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.
11
,
No.
1
,
Febr
uar
y
2021
, pp.
745
~
752
IS
S
N: 20
88
-
8708
,
DOI: 10
.11
591/
ijece
.
v11
i
1
.
pp
745
-
752
745
Journ
al h
om
e
page
:
http:
//
ij
ece.i
aesc
or
e.c
om
A power
ful comp
ar
is
on of
deep le
arnin
g frame
wor
ks for
Arabi
c
sentimen
t a
n
alys
i
s
Youssr
a
Z
ah
idi
1
, Yacine El
Youn
ou
ssi
2
, Y
as
sine
Al
-
Amr
an
i
3
1,2
Inform
at
ion
S
y
stem
and
Softw
are
Engi
n
ee
ring
La
bora
tor
y
,
Abdelmale
k
Essaa
d
i U
nive
rsit
y
,
Morocc
o
3
Te
chno
logi
es
d
e
l
’Inform
a
ti
on
e
t
Modél
isat
ion
d
es
S
y
st
èmes
,
Ab
del
m
al
ek
Essaa
d
i
Univer
si
t
y
,
Morocc
o
Art
ic
le
In
f
o
ABSTR
A
CT
Art
ic
le
history:
Re
cei
ved
Oct
9
, 2
01
9
Re
vised
A
ug 11
,
2020
Accepte
d
Aug
23, 202
0
Dee
p
le
arn
ing
(
DL)
is
a
m
ac
hine
le
arn
ing
(ML)
subdom
ai
n
tha
t
invo
lv
es
al
gorit
hm
s
t
aken
from
the
br
ain
func
t
ion
nam
ed
a
r
ti
fi
cial
n
eu
ral
n
et
works
(AN
Ns
)
.
Rec
ent
l
y
,
DL
appr
oa
che
s
have
g
ai
n
ed
m
aj
or
acco
m
pli
shm
ent
s
ac
ross
var
ious
Arabi
c
nat
ur
al
l
angua
ge
proc
essing
(AN
LP)
ta
s
ks
,
espe
cia
l
l
y
in
the
dom
ai
n
of
Arabi
c
senti
m
ent
an
aly
sis
(AS
A)
.
For
working
on
Arabi
c
SA
,
rese
arc
her
s
ca
n
use
var
ious
DL
li
bra
rie
s
in
the
ir
project
s,
but
without
justi
f
y
ing
the
ir
choi
c
e
or
they
choose
a
group
of
li
br
ari
es
r
ely
i
ng
on
th
ei
r
par
ticula
r
progra
m
m
ing
la
nguage
familiarity
.
W
e
are
basing
i
n
th
is
work
on
Java
and
P
y
tho
n
progra
m
m
ing
la
nguag
es
bec
au
se
they
hav
e
a
l
arg
e
set
of
dee
p
l
ea
rn
ing
l
i
bra
rie
s
th
at
are
ver
y
useful
in
t
he
AS
A
dom
ai
n.
Thi
s
p
ape
r
foc
uses
on
a
compara
ti
ve
an
alys
is
of
diffe
ren
t
val
uab
le
P
y
thon
and
Java
l
ibra
r
ie
s
to
con
cl
ude
th
e
m
ost
rel
ev
ant
and
ro
bust
DL
li
bra
ri
e
s
for
ASA.
Throw
thi
s
compara
t
ive
anal
y
s
i
s,
and
w
e
find
tha
t:
T
ensorF
low,
Theano
,
and
K
er
as
P
y
th
on
f
rameworks
are
ve
r
y
popu
la
r
and
v
er
y
u
sed
in
thi
s
rese
arc
h
dom
ai
n
.
Ke
yw
or
d
s
:
AN
L
P
ASA
DL
Java lib
rar
ie
s
Pyt
hon
li
brarie
s
This
is an
open
acc
ess arti
cl
e
un
der
the
CC
B
Y
-
SA
l
ic
ense
.
Corres
pond
in
g
Aut
h
or
:
Youssra Za
hid
i
,
Inform
at
ion
System
an
d So
ftwar
e
Enginee
ri
ng
La
borato
ry,
Abdelm
al
ek
Essaadi
Un
i
ver
sit
y
,
Tet
uan, Mo
r
oc
co
.
Em
a
il
:
yous
sra
1994zahi
di@
gm
ai
l.co
m
1.
INTROD
U
CTION
Ar
a
bic
sentim
ent
analy
sis
or
o
pi
nion
m
ini
ng
ai
m
s
to
determ
in
e
the
senti
m
ent
po
la
rity
(positi
vity
,
neg
at
ivit
y
,
or
neu
t
rali
ty
)
of
a
wr
it
er.
A
la
rge
var
ie
ty
of
opinions
are
bor
ne
in
po
sts
o
n
di
ff
ere
nt
so
ci
al
m
edia
platfo
rm
s
l
ike
Twitt
er,
You
Tu
be,
I
ns
ta
gra
m
,
Faceboo
k
.
This
fiel
d
of
researc
h
ha
s
recently
at
t
racted
increasin
g
at
te
ntion
[1
,
2]
,
es
pecial
ly
in
Engli
sh
.
Alth
ough
the
Ar
a
bic
la
ngua
ge
is
deem
ed
as
the
m
os
t
us
e
fu
l
la
nguag
e
on so
ci
al
m
edia p
la
tfo
rm
s,
on
ly
s
om
e
wo
rks
ha
ve
re
li
ed
on
ASA
so
fa
r.
Ther
e
is
a
set
of
m
achine
le
a
rn
i
ng
m
od
el
s
powe
rin
g
nat
ur
al
l
angua
g
e
proc
essing
(
NL
P
)
a
pp
li
cat
io
ns.
Re
centl
y,
DL
appr
oac
hes
ha
ve
gai
ned
high
pe
rfor
m
ance
acro
ss
va
ri
ous
NLP
ta
sks
[
3]
.
S
pecifica
ll
y,
it
has
held
go
od
res
ults
in
the
sen
tim
ent
analy
si
s
dom
ai
n
[4
-
7]
,
and
it
is
th
e
sta
te
-
of
-
the
-
art
m
od
el
in
di
ff
ere
nt
la
ngu
age
s
[8
-
10]
w
hile
the sta
te
-
of
-
the
-
art ac
cur
acy
for
AS
A
sti
ll
require
s
am
e
li
or
at
ion
s
.
DL
is
a
p
a
rt
of
the
M
L
fiel
d
con
ce
r
ned
wit
h
a
c
ollec
ti
on
of
al
gorithm
s
ba
se
d
on
the
brai
n
functi
on
na
m
ed
ar
ti
fici
al
n
eu
ral
n
et
work
s
(
ANNs)
.
It
is
a
ML
m
e
t
hod
th
at
te
ach
es
com
pu
te
rs
t
o
do
t
hings
th
at
are
natu
ral
to
hum
ans
by
cre
at
in
g
arc
hitec
ture
m
ade
up
of
an
input
an
d
out
pu
t
la
ye
r
with
var
i
ou
s
hidde
n
la
ye
rs
(en
c
oders
)
between
them
.
N
um
ero
us
te
c
hniqu
es
a
re
a
pp
l
i
ed
wit
h
D
L
li
ke
recurre
nt
neural
netw
orks
(
RNN
)
,
deep ne
ural
n
et
works
(
D
NN),
conv
olu
ti
onal
neural
netw
ork
s
(CN
N
),
l
ong short
-
te
rm
m
em
or
y
(LSTM),
et
c.
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol. 11,
No.
1,
Febr
uar
y
2021 :
7
45
-
752
746
As
par
t
of
our
A
rab
ic
se
ntim
ent
analy
sis
researc
h
pro
je
ct
,
we
at
te
m
pt
to
perform
a
n
in
-
de
pt
h
com
par
at
ive
ev
al
uatio
n
t
o
ac
hieve
a
s
umm
a
ri
z
at
ion
o
f
the
m
os
t
valuab
le
pro
gr
am
m
ing
l
angua
ges,
w
hi
ch
a
re
ab
un
dan
t
i
n
te
r
m
s o
f
ASA
li
br
aries.
We
try
t
o
c
om
par
e thes
e
too
ls
to s
peci
fy t
he
m
os
t use
fu
l
ones
.
Re
centl
y,
the
NLP
c
omm
un
i
ty
has
at
te
nd
ed
num
ero
us
pe
netrati
ons
du
e
to
the
app
li
c
at
ion
of
DL
.
This
la
te
r
has
offer
e
d
sal
ie
nt
a
m
el
iorati
on
s
in
the
dom
ai
n
of
se
ntim
ent
analy
sis
in
En
glish.
H
ow
e
ve
r,
le
ss
researc
h
ha
s
be
en
do
ne
on
e
m
plo
yi
ng
DL
in
ASA
.
Du
e
to
it
s
com
plication
,
m
or
phol
og
ic
al
,
a
nd
sy
ntact
ic
abun
dan
ce
,
Arabic
ha
s
deem
ed
the
m
os
t
di
ff
ic
ult
la
ngua
ge
,
a
nd
it
has
a
lim
i
te
d
nu
m
ber
of
DL
li
br
a
ries
com
par
ed
t
o other fam
ou
s la
ngua
ges
li
ke
E
ngli
sh
.
Choosin
g
the
m
os
t
us
efu
l
lib
ra
ries
is
v
ery
com
plica
te
d
and
nee
ds
an
in
-
dep
t
h
analy
sis.
For
this
reason,
we
rel
y
on
m
any
co
m
par
at
ive
le
vels.
F
or
t
he
pur
po
s
e
of
t
his
wo
rk
,
we
com
par
ed
the
m
os
t
powe
rful
DL
li
braries
f
or
AN
L
P
in
Pyt
hon
a
nd
Ja
va
t
o
sel
ect
the
m
os
t
co
nv
e
nient
gro
up
of
li
bra
ries
that
a
ddresse
s
ou
r
pur
po
se
s
in
the
AS
A d
om
ai
n
.
The
rest
of
this
arti
cl
e
is
div
id
ed
as
fo
ll
ows: Sect
ion
2
sh
ows
the
crit
ic
al
Ja
va
an
d
Pyt
hon
li
br
aries
in
the
A
ra
bic
la
ngua
ge
f
or
DL
.
Sect
ion
3
la
ys
em
ph
asi
s
on
a
detai
le
d
c
omparati
ve
stu
dy
,
an
d
it
al
so
pr
ov
i
des
a
con
cl
us
io
n
a
bout
the
m
os
t
us
ef
ul
dee
p
le
arn
i
ng
li
brarie
s
in
the
fiel
d
of
ASA.
T
he
r
esults
are
de
ba
te
d
in
detai
l i
n
Sect
io
n
4
, a
nd the
wo
rk
is
fi
nish
e
d
with
the
fi
nal
idea
s
i
n
Sect
io
n
5
.
2.
A
R
ABI
C
SE
N
TIMENT
A
N
ALYSIS
LIB
RARIES
FO
R
D
L
This
s
ect
io
n
presents
fo
ur
open
-
s
ource
li
braries
for
Dee
p
Lea
rn
i
ng,
na
m
el
y
Thean
o,
Tens
orFlo
w,
Ker
as
,
a
nd D
ee
pl
ear
ning4j.
2.
1.
Ar
ab
ic
s
u
ppo
r
tin
g
P
yth
on
li
br
aries
f
or
D
L
Nowa
days,
DL
is
the
hott
est
t
rend
i
n
ML
an
d
AI.
We
sel
ect
ed
s
om
e
of
t
he
best
Pyt
hon
li
br
aries
f
or
the Ara
bic
la
ngua
ge, w
hich
is
deem
ed
the
m
os
t
po
we
rful
DL
li
braries i
n AS
A.
Tens
orFlo
w:
it
is
an
op
en
-
s
ource
s
of
t
war
e
li
br
ary
f
or
dataflo
w
an
d
dif
fe
ren
ti
able
pro
gr
a
m
m
ing
on
a
range
of
ta
s
ks
.
It
is
a
sym
bo
li
c
m
at
h
tool
and
is
al
s
o
a
pp
li
ed
f
or
ML
app
li
cat
io
ns
l
ike
ne
u
ral
net
works.
It co
m
es w
it
h
r
obus
t s
upport
f
or
ML
and
DL
,
and the
f
le
xi
bl
e num
erical
co
m
pu
ta
ti
on
core
is em
plo
ye
d
acro
ss
var
i
ou
s
oth
e
r
s
ci
entifi
c
fiel
ds
.
Tens
orFlo
w
li
br
a
ry
can
run
on
m
ulti
ple
G
PU
s
a
nd
C
PU
s
(w
it
h
opti
on
al
SY
CL
and
C
U
DA
e
xt
ensio
ns
f
or
ge
ne
ral
-
ai
m
com
pu
ti
ng
on
gra
phic
s
pro
cessi
ng
unit
s).
It
is
ob
ta
ina
ble
on
64
-
bit
W
i
ndows
,
Li
nux,
a
nd
m
ob
il
e
com
pu
ti
ng
pl
at
fo
rm
s,
co
nta
ining
iO
S
a
nd
an
droid
.
For
m
aking
it
sim
ple
f
or
us
ers
t
o
unde
rs
ta
nd
,
debu
g,
a
nd
op
ti
m
iz
e
T
ens
or
Fl
ow
pro
gr
am
s,
there
is
an
excel
le
nt
gro
up
of
visu
al
i
zat
ion
too
ls
nam
ed
T
ens
or
B
oard
[11]
. F
i
gure
1 pre
sents the
Te
nsorF
l
ow p
yt
hon l
ibrar
y a
rc
hitec
tur
e.
Thea
no
:
[
12
,
13]
,
is
a
cro
s
s
-
platf
or
m
op
e
n
-
s
ource
to
ol
that
per
m
it
s
t
he
resea
rch
e
r
to
evaluate
m
at
he
m
at
ic
a
l
expressi
on
s
,
in
cl
ud
in
g
m
ulti
-
dim
ension
al
ar
rays
w
or
t
hily
.
It
is
a
m
at
he
m
at
ic
al
li
br
ary
,
bu
t
it
was
i
niti
al
ly
c
reated
t
o
facil
it
at
e
researc
h
in
the
D
L
fiel
d
.
Ba
se
d
on
t
he
a
dvanta
ges
of
T
hea
no,
di
ver
s
pack
a
ges
have
been
devel
op
ed,
s
uch
as
K
eras,
Pyl
ear
n2,
Bl
ock
s
,
an
d
Lasag
ne
[
14
]
.
Thea
no
has
m
any
featur
e
s
li
ke
it
pro
vid
es
m
os
t
of
Nu
m
Py’s
f
un
ct
io
nalit
y,
but
ad
ds
aut
oma
ti
c
sy
m
bo
li
c
diff
e
re
ntiat
ion
,
offe
rs
trans
par
e
nt
em
plo
ym
ent
of
a
GPU
al
so
it
do
e
s
the
der
i
va
ti
ves
f
or
func
ti
on
s
with
on
e
or
num
ero
us
inputs
.
Figure
2
pr
e
se
nts
the
Th
ea
no
a
rch
it
ect
ure.
Figure
1.
Ten
s
orflo
w
python
li
br
ary a
rch
it
ec
ture
Figure
2
.
Thea
no Pyt
hon l
ibr
ary
arc
hitec
tur
e
Ker
as
:
It
is
a
t
op
-
le
vel
ne
ur
al
networ
ks
API,
able
to
work
on
to
p
of
Thea
no
,
Te
nsor
Flo
w
,
an
d
ot
he
r
too
ls
.
It
was
c
reated
to
pe
rm
it
rap
id
e
xp
e
ri
m
entat
ion
with
dee
p
ne
ural
ne
tworks
.
Ke
ras
runs
seam
le
ssly
on
CPU
or
G
PU.
It
pro
vid
es
si
m
ple
and
rap
i
d
prot
otyp
ing
a
nd
s
us
ta
in
s
both
rec
urre
nt
netw
ork
s
an
d
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
om
p
En
g
IS
S
N: 20
88
-
8708
A powerf
ul c
omp
ar
is
on o
f
de
ep
le
arni
ng fra
meworks
for
a
r
ab
ic
sentime
nt
analysis
(
Y
ou
s
sra
Zahidi
)
747
conv
olu
ti
onal
netw
orks,
as
well
as
conjuncti
on
s
of
t
he
tw
o.
Fig
ure
3
sho
ws
the
Ker
as
a
rch
it
ec
ture
.
Dep
e
ndin
g
on
the
li
te
ratur
e,
we
f
ound
that
these
li
br
a
ries
are
ve
ry
us
e
d
i
n
the
ASA
fiel
d,
a
nd
m
any
auth
ors
reco
m
m
end
th
e
m
in
their
w
orks
.
Ta
ble
1
hi
gh
li
ghts
a
c
om
par
is
on
of
m
any
valua
ble
dee
p
le
ar
ning
li
braries
in
Ar
a
bic
se
ntim
e
nt an
al
ysi
s.
Figure
3
.
Ke
ra
s
Pyt
hon
li
br
a
r
y archit
ect
ure
Table
1.
C
om
par
iso
n of t
he
m
os
t
us
ed
DL
li
braries i
n AS
A
Librar
y
Creato
r
Has Pre
-
trained
Mod
els
Platf
o
r
m
Interf
ace
So
f
tware
Licence
W
o
rks
in AS
A
Tens
o
rf
lo
w
Go
o
g
le Brain tea
m
Yes
Linu
x
,
m
acO
S,
W
in
d
o
ws,
An
d
roid
Py
th
o
n
(
Ker
as),
C/C
++,
Jav
a,
Go
,
R
Ap
ache 2
.0
[
1
5
]
[16
]
[17
]
[18
]
[
1
9
]
[20
]
[21
]
[22
]
Thean
o
Un
iv
ersité de
Mon
tréal
Thro
u
g
h
Lasagn
e's
m
o
d
el
zo
o
Cro
ss
-
p
latf
o
r
m
Py
th
o
n
(
Ker
as)
BSD licen
se
[
2
3
]
[24
]
[25
]
[26
]
[
2
7
]
[28
]
[29
]
[30
]
Kera
s
Franço
is Ch
o
llet
Yes
Linu
x
,
m
acO
S,
W
in
d
o
ws
Py
th
o
n
,
R
MI
T
licens
e
[
2
0
]
[25
]
[26
]
[29
]
[
3
1
]
[28
]
[32
]
[33
]
2.2.
Ar
ab
ic
su
ppo
r
tin
g
Java
t
oo
lki
t
s
f
or D
L
We
ch
os
e
the
Deep
l
ea
rn
i
ng4j
li
br
a
ry
because
it
is
con
side
red
as
the
m
os
t
us
efu
l
Ja
va
de
ep
le
arn
i
ng
li
br
ary
in
A
ra
bi
c
s
entim
ent
a
naly
sis
:
Deep
l
ea
rn
i
ng4
j
:
It
is
release
d
under
a
pac
he
li
cense
2.0,
cr
eat
ed
m
a
inly
by
a
ML
set
head
qua
rtere
d
in
T
ok
y
o
a
nd
S
an
F
ranci
sco
a
nd led
by
Ad
a
m
G
ibso
n.
It is
a free a
nd cr
oss
-
platf
or
m
too
l
,
an
d
i
t
was des
ign
e
d
to
integ
ra
te
with
S
park
,
Ha
doop,
an
d
oth
e
r
Java
-
bas
ed
di
stribu
te
d
s
of
t
w
are.
It
is
de
sig
ned
f
or
Java
vi
rtual
Ma
chine
a
nd
Java
as
well
a
s
com
pu
te
r
fr
a
m
ewo
r
ks
t
hat
broa
dly
suppo
rt
dee
p
le
ar
ni
ng
al
go
rithm
s.
It
has
pre
-
trai
ne
d
m
od
el
s
an
d
su
sta
i
n
s
CU
DA,
but
it
can
be
m
or
e
qu
ic
ke
ne
d
w
it
h
cuDNN.
T
his
li
br
ary
al
so
of
fe
rs
GPU
s
upport
f
or
the
distri
bu
t
ed
fr
am
ewo
r
k,
an
d
we
c
an
se
le
ct
native
CP
Us
or
GP
U
s
for
our
bac
kend
li
near
al
gebra
proces
ses. F
i
gure
4
s
hows
t
he Dee
pl
earn
in
g4j
arc
hi
te
ct
ur
e
.
Figure
4
.
Dee
pl
earn
in
g4J Ja
va
l
ibrar
y a
rc
hitec
ture
3.
COMP
ARAT
IVE EV
AL
U
ATIO
N
O
F
D
L T
OOLS
In
t
his
pa
rt,
we
will
pr
ese
nt
our
in
-
de
pt
h
c
om
par
at
ive
stud
y
on
va
r
i
ous
le
vels:
fi
rst,
we
s
ho
w
the
m
any
essenti
al
sta
nd
ar
ds
and
the
m
os
t
c
riti
cal
app
li
cat
i
on
area
s.
The
n,
we
rely
on
th
e
GitHub
resu
l
ts
and
the
hist
or
y
of
G
oogle
T
rends
t
o
c
on
cl
ude
the
m
os
t
powe
rful
DL
l
ibrar
ie
s
that
s
at
isfy
our
s
pe
ci
fic
requirem
ents
ve
ry w
el
l.
3.1.
C
ompari
so
n
of the
m
ost
f
am
ou
s
o
pen
-
s
ourc
e
DL
li
b
raries
The
c
om
par
ison
of
these
D
L
to
ols
ca
n
r
el
y
on
di
ff
e
re
nt
crit
eria.
Ta
ble
2
s
ho
ws
a
var
ie
ty
of
par
am
et
ers
ad
opte
d
i
n
this
com
par
at
ive study
:
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol. 11,
No.
1,
Febr
uar
y
2021 :
7
45
-
752
748
Table
2
.
E
val
ua
ti
on
of d
if
fere
nt d
ee
p
le
a
rn
i
ng li
braries
Kera
s
Thean
o
Tens
o
rFlow
Deepl
earnin
g
4
j
W
ritt
en
in
Py
th
o
n
Py
th
o
n
C++, P
y
th
o
n
C++, Jav
a
RB
M/DBNs
Yes
Yes
Yes
Yes
Parallel
Execu
tio
n
Yes (on
ly
po
ss
ib
le with th
e
Tens
o
rFlow b
acke
n
d
)
Yes
Yes
Yes
Do
cu
m
en
tatio
n
The
Ker
as p
roject
h
as rich
d
o
cu
m
en
t
atio
n
,
an
d
a
g
rou
p
o
f
exa
m
p
les ai
m
in
g
at
a
large
n
u
m
b
er
o
f
d
i
f
f
icu
lties
is p
rov
id
ed
[
3
4
]
.
Do
cu
m
en
tatio
n
f
o
r
Thean
o
an
d
Tens
o
rFlow
is
v
er
y
ab
u
n
d
an
t
; it
o
f
f
e
rs
a
large
a
war
en
ess
o
f
the
Tens
o
rf
lo
w
an
d
Thean
o
bas
ics, f
ro
m
ins
tallatio
n
t
o
crea
tin
g
h
u
g
e
n
etwo
rks
on
th
e
d
istrib
u
ted
env
ironm
e
n
t
[
3
5
]
.
The
web
si
te
of
th
e
De
ep
l
earnin
g
4
j
lib
rar
y
of
f
ers
h
elp
f
u
l
ex
a
m
p
les
an
d
d
o
cu
m
en
tatio
n
[
3
5
]
.
Paralleli
z
in
g
tech
n
iq
u
e
su
p
p
o
rt
CUDA Su
p
p
o
rt
Yes
Yes
Yes
Yes
Op
en
MP
Su
p
p
o
rt
On
ly
if
us
in
g
T
h
eano
as
b
acken
d
Yes
No
Yes
Op
en
CL Sup
p
o
rt
O
n
th
e
road
m
ap
f
o
r
th
e
Tens
o
rFlow b
acke
n
d
an
d
u
n
d
er
crea
tio
n
f
o
r
th
e
Thean
o
back
en
d
The b
acken
d
was
b
u
ilt
to
su
p
p
o
rt
Op
en
CL,
b
u
t curre
n
t
su
p
p
o
rt
is d
ef
icien
t
(un
d
er
crea
tio
n
)
On
r
o
ad
m
ap
b
u
t alr
eady
with
SYCL
su
p
p
o
rt
On
r
o
ad
m
ap
3.2.
C
ompari
so
n
by
su
pp
or
ted trea
tmen
t
s
in
DL
and
N
LP
fiel
ds
In
Ta
bles
3
a
nd
4,
we
h
i
gh
li
gh
t
t
he
es
senti
al
s
upporte
d
t
r
eatm
ents
of
DL
to
ols
Te
nsor
Flow,
Ker
as
,
Thea
no
,
a
nd
D
eep
l
ear
ning4j.
We
al
s
o
ai
m
t
o
c
om
par
e
the
m
fo
r
co
nclu
di
ng
the
li
brary
that
s
us
ta
in
s
a
ver
y
high
am
o
un
t
of
cov
e
re
d
ta
sks.
Accor
ding
to
this
ben
c
hma
rk
i
ng
stu
dy,
Pyt
hon
li
br
arie
s
Tenso
rF
l
ow,
Ker
as,
and
T
hea
no
suppo
rt
alm
os
t
t
he
sam
e
ta
sk
s
in
A
NLP
or
de
ep
le
arn
i
ng.
C
om
par
ed
to
the
Deep
l
ear
ni
ng4j
Ja
va
li
br
ary, t
hese t
hr
ee
Pyt
hon l
ib
rar
ie
s s
uppo
rt
a large
num
ber o
f
a
pp
li
cat
ions.
Table
3
.
C
om
par
iso
n of DL
li
br
a
ries in
term
s
of
diff
e
re
nt net
work
t
ypes
Variable
Netwo
rk t
y
p
es
Recu
rr
en
t
Neu
ral
Netwo
rks
(RNN)
Lon
g
-
Sh
o
rt
Ter
m
Me
m
o
ry
Netwo
rks
(L
ST
M)
Recu
rsiv
e
Neu
ral
Netwo
rks
(RNN)
Seq
u
en
ce to
Seq
u
en
ce
(seq
2
seq
)
m
o
d
els
Co
n
v
o
l
u
tio
n
al
n
eu
ral
n
etwo
rks
(CNN)
Bi
-
d
irection
al
LST
Ms
Kera
s
Y
es
Y
es
No
Y
es
Y
es
Y
es
(N
o
Arabic
)
Thean
o
Y
es
No
No
Y
es
(N
o
Arabic
)
Y
es
No
Tens
o
rf
lo
w
Y
es
Y
es
No
Y
es
Y
es
Y
es
(N
o
Arabic
)
Deeplearn
in
g
4
j
Y
es
Y
es
Y
es
Y
es
(N
o
Arabic
)
Y
es
No
Table
4
.
E
val
ua
ti
on
of
DL
to
ols
in
term
s
of
NLP
ta
s
ks
Kera
s
Thean
o
Tens
o
rf
lo
w
Deepl
earnin
g
4
j
Natu
ral
Lang
u
ag
e Pr
o
cess
in
g
Task
s
Arabic Senti
m
en
t
An
aly
sis
Y
es
Y
es
Y
es
Y
es
Machin
e tr
an
slatio
n
Y
es
Y
es
Y
es
No
Co
n
v
ersatio
n
s
/
Qu
estio
n
Ans
wering
(
QA
)
Y
es
Y
es
Y
es
No
Lang
u
ag
e
m
o
d
elin
g
Y
es
No
Y
es
No
Sp
eech recog
n
itio
n
Y
es
Y
es
Y
es
No
Text class
if
icatio
n
Y
es
Y
es
Y
es
Y
es
Text
su
m
m
a
riz
atio
n
Y
es
Y
es
Y
es
No
Text g
en
eration
Y
es
Y
es
Y
es
No
Na
m
ed
Entity
Rec
o
g
n
itio
n
(
NER
)
Y
es
Y
es
Y
es
No
Part
-
Of
-
Sp
ee
ch
ta
g
g
in
g
(
POS
)
Y
es
Y
es
Y
es
Y
es
Dep
en
d
en
cy
parsi
n
g
Y
es
Y
es
No
No
W
o
rd e
m
b
ed
d
in
g
s
Y
es
Y
es
Y
es
Y
es
Se
m
an
tic Ro
le
L
ab
ellin
g
(
SRL)
No
Y
es
No
No
Seq
u
en
ce tagg
in
g
No
No
No
No
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
om
p
En
g
IS
S
N: 20
88
-
8708
A powerf
ul c
omp
ar
is
on o
f
de
ep
le
arni
ng fra
meworks
for
a
r
ab
ic
sentime
nt
analysis
(
Y
ou
s
sra
Zahidi
)
749
3.3.
Co
m
pa
r
ati
ve
eva
lu
ati
on
of
th
e
o
pen
s
oftware
t
ools
f
ocus
ed
on
the
f
orks
,
s
ta
rs
,
c
ommit
s
,
and
c
ontri
butors
r
ecei
ved
b
y the
GitH
ub
c
om
munit
y
The
GitHub
sit
e
n
um
ber
s
a
re
const
antly
var
i
able
.
T
hat
i
s
w
hy
we
will
se
le
ct
the
consulta
ti
on
date
of
th
es
e
pieces
of
inform
at
ion
(
17
/
06
/
20
20
).
T
able
5
s
hows
t
he
GitH
ub
res
ults
.
Re
li
ed
on
the
fo
ll
owin
g
resu
lt
s,
we
deduce
t
hat
Ten
s
orflo
w
is
the m
os
t used,
purs
ued
by
Ke
ras
a
n
d fi
nally
Th
ea
no and
De
epl
ear
ning4j.
Table
5
.
G
it
H
ub
resu
lt
s
Librar
y
Tens
o
rFlow
Kera
s
Thean
o
Deeplearn
in
g
4
j
Lang
u
ag
e
Py
th
o
n
Py
th
o
n
Py
th
o
n
Jav
a
Stars
1
4
6
0
0
0
4
8
7
0
0
9200
1
1
7
0
0
Fo
rks
8
1
7
0
0
1
8
4
0
0
2500
4800
Co
n
tribu
to
rs
2529
818
332
37
Co
m
m
its
8
8
2
7
8
5343
2
8
1
2
0
969
3.4.
C
ompari
so
n
of
l
ibrarie
s
a
cc
ordin
g to Go
ogle
Tren
ds and
GitH
u
b
p
ull
r
eq
uest
h
istor
y
In
Fi
gure
s
5
and
6
,
we
hi
ghli
gh
t
the
res
ults
of
Goo
gle
Trends
a
nd
GitHub
pu
ll
re
quest
history
.
Accor
ding
to
Goo
gle
Tre
nds
histo
ry
res
ults,
we
ca
n
co
ncl
ud
e
t
hat
Ke
ras
and
Tens
orFl
ow
li
braries
a
r
e
ver
y
fam
ous
withi
n
the
use
rs
'
co
m
m
un
it
y.
Fo
c
us
in
g
on
Git
Hub
pu
ll
requ
est
histor
y
res
ults,
Te
ns
orFl
ow
is
consi
der
e
d
t
o b
e the m
os
t com
m
on
ly
u
sed
c
om
par
ed
to t
he
t
hr
ee
o
t
her
li
br
a
ries in
rece
nt years.
Figure
5
.
G
oogl
e Trends
h
ist
ory
Figure
6
.
GitH
ub
p
ull
r
e
quest
h
ist
ory
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol. 11,
No.
1,
Febr
uar
y
2021 :
7
45
-
752
750
4.
RESU
LT
S
AND DI
SCUS
S
ION
In
this
c
om
par
at
ive
eval
uatio
n,
we
s
el
ect
ed
Java
a
nd
Pyt
hon
pr
ogram
m
i
ng
la
ngua
ges
because
they
are v
ery p
opul
ar
a
nd
ha
ve
m
any
use
f
ul
D
L
li
br
a
ries u
sed
in
A
S
A.
I
t
is ver
y
dif
ficult
to
de
du
ce
that o
ne
t
oo
l
is
gr
eat
est
tha
n
oth
e
r
to
ols
,
as
these
to
ols
a
re
ve
ry
val
ua
ble
an
d
hav
e
a
la
rg
e
popul
arit
y
in
these
fiel
ds
.
The follo
wing
par
t
pr
e
sents
our
c
oncl
us
i
ons:
Thea
no
:
is
gr
e
at
for
creati
ng
netw
orks
f
ro
m
a
vaila
ble
com
pone
nts
an
d
re
us
in
g
pre
‐
trai
ne
d
netw
orks
,
bu
t
m
or
e
chall
eng
i
ng
as
far
a
s
the
buil
ding
of
perfect
so
l
ut
ion
s
is
co
ncerned
.
T
he
m
ai
n
disad
van
ta
ge
of
this
li
br
ary
is
fr
e
qu
ently
long
c
ompil
e tim
es w
he
n
c
reati
ng
hu
ge
m
od
el
s.
Tens
orFlo
w:
it
buil
t
to
s
ub
st
it
ute
Thea
no.
The
se
t
wo
tools
are,
i
n
fact,
quit
e
sim
il
ar.
A
ne
ur
al
netw
ork
is
de
cl
ared
as
a
c
om
pu
ta
ti
on
al
gr
a
ph
l
ike
in
Thea
no,
w
hic
h
is
optim
iz
e
d
duri
ng
c
ompil
at
ion.
Tens
orFlo
w,
howe
ver,
has
a
faster
c
ompil
e
tim
e
tha
n
T
heano,
bu
t
it
is
slow
er
than
oth
e
r
li
br
a
ries.
A
sig
nificant
new
feature
is
the
i
m
ple
m
e
ntati
on
of
dat
a
par
al
le
li
s
m
,
wh
ic
h
is
ident
ic
al
to
the
Iterati
ve
Ma
pRed
uce
from
D
eeplea
r
nin
g4
j
.
Ker
as
:
is
easy
to
us
e
an
d
offe
rs
act
ion
a
ble
f
eedb
ac
k
up
on
us
er
er
r
or
,
an
d
it
per
m
it
s
f
ast
prototypi
ng.
Ker
as
li
br
a
ry
sit
s
at
op
T
hea
no
a
nd
Te
nsor
Flow.
Dee
plea
rn
i
ng4j
f
oc
us
e
s
on
Ke
ras
as
it
s
Pyt
ho
n
A
PI
a
nd
i
m
po
rts
m
od
el
s
from
Ker
as
and
t
hroug
h
K
eras
from
Tenso
r
Flo
w
an
d
T
heano
.
Its
pro
gr
am
s
are
generall
y
sm
a
ll
er th
an
t
he
equivale
nt T
heano a
nd Te
nsor
Flo
w progra
m
s.
Deep
le
a
rn
i
ng4j:
pro
vid
es
gre
at
so
luti
on
s
for
beg
in
ne
rs
inte
rested
in
e
xp
l
ori
ng
dee
p
ne
ural
netw
orks
,
appr
opriat
e
for
edu
c
at
ion
al
a
nd
t
rainin
g
obj
ect
ive
s
.
O
n
th
e
oth
e
r
ha
nd,
Tens
orFlo
w
a
nd
Thea
no
tools
are
m
or
e
su
it
able
f
or ex
per
ie
nce
d use
rs who
nee
d
to
h
a
ve
m
uch m
or
e co
ntro
l
over
n
et
work ar
chite
ct
ur
es
.
As
a
su
m
m
ari
zat
ion
of
this
sect
ion
,
each
DL
t
oo
l
is
c
ha
racteri
zed
by
i
ts
be
nef
it
s
in
AN
L
P
ta
s
ks
.
Althou
gh
they
sh
are
al
m
os
t
t
he
sam
e
char
a
ct
erist
ic
s
and
adv
a
ntage
s,
Ten
so
r
flo
w
outpe
r
form
s
oth
er
li
br
aries
accor
ding
t
o
G
oogle
T
rends
a
nd
GitH
ub p
ull
re
qu
est
histo
r
y
.
H
ow
e
ve
r,
when
we
ta
lk
a
bout Arabic Sent
i
m
en
t
analy
sis,
an
d
accor
ding
to
t
he
li
te
ratur
e
a
nd
va
rio
us
great
work
s
in
the
fiel
d
of
ASA,
we
fi
nd
that:
Tens
orFlo
w,
T
heano,
and
Ke
r
as are
ve
ry po
pula
r
a
nd
ver
y
use
d
i
n
this
rese
arch d
om
ai
n.
5.
C
O
NC
L
US
I
O
N
In
this
w
ork,
we
desc
ribe
d
a
var
ie
ty
of
P
yt
hon
an
d
Java
DL
to
ols
that
are
deem
ed
m
os
t
he
lp
fu
l
in
AN
L
P.
Be
si
de
s,
we
at
te
m
pte
d
al
so
to
c
oncl
ud
e
the
m
os
t
powe
rful
a
nd
use
fu
l
DL
li
brari
es
f
or
t
he
ASA
fiel
d.
More
ov
e
r
,
we
hav
e
c
om
pare
d
ea
ch
li
brar
y
us
ing
seve
r
al
aspects.
In
con
cl
us
io
n,
ever
y
D
L
li
brary
is
char
a
ct
e
rized
by
its
a
dv
a
nta
ges
a
nd
be
nef
i
ts
in
ANLP
ta
sk
s.
F
or
this
r
easo
n,
we
ch
ose
this
set
of
l
ibrar
ie
s
because
they
m
ade great res
ults in
t
he
ASA tas
k.
REFERE
NCE
S
[1]
Y.
Al
Am
ran
i,
e
t
al.
,
“
Rec
ov
er
y
of
the
opini
ons
t
hrough
the
spec
i
fic
ities
of
docu
m
ent
s
te
xt
,
”
in
2
019
Inte
rnationa
l
Confe
renc
e
on
Wirel
ess Tec
hno
logi
es,
Embe
dd
e
d
and
Int
el
l
ige
n
t
Syste
ms
,
WITS
2019
,
2019
.
[2]
Y.
Al
Am
ran
i,
M.
Laza
ar
,
and
K.
E.
El
K
adi
rp,
“
Random
fore
st
and
support
v
ector
m
ac
hin
e
base
d
h
y
br
id
appr
o
ach
to
sent
iment anal
y
sis,
” in
Proce
d
ia
Computer
S
cienc
e
,
2018
,
vo
l.
127,
pp
.
511
–
52
0.
[3]
Y.
Za
hid
i,
Y. El Y
ounouss
i,
and
C.
Azroum
ahl
i
,
“
Com
par
at
ive
St
ud
y
of
th
e
Mos
t
Us
efu
l
Arabi
c
-
s
upporti
ng
Natur
al
La
nguag
e
Proce
ss
ing
and
Dee
p
Le
arn
ing
Li
br
ari
es,
”
in
2019
Inte
rnational
Confe
renc
e
on
Optimizat
ion
and
Appl
ic
a
ti
ons,
IC
OA
2019
,
2019.
[4]
S.
Aich,
et
al.
,
“
Convolut
iona
l
neur
al
net
work
-
base
d
m
odel
for
web
-
base
d
te
xt
cl
assifi
catio
n,
”
Inte
rnat
iona
l
Journal
of
Elec
t
rical
and
Computer
Eng
ine
ering
(
IJE
CE)
,
vol. 9,
no.
6
,
p
p
.
5185
-
5191
,
2019
.
[5]
A.
W
ahda
n,
S.
AL
Hantoobi,
S.
A.
Sall
oum
,
and
K.
Shaa
la
n,
“
A
S
y
stematic
Rev
i
ew
of
T
ext
cl
ass
ifi
c
at
ion
R
ese
ar
c
h
base
d
on
Dee
p
Le
arn
ing
Models
in
Arabi
c
La
nguag
e,”
Inter
nati
onal
Journ
al
of
Elec
tri
cal
and
Computer
Engi
ne
ering
(
IJ
ECE
)
,
vol
.
10
,
n
o.
6
,
pp
.
6629
-
6
643,
De
c. 2020
.
[6]
Y.
Al
-
Am
ran
i,
M.
La
zaar,
and
K.
E.
El
kad
iri
,
“
Senti
m
ent
ana
l
ysis
using
supervise
d
cl
assificat
io
n
al
gorit
hm
s,”
i
n
ACM
Int
ernati
o
nal
Conf
ere
nce
Proce
ed
ing
S
eri
es
,
vol
.
Par
t
F12
94
,
2017
.
[7]
Y.
Al
-
Am
ran
i,
M.
La
z
aa
r
,
K.
Eddi
ne
,
and
E.
L.
Kadir
i,
“
Sent
iment
Anal
y
sis
Us
ing
Hy
b
rid
Method
of
Support
Vec
tor
Ma
chi
n
e and
Dec
ision Tr
ee
,
”
J. Theor.
Ap
pl.
In
f. Tec
hnol
.
,
vol. 96
,
no.
7,
p
p.
1886
-
1895
,
2
018.
[8]
G.
Al
-
Bdour,
R.
Al
-
Qurran,
M.
Al
-
A
y
y
oub,
A.
Shatna
wi,
“
Benchm
ark
ing
open
source
dee
p
l
ea
rn
ing
fra
m
eworks,”
Inte
rnation
al
Jo
urnal
of El
e
ct
ri
c
al
and
Comput
er
Engi
n
ee
ring
(
IJE
CE)
,
vol
.
10
,
n
o.
5
,
pp
.
5479
-
5
486
,
Oct
.
2020
.
[9]
N
.
P.
Shetty
,
B
.
Muni
y
al
,
A.
A
nand,
S.
Kum
ar,
and
S.
Prabhu
,
“
Predic
ti
ng
d
ep
ression
using
de
ep
learni
ng
and
ense
m
ble
al
gor
ithm
s
on
raw
twit
te
r
da
ta,”
Int
er
n
ati
onal
Journal
of
Elec
tri
cal
and
Computer
Enginee
ring
(
IJE
CE)
,
vol.
10
,
no
.
4
,
pp
.
3751
–
3756
,
Au
g.
2020
.
[10]
Y.
Al
Am
ran
i,
M.
Laza
ar
,
and
K.
E.
El
Kad
iri,
“
A
novel
h
y
b
ri
d
cl
assifi
cation
appr
oac
h
for
se
nti
m
ent
an
aly
sis
of
te
x
t
docume
nt,
”
Inte
rnat
ion
al
Journa
l
of
El
e
ct
rica
l
and
Computer
Enginee
ring
(
IJE
CE
)
,
vol.
8
,
no
.
6
,
pp.
4554
–
4567
,
2018.
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
om
p
En
g
IS
S
N: 20
88
-
8708
A powerf
ul c
omp
ar
is
on o
f
de
ep
le
arni
ng fra
meworks
for
a
r
ab
ic
sentime
nt
analysis
(
Y
ou
s
sra
Zahidi
)
751
[11]
M.
Abadi
et
a
l.
,
“
Te
nsorF
low:
A
sy
st
em
for
large
-
sca
l
e
m
ac
hi
ne
learni
ng
,
”
in
Proce
ed
ings
of
t
he
12th
USENI
X
Symposium on
Operating
S
yste
ms
Design
and
Imple
menta
ti
on,
OS
D
I
2016
,
pp
.
265
–
283
,
2016
.
[12]
The
Theano
Deve
lopment
Tea
m
et
al.
,
“
The
a
no:
A
Py
thon
f
ramework
for
f
ast
computat
ion
of
m
at
hemati
c
al
expr
essions,”
Co
RR
,
Ma
y
2016
.
[13]
F.
Bastien
et
a
l.
,
“
The
ano:
n
ew
f
e
at
ure
s
and
spe
ed
improvem
ent
s,”
CoRR
,
vo
l. a
bs/
1211.
5,
Nov.
20
12
.
[14]
J.
Bergstra
e
t
a
l.
,
“
The
ano: De
ep Le
arn
ing
on
GP
Us
with
P
y
thon,”
J
.
Ma
ch. Learn.
R
es.
,
vol
.
1
,
pp
.
1
–
48,
2011.
[15]
M.
M.
Fouad
,
A
.
Maha
n
y
,
N
.
Al
joha
ni
,
R
.
A.
Abbasi,
and
S.
U.
Hass
an,
“
ArW
or
dVec
:
eff
i
ci
en
t
word
embeddin
g
m
odel
s for
Arab
ic
twee
ts
,
”
So
ft
Compu
t.
,
vo
l. 24
,
no
.
11
,
pp
.
806
1
–
8068,
2020
.
[16]
R.
Baly
et
al
.
,
“
Com
par
at
ive
E
val
ua
ti
on
of
Senti
m
ent
Anal
y
s
i
s
Methods
Acro
ss
Arabi
c
Dial
e
ct
s,”
in
Proc
edia
Computer
Scien
ce
,
2017,
vol
.
11
7,
pp
.
266
–
273
.
[17]
E.
Om
ara,
M.
Mos
a,
and
N
.
I
sm
ai
l,
“
Dee
p
C
onvolut
ional
Ne
twor
k
for
Arabi
c
Senti
m
en
t
An
aly
s
is,”
in
2018
Proce
ed
ings
of
t
he
Japan
-
Af
rica
Confe
renc
e
on
E
le
c
tronic
s,
Com
municat
ions,
a
n
d
Computati
ons,
JA
C
-
ECC
2018
,
pp.
155
–
159
,
20
19
.
[18]
M.
Alfonse
and
E.
-
S.
M.
E
l
-
Horbat
y
,
“
Opinion
Mining
for
Ara
bic
Dia
lects
on
Twit
ter,”
Eg
y
pt.
Comput.
Sc
i.
J
.
,
vol.
42
,
no
.
4
,
pp
.
52
–
61
,
2018
.
[19]
R.
M.
Alahmar
y
,
H.
Z.
Al
-
Dos
sari,
and
A.
Z.
Emam
,
“
Senti
m
ent
an
aly
sis
of
saudi
diale
ct
usi
ng
dee
p
learni
n
g
te
chn
ique
s,”
in
ICEIC
2019
-
Inte
rnational
Confe
renc
e
o
n
El
ectronics,
Information,
an
d
Comm
uni
cat
ion
,
Auckla
nd,
New
Ze
a
la
nd,
pp.
1
-
6
,
2019.
[20]
H.
Abdell
aou
i
a
nd
M.
Zri
gui
,
“
Us
ing
twee
ts
an
d
emojis
to
build
TE
AD
:
An
Arabi
c
d
at
as
et
for
senti
m
ent
an
aly
s
is,
”
Comput.
y
Sist
.
,
vol.
22
,
no
.
3
,
pp
.
777
–
786
,
2018
.
[21]
R.
Baly
,
e
t
al.,
“
Com
par
at
ive
E
val
ua
ti
on
of
Se
nti
m
ent
Anal
y
si
s
Methods
Acro
ss
Arabi
c
Dial
e
ct
s,”
in
Proc
edia
Computer
Scien
ce
,
vol
.
117
,
pp
.
266
–
273
,
2017
,
.
[22]
A.
El
naga
r
,
L
.
L
ulu,
and
O.
Ei
n
e
a,
“
An
Annotat
e
d
Huge
Data
set
for
Standa
rd
and
Coll
oquia
l
Ara
bic
Reviews
for
Subjec
ti
v
e
Sen
tim
ent
Anal
y
sis,
” in
Pr
oc
edi
a
Computer
Scienc
e
,
vol.
142
,
pp
.
182
–
189
,
2018
.
[23]
S.
Rosentha
l,
N.
Farra
,
and
P.
N
akov,
“
Sem
Eva
l
-
2017
Ta
sk
4:
S
ent
iment
Anal
y
s
is
in
Twit
te
r,”
i
n
Proce
edi
ngs
of
the
11
th
Int
ernat
ional
Workshop
on
Semantic Eva
luat
ion
(
SemE
va
l
-
2017)
,
pp.
502
–
518
,
2017
.
[24]
A.
Barhoumi,
Y.
Estè
v
e,
C.
Alou
lou,
and
L
.
H
.
B
el
guit
h
,
“
Docum
ent
embeddings
for
Arabi
c
Sent
i
m
ent
Anal
y
sis,”
in
Proc
ee
dings
o
f
th
e
F
irst Confe
renc
e
on
Langua
ge
Proc
essing
a
nd
Knowle
dg
e Manageme
nt
,
vo
l.
1988
,
2017
.
[25]
S.
Al
-
Aza
ni
and
E.
S.
El
-
Alf
y
,
“
Emojis
-
base
d
se
nti
m
ent
c
la
ss
ifi
c
at
ion
of
Arab
ic
m
ic
roblogs
using
dee
p
r
ec
urre
n
t
neur
al
n
et
works
,
”
in
2018
Int
e
rnational
Confer
enc
e
on
Computing
Scienc
es
and
Engi
nee
rin
g,
ICCSE
2018
-
Proce
ed
ings
,
pp
.
1
–
6
,
2018
.
[26]
S.
Al
-
Aza
ni
an
d
E.
S.
M.
El
-
Alf
y
,
“
H
y
brid
Dee
p
L
ea
rning
f
or
Senti
m
ent
Polari
t
y
D
eterm
ina
ti
on
of
Arabi
c
Microbl
ogs,”
in
Lect
ure
Note
s
i
n
Computer
Scienc
e
(
inc
lud
ing
subs
erie
s
Lect
ure
Note
s
in
Art
if
i
ci
al
Int
ellige
n
ce
and
Lect
ure
Not
es
in Bi
oin
formatic
s)
,
vol
.
10635
LNCS,
pp.
491
–
500
,
2017
.
[27]
M.
Al
-
Sm
adi
,
B.
Ta
l
afh
a
,
M.
Al
-
A
y
y
oub,
and
Y.
Jara
rweh,
“
Us
ing
long
short
-
t
er
m
m
emory
de
ep
neur
al
net
works
for
aspe
c
t
-
base
d
senti
m
ent
an
aly
s
is of
Arabi
c
rev
i
ews,”
Int
.
J
.
Ma
ch.
Learn.
Cybe
r
n.
,
vo
l.
10
,
no
.
8,
pp.
2163
–
2175
,
Aug.
2019.
[28]
A.
Dahou,
S.
Xiong,
J.
Zhou,
M.
H.
Haddoud,
a
nd
P.
Dua
n,
“
W
ord
embe
ddings
and
con
volut
ional
neur
a
l
net
work
for
Arabi
c
senti
m
ent
class
ifi
cation,”
in
COL
ING
2016
-
26th
Inte
rnation
al
Confe
renc
e
o
n
Computati
onal
Linguisti
cs,
Pro
ce
ed
ings o
f
COL
ING 2016:
Tech
nic
al
Pape
rs
,
pp
.
2418
–
2427
,
20
16
.
[29]
S
.
Al
-
Aza
ni
and
E.
S.
El
-
Alf
y
,
“
Emojis
-
base
d
se
nti
m
ent
c
la
ss
ifi
c
at
ion
of
Arab
ic
m
ic
roblogs
using
dee
p
r
ec
urre
n
t
neur
al
n
et
works
,
”
in
2018
Int
e
rnational
Confer
enc
e
on
Computing
Scienc
es
and
Engi
nee
rin
g,
ICCSE
2018
-
Proce
ed
ings
,
pp
.
1
–
6
,
2018
.
[30]
M.
Shahin,
J.
E
pps,
and
B
.
Ah
m
ed,
“
Autom
at
ic
class
ifi
cation
of
le
xi
ca
l
str
ess
in
Engl
ish
and
Arabi
c
l
angua
g
es
using
dee
p
l
ea
r
ning,
”
in
Proc
e
edi
ngs
of
th
e
A
nnual
Confe
ren
ce
of
the
In
te
r
nati
onal
Sp
ee
ch
Comm
unic
ati
o
n
Associat
ion
,
INT
ER
SP
EE
CH
,
pp.
175
–
179
,
2016
.
[31]
P.
Nakov,
et
a
l.
,
“
Sem
Eva
l
-
2016
Ta
sk
4:
Sen
ti
m
ent
Anal
y
sis
in
Twit
ter,”
in
Proce
ed
ings
of
th
e
10th
Inte
rnat
ion
al
Workshop on
Se
mantic
Ev
alua
tion (
SemE
val
-
20
16)
,
pp.
1
–
18
,
2
016
.
[32]
A.
M.
Ala
y
b
a,
V.
Pala
d
e,
M.
E
ngla
nd,
and
R
.
I
qbal
,
“
A
combined
CNN
and
L
STM
m
odel
for
Arabi
c
sen
ti
m
en
t
ana
l
y
sis,
”
in
Le
ct
ure
Not
es
in
C
omputer
Sci
en
ce
(
inc
ludi
ng
subs
erie
s
Lect
ure
N
ote
s
in
Arti
f
ic
ia
l
Inte
l
li
gen
ce
an
d
Lect
ure
No
te
s in
Bi
o
inf
orm
ati
cs)
,
vol. 11015 LNCS
,
pp.
179
–
191
,
2018
.
[33]
S.
R.
El
-
B
elta
g
y
,
M.
E
l
kalam
aw
y
,
and
A
.
B.
Solim
an
,
“
Nile
TMRG
at
Sem
E
val
-
2017
Ta
sk
4:
Arab
ic
Senti
m
ent
Anal
ysis,”
in
Proc
ee
di
ngs
of
the
11th
I
nte
rnational
Wo
rkshop
on
Semanti
c
E
val
uat
ion
(
SemE
val
-
2017)
,
pp.
790
–
795
,
20
17
.
[34]
B.
J.
Erickson,
P.
Korfia
ti
s,
Z.
Akkus
,
T.
Kline,
and
K.
Philbr
i
ck,
“
Tool
ki
ts
an
d
Li
bra
r
ie
s
for
Dee
p
L
e
arn
ing
,
”
Journal
of
Digital Imagi
ng
,
Spri
nger
New York LLC,
vol
.
30
,
no
.
4
,
pp
.
400
–
405
,
Aug
.
2017
.
[35]
A.
Parva
t
,
J
.
Cha
van,
S
.
Kad
am, S
.
Dev, a
nd
V.
P
at
hak
,
“
A surve
y
of
de
ep
-
l
ea
rn
in
g
fra
m
eworks,” i
n
Proceedi
ngs o
f
the
In
te
rnationa
l
Confe
ren
ce on
I
nve
nt
iv
e
Syst
ems and
Control
,
IC
ISC
2017
,
2017.
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol. 11,
No.
1,
Febr
uar
y
2021 :
7
45
-
752
752
BIOGR
AP
H
I
ES
OF
A
UTH
ORS
You
ss
ra
Z
a
hid
i
is
a
Ph
.
D.
student
in
Com
pute
r
Scie
nc
e,
Infor
m
at
ion
S
y
stem
,
and
Software
Engi
ne
eri
ng
La
b
ora
tor
y
,
Abdelm
al
ek
Essaa
d
i
Univer
sit
y
,
T
et
u
an,
Morocc
o
.
Sh
e
is
a
Com
pute
r
Scie
nc
es
en
gineer,
gra
dua
te
d
i
n
201
7
fro
m
t
he
Nati
onal
Sc
hool
of
Applie
d
Scie
nce
s,
Abdelmale
k
Essaa
di
Univer
sit
y
.
Yacine
El
You
nou
s
si
is
a
Ph
.
D.
doct
or
and
p
rofe
ss
or
of
computer
sci
ences
a
t
the
Na
ti
ona
l
School
of
Applie
d
Sc
ie
nc
es
o
f
Te
tu
an,
Info
r
m
at
ion
S
y
st
em
and
Software
Engi
ne
eri
ng
La
bora
tor
y
,
Ab
del
m
al
ek
Essaa
di
Univer
sit
y
,
T
et
uan
,
Morocc
o
.
He
is
a
superv
isor
of
m
an
y
The
sis
,
and
h
e is pa
rt
of
m
an
y
bo
ard
s of
in
te
rn
at
i
onal
journa
ls
an
d
interna
t
iona
l
c
onfe
ren
c
es.
Yas
sin
e
Al
-
A
mr
ani
is
a
Ph
.
D.
doct
o
r
and
profe
ss
or
of
Com
pute
r
Scie
nce
s
at
the
Multi
dis
ci
pl
ina
r
y
Facult
y
o
f
La
ra
che
,
Abd
emale
k
Essaa
d
i
Univer
sit
y
,
Morocc
o
.
He
is
a
Com
pute
r
Sci
enc
es
E
ng
ine
er
,
and
h
e
is
par
t
of
m
an
y
boar
ds
of
I
nte
rna
ti
ona
l
J
ourna
ls
and
I
nte
rna
ti
ona
l
C
o
nfe
ren
c
es
.
Evaluation Warning : The document was created with Spire.PDF for Python.