Indonesi
an
Journa
l
of El
ect
ri
cal Engineer
ing
an
d
Comp
ut
er
Scie
nce
Vo
l.
23
,
No.
1
,
Ju
ly
2021
, p
p.
1
9
7
~
205
IS
S
N: 25
02
-
4752, DO
I: 10
.11
591/ijeecs
.v
23
.i
1
.
pp
1
9
7
-
205
197
Journ
al h
om
e
page
:
http:
//
ij
eecs.i
aesc
or
e.c
om
Measuri
ng the a
ccur
acy of LS
TM
and BiL
STM mo
dels in t
he
appli
ca
ti
on of
ar
t
ificial
intellig
ence
by app
lying cha
tbot
program
me
Prasnurz
ak
i
An
ki
1
, Alh
ad
i
Bustamam
2
1
,2
Depa
rtment
of
Mathe
m
atics,
Univer
sit
as
Indon
esia
,
Indone
si
a
2
Data
Sc
ie
n
ce
C
ent
re
,
Univ
ersi
tas
Indone
sia, Ind
onesia
Art
ic
le
In
f
o
ABSTR
A
CT
Art
ic
le
history:
Re
cei
ved
Dec
30, 202
0
Re
vised
Ma
y
25
, 2
021
Accepte
d
J
un
4
, 2
021
P
y
thon
progr
amm
e
cont
ai
ns
a
q
uesti
on
and
ans
wer
s
y
stem
th
at
der
ive
d
from
dat
a
sets
th
at
h
ave
used
and
i
m
ple
m
ent
ed
the
cha
tbot
in
thi
s
m
oder
n
era.
where
the
d
at
a
col
l
ec
t
ed
is
in
t
he
form
of
cor
puses
cont
ai
nin
g
ext
ensi
v
e
m
et
ada
t
a
-
ri
ch
fi
ct
ion
al
conve
rsa
ti
ons
der
iv
ed
fr
om
ext
racte
d
f
i
lm
script
s,
comm
only
c
al
l
e
d
cor
ne
ll
m
ovi
e
dialogue
cor
pu
s
.
The
var
ious
m
odel
s
have
bee
n
used
ch
a
tb
ots
in
p
y
thon
pr
ogra
m
m
es,
and
LSTM
and
Bi
L
STM
m
odel
s
were
spec
ifi
c
al
l
y
used
in
thi
s
s
tud
y
.
W
her
e
the
form
of
ac
cur
a
c
y
will
b
e
rep
orte
d
as
a
res
ult
of
the
imple
m
ent
at
ion
of
LS
TM
and
BiL
STM
m
odel
s
in
the
chatbot
prog
ramm
e.
The
pro
gra
m
m
e
per
form
anc
e
will
b
e
i
nflue
nc
ed
b
y
the
data
from
the
m
odel
sele
c
ti
on,
b
ec
ause
the
le
ve
l
of
ac
cur
acy
i
s
det
ermined
b
y
t
he
t
arg
et
progra
m
m
e
bei
ng
ta
k
e
n.
So
th
is
is
th
e
m
ai
n
fa
ct
or
tha
t
de
te
rm
ine
s
which
m
odel
to
choose
.
Based
o
n
conside
ra
ti
ons
req
uire
d
for
choosing
the
pr
ogra
m
m
e
m
odel
,
in
the
end
th
e
LSTM
and
th
e
BiL
STM
m
odel
s
are
chose
n
and
will
be
appl
ie
d
to
the
progra
m
m
e.
B
ase
d
on
th
e
LSTM
and
BiLS
TM
cha
tbot
p
r
ogra
m
m
es
tha
t
have
be
en
t
ested,
it
ca
n
be
conc
lud
ed
th
at
t
he
best
p
ara
m
eters
come
from
a
pa
ir
of
B
iL
ST
M
cha
tbot
s
usin
g
the Bi
L
TSM m
odel
with
a
n
ave
r
age acc
ur
a
c
y
va
lue of
0
.
99
5217.
Ke
yw
or
ds:
Ar
ti
fici
al
intel
li
gen
ce
Bi
LSTM
Chatb
ot
Data sci
ence
LSTM
This
is an
open
acc
ess arti
cl
e
un
der
the
CC
B
Y
-
SA
l
ic
ens
e.
Corres
pond
in
g
Aut
h
or
:
Alha
di Bustam
a
m
Gedu
ng D,
Ka
m
pu
s Baru FM
IF
A
,
Un
i
ver
sit
as Indo
nesia
Dep
ok, Ja
wa
Barat 1
6424,
Indon
e
sia
Em
a
il
: al
had
i@sci
.u
i.ac
.id
1.
INTROD
U
CTION
Chatb
ots
are
a
uto
m
at
ed
syst
em
s
create
d
to
help
us
e
rs
by
a
ns
we
rin
g
t
heir
qu
e
sti
on
s
.
F
or
bu
si
nesses
,
chatb
ots
can
pro
vid
e
a
bette
r
way
to
conne
ct
with
their
custom
ers
and
i
ncr
ease
c
us
to
m
er
sat
isfact
ion
le
vels
.
Custom
ers
get
a
bette
r,
m
or
e
conve
nient
wa
y
to
get
answ
e
rs
to
their
ques
ti
on
s
with
out
wait
ing
on
the
pho
ne
or
sen
ding
fr
e
qu
e
nt
em
a
il
s
[1
]
.
A
rtific
ia
l
intel
li
gen
ce
(
AI
)
has
m
ade
a
n
im
pact
in
ever
y
day
act
ivit
ie
s
by
desig
ning
an
d
prov
i
ding
eva
luati
on
of
s
ophisti
cat
ed
ap
plica
ti
on
s
an
d
de
vices,
w
hich
can
pe
rfor
m
var
i
ou
s
functi
ons.
C
ha
tbo
t
is
an
a
rtif
ic
ia
l
intel
l
igen
ce
pro
gr
am
m
e,
wh
ic
h
is
base
d
on
the
de
velop
m
ent
of
AI,
it
is
hope
d
that t
he
chatb
ot'
s ab
il
ity t
o
i
m
it
a
te
h
um
an
age
nts in conve
rsati
on. C
hatb
ots h
a
ve becom
e so
c
om
m
on
in
their
presence
that
they
can
reduce
ser
vic
e
costs
an
d
ca
n
ha
ndle
m
ult
iple
custom
ers
si
m
ultaneou
sl
y
[2
]
.
Hope
fu
ll
y,
fu
t
ur
e
c
hatb
ots
can
i
m
pr
ove
business
sect
or
per
f
orm
ance
by
increasin
g
custom
er
sat
isfact
io
n
le
vels
by
sa
ving
ti
m
e.
They
will
al
so
sa
ve
custom
er
ser
vice
em
plo
ye
es
tim
e;
custom
er
s
can
us
e
c
hat
bo
ts
t
o
get in
form
at
io
n
that
previ
ou
s
ly
r
eq
uired hu
m
ans
to a
n
s
we
r qu
est
io
ns m
a
nu
al
ly
.
A
fetc
h
m
od
el
con
ta
in
s
se
veral
fo
rm
s
base
d
on
m
at
ches
de
rive
d
f
ro
m
us
e
r
in
pu
t
a
nd
the
chatb
ot
can
gen
e
rate
a
ns
w
ers
base
d
on
th
e
f
or
m
s
that
th
e
us
e
r
has
fill
ed
in
.
Her
e
kn
owle
dge
us
e
d
in
chatb
ots
is
a
f
or
m
of
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
S
ci
,
Vo
l.
23
, N
o.
1
,
Ju
ly
2021
:
1
9
7
-
2
0
5
198
hu
m
an
ha
nd
c
od
e
.
Chat
bo
t
knowle
dge
c
onstr
uction
is
ti
m
e
-
con
s
um
ing
an
d
di
ff
ic
ult.
The
refor
e
,
it
is
very
i
m
po
rtant
to
ha
ve
an
aut
om
ated
knowle
dge
extracti
on
m
ec
han
ism
to
bu
il
d
var
i
ous
form
s
of
chat
bo
ts
[
3].
The
us
e
of
m
od
el
s
that
can
i
m
pr
ove
chatb
ot
pe
rfor
m
ance
in
ans
wer
i
ng
quest
i
ons
autom
at
ic
al
l
y
can
be
con
si
der
e
d
to co
m
par
e
w
hi
ch
tw
o
m
od
el
s
w
il
l i
nf
l
uen
ce
chatb
ot p
e
rfo
r
m
ance and
det
erm
ine w
hich
m
od
el
h
as a
be
tt
er f
it
.
A
syst
em
that
can
recei
ve
fee
db
ac
k
a
nd
res
pond
from
us
ers
and
c
an
kee
p
the
co
nversati
on
goin
g,
is
cal
le
d
a
chatbo
t.
The
encode
r
-
dec
oder
arc
hitec
ture
is
us
ed
i
n
buil
ding
par
t
s
of
the
chatb
ot
[4
]
.
A
chatb
ot
is
a
si
m
ple
robo
t
t
hat
co
ntains
a
pro
gr
am
me
to
answer
quest
io
ns
from
us
ers.
Af
te
r
that,
the
ans
wer
data
will
be
gen
e
rated
f
rom
the
quest
io
ns
aske
d
by
th
e
us
e
r.
The
sem
antic
quest
io
n
-
answerin
g
syst
e
m
has
dev
el
oped
in
wh
ic
h
w
ords
t
hat
are
u
nce
rta
in
are
the
f
or
m
of
t
h
e q
ue
sti
on [
5].
A
ppli
cat
ion
of
quest
ion
an
d
a
nswer
sy
stem
in
the form
o
f
a c
hatb
ot is e
xp
ec
te
d
to a
nswer
these c
halle
nge
s.
A
rtific
ia
l
intel
li
gen
ce
(
Al
)
wh
i
c
h
is
the
l
at
est
te
chn
ologica
l
adv
a
nce
m
ent
is
ver
y
help
fu
l
in
the
dev
el
op
m
ent
of
ne
w
virtua
l
assist
ants
to
be
eff
ic
ie
nt
(
on
li
ne
chatb
ots).
M
eanwhil
e,
the
stud
y
al
so
a
na
ly
ze
s
how
existi
ng
t
echnolo
gical
adv
a
nces
m
ade
on
new
c
hatb
ots
ha
ve
an
i
m
pact
on
fu
tu
re
cust
om
er
s
upport
.
Go
i
ng
for
ward
,
te
chnolo
gical
innov
at
io
ns
i
n
A
I
al
low
c
ha
tbo
ts
to
perf
orm
increasing
l
y
com
plex
ta
sk
s
[
6].
NLP
(
natu
ral
la
nguag
e
proc
essing
)
is
a
m
echan
ism
that
can
be
us
e
d
to
su
pp
or
t
co
m
pu
te
r
m
achi
nes
by
si
m
ulati
ng
hu
m
an
abili
ti
es
that
f
un
ct
ion
to
unde
rsta
nd
la
ngua
ge
[7
]
.
Na
tural
la
ng
uag
e
processi
ng
is
a
no
t
he
r
area
w
her
e
t
he
sta
nce
of
dee
p
le
arn
in
g
ca
n
ha
ve
a
hu
ge
im
pact
on
ex
pe
rim
entat
ion
that
cou
l
d
occ
ur
ov
er
the
nex
t
fe
w
ye
ars
[8
]
.
I
n
NLP
m
od
el
s
,
LS
TM
c
on
si
ders
the
or
der
de
pe
nd
e
nc
e
betwe
e
n
w
ord
se
quences
th
at
the
te
st
will
perfor
m
on
the p
r
ogr
a
m
m
e
to
captu
re d
epe
ndencie
s
in both
t
he
l
ong
an
d
short
-
ra
ng
e
f
or
m
s.
Bi
LSTM
can
pe
rfo
rm
bo
th
directi
onal
scans,
al
lo
wing
sim
ultaneou
s
acce
ss
to
both
con
te
xts
in
f
orwa
rd
a
nd
bac
kwa
r
d
directi
ons.
T
he
re
f
or
e
,
Bi
LS
TM
can
s
olv
e
seq
uen
ce
m
od
el
ta
sk
s
with
bette
r
perfor
m
ance
than
L
STM
[
9].
Ba
sed
on
the
s
tud
y
of
thes
e
r
efere
nces,
t
his
j
ou
rn
al
will
determ
ine
wh
et
he
r
the
Bi
LST
M
m
od
el
will
perform
bette
r
tha
n
t
he LSTM
m
od
el
in use i
n NLP.
The
discu
ssio
n
co
nducted
on
seve
ral
c
hatb
ot
bac
kgr
ound
s
in
dicat
es
tha
t
consum
ers’
pro
blem
s,
in
gen
e
ral,
ca
n
be
presente
d
th
r
ough
se
ver
al
r
ecorde
d
qu
est
i
on
s
that
relat
e
to
va
rio
us
c
on
strai
nts,
s
uc
h
a
s
data
stora
ge
an
d
li
m
it
ed
custom
e
r
serv
ic
e
ho
ur
s
.
In
orde
r
to
prov
i
de
ans
wer
s
giv
e
n
by
consu
m
ers,
a
pr
ogra
m
me
is
need
e
d
to
op
t
i
m
iz
e
the
resul
ts
of
these
s
erv
ic
es.
In
c
onnecti
on
with
this,
the
m
o
deling
t
heory
will
be
discusse
d
in
th
is
j
our
nal.
It
is
us
ed
as
the
ba
sis
wh
e
n
chat
bots
are
de
plo
y
ed
in
quest
io
n
and
a
nswer
sy
stem
s
that
us
e
Pyt
h
on
pro
gr
am
me
s
with
LSTM
an
d
Bi
LSTM
as
m
od
el
s.
Then
,
it
is
exp
ect
ed
that
from
this
researc
h
will
be
seen
a com
par
ison
b
e
tween
the
se
nt
ence
res
pons
e g
ene
rated b
y
the
chat
bo
t wit
h
the
LS
TM
m
od
el
a
nd
the
Bi
LSTM
m
od
el
with
the
sentence
respon
s
e
in
the
data
set
.
The
s
olu
t
ion
m
et
ho
dolo
gy
that
will
be
us
e
d
i
n
this
researc
h
is
m
easur
in
gth
e
accuracy
of
L
STM
and
Bi
L
STM
m
od
el
s
by
app
ly
in
g
c
hatb
ot
pro
gr
a
m
m
e.
In
m
or
e
detai
l,
w
e
will
sta
rt
f
rom
un
der
sta
nd
i
ng
the
bac
kground
of
the
im
po
rta
nce
of
t
he
ro
le
of
the
c
ha
tbo
t
i
n
t
he
qu
est
i
on
a
nd
a
ns
we
r
syst
e
m
,
deter
m
ini
ng
the
ste
ps
f
or
m
aking
a
chatb
ot,
ap
plyi
ng
var
io
us
m
od
el
s
,
m
et
ho
ds,
ap
pl
yi
ng
data
into
the
pro
gram
me,
w
rite
res
ults
and
disc
us
sio
ns
,
t
o
m
ake
con
cl
us
io
ns
,
tha
t
have
been
desc
ribe
d
in
m
or
e
detai
l
from
the
session
1
to
6.
T
he
m
ajo
r
c
ontrib
ut
ion
o
f
t
his
pa
pe
r
is
to d
et
erm
i
ne
th
e
m
os
t
eff
ect
ive
m
od
el
that
can
be
ap
plied
to
t
he
chat
bo
t
pro
gr
am
m
e
based
on
t
he
com
par
i
so
n
of
the
acc
uracy
resu
lt
s
of the t
wo m
od
el
s.
2.
RESEA
R
CH MET
HO
D
As
the
va
rio
us
fo
rm
s
of
chat
bo
ts
inc
reasin
g
ly
integrate
the
desig
n
of
AI
m
echan
ism
s
(s
uch
as
gam
e
theo
ry, d
at
a m
i
ning and
op
ti
m
isa
ti
on
tech
niques)
, th
ey
co
m
ply wit
h
these
netw
orks
’
r
ules an
d dynam
ic
s.
This
form
can
be
chara
ct
erised
by
real
m
ulti
-
act
or
-
base
d
co
nv
e
rsati
on
s
that
re
quire
te
chn
ic
al
res
o
urces,
sp
eci
al
ise
dkno
wled
ge
an
d
c
om
m
un
ic
at
ion
s
kill
s
to
m
a
intai
n
onli
ne
inter
act
ion
s
[
10]
.
Su
m
m
arized
by
the
acronym
AI
,
this
is
a
sci
ence
that
fo
cuses
on
ha
nd
li
ng
the
pro
du
ct
io
n
of
hu
m
an
knowle
dg
e
,
an
d
can
offe
r
to
the
m
achine
the
abili
ty
to
i
m
i
ta
te
hu
m
an
reasonin
g
an
d
intel
li
gen
ce
[1
1].
AI
te
chnolo
gy
can
pro
vid
e
i
m
pr
ovem
ents
to
co
nv
e
rsati
ons
an
d
c
ollaborat
e
betwee
n
hum
ans
and
m
a
chines
.
This
te
chnolo
gy
can
be
use
d
to cr
eat
e
bette
r
interact
io
ns
be
tween
hum
ans
and m
achines [12
]
.
The
L
STM
m
od
el
,
th
e
Bi
LST
M
m
o
del
and
s
ever
al
pairs
of
par
am
et
ers,
it
is
al
so
t
he
gr
ee
dy
m
et
hod
can
be
us
e
d
in
buil
ding
a
c
hatb
ot
pro
gr
a
m
me
by
us
ing
a
set
of
sente
nces
de
rive
d
f
ro
m
the
data
set
.
The
chatb
ot
pro
gr
a
m
me
is
ru
n
ba
sed
on
input
in
the
form
of
com
m
and
s
f
rom
the
us
er,
w
her
e
th
e
re
su
lt
s
of
the
pro
gr
am
me
ar
e
a
colle
ct
ion
of
se
ntences
con
ta
ini
ng
in
form
ation
that
m
at
ches
us
er
inp
ut
base
d
on
th
e
releva
nce
of questi
ons a
nd a
ns
we
rs.
2.1.
S
teps
in
mak
in
g a c
hat
bo
t
Id
e
ntifyi
ng
th
e
identit
y
of
the
data,
in
putt
ing
the
data
ab
ou
t
a
nswe
r
an
d
quest
io
n,
us
in
g
t
he
pro
gr
am
me
of
chatb
ot
an
d
then
e
valuati
ng
the
outp
ut
ar
e
the
4
ste
ps
t
hat
m
us
t
be
do
ne
w
hen
c
rea
ti
ng
a
chatb
ot.
First,
t
he
identit
y
of
the
data,
c
orne
ll
m
ov
ie
dialo
gu
e
c
orp
us
is
us
e
d
as
data,
it
con
ta
in
s
a
col
le
ct
ion
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
Meas
ur
in
g
t
he accu
ra
cy
o
f
LSTM a
nd BiL
STM m
odel
s i
n
t
he
app
li
cati
on of
a
rti
fi
ci
al…
(
Prasnu
r
za
ki
Ank
i
)
199
of
data
in
the
f
or
m
of
a
corp
us
in
wh
ic
h
it
i
nclu
des
a
vast
colle
ct
ion
of
fi
ct
ion
al
co
nv
e
r
sat
ion
s
ric
h
in
m
eta
data
extracte
d
from
the
fil
m
script
[
13
]
.
T
he
data
us
ed
is
from
20
18
in
t
he
f
or
m
of
te
xt
data.
Seco
nd,
in
the
qu
e
sti
on
an
d
a
ns
we
r
syst
e
m
data
in
put
co
nt
ai
ns
se
n
te
nces
that
c
om
e
fr
om
the
fil
m
dialogue,
the
n
th
e
data
functi
ons
as
use
r
input
that
is
entered
into
t
he
pro
gr
am
me
,
then
after
that
it
is
execu
te
d
into
input
fro
m
th
e
pro
gr
am
me
us
er.
T
he
t
hir
d
i
s
the
c
hatb
ot
pro
gr
am
m
e
dev
el
opm
ent.
Four
t
hly,
the
fo
ll
ow
i
ng
co
ns
i
de
rati
on
s
need
to
be
ta
ke
n
to
pr
e
pa
re
thechat
bo
t
pr
ogram
m
e.
A
sequ
ence
-
to
-
se
qu
e
nce
translat
io
n
will
co
m
e
in
sever
al
op
ti
onss
.
Sele
c
ti
ng
the
L
STM
m
od
el
will
ge
ner
at
e
the
m
os
t
accurate
chat
bo
t
r
esp
onse
in
the
e
nd.
Th
e
final
ste
p
in
the
outp
ut ev
al
uatio
nis
to d
et
e
rm
ine w
hethe
r
the
m
od
el
can pro
vid
e
accurate
res
ults o
r no
t.
2.2.
LST
M m
od
el
The
LSTM
m
od
el
net
wor
k
is
known
as
a
m
od
el
that
has
had
in
flue
nce
in
the
past
an
d
sho
ws
the
abili
ty
to
le
arn
from
sequ
entia
l
data
[14].
T
he
im
ple
m
entat
ion
of
enc
ode
r
i
n
this
m
od
el
will
be
co
ntained
i
n
the last
h
i
dd
e
n st
at
e
m
ent o
f
t
he
LSTM
[4
]
.
+
1
=
(
(
ℎ
→
)
ℎ
+
(
→
)
+
1
+
)
for
get gate
(1)
+
1
=
(
(
ℎ
→
)
ℎ
+
(
→
)
+
1
+
)
I
nput
ga
te
(2)
̃
+
1
=
t
a
nh
(
(
ℎ
→
)
ℎ
+
(
→
)
+
1
)
update ca
ndid
at
e
(3)
+
1
=
+
1
ʘ
+
+
1
ʘ
̃
+
1
m
e
m
or
y ce
ll
u
pd
at
e
(4)
+
1
=
(
(
ℎ
→
)
ℎ
+
(
→
)
+
1
+
O
utput
gate
(5)
ℎ
+
1
=
+
1
ʘ
tanh
(
+
1
)
O
utput
(6)
The
operato
r
ʘ
is
incl
ud
e
d
i
n
the
el
em
enta
l
product.
The
resu
lt
of
a
ddin
g
hidden
sta
te
s
that
can
be
i
m
ple
m
ented
in
ℎ
with
.
as
a
represe
ntati
on
of
m
e
m
or
y
cells
is
the
def
i
niti
on
of
th
e
LST
M
m
od
el
.
The
gate
f
or
m
is
th
e
value
of
t
he
m
e
m
or
y
cel
ls
pr
ese
nt
at
eac
h
m
th
tim
e
wh
ic
h
co
ns
ist
s
of
the
s
um
of
th
e
tw
o
qu
al
it
ie
s:
−
1
is
the
value
of
t
he
pr
e
vious
m
e
m
or
y
cel
l,
w
hile
is
the
value
of
the
m
e
m
or
y
c
el
l
after
it
has
change
d,
afte
r
bein
g
cal
culat
ed
from
the
pre
vious
in
pu
t
in
the
c
urre
nt
for
m
and
t
he
pre
vious
hi
dd
e
n
s
ta
te
in
ℎ
−
1
fo
rm
.
Furthe
rm
or
e,
ℎ
is
cal
culat
ed
from
the
cel
l
m
e
m
or
y,
wh
e
re
t
he
non
-
li
near
f
unct
io
n
path
duri
ng
the
upda
te
is
not
passe
d
by
th
e
m
e
m
or
y
cel
l,
so
in
f
or
m
at
io
n
can
be
w
ork
ed
on
over
a
re
m
ote
netwo
r
[
4].
The
pr
ece
ding
hidd
en
sta
te
will
be
determ
ined
by
each
gate
con
t
ro
ll
ed
by
a
weig
hted
ve
ct
or
(e.
g.
(
ℎ
→
)
)
an
d
input
c
urre
nt
(
e.g
.
(
→
)
),
pl
us
a
ve
ct
or
off
set
(e.
g.
).
The
L
ST
M
sta
tus
afte
r
read
i
ng
to
ke
ns
is
re
pr
ese
nte
d
(
ℎ
,
)
in
an
ope
rati
on
that
can
in
form
ally
be
s
um
m
arized
as
(
ℎ
,
)
=
LSTM
〖
(
,
(
ℎ
−
1
,
−
1
)
)
.
And
the
res
ults
ob
ta
in
ed
tha
t
LSTM
can
ou
t
perform
stan
da
r
d
arti
fici
al
neural
netw
orks
i
n
a
va
riet
y
of
pro
blem
s
disp
la
ye
d
in
squa
re
-
sh
a
pe
d
gates
with
dott
ed
e
dges.
the
ne
xt
word
w
m
+
1
is
stripp
e
d
us
in
g
h
m
existi
ng
i
n
the
LSTM
la
ngua
ge
m
od
el
.
LS
TM
ou
t
perfor
m
s
sta
nd
ar
d
r
ecurrent
neura
l
networks
in
var
i
ou
s
pro
blem
s,
su
ch
as lan
guage
m
od
el
li
ng
pro
blem
s [
4].
On
e
of
pa
rall
el
co
m
pu
te
riz
ed
m
od
el
s
is
the
LSTM
m
od
el
.
Parall
el
co
m
pu
ti
ng
i
s
a
ty
pe
of
com
pu
ta
ti
on
in
w
hich
var
i
ous
process
cal
cul
at
ion
s
ca
n
be
c
arr
ie
d
out
sim
ultaneousl
y,
w
hi
le
the
ap
plica
ti
on
of
par
al
le
l
com
puti
ng
can
r
un
al
gorithm
m
or
e
qu
ic
kly
in
the
app
ea
ra
nce
of
the
m
od
el
us
e
d
in
this
st
ud
y
[15
]
,
[
16]
. Base
d on
that st
udy,
we c
hoos
e
p
a
rall
el
co
m
pu
ti
ng m
od
el
s to
r
esea
rc
hed m
or
e d
ee
pl
y.
2.3.
Bi
LST
M
mo
d
el
The
Bi
LSTM
m
od
el
is
a
m
o
del
that
com
bin
es
the
a
dvant
ages
of
t
he
Bi
RNN
m
od
el
a
nd
t
he
LS
TM
m
od
el
[1
7].
T
he
Bi
LSTM
m
od
el
is
us
e
d
to
pro
pa
gate
the
use
of
f
orwa
r
d
and
r
eve
rse
dire
ct
ion
s.
T
he
Bi
LST
M
m
od
el
is
a
two
-
way
netw
ork
us
ed
to
sto
re
fu
t
ur
e
data
a
nd
past
data,
w
hich
is
m
or
e
eff
ect
ive
in
the
LSTM
m
od
el
[18].
I
n
the
feat
ur
e
-
ba
s
ed
m
od
el
,
trai
t
s
relat
ed
t
o
s
ha
pe
knowle
dge
are
processe
d
by
feat
ur
e
suff
ixes
in
the
ne
ur
al
net
work.
Em
bed
di
ng
s
a
re
a
te
ch
nique
us
e
d
to
handle
the
s
pa
rse
m
at
rix
of
the
ba
g
of
wor
ds
.
One
app
li
cat
io
n
of
featur
e
s
uffixe
s
in
neural
net
works
is
that
t
hey
can
be
in
s
erted
by
co
ns
t
ru
ct
in
g
the
in
vi
sible
e
m
bed
di
ngs
of
wo
r
ds
f
ro
m
their
sp
el
li
ng
or
m
or
phology.
O
ne
way
to
do
t
his
is
to
incorp
or
at
e
ad
diti
on
a
l
two
-
way
RNN
la
ye
rs,
one
of
w
hi
ch
is
fo
r
eac
h
wo
r
d
in
the
vo
ca
bula
ry.
T
he
Bi
LSTM
m
od
el
is
one
of
m
any
par
al
le
l
c
om
pu
ta
ti
on
s.
The
fir
st
ste
p
is
to
e
nc
od
e
w
(
)
da
n
the
w
(
)
qu
e
ry
us
in
g
tw
o
L
STMs.
T
his
pr
ocess
is
known
as Bidi
r
ect
ion
al
LS
TM (Bi
LSTM)
ℎ
(
)
=
(
w
(
)
;
θ
(
)
)
(7)
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
S
ci
,
Vo
l.
23
, N
o.
1
,
Ju
ly
2021
:
1
9
7
-
2
0
5
200
ℎ
(
)
=
(
(
)
;
θ
(
)
)
(8)
The
quest
io
ns
are
re
pr
ese
nte
d
by
the
vecto
r
u,
ve
rtic
al
ly
c
om
bin
ing
fin
al
sta
te
s
fr
om
le
ft
to
rig
ht,
and are
r
e
pr
ese
nted by m
at
ching the e
ndin
g
s
ta
te
v
erti
cal
ly
f
ro
m
left t
o
ri
ght
.
=
[
ℎ
(
)
(
)
⃗
⃗
⃗
⃗
⃗
⃗
⃗
⃗
⃗
⃗
⃗
⃗
⃗
⃗
⃗
;
ℎ
(
)
0
⃖
⃗
⃗
⃗
⃗
⃗
⃗
⃗
⃗
⃗
]
(9)
Vecto
r
(
(
)
)
is
the
res
ult
of
a
pply
ing
the
ve
ct
or
u
with
e
qu
at
io
n
ℎ
=
(
,
ℎ
−
1
)
,
=
1
,
2
,
…
,
(
based
on
[
4])
w
hich
has
bee
n
tra
nspose
d,
is
a
weig
ht
m
atr
ix
with
i
nd
e
x
α
,
ℎ
(
)
is
the
res
ult
of
i
m
ple
m
enting
the
hidden
sta
te
with
(
8)
,
a
nd
vector
(
α
̃
)
̃
i
s
a
represe
ntati
on
of
w
hat
is
ex
pected
a
nd
is
cal
culat
ed by,
α
̃
=
(
(
)
)
ℎ
(
)
(10)
α
=
SoftMax
(
α
̃
)
(11)
=
∑
α
ℎ
(
)
=
1
(12)
In
(
11)
,
the v
ec
tor
α
is
the
res
ult
of
the Soft
Ma
x
f
un
ct
io
n
of
α
̃
.
In
(12),
the
se
vecto
rs
can b
e
ar
range
d
equ
al
t
o
the
c
orres
pon
ding
el
e
m
ent
in
h
(
p)
,
a
ssu
m
ing
that
the
ca
nd
i
date’s
answer
(v
ect
or
o
)
is
t
he
s
pan
of
t
he
or
i
gin
al
text.
T
he
sc
or
e
of eac
h
ca
nd
i
date f
or
an
s
wer
a
is ca
lc
ulate
d by the
pro
du
ct
i
n,
̂
=
a
ma
x
.
(13)
2.4.
Gree
dy
meth
od
The
ne
xt
ste
p
afte
r
c
hoos
i
ng
a
m
od
el
is
t
o
determ
ine
the
pro
gram
me
m
et
hod.
The
gr
e
edy
m
et
ho
d
was
ch
os
e
n
as
the p
r
ogram
me
m
et
ho
d
that i
s
u
sed beca
us
e i
t i
s the i
m
ple
m
entat
ion
of
t
he
LSTM m
od
el
, w
he
n
it
is
run,
the
pr
ogram
me
can
proces
s
data
in
a
faster
ti
m
e,
s
o
that
it
can
i
nc
rease
t
he
acc
ur
acy
of
the
se
le
ct
ed
m
od
e
l
[1
9].
W
hile
the
G
reedy
al
go
rithm
is
well
unde
rstoo
d
to
be
able
to
produce
reas
onable
est
im
at
e
s
for
a
wide
cl
ass
of
functi
ons,
it
can
be
see
n
tha
t
it
per
form
s
m
uch
bette
r
than
one
m
igh
t
exp
ect
f
r
om
a
gr
ee
dy
al
gorit
hm
[2
0]
.
Ba
sed
on
[
18]
,
relat
ively
i
ncr
easi
ng
ca
ndidate
so
l
utions
by
try
in
g
to
a
ppr
oach
the
optim
al
so
luti
on
is
G
re
edy
'
s
al
go
rith
m
.
The
gr
eedy
m
et
ho
d
is
a
n
im
ple
m
entat
ion
of
t
he
LSTM
and
Bi
LS
TM
m
od
el
s,
wh
ic
h
is
at
ru
n
tim
e,
the
tim
e
us
e
d
in
proces
sing
dat
a
is
fas
te
r
and
it
can
increase
the
acc
ur
acy
of
the
sel
ect
ed
m
od
el
.
2.5
.
Se
q2
s
eq
mod
el
In
Fig
ur
e
1,
th
e
i
m
po
rta
nt
thi
ng
s
that
m
us
t
be
unde
rstood
about
the
se
q2seq
m
od
el
im
ple
m
entat
ion
are
presente
d
.
The
se
q2
se
q
m
od
el
f
unct
io
ns
to
ge
ner
at
e
va
r
iou
s r
esp
on
ses to
us
e
r
in
pu
t,
s
o
it
can
im
ple
m
ent
a
qu
e
sti
on
a
nd
a
ns
we
r
syst
em
,
wh
e
re
the
Pyt
hon
-
based
Jupyt
er
N
oteb
ook
S
of
t
war
e
is
c
ho
sen
as
a
pr
ogr
a
m
me
that ca
n vie
w p
rogr
am
me
inpu
t and o
utput
[19].
F
igure
1
.
Ge
ne
rati
on of
neura
l respo
ns
es i
n dial
ogue
at
the
e
ncode
r deco
de
r
m
od
el
[
19
]
3.
DA
T
A
I
MPL
EMENT
ATION
I
N
T
HE P
YTHO
N
P
ROGR
AMME
Syst
e
m
at
ic
a
ll
y,
the
program
me
plann
i
ng
i
s
com
piled
in
i
m
ple
m
entin
g
data
f
ro
m
the
Jupyt
er
No
te
book
software
in
the
Py
thon
pro
gr
am
me
as
fo
ll
ow
s:
Firstl
y,
cho
ose
the
appropr
i
at
e
so
ftwa
re,
they
are
choosi
ng
a
pro
gr
am
me
that ca
n
pr
ocess dat
a
well
, h
a
ving
pe
rfor
m
ance in
s
of
t
war
e
data pr
ocessin
g
a
nd havin
g
the
avail
abili
ty
of
s
up
porting
at
tribu
te
s
that
needed
i
n
cre
at
ing
the
pro
gram
m
e.
Secondly
,
the
progr
a
m
me
perform
ance
is
influ
e
nce
d
by
the
s
el
ect
ion
of
m
od
el
that
fo
ll
ows
the
ch
aracte
risti
cs
of
the
data,
s
o
whe
n
choosi
ng
a
m
od
el
and s
uppor
ti
ng
att
rib
utes,
it
is n
ecessa
ry t
o pay at
te
ntion t
o
it
. The outc
om
e o
f
a pr
ogr
a
m
me
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
Meas
ur
in
g
t
he accu
ra
cy
o
f
LSTM a
nd BiL
STM m
odel
s i
n
t
he
app
li
cati
on of
a
rti
fi
ci
al…
(
Prasnu
r
za
ki
Ank
i
)
201
about
it
s
abili
t
y
to
determ
in
e
a
hig
h
or
lo
w
le
vel
of
accuracy
are
the
m
a
j
or
fact
or
s
in
choosin
g
whi
c
h
pro
gr
am
me
to
us
e
d.
S
o,
seei
ng
that
the
LST
M
Mod
el
was
chosen
a
s
the
m
od
el
app
li
ed
to
this
pr
ogra
m
me
,
it
is
becau
se
it
is
in
accord
a
nc
e
with
the
m
od
el
sel
ection
requirem
ents.
In
a
dd
it
io
n
to
the
m
ai
n
factors,
su
pp
or
ti
ng f
act
or
s
a
re also
the
d
et
erm
ining
f
a
ct
or
s wheth
er th
e proces
sin
g
of
data is i
n
li
ne
with the cr
it
eria i
n
qu
e
sti
on.
It
will
m
ake
the
dat
a
ver
i
fiable.
T
he
in
pu
t
se
nte
nce
that
has
be
en
pr
ocesse
d
by
the
seq
2s
e
q
m
od
el
with
oth
e
r
pro
gr
am
me
structur
es
an
d
m
odel
s,
the
n
the
outp
ut
se
ntence
will
be
iss
ue
d,
it
is
as
a
c
hatb
ot
pro
gr
am
me
resu
lt
.
F
inall
y,
es
ta
blish
the
m
e
thod
of
pro
gra
m
me
evaluati
on.
Seve
ral
c
hoic
es
of
ev
al
ua
ti
on
m
et
ho
ds
t
hat
c
an
be
use
d
i
nc
lud
e:
los
s,
acc
ur
acy
,
val_l
os
s
an
d
val
_acc
uracy
,
so
t
hat
from
these
cases
[19]
te
xt classi
ficat
ion eval
uatio
n m
et
ho
ds suc
h as sim
ple b
inar
y detec
ti
on tasks ca
n be c
onsidere
d
it
s
us
e.
Sp
am
detect
ion
f
or
exam
ple,
it
assigns
a
po
s
it
ive
or
ne
gativ
e
la
bel
to
sp
am
accor
ding
to
the
cat
eg
or
y
an
em
ai
l
do
cu
m
ent,
it
shou
l
d
be
a
ble
to
de
fi
ne,
a
nd
al
so
it
m
us
t
be
able
to
identify
a
n
it
e
m
as
sp
a
m
or
not.
As
pr
ese
nted
in
(
14)
:
=
+
+
+
+
(14)
Ba
sed
on
the
t
heory
[
20]
,
the
re
are
seve
ral
t
hings
that
nee
d
to
be
prepa
r
ed
f
or
the
val
ue
of
losses
,
they
inclu
d
e
x
for ob
s
er
vation an
d
sta
te
d i
s for t
he fu
nctio
n of l
os
ses
that is
expres
sed
b
y
(
15)
:
(
ŷ
,
)
=
ℎ
ℎ
(15)
H
e
r
e
i
s
a
c
a
l
c
u
l
a
t
i
o
n
t
o
d
e
t
e
r
m
i
ne
t
h
e
c
l
o
s
e
n
e
s
s
o
f
t
h
e
o
u
t
p
u
t
c
l
a
s
s
i
f
i
c
a
t
i
o
n
.
H
e
r
e
i
s
a
c
a
l
c
ul
a
t
i
o
n
t
o
d
e
t
e
r
m
i
n
e
t
h
e
c
l
o
s
e
n
e
s
s
o
f
t
h
e
o
u
t
p
u
t
c
l
a
s
s
i
f
i
c
a
t
i
o
n
ŷ
=
(
.
+
)
)
t
o
t
h
e
a
c
t
u
a
l
o
u
t
p
u
t
(
y
,
w
h
i
c
h
i
s
0
o
r
1
)
.
The
trai
ning pr
ocess
is used
t
o
cal
culat
e the er
r
or
r
at
e that i
s d
erive
d
in calc
ulate
d
m
od
el
, it
is used
to
observ
e
the
validat
io
n
set
loss
functi
on.
It
is
as
def
in
ed
in
[21]
ab
out
the
m
eaning
of
loss
val
ue
validat
io
n.
The
m
od
el
that
has
been
s
el
ect
ed
an
d
trai
ned,
it
is
then
e
valuated
f
or
it
s
eff
ect
ive
ness
as
a
cl
assifi
cat
ion
ta
s
k.
T
his
c
an
be
done by cal
cul
at
in
g
the
perce
ntage o
f
sam
ples that
hav
e
b
e
en
cl
assifi
e
d,
a
s foll
ow
s:
=
(16)
The
e
qu
at
io
n
(
16)
s
hows
the
m
isc
la
ssific
at
i
on
ca
n
be
cal
culat
ed
an
d
e
quipp
e
d
with
a
c
la
ssific
at
io
n
rate,
wh
e
re is t
he
cal
culat
io
n of t
he
m
od
el
e
r
ror rat
e as
fo
ll
ows:
=
(17)
T
he
l
os
s
f
un
ct
i
on
cal
culat
ed
on
the
prede
fi
ned
m
od
el
vali
dation
set
will
resu
lt
i
n
the
pe
rcen
ta
ge,
a
s
was
the
case
of
[
22]
,
thr
ough
the
gr
a
ph
pr
ese
nted
in
F
igure
2,
we
ca
n
com
par
e
the
losses
the
dataset
of
t
rainin
g
an
d
th
e
validat
io
n
set
of
v
al
_l
os
s.
From
the
gr
ap
h,
ob
ta
ine
d
t
hat
the
res
ult
of
a
va
li
dation
los
s m
ay
be
higher
or lo
we
r
tha
n
t
he
los
s
value o
f
the
tra
ining data
set
, so t
his con
diti
on
is cal
le
d u
nd
erf
it
te
d or o
ve
r
fitt
ed
.
Accuracy
is
t
he
n
c
hosen
f
or
the
e
valuati
on
m
et
ho
d
of
ch
at
bo
t
pro
gr
am
m
e,
after
c
ompari
ng
each
pro
gr
am
m
e’s
e
valutio
n
m
et
ho
d
that
accor
dingly
.
Af
te
r
getti
ng
the
s
uitabil
it
y
value
betwe
en
the
f
or
m
of
two
sentences
,
it
will
be
us
ed
as
t
he
accu
racy
va
lue
that
will
be
us
ed
as
a
chatbo
t
pro
gram
me
evaluati
on
m
et
hod.
The
sel
ect
ed
m
od
el
will
the
n
be
a
ppli
ed
to
fin
d
the
dif
f
eren
ces
of
w
ords
l
ocated
in
sentences
duri
ng
t
he
trai
ning
per
i
od [23
]
.
Figure
2. Com
par
is
on ch
a
rt
of e
po
c
h
t
o
los
s
and ep
oc
h
to
a
ccur
acy
[22]
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
S
ci
,
Vo
l.
23
, N
o.
1
,
Ju
ly
2021
:
1
9
7
-
2
0
5
202
4.
APPL
YING
DA
T
A
I
MPL
EMENT
ATION
I
N
T
HE P
YTHO
N
P
ROGR
AMME
Accuracy
wa
s
chosen
a
s
th
e
m
et
ho
d
of
pro
gr
am
m
e
evaluati
on
i
n
ac
corda
nce
with
the
sel
ect
ed
chatb
ot
pro
gra
m
m
e
after
com
par
i
ng
t
he
e
valuati
on
m
et
ho
ds
of
each
pr
ogr
a
m
m
e.
In
or
de
r
to
st
ud
y
t
he
ne
ed
to
i
m
ple
m
ent
a
c
hatb
ot
in
the
quest
io
n
an
d
a
nswer
syst
em
,
t
his
sect
ion
will
exp
la
in.
The
so
luti
on
m
et
ho
do
l
ogy
that
will
be
us
ed
in
this
rese
arch
sta
rts
fro
m
un
der
sta
nding
the
bac
kgr
ound
of
the
im
po
rta
nce
of
the
ro
le
of
the
chat
bo
t
i
n
the
qu
est
io
n
a
nd
a
ns
we
r
syst
e
m
,
deter
m
ining
the
ste
ps
for
m
aking
a
chat
bo
t,
ap
plyi
ng
va
rio
us
m
od
el
s,
m
et
ho
ds
,
a
pp
ly
in
g
da
ta
into
the
pr
ogram
m
e,
sel
e
ct
ing
eval
uatio
n
m
et
ho
ds
that
hav
e
been
des
cribe
d
in m
or
e d
et
ai
l f
ro
m
the sessio
n 1 to 3
.
4.1.
Descripti
on
of the pr
obl
em
Ma
ny
qu
est
io
ns
will
certai
nly
be
asked
by
con
s
um
ers
base
d
on
certai
n
da
ta
in
the
face
of
dynam
ic
s
in
this
m
od
e
rn
era.
I
n
orde
r
to
pr
ov
i
de
eas
y
acce
ss
a
nd
c
onve
nient
oper
at
ion
al
ti
m
e,
chatb
ots
need
t
o
be
create
d.
So,
th
e
qu
e
sti
on
a
nd
answ
e
r
se
rv
ic
e
betwee
n
hu
m
ans
an
d
chat
bo
t
pro
gr
am
m
es
(m
achines)
can
r
un.
This is as
an i
m
ple
m
entat
ion
of the
quasti
on &
ans
we
r
sys
tem
d
at
a.
4.2.
Pr
ogram
me ma
king
In Fig
ur
e
3,
we
show h
ow to
s
te
ps
by steps
in
progr
am
m
e
m
akin
g
to
buil
d t
he
c
hatb
ot.
Figure
3. Pro
gra
m
m
e
m
aking
flo
wch
a
rt
5.
RESU
LT
S
AND DI
SCUS
S
ION
S
Fr
om
the
chatbo
t
pro
gr
am
m
e
that
has
been
c
reated,
6
file
s
are
gen
e
rated
,
3
f
il
es
fr
om
the
i
m
ple
m
entat
io
n
of
the
L
STM
m
od
el
into
the
chatb
ot
pro
gr
a
m
m
e
and
3
file
s
from
the
Bi
L
STM
m
od
el
into
the
chatb
ot
pro
gra
m
m
e.
In
Table
1,
t
he
LSTM
m
od
el
is
us
e
d
to
te
st
the
pa
ram
et
er
pairs.
Seco
nd
ly
,
e
poch
(
4
diff
e
re
nt
a
m
ounts,
20,
30,
40,
50)
will
be
te
ste
d
with
diff
ere
nt
nu
m
be
rs
of
e
po
c
hs
,
adjustin
g
f
or
oth
e
r
par
am
et
er
pairs,
to
determ
ine
wh
ic
h
pa
ram
e
te
r
pair
is
the
m
os
t
accurate
in
the
two
sel
ect
ed
m
od
el
s
[24].
I
n
Table
1,
the
L
STM
m
od
el
is
us
e
d
to
te
st
the
pa
ram
et
er
pair
s
(F
il
e
1,
Fil
e
2,
an
d
Fil
e
3),
and
Bi
LSTM
m
od
el
(F
il
e
4,
Fil
e
5,
and
Fil
e
6).
The
pur
p
os
e
of
ha
ving
dif
fere
ntv
al
ue
s
in
the
pa
ram
et
er
i
s
to
produce
bette
r
accuracy
an
d
c
om
par
e
the
dif
fer
e
nces
bet
we
en
the
num
ber
of
di
ff
e
ren
ce
s
in
the
sam
e
p
aram
et
er,
wh
et
her
it
can
hav
e
b
et
te
r
outp
ut r
es
ults
than
t
he
te
st
re
su
lt
s on
t
he par
a
m
et
er p
ai
rs
i
n t
he
e
xp
e
rim
ent.
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
Meas
ur
in
g
t
he accu
ra
cy
o
f
LSTM a
nd BiL
STM m
odel
s i
n
t
he
app
li
cati
on of
a
rti
fi
ci
al…
(
Prasnu
r
za
ki
Ank
i
)
203
Table
1.
Data
t
est
ed
on LST
M an
d
Bi
LST
M
chatb
ot
Para
m
eter
Pai
r
File 1, File
4
File 2, File
5
File 3, File
6
size_
lay
er
128
256
512
n
u
m
_
la
y
ers
2
2
2
e
m
b
ed
d
ed
_
size
64
128
256
learnin
g
_
rate
0
.00
1
,
0
.00
1
5
0
.00
1
,
0
.00
1
5
0
.00
1
,
0
.00
1
5
b
atch
_
size
8
16
32
ep
o
ch
2
0
,30
,40
,5
0
2
0
,30
,40
,5
0
2
0
,30
,40
,5
0
Accor
ding
to
the
num
ber
of
pa
ram
et
ers
in
L
STM
Chatbo
t, in
3
diff
e
re
nt
file
s
(w
it
h
each
value
of
the
siz
e_lay
er
par
a
m
et
er
d
iffer
e
nt
in
each
file
,
nam
el
y:
12
8,
25
6,
51
2)
will
be
te
ste
d
fo
r
a
to
ta
l
of
8
tria
ls
each
,
with
4 param
et
ers wit
h
the
nu
m
ber
o
f value
s
in the
sam
e p
aram
et
ers
in each fil
e.
First,
the
siz
e
_l
ay
er
(in
the
f
or
m
of
the
nu
m
ber
of
la
ye
r
siz
es
that
will
be
ap
plie
d
t
o
the
pr
ogram
m
e
,
in
gen
e
ral,
m
ulti
ples
of
2
a
re
us
e
d
su
c
h
as
16,
32
an
d
64,
base
d
on
[
25
]
)
i
n
th
e
program
m
e
t
o
be
discuss
e
d
is
an
a
pp
l
ie
d si
ze
_lay
erw
it
h t
he values
12
8,
25
6
a
nd
51
2.
Seco
nd
,
N
um
_lay
ers
(in
the
f
or
m
of
n
nu
m
ber
of
la
ye
rs
t
ha
t
will
be
ap
pl
ie
d
t
o
the
progra
m
m
e
[2
6
])
in
t
he
pro
gr
am
m
e to b
e
discusse
d
is an a
ppli
ed
siz
e
_lay
er
with a
va
lue of
2
.
Thir
d,
the
em
bed
de
d_siz
e
(i
n
the
f
or
m
of
the
nu
m
ber
of
siz
es
of
t
he
em
bed
de
d
vecto
r
th
at
will
be
ap
plied
to
the
pro
gram
m
e
)
;
in
ge
ne
ral,
m
ulti
ples
of
2
a
re
us
ed
s
uc
h
as
8,
16,
32
and
64
[
26
]
i
n
the
pr
ogram
m
e
to
be discus
sed
is
an
a
ppli
ed
siz
e_lay
er
with
va
lues
128, 25
6
and 51
2.
Finall
y,
the
bat
ch_si
ze
(in
t
he
form
of
a
batc
h
siz
e
that
will
be
a
pp
li
ed
)
in
the
pro
gr
am
m
e
to
be
discu
sse
d
will
aff
ect
pr
ogram
m
e
per
for
m
ance
[24].
T
her
e
fore,
i
n
thi
s
pro
gr
am
m
e,
t
he
batc
h_siz
e
will
be
te
ste
d
with
a v
al
ue
of
8, 16 a
nd 32.
Be
sides
hav
i
ng
th
e
sam
e
para
m
et
ers,
it
will
al
so
be
te
ste
d
f
or
va
rio
us
pa
ram
et
ers
with
a
num
ber
of
values
f
or
different
pa
ram
et
e
rs.
Ba
sed
on
th
e
pr
og
ram
m
e
t
est
con
duct
e
d
in
[27],
one
of
the
param
et
ers
te
ste
d
in
a
di
ff
e
re
nt
num
ber
of
val
ue
s
is
ep
och.
I
n
t
his
pr
ogram
m
e
,
m
anypar
am
eter
s
will
be
te
ste
d
with
a
nu
m
ber
of
diff
e
re
nt v
al
ue
s.
The
Fig
ur
e
4 pr
ese
nts the p
aram
et
er p
ai
r
da
ta
in
each f
il
e,
w
hic
h
is
the
be
st par
am
et
er p
ai
r
of
the
6
file
s
te
st
ed
an
d
will
not
be
rated
if
the
value
is>
1.
0,
this
co
nd
it
ion
is
cal
le
d
an
ov
er
fit
(where
if
the
trai
ning
le
vel
is
ver
y
go
od
i
n
accuracy, but
wh
e
n
te
sti
ng
t
he
resu
lt
s
are n
ot
good).
I
n
Fi
gure 4
will
be
s
el
ect
ed
wh
ic
h
has
the
best
ave
ra
ge
ac
cur
acy
,
w
hi
ch
is
a
scal
e
of
0,
0
to
1,0.
So,
f
r
om
al
l
the
te
sts
that
ha
ve
been
done
,
it
is
ob
ta
ine
d
Param
et
er
Pair
1
from
file
6
is
the
be
st
pa
r
a
m
et
er
pair
of
Bi
LSTM
chatb
ot,
wh
ic
h
is
w
it
h
a
n
aver
a
ge
accu
ra
cy
value
of
0,995
217.
Ba
se
d
on
the
re
fe
rence
,
on
res
ults
of
a
pp
ly
in
g
th
e
Bi
LSTM
m
od
el
in
do
m
ai
n
-
sp
eci
fic
Chinese
w
ord
se
gm
entat
ion
,
t
he
acc
ur
ac
y
rate
is
95.
7415%
or
0,95741
5
[
17]
.
Ba
sed
on
com
par
ison
res
ult
in
this
st
udy
and
t
he
r
esul
ts
of
t
he
a
pp
li
c
at
ion
of
t
he
Bi
LSTM
m
od
el
carr
ie
d
out
by
[17]
i
s
bette
r
the
r
esea
rch re
fer
e
nce.
Figure
4.
Para
m
et
er p
ai
r
c
omparis
on ch
a
rt
ba
sed o
n
file
dat
a res
ults
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
S
ci
,
Vo
l.
23
, N
o.
1
,
Ju
ly
2021
:
1
9
7
-
2
0
5
204
Hope
i
n
order
to
ac
hieve
m
or
e
ef
fecti
ve
r
esult,
to
get
be
tt
er
m
od
el
pe
rfor
m
ance
can
com
bin
e
the
m
od
el
s.
Other
than
that,
we
hope
to
im
pr
ove
the
Bi
LST
M
Chatbo
t
with
m
ake
m
od
ific
at
ion
s
to
the
m
od
e
l
arch
it
ect
uret
ha
t
will
resu
lt
in
bette
r
accu
rac
y
[2
8],
[
29]
.
We
will
al
so
increase
t
he
am
ount
of
data
a
naly
sed,
aim
ing
to
e
nc
oura
ge researc
he
rs
to
pr
opos
e
m
et
ho
ds t
hat
pro
duce
bette
r,
m
or
e eff
i
ci
ent
resu
lt
s
[30].
6.
CONCL
US
I
O
N
Af
te
r
a
pply
ing
the
Bi
LSTM
m
od
el
to
the
chatb
ot,
we
were
able
to
de
du
c
e
from
a
ll
the
t
est
resu
lt
s
of
the
pro
gram
me
that
ha
d
bee
n
c
onduct
ed
w
it
h
a
va
riet
y
of
di
ff
e
ren
t
para
m
et
er
pairs,
then
it
is
obta
ined
th
e
resu
lt
,
if
t
he
P
aram
et
er
Pair
1
(size_lay
e
r
51
2,
nu
m
_lay
ers
2,
em
bed
ded_size
256,
le
arn
in
g_rate
0.001
,
batch
_s
iz
e
32,
epo
c
h
20)
f
rom
Fil
e
6
is
the
best
par
am
et
er
pair
of
Bi
LST
M
Chatb
ot
with
a
n
a
verage
a
ccur
ac
y
value
of
0.9
9521
7.
F
or
f
ut
ur
e
w
ork,
t
he
researc
he
r
s
houl
d
im
pr
ov
e
la
te
st
m
od
el
,
try
ing
t
o
inc
re
ase
the
nu
m
ber
of pr
oport
io
ns
in
the
data to
be st
udie
d,
so as to
pr
oduce
b
et
te
r
r
es
earch
r
es
ults.
ACKN
OWLE
DGE
MENTS
This
resea
rch
i
s
par
tl
y
su
pport
ed
by
DRPM
researc
h
gr
a
nt
2Q2
2020
wit
h
co
ntract
nu
m
ber
NK
B
-
778/UN
2.
RST
/
HK
P.
05.
00
/
2020
from
Un
iversity
of
Ind
onesi
a.
T
he
auth
or
would
li
ke
t
o
tha
nk
the
s
uppo
rt
from
m
e
m
ber
s
of
the
Lab
orat
or
y
of
BACL
(
Bi
on
f
orat
ic
s
and
A
dv
a
nce
d
Com
pu
ti
ng
)
at
the
DS
C
(
De
pa
rtm
ent
of
Ma
them
at
ics
and
Data
S
ci
ence
)
at
the
Faculty
of
Ma
them
a
ti
cs
a
nd
Natu
ral
Sc
ie
nces,
U
ni
versi
ty
of
Ind
on
esi
a.
Ou
r
sp
eci
al
tha
nks
to E
nago (w
w
w.
e
nago.
c
om
)
for
the
E
ng
li
sh re
view of
this
pap
e
r
.
REFERE
NCE
S
[1]
M.
Nuruzz
aman
and
O.
K.
Hus
sain,
“
Intelli
Bo
t
A
Dial
ogue
-
base
d
cha
tbo
t
for
the
insuranc
e
industr
y
,
”
Know
le
dge
-
Based
Syst
ems
,
vol.
196
,
p
.
1
058
10
,
2020
,
doi
:
1
0.
1016/j.knos
y
s.
2020.
105810.
[2]
E.
Adam
opoulou
and
L.
Mous
s
ia
des,
“
Chat
bots
:
Histor
y
,
t
ec
hn
olog
y
,
and
appl
i
ca
t
ions,”
Mac
hi
ne
Learning
wit
h
Appl
ic
a
ti
on
,
vol
.
2,
p.
1
00006
,
20
20,
doi
:
10
.
1016
/j
.
m
lwa.
2020
.
10
0006.
[3]
S.
A
rsovs
ki,
et
al.
,
“
Autom
at
ic
knowledge
ext
r
a
ct
ion
of
an
y
Ch
at
bot
from
conve
rsati
on
,
”
Ex
p
er
t
Syste
ms
Wit
h
Appl
ic
a
ti
ons
,
vo
l.
137
,
p
p
.
343
-
3
48
,
2019
,
doi
:
10
.
1016/j.e
sw
a.
201
9.
07.
014
.
[4]
J.
Ei
senst
ei
n
,
“
Natur
a
l
L
angua
g
e P
roc
essing,
“
MI
T press
,
pp.
137
-
138,
2018
.
[5]
A.
C.
O.
Redd
y
and
K.
Madha
vi,
“
Hier
arc
h
y
b
ase
d
fire
f
l
y
op
tim
iz
ed
k
-
m
ea
ns
cl
uster
ing
for
c
om
ple
x
questi
on
answeri
ng,
”
Ind
onesian
Journal
of
Elec
tric
al
En
gine
ering
a
nd
C
omputer
Scienc
e
(
IJE
ECS)
,
vol.
17,
no.
1,
p.
265
,
2020,
doi
:
10.
11
591/i
jeec
s.v17
.
i
1.
pp
264
-
272.
[6]
E.
Panta
no
and
G.
Pizz
i
,
“
Forec
asti
ng
artificia
l
i
nte
lligen
ce
on
o
nli
ne
customer
a
ss
ista
nce
:
Ev
idence
from
cha
tbo
t
pat
en
ts
ana
l
y
si
s,”
Journal
o
f
Re
ta
il
ing
a
nd
Consum
er
Serv
ices
,
vo
l.
55,
p
.
1
020
96
,
2020,
do
i:
10.
1016/j.jr
e
tc
o
nser.
2020.
10209
6.
[7]
S.
J
.
and
S.
Sw
a
m
y
,
“
A
prior
ca
s
e
stud
y
of
nat
ur
a
l
la
ngu
age
proc
e
ss
ing
on
diffe
r
en
t
dom
ai
n,
”
Int
ernati
onal
Journal
of
Elec
tri
cal
and
Computer
Enginee
ring
(
IJE
CE)
,
vol.
10
,
no
.
5,
p.
4928,
2
020
,
d
oi:
10.
11591
/i
j
ece.
v10i5
.
pp4928
-
4936.
[8]
P.
Pate
l
and
A.
Tha
kkar
,
“
The
upsurge
of
dee
p
le
arn
ing
for
computer
vision
applications,
”
Int
ernati
onal
Journal
o
f
El
e
ct
rica
l
and
C
omputer
Engi
n
e
ering
(
IJE
CE)
,
v
ol.
10
,
no
.
1
,
p
.
5
47,
2020
,
doi
:
10
.
11591/i
j
ece.
v10
i1.
pp538
-
548.
[9]
W
.
Li
,
et
al
.
,
“
Bidi
rec
t
iona
l
L
STM
with
self
-
at
t
ent
ion
m
ec
ha
nism
and
m
ult
i
-
cha
nne
l
fea
t
ure
s
for
senti
m
ent
cl
assifi
ca
t
ion,”
Neurocomputi
ng
,
vol
.
381
,
p
.
64
,
2020,
doi
:
10
.
10
16/j
.
n
euc
om
.
202
0.
01.
006
.
[10]
G.
Murtar
elli,
A.
Gregor
y
,
S.
Rom
ent
i,
“
A
conve
rsati
on
-
b
ase
d
per
spec
ti
v
e
for
shaping
et
hi
ca
l
hum
an
–
m
ac
hine
int
er
ac
t
ions:
Th
e
par
t
ic
ul
ar
cha
l
lenge
of
cha
tbo
ts,
”
Journal
of
Busine
ss
Re
search
,
vol.
129
,
pp
.
92
7
-
935,
2020,
doi
:
10.
1016/j.jbusre
s.2020.
09.
018
.
[11]
S.
Berhi
l,
H.
B
e
nla
hm
ar,
N.
L
ab
ani
,
“
A
rev
ie
w
pape
r
on
artifici
al
intelligence
at
the
service
of
hum
an
resourc
es
m
ana
gement,”
Indone
sian
Journal
of
El
e
ct
ric
al
Engi
ne
ering
and
Computer
Sci
en
ce
(
IJE
ECS)
,
vol
.
18,
no.
1,
p
.
35,
2020,
doi
:
10
.
11
591/i
jeec
s.v18
.
i
1.
pp32
-
40.
[12]
S.
R.
Salkut
i
,
“
A
surve
y
of
big
dat
a
and
m
ac
h
i
ne
le
a
rning,”
In
te
rnational
Journal
of
El
e
ct
ri
ca
l
and
Computer
Engi
ne
ering
(
IJ
ECE
)
,
vol
.
10
,
n
o.
1
,
p
.
2088
-
87
08
,
2020
,
doi
:
10
.
11591/i
j
ece.
v10
i1.
pp575
-
580.
[13]
R.
Chid
ana
nd
a,
“
Cornel
l
Movie
-
Dial
ogs Corpus,
”
Kaggl
e
datas
ets
,
2018.
[14]
M.
Yaqub
,
et
al.
,
“
Modeli
ng
of
a
full
-
sca
l
e
sewage
tre
a
tment
pla
nt
to
pre
dic
t
the
n
utri
entrem
oval
e
ffic
i
ency
using
a
long
short
-
te
rm
m
emory
(LSTM
)
neur
al
ne
twork,
”
Journal
of
Wa
te
r
Proce
ss
Enginee
ring
,
vo
l.
37,
p.
3,
2020
,
do
i
:
10.
1016/j.jwpe
.
2
020.
101388.
[15]
H.
Fang
,
et
al
.
,
“
An
eff
ic
ie
nt
ra
dia
l
basis
func
t
i
ons
m
esh
def
or
m
at
ion
with
gre
e
d
y
al
gori
thm
base
d
on
rec
urre
n
ce
Chole
ske
y
de
co
m
positi
on
and
par
allel
computi
ng
Para
llel
co
m
puti
ng
and
sw
arm
int
el
l
ige
n
ce
base
d
artifi
c
ia
l
int
ellige
n
ce
m
odel
for
m
ult
i
-
ste
p
-
ahe
ad
h
y
drolo
gic
a
l
t
ime
seri
es
pre
di
ct
ion
,
”
Jo
urnal
of
Compu
tat
ional
Phy
si
cs
,
vol.
377
,
p
.
186
,
2019
,
doi
:
10
.
10
16/j
.
j
cp.
2018
.
10.
029.
[16]
J.
M.
Sadle
r,
e
t
al.
,
“
Le
v
era
g
ing
open
source
soft
ware
and
par
al
l
e
l
computing
for
m
odel
pre
dic
t
ive
cont
rol
of
urba
n
dra
ina
g
e
s
y
st
ems
using
EPA
-
S
W
MM
5,
”
Env
ir
onmental
Mode
ll
ing
&
Soft
wa
re
,
vol
.
120,
p
.
11,
2019
,
doi
:
10.
1016/j.e
nvsof
t.
2019.
07
.
009.
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
Meas
ur
in
g
t
he accu
ra
cy
o
f
LSTM a
nd BiL
STM m
odel
s i
n
t
he
app
li
cati
on of
a
rti
fi
ci
al…
(
Prasnu
r
za
ki
Ank
i
)
205
[17]
D.
Shao,
et
al.
,
“
Dom
ai
n
-
Speci
fi
c
Chin
ese
W
ord
Segm
ent
at
ion
B
ase
d
on
B
i
-
Dire
c
ti
onal
Long
-
Sho
rt
T
erm
Mem
or
y
Model,
”
IE
EE A
cc
ess
,
vol. 7, p.
12996,
2019
,
doi
:
10.
1109
/ACCE
SS
.
2019.
2892836.
[18]
I.
Attr
i
and
M.
Dutta,
“
Bi
-
L
in
gual
(
Engl
ish,
Punjabi
)
Sar
ca
s
ti
c
Senti
m
ent
Anal
y
sis
b
y
usi
ng
Cla
ss
ifica
t
io
n
Methods,
”
Int
ernati
onal
Journal
of
Innov
ative
T
ec
hnolog
y
and
Ex
ploring
Eng
in
ee
ring
(
IJI
TEE)
,
vol.
8,
p
.
1386,
2019
.
[19]
D.
Jurafsk
y
and
J.
H.
Marti
n
,
“
Speec
h
and
La
ng
uage
Proce
ss
ing
An
Introduc
ti
on
to
Natur
a
l
La
ng
uage
Proce
ss
ing
,
”
Prent
in
ce
Hal
l
,
2020
.
[20]
L.
Brown
and
S.
Stei
ner
ber
g
er
,
"P
ositi
ve
-
Defi
nit
e
Funct
ions,
Expone
ntial
Su
m
s
and
the
Gree
d
y
Algor
it
hm
:
a
Curious
Phenom
enon,
"
Journal
o
f
Comple
xi
t
y
,
vo
l.
61
,
p
.
9
,
2020
,
doi:
10
.
1016/j.j
c
o.
2020.
101485
.
[21]
C.
Li
m
,
“
An
Eva
luation
of
Mac
hin
e
Learni
ng
Approac
hes
to
Natur
al
L
angu
age
Proce
ss
ing
for
Le
gal
Te
xt
Cla
ss
ifi
c
at
ion
,
”
I
mperial
Col
le
ge
London
,
2019
.
[22]
K.
Kotec
h
a,
et
a
l.
,
”
Dat
a
Sci
enc
e
and
Intelli
g
ent
Applic
a
ti
ons:
Pr
oce
ed
ings
of
IC
D
SIA
2020,
”
Springer
Singapore
,
vol.
52
,
p
.
117
,
2
020
,
doi
:
10
.
100
7/978
-
981
-
15
-
44
74
-
3.
[23]
Y.
Long,
e
t
al.
,
”
Repeat
Paddi
ng:
Bal
an
ci
ng
words
and
sente
nce
le
ng
th
for
l
angua
ge
comprehe
nsion
in
visua
l
questi
on
answeri
ng,
”
In
formation
Scienc
es
,
vol
.
5
29,
pp
.
170
-
1
71
,
2020
,
doi:
10.
10
16/j
.
ins
.
2020.
04
.
034.
[24]
S.
Kal
y
a
,
M.
Kulkar
ni
,
and
K.
S.
Shivapra
k
asha
,
“
Advanc
es
in
Com
m
unic
at
io
n,
Signal
Proce
s
sing,
VLSI,
and
Embedde
d
S
y
st
ems
:
Sele
ct
Proce
ed
ings
of
V
SP
ICE
2019,
”
Springer
Nature
,
vol.
614
,
p.
307,
2019
,
doi:
10.
1007/
978
-
98
1
-
15
-
0626
-
0.
[25]
H.
Yin,
et
al.
,
”In
t
el
l
ige
nt
Data
Engi
n
ee
ring
and
Autom
ate
d
Learni
ng
–
I
DEAL
2019,
2
0th
Int
ern
ationa
l
Confer
ence,
Ma
nche
ster
,
UK
,
N
ovember
14
–
16,
2019,
Proc
ee
di
ngs,
Part
II
,
”
Sp
ringer
Nature
,
v
ol.
11872
,
p
.
18
2,
2019
,
doi
:
10
.
10
07/978
-
3
-
030
-
33
617
-
2.
[26]
L.
Barolli
,
P.
Hell
in
ckx,
T
.
E
nokido,
“
Advanc
es
on
Broad
-
Band
W
ire
le
ss
Com
puti
ng,
Com
m
unic
at
ion
an
d
Applic
a
ti
ons:
Proce
ed
ings
of
the
14th
Inte
rna
t
iona
l
Con
fer
ence
on
Broad
-
Band
W
ireless
Com
puti
ng,
Com
m
unic
at
ion
and
Applic
at
io
ns
(BW
CC
A
-
20
19),
”
Springer
Nature
,
vol.
97,
2019
,
doi:
10.
1007/978
-
3
-
030
-
33506
-
9.
[27]
K.
Fere
da
y
,
e
t
al.
,
”A
compari
son
of
roll
ing
a
ver
age
s
v
ersus
discre
t
e
ti
m
e
ep
ochs
fora
ss
essing
the
wors
t
-
c
ase
sce
nar
io
loc
om
o
tor
demands
of
profe
ss
iona
lsoccer
m
at
ch
-
p
lay
,
”
Journal
of
S
cien
c
e
and
Me
di
ci
ne
in
Sport
,
vol
.
2
,
pp.
765
-
766
,
20
20
,
doi
:
10
.
1016
/j
.
js
ams
.
2020.
01
.
002.
[28]
Y.
Nie,
P.
Ji
an
g,
H.
Zh
ang,
“
A
novel
h
y
brid
m
odel
base
d
o
n
combined
pre
proc
essing
m
et
h
od
and
adva
n
ced
opti
m
iz
ation
a
lg
orit
hm
for
power
loa
d
fore
ca
stin
g,
”
App
li
ed
So
ft
Co
mputing
Journal
,
vol
.
97,
pp
.
16
-
17,
2020,
doi
:
10.
1016/j.a
soc
.
2
020.
106809.
[29]
P.
Sili
tong
a,
e
t
a
l.
,
”Comparison of Dengue Predi
ct
iv
e
Models
De
vel
oped
Us
ing
A
rti
fi
ci
a
l
Neur
al
Network
and
Discriminant
An
aly
s
is wi
th
Sm
al
l
Dat
ase
t
,
”
Appl.
Sci.
,
vol
.
11
,
p
.
15,
2021
,
doi
:
10
.
3390/a
pp11030
943.
[30]
G.
Ardane
sw
ari,
A.
Bust
amam
,
D.
Sarwinda
,
“
I
m
ple
m
ent
at
ion
o
f
plaid
m
odel bi
c
luste
ring
m
et
hod
on
m
ic
ro
arr
a
y
o
f
ca
rc
inoma
and
a
denoma
tumor g
ene
expr
ession
d
at
a
,
”
Journal of Physic
s: Con
fe
r
enc
e
Seri
es
,
vo
l.
893,
p
.
4
,
2017
,
d
oi:
10
.
1088/17
42
-
6596/893/
1/0
12046.
BIOGR
AP
HI
ES OF
A
UTH
ORS
Pr
asn
ur
z
aki
A
nk
i
B.
Sc.
recei
v
ed
the
BS
c
(ho
nour)
degr
ee
in
m
at
hemati
cs
fr
om
Univer
sita
s
Indone
sia
in
2020.
He
is
cur
ren
tly
pursuing
a
m
aste
r’s
degr
ee
in
m
at
hemati
cs
fro
m
Univer
sita
s
Indone
sia.
His
r
ese
arc
h
intere
sts
are
in
th
e
ar
ea
s
of
computation
al
m
at
h
emati
cs,
dat
a
sc
ie
n
ce,
and
ar
ti
fif
icial i
n
te
lligen
ce.
As
soc.
Pr
of
.
Al
had
i
Bu
stama
m,
Ph.
D
.
re
ce
iv
ed
the
BS
c
(ho
nour)
degr
e
e
in
computat
ion
a
l
m
at
hemati
cs
i
n
1996,
the
m
aster’s
degr
ee
i
n
co
m
pute
r
scie
n
ce
from
Univer
sitas
Indone
sia
in
2002,
and
the
P
h.
D
degr
ee
in
b
i
oinformati
cs
fro
m
the
Univer
sity
of
Quee
nsl
and
,
Aus
tra
li
a
,
in
2011.
His
rese
arc
h
foc
us
es
on
high
-
per
form
anc
e
computing
appr
oac
h
es
to
computat
ion
al
m
at
hemati
cs,
c
om
puta
ti
ona
l
bi
olog
y
,
bio
infor
m
at
ic
s,
comput
er
scie
n
ce,
data
scie
n
ce,
and
art
if
ic
i
al
in
te
l
li
g
enc
e
.
Curre
nt
l
y
,
he
is
working
as
an
As
socia
t
e
Profess
or
and
the
Hea
d
of
Bioi
nform
at
i
cs
and
Adv
anc
e
d
Com
puti
ng
La
bora
tor
y
(B
ACL)
at
th
e
Depa
rtment
o
f
Mathe
m
at
i
cs.
He
is
al
so
se
rv
ing
as
the
cha
irman
of
Data
Scie
n
ce
Cent
re
(DS
C)
htt
ps://
dsc
.
ui
.
a
c.
i
d
at Unive
rsi
ta
s
Indone
sia.
Evaluation Warning : The document was created with Spire.PDF for Python.