Indonesi
an
Journa
l
of El
ect
ri
cal Engineer
ing
an
d
Comp
ut
er
Scie
nce
Vo
l.
13
,
No.
3
,
Ma
rch
201
9
, p
p.
902
~
909
IS
S
N: 25
02
-
4752, DO
I: 10
.11
591/ij
eecs
.v1
3
.i
3
.pp
902
-
909
902
Journ
al h
om
e
page
:
http:
//
ia
es
core.c
om/j
ourn
als/i
ndex.
ph
p/ij
eecs
A compa
rative s
t
ud
y o
f
sentim
ent anal
ysis usin
g SVM and
SentiWo
rdNet
Moham
ma
d F
ikri
, R
iy
anar
t
o
S
arno
Depa
rtment
o
f
I
nform
at
ic
s,
Insti
tut
Te
knolog
i
Se
puluh
Nopem
ber
,
Indone
si
a
Art
ic
le
In
f
o
ABSTR
A
CT
Art
ic
le
history:
Re
cei
ved
Ma
y
21
, 201
8
Re
vised
N
ov
13
, 2
018
Accepte
d
Dec
5
, 2
018
Senti
m
ent
an
alys
is
has
grown
ra
pidly
and
impacts
on
the
num
ber
of
services
using
the
int
e
rn
et
popping
up
i
n
Indone
sia.
In
thi
s
rese
ar
ch,
th
e
senti
m
ent
ana
l
y
sis
uses
th
e
rule
-
base
d
m
et
hod
with
the
hel
p
of
Sen
ti
W
ordNet
and
Support
Vec
tor
Mac
hine
(SV
M)
al
gori
thm
with
Te
rm
Freq
uency
-
Inve
rse
Docum
ent
Freq
uency
(
TF
-
IDF
)
as
a
fe
at
ure
ex
t
rac
t
ion
m
et
hod
.
The
d
at
a
as
the
c
ase
stud
y
f
or
the
sent
iment
ana
l
y
sis
is
writ
te
n
in
Indon
esian
la
nguag
e.
Since
th
e
num
ber
of
sent
ences
in
positi
v
e,
n
eg
at
iv
e
and
n
eut
r
a
l
class
es
is
imbala
nc
ed,
the
over
sam
pli
ng
m
et
hod
is
implemente
d.
For
i
m
bal
anc
ed
dat
ase
t,
th
e
rul
e
-
base
d
Senti
W
or
dNet
and
SV
M
al
gorit
hm
ac
h
ie
ve
ac
cur
acie
s
of
56%
and
76%,
respe
c
ti
ve
l
y
.
How
eve
r,
for
t
he
bal
an
ce
d
d
ataset
,
the
rul
e
-
base
d
Senti
W
ordNet
and
SV
M
al
gorit
hm
ac
h
i
eve
a
cc
ura
cies
of
52%
and
89%,
r
espe
ctively
.
Ke
yw
or
ds:
Rule
-
ba
sed
Sentim
ent an
al
ysi
s
Sentiw
ord
net
Suppor
t
v
ect
or m
achine
Wor
dn
et
Copyright
©
201
9
Instit
ut
e
o
f Ad
vanc
ed
Engi
n
ee
r
ing
and
S
cienc
e
.
Al
l
rights re
serv
ed.
Corres
pond
in
g
Aut
h
or
:
Ri
ya
nar
to Sa
r
no,
Dep
a
rtm
ent o
f Info
rm
at
ic
s,
In
sti
tut Te
knol
og
i
Sepulu
h N
op
em
ber
,
Jal
an
Ra
ya
I
T
S
, K
e
puti
h,
Suk
olil
o,
Ko
ta
Sur
abaya, Ja
wa
Ti
m
ur
6
01
11
, In
donesia.
Em
a
il
: riy
anar
to@if
.it
s.ac.id
1.
INTROD
U
CTION
W
it
h
th
e
inc
re
ase
in
t
he
nu
m
ber
of
inte
rn
et
s,
us
ers
can
op
en
a
n
opport
unit
y
to
giv
e
go
od
im
pact
to
an
orga
nizat
io
n
beca
us
e
of
t
he
data
gen
e
ra
te
d
thr
ough
int
ern
et
use
r
act
ivit
y.
These
data
can
be
op
i
nio
ns
or
facts
a
bout
s
om
et
hin
g.
This
r
esearc
h
f
ocu
s
es
on p
ub
li
c op
inions
a
bout p
r
oducts
i
n
the
f
or
m
of
ap
plica
ti
on
s
on
sm
artph
on
e
s.
These
opinio
ns
can
be
furthe
r
analy
zed
f
or
obta
inin
g
c
onsiderati
on
of
the
decisi
on
-
m
aking
in
a
com
pan
y t
hat c
reates t
he
app
li
cat
ion
. A
m
ong t
he
var
io
us
techn
ic
al
an
al
yz
es
, th
e tec
hn
iq
ue
is cal
le
d
senti
m
ent
analy
sis
[1]
.
This
te
chn
i
que
pr
oce
sses
te
xt
do
c
um
ents
in
th
e
form
o
f
op
i
nions
to
gen
e
rate
a
piece
of
inf
or
m
at
ion
so
that
in
form
at
i
on
can
be
us
ed
to
div
i
de
op
i
ni
on
s
into
posit
ive,
ne
gative,
or
neu
tral
opini
on
s
.
In
the
de
velo
pm
e
nt
of
in
form
at
i
on
te
ch
no
l
og
y,
opinio
n
m
ining
is
on
e
of
the
favor
it
e
r
esear
ch
to
pics
in
th
e
fiel
d
of N
at
ur
al
La
ngua
ge Pr
ocessi
ng (NLP
).
In
this
resear
c
h,
we
com
par
e
the
i
m
ple
m
ent
at
ion
of
s
up
e
rvi
sed
m
achine
le
arn
i
ng
a
nd
r
ul
e
-
base
d
f
or
sentim
ent
analy
sis
us
in
g
data
from
Go
ogle
Play
store
a
nd
A
pp
le
Appst
ore
w
ritt
en
in
I
ndonesi
an
la
ngua
ge
.
The
m
et
ho
d
to
get
the
data
is
the
sa
m
e
m
e
thods
as
the
m
et
hod
in
these
sever
al
pa
pe
rs
[2
-
3]
.
Eac
h
pr
oduct
rev
ie
w
a
case
f
old
in
g
proce
ss,
norm
al
iz
a
tio
n
of
punc
tua
ti
on
,
norm
al
izati
on
of
the
sl
ang
w
ord
,
st
opw
ord
rem
ov
al
,
trans
form
ation
int
o
sing
le
li
ne
,
and
t
okenizat
ion
will
be
ca
rr
ie
d
out
as
s
ta
te
d
on
[4
-
5]
.
For
i
m
ple
m
entat
io
ns
us
i
ng
s
uper
vised
m
achine
le
arn
in
g,
we
us
e
Term
Fr
eq
uen
cy
-
I
nv
e
rse
Do
c
um
ent
Fr
equ
e
nc
y
(TF
-
I
DF
)
t
o
conve
rt
te
xt
into
cl
assifi
able
f
eat
ur
es
a
nd
S
uppo
rt
Vecto
r
Ma
chines
to
c
la
ssify
the
pro
cesses.
Senti
Wo
r
dNet
does
not
suppo
rt
la
ngua
nges
oth
e
r
th
a
n
E
ngli
sh
;
w
her
eas
the
la
ngua
ge
of
the
data
i
s
Ind
on
esi
a
n
.
T
he
refor
e
,
transla
ti
ng
the
opinio
ns
into
E
ng
li
s
h
is
need
ed,
s
o
that
resu
lt
of
th
e
transla
ti
on
ca
n
be
done
by
the
an
al
ysi
s
of
the
op
inio
ns
[6
-
7]
.
This
resea
rch
c
on
sist
s
of
sect
ion
2
that
exp
la
ins
the
us
e
d
m
et
hod,
sect
ion
3
t
hat e
xp
la
in
s the
r
es
ults o
f
the
ana
ly
sis, and s
ect
io
n 4 th
at
c
onta
ins t
he
c
on
cl
us
ion
s
.
Evaluation Warning : The document was created with Spire.PDF for Python.
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci
IS
S
N:
25
02
-
4752
A com
pa
r
ative
stud
y
o
f se
ntim
ent an
alysis u
s
ing
SVM
an
d S
entiW
ordNet
...
(
Riy
anar
to
Sa
rn
o
)
903
2.
RESEA
R
CH MET
HO
D
The
m
ai
n
ob
j
e
ct
ive
of
t
his
st
ud
y
is
to
com
par
e
te
xt
cl
assi
ficat
ion
al
gorithm
s
between
us
in
g
a
r
ule
-
base
d
al
gorith
m
with
the
help
of
Senti
Wo
r
dN
et
a
nd
usi
ng
the
com
bin
at
i
on
of
TF
-
I
DF
al
gorithm
and
Suppor
t
Vecto
r
Ma
chine
(SVM)
al
gorithm
.
TF
-
I
D
F
extracts
featur
es
f
r
om
te
xt
to
a
vector
and
S
VM
cl
assifi
es
an
i
m
balanced
da
ta
set
into
the
num
ber
of
posit
ive,
ne
gative,
and
ne
utral
cl
asses.
T
he
res
ults
of
eac
h
al
gorithm
us
e
d
will
be
s
ort
ed
a
nd c
om
par
ed
b
ase
d
on t
he
sc
or
e
s
of
t
he
F
-
S
co
re a
nd
Accuracy
.
2.1.
Data C
onstruc
tio
n
The
data
set
co
ntains
public
op
i
nions
a
bout
so
m
e
app
s
.
T
ho
s
e
opini
ons
are
w
ritt
en
in
Ind
on
esi
a
n
la
nguag
e
an
d
a
re
ta
ke
n
f
r
om
Goo
g
le
Play
Stor
e
a
nd
Apple
AppS
t
or
e
.
T
he
data
co
ns
ist
s
of
“i
d_kom
en”
as
the
identific
at
ion
cod
e
of
c
omm
ents,
“t
it
le
_k
om
en”
as
the
ti
tl
e
of
com
ments,
an
d
“
Kom
en”
as
the
de
ta
il
ed
com
m
ents.
Th
e
sentim
ents
fo
r
each
sente
nc
e
are
dete
rm
in
ed
by
hu
m
ans
into
th
ree
cl
ass
es,
i.e.
posit
ive
cl
ass,
neu
t
r
al
cl
ass,
and
ne
gative
cl
ass.
The
dat
aset
con
ta
in
s
553
se
ntence
s
wh
ic
h
are
25
9
posit
ive
cl
a
ss,
24
1
neg
at
ive
cl
ass,
an
d
53
neu
tral
cl
ass.
T
he
pos
it
ive
cl
ass
an
d
neg
at
ive
cl
ass
are
balance
d;
howe
ver,
the
ne
utra
l
cl
ass
is
im
bala
nced
bec
ause
the
neu
t
ral
cl
as
s
has
fe
wer
se
ntences
tha
n
th
e
oth
e
rs.
T
he
da
ta
is
store
d
i
n
a
data
fr
am
e
that
has
"C
OMME
NT
"
colum
n
fo
r
com
m
ent,
"SENTI
ME
NT"
c
olu
m
n
fo
r
the
correct
sentim
ent,
an
d
"SENT
IMEN
T
_ID" c
olu
m
n
f
or sen
ti
m
ent id,
i.e. "
0"
a
s
negat
ive, "1
"
as
po
sit
ive, and
"
2" a
s n
e
utr
al
.
2.2.
B
ala
ncin
g Data
Ba
la
ncing an u
nb
al
a
nced d
at
aset
is a criti
cal
p
r
ocess
in
m
achine lea
rn
i
ng. T
he
m
et
ho
d us
e
d
this
tim
e
by ove
rsam
pli
ng the m
ino
rit
y cl
ass
[
8]
is s
how
n
in
Fig
ure
1
.
T
he
e
xp
la
na
ti
on
of Fig
ur
e
1
is
as
fo
ll
ows
:
1.
Ma
rk
in
g
t
he
M
inorit
y C
la
ss an
d M
aj
or
it
y C
la
ss.
First,
c
re
at
ing
on
e
col
um
n
na
m
ed
flag
_bal
ance
.
The
n,
m
ar
king
the
m
ino
r
it
y
cl
ass
(n
e
utr
al
)
by
fill
ing
in
the f
la
g_balanc
e
fiel
d wit
h 1 a
nd the m
ajorit
y cl
ass (p
os
it
ive and
neg
at
iv
e)
with
0.
2.
Sp
li
ts i
nto 2
Dat
a fr
am
es.
The
data
w
hic
h
has
1
in
the
flag
_b
al
a
nce
colum
n
be
com
e
the
m
ino
rity
data
f
ram
e
and
the
data
wh
ic
h
has 0 i
n
the
f
la
g_balance
c
olum
n
0
bec
om
e t
he
m
ajo
rity
d
at
a fr
am
e
[9]
.
3.
Re
sa
m
ple Th
e
Mi
no
rity
Cl
ass
D
at
af
ram
e.
The
fir
st
ta
sk
is
ov
e
rsam
pling
the
m
ino
rity
resam
pled
cl
ass
by
us
in
g
the
e
xisti
ng
al
go
rith
m
in
the
sci
kit
-
le
arn
[
9]
.
Af
te
r
that,
re
sam
pl
ing
ra
ndom
ly
un
ti
l
the
num
ber
of
m
ino
rit
y
cl
asses
e
qu
a
ls
the
a
ver
a
ge
nu
m
ber
of
m
ajo
rity
cl
asses.
I
n
this
resea
rc
h,
ne
utral
is
the
m
ino
rity
cl
ass
,
posit
ive
a
nd
neg
at
ive
is
the
m
ajo
rity
class
.
4.
Com
bin
e Th
e
Ma
j
ori
ty
Cl
ass D
at
af
ram
e and
T
he U
ps
am
pled
Mi
nority
.
First, m
erg
in
g bo
t
h data f
ram
e (m
ajo
rity
and
m
ino
rity
).
The
n,
ra
ndom
iz
ing
the se
quence
on th
e
d
at
a
fr
am
e so
t
hat
da
ta
are
m
erg
ed
into
rand
om
.
Figure
1. Ba
la
ncin
g Im
balanced Data
set
Process
2.3.
Pre
proce
ssing D
ata
Be
cause
Ind
onesi
an
la
ng
uag
e
us
i
ng
by
t
he
da
ta
is
inf
or
m
al
,
t
he
pr
e
process
ing
is
do
ne
to
change
t
he
te
xt
into
I
ndon
esi
an
l
an
guage
in
the
f
or
m
al
f
or
m
.
The
f
ol
lo
wing
pr
e
proces
sing st
eps
are
de
scribe
d
as
foll
ow
s:
1.
Enter” C
harac
te
r Norm
alizat
i
on.
Rem
ov
e "
\
n" or enter
on t
he se
ntence
to be
a sin
gle li
ne onl
y.
2.
Lo
wer
case
Norm
al
iz
ation
.
Turn
t
he
se
nte
nce in
t
o
al
l l
owercase
.
3.
Unnecessa
ry C
har
act
er
No
rm
al
iz
at
ion
.
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2502
-
4752
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci,
Vo
l.
13
, N
o.
3
,
Ma
rc
h 201
9
:
902
–
909
904
Delet
e
a
recur
rin
g
char
act
e
r
wh
et
he
r
it
i
s
an
al
ph
abet
or
a
punct
uation.
For
exam
ple
"Set
ia
aaa"
to
"Set
ia
" and “Y
ah.
.
...
"
to
"
Ya
h"
.
4.
Punctuati
on
N
or
m
al
iz
ation
.
Delet
e pun
ct
ua
ti
on
on the
sent
ence.
5.
Slang
Word N
or
m
al
iz
ation
.
Fixed
i
nfo
rm
al
w
ords
a
nd
a
bbre
viati
on
s
.
T
he
fix
us
es
a
m
anu
al
way,
not
s
pellc
heck
i
ng.
T
he
m
anual
way
m
eans
m
a
tc
hin
g
the
wor
d
with
a
has
h
m
ap
con
ta
ini
ng
sla
ng
wor
d
if
the
wor
d
m
a
t
ches
the
key
of
the
sla
ng
wor
d
hash
m
ap
the
n
the
w
ord
is
cha
ng
e
d
i
nto
the
value
of
the
key
of
the
has
h
m
ap.
F
or
exam
ple, abbr
e
viati
on
s
su
c
h
a
s "sp
t"
t
o be "s
eper
ti
" a
nd infor
m
al
w
ords
li
ke
"
pak
e"
to
"
pa
kai".
6.
Stopw
ord
Re
m
ov
al
Delet
e
the wor
ds
t
hat
oft
en
a
pp
ea
r
i
n
eac
h
s
entence.
T
he
t
ype
of
t
he
delet
ed
w
ord
is
a
c
onjun
ct
io
n
w
ord
,
su
c
h
as
"
da
n",
"serta"
,
"se
rta
",
an
d
oth
e
rs
.
Table
1
is
a
n
exam
p
le
of
t
he
prep
ro
ce
ssin
g
re
su
lt
s.
The
or
i
gin
al
te
xt
use
s
Indonesia
n
la
ng
ua
ge
in
the
inf
or
m
al
fo
rm
,
and
the
preprocessi
ng
re
su
lt
s
changes
the
la
nguag
e
of
or
i
gin
al
text
fro
m
infor
m
al
f
or
m
to
f
or
m
al
f
or
m
.
Table
1.
Pr
e
processin
g
Re
s
ul
ts
in
I
ndonesi
a
n
La
ngua
ge
No
Origin
al T
ex
t
Af
ter
Prepro
cess
in
g
T
ex
t
1
BETU
LI
N
DON
G
APLI
KASI
NYA
,
RUSAK
MU
LU
N
IH!
!
!!
b
etu
lin
do
n
g
aplikasin
y
a r
u
sak
m
elu
l
u
nih
2
Ap
p
nya
cr
ash
ter
u
s!!
Tolo
n
g
dip
erbaik
i agar
serv
ice nya
se
m
ak
i
n
baik
ap
p
nya
crash t
eru
s to
lo
n
g
dip
erbaik
i
serv
ice
n
y
a se
m
ak
in
baik
2.4.
Ru
le
-
B
ase
d U
sin
g Sen
tiw
ordn
et
The
pu
rpose
of
this
r
esearc
h
is
t
o
com
par
e
two
dif
fer
e
nt
m
e
tho
ds
a
nd
one
of
the
m
et
ho
ds
is
Senti
Wo
r
dNet
.
Me
an
wh
il
e,
the
process
of
cl
assifi
cat
ion
i
s
dif
fe
ren
t
bec
ause
t
he
S
enti
Wor
dN
et
is
cu
rr
e
ntly
ver
y l
im
i
te
d
an
d no
t y
e
t a
vaila
ble in I
ndonesi
an
la
ng
uag
e
.
1.
Tr
an
sla
te
D
ata
Be
cause
Se
ntiword
net
is
currently
sti
ll
lim
it
ed
to
the
Indonesia
n
la
nguag
e
,
there
f
ore
t
he
data
i
s
translat
ed
int
o
En
glish
la
ngua
ge
.
G
oogle
T
ra
ns
la
te
is
us
ed
as
the
la
nguage
translat
or
t
ool.
The
res
ults
of
this
translat
or
to
ol
can
be
assum
ed
quit
e
well
al
thou
gh
t
her
e
a
r
e
st
il
l
so
m
e
se
ntences
t
hat
do
no
t
have
the
c
orrect
sentence
str
uctur
e
.
2.
Tokeniz
at
io
n
an
d
PO
S Tag
ging
To
ken
iz
at
io
n
is
a
pr
oces
s
to
sp
li
t
on
e
sente
nc
e
into
a
piece
of
the
w
ord
.
A
t
this
pr
ocess
,
the
sente
nce
is split
into u
ni
gr
am
w
hic
h
m
eans s
e
ve
ral p
a
rts consi
sti
ng
of
1 piece
of a
word.
Af
te
r
the
t
ok
e
ni
zat
ion
pr
ocess
,
each
unig
ram
is
determ
ined
the
pa
rt
of
s
pe
ech
[
11]
.
T
here
are
8
par
ts
of
sp
eec
h
wh
ic
h
are
nouns,
pr
onou
ns
,
a
dject
ives
,
ve
rbs,
a
dverb
s
,
pr
e
posit
ion
s
,
co
njunc
ti
on
s
,
and
inter
j
ect
io
ns
.
H
oweve
r,
the
par
t
of
s
pee
ch
ta
g
is
a
Pe
nn
T
ree
bank
P
OS
ta
g
a
nd
Se
nti
W
or
dN
et
on
ly
has
four
ge
ne
ral
P
OS
ta
gs
of
no
un
(
N
),
verb
(
V
),
a
dject
ive
(
A
),
a
nd
a
dv
e
rb
(
R).
Fi
nally
,
co
nv
e
rtin
g
the
P
OS
ta
g
to
Se
nti
W
or
dNet
PO
S
tags
is
necessa
ry
[
12
]
with the
foll
ow
ing
r
ules :
a)
Noun (N
)
If
POS ta
gs
are
'
NN
'
, '
NN
S'
, 'N
N
P'
, '
NN
PS'
,
then
t
he
P
OS t
ags
a
re c
hange
d
int
o
'
N'
.
b)
Verb
(
V)
If
POS ta
gs are
'
VB'
,
'
VBD
'
, 'V
BG'
, '
VBN'
, ‘VBP'
o
r
‘VBZ'
,
the
n
the
P
OS
ta
gs
are
ch
a
ng
ed
int
o
'
V'
.
c)
Adject
iv
e
(
A)
If
POS ta
gs are
'
JJ'
,
'
JJR
'
, o
r
'
J
JS'
, th
en
t
he P
OS
ta
gs
a
re c
ha
ng
e
d
i
nto
'
A'
.
d)
Adver
b
(R)
If
POS ta
gs are
'
RB'
, 'RBR'
, o
r
'
RB
S
'
, th
en
t
he
POS tag
s ar
e
change
d
int
o
'
R'
.
The
la
tt
er
on
this
proce
ss
is
done
le
m
m
atizat
ion
.
L
em
m
at
iz
at
ion
is
a
process
w
her
e
a
w
ord
is
returne
d
to
it
s
basic f
orm
b
ack in ac
co
r
dan
c
e w
it
h
t
he
P
OS t
ag.
3.
Sent
im
en
t Cla
ssific
at
i
on
The
se
ntim
ent
cl
assifi
cat
ion
in
this
stu
dy
use
s
Se
nti
W
or
dnet
an
d
Wo
r
dnet
to
ols.
Sent
iWo
rdnet
is
us
e
d
to
fin
d
the
scor
e
of
each
synset
a
nd
Wor
dn
et
is
us
ed
to
sea
rch
f
or
synon
ym
s
of
each
wor
d
bein
g
a
naly
zed
.
Evaluation Warning : The document was created with Spire.PDF for Python.
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci
IS
S
N:
25
02
-
4752
A com
pa
r
ative
stud
y
o
f se
ntim
ent an
alysis u
s
ing
SVM
an
d S
entiW
ordNet
...
(
Riy
anar
to
Sa
rn
o
)
905
Scores
f
or
eac
h
w
ord
a
re
sea
rch
e
d
us
in
g
Se
nti
W
or
dn
et
ac
cordin
g
to
POS
ta
gs
if
t
he
sc
or
es
a
re
m
or
e
than 0 t
hen it
is t
ake
n,
oth
e
rwi
se it
is b
ypass
ed.
Af
te
r
se
ntim
en
t
scor
es
pe
r
w
ord
are
obta
in
ed,
we
hav
e
t
o
do
a
total
cal
culat
ion
to
ge
t
senti
m
ent
scor
e
for
one
sentence.
T
he
se
m
antic
or
ie
ntati
on
cal
culat
ion
us
es
the
m
et
ho
d
accor
ding
to
E
quat
i
on
(1)
a
nd (2).
=
∑
∈
(1)
=
∑
∈
(2)
Ba
sed
on
E
qua
ti
on
(
1),
(
2),
S
cor
e
positive
is
th
e
final
num
ber
of
posit
ive
sco
res
w
hile
the
S
cor
e
negative
is
the
final
nu
m
ber
of
ne
gative
scor
e
s.
A
nd
n
is
the
num
ber
of
wor
ds
w
hos
e
first
sente
nc
e
value
is
a
bove
0.
The
n,
t
o get se
nti
m
ent v
al
ue
,
Eq
uation
(
3)
is ap
plied.
{
−
≥
0
.
05
−
≤
−
0
.
05
−
0
.
05
<
−
<
0
.
05
(3)
Sentim
ent
ob
ta
ined
va
lue
us
e
s
po
sit
ive
sc
or
e
diff
e
re
nce
an
d
neg
at
iv
e
sc
ore.
If
the
sco
re
di
ff
e
ren
ce
is
gr
eat
er
t
han
0.05
t
hen
t
he
sen
tim
ent
value
is
po
sit
ive.
If
t
he
scor
e
differe
nc
e
is
s
m
al
le
r
e
qu
al
to
-
0.0
5
th
en
the
sentim
ent
value
is
neg
at
ive
.
And
if
t
he
sco
re
dif
fer
e
nce
i
s
sm
a
ll
er
than
0.05
a
nd
great
er
tha
n
-
0.05
t
he
n
t
he
sentim
ent v
al
ue
is ne
utral.
2.5.
S
uper
vis
ed Machine
L
earnin
g
u
sing
SVM
In
this
sect
i
on,
the
T
F
-
IDF
m
et
hod
is
us
e
d
a
s
the
featu
re
e
xtracti
on
proce
ss
f
ro
m
te
xt
to
vect
or
an
d
SV
M i
s
us
e
d
a
s an al
gorithm
f
or text cl
assi
ficat
ion
.
2.5.1
Fe
ature
Extr
act
i
on
using
TF
-
I
DF
Term
Fr
equ
e
nc
y
-
Inve
rse
D
oc
um
ent
Fr
eq
ue
ncy
is
a
m
et
ho
d
f
or
co
nverti
ng
a
do
c
um
ent
(sen
te
nce
)
i
n
a
corpu
s
i
nto
a
sta
ti
sti
ca
ll
y
m
easur
a
ble
wei
ght
in
w
hich
thi
s
weig
ht
repre
sents
how
im
po
rta
nt
the
wor
d
is
in
the
doc
um
ent
or
phrase
[
13
]
.
The
re
are
se
ve
ral
ta
sk
s
t
o
tr
ansfo
rm
a
cor
pu
s
into
a
wei
gh
t
us
in
g
t
he
TF
-
IDF
m
et
ho
d.
a)
To
ken
iz
at
io
n
Do
c
um
ents
that
exist
in
a
corpu
s
a
re
to
ke
nized
into
unig
r
a
m
and
bi
gr
a
m
.
Un
igram
c
on
sist
s
of
on
e
word
a
nd
big
r
a
m
con
sist
s
of
2
word
s
.
The
tok
e
nizat
ion
pr
ocess
can
be
seen
in
Fig
ur
e
2.
Ba
sed
on
Fi
gure
2
,
al
l
of
unigr
am
s
an
d
big
ram
s
are
sti
ll
in
I
ndonesi
an
la
ngua
ge
beca
us
e
t
he
docum
ents
are
wr
it
te
n
i
n
Ind
onesi
an
la
nguag
e
.
Figure
2. To
ke
nizat
ion
Proces
s
of
Do
c
um
ents in In
done
sia
n Lan
guage
b)
Term
Fr
eq
uen
c
ie
s
Term
Fr
eq
uenci
es
(TF)
m
eas
ur
e
s
ho
w
of
te
n
a
w
ord
a
ppear
s
in
a
docum
ent.
It
is
po
ssi
ble
that
a
te
rm
would
a
pp
ea
r
m
uch
m
or
e
tim
es
in
l
ong
doc
um
ents
than
s
horter
ones.
Ter
m
Fr
equ
e
ncies
is
the
t
otal
co
unt
of
a
te
rm
in
a d
oc
um
ent.
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2502
-
4752
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci,
Vo
l.
13
, N
o.
3
,
Ma
rc
h 201
9
:
902
–
909
906
c)
Inverse
Do
c
ume
nt F
reque
ncy
Inverse
D
oc
um
ent
Fr
e
qu
e
nc
y
(IDF)
m
ea
su
res
ho
w
im
portant
a
te
r
m
to
a
doc
um
ent.
Wh
en
cal
culat
ing
te
r
m
fr
eq
uen
ci
es,
assum
ing
that
al
l
te
r
m
s
have
the
sam
e
i
mp
ort
ance
in
a
do
c
um
ent;
w
he
reas
conj
un
ct
io
nal
words
in
Ind
onesi
an
la
ngua
ge
su
ch
as
"da
n"
,
"adala
h",
an
d
"serta"
,
are
ver
y
oft
en
ap
pe
ar
in
sever
al
docum
ents (se
nte
nces
)
the
reb
y
re
du
c
ing
how
im
po
r
ta
nt the wo
r
d
i
s in
a
sen
te
nce.
(
,
)
=
ln
[
(
1
+
)
(
1
+
(
,
)
)
]
+
1
(4)
In
Eq
uatio
n
(4),
IDF(t,
d)
is
t
he
Inverse
Do
c
um
ent
Fr
eq
ue
nc
y
of
a
te
rm
in
a
doc
um
ent,
n
is
the
t
otal
nu
m
ber
of
the
do
c
um
ents,
DF(t,
d)
is
the
nu
m
ber
of
doc
um
ents
with
a
te
rm
(t)
in
it
.
T
he
ef
fect
of
a
ddin
g
“1”
to
the
IDF
in
the
equ
at
i
on
a
bove
is
that
te
rm
s
with
zero
ID
F;
i.e.
,
te
rm
s
that
occur
in
al
l
do
cum
ents
in
a
trai
ning
set
,
will
no
t
be
entirel
y
ign
or
e
d.
T
he
con
sta
nt
“1”
is
add
ed
to
the
nu
m
erator
an
d
de
no
m
inator
of
the
ID
F
as
if
a
n
e
xtra
do
c
um
ent
was
see
n
c
onta
ining
eve
ry
te
rm
in
the
colle
ct
ion
e
xactl
y
once,
w
hich
pr
e
ven
t
s
zero di
visions.
d)
Ca
lc
ulate
TF
-
I
DF
Weig
ht
In
t
he
process
of
cal
culat
ing
weig
hts
us
in
g
the
TF
-
I
DF
m
et
hod
wh
e
re
al
l
the
E
qu
at
io
n
s
us
e
d
a
re
in
accor
da
nce
with
E
quat
ion
(
4),
(5).
This
sect
ion
will
be
exe
m
pl
ifie
d
ho
w
t
he
cal
culat
io
n
of
wei
gh
ts
us
i
ng
the
TF
-
ID
F
m
et
ho
d.
TF
−
ID
F
(
t
,
d
)
=
TF
(
t
,
d
)
x
ID
F
(
t
,
d
)
(5)
e)
Norm
al
iz
e TF
-
ID
F
Weig
ht
Norm
al
iz
a
ti
on
is
done
so
t
hat
the
TF
-
IDF
va
lue
has
a
well
-
bala
nced
wei
gh
t.
N
or
m
al
iz
a
ti
on
is
done
us
in
g
L
2 n
or
m
so
t
hat the
wei
gh
t
of tf
-
idf
for ea
ch
te
rm
h
as
a w
ei
ght
of 0
-
1 sca
le
, s
ee
Eq
ua
ti
on
(6).
=
‖
‖
2
=
√
1
2
+
2
2
+
⋯
+
2
(6)
As
an
e
xam
ple
,
two
doc
um
ents
(D
1
a
nd
D2)
are
com
pu
te
d
the
TF
-
I
DF
va
lues.
Term
s
are
obta
ine
d
us
in
g
t
he
to
ke
nizat
ion
m
et
hod
as
sho
wn
on
Fi
gure
2.
D
F
is
the
doc
um
ent
fr
e
qu
e
nc
y
for
eac
h
Te
r
m
in
a
do
c
um
ent
(Dn
),
I
DF
is
the
i
nv
e
rse
do
c
um
e
nt
fr
e
quency
f
or
eac
h
Te
rm
cal
culat
ed
us
i
ng
Eq
uatio
n
(
4)
.
Ter
m
Fr
e
qu
e
ncy
–
Inv
erse
D
ocu
m
en
t
Fr
eq
uen
cy
(
TF
-
ID
F
)
f
or
e
ach
Term
in
a
do
c
um
ent
(D
n)
is
cal
culat
ed
us
i
ng
Eq
uation
(
5)
and
is
norm
alized
us
in
g
L2
No
rm
as
sh
own
by
Equ
at
i
on
(
6).
The
r
esults
are
expl
ai
ned
in Ta
ble 2.
T
he
term
s ar
e w
ritt
en
in
In
donesia
n
la
ng
uag
e
.
D
1
is “dia
baik
se
kali
”
an
d
D
2
is
“dia j
a
hat se
ka
li
.”
Table
2.
T
F
-
IDF
W
ei
gh
ti
ng
w
it
h
Term
s w
ritt
en
in
In
donesia
n
la
ng
uag
e
Ter
m
TF
DF
IDF
TF
-
I
DF
TF
-
I
DF
(L
2
Nor
m
)
D
1
D
2
D
1
D
2
D
1
D
2
d
ia
1
1
2
1
1
1
0
.35
5
0
.35
5
b
aik
1
0
1
1
.40
5
1
.40
5
0
0
.49
9
0
sek
ali
1
1
2
1
1
1
0
.35
5
0
.35
5
jah
at
0
1
1
1
.40
5
0
1
.40
5
0
0
.49
9
d
ia baik
1
0
1
1
.40
5
1
.40
5
0
0
.49
9
0
b
aik
sek
ali
1
0
1
1
.40
5
1
.40
5
0
0
.49
9
0
d
ia jahat
0
1
1
1
.40
5
0
1
.40
5
0
0
.49
9
jah
at sek
ali
0
1
1
1
.40
5
0
1
.40
5
0
0
.49
9
2.5.2
Sup
po
r
t
Vec
to
r
Machi
ne
Suppor
t
Vecto
r
Ma
chine
(
S
VM)
is
a
cl
assifi
cat
ion
m
eth
od
f
or
li
nea
r
or
no
nlinear
data
by
us
in
g
nonlinea
r
data
m
app
ed
t
o
c
onve
rt
trai
ning
data
to
a
hi
gher
dim
ension.
This
m
et
ho
d
find
hype
r
plane
by
m
axi
m
iz
ing
m
arg
i
n or
distan
ce betwee
n cl
asses
[
14]
, [1
5]
.
Con
si
der
i
ng
t
he
cl
ass
in
cl
assifi
cat
ion
,
the
one
vs
rest
strat
egy
is
i
m
ple
m
ented,
t
his
stra
te
gy
con
sis
ts
in f
it
ti
ng one cl
assifi
er
per cl
ass.
2.6.
Co
m
pari
ng
Resul
ts
Re
su
lt
s
from
t
he
cl
assifi
cat
io
n
of
r
ule
-
base
d
usi
ng
Se
ntiWor
dN
et
a
nd
su
pe
r
vised
m
achine
le
ar
ning
and
us
in
g
S
V
M
al
go
rithm
with
TF
-
I
DF
a
s
featur
e
e
xtra
ct
ion
are
c
ompare
d
by
usi
ng
Re
cal
l,
Pr
eci
sion,
F
-
Score
par
am
eter
s.
Evaluation Warning : The document was created with Spire.PDF for Python.
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci
IS
S
N:
25
02
-
4752
A com
pa
r
ative
stud
y
o
f se
ntim
ent an
alysis u
s
ing
SVM
an
d S
entiW
ordNet
...
(
Riy
anar
to
Sa
rn
o
)
907
Pr
eci
sio
n
is
the
abili
ty
of
a
cl
assifi
cat
ion
m
od
el
to
id
entify
only
the
releva
nt
dat
a
po
i
nts,
se
e
Eq
uation
(
7)
.
Re
cal
l
is
the
abili
ty
of
a
m
od
el
to
fin
d
al
l
th
e
releva
nt
case
s
within
a
data
set
,
see
Eq
uati
on
(
8)
.
F
-
Sc
or
e
is
th
e
har
m
on
ic
m
ean
of
preci
sio
n
and
recall
ta
kin
g
bot
h
m
et
ric
s
into
acc
ount
in
the
E
qu
at
io
n
(9)
.
Accuracy
is th
e quali
ty
o
r
sta
te
o
f
b
ei
ng corr
ect
o
r
preci
se,
see
Eq
uatio
n
(
10).
=
(
+
)
(7)
=
(
+
)
(8)
−
=
2
(
)
(
+
)
(9)
=
+
+
+
+
+
+
+
(10)
The
n
f
or
the
s
plit
between
tr
ai
nin
g
data
an
d
te
sti
ng
da
ta
us
in
g
K
-
F
old
Cros
s
Vali
dation
m
et
ho
d.
S
o
the
data
are
div
ide
d
into
k
f
ol
d
and
the
n
will
be
execu
te
d
c
la
ssific
at
ion
process
as
m
uch
as
the
k
and
f
or
the
te
sti
ng
data
is
sel
ect
ed
from
on
e
of
k
fo
l
d
a
n
d
tr
ai
ning
data
is
fo
ld
w
hich
are
not
us
e
d
a
s
the
data
te
sti
ng
[16]
.
The
sel
ect
ion
of
data
te
sti
ng
per
r
ound
is
sel
ect
ed
in
sequ
e
nce
sta
rtin
g
from
the
fo
ld
s
1,
see
Figure
3
for
th
e
il
lustrati
on
.
F
or e
xam
ple, r
ou
nd 1 is
us
ed
as
data te
sti
ng fol
d 1 a
nd s
o on.
Figure
3.
K
-
F
ol
d
Cr
os
s
-
Vali
da
ti
on
3.
RESU
LT
S
AND A
N
ALYSIS
Im
ple
m
entation
of
this
rese
arch
is
c
reate
d
by
us
in
g
P
yt
ho
n
Pro
gr
a
m
m
ing
Langu
age.
T
he
t
otal
nu
m
ber
of
t
he
dataset
is
55
3
data
with
detai
l
for
posit
ive
c
la
ss
259
data,
neg
at
ive
cl
ass
241
data,
a
nd
neu
t
ral
cl
ass
53
da
ta
.
The
n
s
plit
ti
ng
betwee
n
data
trai
ning
a
nd
da
ta
te
sti
ng
,
we
set
k
=
10
f
or
the
K
-
F
old
Cros
s
-
Vali
dation s
plit
ti
ng
m
et
ho
d.
In
Ta
ble
3,
t
he
resu
lt
s
betwee
n
F
-
Sc
ore
and
Accuracy
ob
ta
ined
us
i
ng
S
V
M
al
go
rithm
c
om
par
ed
to
us
in
g
ru
le
-
base
d
Se
nti
Wo
r
dN
et
is
q
uite
cl
ose
.
S
VM
al
gorit
hm
is
sli
gh
tl
y
bette
r
with
a
n
accuracy
of
76%
an
d
f
-
sc
or
e
51%
. R
ule
-
based Se
ntiWor
dn
et
gets a
ccur
acy
56% a
nd f
-
sc
or
e
48%
.
Table
3.
C
om
par
iso
n of Re
su
l
ts usin
g 1
0
-
Fo
l
d
Cr
os
s
V
al
ida
ti
on
Bef
ore Ba
la
ncing D
at
ase
t
Metho
d
Precisio
n
(
%)
R
ecall (
%
)
F1
-
Sco
re
(%)
Accurac
y
(
%)
SVM
4
8
.74
5
3
.23
5
0
.89
7
5
.75
Ru
le
-
b
ased
SentiWord
Net
4
9
.5
4
6
.42
4
7
.76
5
5
.81
But
that
can
be
seen,
t
her
e
is
a
con
si
der
a
ble
diff
e
re
nce
bet
ween
Acc
ur
ac
y
and
F
-
Sc
or
e
wh
e
n
us
in
g
SV
M al
gorith
m
,
w
it
h
a d
iffe
ren
ce
of
20
% c
an
be
sai
d
t
here i
s an
i
m
balance b
et
wee
n
the
classe
s p
rese
nt
in
the
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2502
-
4752
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci,
Vo
l.
13
, N
o.
3
,
Ma
rc
h 201
9
:
902
–
909
908
dataset
.
F
or
th
e
1
0th
rou
nd
K
-
Fo
l
d
Cr
os
s
-
V
al
idati
on
,
t
he
da
ta
te
sti
ng
f
or
the
neu
tral
cl
as
s
does
no
t
e
xis
t
at
al
l
because
all
of t
he 53 ne
utral c
la
ss d
at
a
has b
ecom
e trai
nin
g data,
see Fi
gur
e 4
.
Figure
4.
U
nde
rf
it
ti
ng
on
Neut
ral Cl
ass
Du
e
to
the
unde
rf
it
ti
ng
dataset
,
the
dataset
is
balanced
us
ing
balanci
ng
m
et
ho
d
as
sh
own
by
Figure
1.
During
t
he
pr
ocess
of
balanci
ng
t
he
dataset
,
the
a
m
ou
nt
of
ne
utr
al
cl
ass
data
in
creases
to
250
data
du
e
to
t
he
a
verage
of posit
ive
and n
e
gative
dat
a.
Table
4
s
how
s
that
the
resu
lt
s
of
S
VM
al
gori
thm
with
the
TF
-
I
DF
m
et
ho
d
as
featur
e
e
xtra
ct
ion
(
F
-
scor
e
83
%
a
nd
Accu
racy
89
%)
are
bette
r
than
the
re
su
lt
s
of
R
ule
-
base
d
Senti
W
ord
Ne
t
(
F
-
Sc
or
e
50
%
and
Accuracy
51%
).
T
he
re
su
lt
s
from
Table
3
and
Table
4
c
an
be
com
par
e
d,
the
bala
nce
d
dataset
s
get
bette
r
resu
lt
s
wh
e
n
usi
ng
SV
M
al
gorithm
with
TF
-
I
DF
a
s
feat
ure
ext
ractor
,
si
nce
it
increas
e
s
the
Ac
cu
racy
an
d
F
-
Score
because
the
ne
utral
cl
ass
ha
s
been
balance
d
;
how
ever,
the
Se
ntiWor
dN
et
ru
le
-
base
d
al
go
rith
m
ha
s
decr
ease
d bo
t
h i
n
Acc
ur
acy
a
nd
F
-
Sc
ore.
T
he
ex
pe
rim
ent
fo
un
d
the a
ver
a
ge
num
ber
of
word
s which
w
ere not
in
the
synsets
was
573
w
ords
.
T
her
e
fore,
t
he
r
ule
-
ba
sed
Se
nti
W
or
dN
et
c
onside
rin
g
th
os
e
m
issi
ng
573
s
ynset
s
can in
c
rease
th
e accu
racy t
o
a
bout
20%.
Table
4.
C
om
par
iso
n of Re
su
l
ts usin
g 1
0
-
Fo
l
d
Cr
os
s
V
al
ida
ti
on
a
fter Bal
a
ncin
g Data
set
Meth
od
Precisio
n
(
%)
Recall (%
)
F1
-
Sco
re
(%)
Accurac
y
(
%)
SVM
8
2
.02
8
5
.45
8
3
.69
8
9
.06
Ru
le
-
b
ased
SentiWord
Net
5
1
.34
4
9
.65
5
0
.45
5
1
.59
4.
CONCL
US
I
O
N
Ba
sed on t
he r
esults o
f
t
he
cl
assifi
cat
ion
us
i
ng S
VM and
r
ule
-
based,
we c
an
c
on
cl
ud
e
t
hat:
1.
Ba
la
ncing
data
set
s
can
im
pr
ove
both
Acc
uracy
and
F
-
Sc
ore
achieve
d
by
SV
M
al
gorith
m
with
TF
-
IDF
a
s
featur
e
e
xtracti
on
m
et
ho
d
;
ho
wev
e
r
bala
ncing
dataset
s
can
dec
rease
bot
h
Acc
ur
acy
a
nd
F
-
Score
res
ul
te
d
by
the
r
uled
-
ba
sed
S
enti
Wo
r
dNet
.
2.
SV
M
al
gorith
m
with
TF
-
I
D
F
as
featur
e
e
xt
racti
on
m
et
ho
d
achie
ves
bett
er
res
ults
than
tho
se
re
su
lt
ed
by
the
r
ule
-
base
d Senti
Wo
r
dNet
.
3.
Ther
e
a
re
sti
ll
m
any
wo
r
ds
that
do
not
have
synset
because
Indonesia
n
vo
ca
bula
ry
is
sti
ll
incom
plete
.
U
sin
g
Se
nti
WordNet a
nd transl
at
or
to
ols a
re
sti
ll
n
ot go
od
e
nough
f
or tra
nsl
at
ing
In
donesi
an
in
t
o
E
ngli
sh
.
ACKN
OWLE
DGE
MENTS
The
aut
hors
w
ou
l
d
li
ke
to
thank
to
I
ns
ti
tut
Teknolo
gi
Sepu
l
uh
Nopem
ber,
Direkt
ora
t
Rise
t
da
n
Pen
gabdia
n
Masy
ar
ak
at,
Direkt
orat
Jend
e
ra
l
Pen
gu
atan
Rise
t
dan
Pen
ge
mba
ngan
,
the
Mi
ni
stry
of
Re
search
, Tec
hnology, a
nd
Higher
Educat
i
on of
Ind
on
e
sia
for
fina
ncin
g t
he
resea
rc
h.
REFERE
NCE
S
[1]
B.
Pang
and
L.
Le
e
.
Opinion
Mi
ning
and
Senti
m
ent
Anal
y
sis
.
Fo
und.
Tr
ends®
InformatioP
ang,
B
.
,
Lee
,
L.
(
2006)
.
Opin.
Min
.
S
ent
i
m.
Anal.
Found.
Tr
ends®
Inf.
R
etr
ie
val,
1(
2)
,
91
–
231.
doi10
.
1561/
1500000001n
Retr.
,
vo
l.
1,
no.
2
,
pp.
91
–
231
,
200
6.
[2]
M.
R.
Islam
.
N
um
eri
c
rating
of
Apps
on
Google
Pla
y
Store
b
y
senti
m
ent
an
aly
s
is
on
user
rev
i
e
ws
.
1st
Int.
Con
f.
El
e
ct
r.
Eng. Inf. Com
mun.
Techn
ol.
IC
EE
ICT 201
4
,
pp
.
1
–
4
,
2014
.
[3]
E.
Guzm
an
and
W
.
Maa
le
j
.
How
do
users
li
ke
this
fea
ture
?
A
fine
gra
ine
d
sen
ti
m
e
nt
anal
y
sis
of
A
pp
rev
ie
ws
.
201
4
IEE
E
22nd
In
t.
Re
quir.
Eng
.
Co
nf.
RE 2014
-
Pro
c.
,
pp.
153
–
162,
2014.
[4]
A.
R.
Nara
dh
ip
a
and
A.
Purw
ariant
i
.
Senti
m
ent
Cla
ss
ifi
c
at
ion
fo
r
Indone
sian
Me
ss
age
s
in
Socia
l
Media
.
In
t.
Conf
.
El
e
ct
r.
Eng. Inf
o
rm
ati
cs
,
no
.
Jul
y,
pp
.
2
–
5
,
2011
.
[5]
D.
A
y
u
and
K.
Khotimah
.
Sentim
ent
Detect
ion
of
Com
m
ent
Ti
t
le
s
in
Booking
.
com
Us
ing
Probabil
ist
ic
L
aten
t
Sem
ant
ic
Ana
l
y
s
is
.
2018
6
th
In
t.
Conf.
In
f. Commun.
Techno
l.
,
vo
l
.
0
,
no
.
c
,
pp
.
51
4
–
519,
2018
.
Fo
ld
10
Ju
m
lah
dat
a tr
a
in
in
g
po
sitif
:
217
Ju
m
lah
dat
a tr
a
in
in
g
neg
atif
:
228
Ju
m
lah
dat
a tr
a
in
in
g
neu
tral
:
53
Evaluation Warning : The document was created with Spire.PDF for Python.
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci
IS
S
N:
25
02
-
4752
A com
pa
r
ative
stud
y
o
f se
ntim
ent an
alysis u
s
ing
SVM
an
d S
entiW
ordNet
...
(
Riy
anar
to
Sa
rn
o
)
909
[6]
N.
Farra
,
E
.
Cha
ll
ita,
R
.
A.
As
si,
and
H.
Hajj
.
Se
nte
nc
e
-
le
v
el
and
document
-
le
v
el
senti
m
ent
m
ini
n
g
for
ara
bic
te
xt
s
.
Proc.
-
IE
EE Int
.
Conf.
Data
Min
i
ng
,
ICDM
,
pp.
1
114
–
1119,
2010
.
[7]
K.
Dene
cke
.
Usi
ng
Senti
W
ordNet
for
m
ult
il
ing
ual
senti
m
ent
a
naly
s
is
.
Proc.
-
In
t.
Conf.
Data
E
ng.
,
pp.
507
–
51
2,
2008.
[8]
A.
Sun,
E.
P.
Li
m
,
and
Y.
Li
u
.
On
strat
egi
es
fo
r
imbala
nce
d
t
e
xt
cl
assifi
ca
t
ion
using
SVM:
A
c
om
par
at
ive
stu
d
y
.
Dec
is.
Supp
ort S
yst.
,
vol
.
48
,
no
.
1,
pp
.
191
–
201
,
2009.
[9]
W
.
McKinne
y
.
Data
Struct
ur
es
for
Stat
isti
ca
l
C
om
puti
ng
in
P
y
t
hon,
”
Proc
.
9th
Py
thon
Sc
i.
Conf
.
,
vol
.
1697900,
no.
Scip
y
,
pp
.
51
–
56
,
2010
.
[10]
F.
Pedre
gosa
et
al.
Sci
kit
-
le
arn
:
Mac
hine
Learni
ng
in
P
y
thon
G
aë
l
Varoqua
ux
.
J.
Ma
ch.
Learn.
Re
s.
,
vol
.
12
,
p
p.
2825
–
2830,
201
1.
[11]
S.
Bird, E.
Kl
ei
n
,
and
E
.
Lope
r
,
Natural
languag
e
proce
ss
ing
wit
h
Python
.
2009.
[12]
B.
S.
Rinty
arn
a
and
R.
Sarno
.
Adapte
d
weigh
te
d
gra
ph
for
W
or
d
Sense
Disam
b
igua
t
ion
.
2016
4th
Int.
Conf.
Inf
.
Comm
un.
Techn
ol.
ICoICT 2016
,
vol
.
4
,
no
.
c, 20
16.
[13]
H.
C.
W
u,
R.
W
.
P.
Luk,
K
.
F.
W
ong,
and
K
.
L
.
Kw
ok
.
Int
e
rpre
ti
ng
TF
-
IDF
te
rm
weight
s
a
s
m
aki
ng
rel
ev
a
nce
dec
isions
.
ACM
Tr
ans.
Inf.
S
yst.
,
vol.
26
,
no
.
3
,
pp
.
1
–
37
,
2008
.
[14]
B.
Y.
Prata
m
a
and
R.
Sarno
.
Personali
t
y
class
ifica
t
ion
base
d
on
Twit
ter
te
xt
usin
g
Naive
Ba
y
es,
KN
N
and
SV
M
.
Proc.
2015
Int
.
Conf.
Data
Softw
.
Eng
.
ICOD
S
E
2015
,
pp.
170
–
174,
2016
.
[15]
F.
H.
Rac
hm
an,
R.
Sarno,
and
C
.
Fati
ch
ah
.
Mus
i
c
emotion
class
ifi
cation
base
d
o
n
l
y
r
ic
s
-
audi
o
us
ing
cor
pus
base
d
emotion
.
In
t. J. Elect
r.
Comput.
Eng.
,
vol
.
8
,
no
.
3,
pp
.
1720
–
173
0,
2018
.
[16]
M.
Jupri
and
R.
Sarno
.
Ta
xp
a
y
er
complia
nc
e
class
ifi
cation
using
C4.
5,
SV
M,
KNN,
Naive
Ba
y
es
and
MLP
.
in
201
8
Inte
rnational
Co
nfe
renc
e
on
Info
rm
ati
on
and
Co
mm
unic
ati
ons Tec
hnolog
y
(
ICOIACT)
,
2018,
pp.
297
–
303.
BIOGR
AP
HI
ES OF
A
UTH
ORS
Moh
am
m
ad
Fikr
i
is
now
f
ourth
ye
ar
stu
den
t
of
Inf
orm
at
ic
s
Dep
art
m
ent
at
I
ns
ti
tut
Tek
no
l
og
i
Sepulu
h
Nopem
ber
.
His
c
urren
t
interest
s
a
re
in
Te
xt
Re
trie
val
an
d
Im
age
Re
trie
val.
E
-
m
ail: fikr
i.m
oh
am
m
ad1
5@
m
hs
.if.its.ac.i
d
Ri
ya
nar
to
Sa
r
no
receive
d
M
.Sc
an
d
P
h.D
in
Com
pu
te
r
S
ci
ence
from
the
Un
i
ver
sit
y
of
Brunswic
k
Ca
nad
a
in
1988
a
nd
19
92.
I
n
20
03
he
was
pr
om
oted
to
a
F
ull
Pr
ofess
or.
His
te
aching
a
nd
r
esearch
i
nteres
ts
includes
In
t
ern
et
of
T
hing
s,
Process
Aware
I
nfor
m
at
ion
Sys
tem
s,
In
te
ll
igent Syste
m
s an
d B
us
i
ness P
ro
ces
s Mana
ge
m
ent.
E
-
m
ail: riy
anart
o@
if.it
s.ac
.id
Evaluation Warning : The document was created with Spire.PDF for Python.