Indonesian
J
our
nal
of
Electrical
Engineering
and
Computer
Science
V
ol.
22,
No.
3,
June
2021,
pp.
1731
1738
ISSN:
2502-4752,
DOI:
10.11591/ijeecs.v22i3.pp1731-1738
r
1731
A
new
appr
oach
f
or
extracting
and
scoring
aspect
using
SentiW
ordNet
T
uan
Anh
T
ran,
J
arunee
Duangsuwan,
W
iphada
W
ettayaprasit
Artificial
Intelligence
Research
Lab,
Department
of
Computer
Science,
Di
vision
of
Computational
Scienc
e,
F
aculty
of
Science,
Prince
of
Songkla
Uni
v
ersity
,
Songkhla,
Thailand
Article
Inf
o
Article
history:
Recei
v
ed
Feb
10,
2021
Re
vised
Mar
17,
2021
Accepted
Mar
20,
2021
K
eyw
ords:
Aspect
e
xtraction
Aspect
scoring
Score
le
v
el
SentiW
ordNet
ABSTRA
CT
Aspect-based
online
information
on
social
media
plays
a
vital
role
in
influencing
people’
s
opinions
when
consumers
concern
with
their
decisions
to
mak
e
a
purchase,
or
companies
intend
to
pursue
opinions
on
their
product
or
ser
vices.
Determining
aspect-based
opinions
from
the
online
information
is
necessary
for
b
usiness
intelligence
to
support
users
in
reaching
their
objecti
v
es.
In
this
study
,
we
propose
the
ne
w
aspect
e
xtraction
and
scoring
system
which
has
three
procedures.
The
first
procedure
is
normalizing
and
tagging
part-of-speech
for
sentences
of
datasets.
The
second
procedure
is
e
xtracting
aspects
with
pattern
rules.
The
third
procedure
is
assigning
scores
for
aspects
with
Sent
iW
ordNet.
In
the
e
xperiments,
benchmark
datasets
of
customer
re
vi
e
ws
are
used
for
e
v
aluation.
The
performance
e
v
aluation
of
our
proposed
system
sho
ws
that
our
proposed
system
has
high
accurac
y
when
compared
to
other
systems.
This
is
an
open
access
article
under
the
CC
BY
-SA
license
.
Corresponding
A
uthor:
W
iphada
W
ettayaprasit
Department
of
Computer
Science,
Di
vision
of
Computational
Science
F
aculty
of
Science,
Prince
of
Songkla
Uni
v
ersity
,
Songkhla,
Thailand
Email:
wiphada.w@psu.ac.th
1.
INTR
ODUCTION
No
w
adays,
the
digital
era
af
fects
humans’
beha
viors
in
choosing
reference
resources
to
decide
t
heir
decisions.
The
online
information
usually
compos
es
of
opini
o
ns
or
feeli
ngs
e
xpressed
b
y
the
Internet
users
about
services,
healthcare,
products,
politics,
etc.
Determining
and
understanding
the
Internet
users’
opinions
(e.g.,
happ
y
or
unhapp
y)
using
sentiment
analysis
is
the
vital
k
e
y-role
to
apply
to
mark
eting,
and
making
decisions
or
recommendations
[1–3].
In
te
xtual
online
information,
the
users
usually
mention
about
opinions
or
feelings.
These
attr
ib
utes
are
called
aspects,
and
the
phase
to
e
xtract
the
useful
aspects
from
the
online
information
is
called
aspect
e
xtraction
[4–8].
In
the
pre
vious
w
orks,
most
of
these
studies
e
xtracted
aspects
from
customers’
re
vie
ws
and
did
not
sho
w
ho
w
much
satisfied
or
dissatisfied
the
Internet
users
mentioned
in
re
vie
ws
for
the
aspects.
T
o
determine
and
understand
ho
w
much
satisfied
or
dissatisfied
the
Internet
users
mention
for
aspects
i
s
useful
to
mak
e
decisions.
In
this
study
,
we
propose
aspect
e
xtraction
and
scoring
system
(AESS)
to
e
xtract
and
score
aspects
which
become
the
kno
wledgebase.
Datasets
from
independent
domains
(e.g.,
services,
products,
etc.)
are
the
input
of
the
AESS.
The
pre-processing
phase
is
normalizing
and
tagging
part-of-speech
(POS).
The
AESS
uses
pattern
rules
to
e
xtract
aspects
from
datasets.
SentiW
ordNet
is
used
to
assign
score
le
v
els
for
aspects.
The
output
is
the
scored
aspect
kno
wledgebase
which
sho
ws
satisfied
le
v
els
of
the
users
as
well.
The
rest
of
the
paper
is
or
g
anized
as
the
follo
wing:
The
related
w
orks
are
presented
in
section
2.
The
J
ournal
homepage:
http://ijeecs.iaescor
e
.com
Evaluation Warning : The document was created with Spire.PDF for Python.
1732
r
ISSN:
2502-4752
architecture
of
the
proposed
AESS
system
is
discussed
in
section
3.
The
e
xperimental
results
and
e
v
aluation
are
e
xplained
in
section
4.
Finally
,
the
conclusion
is
gi
v
en
in
section
5.
2.
RELA
TED
W
ORK
T
o
e
xtract
aspect,
W
ei
et
al.
[9]
proposed
semantic-based
product
feature
e
xtraction
(SPE)
method
which
used
the
association
rule
mining
algorithm
to
e
xtract
aspects.
Qiu
et
al.
[10]
presented
a
double-propag
ation
(DP)
algorithm
which
used
dependenc
y
relations
among
constituencies
in
a
sentence
to
e
xtract
aspects.
Liu
et
al.
[4]
e
xtended
more
dependenc
y
relations
(DP
+
)
to
e
xtract
aspects.
Rana
and
Cheah
[11]
proposed
a
tw
o-fold
rules-based
model
(TF-RBM)
which
used
sequential
pattern
rules
to
e
xtract
aspects.
Mataoui
et
al.
[12]
introduced
a
method
for
the
Arabic
language
by
using
s
yntactic
rules
in
order
to
e
xtract
aspects.
Rana
and
Cheah
[13]
proposed
a
sequential
pattern
rules-based
approach
(SPR)
to
automatically
produce
sequential
pattern
rules
to
e
xtract
aspects.
Poria
et
al.
[14]
suggested
rules
and
dependenc
y
trees
to
e
xtract
aspects
(e
xplicit
and
implicit).
Meanwhil
e,
Alqaryouti
et
al.
[15]
used
rules
to
e
xtract
aspects
(e
xplicit
and
implicit)
from
go
v
ernment
re
vie
ws.
F
or
aspect
scoring,
Kherw
a
et
al.
[16]
assigned
a
s
core
for
an
aspect
by
calculating
an
a
v
erage
score
of
opinion
w
ords
from
SentiW
ordNet
where
these
opinion
w
ords
and
that
aspect
co-occurred.
Asghar
et
al.
[17]
chose
the
highest
score
in
three
scores
(positi
v
e,
ne
g
ati
v
e,
objecti
v
e)
of
an
opinion
w
ord
which
were
respecti
v
e
a
v
erage
scores
of
all
synsets
of
t
h
a
t
opinion
w
ord
from
SentiW
ordNet.
Xu
et
al.
[18]
used
frequenc
y
and
a
dictionary
to
calcul
ate
scores.
Jmal
and
F
aiz
[19]
calculated
a
score
by
using
the
popularity
of
a
frequenc
y
for
one
aspect
on
T
witter
and
scores
(ne
g
ati
v
e,
positi
v
e,
neutral)
from
SentiW
ordNet
of
w
ords
(v
erb/adjecti
v
e)
related
to
the
aspect.
The
frequenc
y
of
the
aspect
w
as
estimated
in
the
dataset.
Meanwhile,
Mahesw
ari
and
Dhenakaran
[20]
used
a
dictionary
for
opinion
w
ords
and
Fuzzy
rules.
3.
PR
OPOSED
METHODOLOGY
T
o
automatically
e
xtract
and
score
aspects
from
datasets,
the
AESS
is
proposed
and
illustrated
in
Figure
1.
The
AESS
system
has
three
procedures:
1)
pre-processing,
2)
aspect
e
xtraction
using
pattern
rules
and
W
ord2V
ec,
and
3)
aspec
t
scoring
using
SentiW
ordNet.
The
input
of
the
system
is
datasets
such
as
product
re
vie
ws.
The
output
of
the
system
is
the
scored
aspect
kno
wledgebase
which
can
be
represented
in
graphics.
Figure
1.
An
architecture
of
AESS
3.1.
Pr
e-pr
ocessing
This
procedure
aims
to
normalize
and
tag
POS
for
sentences
of
datasets.
The
details
are
1)
eli
minating
special
characters
in
the
te
xt
of
social
media,
e.g.,
HyperT
e
xt
markup
language
(HTML)
tags,
a
pair
of
quotations,
2)
correcting
misspelt
w
ords,
and
3)
tagging
POS
for
te
xt.
3.2.
Aspect
extraction
This
procedure
is
used
to
e
xtract
aspects
with
opinion
w
ords
and
intensifier
w
ords
from
datasets
using
pattern
rules.
There
are
tw
o
main
steps:
1)
aspect
candidates
e
xtraction,
and
2)
aspect
pruning.
Let
a
be
an
aspect,
ow
be
an
opinion
w
ord
in
the
(opinion
le
xicons)
OL,
and
iw
be
an
intensifier
w
ord.
Let
neg
be
a
ne
g
ation
status
which
sho
ws
a
ne
g
ation
w
ord
e
xisting
in
a
sentence
with
an
opinion
w
ord
where
neg
2
f
T
r
ue;
F
al
se
g
.
Definition
1:
Sentence
based
on
aspect-opinion-intensifier
(SA
OI)
is
a
set
which
members
ha
v
e
a
quadruple
<
a
,
ow
,
iw
,
neg
>
in
the
sentence
as
sho
wn
in
(1)
SA
OI
=
<
a
i
;
ow
i
;
iw
i
;
neg
i
>
(1)
where
i
is
an
inde
x
of
an
e
xtracted
aspect,
1
i
n,
n
is
the
number
of
e
xtracted
aspects.
Indonesian
J
Elec
Eng
&
Comp
Sci,
V
ol.
22,
No.
3,
June
2021
:
1731
–
1738
Evaluation Warning : The document was created with Spire.PDF for Python.
Indonesian
J
Elec
Eng
&
Comp
Sci
ISSN:
2502-4752
r
1733
-
Step
1.
aspect
candidates
e
xtraction.
This
step
will
e
xtract
aspect
candidates
from
datasets
by
using
the
pattern
rules
and
the
OL
dictionary
(Bing
Liu’
s
opinion
le
xicon
[21]
and
MPQA
’
s
opinion
le
xicon
[22]).
After
e
xtracting,
the
aspects,
opinion
w
ords,
and
intensifier
w
ords
are
sa
v
ed
in
SA
OI.
The
pattern
rules
are
determined
by
using
the
relationship
between
aspect
and
opinion
w
ords.
The
relationships
based
on
a
syntactic
structure
are
determined
from
the
dependenc
y
tree
[23].
Some
e
xamples
of
the
patt
ern
rules
are
in
T
able
1.
There
are
opinion
w
ord(s)
in
italic,
aspect(s)
in
bold,
co-reference
w
ord(s)
in
italic
bold,
optional
w
ords
in
brack
ets,
and
a
subscript
sho
wing
positions
for
a
constituent
in
a
sentence
(e.g.,
“a”,
“b”,
etc.).
-
Step
2.
aspect
pruning.
Thi
s
step
eliminates
the
irrele
v
ant
aspects
by
using
the
cosine
similarity
and
W
ord2V
ec
(W
ord2V
ec
is
pro
vided
by
SpaCy
[24]).
T
able
1.
Some
e
xamples
of
pattern
rules
for
aspect
e
xtraction
P
attern
No.
Syntax-based
P
attern
Rule
P
attern
No.
Syntax-based
P
attern
Rule
S1
AP
+
CN
S6
CN
+
RCl
+
V2A
+
AP
(Note:
RCl
is
an
y
pattern)
S2
CN
+
RCl
S7
CN
a
+
V
+
(Prep)
+
CN
b
+
V2A
+
AP
Note:
Prep
is
“by”;
V
is
V+ed
/
V+ing
S3
V2A
+
(Adv)
+
A2
+
NP
S8
Pron
1
+
V/V2A
+
CN
+
(Adv)
+
Conj
+
Pron
2
+
V2A
+
AP
Note:
Pron
2
is
a
co-reference
of
CN
S4
CN
+
V2A
+
AP
S9
(Adv)
+
V2
+
NP
S5
CN
a
+
V2A
+
CN
b
S10
CN
+
V2A
+
V
(Note:
V
is
V+ed
/
V+ing)
3.3.
Aspect
scoring
A
goal
of
this
procedure
is
to
score
aspects
by
using
SentiW
ordNet
(SentiW
ordNet
which
is
a
l
e
xi
cal
resource
is
automatically
annotated
“positi
vity”
and
“ne
g
ati
vity”
scores
for
all
of
synsets
[25]).
Definition
2:
Opinion
v
alue
of
an
opinion
w
ord
(
O
V
)
is
an
a
v
erage
of
all
synsets
v
alues
for
an
opinion
w
ord
(
ow
)
which
are
retrie
v
ed
from
SentiW
ordNet
as
sho
wn
in
(2)
O
V
=
(
P
p
i
=1
P
V
i
=
p
,
if
ow
2
OLP
P
p
i
=1
N
V
i
=
p
,
if
ow
2
OLN
(2)
where
p
is
a
number
of
entries
(synsets)
for
ow
in
Senti
W
ordNet,
P
V
i
is
the
i
th
positi
v
e
v
alue,
N
V
i
is
the
i
th
ne
g
ati
v
e
v
alue,
OLP
is
a
set
of
Opinion
Le
xicons
in
Positi
v
e
(e.g.,
“good”,
“great”,
etc.),
and
OLN
is
a
set
of
Opinion
Le
xicons
in
Ne
g
ati
v
e
(e.g.,“bad”,
“hate”,
etc.)
(OL
=
OLP
[
OLN).
Definition
3:
Sentence
polarity
(
S
P
ol
)
is
a
v
alue
which
is
aggre
g
ated
from
a
ne
g
ation
status
neg
and
an
opinion
w
ord
(
ow
)
in
a
sentence
as
sho
wn
in
(3)
S
P
ol
=
(
+1
,
if
pol
ar
ity
ow
L
neg
=
T
r
ue
1
,
if
pol
ar
ity
ow
L
neg
=
F
al
se
(3)
where
neg
is
a
ne
g
ation
w
ord
e
xists
in
a
sentence
or
not,
and
pol
ar
ity
ow
is
a
polarity
of
an
opinion
w
ord
(
pol
ar
ity
ow
is
equal
to
T
r
ue
if
an
opinion
w
ord
(
ow
)
is
positi
v
e.
pol
ar
ity
ow
is
equal
to
F
al
se
if
an
opinion
w
ord
(
ow
)
is
ne
g
ati
v
e).
F
or
e
xample,
from
the
sentence
“
A
picture
is
not
beautiful”,
an
opinion
w
ord
“beautiful”
is
positi
v
e.
Polarity
of
“beautiful”
is
determined
T
rue.
A
ne
g
ation
w
ord
is
“not”.
neg
for
“not”
is
T
rue.
W
ith
pol
ar
ity
ow
=
T
r
ue
and
neg
=
T
r
ue
,
pol
ar
ity
ow
L
neg
and
S
P
ol
equal
to
F
alse
and
-1,
respecti
v
ely
.
Let
I
V
iw
be
an
Intensifier
V
alue
of
an
intensifier
w
ord
(
iw
).
I
V
iw
is
pre-defined
by
users
in
T
able
2
and
has
the
v
alue
in
[-1,
1].
Definition
4:
SA
OI
score
for
an
aspect
(
S
S
cor
e
a
)
is
a
v
alue
which
is
aggre
g
ated
from
v
alues
of
an
opinion
w
ord,
an
intensifier
w
ord,
and
ne
g
ation
e
xpressed
by
users’
opinions
for
aspect
a
in
one
sentence
as
sho
wn
in
(4).
S
S
cor
e
a
=
S
P
ol
(
I
V
iw
O
V
)
+
O
V
(4)
F
or
e
xample,
SA
OI
Score
for
an
aspect
S
S
cor
e
a
for
a
quadruple
(“speed”,
“good”,
“so”,
F
alse)
in
T
able
3
(
i
=
1)
from
the
sentence
“The
speed
is
so
good”
is
calculated
with
F
ormula
(4)
as
follo
ws:
“good”
is
positi
v
e
opinion
(i.e.
pol
ar
ity
ow
=
T
r
ue
).
S
P
ol
=
+1
because
neg
=
F
al
s
e
and
pol
ar
ity
ow
=
T
r
ue
.
A
ne
w
appr
oac
h
for
e
xtr
acting
and
scoring
aspect
using
SentiW
or
dNet
(T
uan
Anh
T
r
an)
Evaluation Warning : The document was created with Spire.PDF for Python.
1734
r
ISSN:
2502-4752
Intensifier
w
ord
“so”
has
intensifier
v
alue
0.45
(i.e.
I
V
iw
=
0
:
45
).
O
V
for
“good”
is
an
a
v
erage
score
which
is
retrie
v
ed
from
SentiW
ordNet
and
is
equal
to
0.70.
Hence,
S
S
cor
e
a
=
(+1)
x
[(0.45
x
0.70)
+
0.70]
=
1.02.
The
e
xample
of
S
S
cor
e
a
for
aspects
are
sho
wn
in
T
able
3.
T
able
2.
Intensifier
v
alues
(
I
V
iw
)
for
intensifier
w
ords
(
iw
)
Intensifier
w
ord(s)
Intensifier
V
alue
Intensifier
w
ord(s)
Intensifier
V
alue
(
iw
)
(
I
V
iw
)
(
iw
)
(
I
V
iw
)
a
wfully
,
critically
-1.00
altogether
,
so
0.45
dangerously
,
dreadfully
,
hopelessly
-0.70
primarily
,
v
ery
0.50
bitterly
,
horribly
,
strikingly
-0.50
highly
0.55
terribly
,
violently
-0.50
lar
gely
,
reasonably
0.60
suspiciously
,
slightly
-0.40
greatly
0.65
some
what
-0.25
hugely
,
surprisingly
,
totally
,
utterly
0.70
mildly
,
quite
-0.20
fully
,
mainly
,
deeply
0.70
f
aintly
0.10
especially
,
particularly
,
predominantly
0.75
really
,
purely
0.15
amazingly
,
e
xceedingly
,
e
xtremely
0.80
remarkably
,
nearly
,
partly
0.20
incredibly
,
seriously
,
unbelie
v
ably
0.80
pretty
,
rather
,
roughly
0.20
w
onderfully
,
e
xclusi
v
ely
0.80
simply
0.25
entirely
,
almost,
mostly
0.90
f
airly
,
moderately
0.30
absolutely
,
completely
,
perfectly
1.00
T
able
3.
Examples
of
S
S
cor
e
a
and
score
le
v
el
for
aspects
SA
OI
S
P
o
l
I
V
iw
O
V
S
S
cor
e
a
score
le
v
el
i
a
i
ow
i
iw
i
neg
i
number
name
1
speed
good
so
F
alse
+1
0.45
0.70
1.02
+2
v
ery
satisfied
2
battery
good
“”
T
rue
-1
0
0.70
-0.70
-2
v
ery
dissatisfied
3
battery
lo
v
ed
“”
F
alse
+1
0
0.71
0.71
+2
v
ery
satisfied
4
speed
bad
“”
T
rue
+1
0
0.66
0.66
+1
satisfied
5
battery
cool
quite
F
alse
+1
-0.20
0.29
0.23
+1
satisfied
6
battery
good
“”
F
alse
+1
0
0.70
0.70
+2
v
ery
satisfied
7
speed
bad
“”
F
alse
-1
0
0.66
0.66
+1
satisfied
Definition
5:
Score
le
v
el
is
a
pair
of
tw
o
data
(number
,
name)
in
which
“number”
is
an
inte
ger
number
in
[-3,
+3],
and
“name”
is
(“the
most
dissatisfied”,
“very
dissatisfied”,
“dissatisfied”,
“s
o
so”,
“satisfied”,
“very
satisfied”,
“the
most
satisfied”)
.
Relations
between
number
and
name
are
f
(-3,
“the
most
dissatisfied”),
(-2,
“v
ery
dissatisfied”),
(-1,
“dissatisfied”),
(0,
“so
so”),
(+1,
“satisfied”),
(+2,
“v
ery
satisfied”),
(+3,
“the
most
satisfied”)
g
.
Score
le
v
el
for
an
aspect
a
(
S
L
a
)
is
determined
by
using
S
S
cor
e
a
as
sho
wn
in
(5)
S
L
a
=
8
>
>
>
>
>
>
>
>
>
>
>
<
>
>
>
>
>
>
>
>
>
>
>
:
(-3,
“the
most
dissatisfied”)
,
if
SScor
e
a
<
=
1
:
4
(-2,
“v
ery
dissatisfied”)
,
if
SScor
e
a
2
(
1
:
4
;
0
:
7]
(-1,
“dissatisfied”)
,
if
SScor
e
a
2
(
0
:
7
;
0)
(0,
“so
so”)
,
if
SScor
e
a
=
0
(+1,
“satisfied”)
,
if
SScor
e
a
2
(0
;
0
:
7)
(+2,
“v
ery
satisfied”)
,
if
SScor
e
a
2
[0
:
7
;
1
:
4)
(+3,
“the
most
satisfied”)
,
if
SScor
e
a
>
=
1
:
4
(5)
note
that
minimum
and
maximum
scores
of
S
S
cor
e
a
are
-2
and
+2,
respecti
v
ely
.
F
or
e
xample,
S
S
cor
e
a
for
“speed”
in
the
pre
vious
e
xample
(
i
=
1
in
T
able
3)
equals
to
1.02
.
Score
le
v
el
for
“speed”
is
(+2,
“v
ery
satisfied”).
Score
le
v
els
for
all
of
aspects
are
sho
wn
in
the
last
tw
o
columns
of
T
able
3.
Definition
6:
Scored
aspect
kno
wledgebase
(Sakb)
is
a
set
which
members
ha
v
e
an
octuple
<
a
,
l
3
;
l
2
;
l
1
;
l
0
;
l
+1
;
l
+2
;
l
+3
>
as
sho
wn
in
(6)
Sakb
=
<
a
k
;
l
3
;
l
2
;
l
1
;
l
0
;
l
+1
;
l
+2
;
l
+3
>
(6)
where
k
is
an
inde
x
of
an
aspect
(none
redundant),
1
k
m,
m
is
the
number
of
none
redundant
aspects,
l
name
is
a
frequenc
y
of
score
le
v
el
for
aspect
a
k
.
Indonesian
J
Elec
Eng
&
Comp
Sci,
V
ol.
22,
No.
3,
June
2021
:
1731
–
1738
Evaluation Warning : The document was created with Spire.PDF for Python.
Indonesian
J
Elec
Eng
&
Comp
Sci
ISSN:
2502-4752
r
1735
Algorithm
1
e
xplains
aspect
e
xtraction
and
scoring
from
a
dataset.
Line
1
is
used
to
e
xtract
aspect
and
other
information
by
using
pattern
rules
and
sa
v
e
into
SA
OI.
Line
2
is
used
to
eliminate
irrele
v
ant
aspects
by
using
the
cosine
similarity
and
W
ord2V
ec.
Li
ne
3
is
used
to
initialize
Scored
Aspect
Kno
wledgebase
(Sakb).
In
lines
4-10,
the
algorithm
scores
aspects
in
SA
OI.
If
an
aspect
a
i
is
not
in
Sakb,
then
a
ne
w
aspect
a
i
is
added
to
Sakb
.
l
number
v
alues
are
equal
to
0
for
initialization.
S
S
cor
e
a
i
is
calculated
for
aspect
a
i
.
The
score
le
v
el
for
aspect
a
i
(
S
L
a
i
)
is
calculated
by
using
S
S
cor
e
a
i
in
(5).
A
frequenc
y
of
score
le
v
el
for
aspect
a
i
at
l
number
v
alue
is
increased
by
1.
On
line
11,
the
algorithm
returns
Sakb
.
Algorithm
1:
Aspect
e
xtraction
and
scoring
Input
:
Dataset
D,
P
attern
rules,
opinion
le
xicon
OL,
W
ord2V
ec,
SentiW
ordNet,
Intensifier
v
alues
list
Output
:
Scored
Aspect
Kno
wledgebase
Sakb
=
<
a
k
,
l
3
;
l
2
;
l
1
;
l
0
;
l
+1
;
l
+2
;
l
+3
>
1
SA
OI
e
xtract
aspect,
opinion
w
ord,
and
intensifier
from
D
using
pattern
rules
and
OL
2
SA
OI
eliminate
irrele
v
ant
aspects
from
SA
OI
using
the
cosine
similarity
and
W
ord2V
ec
3
Sakb
;
4
f
or
eac
h
aspect
a
i
in
SA
OI
do
5
if
a
i
is
not
in
Sakb
then
6
add
ne
w
a
i
to
Sakb
7
l
number
0
8
S
S
cor
e
a
i
S
P
ol
(
I
V
iw
i
O
V
)
+
O
V
9
calculate
score
le
v
el
for
a
i
(
S
L
a
i
)
by
using
S
S
cor
e
a
i
in
(5)
10
increment
the
l
number
of
aspect
a
i
by
one
11
r
etur
n
the
Scored
Aspect
Kno
wledgebase
(Sakb)
F
or
e
xample,
the
Aspect
Extr
action
and
Scoring
algorithm
has
been
applied
to
T
able
3.
The
result
has
tw
o
tuples.
T
w
o
aspects
of
the
result
are
speed
and
battery
.
The
Sakb
v
alues
of
l
number
for
aspect
speed
are
<
speed,
0,
0,
0,
0,
2,
1,
0
>
.
The
Sakb
v
alues
of
l
number
for
aspect
battery
are
<
battery
,
0,
1,
0,
0,
1,
2,
0
>
.
4.
RESUL
T
AND
DISCUSSION
In
this
study
,
we
used
tw
o
benchmark
datasets
to
conduct
our
e
xperiment.
The
first
dataset
[4]
has
three
re
vie
wed
domains
(computer
,
speak
er
,
and
router).
The
second
dataset
[21]
has
fi
v
e
re
vie
wed
domains
(Canon
camera,
MP3
player
,
Nokia
cellphone,
Nik
on
camera,and
D
VD
player).
Each
re
vie
wed
domain
is
described
with
the
format
re
vie
wed
domain
[total
o
f
sentences/
total
of
aspects]
as
the
follo
wing:
Computer
[531/
354],
Speak
er
[689/
440],
Router
[879/
307],
Canon
camera
[597/
237],
MP3
player
[1,716/
674],
Nokia
cellphone
[546/
302],
Nik
on
camera
[346/
174],
and
D
VD
player
[740/
296].
In
our
e
xperiment,
the
result
is
the
scored
aspect
kno
wledgebase
which
is
used
to
represent
with
graphical
charts.
In
addition,
we
compare
the
proposed
method
with
other
approaches
by
using
three
measures
(Precision,
Recall,
and
F1-score)
[13,
21].
The
formul
as
are
P
r
ecision
=
T
P
=
(
T
P
+
F
P
)
,
R
ecal
l
=
T
P
=
(
T
P
+
F
N
)
,
and
F1-scor
e
=
(2
P
R
)
=
(P
+
R),
where
T
P
is
j
E
\
A
j
,
F
P
is
j
E
n
A
j
,
and
F
N
is
j
A
n
E
j
.
Note
that
E
is
the
set
of
e
xtracted
aspects,
and
A
is
the
set
of
annotated
aspects
in
datasets.
Figure
2
sho
ws
comparisons
of
the
performance
e
xperimented
with
three
measures
(Precision,
Recall,
and
F1-score).
The
comparisons
are
semantic-based
product
feature
e
xtrac
tion
(SPE)
[9],
double
propag
ation
(DP)
[10],
DP
+
[4],
tw
o-f
o
l
d
rule-based
model
(TF-RBM)
[11],
sequential
pattern
rule
(SPR)
[13],
and
the
proposed
AESS.
From
Figure
2,
our
proposed
method
AESS
has
the
highest
precision
for
all
of
the
re
vie
wed
domains.
In
terms
of
F1-score,
the
proposed
method
sho
ws
the
highest
result
for
Computer
,
Speak
er
,
Canon
camera,
and
Mp3
player
with
the
v
alues
0.80,
0.74,
0.93,
and
0.83,
respecti
v
ely
.
Furthermore,
from
AESS
system
Figure
3
sho
ws
some
e
xamples
of
graphical
charts
for
Computer
re
vie
wed.
Figure
3a
sho
ws
all
aspects
score
with
so
so
80%,
satisfied
12%,
dissatisfied
7%,
and
the
most
dissatisfied
1%.
Figure
3b
sho
ws
“screen
quality”
aspect
score
with
satisfied
75%
and
dissatisfied
25%.
A
ne
w
appr
oac
h
for
e
xtr
acting
and
scoring
aspect
using
SentiW
or
dNet
(T
uan
Anh
T
r
an)
Evaluation Warning : The document was created with Spire.PDF for Python.
1736
r
ISSN:
2502-4752
(a)
(b)
(c)
(d)
(e)
(f)
(g)
(h)
Figure
2.
The
comparison
of
approaches
for
re
vie
wed
domains:
(a)
computer
,
(b)
speak
er
,
(c)
router
,
(d)
canon
camera,
(e)
mp3
player
,
(f)
nokia
cellphone,
(g)
nik
on
camera,
and
(h)
D
VD
player
(a)
(b)
Figure
3.
Graphical
charts
representing
for
computer
re
vie
wed
domain:
(a)
all
aspects
score
and
(b)
“screen
quality”
aspect
score
Indonesian
J
Elec
Eng
&
Comp
Sci,
V
ol.
22,
No.
3,
June
2021
:
1731
–
1738
Evaluation Warning : The document was created with Spire.PDF for Python.
Indonesian
J
Elec
Eng
&
Comp
Sci
ISSN:
2502-4752
r
1737
5.
CONCLUSION
Customer
satisf
action
or
dissatisf
action
feedback
is
really
important
for
b
usiness
intelligent
system
s.
W
e
propos
ed
the
ne
w
aspect
e
xtraction
and
scoring
system
(AESS)
to
represent
the
satisf
action
or
dissatisf
action
of
the
consumers
in
graphical
format.
The
input
of
the
AESS
is
the
te
xtual
online
data.
The
output
of
the
AESS
is
the
score
of
the
a
spect
kno
wledgebase.
The
aspect
kno
wledgebase
is
e
xtracted
by
using
pattern
rules
and
assigned
score
le
v
els
with
SentiW
ordNet.
From
the
benchmark
datasets,
the
proposed
AESS
has
a
v
ery
high
performance
when
compared
to
other
approaches.
The
proposed
AESS
could
be
applied
to
independent
domains
(e.g.,
services,
products,
etc.).
Moreo
v
er
,
the
proposed
AESS
does
not
need
an
y
annotated
data.
In
future
w
ork,
we
ha
v
e
a
plan
to
retrie
v
e
scores
from
dif
ferent
le
xical
resources.
A
CKNO
WLEDGEMENT
This
w
ork
w
as
supported
by
Thailand’
s
Education
Hub
for
the
Southern
Re
gion
of
ASEAN
Countries
(TEH-A
C)
and
PSU.GS.
Financial
Support
for
Thesis
(Fiscal
Y
ear:
2019).
REFERENCES
[1]
A.
A.
Jihad
and
A.
S.
Abdalkafor
,
“
A
frame
w
ork
for
sentiment
analysis
in
Arabic
te
xt,
”
Indonesian
J
ournal
of
Electrical
Engineering
and
Computer
Science
,
v
ol.
16,
no.
3,
pp.
1482–1489,
2019.
[2]
M.
A.
A
yu,
S.
S.
W
ijaya,
and
T
.
Mantoro,
“
An
automatic
le
xicon
generation
for
Indonesian
ne
ws
sentiment
analysis:
a
case
on
go
v
ernor
elect
ions
in
Indonesia,
”
Indonesian
J
ournal
of
Electrical
Engineering
and
Computer
Science
,
v
ol.
16,
no.
3,
pp.
1555–1561,
2019.
[3]
B.
Gliw
a,
A.
Zygmunt,
and
M.
Dabro
wski,
“Building
sentiment
le
xicons
based
on
recommending
services
for
the
Polish
language,
”
Computer
Science
,
v
ol.
17,
no.
2,
pp.
163–185,
2016.
[4]
Q.
Liu,
Z.
Gao,
B.
Liu,
and
Y
.
Zhang,
“
Automated
rule
selection
for
opinion
tar
get
e
xtrac
tion,
”
Knowledg
e-Based
Systems
,
v
ol.
104,
no.
15,
pp.
74–88,
2016.
[5]
L.
Bing,
Sentiment
Analysis:
Mining
Opinions,
Sentiments,
and
Emotions
.
Ne
w
Y
ork,
NY
10013-2473,
USA:
Cambridge
Uni
v
ersity
Press,
2015.
[6]
M.
Pontiki,
D.
Galanis,
H.
P
apageor
giou,
S.
Manandhar
,
and
I.
Androutsopoulos,
“SemEv
al-2015
task
12:
Aspect
based
sentiment
analysis,
”
in
Pr
oceedings
of
the
9th
International
W
orkshop
on
Semantic
E
valuation
.
Den
v
er
,
Colorado:
Association
for
Computational
Linguistics,
Jun.
2015,
pp.
486–495.
[7]
Q.
Liu,
Z.
Gao,
B.
Liu,
and
Y
.
Zhang,
“
Automated
rule
selection
for
aspect
e
xtraction
in
opinion
mining,
”
in
Pr
oceedings
of
the
24th
International
Confer
ence
on
Artificial
Intellig
ence
,
2015,
pp.
1291—-1297.
[8]
M.
Pontiki,
D.
Galanis,
J.
P
a
vlopoulos,
H.
P
apageor
giou,
I.
Androutsopoulos,
and
S.
Manandhar
,
“SemEv
al-2014
task
4:
Aspect
based
sentiment
analysis,
”
in
Pr
oceedings
of
the
8th
International
W
orkshop
on
Semantic
Evaluation
.
Dublin,
Ireland:
Association
for
Computational
Linguistics,
Aug.
2014,
pp.
27–35.
[9]
C.-P
.
W
ei,
Y
.-M.
Chen,
C.-S.
Y
ang,
and
C.
C.
Y
ang,
“Under
standing
what
concerns
consumers:
a
semantic
approach
to
product
feature
e
xtraction
from
consumer
re
vie
ws,
”
Information
Systems
and
e-Business
Mana
g
ement
,
v
ol.
8,
no.
2,
pp.
149–167,
2010.
[10]
G.
Qiu,
B.
Liu,
J.
Bu,
and
C.
Chen,
“Opinion
w
ord
e
xpansion
and
tar
get
e
xtraction
through
double
propag
ation,
”
Computational
Linguistics
,
v
ol.
37,
no.
1,
pp.
9–27,
2011.
[11]
T
.
A.
Rana
and
Y
.-N.
Cheah,
“
A
tw
o-fold
rule-based
model
for
aspect
e
xtraction,
”
Expert
Systems
with
Applications
,
v
ol.
89,
no.
15,
pp.
273–285,
2017.
[12]
M.
Mataoui,
T
.
E.
B.
Hacine,
I.
T
ellache,
A.
Bakhtouchi,
and
O.
Zelmati,
“
A
ne
w
syntax-based
aspect
detection
approach
for
sentiment
analysis
in
Arabic
re
vie
ws,
”
in
Pr
oceedings
of
the
2nd
International
Confer
ence
on
Natur
al
Langua
g
e
and
Speec
h
Pr
ocessing
(ICNLSP)
,
2018,
pp.
1–6.
[13]
T
.
A.
Rana
and
Y
.-N.
Cheah,
“Sequential
patterns
rule-based
approach
for
opinion
tar
get
e
xtraction
from
customer
re
vie
ws,
”
J
ournal
of
Information
Science
,
v
ol.
45,
no.
5,
pp.
643–655,
2019.
[14]
S.
Poria,
E.
Cambria,
L.-W
.
K
u,
C.
Gui,
and
A.
Gelb
ukh,
“
A
rule-based
approach
to
aspect
e
xtraction
from
product
re
vie
ws,
”
in
Pr
oceedings
of
the
2nd
W
orkshop
on
Natur
al
Langua
g
e
Pr
ocessing
for
Social
Media
(SocialNLP)
,
2014,
pp.
28–37.
[15]
O.
Alqaryouti,
N.
Siyam,
A.
A.
Monem,
and
K.
Shaalan,
“
Aspect-based
s
entiment
analysis
using
smart
go
v
ernment
re
vie
w
data,
”
Applied
Computing
and
Informatics
,
2020.
[16]
P
.
Kherw
a,
A.
Sachde
v
a,
D.
Mahajan,
N.
P
ande,
and
P
.
K.
Singh,
“
An
approach
to
w
ards
comprehensi
v
e
sentimental
data
analysis
and
opinion
mining,
”
in
2014
IEEE
International
Advance
Computing
Confer
ence
(IA
CC)
,
2014,
pp.
606–612.
A
ne
w
appr
oac
h
for
e
xtr
acting
and
scoring
aspect
using
SentiW
or
dNet
(T
uan
Anh
T
r
an)
Evaluation Warning : The document was created with Spire.PDF for Python.
1738
r
ISSN:
2502-4752
[17]
M.
Z.
Asghar
,
A.
Khan,
S.
R.
Zahra,
S.
Ahmad,
and
F
.
M.
K
undi,
“
Aspect-based
opinion
mining
frame
w
ork
using
heuristic
patterns,
”
Cluster
Computing
,
v
ol.
22,
pp.
7181–7199,
2017.
[18]
X.
Xu,
T
.
Meng,
and
X.
Cheng,
“
As
pect-based
e
xtracti
v
e
summarization
of
online
re
vie
ws,
”
pp.
968–975,
2011.
[19]
J.
Jmal
and
R.
F
aiz,
“Custom
er
re
vie
w
summarization
approach
using
T
witter
and
SentiW
ordNet,
”
in
Pr
oceedings
of
the
3r
d
International
Confer
ence
on
W
eb
Intellig
ence
,
Mining
and
Semantics
,
USA,
2013.
[20]
S.
U.
Mahesw
ari
and
S.
S.
Dhenakaran,
“
Aspect
based
fuzzy
logic
sentiment
analysis
on
social
media
big
data,
”
in
2020
International
Confer
ence
on
Communication
and
Signal
Pr
ocessing
(ICCSP)
,
2020,
pp.
0971–0975.
[21]
M.
Hu
and
B.
Liu,
“Mining
and
summarizing
customer
re
vie
ws,
”
in
Pr
oceedings
of
the
10th
International
Confer
ence
on
Knowledg
e
Disco
very
and
Data
Mining
(SIGKDD)
.
USA:
A
CM,
2004,
pp.
168–177.
[22]
T
.
W
ilson,
J.
W
iebe,
and
P
.
Hof
fma
nn,
“Recognizing
conte
xtual
polarity
in
phrase-le
v
el
sentiment
analysis,
”
pp.
347—-354,
2005.
[23]
A.
Ranta,
Gr
ammatical
F
r
ame
work:
Pr
o
gr
amming
with
Multilingual
Gr
ammar
s
.
CSLI
Publications,
Center
for
the
Study
of
Language
and
Information,
2011.
[24]
Spac
y
,
“Spac
y
guides,
”
2020.
[Onl
ine].
A
v
ailable:
https://spac
y
.io/
[25]
S.
Baccianella,
A.
Esuli,
and
F
.
Sebastiani,
“Sentiw
ordnet
3.0:
An
enhanced
le
xical
resource
for
sentiment
analysis
and
opinion
mining.
”
European
Language
Resources
Association,
2010.
BIOGRAPHIES
OF
A
UTHORS
T
uan
Anh
T
ran
,
i
s
a
PhD.
candidate
at
Department
of
Computer
Science,
Di
vision
of
Computational
Science,
F
aculty
of
Science,
Prince
of
Songkla
Uni
v
ersity
,
Thailand.
He
got
his
BSc.
in
Information
T
echnology
from
Hue
Uni
v
ersity
of
Education,
MSc.
i
n
Information
System
from
Ho
Chi
Minh
city
Uni
v
ersity
of
Sciences,
V
ietnam.
His
research
interest
is
Natural
Language
Processing.
J
arunee
Duangsuwan
,
is
an
assistant
professor
at
Department
of
Computer
Science,
Di
vision
of
Computational
Science,
F
aculty
of
Science,
Prince
of
Songkla
Uni
v
ersity
,
Thailand.
She
got
BSc.,
MSc.,
PhD.
in
Computer
Science
from
Chiang
Mai
Uni
v
ersity
,
Prince
of
Songkla
Uni
v
ersity
,
and
Uni
v
ersity
of
Reading,
UK,
respecti
v
ely
.
Her
research
interests
are
Natural
Language
Processing
and
Machine
Learning.
W
iphada
W
ettayaprasit
,
is
an
assistant
professor
at
Department
of
Computer
Science,
Di
vision
of
Computational
Science,
F
aculty
of
Science,
Prince
of
Songkla
Uni
v
ersity
,
Thailand.
She
got
BSc.,
MSc.,
PhD.
in
Computer
Science
from
Prince
of
Songkla
Uni
v
ersity
,
Uni
v
ersity
of
Missouri-Columbia,
USA,
and
Chulalongk
orn
Uni
v
ersity
,
respecti
v
ely
.
Her
research
interests
are
Artificial
Intelligence,
Neural
Netw
orks
and
Machine
Learning.
Indonesian
J
Elec
Eng
&
Comp
Sci,
V
ol.
22,
No.
3,
June
2021
:
1731
–
1738
Evaluation Warning : The document was created with Spire.PDF for Python.