Int
ern
at
i
onal
Journ
al of Ele
ctrical
an
d
Co
mput
er
En
gin
eeri
ng
(IJ
E
C
E)
Vo
l.
15
,
No.
1
,
Febr
uary
20
25
, pp.
755
~
766
IS
S
N:
20
88
-
8708
, DO
I: 10
.11
591/ij
ece.v
15
i
1
.
pp
755
-
766
755
Journ
al h
om
e
page
:
http:
//
ij
ece.i
aesc
or
e.c
om
An
im
pr
oved re
ptil
e searc
h
algorit
hm
-
bas
ed
m
achine le
arning
for senti
ment an
alys
i
s
Nitesh
Su
re
j
a
1
, Nandi
ni
M
.
Chaud
ha
ri
2
, J
alpa Bh
att
1
, T
usha
r
D
e
sa
i
1
, Vruti
Pa
ri
k
h
1
,
Sonia
Pa
nesar
1
,
Heli
Su
rej
a
3
, Jahn
av
i
Kh
ar
va
1
1
Dep
artm
en
t of
Co
m
p
u
ter
Sci
en
ce
an
d
Eng
in
eering
,
Kri
sh
n
a Scho
o
l of E
m
ergin
g
T
echn
o
lo
g
y
and
App
lied
Research
,
D
rs.
Kiran
an
d
Pallav
i Patel
Glob
al Univ
ersity
,
V
ad
o
d
ara,
I
n
d
ia
2
Dep
artm
en
t of
I
n
f
o
rm
atio
n
T
echn
o
lo
g
y
,
Krish
n
a Scho
o
l of E
m
ergin
g
Te
c
h
n
o
lo
g
y
and
App
li
ed
Research
,
Drs.
Kiran
an
d
Pallav
i
Patel
Glo
b
al Univ
ersity
,
Vad
o
d
ara,
Ind
ia
3
Dep
artm
en
t of
Co
m
p
u
ter
Sci
en
ce
an
d
Eng
in
eering
,
Bab
aria
Ins
titu
te of
T
echn
o
lo
g
y
,
Gu
jarat
T
echn
o
lo
g
ical Univ
ersity
,
Vad
o
d
ara
,
Ind
ia
Art
ic
le
In
f
o
ABSTR
A
CT
Art
ic
le
history:
Re
cei
ved
A
pr
30, 202
4
Re
vised
A
ug 16, 2
024
Accepte
d
Oct
1,
2024
The
rap
id
grow
th
of
mobile
t
e
chnol
ogi
es
has
tra
nsformed
soc
ia
l
media
,
ma
king
it
cru
cial
for
expr
essin
g
em
ot
ions
and
thought
s.
When
ma
king
signifi
c
ant
de
cis
ions,
business
es
and
gov
ern
me
nts
ca
n
b
e
nef
it
from
under
standi
ng
p
ubli
c
opini
on
.
Thi
s
infor
ma
t
io
n
ma
k
es
sent
iment
ana
lysis
vit
al
for
under
s
ta
nding
public
senti
me
n
t
pola
r
i
ty.
Thi
s
study
deve
lops
a
hyper
tune
d
d
e
ep
learni
ng
mo
del
with
sw
ar
m
int
e
ll
ig
ence
and
ma
ny
appr
oac
h
es
for
senti
me
n
t
an
al
y
sis.
convo
lut
ion
al
neur
al
net
wo
rk
(CNN
),
bidi
re
ct
ion
al
encoder
r
epr
ese
nt
at
i
ons
from
tra
nsfo
rme
rs
(BER
T),
l
ong
short
-
te
rm
me
mory
(
L
STM),
CNN
-
LSTM,
BERT
-
LST
M,
and
BERT
-
C
NN
are
the
six
de
ep
learni
n
g
mod
el
s
of
the
senti
me
n
t
analysis
using
dee
p
lea
rning
wi
th
rei
nforc
ed
l
ea
rn
ing
base
d
on
r
ept
ile
sea
r
ch
a
l
gorit
hm
(SA
-
DLRLRSA)
mode
l
.
The
re
pti
le
s
ea
rch
algorithm,
an
e
n
hanc
ed
sw
arm
int
ellige
n
ce
al
gorit
h
m
(SIA
),
opt
im
i
ze
s
d
ee
p
learni
ng
mode
l
hyper
par
amete
rs
.
Wor
d2Vec
wor
d
em
b
ed
ding
is
used
to
conve
r
t
te
x
tual
inpu
t
s
eque
nc
es
to
rep
rese
nt
at
iv
e
e
mbe
dding
spa
ces
.
Pre
-
traine
d
W
ord2Vec
e
mbe
d
ding
is
al
so
used
to
addr
e
ss
issue
of
u
nbal
an
ce
d
da
tas
et
s.
Expe
ri
mental
resul
ts
dem
onstra
te
th
at
the
SA
-
DLRL
RS
A
mode
l
works
best
with
a
c
cur
acie
s
o
f
93.
1%,
94
.
7%,
96.
8%,
96
.
3%,
97.
2%,
and
98.
3%
uti
lizing
C
NN
,
LSTM,
BERT
,
CNN
-
LSTM,
BER
T
-
CNN
,
and
BERT
-
L
STM.
Ke
yw
or
d
s
:
Deep l
ear
ning
M
ac
hin
e lea
rn
i
ng
Re
inforceme
nt
learni
ng
Re
ptil
e search
al
gorithm
Sentiment
anal
ys
is
So
ci
al
me
dia
Sw
ar
m
i
ntell
igence
This
is an
open
acc
ess arti
cl
e
un
der
the
CC
BY
-
SA
l
ic
ense
.
Corres
pond
in
g
Aut
h
or
:
Nite
sh
Surej
a
Dep
a
rtme
nt of
Com
pu
te
r
Scie
nce
a
nd
E
ng
i
ne
erin
g,
Kr
is
hn
a Scho
ol of
E
mer
ging Tec
hn
ology an
d A
ppli
ed
Re
search
,
Dr
s
. Kira
n
a
nd
Pall
avi Patel
G
l
ob
al
U
ni
ver
sit
y
Vado
dar
a
, Guj
arat,
391410,
Ind
ia
Emai
l:
nm
s
ur
ej
a@gma
il
.co
m
1.
INTROD
U
CTION
Re
cent
interest
in
sentime
nt
a
nalysis
has
gro
wn
due
to
it
s
man
y
us
es.
O
pi
nio
n
mini
ng,
or
se
ntime
nt
analysis,
u
se
s
natu
ral la
ngua
ge pr
ocessin
g and dee
p
le
ar
ni
ng to unc
over
s
ub
je
ct
ive i
nformat
ion
a
nd em
otion
a
l
sta
te
s.
Se
ntiment
a
nalysis
de
te
rmin
es
if
w
ritt
en
messa
ge
s
are
posit
ive,
ne
gative,
or
neu
t
ral
[
1]
.
In
rece
nt
year
s
,
s
ocial
media
ha
s
bec
om
e
vital
to
da
il
y
li
fe.
Peopl
e
ex
pr
e
ss
t
heir
feeli
ngs
on
T
witt
er,
M
et
a
(formerly
Faceb
ook),
I
nst
agr
am
,
an
d
oth
e
r
pu
blic
platfo
rms.
Th
eref
or
e
,
so
ci
al
media
te
xt
analysis
ma
y
assist
unde
rstan
d
public
opinio
ns
[2]
.
B
y
rev
ie
wing
c
us
to
me
r
re
v
ie
w
s,
bus
iness
owners
can
ide
ntif
y
pro
duct
impro
veme
nts.
Add
it
io
nally,
po
li
ti
cal
bod
ie
s
can use
sen
ti
m
ent an
al
ys
is t
o creat
e act
ion pl
ans
[3]
.
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
omp E
ng,
V
ol.
15
, No
.
1
,
Febr
uary
20
25
:
755
-
766
756
Sentiment
a
nal
ys
is
(SA
)
is
a
biggest
an
d
ha
rd
est
ta
s
k
in
ar
ti
fici
al
intel
l
igence
(
AI)
.
T
he
sy
ste
m
us
es
arti
fici
al
meth
od
s
to
rec
ogni
ze
psyc
ho
l
og
i
cal
inf
ormat
io
n
incl
ud
i
ng
at
ti
tud
es,
pe
rs
pec
ti
ves
,
a
nd
m
oods
in
blogs,
ne
ws
ar
ti
cl
es,
and
s
oc
ia
l
media
posts
[
4]
.
S
ocial
media
a
nalysis
re
qu
i
res
man
agin
g
a
nd
pro
cessi
ng
enorm
ous
am
ounts
of
co
nten
t.
Lar
ge
am
ounts
of
c
on
te
nt
wer
e
sh
a
red
and
ge
ner
at
e
d
instantl
y,
re
quirin
g
eff
ic
ie
nt
c
onte
nt
ma
na
geme
nt
.
The
c
on
te
n
t processi
ng
ap
proac
h
must
al
s
o
be
c
onside
re
d
beca
us
e
t
he
c
on
te
xts
wer
e
not
sta
ndar
dized
li
ke
the
pr
e
valent
da
ta
[5]
.
T
his
work
pro
poses
a
cutti
ng
-
e
dge
sentime
nt
a
nalysi
s
method
us
in
g
s
warm
i
ntell
igence
(SI)
a
nd
de
ep
le
a
rn
i
ng.
A
n
upgra
ded
rept
il
e
search
al
go
rithm
(RS
A
)
is
use
d
with
six
dee
p
le
arn
in
g
m
ode
ls:
convo
l
utional
neural
netw
ork
(C
NN)
,
bi
directi
onal
enc
od
e
r
re
present
at
ion
s
from
tra
nsfo
rm
ers
(BERT
)
,
l
ong
short
-
te
rm
memor
y
(L
ST
M
)
,
CN
N
-
LST
M
,
BER
T
-
LS
T
M
,
a
nd
BER
T
-
CNN.
Re
inforceme
nt
le
arn
i
ng
(RL
)
fine
-
tu
nes
al
l
deep
le
ar
ning
model
hyperpa
rameters
to
im
pro
ve
RS
A
[6]
,
[
7]
.
The
W
ord
2Ve
c
w
ord
em
be
ddin
g
te
c
hniq
ue
is
us
e
d.
T
his
stu
dy
a
ddres
s
es
im
balance
d
dataset
s
us
in
g
data
augmentat
io
n.
Fo
ll
owin
g
pa
r
agr
a
phs
re
view
seve
ral
well
-
est
ablishe
d
de
ep
le
arn
i
ng
me
thods
integ
rate
d
with
swarm intel
li
ge
nce
(S
I
) for se
ntiment a
nalysi
s.
H
a
l
a
w
a
n
i
e
t
al
.
[8]
u
s
e
H
a
r
r
i
s
H
a
w
k
s
o
p
t
i
m
i
z
a
t
i
o
n
a
nd
d
e
e
p
l
e
a
r
n
i
n
g
f
o
r
s
e
n
t
i
m
e
n
t
a
n
a
l
y
s
i
s
.
T
h
e
a
u
t
o
m
a
t
e
d
s
e
n
t
i
m
e
n
t
a
na
l
y
s
i
s
i
n
s
o
c
i
a
l
m
e
di
a
u
s
i
n
g
H
a
r
r
i
s
H
a
w
ks
o
p
t
i
m
i
z
a
t
i
o
n
w
i
t
h
d
e
e
p
l
e
a
r
n
i
n
g
(
A
S
A
S
M
-
H
H
O
D
L
)
m
o
d
e
l
h
a
d
8
4
.
2
5
%
,
9
5
.
5
0
%
,
a
n
d
8
8
.
7
5
%
a
c
c
u
r
a
c
y
o
n
S
e
n
t
i
m
e
n
t
1
4
0
,
T
w
e
e
t
s
A
i
r
l
i
n
e
,
a
n
d
T
w
e
e
t
s
s
e
m
i
n
a
l
d
a
t
a
s
e
t
s
.
E
l
e
c
t
r
o
e
n
c
e
p
h
a
l
o
g
r
a
p
h
y
(
E
E
G
)
s
i
g
n
a
l
s
w
e
r
e
u
s
e
d
t
o
c
r
e
a
t
e
a
n
e
m
o
t
i
o
n
r
e
c
o
gn
i
t
i
o
n
s
y
s
t
e
m
i
n
s
t
u
dy
[9]
u
s
i
n
g
t
h
e
S
h
a
n
g
h
a
i
J
i
a
o
T
o
n
g
U
n
i
v
e
r
s
i
t
y
(
S
J
T
U
)
d
a
t
a
s
e
t
.
B
i
n
a
r
y
m
o
t
h
f
l
a
m
e
o
p
t
i
m
i
z
a
t
i
o
n
(
B
M
F
O
)
s
e
l
e
c
t
e
d
f
e
a
t
u
r
e
s
a
n
d
C
N
N
c
l
a
s
s
i
f
i
e
d
t
h
e
m
.
T
h
e
a
l
g
o
r
i
t
h
m
w
a
s
95
.
0
0
%
a
c
c
u
r
a
t
e
.
A
u
t
h
o
r
s
d
e
s
i
gn
e
d
a
h
y
b
r
i
d
t
w
e
e
t
s
e
n
t
i
m
e
n
t
a
n
a
l
ys
i
s
a
l
g
o
r
i
t
h
m
i
n
[
1
0
]
.
T
u
n
i
c
a
t
e
s
w
a
r
m
a
l
g
o
r
i
t
h
m
(
T
S
A
)
i
m
p
r
o
v
e
d
s
c
a
l
a
b
i
l
i
t
y
a
n
d
p
r
o
c
e
s
s
i
n
g
s
p
e
e
d
i
n
t
h
e
e
x
p
e
r
i
m
e
n
t
.
S
i
m
ul
a
t
i
o
n
a
n
n
e
a
l
i
n
g
(
S
A
)
a
n
d
b
i
t
w
i
s
e
o
p
e
r
a
t
i
o
n
s
a
r
e
u
s
e
d
i
n
t
h
e
h
y
b
r
i
d
H
H
O
m
e
t
h
o
d
t
o
s
o
l
v
e
l
o
c
a
l
op
t
i
m
a
i
n
t
h
i
s
w
o
r
k
.
T
h
e
m
o
d
e
l
h
a
d
9
6
.
3
7
%
p
r
e
c
i
s
i
o
n
.
P
a
r
t
i
c
l
e
s
w
a
r
m
o
p
t
i
m
i
z
a
t
i
o
n
(
P
S
O
)
,
g
e
n
e
t
i
c
a
l
g
o
r
i
t
h
m
s
(
G
A
)
,
a
n
d
d
e
c
i
s
i
on
t
r
e
e
(
D
T
)
c
l
a
s
s
i
f
i
e
r
w
e
r
e
u
t
i
l
i
z
e
d
i
n
r
e
f
e
r
e
n
c
e
[11]
.
The
meth
od
ha
d
90.00%
pr
eci
sion
.
G
A,
PSO,
a
nd
deci
sion
t
rees
were
us
e
d
to
crea
te
a
hybri
d
Twitt
er
s
pam
detect
ion
sy
s
te
m
in
[
12]
.
The
y
c
reate
over
600
mil
li
on
tw
eet
s
an
d
e
xtrac
t
at
tribu
te
s
t
o
detect
sp
am
in
real
ti
me
us
in
g
unif
or
m
resou
rce
l
ocato
r
(U
RL
)
secur
it
y.
The
hybri
d
G
A
-
PS
O
-
DT
meth
od
is
over
90.00%
accu
r
at
e.
A
dee
p
le
arn
i
ng
m
odel
na
med
bi
directi
onal
lo
ng
sho
rt
-
te
rm
mem
ory
with
te
xt
conv
olu
ti
onal
sel
f
-
at
te
ntio
n
(
Bi
LSTM
-
TCS
A)
was
de
velo
ped
f
or
sho
rt
te
xt
sentime
nt
analysis
i
n
[13]
.
T
his
model
us
e
s
bi
di
recti
on
al
lo
ng
sh
ort
-
te
r
m
me
mory
(BiLST
M
)
,
te
xt
c
onvo
luti
on
al
neural
netw
ork
(
Text
CNN
)
,
and
s
el
f
-
a
tt
ention.
E
nhance
d
impro
ved
pa
rtic
le
swa
rm
opti
miza
ti
on
(IPS
O)
opti
mize
d
the
hype
rp
a
ram
et
ers
.
Using
a
ge
ne
r
at
ive
ad
versari
al
netw
ork
(
G
AN),
a
la
r
ge
a
moun
t
of
upda
te
d
te
xt
was
cr
eat
ed,
im
pro
vi
ng
the
model’s
resil
ie
nce.
Af
te
r
processin
g,
the
Bi
LSTM
mod
el
yielded
gl
obal
sema
ntic
insig
hts
a
nd
94.
59%
accurac
y
on
th
e
hote
l
re
view
s
dataset
.
A
rabi
c
T
witt
er
se
ntiment
a
nalysis
us
in
g
PSO
a
nd
dee
p
le
ar
ning
(DL
)
was
pr
e
sente
d
in
st
udy
[14
]
.
The
bi
directi
on
al
gate
d
rec
urren
t
unit
(Bi
GRU
)
cl
assifi
e
r
cl
assifi
es
at
ti
tud
es
.
Qu
a
ntum
P
SO
(Q
P
SO)
opti
mize
s hy
perpara
mete
rs.
Hybr
i
d
-
flash
bu
tt
er
fly
opti
miza
ti
on
with
dee
p
le
ar
nin
g
-
based
sent
iment
anal
ys
i
s
[15
]
wa
s
dev
el
op
e
d.
O
n
the
Ca
non
da
ta
set
,
hybr
i
d
f
lowe
r
bee
op
ti
miza
ti
on
with
deep
le
ar
ning
sentime
nt
an
al
ys
is
(H
FB
O
-
DL
SA)
ha
d
97.
66%
pr
e
ci
sio
n.
A
n
inno
vative
s
of
t
war
e
te
c
hn
i
que
for
anal
yzin
g
em
oji
emotion
s
was
dev
el
op
e
d
in
[
16]
.
Vi
deos
a
nd
ima
ges
are
noise
-
filt
ered
fir
st.
Jie
bas
vo
ca
bu
la
ry
wa
s
e
xp
and
e
d
by
seg
m
enting
En
glish
te
xt
with
e
mo
ji
a
nd
i
nter
net
sla
ng
.
E
m
ojis
sta
rted
as
te
xt.
A
recurr
ent
ne
ur
al
net
work
(R
NN
)
cl
assifi
es
emot
ion
s
a
s
po
sit
ive,
e
xtre
mely
po
sit
ive
,
neu
tr
al
,
ne
gative,
a
nd
ve
ry
ne
gat
ive
us
i
ng
the
fu
z
z
y
bu
tt
er
fly
opti
miza
ti
on
(F
B
O)
al
gorithm
.
This
cat
e
gori
zat
ion
us
es
L
STM.
T
he
re
comme
nded
s
entime
nt
analysis
m
od
el
outpe
rforms
c
urren
t
meth
ods
.
P
rod
uct
re
vie
w
se
ntiments
a
re
cat
eg
ori
zed
us
in
g
t
he
a
dapt
ive
par
ti
cl
e gre
y w
olf o
ptimi
zer
with
deep le
ar
ning
base
d
se
nt
iment anal
ys
is
(APG
W
O
-
DL
SA
)
in
[
17]
.
The
APG
WO
-
DLSA
model
ob
ta
ine
d
94.
77
%
accu
rac
y
on
the
C
PAA
dat
aset
an
d
85.
31%
on
the
AP
dataset
.
Alza
qe
bah
et
al.
[18
]
present
a
n
i
mpro
ve
d
sal
p
swarm
meth
od
(
SS
A)
f
or
A
ra
bic
sentime
nt
analysis
featur
e
selec
ti
on. Wi
th
80.
00% accu
racy, th
e SS
A
outpe
r
f
ormed
t
he
P
SO
an
d
gr
e
y wo
lf
op
ti
miza
ti
on (GW
O)
.
M
as
hr
a
qi
a
nd
Halawa
ni
[
19]
co
ns
tr
ucted
drag
onfly
opti
m
iz
at
ion
with
de
ep
le
ar
ning
e
nab
le
d
Ar
a
bic
tweet
sentime
nt
a
nal
ys
is
.
T
he
te
r
m
f
reque
ncy
-
i
nv
e
rse
docu
m
ent
f
reque
ncy
(TF
-
I
DF)
m
odel
ge
ne
rates
featur
e
vecto
rs.
Atte
nt
ion
-
base
d
bid
i
r
ect
ion
al
lo
ng
s
hort
-
te
r
m
me
m
ory
(
ABL
ST
M)
cl
assifi
es
sen
ti
ment.
Diff
e
re
ntial
flo
wer
opti
miza
ti
on
(DFO
)
opti
mize
s
A
BLSTM
hype
rp
a
rameters
la
st.
O
n
the
se
mEv
al
2017
da
ta
set
,
the
dif
fer
e
ntial
flow
e
r
opti
miza
ti
on
with
deep
le
ar
ning
sentime
nt
analysis
a
nd
at
te
ntion
te
ch
niq
ue
(D
F
O
DL
-
SAA
T)
m
odel
is
92.
00%
accu
rat
e.
Lo
g
te
r
m
f
r
equ
e
nc
y
-
base
d
modifie
d
in
ve
rse
cl
ass
fr
e
quenc
y
(LFMI
)
was
use
d
to
e
xtract
featur
e
s
in
st
udy
[
20]
.
T
he
f
eat
ur
e
was
c
hose
n
via
Le
vy
fligh
t
-
base
d
may
fly
op
ti
miza
ti
on.
The
sel
ect
ed
data
is
use
d
t
o
bu
il
d
the
e
nh
a
nce
d
local
searc
h
w
hale
opti
miza
ti
on
-
base
d
impro
ved
l
oc
al
search
w
ha
le
op
ti
miza
ti
on
with
l
ong
short
-
te
r
m
memor
y
(
IL
W
-
L
ST
M
)
m
od
el
.
The
ILW
-
LS
TM m
et
hod has
97%
pr
e
ci
sio
n.
I
n
[21]
, a
s
warm int
el
li
gen
ce al
gor
it
hm
call
ed
s
oc
ia
l spider al
gor
it
hm
(S
S
A)
is
us
e
d
for
t
he
sentime
nt
a
nalysis
within
T
witt
er
data.
Decisi
on
t
re
e,
n
aï
ve
B
ay
es,
S
V
M
,
an
d
K
N
N
a
re
the
oth
e
r
cl
ass
ifie
rs
us
e
d
in
t
his
ap
proac
h.
SSA
has
pro
duced
ve
ry
good
res
ults
in
c
ompa
rison
with
oth
e
r
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
omp E
ng
IS
S
N:
20
88
-
8708
An
i
mp
r
ove
d
r
eptil
e sear
c
h
al
go
rit
hm
-
base
d
m
ach
i
ne
le
arn
ing
f
or
se
ntime
nt an
alysis
(
Ni
te
sh
Surej
a
)
757
cl
assifi
ers.
I
n
[
22]
,
a
n
op
ti
mi
zat
ion
ap
proac
h
with
a
nt
li
on
op
ti
miza
ti
on
(
ALO)
a
nd
mo
t
h
flame
opti
miza
ti
on
(MFO)
we
re
de
sign
e
d
f
or
th
e
hate
s
peec
h
analysis
pr
ob
l
em.
The
a
ppr
oach
ac
hieve
d
acc
ur
ac
y
val
ue
w
a
s
92.1% a
nd 90.7%
with
AL
O, an
d MF
O resp
ect
ively.
The
remai
ning
sect
ion
s
of
t
he
pa
per
are
or
gan
iz
e
d
i
n
the
f
ollow
i
ng
ma
nn
e
r:
sect
io
n
2
outl
ines
the
pro
po
se
d
senti
ment
a
naly
sis
us
in
g
dee
p
le
a
rn
i
ng
wit
h
rein
forced
le
ar
ning
base
d
on
re
pt
il
e
search
al
gorith
m
(S
A
-
DLRLRS
A)
m
od
el
.
Sect
ion
3
pro
vid
e
s
a
com
prehe
ns
i
ve
ov
e
r
view
of
the
perf
or
ma
nc
e
eval
uation
of
the
pro
po
se
d
a
ppr
oach. Fi
n
al
ly,
s
ect
ion
4
se
rv
e
s
as the
conclu
di
ng
sect
ion o
f
t
he
e
ntire
work.
2.
M
ATERI
AL
AND ME
TH
ODS
This
st
udy
intr
oduces
a
ne
w
SA
-
DLRLRS
A
model
t
o
cl
a
ssify
s
ocial
me
dia
se
ntiments.
S
ocial
me
di
a
te
xt
is
pr
imari
ly
trans
forme
d
into
us
e
fu
l
da
ta
by
S
A
-
DL
RLR
SA
.
T
he
SA
-
DLRLRS
A
ap
proac
h
re
du
ce
s
data
-
pre
-
proc
e
ssing
-
de
pe
nde
nt lan
gu
a
ge p
r
ocessin
g wit
h Wor
d2Vec
w
ord
em
beddin
g.
2.1.
Prep
ar
at
i
on
of
d
ata
Data
prepa
rati
on
rem
oves
unwan
te
d
an
d
no
isy
data.
This
study
incl
ud
e
s
pr
e
-
pr
ocessin
g
ta
sk
s
su
c
h
as,
pe
rfo
rming
to
ke
nizat
ion
to
co
nvert
t
ext
into
a
w
ord
li
st,
strea
mli
nin
g
S
A
via
minimi
zi
ng
r
oot
proliferati
on,
doin
g
case
c
onve
rsion,
perfor
ming
punct
uation
r
em
oval
fro
m
the
te
xt,
performin
g
st
op
words
rem
ov
al
f
r
om
the
te
xt.
Ne
ura
l
networ
k
-
base
d
nat
ur
al
la
ng
uag
e
proces
sin
g
(
NLP)
m
ode
ls
are
popula
r
du
e
to
their acc
ur
ac
y.
Howe
ver, m
os
t
N
LP
tech
niqu
es p
e
rform
poorl
y on la
r
ge
dat
aset
s and r
e
quire w
ord
e
m
beddin
g
for
te
xt
ual
da
ta
set
s.
To
im
pro
ve
s
ys
te
m
performa
nce
a
nd
proce
ssin
g
sp
ee
d,
we
use
d
Wod
2V
ec
wor
d
embe
dd
i
ng.
Si
x
dif
fer
e
nt
dee
p
le
a
rn
i
ng
models
s
uc
h
as
CNN
[23]
,
LST
M
[
24]
,
BER
T
[
25]
,
C
NN
-
L
STM
,
BER
T
-
LST
M
,
and BER
T
-
C
N
N
a
re
us
ed
in
t
his stu
dy to
ac
cur
at
el
y cl
assif
y
se
ntiments
on s
ocial
me
dia.
2.2.
Hyper
p
ar
am
eter
t
unin
g
usin
g re
pt
il
e
sea
rc
h
algorit
hm
An
im
prov
e
d
rep
ti
le
sear
ch
al
gorithm
(RS
A)
adj
us
ts
th
e
se
m
od
el
s
hy
perpara
mete
rs
to
im
prov
e
cl
assifi
cat
ion
.
The
RSA
al
go
rithm,
pr
ese
nte
d
by
a
bual
iga
h
mimi
cs
the
hu
nting
beh
a
vi
or
of
cr
ocodile
s
in
the
wild
[7]
.
Cr
oc
od
il
es
m
ay
hu
nt
on
la
nd
a
nd
in
wate
r
as
a
mphibian
s.
T
he
basic
RS
A
al
gorithm
co
ntains
th
ree
ste
ps
.
2.2.1. Ini
tializ
at
i
on
p
hase
The
sta
rting
s
ol
ution
of
the
RSA
is
pro
du
ced
rand
om
ly
thr
ough
the
a
pp
li
cat
io
n
of
t
he
e
qu
at
i
on
1
=
+
×
(
−
)
.
In
this
set
ti
ng
,
1
represe
nts
the
it
h
sta
rtin
g
in
div
id
ual,
wh
e
reas
LB
ound
a
nd
U
Bo
un
d
ref
e
r
t
o
the
l
ow
e
r
a
nd
uppe
r
li
mit
s,
respec
ti
vely.
Also,
it
de
note
s
the
c
urren
t
it
erati
on
c
ount,
IT
re
pr
ese
nts t
he
ma
xim
um i
te
rati
on cou
nt.
2.2.2. E
ncir
cl
ing
ph
as
e
(expl
oratio
n
)
Croc
od
il
es
walk
hi
gh
an
d
wide
du
rin
g
gl
obal
search
.
Cu
rrent
num
ber
of
it
erat
ion
s
deter
mines
RS
A
search
strat
eg
y.
RS
A
walks
high
w
hen
IT
is
0.2
5
or
le
ss.
T
he
RS
A
sprawl
walks
w
hen
it
is
le
ss
tha
n
0.25 ti
mes the
IT
or lar
ger
t
ha
n
it
. T
he foll
owin
g
mat
hema
ti
cal
mo
dels
de
scribe t
he
m
ec
han
is
m:
+
1
=
{
−
×
−
×
,
≤
4
×
×
×
,
≤
4
>
4
(1)
=
×
(2)
=
−
+
(3)
=
2
×
1
×
(
1
−
1
)
(4)
=
+
−
(
)
×
(
−
)
+
(5)
=
1
∑
=
1
(6)
w
he
re,
repres
ent
the
c
urren
t
best
so
l
utio
n,
α
is
a
co
ns
ta
nt
of
0.1,
co
ntr
ols
e
xp
l
or
at
io
n
rate,
is
a
rand
om
ly
sel
e
ct
ed
i
nd
i
vidua
l.
T
o
a
void
th
e
de
nomi
nator
f
rom
bei
ng
z
ero,
t
he
re
qu
i
r
ed
mi
nimal
va
lue
is
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
omp E
ng,
V
ol.
15
, No
.
1
,
Febr
uary
20
25
:
755
-
766
758
denoted
as
ε
.
T
he
1
is
a
ra
ndom
numb
e
r
fro
m
-
1
to
1.
T
he
co
ns
ta
nt
β
is
s
et
to
0.1
a
nd,
is
a
0
–
1
ra
ndom
numb
e
r.
The
huntin
g
operato
r
in
the
ℎ
so
l
ution
is
de
no
te
d
a
s
w
hich
is
cal
culat
ed
us
in
g
(
2).
Ev
olu
ti
on
a
r
y
s
ense
(
E
VS
)
is
a
rand
om
rati
o
betwee
n
[
2,
-
2]
desc
ribe
t
he
pro
ba
bili
ty
of
decr
easi
ng
va
lues
th
r
oughout
the
it
erati
on
s,
cal
c
ulate
d
by
(4).
Bi
t
cor
res
pond
ing
to
the
difference
betwee
n
the
posit
ion
of
t
he
be
st
-
obt
ai
ne
d
so
luti
on
a
nd
th
e
posit
ion
of
t
he
c
urren
t
s
olut
ion
,
cal
c
ulate
d
by
(5).
M
sta
nd
s
to
the
mea
n
posit
io
ns
of
t
he
ℎ
so
luti
on,
comp
uted b
y (6).
2.2.3
.
Hun
tin
g
ph
as
e
(expl
oi
tation
)
I
n
R
S
A
,
c
r
oc
o
di
l
e
s
us
e
t
w
o
s
t
r
a
t
e
gi
e
s
f
or
f
or
a
gi
n
g:
hu
nt
i
ng
c
o
or
di
na
t
i
on
a
nd
c
oo
pe
r
a
t
i
on
.
W
he
n
it
<
0
.
75
a
nd
it
≥
0
.
5
,
th
e
R
S
A
pe
r
f
or
m
s
h
un
t
i
ng
c
o
or
di
n
a
t
i
on
.
W
he
n
it
<
a
nd
it
≥
0
.
75
,
a
hu
nt
i
n
g
c
oo
pe
r
a
t
i
on
s
t
r
a
t
e
gy
i
s
e
m
pl
o
ye
d
by
t
he
R
S
A
.
T
he
po
s
i
t
i
on
u
pd
a
t
i
ng
i
n
t
he
h
un
t
i
ng
ph
a
s
e
i
s
do
ne
a
s
(
7
)
:
+
1
=
{
−
×
,
≤
4
>
2
×
×
−
×
,
≤
>
3
4
(7)
RSA
gen
e
rates
the
init
ia
l
popu
la
ti
on
ra
ndoml
y
in
t
he
sea
rch
s
pace
first
and
the
n
ch
ooses
diff
e
re
nt
search
strat
egies
de
pe
nd
i
ng
on
the
num
ber
of
it
era
ti
on
s.
T
he
ps
e
udoc
od
e
for
th
e
RSA
is
sho
wn
i
n
F
ig
ur
e
1.
T
he
RSA
imp
r
ov
es
cl
assifi
er
eff
i
ci
ency
with
a
fitness
f
unct
ion.
It
assig
ns
good
-
perf
or
mi
ng
s
olu
ti
on
s
a
valu
e
gr
eat
er
tha
n
ze
ro. T
he fit
ness funct
io
n used i
n
this
scena
rio
was red
ucin
g
c
la
ssific
at
ion
e
r
ror rat
e.
(
)
=
∗
100
(8)
1
In
i
t
i
al
i
ze R
SA
p
ara
m
et
er
s
,
crea
t
e i
n
i
t
i
a
l
p
o
p
u
l
a
t
i
o
n
ra
n
d
o
m
l
y
2
W
h
i
l
e
i
t
<
I
T
3
Cal
cu
l
a
t
e t
h
e F
i
t
n
e
s
s
o
f eac
h
s
o
l
u
t
i
o
n
s
4
Fi
n
d
t
h
e Be
s
t
s
o
l
u
t
i
o
n
s
o
f
ar
5
U
p
d
a
t
e t
h
e
E
V
S
u
s
i
n
g
(2
).
6
Fo
r (i = 1
t
o
N
)
7
Fo
r (j = 1 t
o
N
)
8
Cal
cu
l
at
e
,
B
,
R
u
s
i
n
g
(3
),
(4
) a
n
d
(
6
)
9
U
p
d
a
t
e P
o
s
i
t
i
o
n
of c
ro
c
o
d
i
l
e
u
s
i
n
g
(1
) t
o
(
8
)
10
E
n
d
F
o
r
11
it
=
it
+
1
12
E
n
d
W
h
i
l
e
13
Ret
u
rn
t
h
e b
es
t
p
o
s
i
t
i
o
n
a
n
d
fi
t
n
es
s
Figure
1. Pse
udoc
ode
of the
RSA al
gorithm
2.3.
Rein
f
orc
ement
l
e
arnin
g
Re
inforceme
nt
le
arn
in
g
has
fou
nd
e
xten
s
ive
ap
plica
ti
on
in
var
i
ous
fiel
ds
for
probl
em
-
so
l
ving.
Re
inforceme
nt
learni
ng
(RL)
is base
d on
t
he
idea that a
n
a
ge
nt
cha
nges
th
e stat
e o
f
the e
nv
i
ronme
nt
by
act
ing
on
it
a
nd
rece
ives
a
re
ward
base
d
on
t
he
resu
lt
s
of
t
he
act
ion
.
The
t
wo
disti
nct
ki
nd
s
of
rei
nforce
ment
le
arn
in
g
(RL)
are
val
ue
a
nd
po
li
cy
-
based
le
arn
i
ng.
Q
le
arn
i
ng
(
QL)
is
a
value
-
ba
sed
RL
meth
od.
It
is
a
model
-
f
ree,
w
hich
mea
ns
th
at
the
age
nt
le
arn
s
how
to
m
ake
the
ri
gh
t
c
ho
ic
es
in
a
M
a
rko
vian
domai
n
[
26]
.
The
a
gen
t
perf
orms
the
act
io
n
w
it
h
the
highest
ex
pected
Ql
valu
e
duri
ng
le
ar
ning.
one
-
ste
p
Q
le
ar
ning
is
a
very
sim
ple
ty
pe
of
Q
le
ar
ning.
I
n
this,
Ql
va
lue
is
c
hange
d
i
n
a
sin
gle
st
ep
acco
rd
i
ng
t
o
t
he
sta
te
-
act
ion
pair.
This
wor
k
em
ploys
a
one
-
ste
p
Q
le
ar
ning
method
ology.
Each
sta
te
-
ac
ti
on
pai
rs
rew
a
r
d
updates
t
he
Q
ta
ble
con
ti
nu
ously
usi
ng (9).
(
,
)
←
(
1
−
)
(
,
)
+
(
+
1
+
(
+
1
,
+
1
)
)
(9)
The
s
ymb
ols
an
d
Lr
de
no
te
the
disc
ount
f
act
or
a
nd
rate
of
le
a
rn
i
ng,
res
pecti
vely
.
Bot
h
nu
mb
e
rs
a
re
within
the
range
of
0
to
1.
T
he
(
,
)
ref
e
rs
to
the
Ql
va
lue
obta
ined
by
pe
rformi
ng
a
ct
ion
in
the
current
sta
te
.
O
n
t
he
oth
e
r
hand,
(
+
1
,
+
1
)
re
pr
ese
nts
t
he
hig
he
st
antic
ipat
ed
Ql
val
ue
i
n
the
Q
ta
ble
wh
e
n
exec
utin
g
act
io
n
+
1
in
sta
te
+
1
.
It
is
cr
ucial
to
note
t
hat
a
n
inc
reased
rate
of
le
ar
ning
(
)
pro
mp
t
s
the
al
gorithm
to
acq
uire
knowle
dge
f
rom
the
antic
ipate
d
Ql
value
,
w
he
reas
a
dec
rea
sed
rate
of
le
a
rn
i
ng
prom
pts
the
al
gorithm
t
o
capi
ta
li
ze
on
the p
r
evio
us
Ql
valu
e.
The
refore
,
t
he
rate of
le
ar
ni
ng
is used
t
o
s
trike
a
balance
betwe
en
util
iz
ing
ex
plo
it
ing
kn
ow
l
edg
e
an
d
e
xp
l
or
i
ng
ne
w
op
port
un
it
ie
s.
Q
le
arn
i
ng
ps
e
udo
-
cod
e
is
sh
ow
n
i
n
Fig
ure
2.
Q
le
a
rn
i
ng
ra
ndom
l
y
a
s
sign
s
values
to
the
Q
a
nd
re
w
ard
ta
bles
.
A
st
at
e
is
the
n
ra
ndoml
y
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
omp E
ng
IS
S
N:
20
88
-
8708
An
i
mp
r
ove
d
r
eptil
e sear
c
h
al
go
rit
hm
-
base
d
m
ach
i
ne
le
arn
ing
f
or
se
ntime
nt an
alysis
(
Ni
te
sh
Surej
a
)
759
chosen
by
the
al
gorithm.
A
s
per
li
ne
s
4
and
5,
t
he
al
gorith
m
ma
xim
iz
es
the
sta
te
s
fu
t
ur
e
r
ewa
rd.
This
modifie
s th
e Q t
able,
rew
a
rd t
able, a
nd n
e
w st
at
e.
1
In
i
t
i
al
i
ze
Q
-
t
a
b
l
e a
n
d
rew
ar
d
t
ab
l
e
w
i
t
h
ra
n
d
o
ml
y
2
Ch
o
s
e r
an
d
o
m s
t
a
t
e
3
W
h
i
l
e (
T
erm
i
n
at
i
o
n
c
ri
t
eri
a n
o
t
m
et
)
4
Ch
o
o
s
e t
h
e b
es
t
ac
t
i
o
n
fo
r t
h
e c
u
rre
n
t
s
t
at
e
fro
m
Q
t
a
b
l
e
5
E
x
ec
u
t
e t
h
e ac
t
i
o
n
an
d
t
h
e re
w
ar
d
+
1
6
G
et
t
h
e
n
e
w
s
t
a
t
e
+
1
7
U
p
d
at
e
Q
t
a
b
l
e
u
s
i
n
g
(9)
8
←
+
1
9
E
n
d
W
h
i
l
e
Figure
2. Pse
udoc
ode
of the
Q
le
ar
ning al
gorith
m
2.4.
The
devel
op
men
t of t
he
proposed
S
A
-
DLRL
RSA
2.4.1.
M
ot
i
vat
ion
The
t
yp
ic
al
R
SA
te
c
hniq
ue
fin
ds
s
olu
ti
ons
thr
ough
ex
pl
or
at
io
n
a
nd
e
xploit
at
ion.
Indi
viduals
us
e
eff
ic
ie
nt
an
d
be
ll
y
-
wal
king
m
et
hods
to
ex
pl
or
e
new
a
ns
we
rs.
H
unti
ng
operati
on
s
a
re
c
oor
din
at
e
d
t
o
find
the
best
op
ti
m
al
s
olu
ti
ons
duri
ng
e
xploit
at
ion
.
The
al
gorith
ms
ca
pacit
y
t
o
cha
nge
direct
ion
is
li
mit
ed
because
exp
l
or
at
io
n
oc
cur
s
in
t
he
fir
st
half
of
it
era
ti
on
s
a
nd
e
xp
l
oitat
ion
i
n
the
seco
nd.
RS
A
s
ina
bili
ty
to
adjust
it
erati
vely
ma
kes
it
pr
on
e
t
o
l
ocal
opti
ma.
T
hus,
a
de
fi
ned
sea
rc
h
pat
te
rn
does
no
t
gu
a
ra
ntee
the
op
ti
mal
value
.
Re
in
f
orcement
le
ar
ning
an
d
ada
ptive
sea
rch
fin
d
t
he
global
mi
nim
um
ef
fici
ently.
Ra
nd
om
oppo
sit
ion
-
base
d
le
ar
ning
incr
eases
p
op
ul
at
ion
v
ariat
io
n
to
fin
d
al
te
rn
a
ti
ve
an
swe
rs.
T
hese f
eat
ures
of
th
e
RL
an
d
R
OBL
mo
ti
vate
d us
t
o use t
hem for
impro
ving the
RSA
for
e
ff
ic
i
ent se
ntiment a
nalysis.
2.4.2. T
he S
A
-
DLRL
RSA s
t
ructure
SA
-
DLRLRS
A
us
es
t
he
enti
re
searc
h
sp
ac
e
as
it
s
e
nv
iro
nme
nt
an
d
e
very
so
l
ution
(in
di
vid
ual
)
as
an
RL
trai
ni
ng
a
ge
nt.
The
Q
le
a
rn
i
ng
al
gorit
hm
s
witc
hes
bet
ween
e
xp
l
or
at
i
on
a
nd
e
xp
l
oitat
ion
.
T
he
Q
va
lue
of
the
sta
te
-
act
io
n
pair
is
updat
ed
by
the
Q
le
arn
i
ng
al
gorith
m
us
i
ng
the
hi
gh
e
st
fitness
va
lue
an
d
t
he
a
ver
a
ge
fitness
val
ue
f
rom
ea
rlie
r
it
erati
on
s
.
A
ta
bl
e
de
scri
bed
as
a
re
ward
ta
bl
e
is
us
e
d
to
gi
ve
the
punis
hme
nts
or
incenti
ves
to
the
s
olu
ti
ons
(a
gen
ts
)
base
d
on
it
s
act
io
ns
a
nd
sta
tus.
T
he
pro
po
se
d
SA
with
RL
an
d
r
andom
sta
te
le
arn
in
g
c
on
sist
s
of
th
ree
act
ion
s
that
a
r
e
deter
mine
d
by
the
rate
of
th
e
ex
plorat
ion
∅
:
increasin
g
t
he
rate
of
t
he
e
xp
l
or
a
ti
on
,
decre
asi
ng
the
rate
of
the
ex
plorat
io
n,
or
mainta
in
ing
c
urre
nt
rat
e.
I
n
the
f
ollo
wing
it
erati
on
value
of
∅
is
adj
us
te
d
c
onside
rin
g
the
cu
rr
e
nt
hi
gh
e
st
fitnes
s
a
nd
c
umulat
ive
aver
a
ge
fitness
us
i
ng
(10).
∅
+
1
in
dicat
es
the
rate
of
e
xplorati
on
i
n
the
fo
ll
owin
g
it
era
ti
on
,
∆
represe
nts
the
inc
rem
ental
val
ue,
an
d
(
)
represe
nts
the
fitne
ss
of
the
be
st
po
sit
io
n
in
the
cu
rr
e
nt
it
erati
on
.
T
he
M
r
epr
ese
nts
the
mean
fitness
of
the
fit
s
olu
ti
on
(in
div
id
uals
)
fou
nd
t
hu
s
fa
r,
co
mputed
usi
ng
(
11)
.
U
p
t
o
this
poi
nt,
n
it
erati
on
s
ha
ve
been
done.
T
o
cal
c
ulate
the
weigh
te
d
facto
r
f
or
the
fitt
est
ind
iv
i
du
al
at
it
erati
on
,
use
the
form
ul
a
=
/
.
∅
+
1
=
{
∅
∗
(
1
+
∆
)
(
)
>
∅
∗
(
1
−
∆
)
(
)
<
∅
ℎ
(10)
=
1
∑
=
1
(11)
He
re
,
represe
nt
s
the
c
urren
t
it
erati
on
a
nd
IT
represe
nts
t
he
t
otal
numb
e
r
of
it
erati
ons.
It
is
im
portant
to
note
that
the
m
os
t
phys
ic
al
ly
fit
i
ndivid
uals
i
n
re
cent
ti
mes
ha
ve
a
gr
eat
e
r
im
pa
ct
on
t
he
cal
c
ulati
on
of
t
he
va
lue
of
M
.
s
pecifica
ll
y,
if
t
he
ac
hiev
ed
fitness
is
hi
gh
e
r
tha
n
the
ave
rag
e
fitne
s
s,
the
al
gorith
m
s
hould
f
oc
us
on
a
small
er
sea
rch
sp
ace
an
d
im
pro
ve
the
acq
uire
d
s
olu
ti
ons
.
Alte
r
nativel
y,
the
al
gorith
m
ex
pa
nds
it
s
searc
h
reg
i
on
in
orde
r
to
disco
ve
r
novel
s
olu
ti
ons
a
nd
pre
ven
t
l
oc
al
op
ti
ma
.
I
n
s
um
ma
r
y,
t
he
first
sce
nar
io
de
scribe
d
in
(10
)
t
yp
ic
al
l
y
occ
urs
w
hen
the
a
ge
nt
ac
hie
ves
a
hi
gh
e
r
le
vel
of
fitness
over
the
mean
fitness.
I
n
the
se
cond
sit
uation,
t
he
agen
t
’s
fitne
ss
sta
rts
to
decli
ne
in
c
ompar
ison
t
o
the
pri
or
a
gen
t
’s
e
xperie
nce.
T
he
SA
-
DLRLRS
A has
thr
ee
stat
es,
de
no
te
d
as
st
=
{
1
,
−
1
,
0
)
wh
i
ch
c
orres
pond
to the acti
viti
es d
esc
ribe
d
in
(12).
The
re
ward
ta
ble
i
n
t
his
w
ork
a
ssig
ns
a
po
sit
ive
val
ue
of
(+1)
t
o
sta
te
=
1
and
a
neg
at
ive
value
of
(
-
1)
to
al
l
ot
her
sta
te
s.
I
f
the
fitne
ss
gain
ed
at
it
erati
on,
it
is
good
tha
n
the
mea
n
fit
ness
of
the
la
st
−
1
it
erati
on
s,
the
n
the
pr
ese
nt
sta
te
is
eq
ual
to
1.
In
(
13)
il
lustrate
s
t
he
rew
a
r
d
a
ppr
oac
h.
He
re,
is
the
sta
t
e
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
omp E
ng,
V
ol.
15
, No
.
1
,
Febr
uary
20
25
:
755
-
766
760
achieve
d
by
t
he
in
div
i
du
al
(ag
e
nt)
at
it
er
at
io
n
it
.
F
ur
th
erm
or
e,
t
he
s
uggeste
d
S
A
-
DLRLRS
A
m
et
hod
pr
eci
sel
y
adj
ust
s
the
rate
of
le
arn
in
g
acc
ordi
ng
t
o
the
a
ccum
ulate
d
pe
rformance
,
as
this
fact
or
great
ly
influ
e
nces
t
he
at
ta
inment
of
t
he
ideal
so
l
ution.
Wh
e
n
t
he
rate
of
le
ar
ning
is
near
to
on
e,
the
fr
e
sh
col
le
ct
ed
inf
or
mati
on
sign
i
ficantl
y
i
nfl
uen
ces
the
f
uture
re
wa
rd.
At
a
l
ow
le
arn
i
ng
rate,
the
value
of
e
xisti
ng
inf
or
mati
on
surpass
es
that
of
new
l
y
ac
qu
i
re
d
in
formati
on.
In
orde
r
to
opti
mize
the
ou
tc
ome
,
the
le
arn
i
ng
rate
is
dynamical
ly
decr
ease
d
at
e
ach
i
te
rati
on
usi
ng
(14
).
Her
e
,
an
d
represe
nt
the
sta
rtin
g
and
final
values
of the
le
arn
i
ng r
at
e,
res
pecti
vely
.
=
(
(
)
−
)
,
(
)
=
{
1
>
1
−
1
<
1
0
ℎ
(12)
=
{
+
1
=
1
−
1
ℎ
(13)
=
+
2
−
+
2
∙
cos
(
(
1
−
)
)
)
(14)
The
ra
ndom
opposit
io
n
-
base
d
le
a
rn
i
ng
(ROBL)
te
ch
nique
is
inc
orp
orat
ed
i
nto
the
S
A
-
DLRLRS
A
al
gorithm
to
dyna
m
ic
al
ly
as
s
ist
in
a
voidin
g
the
prob
le
m
of
bein
g
stuc
k
in
sub
op
ti
mal
s
olu
ti
ons.
ROB
L
is
a
te
chn
iq
ue
est
a
blishe
d
by
[
27]
that use r
a
ndomi
zat
ion
to
e
nhance t
he
pe
rformance
of
opti
mize
d
bee life
(
OBL
)
methods
def
i
ne
d
as:
=
+
−
×
,
=
1
,
2
,
…
…
…
.
,
.
He
re
,
a
nd
deno
te
the
antit
hetic
al
an
d
init
ia
l
so
luti
on
s
,
wh
e
reas
and
represent
the
mi
nimum
an
d
ma
xim
um
li
mit
s
of
t
he
var
ia
bles.
Fig
ure
3
dep
ic
ts
th
e
propose
d
S
A
-
DLRLR
SA
a
nd
pro
vid
es
a
more
com
preh
ensive
e
xp
la
na
ti
on
of
how
t
h
e alg
ori
thm e
xplo
res
th
e g
lo
bal s
olu
ti
on.
1
In
i
t
i
al
i
ze R
SA
p
ara
m
et
er
s
,
crea
t
e i
n
i
t
i
a
l
p
o
p
u
l
a
t
i
o
n
ra
n
d
o
m
l
y
.
Set
t
h
e
s
t
a
t
e
s
t
= (
st
1
,
s
t
2
,
s
t
3
}
a
n
d
a
ct
i
o
n
at
= (
at
1
,
a
t
2
,
a
t
3
).
In
i
t
i
al
ze
Q
t
a
b
l
e an
d
r
ew
ar
d
t
ab
l
e,
ra
n
d
o
m
l
y
se
l
ec
t
p
res
en
t
s
t
a
t
e
2
W
h
i
l
e
i
t
<
I
T
3
Cal
cu
l
a
t
e t
h
e F
i
t
n
e
s
s
o
f eac
h
s
o
l
u
t
i
o
n
s
4
Fi
n
d
t
h
e B
es
t
s
o
l
u
t
i
o
n
s
o
far
5
U
p
d
a
t
e t
h
e
E
V
S
u
s
i
n
g
(2
).
6
U
p
d
a
t
e ra
t
e of
ex
p
l
o
rat
i
o
n
φ
u
s
i
n
g
(
9
)
7
Fo
r (
i
= 1
t
o
N
)
8
Fo
r (
j
= 1 t
o
N
)
9
If
r
a
n
d
<
∅
10
If
r
a
n
d
< 0.
5
11
U
p
d
at
e
cro
co
d
i
l
e
s
’ Po
s
i
t
i
o
n
u
s
i
n
g
fir
s
t
par
t
o
f
(1
)
12
Else
13
U
p
d
at
e
cro
co
d
i
l
e
s
’Po
s
i
t
i
o
n
u
s
i
n
g
se
co
n
d
par
t
o
f
(1
)
14
E
n
d
I
f
15
E
l
s
e
16
If
r
a
n
d
< 0.
5
17
U
p
d
at
e cr
o
c
o
d
i
l
es
’ P
o
s
i
t
i
o
n
u
s
i
n
g
first p
art
o
f
(7
)
18
Else
19
U
p
d
at
e cr
o
c
o
d
i
l
es
’ P
o
s
i
t
i
o
n
u
s
i
n
g
sec
o
n
d
p
art
o
f
(7
)
20
E
n
d
If
21
E
n
d
If
22
Co
mp
u
t
e
(
+
1
)
u
s
i
n
g
R
O
B
L
23
E
n
d
Fo
r
24
If
(
(
+
1
)
)
i
s
b
et
t
er
t
h
an
(
(
+
1
)
)
25
(
+
1
)
=
(
+
1
)
26
E
n
d
If
27
U
p
d
at
e
t
h
e pr
es
e
n
t
s
t
a
t
e
u
s
i
n
g
(
1
2
)
28
Co
mp
u
t
e re
w
ar
d
u
s
i
n
g
(1
3
)
29
U
p
d
at
e
Q
t
ab
l
e
u
s
i
n
g
(
9
)
30
E
n
d
F
o
r
31
it
=
it
+
1
32
U
p
d
a
t
e ra
t
e
o
f le
arn
i
n
g
u
s
i
n
g
(1
4
)
33
E
n
d
W
h
i
l
e
34
Ret
u
rn
t
h
e b
es
t
p
o
s
i
t
i
o
n
a
n
d
fi
t
n
es
s
Figure
3. Pro
pose
d
S
A
-
DLR
LRSA al
gorith
m
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
omp E
ng
IS
S
N:
20
88
-
8708
An
i
mp
r
ove
d
r
eptil
e sear
c
h
al
go
rit
hm
-
base
d
m
ach
i
ne
le
arn
ing
f
or
se
ntime
nt an
alysis
(
Ni
te
sh
Surej
a
)
761
The
al
gorith
m
sta
rts
by
asse
ssing
in
div
i
dual
fitness
a
nd
fin
ding
t
he
be
st
so
l
ution.
Ne
xt,
t
he
m
os
t
adv
a
ntage
ous
act
ion
f
r
om
th
e
Q
ta
ble
is
sel
ect
ed,
a
nd
t
he
exp
l
or
at
io
n
rate
is
cha
ng
e
d
usi
ng
(
10).
T
he
rand
om
numb
e
r
an
d
e
xp
l
or
at
io
n
rate
dete
rmin
e
whet
her
ex
plorat
ion
or
e
xp
l
oitat
ion
ge
ne
rates
a
ne
w
s
olu
ti
on
.
T
he
recently
acq
ui
red
so
l
utio
n
is
us
e
d
t
o
c
ompu
te
t
he
rever
s
e
so
l
ution
usi
ng
R
OBL.
A
fter
t
hen,
the
el
it
ism
mecha
nism
determines
t
he
be
st
so
luti
on
f
rom
the
re
ver
se
and
ne
w
so
l
ution
s
.
F
ollow
i
ng
this,
updates
oc
cur
t
o
the
Q
ta
ble,
re
ward
ta
ble,
a
nd
prese
nt
sta
te
.
The
rate
of
l
earn
i
ng
is
m
odifie
d
a
fter
ea
ch
it
erati
on
u
nt
il
th
e
stoppin
g req
uir
ements a
re
met.
3.
RESU
LT
S
AND DI
SCUS
S
ION
3.1
.
As
sessme
nt
i
ndic
ators
All
the
a
ssess
ment
i
nd
ic
at
or
s
us
e
d
in
t
his
work
ar
e
descri
bed
in
Table
1,
w
her
e
tr
ue
po
sit
ives
(
TP),
true
ne
gatives
(TN),
f
al
se
po
sit
ives
(
FP),
a
nd
false
ne
gatives
(FN
)
a
re
us
e
d
t
o
c
al
cul
at
e
their
value
s.
T
he
model i
s e
valu
at
ed
usi
ng acc
ur
ac
y, F
1
sc
or
e
, precisi
on, a
nd
r
ecal
l met
rics.
Table
1.
Asses
sment
i
nd
ic
at
ors
Metr
ic
Descripti
o
n
Fo
rm
u
la
Accuracy
Prop
o
rtion
of corr
e
ctly
classified
ins
ta
n
ces
am
o
n
g
the
to
tal ins
tan
ces.
=
(
+
)
/
(
+
+
+
)
Precisio
n
Prop
o
rtion
of tr
u
e
p
o
sitiv
e predictio
n
s am
o
n
g
all
p
o
sitiv
e predictio
n
s.
=
/
(
+
)
Recall
Prop
o
rtion
of tr
u
e
p
o
sitiv
e predictio
n
s am
o
n
g
all
actu
al po
sitiv
e ins
tan
ces.
=
/
(
+
)
F1
Score
Har
m
o
n
ic m
e
an
of
precisio
n
and
r
eca
ll,
b
alan
cin
g
b
etween p
recisio
n
an
d
r
ecall.
1
=
(
2
×
∗
)
/
(
+
)
3.2
.
Ex
peri
ment
al results
Table
2
s
how
s
the
pa
rameter
com
bin
at
io
ns
of
al
l
cl
assifi
ers
in
this
i
nvest
igati
on
.
Tot
al
po
sit
ive,
neu
t
ral,
a
nd
ne
gative
tw
eet
s
a
re
22,93
7,
21,
938,
a
nd
5,1
83,
resp
ect
ivel
y.
T
his
project
ai
m
s
to
c
reate
a
B
i
tc
oin
sentime
nt
a
nal
ys
is
m
odel
us
i
ng
dee
p/mac
hi
ne
l
ear
ning
me
thods.
I
n
order
to
e
nh
a
nce
m
od
el
e
ff
ect
i
veness,
we
enh
a
nce
d
de
ep
le
arn
in
g/mac
hi
ne
le
arn
i
ng
pa
rameters
util
iz
i
ng
a
n
e
nh
a
nce
d
re
ptil
e
searc
h
al
gorith
m
(R
SA
)
.
We
a
dd
e
d
rein
forceme
nt
le
ar
ning
to
the
RS
A
al
go
rithm
t
o
boos
t
it
s
e
xplorati
on
a
nd
use
of
i
nfo
rmati
on.
The
models
accu
ra
cy,
F
1
sc
or
e,
pr
eci
sio
n,
a
nd
recall
were
evaluated
.
T
hese
metri
cs
pro
vid
e
a
co
mp
le
te
assessme
nt
of
the
m
odel
s
a
bi
li
ty
to
cl
assif
y
t
weet
se
ntiments
a
s
posit
ive,
ne
gative,
or
ne
utral.
O
ur
st
udy
fou
nd
si
gn
i
ficant
dif
fer
e
nces
in
de
e
p
le
ar
nin
g
(DL)
or
m
achine
le
ar
ning
(
M
L
)
al
gorit
hm
e
ff
ic
ac
y.
T
able
3
com
par
es
m
odel
s p
er
f
or
ma
nc
e facto
rs.
Table
2.
Para
m
et
er s
et
ti
ng of t
he
m
odel
s
Para
m
eters
Valu
es
CNN
LST
M
BERT
CNN
-
LS
TM
BERT
-
CN
N
BERT
-
LST
M
Act.
f
u
n
ctio
n
So
ftm
ax
So
ftm
ax
So
ftm
ax
So
ftm
ax
So
ftm
ax
So
ftm
ax
Batch
size
128
128
128
128
128
128
Op
tim
ize
r
Ad
am
Ad
am
Ad
am
Ad
am
Ad
am
Ad
am
Epo
ch
s
10
10
10
10
10
10
Lear
n
in
g
r
ate
--
-
-
--
-
-
0
.00
1
--
-
--
-
-
Table
3.
Res
ults at
ta
ined b
y
th
e
m
odel
s
Mod
els
Accuracy
(
%)
Precisio
n
(
%)
Recall (%
)
F1
sco
re
(%
)
CNN
9
3
.10
9
3
.96
9
2
.10
9
3
.02
LST
M
9
4
.97
9
6
.19
9
5
.85
9
6
.02
BERT
9
6
.84
9
7
.44
9
7
.18
9
7
.31
CNN
-
LS
TM
9
6
.38
9
6
.44
9
6
.22
9
6
.33
BERT
-
CN
N
9
7
.22
9
7
.12
9
6
.90
9
7
.01
BERT
-
LST
M
9
8
.32
9
7
.28
9
7
.18
9
7
.23
Figur
e
4
s
hows
the
graphs
twe
et
numerically.
Figur
e
5
s
hows
the
res
ults
graphically.
Figur
e
6
s
hows
the
CNN
class
ifi
er
achieved
96.
5%
accuracy,
93.
9%
precis
ion,
92
.
1%
recall,
and
93.
0%
F1
s
core
.
Figur
e
7
s
hows
tha
t
the
LS
TM
classifier
had
94.
9%
accuracy,
96.
1
%
precis
ion,
9
5.
8%
recall,
and
96.
0%
F1
s
core.
Figur
e
8
s
hows
that
the
BERT
class
ifi
er
had
96.
8%
accuracy,
97.
4
%
precis
ion,
9
7.
1%
recall,
and
97.
3%
F1
s
c
ore.
The
CNN
-
LS
TM
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
omp E
ng,
V
ol.
15
, No
.
1
,
Febr
uary
20
25
:
755
-
766
762
class
ifi
er
in
Fig
ure
9
had
96.
3
%
accuracy,
96.
4%
precis
ion,
9
6.
9%
recall,
an
d
97.
0%
F1
s
core.
Figur
e
10
s
h
ows
the
BERT
-
CN
N
cla
s
s
ifi
ers
97
.
2%
accuracy,
97.
1%
precis
ion,
92.
1%
recall,
and
9
3.
0%
F1
s
core.
Figur
e
11
s
how
s
that
the
BERT
-
LS
T
M
class
i
fier
ou
tperf
ormed
all
other
models
.
The
model
had
98.
3%
accuracy,
97.
2%
precis
ion,
97.
1%
recall,
a
nd
97.
2%
F1
s
c
ore.
BERT
-
LSTM
was
re
s
ilient
and
eff
ective
at
ass
es
s
ing
Bitcoin
tweet
s
ent
iment.
Our
res
earch
also
compares
our
findi
ngs
to
earlier
s
tudies
[16]
–
[18]
,
[21]
.
Our
w
ork
s
hows
tha
t
deep
learning/mac
hi
ne
learning
impr
oves
class
ifi
c
ation
accurac
y
through
real
-
time
tweet
detec
tion
and
analysis
.
It
is
also
impor
tant
t
o
highlight
constr
aints
like
dataset
s
ize
and
bi
ases
that
may
r
estr
ict
outcomes
of
the
res
earch
.
This
may
requir
e
intelligent
hype
rpar
ameter
adjust
ing,
us
ing
domain
-
s
pecific
characteris
tics
based
on
ex
pert
knowledg
e,
or
adding
extern
al
data
to
augme
nt
the
dataset.
We
can
impr
ov
e
s
entiment
ana
lys
is
by
incorpor
ating
these
variables
and im
proving
categorizati
on
a
lgori
thms
.
Figure
4. N
umber
of T
weets
Figure
5. Re
su
l
ts o
btained
Figure
6.
CN
N
conf
us
io
n mat
rix
Figure
7. LST
M
c
onf
us
io
n m
at
rix
0
5000
10000
15000
20000
25000
Positi
ve
Neutral
Negativ
e
Number
of
Tw
eets
Tw
eet Types
T
w
ee
ts
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
omp E
ng
IS
S
N:
20
88
-
8708
An
i
mp
r
ove
d
r
eptil
e sear
c
h
al
go
rit
hm
-
base
d
m
ach
i
ne
le
arn
ing
f
or
se
ntime
nt an
alysis
(
Ni
te
sh
Surej
a
)
763
Figure
8. BER
T
co
nfusi
on m
at
rix
Figure
9.
CN
N
-
LST
M
c
onf
usi
on
matri
x
Figure
10.
BE
RT
-
CN
N
c
onf
us
io
n mat
rix
Figure
11. BE
RT
-
LST
M
c
on
fu
si
on matri
x
4.
CONCL
US
I
O
N
This
wor
k
pr
es
ents
the
de
vel
opme
nt
of
t
he
S
A
-
DLRLRS
A
al
gorithm
f
or
s
entime
nt
cl
assi
ficat
ion
of
Bi
tc
oin
tweet
s
.
T
he
obje
ct
ive
of
the
S
A
-
DLRLR
SA
te
chn
i
qu
e
is
t
o
de
velo
p
a
n
aut
om
at
ed
a
r
ti
fici
al
intel
li
gen
ce
m
od
el
that
acc
urat
el
y
cl
assi
fies
the
tweet
s
a
s
po
sit
ive
,
ne
gative,
or
neut
ral
in
te
r
ms
of
thei
r
sentime
nt
t
oward
s
Bi
tc
oin
s
.
The
S
A
-
DL
RLR
SA
te
c
hniqu
e
co
ns
ist
s
of
f
our
sta
ge
s:
data
pr
e
pa
r
at
ion
,
pr
e
processi
ng,
sentime
nt
cl
ass
ific
at
ion
bas
ed
on
dee
p
le
arn
i
ng
or
mac
hin
e
le
a
rn
i
ng,
an
d
hyperpa
r
amet
er
tun
in
g
base
d
on
im
prov
e
d
RSA.
Th
e
RS
A
te
c
hn
i
qu
e
is
em
ployed
f
or
hy
perpara
mete
r
tu
nnin
g
in
order
to
enh
a
nce
the
r
esults
of
the
DL
or
M
L
al
gorithms
.
T
he
eff
ect
ive
ness
of
the
S
A
-
DL
RLR
SA
a
ppr
oa
ch
is
confirme
d
by
te
sti
ng
it
on
t
he
Bi
tc
oin
tweet
s
dataset
obta
in
ed
from
the
K
agg
le
re
posit
ory.
Th
e
e
xperi
mental
resu
lt
s
s
howe
d
that
the
S
A
-
D
LRLR
SA
met
hod
outpe
rformed
ot
her
pres
ent
al
gorith
ms
in
se
ve
ral
me
asur
e
s.
This
a
nal
ys
is
offer
s
a
valua
ble
i
ns
ig
ht
i
nto
t
he
pu
blics
emotio
ns
to
wa
rd
s
Bi
tc
oins
on
Twitt
er,
e
na
bling
a
bette
r
c
ompre
he
ns
io
n
a
nd
e
va
luati
on
of
it
s
i
nfl
ue
nce
a
nd
pe
rcep
ti
on
a
mon
g
us
er
s.
I
n
the
fu
t
ur
e,
the
pres
ented
model
has
t
he
po
te
ntial
to
be
exp
a
nded
for
c
la
ssify
in
g
vie
w
s
relat
ed
to
var
iou
s
to
pics.
F
urt
hermo
re,
the
r
e
are
sever
al
ad
diti
on
al
strat
egie
s
that
can
be
em
ployed
t
o
e
nhance
th
e
pe
rfo
rma
nc
e
of
t
he
sug
gested
SA
-
DLRLRS
A
m
odel
.
REFERE
NCE
S
[1]
C.
A.
Igles
ias
an
d
A.
Moren
o
,
“Sen
tim
en
t
an
aly
sis
for
so
cial
m
ed
ia
,”
A
p
p
lied
S
cien
ces
,
v
o
l.
9
,
n
o
.
2
3
,
p
p
.
1
–
4
,
No
v
.
2
0
1
9
,
d
o
i: 10
.3390
/ap
p
9
2
3
5
0
3
7
.
[2]
Q.
Tul
et
al.
,
“Sent
im
en
t analy
sis
us
in
g
deep
lear
n
in
g
tec
h
n
iq
u
es: a r
ev
iew,
”
Inter
n
a
tio
n
a
l Journ
a
l of
Adva
n
ced C
o
mp
u
ter S
cien
ce
a
n
d
A
p
p
lica
tio
n
s
,
v
o
l.
8
,
n
o
.
6
,
p
p
.
4
2
4
–
4
3
3
,
Jan
.
2
0
1
7
,
do
i: 10
.14
5
6
9
/IJACSA.2
0
1
7
.08
0
6
5
7
.
[3]
D.
Li,
R
.
Rzep
k
a,
M.
Ptaszy
n
sk
i,
an
d
K.
Araki,
“H
EM
OS:
a
n
o
v
el
d
eep
l
earnin
g
-
b
ased
fine
-
g
rained
h
u
m
o
r
d
etectin
g
m
eth
o
d
for
sen
tim
en
t
an
aly
sis
o
f
so
cial
m
ed
ia,”
Info
rma
tio
n
Pro
cess
in
g
a
n
d
Ma
n
a
g
ement
,
v
o
l.
5
7
,
n
o
.
6
,
No
v
.
2
0
2
0
,
d
o
i: 10
.1016
/j.ipm.20
2
0
.1
0
2
2
9
0
.
[4]
M.
Sin
an
et
a
l.
,
“
An
aly
sis
o
f
th
e
m
ath
em
atical
m
o
d
el
o
f
cu
tan
eo
u
s
leis
h
m
an
iasis
d
iseas
e,
”
Alexan
d
ria
Eng
i
n
eerin
g
Jo
u
r
n
a
l
,
v
o
l.
7
2
,
p
p
.
1
1
7
–
1
3
4
,
Ju
n
.
2
0
2
3
,
d
o
i:
10
.10
1
6
/j.aej.
2
0
2
3
.03
.06
5
.
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
omp E
ng,
V
ol.
15
, No
.
1
,
Febr
uary
20
25
:
755
-
766
764
[5]
A.
R.
Path
ak
,
B.
Ag
arwa
l,
M
.
P
an
d
ey
,
an
d
S.
R
au
taray,
“
Ap
p
licatio
n
o
f
d
eep
lea
r
n
in
g
ap
p
roach
es
for
se
n
tim
en
t
an
aly
sis
,
”
in
Algo
rith
ms fo
r Intellig
en
t Sys
tems
,
2
0
2
0
,
p
p
.
1
–
3
1
.
[6]
D.
W
u
,
S.
W
an
g
,
Q.
Liu,
L
.
Ab
u
alig
ah
,
an
d
H.
Ji
a,
“An
i
m
p
rov
ed
teachin
g
-
learnin
g
-
b
ased
o
p
tim
izati
o
n
alg
o
rithm
with
reinforce
m
en
t
lear
n
in
g
strateg
y
for
so
lv
in
g
o
p
tim
izatio
n
p
rob
lem
s,”
Compu
ta
tio
n
a
l
Intellig
en
ce
a
n
d
Ne
u
ro
s
cien
ce
,
v
o
l.
2
0
2
2
,
p
p
.
1
–
2
4
,
Mar
.
2
0
2
2
,
d
o
i: 10
.11
5
5
/2
0
2
2
/1
5
3
5
9
5
7
.
[7]
L.
Ab
u
alig
ah
,
M.
A.
Elazi
z,
P.
Su
m
ari,
Z
.
W
.
Gee
m
,
an
d
A.
H.
Gan
d
o
m
i,
“Rep
tile
sea
rc
h
alg
o
rithm
(RSA
):
a
n
atu
re
-
in
sp
ire
d
m
eta
-
h
eu
ristic op
tim
ize
r,
”
Expert
Sys
tems with
A
p
p
lica
tio
n
s
,
v
o
l.
1
9
1
,
Ap
r.
20
2
2
,
d
o
i: 10
.101
6
/j.eswa.20
2
1
.11
6
1
5
8
.
[8]
H.
T
.
Hal
awani,
A.
M
.
Mash
raqi,
S.
K.
Bad
r,
an
d
S.
Alk
h
alaf,
“Au
to
m
ated
sen
ti
m
en
t
a
n
aly
sis
in
so
cial
m
ed
ia
u
sin
g
Har
ri
s
Hawks
o
p
tim
isa
ti
o
n
an
d
d
eep
learnin
g
tech
n
iq
u
es,”
Alexan
d
ria
Eng
in
eerin
g
Jo
u
rnal
,
v
o
l.
8
0
,
p
p
.
4
3
3
–
4
4
3
,
Oct.
2
0
2
3
,
d
o
i: 10
.1016
/j.aej.
2
0
2
3
.0
8
.06
2
.
[9]
T.
A.
Tuib
,
B.
H.
Sao
u
d
i,
Y.
M.
Hu
ss
ein
,
T.
H.
Mand
eel,
an
d
F.
T.
Al
-
Dh
ief,
“Co
n
v
o
lu
ti
o
n
al
n
eu
ral
n
etwo
rk
with
b
in
ary
m
o
t
h
flame
o
p
tim
iz
atio
n
for
em
o
tio
n
d
etect
io
n
in
electroen
cep
h
alo
g
ram
,”
IA
ES
I
n
tern
a
tio
n
a
l
Jo
u
rnal
o
f
Artificia
l
Inte
llig
en
ce
,
v
o
l.
1
3
,
n
o
.
1
,
p
p
.
1
1
7
2
–
1
1
7
8
,
Mar
.
20
2
4
,
d
o
i
: 10
.11
5
9
1
/ijai.v1
3
.i1.p
p
1
1
7
2
-
1
1
7
8
.
[10
]
R.
Seth
an
d
A.
Sh
araff,
“S
en
tim
en
t
d
ata
an
a
ly
sis
for
d
etectin
g
so
cial
sen
se
afte
r
COV
ID
-
1
9
u
sin
g
h
y
b
rid
o
p
t
im
izatio
n
m
eth
o
d
,”
S
N Co
mp
u
ter S
cien
ce
,
v
o
l.
4
,
n
o
.
5
,
J
u
l.
2
0
2
3
,
d
o
i: 10
.1
0
0
7
/s429
7
9
-
023
-
0
2
0
1
7
-
3.
[11
]
S.
M
.
Nag
a
rajan
a
n
d
U.
D.
Gan
d
h
i,
“Clas
sify
in
g
strea
m
in
g
o
f
Twit
ter
d
ata
b
ased
o
n
sen
tim
en
t
an
aly
sis
u
sin
g
h
y
b
ridizatio
n
,”
Neu
ra
l Co
mp
u
tin
g
an
d
A
p
p
lica
tio
n
s
,
v
o
l.
3
1
,
n
o
.
5
,
p
p
.
1
4
2
5
–
1
4
3
3
,
May 2
0
1
9
,
d
o
i: 10
.1
0
0
7
/
s0
0
5
2
1
-
018
-
3
4
7
6
-
3.
[12
]
N.
S.
Muru
g
an
an
d
G.
Ush
a
De
v
i,
“Detecting
str
eam
in
g
o
f
T
witt
er
sp
am
u
sin
g
h
y
b
rid
m
eth
o
d
,”
Wir
eless
Perso
n
a
l
Co
mmu
n
ica
tio
n
s
,
v
o
l.
1
0
3
,
n
o
.
2
,
p
p
.
1
3
5
3
–
1
3
7
4
,
No
v
.
2
0
1
8
,
d
o
i: 1
0
.10
0
7
/s1
1
2
7
7
-
018
-
5
5
1
3
-
z.
[13
]
Y.
Yu
e
,
Y
.
Pen
g
,
an
d
D.
W
an
g
,
“
D
eep
learnin
g
sh
o
rt
tex
t
sen
tim
en
t
an
aly
sis
b
ased
o
n
im
p
rov
ed
p
article
sw
arm
o
p
tim
izatio
n
,”
Electro
n
ics
,
v
o
l.
1
2
,
n
o
.
1
9
,
p
p
.
1
–
2
3
,
Oct
.
2
0
2
3
,
d
o
i: 10
.33
9
0
/electron
ics
1
2
1
9
4
1
1
9
.
[14
]
B.
B.
Al
-
o
n
azi
et
a
l.
,
“Qu
an
tu
m
p
article
swa
rm
o
p
t
im
izatio
n
with
d
eep
learnin
g
-
b
ased
Arabic
tweets
se
n
tim
en
t
an
aly
sis
,”
Co
mp
u
ters
,
Ma
teria
ls {
\
and
}
Co
n
tin
u
a
,
v
o
l.
7
5
,
n
o
.
2
,
p
p
.
2
5
7
5
–
2
5
9
1
,
Jan
.
2
0
2
3
,
d
o
i: 10
.3
2
6
0
4
/cm
c.20
2
3
.03
3
5
3
1
.
[15
]
P
.
M
a
n
j
u
l
a
,
M
.
M
a
r
a
g
a
t
h
a
r
a
j
a
n
,
P
.
R
a
j
p
u
t
,
S
.
P
.
R
.
K
a
m
a
l
a
,
N
.
K
o
p
p
e
r
u
n
d
e
v
i
,
a
n
d
S
.
N
.
S
a
n
g
e
e
t
h
a
a
,
“
S
e
n
t
i
m
e
n
t
a
n
a
l
y
s
i
s
f
o
r
o
n
l
i
n
e
p
r
o
d
u
c
t
r
e
v
i
e
w
s
a
n
d
r
e
c
o
m
m
e
n
d
a
t
i
o
n
u
s
i
n
g
d
e
e
p
l
e
a
r
n
i
n
g
b
a
s
e
d
o
p
t
i
m
i
z
a
t
i
o
n
a
l
g
o
r
i
t
h
m
,
”
I
n
t
e
r
n
a
t
i
o
n
a
l
J
o
u
r
n
a
l
o
n
R
e
c
e
n
t
a
n
d
I
n
n
o
v
a
t
i
o
n
T
r
e
n
d
s
i
n
C
o
m
p
u
t
i
n
g
a
n
d
C
o
m
m
u
n
i
c
a
t
i
o
n
,
v
o
l
.
1
1
,
n
o
.
9
,
p
p
.
3
6
2
9
–
3
6
4
0
,
N
o
v
.
2
0
2
3
,
d
o
i
:
1
0
.
1
7
7
6
2
/
i
j
r
i
t
c
c
.
v
1
1
i
9
.
9
5
8
5
.
[16
]
J.
Ven
k
atar
am
an
an
d
L.
Moh
an
d
o
s
s,
“FBO
‐RNN:
fu
zzy
b
u
tterf
ly
o
p
ti
m
izatio
n
‐bas
ed
R
NN‐LS
T
M
for
ex
tracting
sen
tim
en
ts
fr
o
m
Twitte
r
e
m
o
ji
d
atab
ase,”
Co
n
cu
rr
en
cy
a
n
d
C
o
mp
u
ta
tio
n
:
Pra
ctice
a
n
d
Experi
en
ce
,
v
o
l.
3
5
,
n
o
.
1
2
,
May
2
0
2
3
,
d
o
i: 10
.1002
/cp
e.7
6
8
3
.
[17
]
D
.
Elang
o
v
an
an
d
V.
Su
b
ed
h
a,
“Ad
ap
tiv
e
p
article
g
re
y
wo
lf
o
p
tim
i
zer
with
d
eep
lea
rnin
g
-
b
ased
sen
tim
en
t
an
aly
sis
o
n
o
n
lin
e
p
rod
u
ct
reviews,”
Eng
in
eerin
g
,
Tec
h
n
o
lo
g
y
{
\
and
}
A
p
p
lied
S
cien
ce
Re
sea
rch
,
v
o
l.
1
3
,
n
o
.
3
,
p
p
.
1
0
9
8
9
–
1
0
9
9
3
,
Ju
n
.
2
0
2
3
,
d
o
i: 10
.4808
4
/e
tas
r.
5
7
8
7
.
[18
]
A.
Alzaqeb
ah
,
B.
Sm
ad
i,
an
d
B.
H.
Ham
m
o
,
“Ar
ab
ic
sen
tim
en
t
an
aly
sis
b
ased
o
n
salp
sw
arm
alg
o
rithm
wit
h
S
-
sh
ap
ed
trans
fe
r
fun
ctio
n
s,”
in
2
0
2
0
1
1
th
Inter
n
a
tio
n
a
l
Co
n
feren
ce
o
n
Info
rma
tio
n
a
n
d
Commun
ica
tio
n
S
ystems
(I
C
ICS
)
,
Ap
r.
2
0
2
0
,
p
p
.
1
7
9
–
1
8
4
,
d
o
i: 10
.1109
/ICICS4
9
4
6
9
.20
2
0
.23
9
5
0
7
.
[19
]
A.
M
.
Mash
r
aq
i
an
d
H.
T.
Hal
awani
,
“Dra
g
o
n
fly
o
p
tim
izatio
n
with
d
ee
p
learnin
g
en
ab
led
sen
tim
en
t
an
aly
sis
for
Arabic
t
weet
s,”
Co
mp
u
ter S
ystems
Scien
ce an
d
E
n
g
in
eerin
g
,
v
o
l.
4
6
,
n
o
.
2
,
p
p
.
2
5
5
5
–
2
5
7
0
,
Jan
.
2
0
2
3
,
d
o
i: 10
.
3
2
6
0
4
/css
e.20
2
3
.0
3
1
2
4
6
.
[20
]
L.
R
.
K
ros
u
ri
an
d
R.
S
.
Aravap
alli,
“Featu
re
lev
el
f
in
e
g
rained
sen
tim
en
t
an
aly
sis
u
sin
g
b
o
o
sted
lo
n
g
sh
o
rt
-
term
m
e
m
o
ry
with
im
p
rov
ised
local s
earc
h
whale
op
tim
i
zatio
n
,”
PeerJ
Co
mp
u
ter S
cien
ce
,
v
o
l.
9
,
p
p
.
1
–
2
4
,
Ap
r.
20
2
3
,
d
o
i: 10
.7
7
1
7
/p
eerj
-
cs.1
3
3
6
.
[21
]
Bay
d
o
g
an
an
d
B.
Alatas,
“Sen
ti
m
en
t
an
aly
sis
in
so
cial
n
etwo
rks
u
sin
g
so
cial
sp
id
er
o
p
tim
iz
atio
n
alg
o
rithm,
”
Teh
n
icki
Vjesn
ik
-
Tech
n
ica
l Gaz
ette
,
vo
l.
2
8
,
n
o
.
6
,
p
p
.
1
9
4
3
–
1
9
5
1
,
Dec.
2
0
2
1
,
d
o
i: 10
.1
7
5
5
9
/TV
-
2
0
2
0
0
6
1
4
1
7
2
4
4
5
.
[22
]
C
.
B
a
y
d
o
g
a
n
a
n
d
B
.
A
l
a
t
a
s
,
“
M
e
t
a
h
e
u
r
i
s
t
i
c
a
n
t
l
i
o
n
a
n
d
m
o
t
h
f
l
a
m
e
o
p
t
i
m
i
z
a
t
i
o
n
-
b
a
s
e
d
n
o
v
e
l
a
p
p
r
o
a
c
h
f
o
r
a
u
t
o
m
a
t
i
c
d
e
t
e
c
t
i
o
n
o
f
h
a
t
e
s
p
e
e
c
h
i
n
o
n
l
i
n
e
s
o
c
i
a
l
n
e
t
w
o
r
k
s
,
”
I
E
E
E
A
c
c
e
s
s
,
v
o
l
.
9
,
p
p
.
1
1
0
0
4
7
–
1
1
0
0
6
2
,
J
a
n
.
2
0
2
1
,
d
o
i
:
1
0
.
1
1
0
9
/
A
C
C
E
S
S
.
2
0
2
1
.
3
1
0
2
2
7
7.
[23
]
U.
D
.
Gan
d
h
i,
P.
Mala
rvizh
i
Ku
m
ar,
G.
Ch
an
d
ra
B
ab
u
,
an
d
G.
Kart
h
ick
,
“Sen
tim
en
t
an
aly
sis
o
n
Twitt
er
d
ata
b
y
u
sin
g
co
n
v
o
lu
ti
o
n
al
n
eu
ral
n
etwo
rk
(C
NN
)
an
d
lo
n
g
sh
o
rt
t
erm
m
e
m
o
ry
(
LS
TM
)
,”
Wir
eless
P
ers
o
n
a
l
Co
mmu
n
i
ca
tio
n
s
,
p
p
.
1
–
1
0
,
May 2
0
2
1
,
d
o
i: 10
.
1
0
0
7
/s
1
1
2
7
7
-
0
2
1
-
0
8
5
8
0
-
3.
[24
]
Y.
Yu
,
X.
Si,
C.
Hu
,
an
d
J
.
Zhan
g
,
“A
r
ev
iew
o
f
rec
u
rr
en
t
n
eu
ral
n
etw
o
rks
:
LST
M
cells
an
d
n
etwo
rk
arch
itectu
res,”
Neu
ra
l
Co
mp
u
ta
tio
n
,
v
o
l.
3
1
,
n
o
.
7
,
p
p
.
1
2
3
5
–
1
2
7
0
,
Ju
l.
2
0
1
9
,
d
o
i: 10
.11
6
2
/n
eco_
a
_
0
1
1
9
9
.
[25
]
M.
Sch
u
ster
an
d
K.
K.
Paliwal
,
“B
id
irection
al
recurr
en
t
n
eu
ral
n
etwo
rks
,”
IE
E
E
Tra
n
sa
ctio
n
s
o
n
S
i
g
n
a
l
Pro
cess
in
g
,
v
o
l.
4
5
,
n
o
.
1
1
,
p
p
.
2
6
7
3
–
2
6
8
1
,
Jan
.
1
9
9
7
,
d
o
i
: 10
.11
0
9
/
7
8
.650093
.
[26
]
B.
M.
K
ay
h
an
an
d
G.
Yild
iz,
“Rein
f
o
rcem
en
t
lea
rnin
g
ap
p
licatio
n
s
to
m
achi
n
e
sch
ed
u
lin
g
p
rob
lem
s:
a
co
m
p
re
h
en
siv
e
literature
review,”
Jo
u
rn
a
l o
f I
n
tellig
en
t Ma
n
u
fa
ctu
rin
g
,
v
o
l.
3
4
,
n
o
.
3
,
p
p
.
9
0
5
–
9
2
9
,
Mar
.
20
2
3
,
d
o
i: 10.
1
0
0
7
/s108
4
5
-
021
-
0
1
8
4
7
-
3.
[27
]
W
.
Lon
g
,
J.
Jiao
,
X.
Liang
,
S.
Cai,
a
n
d
M.
Xu
,
“A
ran
d
o
m
o
p
p
o
sitio
n
-
b
a
sed
learnin
g
g
rey
wo
lf
o
p
tim
i
zer,”
I
EE
E
Acce
ss
,
v
o
l.
7
,
p
p
.
1
1
3
8
1
0
–
1
1
3
8
2
5
,
Jan
.
2
0
1
9
,
d
o
i: 1
0
.11
0
9
/ACC
ESS.
2
0
1
9
.2934
9
9
4
.
BIOGR
AP
HI
ES OF
A
UTH
ORS
Nite
sh
Su
re
ja
is
cur
r
ent
ly
e
mpl
oyed
wi
th
KP
GU
Univer
s
it
y,
Vadod
ara,
an
d
Gujar
at,
Indi
a.
He
is
working
a
s
the
direct
or
a
t
the
Krishna
Sch
ool
of
Emerging
Te
chno
logy
and
Appli
ed
Re
sea
rch
,
one
of
the
consti
tu
ent
insti
tutes
of
KP
GU
.
His
rese
ar
ch
intere
st
s
inc
lud
e
m
ac
h
in
e
l
e
arn
ing,
data
scie
n
ce,
art
if
icial
intelligen
ce
,
and
al
gor
it
hms
inspire
d
by
nat
ure
.
H
e
h
as
b
ee
n
a
teac
h
er
fo
r
more
tha
n
23
y
ea
rs.
Publicati
on
s
with
SC
OP
US
and
Googl
e
inde
xes
have
p
ubli
shed
his
stu
dy
find
ings.
H
e
has
me
n
tore
d
t
wo
Ph.D.
ca
nd
i
dat
es
and
is
cur
re
nt
ly
m
ent
or
ing
fiv
e
Ph.D.
c
a
ndida
t
es.
He
c
an
be contacted at
nmsurej
a@gm
ail.
co
m
.
Evaluation Warning : The document was created with Spire.PDF for Python.