gga
r
w
a
l,
S
.
C
ha
dha
,
a
nd
H
.
S
.
P
a
nnu,
“
T
w
it
te
r
f
or
di
s
a
s
te
r
r
e
li
e
f
th
r
ough
s
e
nt
im
e
nt
a
na
ly
s
is
f
or
C
O
V
I
D
-
19
a
nd na
tu
r
a
l
ha
z
a
r
d c
r
is
e
s
,”
I
nt
e
r
nat
io
nal
J
our
nal
of
D
i
s
as
te
r
R
is
k
R
e
duc
ti
on
, vol
. 55, M
a
r
. 2021,
doi
:
10.1016/j
.i
jd
r
r
.2021.102
101.
[
23]
S
.
D
e
b
a
nd
A
.
K
.
C
ha
nda
,
“
C
ompa
r
a
ti
ve
a
na
ly
s
is
of
c
ont
e
xt
ua
l
a
nd
c
ont
e
xt
-
f
r
e
e
e
mbe
ddi
ng
s
in
di
s
a
s
te
r
pr
e
di
c
ti
on
f
r
om
T
w
i
tt
e
r
da
ta
,”
M
ac
hi
ne
L
e
ar
ni
ng w
it
h A
ppl
ic
at
io
ns
, vol
. 7, M
a
r
. 2022,
doi
:
10.1016/j
.ml
w
a
.2022.100253.
[
24]
M
.
A
.
S
.
N
a
s
ut
io
n
a
nd
E
.
B
.
S
e
ti
a
w
a
n, “
E
nha
nc
in
g
c
ybe
r
bul
ly
i
ng
de
te
c
ti
on
on
I
ndone
s
ia
n T
w
it
te
r
:
le
ve
r
a
gi
ng
f
a
s
t
te
xt
f
or
f
e
a
tu
r
e
e
xpa
ns
io
n
a
nd
hybr
id
a
ppr
oa
c
h
a
ppl
yi
ng C
N
N
a
nd B
iL
S
T
M
,”
R
e
v
ue
d’
I
nt
e
ll
ig
e
nc
e
A
r
ti
fi
c
ie
ll
e
,
vol
.
37,
no.
4,
pp.
929
–
936,
2
023,
doi
:
10.18280/r
ia
.370413.
[
25]
S
.
B
e
nge
s
i,
T
.
O
la
dunni,
R
.
O
lu
s
e
gun,
a
nd
H
.
A
udu,
“
A
ma
c
hi
ne
le
a
r
ni
ng
-
s
e
nt
im
e
nt
a
na
ly
s
is
on
monke
ypox
out
br
e
a
k:
a
n
e
xt
e
ns
iv
e
da
ta
s
e
t
to
s
how
th
e
pol
a
r
it
y
of
publ
ic
opi
ni
on
f
r
om
T
w
it
te
r
twe
e
ts
,”
I
E
E
E
A
c
c
e
s
s
,
vol
.
11,
pp.
11811
–
11826,
2023,
doi
:
10.1109/AC
C
E
S
S
.2023.3242290.
[
26]
M
. I
br
a
hi
m, S
.
G
a
uc
h, O
. S
a
lm
a
n,
a
nd M
. A
lq
a
ht
a
ni
, “
A
n a
ut
o
ma
te
d me
th
od t
o e
nr
ic
h
c
ons
ume
r
he
a
lt
h voc
a
bul
a
r
ie
s
us
in
g
G
l
oV
e
w
or
d
e
mbe
ddi
ngs
a
nd
a
n
a
uxi
li
a
r
y
le
xi
c
a
l
r
e
s
our
c
e
,”
P
e
e
r
J
C
om
put
e
r
Sc
ie
nc
e
,
vol
.
7,
pp.
1
–
28,
A
ug.
2021,
doi
:
10.7717/p
e
e
r
j
-
c
s
.668.
[
27]
D
. V
id
ya
bha
r
a
th
i,
V
. M
oha
nr
a
j,
J
. S
. K
uma
r
, a
nd Y
. S
ur
e
s
h, “
A
c
hi
e
vi
ng ge
ne
r
a
li
z
a
ti
on of
de
e
p l
e
a
r
ni
ng mode
ls
i
n a
qui
c
k w
a
y b
y
a
da
pt
in
g
T
-
H
T
R
le
a
r
ni
ng
r
a
te
s
c
he
dul
e
r
,”
P
e
r
s
onal
and
U
bi
q
ui
to
us
C
om
put
in
g
,
vol
.
27,
no.
3,
pp.
1335
–
1353,
A
ug.
2023,
doi
:
10.1007/s
00779
-
021
-
01587
-
4.
B
I
OG
RA
P
HI
E
S
OF
AU
T
HO
RS
Suha
s
Bha
r
a
dw
a
j
R
a
jendr
a
i
s
w
o
r
k
i
n
g
as
a
s
s
i
s
t
an
t
p
ro
fe
s
s
o
r
i
n
t
h
e
D
e
p
art
me
n
t
o
f
In
fo
rma
t
i
o
n
Sc
i
en
ce
a
n
d
E
n
g
i
n
eer
i
n
g
at
T
h
e
N
a
t
i
o
n
a
l
I
n
s
t
i
t
u
t
e
o
f
E
n
g
i
n
eeri
n
g
,
M
y
s
u
ru
,
s
i
n
ce
2
0
1
6
.
O
b
t
a
i
n
e
d
B.
E
.
fro
m
Mal
n
ad
Co
l
l
e
g
e
o
f
E
n
g
i
n
eeri
n
g
H
as
s
an
i
n
t
h
e
y
ear
2
0
1
4
a
n
d
M.
T
ech
.
fro
m
T
h
e
N
at
i
o
n
al
In
s
t
i
t
u
t
e
o
f
E
n
g
i
n
eer
i
n
g
,
My
s
u
ru
i
n
t
h
e
y
ear
2
0
1
6
.
Cu
rren
t
l
y
p
u
r
s
u
i
n
g
P
h
.
D
.
u
n
d
er
V
i
s
v
e
s
aray
a
T
e
ch
n
o
l
o
g
i
ca
l
U
n
i
v
e
rs
i
t
y
i
n
t
h
e
area
o
f
cy
b
er
s
ecu
r
i
t
y
.
H
i
s
area
o
f
res
earch
i
n
t
eres
t
s
are
cy
b
er
s
ec
u
ri
t
y
,
mach
i
n
e
l
earn
i
n
g
,
s
o
c
i
al
n
e
t
w
o
rk
s
,
co
mp
u
t
er
fo
ren
s
i
c
s
.
H
e
can
b
e
co
n
t
ac
t
e
d
at
emai
l
:
s
u
h
a
s
b
r.
re
s
earc
h
@
g
mai
l
.
co
m.
Evaluation Warning : The document was created with Spire.PDF for Python.