IAES
Inter
national
J
our
nal
of
Articial
Intelligence
(IJ-AI)
V
ol.
13,
No.
4,
December
2024,
pp.
4688
∼
4694
ISSN:
2252-8938,
DOI:
10.11591/ijai.v13.i4.pp4688-4694
❒
4688
Indonesian
news
article
authorship
attrib
ution
multilabel
multiclass
classication
using
IndoBER
T
Kar
en
Etania
Saputra,
Riccosan
Department
of
Computer
Science,
School
of
Computer
Science,
Bina
Nusantara
Uni
v
ersity
,
Jakarta,
Indonesia
Article
Inf
o
Article
history:
Recei
v
ed
Dec
6,
2023
Re
vised
Mar
14,
2024
Accepted
Apr
17,
2024
K
eyw
ords:
Authorship
attrib
ution
Deep
learning
IndoBER
T
Indonesian
language
Multiclass
classication
Multilabel
classication
ABSTRA
CT
Recent
de
v
elopments
in
technology
ha
v
e
made
it
e
asier
to
produce
digital
con-
tent,
especially
te
xtual
articles.
But,
it
has
a
ne
g
ati
v
e
impact
in
the
form
of
a
rising
public
sk
epticism
of
digital
data
due
to
plagiaris
m.
Indonesia,
one
of
the
w
orld’
s
most
populous
countries,
is
not
resistant
to
this
problem.
T
o
resolv
e
it,
the
authorship
attrib
ution
(AA)
task
must
be
e
x
ecuted.
Ho
we
v
er
,
there
has
been
little
in
v
estig
ation
on
AA
for
Indonesian
articles.
As
a
result,
this
research
applies
the
AA
task
to
an
Indonesian
digital
ne
ws
articles
dataset.
Continuing
the
pre
vious
research,
dataset
modication
w
as
carried
out
to
increase
data
com-
ple
xity
by
adding
a
ne
w
class,
namely
the
author’
s
gender
,
and
a
lso
by
balancing
the
distrib
ution
of
data
v
ersus
labels
to
minimize
potential
o
v
ertting,
and
model
h
yper
-parameter
congurations
were
carried
out
to
enhance
the
results
g
ained.
This
research
successfully
applied
the
IndoBER
T
model
to
the
Indonesian
AA
task,
yielding
results
in
the
form
of
precision
=
0.92,
recall
=
0.90,
and
F1-score
=
0.91.
These
results
indicate
that
the
Indonesian
AA
task
has
a
lot
of
potential
for
de
v
elopment
since
it
identies
writing
patterns
that
may
benet
the
forensic
eld,
detect
plagiarism,
and
analyze
Indonesian
te
xts.
This
is
an
open
access
article
under
the
CC
BY
-SA
license
.
Corresponding
A
uthor:
Karen
Etania
Saputra
Department
of
Computer
Science,
School
of
Computer
Science,
Bina
Nusantara
Uni
v
ersity
Bandung
Campus,
Jakarta
11480,
Indonesia
Email:
karen.sapurta@binus.ac.id
1.
INTR
ODUCTION
T
echnological
adv
ancements
ha
v
e
made
it
easier
to
generate
content,
particularly
articles
and
informa-
tion,
with
a
rising
v
olume
and
f
aster
transmission
via
the
internet
as
a
channel
with
a
wide
range
of
topics
[1],
minimizing
the
need
for
ph
ysical
forms.
This
causes
the
issue
of
a
lack
of
trust
from
the
public
in
the
substance
of
digital
se
gments,
as
opposed
to
traditional
ne
ws
and
journalism,
which
has
tight
selection
and
re
gulation
processes
[2].
The
problem
e
xists
because
of
plagiarism,
an
act
of
duplicating
an
e
xisting
idea,
method,
results,
or
some
data
without
displaying
the
original
author
and
its
source
[3].
Until
no
w
,
plagiarism
has
become
a
root
problem
in
man
y
countries
as
technology
adv
ances
and
data
gets
easier
to
be
accessed.
Indonesia,
as
one
of
the
most
populated
countries,
also
f
ace
this
problem
and
resulting
in
the
creation
of
man
y
digital
articles
with
unkno
wn
authors
and
the
origins,
thus
ha
ving
the
ne
g
ati
v
e
impact
of
plagiarism,
such
as
deception
in
searching
o
w
to
get
genuine
information
and
spreading
misinformation
to
the
public
[4].
T
o
answer
the
problems,
article
classication
task
using
author
is
critical
to
implement,
especially
in
Indonesian
articles.
Article
classication
is
a
part
of
the
te
xt
analysis
task,
namely
authorship
attrib
ution
(AA),
which
has
gro
wn
into
an
interesting
research
topic
that
has
dra
wn
a
l
ot
of
dif
ferent
sectors,
including
media,
go
v
ernment,
economics,
and
education
[5].
AA
is
a
technique
for
identifying
stylometry
and
linguistic
patterns
J
ournal
homepage:
http://ijai.iaescor
e
.com
Evaluation Warning : The document was created with Spire.PDF for Python.
Int
J
Artif
Intell
ISSN:
2252-8938
❒
4689
in
documents
written
by
humans
[6].
AA
is
part
of
the
te
xt
analysis
task
since
it
seeks
to
process
v
ast
v
olumes
of
te
xtual
data
to
detect
plagiarism
and
nd
the
original
article’
s
author
.
The
AA
task
is
supported
by
the
author’
s
personality
traits
in
the
form
of
distinct
and
unique
linguistics
in
each
article
as
the
main
data,
which
is
an
important
f
actor
in
determining
the
originality
and
source
of
an
article
[7].
The
AA
task
in
this
research
is
inseparable
from
its
benets
of
dynamic
interdisciplinary
inte
gration
in
the
modern
digital
w
orld,
the
ability
to
obtain
important
and
in-depth
information,
and
its
wide
contrib
ution
to
v
arious
elds
(forensic
in
v
estig
ations,
digital
content
analysis,
and
plagiarism
detection)
[8].
There
are
barriers
to
the
AA
implementations,
including
anon
ymized
documents,
article
lengths,
art
i-
cles
with
numerous
authors
[8],
the
need
for
an
e
xcessi
v
e
quantity
of
data
[9],
articles
used
by
other
parties
that
e
xtend
t
he
search
o
w
[10],
and
the
quality
of
author
writing,
which
continues
to
impro
v
e
[11].
These
barriers
mak
e
tracing
the
origin
of
digital
articles
a
complicated
and
time-consuming
task.
T
o
answer
these
issues,
it
is
possible
to
increase
the
quality
of
the
dataset
used
in
the
research
by
increasing
or
lo
wering
the
amount
of
data,
balancing
data
distrib
ution,
and
determining
the
rele
v
ant
literary
components
for
the
classication
process.
This
research
emplo
ys
a
dataset
of
digital
articles
from
ne
ws
portal
websites
from
earlier
research
on
automatic
Indonesian
AA
using
transformers
[12].
F
or
greater
comple
xity
,
the
dataset
w
as
modied
by
including
a
sec-
ond
feature
or
class,
author
gender
.
Gender
elements
are
included
due
to
the
dif
ference
in
language
styles
used
by
men
and
w
omen,
with
male
authors
using
more
formal
la
n
guage
forms
and
female
authors
combining
the
formal,
informal,
and
more
f
amiliar
gurati
v
e
language
forms
[13].
After
modifyi
n
g
the
dataset,
this
research
implement
multi-label
multi-class
methods
for
classic
a-
tion.
Multi-label
classication
is
the
classication
of
data
that
has
more
than
one
sort
of
class
designated
for
it
[14],
which
in
this
case
are
the
author’
s
name
and
gender
.
Multi-class
classication
is
the
classication
of
data
utilizing
more
than
one
class
[15],
with
e
v
ery
point
of
data
only
corresponding
to
a
single
class
to
a
v
oid
o
v
erlapping.
Another
modication
to
the
dataset
w
as
a
reduction
in
author’
s
name
class
to
balance
the
distri-
b
ution
of
data
ag
ainst
the
total
classes.
This
research
achie
v
ed
a
higher
le
v
el
of
data
comple
xity
by
modifying
the
dataset,
allo
wing
for
impro
v
ed
outcomes
when
implementing
AA
tasks
on
Indonesian
ne
ws
articles.
In
this
research,
AA
tasks
for
authorship
classication
of
Indonesian
digital
ne
ws
items
are
i
mple-
mented
using
the
deep
learning
approach,
as
in
earlier
research
[12].
Se
v
eral
prior
research
utilized
deep
learning
methods
for
AA
tasks
and
pro
v
en
their
ef
fecti
v
eness.
i)
the
implementation
of
AA
with
con
v
olutional
neural
netw
ork
(CNN)
and
POS-Eliding
[16]
ef
fecti
v
ely
impro
v
ed
the
performance
of
the
AA
task,
resulting
in
better
generalization
of
te
xtual
information
of
research
articles.
ii)
AA
for
neural
te
xt
generation
[17]
in-
v
estig
ated
the
AA
problem
with
se
v
eral
deep
learning
models
and
sho
wed
the
ur
genc
y
that
machine-generated
te
xts
are
e
v
er
more
dif
cult
to
detec
t
by
humans
because
their
better
quality
creates
the
potential
for
fraud
in
the
form
of
misleading
information
generation,
demanding
the
de
v
elopment
of
AA
technologies.
iii)
the
de
v
elopment
of
the
bidirect
ional
encoder
representation
from
transformers
for
authorship
attrib
ution
(BertAA)
model
by
F
abien
et
al.
[18]
utilized
the
BER
T
base
model
[19]
and
ne-tuned
for
AA
tasks
with
data
in
the
form
of
Enron
emails,
blog
posts,
and
mo
vie
re
vie
ws,
achie
ving
state-of-the-art
(SO
T
A)
on
model
accurac
y
of
93%.
Ho
we
v
er
,
most
earlier
research
on
AA
with
deep
learning
approaches
made
use
of
English
datasets,
hence
there
is
a
shortage
of
local
research
on
AA
tasks
with
Indonesian
datasets.
As
a
result,
the
purpose
of
this
research
is
to
contrib
ute
to
the
eld
of
AA
tasks
by
applying
the
deep
learning
method
to
Indonesian
digital
articles
in
particular
.
This
research
utilizes
the
transformers
architecture
[20],
with
Indonesian
bidirectional
encoder
r
epre-
sentation
from
transformers
(IndoBER
T)
[21]
as
the
base
model.
IndoBER
T
is
a
specialized
model
for
natural
language
processing
(NLP)
tasks
based
on
the
bidirectional
encoder
representation
from
transformers
(BER
T)
model
[19].
Implementing
the
IndoBER
T
model
[21]
cannot
be
separated
from
its
ability
to
focus
on
Indonesian
datasets.
This
is
ef
fecti
v
ely
understood
by
an
in-depth
re
vie
w
of
pre
vious
research
that
conrms
IndoBER
T’
s
adv
antages
for
specic
Indonesian-language
datasets
[12],
[22],
[23].
As
an
outcome,
the
use
of
IndoBER
T
will
benet
this
research
emplo
ying
an
Indonesian
ne
ws
article
dataset.
Etania
et
al.
[12]
achie
v
ed
model
e
v
al-
uation
ndings
in
the
form
of
test
predict
accurac
y
of
0.70
(70%),
and
T
op-K
accurac
y
of
0.84
(84%).
Although
e
xceeding
the
accurac
y
v
alues
obtained,
the
model
remains
short
of
t
he
BertAA
model,
which
has
an
accurac
y
score
of
93%
[18].
The
pre
vious
research’
s
model
f
ailed
to
achie
v
e
its
objecti
v
e
due
to
an
o
v
ert
model,
which
happens
when
the
amount
of
data
is
not
suf
cient
for
the
number
of
classes.
On
that
basis,
this
research
reap-
plied
the
IndoBER
T
model
[21].
As
a
result,
changes
were
made
to
the
dataset
used
in
this
research,
which
included
adding
a
feature
and
balancing
the
data
distrib
ution
compared
to
the
classes.
This
research
e
xperiment
contrib
utes
by
sho
wing
that
AA
tasks
with
deep
learning
methods
for
Indonesian
ne
ws
article
author
ship
attrib
ution
multilabel
multiclass
...
(Kar
en
Etania
Saputr
a)
Evaluation Warning : The document was created with Spire.PDF for Python.
4690
❒
ISSN:
2252-8938
Indonesian-language
articles
yield
better
and
more
signicant
results;
the
implementation
of
AA
tasks
can
be
focused
for
Indonesian-language
digital
ne
ws
articles;
and
the
application
of
multi-label
multi-class
meth-
ods
in
AA
tasks
for
Indonesian-language
ne
ws
articles
enhances
the
performance
of
the
IndoBER
T
model.
The
research
se
gment
is
separated
into
four
parts
.
Section
1,
introduction:
contains
the
problem
statement,
an
o
v
ervie
w
of
earlier
research,
the
source
of
the
dataset
utilized,
and
the
r
esearch’
s
objecti
v
es.
Section
2,
method:
co
v
ers
the
research
o
w
used
in
this
article.
Section
3,
result
and
discussion:
pro
vides
an
in-depth
e
xplanation
of
the
analysis
of
the
study
e
xperiment
results.
Section
4,
conclusion:
presents
the
research’
s
conclusions
along
with
recommendations
for
further
research.
2.
METHOD
Figure
1
illustrates
the
research
process.
Continuing
the
pre
vious
research
[12],
the
dataset
w
as
ob-
tained
in
the
data
scraping
stage,
i
n
the
form
of
digital
ne
ws
articles
in
the
Indonesian
language
retrie
v
ed
from
the
author’
s
site
on
the
ne
ws
website.
Once
the
g
athering
process
is
completed,
the
data
is
stored
in
a
format
that
ts
this
research.
In
this
research,
a
class
cate
gory
,
nota
bly
the
author’
s
gender
,
w
as
added
to
increase
the
data’
s
comple
xity
as
the
second
label.
After
data
collection,
the
second
stage
is
data
pre-processing.
This
stage
co
v
ers
sub-processes
for
cleaning
the
collected
data,
such
as
the
remo
v
al
of
special
symbols,
URLs
or
h
yperlinks,
and
irrele
v
ant
information
lik
e
picture
captions,
adv
ertisements,
or
comments.
After
the
data
has
been
properly
cleaned,
it
is
then
shuf
ed
to
a
random
representation
and
a
v
oid
bias.
Figure
1.
Research
o
w
Additionally
,
the
utilized
dataset
is
separat
ed
into
3
parts:
training,
v
alid,
and
test
data.
The
rst
one
is
training
data
for
the
model
training
phase;
the
second
is
v
alid
data
used
for
model
parameter
tuning;
the
last
is
test
data
used
to
e
v
aluate
the
utilized
model’
s
performance.
Because
IndoBER
T
can
read
more
than
one
pattern
from
the
gi
v
en
data,
data
splitting
is
emplo
yed
to
generate
more
di
v
erse
training
data.
Out
of
the
total
2000
collected
data,
the
primary
dataset
is
di
vided
into
1300
training
data,
455
v
alidation
data,
and
245
test
data.
Furthermore,
each
of
the
content
and
label
columns
are
mer
ged
to
or
g
anize
the
data
in
a
suitable
form
for
training
the
multilabel
multiclass
model.
The
nal
step
in
v
olv
es
encoding
the
data
using
the
one-hot
encoding
method.
One-hot
encoding
is
used
t
o
transform
the
te
xt
into
binary
numeric
v
ectors
that
can
be
processed
by
the
model
[24],
[25].
Each
unique
w
ord
in
the
dataset
is
mapped
to
a
binary
v
ector
of
the
same
size,
enabling
the
model
to
understand
and
process
the
data
in
a
numeric
format.
W
ith
the
help
of
this
data
preprocessing
pipeline,
the
model
will
be
able
to
learn
and
accompan
ying
labels
to
predict
the
author
and
gender
with
accurac
y
.
F
or
the
training
phase,
the
IndoBER
T
model
w
as
used,
which
has
pro
v
en
to
be
ef
fecti
v
e
in
processing
the
Indonesian
language.
Before
be
ginning
training,
the
data
is
tok
enized
with
the
Bert
tok
enizer
,
which
sepa-
rates
the
te
xt
into
fragments
such
as
w
ords
or
sub-w
ords.
The
tok
enization
steps
prepare
the
data
to
ensure
the
model
can
utilize
it.
The
h
yperparameters
that
emplo
yed
in
the
training
process
are:
10
epochs,
8
batch
size,
3e-5
learning
rate,
0.01
weight
decay
,
and
100
logging
steps.
After
the
trai
ning
process,
the
model
is
e
v
aluated
using
classication
report
in
the
form
of
accurac
y
and
F1-score.
The
F1-score
calculate
using
precision
and
re-
call
v
alues
to
of
fer
a
comprehensi
v
e
e
v
aluation
of
the
model’
s
performance.
Accurac
y
measures
the
percentage
of
model’
s
accurate
predictions
based
on
total
predictions,
meanwhile
the
F1-score
measures
the
percentage
of
model’
s
correct
predictions
based
on
total
predictions.
F
ollo
wing
these
steps,
this
research
could
pro
vide
an
ef
fecti
v
e
and
precise
multi-label
multi-class
classication
model
for
Indonesian
ne
ws
i
tems,
allo
wing
for
impro
v
ed
information
structure
and
analysis
in
the
conte
xt
of
Indonesian
ne
ws
article.
Int
J
Artif
Intell,
V
ol.
13,
No.
4,
December
2024:
4688–4694
Evaluation Warning : The document was created with Spire.PDF for Python.
Int
J
Artif
Intell
ISSN:
2252-8938
❒
4691
3.
RESUL
TS
AND
DISCUSSION
The
dataset
for
this
research
is
from
earlier
research
[12],
that
contained
2,000
Indonesian
digita
l
ne
ws
articles
in
the
form
of
comma
separated
v
alue
(C
SV)
les.
F
or
each
author
in
the
dataset
contrib
uted
100
unique
articles
to
g
ain
a
more
wide
range
of
writing
styles
and
patterns.
The
number
of
authors
in
the
dataset
has
been
reduced
from
80
in
the
pre
vious
research
[12]
to
20
authors
in
this
research
to
minimize
the
potential
of
model
o
v
ertting.
T
o
increase
the
comple
xity
of
the
data
used
in
this
research,
a
ne
w
class
w
as
added,
namely
author
gender
,
to
mak
e
the
research
multi
-labeled.
Due
to
its
multi-label
approach,
the
utilized
model
can
determine
rele
v
ant
authors
with
high
accurac
y
,
e
v
en
when
assigning
man
y
dif
ferent
author
classes
to
each
digital
ne
ws
article.
In
this
research,
the
model
’
s
h
yper
-parameters
ha
v
e
been
adapted
at
learning
rate
(LR)
=
3e-
5,
batch
size
=
8,
weight
decay
=
0.01,
logging
steps
=
100,
and
training
epoch
=
10,
t
o
impro
v
e
the
model’
s
performance
in
minimizing
o
v
ertting
and
obtaining
more
optimal
results
than
pre
vious
research.
F
or
the
training
stage,
the
sigmoid
acti
v
ation
function
w
as
used
for
the
multi-label
classication
task.
The
model’
s
training
results
yield
impressi
v
e
per
formance
metrics
as
seen
in
T
able
1.
The
results
are
in
the
form
of
training
accurac
y
=
0.8879,
the
F1-score
=
0.9360,
the
training
loss
=0.0187,
and
the
v
alidation
loss
=
0.0449.
These
e
v
aluation
metrics
indicate
the
model’
s
rob
ustness
and
ef
fecti
v
eness
in
classifying
the
ne
ws
articles
and
attrib
uting
them
to
their
respecti
v
e
authors.
The
model
demonstrates
high
accurac
y
and
F1-
score,
implying
that
the
model
utilized
can
correctly
identify
authors’
distinct
writing
styles
and
characteristics.
Additionally
,
a
comparison
of
model
training
results,
sho
wn
in
T
able
2,
indicates
this
research
model
succeeded
in
o
v
ercoming
the
o
v
ertting
problem
in
pre
vious
research
[12].
This
can
be
seen
through
the
v
alidation
loss
v
alue
which
is
decreased
to
0.0449
and
the
training
accurac
y
which
is
slightly
increased
to
0.8879.
These
results
were
supported
by
modications
to
the
dataset
utilized
and
adding
additional
training
epochs.
T
able
1.
Model
training
result
Epoch
T
raining
loss
V
alidation
loss
F1
score
T
raining
accurac
y
1
-
5
...
...
...
...
6
0.0318
0.0502
0.9319
0.8747
7
0.0248
0.0461
0.9354
0.8901
8
0.0215
0.0449
0.9371
0.8923
9
0.0198
0.0449
0.9360
0.8879
10
0.0187
0.0449
0.9360
0.8879
T
able
2.
Model
training
result
comparison
T
otal
training
epoch
T
raining
loss
V
alidation
loss
T
raining
accurac
y
Earlier
research
[12]
8
0.062
0.992
0.84
This
research
10
0.0187
0.0449
0.8879
After
the
completion
of
the
training
process,
an
e
v
aluation
is
conducted
to
assess
the
model’
s
perfor
-
mance
in
the
form
of
classication
report
that
pro
vides
v
aluable
insights
into
k
e
y
metrics,
including
precision,
recall,
and
F1-score,
which
play
s
ignicant
roles
in
g
auging
the
model’
s
ef
fecti
v
eness
in
classifying
ne
ws
ar
-
ticles
accurately
.
Precision
measures
the
proportion
of
true
positi
v
e
predictions
(correctly
predicted
positi
v
e
cases)
out
of
all
positi
v
e
forecasts
made
by
the
model.
Recall
represents
the
proportion
of
positi
v
e
instances
in
the
dataset
that
the
classier
correctly
identied
among
all
the
actual
positi
v
e
cases.
It
sho
wcases
the
model’
s
capability
to
capture
all
rele
v
ant
instances
belonging
to
a
specic
class.
The
F1-score
serv
es
as
an
o
v
erall
mea-
sure
of
a
model’
s
performance,
balancing
both
the
precision
and
recall
scores.
F1-score
combines
these
tw
o
metrics
to
pro
vide
a
comprehensi
v
e
e
v
aluation
of
the
model’
s
accurac
y
in
predicting
both
positi
v
e
and
ne
g
ati
v
e
cases.
The
formula
that
serv
es
as
the
foundation
for
the
e
v
aluation
calculation
is
calculated
as
follo
ws.
pr
ecision
=
T
P
T
P
+
F
P
(1)
r
ecal
l
=
T
P
T
P
+
F
N
(2)
F
1
scor
e
=
T
P
1
2
+
(
F
P
+
F
N
)
(3)
Indonesian
ne
ws
article
author
ship
attrib
ution
multilabel
multiclass
...
(Kar
en
Etania
Saputr
a)
Evaluation Warning : The document was created with Spire.PDF for Python.
4692
❒
ISSN:
2252-8938
When
e
v
aluating
the
model
in
the
conte
xt
of
AA,
the
assessment
measures
indicated
in
the
pre
vious
formulations
are
cruci
al.
The
counts
of
true
positi
v
e
(TP),
f
alse
positi
v
e
(FP),
and
f
alse
ne
g
at
i
v
e
(FN)
instances
are
used
to
calculate
these
metrics.
Where
TP
denotes
the
number
of
accurate
positi
v
e
predictions
made
by
the
model,
FP
denotes
the
number
of
inaccurate
positi
v
e
predictions,
and
FN
denotes
the
number
of
positi
v
e
instances
that
were
mistak
enly
classied
as
ne
g
ati
v
e.
T
able
3
presents
a
detailed
breakdo
wn
of
the
e
v
aluation
ndings
for
each
label
(article’
s
author),
re
v
ealing
distinct
performance
le
v
els
across
dif
ferent
authors.
Notably
,
Di
v
a
Luana
Putri
e
xhibits
the
lo
west
e
v
aluation
v
alues
among
all
labels,
with
a
precision
of
0.50,
recall
of
0.40,
and
F1
score
of
0.44.
These
results
suggest
that
the
author’
s
language
choices
might
be
less
di
v
erse,
leading
to
comparati
v
ely
less
clarity
in
the
AA
task.
On
the
other
hand,
author
Abdul
Haris
Maulana
and
author
Ja
w
ahir
Gusta
v
Rizal
recei
v
e
the
highest
ratings
for
their
labels,
achie
ving
precision,
recall,
and
F1-
scores
of
1.
These
e
xceptional
scores
indicate
a
a
wless
performance
in
the
classication
and
attrib
ution
of
articles
to
those
authors.
T
able
3.
Classication
report
result
Label
Precision
Recall
F1
Score
Aisyah
Sekar
A
yu
Maharani
0.80
0.92
0.86
Di
v
a
Luana
Putri
0.50
0.40
0.44
Fika
Nurul
Ulya
0.75
0.75
0.75
Kiki
Satri
0.69
0.82
0.75
Lely
Maulida
1.00
1.00
1.00
Nadia
F
aradiba
0.79
1.00
0.88
Nur
Fitriatus
Shalihah
0.83
0.59
0.69
Rintan
Puspita
Sari
0.94
1.00
0.97
Silmi
Nurul
Utami
0.88
0.70
0.78
V
an
ya
Karunia
Mulia
Putri
1.00
0.86
0.92
Abdul
Haris
Maulana
1.00
1.00
1.00
Ja
w
ahir
Gusta
v
Rizal
1.00
1.00
1.00
Labib
Zamani
0.93
1.00
0.96
M
Adika
F
aris
Ihsan
1.00
1.00
1.00
Muhammad
Idris
1.00
0.91
0.95
Nicholas
Ryan
Aditya
1.00
0.86
0.92
Nur
Jamal
Shaid
0.90
1.00
0.95
Reza
Agustian
1.00
0.85
0.92
Rully
R
Ramli
1.00
0.57
0.73
Reza
K
urnia
Darma
w
an
0.93
0.88
0.90
L
0.97
0.93
0.95
P
0.94
0.97
0.95
Micro
a
v
erage
0.92
0.90
0.91
The
dif
ferences
in
assessment
metrics
across
authors
underline
each
author’
s
unique
writing
st
yle
and
language.
While
linguistic
patterns
may
be
more
dif
cult
to
disco
v
er
in
certain
authors,
resulting
in
lo
wer
e
v
aluation
scores,
for
others,
the
model
may
be
able
to
generate
accurate
predictions
with
high
precision
and
recall.
This
performance
v
ariance
sho
ws
the
dif
culties
of
AA
task,
which
requires
a
deep
e
xamination
of
each
author’
s
unique
linguistic
preferences
and
writing
characteristics.
The
analysi
s
of
gender
labels
sho
ws
remarkable
gender
classication
skills,
with
an
a
v
erage
accurac
y
,
recall,
and
F1-score
of
0.95.
This
high
accurac
y
in
identifying
gender
is
crucial
as
it
sho
ws
the
model’
s
ability
to
dif
ferentiate
gender
-based
writing
patterns
in
ne
ws
articles.
T
able
3
sho
ws
the
results
of
model
e
v
aluation
usi
ng
a
classication
report,
containing
precision,
recall,
and
F1-score
v
alues,
as
well
as
the
a
v
erage
of
each
cate
gory
calculated
with
the
micro-
a
v
eraging
approach.
The
model’
s
a
v
erage
precision
score
of
0.92
indic
ates
that
the
model
accurately
detects
and
labels
the
data
that
belong
to
a
certain
cate
gory
.
The
model’
s
a
v
erage
recall
score
of
0.90
indicates
that
the
model
ef
fecti
v
ely
retrie
v
ed
e
v
ents
from
dif
ferent
classes.
The
classication
report
in
T
abel
3
sho
ws
a
micro-
a
v
erage
F1-score
of
0.91,
sho
wing
a
balanced
performance
in
terms
of
recall
and
precision.
These
note
w
orth
y
micro-a
v
erage
results
indicate
the
model
utilized,
namely
IndoBER
T
,
has
performed
well
in
classifying
ne
ws
articles
and
assigning
them
to
their
authors,
especially
in
this
research
with
Indonesian
digital
ne
ws
article
dataset
[12].
The
model’
s
ability
to
accurately
project
positi
v
e
occurrences
while
limiting
FP
and
FN
is
sho
wn
through
the
model’
s
high
precision,
recall,
and
F1-score,
pro
ving
its
rob
ustness
and
dependability
in
the
Indonesian
AA
task.
T
able
3
about
e
v
aluation
details
not
only
pro
vides
v
aluable
insight
into
the
model’
s
performance
b
ut
also
acts
as
a
critical
reference
point
for
further
research
and
study
in
the
eld
Int
J
Artif
Intell,
V
ol.
13,
No.
4,
December
2024:
4688–4694
Evaluation Warning : The document was created with Spire.PDF for Python.
Int
J
Artif
Intell
ISSN:
2252-8938
❒
4693
of
Indonesian
AA.
The
metrics
used
in
this
research
can
be
implemented
by
other
researchers
to
e
v
aluate
the
performance
of
dif
ferent
models,
nd
areas
for
adv
ancement,
and
mak
e
data-dri
v
en
decisions
that
optimize
the
model’
s
performance.
4.
CONCLUSION
This
research
successfully
o
v
ercomes
the
problems
f
aced
in
earlier
research
by
modifying
the
dataset
utilized
in
tw
o
w
ays:
i)
adding
a
ne
w
class
in
the
form
of
the
author’
s
gender
and
ii)
reducing
the
number
of
author
labels,
which
equalizes
the
amount
of
data
v
ersus
cla
ss.
In
addition,
the
result
of
the
research
is
tightly
associated
with
the
proper
model’
s
h
yper
-parameters
tuning.
The
e
v
aluation
results
demonstrates
the
model’
s
success
in
classifying
Indonesian
digital
ne
ws
articles
and
identifying
their
authors,
with
outstanding
performance,
the
precision
0.92,
recall
0.90,
and
F1-score
0.91.
These
results
will
act
as
a
benchmark
for
upcoming
research
and
de
v
elopment
in
Indonesian
AA
as
the
eld
continues
to
e
v
olv
e.
Ho
we
v
er
,
despite
its
e
xcellent
results,
the
model’
s
accurac
y
is
still
belo
w
.
This
research
still
has
potential
for
more
adv
ancement,
for
e
xample
utilizing
Indonesian
datasets
with
higher
comple
xity
and
lar
ger
amounts
and
implementing
better
model
para
meters
conguration
to
obtain
more
optimal
results.
Researchers
can
use
the
implemented
metrics
in
this
research
to
compare
models,
enhance
performance
by
making
data-dri
v
en
decisions,
or
identify
areas
for
impro
v
ements
in
Indonesian
AA.
REFERENCES
[1]
H.
S.
A.
-Ash,
M.
F
.
Putri,
P
.
Mursanto,
and
A.
Bustamam,
“Ensemble
learning
approach
on
Indonesian
f
ak
e
ne
ws
classication,
”
in
2019
3rd
International
Conference
on
Informatics
and
Computational
Sciences
(ICICoS)
,
2019,
pp.
1–6,
doi:
10.1109/ICI-
CoS48119.2019.8982409.
[2]
R.
Baly
et
al.
,
“What
w
as
written
vs.
who
read
it:
ne
ws
media
proling
using
te
xt
analysis
and
social
media
cont
e
xt
,
”
in
Proceedings
of
the
58th
Annual
Meeting
of
the
Association
for
Computational
Linguistics
,
2020,
pp.
3364–3374,
doi:
10.18653/v1/2020.acl-
main.308.
[3]
P
.
Abbasi,
J.
Y
oose-Lebni,
A.
Jalali,
A.
Ziapour
,
and
P
.
Nouri,
“Causes
of
the
plagiarism:
a
grounded
theory
study
,
”
Nursing
Ethics
,
v
ol.
28,
no.
2,
pp.
282–296,
2021,
doi:
10.1177/0969733020945753.
[4]
T
.
F
olt
´
ynek,
N.
Meuschk
e,
and
B.
Gipp,
“
Academic
plagiarism
detection:
a
systematic
literature
re
vie
w
,
”
A
CM
Computing
Surv
e
ys
,
v
ol.
52,
no.
6,
pp.
1–42,
2020,
doi:
10.1145/3345317.
[5]
P
.
Barber
´
a,
A.
E.
Bo
ydstun,
S.
Linn,
R.
McMahon,
and
J.
Nagler
,
“
Automated
te
xt
classication
of
ne
ws
articles:
a
practical
guide,
”
Political
Analysis
,
v
ol.
29,
no.
1,
pp.
19–42,
2021,
doi:
10.1017/pan.2020.8.
[6]
T
.
Schuster
,
R.
Schuster
,
D.
J.
Shah,
and
R.
Barzilay
,
“The
limitations
of
stylometry
for
detecting
machine-generated
f
ak
e
ne
ws,
”
Computational
Linguistics
,
v
ol.
46,
no.
2,
pp.
499–510,
2020,
doi:
10.1162/COLI
a
00380.
[7]
D.
D.
Maharani,
“
A
sociolinguistics
analysis
of
language
s
tyle
in
wild
child,
”
Bachelor
Thesis
,
Department
of
English
Education,
Uni
v
ersitas
Muhammadiyah
Mataram,
Mataram,
Indonesia,
2019.
[8]
F
.
Iqbal,
M.
Debbabi,
and
B.
C.
M.
Fung,
Machine
learning
for
authorship
attrib
ution
and
c
yber
forensics
.
Cham:
Springer
,
2020,
doi:
10.1007/978-3-030-61675-5.
[9]
U.
Mahor
and
A.
K
umar
,
“
A
comparati
v
e
study
of
stylometric
characteristics
in
authorship
attrib
ution,
”
in
Information
and
Com-
munication
T
echnology
for
Compet
iti
v
e
Strate
gies
(ICTCS
2021)
,
Sing
apore:
Springer
,
2023,
pp.
71–81,
doi:
10.1007/978-981-19-
0095-2
8.
[10]
V
.
Kalgutkar
,
R.
Kaur
,
H.
Gonz
alez,
N.
Stakhano
v
a,
and
A.
Matyukhina,
“Code
authorship
attrib
ution:
Methods
and
challenges,
”
A
CM
Computing
Surv
e
ys
,
v
ol.
52,
no.
1,
pp.
1–36,
2019,
doi:
10.1145/3292577.
[11]
P
.
Rizvi,
“The
problem
of
microattrib
ution,
”
Digital
Scholarship
in
the
Humanities
,
v
ol.
34,
no.
3,
pp.
606–615,
2018,
doi:
10.1093/digitalsh/fqy066.
[12]
S.
K.
Etania,
Riccosan,
and
A.
Cho
w
anda,
“
Automatic
Indonesian
authorship
attrib
ution
recognition
using
transformer
,
”
ICIC
Express
Letters
,
v
ol.
17,
no.
5,
pp.
497–503,
2023,
doi:
10.24507/icicel.17.05.497.
[13]
F
.
T
og
atorop
and
D.
Y
.
Saragi
h,
“The
dif
ference
between
male
and
female
language
style
of
students
grade
viii,
”
Journal
on
Educa-
tion
,
v
ol.
5,
no.
3,
pp.
9471–9479,
2023,
doi:
10.31004/joe.v5i3.1762.
[14]
M.
L.
Zhang,
Y
.
K.
Li,
H.
Y
ang,
and
X.
Y
.
Liu,
“T
o
w
ards
class-i
mbalance
a
w
are
multi-label
learning,
”
IEEE
T
ransactions
on
Cybernetics
,
v
ol.
52,
no.
6,
pp.
4459–4471,
2022,
doi:
10.1109/TCYB.2020.3027509.
[15]
R.
M.
Mathe
w
and
R.
Gunasundari,
“
A
re
vie
w
on
handling
multiclass
imbalanced
data
classication
in
education
domain,
”
in
2021
International
Conference
on
Adv
ance
Computing
and
Inno
v
ati
v
e
T
echnologies
in
Engineering
(ICA
CITE)
,
2021,
pp.
752–755,
doi:
10.1109/ICA
CITE51222.2021.9404626.
[16]
J.
Hitschler
,
E.
V
.
D.
Ber
g,
and
I.
Rehbein,
“
Authorship
attrib
ution
with
con
v
olutional
neural
netw
orks
and
POS-eliding,
”
in
Pro-
ceedings
of
the
W
orkshop
on
Stylistic
V
ariation
,
2017,
pp.
53–58,
doi:
10.18653/v1/W17-4907.
[17]
A.
Uchendu,
T
.
Le,
K.
Shu,
and
D.
Lee,
“
Authorship
attrib
ution
for
neural
te
xt
generation,
”
in
Proceedings
of
the
2020
Conference
on
Empirical
Methods
in
Natural
Language
Processing
(EMNLP)
,
2020,
pp.
8384–8395,
doi:
10.18653/v1/2020.emnlp-main.673.
[18]
M.
F
abien,
E.
V
illatoro-T
ello,
P
.
Motlicek,
and
S.
P
arida,
“BertAA:
BER
T
ne-tuning
for
Authorship
Attrib
ution,
”
in
Proceedings
of
the
17th
International
Conference
on
Natural
Language
Processing
(ICON)
,
2020,
pp.
127–137.
[19]
J.
De
vlin,
M.
W
.
Chang,
K.
Lee,
and
K.
T
outano
v
a,
“BER
T
:
Pre-traini
ng
of
deep
bidirectional
transformers
for
language
under
-
standing,
”
in
2019
Conference
of
the
North
American
Chapter
of
the
Association
for
Computational
Linguistics:
Human
Language
Indonesian
ne
ws
article
author
ship
attrib
ution
multilabel
multiclass
...
(Kar
en
Etania
Saputr
a)
Evaluation Warning : The document was created with Spire.PDF for Python.
4694
❒
ISSN:
2252-8938
T
echnologies
,
2019,
v
ol.
1,
pp.
4171–4186.
[20]
A.
V
asw
ani
et
al.
,
“
Attention
is
all
you
need,
”
in
31st
Conference
on
Neural
Information
Processing
System
s
(NIPS
2017)
,
2017,
pp.
1–15.
[21]
F
.
K
oto,
A.
Rahimi,
J.
H.
Lau,
and
T
.
Baldwin,
“IndoLEM
and
IndoBER
T
:
a
benchmark
dataset
and
pre-trained
lan-
guage
model
for
Indonesian
NLP
,
”
in
28th
International
Conference
on
Comput
ational
Linguistics
,
2020,
pp.
757–770,
doi:
10.18653/v1/2020.coling-main.66.
[22]
B
.
Juarto
and
Y
ulianto,
“Indonesian
ne
ws
classication
using
IndoBert,
”
International
Journal
of
Int
elligent
Systems
and
Applica-
tions
in
Engineering
,
v
ol.
11,
no.
2,
pp.
454–460,
2023.
[23]
D.
U.
K.
Putri
and
D.
N.
Pratomo,
“Clickbait
detection
of
Indonesi
an
ne
ws
headlines
using
ne-tune
bidirectional
encoder
represen-
tations
from
transformers
(BER
T),
”
Inform:
Jurnal
Ilmiah
Bidang
T
eknologi
Informasi
dan
K
omunikasi
,
v
ol.
7,
no.
2,
pp.
162-168,
2022,
doi:
10.25139/inform.v7i2.4686.
[24]
S
.
H.
K
ong,
L.
M.
T
an,
K.
H.
Gan,
and
N.
H.
Samsudin,
“F
ak
e
ne
ws
detection
using
deep
learning,
”
in
2020
IEEE
10th
Symposium
on
Computer
Applications
&
Industrial
Electronics
(ISCAIE)
,
2020,
pp.
102–107,
doi:
10.1109/ISCAIE47305.2020.9108841.
[25]
M
.
K.
Dahouda
and
I.
Joe,
“
A
deep-learned
embedding
technique
for
cate
gorical
features
encoding,
”
IEEE
Access
,
v
ol.
9,
pp.
114381–114391,
2021,
doi:
10.1109/A
CCESS.2021.3104357.
BIOGRAPHIES
OF
A
UTHORS
Kar
en
Etania
Saputra
recei
v
ed
her
master’
s
de
gree
in
Computer
Science
from
Bina
Nusantara
Uni
v
ersity
.
She
is
a
full-time
le
cturer
in
the
computer
science
program
at
Bina
Nusantara
Uni
v
ersity’
s
Bandung
Campus.
Her
primary
research
focus
lies
in
the
eld
of
deep
learning.
She
can
be
contacted
at
email:
karen.saputra@binus.ac.id.
Riccosan
obtained
his
master’
s
de
gree
in
Computer
Science
from
Bina
Nusantara
Uni
v
er
-
sity
.
He
currently
serv
es
as
a
full-time
lecturer
in
the
computer
science
program
at
Bina
Nusantara
Uni
v
ersity’
s
Bandung
Campus.
His
primary
research
interest
re
v
olv
es
around
emotions
and
sentiment
analysis,
delving
into
the
f
ascinating
w
orld
of
understanding
human
emotions
through
technology
.
He
can
be
contacted
at
email:
riccosan@binus.ac.id.
Int
J
Artif
Intell,
V
ol.
13,
No.
4,
December
2024:
4688–4694
Evaluation Warning : The document was created with Spire.PDF for Python.